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 = 2;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
89 int sysctl_tcp_stdurg __read_mostly;
90 int sysctl_tcp_rfc1337 __read_mostly;
91 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
92 int sysctl_tcp_frto __read_mostly = 2;
93 int sysctl_tcp_frto_response __read_mostly;
94 int sysctl_tcp_nometrics_save __read_mostly;
96 int sysctl_tcp_thin_dupack __read_mostly;
98 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
99 int sysctl_tcp_abc __read_mostly;
101 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
102 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
103 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
104 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
105 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
106 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
107 #define FLAG_ECE 0x40 /* ECE in this ACK */
108 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
111 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
114 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
120 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
122 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
123 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 /* Adapt the MSS value used to make delayed ack decision to the
128 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
130 struct inet_connection_sock *icsk = inet_csk(sk);
131 const unsigned int lss = icsk->icsk_ack.last_seg_size;
134 icsk->icsk_ack.last_seg_size = 0;
136 /* skb->len may jitter because of SACKs, even if peer
137 * sends good full-sized frames.
139 len = skb_shinfo(skb)->gso_size ? : skb->len;
140 if (len >= icsk->icsk_ack.rcv_mss) {
141 icsk->icsk_ack.rcv_mss = len;
143 /* Otherwise, we make more careful check taking into account,
144 * that SACKs block is variable.
146 * "len" is invariant segment length, including TCP header.
148 len += skb->data - skb_transport_header(skb);
149 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
150 /* If PSH is not set, packet should be
151 * full sized, provided peer TCP is not badly broken.
152 * This observation (if it is correct 8)) allows
153 * to handle super-low mtu links fairly.
155 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
156 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
157 /* Subtract also invariant (if peer is RFC compliant),
158 * tcp header plus fixed timestamp option length.
159 * Resulting "len" is MSS free of SACK jitter.
161 len -= tcp_sk(sk)->tcp_header_len;
162 icsk->icsk_ack.last_seg_size = len;
164 icsk->icsk_ack.rcv_mss = len;
168 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
170 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
174 static void tcp_incr_quickack(struct sock *sk)
176 struct inet_connection_sock *icsk = inet_csk(sk);
177 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
181 if (quickacks > icsk->icsk_ack.quick)
182 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
185 static void tcp_enter_quickack_mode(struct sock *sk)
187 struct inet_connection_sock *icsk = inet_csk(sk);
188 tcp_incr_quickack(sk);
189 icsk->icsk_ack.pingpong = 0;
190 icsk->icsk_ack.ato = TCP_ATO_MIN;
193 /* Send ACKs quickly, if "quick" count is not exhausted
194 * and the session is not interactive.
197 static inline int tcp_in_quickack_mode(const struct sock *sk)
199 const struct inet_connection_sock *icsk = inet_csk(sk);
200 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
205 if (tp->ecn_flags & TCP_ECN_OK)
206 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
211 if (tcp_hdr(skb)->cwr)
212 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
217 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
220 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
222 if (tp->ecn_flags & TCP_ECN_OK) {
223 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
224 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
225 /* Funny extension: if ECT is not set on a segment,
226 * it is surely retransmit. It is not in ECN RFC,
227 * but Linux follows this rule. */
228 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
229 tcp_enter_quickack_mode((struct sock *)tp);
233 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
235 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
236 tp->ecn_flags &= ~TCP_ECN_OK;
239 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
241 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
242 tp->ecn_flags &= ~TCP_ECN_OK;
245 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
247 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
252 /* Buffer size and advertised window tuning.
254 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
257 static void tcp_fixup_sndbuf(struct sock *sk)
259 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
260 sizeof(struct sk_buff);
262 if (sk->sk_sndbuf < 3 * sndmem) {
263 sk->sk_sndbuf = 3 * sndmem;
264 if (sk->sk_sndbuf > sysctl_tcp_wmem[2])
265 sk->sk_sndbuf = sysctl_tcp_wmem[2];
269 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
271 * All tcp_full_space() is split to two parts: "network" buffer, allocated
272 * forward and advertised in receiver window (tp->rcv_wnd) and
273 * "application buffer", required to isolate scheduling/application
274 * latencies from network.
275 * window_clamp is maximal advertised window. It can be less than
276 * tcp_full_space(), in this case tcp_full_space() - window_clamp
277 * is reserved for "application" buffer. The less window_clamp is
278 * the smoother our behaviour from viewpoint of network, but the lower
279 * throughput and the higher sensitivity of the connection to losses. 8)
281 * rcv_ssthresh is more strict window_clamp used at "slow start"
282 * phase to predict further behaviour of this connection.
283 * It is used for two goals:
284 * - to enforce header prediction at sender, even when application
285 * requires some significant "application buffer". It is check #1.
286 * - to prevent pruning of receive queue because of misprediction
287 * of receiver window. Check #2.
289 * The scheme does not work when sender sends good segments opening
290 * window and then starts to feed us spaghetti. But it should work
291 * in common situations. Otherwise, we have to rely on queue collapsing.
294 /* Slow part of check#2. */
295 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
297 struct tcp_sock *tp = tcp_sk(sk);
299 int truesize = tcp_win_from_space(skb->truesize) >> 1;
300 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
302 while (tp->rcv_ssthresh <= window) {
303 if (truesize <= skb->len)
304 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
312 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
314 struct tcp_sock *tp = tcp_sk(sk);
317 if (tp->rcv_ssthresh < tp->window_clamp &&
318 (int)tp->rcv_ssthresh < tcp_space(sk) &&
319 !tcp_memory_pressure) {
322 /* Check #2. Increase window, if skb with such overhead
323 * will fit to rcvbuf in future.
325 if (tcp_win_from_space(skb->truesize) <= skb->len)
326 incr = 2 * tp->advmss;
328 incr = __tcp_grow_window(sk, skb);
331 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
333 inet_csk(sk)->icsk_ack.quick |= 1;
338 /* 3. Tuning rcvbuf, when connection enters established state. */
340 static void tcp_fixup_rcvbuf(struct sock *sk)
342 struct tcp_sock *tp = tcp_sk(sk);
343 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
345 /* Try to select rcvbuf so that 4 mss-sized segments
346 * will fit to window and corresponding skbs will fit to our rcvbuf.
347 * (was 3; 4 is minimum to allow fast retransmit to work.)
349 while (tcp_win_from_space(rcvmem) < tp->advmss)
351 if (sk->sk_rcvbuf < 4 * rcvmem)
352 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
355 /* 4. Try to fixup all. It is made immediately after connection enters
358 static void tcp_init_buffer_space(struct sock *sk)
360 struct tcp_sock *tp = tcp_sk(sk);
363 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
364 tcp_fixup_rcvbuf(sk);
365 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
366 tcp_fixup_sndbuf(sk);
368 tp->rcvq_space.space = tp->rcv_wnd;
370 maxwin = tcp_full_space(sk);
372 if (tp->window_clamp >= maxwin) {
373 tp->window_clamp = maxwin;
375 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
376 tp->window_clamp = max(maxwin -
377 (maxwin >> sysctl_tcp_app_win),
381 /* Force reservation of one segment. */
382 if (sysctl_tcp_app_win &&
383 tp->window_clamp > 2 * tp->advmss &&
384 tp->window_clamp + tp->advmss > maxwin)
385 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
387 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
388 tp->snd_cwnd_stamp = tcp_time_stamp;
391 /* 5. Recalculate window clamp after socket hit its memory bounds. */
392 static void tcp_clamp_window(struct sock *sk)
394 struct tcp_sock *tp = tcp_sk(sk);
395 struct inet_connection_sock *icsk = inet_csk(sk);
397 icsk->icsk_ack.quick = 0;
399 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
400 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
401 !tcp_memory_pressure &&
402 atomic_long_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
403 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
406 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
407 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
410 /* Initialize RCV_MSS value.
411 * RCV_MSS is an our guess about MSS used by the peer.
412 * We haven't any direct information about the MSS.
413 * It's better to underestimate the RCV_MSS rather than overestimate.
414 * Overestimations make us ACKing less frequently than needed.
415 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
417 void tcp_initialize_rcv_mss(struct sock *sk)
419 struct tcp_sock *tp = tcp_sk(sk);
420 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
422 hint = min(hint, tp->rcv_wnd / 2);
423 hint = min(hint, TCP_MSS_DEFAULT);
424 hint = max(hint, TCP_MIN_MSS);
426 inet_csk(sk)->icsk_ack.rcv_mss = hint;
428 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
430 /* Receiver "autotuning" code.
432 * The algorithm for RTT estimation w/o timestamps is based on
433 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
434 * <http://public.lanl.gov/radiant/pubs.html#DRS>
436 * More detail on this code can be found at
437 * <http://staff.psc.edu/jheffner/>,
438 * though this reference is out of date. A new paper
441 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
443 u32 new_sample = tp->rcv_rtt_est.rtt;
449 if (new_sample != 0) {
450 /* If we sample in larger samples in the non-timestamp
451 * case, we could grossly overestimate the RTT especially
452 * with chatty applications or bulk transfer apps which
453 * are stalled on filesystem I/O.
455 * Also, since we are only going for a minimum in the
456 * non-timestamp case, we do not smooth things out
457 * else with timestamps disabled convergence takes too
461 m -= (new_sample >> 3);
463 } else if (m < new_sample)
466 /* No previous measure. */
470 if (tp->rcv_rtt_est.rtt != new_sample)
471 tp->rcv_rtt_est.rtt = new_sample;
474 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
476 if (tp->rcv_rtt_est.time == 0)
478 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
480 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
483 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
484 tp->rcv_rtt_est.time = tcp_time_stamp;
487 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
488 const struct sk_buff *skb)
490 struct tcp_sock *tp = tcp_sk(sk);
491 if (tp->rx_opt.rcv_tsecr &&
492 (TCP_SKB_CB(skb)->end_seq -
493 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
494 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
498 * This function should be called every time data is copied to user space.
499 * It calculates the appropriate TCP receive buffer space.
501 void tcp_rcv_space_adjust(struct sock *sk)
503 struct tcp_sock *tp = tcp_sk(sk);
507 if (tp->rcvq_space.time == 0)
510 time = tcp_time_stamp - tp->rcvq_space.time;
511 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
514 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
516 space = max(tp->rcvq_space.space, space);
518 if (tp->rcvq_space.space != space) {
521 tp->rcvq_space.space = space;
523 if (sysctl_tcp_moderate_rcvbuf &&
524 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
525 int new_clamp = space;
527 /* Receive space grows, normalize in order to
528 * take into account packet headers and sk_buff
529 * structure overhead.
534 rcvmem = (tp->advmss + MAX_TCP_HEADER +
535 16 + sizeof(struct sk_buff));
536 while (tcp_win_from_space(rcvmem) < tp->advmss)
539 space = min(space, sysctl_tcp_rmem[2]);
540 if (space > sk->sk_rcvbuf) {
541 sk->sk_rcvbuf = space;
543 /* Make the window clamp follow along. */
544 tp->window_clamp = new_clamp;
550 tp->rcvq_space.seq = tp->copied_seq;
551 tp->rcvq_space.time = tcp_time_stamp;
554 /* There is something which you must keep in mind when you analyze the
555 * behavior of the tp->ato delayed ack timeout interval. When a
556 * connection starts up, we want to ack as quickly as possible. The
557 * problem is that "good" TCP's do slow start at the beginning of data
558 * transmission. The means that until we send the first few ACK's the
559 * sender will sit on his end and only queue most of his data, because
560 * he can only send snd_cwnd unacked packets at any given time. For
561 * each ACK we send, he increments snd_cwnd and transmits more of his
564 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
566 struct tcp_sock *tp = tcp_sk(sk);
567 struct inet_connection_sock *icsk = inet_csk(sk);
570 inet_csk_schedule_ack(sk);
572 tcp_measure_rcv_mss(sk, skb);
574 tcp_rcv_rtt_measure(tp);
576 now = tcp_time_stamp;
578 if (!icsk->icsk_ack.ato) {
579 /* The _first_ data packet received, initialize
580 * delayed ACK engine.
582 tcp_incr_quickack(sk);
583 icsk->icsk_ack.ato = TCP_ATO_MIN;
585 int m = now - icsk->icsk_ack.lrcvtime;
587 if (m <= TCP_ATO_MIN / 2) {
588 /* The fastest case is the first. */
589 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
590 } else if (m < icsk->icsk_ack.ato) {
591 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
592 if (icsk->icsk_ack.ato > icsk->icsk_rto)
593 icsk->icsk_ack.ato = icsk->icsk_rto;
594 } else if (m > icsk->icsk_rto) {
595 /* Too long gap. Apparently sender failed to
596 * restart window, so that we send ACKs quickly.
598 tcp_incr_quickack(sk);
602 icsk->icsk_ack.lrcvtime = now;
604 TCP_ECN_check_ce(tp, skb);
607 tcp_grow_window(sk, skb);
610 /* Called to compute a smoothed rtt estimate. The data fed to this
611 * routine either comes from timestamps, or from segments that were
612 * known _not_ to have been retransmitted [see Karn/Partridge
613 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
614 * piece by Van Jacobson.
615 * NOTE: the next three routines used to be one big routine.
616 * To save cycles in the RFC 1323 implementation it was better to break
617 * it up into three procedures. -- erics
619 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
621 struct tcp_sock *tp = tcp_sk(sk);
622 long m = mrtt; /* RTT */
624 /* The following amusing code comes from Jacobson's
625 * article in SIGCOMM '88. Note that rtt and mdev
626 * are scaled versions of rtt and mean deviation.
627 * This is designed to be as fast as possible
628 * m stands for "measurement".
630 * On a 1990 paper the rto value is changed to:
631 * RTO = rtt + 4 * mdev
633 * Funny. This algorithm seems to be very broken.
634 * These formulae increase RTO, when it should be decreased, increase
635 * too slowly, when it should be increased quickly, decrease too quickly
636 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
637 * does not matter how to _calculate_ it. Seems, it was trap
638 * that VJ failed to avoid. 8)
643 m -= (tp->srtt >> 3); /* m is now error in rtt est */
644 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
646 m = -m; /* m is now abs(error) */
647 m -= (tp->mdev >> 2); /* similar update on mdev */
648 /* This is similar to one of Eifel findings.
649 * Eifel blocks mdev updates when rtt decreases.
650 * This solution is a bit different: we use finer gain
651 * for mdev in this case (alpha*beta).
652 * Like Eifel it also prevents growth of rto,
653 * but also it limits too fast rto decreases,
654 * happening in pure Eifel.
659 m -= (tp->mdev >> 2); /* similar update on mdev */
661 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
662 if (tp->mdev > tp->mdev_max) {
663 tp->mdev_max = tp->mdev;
664 if (tp->mdev_max > tp->rttvar)
665 tp->rttvar = tp->mdev_max;
667 if (after(tp->snd_una, tp->rtt_seq)) {
668 if (tp->mdev_max < tp->rttvar)
669 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
670 tp->rtt_seq = tp->snd_nxt;
671 tp->mdev_max = tcp_rto_min(sk);
674 /* no previous measure. */
675 tp->srtt = m << 3; /* take the measured time to be rtt */
676 tp->mdev = m << 1; /* make sure rto = 3*rtt */
677 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
678 tp->rtt_seq = tp->snd_nxt;
682 /* Calculate rto without backoff. This is the second half of Van Jacobson's
683 * routine referred to above.
685 static inline void tcp_set_rto(struct sock *sk)
687 const struct tcp_sock *tp = tcp_sk(sk);
688 /* Old crap is replaced with new one. 8)
691 * 1. If rtt variance happened to be less 50msec, it is hallucination.
692 * It cannot be less due to utterly erratic ACK generation made
693 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
694 * to do with delayed acks, because at cwnd>2 true delack timeout
695 * is invisible. Actually, Linux-2.4 also generates erratic
696 * ACKs in some circumstances.
698 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
700 /* 2. Fixups made earlier cannot be right.
701 * If we do not estimate RTO correctly without them,
702 * all the algo is pure shit and should be replaced
703 * with correct one. It is exactly, which we pretend to do.
706 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
707 * guarantees that rto is higher.
712 /* Save metrics learned by this TCP session.
713 This function is called only, when TCP finishes successfully
714 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
716 void tcp_update_metrics(struct sock *sk)
718 struct tcp_sock *tp = tcp_sk(sk);
719 struct dst_entry *dst = __sk_dst_get(sk);
721 if (sysctl_tcp_nometrics_save)
726 if (dst && (dst->flags & DST_HOST)) {
727 const struct inet_connection_sock *icsk = inet_csk(sk);
731 if (icsk->icsk_backoff || !tp->srtt) {
732 /* This session failed to estimate rtt. Why?
733 * Probably, no packets returned in time.
736 if (!(dst_metric_locked(dst, RTAX_RTT)))
737 dst_metric_set(dst, RTAX_RTT, 0);
741 rtt = dst_metric_rtt(dst, RTAX_RTT);
744 /* If newly calculated rtt larger than stored one,
745 * store new one. Otherwise, use EWMA. Remember,
746 * rtt overestimation is always better than underestimation.
748 if (!(dst_metric_locked(dst, RTAX_RTT))) {
750 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
752 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
755 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
760 /* Scale deviation to rttvar fixed point */
765 var = dst_metric_rtt(dst, RTAX_RTTVAR);
769 var -= (var - m) >> 2;
771 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
774 if (tcp_in_initial_slowstart(tp)) {
775 /* Slow start still did not finish. */
776 if (dst_metric(dst, RTAX_SSTHRESH) &&
777 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
778 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
779 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
780 if (!dst_metric_locked(dst, RTAX_CWND) &&
781 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
782 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
783 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
784 icsk->icsk_ca_state == TCP_CA_Open) {
785 /* Cong. avoidance phase, cwnd is reliable. */
786 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
787 dst_metric_set(dst, RTAX_SSTHRESH,
788 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
789 if (!dst_metric_locked(dst, RTAX_CWND))
790 dst_metric_set(dst, RTAX_CWND,
791 (dst_metric(dst, RTAX_CWND) +
794 /* Else slow start did not finish, cwnd is non-sense,
795 ssthresh may be also invalid.
797 if (!dst_metric_locked(dst, RTAX_CWND))
798 dst_metric_set(dst, RTAX_CWND,
799 (dst_metric(dst, RTAX_CWND) +
800 tp->snd_ssthresh) >> 1);
801 if (dst_metric(dst, RTAX_SSTHRESH) &&
802 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
803 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
804 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
807 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
808 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
809 tp->reordering != sysctl_tcp_reordering)
810 dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
815 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
817 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
820 cwnd = TCP_INIT_CWND;
821 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
824 /* Set slow start threshold and cwnd not falling to slow start */
825 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
827 struct tcp_sock *tp = tcp_sk(sk);
828 const struct inet_connection_sock *icsk = inet_csk(sk);
830 tp->prior_ssthresh = 0;
832 if (icsk->icsk_ca_state < TCP_CA_CWR) {
835 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
836 tp->snd_cwnd = min(tp->snd_cwnd,
837 tcp_packets_in_flight(tp) + 1U);
838 tp->snd_cwnd_cnt = 0;
839 tp->high_seq = tp->snd_nxt;
840 tp->snd_cwnd_stamp = tcp_time_stamp;
841 TCP_ECN_queue_cwr(tp);
843 tcp_set_ca_state(sk, TCP_CA_CWR);
848 * Packet counting of FACK is based on in-order assumptions, therefore TCP
849 * disables it when reordering is detected
851 static void tcp_disable_fack(struct tcp_sock *tp)
853 /* RFC3517 uses different metric in lost marker => reset on change */
855 tp->lost_skb_hint = NULL;
856 tp->rx_opt.sack_ok &= ~2;
859 /* Take a notice that peer is sending D-SACKs */
860 static void tcp_dsack_seen(struct tcp_sock *tp)
862 tp->rx_opt.sack_ok |= 4;
865 /* Initialize metrics on socket. */
867 static void tcp_init_metrics(struct sock *sk)
869 struct tcp_sock *tp = tcp_sk(sk);
870 struct dst_entry *dst = __sk_dst_get(sk);
877 if (dst_metric_locked(dst, RTAX_CWND))
878 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
879 if (dst_metric(dst, RTAX_SSTHRESH)) {
880 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
881 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
882 tp->snd_ssthresh = tp->snd_cwnd_clamp;
884 /* ssthresh may have been reduced unnecessarily during.
885 * 3WHS. Restore it back to its initial default.
887 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
889 if (dst_metric(dst, RTAX_REORDERING) &&
890 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
891 tcp_disable_fack(tp);
892 tp->reordering = dst_metric(dst, RTAX_REORDERING);
895 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
898 /* Initial rtt is determined from SYN,SYN-ACK.
899 * The segment is small and rtt may appear much
900 * less than real one. Use per-dst memory
901 * to make it more realistic.
903 * A bit of theory. RTT is time passed after "normal" sized packet
904 * is sent until it is ACKed. In normal circumstances sending small
905 * packets force peer to delay ACKs and calculation is correct too.
906 * The algorithm is adaptive and, provided we follow specs, it
907 * NEVER underestimate RTT. BUT! If peer tries to make some clever
908 * tricks sort of "quick acks" for time long enough to decrease RTT
909 * to low value, and then abruptly stops to do it and starts to delay
910 * ACKs, wait for troubles.
912 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
913 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
914 tp->rtt_seq = tp->snd_nxt;
916 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
917 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
918 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
923 /* RFC2988bis: We've failed to get a valid RTT sample from
924 * 3WHS. This is most likely due to retransmission,
925 * including spurious one. Reset the RTO back to 3secs
926 * from the more aggressive 1sec to avoid more spurious
929 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
930 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
932 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
933 * retransmitted. In light of RFC2988bis' more aggressive 1sec
934 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
935 * retransmission has occurred.
937 if (tp->total_retrans > 1)
940 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
941 tp->snd_cwnd_stamp = tcp_time_stamp;
944 static void tcp_update_reordering(struct sock *sk, const int metric,
947 struct tcp_sock *tp = tcp_sk(sk);
948 if (metric > tp->reordering) {
951 tp->reordering = min(TCP_MAX_REORDERING, metric);
953 /* This exciting event is worth to be remembered. 8) */
955 mib_idx = LINUX_MIB_TCPTSREORDER;
956 else if (tcp_is_reno(tp))
957 mib_idx = LINUX_MIB_TCPRENOREORDER;
958 else if (tcp_is_fack(tp))
959 mib_idx = LINUX_MIB_TCPFACKREORDER;
961 mib_idx = LINUX_MIB_TCPSACKREORDER;
963 NET_INC_STATS_BH(sock_net(sk), mib_idx);
964 #if FASTRETRANS_DEBUG > 1
965 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
966 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
970 tp->undo_marker ? tp->undo_retrans : 0);
972 tcp_disable_fack(tp);
976 /* This must be called before lost_out is incremented */
977 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
979 if ((tp->retransmit_skb_hint == NULL) ||
980 before(TCP_SKB_CB(skb)->seq,
981 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
982 tp->retransmit_skb_hint = skb;
985 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
986 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
989 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
991 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
992 tcp_verify_retransmit_hint(tp, skb);
994 tp->lost_out += tcp_skb_pcount(skb);
995 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
999 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1000 struct sk_buff *skb)
1002 tcp_verify_retransmit_hint(tp, skb);
1004 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1005 tp->lost_out += tcp_skb_pcount(skb);
1006 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1010 /* This procedure tags the retransmission queue when SACKs arrive.
1012 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1013 * Packets in queue with these bits set are counted in variables
1014 * sacked_out, retrans_out and lost_out, correspondingly.
1016 * Valid combinations are:
1017 * Tag InFlight Description
1018 * 0 1 - orig segment is in flight.
1019 * S 0 - nothing flies, orig reached receiver.
1020 * L 0 - nothing flies, orig lost by net.
1021 * R 2 - both orig and retransmit are in flight.
1022 * L|R 1 - orig is lost, retransmit is in flight.
1023 * S|R 1 - orig reached receiver, retrans is still in flight.
1024 * (L|S|R is logically valid, it could occur when L|R is sacked,
1025 * but it is equivalent to plain S and code short-curcuits it to S.
1026 * L|S is logically invalid, it would mean -1 packet in flight 8))
1028 * These 6 states form finite state machine, controlled by the following events:
1029 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1030 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1031 * 3. Loss detection event of one of three flavors:
1032 * A. Scoreboard estimator decided the packet is lost.
1033 * A'. Reno "three dupacks" marks head of queue lost.
1034 * A''. Its FACK modfication, head until snd.fack is lost.
1035 * B. SACK arrives sacking data transmitted after never retransmitted
1036 * hole was sent out.
1037 * C. SACK arrives sacking SND.NXT at the moment, when the
1038 * segment was retransmitted.
1039 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1041 * It is pleasant to note, that state diagram turns out to be commutative,
1042 * so that we are allowed not to be bothered by order of our actions,
1043 * when multiple events arrive simultaneously. (see the function below).
1045 * Reordering detection.
1046 * --------------------
1047 * Reordering metric is maximal distance, which a packet can be displaced
1048 * in packet stream. With SACKs we can estimate it:
1050 * 1. SACK fills old hole and the corresponding segment was not
1051 * ever retransmitted -> reordering. Alas, we cannot use it
1052 * when segment was retransmitted.
1053 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1054 * for retransmitted and already SACKed segment -> reordering..
1055 * Both of these heuristics are not used in Loss state, when we cannot
1056 * account for retransmits accurately.
1058 * SACK block validation.
1059 * ----------------------
1061 * SACK block range validation checks that the received SACK block fits to
1062 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1063 * Note that SND.UNA is not included to the range though being valid because
1064 * it means that the receiver is rather inconsistent with itself reporting
1065 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1066 * perfectly valid, however, in light of RFC2018 which explicitly states
1067 * that "SACK block MUST reflect the newest segment. Even if the newest
1068 * segment is going to be discarded ...", not that it looks very clever
1069 * in case of head skb. Due to potentional receiver driven attacks, we
1070 * choose to avoid immediate execution of a walk in write queue due to
1071 * reneging and defer head skb's loss recovery to standard loss recovery
1072 * procedure that will eventually trigger (nothing forbids us doing this).
1074 * Implements also blockage to start_seq wrap-around. Problem lies in the
1075 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1076 * there's no guarantee that it will be before snd_nxt (n). The problem
1077 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1080 * <- outs wnd -> <- wrapzone ->
1081 * u e n u_w e_w s n_w
1083 * |<------------+------+----- TCP seqno space --------------+---------->|
1084 * ...-- <2^31 ->| |<--------...
1085 * ...---- >2^31 ------>| |<--------...
1087 * Current code wouldn't be vulnerable but it's better still to discard such
1088 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1089 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1090 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1091 * equal to the ideal case (infinite seqno space without wrap caused issues).
1093 * With D-SACK the lower bound is extended to cover sequence space below
1094 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1095 * again, D-SACK block must not to go across snd_una (for the same reason as
1096 * for the normal SACK blocks, explained above). But there all simplicity
1097 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1098 * fully below undo_marker they do not affect behavior in anyway and can
1099 * therefore be safely ignored. In rare cases (which are more or less
1100 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1101 * fragmentation and packet reordering past skb's retransmission. To consider
1102 * them correctly, the acceptable range must be extended even more though
1103 * the exact amount is rather hard to quantify. However, tp->max_window can
1104 * be used as an exaggerated estimate.
1106 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1107 u32 start_seq, u32 end_seq)
1109 /* Too far in future, or reversed (interpretation is ambiguous) */
1110 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1113 /* Nasty start_seq wrap-around check (see comments above) */
1114 if (!before(start_seq, tp->snd_nxt))
1117 /* In outstanding window? ...This is valid exit for D-SACKs too.
1118 * start_seq == snd_una is non-sensical (see comments above)
1120 if (after(start_seq, tp->snd_una))
1123 if (!is_dsack || !tp->undo_marker)
1126 /* ...Then it's D-SACK, and must reside below snd_una completely */
1127 if (after(end_seq, tp->snd_una))
1130 if (!before(start_seq, tp->undo_marker))
1134 if (!after(end_seq, tp->undo_marker))
1137 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1138 * start_seq < undo_marker and end_seq >= undo_marker.
1140 return !before(start_seq, end_seq - tp->max_window);
1143 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1144 * Event "C". Later note: FACK people cheated me again 8), we have to account
1145 * for reordering! Ugly, but should help.
1147 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1148 * less than what is now known to be received by the other end (derived from
1149 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1150 * retransmitted skbs to avoid some costly processing per ACKs.
1152 static void tcp_mark_lost_retrans(struct sock *sk)
1154 const struct inet_connection_sock *icsk = inet_csk(sk);
1155 struct tcp_sock *tp = tcp_sk(sk);
1156 struct sk_buff *skb;
1158 u32 new_low_seq = tp->snd_nxt;
1159 u32 received_upto = tcp_highest_sack_seq(tp);
1161 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1162 !after(received_upto, tp->lost_retrans_low) ||
1163 icsk->icsk_ca_state != TCP_CA_Recovery)
1166 tcp_for_write_queue(skb, sk) {
1167 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1169 if (skb == tcp_send_head(sk))
1171 if (cnt == tp->retrans_out)
1173 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1176 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1179 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1180 * constraint here (see above) but figuring out that at
1181 * least tp->reordering SACK blocks reside between ack_seq
1182 * and received_upto is not easy task to do cheaply with
1183 * the available datastructures.
1185 * Whether FACK should check here for tp->reordering segs
1186 * in-between one could argue for either way (it would be
1187 * rather simple to implement as we could count fack_count
1188 * during the walk and do tp->fackets_out - fack_count).
1190 if (after(received_upto, ack_seq)) {
1191 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1192 tp->retrans_out -= tcp_skb_pcount(skb);
1194 tcp_skb_mark_lost_uncond_verify(tp, skb);
1195 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1197 if (before(ack_seq, new_low_seq))
1198 new_low_seq = ack_seq;
1199 cnt += tcp_skb_pcount(skb);
1203 if (tp->retrans_out)
1204 tp->lost_retrans_low = new_low_seq;
1207 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1208 struct tcp_sack_block_wire *sp, int num_sacks,
1211 struct tcp_sock *tp = tcp_sk(sk);
1212 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1213 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1216 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1219 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1220 } else if (num_sacks > 1) {
1221 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1222 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1224 if (!after(end_seq_0, end_seq_1) &&
1225 !before(start_seq_0, start_seq_1)) {
1228 NET_INC_STATS_BH(sock_net(sk),
1229 LINUX_MIB_TCPDSACKOFORECV);
1233 /* D-SACK for already forgotten data... Do dumb counting. */
1234 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1235 !after(end_seq_0, prior_snd_una) &&
1236 after(end_seq_0, tp->undo_marker))
1242 struct tcp_sacktag_state {
1248 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1249 * the incoming SACK may not exactly match but we can find smaller MSS
1250 * aligned portion of it that matches. Therefore we might need to fragment
1251 * which may fail and creates some hassle (caller must handle error case
1254 * FIXME: this could be merged to shift decision code
1256 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1257 u32 start_seq, u32 end_seq)
1260 unsigned int pkt_len;
1263 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1264 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1266 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1267 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1268 mss = tcp_skb_mss(skb);
1269 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1272 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1276 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1281 /* Round if necessary so that SACKs cover only full MSSes
1282 * and/or the remaining small portion (if present)
1284 if (pkt_len > mss) {
1285 unsigned int new_len = (pkt_len / mss) * mss;
1286 if (!in_sack && new_len < pkt_len) {
1288 if (new_len > skb->len)
1293 err = tcp_fragment(sk, skb, pkt_len, mss);
1301 static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1302 struct tcp_sacktag_state *state,
1303 int dup_sack, int pcount)
1305 struct tcp_sock *tp = tcp_sk(sk);
1306 u8 sacked = TCP_SKB_CB(skb)->sacked;
1307 int fack_count = state->fack_count;
1309 /* Account D-SACK for retransmitted packet. */
1310 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1311 if (tp->undo_marker && tp->undo_retrans &&
1312 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1314 if (sacked & TCPCB_SACKED_ACKED)
1315 state->reord = min(fack_count, state->reord);
1318 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1319 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1322 if (!(sacked & TCPCB_SACKED_ACKED)) {
1323 if (sacked & TCPCB_SACKED_RETRANS) {
1324 /* If the segment is not tagged as lost,
1325 * we do not clear RETRANS, believing
1326 * that retransmission is still in flight.
1328 if (sacked & TCPCB_LOST) {
1329 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1330 tp->lost_out -= pcount;
1331 tp->retrans_out -= pcount;
1334 if (!(sacked & TCPCB_RETRANS)) {
1335 /* New sack for not retransmitted frame,
1336 * which was in hole. It is reordering.
1338 if (before(TCP_SKB_CB(skb)->seq,
1339 tcp_highest_sack_seq(tp)))
1340 state->reord = min(fack_count,
1343 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1344 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1345 state->flag |= FLAG_ONLY_ORIG_SACKED;
1348 if (sacked & TCPCB_LOST) {
1349 sacked &= ~TCPCB_LOST;
1350 tp->lost_out -= pcount;
1354 sacked |= TCPCB_SACKED_ACKED;
1355 state->flag |= FLAG_DATA_SACKED;
1356 tp->sacked_out += pcount;
1358 fack_count += pcount;
1360 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1361 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1362 before(TCP_SKB_CB(skb)->seq,
1363 TCP_SKB_CB(tp->lost_skb_hint)->seq))
1364 tp->lost_cnt_hint += pcount;
1366 if (fack_count > tp->fackets_out)
1367 tp->fackets_out = fack_count;
1370 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1371 * frames and clear it. undo_retrans is decreased above, L|R frames
1372 * are accounted above as well.
1374 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1375 sacked &= ~TCPCB_SACKED_RETRANS;
1376 tp->retrans_out -= pcount;
1382 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1383 struct tcp_sacktag_state *state,
1384 unsigned int pcount, int shifted, int mss,
1387 struct tcp_sock *tp = tcp_sk(sk);
1388 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1392 if (skb == tp->lost_skb_hint)
1393 tp->lost_cnt_hint += pcount;
1395 TCP_SKB_CB(prev)->end_seq += shifted;
1396 TCP_SKB_CB(skb)->seq += shifted;
1398 skb_shinfo(prev)->gso_segs += pcount;
1399 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1400 skb_shinfo(skb)->gso_segs -= pcount;
1402 /* When we're adding to gso_segs == 1, gso_size will be zero,
1403 * in theory this shouldn't be necessary but as long as DSACK
1404 * code can come after this skb later on it's better to keep
1405 * setting gso_size to something.
1407 if (!skb_shinfo(prev)->gso_size) {
1408 skb_shinfo(prev)->gso_size = mss;
1409 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1412 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1413 if (skb_shinfo(skb)->gso_segs <= 1) {
1414 skb_shinfo(skb)->gso_size = 0;
1415 skb_shinfo(skb)->gso_type = 0;
1418 /* We discard results */
1419 tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
1421 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1422 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1425 BUG_ON(!tcp_skb_pcount(skb));
1426 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1430 /* Whole SKB was eaten :-) */
1432 if (skb == tp->retransmit_skb_hint)
1433 tp->retransmit_skb_hint = prev;
1434 if (skb == tp->scoreboard_skb_hint)
1435 tp->scoreboard_skb_hint = prev;
1436 if (skb == tp->lost_skb_hint) {
1437 tp->lost_skb_hint = prev;
1438 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1441 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1442 if (skb == tcp_highest_sack(sk))
1443 tcp_advance_highest_sack(sk, skb);
1445 tcp_unlink_write_queue(skb, sk);
1446 sk_wmem_free_skb(sk, skb);
1448 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1453 /* I wish gso_size would have a bit more sane initialization than
1454 * something-or-zero which complicates things
1456 static int tcp_skb_seglen(struct sk_buff *skb)
1458 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1461 /* Shifting pages past head area doesn't work */
1462 static int skb_can_shift(struct sk_buff *skb)
1464 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1467 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1470 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1471 struct tcp_sacktag_state *state,
1472 u32 start_seq, u32 end_seq,
1475 struct tcp_sock *tp = tcp_sk(sk);
1476 struct sk_buff *prev;
1482 if (!sk_can_gso(sk))
1485 /* Normally R but no L won't result in plain S */
1487 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1489 if (!skb_can_shift(skb))
1491 /* This frame is about to be dropped (was ACKed). */
1492 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1495 /* Can only happen with delayed DSACK + discard craziness */
1496 if (unlikely(skb == tcp_write_queue_head(sk)))
1498 prev = tcp_write_queue_prev(sk, skb);
1500 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1503 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1504 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1508 pcount = tcp_skb_pcount(skb);
1509 mss = tcp_skb_seglen(skb);
1511 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1512 * drop this restriction as unnecessary
1514 if (mss != tcp_skb_seglen(prev))
1517 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1519 /* CHECKME: This is non-MSS split case only?, this will
1520 * cause skipped skbs due to advancing loop btw, original
1521 * has that feature too
1523 if (tcp_skb_pcount(skb) <= 1)
1526 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1528 /* TODO: head merge to next could be attempted here
1529 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1530 * though it might not be worth of the additional hassle
1532 * ...we can probably just fallback to what was done
1533 * previously. We could try merging non-SACKed ones
1534 * as well but it probably isn't going to buy off
1535 * because later SACKs might again split them, and
1536 * it would make skb timestamp tracking considerably
1542 len = end_seq - TCP_SKB_CB(skb)->seq;
1544 BUG_ON(len > skb->len);
1546 /* MSS boundaries should be honoured or else pcount will
1547 * severely break even though it makes things bit trickier.
1548 * Optimize common case to avoid most of the divides
1550 mss = tcp_skb_mss(skb);
1552 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1553 * drop this restriction as unnecessary
1555 if (mss != tcp_skb_seglen(prev))
1560 } else if (len < mss) {
1568 if (!skb_shift(prev, skb, len))
1570 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1573 /* Hole filled allows collapsing with the next as well, this is very
1574 * useful when hole on every nth skb pattern happens
1576 if (prev == tcp_write_queue_tail(sk))
1578 skb = tcp_write_queue_next(sk, prev);
1580 if (!skb_can_shift(skb) ||
1581 (skb == tcp_send_head(sk)) ||
1582 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1583 (mss != tcp_skb_seglen(skb)))
1587 if (skb_shift(prev, skb, len)) {
1588 pcount += tcp_skb_pcount(skb);
1589 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1593 state->fack_count += pcount;
1600 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1604 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1605 struct tcp_sack_block *next_dup,
1606 struct tcp_sacktag_state *state,
1607 u32 start_seq, u32 end_seq,
1610 struct tcp_sock *tp = tcp_sk(sk);
1611 struct sk_buff *tmp;
1613 tcp_for_write_queue_from(skb, sk) {
1615 int dup_sack = dup_sack_in;
1617 if (skb == tcp_send_head(sk))
1620 /* queue is in-order => we can short-circuit the walk early */
1621 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1624 if ((next_dup != NULL) &&
1625 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1626 in_sack = tcp_match_skb_to_sack(sk, skb,
1627 next_dup->start_seq,
1633 /* skb reference here is a bit tricky to get right, since
1634 * shifting can eat and free both this skb and the next,
1635 * so not even _safe variant of the loop is enough.
1638 tmp = tcp_shift_skb_data(sk, skb, state,
1639 start_seq, end_seq, dup_sack);
1648 in_sack = tcp_match_skb_to_sack(sk, skb,
1654 if (unlikely(in_sack < 0))
1658 TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1661 tcp_skb_pcount(skb));
1663 if (!before(TCP_SKB_CB(skb)->seq,
1664 tcp_highest_sack_seq(tp)))
1665 tcp_advance_highest_sack(sk, skb);
1668 state->fack_count += tcp_skb_pcount(skb);
1673 /* Avoid all extra work that is being done by sacktag while walking in
1676 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1677 struct tcp_sacktag_state *state,
1680 tcp_for_write_queue_from(skb, sk) {
1681 if (skb == tcp_send_head(sk))
1684 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1687 state->fack_count += tcp_skb_pcount(skb);
1692 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1694 struct tcp_sack_block *next_dup,
1695 struct tcp_sacktag_state *state,
1698 if (next_dup == NULL)
1701 if (before(next_dup->start_seq, skip_to_seq)) {
1702 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1703 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1704 next_dup->start_seq, next_dup->end_seq,
1711 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1713 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1717 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1720 const struct inet_connection_sock *icsk = inet_csk(sk);
1721 struct tcp_sock *tp = tcp_sk(sk);
1722 unsigned char *ptr = (skb_transport_header(ack_skb) +
1723 TCP_SKB_CB(ack_skb)->sacked);
1724 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1725 struct tcp_sack_block sp[TCP_NUM_SACKS];
1726 struct tcp_sack_block *cache;
1727 struct tcp_sacktag_state state;
1728 struct sk_buff *skb;
1729 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1731 int found_dup_sack = 0;
1733 int first_sack_index;
1736 state.reord = tp->packets_out;
1738 if (!tp->sacked_out) {
1739 if (WARN_ON(tp->fackets_out))
1740 tp->fackets_out = 0;
1741 tcp_highest_sack_reset(sk);
1744 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1745 num_sacks, prior_snd_una);
1747 state.flag |= FLAG_DSACKING_ACK;
1749 /* Eliminate too old ACKs, but take into
1750 * account more or less fresh ones, they can
1751 * contain valid SACK info.
1753 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1756 if (!tp->packets_out)
1760 first_sack_index = 0;
1761 for (i = 0; i < num_sacks; i++) {
1762 int dup_sack = !i && found_dup_sack;
1764 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1765 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1767 if (!tcp_is_sackblock_valid(tp, dup_sack,
1768 sp[used_sacks].start_seq,
1769 sp[used_sacks].end_seq)) {
1773 if (!tp->undo_marker)
1774 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1776 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1778 /* Don't count olds caused by ACK reordering */
1779 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1780 !after(sp[used_sacks].end_seq, tp->snd_una))
1782 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1785 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1787 first_sack_index = -1;
1791 /* Ignore very old stuff early */
1792 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1798 /* order SACK blocks to allow in order walk of the retrans queue */
1799 for (i = used_sacks - 1; i > 0; i--) {
1800 for (j = 0; j < i; j++) {
1801 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1802 swap(sp[j], sp[j + 1]);
1804 /* Track where the first SACK block goes to */
1805 if (j == first_sack_index)
1806 first_sack_index = j + 1;
1811 skb = tcp_write_queue_head(sk);
1812 state.fack_count = 0;
1815 if (!tp->sacked_out) {
1816 /* It's already past, so skip checking against it */
1817 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1819 cache = tp->recv_sack_cache;
1820 /* Skip empty blocks in at head of the cache */
1821 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1826 while (i < used_sacks) {
1827 u32 start_seq = sp[i].start_seq;
1828 u32 end_seq = sp[i].end_seq;
1829 int dup_sack = (found_dup_sack && (i == first_sack_index));
1830 struct tcp_sack_block *next_dup = NULL;
1832 if (found_dup_sack && ((i + 1) == first_sack_index))
1833 next_dup = &sp[i + 1];
1835 /* Event "B" in the comment above. */
1836 if (after(end_seq, tp->high_seq))
1837 state.flag |= FLAG_DATA_LOST;
1839 /* Skip too early cached blocks */
1840 while (tcp_sack_cache_ok(tp, cache) &&
1841 !before(start_seq, cache->end_seq))
1844 /* Can skip some work by looking recv_sack_cache? */
1845 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1846 after(end_seq, cache->start_seq)) {
1849 if (before(start_seq, cache->start_seq)) {
1850 skb = tcp_sacktag_skip(skb, sk, &state,
1852 skb = tcp_sacktag_walk(skb, sk, next_dup,
1859 /* Rest of the block already fully processed? */
1860 if (!after(end_seq, cache->end_seq))
1863 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1867 /* ...tail remains todo... */
1868 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1869 /* ...but better entrypoint exists! */
1870 skb = tcp_highest_sack(sk);
1873 state.fack_count = tp->fackets_out;
1878 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1879 /* Check overlap against next cached too (past this one already) */
1884 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1885 skb = tcp_highest_sack(sk);
1888 state.fack_count = tp->fackets_out;
1890 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1893 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1894 start_seq, end_seq, dup_sack);
1897 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1898 * due to in-order walk
1900 if (after(end_seq, tp->frto_highmark))
1901 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1906 /* Clear the head of the cache sack blocks so we can skip it next time */
1907 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1908 tp->recv_sack_cache[i].start_seq = 0;
1909 tp->recv_sack_cache[i].end_seq = 0;
1911 for (j = 0; j < used_sacks; j++)
1912 tp->recv_sack_cache[i++] = sp[j];
1914 tcp_mark_lost_retrans(sk);
1916 tcp_verify_left_out(tp);
1918 if ((state.reord < tp->fackets_out) &&
1919 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1920 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1921 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1925 #if FASTRETRANS_DEBUG > 0
1926 WARN_ON((int)tp->sacked_out < 0);
1927 WARN_ON((int)tp->lost_out < 0);
1928 WARN_ON((int)tp->retrans_out < 0);
1929 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1934 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1935 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1937 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1941 holes = max(tp->lost_out, 1U);
1942 holes = min(holes, tp->packets_out);
1944 if ((tp->sacked_out + holes) > tp->packets_out) {
1945 tp->sacked_out = tp->packets_out - holes;
1951 /* If we receive more dupacks than we expected counting segments
1952 * in assumption of absent reordering, interpret this as reordering.
1953 * The only another reason could be bug in receiver TCP.
1955 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1957 struct tcp_sock *tp = tcp_sk(sk);
1958 if (tcp_limit_reno_sacked(tp))
1959 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1962 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1964 static void tcp_add_reno_sack(struct sock *sk)
1966 struct tcp_sock *tp = tcp_sk(sk);
1968 tcp_check_reno_reordering(sk, 0);
1969 tcp_verify_left_out(tp);
1972 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1974 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1976 struct tcp_sock *tp = tcp_sk(sk);
1979 /* One ACK acked hole. The rest eat duplicate ACKs. */
1980 if (acked - 1 >= tp->sacked_out)
1983 tp->sacked_out -= acked - 1;
1985 tcp_check_reno_reordering(sk, acked);
1986 tcp_verify_left_out(tp);
1989 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1994 static int tcp_is_sackfrto(const struct tcp_sock *tp)
1996 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1999 /* F-RTO can only be used if TCP has never retransmitted anything other than
2000 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2002 int tcp_use_frto(struct sock *sk)
2004 const struct tcp_sock *tp = tcp_sk(sk);
2005 const struct inet_connection_sock *icsk = inet_csk(sk);
2006 struct sk_buff *skb;
2008 if (!sysctl_tcp_frto)
2011 /* MTU probe and F-RTO won't really play nicely along currently */
2012 if (icsk->icsk_mtup.probe_size)
2015 if (tcp_is_sackfrto(tp))
2018 /* Avoid expensive walking of rexmit queue if possible */
2019 if (tp->retrans_out > 1)
2022 skb = tcp_write_queue_head(sk);
2023 if (tcp_skb_is_last(sk, skb))
2025 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2026 tcp_for_write_queue_from(skb, sk) {
2027 if (skb == tcp_send_head(sk))
2029 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2031 /* Short-circuit when first non-SACKed skb has been checked */
2032 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2038 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2039 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2040 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2041 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2042 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2043 * bits are handled if the Loss state is really to be entered (in
2044 * tcp_enter_frto_loss).
2046 * Do like tcp_enter_loss() would; when RTO expires the second time it
2048 * "Reduce ssthresh if it has not yet been made inside this window."
2050 void tcp_enter_frto(struct sock *sk)
2052 const struct inet_connection_sock *icsk = inet_csk(sk);
2053 struct tcp_sock *tp = tcp_sk(sk);
2054 struct sk_buff *skb;
2056 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2057 tp->snd_una == tp->high_seq ||
2058 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2059 !icsk->icsk_retransmits)) {
2060 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2061 /* Our state is too optimistic in ssthresh() call because cwnd
2062 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2063 * recovery has not yet completed. Pattern would be this: RTO,
2064 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2066 * RFC4138 should be more specific on what to do, even though
2067 * RTO is quite unlikely to occur after the first Cumulative ACK
2068 * due to back-off and complexity of triggering events ...
2070 if (tp->frto_counter) {
2072 stored_cwnd = tp->snd_cwnd;
2074 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2075 tp->snd_cwnd = stored_cwnd;
2077 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2079 /* ... in theory, cong.control module could do "any tricks" in
2080 * ssthresh(), which means that ca_state, lost bits and lost_out
2081 * counter would have to be faked before the call occurs. We
2082 * consider that too expensive, unlikely and hacky, so modules
2083 * using these in ssthresh() must deal these incompatibility
2084 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2086 tcp_ca_event(sk, CA_EVENT_FRTO);
2089 tp->undo_marker = tp->snd_una;
2090 tp->undo_retrans = 0;
2092 skb = tcp_write_queue_head(sk);
2093 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2094 tp->undo_marker = 0;
2095 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2096 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2097 tp->retrans_out -= tcp_skb_pcount(skb);
2099 tcp_verify_left_out(tp);
2101 /* Too bad if TCP was application limited */
2102 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2104 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2105 * The last condition is necessary at least in tp->frto_counter case.
2107 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2108 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2109 after(tp->high_seq, tp->snd_una)) {
2110 tp->frto_highmark = tp->high_seq;
2112 tp->frto_highmark = tp->snd_nxt;
2114 tcp_set_ca_state(sk, TCP_CA_Disorder);
2115 tp->high_seq = tp->snd_nxt;
2116 tp->frto_counter = 1;
2119 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2120 * which indicates that we should follow the traditional RTO recovery,
2121 * i.e. mark everything lost and do go-back-N retransmission.
2123 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2125 struct tcp_sock *tp = tcp_sk(sk);
2126 struct sk_buff *skb;
2129 tp->retrans_out = 0;
2130 if (tcp_is_reno(tp))
2131 tcp_reset_reno_sack(tp);
2133 tcp_for_write_queue(skb, sk) {
2134 if (skb == tcp_send_head(sk))
2137 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2139 * Count the retransmission made on RTO correctly (only when
2140 * waiting for the first ACK and did not get it)...
2142 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2143 /* For some reason this R-bit might get cleared? */
2144 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2145 tp->retrans_out += tcp_skb_pcount(skb);
2146 /* ...enter this if branch just for the first segment */
2147 flag |= FLAG_DATA_ACKED;
2149 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2150 tp->undo_marker = 0;
2151 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2154 /* Marking forward transmissions that were made after RTO lost
2155 * can cause unnecessary retransmissions in some scenarios,
2156 * SACK blocks will mitigate that in some but not in all cases.
2157 * We used to not mark them but it was causing break-ups with
2158 * receivers that do only in-order receival.
2160 * TODO: we could detect presence of such receiver and select
2161 * different behavior per flow.
2163 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2164 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2165 tp->lost_out += tcp_skb_pcount(skb);
2166 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2169 tcp_verify_left_out(tp);
2171 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2172 tp->snd_cwnd_cnt = 0;
2173 tp->snd_cwnd_stamp = tcp_time_stamp;
2174 tp->frto_counter = 0;
2175 tp->bytes_acked = 0;
2177 tp->reordering = min_t(unsigned int, tp->reordering,
2178 sysctl_tcp_reordering);
2179 tcp_set_ca_state(sk, TCP_CA_Loss);
2180 tp->high_seq = tp->snd_nxt;
2181 TCP_ECN_queue_cwr(tp);
2183 tcp_clear_all_retrans_hints(tp);
2186 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2188 tp->retrans_out = 0;
2191 tp->undo_marker = 0;
2192 tp->undo_retrans = 0;
2195 void tcp_clear_retrans(struct tcp_sock *tp)
2197 tcp_clear_retrans_partial(tp);
2199 tp->fackets_out = 0;
2203 /* Enter Loss state. If "how" is not zero, forget all SACK information
2204 * and reset tags completely, otherwise preserve SACKs. If receiver
2205 * dropped its ofo queue, we will know this due to reneging detection.
2207 void tcp_enter_loss(struct sock *sk, int how)
2209 const struct inet_connection_sock *icsk = inet_csk(sk);
2210 struct tcp_sock *tp = tcp_sk(sk);
2211 struct sk_buff *skb;
2213 /* Reduce ssthresh if it has not yet been made inside this window. */
2214 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2215 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2216 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2217 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2218 tcp_ca_event(sk, CA_EVENT_LOSS);
2221 tp->snd_cwnd_cnt = 0;
2222 tp->snd_cwnd_stamp = tcp_time_stamp;
2224 tp->bytes_acked = 0;
2225 tcp_clear_retrans_partial(tp);
2227 if (tcp_is_reno(tp))
2228 tcp_reset_reno_sack(tp);
2231 /* Push undo marker, if it was plain RTO and nothing
2232 * was retransmitted. */
2233 tp->undo_marker = tp->snd_una;
2236 tp->fackets_out = 0;
2238 tcp_clear_all_retrans_hints(tp);
2240 tcp_for_write_queue(skb, sk) {
2241 if (skb == tcp_send_head(sk))
2244 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2245 tp->undo_marker = 0;
2246 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2247 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2248 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2249 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2250 tp->lost_out += tcp_skb_pcount(skb);
2251 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2254 tcp_verify_left_out(tp);
2256 tp->reordering = min_t(unsigned int, tp->reordering,
2257 sysctl_tcp_reordering);
2258 tcp_set_ca_state(sk, TCP_CA_Loss);
2259 tp->high_seq = tp->snd_nxt;
2260 TCP_ECN_queue_cwr(tp);
2261 /* Abort F-RTO algorithm if one is in progress */
2262 tp->frto_counter = 0;
2265 /* If ACK arrived pointing to a remembered SACK, it means that our
2266 * remembered SACKs do not reflect real state of receiver i.e.
2267 * receiver _host_ is heavily congested (or buggy).
2269 * Do processing similar to RTO timeout.
2271 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2273 if (flag & FLAG_SACK_RENEGING) {
2274 struct inet_connection_sock *icsk = inet_csk(sk);
2275 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2277 tcp_enter_loss(sk, 1);
2278 icsk->icsk_retransmits++;
2279 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2280 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2281 icsk->icsk_rto, TCP_RTO_MAX);
2287 static inline int tcp_fackets_out(struct tcp_sock *tp)
2289 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2292 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2293 * counter when SACK is enabled (without SACK, sacked_out is used for
2296 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2297 * segments up to the highest received SACK block so far and holes in
2300 * With reordering, holes may still be in flight, so RFC3517 recovery
2301 * uses pure sacked_out (total number of SACKed segments) even though
2302 * it violates the RFC that uses duplicate ACKs, often these are equal
2303 * but when e.g. out-of-window ACKs or packet duplication occurs,
2304 * they differ. Since neither occurs due to loss, TCP should really
2307 static inline int tcp_dupack_heuristics(struct tcp_sock *tp)
2309 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2312 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2314 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2317 static inline int tcp_head_timedout(struct sock *sk)
2319 struct tcp_sock *tp = tcp_sk(sk);
2321 return tp->packets_out &&
2322 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2325 /* Linux NewReno/SACK/FACK/ECN state machine.
2326 * --------------------------------------
2328 * "Open" Normal state, no dubious events, fast path.
2329 * "Disorder" In all the respects it is "Open",
2330 * but requires a bit more attention. It is entered when
2331 * we see some SACKs or dupacks. It is split of "Open"
2332 * mainly to move some processing from fast path to slow one.
2333 * "CWR" CWND was reduced due to some Congestion Notification event.
2334 * It can be ECN, ICMP source quench, local device congestion.
2335 * "Recovery" CWND was reduced, we are fast-retransmitting.
2336 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2338 * tcp_fastretrans_alert() is entered:
2339 * - each incoming ACK, if state is not "Open"
2340 * - when arrived ACK is unusual, namely:
2345 * Counting packets in flight is pretty simple.
2347 * in_flight = packets_out - left_out + retrans_out
2349 * packets_out is SND.NXT-SND.UNA counted in packets.
2351 * retrans_out is number of retransmitted segments.
2353 * left_out is number of segments left network, but not ACKed yet.
2355 * left_out = sacked_out + lost_out
2357 * sacked_out: Packets, which arrived to receiver out of order
2358 * and hence not ACKed. With SACKs this number is simply
2359 * amount of SACKed data. Even without SACKs
2360 * it is easy to give pretty reliable estimate of this number,
2361 * counting duplicate ACKs.
2363 * lost_out: Packets lost by network. TCP has no explicit
2364 * "loss notification" feedback from network (for now).
2365 * It means that this number can be only _guessed_.
2366 * Actually, it is the heuristics to predict lossage that
2367 * distinguishes different algorithms.
2369 * F.e. after RTO, when all the queue is considered as lost,
2370 * lost_out = packets_out and in_flight = retrans_out.
2372 * Essentially, we have now two algorithms counting
2375 * FACK: It is the simplest heuristics. As soon as we decided
2376 * that something is lost, we decide that _all_ not SACKed
2377 * packets until the most forward SACK are lost. I.e.
2378 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2379 * It is absolutely correct estimate, if network does not reorder
2380 * packets. And it loses any connection to reality when reordering
2381 * takes place. We use FACK by default until reordering
2382 * is suspected on the path to this destination.
2384 * NewReno: when Recovery is entered, we assume that one segment
2385 * is lost (classic Reno). While we are in Recovery and
2386 * a partial ACK arrives, we assume that one more packet
2387 * is lost (NewReno). This heuristics are the same in NewReno
2390 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2391 * deflation etc. CWND is real congestion window, never inflated, changes
2392 * only according to classic VJ rules.
2394 * Really tricky (and requiring careful tuning) part of algorithm
2395 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2396 * The first determines the moment _when_ we should reduce CWND and,
2397 * hence, slow down forward transmission. In fact, it determines the moment
2398 * when we decide that hole is caused by loss, rather than by a reorder.
2400 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2401 * holes, caused by lost packets.
2403 * And the most logically complicated part of algorithm is undo
2404 * heuristics. We detect false retransmits due to both too early
2405 * fast retransmit (reordering) and underestimated RTO, analyzing
2406 * timestamps and D-SACKs. When we detect that some segments were
2407 * retransmitted by mistake and CWND reduction was wrong, we undo
2408 * window reduction and abort recovery phase. This logic is hidden
2409 * inside several functions named tcp_try_undo_<something>.
2412 /* This function decides, when we should leave Disordered state
2413 * and enter Recovery phase, reducing congestion window.
2415 * Main question: may we further continue forward transmission
2416 * with the same cwnd?
2418 static int tcp_time_to_recover(struct sock *sk)
2420 struct tcp_sock *tp = tcp_sk(sk);
2423 /* Do not perform any recovery during F-RTO algorithm */
2424 if (tp->frto_counter)
2427 /* Trick#1: The loss is proven. */
2431 /* Not-A-Trick#2 : Classic rule... */
2432 if (tcp_dupack_heuristics(tp) > tp->reordering)
2435 /* Trick#3 : when we use RFC2988 timer restart, fast
2436 * retransmit can be triggered by timeout of queue head.
2438 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2441 /* Trick#4: It is still not OK... But will it be useful to delay
2444 packets_out = tp->packets_out;
2445 if (packets_out <= tp->reordering &&
2446 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2447 !tcp_may_send_now(sk)) {
2448 /* We have nothing to send. This connection is limited
2449 * either by receiver window or by application.
2454 /* If a thin stream is detected, retransmit after first
2455 * received dupack. Employ only if SACK is supported in order
2456 * to avoid possible corner-case series of spurious retransmissions
2457 * Use only if there are no unsent data.
2459 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2460 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2461 tcp_is_sack(tp) && !tcp_send_head(sk))
2467 /* New heuristics: it is possible only after we switched to restart timer
2468 * each time when something is ACKed. Hence, we can detect timed out packets
2469 * during fast retransmit without falling to slow start.
2471 * Usefulness of this as is very questionable, since we should know which of
2472 * the segments is the next to timeout which is relatively expensive to find
2473 * in general case unless we add some data structure just for that. The
2474 * current approach certainly won't find the right one too often and when it
2475 * finally does find _something_ it usually marks large part of the window
2476 * right away (because a retransmission with a larger timestamp blocks the
2477 * loop from advancing). -ij
2479 static void tcp_timeout_skbs(struct sock *sk)
2481 struct tcp_sock *tp = tcp_sk(sk);
2482 struct sk_buff *skb;
2484 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2487 skb = tp->scoreboard_skb_hint;
2488 if (tp->scoreboard_skb_hint == NULL)
2489 skb = tcp_write_queue_head(sk);
2491 tcp_for_write_queue_from(skb, sk) {
2492 if (skb == tcp_send_head(sk))
2494 if (!tcp_skb_timedout(sk, skb))
2497 tcp_skb_mark_lost(tp, skb);
2500 tp->scoreboard_skb_hint = skb;
2502 tcp_verify_left_out(tp);
2505 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2506 * is against sacked "cnt", otherwise it's against facked "cnt"
2508 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2510 struct tcp_sock *tp = tcp_sk(sk);
2511 struct sk_buff *skb;
2516 WARN_ON(packets > tp->packets_out);
2517 if (tp->lost_skb_hint) {
2518 skb = tp->lost_skb_hint;
2519 cnt = tp->lost_cnt_hint;
2520 /* Head already handled? */
2521 if (mark_head && skb != tcp_write_queue_head(sk))
2524 skb = tcp_write_queue_head(sk);
2528 tcp_for_write_queue_from(skb, sk) {
2529 if (skb == tcp_send_head(sk))
2531 /* TODO: do this better */
2532 /* this is not the most efficient way to do this... */
2533 tp->lost_skb_hint = skb;
2534 tp->lost_cnt_hint = cnt;
2536 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2540 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2541 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2542 cnt += tcp_skb_pcount(skb);
2544 if (cnt > packets) {
2545 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2546 (oldcnt >= packets))
2549 mss = skb_shinfo(skb)->gso_size;
2550 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2556 tcp_skb_mark_lost(tp, skb);
2561 tcp_verify_left_out(tp);
2564 /* Account newly detected lost packet(s) */
2566 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2568 struct tcp_sock *tp = tcp_sk(sk);
2570 if (tcp_is_reno(tp)) {
2571 tcp_mark_head_lost(sk, 1, 1);
2572 } else if (tcp_is_fack(tp)) {
2573 int lost = tp->fackets_out - tp->reordering;
2576 tcp_mark_head_lost(sk, lost, 0);
2578 int sacked_upto = tp->sacked_out - tp->reordering;
2579 if (sacked_upto >= 0)
2580 tcp_mark_head_lost(sk, sacked_upto, 0);
2581 else if (fast_rexmit)
2582 tcp_mark_head_lost(sk, 1, 1);
2585 tcp_timeout_skbs(sk);
2588 /* CWND moderation, preventing bursts due to too big ACKs
2589 * in dubious situations.
2591 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2593 tp->snd_cwnd = min(tp->snd_cwnd,
2594 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2595 tp->snd_cwnd_stamp = tcp_time_stamp;
2598 /* Lower bound on congestion window is slow start threshold
2599 * unless congestion avoidance choice decides to overide it.
2601 static inline u32 tcp_cwnd_min(const struct sock *sk)
2603 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2605 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2608 /* Decrease cwnd each second ack. */
2609 static void tcp_cwnd_down(struct sock *sk, int flag)
2611 struct tcp_sock *tp = tcp_sk(sk);
2612 int decr = tp->snd_cwnd_cnt + 1;
2614 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2615 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2616 tp->snd_cwnd_cnt = decr & 1;
2619 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2620 tp->snd_cwnd -= decr;
2622 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2623 tp->snd_cwnd_stamp = tcp_time_stamp;
2627 /* Nothing was retransmitted or returned timestamp is less
2628 * than timestamp of the first retransmission.
2630 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2632 return !tp->retrans_stamp ||
2633 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2634 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2637 /* Undo procedures. */
2639 #if FASTRETRANS_DEBUG > 1
2640 static void DBGUNDO(struct sock *sk, const char *msg)
2642 struct tcp_sock *tp = tcp_sk(sk);
2643 struct inet_sock *inet = inet_sk(sk);
2645 if (sk->sk_family == AF_INET) {
2646 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2648 &inet->inet_daddr, ntohs(inet->inet_dport),
2649 tp->snd_cwnd, tcp_left_out(tp),
2650 tp->snd_ssthresh, tp->prior_ssthresh,
2653 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2654 else if (sk->sk_family == AF_INET6) {
2655 struct ipv6_pinfo *np = inet6_sk(sk);
2656 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2658 &np->daddr, ntohs(inet->inet_dport),
2659 tp->snd_cwnd, tcp_left_out(tp),
2660 tp->snd_ssthresh, tp->prior_ssthresh,
2666 #define DBGUNDO(x...) do { } while (0)
2669 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2671 struct tcp_sock *tp = tcp_sk(sk);
2673 if (tp->prior_ssthresh) {
2674 const struct inet_connection_sock *icsk = inet_csk(sk);
2676 if (icsk->icsk_ca_ops->undo_cwnd)
2677 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2679 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2681 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2682 tp->snd_ssthresh = tp->prior_ssthresh;
2683 TCP_ECN_withdraw_cwr(tp);
2686 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2688 tp->snd_cwnd_stamp = tcp_time_stamp;
2691 static inline int tcp_may_undo(struct tcp_sock *tp)
2693 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2696 /* People celebrate: "We love our President!" */
2697 static int tcp_try_undo_recovery(struct sock *sk)
2699 struct tcp_sock *tp = tcp_sk(sk);
2701 if (tcp_may_undo(tp)) {
2704 /* Happy end! We did not retransmit anything
2705 * or our original transmission succeeded.
2707 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2708 tcp_undo_cwr(sk, true);
2709 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2710 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2712 mib_idx = LINUX_MIB_TCPFULLUNDO;
2714 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2715 tp->undo_marker = 0;
2717 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2718 /* Hold old state until something *above* high_seq
2719 * is ACKed. For Reno it is MUST to prevent false
2720 * fast retransmits (RFC2582). SACK TCP is safe. */
2721 tcp_moderate_cwnd(tp);
2724 tcp_set_ca_state(sk, TCP_CA_Open);
2728 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2729 static void tcp_try_undo_dsack(struct sock *sk)
2731 struct tcp_sock *tp = tcp_sk(sk);
2733 if (tp->undo_marker && !tp->undo_retrans) {
2734 DBGUNDO(sk, "D-SACK");
2735 tcp_undo_cwr(sk, true);
2736 tp->undo_marker = 0;
2737 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2741 /* We can clear retrans_stamp when there are no retransmissions in the
2742 * window. It would seem that it is trivially available for us in
2743 * tp->retrans_out, however, that kind of assumptions doesn't consider
2744 * what will happen if errors occur when sending retransmission for the
2745 * second time. ...It could the that such segment has only
2746 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2747 * the head skb is enough except for some reneging corner cases that
2748 * are not worth the effort.
2750 * Main reason for all this complexity is the fact that connection dying
2751 * time now depends on the validity of the retrans_stamp, in particular,
2752 * that successive retransmissions of a segment must not advance
2753 * retrans_stamp under any conditions.
2755 static int tcp_any_retrans_done(struct sock *sk)
2757 struct tcp_sock *tp = tcp_sk(sk);
2758 struct sk_buff *skb;
2760 if (tp->retrans_out)
2763 skb = tcp_write_queue_head(sk);
2764 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2770 /* Undo during fast recovery after partial ACK. */
2772 static int tcp_try_undo_partial(struct sock *sk, int acked)
2774 struct tcp_sock *tp = tcp_sk(sk);
2775 /* Partial ACK arrived. Force Hoe's retransmit. */
2776 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2778 if (tcp_may_undo(tp)) {
2779 /* Plain luck! Hole if filled with delayed
2780 * packet, rather than with a retransmit.
2782 if (!tcp_any_retrans_done(sk))
2783 tp->retrans_stamp = 0;
2785 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2788 tcp_undo_cwr(sk, false);
2789 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2791 /* So... Do not make Hoe's retransmit yet.
2792 * If the first packet was delayed, the rest
2793 * ones are most probably delayed as well.
2800 /* Undo during loss recovery after partial ACK. */
2801 static int tcp_try_undo_loss(struct sock *sk)
2803 struct tcp_sock *tp = tcp_sk(sk);
2805 if (tcp_may_undo(tp)) {
2806 struct sk_buff *skb;
2807 tcp_for_write_queue(skb, sk) {
2808 if (skb == tcp_send_head(sk))
2810 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2813 tcp_clear_all_retrans_hints(tp);
2815 DBGUNDO(sk, "partial loss");
2817 tcp_undo_cwr(sk, true);
2818 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2819 inet_csk(sk)->icsk_retransmits = 0;
2820 tp->undo_marker = 0;
2821 if (tcp_is_sack(tp))
2822 tcp_set_ca_state(sk, TCP_CA_Open);
2828 static inline void tcp_complete_cwr(struct sock *sk)
2830 struct tcp_sock *tp = tcp_sk(sk);
2831 /* Do not moderate cwnd if it's already undone in cwr or recovery */
2832 if (tp->undo_marker && tp->snd_cwnd > tp->snd_ssthresh) {
2833 tp->snd_cwnd = tp->snd_ssthresh;
2834 tp->snd_cwnd_stamp = tcp_time_stamp;
2836 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2839 static void tcp_try_keep_open(struct sock *sk)
2841 struct tcp_sock *tp = tcp_sk(sk);
2842 int state = TCP_CA_Open;
2844 if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker)
2845 state = TCP_CA_Disorder;
2847 if (inet_csk(sk)->icsk_ca_state != state) {
2848 tcp_set_ca_state(sk, state);
2849 tp->high_seq = tp->snd_nxt;
2853 static void tcp_try_to_open(struct sock *sk, int flag)
2855 struct tcp_sock *tp = tcp_sk(sk);
2857 tcp_verify_left_out(tp);
2859 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2860 tp->retrans_stamp = 0;
2862 if (flag & FLAG_ECE)
2863 tcp_enter_cwr(sk, 1);
2865 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2866 tcp_try_keep_open(sk);
2867 tcp_moderate_cwnd(tp);
2869 tcp_cwnd_down(sk, flag);
2873 static void tcp_mtup_probe_failed(struct sock *sk)
2875 struct inet_connection_sock *icsk = inet_csk(sk);
2877 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2878 icsk->icsk_mtup.probe_size = 0;
2881 static void tcp_mtup_probe_success(struct sock *sk)
2883 struct tcp_sock *tp = tcp_sk(sk);
2884 struct inet_connection_sock *icsk = inet_csk(sk);
2886 /* FIXME: breaks with very large cwnd */
2887 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2888 tp->snd_cwnd = tp->snd_cwnd *
2889 tcp_mss_to_mtu(sk, tp->mss_cache) /
2890 icsk->icsk_mtup.probe_size;
2891 tp->snd_cwnd_cnt = 0;
2892 tp->snd_cwnd_stamp = tcp_time_stamp;
2893 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2895 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2896 icsk->icsk_mtup.probe_size = 0;
2897 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2900 /* Do a simple retransmit without using the backoff mechanisms in
2901 * tcp_timer. This is used for path mtu discovery.
2902 * The socket is already locked here.
2904 void tcp_simple_retransmit(struct sock *sk)
2906 const struct inet_connection_sock *icsk = inet_csk(sk);
2907 struct tcp_sock *tp = tcp_sk(sk);
2908 struct sk_buff *skb;
2909 unsigned int mss = tcp_current_mss(sk);
2910 u32 prior_lost = tp->lost_out;
2912 tcp_for_write_queue(skb, sk) {
2913 if (skb == tcp_send_head(sk))
2915 if (tcp_skb_seglen(skb) > mss &&
2916 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2917 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2918 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2919 tp->retrans_out -= tcp_skb_pcount(skb);
2921 tcp_skb_mark_lost_uncond_verify(tp, skb);
2925 tcp_clear_retrans_hints_partial(tp);
2927 if (prior_lost == tp->lost_out)
2930 if (tcp_is_reno(tp))
2931 tcp_limit_reno_sacked(tp);
2933 tcp_verify_left_out(tp);
2935 /* Don't muck with the congestion window here.
2936 * Reason is that we do not increase amount of _data_
2937 * in network, but units changed and effective
2938 * cwnd/ssthresh really reduced now.
2940 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2941 tp->high_seq = tp->snd_nxt;
2942 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2943 tp->prior_ssthresh = 0;
2944 tp->undo_marker = 0;
2945 tcp_set_ca_state(sk, TCP_CA_Loss);
2947 tcp_xmit_retransmit_queue(sk);
2949 EXPORT_SYMBOL(tcp_simple_retransmit);
2951 /* Process an event, which can update packets-in-flight not trivially.
2952 * Main goal of this function is to calculate new estimate for left_out,
2953 * taking into account both packets sitting in receiver's buffer and
2954 * packets lost by network.
2956 * Besides that it does CWND reduction, when packet loss is detected
2957 * and changes state of machine.
2959 * It does _not_ decide what to send, it is made in function
2960 * tcp_xmit_retransmit_queue().
2962 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2964 struct inet_connection_sock *icsk = inet_csk(sk);
2965 struct tcp_sock *tp = tcp_sk(sk);
2966 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2967 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2968 (tcp_fackets_out(tp) > tp->reordering));
2969 int fast_rexmit = 0, mib_idx;
2971 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2973 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2974 tp->fackets_out = 0;
2976 /* Now state machine starts.
2977 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2978 if (flag & FLAG_ECE)
2979 tp->prior_ssthresh = 0;
2981 /* B. In all the states check for reneging SACKs. */
2982 if (tcp_check_sack_reneging(sk, flag))
2985 /* C. Process data loss notification, provided it is valid. */
2986 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2987 before(tp->snd_una, tp->high_seq) &&
2988 icsk->icsk_ca_state != TCP_CA_Open &&
2989 tp->fackets_out > tp->reordering) {
2990 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0);
2991 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
2994 /* D. Check consistency of the current state. */
2995 tcp_verify_left_out(tp);
2997 /* E. Check state exit conditions. State can be terminated
2998 * when high_seq is ACKed. */
2999 if (icsk->icsk_ca_state == TCP_CA_Open) {
3000 WARN_ON(tp->retrans_out != 0);
3001 tp->retrans_stamp = 0;
3002 } else if (!before(tp->snd_una, tp->high_seq)) {
3003 switch (icsk->icsk_ca_state) {
3005 icsk->icsk_retransmits = 0;
3006 if (tcp_try_undo_recovery(sk))
3011 /* CWR is to be held something *above* high_seq
3012 * is ACKed for CWR bit to reach receiver. */
3013 if (tp->snd_una != tp->high_seq) {
3014 tcp_complete_cwr(sk);
3015 tcp_set_ca_state(sk, TCP_CA_Open);
3019 case TCP_CA_Disorder:
3020 tcp_try_undo_dsack(sk);
3021 if (!tp->undo_marker ||
3022 /* For SACK case do not Open to allow to undo
3023 * catching for all duplicate ACKs. */
3024 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
3025 tp->undo_marker = 0;
3026 tcp_set_ca_state(sk, TCP_CA_Open);
3030 case TCP_CA_Recovery:
3031 if (tcp_is_reno(tp))
3032 tcp_reset_reno_sack(tp);
3033 if (tcp_try_undo_recovery(sk))
3035 tcp_complete_cwr(sk);
3040 /* F. Process state. */
3041 switch (icsk->icsk_ca_state) {
3042 case TCP_CA_Recovery:
3043 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3044 if (tcp_is_reno(tp) && is_dupack)
3045 tcp_add_reno_sack(sk);
3047 do_lost = tcp_try_undo_partial(sk, pkts_acked);
3050 if (flag & FLAG_DATA_ACKED)
3051 icsk->icsk_retransmits = 0;
3052 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3053 tcp_reset_reno_sack(tp);
3054 if (!tcp_try_undo_loss(sk)) {
3055 tcp_moderate_cwnd(tp);
3056 tcp_xmit_retransmit_queue(sk);
3059 if (icsk->icsk_ca_state != TCP_CA_Open)
3061 /* Loss is undone; fall through to processing in Open state. */
3063 if (tcp_is_reno(tp)) {
3064 if (flag & FLAG_SND_UNA_ADVANCED)
3065 tcp_reset_reno_sack(tp);
3067 tcp_add_reno_sack(sk);
3070 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3071 tcp_try_undo_dsack(sk);
3073 if (!tcp_time_to_recover(sk)) {
3074 tcp_try_to_open(sk, flag);
3078 /* MTU probe failure: don't reduce cwnd */
3079 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3080 icsk->icsk_mtup.probe_size &&
3081 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3082 tcp_mtup_probe_failed(sk);
3083 /* Restores the reduction we did in tcp_mtup_probe() */
3085 tcp_simple_retransmit(sk);
3089 /* Otherwise enter Recovery state */
3091 if (tcp_is_reno(tp))
3092 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3094 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3096 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3098 tp->high_seq = tp->snd_nxt;
3099 tp->prior_ssthresh = 0;
3100 tp->undo_marker = tp->snd_una;
3101 tp->undo_retrans = tp->retrans_out;
3103 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3104 if (!(flag & FLAG_ECE))
3105 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3106 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3107 TCP_ECN_queue_cwr(tp);
3110 tp->bytes_acked = 0;
3111 tp->snd_cwnd_cnt = 0;
3112 tcp_set_ca_state(sk, TCP_CA_Recovery);
3116 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3117 tcp_update_scoreboard(sk, fast_rexmit);
3118 tcp_cwnd_down(sk, flag);
3119 tcp_xmit_retransmit_queue(sk);
3122 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3124 tcp_rtt_estimator(sk, seq_rtt);
3126 inet_csk(sk)->icsk_backoff = 0;
3128 EXPORT_SYMBOL(tcp_valid_rtt_meas);
3130 /* Read draft-ietf-tcplw-high-performance before mucking
3131 * with this code. (Supersedes RFC1323)
3133 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3135 /* RTTM Rule: A TSecr value received in a segment is used to
3136 * update the averaged RTT measurement only if the segment
3137 * acknowledges some new data, i.e., only if it advances the
3138 * left edge of the send window.
3140 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3141 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3143 * Changed: reset backoff as soon as we see the first valid sample.
3144 * If we do not, we get strongly overestimated rto. With timestamps
3145 * samples are accepted even from very old segments: f.e., when rtt=1
3146 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3147 * answer arrives rto becomes 120 seconds! If at least one of segments
3148 * in window is lost... Voila. --ANK (010210)
3150 struct tcp_sock *tp = tcp_sk(sk);
3152 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3155 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3157 /* We don't have a timestamp. Can only use
3158 * packets that are not retransmitted to determine
3159 * rtt estimates. Also, we must not reset the
3160 * backoff for rto until we get a non-retransmitted
3161 * packet. This allows us to deal with a situation
3162 * where the network delay has increased suddenly.
3163 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3166 if (flag & FLAG_RETRANS_DATA_ACKED)
3169 tcp_valid_rtt_meas(sk, seq_rtt);
3172 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3175 const struct tcp_sock *tp = tcp_sk(sk);
3176 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3177 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3178 tcp_ack_saw_tstamp(sk, flag);
3179 else if (seq_rtt >= 0)
3180 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3183 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3185 const struct inet_connection_sock *icsk = inet_csk(sk);
3186 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3187 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3190 /* Restart timer after forward progress on connection.
3191 * RFC2988 recommends to restart timer to now+rto.
3193 static void tcp_rearm_rto(struct sock *sk)
3195 struct tcp_sock *tp = tcp_sk(sk);
3197 if (!tp->packets_out) {
3198 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3200 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3201 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3205 /* If we get here, the whole TSO packet has not been acked. */
3206 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3208 struct tcp_sock *tp = tcp_sk(sk);
3211 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3213 packets_acked = tcp_skb_pcount(skb);
3214 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3216 packets_acked -= tcp_skb_pcount(skb);
3218 if (packets_acked) {
3219 BUG_ON(tcp_skb_pcount(skb) == 0);
3220 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3223 return packets_acked;
3226 /* Remove acknowledged frames from the retransmission queue. If our packet
3227 * is before the ack sequence we can discard it as it's confirmed to have
3228 * arrived at the other end.
3230 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3233 struct tcp_sock *tp = tcp_sk(sk);
3234 const struct inet_connection_sock *icsk = inet_csk(sk);
3235 struct sk_buff *skb;
3236 u32 now = tcp_time_stamp;
3237 int fully_acked = 1;
3240 u32 reord = tp->packets_out;
3241 u32 prior_sacked = tp->sacked_out;
3243 s32 ca_seq_rtt = -1;
3244 ktime_t last_ackt = net_invalid_timestamp();
3246 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3247 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3249 u8 sacked = scb->sacked;
3251 /* Determine how many packets and what bytes were acked, tso and else */
3252 if (after(scb->end_seq, tp->snd_una)) {
3253 if (tcp_skb_pcount(skb) == 1 ||
3254 !after(tp->snd_una, scb->seq))
3257 acked_pcount = tcp_tso_acked(sk, skb);
3263 acked_pcount = tcp_skb_pcount(skb);
3266 if (sacked & TCPCB_RETRANS) {
3267 if (sacked & TCPCB_SACKED_RETRANS)
3268 tp->retrans_out -= acked_pcount;
3269 flag |= FLAG_RETRANS_DATA_ACKED;
3272 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3273 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3275 ca_seq_rtt = now - scb->when;
3276 last_ackt = skb->tstamp;
3278 seq_rtt = ca_seq_rtt;
3280 if (!(sacked & TCPCB_SACKED_ACKED))
3281 reord = min(pkts_acked, reord);
3284 if (sacked & TCPCB_SACKED_ACKED)
3285 tp->sacked_out -= acked_pcount;
3286 if (sacked & TCPCB_LOST)
3287 tp->lost_out -= acked_pcount;
3289 tp->packets_out -= acked_pcount;
3290 pkts_acked += acked_pcount;
3292 /* Initial outgoing SYN's get put onto the write_queue
3293 * just like anything else we transmit. It is not
3294 * true data, and if we misinform our callers that
3295 * this ACK acks real data, we will erroneously exit
3296 * connection startup slow start one packet too
3297 * quickly. This is severely frowned upon behavior.
3299 if (!(scb->flags & TCPHDR_SYN)) {
3300 flag |= FLAG_DATA_ACKED;
3302 flag |= FLAG_SYN_ACKED;
3303 tp->retrans_stamp = 0;
3309 tcp_unlink_write_queue(skb, sk);
3310 sk_wmem_free_skb(sk, skb);
3311 tp->scoreboard_skb_hint = NULL;
3312 if (skb == tp->retransmit_skb_hint)
3313 tp->retransmit_skb_hint = NULL;
3314 if (skb == tp->lost_skb_hint)
3315 tp->lost_skb_hint = NULL;
3318 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3319 tp->snd_up = tp->snd_una;
3321 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3322 flag |= FLAG_SACK_RENEGING;
3324 if (flag & FLAG_ACKED) {
3325 const struct tcp_congestion_ops *ca_ops
3326 = inet_csk(sk)->icsk_ca_ops;
3328 if (unlikely(icsk->icsk_mtup.probe_size &&
3329 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3330 tcp_mtup_probe_success(sk);
3333 tcp_ack_update_rtt(sk, flag, seq_rtt);
3336 if (tcp_is_reno(tp)) {
3337 tcp_remove_reno_sacks(sk, pkts_acked);
3341 /* Non-retransmitted hole got filled? That's reordering */
3342 if (reord < prior_fackets)
3343 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3345 delta = tcp_is_fack(tp) ? pkts_acked :
3346 prior_sacked - tp->sacked_out;
3347 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3350 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3352 if (ca_ops->pkts_acked) {
3355 /* Is the ACK triggering packet unambiguous? */
3356 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3357 /* High resolution needed and available? */
3358 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3359 !ktime_equal(last_ackt,
3360 net_invalid_timestamp()))
3361 rtt_us = ktime_us_delta(ktime_get_real(),
3363 else if (ca_seq_rtt >= 0)
3364 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3367 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3371 #if FASTRETRANS_DEBUG > 0
3372 WARN_ON((int)tp->sacked_out < 0);
3373 WARN_ON((int)tp->lost_out < 0);
3374 WARN_ON((int)tp->retrans_out < 0);
3375 if (!tp->packets_out && tcp_is_sack(tp)) {
3376 icsk = inet_csk(sk);
3378 printk(KERN_DEBUG "Leak l=%u %d\n",
3379 tp->lost_out, icsk->icsk_ca_state);
3382 if (tp->sacked_out) {
3383 printk(KERN_DEBUG "Leak s=%u %d\n",
3384 tp->sacked_out, icsk->icsk_ca_state);
3387 if (tp->retrans_out) {
3388 printk(KERN_DEBUG "Leak r=%u %d\n",
3389 tp->retrans_out, icsk->icsk_ca_state);
3390 tp->retrans_out = 0;
3397 static void tcp_ack_probe(struct sock *sk)
3399 const struct tcp_sock *tp = tcp_sk(sk);
3400 struct inet_connection_sock *icsk = inet_csk(sk);
3402 /* Was it a usable window open? */
3404 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3405 icsk->icsk_backoff = 0;
3406 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3407 /* Socket must be waked up by subsequent tcp_data_snd_check().
3408 * This function is not for random using!
3411 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3412 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3417 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3419 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3420 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3423 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3425 const struct tcp_sock *tp = tcp_sk(sk);
3426 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3427 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3430 /* Check that window update is acceptable.
3431 * The function assumes that snd_una<=ack<=snd_next.
3433 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3434 const u32 ack, const u32 ack_seq,
3437 return after(ack, tp->snd_una) ||
3438 after(ack_seq, tp->snd_wl1) ||
3439 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3442 /* Update our send window.
3444 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3445 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3447 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3450 struct tcp_sock *tp = tcp_sk(sk);
3452 u32 nwin = ntohs(tcp_hdr(skb)->window);
3454 if (likely(!tcp_hdr(skb)->syn))
3455 nwin <<= tp->rx_opt.snd_wscale;
3457 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3458 flag |= FLAG_WIN_UPDATE;
3459 tcp_update_wl(tp, ack_seq);
3461 if (tp->snd_wnd != nwin) {
3464 /* Note, it is the only place, where
3465 * fast path is recovered for sending TCP.
3468 tcp_fast_path_check(sk);
3470 if (nwin > tp->max_window) {
3471 tp->max_window = nwin;
3472 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3482 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3483 * continue in congestion avoidance.
3485 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3487 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3488 tp->snd_cwnd_cnt = 0;
3489 tp->bytes_acked = 0;
3490 TCP_ECN_queue_cwr(tp);
3491 tcp_moderate_cwnd(tp);
3494 /* A conservative spurious RTO response algorithm: reduce cwnd using
3495 * rate halving and continue in congestion avoidance.
3497 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3499 tcp_enter_cwr(sk, 0);
3502 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3504 if (flag & FLAG_ECE)
3505 tcp_ratehalving_spur_to_response(sk);
3507 tcp_undo_cwr(sk, true);
3510 /* F-RTO spurious RTO detection algorithm (RFC4138)
3512 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3513 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3514 * window (but not to or beyond highest sequence sent before RTO):
3515 * On First ACK, send two new segments out.
3516 * On Second ACK, RTO was likely spurious. Do spurious response (response
3517 * algorithm is not part of the F-RTO detection algorithm
3518 * given in RFC4138 but can be selected separately).
3519 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3520 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3521 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3522 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3524 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3525 * original window even after we transmit two new data segments.
3528 * on first step, wait until first cumulative ACK arrives, then move to
3529 * the second step. In second step, the next ACK decides.
3531 * F-RTO is implemented (mainly) in four functions:
3532 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3533 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3534 * called when tcp_use_frto() showed green light
3535 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3536 * - tcp_enter_frto_loss() is called if there is not enough evidence
3537 * to prove that the RTO is indeed spurious. It transfers the control
3538 * from F-RTO to the conventional RTO recovery
3540 static int tcp_process_frto(struct sock *sk, int flag)
3542 struct tcp_sock *tp = tcp_sk(sk);
3544 tcp_verify_left_out(tp);
3546 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3547 if (flag & FLAG_DATA_ACKED)
3548 inet_csk(sk)->icsk_retransmits = 0;
3550 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3551 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3552 tp->undo_marker = 0;
3554 if (!before(tp->snd_una, tp->frto_highmark)) {
3555 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3559 if (!tcp_is_sackfrto(tp)) {
3560 /* RFC4138 shortcoming in step 2; should also have case c):
3561 * ACK isn't duplicate nor advances window, e.g., opposite dir
3564 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3567 if (!(flag & FLAG_DATA_ACKED)) {
3568 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3573 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3574 /* Prevent sending of new data. */
3575 tp->snd_cwnd = min(tp->snd_cwnd,
3576 tcp_packets_in_flight(tp));
3580 if ((tp->frto_counter >= 2) &&
3581 (!(flag & FLAG_FORWARD_PROGRESS) ||
3582 ((flag & FLAG_DATA_SACKED) &&
3583 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3584 /* RFC4138 shortcoming (see comment above) */
3585 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3586 (flag & FLAG_NOT_DUP))
3589 tcp_enter_frto_loss(sk, 3, flag);
3594 if (tp->frto_counter == 1) {
3595 /* tcp_may_send_now needs to see updated state */
3596 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3597 tp->frto_counter = 2;
3599 if (!tcp_may_send_now(sk))
3600 tcp_enter_frto_loss(sk, 2, flag);
3604 switch (sysctl_tcp_frto_response) {
3606 tcp_undo_spur_to_response(sk, flag);
3609 tcp_conservative_spur_to_response(tp);
3612 tcp_ratehalving_spur_to_response(sk);
3615 tp->frto_counter = 0;
3616 tp->undo_marker = 0;
3617 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3622 /* This routine deals with incoming acks, but not outgoing ones. */
3623 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3625 struct inet_connection_sock *icsk = inet_csk(sk);
3626 struct tcp_sock *tp = tcp_sk(sk);
3627 u32 prior_snd_una = tp->snd_una;
3628 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3629 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3630 u32 prior_in_flight;
3635 /* If the ack is older than previous acks
3636 * then we can probably ignore it.
3638 if (before(ack, prior_snd_una))
3641 /* If the ack includes data we haven't sent yet, discard
3642 * this segment (RFC793 Section 3.9).
3644 if (after(ack, tp->snd_nxt))
3647 if (after(ack, prior_snd_una))
3648 flag |= FLAG_SND_UNA_ADVANCED;
3650 if (sysctl_tcp_abc) {
3651 if (icsk->icsk_ca_state < TCP_CA_CWR)
3652 tp->bytes_acked += ack - prior_snd_una;
3653 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3654 /* we assume just one segment left network */
3655 tp->bytes_acked += min(ack - prior_snd_una,
3659 prior_fackets = tp->fackets_out;
3660 prior_in_flight = tcp_packets_in_flight(tp);
3662 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3663 /* Window is constant, pure forward advance.
3664 * No more checks are required.
3665 * Note, we use the fact that SND.UNA>=SND.WL2.
3667 tcp_update_wl(tp, ack_seq);
3669 flag |= FLAG_WIN_UPDATE;
3671 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3673 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3675 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3678 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3680 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3682 if (TCP_SKB_CB(skb)->sacked)
3683 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3685 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3688 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3691 /* We passed data and got it acked, remove any soft error
3692 * log. Something worked...
3694 sk->sk_err_soft = 0;
3695 icsk->icsk_probes_out = 0;
3696 tp->rcv_tstamp = tcp_time_stamp;
3697 prior_packets = tp->packets_out;
3701 /* See if we can take anything off of the retransmit queue. */
3702 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3704 if (tp->frto_counter)
3705 frto_cwnd = tcp_process_frto(sk, flag);
3706 /* Guarantee sacktag reordering detection against wrap-arounds */
3707 if (before(tp->frto_highmark, tp->snd_una))
3708 tp->frto_highmark = 0;
3710 if (tcp_ack_is_dubious(sk, flag)) {
3711 /* Advance CWND, if state allows this. */
3712 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3713 tcp_may_raise_cwnd(sk, flag))
3714 tcp_cong_avoid(sk, ack, prior_in_flight);
3715 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3718 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3719 tcp_cong_avoid(sk, ack, prior_in_flight);
3722 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3723 dst_confirm(__sk_dst_get(sk));
3728 /* If this ack opens up a zero window, clear backoff. It was
3729 * being used to time the probes, and is probably far higher than
3730 * it needs to be for normal retransmission.
3732 if (tcp_send_head(sk))
3737 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3741 if (TCP_SKB_CB(skb)->sacked) {
3742 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3743 if (icsk->icsk_ca_state == TCP_CA_Open)
3744 tcp_try_keep_open(sk);
3747 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3751 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3752 * But, this can also be called on packets in the established flow when
3753 * the fast version below fails.
3755 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3756 u8 **hvpp, int estab)
3759 struct tcphdr *th = tcp_hdr(skb);
3760 int length = (th->doff * 4) - sizeof(struct tcphdr);
3762 ptr = (unsigned char *)(th + 1);
3763 opt_rx->saw_tstamp = 0;
3765 while (length > 0) {
3766 int opcode = *ptr++;
3772 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3777 if (opsize < 2) /* "silly options" */
3779 if (opsize > length)
3780 return; /* don't parse partial options */
3783 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3784 u16 in_mss = get_unaligned_be16(ptr);
3786 if (opt_rx->user_mss &&
3787 opt_rx->user_mss < in_mss)
3788 in_mss = opt_rx->user_mss;
3789 opt_rx->mss_clamp = in_mss;
3794 if (opsize == TCPOLEN_WINDOW && th->syn &&
3795 !estab && sysctl_tcp_window_scaling) {
3796 __u8 snd_wscale = *(__u8 *)ptr;
3797 opt_rx->wscale_ok = 1;
3798 if (snd_wscale > 14) {
3799 if (net_ratelimit())
3800 printk(KERN_INFO "tcp_parse_options: Illegal window "
3801 "scaling value %d >14 received.\n",
3805 opt_rx->snd_wscale = snd_wscale;
3808 case TCPOPT_TIMESTAMP:
3809 if ((opsize == TCPOLEN_TIMESTAMP) &&
3810 ((estab && opt_rx->tstamp_ok) ||
3811 (!estab && sysctl_tcp_timestamps))) {
3812 opt_rx->saw_tstamp = 1;
3813 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3814 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3817 case TCPOPT_SACK_PERM:
3818 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3819 !estab && sysctl_tcp_sack) {
3820 opt_rx->sack_ok = 1;
3821 tcp_sack_reset(opt_rx);
3826 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3827 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3829 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3832 #ifdef CONFIG_TCP_MD5SIG
3835 * The MD5 Hash has already been
3836 * checked (see tcp_v{4,6}_do_rcv()).
3841 /* This option is variable length.
3844 case TCPOLEN_COOKIE_BASE:
3845 /* not yet implemented */
3847 case TCPOLEN_COOKIE_PAIR:
3848 /* not yet implemented */
3850 case TCPOLEN_COOKIE_MIN+0:
3851 case TCPOLEN_COOKIE_MIN+2:
3852 case TCPOLEN_COOKIE_MIN+4:
3853 case TCPOLEN_COOKIE_MIN+6:
3854 case TCPOLEN_COOKIE_MAX:
3855 /* 16-bit multiple */
3856 opt_rx->cookie_plus = opsize;
3871 EXPORT_SYMBOL(tcp_parse_options);
3873 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3875 __be32 *ptr = (__be32 *)(th + 1);
3877 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3878 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3879 tp->rx_opt.saw_tstamp = 1;
3881 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3883 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3889 /* Fast parse options. This hopes to only see timestamps.
3890 * If it is wrong it falls back on tcp_parse_options().
3892 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3893 struct tcp_sock *tp, u8 **hvpp)
3895 /* In the spirit of fast parsing, compare doff directly to constant
3896 * values. Because equality is used, short doff can be ignored here.
3898 if (th->doff == (sizeof(*th) / 4)) {
3899 tp->rx_opt.saw_tstamp = 0;
3901 } else if (tp->rx_opt.tstamp_ok &&
3902 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3903 if (tcp_parse_aligned_timestamp(tp, th))
3906 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3910 #ifdef CONFIG_TCP_MD5SIG
3912 * Parse MD5 Signature option
3914 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3916 int length = (th->doff << 2) - sizeof (*th);
3917 u8 *ptr = (u8*)(th + 1);
3919 /* If the TCP option is too short, we can short cut */
3920 if (length < TCPOLEN_MD5SIG)
3923 while (length > 0) {
3924 int opcode = *ptr++;
3935 if (opsize < 2 || opsize > length)
3937 if (opcode == TCPOPT_MD5SIG)
3938 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3945 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3948 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3950 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3951 tp->rx_opt.ts_recent_stamp = get_seconds();
3954 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3956 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3957 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3958 * extra check below makes sure this can only happen
3959 * for pure ACK frames. -DaveM
3961 * Not only, also it occurs for expired timestamps.
3964 if (tcp_paws_check(&tp->rx_opt, 0))
3965 tcp_store_ts_recent(tp);
3969 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3971 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3972 * it can pass through stack. So, the following predicate verifies that
3973 * this segment is not used for anything but congestion avoidance or
3974 * fast retransmit. Moreover, we even are able to eliminate most of such
3975 * second order effects, if we apply some small "replay" window (~RTO)
3976 * to timestamp space.
3978 * All these measures still do not guarantee that we reject wrapped ACKs
3979 * on networks with high bandwidth, when sequence space is recycled fastly,
3980 * but it guarantees that such events will be very rare and do not affect
3981 * connection seriously. This doesn't look nice, but alas, PAWS is really
3984 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3985 * states that events when retransmit arrives after original data are rare.
3986 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3987 * the biggest problem on large power networks even with minor reordering.
3988 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3989 * up to bandwidth of 18Gigabit/sec. 8) ]
3992 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3994 struct tcp_sock *tp = tcp_sk(sk);
3995 struct tcphdr *th = tcp_hdr(skb);
3996 u32 seq = TCP_SKB_CB(skb)->seq;
3997 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3999 return (/* 1. Pure ACK with correct sequence number. */
4000 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4002 /* 2. ... and duplicate ACK. */
4003 ack == tp->snd_una &&
4005 /* 3. ... and does not update window. */
4006 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4008 /* 4. ... and sits in replay window. */
4009 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4012 static inline int tcp_paws_discard(const struct sock *sk,
4013 const struct sk_buff *skb)
4015 const struct tcp_sock *tp = tcp_sk(sk);
4017 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4018 !tcp_disordered_ack(sk, skb);
4021 /* Check segment sequence number for validity.
4023 * Segment controls are considered valid, if the segment
4024 * fits to the window after truncation to the window. Acceptability
4025 * of data (and SYN, FIN, of course) is checked separately.
4026 * See tcp_data_queue(), for example.
4028 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4029 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4030 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4031 * (borrowed from freebsd)
4034 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
4036 return !before(end_seq, tp->rcv_wup) &&
4037 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4040 /* When we get a reset we do this. */
4041 static void tcp_reset(struct sock *sk)
4043 /* We want the right error as BSD sees it (and indeed as we do). */
4044 switch (sk->sk_state) {
4046 sk->sk_err = ECONNREFUSED;
4048 case TCP_CLOSE_WAIT:
4054 sk->sk_err = ECONNRESET;
4056 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4059 if (!sock_flag(sk, SOCK_DEAD))
4060 sk->sk_error_report(sk);
4066 * Process the FIN bit. This now behaves as it is supposed to work
4067 * and the FIN takes effect when it is validly part of sequence
4068 * space. Not before when we get holes.
4070 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4071 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4074 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4075 * close and we go into CLOSING (and later onto TIME-WAIT)
4077 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4079 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
4081 struct tcp_sock *tp = tcp_sk(sk);
4083 inet_csk_schedule_ack(sk);
4085 sk->sk_shutdown |= RCV_SHUTDOWN;
4086 sock_set_flag(sk, SOCK_DONE);
4088 switch (sk->sk_state) {
4090 case TCP_ESTABLISHED:
4091 /* Move to CLOSE_WAIT */
4092 tcp_set_state(sk, TCP_CLOSE_WAIT);
4093 inet_csk(sk)->icsk_ack.pingpong = 1;
4096 case TCP_CLOSE_WAIT:
4098 /* Received a retransmission of the FIN, do
4103 /* RFC793: Remain in the LAST-ACK state. */
4107 /* This case occurs when a simultaneous close
4108 * happens, we must ack the received FIN and
4109 * enter the CLOSING state.
4112 tcp_set_state(sk, TCP_CLOSING);
4115 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4117 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4120 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4121 * cases we should never reach this piece of code.
4123 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4124 __func__, sk->sk_state);
4128 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4129 * Probably, we should reset in this case. For now drop them.
4131 __skb_queue_purge(&tp->out_of_order_queue);
4132 if (tcp_is_sack(tp))
4133 tcp_sack_reset(&tp->rx_opt);
4136 if (!sock_flag(sk, SOCK_DEAD)) {
4137 sk->sk_state_change(sk);
4139 /* Do not send POLL_HUP for half duplex close. */
4140 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4141 sk->sk_state == TCP_CLOSE)
4142 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4144 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4148 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4151 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4152 if (before(seq, sp->start_seq))
4153 sp->start_seq = seq;
4154 if (after(end_seq, sp->end_seq))
4155 sp->end_seq = end_seq;
4161 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4163 struct tcp_sock *tp = tcp_sk(sk);
4165 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4168 if (before(seq, tp->rcv_nxt))
4169 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4171 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4173 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4175 tp->rx_opt.dsack = 1;
4176 tp->duplicate_sack[0].start_seq = seq;
4177 tp->duplicate_sack[0].end_seq = end_seq;
4181 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4183 struct tcp_sock *tp = tcp_sk(sk);
4185 if (!tp->rx_opt.dsack)
4186 tcp_dsack_set(sk, seq, end_seq);
4188 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4191 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4193 struct tcp_sock *tp = tcp_sk(sk);
4195 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4196 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4197 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4198 tcp_enter_quickack_mode(sk);
4200 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4201 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4203 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4204 end_seq = tp->rcv_nxt;
4205 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4212 /* These routines update the SACK block as out-of-order packets arrive or
4213 * in-order packets close up the sequence space.
4215 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4218 struct tcp_sack_block *sp = &tp->selective_acks[0];
4219 struct tcp_sack_block *swalk = sp + 1;
4221 /* See if the recent change to the first SACK eats into
4222 * or hits the sequence space of other SACK blocks, if so coalesce.
4224 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4225 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4228 /* Zap SWALK, by moving every further SACK up by one slot.
4229 * Decrease num_sacks.
4231 tp->rx_opt.num_sacks--;
4232 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4236 this_sack++, swalk++;
4240 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4242 struct tcp_sock *tp = tcp_sk(sk);
4243 struct tcp_sack_block *sp = &tp->selective_acks[0];
4244 int cur_sacks = tp->rx_opt.num_sacks;
4250 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4251 if (tcp_sack_extend(sp, seq, end_seq)) {
4252 /* Rotate this_sack to the first one. */
4253 for (; this_sack > 0; this_sack--, sp--)
4254 swap(*sp, *(sp - 1));
4256 tcp_sack_maybe_coalesce(tp);
4261 /* Could not find an adjacent existing SACK, build a new one,
4262 * put it at the front, and shift everyone else down. We
4263 * always know there is at least one SACK present already here.
4265 * If the sack array is full, forget about the last one.
4267 if (this_sack >= TCP_NUM_SACKS) {
4269 tp->rx_opt.num_sacks--;
4272 for (; this_sack > 0; this_sack--, sp--)
4276 /* Build the new head SACK, and we're done. */
4277 sp->start_seq = seq;
4278 sp->end_seq = end_seq;
4279 tp->rx_opt.num_sacks++;
4282 /* RCV.NXT advances, some SACKs should be eaten. */
4284 static void tcp_sack_remove(struct tcp_sock *tp)
4286 struct tcp_sack_block *sp = &tp->selective_acks[0];
4287 int num_sacks = tp->rx_opt.num_sacks;
4290 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4291 if (skb_queue_empty(&tp->out_of_order_queue)) {
4292 tp->rx_opt.num_sacks = 0;
4296 for (this_sack = 0; this_sack < num_sacks;) {
4297 /* Check if the start of the sack is covered by RCV.NXT. */
4298 if (!before(tp->rcv_nxt, sp->start_seq)) {
4301 /* RCV.NXT must cover all the block! */
4302 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4304 /* Zap this SACK, by moving forward any other SACKS. */
4305 for (i=this_sack+1; i < num_sacks; i++)
4306 tp->selective_acks[i-1] = tp->selective_acks[i];
4313 tp->rx_opt.num_sacks = num_sacks;
4316 /* This one checks to see if we can put data from the
4317 * out_of_order queue into the receive_queue.
4319 static void tcp_ofo_queue(struct sock *sk)
4321 struct tcp_sock *tp = tcp_sk(sk);
4322 __u32 dsack_high = tp->rcv_nxt;
4323 struct sk_buff *skb;
4325 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4326 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4329 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4330 __u32 dsack = dsack_high;
4331 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4332 dsack_high = TCP_SKB_CB(skb)->end_seq;
4333 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4336 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4337 SOCK_DEBUG(sk, "ofo packet was already received\n");
4338 __skb_unlink(skb, &tp->out_of_order_queue);
4342 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4343 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4344 TCP_SKB_CB(skb)->end_seq);
4346 __skb_unlink(skb, &tp->out_of_order_queue);
4347 __skb_queue_tail(&sk->sk_receive_queue, skb);
4348 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4349 if (tcp_hdr(skb)->fin)
4350 tcp_fin(skb, sk, tcp_hdr(skb));
4354 static int tcp_prune_ofo_queue(struct sock *sk);
4355 static int tcp_prune_queue(struct sock *sk);
4357 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4359 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4360 !sk_rmem_schedule(sk, size)) {
4362 if (tcp_prune_queue(sk) < 0)
4365 if (!sk_rmem_schedule(sk, size)) {
4366 if (!tcp_prune_ofo_queue(sk))
4369 if (!sk_rmem_schedule(sk, size))
4376 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4378 struct tcphdr *th = tcp_hdr(skb);
4379 struct tcp_sock *tp = tcp_sk(sk);
4382 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4386 __skb_pull(skb, th->doff * 4);
4388 TCP_ECN_accept_cwr(tp, skb);
4390 tp->rx_opt.dsack = 0;
4392 /* Queue data for delivery to the user.
4393 * Packets in sequence go to the receive queue.
4394 * Out of sequence packets to the out_of_order_queue.
4396 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4397 if (tcp_receive_window(tp) == 0)
4400 /* Ok. In sequence. In window. */
4401 if (tp->ucopy.task == current &&
4402 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4403 sock_owned_by_user(sk) && !tp->urg_data) {
4404 int chunk = min_t(unsigned int, skb->len,
4407 __set_current_state(TASK_RUNNING);
4410 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4411 tp->ucopy.len -= chunk;
4412 tp->copied_seq += chunk;
4413 eaten = (chunk == skb->len);
4414 tcp_rcv_space_adjust(sk);
4422 tcp_try_rmem_schedule(sk, skb->truesize))
4425 skb_set_owner_r(skb, sk);
4426 __skb_queue_tail(&sk->sk_receive_queue, skb);
4428 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4430 tcp_event_data_recv(sk, skb);
4432 tcp_fin(skb, sk, th);
4434 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4437 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4438 * gap in queue is filled.
4440 if (skb_queue_empty(&tp->out_of_order_queue))
4441 inet_csk(sk)->icsk_ack.pingpong = 0;
4444 if (tp->rx_opt.num_sacks)
4445 tcp_sack_remove(tp);
4447 tcp_fast_path_check(sk);
4451 else if (!sock_flag(sk, SOCK_DEAD))
4452 sk->sk_data_ready(sk, 0);
4456 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4457 /* A retransmit, 2nd most common case. Force an immediate ack. */
4458 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4459 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4462 tcp_enter_quickack_mode(sk);
4463 inet_csk_schedule_ack(sk);
4469 /* Out of window. F.e. zero window probe. */
4470 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4473 tcp_enter_quickack_mode(sk);
4475 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4476 /* Partial packet, seq < rcv_next < end_seq */
4477 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4478 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4479 TCP_SKB_CB(skb)->end_seq);
4481 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4483 /* If window is closed, drop tail of packet. But after
4484 * remembering D-SACK for its head made in previous line.
4486 if (!tcp_receive_window(tp))
4491 TCP_ECN_check_ce(tp, skb);
4493 if (tcp_try_rmem_schedule(sk, skb->truesize))
4496 /* Disable header prediction. */
4498 inet_csk_schedule_ack(sk);
4500 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4501 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4503 skb_set_owner_r(skb, sk);
4505 if (!skb_peek(&tp->out_of_order_queue)) {
4506 /* Initial out of order segment, build 1 SACK. */
4507 if (tcp_is_sack(tp)) {
4508 tp->rx_opt.num_sacks = 1;
4509 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4510 tp->selective_acks[0].end_seq =
4511 TCP_SKB_CB(skb)->end_seq;
4513 __skb_queue_head(&tp->out_of_order_queue, skb);
4515 struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue);
4516 u32 seq = TCP_SKB_CB(skb)->seq;
4517 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4519 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4520 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4522 if (!tp->rx_opt.num_sacks ||
4523 tp->selective_acks[0].end_seq != seq)
4526 /* Common case: data arrive in order after hole. */
4527 tp->selective_acks[0].end_seq = end_seq;
4531 /* Find place to insert this segment. */
4533 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4535 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4539 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4542 /* Do skb overlap to previous one? */
4543 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4544 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4545 /* All the bits are present. Drop. */
4547 tcp_dsack_set(sk, seq, end_seq);
4550 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4551 /* Partial overlap. */
4552 tcp_dsack_set(sk, seq,
4553 TCP_SKB_CB(skb1)->end_seq);
4555 if (skb_queue_is_first(&tp->out_of_order_queue,
4559 skb1 = skb_queue_prev(
4560 &tp->out_of_order_queue,
4565 __skb_queue_head(&tp->out_of_order_queue, skb);
4567 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4569 /* And clean segments covered by new one as whole. */
4570 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4571 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4573 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4575 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4576 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4580 __skb_unlink(skb1, &tp->out_of_order_queue);
4581 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4582 TCP_SKB_CB(skb1)->end_seq);
4587 if (tcp_is_sack(tp))
4588 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4592 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4593 struct sk_buff_head *list)
4595 struct sk_buff *next = NULL;
4597 if (!skb_queue_is_last(list, skb))
4598 next = skb_queue_next(list, skb);
4600 __skb_unlink(skb, list);
4602 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4607 /* Collapse contiguous sequence of skbs head..tail with
4608 * sequence numbers start..end.
4610 * If tail is NULL, this means until the end of the list.
4612 * Segments with FIN/SYN are not collapsed (only because this
4616 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4617 struct sk_buff *head, struct sk_buff *tail,
4620 struct sk_buff *skb, *n;
4623 /* First, check that queue is collapsible and find
4624 * the point where collapsing can be useful. */
4628 skb_queue_walk_from_safe(list, skb, n) {
4631 /* No new bits? It is possible on ofo queue. */
4632 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4633 skb = tcp_collapse_one(sk, skb, list);
4639 /* The first skb to collapse is:
4641 * - bloated or contains data before "start" or
4642 * overlaps to the next one.
4644 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4645 (tcp_win_from_space(skb->truesize) > skb->len ||
4646 before(TCP_SKB_CB(skb)->seq, start))) {
4647 end_of_skbs = false;
4651 if (!skb_queue_is_last(list, skb)) {
4652 struct sk_buff *next = skb_queue_next(list, skb);
4654 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4655 end_of_skbs = false;
4660 /* Decided to skip this, advance start seq. */
4661 start = TCP_SKB_CB(skb)->end_seq;
4663 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4666 while (before(start, end)) {
4667 struct sk_buff *nskb;
4668 unsigned int header = skb_headroom(skb);
4669 int copy = SKB_MAX_ORDER(header, 0);
4671 /* Too big header? This can happen with IPv6. */
4674 if (end - start < copy)
4676 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4680 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4681 skb_set_network_header(nskb, (skb_network_header(skb) -
4683 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4685 skb_reserve(nskb, header);
4686 memcpy(nskb->head, skb->head, header);
4687 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4688 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4689 __skb_queue_before(list, skb, nskb);
4690 skb_set_owner_r(nskb, sk);
4692 /* Copy data, releasing collapsed skbs. */
4694 int offset = start - TCP_SKB_CB(skb)->seq;
4695 int size = TCP_SKB_CB(skb)->end_seq - start;
4699 size = min(copy, size);
4700 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4702 TCP_SKB_CB(nskb)->end_seq += size;
4706 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4707 skb = tcp_collapse_one(sk, skb, list);
4710 tcp_hdr(skb)->syn ||
4718 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4719 * and tcp_collapse() them until all the queue is collapsed.
4721 static void tcp_collapse_ofo_queue(struct sock *sk)
4723 struct tcp_sock *tp = tcp_sk(sk);
4724 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4725 struct sk_buff *head;
4731 start = TCP_SKB_CB(skb)->seq;
4732 end = TCP_SKB_CB(skb)->end_seq;
4736 struct sk_buff *next = NULL;
4738 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4739 next = skb_queue_next(&tp->out_of_order_queue, skb);
4742 /* Segment is terminated when we see gap or when
4743 * we are at the end of all the queue. */
4745 after(TCP_SKB_CB(skb)->seq, end) ||
4746 before(TCP_SKB_CB(skb)->end_seq, start)) {
4747 tcp_collapse(sk, &tp->out_of_order_queue,
4748 head, skb, start, end);
4752 /* Start new segment */
4753 start = TCP_SKB_CB(skb)->seq;
4754 end = TCP_SKB_CB(skb)->end_seq;
4756 if (before(TCP_SKB_CB(skb)->seq, start))
4757 start = TCP_SKB_CB(skb)->seq;
4758 if (after(TCP_SKB_CB(skb)->end_seq, end))
4759 end = TCP_SKB_CB(skb)->end_seq;
4765 * Purge the out-of-order queue.
4766 * Return true if queue was pruned.
4768 static int tcp_prune_ofo_queue(struct sock *sk)
4770 struct tcp_sock *tp = tcp_sk(sk);
4773 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4774 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4775 __skb_queue_purge(&tp->out_of_order_queue);
4777 /* Reset SACK state. A conforming SACK implementation will
4778 * do the same at a timeout based retransmit. When a connection
4779 * is in a sad state like this, we care only about integrity
4780 * of the connection not performance.
4782 if (tp->rx_opt.sack_ok)
4783 tcp_sack_reset(&tp->rx_opt);
4790 /* Reduce allocated memory if we can, trying to get
4791 * the socket within its memory limits again.
4793 * Return less than zero if we should start dropping frames
4794 * until the socket owning process reads some of the data
4795 * to stabilize the situation.
4797 static int tcp_prune_queue(struct sock *sk)
4799 struct tcp_sock *tp = tcp_sk(sk);
4801 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4803 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4805 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4806 tcp_clamp_window(sk);
4807 else if (tcp_memory_pressure)
4808 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4810 tcp_collapse_ofo_queue(sk);
4811 if (!skb_queue_empty(&sk->sk_receive_queue))
4812 tcp_collapse(sk, &sk->sk_receive_queue,
4813 skb_peek(&sk->sk_receive_queue),
4815 tp->copied_seq, tp->rcv_nxt);
4818 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4821 /* Collapsing did not help, destructive actions follow.
4822 * This must not ever occur. */
4824 tcp_prune_ofo_queue(sk);
4826 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4829 /* If we are really being abused, tell the caller to silently
4830 * drop receive data on the floor. It will get retransmitted
4831 * and hopefully then we'll have sufficient space.
4833 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4835 /* Massive buffer overcommit. */
4840 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4841 * As additional protections, we do not touch cwnd in retransmission phases,
4842 * and if application hit its sndbuf limit recently.
4844 void tcp_cwnd_application_limited(struct sock *sk)
4846 struct tcp_sock *tp = tcp_sk(sk);
4848 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4849 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4850 /* Limited by application or receiver window. */
4851 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4852 u32 win_used = max(tp->snd_cwnd_used, init_win);
4853 if (win_used < tp->snd_cwnd) {
4854 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4855 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4857 tp->snd_cwnd_used = 0;
4859 tp->snd_cwnd_stamp = tcp_time_stamp;
4862 static int tcp_should_expand_sndbuf(struct sock *sk)
4864 struct tcp_sock *tp = tcp_sk(sk);
4866 /* If the user specified a specific send buffer setting, do
4869 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4872 /* If we are under global TCP memory pressure, do not expand. */
4873 if (tcp_memory_pressure)
4876 /* If we are under soft global TCP memory pressure, do not expand. */
4877 if (atomic_long_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4880 /* If we filled the congestion window, do not expand. */
4881 if (tp->packets_out >= tp->snd_cwnd)
4887 /* When incoming ACK allowed to free some skb from write_queue,
4888 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4889 * on the exit from tcp input handler.
4891 * PROBLEM: sndbuf expansion does not work well with largesend.
4893 static void tcp_new_space(struct sock *sk)
4895 struct tcp_sock *tp = tcp_sk(sk);
4897 if (tcp_should_expand_sndbuf(sk)) {
4898 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4899 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4900 int demanded = max_t(unsigned int, tp->snd_cwnd,
4901 tp->reordering + 1);
4902 sndmem *= 2 * demanded;
4903 if (sndmem > sk->sk_sndbuf)
4904 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4905 tp->snd_cwnd_stamp = tcp_time_stamp;
4908 sk->sk_write_space(sk);
4911 static void tcp_check_space(struct sock *sk)
4913 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4914 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4915 if (sk->sk_socket &&
4916 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4921 static inline void tcp_data_snd_check(struct sock *sk)
4923 tcp_push_pending_frames(sk);
4924 tcp_check_space(sk);
4928 * Check if sending an ack is needed.
4930 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4932 struct tcp_sock *tp = tcp_sk(sk);
4934 /* More than one full frame received... */
4935 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4936 /* ... and right edge of window advances far enough.
4937 * (tcp_recvmsg() will send ACK otherwise). Or...
4939 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4940 /* We ACK each frame or... */
4941 tcp_in_quickack_mode(sk) ||
4942 /* We have out of order data. */
4943 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4944 /* Then ack it now */
4947 /* Else, send delayed ack. */
4948 tcp_send_delayed_ack(sk);
4952 static inline void tcp_ack_snd_check(struct sock *sk)
4954 if (!inet_csk_ack_scheduled(sk)) {
4955 /* We sent a data segment already. */
4958 __tcp_ack_snd_check(sk, 1);
4962 * This routine is only called when we have urgent data
4963 * signaled. Its the 'slow' part of tcp_urg. It could be
4964 * moved inline now as tcp_urg is only called from one
4965 * place. We handle URGent data wrong. We have to - as
4966 * BSD still doesn't use the correction from RFC961.
4967 * For 1003.1g we should support a new option TCP_STDURG to permit
4968 * either form (or just set the sysctl tcp_stdurg).
4971 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4973 struct tcp_sock *tp = tcp_sk(sk);
4974 u32 ptr = ntohs(th->urg_ptr);
4976 if (ptr && !sysctl_tcp_stdurg)
4978 ptr += ntohl(th->seq);
4980 /* Ignore urgent data that we've already seen and read. */
4981 if (after(tp->copied_seq, ptr))
4984 /* Do not replay urg ptr.
4986 * NOTE: interesting situation not covered by specs.
4987 * Misbehaving sender may send urg ptr, pointing to segment,
4988 * which we already have in ofo queue. We are not able to fetch
4989 * such data and will stay in TCP_URG_NOTYET until will be eaten
4990 * by recvmsg(). Seems, we are not obliged to handle such wicked
4991 * situations. But it is worth to think about possibility of some
4992 * DoSes using some hypothetical application level deadlock.
4994 if (before(ptr, tp->rcv_nxt))
4997 /* Do we already have a newer (or duplicate) urgent pointer? */
4998 if (tp->urg_data && !after(ptr, tp->urg_seq))
5001 /* Tell the world about our new urgent pointer. */
5004 /* We may be adding urgent data when the last byte read was
5005 * urgent. To do this requires some care. We cannot just ignore
5006 * tp->copied_seq since we would read the last urgent byte again
5007 * as data, nor can we alter copied_seq until this data arrives
5008 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5010 * NOTE. Double Dutch. Rendering to plain English: author of comment
5011 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5012 * and expect that both A and B disappear from stream. This is _wrong_.
5013 * Though this happens in BSD with high probability, this is occasional.
5014 * Any application relying on this is buggy. Note also, that fix "works"
5015 * only in this artificial test. Insert some normal data between A and B and we will
5016 * decline of BSD again. Verdict: it is better to remove to trap
5019 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5020 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5021 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5023 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5024 __skb_unlink(skb, &sk->sk_receive_queue);
5029 tp->urg_data = TCP_URG_NOTYET;
5032 /* Disable header prediction. */
5036 /* This is the 'fast' part of urgent handling. */
5037 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
5039 struct tcp_sock *tp = tcp_sk(sk);
5041 /* Check if we get a new urgent pointer - normally not. */
5043 tcp_check_urg(sk, th);
5045 /* Do we wait for any urgent data? - normally not... */
5046 if (tp->urg_data == TCP_URG_NOTYET) {
5047 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5050 /* Is the urgent pointer pointing into this packet? */
5051 if (ptr < skb->len) {
5053 if (skb_copy_bits(skb, ptr, &tmp, 1))
5055 tp->urg_data = TCP_URG_VALID | tmp;
5056 if (!sock_flag(sk, SOCK_DEAD))
5057 sk->sk_data_ready(sk, 0);
5062 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5064 struct tcp_sock *tp = tcp_sk(sk);
5065 int chunk = skb->len - hlen;
5069 if (skb_csum_unnecessary(skb))
5070 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5072 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5076 tp->ucopy.len -= chunk;
5077 tp->copied_seq += chunk;
5078 tcp_rcv_space_adjust(sk);
5085 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5086 struct sk_buff *skb)
5090 if (sock_owned_by_user(sk)) {
5092 result = __tcp_checksum_complete(skb);
5095 result = __tcp_checksum_complete(skb);
5100 static inline int tcp_checksum_complete_user(struct sock *sk,
5101 struct sk_buff *skb)
5103 return !skb_csum_unnecessary(skb) &&
5104 __tcp_checksum_complete_user(sk, skb);
5107 #ifdef CONFIG_NET_DMA
5108 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5111 struct tcp_sock *tp = tcp_sk(sk);
5112 int chunk = skb->len - hlen;
5114 int copied_early = 0;
5116 if (tp->ucopy.wakeup)
5119 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5120 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5122 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5124 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5126 tp->ucopy.iov, chunk,
5127 tp->ucopy.pinned_list);
5132 tp->ucopy.dma_cookie = dma_cookie;
5135 tp->ucopy.len -= chunk;
5136 tp->copied_seq += chunk;
5137 tcp_rcv_space_adjust(sk);
5139 if ((tp->ucopy.len == 0) ||
5140 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5141 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5142 tp->ucopy.wakeup = 1;
5143 sk->sk_data_ready(sk, 0);
5145 } else if (chunk > 0) {
5146 tp->ucopy.wakeup = 1;
5147 sk->sk_data_ready(sk, 0);
5150 return copied_early;
5152 #endif /* CONFIG_NET_DMA */
5154 /* Does PAWS and seqno based validation of an incoming segment, flags will
5155 * play significant role here.
5157 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5158 struct tcphdr *th, int syn_inerr)
5161 struct tcp_sock *tp = tcp_sk(sk);
5163 /* RFC1323: H1. Apply PAWS check first. */
5164 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5165 tp->rx_opt.saw_tstamp &&
5166 tcp_paws_discard(sk, skb)) {
5168 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5169 tcp_send_dupack(sk, skb);
5172 /* Reset is accepted even if it did not pass PAWS. */
5175 /* Step 1: check sequence number */
5176 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5177 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5178 * (RST) segments are validated by checking their SEQ-fields."
5179 * And page 69: "If an incoming segment is not acceptable,
5180 * an acknowledgment should be sent in reply (unless the RST
5181 * bit is set, if so drop the segment and return)".
5184 tcp_send_dupack(sk, skb);
5188 /* Step 2: check RST bit */
5194 /* ts_recent update must be made after we are sure that the packet
5197 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5199 /* step 3: check security and precedence [ignored] */
5201 /* step 4: Check for a SYN in window. */
5202 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5204 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5205 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5218 * TCP receive function for the ESTABLISHED state.
5220 * It is split into a fast path and a slow path. The fast path is
5222 * - A zero window was announced from us - zero window probing
5223 * is only handled properly in the slow path.
5224 * - Out of order segments arrived.
5225 * - Urgent data is expected.
5226 * - There is no buffer space left
5227 * - Unexpected TCP flags/window values/header lengths are received
5228 * (detected by checking the TCP header against pred_flags)
5229 * - Data is sent in both directions. Fast path only supports pure senders
5230 * or pure receivers (this means either the sequence number or the ack
5231 * value must stay constant)
5232 * - Unexpected TCP option.
5234 * When these conditions are not satisfied it drops into a standard
5235 * receive procedure patterned after RFC793 to handle all cases.
5236 * The first three cases are guaranteed by proper pred_flags setting,
5237 * the rest is checked inline. Fast processing is turned on in
5238 * tcp_data_queue when everything is OK.
5240 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5241 struct tcphdr *th, unsigned len)
5243 struct tcp_sock *tp = tcp_sk(sk);
5247 * Header prediction.
5248 * The code loosely follows the one in the famous
5249 * "30 instruction TCP receive" Van Jacobson mail.
5251 * Van's trick is to deposit buffers into socket queue
5252 * on a device interrupt, to call tcp_recv function
5253 * on the receive process context and checksum and copy
5254 * the buffer to user space. smart...
5256 * Our current scheme is not silly either but we take the
5257 * extra cost of the net_bh soft interrupt processing...
5258 * We do checksum and copy also but from device to kernel.
5261 tp->rx_opt.saw_tstamp = 0;
5263 /* pred_flags is 0xS?10 << 16 + snd_wnd
5264 * if header_prediction is to be made
5265 * 'S' will always be tp->tcp_header_len >> 2
5266 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5267 * turn it off (when there are holes in the receive
5268 * space for instance)
5269 * PSH flag is ignored.
5272 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5273 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5274 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5275 int tcp_header_len = tp->tcp_header_len;
5277 /* Timestamp header prediction: tcp_header_len
5278 * is automatically equal to th->doff*4 due to pred_flags
5282 /* Check timestamp */
5283 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5284 /* No? Slow path! */
5285 if (!tcp_parse_aligned_timestamp(tp, th))
5288 /* If PAWS failed, check it more carefully in slow path */
5289 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5292 /* DO NOT update ts_recent here, if checksum fails
5293 * and timestamp was corrupted part, it will result
5294 * in a hung connection since we will drop all
5295 * future packets due to the PAWS test.
5299 if (len <= tcp_header_len) {
5300 /* Bulk data transfer: sender */
5301 if (len == tcp_header_len) {
5302 /* Predicted packet is in window by definition.
5303 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5304 * Hence, check seq<=rcv_wup reduces to:
5306 if (tcp_header_len ==
5307 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5308 tp->rcv_nxt == tp->rcv_wup)
5309 tcp_store_ts_recent(tp);
5311 /* We know that such packets are checksummed
5314 tcp_ack(sk, skb, 0);
5316 tcp_data_snd_check(sk);
5318 } else { /* Header too small */
5319 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5324 int copied_early = 0;
5326 if (tp->copied_seq == tp->rcv_nxt &&
5327 len - tcp_header_len <= tp->ucopy.len) {
5328 #ifdef CONFIG_NET_DMA
5329 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5334 if (tp->ucopy.task == current &&
5335 sock_owned_by_user(sk) && !copied_early) {
5336 __set_current_state(TASK_RUNNING);
5338 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5342 /* Predicted packet is in window by definition.
5343 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5344 * Hence, check seq<=rcv_wup reduces to:
5346 if (tcp_header_len ==
5347 (sizeof(struct tcphdr) +
5348 TCPOLEN_TSTAMP_ALIGNED) &&
5349 tp->rcv_nxt == tp->rcv_wup)
5350 tcp_store_ts_recent(tp);
5352 tcp_rcv_rtt_measure_ts(sk, skb);
5354 __skb_pull(skb, tcp_header_len);
5355 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5356 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5359 tcp_cleanup_rbuf(sk, skb->len);
5362 if (tcp_checksum_complete_user(sk, skb))
5365 /* Predicted packet is in window by definition.
5366 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5367 * Hence, check seq<=rcv_wup reduces to:
5369 if (tcp_header_len ==
5370 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5371 tp->rcv_nxt == tp->rcv_wup)
5372 tcp_store_ts_recent(tp);
5374 tcp_rcv_rtt_measure_ts(sk, skb);
5376 if ((int)skb->truesize > sk->sk_forward_alloc)
5379 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5381 /* Bulk data transfer: receiver */
5382 __skb_pull(skb, tcp_header_len);
5383 __skb_queue_tail(&sk->sk_receive_queue, skb);
5384 skb_set_owner_r(skb, sk);
5385 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5388 tcp_event_data_recv(sk, skb);
5390 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5391 /* Well, only one small jumplet in fast path... */
5392 tcp_ack(sk, skb, FLAG_DATA);
5393 tcp_data_snd_check(sk);
5394 if (!inet_csk_ack_scheduled(sk))
5398 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5399 __tcp_ack_snd_check(sk, 0);
5401 #ifdef CONFIG_NET_DMA
5403 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5409 sk->sk_data_ready(sk, 0);
5415 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5419 * Standard slow path.
5422 res = tcp_validate_incoming(sk, skb, th, 1);
5427 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5430 tcp_rcv_rtt_measure_ts(sk, skb);
5432 /* Process urgent data. */
5433 tcp_urg(sk, skb, th);
5435 /* step 7: process the segment text */
5436 tcp_data_queue(sk, skb);
5438 tcp_data_snd_check(sk);
5439 tcp_ack_snd_check(sk);
5443 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5449 EXPORT_SYMBOL(tcp_rcv_established);
5451 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5452 struct tcphdr *th, unsigned len)
5455 struct inet_connection_sock *icsk = inet_csk(sk);
5456 struct tcp_sock *tp = tcp_sk(sk);
5457 struct tcp_cookie_values *cvp = tp->cookie_values;
5458 int saved_clamp = tp->rx_opt.mss_clamp;
5460 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5464 * "If the state is SYN-SENT then
5465 * first check the ACK bit
5466 * If the ACK bit is set
5467 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5468 * a reset (unless the RST bit is set, if so drop
5469 * the segment and return)"
5471 * We do not send data with SYN, so that RFC-correct
5474 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5475 goto reset_and_undo;
5477 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5478 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5480 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5481 goto reset_and_undo;
5484 /* Now ACK is acceptable.
5486 * "If the RST bit is set
5487 * If the ACK was acceptable then signal the user "error:
5488 * connection reset", drop the segment, enter CLOSED state,
5489 * delete TCB, and return."
5498 * "fifth, if neither of the SYN or RST bits is set then
5499 * drop the segment and return."
5505 goto discard_and_undo;
5508 * "If the SYN bit is on ...
5509 * are acceptable then ...
5510 * (our SYN has been ACKed), change the connection
5511 * state to ESTABLISHED..."
5514 TCP_ECN_rcv_synack(tp, th);
5516 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5517 tcp_ack(sk, skb, FLAG_SLOWPATH);
5519 /* Ok.. it's good. Set up sequence numbers and
5520 * move to established.
5522 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5523 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5525 /* RFC1323: The window in SYN & SYN/ACK segments is
5528 tp->snd_wnd = ntohs(th->window);
5529 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5531 if (!tp->rx_opt.wscale_ok) {
5532 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5533 tp->window_clamp = min(tp->window_clamp, 65535U);
5536 if (tp->rx_opt.saw_tstamp) {
5537 tp->rx_opt.tstamp_ok = 1;
5538 tp->tcp_header_len =
5539 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5540 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5541 tcp_store_ts_recent(tp);
5543 tp->tcp_header_len = sizeof(struct tcphdr);
5546 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5547 tcp_enable_fack(tp);
5550 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5551 tcp_initialize_rcv_mss(sk);
5553 /* Remember, tcp_poll() does not lock socket!
5554 * Change state from SYN-SENT only after copied_seq
5555 * is initialized. */
5556 tp->copied_seq = tp->rcv_nxt;
5559 cvp->cookie_pair_size > 0 &&
5560 tp->rx_opt.cookie_plus > 0) {
5561 int cookie_size = tp->rx_opt.cookie_plus
5562 - TCPOLEN_COOKIE_BASE;
5563 int cookie_pair_size = cookie_size
5564 + cvp->cookie_desired;
5566 /* A cookie extension option was sent and returned.
5567 * Note that each incoming SYNACK replaces the
5568 * Responder cookie. The initial exchange is most
5569 * fragile, as protection against spoofing relies
5570 * entirely upon the sequence and timestamp (above).
5571 * This replacement strategy allows the correct pair to
5572 * pass through, while any others will be filtered via
5573 * Responder verification later.
5575 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5576 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5577 hash_location, cookie_size);
5578 cvp->cookie_pair_size = cookie_pair_size;
5583 tcp_set_state(sk, TCP_ESTABLISHED);
5585 security_inet_conn_established(sk, skb);
5587 /* Make sure socket is routed, for correct metrics. */
5588 icsk->icsk_af_ops->rebuild_header(sk);
5590 tcp_init_metrics(sk);
5592 tcp_init_congestion_control(sk);
5594 /* Prevent spurious tcp_cwnd_restart() on first data
5597 tp->lsndtime = tcp_time_stamp;
5599 tcp_init_buffer_space(sk);
5601 if (sock_flag(sk, SOCK_KEEPOPEN))
5602 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5604 if (!tp->rx_opt.snd_wscale)
5605 __tcp_fast_path_on(tp, tp->snd_wnd);
5609 if (!sock_flag(sk, SOCK_DEAD)) {
5610 sk->sk_state_change(sk);
5611 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5614 if (sk->sk_write_pending ||
5615 icsk->icsk_accept_queue.rskq_defer_accept ||
5616 icsk->icsk_ack.pingpong) {
5617 /* Save one ACK. Data will be ready after
5618 * several ticks, if write_pending is set.
5620 * It may be deleted, but with this feature tcpdumps
5621 * look so _wonderfully_ clever, that I was not able
5622 * to stand against the temptation 8) --ANK
5624 inet_csk_schedule_ack(sk);
5625 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5626 icsk->icsk_ack.ato = TCP_ATO_MIN;
5627 tcp_incr_quickack(sk);
5628 tcp_enter_quickack_mode(sk);
5629 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5630 TCP_DELACK_MAX, TCP_RTO_MAX);
5641 /* No ACK in the segment */
5645 * "If the RST bit is set
5647 * Otherwise (no ACK) drop the segment and return."
5650 goto discard_and_undo;
5654 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5655 tcp_paws_reject(&tp->rx_opt, 0))
5656 goto discard_and_undo;
5659 /* We see SYN without ACK. It is attempt of
5660 * simultaneous connect with crossed SYNs.
5661 * Particularly, it can be connect to self.
5663 tcp_set_state(sk, TCP_SYN_RECV);
5665 if (tp->rx_opt.saw_tstamp) {
5666 tp->rx_opt.tstamp_ok = 1;
5667 tcp_store_ts_recent(tp);
5668 tp->tcp_header_len =
5669 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5671 tp->tcp_header_len = sizeof(struct tcphdr);
5674 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5675 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5677 /* RFC1323: The window in SYN & SYN/ACK segments is
5680 tp->snd_wnd = ntohs(th->window);
5681 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5682 tp->max_window = tp->snd_wnd;
5684 TCP_ECN_rcv_syn(tp, th);
5687 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5688 tcp_initialize_rcv_mss(sk);
5690 tcp_send_synack(sk);
5692 /* Note, we could accept data and URG from this segment.
5693 * There are no obstacles to make this.
5695 * However, if we ignore data in ACKless segments sometimes,
5696 * we have no reasons to accept it sometimes.
5697 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5698 * is not flawless. So, discard packet for sanity.
5699 * Uncomment this return to process the data.
5706 /* "fifth, if neither of the SYN or RST bits is set then
5707 * drop the segment and return."
5711 tcp_clear_options(&tp->rx_opt);
5712 tp->rx_opt.mss_clamp = saved_clamp;
5716 tcp_clear_options(&tp->rx_opt);
5717 tp->rx_opt.mss_clamp = saved_clamp;
5722 * This function implements the receiving procedure of RFC 793 for
5723 * all states except ESTABLISHED and TIME_WAIT.
5724 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5725 * address independent.
5728 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5729 struct tcphdr *th, unsigned len)
5731 struct tcp_sock *tp = tcp_sk(sk);
5732 struct inet_connection_sock *icsk = inet_csk(sk);
5736 tp->rx_opt.saw_tstamp = 0;
5738 switch (sk->sk_state) {
5750 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5753 /* Now we have several options: In theory there is
5754 * nothing else in the frame. KA9Q has an option to
5755 * send data with the syn, BSD accepts data with the
5756 * syn up to the [to be] advertised window and
5757 * Solaris 2.1 gives you a protocol error. For now
5758 * we just ignore it, that fits the spec precisely
5759 * and avoids incompatibilities. It would be nice in
5760 * future to drop through and process the data.
5762 * Now that TTCP is starting to be used we ought to
5764 * But, this leaves one open to an easy denial of
5765 * service attack, and SYN cookies can't defend
5766 * against this problem. So, we drop the data
5767 * in the interest of security over speed unless
5768 * it's still in use.
5776 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5780 /* Do step6 onward by hand. */
5781 tcp_urg(sk, skb, th);
5783 tcp_data_snd_check(sk);
5787 res = tcp_validate_incoming(sk, skb, th, 0);
5791 /* step 5: check the ACK field */
5793 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5795 switch (sk->sk_state) {
5798 tp->copied_seq = tp->rcv_nxt;
5800 tcp_set_state(sk, TCP_ESTABLISHED);
5801 sk->sk_state_change(sk);
5803 /* Note, that this wakeup is only for marginal
5804 * crossed SYN case. Passively open sockets
5805 * are not waked up, because sk->sk_sleep ==
5806 * NULL and sk->sk_socket == NULL.
5810 SOCK_WAKE_IO, POLL_OUT);
5812 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5813 tp->snd_wnd = ntohs(th->window) <<
5814 tp->rx_opt.snd_wscale;
5815 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5817 if (tp->rx_opt.tstamp_ok)
5818 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5820 /* Make sure socket is routed, for
5823 icsk->icsk_af_ops->rebuild_header(sk);
5825 tcp_init_metrics(sk);
5827 tcp_init_congestion_control(sk);
5829 /* Prevent spurious tcp_cwnd_restart() on
5830 * first data packet.
5832 tp->lsndtime = tcp_time_stamp;
5835 tcp_initialize_rcv_mss(sk);
5836 tcp_init_buffer_space(sk);
5837 tcp_fast_path_on(tp);
5844 if (tp->snd_una == tp->write_seq) {
5845 tcp_set_state(sk, TCP_FIN_WAIT2);
5846 sk->sk_shutdown |= SEND_SHUTDOWN;
5847 dst_confirm(__sk_dst_get(sk));
5849 if (!sock_flag(sk, SOCK_DEAD))
5850 /* Wake up lingering close() */
5851 sk->sk_state_change(sk);
5855 if (tp->linger2 < 0 ||
5856 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5857 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5859 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5863 tmo = tcp_fin_time(sk);
5864 if (tmo > TCP_TIMEWAIT_LEN) {
5865 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5866 } else if (th->fin || sock_owned_by_user(sk)) {
5867 /* Bad case. We could lose such FIN otherwise.
5868 * It is not a big problem, but it looks confusing
5869 * and not so rare event. We still can lose it now,
5870 * if it spins in bh_lock_sock(), but it is really
5873 inet_csk_reset_keepalive_timer(sk, tmo);
5875 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5883 if (tp->snd_una == tp->write_seq) {
5884 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5890 if (tp->snd_una == tp->write_seq) {
5891 tcp_update_metrics(sk);
5900 /* step 6: check the URG bit */
5901 tcp_urg(sk, skb, th);
5903 /* step 7: process the segment text */
5904 switch (sk->sk_state) {
5905 case TCP_CLOSE_WAIT:
5908 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5912 /* RFC 793 says to queue data in these states,
5913 * RFC 1122 says we MUST send a reset.
5914 * BSD 4.4 also does reset.
5916 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5917 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5918 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5919 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5925 case TCP_ESTABLISHED:
5926 tcp_data_queue(sk, skb);
5931 /* tcp_data could move socket to TIME-WAIT */
5932 if (sk->sk_state != TCP_CLOSE) {
5933 tcp_data_snd_check(sk);
5934 tcp_ack_snd_check(sk);
5943 EXPORT_SYMBOL(tcp_rcv_state_process);