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
65 #include <linux/slab.h>
66 #include <linux/module.h>
67 #include <linux/sysctl.h>
68 #include <linux/kernel.h>
71 #include <net/inet_common.h>
72 #include <linux/ipsec.h>
73 #include <asm/unaligned.h>
74 #include <net/netdma.h>
76 int sysctl_tcp_timestamps __read_mostly = 1;
77 int sysctl_tcp_window_scaling __read_mostly = 1;
78 int sysctl_tcp_sack __read_mostly = 1;
79 int sysctl_tcp_fack __read_mostly = 1;
80 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
81 EXPORT_SYMBOL(sysctl_tcp_reordering);
82 int sysctl_tcp_ecn __read_mostly = 2;
83 EXPORT_SYMBOL(sysctl_tcp_ecn);
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;
96 int sysctl_tcp_frto_response __read_mostly;
97 int sysctl_tcp_nometrics_save __read_mostly;
99 int sysctl_tcp_thin_dupack __read_mostly;
101 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
102 int sysctl_tcp_abc __read_mostly;
104 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
105 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
106 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
108 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
109 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
110 #define FLAG_ECE 0x40 /* ECE in this ACK */
111 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
112 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
113 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
114 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
115 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
116 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
117 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
119 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
120 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
121 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
122 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
123 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
125 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
126 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
128 /* Adapt the MSS value used to make delayed ack decision to the
131 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
133 struct inet_connection_sock *icsk = inet_csk(sk);
134 const unsigned int lss = icsk->icsk_ack.last_seg_size;
137 icsk->icsk_ack.last_seg_size = 0;
139 /* skb->len may jitter because of SACKs, even if peer
140 * sends good full-sized frames.
142 len = skb_shinfo(skb)->gso_size ? : skb->len;
143 if (len >= icsk->icsk_ack.rcv_mss) {
144 icsk->icsk_ack.rcv_mss = len;
146 /* Otherwise, we make more careful check taking into account,
147 * that SACKs block is variable.
149 * "len" is invariant segment length, including TCP header.
151 len += skb->data - skb_transport_header(skb);
152 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
153 /* If PSH is not set, packet should be
154 * full sized, provided peer TCP is not badly broken.
155 * This observation (if it is correct 8)) allows
156 * to handle super-low mtu links fairly.
158 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
159 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
160 /* Subtract also invariant (if peer is RFC compliant),
161 * tcp header plus fixed timestamp option length.
162 * Resulting "len" is MSS free of SACK jitter.
164 len -= tcp_sk(sk)->tcp_header_len;
165 icsk->icsk_ack.last_seg_size = len;
167 icsk->icsk_ack.rcv_mss = len;
171 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
172 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
173 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
177 static void tcp_incr_quickack(struct sock *sk)
179 struct inet_connection_sock *icsk = inet_csk(sk);
180 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
184 if (quickacks > icsk->icsk_ack.quick)
185 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
188 static void tcp_enter_quickack_mode(struct sock *sk)
190 struct inet_connection_sock *icsk = inet_csk(sk);
191 tcp_incr_quickack(sk);
192 icsk->icsk_ack.pingpong = 0;
193 icsk->icsk_ack.ato = TCP_ATO_MIN;
196 /* Send ACKs quickly, if "quick" count is not exhausted
197 * and the session is not interactive.
200 static inline int tcp_in_quickack_mode(const struct sock *sk)
202 const struct inet_connection_sock *icsk = inet_csk(sk);
203 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
206 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
208 if (tp->ecn_flags & TCP_ECN_OK)
209 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
212 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
214 if (tcp_hdr(skb)->cwr)
215 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
218 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
220 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
223 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
225 if (tp->ecn_flags & TCP_ECN_OK) {
226 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
227 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
228 /* Funny extension: if ECT is not set on a segment,
229 * it is surely retransmit. It is not in ECN RFC,
230 * but Linux follows this rule. */
231 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
232 tcp_enter_quickack_mode((struct sock *)tp);
236 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
238 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
239 tp->ecn_flags &= ~TCP_ECN_OK;
242 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
244 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
245 tp->ecn_flags &= ~TCP_ECN_OK;
248 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
250 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
255 /* Buffer size and advertised window tuning.
257 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
260 static void tcp_fixup_sndbuf(struct sock *sk)
262 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
263 sizeof(struct sk_buff);
265 if (sk->sk_sndbuf < 3 * sndmem) {
266 sk->sk_sndbuf = 3 * sndmem;
267 if (sk->sk_sndbuf > sysctl_tcp_wmem[2])
268 sk->sk_sndbuf = sysctl_tcp_wmem[2];
272 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
274 * All tcp_full_space() is split to two parts: "network" buffer, allocated
275 * forward and advertised in receiver window (tp->rcv_wnd) and
276 * "application buffer", required to isolate scheduling/application
277 * latencies from network.
278 * window_clamp is maximal advertised window. It can be less than
279 * tcp_full_space(), in this case tcp_full_space() - window_clamp
280 * is reserved for "application" buffer. The less window_clamp is
281 * the smoother our behaviour from viewpoint of network, but the lower
282 * throughput and the higher sensitivity of the connection to losses. 8)
284 * rcv_ssthresh is more strict window_clamp used at "slow start"
285 * phase to predict further behaviour of this connection.
286 * It is used for two goals:
287 * - to enforce header prediction at sender, even when application
288 * requires some significant "application buffer". It is check #1.
289 * - to prevent pruning of receive queue because of misprediction
290 * of receiver window. Check #2.
292 * The scheme does not work when sender sends good segments opening
293 * window and then starts to feed us spaghetti. But it should work
294 * in common situations. Otherwise, we have to rely on queue collapsing.
297 /* Slow part of check#2. */
298 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
300 struct tcp_sock *tp = tcp_sk(sk);
302 int truesize = tcp_win_from_space(skb->truesize) >> 1;
303 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
305 while (tp->rcv_ssthresh <= window) {
306 if (truesize <= skb->len)
307 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
315 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
317 struct tcp_sock *tp = tcp_sk(sk);
320 if (tp->rcv_ssthresh < tp->window_clamp &&
321 (int)tp->rcv_ssthresh < tcp_space(sk) &&
322 !tcp_memory_pressure) {
325 /* Check #2. Increase window, if skb with such overhead
326 * will fit to rcvbuf in future.
328 if (tcp_win_from_space(skb->truesize) <= skb->len)
329 incr = 2 * tp->advmss;
331 incr = __tcp_grow_window(sk, skb);
334 incr = max_t(int, incr, 2 * skb->len);
335 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
337 inet_csk(sk)->icsk_ack.quick |= 1;
342 /* 3. Tuning rcvbuf, when connection enters established state. */
344 static void tcp_fixup_rcvbuf(struct sock *sk)
346 struct tcp_sock *tp = tcp_sk(sk);
347 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
349 /* Try to select rcvbuf so that 4 mss-sized segments
350 * will fit to window and corresponding skbs will fit to our rcvbuf.
351 * (was 3; 4 is minimum to allow fast retransmit to work.)
353 while (tcp_win_from_space(rcvmem) < tp->advmss)
355 if (sk->sk_rcvbuf < 4 * rcvmem)
356 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
359 /* 4. Try to fixup all. It is made immediately after connection enters
362 static void tcp_init_buffer_space(struct sock *sk)
364 struct tcp_sock *tp = tcp_sk(sk);
367 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
368 tcp_fixup_rcvbuf(sk);
369 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
370 tcp_fixup_sndbuf(sk);
372 tp->rcvq_space.space = tp->rcv_wnd;
374 maxwin = tcp_full_space(sk);
376 if (tp->window_clamp >= maxwin) {
377 tp->window_clamp = maxwin;
379 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
380 tp->window_clamp = max(maxwin -
381 (maxwin >> sysctl_tcp_app_win),
385 /* Force reservation of one segment. */
386 if (sysctl_tcp_app_win &&
387 tp->window_clamp > 2 * tp->advmss &&
388 tp->window_clamp + tp->advmss > maxwin)
389 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
391 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
392 tp->snd_cwnd_stamp = tcp_time_stamp;
395 /* 5. Recalculate window clamp after socket hit its memory bounds. */
396 static void tcp_clamp_window(struct sock *sk)
398 struct tcp_sock *tp = tcp_sk(sk);
399 struct inet_connection_sock *icsk = inet_csk(sk);
401 icsk->icsk_ack.quick = 0;
403 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
404 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
405 !tcp_memory_pressure &&
406 atomic_long_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
407 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
410 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
411 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
414 /* Initialize RCV_MSS value.
415 * RCV_MSS is an our guess about MSS used by the peer.
416 * We haven't any direct information about the MSS.
417 * It's better to underestimate the RCV_MSS rather than overestimate.
418 * Overestimations make us ACKing less frequently than needed.
419 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
421 void tcp_initialize_rcv_mss(struct sock *sk)
423 struct tcp_sock *tp = tcp_sk(sk);
424 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
426 hint = min(hint, tp->rcv_wnd / 2);
427 hint = min(hint, TCP_MSS_DEFAULT);
428 hint = max(hint, TCP_MIN_MSS);
430 inet_csk(sk)->icsk_ack.rcv_mss = hint;
432 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
434 /* Receiver "autotuning" code.
436 * The algorithm for RTT estimation w/o timestamps is based on
437 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
438 * <http://public.lanl.gov/radiant/pubs.html#DRS>
440 * More detail on this code can be found at
441 * <http://staff.psc.edu/jheffner/>,
442 * though this reference is out of date. A new paper
445 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
447 u32 new_sample = tp->rcv_rtt_est.rtt;
453 if (new_sample != 0) {
454 /* If we sample in larger samples in the non-timestamp
455 * case, we could grossly overestimate the RTT especially
456 * with chatty applications or bulk transfer apps which
457 * are stalled on filesystem I/O.
459 * Also, since we are only going for a minimum in the
460 * non-timestamp case, we do not smooth things out
461 * else with timestamps disabled convergence takes too
465 m -= (new_sample >> 3);
473 /* No previous measure. */
477 if (tp->rcv_rtt_est.rtt != new_sample)
478 tp->rcv_rtt_est.rtt = new_sample;
481 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
483 if (tp->rcv_rtt_est.time == 0)
485 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
487 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
490 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
491 tp->rcv_rtt_est.time = tcp_time_stamp;
494 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
495 const struct sk_buff *skb)
497 struct tcp_sock *tp = tcp_sk(sk);
498 if (tp->rx_opt.rcv_tsecr &&
499 (TCP_SKB_CB(skb)->end_seq -
500 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
501 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
505 * This function should be called every time data is copied to user space.
506 * It calculates the appropriate TCP receive buffer space.
508 void tcp_rcv_space_adjust(struct sock *sk)
510 struct tcp_sock *tp = tcp_sk(sk);
514 if (tp->rcvq_space.time == 0)
517 time = tcp_time_stamp - tp->rcvq_space.time;
518 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
521 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
523 space = max(tp->rcvq_space.space, space);
525 if (tp->rcvq_space.space != space) {
528 tp->rcvq_space.space = space;
530 if (sysctl_tcp_moderate_rcvbuf &&
531 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
532 int new_clamp = space;
534 /* Receive space grows, normalize in order to
535 * take into account packet headers and sk_buff
536 * structure overhead.
541 rcvmem = (tp->advmss + MAX_TCP_HEADER +
542 16 + sizeof(struct sk_buff));
543 while (tcp_win_from_space(rcvmem) < tp->advmss)
546 space = min(space, sysctl_tcp_rmem[2]);
547 if (space > sk->sk_rcvbuf) {
548 sk->sk_rcvbuf = space;
550 /* Make the window clamp follow along. */
551 tp->window_clamp = new_clamp;
557 tp->rcvq_space.seq = tp->copied_seq;
558 tp->rcvq_space.time = tcp_time_stamp;
561 /* There is something which you must keep in mind when you analyze the
562 * behavior of the tp->ato delayed ack timeout interval. When a
563 * connection starts up, we want to ack as quickly as possible. The
564 * problem is that "good" TCP's do slow start at the beginning of data
565 * transmission. The means that until we send the first few ACK's the
566 * sender will sit on his end and only queue most of his data, because
567 * he can only send snd_cwnd unacked packets at any given time. For
568 * each ACK we send, he increments snd_cwnd and transmits more of his
571 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
573 struct tcp_sock *tp = tcp_sk(sk);
574 struct inet_connection_sock *icsk = inet_csk(sk);
577 inet_csk_schedule_ack(sk);
579 tcp_measure_rcv_mss(sk, skb);
581 tcp_rcv_rtt_measure(tp);
583 now = tcp_time_stamp;
585 if (!icsk->icsk_ack.ato) {
586 /* The _first_ data packet received, initialize
587 * delayed ACK engine.
589 tcp_incr_quickack(sk);
590 icsk->icsk_ack.ato = TCP_ATO_MIN;
592 int m = now - icsk->icsk_ack.lrcvtime;
594 if (m <= TCP_ATO_MIN / 2) {
595 /* The fastest case is the first. */
596 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
597 } else if (m < icsk->icsk_ack.ato) {
598 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
599 if (icsk->icsk_ack.ato > icsk->icsk_rto)
600 icsk->icsk_ack.ato = icsk->icsk_rto;
601 } else if (m > icsk->icsk_rto) {
602 /* Too long gap. Apparently sender failed to
603 * restart window, so that we send ACKs quickly.
605 tcp_incr_quickack(sk);
609 icsk->icsk_ack.lrcvtime = now;
611 TCP_ECN_check_ce(tp, skb);
614 tcp_grow_window(sk, skb);
617 /* Called to compute a smoothed rtt estimate. The data fed to this
618 * routine either comes from timestamps, or from segments that were
619 * known _not_ to have been retransmitted [see Karn/Partridge
620 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
621 * piece by Van Jacobson.
622 * NOTE: the next three routines used to be one big routine.
623 * To save cycles in the RFC 1323 implementation it was better to break
624 * it up into three procedures. -- erics
626 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
628 struct tcp_sock *tp = tcp_sk(sk);
629 long m = mrtt; /* RTT */
631 /* The following amusing code comes from Jacobson's
632 * article in SIGCOMM '88. Note that rtt and mdev
633 * are scaled versions of rtt and mean deviation.
634 * This is designed to be as fast as possible
635 * m stands for "measurement".
637 * On a 1990 paper the rto value is changed to:
638 * RTO = rtt + 4 * mdev
640 * Funny. This algorithm seems to be very broken.
641 * These formulae increase RTO, when it should be decreased, increase
642 * too slowly, when it should be increased quickly, decrease too quickly
643 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
644 * does not matter how to _calculate_ it. Seems, it was trap
645 * that VJ failed to avoid. 8)
650 m -= (tp->srtt >> 3); /* m is now error in rtt est */
651 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
653 m = -m; /* m is now abs(error) */
654 m -= (tp->mdev >> 2); /* similar update on mdev */
655 /* This is similar to one of Eifel findings.
656 * Eifel blocks mdev updates when rtt decreases.
657 * This solution is a bit different: we use finer gain
658 * for mdev in this case (alpha*beta).
659 * Like Eifel it also prevents growth of rto,
660 * but also it limits too fast rto decreases,
661 * happening in pure Eifel.
666 m -= (tp->mdev >> 2); /* similar update on mdev */
668 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
669 if (tp->mdev > tp->mdev_max) {
670 tp->mdev_max = tp->mdev;
671 if (tp->mdev_max > tp->rttvar)
672 tp->rttvar = tp->mdev_max;
674 if (after(tp->snd_una, tp->rtt_seq)) {
675 if (tp->mdev_max < tp->rttvar)
676 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
677 tp->rtt_seq = tp->snd_nxt;
678 tp->mdev_max = tcp_rto_min(sk);
681 /* no previous measure. */
682 tp->srtt = m << 3; /* take the measured time to be rtt */
683 tp->mdev = m << 1; /* make sure rto = 3*rtt */
684 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
685 tp->rtt_seq = tp->snd_nxt;
689 /* Calculate rto without backoff. This is the second half of Van Jacobson's
690 * routine referred to above.
692 static inline void tcp_set_rto(struct sock *sk)
694 const struct tcp_sock *tp = tcp_sk(sk);
695 /* Old crap is replaced with new one. 8)
698 * 1. If rtt variance happened to be less 50msec, it is hallucination.
699 * It cannot be less due to utterly erratic ACK generation made
700 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
701 * to do with delayed acks, because at cwnd>2 true delack timeout
702 * is invisible. Actually, Linux-2.4 also generates erratic
703 * ACKs in some circumstances.
705 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
707 /* 2. Fixups made earlier cannot be right.
708 * If we do not estimate RTO correctly without them,
709 * all the algo is pure shit and should be replaced
710 * with correct one. It is exactly, which we pretend to do.
713 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
714 * guarantees that rto is higher.
719 /* Save metrics learned by this TCP session.
720 This function is called only, when TCP finishes successfully
721 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
723 void tcp_update_metrics(struct sock *sk)
725 struct tcp_sock *tp = tcp_sk(sk);
726 struct dst_entry *dst = __sk_dst_get(sk);
728 if (sysctl_tcp_nometrics_save)
733 if (dst && (dst->flags & DST_HOST)) {
734 const struct inet_connection_sock *icsk = inet_csk(sk);
738 if (icsk->icsk_backoff || !tp->srtt) {
739 /* This session failed to estimate rtt. Why?
740 * Probably, no packets returned in time.
743 if (!(dst_metric_locked(dst, RTAX_RTT)))
744 dst_metric_set(dst, RTAX_RTT, 0);
748 rtt = dst_metric_rtt(dst, RTAX_RTT);
751 /* If newly calculated rtt larger than stored one,
752 * store new one. Otherwise, use EWMA. Remember,
753 * rtt overestimation is always better than underestimation.
755 if (!(dst_metric_locked(dst, RTAX_RTT))) {
757 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
759 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
762 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
767 /* Scale deviation to rttvar fixed point */
772 var = dst_metric_rtt(dst, RTAX_RTTVAR);
776 var -= (var - m) >> 2;
778 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
781 if (tcp_in_initial_slowstart(tp)) {
782 /* Slow start still did not finish. */
783 if (dst_metric(dst, RTAX_SSTHRESH) &&
784 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
785 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
786 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
787 if (!dst_metric_locked(dst, RTAX_CWND) &&
788 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
789 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
790 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
791 icsk->icsk_ca_state == TCP_CA_Open) {
792 /* Cong. avoidance phase, cwnd is reliable. */
793 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
794 dst_metric_set(dst, RTAX_SSTHRESH,
795 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
796 if (!dst_metric_locked(dst, RTAX_CWND))
797 dst_metric_set(dst, RTAX_CWND,
798 (dst_metric(dst, RTAX_CWND) +
801 /* Else slow start did not finish, cwnd is non-sense,
802 ssthresh may be also invalid.
804 if (!dst_metric_locked(dst, RTAX_CWND))
805 dst_metric_set(dst, RTAX_CWND,
806 (dst_metric(dst, RTAX_CWND) +
807 tp->snd_ssthresh) >> 1);
808 if (dst_metric(dst, RTAX_SSTHRESH) &&
809 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
810 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
811 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
814 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
815 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
816 tp->reordering != sysctl_tcp_reordering)
817 dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
822 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
824 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
827 cwnd = TCP_INIT_CWND;
828 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
831 /* Set slow start threshold and cwnd not falling to slow start */
832 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
834 struct tcp_sock *tp = tcp_sk(sk);
835 const struct inet_connection_sock *icsk = inet_csk(sk);
837 tp->prior_ssthresh = 0;
839 if (icsk->icsk_ca_state < TCP_CA_CWR) {
842 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
843 tp->snd_cwnd = min(tp->snd_cwnd,
844 tcp_packets_in_flight(tp) + 1U);
845 tp->snd_cwnd_cnt = 0;
846 tp->high_seq = tp->snd_nxt;
847 tp->snd_cwnd_stamp = tcp_time_stamp;
848 TCP_ECN_queue_cwr(tp);
850 tcp_set_ca_state(sk, TCP_CA_CWR);
855 * Packet counting of FACK is based on in-order assumptions, therefore TCP
856 * disables it when reordering is detected
858 static void tcp_disable_fack(struct tcp_sock *tp)
860 /* RFC3517 uses different metric in lost marker => reset on change */
862 tp->lost_skb_hint = NULL;
863 tp->rx_opt.sack_ok &= ~2;
866 /* Take a notice that peer is sending D-SACKs */
867 static void tcp_dsack_seen(struct tcp_sock *tp)
869 tp->rx_opt.sack_ok |= 4;
872 /* Initialize metrics on socket. */
874 static void tcp_init_metrics(struct sock *sk)
876 struct tcp_sock *tp = tcp_sk(sk);
877 struct dst_entry *dst = __sk_dst_get(sk);
884 if (dst_metric_locked(dst, RTAX_CWND))
885 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
886 if (dst_metric(dst, RTAX_SSTHRESH)) {
887 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
888 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
889 tp->snd_ssthresh = tp->snd_cwnd_clamp;
891 if (dst_metric(dst, RTAX_REORDERING) &&
892 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
893 tcp_disable_fack(tp);
894 tp->reordering = dst_metric(dst, RTAX_REORDERING);
897 if (dst_metric(dst, RTAX_RTT) == 0)
900 if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
903 /* Initial rtt is determined from SYN,SYN-ACK.
904 * The segment is small and rtt may appear much
905 * less than real one. Use per-dst memory
906 * to make it more realistic.
908 * A bit of theory. RTT is time passed after "normal" sized packet
909 * is sent until it is ACKed. In normal circumstances sending small
910 * packets force peer to delay ACKs and calculation is correct too.
911 * The algorithm is adaptive and, provided we follow specs, it
912 * NEVER underestimate RTT. BUT! If peer tries to make some clever
913 * tricks sort of "quick acks" for time long enough to decrease RTT
914 * to low value, and then abruptly stops to do it and starts to delay
915 * ACKs, wait for troubles.
917 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
918 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
919 tp->rtt_seq = tp->snd_nxt;
921 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
922 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
923 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
926 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) {
928 /* Play conservative. If timestamps are not
929 * supported, TCP will fail to recalculate correct
930 * rtt, if initial rto is too small. FORGET ALL AND RESET!
932 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
934 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
935 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
938 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
939 tp->snd_cwnd_stamp = tcp_time_stamp;
942 static void tcp_update_reordering(struct sock *sk, const int metric,
945 struct tcp_sock *tp = tcp_sk(sk);
946 if (metric > tp->reordering) {
949 tp->reordering = min(TCP_MAX_REORDERING, metric);
951 /* This exciting event is worth to be remembered. 8) */
953 mib_idx = LINUX_MIB_TCPTSREORDER;
954 else if (tcp_is_reno(tp))
955 mib_idx = LINUX_MIB_TCPRENOREORDER;
956 else if (tcp_is_fack(tp))
957 mib_idx = LINUX_MIB_TCPFACKREORDER;
959 mib_idx = LINUX_MIB_TCPSACKREORDER;
961 NET_INC_STATS_BH(sock_net(sk), mib_idx);
962 #if FASTRETRANS_DEBUG > 1
963 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
964 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
968 tp->undo_marker ? tp->undo_retrans : 0);
970 tcp_disable_fack(tp);
974 /* This must be called before lost_out is incremented */
975 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
977 if ((tp->retransmit_skb_hint == NULL) ||
978 before(TCP_SKB_CB(skb)->seq,
979 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
980 tp->retransmit_skb_hint = skb;
983 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
984 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
987 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
989 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
990 tcp_verify_retransmit_hint(tp, skb);
992 tp->lost_out += tcp_skb_pcount(skb);
993 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
997 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1000 tcp_verify_retransmit_hint(tp, skb);
1002 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1003 tp->lost_out += tcp_skb_pcount(skb);
1004 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1008 /* This procedure tags the retransmission queue when SACKs arrive.
1010 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1011 * Packets in queue with these bits set are counted in variables
1012 * sacked_out, retrans_out and lost_out, correspondingly.
1014 * Valid combinations are:
1015 * Tag InFlight Description
1016 * 0 1 - orig segment is in flight.
1017 * S 0 - nothing flies, orig reached receiver.
1018 * L 0 - nothing flies, orig lost by net.
1019 * R 2 - both orig and retransmit are in flight.
1020 * L|R 1 - orig is lost, retransmit is in flight.
1021 * S|R 1 - orig reached receiver, retrans is still in flight.
1022 * (L|S|R is logically valid, it could occur when L|R is sacked,
1023 * but it is equivalent to plain S and code short-curcuits it to S.
1024 * L|S is logically invalid, it would mean -1 packet in flight 8))
1026 * These 6 states form finite state machine, controlled by the following events:
1027 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1028 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1029 * 3. Loss detection event of one of three flavors:
1030 * A. Scoreboard estimator decided the packet is lost.
1031 * A'. Reno "three dupacks" marks head of queue lost.
1032 * A''. Its FACK modfication, head until snd.fack is lost.
1033 * B. SACK arrives sacking data transmitted after never retransmitted
1034 * hole was sent out.
1035 * C. SACK arrives sacking SND.NXT at the moment, when the
1036 * segment was retransmitted.
1037 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1039 * It is pleasant to note, that state diagram turns out to be commutative,
1040 * so that we are allowed not to be bothered by order of our actions,
1041 * when multiple events arrive simultaneously. (see the function below).
1043 * Reordering detection.
1044 * --------------------
1045 * Reordering metric is maximal distance, which a packet can be displaced
1046 * in packet stream. With SACKs we can estimate it:
1048 * 1. SACK fills old hole and the corresponding segment was not
1049 * ever retransmitted -> reordering. Alas, we cannot use it
1050 * when segment was retransmitted.
1051 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1052 * for retransmitted and already SACKed segment -> reordering..
1053 * Both of these heuristics are not used in Loss state, when we cannot
1054 * account for retransmits accurately.
1056 * SACK block validation.
1057 * ----------------------
1059 * SACK block range validation checks that the received SACK block fits to
1060 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1061 * Note that SND.UNA is not included to the range though being valid because
1062 * it means that the receiver is rather inconsistent with itself reporting
1063 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1064 * perfectly valid, however, in light of RFC2018 which explicitly states
1065 * that "SACK block MUST reflect the newest segment. Even if the newest
1066 * segment is going to be discarded ...", not that it looks very clever
1067 * in case of head skb. Due to potentional receiver driven attacks, we
1068 * choose to avoid immediate execution of a walk in write queue due to
1069 * reneging and defer head skb's loss recovery to standard loss recovery
1070 * procedure that will eventually trigger (nothing forbids us doing this).
1072 * Implements also blockage to start_seq wrap-around. Problem lies in the
1073 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1074 * there's no guarantee that it will be before snd_nxt (n). The problem
1075 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1078 * <- outs wnd -> <- wrapzone ->
1079 * u e n u_w e_w s n_w
1081 * |<------------+------+----- TCP seqno space --------------+---------->|
1082 * ...-- <2^31 ->| |<--------...
1083 * ...---- >2^31 ------>| |<--------...
1085 * Current code wouldn't be vulnerable but it's better still to discard such
1086 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1087 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1088 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1089 * equal to the ideal case (infinite seqno space without wrap caused issues).
1091 * With D-SACK the lower bound is extended to cover sequence space below
1092 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1093 * again, D-SACK block must not to go across snd_una (for the same reason as
1094 * for the normal SACK blocks, explained above). But there all simplicity
1095 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1096 * fully below undo_marker they do not affect behavior in anyway and can
1097 * therefore be safely ignored. In rare cases (which are more or less
1098 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1099 * fragmentation and packet reordering past skb's retransmission. To consider
1100 * them correctly, the acceptable range must be extended even more though
1101 * the exact amount is rather hard to quantify. However, tp->max_window can
1102 * be used as an exaggerated estimate.
1104 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1105 u32 start_seq, u32 end_seq)
1107 /* Too far in future, or reversed (interpretation is ambiguous) */
1108 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1111 /* Nasty start_seq wrap-around check (see comments above) */
1112 if (!before(start_seq, tp->snd_nxt))
1115 /* In outstanding window? ...This is valid exit for D-SACKs too.
1116 * start_seq == snd_una is non-sensical (see comments above)
1118 if (after(start_seq, tp->snd_una))
1121 if (!is_dsack || !tp->undo_marker)
1124 /* ...Then it's D-SACK, and must reside below snd_una completely */
1125 if (after(end_seq, tp->snd_una))
1128 if (!before(start_seq, tp->undo_marker))
1132 if (!after(end_seq, tp->undo_marker))
1135 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1136 * start_seq < undo_marker and end_seq >= undo_marker.
1138 return !before(start_seq, end_seq - tp->max_window);
1141 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1142 * Event "C". Later note: FACK people cheated me again 8), we have to account
1143 * for reordering! Ugly, but should help.
1145 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1146 * less than what is now known to be received by the other end (derived from
1147 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1148 * retransmitted skbs to avoid some costly processing per ACKs.
1150 static void tcp_mark_lost_retrans(struct sock *sk)
1152 const struct inet_connection_sock *icsk = inet_csk(sk);
1153 struct tcp_sock *tp = tcp_sk(sk);
1154 struct sk_buff *skb;
1156 u32 new_low_seq = tp->snd_nxt;
1157 u32 received_upto = tcp_highest_sack_seq(tp);
1159 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1160 !after(received_upto, tp->lost_retrans_low) ||
1161 icsk->icsk_ca_state != TCP_CA_Recovery)
1164 tcp_for_write_queue(skb, sk) {
1165 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1167 if (skb == tcp_send_head(sk))
1169 if (cnt == tp->retrans_out)
1171 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1174 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1177 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1178 * constraint here (see above) but figuring out that at
1179 * least tp->reordering SACK blocks reside between ack_seq
1180 * and received_upto is not easy task to do cheaply with
1181 * the available datastructures.
1183 * Whether FACK should check here for tp->reordering segs
1184 * in-between one could argue for either way (it would be
1185 * rather simple to implement as we could count fack_count
1186 * during the walk and do tp->fackets_out - fack_count).
1188 if (after(received_upto, ack_seq)) {
1189 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1190 tp->retrans_out -= tcp_skb_pcount(skb);
1192 tcp_skb_mark_lost_uncond_verify(tp, skb);
1193 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1195 if (before(ack_seq, new_low_seq))
1196 new_low_seq = ack_seq;
1197 cnt += tcp_skb_pcount(skb);
1201 if (tp->retrans_out)
1202 tp->lost_retrans_low = new_low_seq;
1205 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1206 struct tcp_sack_block_wire *sp, int num_sacks,
1209 struct tcp_sock *tp = tcp_sk(sk);
1210 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1211 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1214 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1217 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1218 } else if (num_sacks > 1) {
1219 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1220 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1222 if (!after(end_seq_0, end_seq_1) &&
1223 !before(start_seq_0, start_seq_1)) {
1226 NET_INC_STATS_BH(sock_net(sk),
1227 LINUX_MIB_TCPDSACKOFORECV);
1231 /* D-SACK for already forgotten data... Do dumb counting. */
1232 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1233 !after(end_seq_0, prior_snd_una) &&
1234 after(end_seq_0, tp->undo_marker))
1240 struct tcp_sacktag_state {
1246 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1247 * the incoming SACK may not exactly match but we can find smaller MSS
1248 * aligned portion of it that matches. Therefore we might need to fragment
1249 * which may fail and creates some hassle (caller must handle error case
1252 * FIXME: this could be merged to shift decision code
1254 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1255 u32 start_seq, u32 end_seq)
1258 unsigned int pkt_len;
1261 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1262 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1264 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1265 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1266 mss = tcp_skb_mss(skb);
1267 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1270 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1274 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1279 /* Round if necessary so that SACKs cover only full MSSes
1280 * and/or the remaining small portion (if present)
1282 if (pkt_len > mss) {
1283 unsigned int new_len = (pkt_len / mss) * mss;
1284 if (!in_sack && new_len < pkt_len) {
1286 if (new_len > skb->len)
1291 err = tcp_fragment(sk, skb, pkt_len, mss);
1299 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1300 static u8 tcp_sacktag_one(struct sock *sk,
1301 struct tcp_sacktag_state *state, u8 sacked,
1302 u32 start_seq, u32 end_seq,
1303 int dup_sack, int pcount)
1305 struct tcp_sock *tp = tcp_sk(sk);
1306 int fack_count = state->fack_count;
1308 /* Account D-SACK for retransmitted packet. */
1309 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1310 if (tp->undo_marker && tp->undo_retrans &&
1311 after(end_seq, tp->undo_marker))
1313 if (sacked & TCPCB_SACKED_ACKED)
1314 state->reord = min(fack_count, state->reord);
1317 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1318 if (!after(end_seq, tp->snd_una))
1321 if (!(sacked & TCPCB_SACKED_ACKED)) {
1322 if (sacked & TCPCB_SACKED_RETRANS) {
1323 /* If the segment is not tagged as lost,
1324 * we do not clear RETRANS, believing
1325 * that retransmission is still in flight.
1327 if (sacked & TCPCB_LOST) {
1328 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1329 tp->lost_out -= pcount;
1330 tp->retrans_out -= pcount;
1333 if (!(sacked & TCPCB_RETRANS)) {
1334 /* New sack for not retransmitted frame,
1335 * which was in hole. It is reordering.
1337 if (before(start_seq,
1338 tcp_highest_sack_seq(tp)))
1339 state->reord = min(fack_count,
1342 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1343 if (!after(end_seq, tp->frto_highmark))
1344 state->flag |= FLAG_ONLY_ORIG_SACKED;
1347 if (sacked & TCPCB_LOST) {
1348 sacked &= ~TCPCB_LOST;
1349 tp->lost_out -= pcount;
1353 sacked |= TCPCB_SACKED_ACKED;
1354 state->flag |= FLAG_DATA_SACKED;
1355 tp->sacked_out += pcount;
1357 fack_count += pcount;
1359 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1360 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1361 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1362 tp->lost_cnt_hint += pcount;
1364 if (fack_count > tp->fackets_out)
1365 tp->fackets_out = fack_count;
1368 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1369 * frames and clear it. undo_retrans is decreased above, L|R frames
1370 * are accounted above as well.
1372 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1373 sacked &= ~TCPCB_SACKED_RETRANS;
1374 tp->retrans_out -= pcount;
1380 /* Shift newly-SACKed bytes from this skb to the immediately previous
1381 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1383 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1384 struct tcp_sacktag_state *state,
1385 unsigned int pcount, int shifted, int mss,
1388 struct tcp_sock *tp = tcp_sk(sk);
1389 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1390 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1391 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1395 /* Adjust counters and hints for the newly sacked sequence
1396 * range but discard the return value since prev is already
1397 * marked. We must tag the range first because the seq
1398 * advancement below implicitly advances
1399 * tcp_highest_sack_seq() when skb is highest_sack.
1401 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1402 start_seq, end_seq, dup_sack, pcount);
1404 if (skb == tp->lost_skb_hint)
1405 tp->lost_cnt_hint += pcount;
1407 TCP_SKB_CB(prev)->end_seq += shifted;
1408 TCP_SKB_CB(skb)->seq += shifted;
1410 skb_shinfo(prev)->gso_segs += pcount;
1411 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1412 skb_shinfo(skb)->gso_segs -= pcount;
1414 /* When we're adding to gso_segs == 1, gso_size will be zero,
1415 * in theory this shouldn't be necessary but as long as DSACK
1416 * code can come after this skb later on it's better to keep
1417 * setting gso_size to something.
1419 if (!skb_shinfo(prev)->gso_size) {
1420 skb_shinfo(prev)->gso_size = mss;
1421 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1424 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1425 if (skb_shinfo(skb)->gso_segs <= 1) {
1426 skb_shinfo(skb)->gso_size = 0;
1427 skb_shinfo(skb)->gso_type = 0;
1430 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1431 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1434 BUG_ON(!tcp_skb_pcount(skb));
1435 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1439 /* Whole SKB was eaten :-) */
1441 if (skb == tp->retransmit_skb_hint)
1442 tp->retransmit_skb_hint = prev;
1443 if (skb == tp->scoreboard_skb_hint)
1444 tp->scoreboard_skb_hint = prev;
1445 if (skb == tp->lost_skb_hint) {
1446 tp->lost_skb_hint = prev;
1447 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1450 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1451 if (skb == tcp_highest_sack(sk))
1452 tcp_advance_highest_sack(sk, skb);
1454 tcp_unlink_write_queue(skb, sk);
1455 sk_wmem_free_skb(sk, skb);
1457 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1462 /* I wish gso_size would have a bit more sane initialization than
1463 * something-or-zero which complicates things
1465 static int tcp_skb_seglen(struct sk_buff *skb)
1467 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1470 /* Shifting pages past head area doesn't work */
1471 static int skb_can_shift(struct sk_buff *skb)
1473 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1476 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1479 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1480 struct tcp_sacktag_state *state,
1481 u32 start_seq, u32 end_seq,
1484 struct tcp_sock *tp = tcp_sk(sk);
1485 struct sk_buff *prev;
1491 if (!sk_can_gso(sk))
1494 /* Normally R but no L won't result in plain S */
1496 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1498 if (!skb_can_shift(skb))
1500 /* This frame is about to be dropped (was ACKed). */
1501 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1504 /* Can only happen with delayed DSACK + discard craziness */
1505 if (unlikely(skb == tcp_write_queue_head(sk)))
1507 prev = tcp_write_queue_prev(sk, skb);
1509 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1512 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1513 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1517 pcount = tcp_skb_pcount(skb);
1518 mss = tcp_skb_seglen(skb);
1520 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1521 * drop this restriction as unnecessary
1523 if (mss != tcp_skb_seglen(prev))
1526 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1528 /* CHECKME: This is non-MSS split case only?, this will
1529 * cause skipped skbs due to advancing loop btw, original
1530 * has that feature too
1532 if (tcp_skb_pcount(skb) <= 1)
1535 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1537 /* TODO: head merge to next could be attempted here
1538 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1539 * though it might not be worth of the additional hassle
1541 * ...we can probably just fallback to what was done
1542 * previously. We could try merging non-SACKed ones
1543 * as well but it probably isn't going to buy off
1544 * because later SACKs might again split them, and
1545 * it would make skb timestamp tracking considerably
1551 len = end_seq - TCP_SKB_CB(skb)->seq;
1553 BUG_ON(len > skb->len);
1555 /* MSS boundaries should be honoured or else pcount will
1556 * severely break even though it makes things bit trickier.
1557 * Optimize common case to avoid most of the divides
1559 mss = tcp_skb_mss(skb);
1561 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1562 * drop this restriction as unnecessary
1564 if (mss != tcp_skb_seglen(prev))
1569 } else if (len < mss) {
1577 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1578 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1581 if (!skb_shift(prev, skb, len))
1583 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1586 /* Hole filled allows collapsing with the next as well, this is very
1587 * useful when hole on every nth skb pattern happens
1589 if (prev == tcp_write_queue_tail(sk))
1591 skb = tcp_write_queue_next(sk, prev);
1593 if (!skb_can_shift(skb) ||
1594 (skb == tcp_send_head(sk)) ||
1595 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1596 (mss != tcp_skb_seglen(skb)))
1600 if (skb_shift(prev, skb, len)) {
1601 pcount += tcp_skb_pcount(skb);
1602 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1606 state->fack_count += pcount;
1613 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1617 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1618 struct tcp_sack_block *next_dup,
1619 struct tcp_sacktag_state *state,
1620 u32 start_seq, u32 end_seq,
1623 struct tcp_sock *tp = tcp_sk(sk);
1624 struct sk_buff *tmp;
1626 tcp_for_write_queue_from(skb, sk) {
1628 int dup_sack = dup_sack_in;
1630 if (skb == tcp_send_head(sk))
1633 /* queue is in-order => we can short-circuit the walk early */
1634 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1637 if ((next_dup != NULL) &&
1638 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1639 in_sack = tcp_match_skb_to_sack(sk, skb,
1640 next_dup->start_seq,
1646 /* skb reference here is a bit tricky to get right, since
1647 * shifting can eat and free both this skb and the next,
1648 * so not even _safe variant of the loop is enough.
1651 tmp = tcp_shift_skb_data(sk, skb, state,
1652 start_seq, end_seq, dup_sack);
1661 in_sack = tcp_match_skb_to_sack(sk, skb,
1667 if (unlikely(in_sack < 0))
1671 TCP_SKB_CB(skb)->sacked =
1674 TCP_SKB_CB(skb)->sacked,
1675 TCP_SKB_CB(skb)->seq,
1676 TCP_SKB_CB(skb)->end_seq,
1678 tcp_skb_pcount(skb));
1680 if (!before(TCP_SKB_CB(skb)->seq,
1681 tcp_highest_sack_seq(tp)))
1682 tcp_advance_highest_sack(sk, skb);
1685 state->fack_count += tcp_skb_pcount(skb);
1690 /* Avoid all extra work that is being done by sacktag while walking in
1693 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1694 struct tcp_sacktag_state *state,
1697 tcp_for_write_queue_from(skb, sk) {
1698 if (skb == tcp_send_head(sk))
1701 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1704 state->fack_count += tcp_skb_pcount(skb);
1709 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1711 struct tcp_sack_block *next_dup,
1712 struct tcp_sacktag_state *state,
1715 if (next_dup == NULL)
1718 if (before(next_dup->start_seq, skip_to_seq)) {
1719 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1720 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1721 next_dup->start_seq, next_dup->end_seq,
1728 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1730 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1734 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1737 const struct inet_connection_sock *icsk = inet_csk(sk);
1738 struct tcp_sock *tp = tcp_sk(sk);
1739 unsigned char *ptr = (skb_transport_header(ack_skb) +
1740 TCP_SKB_CB(ack_skb)->sacked);
1741 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1742 struct tcp_sack_block sp[TCP_NUM_SACKS];
1743 struct tcp_sack_block *cache;
1744 struct tcp_sacktag_state state;
1745 struct sk_buff *skb;
1746 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1748 int found_dup_sack = 0;
1750 int first_sack_index;
1753 state.reord = tp->packets_out;
1755 if (!tp->sacked_out) {
1756 if (WARN_ON(tp->fackets_out))
1757 tp->fackets_out = 0;
1758 tcp_highest_sack_reset(sk);
1761 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1762 num_sacks, prior_snd_una);
1764 state.flag |= FLAG_DSACKING_ACK;
1766 /* Eliminate too old ACKs, but take into
1767 * account more or less fresh ones, they can
1768 * contain valid SACK info.
1770 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1773 if (!tp->packets_out)
1777 first_sack_index = 0;
1778 for (i = 0; i < num_sacks; i++) {
1779 int dup_sack = !i && found_dup_sack;
1781 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1782 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1784 if (!tcp_is_sackblock_valid(tp, dup_sack,
1785 sp[used_sacks].start_seq,
1786 sp[used_sacks].end_seq)) {
1790 if (!tp->undo_marker)
1791 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1793 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1795 /* Don't count olds caused by ACK reordering */
1796 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1797 !after(sp[used_sacks].end_seq, tp->snd_una))
1799 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1802 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1804 first_sack_index = -1;
1808 /* Ignore very old stuff early */
1809 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1815 /* order SACK blocks to allow in order walk of the retrans queue */
1816 for (i = used_sacks - 1; i > 0; i--) {
1817 for (j = 0; j < i; j++) {
1818 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1819 swap(sp[j], sp[j + 1]);
1821 /* Track where the first SACK block goes to */
1822 if (j == first_sack_index)
1823 first_sack_index = j + 1;
1828 skb = tcp_write_queue_head(sk);
1829 state.fack_count = 0;
1832 if (!tp->sacked_out) {
1833 /* It's already past, so skip checking against it */
1834 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1836 cache = tp->recv_sack_cache;
1837 /* Skip empty blocks in at head of the cache */
1838 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1843 while (i < used_sacks) {
1844 u32 start_seq = sp[i].start_seq;
1845 u32 end_seq = sp[i].end_seq;
1846 int dup_sack = (found_dup_sack && (i == first_sack_index));
1847 struct tcp_sack_block *next_dup = NULL;
1849 if (found_dup_sack && ((i + 1) == first_sack_index))
1850 next_dup = &sp[i + 1];
1852 /* Event "B" in the comment above. */
1853 if (after(end_seq, tp->high_seq))
1854 state.flag |= FLAG_DATA_LOST;
1856 /* Skip too early cached blocks */
1857 while (tcp_sack_cache_ok(tp, cache) &&
1858 !before(start_seq, cache->end_seq))
1861 /* Can skip some work by looking recv_sack_cache? */
1862 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1863 after(end_seq, cache->start_seq)) {
1866 if (before(start_seq, cache->start_seq)) {
1867 skb = tcp_sacktag_skip(skb, sk, &state,
1869 skb = tcp_sacktag_walk(skb, sk, next_dup,
1876 /* Rest of the block already fully processed? */
1877 if (!after(end_seq, cache->end_seq))
1880 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1884 /* ...tail remains todo... */
1885 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1886 /* ...but better entrypoint exists! */
1887 skb = tcp_highest_sack(sk);
1890 state.fack_count = tp->fackets_out;
1895 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1896 /* Check overlap against next cached too (past this one already) */
1901 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1902 skb = tcp_highest_sack(sk);
1905 state.fack_count = tp->fackets_out;
1907 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1910 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1911 start_seq, end_seq, dup_sack);
1914 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1915 * due to in-order walk
1917 if (after(end_seq, tp->frto_highmark))
1918 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1923 /* Clear the head of the cache sack blocks so we can skip it next time */
1924 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1925 tp->recv_sack_cache[i].start_seq = 0;
1926 tp->recv_sack_cache[i].end_seq = 0;
1928 for (j = 0; j < used_sacks; j++)
1929 tp->recv_sack_cache[i++] = sp[j];
1931 tcp_mark_lost_retrans(sk);
1933 tcp_verify_left_out(tp);
1935 if ((state.reord < tp->fackets_out) &&
1936 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1937 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1938 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1942 #if FASTRETRANS_DEBUG > 0
1943 WARN_ON((int)tp->sacked_out < 0);
1944 WARN_ON((int)tp->lost_out < 0);
1945 WARN_ON((int)tp->retrans_out < 0);
1946 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1951 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1952 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1954 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1958 holes = max(tp->lost_out, 1U);
1959 holes = min(holes, tp->packets_out);
1961 if ((tp->sacked_out + holes) > tp->packets_out) {
1962 tp->sacked_out = tp->packets_out - holes;
1968 /* If we receive more dupacks than we expected counting segments
1969 * in assumption of absent reordering, interpret this as reordering.
1970 * The only another reason could be bug in receiver TCP.
1972 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1974 struct tcp_sock *tp = tcp_sk(sk);
1975 if (tcp_limit_reno_sacked(tp))
1976 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1979 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1981 static void tcp_add_reno_sack(struct sock *sk)
1983 struct tcp_sock *tp = tcp_sk(sk);
1985 tcp_check_reno_reordering(sk, 0);
1986 tcp_verify_left_out(tp);
1989 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1991 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1993 struct tcp_sock *tp = tcp_sk(sk);
1996 /* One ACK acked hole. The rest eat duplicate ACKs. */
1997 if (acked - 1 >= tp->sacked_out)
2000 tp->sacked_out -= acked - 1;
2002 tcp_check_reno_reordering(sk, acked);
2003 tcp_verify_left_out(tp);
2006 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2011 static int tcp_is_sackfrto(const struct tcp_sock *tp)
2013 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2016 /* F-RTO can only be used if TCP has never retransmitted anything other than
2017 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2019 int tcp_use_frto(struct sock *sk)
2021 const struct tcp_sock *tp = tcp_sk(sk);
2022 const struct inet_connection_sock *icsk = inet_csk(sk);
2023 struct sk_buff *skb;
2025 if (!sysctl_tcp_frto)
2028 /* MTU probe and F-RTO won't really play nicely along currently */
2029 if (icsk->icsk_mtup.probe_size)
2032 if (tcp_is_sackfrto(tp))
2035 /* Avoid expensive walking of rexmit queue if possible */
2036 if (tp->retrans_out > 1)
2039 skb = tcp_write_queue_head(sk);
2040 if (tcp_skb_is_last(sk, skb))
2042 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2043 tcp_for_write_queue_from(skb, sk) {
2044 if (skb == tcp_send_head(sk))
2046 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2048 /* Short-circuit when first non-SACKed skb has been checked */
2049 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2055 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2056 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2057 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2058 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2059 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2060 * bits are handled if the Loss state is really to be entered (in
2061 * tcp_enter_frto_loss).
2063 * Do like tcp_enter_loss() would; when RTO expires the second time it
2065 * "Reduce ssthresh if it has not yet been made inside this window."
2067 void tcp_enter_frto(struct sock *sk)
2069 const struct inet_connection_sock *icsk = inet_csk(sk);
2070 struct tcp_sock *tp = tcp_sk(sk);
2071 struct sk_buff *skb;
2073 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2074 tp->snd_una == tp->high_seq ||
2075 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2076 !icsk->icsk_retransmits)) {
2077 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2078 /* Our state is too optimistic in ssthresh() call because cwnd
2079 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2080 * recovery has not yet completed. Pattern would be this: RTO,
2081 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2083 * RFC4138 should be more specific on what to do, even though
2084 * RTO is quite unlikely to occur after the first Cumulative ACK
2085 * due to back-off and complexity of triggering events ...
2087 if (tp->frto_counter) {
2089 stored_cwnd = tp->snd_cwnd;
2091 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2092 tp->snd_cwnd = stored_cwnd;
2094 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2096 /* ... in theory, cong.control module could do "any tricks" in
2097 * ssthresh(), which means that ca_state, lost bits and lost_out
2098 * counter would have to be faked before the call occurs. We
2099 * consider that too expensive, unlikely and hacky, so modules
2100 * using these in ssthresh() must deal these incompatibility
2101 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2103 tcp_ca_event(sk, CA_EVENT_FRTO);
2106 tp->undo_marker = tp->snd_una;
2107 tp->undo_retrans = 0;
2109 skb = tcp_write_queue_head(sk);
2110 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2111 tp->undo_marker = 0;
2112 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2113 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2114 tp->retrans_out -= tcp_skb_pcount(skb);
2116 tcp_verify_left_out(tp);
2118 /* Too bad if TCP was application limited */
2119 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2121 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2122 * The last condition is necessary at least in tp->frto_counter case.
2124 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2125 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2126 after(tp->high_seq, tp->snd_una)) {
2127 tp->frto_highmark = tp->high_seq;
2129 tp->frto_highmark = tp->snd_nxt;
2131 tcp_set_ca_state(sk, TCP_CA_Disorder);
2132 tp->high_seq = tp->snd_nxt;
2133 tp->frto_counter = 1;
2136 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2137 * which indicates that we should follow the traditional RTO recovery,
2138 * i.e. mark everything lost and do go-back-N retransmission.
2140 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2142 struct tcp_sock *tp = tcp_sk(sk);
2143 struct sk_buff *skb;
2146 tp->retrans_out = 0;
2147 if (tcp_is_reno(tp))
2148 tcp_reset_reno_sack(tp);
2150 tcp_for_write_queue(skb, sk) {
2151 if (skb == tcp_send_head(sk))
2154 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2156 * Count the retransmission made on RTO correctly (only when
2157 * waiting for the first ACK and did not get it)...
2159 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2160 /* For some reason this R-bit might get cleared? */
2161 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2162 tp->retrans_out += tcp_skb_pcount(skb);
2163 /* ...enter this if branch just for the first segment */
2164 flag |= FLAG_DATA_ACKED;
2166 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2167 tp->undo_marker = 0;
2168 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2171 /* Marking forward transmissions that were made after RTO lost
2172 * can cause unnecessary retransmissions in some scenarios,
2173 * SACK blocks will mitigate that in some but not in all cases.
2174 * We used to not mark them but it was causing break-ups with
2175 * receivers that do only in-order receival.
2177 * TODO: we could detect presence of such receiver and select
2178 * different behavior per flow.
2180 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2181 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2182 tp->lost_out += tcp_skb_pcount(skb);
2183 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2186 tcp_verify_left_out(tp);
2188 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2189 tp->snd_cwnd_cnt = 0;
2190 tp->snd_cwnd_stamp = tcp_time_stamp;
2191 tp->frto_counter = 0;
2192 tp->bytes_acked = 0;
2194 tp->reordering = min_t(unsigned int, tp->reordering,
2195 sysctl_tcp_reordering);
2196 tcp_set_ca_state(sk, TCP_CA_Loss);
2197 tp->high_seq = tp->snd_nxt;
2198 TCP_ECN_queue_cwr(tp);
2200 tcp_clear_all_retrans_hints(tp);
2203 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2205 tp->retrans_out = 0;
2208 tp->undo_marker = 0;
2209 tp->undo_retrans = 0;
2212 void tcp_clear_retrans(struct tcp_sock *tp)
2214 tcp_clear_retrans_partial(tp);
2216 tp->fackets_out = 0;
2220 /* Enter Loss state. If "how" is not zero, forget all SACK information
2221 * and reset tags completely, otherwise preserve SACKs. If receiver
2222 * dropped its ofo queue, we will know this due to reneging detection.
2224 void tcp_enter_loss(struct sock *sk, int how)
2226 const struct inet_connection_sock *icsk = inet_csk(sk);
2227 struct tcp_sock *tp = tcp_sk(sk);
2228 struct sk_buff *skb;
2230 /* Reduce ssthresh if it has not yet been made inside this window. */
2231 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2232 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2233 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2234 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2235 tcp_ca_event(sk, CA_EVENT_LOSS);
2238 tp->snd_cwnd_cnt = 0;
2239 tp->snd_cwnd_stamp = tcp_time_stamp;
2241 tp->bytes_acked = 0;
2242 tcp_clear_retrans_partial(tp);
2244 if (tcp_is_reno(tp))
2245 tcp_reset_reno_sack(tp);
2247 tp->undo_marker = tp->snd_una;
2250 tp->fackets_out = 0;
2252 tcp_clear_all_retrans_hints(tp);
2254 tcp_for_write_queue(skb, sk) {
2255 if (skb == tcp_send_head(sk))
2258 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2259 tp->undo_marker = 0;
2260 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2261 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2262 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2263 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2264 tp->lost_out += tcp_skb_pcount(skb);
2265 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2268 tcp_verify_left_out(tp);
2270 tp->reordering = min_t(unsigned int, tp->reordering,
2271 sysctl_tcp_reordering);
2272 tcp_set_ca_state(sk, TCP_CA_Loss);
2273 tp->high_seq = tp->snd_nxt;
2274 TCP_ECN_queue_cwr(tp);
2275 /* Abort F-RTO algorithm if one is in progress */
2276 tp->frto_counter = 0;
2279 /* If ACK arrived pointing to a remembered SACK, it means that our
2280 * remembered SACKs do not reflect real state of receiver i.e.
2281 * receiver _host_ is heavily congested (or buggy).
2283 * Do processing similar to RTO timeout.
2285 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2287 if (flag & FLAG_SACK_RENEGING) {
2288 struct inet_connection_sock *icsk = inet_csk(sk);
2289 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2291 tcp_enter_loss(sk, 1);
2292 icsk->icsk_retransmits++;
2293 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2294 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2295 icsk->icsk_rto, TCP_RTO_MAX);
2301 static inline int tcp_fackets_out(struct tcp_sock *tp)
2303 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2306 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2307 * counter when SACK is enabled (without SACK, sacked_out is used for
2310 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2311 * segments up to the highest received SACK block so far and holes in
2314 * With reordering, holes may still be in flight, so RFC3517 recovery
2315 * uses pure sacked_out (total number of SACKed segments) even though
2316 * it violates the RFC that uses duplicate ACKs, often these are equal
2317 * but when e.g. out-of-window ACKs or packet duplication occurs,
2318 * they differ. Since neither occurs due to loss, TCP should really
2321 static inline int tcp_dupack_heuristics(struct tcp_sock *tp)
2323 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2326 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2328 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2331 static inline int tcp_head_timedout(struct sock *sk)
2333 struct tcp_sock *tp = tcp_sk(sk);
2335 return tp->packets_out &&
2336 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2339 /* Linux NewReno/SACK/FACK/ECN state machine.
2340 * --------------------------------------
2342 * "Open" Normal state, no dubious events, fast path.
2343 * "Disorder" In all the respects it is "Open",
2344 * but requires a bit more attention. It is entered when
2345 * we see some SACKs or dupacks. It is split of "Open"
2346 * mainly to move some processing from fast path to slow one.
2347 * "CWR" CWND was reduced due to some Congestion Notification event.
2348 * It can be ECN, ICMP source quench, local device congestion.
2349 * "Recovery" CWND was reduced, we are fast-retransmitting.
2350 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2352 * tcp_fastretrans_alert() is entered:
2353 * - each incoming ACK, if state is not "Open"
2354 * - when arrived ACK is unusual, namely:
2359 * Counting packets in flight is pretty simple.
2361 * in_flight = packets_out - left_out + retrans_out
2363 * packets_out is SND.NXT-SND.UNA counted in packets.
2365 * retrans_out is number of retransmitted segments.
2367 * left_out is number of segments left network, but not ACKed yet.
2369 * left_out = sacked_out + lost_out
2371 * sacked_out: Packets, which arrived to receiver out of order
2372 * and hence not ACKed. With SACKs this number is simply
2373 * amount of SACKed data. Even without SACKs
2374 * it is easy to give pretty reliable estimate of this number,
2375 * counting duplicate ACKs.
2377 * lost_out: Packets lost by network. TCP has no explicit
2378 * "loss notification" feedback from network (for now).
2379 * It means that this number can be only _guessed_.
2380 * Actually, it is the heuristics to predict lossage that
2381 * distinguishes different algorithms.
2383 * F.e. after RTO, when all the queue is considered as lost,
2384 * lost_out = packets_out and in_flight = retrans_out.
2386 * Essentially, we have now two algorithms counting
2389 * FACK: It is the simplest heuristics. As soon as we decided
2390 * that something is lost, we decide that _all_ not SACKed
2391 * packets until the most forward SACK are lost. I.e.
2392 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2393 * It is absolutely correct estimate, if network does not reorder
2394 * packets. And it loses any connection to reality when reordering
2395 * takes place. We use FACK by default until reordering
2396 * is suspected on the path to this destination.
2398 * NewReno: when Recovery is entered, we assume that one segment
2399 * is lost (classic Reno). While we are in Recovery and
2400 * a partial ACK arrives, we assume that one more packet
2401 * is lost (NewReno). This heuristics are the same in NewReno
2404 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2405 * deflation etc. CWND is real congestion window, never inflated, changes
2406 * only according to classic VJ rules.
2408 * Really tricky (and requiring careful tuning) part of algorithm
2409 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2410 * The first determines the moment _when_ we should reduce CWND and,
2411 * hence, slow down forward transmission. In fact, it determines the moment
2412 * when we decide that hole is caused by loss, rather than by a reorder.
2414 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2415 * holes, caused by lost packets.
2417 * And the most logically complicated part of algorithm is undo
2418 * heuristics. We detect false retransmits due to both too early
2419 * fast retransmit (reordering) and underestimated RTO, analyzing
2420 * timestamps and D-SACKs. When we detect that some segments were
2421 * retransmitted by mistake and CWND reduction was wrong, we undo
2422 * window reduction and abort recovery phase. This logic is hidden
2423 * inside several functions named tcp_try_undo_<something>.
2426 /* This function decides, when we should leave Disordered state
2427 * and enter Recovery phase, reducing congestion window.
2429 * Main question: may we further continue forward transmission
2430 * with the same cwnd?
2432 static int tcp_time_to_recover(struct sock *sk)
2434 struct tcp_sock *tp = tcp_sk(sk);
2437 /* Do not perform any recovery during F-RTO algorithm */
2438 if (tp->frto_counter)
2441 /* Trick#1: The loss is proven. */
2445 /* Not-A-Trick#2 : Classic rule... */
2446 if (tcp_dupack_heuristics(tp) > tp->reordering)
2449 /* Trick#3 : when we use RFC2988 timer restart, fast
2450 * retransmit can be triggered by timeout of queue head.
2452 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2455 /* Trick#4: It is still not OK... But will it be useful to delay
2458 packets_out = tp->packets_out;
2459 if (packets_out <= tp->reordering &&
2460 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2461 !tcp_may_send_now(sk)) {
2462 /* We have nothing to send. This connection is limited
2463 * either by receiver window or by application.
2468 /* If a thin stream is detected, retransmit after first
2469 * received dupack. Employ only if SACK is supported in order
2470 * to avoid possible corner-case series of spurious retransmissions
2471 * Use only if there are no unsent data.
2473 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2474 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2475 tcp_is_sack(tp) && !tcp_send_head(sk))
2481 /* New heuristics: it is possible only after we switched to restart timer
2482 * each time when something is ACKed. Hence, we can detect timed out packets
2483 * during fast retransmit without falling to slow start.
2485 * Usefulness of this as is very questionable, since we should know which of
2486 * the segments is the next to timeout which is relatively expensive to find
2487 * in general case unless we add some data structure just for that. The
2488 * current approach certainly won't find the right one too often and when it
2489 * finally does find _something_ it usually marks large part of the window
2490 * right away (because a retransmission with a larger timestamp blocks the
2491 * loop from advancing). -ij
2493 static void tcp_timeout_skbs(struct sock *sk)
2495 struct tcp_sock *tp = tcp_sk(sk);
2496 struct sk_buff *skb;
2498 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2501 skb = tp->scoreboard_skb_hint;
2502 if (tp->scoreboard_skb_hint == NULL)
2503 skb = tcp_write_queue_head(sk);
2505 tcp_for_write_queue_from(skb, sk) {
2506 if (skb == tcp_send_head(sk))
2508 if (!tcp_skb_timedout(sk, skb))
2511 tcp_skb_mark_lost(tp, skb);
2514 tp->scoreboard_skb_hint = skb;
2516 tcp_verify_left_out(tp);
2519 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2520 * is against sacked "cnt", otherwise it's against facked "cnt"
2522 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2524 struct tcp_sock *tp = tcp_sk(sk);
2525 struct sk_buff *skb;
2530 WARN_ON(packets > tp->packets_out);
2531 if (tp->lost_skb_hint) {
2532 skb = tp->lost_skb_hint;
2533 cnt = tp->lost_cnt_hint;
2534 /* Head already handled? */
2535 if (mark_head && skb != tcp_write_queue_head(sk))
2538 skb = tcp_write_queue_head(sk);
2542 tcp_for_write_queue_from(skb, sk) {
2543 if (skb == tcp_send_head(sk))
2545 /* TODO: do this better */
2546 /* this is not the most efficient way to do this... */
2547 tp->lost_skb_hint = skb;
2548 tp->lost_cnt_hint = cnt;
2550 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2554 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2555 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2556 cnt += tcp_skb_pcount(skb);
2558 if (cnt > packets) {
2559 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2560 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2561 (oldcnt >= packets))
2564 mss = skb_shinfo(skb)->gso_size;
2565 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2571 tcp_skb_mark_lost(tp, skb);
2576 tcp_verify_left_out(tp);
2579 /* Account newly detected lost packet(s) */
2581 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2583 struct tcp_sock *tp = tcp_sk(sk);
2585 if (tcp_is_reno(tp)) {
2586 tcp_mark_head_lost(sk, 1, 1);
2587 } else if (tcp_is_fack(tp)) {
2588 int lost = tp->fackets_out - tp->reordering;
2591 tcp_mark_head_lost(sk, lost, 0);
2593 int sacked_upto = tp->sacked_out - tp->reordering;
2594 if (sacked_upto >= 0)
2595 tcp_mark_head_lost(sk, sacked_upto, 0);
2596 else if (fast_rexmit)
2597 tcp_mark_head_lost(sk, 1, 1);
2600 tcp_timeout_skbs(sk);
2603 /* CWND moderation, preventing bursts due to too big ACKs
2604 * in dubious situations.
2606 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2608 tp->snd_cwnd = min(tp->snd_cwnd,
2609 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2610 tp->snd_cwnd_stamp = tcp_time_stamp;
2613 /* Lower bound on congestion window is slow start threshold
2614 * unless congestion avoidance choice decides to overide it.
2616 static inline u32 tcp_cwnd_min(const struct sock *sk)
2618 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2620 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2623 /* Decrease cwnd each second ack. */
2624 static void tcp_cwnd_down(struct sock *sk, int flag)
2626 struct tcp_sock *tp = tcp_sk(sk);
2627 int decr = tp->snd_cwnd_cnt + 1;
2629 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2630 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2631 tp->snd_cwnd_cnt = decr & 1;
2634 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2635 tp->snd_cwnd -= decr;
2637 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2638 tp->snd_cwnd_stamp = tcp_time_stamp;
2642 /* Nothing was retransmitted or returned timestamp is less
2643 * than timestamp of the first retransmission.
2645 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2647 return !tp->retrans_stamp ||
2648 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2649 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2652 /* Undo procedures. */
2654 #if FASTRETRANS_DEBUG > 1
2655 static void DBGUNDO(struct sock *sk, const char *msg)
2657 struct tcp_sock *tp = tcp_sk(sk);
2658 struct inet_sock *inet = inet_sk(sk);
2660 if (sk->sk_family == AF_INET) {
2661 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2663 &inet->inet_daddr, ntohs(inet->inet_dport),
2664 tp->snd_cwnd, tcp_left_out(tp),
2665 tp->snd_ssthresh, tp->prior_ssthresh,
2668 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2669 else if (sk->sk_family == AF_INET6) {
2670 struct ipv6_pinfo *np = inet6_sk(sk);
2671 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2673 &np->daddr, ntohs(inet->inet_dport),
2674 tp->snd_cwnd, tcp_left_out(tp),
2675 tp->snd_ssthresh, tp->prior_ssthresh,
2681 #define DBGUNDO(x...) do { } while (0)
2684 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2686 struct tcp_sock *tp = tcp_sk(sk);
2688 if (tp->prior_ssthresh) {
2689 const struct inet_connection_sock *icsk = inet_csk(sk);
2691 if (icsk->icsk_ca_ops->undo_cwnd)
2692 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2694 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2696 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2697 tp->snd_ssthresh = tp->prior_ssthresh;
2698 TCP_ECN_withdraw_cwr(tp);
2701 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2703 tp->snd_cwnd_stamp = tcp_time_stamp;
2706 static inline int tcp_may_undo(struct tcp_sock *tp)
2708 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2711 /* People celebrate: "We love our President!" */
2712 static int tcp_try_undo_recovery(struct sock *sk)
2714 struct tcp_sock *tp = tcp_sk(sk);
2716 if (tcp_may_undo(tp)) {
2719 /* Happy end! We did not retransmit anything
2720 * or our original transmission succeeded.
2722 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2723 tcp_undo_cwr(sk, true);
2724 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2725 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2727 mib_idx = LINUX_MIB_TCPFULLUNDO;
2729 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2730 tp->undo_marker = 0;
2732 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2733 /* Hold old state until something *above* high_seq
2734 * is ACKed. For Reno it is MUST to prevent false
2735 * fast retransmits (RFC2582). SACK TCP is safe. */
2736 tcp_moderate_cwnd(tp);
2739 tcp_set_ca_state(sk, TCP_CA_Open);
2743 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2744 static void tcp_try_undo_dsack(struct sock *sk)
2746 struct tcp_sock *tp = tcp_sk(sk);
2748 if (tp->undo_marker && !tp->undo_retrans) {
2749 DBGUNDO(sk, "D-SACK");
2750 tcp_undo_cwr(sk, true);
2751 tp->undo_marker = 0;
2752 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2756 /* We can clear retrans_stamp when there are no retransmissions in the
2757 * window. It would seem that it is trivially available for us in
2758 * tp->retrans_out, however, that kind of assumptions doesn't consider
2759 * what will happen if errors occur when sending retransmission for the
2760 * second time. ...It could the that such segment has only
2761 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2762 * the head skb is enough except for some reneging corner cases that
2763 * are not worth the effort.
2765 * Main reason for all this complexity is the fact that connection dying
2766 * time now depends on the validity of the retrans_stamp, in particular,
2767 * that successive retransmissions of a segment must not advance
2768 * retrans_stamp under any conditions.
2770 static int tcp_any_retrans_done(struct sock *sk)
2772 struct tcp_sock *tp = tcp_sk(sk);
2773 struct sk_buff *skb;
2775 if (tp->retrans_out)
2778 skb = tcp_write_queue_head(sk);
2779 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2785 /* Undo during fast recovery after partial ACK. */
2787 static int tcp_try_undo_partial(struct sock *sk, int acked)
2789 struct tcp_sock *tp = tcp_sk(sk);
2790 /* Partial ACK arrived. Force Hoe's retransmit. */
2791 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2793 if (tcp_may_undo(tp)) {
2794 /* Plain luck! Hole if filled with delayed
2795 * packet, rather than with a retransmit.
2797 if (!tcp_any_retrans_done(sk))
2798 tp->retrans_stamp = 0;
2800 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2803 tcp_undo_cwr(sk, false);
2804 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2806 /* So... Do not make Hoe's retransmit yet.
2807 * If the first packet was delayed, the rest
2808 * ones are most probably delayed as well.
2815 /* Undo during loss recovery after partial ACK. */
2816 static int tcp_try_undo_loss(struct sock *sk)
2818 struct tcp_sock *tp = tcp_sk(sk);
2820 if (tcp_may_undo(tp)) {
2821 struct sk_buff *skb;
2822 tcp_for_write_queue(skb, sk) {
2823 if (skb == tcp_send_head(sk))
2825 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2828 tcp_clear_all_retrans_hints(tp);
2830 DBGUNDO(sk, "partial loss");
2832 tcp_undo_cwr(sk, true);
2833 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2834 inet_csk(sk)->icsk_retransmits = 0;
2835 tp->undo_marker = 0;
2836 if (tcp_is_sack(tp))
2837 tcp_set_ca_state(sk, TCP_CA_Open);
2843 static inline void tcp_complete_cwr(struct sock *sk)
2845 struct tcp_sock *tp = tcp_sk(sk);
2846 /* Do not moderate cwnd if it's already undone in cwr or recovery */
2847 if (tp->undo_marker && tp->snd_cwnd > tp->snd_ssthresh) {
2848 tp->snd_cwnd = tp->snd_ssthresh;
2849 tp->snd_cwnd_stamp = tcp_time_stamp;
2851 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2854 static void tcp_try_keep_open(struct sock *sk)
2856 struct tcp_sock *tp = tcp_sk(sk);
2857 int state = TCP_CA_Open;
2859 if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker)
2860 state = TCP_CA_Disorder;
2862 if (inet_csk(sk)->icsk_ca_state != state) {
2863 tcp_set_ca_state(sk, state);
2864 tp->high_seq = tp->snd_nxt;
2868 static void tcp_try_to_open(struct sock *sk, int flag)
2870 struct tcp_sock *tp = tcp_sk(sk);
2872 tcp_verify_left_out(tp);
2874 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2875 tp->retrans_stamp = 0;
2877 if (flag & FLAG_ECE)
2878 tcp_enter_cwr(sk, 1);
2880 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2881 tcp_try_keep_open(sk);
2882 tcp_moderate_cwnd(tp);
2884 tcp_cwnd_down(sk, flag);
2888 static void tcp_mtup_probe_failed(struct sock *sk)
2890 struct inet_connection_sock *icsk = inet_csk(sk);
2892 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2893 icsk->icsk_mtup.probe_size = 0;
2896 static void tcp_mtup_probe_success(struct sock *sk)
2898 struct tcp_sock *tp = tcp_sk(sk);
2899 struct inet_connection_sock *icsk = inet_csk(sk);
2901 /* FIXME: breaks with very large cwnd */
2902 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2903 tp->snd_cwnd = tp->snd_cwnd *
2904 tcp_mss_to_mtu(sk, tp->mss_cache) /
2905 icsk->icsk_mtup.probe_size;
2906 tp->snd_cwnd_cnt = 0;
2907 tp->snd_cwnd_stamp = tcp_time_stamp;
2908 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2910 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2911 icsk->icsk_mtup.probe_size = 0;
2912 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2915 /* Do a simple retransmit without using the backoff mechanisms in
2916 * tcp_timer. This is used for path mtu discovery.
2917 * The socket is already locked here.
2919 void tcp_simple_retransmit(struct sock *sk)
2921 const struct inet_connection_sock *icsk = inet_csk(sk);
2922 struct tcp_sock *tp = tcp_sk(sk);
2923 struct sk_buff *skb;
2924 unsigned int mss = tcp_current_mss(sk);
2925 u32 prior_lost = tp->lost_out;
2927 tcp_for_write_queue(skb, sk) {
2928 if (skb == tcp_send_head(sk))
2930 if (tcp_skb_seglen(skb) > mss &&
2931 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2932 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2933 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2934 tp->retrans_out -= tcp_skb_pcount(skb);
2936 tcp_skb_mark_lost_uncond_verify(tp, skb);
2940 tcp_clear_retrans_hints_partial(tp);
2942 if (prior_lost == tp->lost_out)
2945 if (tcp_is_reno(tp))
2946 tcp_limit_reno_sacked(tp);
2948 tcp_verify_left_out(tp);
2950 /* Don't muck with the congestion window here.
2951 * Reason is that we do not increase amount of _data_
2952 * in network, but units changed and effective
2953 * cwnd/ssthresh really reduced now.
2955 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2956 tp->high_seq = tp->snd_nxt;
2957 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2958 tp->prior_ssthresh = 0;
2959 tp->undo_marker = 0;
2960 tcp_set_ca_state(sk, TCP_CA_Loss);
2962 tcp_xmit_retransmit_queue(sk);
2964 EXPORT_SYMBOL(tcp_simple_retransmit);
2966 /* Process an event, which can update packets-in-flight not trivially.
2967 * Main goal of this function is to calculate new estimate for left_out,
2968 * taking into account both packets sitting in receiver's buffer and
2969 * packets lost by network.
2971 * Besides that it does CWND reduction, when packet loss is detected
2972 * and changes state of machine.
2974 * It does _not_ decide what to send, it is made in function
2975 * tcp_xmit_retransmit_queue().
2977 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2979 struct inet_connection_sock *icsk = inet_csk(sk);
2980 struct tcp_sock *tp = tcp_sk(sk);
2981 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2982 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2983 (tcp_fackets_out(tp) > tp->reordering));
2984 int fast_rexmit = 0, mib_idx;
2986 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2988 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2989 tp->fackets_out = 0;
2991 /* Now state machine starts.
2992 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2993 if (flag & FLAG_ECE)
2994 tp->prior_ssthresh = 0;
2996 /* B. In all the states check for reneging SACKs. */
2997 if (tcp_check_sack_reneging(sk, flag))
3000 /* C. Process data loss notification, provided it is valid. */
3001 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
3002 before(tp->snd_una, tp->high_seq) &&
3003 icsk->icsk_ca_state != TCP_CA_Open &&
3004 tp->fackets_out > tp->reordering) {
3005 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0);
3006 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
3009 /* D. Check consistency of the current state. */
3010 tcp_verify_left_out(tp);
3012 /* E. Check state exit conditions. State can be terminated
3013 * when high_seq is ACKed. */
3014 if (icsk->icsk_ca_state == TCP_CA_Open) {
3015 WARN_ON(tp->retrans_out != 0);
3016 tp->retrans_stamp = 0;
3017 } else if (!before(tp->snd_una, tp->high_seq)) {
3018 switch (icsk->icsk_ca_state) {
3020 icsk->icsk_retransmits = 0;
3021 if (tcp_try_undo_recovery(sk))
3026 /* CWR is to be held something *above* high_seq
3027 * is ACKed for CWR bit to reach receiver. */
3028 if (tp->snd_una != tp->high_seq) {
3029 tcp_complete_cwr(sk);
3030 tcp_set_ca_state(sk, TCP_CA_Open);
3034 case TCP_CA_Disorder:
3035 tcp_try_undo_dsack(sk);
3036 if (!tp->undo_marker ||
3037 /* For SACK case do not Open to allow to undo
3038 * catching for all duplicate ACKs. */
3039 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
3040 tp->undo_marker = 0;
3041 tcp_set_ca_state(sk, TCP_CA_Open);
3045 case TCP_CA_Recovery:
3046 if (tcp_is_reno(tp))
3047 tcp_reset_reno_sack(tp);
3048 if (tcp_try_undo_recovery(sk))
3050 tcp_complete_cwr(sk);
3055 /* F. Process state. */
3056 switch (icsk->icsk_ca_state) {
3057 case TCP_CA_Recovery:
3058 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3059 if (tcp_is_reno(tp) && is_dupack)
3060 tcp_add_reno_sack(sk);
3062 do_lost = tcp_try_undo_partial(sk, pkts_acked);
3065 if (flag & FLAG_DATA_ACKED)
3066 icsk->icsk_retransmits = 0;
3067 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3068 tcp_reset_reno_sack(tp);
3069 if (!tcp_try_undo_loss(sk)) {
3070 tcp_moderate_cwnd(tp);
3071 tcp_xmit_retransmit_queue(sk);
3074 if (icsk->icsk_ca_state != TCP_CA_Open)
3076 /* Loss is undone; fall through to processing in Open state. */
3078 if (tcp_is_reno(tp)) {
3079 if (flag & FLAG_SND_UNA_ADVANCED)
3080 tcp_reset_reno_sack(tp);
3082 tcp_add_reno_sack(sk);
3085 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3086 tcp_try_undo_dsack(sk);
3088 if (!tcp_time_to_recover(sk)) {
3089 tcp_try_to_open(sk, flag);
3093 /* MTU probe failure: don't reduce cwnd */
3094 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3095 icsk->icsk_mtup.probe_size &&
3096 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3097 tcp_mtup_probe_failed(sk);
3098 /* Restores the reduction we did in tcp_mtup_probe() */
3100 tcp_simple_retransmit(sk);
3104 /* Otherwise enter Recovery state */
3106 if (tcp_is_reno(tp))
3107 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3109 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3111 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3113 tp->high_seq = tp->snd_nxt;
3114 tp->prior_ssthresh = 0;
3115 tp->undo_marker = tp->snd_una;
3116 tp->undo_retrans = tp->retrans_out;
3118 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3119 if (!(flag & FLAG_ECE))
3120 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3121 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3122 TCP_ECN_queue_cwr(tp);
3125 tp->bytes_acked = 0;
3126 tp->snd_cwnd_cnt = 0;
3127 tcp_set_ca_state(sk, TCP_CA_Recovery);
3131 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3132 tcp_update_scoreboard(sk, fast_rexmit);
3133 tcp_cwnd_down(sk, flag);
3134 tcp_xmit_retransmit_queue(sk);
3137 static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3139 tcp_rtt_estimator(sk, seq_rtt);
3141 inet_csk(sk)->icsk_backoff = 0;
3144 /* Read draft-ietf-tcplw-high-performance before mucking
3145 * with this code. (Supersedes RFC1323)
3147 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3149 /* RTTM Rule: A TSecr value received in a segment is used to
3150 * update the averaged RTT measurement only if the segment
3151 * acknowledges some new data, i.e., only if it advances the
3152 * left edge of the send window.
3154 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3155 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3157 * Changed: reset backoff as soon as we see the first valid sample.
3158 * If we do not, we get strongly overestimated rto. With timestamps
3159 * samples are accepted even from very old segments: f.e., when rtt=1
3160 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3161 * answer arrives rto becomes 120 seconds! If at least one of segments
3162 * in window is lost... Voila. --ANK (010210)
3164 struct tcp_sock *tp = tcp_sk(sk);
3166 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3169 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3171 /* We don't have a timestamp. Can only use
3172 * packets that are not retransmitted to determine
3173 * rtt estimates. Also, we must not reset the
3174 * backoff for rto until we get a non-retransmitted
3175 * packet. This allows us to deal with a situation
3176 * where the network delay has increased suddenly.
3177 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3180 if (flag & FLAG_RETRANS_DATA_ACKED)
3183 tcp_valid_rtt_meas(sk, seq_rtt);
3186 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3189 const struct tcp_sock *tp = tcp_sk(sk);
3190 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3191 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3192 tcp_ack_saw_tstamp(sk, flag);
3193 else if (seq_rtt >= 0)
3194 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3197 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3199 const struct inet_connection_sock *icsk = inet_csk(sk);
3200 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3201 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3204 /* Restart timer after forward progress on connection.
3205 * RFC2988 recommends to restart timer to now+rto.
3207 static void tcp_rearm_rto(struct sock *sk)
3209 struct tcp_sock *tp = tcp_sk(sk);
3211 if (!tp->packets_out) {
3212 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3214 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3215 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3219 /* If we get here, the whole TSO packet has not been acked. */
3220 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3222 struct tcp_sock *tp = tcp_sk(sk);
3225 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3227 packets_acked = tcp_skb_pcount(skb);
3228 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3230 packets_acked -= tcp_skb_pcount(skb);
3232 if (packets_acked) {
3233 BUG_ON(tcp_skb_pcount(skb) == 0);
3234 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3237 return packets_acked;
3240 /* Remove acknowledged frames from the retransmission queue. If our packet
3241 * is before the ack sequence we can discard it as it's confirmed to have
3242 * arrived at the other end.
3244 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3247 struct tcp_sock *tp = tcp_sk(sk);
3248 const struct inet_connection_sock *icsk = inet_csk(sk);
3249 struct sk_buff *skb;
3250 u32 now = tcp_time_stamp;
3251 int fully_acked = 1;
3254 u32 reord = tp->packets_out;
3255 u32 prior_sacked = tp->sacked_out;
3257 s32 ca_seq_rtt = -1;
3258 ktime_t last_ackt = net_invalid_timestamp();
3260 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3261 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3263 u8 sacked = scb->sacked;
3265 /* Determine how many packets and what bytes were acked, tso and else */
3266 if (after(scb->end_seq, tp->snd_una)) {
3267 if (tcp_skb_pcount(skb) == 1 ||
3268 !after(tp->snd_una, scb->seq))
3271 acked_pcount = tcp_tso_acked(sk, skb);
3277 acked_pcount = tcp_skb_pcount(skb);
3280 if (sacked & TCPCB_RETRANS) {
3281 if (sacked & TCPCB_SACKED_RETRANS)
3282 tp->retrans_out -= acked_pcount;
3283 flag |= FLAG_RETRANS_DATA_ACKED;
3286 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3287 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3289 ca_seq_rtt = now - scb->when;
3290 last_ackt = skb->tstamp;
3292 seq_rtt = ca_seq_rtt;
3294 if (!(sacked & TCPCB_SACKED_ACKED))
3295 reord = min(pkts_acked, reord);
3298 if (sacked & TCPCB_SACKED_ACKED)
3299 tp->sacked_out -= acked_pcount;
3300 if (sacked & TCPCB_LOST)
3301 tp->lost_out -= acked_pcount;
3303 tp->packets_out -= acked_pcount;
3304 pkts_acked += acked_pcount;
3306 /* Initial outgoing SYN's get put onto the write_queue
3307 * just like anything else we transmit. It is not
3308 * true data, and if we misinform our callers that
3309 * this ACK acks real data, we will erroneously exit
3310 * connection startup slow start one packet too
3311 * quickly. This is severely frowned upon behavior.
3313 if (!(scb->flags & TCPHDR_SYN)) {
3314 flag |= FLAG_DATA_ACKED;
3316 flag |= FLAG_SYN_ACKED;
3317 tp->retrans_stamp = 0;
3323 tcp_unlink_write_queue(skb, sk);
3324 sk_wmem_free_skb(sk, skb);
3325 tp->scoreboard_skb_hint = NULL;
3326 if (skb == tp->retransmit_skb_hint)
3327 tp->retransmit_skb_hint = NULL;
3328 if (skb == tp->lost_skb_hint)
3329 tp->lost_skb_hint = NULL;
3332 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3333 tp->snd_up = tp->snd_una;
3335 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3336 flag |= FLAG_SACK_RENEGING;
3338 if (flag & FLAG_ACKED) {
3339 const struct tcp_congestion_ops *ca_ops
3340 = inet_csk(sk)->icsk_ca_ops;
3342 if (unlikely(icsk->icsk_mtup.probe_size &&
3343 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3344 tcp_mtup_probe_success(sk);
3347 tcp_ack_update_rtt(sk, flag, seq_rtt);
3350 if (tcp_is_reno(tp)) {
3351 tcp_remove_reno_sacks(sk, pkts_acked);
3355 /* Non-retransmitted hole got filled? That's reordering */
3356 if (reord < prior_fackets)
3357 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3359 delta = tcp_is_fack(tp) ? pkts_acked :
3360 prior_sacked - tp->sacked_out;
3361 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3364 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3366 if (ca_ops->pkts_acked) {
3369 /* Is the ACK triggering packet unambiguous? */
3370 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3371 /* High resolution needed and available? */
3372 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3373 !ktime_equal(last_ackt,
3374 net_invalid_timestamp()))
3375 rtt_us = ktime_us_delta(ktime_get_real(),
3377 else if (ca_seq_rtt >= 0)
3378 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3381 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3385 #if FASTRETRANS_DEBUG > 0
3386 WARN_ON((int)tp->sacked_out < 0);
3387 WARN_ON((int)tp->lost_out < 0);
3388 WARN_ON((int)tp->retrans_out < 0);
3389 if (!tp->packets_out && tcp_is_sack(tp)) {
3390 icsk = inet_csk(sk);
3392 printk(KERN_DEBUG "Leak l=%u %d\n",
3393 tp->lost_out, icsk->icsk_ca_state);
3396 if (tp->sacked_out) {
3397 printk(KERN_DEBUG "Leak s=%u %d\n",
3398 tp->sacked_out, icsk->icsk_ca_state);
3401 if (tp->retrans_out) {
3402 printk(KERN_DEBUG "Leak r=%u %d\n",
3403 tp->retrans_out, icsk->icsk_ca_state);
3404 tp->retrans_out = 0;
3411 static void tcp_ack_probe(struct sock *sk)
3413 const struct tcp_sock *tp = tcp_sk(sk);
3414 struct inet_connection_sock *icsk = inet_csk(sk);
3416 /* Was it a usable window open? */
3418 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3419 icsk->icsk_backoff = 0;
3420 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3421 /* Socket must be waked up by subsequent tcp_data_snd_check().
3422 * This function is not for random using!
3425 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3426 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3431 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3433 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3434 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3437 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3439 const struct tcp_sock *tp = tcp_sk(sk);
3440 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3441 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3444 /* Check that window update is acceptable.
3445 * The function assumes that snd_una<=ack<=snd_next.
3447 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3448 const u32 ack, const u32 ack_seq,
3451 return after(ack, tp->snd_una) ||
3452 after(ack_seq, tp->snd_wl1) ||
3453 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3456 /* Update our send window.
3458 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3459 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3461 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3464 struct tcp_sock *tp = tcp_sk(sk);
3466 u32 nwin = ntohs(tcp_hdr(skb)->window);
3468 if (likely(!tcp_hdr(skb)->syn))
3469 nwin <<= tp->rx_opt.snd_wscale;
3471 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3472 flag |= FLAG_WIN_UPDATE;
3473 tcp_update_wl(tp, ack_seq);
3475 if (tp->snd_wnd != nwin) {
3478 /* Note, it is the only place, where
3479 * fast path is recovered for sending TCP.
3482 tcp_fast_path_check(sk);
3484 if (nwin > tp->max_window) {
3485 tp->max_window = nwin;
3486 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3496 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3497 * continue in congestion avoidance.
3499 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3501 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3502 tp->snd_cwnd_cnt = 0;
3503 tp->bytes_acked = 0;
3504 TCP_ECN_queue_cwr(tp);
3505 tcp_moderate_cwnd(tp);
3508 /* A conservative spurious RTO response algorithm: reduce cwnd using
3509 * rate halving and continue in congestion avoidance.
3511 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3513 tcp_enter_cwr(sk, 0);
3516 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3518 if (flag & FLAG_ECE)
3519 tcp_ratehalving_spur_to_response(sk);
3521 tcp_undo_cwr(sk, true);
3524 /* F-RTO spurious RTO detection algorithm (RFC4138)
3526 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3527 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3528 * window (but not to or beyond highest sequence sent before RTO):
3529 * On First ACK, send two new segments out.
3530 * On Second ACK, RTO was likely spurious. Do spurious response (response
3531 * algorithm is not part of the F-RTO detection algorithm
3532 * given in RFC4138 but can be selected separately).
3533 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3534 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3535 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3536 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3538 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3539 * original window even after we transmit two new data segments.
3542 * on first step, wait until first cumulative ACK arrives, then move to
3543 * the second step. In second step, the next ACK decides.
3545 * F-RTO is implemented (mainly) in four functions:
3546 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3547 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3548 * called when tcp_use_frto() showed green light
3549 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3550 * - tcp_enter_frto_loss() is called if there is not enough evidence
3551 * to prove that the RTO is indeed spurious. It transfers the control
3552 * from F-RTO to the conventional RTO recovery
3554 static int tcp_process_frto(struct sock *sk, int flag)
3556 struct tcp_sock *tp = tcp_sk(sk);
3558 tcp_verify_left_out(tp);
3560 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3561 if (flag & FLAG_DATA_ACKED)
3562 inet_csk(sk)->icsk_retransmits = 0;
3564 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3565 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3566 tp->undo_marker = 0;
3568 if (!before(tp->snd_una, tp->frto_highmark)) {
3569 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3573 if (!tcp_is_sackfrto(tp)) {
3574 /* RFC4138 shortcoming in step 2; should also have case c):
3575 * ACK isn't duplicate nor advances window, e.g., opposite dir
3578 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3581 if (!(flag & FLAG_DATA_ACKED)) {
3582 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3587 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3588 if (!tcp_packets_in_flight(tp)) {
3589 tcp_enter_frto_loss(sk, 2, flag);
3593 /* Prevent sending of new data. */
3594 tp->snd_cwnd = min(tp->snd_cwnd,
3595 tcp_packets_in_flight(tp));
3599 if ((tp->frto_counter >= 2) &&
3600 (!(flag & FLAG_FORWARD_PROGRESS) ||
3601 ((flag & FLAG_DATA_SACKED) &&
3602 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3603 /* RFC4138 shortcoming (see comment above) */
3604 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3605 (flag & FLAG_NOT_DUP))
3608 tcp_enter_frto_loss(sk, 3, flag);
3613 if (tp->frto_counter == 1) {
3614 /* tcp_may_send_now needs to see updated state */
3615 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3616 tp->frto_counter = 2;
3618 if (!tcp_may_send_now(sk))
3619 tcp_enter_frto_loss(sk, 2, flag);
3623 switch (sysctl_tcp_frto_response) {
3625 tcp_undo_spur_to_response(sk, flag);
3628 tcp_conservative_spur_to_response(tp);
3631 tcp_ratehalving_spur_to_response(sk);
3634 tp->frto_counter = 0;
3635 tp->undo_marker = 0;
3636 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3641 /* RFC 5961 7 [ACK Throttling] */
3642 static void tcp_send_challenge_ack(struct sock *sk)
3644 /* unprotected vars, we dont care of overwrites */
3645 static u32 challenge_timestamp;
3646 static unsigned int challenge_count;
3647 u32 now = jiffies / HZ;
3649 if (now != challenge_timestamp) {
3650 challenge_timestamp = now;
3651 challenge_count = 0;
3653 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3654 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3659 /* This routine deals with incoming acks, but not outgoing ones. */
3660 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3662 struct inet_connection_sock *icsk = inet_csk(sk);
3663 struct tcp_sock *tp = tcp_sk(sk);
3664 u32 prior_snd_una = tp->snd_una;
3665 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3666 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3667 u32 prior_in_flight;
3672 /* If the ack is older than previous acks
3673 * then we can probably ignore it.
3675 if (before(ack, prior_snd_una)) {
3676 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3677 if (before(ack, prior_snd_una - tp->max_window)) {
3678 tcp_send_challenge_ack(sk);
3684 /* If the ack includes data we haven't sent yet, discard
3685 * this segment (RFC793 Section 3.9).
3687 if (after(ack, tp->snd_nxt))
3690 if (after(ack, prior_snd_una))
3691 flag |= FLAG_SND_UNA_ADVANCED;
3693 if (sysctl_tcp_abc) {
3694 if (icsk->icsk_ca_state < TCP_CA_CWR)
3695 tp->bytes_acked += ack - prior_snd_una;
3696 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3697 /* we assume just one segment left network */
3698 tp->bytes_acked += min(ack - prior_snd_una,
3702 prior_fackets = tp->fackets_out;
3703 prior_in_flight = tcp_packets_in_flight(tp);
3705 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3706 /* Window is constant, pure forward advance.
3707 * No more checks are required.
3708 * Note, we use the fact that SND.UNA>=SND.WL2.
3710 tcp_update_wl(tp, ack_seq);
3712 flag |= FLAG_WIN_UPDATE;
3714 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3716 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3718 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3721 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3723 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3725 if (TCP_SKB_CB(skb)->sacked)
3726 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3728 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3731 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3734 /* We passed data and got it acked, remove any soft error
3735 * log. Something worked...
3737 sk->sk_err_soft = 0;
3738 icsk->icsk_probes_out = 0;
3739 tp->rcv_tstamp = tcp_time_stamp;
3740 prior_packets = tp->packets_out;
3744 /* See if we can take anything off of the retransmit queue. */
3745 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3747 if (tp->frto_counter)
3748 frto_cwnd = tcp_process_frto(sk, flag);
3749 /* Guarantee sacktag reordering detection against wrap-arounds */
3750 if (before(tp->frto_highmark, tp->snd_una))
3751 tp->frto_highmark = 0;
3753 if (tcp_ack_is_dubious(sk, flag)) {
3754 /* Advance CWND, if state allows this. */
3755 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3756 tcp_may_raise_cwnd(sk, flag))
3757 tcp_cong_avoid(sk, ack, prior_in_flight);
3758 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3761 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3762 tcp_cong_avoid(sk, ack, prior_in_flight);
3765 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3766 dst_confirm(__sk_dst_get(sk));
3771 /* If this ack opens up a zero window, clear backoff. It was
3772 * being used to time the probes, and is probably far higher than
3773 * it needs to be for normal retransmission.
3775 if (tcp_send_head(sk))
3780 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3784 if (TCP_SKB_CB(skb)->sacked) {
3785 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3786 if (icsk->icsk_ca_state == TCP_CA_Open)
3787 tcp_try_keep_open(sk);
3790 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3794 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3795 * But, this can also be called on packets in the established flow when
3796 * the fast version below fails.
3798 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3799 u8 **hvpp, int estab)
3802 struct tcphdr *th = tcp_hdr(skb);
3803 int length = (th->doff * 4) - sizeof(struct tcphdr);
3805 ptr = (unsigned char *)(th + 1);
3806 opt_rx->saw_tstamp = 0;
3808 while (length > 0) {
3809 int opcode = *ptr++;
3815 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3820 if (opsize < 2) /* "silly options" */
3822 if (opsize > length)
3823 return; /* don't parse partial options */
3826 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3827 u16 in_mss = get_unaligned_be16(ptr);
3829 if (opt_rx->user_mss &&
3830 opt_rx->user_mss < in_mss)
3831 in_mss = opt_rx->user_mss;
3832 opt_rx->mss_clamp = in_mss;
3837 if (opsize == TCPOLEN_WINDOW && th->syn &&
3838 !estab && sysctl_tcp_window_scaling) {
3839 __u8 snd_wscale = *(__u8 *)ptr;
3840 opt_rx->wscale_ok = 1;
3841 if (snd_wscale > 14) {
3842 if (net_ratelimit())
3843 printk(KERN_INFO "tcp_parse_options: Illegal window "
3844 "scaling value %d >14 received.\n",
3848 opt_rx->snd_wscale = snd_wscale;
3851 case TCPOPT_TIMESTAMP:
3852 if ((opsize == TCPOLEN_TIMESTAMP) &&
3853 ((estab && opt_rx->tstamp_ok) ||
3854 (!estab && sysctl_tcp_timestamps))) {
3855 opt_rx->saw_tstamp = 1;
3856 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3857 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3860 case TCPOPT_SACK_PERM:
3861 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3862 !estab && sysctl_tcp_sack) {
3863 opt_rx->sack_ok = 1;
3864 tcp_sack_reset(opt_rx);
3869 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3870 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3872 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3875 #ifdef CONFIG_TCP_MD5SIG
3878 * The MD5 Hash has already been
3879 * checked (see tcp_v{4,6}_do_rcv()).
3884 /* This option is variable length.
3887 case TCPOLEN_COOKIE_BASE:
3888 /* not yet implemented */
3890 case TCPOLEN_COOKIE_PAIR:
3891 /* not yet implemented */
3893 case TCPOLEN_COOKIE_MIN+0:
3894 case TCPOLEN_COOKIE_MIN+2:
3895 case TCPOLEN_COOKIE_MIN+4:
3896 case TCPOLEN_COOKIE_MIN+6:
3897 case TCPOLEN_COOKIE_MAX:
3898 /* 16-bit multiple */
3899 opt_rx->cookie_plus = opsize;
3914 EXPORT_SYMBOL(tcp_parse_options);
3916 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3918 __be32 *ptr = (__be32 *)(th + 1);
3920 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3921 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3922 tp->rx_opt.saw_tstamp = 1;
3924 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3926 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3932 /* Fast parse options. This hopes to only see timestamps.
3933 * If it is wrong it falls back on tcp_parse_options().
3935 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3936 struct tcp_sock *tp, u8 **hvpp)
3938 /* In the spirit of fast parsing, compare doff directly to constant
3939 * values. Because equality is used, short doff can be ignored here.
3941 if (th->doff == (sizeof(*th) / 4)) {
3942 tp->rx_opt.saw_tstamp = 0;
3944 } else if (tp->rx_opt.tstamp_ok &&
3945 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3946 if (tcp_parse_aligned_timestamp(tp, th))
3949 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3953 #ifdef CONFIG_TCP_MD5SIG
3955 * Parse MD5 Signature option
3957 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3959 int length = (th->doff << 2) - sizeof (*th);
3960 u8 *ptr = (u8*)(th + 1);
3962 /* If the TCP option is too short, we can short cut */
3963 if (length < TCPOLEN_MD5SIG)
3966 while (length > 0) {
3967 int opcode = *ptr++;
3978 if (opsize < 2 || opsize > length)
3980 if (opcode == TCPOPT_MD5SIG)
3981 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3988 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3991 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3993 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3994 tp->rx_opt.ts_recent_stamp = get_seconds();
3997 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3999 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
4000 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
4001 * extra check below makes sure this can only happen
4002 * for pure ACK frames. -DaveM
4004 * Not only, also it occurs for expired timestamps.
4007 if (tcp_paws_check(&tp->rx_opt, 0))
4008 tcp_store_ts_recent(tp);
4012 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4014 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4015 * it can pass through stack. So, the following predicate verifies that
4016 * this segment is not used for anything but congestion avoidance or
4017 * fast retransmit. Moreover, we even are able to eliminate most of such
4018 * second order effects, if we apply some small "replay" window (~RTO)
4019 * to timestamp space.
4021 * All these measures still do not guarantee that we reject wrapped ACKs
4022 * on networks with high bandwidth, when sequence space is recycled fastly,
4023 * but it guarantees that such events will be very rare and do not affect
4024 * connection seriously. This doesn't look nice, but alas, PAWS is really
4027 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4028 * states that events when retransmit arrives after original data are rare.
4029 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4030 * the biggest problem on large power networks even with minor reordering.
4031 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4032 * up to bandwidth of 18Gigabit/sec. 8) ]
4035 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4037 struct tcp_sock *tp = tcp_sk(sk);
4038 struct tcphdr *th = tcp_hdr(skb);
4039 u32 seq = TCP_SKB_CB(skb)->seq;
4040 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4042 return (/* 1. Pure ACK with correct sequence number. */
4043 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4045 /* 2. ... and duplicate ACK. */
4046 ack == tp->snd_una &&
4048 /* 3. ... and does not update window. */
4049 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4051 /* 4. ... and sits in replay window. */
4052 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4055 static inline int tcp_paws_discard(const struct sock *sk,
4056 const struct sk_buff *skb)
4058 const struct tcp_sock *tp = tcp_sk(sk);
4060 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4061 !tcp_disordered_ack(sk, skb);
4064 /* Check segment sequence number for validity.
4066 * Segment controls are considered valid, if the segment
4067 * fits to the window after truncation to the window. Acceptability
4068 * of data (and SYN, FIN, of course) is checked separately.
4069 * See tcp_data_queue(), for example.
4071 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4072 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4073 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4074 * (borrowed from freebsd)
4077 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
4079 return !before(end_seq, tp->rcv_wup) &&
4080 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4083 /* When we get a reset we do this. */
4084 static void tcp_reset(struct sock *sk)
4086 /* We want the right error as BSD sees it (and indeed as we do). */
4087 switch (sk->sk_state) {
4089 sk->sk_err = ECONNREFUSED;
4091 case TCP_CLOSE_WAIT:
4097 sk->sk_err = ECONNRESET;
4099 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4102 if (!sock_flag(sk, SOCK_DEAD))
4103 sk->sk_error_report(sk);
4109 * Process the FIN bit. This now behaves as it is supposed to work
4110 * and the FIN takes effect when it is validly part of sequence
4111 * space. Not before when we get holes.
4113 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4114 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4117 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4118 * close and we go into CLOSING (and later onto TIME-WAIT)
4120 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4122 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
4124 struct tcp_sock *tp = tcp_sk(sk);
4126 inet_csk_schedule_ack(sk);
4128 sk->sk_shutdown |= RCV_SHUTDOWN;
4129 sock_set_flag(sk, SOCK_DONE);
4131 switch (sk->sk_state) {
4133 case TCP_ESTABLISHED:
4134 /* Move to CLOSE_WAIT */
4135 tcp_set_state(sk, TCP_CLOSE_WAIT);
4136 inet_csk(sk)->icsk_ack.pingpong = 1;
4139 case TCP_CLOSE_WAIT:
4141 /* Received a retransmission of the FIN, do
4146 /* RFC793: Remain in the LAST-ACK state. */
4150 /* This case occurs when a simultaneous close
4151 * happens, we must ack the received FIN and
4152 * enter the CLOSING state.
4155 tcp_set_state(sk, TCP_CLOSING);
4158 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4160 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4163 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4164 * cases we should never reach this piece of code.
4166 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4167 __func__, sk->sk_state);
4171 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4172 * Probably, we should reset in this case. For now drop them.
4174 __skb_queue_purge(&tp->out_of_order_queue);
4175 if (tcp_is_sack(tp))
4176 tcp_sack_reset(&tp->rx_opt);
4179 if (!sock_flag(sk, SOCK_DEAD)) {
4180 sk->sk_state_change(sk);
4182 /* Do not send POLL_HUP for half duplex close. */
4183 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4184 sk->sk_state == TCP_CLOSE)
4185 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4187 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4191 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4194 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4195 if (before(seq, sp->start_seq))
4196 sp->start_seq = seq;
4197 if (after(end_seq, sp->end_seq))
4198 sp->end_seq = end_seq;
4204 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4206 struct tcp_sock *tp = tcp_sk(sk);
4208 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4211 if (before(seq, tp->rcv_nxt))
4212 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4214 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4216 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4218 tp->rx_opt.dsack = 1;
4219 tp->duplicate_sack[0].start_seq = seq;
4220 tp->duplicate_sack[0].end_seq = end_seq;
4224 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4226 struct tcp_sock *tp = tcp_sk(sk);
4228 if (!tp->rx_opt.dsack)
4229 tcp_dsack_set(sk, seq, end_seq);
4231 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4234 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4236 struct tcp_sock *tp = tcp_sk(sk);
4238 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4239 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4240 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4241 tcp_enter_quickack_mode(sk);
4243 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4244 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4246 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4247 end_seq = tp->rcv_nxt;
4248 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4255 /* These routines update the SACK block as out-of-order packets arrive or
4256 * in-order packets close up the sequence space.
4258 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4261 struct tcp_sack_block *sp = &tp->selective_acks[0];
4262 struct tcp_sack_block *swalk = sp + 1;
4264 /* See if the recent change to the first SACK eats into
4265 * or hits the sequence space of other SACK blocks, if so coalesce.
4267 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4268 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4271 /* Zap SWALK, by moving every further SACK up by one slot.
4272 * Decrease num_sacks.
4274 tp->rx_opt.num_sacks--;
4275 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4279 this_sack++, swalk++;
4283 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4285 struct tcp_sock *tp = tcp_sk(sk);
4286 struct tcp_sack_block *sp = &tp->selective_acks[0];
4287 int cur_sacks = tp->rx_opt.num_sacks;
4293 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4294 if (tcp_sack_extend(sp, seq, end_seq)) {
4295 /* Rotate this_sack to the first one. */
4296 for (; this_sack > 0; this_sack--, sp--)
4297 swap(*sp, *(sp - 1));
4299 tcp_sack_maybe_coalesce(tp);
4304 /* Could not find an adjacent existing SACK, build a new one,
4305 * put it at the front, and shift everyone else down. We
4306 * always know there is at least one SACK present already here.
4308 * If the sack array is full, forget about the last one.
4310 if (this_sack >= TCP_NUM_SACKS) {
4312 tp->rx_opt.num_sacks--;
4315 for (; this_sack > 0; this_sack--, sp--)
4319 /* Build the new head SACK, and we're done. */
4320 sp->start_seq = seq;
4321 sp->end_seq = end_seq;
4322 tp->rx_opt.num_sacks++;
4325 /* RCV.NXT advances, some SACKs should be eaten. */
4327 static void tcp_sack_remove(struct tcp_sock *tp)
4329 struct tcp_sack_block *sp = &tp->selective_acks[0];
4330 int num_sacks = tp->rx_opt.num_sacks;
4333 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4334 if (skb_queue_empty(&tp->out_of_order_queue)) {
4335 tp->rx_opt.num_sacks = 0;
4339 for (this_sack = 0; this_sack < num_sacks;) {
4340 /* Check if the start of the sack is covered by RCV.NXT. */
4341 if (!before(tp->rcv_nxt, sp->start_seq)) {
4344 /* RCV.NXT must cover all the block! */
4345 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4347 /* Zap this SACK, by moving forward any other SACKS. */
4348 for (i=this_sack+1; i < num_sacks; i++)
4349 tp->selective_acks[i-1] = tp->selective_acks[i];
4356 tp->rx_opt.num_sacks = num_sacks;
4359 /* This one checks to see if we can put data from the
4360 * out_of_order queue into the receive_queue.
4362 static void tcp_ofo_queue(struct sock *sk)
4364 struct tcp_sock *tp = tcp_sk(sk);
4365 __u32 dsack_high = tp->rcv_nxt;
4366 struct sk_buff *skb;
4368 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4369 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4372 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4373 __u32 dsack = dsack_high;
4374 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4375 dsack_high = TCP_SKB_CB(skb)->end_seq;
4376 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4379 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4380 SOCK_DEBUG(sk, "ofo packet was already received\n");
4381 __skb_unlink(skb, &tp->out_of_order_queue);
4385 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4386 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4387 TCP_SKB_CB(skb)->end_seq);
4389 __skb_unlink(skb, &tp->out_of_order_queue);
4390 __skb_queue_tail(&sk->sk_receive_queue, skb);
4391 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4392 if (tcp_hdr(skb)->fin)
4393 tcp_fin(skb, sk, tcp_hdr(skb));
4397 static int tcp_prune_ofo_queue(struct sock *sk);
4398 static int tcp_prune_queue(struct sock *sk);
4400 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4402 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4403 !sk_rmem_schedule(sk, size)) {
4405 if (tcp_prune_queue(sk) < 0)
4408 if (!sk_rmem_schedule(sk, size)) {
4409 if (!tcp_prune_ofo_queue(sk))
4412 if (!sk_rmem_schedule(sk, size))
4419 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4421 struct tcphdr *th = tcp_hdr(skb);
4422 struct tcp_sock *tp = tcp_sk(sk);
4425 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4429 __skb_pull(skb, th->doff * 4);
4431 TCP_ECN_accept_cwr(tp, skb);
4433 tp->rx_opt.dsack = 0;
4435 /* Queue data for delivery to the user.
4436 * Packets in sequence go to the receive queue.
4437 * Out of sequence packets to the out_of_order_queue.
4439 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4440 if (tcp_receive_window(tp) == 0)
4443 /* Ok. In sequence. In window. */
4444 if (tp->ucopy.task == current &&
4445 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4446 sock_owned_by_user(sk) && !tp->urg_data) {
4447 int chunk = min_t(unsigned int, skb->len,
4450 __set_current_state(TASK_RUNNING);
4453 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4454 tp->ucopy.len -= chunk;
4455 tp->copied_seq += chunk;
4456 eaten = (chunk == skb->len);
4457 tcp_rcv_space_adjust(sk);
4465 tcp_try_rmem_schedule(sk, skb->truesize))
4468 skb_set_owner_r(skb, sk);
4469 __skb_queue_tail(&sk->sk_receive_queue, skb);
4471 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4473 tcp_event_data_recv(sk, skb);
4475 tcp_fin(skb, sk, th);
4477 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4480 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4481 * gap in queue is filled.
4483 if (skb_queue_empty(&tp->out_of_order_queue))
4484 inet_csk(sk)->icsk_ack.pingpong = 0;
4487 if (tp->rx_opt.num_sacks)
4488 tcp_sack_remove(tp);
4490 tcp_fast_path_check(sk);
4494 else if (!sock_flag(sk, SOCK_DEAD))
4495 sk->sk_data_ready(sk, 0);
4499 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4500 /* A retransmit, 2nd most common case. Force an immediate ack. */
4501 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4502 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4505 tcp_enter_quickack_mode(sk);
4506 inet_csk_schedule_ack(sk);
4512 /* Out of window. F.e. zero window probe. */
4513 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4516 tcp_enter_quickack_mode(sk);
4518 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4519 /* Partial packet, seq < rcv_next < end_seq */
4520 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4521 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4522 TCP_SKB_CB(skb)->end_seq);
4524 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4526 /* If window is closed, drop tail of packet. But after
4527 * remembering D-SACK for its head made in previous line.
4529 if (!tcp_receive_window(tp))
4534 TCP_ECN_check_ce(tp, skb);
4536 if (tcp_try_rmem_schedule(sk, skb->truesize))
4539 /* Disable header prediction. */
4541 inet_csk_schedule_ack(sk);
4543 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4544 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4546 skb_set_owner_r(skb, sk);
4548 if (!skb_peek(&tp->out_of_order_queue)) {
4549 /* Initial out of order segment, build 1 SACK. */
4550 if (tcp_is_sack(tp)) {
4551 tp->rx_opt.num_sacks = 1;
4552 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4553 tp->selective_acks[0].end_seq =
4554 TCP_SKB_CB(skb)->end_seq;
4556 __skb_queue_head(&tp->out_of_order_queue, skb);
4558 struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue);
4559 u32 seq = TCP_SKB_CB(skb)->seq;
4560 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4562 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4563 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4565 if (!tp->rx_opt.num_sacks ||
4566 tp->selective_acks[0].end_seq != seq)
4569 /* Common case: data arrive in order after hole. */
4570 tp->selective_acks[0].end_seq = end_seq;
4574 /* Find place to insert this segment. */
4576 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4578 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4582 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4585 /* Do skb overlap to previous one? */
4586 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4587 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4588 /* All the bits are present. Drop. */
4590 tcp_dsack_set(sk, seq, end_seq);
4593 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4594 /* Partial overlap. */
4595 tcp_dsack_set(sk, seq,
4596 TCP_SKB_CB(skb1)->end_seq);
4598 if (skb_queue_is_first(&tp->out_of_order_queue,
4602 skb1 = skb_queue_prev(
4603 &tp->out_of_order_queue,
4608 __skb_queue_head(&tp->out_of_order_queue, skb);
4610 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4612 /* And clean segments covered by new one as whole. */
4613 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4614 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4616 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4618 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4619 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4623 __skb_unlink(skb1, &tp->out_of_order_queue);
4624 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4625 TCP_SKB_CB(skb1)->end_seq);
4630 if (tcp_is_sack(tp))
4631 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4635 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4636 struct sk_buff_head *list)
4638 struct sk_buff *next = NULL;
4640 if (!skb_queue_is_last(list, skb))
4641 next = skb_queue_next(list, skb);
4643 __skb_unlink(skb, list);
4645 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4650 /* Collapse contiguous sequence of skbs head..tail with
4651 * sequence numbers start..end.
4653 * If tail is NULL, this means until the end of the list.
4655 * Segments with FIN/SYN are not collapsed (only because this
4659 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4660 struct sk_buff *head, struct sk_buff *tail,
4663 struct sk_buff *skb, *n;
4666 /* First, check that queue is collapsible and find
4667 * the point where collapsing can be useful. */
4671 skb_queue_walk_from_safe(list, skb, n) {
4674 /* No new bits? It is possible on ofo queue. */
4675 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4676 skb = tcp_collapse_one(sk, skb, list);
4682 /* The first skb to collapse is:
4684 * - bloated or contains data before "start" or
4685 * overlaps to the next one.
4687 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4688 (tcp_win_from_space(skb->truesize) > skb->len ||
4689 before(TCP_SKB_CB(skb)->seq, start))) {
4690 end_of_skbs = false;
4694 if (!skb_queue_is_last(list, skb)) {
4695 struct sk_buff *next = skb_queue_next(list, skb);
4697 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4698 end_of_skbs = false;
4703 /* Decided to skip this, advance start seq. */
4704 start = TCP_SKB_CB(skb)->end_seq;
4706 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4709 while (before(start, end)) {
4710 struct sk_buff *nskb;
4711 unsigned int header = skb_headroom(skb);
4712 int copy = SKB_MAX_ORDER(header, 0);
4714 /* Too big header? This can happen with IPv6. */
4717 if (end - start < copy)
4719 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4723 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4724 skb_set_network_header(nskb, (skb_network_header(skb) -
4726 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4728 skb_reserve(nskb, header);
4729 memcpy(nskb->head, skb->head, header);
4730 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4731 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4732 __skb_queue_before(list, skb, nskb);
4733 skb_set_owner_r(nskb, sk);
4735 /* Copy data, releasing collapsed skbs. */
4737 int offset = start - TCP_SKB_CB(skb)->seq;
4738 int size = TCP_SKB_CB(skb)->end_seq - start;
4742 size = min(copy, size);
4743 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4745 TCP_SKB_CB(nskb)->end_seq += size;
4749 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4750 skb = tcp_collapse_one(sk, skb, list);
4753 tcp_hdr(skb)->syn ||
4761 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4762 * and tcp_collapse() them until all the queue is collapsed.
4764 static void tcp_collapse_ofo_queue(struct sock *sk)
4766 struct tcp_sock *tp = tcp_sk(sk);
4767 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4768 struct sk_buff *head;
4774 start = TCP_SKB_CB(skb)->seq;
4775 end = TCP_SKB_CB(skb)->end_seq;
4779 struct sk_buff *next = NULL;
4781 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4782 next = skb_queue_next(&tp->out_of_order_queue, skb);
4785 /* Segment is terminated when we see gap or when
4786 * we are at the end of all the queue. */
4788 after(TCP_SKB_CB(skb)->seq, end) ||
4789 before(TCP_SKB_CB(skb)->end_seq, start)) {
4790 tcp_collapse(sk, &tp->out_of_order_queue,
4791 head, skb, start, end);
4795 /* Start new segment */
4796 start = TCP_SKB_CB(skb)->seq;
4797 end = TCP_SKB_CB(skb)->end_seq;
4799 if (before(TCP_SKB_CB(skb)->seq, start))
4800 start = TCP_SKB_CB(skb)->seq;
4801 if (after(TCP_SKB_CB(skb)->end_seq, end))
4802 end = TCP_SKB_CB(skb)->end_seq;
4808 * Purge the out-of-order queue.
4809 * Return true if queue was pruned.
4811 static int tcp_prune_ofo_queue(struct sock *sk)
4813 struct tcp_sock *tp = tcp_sk(sk);
4816 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4817 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4818 __skb_queue_purge(&tp->out_of_order_queue);
4820 /* Reset SACK state. A conforming SACK implementation will
4821 * do the same at a timeout based retransmit. When a connection
4822 * is in a sad state like this, we care only about integrity
4823 * of the connection not performance.
4825 if (tp->rx_opt.sack_ok)
4826 tcp_sack_reset(&tp->rx_opt);
4833 /* Reduce allocated memory if we can, trying to get
4834 * the socket within its memory limits again.
4836 * Return less than zero if we should start dropping frames
4837 * until the socket owning process reads some of the data
4838 * to stabilize the situation.
4840 static int tcp_prune_queue(struct sock *sk)
4842 struct tcp_sock *tp = tcp_sk(sk);
4844 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4846 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4848 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4849 tcp_clamp_window(sk);
4850 else if (tcp_memory_pressure)
4851 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4853 tcp_collapse_ofo_queue(sk);
4854 if (!skb_queue_empty(&sk->sk_receive_queue))
4855 tcp_collapse(sk, &sk->sk_receive_queue,
4856 skb_peek(&sk->sk_receive_queue),
4858 tp->copied_seq, tp->rcv_nxt);
4861 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4864 /* Collapsing did not help, destructive actions follow.
4865 * This must not ever occur. */
4867 tcp_prune_ofo_queue(sk);
4869 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4872 /* If we are really being abused, tell the caller to silently
4873 * drop receive data on the floor. It will get retransmitted
4874 * and hopefully then we'll have sufficient space.
4876 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4878 /* Massive buffer overcommit. */
4883 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4884 * As additional protections, we do not touch cwnd in retransmission phases,
4885 * and if application hit its sndbuf limit recently.
4887 void tcp_cwnd_application_limited(struct sock *sk)
4889 struct tcp_sock *tp = tcp_sk(sk);
4891 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4892 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4893 /* Limited by application or receiver window. */
4894 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4895 u32 win_used = max(tp->snd_cwnd_used, init_win);
4896 if (win_used < tp->snd_cwnd) {
4897 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4898 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4900 tp->snd_cwnd_used = 0;
4902 tp->snd_cwnd_stamp = tcp_time_stamp;
4905 static int tcp_should_expand_sndbuf(struct sock *sk)
4907 struct tcp_sock *tp = tcp_sk(sk);
4909 /* If the user specified a specific send buffer setting, do
4912 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4915 /* If we are under global TCP memory pressure, do not expand. */
4916 if (tcp_memory_pressure)
4919 /* If we are under soft global TCP memory pressure, do not expand. */
4920 if (atomic_long_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4923 /* If we filled the congestion window, do not expand. */
4924 if (tp->packets_out >= tp->snd_cwnd)
4930 /* When incoming ACK allowed to free some skb from write_queue,
4931 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4932 * on the exit from tcp input handler.
4934 * PROBLEM: sndbuf expansion does not work well with largesend.
4936 static void tcp_new_space(struct sock *sk)
4938 struct tcp_sock *tp = tcp_sk(sk);
4940 if (tcp_should_expand_sndbuf(sk)) {
4941 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4942 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4943 int demanded = max_t(unsigned int, tp->snd_cwnd,
4944 tp->reordering + 1);
4945 sndmem *= 2 * demanded;
4946 if (sndmem > sk->sk_sndbuf)
4947 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4948 tp->snd_cwnd_stamp = tcp_time_stamp;
4951 sk->sk_write_space(sk);
4954 static void tcp_check_space(struct sock *sk)
4956 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4957 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4958 if (sk->sk_socket &&
4959 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4964 static inline void tcp_data_snd_check(struct sock *sk)
4966 tcp_push_pending_frames(sk);
4967 tcp_check_space(sk);
4971 * Check if sending an ack is needed.
4973 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4975 struct tcp_sock *tp = tcp_sk(sk);
4977 /* More than one full frame received... */
4978 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4979 /* ... and right edge of window advances far enough.
4980 * (tcp_recvmsg() will send ACK otherwise). Or...
4982 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4983 /* We ACK each frame or... */
4984 tcp_in_quickack_mode(sk) ||
4985 /* We have out of order data. */
4986 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4987 /* Then ack it now */
4990 /* Else, send delayed ack. */
4991 tcp_send_delayed_ack(sk);
4995 static inline void tcp_ack_snd_check(struct sock *sk)
4997 if (!inet_csk_ack_scheduled(sk)) {
4998 /* We sent a data segment already. */
5001 __tcp_ack_snd_check(sk, 1);
5005 * This routine is only called when we have urgent data
5006 * signaled. Its the 'slow' part of tcp_urg. It could be
5007 * moved inline now as tcp_urg is only called from one
5008 * place. We handle URGent data wrong. We have to - as
5009 * BSD still doesn't use the correction from RFC961.
5010 * For 1003.1g we should support a new option TCP_STDURG to permit
5011 * either form (or just set the sysctl tcp_stdurg).
5014 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
5016 struct tcp_sock *tp = tcp_sk(sk);
5017 u32 ptr = ntohs(th->urg_ptr);
5019 if (ptr && !sysctl_tcp_stdurg)
5021 ptr += ntohl(th->seq);
5023 /* Ignore urgent data that we've already seen and read. */
5024 if (after(tp->copied_seq, ptr))
5027 /* Do not replay urg ptr.
5029 * NOTE: interesting situation not covered by specs.
5030 * Misbehaving sender may send urg ptr, pointing to segment,
5031 * which we already have in ofo queue. We are not able to fetch
5032 * such data and will stay in TCP_URG_NOTYET until will be eaten
5033 * by recvmsg(). Seems, we are not obliged to handle such wicked
5034 * situations. But it is worth to think about possibility of some
5035 * DoSes using some hypothetical application level deadlock.
5037 if (before(ptr, tp->rcv_nxt))
5040 /* Do we already have a newer (or duplicate) urgent pointer? */
5041 if (tp->urg_data && !after(ptr, tp->urg_seq))
5044 /* Tell the world about our new urgent pointer. */
5047 /* We may be adding urgent data when the last byte read was
5048 * urgent. To do this requires some care. We cannot just ignore
5049 * tp->copied_seq since we would read the last urgent byte again
5050 * as data, nor can we alter copied_seq until this data arrives
5051 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5053 * NOTE. Double Dutch. Rendering to plain English: author of comment
5054 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5055 * and expect that both A and B disappear from stream. This is _wrong_.
5056 * Though this happens in BSD with high probability, this is occasional.
5057 * Any application relying on this is buggy. Note also, that fix "works"
5058 * only in this artificial test. Insert some normal data between A and B and we will
5059 * decline of BSD again. Verdict: it is better to remove to trap
5062 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5063 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5064 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5066 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5067 __skb_unlink(skb, &sk->sk_receive_queue);
5072 tp->urg_data = TCP_URG_NOTYET;
5075 /* Disable header prediction. */
5079 /* This is the 'fast' part of urgent handling. */
5080 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
5082 struct tcp_sock *tp = tcp_sk(sk);
5084 /* Check if we get a new urgent pointer - normally not. */
5086 tcp_check_urg(sk, th);
5088 /* Do we wait for any urgent data? - normally not... */
5089 if (tp->urg_data == TCP_URG_NOTYET) {
5090 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5093 /* Is the urgent pointer pointing into this packet? */
5094 if (ptr < skb->len) {
5096 if (skb_copy_bits(skb, ptr, &tmp, 1))
5098 tp->urg_data = TCP_URG_VALID | tmp;
5099 if (!sock_flag(sk, SOCK_DEAD))
5100 sk->sk_data_ready(sk, 0);
5105 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5107 struct tcp_sock *tp = tcp_sk(sk);
5108 int chunk = skb->len - hlen;
5112 if (skb_csum_unnecessary(skb))
5113 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5115 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5119 tp->ucopy.len -= chunk;
5120 tp->copied_seq += chunk;
5121 tcp_rcv_space_adjust(sk);
5128 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5129 struct sk_buff *skb)
5133 if (sock_owned_by_user(sk)) {
5135 result = __tcp_checksum_complete(skb);
5138 result = __tcp_checksum_complete(skb);
5143 static inline int tcp_checksum_complete_user(struct sock *sk,
5144 struct sk_buff *skb)
5146 return !skb_csum_unnecessary(skb) &&
5147 __tcp_checksum_complete_user(sk, skb);
5150 #ifdef CONFIG_NET_DMA
5151 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5154 struct tcp_sock *tp = tcp_sk(sk);
5155 int chunk = skb->len - hlen;
5157 int copied_early = 0;
5159 if (tp->ucopy.wakeup)
5162 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5163 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5165 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5167 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5169 tp->ucopy.iov, chunk,
5170 tp->ucopy.pinned_list);
5175 tp->ucopy.dma_cookie = dma_cookie;
5178 tp->ucopy.len -= chunk;
5179 tp->copied_seq += chunk;
5180 tcp_rcv_space_adjust(sk);
5182 if ((tp->ucopy.len == 0) ||
5183 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5184 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5185 tp->ucopy.wakeup = 1;
5186 sk->sk_data_ready(sk, 0);
5188 } else if (chunk > 0) {
5189 tp->ucopy.wakeup = 1;
5190 sk->sk_data_ready(sk, 0);
5193 return copied_early;
5195 #endif /* CONFIG_NET_DMA */
5197 /* Does PAWS and seqno based validation of an incoming segment, flags will
5198 * play significant role here.
5200 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5201 struct tcphdr *th, int syn_inerr)
5204 struct tcp_sock *tp = tcp_sk(sk);
5206 /* RFC1323: H1. Apply PAWS check first. */
5207 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5208 tp->rx_opt.saw_tstamp &&
5209 tcp_paws_discard(sk, skb)) {
5211 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5212 tcp_send_dupack(sk, skb);
5215 /* Reset is accepted even if it did not pass PAWS. */
5218 /* Step 1: check sequence number */
5219 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5220 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5221 * (RST) segments are validated by checking their SEQ-fields."
5222 * And page 69: "If an incoming segment is not acceptable,
5223 * an acknowledgment should be sent in reply (unless the RST
5224 * bit is set, if so drop the segment and return)".
5229 tcp_send_dupack(sk, skb);
5234 /* Step 2: check RST bit */
5237 * If sequence number exactly matches RCV.NXT, then
5238 * RESET the connection
5240 * Send a challenge ACK
5242 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5245 tcp_send_challenge_ack(sk);
5249 /* step 3: check security and precedence [ignored] */
5251 /* step 4: Check for a SYN
5252 * RFC 5691 4.2 : Send a challenge ack
5257 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5258 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5259 tcp_send_challenge_ack(sk);
5271 * TCP receive function for the ESTABLISHED state.
5273 * It is split into a fast path and a slow path. The fast path is
5275 * - A zero window was announced from us - zero window probing
5276 * is only handled properly in the slow path.
5277 * - Out of order segments arrived.
5278 * - Urgent data is expected.
5279 * - There is no buffer space left
5280 * - Unexpected TCP flags/window values/header lengths are received
5281 * (detected by checking the TCP header against pred_flags)
5282 * - Data is sent in both directions. Fast path only supports pure senders
5283 * or pure receivers (this means either the sequence number or the ack
5284 * value must stay constant)
5285 * - Unexpected TCP option.
5287 * When these conditions are not satisfied it drops into a standard
5288 * receive procedure patterned after RFC793 to handle all cases.
5289 * The first three cases are guaranteed by proper pred_flags setting,
5290 * the rest is checked inline. Fast processing is turned on in
5291 * tcp_data_queue when everything is OK.
5293 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5294 struct tcphdr *th, unsigned len)
5296 struct tcp_sock *tp = tcp_sk(sk);
5299 * Header prediction.
5300 * The code loosely follows the one in the famous
5301 * "30 instruction TCP receive" Van Jacobson mail.
5303 * Van's trick is to deposit buffers into socket queue
5304 * on a device interrupt, to call tcp_recv function
5305 * on the receive process context and checksum and copy
5306 * the buffer to user space. smart...
5308 * Our current scheme is not silly either but we take the
5309 * extra cost of the net_bh soft interrupt processing...
5310 * We do checksum and copy also but from device to kernel.
5313 tp->rx_opt.saw_tstamp = 0;
5315 /* pred_flags is 0xS?10 << 16 + snd_wnd
5316 * if header_prediction is to be made
5317 * 'S' will always be tp->tcp_header_len >> 2
5318 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5319 * turn it off (when there are holes in the receive
5320 * space for instance)
5321 * PSH flag is ignored.
5324 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5325 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5326 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5327 int tcp_header_len = tp->tcp_header_len;
5329 /* Timestamp header prediction: tcp_header_len
5330 * is automatically equal to th->doff*4 due to pred_flags
5334 /* Check timestamp */
5335 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5336 /* No? Slow path! */
5337 if (!tcp_parse_aligned_timestamp(tp, th))
5340 /* If PAWS failed, check it more carefully in slow path */
5341 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5344 /* DO NOT update ts_recent here, if checksum fails
5345 * and timestamp was corrupted part, it will result
5346 * in a hung connection since we will drop all
5347 * future packets due to the PAWS test.
5351 if (len <= tcp_header_len) {
5352 /* Bulk data transfer: sender */
5353 if (len == tcp_header_len) {
5354 /* Predicted packet is in window by definition.
5355 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5356 * Hence, check seq<=rcv_wup reduces to:
5358 if (tcp_header_len ==
5359 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5360 tp->rcv_nxt == tp->rcv_wup)
5361 tcp_store_ts_recent(tp);
5363 /* We know that such packets are checksummed
5366 tcp_ack(sk, skb, 0);
5368 tcp_data_snd_check(sk);
5370 } else { /* Header too small */
5371 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5376 int copied_early = 0;
5378 if (tp->copied_seq == tp->rcv_nxt &&
5379 len - tcp_header_len <= tp->ucopy.len) {
5380 #ifdef CONFIG_NET_DMA
5381 if (tp->ucopy.task == current &&
5382 sock_owned_by_user(sk) &&
5383 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5388 if (tp->ucopy.task == current &&
5389 sock_owned_by_user(sk) && !copied_early) {
5390 __set_current_state(TASK_RUNNING);
5392 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5396 /* Predicted packet is in window by definition.
5397 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5398 * Hence, check seq<=rcv_wup reduces to:
5400 if (tcp_header_len ==
5401 (sizeof(struct tcphdr) +
5402 TCPOLEN_TSTAMP_ALIGNED) &&
5403 tp->rcv_nxt == tp->rcv_wup)
5404 tcp_store_ts_recent(tp);
5406 tcp_rcv_rtt_measure_ts(sk, skb);
5408 __skb_pull(skb, tcp_header_len);
5409 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5410 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5413 tcp_cleanup_rbuf(sk, skb->len);
5416 if (tcp_checksum_complete_user(sk, skb))
5419 if ((int)skb->truesize > sk->sk_forward_alloc)
5422 /* Predicted packet is in window by definition.
5423 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5424 * Hence, check seq<=rcv_wup reduces to:
5426 if (tcp_header_len ==
5427 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5428 tp->rcv_nxt == tp->rcv_wup)
5429 tcp_store_ts_recent(tp);
5431 tcp_rcv_rtt_measure_ts(sk, skb);
5433 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5435 /* Bulk data transfer: receiver */
5436 __skb_pull(skb, tcp_header_len);
5437 __skb_queue_tail(&sk->sk_receive_queue, skb);
5438 skb_set_owner_r(skb, sk);
5439 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5442 tcp_event_data_recv(sk, skb);
5444 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5445 /* Well, only one small jumplet in fast path... */
5446 tcp_ack(sk, skb, FLAG_DATA);
5447 tcp_data_snd_check(sk);
5448 if (!inet_csk_ack_scheduled(sk))
5452 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5453 __tcp_ack_snd_check(sk, 0);
5455 #ifdef CONFIG_NET_DMA
5457 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5463 sk->sk_data_ready(sk, 0);
5469 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5473 * Standard slow path.
5476 if (!tcp_validate_incoming(sk, skb, th, 1))
5480 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5483 /* ts_recent update must be made after we are sure that the packet
5486 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5488 tcp_rcv_rtt_measure_ts(sk, skb);
5490 /* Process urgent data. */
5491 tcp_urg(sk, skb, th);
5493 /* step 7: process the segment text */
5494 tcp_data_queue(sk, skb);
5496 tcp_data_snd_check(sk);
5497 tcp_ack_snd_check(sk);
5501 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5507 EXPORT_SYMBOL(tcp_rcv_established);
5509 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5510 struct tcphdr *th, unsigned len)
5513 struct inet_connection_sock *icsk = inet_csk(sk);
5514 struct tcp_sock *tp = tcp_sk(sk);
5515 struct tcp_cookie_values *cvp = tp->cookie_values;
5516 int saved_clamp = tp->rx_opt.mss_clamp;
5518 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5522 * "If the state is SYN-SENT then
5523 * first check the ACK bit
5524 * If the ACK bit is set
5525 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5526 * a reset (unless the RST bit is set, if so drop
5527 * the segment and return)"
5529 * We do not send data with SYN, so that RFC-correct
5532 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5533 goto reset_and_undo;
5535 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5536 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5538 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5539 goto reset_and_undo;
5542 /* Now ACK is acceptable.
5544 * "If the RST bit is set
5545 * If the ACK was acceptable then signal the user "error:
5546 * connection reset", drop the segment, enter CLOSED state,
5547 * delete TCB, and return."
5556 * "fifth, if neither of the SYN or RST bits is set then
5557 * drop the segment and return."
5563 goto discard_and_undo;
5566 * "If the SYN bit is on ...
5567 * are acceptable then ...
5568 * (our SYN has been ACKed), change the connection
5569 * state to ESTABLISHED..."
5572 TCP_ECN_rcv_synack(tp, th);
5574 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5575 tcp_ack(sk, skb, FLAG_SLOWPATH);
5577 /* Ok.. it's good. Set up sequence numbers and
5578 * move to established.
5580 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5581 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5583 /* RFC1323: The window in SYN & SYN/ACK segments is
5586 tp->snd_wnd = ntohs(th->window);
5587 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5589 if (!tp->rx_opt.wscale_ok) {
5590 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5591 tp->window_clamp = min(tp->window_clamp, 65535U);
5594 if (tp->rx_opt.saw_tstamp) {
5595 tp->rx_opt.tstamp_ok = 1;
5596 tp->tcp_header_len =
5597 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5598 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5599 tcp_store_ts_recent(tp);
5601 tp->tcp_header_len = sizeof(struct tcphdr);
5604 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5605 tcp_enable_fack(tp);
5608 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5609 tcp_initialize_rcv_mss(sk);
5611 /* Remember, tcp_poll() does not lock socket!
5612 * Change state from SYN-SENT only after copied_seq
5613 * is initialized. */
5614 tp->copied_seq = tp->rcv_nxt;
5617 cvp->cookie_pair_size > 0 &&
5618 tp->rx_opt.cookie_plus > 0) {
5619 int cookie_size = tp->rx_opt.cookie_plus
5620 - TCPOLEN_COOKIE_BASE;
5621 int cookie_pair_size = cookie_size
5622 + cvp->cookie_desired;
5624 /* A cookie extension option was sent and returned.
5625 * Note that each incoming SYNACK replaces the
5626 * Responder cookie. The initial exchange is most
5627 * fragile, as protection against spoofing relies
5628 * entirely upon the sequence and timestamp (above).
5629 * This replacement strategy allows the correct pair to
5630 * pass through, while any others will be filtered via
5631 * Responder verification later.
5633 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5634 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5635 hash_location, cookie_size);
5636 cvp->cookie_pair_size = cookie_pair_size;
5641 tcp_set_state(sk, TCP_ESTABLISHED);
5643 security_inet_conn_established(sk, skb);
5645 /* Make sure socket is routed, for correct metrics. */
5646 icsk->icsk_af_ops->rebuild_header(sk);
5648 tcp_init_metrics(sk);
5650 tcp_init_congestion_control(sk);
5652 /* Prevent spurious tcp_cwnd_restart() on first data
5655 tp->lsndtime = tcp_time_stamp;
5657 tcp_init_buffer_space(sk);
5659 if (sock_flag(sk, SOCK_KEEPOPEN))
5660 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5662 if (!tp->rx_opt.snd_wscale)
5663 __tcp_fast_path_on(tp, tp->snd_wnd);
5667 if (!sock_flag(sk, SOCK_DEAD)) {
5668 sk->sk_state_change(sk);
5669 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5672 if (sk->sk_write_pending ||
5673 icsk->icsk_accept_queue.rskq_defer_accept ||
5674 icsk->icsk_ack.pingpong) {
5675 /* Save one ACK. Data will be ready after
5676 * several ticks, if write_pending is set.
5678 * It may be deleted, but with this feature tcpdumps
5679 * look so _wonderfully_ clever, that I was not able
5680 * to stand against the temptation 8) --ANK
5682 inet_csk_schedule_ack(sk);
5683 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5684 icsk->icsk_ack.ato = TCP_ATO_MIN;
5685 tcp_incr_quickack(sk);
5686 tcp_enter_quickack_mode(sk);
5687 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5688 TCP_DELACK_MAX, TCP_RTO_MAX);
5699 /* No ACK in the segment */
5703 * "If the RST bit is set
5705 * Otherwise (no ACK) drop the segment and return."
5708 goto discard_and_undo;
5712 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5713 tcp_paws_reject(&tp->rx_opt, 0))
5714 goto discard_and_undo;
5717 /* We see SYN without ACK. It is attempt of
5718 * simultaneous connect with crossed SYNs.
5719 * Particularly, it can be connect to self.
5721 tcp_set_state(sk, TCP_SYN_RECV);
5723 if (tp->rx_opt.saw_tstamp) {
5724 tp->rx_opt.tstamp_ok = 1;
5725 tcp_store_ts_recent(tp);
5726 tp->tcp_header_len =
5727 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5729 tp->tcp_header_len = sizeof(struct tcphdr);
5732 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5733 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5735 /* RFC1323: The window in SYN & SYN/ACK segments is
5738 tp->snd_wnd = ntohs(th->window);
5739 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5740 tp->max_window = tp->snd_wnd;
5742 TCP_ECN_rcv_syn(tp, th);
5745 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5746 tcp_initialize_rcv_mss(sk);
5748 tcp_send_synack(sk);
5750 /* Note, we could accept data and URG from this segment.
5751 * There are no obstacles to make this.
5753 * However, if we ignore data in ACKless segments sometimes,
5754 * we have no reasons to accept it sometimes.
5755 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5756 * is not flawless. So, discard packet for sanity.
5757 * Uncomment this return to process the data.
5764 /* "fifth, if neither of the SYN or RST bits is set then
5765 * drop the segment and return."
5769 tcp_clear_options(&tp->rx_opt);
5770 tp->rx_opt.mss_clamp = saved_clamp;
5774 tcp_clear_options(&tp->rx_opt);
5775 tp->rx_opt.mss_clamp = saved_clamp;
5780 * This function implements the receiving procedure of RFC 793 for
5781 * all states except ESTABLISHED and TIME_WAIT.
5782 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5783 * address independent.
5786 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5787 struct tcphdr *th, unsigned len)
5789 struct tcp_sock *tp = tcp_sk(sk);
5790 struct inet_connection_sock *icsk = inet_csk(sk);
5793 tp->rx_opt.saw_tstamp = 0;
5795 switch (sk->sk_state) {
5809 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5812 /* Now we have several options: In theory there is
5813 * nothing else in the frame. KA9Q has an option to
5814 * send data with the syn, BSD accepts data with the
5815 * syn up to the [to be] advertised window and
5816 * Solaris 2.1 gives you a protocol error. For now
5817 * we just ignore it, that fits the spec precisely
5818 * and avoids incompatibilities. It would be nice in
5819 * future to drop through and process the data.
5821 * Now that TTCP is starting to be used we ought to
5823 * But, this leaves one open to an easy denial of
5824 * service attack, and SYN cookies can't defend
5825 * against this problem. So, we drop the data
5826 * in the interest of security over speed unless
5827 * it's still in use.
5835 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5839 /* Do step6 onward by hand. */
5840 tcp_urg(sk, skb, th);
5842 tcp_data_snd_check(sk);
5846 if (!tcp_validate_incoming(sk, skb, th, 0))
5849 /* step 5: check the ACK field */
5851 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5853 switch (sk->sk_state) {
5856 tp->copied_seq = tp->rcv_nxt;
5858 tcp_set_state(sk, TCP_ESTABLISHED);
5859 sk->sk_state_change(sk);
5861 /* Note, that this wakeup is only for marginal
5862 * crossed SYN case. Passively open sockets
5863 * are not waked up, because sk->sk_sleep ==
5864 * NULL and sk->sk_socket == NULL.
5868 SOCK_WAKE_IO, POLL_OUT);
5870 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5871 tp->snd_wnd = ntohs(th->window) <<
5872 tp->rx_opt.snd_wscale;
5873 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5875 /* tcp_ack considers this ACK as duplicate
5876 * and does not calculate rtt.
5879 tcp_ack_update_rtt(sk, 0, 0);
5881 if (tp->rx_opt.tstamp_ok)
5882 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5884 /* Make sure socket is routed, for
5887 icsk->icsk_af_ops->rebuild_header(sk);
5889 tcp_init_metrics(sk);
5891 tcp_init_congestion_control(sk);
5893 /* Prevent spurious tcp_cwnd_restart() on
5894 * first data packet.
5896 tp->lsndtime = tcp_time_stamp;
5899 tcp_initialize_rcv_mss(sk);
5900 tcp_init_buffer_space(sk);
5901 tcp_fast_path_on(tp);
5908 if (tp->snd_una == tp->write_seq) {
5909 tcp_set_state(sk, TCP_FIN_WAIT2);
5910 sk->sk_shutdown |= SEND_SHUTDOWN;
5911 dst_confirm(__sk_dst_get(sk));
5913 if (!sock_flag(sk, SOCK_DEAD))
5914 /* Wake up lingering close() */
5915 sk->sk_state_change(sk);
5919 if (tp->linger2 < 0 ||
5920 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5921 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5923 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5927 tmo = tcp_fin_time(sk);
5928 if (tmo > TCP_TIMEWAIT_LEN) {
5929 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5930 } else if (th->fin || sock_owned_by_user(sk)) {
5931 /* Bad case. We could lose such FIN otherwise.
5932 * It is not a big problem, but it looks confusing
5933 * and not so rare event. We still can lose it now,
5934 * if it spins in bh_lock_sock(), but it is really
5937 inet_csk_reset_keepalive_timer(sk, tmo);
5939 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5947 if (tp->snd_una == tp->write_seq) {
5948 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5954 if (tp->snd_una == tp->write_seq) {
5955 tcp_update_metrics(sk);
5964 /* ts_recent update must be made after we are sure that the packet
5967 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5969 /* step 6: check the URG bit */
5970 tcp_urg(sk, skb, th);
5972 /* step 7: process the segment text */
5973 switch (sk->sk_state) {
5974 case TCP_CLOSE_WAIT:
5977 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5981 /* RFC 793 says to queue data in these states,
5982 * RFC 1122 says we MUST send a reset.
5983 * BSD 4.4 also does reset.
5985 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5986 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5987 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5988 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5994 case TCP_ESTABLISHED:
5995 tcp_data_queue(sk, skb);
6000 /* tcp_data could move socket to TIME-WAIT */
6001 if (sk->sk_state != TCP_CLOSE) {
6002 tcp_data_snd_check(sk);
6003 tcp_ack_snd_check(sk);
6012 EXPORT_SYMBOL(tcp_rcv_state_process);