tcp_input.c revision 974c12360dfe6ab01201fe9e708e7755c413f8b6
1/*
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.
5 *
6 *		Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors:	Ross Biro
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>
19 */
20
21/*
22 * Changes:
23 *		Pedro Roque	:	Fast Retransmit/Recovery.
24 *					Two receive queues.
25 *					Retransmit queue handled by TCP.
26 *					Better retransmit timer handling.
27 *					New congestion avoidance.
28 *					Header prediction.
29 *					Variable renaming.
30 *
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
44 *					timestamps.
45 *		Andrey Savochkin:	Check sequence numbers correctly when
46 *					removing SACKs due to in sequence incoming
47 *					data segments.
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
55 *					fast path.
56 *		J Hadi Salim:		ECN support
57 *	 	Andrei Gurtov,
58 *		Pasi Sarolahti,
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
62 */
63
64#include <linux/mm.h>
65#include <linux/slab.h>
66#include <linux/module.h>
67#include <linux/sysctl.h>
68#include <linux/kernel.h>
69#include <net/dst.h>
70#include <net/tcp.h>
71#include <net/inet_common.h>
72#include <linux/ipsec.h>
73#include <asm/unaligned.h>
74#include <net/netdma.h>
75
76int sysctl_tcp_timestamps __read_mostly = 1;
77int sysctl_tcp_window_scaling __read_mostly = 1;
78int sysctl_tcp_sack __read_mostly = 1;
79int sysctl_tcp_fack __read_mostly = 1;
80int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
81EXPORT_SYMBOL(sysctl_tcp_reordering);
82int sysctl_tcp_ecn __read_mostly = 2;
83EXPORT_SYMBOL(sysctl_tcp_ecn);
84int sysctl_tcp_dsack __read_mostly = 1;
85int sysctl_tcp_app_win __read_mostly = 31;
86int sysctl_tcp_adv_win_scale __read_mostly = 2;
87EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
88
89int sysctl_tcp_stdurg __read_mostly;
90int sysctl_tcp_rfc1337 __read_mostly;
91int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
92int sysctl_tcp_frto __read_mostly = 2;
93int sysctl_tcp_frto_response __read_mostly;
94int sysctl_tcp_nometrics_save __read_mostly;
95
96int sysctl_tcp_thin_dupack __read_mostly;
97
98int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
99int sysctl_tcp_abc __read_mostly;
100
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_SLOWPATH		0x100 /* Do not skip RFC checks for window update.*/
109#define FLAG_ONLY_ORIG_SACKED	0x200 /* SACKs only non-rexmit sent before RTO */
110#define FLAG_SND_UNA_ADVANCED	0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
111#define FLAG_DSACKING_ACK	0x800 /* SACK blocks contained D-SACK info */
112#define FLAG_NONHEAD_RETRANS_ACKED	0x1000 /* Non-head rexmitted data was ACKed */
113#define FLAG_SACK_RENEGING	0x2000 /* snd_una advanced to a sacked seq */
114
115#define FLAG_ACKED		(FLAG_DATA_ACKED|FLAG_SYN_ACKED)
116#define FLAG_NOT_DUP		(FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
117#define FLAG_CA_ALERT		(FLAG_DATA_SACKED|FLAG_ECE)
118#define FLAG_FORWARD_PROGRESS	(FLAG_ACKED|FLAG_DATA_SACKED)
119#define FLAG_ANY_PROGRESS	(FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
120
121#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
123
124/* Adapt the MSS value used to make delayed ack decision to the
125 * real world.
126 */
127static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
128{
129	struct inet_connection_sock *icsk = inet_csk(sk);
130	const unsigned int lss = icsk->icsk_ack.last_seg_size;
131	unsigned int len;
132
133	icsk->icsk_ack.last_seg_size = 0;
134
135	/* skb->len may jitter because of SACKs, even if peer
136	 * sends good full-sized frames.
137	 */
138	len = skb_shinfo(skb)->gso_size ? : skb->len;
139	if (len >= icsk->icsk_ack.rcv_mss) {
140		icsk->icsk_ack.rcv_mss = len;
141	} else {
142		/* Otherwise, we make more careful check taking into account,
143		 * that SACKs block is variable.
144		 *
145		 * "len" is invariant segment length, including TCP header.
146		 */
147		len += skb->data - skb_transport_header(skb);
148		if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
149		    /* If PSH is not set, packet should be
150		     * full sized, provided peer TCP is not badly broken.
151		     * This observation (if it is correct 8)) allows
152		     * to handle super-low mtu links fairly.
153		     */
154		    (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
155		     !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
156			/* Subtract also invariant (if peer is RFC compliant),
157			 * tcp header plus fixed timestamp option length.
158			 * Resulting "len" is MSS free of SACK jitter.
159			 */
160			len -= tcp_sk(sk)->tcp_header_len;
161			icsk->icsk_ack.last_seg_size = len;
162			if (len == lss) {
163				icsk->icsk_ack.rcv_mss = len;
164				return;
165			}
166		}
167		if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
168			icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
169		icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
170	}
171}
172
173static void tcp_incr_quickack(struct sock *sk)
174{
175	struct inet_connection_sock *icsk = inet_csk(sk);
176	unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
177
178	if (quickacks == 0)
179		quickacks = 2;
180	if (quickacks > icsk->icsk_ack.quick)
181		icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
182}
183
184static void tcp_enter_quickack_mode(struct sock *sk)
185{
186	struct inet_connection_sock *icsk = inet_csk(sk);
187	tcp_incr_quickack(sk);
188	icsk->icsk_ack.pingpong = 0;
189	icsk->icsk_ack.ato = TCP_ATO_MIN;
190}
191
192/* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
194 */
195
196static inline int tcp_in_quickack_mode(const struct sock *sk)
197{
198	const struct inet_connection_sock *icsk = inet_csk(sk);
199	return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
200}
201
202static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
203{
204	if (tp->ecn_flags & TCP_ECN_OK)
205		tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
206}
207
208static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
209{
210	if (tcp_hdr(skb)->cwr)
211		tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
212}
213
214static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
215{
216	tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
217}
218
219static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
220{
221	if (!(tp->ecn_flags & TCP_ECN_OK))
222		return;
223
224	switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
225	case INET_ECN_NOT_ECT:
226		/* Funny extension: if ECT is not set on a segment,
227		 * and we already seen ECT on a previous segment,
228		 * it is probably a retransmit.
229		 */
230		if (tp->ecn_flags & TCP_ECN_SEEN)
231			tcp_enter_quickack_mode((struct sock *)tp);
232		break;
233	case INET_ECN_CE:
234		tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
235		/* fallinto */
236	default:
237		tp->ecn_flags |= TCP_ECN_SEEN;
238	}
239}
240
241static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
242{
243	if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
244		tp->ecn_flags &= ~TCP_ECN_OK;
245}
246
247static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
248{
249	if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
250		tp->ecn_flags &= ~TCP_ECN_OK;
251}
252
253static inline int TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
254{
255	if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
256		return 1;
257	return 0;
258}
259
260/* Buffer size and advertised window tuning.
261 *
262 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
263 */
264
265static void tcp_fixup_sndbuf(struct sock *sk)
266{
267	int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
268
269	sndmem *= TCP_INIT_CWND;
270	if (sk->sk_sndbuf < sndmem)
271		sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
272}
273
274/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
275 *
276 * All tcp_full_space() is split to two parts: "network" buffer, allocated
277 * forward and advertised in receiver window (tp->rcv_wnd) and
278 * "application buffer", required to isolate scheduling/application
279 * latencies from network.
280 * window_clamp is maximal advertised window. It can be less than
281 * tcp_full_space(), in this case tcp_full_space() - window_clamp
282 * is reserved for "application" buffer. The less window_clamp is
283 * the smoother our behaviour from viewpoint of network, but the lower
284 * throughput and the higher sensitivity of the connection to losses. 8)
285 *
286 * rcv_ssthresh is more strict window_clamp used at "slow start"
287 * phase to predict further behaviour of this connection.
288 * It is used for two goals:
289 * - to enforce header prediction at sender, even when application
290 *   requires some significant "application buffer". It is check #1.
291 * - to prevent pruning of receive queue because of misprediction
292 *   of receiver window. Check #2.
293 *
294 * The scheme does not work when sender sends good segments opening
295 * window and then starts to feed us spaghetti. But it should work
296 * in common situations. Otherwise, we have to rely on queue collapsing.
297 */
298
299/* Slow part of check#2. */
300static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
301{
302	struct tcp_sock *tp = tcp_sk(sk);
303	/* Optimize this! */
304	int truesize = tcp_win_from_space(skb->truesize) >> 1;
305	int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
306
307	while (tp->rcv_ssthresh <= window) {
308		if (truesize <= skb->len)
309			return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
310
311		truesize >>= 1;
312		window >>= 1;
313	}
314	return 0;
315}
316
317static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
318{
319	struct tcp_sock *tp = tcp_sk(sk);
320
321	/* Check #1 */
322	if (tp->rcv_ssthresh < tp->window_clamp &&
323	    (int)tp->rcv_ssthresh < tcp_space(sk) &&
324	    !sk_under_memory_pressure(sk)) {
325		int incr;
326
327		/* Check #2. Increase window, if skb with such overhead
328		 * will fit to rcvbuf in future.
329		 */
330		if (tcp_win_from_space(skb->truesize) <= skb->len)
331			incr = 2 * tp->advmss;
332		else
333			incr = __tcp_grow_window(sk, skb);
334
335		if (incr) {
336			tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
337					       tp->window_clamp);
338			inet_csk(sk)->icsk_ack.quick |= 1;
339		}
340	}
341}
342
343/* 3. Tuning rcvbuf, when connection enters established state. */
344
345static void tcp_fixup_rcvbuf(struct sock *sk)
346{
347	u32 mss = tcp_sk(sk)->advmss;
348	u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
349	int rcvmem;
350
351	/* Limit to 10 segments if mss <= 1460,
352	 * or 14600/mss segments, with a minimum of two segments.
353	 */
354	if (mss > 1460)
355		icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
356
357	rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
358	while (tcp_win_from_space(rcvmem) < mss)
359		rcvmem += 128;
360
361	rcvmem *= icwnd;
362
363	if (sk->sk_rcvbuf < rcvmem)
364		sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
365}
366
367/* 4. Try to fixup all. It is made immediately after connection enters
368 *    established state.
369 */
370static void tcp_init_buffer_space(struct sock *sk)
371{
372	struct tcp_sock *tp = tcp_sk(sk);
373	int maxwin;
374
375	if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
376		tcp_fixup_rcvbuf(sk);
377	if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
378		tcp_fixup_sndbuf(sk);
379
380	tp->rcvq_space.space = tp->rcv_wnd;
381
382	maxwin = tcp_full_space(sk);
383
384	if (tp->window_clamp >= maxwin) {
385		tp->window_clamp = maxwin;
386
387		if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
388			tp->window_clamp = max(maxwin -
389					       (maxwin >> sysctl_tcp_app_win),
390					       4 * tp->advmss);
391	}
392
393	/* Force reservation of one segment. */
394	if (sysctl_tcp_app_win &&
395	    tp->window_clamp > 2 * tp->advmss &&
396	    tp->window_clamp + tp->advmss > maxwin)
397		tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
398
399	tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
400	tp->snd_cwnd_stamp = tcp_time_stamp;
401}
402
403/* 5. Recalculate window clamp after socket hit its memory bounds. */
404static void tcp_clamp_window(struct sock *sk)
405{
406	struct tcp_sock *tp = tcp_sk(sk);
407	struct inet_connection_sock *icsk = inet_csk(sk);
408
409	icsk->icsk_ack.quick = 0;
410
411	if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
412	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
413	    !sk_under_memory_pressure(sk) &&
414	    sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
415		sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
416				    sysctl_tcp_rmem[2]);
417	}
418	if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
419		tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
420}
421
422/* Initialize RCV_MSS value.
423 * RCV_MSS is an our guess about MSS used by the peer.
424 * We haven't any direct information about the MSS.
425 * It's better to underestimate the RCV_MSS rather than overestimate.
426 * Overestimations make us ACKing less frequently than needed.
427 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
428 */
429void tcp_initialize_rcv_mss(struct sock *sk)
430{
431	const struct tcp_sock *tp = tcp_sk(sk);
432	unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
433
434	hint = min(hint, tp->rcv_wnd / 2);
435	hint = min(hint, TCP_MSS_DEFAULT);
436	hint = max(hint, TCP_MIN_MSS);
437
438	inet_csk(sk)->icsk_ack.rcv_mss = hint;
439}
440EXPORT_SYMBOL(tcp_initialize_rcv_mss);
441
442/* Receiver "autotuning" code.
443 *
444 * The algorithm for RTT estimation w/o timestamps is based on
445 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
446 * <http://public.lanl.gov/radiant/pubs.html#DRS>
447 *
448 * More detail on this code can be found at
449 * <http://staff.psc.edu/jheffner/>,
450 * though this reference is out of date.  A new paper
451 * is pending.
452 */
453static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
454{
455	u32 new_sample = tp->rcv_rtt_est.rtt;
456	long m = sample;
457
458	if (m == 0)
459		m = 1;
460
461	if (new_sample != 0) {
462		/* If we sample in larger samples in the non-timestamp
463		 * case, we could grossly overestimate the RTT especially
464		 * with chatty applications or bulk transfer apps which
465		 * are stalled on filesystem I/O.
466		 *
467		 * Also, since we are only going for a minimum in the
468		 * non-timestamp case, we do not smooth things out
469		 * else with timestamps disabled convergence takes too
470		 * long.
471		 */
472		if (!win_dep) {
473			m -= (new_sample >> 3);
474			new_sample += m;
475		} else if (m < new_sample)
476			new_sample = m << 3;
477	} else {
478		/* No previous measure. */
479		new_sample = m << 3;
480	}
481
482	if (tp->rcv_rtt_est.rtt != new_sample)
483		tp->rcv_rtt_est.rtt = new_sample;
484}
485
486static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
487{
488	if (tp->rcv_rtt_est.time == 0)
489		goto new_measure;
490	if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
491		return;
492	tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
493
494new_measure:
495	tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
496	tp->rcv_rtt_est.time = tcp_time_stamp;
497}
498
499static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
500					  const struct sk_buff *skb)
501{
502	struct tcp_sock *tp = tcp_sk(sk);
503	if (tp->rx_opt.rcv_tsecr &&
504	    (TCP_SKB_CB(skb)->end_seq -
505	     TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
506		tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
507}
508
509/*
510 * This function should be called every time data is copied to user space.
511 * It calculates the appropriate TCP receive buffer space.
512 */
513void tcp_rcv_space_adjust(struct sock *sk)
514{
515	struct tcp_sock *tp = tcp_sk(sk);
516	int time;
517	int space;
518
519	if (tp->rcvq_space.time == 0)
520		goto new_measure;
521
522	time = tcp_time_stamp - tp->rcvq_space.time;
523	if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
524		return;
525
526	space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
527
528	space = max(tp->rcvq_space.space, space);
529
530	if (tp->rcvq_space.space != space) {
531		int rcvmem;
532
533		tp->rcvq_space.space = space;
534
535		if (sysctl_tcp_moderate_rcvbuf &&
536		    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
537			int new_clamp = space;
538
539			/* Receive space grows, normalize in order to
540			 * take into account packet headers and sk_buff
541			 * structure overhead.
542			 */
543			space /= tp->advmss;
544			if (!space)
545				space = 1;
546			rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
547			while (tcp_win_from_space(rcvmem) < tp->advmss)
548				rcvmem += 128;
549			space *= rcvmem;
550			space = min(space, sysctl_tcp_rmem[2]);
551			if (space > sk->sk_rcvbuf) {
552				sk->sk_rcvbuf = space;
553
554				/* Make the window clamp follow along.  */
555				tp->window_clamp = new_clamp;
556			}
557		}
558	}
559
560new_measure:
561	tp->rcvq_space.seq = tp->copied_seq;
562	tp->rcvq_space.time = tcp_time_stamp;
563}
564
565/* There is something which you must keep in mind when you analyze the
566 * behavior of the tp->ato delayed ack timeout interval.  When a
567 * connection starts up, we want to ack as quickly as possible.  The
568 * problem is that "good" TCP's do slow start at the beginning of data
569 * transmission.  The means that until we send the first few ACK's the
570 * sender will sit on his end and only queue most of his data, because
571 * he can only send snd_cwnd unacked packets at any given time.  For
572 * each ACK we send, he increments snd_cwnd and transmits more of his
573 * queue.  -DaveM
574 */
575static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
576{
577	struct tcp_sock *tp = tcp_sk(sk);
578	struct inet_connection_sock *icsk = inet_csk(sk);
579	u32 now;
580
581	inet_csk_schedule_ack(sk);
582
583	tcp_measure_rcv_mss(sk, skb);
584
585	tcp_rcv_rtt_measure(tp);
586
587	now = tcp_time_stamp;
588
589	if (!icsk->icsk_ack.ato) {
590		/* The _first_ data packet received, initialize
591		 * delayed ACK engine.
592		 */
593		tcp_incr_quickack(sk);
594		icsk->icsk_ack.ato = TCP_ATO_MIN;
595	} else {
596		int m = now - icsk->icsk_ack.lrcvtime;
597
598		if (m <= TCP_ATO_MIN / 2) {
599			/* The fastest case is the first. */
600			icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
601		} else if (m < icsk->icsk_ack.ato) {
602			icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
603			if (icsk->icsk_ack.ato > icsk->icsk_rto)
604				icsk->icsk_ack.ato = icsk->icsk_rto;
605		} else if (m > icsk->icsk_rto) {
606			/* Too long gap. Apparently sender failed to
607			 * restart window, so that we send ACKs quickly.
608			 */
609			tcp_incr_quickack(sk);
610			sk_mem_reclaim(sk);
611		}
612	}
613	icsk->icsk_ack.lrcvtime = now;
614
615	TCP_ECN_check_ce(tp, skb);
616
617	if (skb->len >= 128)
618		tcp_grow_window(sk, skb);
619}
620
621/* Called to compute a smoothed rtt estimate. The data fed to this
622 * routine either comes from timestamps, or from segments that were
623 * known _not_ to have been retransmitted [see Karn/Partridge
624 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
625 * piece by Van Jacobson.
626 * NOTE: the next three routines used to be one big routine.
627 * To save cycles in the RFC 1323 implementation it was better to break
628 * it up into three procedures. -- erics
629 */
630static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
631{
632	struct tcp_sock *tp = tcp_sk(sk);
633	long m = mrtt; /* RTT */
634
635	/*	The following amusing code comes from Jacobson's
636	 *	article in SIGCOMM '88.  Note that rtt and mdev
637	 *	are scaled versions of rtt and mean deviation.
638	 *	This is designed to be as fast as possible
639	 *	m stands for "measurement".
640	 *
641	 *	On a 1990 paper the rto value is changed to:
642	 *	RTO = rtt + 4 * mdev
643	 *
644	 * Funny. This algorithm seems to be very broken.
645	 * These formulae increase RTO, when it should be decreased, increase
646	 * too slowly, when it should be increased quickly, decrease too quickly
647	 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
648	 * does not matter how to _calculate_ it. Seems, it was trap
649	 * that VJ failed to avoid. 8)
650	 */
651	if (m == 0)
652		m = 1;
653	if (tp->srtt != 0) {
654		m -= (tp->srtt >> 3);	/* m is now error in rtt est */
655		tp->srtt += m;		/* rtt = 7/8 rtt + 1/8 new */
656		if (m < 0) {
657			m = -m;		/* m is now abs(error) */
658			m -= (tp->mdev >> 2);   /* similar update on mdev */
659			/* This is similar to one of Eifel findings.
660			 * Eifel blocks mdev updates when rtt decreases.
661			 * This solution is a bit different: we use finer gain
662			 * for mdev in this case (alpha*beta).
663			 * Like Eifel it also prevents growth of rto,
664			 * but also it limits too fast rto decreases,
665			 * happening in pure Eifel.
666			 */
667			if (m > 0)
668				m >>= 3;
669		} else {
670			m -= (tp->mdev >> 2);   /* similar update on mdev */
671		}
672		tp->mdev += m;	    	/* mdev = 3/4 mdev + 1/4 new */
673		if (tp->mdev > tp->mdev_max) {
674			tp->mdev_max = tp->mdev;
675			if (tp->mdev_max > tp->rttvar)
676				tp->rttvar = tp->mdev_max;
677		}
678		if (after(tp->snd_una, tp->rtt_seq)) {
679			if (tp->mdev_max < tp->rttvar)
680				tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
681			tp->rtt_seq = tp->snd_nxt;
682			tp->mdev_max = tcp_rto_min(sk);
683		}
684	} else {
685		/* no previous measure. */
686		tp->srtt = m << 3;	/* take the measured time to be rtt */
687		tp->mdev = m << 1;	/* make sure rto = 3*rtt */
688		tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
689		tp->rtt_seq = tp->snd_nxt;
690	}
691}
692
693/* Calculate rto without backoff.  This is the second half of Van Jacobson's
694 * routine referred to above.
695 */
696static inline void tcp_set_rto(struct sock *sk)
697{
698	const struct tcp_sock *tp = tcp_sk(sk);
699	/* Old crap is replaced with new one. 8)
700	 *
701	 * More seriously:
702	 * 1. If rtt variance happened to be less 50msec, it is hallucination.
703	 *    It cannot be less due to utterly erratic ACK generation made
704	 *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
705	 *    to do with delayed acks, because at cwnd>2 true delack timeout
706	 *    is invisible. Actually, Linux-2.4 also generates erratic
707	 *    ACKs in some circumstances.
708	 */
709	inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
710
711	/* 2. Fixups made earlier cannot be right.
712	 *    If we do not estimate RTO correctly without them,
713	 *    all the algo is pure shit and should be replaced
714	 *    with correct one. It is exactly, which we pretend to do.
715	 */
716
717	/* NOTE: clamping at TCP_RTO_MIN is not required, current algo
718	 * guarantees that rto is higher.
719	 */
720	tcp_bound_rto(sk);
721}
722
723/* Save metrics learned by this TCP session.
724   This function is called only, when TCP finishes successfully
725   i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
726 */
727void tcp_update_metrics(struct sock *sk)
728{
729	struct tcp_sock *tp = tcp_sk(sk);
730	struct dst_entry *dst = __sk_dst_get(sk);
731
732	if (sysctl_tcp_nometrics_save)
733		return;
734
735	dst_confirm(dst);
736
737	if (dst && (dst->flags & DST_HOST)) {
738		const struct inet_connection_sock *icsk = inet_csk(sk);
739		int m;
740		unsigned long rtt;
741
742		if (icsk->icsk_backoff || !tp->srtt) {
743			/* This session failed to estimate rtt. Why?
744			 * Probably, no packets returned in time.
745			 * Reset our results.
746			 */
747			if (!(dst_metric_locked(dst, RTAX_RTT)))
748				dst_metric_set(dst, RTAX_RTT, 0);
749			return;
750		}
751
752		rtt = dst_metric_rtt(dst, RTAX_RTT);
753		m = rtt - tp->srtt;
754
755		/* If newly calculated rtt larger than stored one,
756		 * store new one. Otherwise, use EWMA. Remember,
757		 * rtt overestimation is always better than underestimation.
758		 */
759		if (!(dst_metric_locked(dst, RTAX_RTT))) {
760			if (m <= 0)
761				set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
762			else
763				set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
764		}
765
766		if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
767			unsigned long var;
768			if (m < 0)
769				m = -m;
770
771			/* Scale deviation to rttvar fixed point */
772			m >>= 1;
773			if (m < tp->mdev)
774				m = tp->mdev;
775
776			var = dst_metric_rtt(dst, RTAX_RTTVAR);
777			if (m >= var)
778				var = m;
779			else
780				var -= (var - m) >> 2;
781
782			set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
783		}
784
785		if (tcp_in_initial_slowstart(tp)) {
786			/* Slow start still did not finish. */
787			if (dst_metric(dst, RTAX_SSTHRESH) &&
788			    !dst_metric_locked(dst, RTAX_SSTHRESH) &&
789			    (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
790				dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
791			if (!dst_metric_locked(dst, RTAX_CWND) &&
792			    tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
793				dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
794		} else if (tp->snd_cwnd > tp->snd_ssthresh &&
795			   icsk->icsk_ca_state == TCP_CA_Open) {
796			/* Cong. avoidance phase, cwnd is reliable. */
797			if (!dst_metric_locked(dst, RTAX_SSTHRESH))
798				dst_metric_set(dst, RTAX_SSTHRESH,
799					       max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
800			if (!dst_metric_locked(dst, RTAX_CWND))
801				dst_metric_set(dst, RTAX_CWND,
802					       (dst_metric(dst, RTAX_CWND) +
803						tp->snd_cwnd) >> 1);
804		} else {
805			/* Else slow start did not finish, cwnd is non-sense,
806			   ssthresh may be also invalid.
807			 */
808			if (!dst_metric_locked(dst, RTAX_CWND))
809				dst_metric_set(dst, RTAX_CWND,
810					       (dst_metric(dst, RTAX_CWND) +
811						tp->snd_ssthresh) >> 1);
812			if (dst_metric(dst, RTAX_SSTHRESH) &&
813			    !dst_metric_locked(dst, RTAX_SSTHRESH) &&
814			    tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
815				dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
816		}
817
818		if (!dst_metric_locked(dst, RTAX_REORDERING)) {
819			if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
820			    tp->reordering != sysctl_tcp_reordering)
821				dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
822		}
823	}
824}
825
826__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
827{
828	__u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
829
830	if (!cwnd)
831		cwnd = TCP_INIT_CWND;
832	return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
833}
834
835/* Set slow start threshold and cwnd not falling to slow start */
836void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
837{
838	struct tcp_sock *tp = tcp_sk(sk);
839	const struct inet_connection_sock *icsk = inet_csk(sk);
840
841	tp->prior_ssthresh = 0;
842	tp->bytes_acked = 0;
843	if (icsk->icsk_ca_state < TCP_CA_CWR) {
844		tp->undo_marker = 0;
845		if (set_ssthresh)
846			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
847		tp->snd_cwnd = min(tp->snd_cwnd,
848				   tcp_packets_in_flight(tp) + 1U);
849		tp->snd_cwnd_cnt = 0;
850		tp->high_seq = tp->snd_nxt;
851		tp->snd_cwnd_stamp = tcp_time_stamp;
852		TCP_ECN_queue_cwr(tp);
853
854		tcp_set_ca_state(sk, TCP_CA_CWR);
855	}
856}
857
858/*
859 * Packet counting of FACK is based on in-order assumptions, therefore TCP
860 * disables it when reordering is detected
861 */
862static void tcp_disable_fack(struct tcp_sock *tp)
863{
864	/* RFC3517 uses different metric in lost marker => reset on change */
865	if (tcp_is_fack(tp))
866		tp->lost_skb_hint = NULL;
867	tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
868}
869
870/* Take a notice that peer is sending D-SACKs */
871static void tcp_dsack_seen(struct tcp_sock *tp)
872{
873	tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
874}
875
876/* Initialize metrics on socket. */
877
878static void tcp_init_metrics(struct sock *sk)
879{
880	struct tcp_sock *tp = tcp_sk(sk);
881	struct dst_entry *dst = __sk_dst_get(sk);
882
883	if (dst == NULL)
884		goto reset;
885
886	dst_confirm(dst);
887
888	if (dst_metric_locked(dst, RTAX_CWND))
889		tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
890	if (dst_metric(dst, RTAX_SSTHRESH)) {
891		tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
892		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
893			tp->snd_ssthresh = tp->snd_cwnd_clamp;
894	} else {
895		/* ssthresh may have been reduced unnecessarily during.
896		 * 3WHS. Restore it back to its initial default.
897		 */
898		tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
899	}
900	if (dst_metric(dst, RTAX_REORDERING) &&
901	    tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
902		tcp_disable_fack(tp);
903		tp->reordering = dst_metric(dst, RTAX_REORDERING);
904	}
905
906	if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
907		goto reset;
908
909	/* Initial rtt is determined from SYN,SYN-ACK.
910	 * The segment is small and rtt may appear much
911	 * less than real one. Use per-dst memory
912	 * to make it more realistic.
913	 *
914	 * A bit of theory. RTT is time passed after "normal" sized packet
915	 * is sent until it is ACKed. In normal circumstances sending small
916	 * packets force peer to delay ACKs and calculation is correct too.
917	 * The algorithm is adaptive and, provided we follow specs, it
918	 * NEVER underestimate RTT. BUT! If peer tries to make some clever
919	 * tricks sort of "quick acks" for time long enough to decrease RTT
920	 * to low value, and then abruptly stops to do it and starts to delay
921	 * ACKs, wait for troubles.
922	 */
923	if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
924		tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
925		tp->rtt_seq = tp->snd_nxt;
926	}
927	if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
928		tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
929		tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
930	}
931	tcp_set_rto(sk);
932reset:
933	if (tp->srtt == 0) {
934		/* RFC2988bis: We've failed to get a valid RTT sample from
935		 * 3WHS. This is most likely due to retransmission,
936		 * including spurious one. Reset the RTO back to 3secs
937		 * from the more aggressive 1sec to avoid more spurious
938		 * retransmission.
939		 */
940		tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
941		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
942	}
943	/* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
944	 * retransmitted. In light of RFC2988bis' more aggressive 1sec
945	 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
946	 * retransmission has occurred.
947	 */
948	if (tp->total_retrans > 1)
949		tp->snd_cwnd = 1;
950	else
951		tp->snd_cwnd = tcp_init_cwnd(tp, dst);
952	tp->snd_cwnd_stamp = tcp_time_stamp;
953}
954
955static void tcp_update_reordering(struct sock *sk, const int metric,
956				  const int ts)
957{
958	struct tcp_sock *tp = tcp_sk(sk);
959	if (metric > tp->reordering) {
960		int mib_idx;
961
962		tp->reordering = min(TCP_MAX_REORDERING, metric);
963
964		/* This exciting event is worth to be remembered. 8) */
965		if (ts)
966			mib_idx = LINUX_MIB_TCPTSREORDER;
967		else if (tcp_is_reno(tp))
968			mib_idx = LINUX_MIB_TCPRENOREORDER;
969		else if (tcp_is_fack(tp))
970			mib_idx = LINUX_MIB_TCPFACKREORDER;
971		else
972			mib_idx = LINUX_MIB_TCPSACKREORDER;
973
974		NET_INC_STATS_BH(sock_net(sk), mib_idx);
975#if FASTRETRANS_DEBUG > 1
976		printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
977		       tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
978		       tp->reordering,
979		       tp->fackets_out,
980		       tp->sacked_out,
981		       tp->undo_marker ? tp->undo_retrans : 0);
982#endif
983		tcp_disable_fack(tp);
984	}
985}
986
987/* This must be called before lost_out is incremented */
988static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
989{
990	if ((tp->retransmit_skb_hint == NULL) ||
991	    before(TCP_SKB_CB(skb)->seq,
992		   TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
993		tp->retransmit_skb_hint = skb;
994
995	if (!tp->lost_out ||
996	    after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
997		tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
998}
999
1000static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
1001{
1002	if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1003		tcp_verify_retransmit_hint(tp, skb);
1004
1005		tp->lost_out += tcp_skb_pcount(skb);
1006		TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1007	}
1008}
1009
1010static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1011					    struct sk_buff *skb)
1012{
1013	tcp_verify_retransmit_hint(tp, skb);
1014
1015	if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1016		tp->lost_out += tcp_skb_pcount(skb);
1017		TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1018	}
1019}
1020
1021/* This procedure tags the retransmission queue when SACKs arrive.
1022 *
1023 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1024 * Packets in queue with these bits set are counted in variables
1025 * sacked_out, retrans_out and lost_out, correspondingly.
1026 *
1027 * Valid combinations are:
1028 * Tag  InFlight	Description
1029 * 0	1		- orig segment is in flight.
1030 * S	0		- nothing flies, orig reached receiver.
1031 * L	0		- nothing flies, orig lost by net.
1032 * R	2		- both orig and retransmit are in flight.
1033 * L|R	1		- orig is lost, retransmit is in flight.
1034 * S|R  1		- orig reached receiver, retrans is still in flight.
1035 * (L|S|R is logically valid, it could occur when L|R is sacked,
1036 *  but it is equivalent to plain S and code short-curcuits it to S.
1037 *  L|S is logically invalid, it would mean -1 packet in flight 8))
1038 *
1039 * These 6 states form finite state machine, controlled by the following events:
1040 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1041 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1042 * 3. Loss detection event of two flavors:
1043 *	A. Scoreboard estimator decided the packet is lost.
1044 *	   A'. Reno "three dupacks" marks head of queue lost.
1045 *	   A''. Its FACK modification, head until snd.fack is lost.
1046 *	B. SACK arrives sacking SND.NXT at the moment, when the
1047 *	   segment was retransmitted.
1048 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1049 *
1050 * It is pleasant to note, that state diagram turns out to be commutative,
1051 * so that we are allowed not to be bothered by order of our actions,
1052 * when multiple events arrive simultaneously. (see the function below).
1053 *
1054 * Reordering detection.
1055 * --------------------
1056 * Reordering metric is maximal distance, which a packet can be displaced
1057 * in packet stream. With SACKs we can estimate it:
1058 *
1059 * 1. SACK fills old hole and the corresponding segment was not
1060 *    ever retransmitted -> reordering. Alas, we cannot use it
1061 *    when segment was retransmitted.
1062 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1063 *    for retransmitted and already SACKed segment -> reordering..
1064 * Both of these heuristics are not used in Loss state, when we cannot
1065 * account for retransmits accurately.
1066 *
1067 * SACK block validation.
1068 * ----------------------
1069 *
1070 * SACK block range validation checks that the received SACK block fits to
1071 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1072 * Note that SND.UNA is not included to the range though being valid because
1073 * it means that the receiver is rather inconsistent with itself reporting
1074 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1075 * perfectly valid, however, in light of RFC2018 which explicitly states
1076 * that "SACK block MUST reflect the newest segment.  Even if the newest
1077 * segment is going to be discarded ...", not that it looks very clever
1078 * in case of head skb. Due to potentional receiver driven attacks, we
1079 * choose to avoid immediate execution of a walk in write queue due to
1080 * reneging and defer head skb's loss recovery to standard loss recovery
1081 * procedure that will eventually trigger (nothing forbids us doing this).
1082 *
1083 * Implements also blockage to start_seq wrap-around. Problem lies in the
1084 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1085 * there's no guarantee that it will be before snd_nxt (n). The problem
1086 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1087 * wrap (s_w):
1088 *
1089 *         <- outs wnd ->                          <- wrapzone ->
1090 *         u     e      n                         u_w   e_w  s n_w
1091 *         |     |      |                          |     |   |  |
1092 * |<------------+------+----- TCP seqno space --------------+---------->|
1093 * ...-- <2^31 ->|                                           |<--------...
1094 * ...---- >2^31 ------>|                                    |<--------...
1095 *
1096 * Current code wouldn't be vulnerable but it's better still to discard such
1097 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1098 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1099 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1100 * equal to the ideal case (infinite seqno space without wrap caused issues).
1101 *
1102 * With D-SACK the lower bound is extended to cover sequence space below
1103 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1104 * again, D-SACK block must not to go across snd_una (for the same reason as
1105 * for the normal SACK blocks, explained above). But there all simplicity
1106 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1107 * fully below undo_marker they do not affect behavior in anyway and can
1108 * therefore be safely ignored. In rare cases (which are more or less
1109 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1110 * fragmentation and packet reordering past skb's retransmission. To consider
1111 * them correctly, the acceptable range must be extended even more though
1112 * the exact amount is rather hard to quantify. However, tp->max_window can
1113 * be used as an exaggerated estimate.
1114 */
1115static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1116				  u32 start_seq, u32 end_seq)
1117{
1118	/* Too far in future, or reversed (interpretation is ambiguous) */
1119	if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1120		return 0;
1121
1122	/* Nasty start_seq wrap-around check (see comments above) */
1123	if (!before(start_seq, tp->snd_nxt))
1124		return 0;
1125
1126	/* In outstanding window? ...This is valid exit for D-SACKs too.
1127	 * start_seq == snd_una is non-sensical (see comments above)
1128	 */
1129	if (after(start_seq, tp->snd_una))
1130		return 1;
1131
1132	if (!is_dsack || !tp->undo_marker)
1133		return 0;
1134
1135	/* ...Then it's D-SACK, and must reside below snd_una completely */
1136	if (after(end_seq, tp->snd_una))
1137		return 0;
1138
1139	if (!before(start_seq, tp->undo_marker))
1140		return 1;
1141
1142	/* Too old */
1143	if (!after(end_seq, tp->undo_marker))
1144		return 0;
1145
1146	/* Undo_marker boundary crossing (overestimates a lot). Known already:
1147	 *   start_seq < undo_marker and end_seq >= undo_marker.
1148	 */
1149	return !before(start_seq, end_seq - tp->max_window);
1150}
1151
1152/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1153 * Event "B". Later note: FACK people cheated me again 8), we have to account
1154 * for reordering! Ugly, but should help.
1155 *
1156 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1157 * less than what is now known to be received by the other end (derived from
1158 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1159 * retransmitted skbs to avoid some costly processing per ACKs.
1160 */
1161static void tcp_mark_lost_retrans(struct sock *sk)
1162{
1163	const struct inet_connection_sock *icsk = inet_csk(sk);
1164	struct tcp_sock *tp = tcp_sk(sk);
1165	struct sk_buff *skb;
1166	int cnt = 0;
1167	u32 new_low_seq = tp->snd_nxt;
1168	u32 received_upto = tcp_highest_sack_seq(tp);
1169
1170	if (!tcp_is_fack(tp) || !tp->retrans_out ||
1171	    !after(received_upto, tp->lost_retrans_low) ||
1172	    icsk->icsk_ca_state != TCP_CA_Recovery)
1173		return;
1174
1175	tcp_for_write_queue(skb, sk) {
1176		u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1177
1178		if (skb == tcp_send_head(sk))
1179			break;
1180		if (cnt == tp->retrans_out)
1181			break;
1182		if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1183			continue;
1184
1185		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1186			continue;
1187
1188		/* TODO: We would like to get rid of tcp_is_fack(tp) only
1189		 * constraint here (see above) but figuring out that at
1190		 * least tp->reordering SACK blocks reside between ack_seq
1191		 * and received_upto is not easy task to do cheaply with
1192		 * the available datastructures.
1193		 *
1194		 * Whether FACK should check here for tp->reordering segs
1195		 * in-between one could argue for either way (it would be
1196		 * rather simple to implement as we could count fack_count
1197		 * during the walk and do tp->fackets_out - fack_count).
1198		 */
1199		if (after(received_upto, ack_seq)) {
1200			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1201			tp->retrans_out -= tcp_skb_pcount(skb);
1202
1203			tcp_skb_mark_lost_uncond_verify(tp, skb);
1204			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1205		} else {
1206			if (before(ack_seq, new_low_seq))
1207				new_low_seq = ack_seq;
1208			cnt += tcp_skb_pcount(skb);
1209		}
1210	}
1211
1212	if (tp->retrans_out)
1213		tp->lost_retrans_low = new_low_seq;
1214}
1215
1216static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1217			   struct tcp_sack_block_wire *sp, int num_sacks,
1218			   u32 prior_snd_una)
1219{
1220	struct tcp_sock *tp = tcp_sk(sk);
1221	u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1222	u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1223	int dup_sack = 0;
1224
1225	if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1226		dup_sack = 1;
1227		tcp_dsack_seen(tp);
1228		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1229	} else if (num_sacks > 1) {
1230		u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1231		u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1232
1233		if (!after(end_seq_0, end_seq_1) &&
1234		    !before(start_seq_0, start_seq_1)) {
1235			dup_sack = 1;
1236			tcp_dsack_seen(tp);
1237			NET_INC_STATS_BH(sock_net(sk),
1238					LINUX_MIB_TCPDSACKOFORECV);
1239		}
1240	}
1241
1242	/* D-SACK for already forgotten data... Do dumb counting. */
1243	if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1244	    !after(end_seq_0, prior_snd_una) &&
1245	    after(end_seq_0, tp->undo_marker))
1246		tp->undo_retrans--;
1247
1248	return dup_sack;
1249}
1250
1251struct tcp_sacktag_state {
1252	int reord;
1253	int fack_count;
1254	int flag;
1255};
1256
1257/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1258 * the incoming SACK may not exactly match but we can find smaller MSS
1259 * aligned portion of it that matches. Therefore we might need to fragment
1260 * which may fail and creates some hassle (caller must handle error case
1261 * returns).
1262 *
1263 * FIXME: this could be merged to shift decision code
1264 */
1265static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1266				 u32 start_seq, u32 end_seq)
1267{
1268	int in_sack, err;
1269	unsigned int pkt_len;
1270	unsigned int mss;
1271
1272	in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1273		  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1274
1275	if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1276	    after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1277		mss = tcp_skb_mss(skb);
1278		in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1279
1280		if (!in_sack) {
1281			pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1282			if (pkt_len < mss)
1283				pkt_len = mss;
1284		} else {
1285			pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1286			if (pkt_len < mss)
1287				return -EINVAL;
1288		}
1289
1290		/* Round if necessary so that SACKs cover only full MSSes
1291		 * and/or the remaining small portion (if present)
1292		 */
1293		if (pkt_len > mss) {
1294			unsigned int new_len = (pkt_len / mss) * mss;
1295			if (!in_sack && new_len < pkt_len) {
1296				new_len += mss;
1297				if (new_len > skb->len)
1298					return 0;
1299			}
1300			pkt_len = new_len;
1301		}
1302		err = tcp_fragment(sk, skb, pkt_len, mss);
1303		if (err < 0)
1304			return err;
1305	}
1306
1307	return in_sack;
1308}
1309
1310static u8 tcp_sacktag_one(const struct sk_buff *skb, struct sock *sk,
1311			  struct tcp_sacktag_state *state,
1312			  int dup_sack, int pcount)
1313{
1314	struct tcp_sock *tp = tcp_sk(sk);
1315	u8 sacked = TCP_SKB_CB(skb)->sacked;
1316	int fack_count = state->fack_count;
1317
1318	/* Account D-SACK for retransmitted packet. */
1319	if (dup_sack && (sacked & TCPCB_RETRANS)) {
1320		if (tp->undo_marker && tp->undo_retrans &&
1321		    after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1322			tp->undo_retrans--;
1323		if (sacked & TCPCB_SACKED_ACKED)
1324			state->reord = min(fack_count, state->reord);
1325	}
1326
1327	/* Nothing to do; acked frame is about to be dropped (was ACKed). */
1328	if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1329		return sacked;
1330
1331	if (!(sacked & TCPCB_SACKED_ACKED)) {
1332		if (sacked & TCPCB_SACKED_RETRANS) {
1333			/* If the segment is not tagged as lost,
1334			 * we do not clear RETRANS, believing
1335			 * that retransmission is still in flight.
1336			 */
1337			if (sacked & TCPCB_LOST) {
1338				sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1339				tp->lost_out -= pcount;
1340				tp->retrans_out -= pcount;
1341			}
1342		} else {
1343			if (!(sacked & TCPCB_RETRANS)) {
1344				/* New sack for not retransmitted frame,
1345				 * which was in hole. It is reordering.
1346				 */
1347				if (before(TCP_SKB_CB(skb)->seq,
1348					   tcp_highest_sack_seq(tp)))
1349					state->reord = min(fack_count,
1350							   state->reord);
1351
1352				/* SACK enhanced F-RTO (RFC4138; Appendix B) */
1353				if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1354					state->flag |= FLAG_ONLY_ORIG_SACKED;
1355			}
1356
1357			if (sacked & TCPCB_LOST) {
1358				sacked &= ~TCPCB_LOST;
1359				tp->lost_out -= pcount;
1360			}
1361		}
1362
1363		sacked |= TCPCB_SACKED_ACKED;
1364		state->flag |= FLAG_DATA_SACKED;
1365		tp->sacked_out += pcount;
1366
1367		fack_count += pcount;
1368
1369		/* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1370		if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1371		    before(TCP_SKB_CB(skb)->seq,
1372			   TCP_SKB_CB(tp->lost_skb_hint)->seq))
1373			tp->lost_cnt_hint += pcount;
1374
1375		if (fack_count > tp->fackets_out)
1376			tp->fackets_out = fack_count;
1377	}
1378
1379	/* D-SACK. We can detect redundant retransmission in S|R and plain R
1380	 * frames and clear it. undo_retrans is decreased above, L|R frames
1381	 * are accounted above as well.
1382	 */
1383	if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1384		sacked &= ~TCPCB_SACKED_RETRANS;
1385		tp->retrans_out -= pcount;
1386	}
1387
1388	return sacked;
1389}
1390
1391static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1392			   struct tcp_sacktag_state *state,
1393			   unsigned int pcount, int shifted, int mss,
1394			   int dup_sack)
1395{
1396	struct tcp_sock *tp = tcp_sk(sk);
1397	struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1398
1399	BUG_ON(!pcount);
1400
1401	if (skb == tp->lost_skb_hint)
1402		tp->lost_cnt_hint += pcount;
1403
1404	TCP_SKB_CB(prev)->end_seq += shifted;
1405	TCP_SKB_CB(skb)->seq += shifted;
1406
1407	skb_shinfo(prev)->gso_segs += pcount;
1408	BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1409	skb_shinfo(skb)->gso_segs -= pcount;
1410
1411	/* When we're adding to gso_segs == 1, gso_size will be zero,
1412	 * in theory this shouldn't be necessary but as long as DSACK
1413	 * code can come after this skb later on it's better to keep
1414	 * setting gso_size to something.
1415	 */
1416	if (!skb_shinfo(prev)->gso_size) {
1417		skb_shinfo(prev)->gso_size = mss;
1418		skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1419	}
1420
1421	/* CHECKME: To clear or not to clear? Mimics normal skb currently */
1422	if (skb_shinfo(skb)->gso_segs <= 1) {
1423		skb_shinfo(skb)->gso_size = 0;
1424		skb_shinfo(skb)->gso_type = 0;
1425	}
1426
1427	/* We discard results */
1428	tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
1429
1430	/* Difference in this won't matter, both ACKed by the same cumul. ACK */
1431	TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1432
1433	if (skb->len > 0) {
1434		BUG_ON(!tcp_skb_pcount(skb));
1435		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1436		return 0;
1437	}
1438
1439	/* Whole SKB was eaten :-) */
1440
1441	if (skb == tp->retransmit_skb_hint)
1442		tp->retransmit_skb_hint = prev;
1443	if (skb == tp->scoreboard_skb_hint)
1444		tp->scoreboard_skb_hint = prev;
1445	if (skb == tp->lost_skb_hint) {
1446		tp->lost_skb_hint = prev;
1447		tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1448	}
1449
1450	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1451	if (skb == tcp_highest_sack(sk))
1452		tcp_advance_highest_sack(sk, skb);
1453
1454	tcp_unlink_write_queue(skb, sk);
1455	sk_wmem_free_skb(sk, skb);
1456
1457	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1458
1459	return 1;
1460}
1461
1462/* I wish gso_size would have a bit more sane initialization than
1463 * something-or-zero which complicates things
1464 */
1465static int tcp_skb_seglen(const struct sk_buff *skb)
1466{
1467	return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1468}
1469
1470/* Shifting pages past head area doesn't work */
1471static int skb_can_shift(const struct sk_buff *skb)
1472{
1473	return !skb_headlen(skb) && skb_is_nonlinear(skb);
1474}
1475
1476/* Try collapsing SACK blocks spanning across multiple skbs to a single
1477 * skb.
1478 */
1479static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1480					  struct tcp_sacktag_state *state,
1481					  u32 start_seq, u32 end_seq,
1482					  int dup_sack)
1483{
1484	struct tcp_sock *tp = tcp_sk(sk);
1485	struct sk_buff *prev;
1486	int mss;
1487	int pcount = 0;
1488	int len;
1489	int in_sack;
1490
1491	if (!sk_can_gso(sk))
1492		goto fallback;
1493
1494	/* Normally R but no L won't result in plain S */
1495	if (!dup_sack &&
1496	    (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1497		goto fallback;
1498	if (!skb_can_shift(skb))
1499		goto fallback;
1500	/* This frame is about to be dropped (was ACKed). */
1501	if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1502		goto fallback;
1503
1504	/* Can only happen with delayed DSACK + discard craziness */
1505	if (unlikely(skb == tcp_write_queue_head(sk)))
1506		goto fallback;
1507	prev = tcp_write_queue_prev(sk, skb);
1508
1509	if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1510		goto fallback;
1511
1512	in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1513		  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1514
1515	if (in_sack) {
1516		len = skb->len;
1517		pcount = tcp_skb_pcount(skb);
1518		mss = tcp_skb_seglen(skb);
1519
1520		/* TODO: Fix DSACKs to not fragment already SACKed and we can
1521		 * drop this restriction as unnecessary
1522		 */
1523		if (mss != tcp_skb_seglen(prev))
1524			goto fallback;
1525	} else {
1526		if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1527			goto noop;
1528		/* CHECKME: This is non-MSS split case only?, this will
1529		 * cause skipped skbs due to advancing loop btw, original
1530		 * has that feature too
1531		 */
1532		if (tcp_skb_pcount(skb) <= 1)
1533			goto noop;
1534
1535		in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1536		if (!in_sack) {
1537			/* TODO: head merge to next could be attempted here
1538			 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1539			 * though it might not be worth of the additional hassle
1540			 *
1541			 * ...we can probably just fallback to what was done
1542			 * previously. We could try merging non-SACKed ones
1543			 * as well but it probably isn't going to buy off
1544			 * because later SACKs might again split them, and
1545			 * it would make skb timestamp tracking considerably
1546			 * harder problem.
1547			 */
1548			goto fallback;
1549		}
1550
1551		len = end_seq - TCP_SKB_CB(skb)->seq;
1552		BUG_ON(len < 0);
1553		BUG_ON(len > skb->len);
1554
1555		/* MSS boundaries should be honoured or else pcount will
1556		 * severely break even though it makes things bit trickier.
1557		 * Optimize common case to avoid most of the divides
1558		 */
1559		mss = tcp_skb_mss(skb);
1560
1561		/* TODO: Fix DSACKs to not fragment already SACKed and we can
1562		 * drop this restriction as unnecessary
1563		 */
1564		if (mss != tcp_skb_seglen(prev))
1565			goto fallback;
1566
1567		if (len == mss) {
1568			pcount = 1;
1569		} else if (len < mss) {
1570			goto noop;
1571		} else {
1572			pcount = len / mss;
1573			len = pcount * mss;
1574		}
1575	}
1576
1577	if (!skb_shift(prev, skb, len))
1578		goto fallback;
1579	if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1580		goto out;
1581
1582	/* Hole filled allows collapsing with the next as well, this is very
1583	 * useful when hole on every nth skb pattern happens
1584	 */
1585	if (prev == tcp_write_queue_tail(sk))
1586		goto out;
1587	skb = tcp_write_queue_next(sk, prev);
1588
1589	if (!skb_can_shift(skb) ||
1590	    (skb == tcp_send_head(sk)) ||
1591	    ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1592	    (mss != tcp_skb_seglen(skb)))
1593		goto out;
1594
1595	len = skb->len;
1596	if (skb_shift(prev, skb, len)) {
1597		pcount += tcp_skb_pcount(skb);
1598		tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1599	}
1600
1601out:
1602	state->fack_count += pcount;
1603	return prev;
1604
1605noop:
1606	return skb;
1607
1608fallback:
1609	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1610	return NULL;
1611}
1612
1613static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1614					struct tcp_sack_block *next_dup,
1615					struct tcp_sacktag_state *state,
1616					u32 start_seq, u32 end_seq,
1617					int dup_sack_in)
1618{
1619	struct tcp_sock *tp = tcp_sk(sk);
1620	struct sk_buff *tmp;
1621
1622	tcp_for_write_queue_from(skb, sk) {
1623		int in_sack = 0;
1624		int dup_sack = dup_sack_in;
1625
1626		if (skb == tcp_send_head(sk))
1627			break;
1628
1629		/* queue is in-order => we can short-circuit the walk early */
1630		if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1631			break;
1632
1633		if ((next_dup != NULL) &&
1634		    before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1635			in_sack = tcp_match_skb_to_sack(sk, skb,
1636							next_dup->start_seq,
1637							next_dup->end_seq);
1638			if (in_sack > 0)
1639				dup_sack = 1;
1640		}
1641
1642		/* skb reference here is a bit tricky to get right, since
1643		 * shifting can eat and free both this skb and the next,
1644		 * so not even _safe variant of the loop is enough.
1645		 */
1646		if (in_sack <= 0) {
1647			tmp = tcp_shift_skb_data(sk, skb, state,
1648						 start_seq, end_seq, dup_sack);
1649			if (tmp != NULL) {
1650				if (tmp != skb) {
1651					skb = tmp;
1652					continue;
1653				}
1654
1655				in_sack = 0;
1656			} else {
1657				in_sack = tcp_match_skb_to_sack(sk, skb,
1658								start_seq,
1659								end_seq);
1660			}
1661		}
1662
1663		if (unlikely(in_sack < 0))
1664			break;
1665
1666		if (in_sack) {
1667			TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1668								  state,
1669								  dup_sack,
1670								  tcp_skb_pcount(skb));
1671
1672			if (!before(TCP_SKB_CB(skb)->seq,
1673				    tcp_highest_sack_seq(tp)))
1674				tcp_advance_highest_sack(sk, skb);
1675		}
1676
1677		state->fack_count += tcp_skb_pcount(skb);
1678	}
1679	return skb;
1680}
1681
1682/* Avoid all extra work that is being done by sacktag while walking in
1683 * a normal way
1684 */
1685static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1686					struct tcp_sacktag_state *state,
1687					u32 skip_to_seq)
1688{
1689	tcp_for_write_queue_from(skb, sk) {
1690		if (skb == tcp_send_head(sk))
1691			break;
1692
1693		if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1694			break;
1695
1696		state->fack_count += tcp_skb_pcount(skb);
1697	}
1698	return skb;
1699}
1700
1701static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1702						struct sock *sk,
1703						struct tcp_sack_block *next_dup,
1704						struct tcp_sacktag_state *state,
1705						u32 skip_to_seq)
1706{
1707	if (next_dup == NULL)
1708		return skb;
1709
1710	if (before(next_dup->start_seq, skip_to_seq)) {
1711		skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1712		skb = tcp_sacktag_walk(skb, sk, NULL, state,
1713				       next_dup->start_seq, next_dup->end_seq,
1714				       1);
1715	}
1716
1717	return skb;
1718}
1719
1720static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1721{
1722	return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1723}
1724
1725static int
1726tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1727			u32 prior_snd_una)
1728{
1729	const struct inet_connection_sock *icsk = inet_csk(sk);
1730	struct tcp_sock *tp = tcp_sk(sk);
1731	const unsigned char *ptr = (skb_transport_header(ack_skb) +
1732				    TCP_SKB_CB(ack_skb)->sacked);
1733	struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1734	struct tcp_sack_block sp[TCP_NUM_SACKS];
1735	struct tcp_sack_block *cache;
1736	struct tcp_sacktag_state state;
1737	struct sk_buff *skb;
1738	int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1739	int used_sacks;
1740	int found_dup_sack = 0;
1741	int i, j;
1742	int first_sack_index;
1743
1744	state.flag = 0;
1745	state.reord = tp->packets_out;
1746
1747	if (!tp->sacked_out) {
1748		if (WARN_ON(tp->fackets_out))
1749			tp->fackets_out = 0;
1750		tcp_highest_sack_reset(sk);
1751	}
1752
1753	found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1754					 num_sacks, prior_snd_una);
1755	if (found_dup_sack)
1756		state.flag |= FLAG_DSACKING_ACK;
1757
1758	/* Eliminate too old ACKs, but take into
1759	 * account more or less fresh ones, they can
1760	 * contain valid SACK info.
1761	 */
1762	if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1763		return 0;
1764
1765	if (!tp->packets_out)
1766		goto out;
1767
1768	used_sacks = 0;
1769	first_sack_index = 0;
1770	for (i = 0; i < num_sacks; i++) {
1771		int dup_sack = !i && found_dup_sack;
1772
1773		sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1774		sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1775
1776		if (!tcp_is_sackblock_valid(tp, dup_sack,
1777					    sp[used_sacks].start_seq,
1778					    sp[used_sacks].end_seq)) {
1779			int mib_idx;
1780
1781			if (dup_sack) {
1782				if (!tp->undo_marker)
1783					mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1784				else
1785					mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1786			} else {
1787				/* Don't count olds caused by ACK reordering */
1788				if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1789				    !after(sp[used_sacks].end_seq, tp->snd_una))
1790					continue;
1791				mib_idx = LINUX_MIB_TCPSACKDISCARD;
1792			}
1793
1794			NET_INC_STATS_BH(sock_net(sk), mib_idx);
1795			if (i == 0)
1796				first_sack_index = -1;
1797			continue;
1798		}
1799
1800		/* Ignore very old stuff early */
1801		if (!after(sp[used_sacks].end_seq, prior_snd_una))
1802			continue;
1803
1804		used_sacks++;
1805	}
1806
1807	/* order SACK blocks to allow in order walk of the retrans queue */
1808	for (i = used_sacks - 1; i > 0; i--) {
1809		for (j = 0; j < i; j++) {
1810			if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1811				swap(sp[j], sp[j + 1]);
1812
1813				/* Track where the first SACK block goes to */
1814				if (j == first_sack_index)
1815					first_sack_index = j + 1;
1816			}
1817		}
1818	}
1819
1820	skb = tcp_write_queue_head(sk);
1821	state.fack_count = 0;
1822	i = 0;
1823
1824	if (!tp->sacked_out) {
1825		/* It's already past, so skip checking against it */
1826		cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1827	} else {
1828		cache = tp->recv_sack_cache;
1829		/* Skip empty blocks in at head of the cache */
1830		while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1831		       !cache->end_seq)
1832			cache++;
1833	}
1834
1835	while (i < used_sacks) {
1836		u32 start_seq = sp[i].start_seq;
1837		u32 end_seq = sp[i].end_seq;
1838		int dup_sack = (found_dup_sack && (i == first_sack_index));
1839		struct tcp_sack_block *next_dup = NULL;
1840
1841		if (found_dup_sack && ((i + 1) == first_sack_index))
1842			next_dup = &sp[i + 1];
1843
1844		/* Skip too early cached blocks */
1845		while (tcp_sack_cache_ok(tp, cache) &&
1846		       !before(start_seq, cache->end_seq))
1847			cache++;
1848
1849		/* Can skip some work by looking recv_sack_cache? */
1850		if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1851		    after(end_seq, cache->start_seq)) {
1852
1853			/* Head todo? */
1854			if (before(start_seq, cache->start_seq)) {
1855				skb = tcp_sacktag_skip(skb, sk, &state,
1856						       start_seq);
1857				skb = tcp_sacktag_walk(skb, sk, next_dup,
1858						       &state,
1859						       start_seq,
1860						       cache->start_seq,
1861						       dup_sack);
1862			}
1863
1864			/* Rest of the block already fully processed? */
1865			if (!after(end_seq, cache->end_seq))
1866				goto advance_sp;
1867
1868			skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1869						       &state,
1870						       cache->end_seq);
1871
1872			/* ...tail remains todo... */
1873			if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1874				/* ...but better entrypoint exists! */
1875				skb = tcp_highest_sack(sk);
1876				if (skb == NULL)
1877					break;
1878				state.fack_count = tp->fackets_out;
1879				cache++;
1880				goto walk;
1881			}
1882
1883			skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1884			/* Check overlap against next cached too (past this one already) */
1885			cache++;
1886			continue;
1887		}
1888
1889		if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1890			skb = tcp_highest_sack(sk);
1891			if (skb == NULL)
1892				break;
1893			state.fack_count = tp->fackets_out;
1894		}
1895		skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1896
1897walk:
1898		skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1899				       start_seq, end_seq, dup_sack);
1900
1901advance_sp:
1902		/* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1903		 * due to in-order walk
1904		 */
1905		if (after(end_seq, tp->frto_highmark))
1906			state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1907
1908		i++;
1909	}
1910
1911	/* Clear the head of the cache sack blocks so we can skip it next time */
1912	for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1913		tp->recv_sack_cache[i].start_seq = 0;
1914		tp->recv_sack_cache[i].end_seq = 0;
1915	}
1916	for (j = 0; j < used_sacks; j++)
1917		tp->recv_sack_cache[i++] = sp[j];
1918
1919	tcp_mark_lost_retrans(sk);
1920
1921	tcp_verify_left_out(tp);
1922
1923	if ((state.reord < tp->fackets_out) &&
1924	    ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1925	    (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1926		tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1927
1928out:
1929
1930#if FASTRETRANS_DEBUG > 0
1931	WARN_ON((int)tp->sacked_out < 0);
1932	WARN_ON((int)tp->lost_out < 0);
1933	WARN_ON((int)tp->retrans_out < 0);
1934	WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1935#endif
1936	return state.flag;
1937}
1938
1939/* Limits sacked_out so that sum with lost_out isn't ever larger than
1940 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1941 */
1942static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1943{
1944	u32 holes;
1945
1946	holes = max(tp->lost_out, 1U);
1947	holes = min(holes, tp->packets_out);
1948
1949	if ((tp->sacked_out + holes) > tp->packets_out) {
1950		tp->sacked_out = tp->packets_out - holes;
1951		return 1;
1952	}
1953	return 0;
1954}
1955
1956/* If we receive more dupacks than we expected counting segments
1957 * in assumption of absent reordering, interpret this as reordering.
1958 * The only another reason could be bug in receiver TCP.
1959 */
1960static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1961{
1962	struct tcp_sock *tp = tcp_sk(sk);
1963	if (tcp_limit_reno_sacked(tp))
1964		tcp_update_reordering(sk, tp->packets_out + addend, 0);
1965}
1966
1967/* Emulate SACKs for SACKless connection: account for a new dupack. */
1968
1969static void tcp_add_reno_sack(struct sock *sk)
1970{
1971	struct tcp_sock *tp = tcp_sk(sk);
1972	tp->sacked_out++;
1973	tcp_check_reno_reordering(sk, 0);
1974	tcp_verify_left_out(tp);
1975}
1976
1977/* Account for ACK, ACKing some data in Reno Recovery phase. */
1978
1979static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1980{
1981	struct tcp_sock *tp = tcp_sk(sk);
1982
1983	if (acked > 0) {
1984		/* One ACK acked hole. The rest eat duplicate ACKs. */
1985		if (acked - 1 >= tp->sacked_out)
1986			tp->sacked_out = 0;
1987		else
1988			tp->sacked_out -= acked - 1;
1989	}
1990	tcp_check_reno_reordering(sk, acked);
1991	tcp_verify_left_out(tp);
1992}
1993
1994static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1995{
1996	tp->sacked_out = 0;
1997}
1998
1999static int tcp_is_sackfrto(const struct tcp_sock *tp)
2000{
2001	return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2002}
2003
2004/* F-RTO can only be used if TCP has never retransmitted anything other than
2005 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2006 */
2007int tcp_use_frto(struct sock *sk)
2008{
2009	const struct tcp_sock *tp = tcp_sk(sk);
2010	const struct inet_connection_sock *icsk = inet_csk(sk);
2011	struct sk_buff *skb;
2012
2013	if (!sysctl_tcp_frto)
2014		return 0;
2015
2016	/* MTU probe and F-RTO won't really play nicely along currently */
2017	if (icsk->icsk_mtup.probe_size)
2018		return 0;
2019
2020	if (tcp_is_sackfrto(tp))
2021		return 1;
2022
2023	/* Avoid expensive walking of rexmit queue if possible */
2024	if (tp->retrans_out > 1)
2025		return 0;
2026
2027	skb = tcp_write_queue_head(sk);
2028	if (tcp_skb_is_last(sk, skb))
2029		return 1;
2030	skb = tcp_write_queue_next(sk, skb);	/* Skips head */
2031	tcp_for_write_queue_from(skb, sk) {
2032		if (skb == tcp_send_head(sk))
2033			break;
2034		if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2035			return 0;
2036		/* Short-circuit when first non-SACKed skb has been checked */
2037		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2038			break;
2039	}
2040	return 1;
2041}
2042
2043/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2044 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2045 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2046 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2047 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2048 * bits are handled if the Loss state is really to be entered (in
2049 * tcp_enter_frto_loss).
2050 *
2051 * Do like tcp_enter_loss() would; when RTO expires the second time it
2052 * does:
2053 *  "Reduce ssthresh if it has not yet been made inside this window."
2054 */
2055void tcp_enter_frto(struct sock *sk)
2056{
2057	const struct inet_connection_sock *icsk = inet_csk(sk);
2058	struct tcp_sock *tp = tcp_sk(sk);
2059	struct sk_buff *skb;
2060
2061	if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2062	    tp->snd_una == tp->high_seq ||
2063	    ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2064	     !icsk->icsk_retransmits)) {
2065		tp->prior_ssthresh = tcp_current_ssthresh(sk);
2066		/* Our state is too optimistic in ssthresh() call because cwnd
2067		 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2068		 * recovery has not yet completed. Pattern would be this: RTO,
2069		 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2070		 * up here twice).
2071		 * RFC4138 should be more specific on what to do, even though
2072		 * RTO is quite unlikely to occur after the first Cumulative ACK
2073		 * due to back-off and complexity of triggering events ...
2074		 */
2075		if (tp->frto_counter) {
2076			u32 stored_cwnd;
2077			stored_cwnd = tp->snd_cwnd;
2078			tp->snd_cwnd = 2;
2079			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2080			tp->snd_cwnd = stored_cwnd;
2081		} else {
2082			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2083		}
2084		/* ... in theory, cong.control module could do "any tricks" in
2085		 * ssthresh(), which means that ca_state, lost bits and lost_out
2086		 * counter would have to be faked before the call occurs. We
2087		 * consider that too expensive, unlikely and hacky, so modules
2088		 * using these in ssthresh() must deal these incompatibility
2089		 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2090		 */
2091		tcp_ca_event(sk, CA_EVENT_FRTO);
2092	}
2093
2094	tp->undo_marker = tp->snd_una;
2095	tp->undo_retrans = 0;
2096
2097	skb = tcp_write_queue_head(sk);
2098	if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2099		tp->undo_marker = 0;
2100	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2101		TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2102		tp->retrans_out -= tcp_skb_pcount(skb);
2103	}
2104	tcp_verify_left_out(tp);
2105
2106	/* Too bad if TCP was application limited */
2107	tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2108
2109	/* Earlier loss recovery underway (see RFC4138; Appendix B).
2110	 * The last condition is necessary at least in tp->frto_counter case.
2111	 */
2112	if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2113	    ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2114	    after(tp->high_seq, tp->snd_una)) {
2115		tp->frto_highmark = tp->high_seq;
2116	} else {
2117		tp->frto_highmark = tp->snd_nxt;
2118	}
2119	tcp_set_ca_state(sk, TCP_CA_Disorder);
2120	tp->high_seq = tp->snd_nxt;
2121	tp->frto_counter = 1;
2122}
2123
2124/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2125 * which indicates that we should follow the traditional RTO recovery,
2126 * i.e. mark everything lost and do go-back-N retransmission.
2127 */
2128static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2129{
2130	struct tcp_sock *tp = tcp_sk(sk);
2131	struct sk_buff *skb;
2132
2133	tp->lost_out = 0;
2134	tp->retrans_out = 0;
2135	if (tcp_is_reno(tp))
2136		tcp_reset_reno_sack(tp);
2137
2138	tcp_for_write_queue(skb, sk) {
2139		if (skb == tcp_send_head(sk))
2140			break;
2141
2142		TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2143		/*
2144		 * Count the retransmission made on RTO correctly (only when
2145		 * waiting for the first ACK and did not get it)...
2146		 */
2147		if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2148			/* For some reason this R-bit might get cleared? */
2149			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2150				tp->retrans_out += tcp_skb_pcount(skb);
2151			/* ...enter this if branch just for the first segment */
2152			flag |= FLAG_DATA_ACKED;
2153		} else {
2154			if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2155				tp->undo_marker = 0;
2156			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2157		}
2158
2159		/* Marking forward transmissions that were made after RTO lost
2160		 * can cause unnecessary retransmissions in some scenarios,
2161		 * SACK blocks will mitigate that in some but not in all cases.
2162		 * We used to not mark them but it was causing break-ups with
2163		 * receivers that do only in-order receival.
2164		 *
2165		 * TODO: we could detect presence of such receiver and select
2166		 * different behavior per flow.
2167		 */
2168		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2169			TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2170			tp->lost_out += tcp_skb_pcount(skb);
2171			tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2172		}
2173	}
2174	tcp_verify_left_out(tp);
2175
2176	tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2177	tp->snd_cwnd_cnt = 0;
2178	tp->snd_cwnd_stamp = tcp_time_stamp;
2179	tp->frto_counter = 0;
2180	tp->bytes_acked = 0;
2181
2182	tp->reordering = min_t(unsigned int, tp->reordering,
2183			       sysctl_tcp_reordering);
2184	tcp_set_ca_state(sk, TCP_CA_Loss);
2185	tp->high_seq = tp->snd_nxt;
2186	TCP_ECN_queue_cwr(tp);
2187
2188	tcp_clear_all_retrans_hints(tp);
2189}
2190
2191static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2192{
2193	tp->retrans_out = 0;
2194	tp->lost_out = 0;
2195
2196	tp->undo_marker = 0;
2197	tp->undo_retrans = 0;
2198}
2199
2200void tcp_clear_retrans(struct tcp_sock *tp)
2201{
2202	tcp_clear_retrans_partial(tp);
2203
2204	tp->fackets_out = 0;
2205	tp->sacked_out = 0;
2206}
2207
2208/* Enter Loss state. If "how" is not zero, forget all SACK information
2209 * and reset tags completely, otherwise preserve SACKs. If receiver
2210 * dropped its ofo queue, we will know this due to reneging detection.
2211 */
2212void tcp_enter_loss(struct sock *sk, int how)
2213{
2214	const struct inet_connection_sock *icsk = inet_csk(sk);
2215	struct tcp_sock *tp = tcp_sk(sk);
2216	struct sk_buff *skb;
2217
2218	/* Reduce ssthresh if it has not yet been made inside this window. */
2219	if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2220	    (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2221		tp->prior_ssthresh = tcp_current_ssthresh(sk);
2222		tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2223		tcp_ca_event(sk, CA_EVENT_LOSS);
2224	}
2225	tp->snd_cwnd	   = 1;
2226	tp->snd_cwnd_cnt   = 0;
2227	tp->snd_cwnd_stamp = tcp_time_stamp;
2228
2229	tp->bytes_acked = 0;
2230	tcp_clear_retrans_partial(tp);
2231
2232	if (tcp_is_reno(tp))
2233		tcp_reset_reno_sack(tp);
2234
2235	if (!how) {
2236		/* Push undo marker, if it was plain RTO and nothing
2237		 * was retransmitted. */
2238		tp->undo_marker = tp->snd_una;
2239	} else {
2240		tp->sacked_out = 0;
2241		tp->fackets_out = 0;
2242	}
2243	tcp_clear_all_retrans_hints(tp);
2244
2245	tcp_for_write_queue(skb, sk) {
2246		if (skb == tcp_send_head(sk))
2247			break;
2248
2249		if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2250			tp->undo_marker = 0;
2251		TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2252		if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2253			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2254			TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2255			tp->lost_out += tcp_skb_pcount(skb);
2256			tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2257		}
2258	}
2259	tcp_verify_left_out(tp);
2260
2261	tp->reordering = min_t(unsigned int, tp->reordering,
2262			       sysctl_tcp_reordering);
2263	tcp_set_ca_state(sk, TCP_CA_Loss);
2264	tp->high_seq = tp->snd_nxt;
2265	TCP_ECN_queue_cwr(tp);
2266	/* Abort F-RTO algorithm if one is in progress */
2267	tp->frto_counter = 0;
2268}
2269
2270/* If ACK arrived pointing to a remembered SACK, it means that our
2271 * remembered SACKs do not reflect real state of receiver i.e.
2272 * receiver _host_ is heavily congested (or buggy).
2273 *
2274 * Do processing similar to RTO timeout.
2275 */
2276static int tcp_check_sack_reneging(struct sock *sk, int flag)
2277{
2278	if (flag & FLAG_SACK_RENEGING) {
2279		struct inet_connection_sock *icsk = inet_csk(sk);
2280		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2281
2282		tcp_enter_loss(sk, 1);
2283		icsk->icsk_retransmits++;
2284		tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2285		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2286					  icsk->icsk_rto, TCP_RTO_MAX);
2287		return 1;
2288	}
2289	return 0;
2290}
2291
2292static inline int tcp_fackets_out(const struct tcp_sock *tp)
2293{
2294	return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2295}
2296
2297/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2298 * counter when SACK is enabled (without SACK, sacked_out is used for
2299 * that purpose).
2300 *
2301 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2302 * segments up to the highest received SACK block so far and holes in
2303 * between them.
2304 *
2305 * With reordering, holes may still be in flight, so RFC3517 recovery
2306 * uses pure sacked_out (total number of SACKed segments) even though
2307 * it violates the RFC that uses duplicate ACKs, often these are equal
2308 * but when e.g. out-of-window ACKs or packet duplication occurs,
2309 * they differ. Since neither occurs due to loss, TCP should really
2310 * ignore them.
2311 */
2312static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2313{
2314	return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2315}
2316
2317static inline int tcp_skb_timedout(const struct sock *sk,
2318				   const struct sk_buff *skb)
2319{
2320	return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2321}
2322
2323static inline int tcp_head_timedout(const struct sock *sk)
2324{
2325	const struct tcp_sock *tp = tcp_sk(sk);
2326
2327	return tp->packets_out &&
2328	       tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2329}
2330
2331/* Linux NewReno/SACK/FACK/ECN state machine.
2332 * --------------------------------------
2333 *
2334 * "Open"	Normal state, no dubious events, fast path.
2335 * "Disorder"   In all the respects it is "Open",
2336 *		but requires a bit more attention. It is entered when
2337 *		we see some SACKs or dupacks. It is split of "Open"
2338 *		mainly to move some processing from fast path to slow one.
2339 * "CWR"	CWND was reduced due to some Congestion Notification event.
2340 *		It can be ECN, ICMP source quench, local device congestion.
2341 * "Recovery"	CWND was reduced, we are fast-retransmitting.
2342 * "Loss"	CWND was reduced due to RTO timeout or SACK reneging.
2343 *
2344 * tcp_fastretrans_alert() is entered:
2345 * - each incoming ACK, if state is not "Open"
2346 * - when arrived ACK is unusual, namely:
2347 *	* SACK
2348 *	* Duplicate ACK.
2349 *	* ECN ECE.
2350 *
2351 * Counting packets in flight is pretty simple.
2352 *
2353 *	in_flight = packets_out - left_out + retrans_out
2354 *
2355 *	packets_out is SND.NXT-SND.UNA counted in packets.
2356 *
2357 *	retrans_out is number of retransmitted segments.
2358 *
2359 *	left_out is number of segments left network, but not ACKed yet.
2360 *
2361 *		left_out = sacked_out + lost_out
2362 *
2363 *     sacked_out: Packets, which arrived to receiver out of order
2364 *		   and hence not ACKed. With SACKs this number is simply
2365 *		   amount of SACKed data. Even without SACKs
2366 *		   it is easy to give pretty reliable estimate of this number,
2367 *		   counting duplicate ACKs.
2368 *
2369 *       lost_out: Packets lost by network. TCP has no explicit
2370 *		   "loss notification" feedback from network (for now).
2371 *		   It means that this number can be only _guessed_.
2372 *		   Actually, it is the heuristics to predict lossage that
2373 *		   distinguishes different algorithms.
2374 *
2375 *	F.e. after RTO, when all the queue is considered as lost,
2376 *	lost_out = packets_out and in_flight = retrans_out.
2377 *
2378 *		Essentially, we have now two algorithms counting
2379 *		lost packets.
2380 *
2381 *		FACK: It is the simplest heuristics. As soon as we decided
2382 *		that something is lost, we decide that _all_ not SACKed
2383 *		packets until the most forward SACK are lost. I.e.
2384 *		lost_out = fackets_out - sacked_out and left_out = fackets_out.
2385 *		It is absolutely correct estimate, if network does not reorder
2386 *		packets. And it loses any connection to reality when reordering
2387 *		takes place. We use FACK by default until reordering
2388 *		is suspected on the path to this destination.
2389 *
2390 *		NewReno: when Recovery is entered, we assume that one segment
2391 *		is lost (classic Reno). While we are in Recovery and
2392 *		a partial ACK arrives, we assume that one more packet
2393 *		is lost (NewReno). This heuristics are the same in NewReno
2394 *		and SACK.
2395 *
2396 *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2397 *  deflation etc. CWND is real congestion window, never inflated, changes
2398 *  only according to classic VJ rules.
2399 *
2400 * Really tricky (and requiring careful tuning) part of algorithm
2401 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2402 * The first determines the moment _when_ we should reduce CWND and,
2403 * hence, slow down forward transmission. In fact, it determines the moment
2404 * when we decide that hole is caused by loss, rather than by a reorder.
2405 *
2406 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2407 * holes, caused by lost packets.
2408 *
2409 * And the most logically complicated part of algorithm is undo
2410 * heuristics. We detect false retransmits due to both too early
2411 * fast retransmit (reordering) and underestimated RTO, analyzing
2412 * timestamps and D-SACKs. When we detect that some segments were
2413 * retransmitted by mistake and CWND reduction was wrong, we undo
2414 * window reduction and abort recovery phase. This logic is hidden
2415 * inside several functions named tcp_try_undo_<something>.
2416 */
2417
2418/* This function decides, when we should leave Disordered state
2419 * and enter Recovery phase, reducing congestion window.
2420 *
2421 * Main question: may we further continue forward transmission
2422 * with the same cwnd?
2423 */
2424static int tcp_time_to_recover(struct sock *sk)
2425{
2426	struct tcp_sock *tp = tcp_sk(sk);
2427	__u32 packets_out;
2428
2429	/* Do not perform any recovery during F-RTO algorithm */
2430	if (tp->frto_counter)
2431		return 0;
2432
2433	/* Trick#1: The loss is proven. */
2434	if (tp->lost_out)
2435		return 1;
2436
2437	/* Not-A-Trick#2 : Classic rule... */
2438	if (tcp_dupack_heuristics(tp) > tp->reordering)
2439		return 1;
2440
2441	/* Trick#3 : when we use RFC2988 timer restart, fast
2442	 * retransmit can be triggered by timeout of queue head.
2443	 */
2444	if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2445		return 1;
2446
2447	/* Trick#4: It is still not OK... But will it be useful to delay
2448	 * recovery more?
2449	 */
2450	packets_out = tp->packets_out;
2451	if (packets_out <= tp->reordering &&
2452	    tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2453	    !tcp_may_send_now(sk)) {
2454		/* We have nothing to send. This connection is limited
2455		 * either by receiver window or by application.
2456		 */
2457		return 1;
2458	}
2459
2460	/* If a thin stream is detected, retransmit after first
2461	 * received dupack. Employ only if SACK is supported in order
2462	 * to avoid possible corner-case series of spurious retransmissions
2463	 * Use only if there are no unsent data.
2464	 */
2465	if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2466	    tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2467	    tcp_is_sack(tp) && !tcp_send_head(sk))
2468		return 1;
2469
2470	return 0;
2471}
2472
2473/* New heuristics: it is possible only after we switched to restart timer
2474 * each time when something is ACKed. Hence, we can detect timed out packets
2475 * during fast retransmit without falling to slow start.
2476 *
2477 * Usefulness of this as is very questionable, since we should know which of
2478 * the segments is the next to timeout which is relatively expensive to find
2479 * in general case unless we add some data structure just for that. The
2480 * current approach certainly won't find the right one too often and when it
2481 * finally does find _something_ it usually marks large part of the window
2482 * right away (because a retransmission with a larger timestamp blocks the
2483 * loop from advancing). -ij
2484 */
2485static void tcp_timeout_skbs(struct sock *sk)
2486{
2487	struct tcp_sock *tp = tcp_sk(sk);
2488	struct sk_buff *skb;
2489
2490	if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2491		return;
2492
2493	skb = tp->scoreboard_skb_hint;
2494	if (tp->scoreboard_skb_hint == NULL)
2495		skb = tcp_write_queue_head(sk);
2496
2497	tcp_for_write_queue_from(skb, sk) {
2498		if (skb == tcp_send_head(sk))
2499			break;
2500		if (!tcp_skb_timedout(sk, skb))
2501			break;
2502
2503		tcp_skb_mark_lost(tp, skb);
2504	}
2505
2506	tp->scoreboard_skb_hint = skb;
2507
2508	tcp_verify_left_out(tp);
2509}
2510
2511/* Detect loss in event "A" above by marking head of queue up as lost.
2512 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2513 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2514 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2515 * the maximum SACKed segments to pass before reaching this limit.
2516 */
2517static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2518{
2519	struct tcp_sock *tp = tcp_sk(sk);
2520	struct sk_buff *skb;
2521	int cnt, oldcnt;
2522	int err;
2523	unsigned int mss;
2524	/* Use SACK to deduce losses of new sequences sent during recovery */
2525	const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2526
2527	WARN_ON(packets > tp->packets_out);
2528	if (tp->lost_skb_hint) {
2529		skb = tp->lost_skb_hint;
2530		cnt = tp->lost_cnt_hint;
2531		/* Head already handled? */
2532		if (mark_head && skb != tcp_write_queue_head(sk))
2533			return;
2534	} else {
2535		skb = tcp_write_queue_head(sk);
2536		cnt = 0;
2537	}
2538
2539	tcp_for_write_queue_from(skb, sk) {
2540		if (skb == tcp_send_head(sk))
2541			break;
2542		/* TODO: do this better */
2543		/* this is not the most efficient way to do this... */
2544		tp->lost_skb_hint = skb;
2545		tp->lost_cnt_hint = cnt;
2546
2547		if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2548			break;
2549
2550		oldcnt = cnt;
2551		if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2552		    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2553			cnt += tcp_skb_pcount(skb);
2554
2555		if (cnt > packets) {
2556			if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2557			    (oldcnt >= packets))
2558				break;
2559
2560			mss = skb_shinfo(skb)->gso_size;
2561			err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2562			if (err < 0)
2563				break;
2564			cnt = packets;
2565		}
2566
2567		tcp_skb_mark_lost(tp, skb);
2568
2569		if (mark_head)
2570			break;
2571	}
2572	tcp_verify_left_out(tp);
2573}
2574
2575/* Account newly detected lost packet(s) */
2576
2577static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2578{
2579	struct tcp_sock *tp = tcp_sk(sk);
2580
2581	if (tcp_is_reno(tp)) {
2582		tcp_mark_head_lost(sk, 1, 1);
2583	} else if (tcp_is_fack(tp)) {
2584		int lost = tp->fackets_out - tp->reordering;
2585		if (lost <= 0)
2586			lost = 1;
2587		tcp_mark_head_lost(sk, lost, 0);
2588	} else {
2589		int sacked_upto = tp->sacked_out - tp->reordering;
2590		if (sacked_upto >= 0)
2591			tcp_mark_head_lost(sk, sacked_upto, 0);
2592		else if (fast_rexmit)
2593			tcp_mark_head_lost(sk, 1, 1);
2594	}
2595
2596	tcp_timeout_skbs(sk);
2597}
2598
2599/* CWND moderation, preventing bursts due to too big ACKs
2600 * in dubious situations.
2601 */
2602static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2603{
2604	tp->snd_cwnd = min(tp->snd_cwnd,
2605			   tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2606	tp->snd_cwnd_stamp = tcp_time_stamp;
2607}
2608
2609/* Lower bound on congestion window is slow start threshold
2610 * unless congestion avoidance choice decides to overide it.
2611 */
2612static inline u32 tcp_cwnd_min(const struct sock *sk)
2613{
2614	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2615
2616	return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2617}
2618
2619/* Decrease cwnd each second ack. */
2620static void tcp_cwnd_down(struct sock *sk, int flag)
2621{
2622	struct tcp_sock *tp = tcp_sk(sk);
2623	int decr = tp->snd_cwnd_cnt + 1;
2624
2625	if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2626	    (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2627		tp->snd_cwnd_cnt = decr & 1;
2628		decr >>= 1;
2629
2630		if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2631			tp->snd_cwnd -= decr;
2632
2633		tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2634		tp->snd_cwnd_stamp = tcp_time_stamp;
2635	}
2636}
2637
2638/* Nothing was retransmitted or returned timestamp is less
2639 * than timestamp of the first retransmission.
2640 */
2641static inline int tcp_packet_delayed(const struct tcp_sock *tp)
2642{
2643	return !tp->retrans_stamp ||
2644		(tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2645		 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2646}
2647
2648/* Undo procedures. */
2649
2650#if FASTRETRANS_DEBUG > 1
2651static void DBGUNDO(struct sock *sk, const char *msg)
2652{
2653	struct tcp_sock *tp = tcp_sk(sk);
2654	struct inet_sock *inet = inet_sk(sk);
2655
2656	if (sk->sk_family == AF_INET) {
2657		printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2658		       msg,
2659		       &inet->inet_daddr, ntohs(inet->inet_dport),
2660		       tp->snd_cwnd, tcp_left_out(tp),
2661		       tp->snd_ssthresh, tp->prior_ssthresh,
2662		       tp->packets_out);
2663	}
2664#if IS_ENABLED(CONFIG_IPV6)
2665	else if (sk->sk_family == AF_INET6) {
2666		struct ipv6_pinfo *np = inet6_sk(sk);
2667		printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2668		       msg,
2669		       &np->daddr, ntohs(inet->inet_dport),
2670		       tp->snd_cwnd, tcp_left_out(tp),
2671		       tp->snd_ssthresh, tp->prior_ssthresh,
2672		       tp->packets_out);
2673	}
2674#endif
2675}
2676#else
2677#define DBGUNDO(x...) do { } while (0)
2678#endif
2679
2680static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2681{
2682	struct tcp_sock *tp = tcp_sk(sk);
2683
2684	if (tp->prior_ssthresh) {
2685		const struct inet_connection_sock *icsk = inet_csk(sk);
2686
2687		if (icsk->icsk_ca_ops->undo_cwnd)
2688			tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2689		else
2690			tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2691
2692		if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2693			tp->snd_ssthresh = tp->prior_ssthresh;
2694			TCP_ECN_withdraw_cwr(tp);
2695		}
2696	} else {
2697		tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2698	}
2699	tp->snd_cwnd_stamp = tcp_time_stamp;
2700}
2701
2702static inline int tcp_may_undo(const struct tcp_sock *tp)
2703{
2704	return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2705}
2706
2707/* People celebrate: "We love our President!" */
2708static int tcp_try_undo_recovery(struct sock *sk)
2709{
2710	struct tcp_sock *tp = tcp_sk(sk);
2711
2712	if (tcp_may_undo(tp)) {
2713		int mib_idx;
2714
2715		/* Happy end! We did not retransmit anything
2716		 * or our original transmission succeeded.
2717		 */
2718		DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2719		tcp_undo_cwr(sk, true);
2720		if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2721			mib_idx = LINUX_MIB_TCPLOSSUNDO;
2722		else
2723			mib_idx = LINUX_MIB_TCPFULLUNDO;
2724
2725		NET_INC_STATS_BH(sock_net(sk), mib_idx);
2726		tp->undo_marker = 0;
2727	}
2728	if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2729		/* Hold old state until something *above* high_seq
2730		 * is ACKed. For Reno it is MUST to prevent false
2731		 * fast retransmits (RFC2582). SACK TCP is safe. */
2732		tcp_moderate_cwnd(tp);
2733		return 1;
2734	}
2735	tcp_set_ca_state(sk, TCP_CA_Open);
2736	return 0;
2737}
2738
2739/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2740static void tcp_try_undo_dsack(struct sock *sk)
2741{
2742	struct tcp_sock *tp = tcp_sk(sk);
2743
2744	if (tp->undo_marker && !tp->undo_retrans) {
2745		DBGUNDO(sk, "D-SACK");
2746		tcp_undo_cwr(sk, true);
2747		tp->undo_marker = 0;
2748		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2749	}
2750}
2751
2752/* We can clear retrans_stamp when there are no retransmissions in the
2753 * window. It would seem that it is trivially available for us in
2754 * tp->retrans_out, however, that kind of assumptions doesn't consider
2755 * what will happen if errors occur when sending retransmission for the
2756 * second time. ...It could the that such segment has only
2757 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2758 * the head skb is enough except for some reneging corner cases that
2759 * are not worth the effort.
2760 *
2761 * Main reason for all this complexity is the fact that connection dying
2762 * time now depends on the validity of the retrans_stamp, in particular,
2763 * that successive retransmissions of a segment must not advance
2764 * retrans_stamp under any conditions.
2765 */
2766static int tcp_any_retrans_done(const struct sock *sk)
2767{
2768	const struct tcp_sock *tp = tcp_sk(sk);
2769	struct sk_buff *skb;
2770
2771	if (tp->retrans_out)
2772		return 1;
2773
2774	skb = tcp_write_queue_head(sk);
2775	if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2776		return 1;
2777
2778	return 0;
2779}
2780
2781/* Undo during fast recovery after partial ACK. */
2782
2783static int tcp_try_undo_partial(struct sock *sk, int acked)
2784{
2785	struct tcp_sock *tp = tcp_sk(sk);
2786	/* Partial ACK arrived. Force Hoe's retransmit. */
2787	int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2788
2789	if (tcp_may_undo(tp)) {
2790		/* Plain luck! Hole if filled with delayed
2791		 * packet, rather than with a retransmit.
2792		 */
2793		if (!tcp_any_retrans_done(sk))
2794			tp->retrans_stamp = 0;
2795
2796		tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2797
2798		DBGUNDO(sk, "Hoe");
2799		tcp_undo_cwr(sk, false);
2800		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2801
2802		/* So... Do not make Hoe's retransmit yet.
2803		 * If the first packet was delayed, the rest
2804		 * ones are most probably delayed as well.
2805		 */
2806		failed = 0;
2807	}
2808	return failed;
2809}
2810
2811/* Undo during loss recovery after partial ACK. */
2812static int tcp_try_undo_loss(struct sock *sk)
2813{
2814	struct tcp_sock *tp = tcp_sk(sk);
2815
2816	if (tcp_may_undo(tp)) {
2817		struct sk_buff *skb;
2818		tcp_for_write_queue(skb, sk) {
2819			if (skb == tcp_send_head(sk))
2820				break;
2821			TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2822		}
2823
2824		tcp_clear_all_retrans_hints(tp);
2825
2826		DBGUNDO(sk, "partial loss");
2827		tp->lost_out = 0;
2828		tcp_undo_cwr(sk, true);
2829		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2830		inet_csk(sk)->icsk_retransmits = 0;
2831		tp->undo_marker = 0;
2832		if (tcp_is_sack(tp))
2833			tcp_set_ca_state(sk, TCP_CA_Open);
2834		return 1;
2835	}
2836	return 0;
2837}
2838
2839static inline void tcp_complete_cwr(struct sock *sk)
2840{
2841	struct tcp_sock *tp = tcp_sk(sk);
2842
2843	/* Do not moderate cwnd if it's already undone in cwr or recovery. */
2844	if (tp->undo_marker) {
2845		if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR)
2846			tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2847		else /* PRR */
2848			tp->snd_cwnd = tp->snd_ssthresh;
2849		tp->snd_cwnd_stamp = tcp_time_stamp;
2850	}
2851	tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2852}
2853
2854static void tcp_try_keep_open(struct sock *sk)
2855{
2856	struct tcp_sock *tp = tcp_sk(sk);
2857	int state = TCP_CA_Open;
2858
2859	if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2860		state = TCP_CA_Disorder;
2861
2862	if (inet_csk(sk)->icsk_ca_state != state) {
2863		tcp_set_ca_state(sk, state);
2864		tp->high_seq = tp->snd_nxt;
2865	}
2866}
2867
2868static void tcp_try_to_open(struct sock *sk, int flag)
2869{
2870	struct tcp_sock *tp = tcp_sk(sk);
2871
2872	tcp_verify_left_out(tp);
2873
2874	if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2875		tp->retrans_stamp = 0;
2876
2877	if (flag & FLAG_ECE)
2878		tcp_enter_cwr(sk, 1);
2879
2880	if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2881		tcp_try_keep_open(sk);
2882		if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2883			tcp_moderate_cwnd(tp);
2884	} else {
2885		tcp_cwnd_down(sk, flag);
2886	}
2887}
2888
2889static void tcp_mtup_probe_failed(struct sock *sk)
2890{
2891	struct inet_connection_sock *icsk = inet_csk(sk);
2892
2893	icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2894	icsk->icsk_mtup.probe_size = 0;
2895}
2896
2897static void tcp_mtup_probe_success(struct sock *sk)
2898{
2899	struct tcp_sock *tp = tcp_sk(sk);
2900	struct inet_connection_sock *icsk = inet_csk(sk);
2901
2902	/* FIXME: breaks with very large cwnd */
2903	tp->prior_ssthresh = tcp_current_ssthresh(sk);
2904	tp->snd_cwnd = tp->snd_cwnd *
2905		       tcp_mss_to_mtu(sk, tp->mss_cache) /
2906		       icsk->icsk_mtup.probe_size;
2907	tp->snd_cwnd_cnt = 0;
2908	tp->snd_cwnd_stamp = tcp_time_stamp;
2909	tp->snd_ssthresh = tcp_current_ssthresh(sk);
2910
2911	icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2912	icsk->icsk_mtup.probe_size = 0;
2913	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2914}
2915
2916/* Do a simple retransmit without using the backoff mechanisms in
2917 * tcp_timer. This is used for path mtu discovery.
2918 * The socket is already locked here.
2919 */
2920void tcp_simple_retransmit(struct sock *sk)
2921{
2922	const struct inet_connection_sock *icsk = inet_csk(sk);
2923	struct tcp_sock *tp = tcp_sk(sk);
2924	struct sk_buff *skb;
2925	unsigned int mss = tcp_current_mss(sk);
2926	u32 prior_lost = tp->lost_out;
2927
2928	tcp_for_write_queue(skb, sk) {
2929		if (skb == tcp_send_head(sk))
2930			break;
2931		if (tcp_skb_seglen(skb) > mss &&
2932		    !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2933			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2934				TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2935				tp->retrans_out -= tcp_skb_pcount(skb);
2936			}
2937			tcp_skb_mark_lost_uncond_verify(tp, skb);
2938		}
2939	}
2940
2941	tcp_clear_retrans_hints_partial(tp);
2942
2943	if (prior_lost == tp->lost_out)
2944		return;
2945
2946	if (tcp_is_reno(tp))
2947		tcp_limit_reno_sacked(tp);
2948
2949	tcp_verify_left_out(tp);
2950
2951	/* Don't muck with the congestion window here.
2952	 * Reason is that we do not increase amount of _data_
2953	 * in network, but units changed and effective
2954	 * cwnd/ssthresh really reduced now.
2955	 */
2956	if (icsk->icsk_ca_state != TCP_CA_Loss) {
2957		tp->high_seq = tp->snd_nxt;
2958		tp->snd_ssthresh = tcp_current_ssthresh(sk);
2959		tp->prior_ssthresh = 0;
2960		tp->undo_marker = 0;
2961		tcp_set_ca_state(sk, TCP_CA_Loss);
2962	}
2963	tcp_xmit_retransmit_queue(sk);
2964}
2965EXPORT_SYMBOL(tcp_simple_retransmit);
2966
2967/* This function implements the PRR algorithm, specifcally the PRR-SSRB
2968 * (proportional rate reduction with slow start reduction bound) as described in
2969 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2970 * It computes the number of packets to send (sndcnt) based on packets newly
2971 * delivered:
2972 *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2973 *	cwnd reductions across a full RTT.
2974 *   2) If packets in flight is lower than ssthresh (such as due to excess
2975 *	losses and/or application stalls), do not perform any further cwnd
2976 *	reductions, but instead slow start up to ssthresh.
2977 */
2978static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
2979					int fast_rexmit, int flag)
2980{
2981	struct tcp_sock *tp = tcp_sk(sk);
2982	int sndcnt = 0;
2983	int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2984
2985	if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2986		u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2987			       tp->prior_cwnd - 1;
2988		sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2989	} else {
2990		sndcnt = min_t(int, delta,
2991			       max_t(int, tp->prr_delivered - tp->prr_out,
2992				     newly_acked_sacked) + 1);
2993	}
2994
2995	sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2996	tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2997}
2998
2999/* Process an event, which can update packets-in-flight not trivially.
3000 * Main goal of this function is to calculate new estimate for left_out,
3001 * taking into account both packets sitting in receiver's buffer and
3002 * packets lost by network.
3003 *
3004 * Besides that it does CWND reduction, when packet loss is detected
3005 * and changes state of machine.
3006 *
3007 * It does _not_ decide what to send, it is made in function
3008 * tcp_xmit_retransmit_queue().
3009 */
3010static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
3011				  int newly_acked_sacked, bool is_dupack,
3012				  int flag)
3013{
3014	struct inet_connection_sock *icsk = inet_csk(sk);
3015	struct tcp_sock *tp = tcp_sk(sk);
3016	int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
3017				    (tcp_fackets_out(tp) > tp->reordering));
3018	int fast_rexmit = 0, mib_idx;
3019
3020	if (WARN_ON(!tp->packets_out && tp->sacked_out))
3021		tp->sacked_out = 0;
3022	if (WARN_ON(!tp->sacked_out && tp->fackets_out))
3023		tp->fackets_out = 0;
3024
3025	/* Now state machine starts.
3026	 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3027	if (flag & FLAG_ECE)
3028		tp->prior_ssthresh = 0;
3029
3030	/* B. In all the states check for reneging SACKs. */
3031	if (tcp_check_sack_reneging(sk, flag))
3032		return;
3033
3034	/* C. Check consistency of the current state. */
3035	tcp_verify_left_out(tp);
3036
3037	/* D. Check state exit conditions. State can be terminated
3038	 *    when high_seq is ACKed. */
3039	if (icsk->icsk_ca_state == TCP_CA_Open) {
3040		WARN_ON(tp->retrans_out != 0);
3041		tp->retrans_stamp = 0;
3042	} else if (!before(tp->snd_una, tp->high_seq)) {
3043		switch (icsk->icsk_ca_state) {
3044		case TCP_CA_Loss:
3045			icsk->icsk_retransmits = 0;
3046			if (tcp_try_undo_recovery(sk))
3047				return;
3048			break;
3049
3050		case TCP_CA_CWR:
3051			/* CWR is to be held something *above* high_seq
3052			 * is ACKed for CWR bit to reach receiver. */
3053			if (tp->snd_una != tp->high_seq) {
3054				tcp_complete_cwr(sk);
3055				tcp_set_ca_state(sk, TCP_CA_Open);
3056			}
3057			break;
3058
3059		case TCP_CA_Recovery:
3060			if (tcp_is_reno(tp))
3061				tcp_reset_reno_sack(tp);
3062			if (tcp_try_undo_recovery(sk))
3063				return;
3064			tcp_complete_cwr(sk);
3065			break;
3066		}
3067	}
3068
3069	/* E. Process state. */
3070	switch (icsk->icsk_ca_state) {
3071	case TCP_CA_Recovery:
3072		if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3073			if (tcp_is_reno(tp) && is_dupack)
3074				tcp_add_reno_sack(sk);
3075		} else
3076			do_lost = tcp_try_undo_partial(sk, pkts_acked);
3077		break;
3078	case TCP_CA_Loss:
3079		if (flag & FLAG_DATA_ACKED)
3080			icsk->icsk_retransmits = 0;
3081		if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3082			tcp_reset_reno_sack(tp);
3083		if (!tcp_try_undo_loss(sk)) {
3084			tcp_moderate_cwnd(tp);
3085			tcp_xmit_retransmit_queue(sk);
3086			return;
3087		}
3088		if (icsk->icsk_ca_state != TCP_CA_Open)
3089			return;
3090		/* Loss is undone; fall through to processing in Open state. */
3091	default:
3092		if (tcp_is_reno(tp)) {
3093			if (flag & FLAG_SND_UNA_ADVANCED)
3094				tcp_reset_reno_sack(tp);
3095			if (is_dupack)
3096				tcp_add_reno_sack(sk);
3097		}
3098
3099		if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3100			tcp_try_undo_dsack(sk);
3101
3102		if (!tcp_time_to_recover(sk)) {
3103			tcp_try_to_open(sk, flag);
3104			return;
3105		}
3106
3107		/* MTU probe failure: don't reduce cwnd */
3108		if (icsk->icsk_ca_state < TCP_CA_CWR &&
3109		    icsk->icsk_mtup.probe_size &&
3110		    tp->snd_una == tp->mtu_probe.probe_seq_start) {
3111			tcp_mtup_probe_failed(sk);
3112			/* Restores the reduction we did in tcp_mtup_probe() */
3113			tp->snd_cwnd++;
3114			tcp_simple_retransmit(sk);
3115			return;
3116		}
3117
3118		/* Otherwise enter Recovery state */
3119
3120		if (tcp_is_reno(tp))
3121			mib_idx = LINUX_MIB_TCPRENORECOVERY;
3122		else
3123			mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3124
3125		NET_INC_STATS_BH(sock_net(sk), mib_idx);
3126
3127		tp->high_seq = tp->snd_nxt;
3128		tp->prior_ssthresh = 0;
3129		tp->undo_marker = tp->snd_una;
3130		tp->undo_retrans = tp->retrans_out;
3131
3132		if (icsk->icsk_ca_state < TCP_CA_CWR) {
3133			if (!(flag & FLAG_ECE))
3134				tp->prior_ssthresh = tcp_current_ssthresh(sk);
3135			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3136			TCP_ECN_queue_cwr(tp);
3137		}
3138
3139		tp->bytes_acked = 0;
3140		tp->snd_cwnd_cnt = 0;
3141		tp->prior_cwnd = tp->snd_cwnd;
3142		tp->prr_delivered = 0;
3143		tp->prr_out = 0;
3144		tcp_set_ca_state(sk, TCP_CA_Recovery);
3145		fast_rexmit = 1;
3146	}
3147
3148	if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3149		tcp_update_scoreboard(sk, fast_rexmit);
3150	tp->prr_delivered += newly_acked_sacked;
3151	tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3152	tcp_xmit_retransmit_queue(sk);
3153}
3154
3155void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3156{
3157	tcp_rtt_estimator(sk, seq_rtt);
3158	tcp_set_rto(sk);
3159	inet_csk(sk)->icsk_backoff = 0;
3160}
3161EXPORT_SYMBOL(tcp_valid_rtt_meas);
3162
3163/* Read draft-ietf-tcplw-high-performance before mucking
3164 * with this code. (Supersedes RFC1323)
3165 */
3166static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3167{
3168	/* RTTM Rule: A TSecr value received in a segment is used to
3169	 * update the averaged RTT measurement only if the segment
3170	 * acknowledges some new data, i.e., only if it advances the
3171	 * left edge of the send window.
3172	 *
3173	 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3174	 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3175	 *
3176	 * Changed: reset backoff as soon as we see the first valid sample.
3177	 * If we do not, we get strongly overestimated rto. With timestamps
3178	 * samples are accepted even from very old segments: f.e., when rtt=1
3179	 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3180	 * answer arrives rto becomes 120 seconds! If at least one of segments
3181	 * in window is lost... Voila.	 			--ANK (010210)
3182	 */
3183	struct tcp_sock *tp = tcp_sk(sk);
3184
3185	tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3186}
3187
3188static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3189{
3190	/* We don't have a timestamp. Can only use
3191	 * packets that are not retransmitted to determine
3192	 * rtt estimates. Also, we must not reset the
3193	 * backoff for rto until we get a non-retransmitted
3194	 * packet. This allows us to deal with a situation
3195	 * where the network delay has increased suddenly.
3196	 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3197	 */
3198
3199	if (flag & FLAG_RETRANS_DATA_ACKED)
3200		return;
3201
3202	tcp_valid_rtt_meas(sk, seq_rtt);
3203}
3204
3205static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3206				      const s32 seq_rtt)
3207{
3208	const struct tcp_sock *tp = tcp_sk(sk);
3209	/* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3210	if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3211		tcp_ack_saw_tstamp(sk, flag);
3212	else if (seq_rtt >= 0)
3213		tcp_ack_no_tstamp(sk, seq_rtt, flag);
3214}
3215
3216static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3217{
3218	const struct inet_connection_sock *icsk = inet_csk(sk);
3219	icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3220	tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3221}
3222
3223/* Restart timer after forward progress on connection.
3224 * RFC2988 recommends to restart timer to now+rto.
3225 */
3226static void tcp_rearm_rto(struct sock *sk)
3227{
3228	const struct tcp_sock *tp = tcp_sk(sk);
3229
3230	if (!tp->packets_out) {
3231		inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3232	} else {
3233		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3234					  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3235	}
3236}
3237
3238/* If we get here, the whole TSO packet has not been acked. */
3239static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3240{
3241	struct tcp_sock *tp = tcp_sk(sk);
3242	u32 packets_acked;
3243
3244	BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3245
3246	packets_acked = tcp_skb_pcount(skb);
3247	if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3248		return 0;
3249	packets_acked -= tcp_skb_pcount(skb);
3250
3251	if (packets_acked) {
3252		BUG_ON(tcp_skb_pcount(skb) == 0);
3253		BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3254	}
3255
3256	return packets_acked;
3257}
3258
3259/* Remove acknowledged frames from the retransmission queue. If our packet
3260 * is before the ack sequence we can discard it as it's confirmed to have
3261 * arrived at the other end.
3262 */
3263static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3264			       u32 prior_snd_una)
3265{
3266	struct tcp_sock *tp = tcp_sk(sk);
3267	const struct inet_connection_sock *icsk = inet_csk(sk);
3268	struct sk_buff *skb;
3269	u32 now = tcp_time_stamp;
3270	int fully_acked = 1;
3271	int flag = 0;
3272	u32 pkts_acked = 0;
3273	u32 reord = tp->packets_out;
3274	u32 prior_sacked = tp->sacked_out;
3275	s32 seq_rtt = -1;
3276	s32 ca_seq_rtt = -1;
3277	ktime_t last_ackt = net_invalid_timestamp();
3278
3279	while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3280		struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3281		u32 acked_pcount;
3282		u8 sacked = scb->sacked;
3283
3284		/* Determine how many packets and what bytes were acked, tso and else */
3285		if (after(scb->end_seq, tp->snd_una)) {
3286			if (tcp_skb_pcount(skb) == 1 ||
3287			    !after(tp->snd_una, scb->seq))
3288				break;
3289
3290			acked_pcount = tcp_tso_acked(sk, skb);
3291			if (!acked_pcount)
3292				break;
3293
3294			fully_acked = 0;
3295		} else {
3296			acked_pcount = tcp_skb_pcount(skb);
3297		}
3298
3299		if (sacked & TCPCB_RETRANS) {
3300			if (sacked & TCPCB_SACKED_RETRANS)
3301				tp->retrans_out -= acked_pcount;
3302			flag |= FLAG_RETRANS_DATA_ACKED;
3303			ca_seq_rtt = -1;
3304			seq_rtt = -1;
3305			if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3306				flag |= FLAG_NONHEAD_RETRANS_ACKED;
3307		} else {
3308			ca_seq_rtt = now - scb->when;
3309			last_ackt = skb->tstamp;
3310			if (seq_rtt < 0) {
3311				seq_rtt = ca_seq_rtt;
3312			}
3313			if (!(sacked & TCPCB_SACKED_ACKED))
3314				reord = min(pkts_acked, reord);
3315		}
3316
3317		if (sacked & TCPCB_SACKED_ACKED)
3318			tp->sacked_out -= acked_pcount;
3319		if (sacked & TCPCB_LOST)
3320			tp->lost_out -= acked_pcount;
3321
3322		tp->packets_out -= acked_pcount;
3323		pkts_acked += acked_pcount;
3324
3325		/* Initial outgoing SYN's get put onto the write_queue
3326		 * just like anything else we transmit.  It is not
3327		 * true data, and if we misinform our callers that
3328		 * this ACK acks real data, we will erroneously exit
3329		 * connection startup slow start one packet too
3330		 * quickly.  This is severely frowned upon behavior.
3331		 */
3332		if (!(scb->tcp_flags & TCPHDR_SYN)) {
3333			flag |= FLAG_DATA_ACKED;
3334		} else {
3335			flag |= FLAG_SYN_ACKED;
3336			tp->retrans_stamp = 0;
3337		}
3338
3339		if (!fully_acked)
3340			break;
3341
3342		tcp_unlink_write_queue(skb, sk);
3343		sk_wmem_free_skb(sk, skb);
3344		tp->scoreboard_skb_hint = NULL;
3345		if (skb == tp->retransmit_skb_hint)
3346			tp->retransmit_skb_hint = NULL;
3347		if (skb == tp->lost_skb_hint)
3348			tp->lost_skb_hint = NULL;
3349	}
3350
3351	if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3352		tp->snd_up = tp->snd_una;
3353
3354	if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3355		flag |= FLAG_SACK_RENEGING;
3356
3357	if (flag & FLAG_ACKED) {
3358		const struct tcp_congestion_ops *ca_ops
3359			= inet_csk(sk)->icsk_ca_ops;
3360
3361		if (unlikely(icsk->icsk_mtup.probe_size &&
3362			     !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3363			tcp_mtup_probe_success(sk);
3364		}
3365
3366		tcp_ack_update_rtt(sk, flag, seq_rtt);
3367		tcp_rearm_rto(sk);
3368
3369		if (tcp_is_reno(tp)) {
3370			tcp_remove_reno_sacks(sk, pkts_acked);
3371		} else {
3372			int delta;
3373
3374			/* Non-retransmitted hole got filled? That's reordering */
3375			if (reord < prior_fackets)
3376				tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3377
3378			delta = tcp_is_fack(tp) ? pkts_acked :
3379						  prior_sacked - tp->sacked_out;
3380			tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3381		}
3382
3383		tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3384
3385		if (ca_ops->pkts_acked) {
3386			s32 rtt_us = -1;
3387
3388			/* Is the ACK triggering packet unambiguous? */
3389			if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3390				/* High resolution needed and available? */
3391				if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3392				    !ktime_equal(last_ackt,
3393						 net_invalid_timestamp()))
3394					rtt_us = ktime_us_delta(ktime_get_real(),
3395								last_ackt);
3396				else if (ca_seq_rtt >= 0)
3397					rtt_us = jiffies_to_usecs(ca_seq_rtt);
3398			}
3399
3400			ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3401		}
3402	}
3403
3404#if FASTRETRANS_DEBUG > 0
3405	WARN_ON((int)tp->sacked_out < 0);
3406	WARN_ON((int)tp->lost_out < 0);
3407	WARN_ON((int)tp->retrans_out < 0);
3408	if (!tp->packets_out && tcp_is_sack(tp)) {
3409		icsk = inet_csk(sk);
3410		if (tp->lost_out) {
3411			printk(KERN_DEBUG "Leak l=%u %d\n",
3412			       tp->lost_out, icsk->icsk_ca_state);
3413			tp->lost_out = 0;
3414		}
3415		if (tp->sacked_out) {
3416			printk(KERN_DEBUG "Leak s=%u %d\n",
3417			       tp->sacked_out, icsk->icsk_ca_state);
3418			tp->sacked_out = 0;
3419		}
3420		if (tp->retrans_out) {
3421			printk(KERN_DEBUG "Leak r=%u %d\n",
3422			       tp->retrans_out, icsk->icsk_ca_state);
3423			tp->retrans_out = 0;
3424		}
3425	}
3426#endif
3427	return flag;
3428}
3429
3430static void tcp_ack_probe(struct sock *sk)
3431{
3432	const struct tcp_sock *tp = tcp_sk(sk);
3433	struct inet_connection_sock *icsk = inet_csk(sk);
3434
3435	/* Was it a usable window open? */
3436
3437	if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3438		icsk->icsk_backoff = 0;
3439		inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3440		/* Socket must be waked up by subsequent tcp_data_snd_check().
3441		 * This function is not for random using!
3442		 */
3443	} else {
3444		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3445					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3446					  TCP_RTO_MAX);
3447	}
3448}
3449
3450static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3451{
3452	return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3453		inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3454}
3455
3456static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3457{
3458	const struct tcp_sock *tp = tcp_sk(sk);
3459	return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3460		!((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3461}
3462
3463/* Check that window update is acceptable.
3464 * The function assumes that snd_una<=ack<=snd_next.
3465 */
3466static inline int tcp_may_update_window(const struct tcp_sock *tp,
3467					const u32 ack, const u32 ack_seq,
3468					const u32 nwin)
3469{
3470	return	after(ack, tp->snd_una) ||
3471		after(ack_seq, tp->snd_wl1) ||
3472		(ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3473}
3474
3475/* Update our send window.
3476 *
3477 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3478 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3479 */
3480static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3481				 u32 ack_seq)
3482{
3483	struct tcp_sock *tp = tcp_sk(sk);
3484	int flag = 0;
3485	u32 nwin = ntohs(tcp_hdr(skb)->window);
3486
3487	if (likely(!tcp_hdr(skb)->syn))
3488		nwin <<= tp->rx_opt.snd_wscale;
3489
3490	if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3491		flag |= FLAG_WIN_UPDATE;
3492		tcp_update_wl(tp, ack_seq);
3493
3494		if (tp->snd_wnd != nwin) {
3495			tp->snd_wnd = nwin;
3496
3497			/* Note, it is the only place, where
3498			 * fast path is recovered for sending TCP.
3499			 */
3500			tp->pred_flags = 0;
3501			tcp_fast_path_check(sk);
3502
3503			if (nwin > tp->max_window) {
3504				tp->max_window = nwin;
3505				tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3506			}
3507		}
3508	}
3509
3510	tp->snd_una = ack;
3511
3512	return flag;
3513}
3514
3515/* A very conservative spurious RTO response algorithm: reduce cwnd and
3516 * continue in congestion avoidance.
3517 */
3518static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3519{
3520	tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3521	tp->snd_cwnd_cnt = 0;
3522	tp->bytes_acked = 0;
3523	TCP_ECN_queue_cwr(tp);
3524	tcp_moderate_cwnd(tp);
3525}
3526
3527/* A conservative spurious RTO response algorithm: reduce cwnd using
3528 * rate halving and continue in congestion avoidance.
3529 */
3530static void tcp_ratehalving_spur_to_response(struct sock *sk)
3531{
3532	tcp_enter_cwr(sk, 0);
3533}
3534
3535static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3536{
3537	if (flag & FLAG_ECE)
3538		tcp_ratehalving_spur_to_response(sk);
3539	else
3540		tcp_undo_cwr(sk, true);
3541}
3542
3543/* F-RTO spurious RTO detection algorithm (RFC4138)
3544 *
3545 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3546 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3547 * window (but not to or beyond highest sequence sent before RTO):
3548 *   On First ACK,  send two new segments out.
3549 *   On Second ACK, RTO was likely spurious. Do spurious response (response
3550 *                  algorithm is not part of the F-RTO detection algorithm
3551 *                  given in RFC4138 but can be selected separately).
3552 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3553 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3554 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3555 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3556 *
3557 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3558 * original window even after we transmit two new data segments.
3559 *
3560 * SACK version:
3561 *   on first step, wait until first cumulative ACK arrives, then move to
3562 *   the second step. In second step, the next ACK decides.
3563 *
3564 * F-RTO is implemented (mainly) in four functions:
3565 *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3566 *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3567 *     called when tcp_use_frto() showed green light
3568 *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3569 *   - tcp_enter_frto_loss() is called if there is not enough evidence
3570 *     to prove that the RTO is indeed spurious. It transfers the control
3571 *     from F-RTO to the conventional RTO recovery
3572 */
3573static int tcp_process_frto(struct sock *sk, int flag)
3574{
3575	struct tcp_sock *tp = tcp_sk(sk);
3576
3577	tcp_verify_left_out(tp);
3578
3579	/* Duplicate the behavior from Loss state (fastretrans_alert) */
3580	if (flag & FLAG_DATA_ACKED)
3581		inet_csk(sk)->icsk_retransmits = 0;
3582
3583	if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3584	    ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3585		tp->undo_marker = 0;
3586
3587	if (!before(tp->snd_una, tp->frto_highmark)) {
3588		tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3589		return 1;
3590	}
3591
3592	if (!tcp_is_sackfrto(tp)) {
3593		/* RFC4138 shortcoming in step 2; should also have case c):
3594		 * ACK isn't duplicate nor advances window, e.g., opposite dir
3595		 * data, winupdate
3596		 */
3597		if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3598			return 1;
3599
3600		if (!(flag & FLAG_DATA_ACKED)) {
3601			tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3602					    flag);
3603			return 1;
3604		}
3605	} else {
3606		if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3607			/* Prevent sending of new data. */
3608			tp->snd_cwnd = min(tp->snd_cwnd,
3609					   tcp_packets_in_flight(tp));
3610			return 1;
3611		}
3612
3613		if ((tp->frto_counter >= 2) &&
3614		    (!(flag & FLAG_FORWARD_PROGRESS) ||
3615		     ((flag & FLAG_DATA_SACKED) &&
3616		      !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3617			/* RFC4138 shortcoming (see comment above) */
3618			if (!(flag & FLAG_FORWARD_PROGRESS) &&
3619			    (flag & FLAG_NOT_DUP))
3620				return 1;
3621
3622			tcp_enter_frto_loss(sk, 3, flag);
3623			return 1;
3624		}
3625	}
3626
3627	if (tp->frto_counter == 1) {
3628		/* tcp_may_send_now needs to see updated state */
3629		tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3630		tp->frto_counter = 2;
3631
3632		if (!tcp_may_send_now(sk))
3633			tcp_enter_frto_loss(sk, 2, flag);
3634
3635		return 1;
3636	} else {
3637		switch (sysctl_tcp_frto_response) {
3638		case 2:
3639			tcp_undo_spur_to_response(sk, flag);
3640			break;
3641		case 1:
3642			tcp_conservative_spur_to_response(tp);
3643			break;
3644		default:
3645			tcp_ratehalving_spur_to_response(sk);
3646			break;
3647		}
3648		tp->frto_counter = 0;
3649		tp->undo_marker = 0;
3650		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3651	}
3652	return 0;
3653}
3654
3655/* This routine deals with incoming acks, but not outgoing ones. */
3656static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3657{
3658	struct inet_connection_sock *icsk = inet_csk(sk);
3659	struct tcp_sock *tp = tcp_sk(sk);
3660	u32 prior_snd_una = tp->snd_una;
3661	u32 ack_seq = TCP_SKB_CB(skb)->seq;
3662	u32 ack = TCP_SKB_CB(skb)->ack_seq;
3663	bool is_dupack = false;
3664	u32 prior_in_flight;
3665	u32 prior_fackets;
3666	int prior_packets;
3667	int prior_sacked = tp->sacked_out;
3668	int pkts_acked = 0;
3669	int newly_acked_sacked = 0;
3670	int frto_cwnd = 0;
3671
3672	/* If the ack is older than previous acks
3673	 * then we can probably ignore it.
3674	 */
3675	if (before(ack, prior_snd_una))
3676		goto old_ack;
3677
3678	/* If the ack includes data we haven't sent yet, discard
3679	 * this segment (RFC793 Section 3.9).
3680	 */
3681	if (after(ack, tp->snd_nxt))
3682		goto invalid_ack;
3683
3684	if (after(ack, prior_snd_una))
3685		flag |= FLAG_SND_UNA_ADVANCED;
3686
3687	if (sysctl_tcp_abc) {
3688		if (icsk->icsk_ca_state < TCP_CA_CWR)
3689			tp->bytes_acked += ack - prior_snd_una;
3690		else if (icsk->icsk_ca_state == TCP_CA_Loss)
3691			/* we assume just one segment left network */
3692			tp->bytes_acked += min(ack - prior_snd_una,
3693					       tp->mss_cache);
3694	}
3695
3696	prior_fackets = tp->fackets_out;
3697	prior_in_flight = tcp_packets_in_flight(tp);
3698
3699	if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3700		/* Window is constant, pure forward advance.
3701		 * No more checks are required.
3702		 * Note, we use the fact that SND.UNA>=SND.WL2.
3703		 */
3704		tcp_update_wl(tp, ack_seq);
3705		tp->snd_una = ack;
3706		flag |= FLAG_WIN_UPDATE;
3707
3708		tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3709
3710		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3711	} else {
3712		if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3713			flag |= FLAG_DATA;
3714		else
3715			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3716
3717		flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3718
3719		if (TCP_SKB_CB(skb)->sacked)
3720			flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3721
3722		if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3723			flag |= FLAG_ECE;
3724
3725		tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3726	}
3727
3728	/* We passed data and got it acked, remove any soft error
3729	 * log. Something worked...
3730	 */
3731	sk->sk_err_soft = 0;
3732	icsk->icsk_probes_out = 0;
3733	tp->rcv_tstamp = tcp_time_stamp;
3734	prior_packets = tp->packets_out;
3735	if (!prior_packets)
3736		goto no_queue;
3737
3738	/* See if we can take anything off of the retransmit queue. */
3739	flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3740
3741	pkts_acked = prior_packets - tp->packets_out;
3742	newly_acked_sacked = (prior_packets - prior_sacked) -
3743			     (tp->packets_out - tp->sacked_out);
3744
3745	if (tp->frto_counter)
3746		frto_cwnd = tcp_process_frto(sk, flag);
3747	/* Guarantee sacktag reordering detection against wrap-arounds */
3748	if (before(tp->frto_highmark, tp->snd_una))
3749		tp->frto_highmark = 0;
3750
3751	if (tcp_ack_is_dubious(sk, flag)) {
3752		/* Advance CWND, if state allows this. */
3753		if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3754		    tcp_may_raise_cwnd(sk, flag))
3755			tcp_cong_avoid(sk, ack, prior_in_flight);
3756		is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3757		tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3758				      is_dupack, flag);
3759	} else {
3760		if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3761			tcp_cong_avoid(sk, ack, prior_in_flight);
3762	}
3763
3764	if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3765		dst_confirm(__sk_dst_get(sk));
3766
3767	return 1;
3768
3769no_queue:
3770	/* If data was DSACKed, see if we can undo a cwnd reduction. */
3771	if (flag & FLAG_DSACKING_ACK)
3772		tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3773				      is_dupack, flag);
3774	/* If this ack opens up a zero window, clear backoff.  It was
3775	 * being used to time the probes, and is probably far higher than
3776	 * it needs to be for normal retransmission.
3777	 */
3778	if (tcp_send_head(sk))
3779		tcp_ack_probe(sk);
3780	return 1;
3781
3782invalid_ack:
3783	SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3784	return -1;
3785
3786old_ack:
3787	/* If data was SACKed, tag it and see if we should send more data.
3788	 * If data was DSACKed, see if we can undo a cwnd reduction.
3789	 */
3790	if (TCP_SKB_CB(skb)->sacked) {
3791		flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3792		newly_acked_sacked = tp->sacked_out - prior_sacked;
3793		tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3794				      is_dupack, flag);
3795	}
3796
3797	SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3798	return 0;
3799}
3800
3801/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3802 * But, this can also be called on packets in the established flow when
3803 * the fast version below fails.
3804 */
3805void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3806		       const u8 **hvpp, int estab)
3807{
3808	const unsigned char *ptr;
3809	const struct tcphdr *th = tcp_hdr(skb);
3810	int length = (th->doff * 4) - sizeof(struct tcphdr);
3811
3812	ptr = (const unsigned char *)(th + 1);
3813	opt_rx->saw_tstamp = 0;
3814
3815	while (length > 0) {
3816		int opcode = *ptr++;
3817		int opsize;
3818
3819		switch (opcode) {
3820		case TCPOPT_EOL:
3821			return;
3822		case TCPOPT_NOP:	/* Ref: RFC 793 section 3.1 */
3823			length--;
3824			continue;
3825		default:
3826			opsize = *ptr++;
3827			if (opsize < 2) /* "silly options" */
3828				return;
3829			if (opsize > length)
3830				return;	/* don't parse partial options */
3831			switch (opcode) {
3832			case TCPOPT_MSS:
3833				if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3834					u16 in_mss = get_unaligned_be16(ptr);
3835					if (in_mss) {
3836						if (opt_rx->user_mss &&
3837						    opt_rx->user_mss < in_mss)
3838							in_mss = opt_rx->user_mss;
3839						opt_rx->mss_clamp = in_mss;
3840					}
3841				}
3842				break;
3843			case TCPOPT_WINDOW:
3844				if (opsize == TCPOLEN_WINDOW && th->syn &&
3845				    !estab && sysctl_tcp_window_scaling) {
3846					__u8 snd_wscale = *(__u8 *)ptr;
3847					opt_rx->wscale_ok = 1;
3848					if (snd_wscale > 14) {
3849						if (net_ratelimit())
3850							printk(KERN_INFO "tcp_parse_options: Illegal window "
3851							       "scaling value %d >14 received.\n",
3852							       snd_wscale);
3853						snd_wscale = 14;
3854					}
3855					opt_rx->snd_wscale = snd_wscale;
3856				}
3857				break;
3858			case TCPOPT_TIMESTAMP:
3859				if ((opsize == TCPOLEN_TIMESTAMP) &&
3860				    ((estab && opt_rx->tstamp_ok) ||
3861				     (!estab && sysctl_tcp_timestamps))) {
3862					opt_rx->saw_tstamp = 1;
3863					opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3864					opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3865				}
3866				break;
3867			case TCPOPT_SACK_PERM:
3868				if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3869				    !estab && sysctl_tcp_sack) {
3870					opt_rx->sack_ok = TCP_SACK_SEEN;
3871					tcp_sack_reset(opt_rx);
3872				}
3873				break;
3874
3875			case TCPOPT_SACK:
3876				if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3877				   !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3878				   opt_rx->sack_ok) {
3879					TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3880				}
3881				break;
3882#ifdef CONFIG_TCP_MD5SIG
3883			case TCPOPT_MD5SIG:
3884				/*
3885				 * The MD5 Hash has already been
3886				 * checked (see tcp_v{4,6}_do_rcv()).
3887				 */
3888				break;
3889#endif
3890			case TCPOPT_COOKIE:
3891				/* This option is variable length.
3892				 */
3893				switch (opsize) {
3894				case TCPOLEN_COOKIE_BASE:
3895					/* not yet implemented */
3896					break;
3897				case TCPOLEN_COOKIE_PAIR:
3898					/* not yet implemented */
3899					break;
3900				case TCPOLEN_COOKIE_MIN+0:
3901				case TCPOLEN_COOKIE_MIN+2:
3902				case TCPOLEN_COOKIE_MIN+4:
3903				case TCPOLEN_COOKIE_MIN+6:
3904				case TCPOLEN_COOKIE_MAX:
3905					/* 16-bit multiple */
3906					opt_rx->cookie_plus = opsize;
3907					*hvpp = ptr;
3908					break;
3909				default:
3910					/* ignore option */
3911					break;
3912				}
3913				break;
3914			}
3915
3916			ptr += opsize-2;
3917			length -= opsize;
3918		}
3919	}
3920}
3921EXPORT_SYMBOL(tcp_parse_options);
3922
3923static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3924{
3925	const __be32 *ptr = (const __be32 *)(th + 1);
3926
3927	if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3928			  | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3929		tp->rx_opt.saw_tstamp = 1;
3930		++ptr;
3931		tp->rx_opt.rcv_tsval = ntohl(*ptr);
3932		++ptr;
3933		tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3934		return 1;
3935	}
3936	return 0;
3937}
3938
3939/* Fast parse options. This hopes to only see timestamps.
3940 * If it is wrong it falls back on tcp_parse_options().
3941 */
3942static int tcp_fast_parse_options(const struct sk_buff *skb,
3943				  const struct tcphdr *th,
3944				  struct tcp_sock *tp, const u8 **hvpp)
3945{
3946	/* In the spirit of fast parsing, compare doff directly to constant
3947	 * values.  Because equality is used, short doff can be ignored here.
3948	 */
3949	if (th->doff == (sizeof(*th) / 4)) {
3950		tp->rx_opt.saw_tstamp = 0;
3951		return 0;
3952	} else if (tp->rx_opt.tstamp_ok &&
3953		   th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3954		if (tcp_parse_aligned_timestamp(tp, th))
3955			return 1;
3956	}
3957	tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3958	return 1;
3959}
3960
3961#ifdef CONFIG_TCP_MD5SIG
3962/*
3963 * Parse MD5 Signature option
3964 */
3965const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3966{
3967	int length = (th->doff << 2) - sizeof(*th);
3968	const u8 *ptr = (const u8 *)(th + 1);
3969
3970	/* If the TCP option is too short, we can short cut */
3971	if (length < TCPOLEN_MD5SIG)
3972		return NULL;
3973
3974	while (length > 0) {
3975		int opcode = *ptr++;
3976		int opsize;
3977
3978		switch(opcode) {
3979		case TCPOPT_EOL:
3980			return NULL;
3981		case TCPOPT_NOP:
3982			length--;
3983			continue;
3984		default:
3985			opsize = *ptr++;
3986			if (opsize < 2 || opsize > length)
3987				return NULL;
3988			if (opcode == TCPOPT_MD5SIG)
3989				return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3990		}
3991		ptr += opsize - 2;
3992		length -= opsize;
3993	}
3994	return NULL;
3995}
3996EXPORT_SYMBOL(tcp_parse_md5sig_option);
3997#endif
3998
3999static inline void tcp_store_ts_recent(struct tcp_sock *tp)
4000{
4001	tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
4002	tp->rx_opt.ts_recent_stamp = get_seconds();
4003}
4004
4005static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
4006{
4007	if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
4008		/* PAWS bug workaround wrt. ACK frames, the PAWS discard
4009		 * extra check below makes sure this can only happen
4010		 * for pure ACK frames.  -DaveM
4011		 *
4012		 * Not only, also it occurs for expired timestamps.
4013		 */
4014
4015		if (tcp_paws_check(&tp->rx_opt, 0))
4016			tcp_store_ts_recent(tp);
4017	}
4018}
4019
4020/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4021 *
4022 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4023 * it can pass through stack. So, the following predicate verifies that
4024 * this segment is not used for anything but congestion avoidance or
4025 * fast retransmit. Moreover, we even are able to eliminate most of such
4026 * second order effects, if we apply some small "replay" window (~RTO)
4027 * to timestamp space.
4028 *
4029 * All these measures still do not guarantee that we reject wrapped ACKs
4030 * on networks with high bandwidth, when sequence space is recycled fastly,
4031 * but it guarantees that such events will be very rare and do not affect
4032 * connection seriously. This doesn't look nice, but alas, PAWS is really
4033 * buggy extension.
4034 *
4035 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4036 * states that events when retransmit arrives after original data are rare.
4037 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4038 * the biggest problem on large power networks even with minor reordering.
4039 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4040 * up to bandwidth of 18Gigabit/sec. 8) ]
4041 */
4042
4043static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4044{
4045	const struct tcp_sock *tp = tcp_sk(sk);
4046	const struct tcphdr *th = tcp_hdr(skb);
4047	u32 seq = TCP_SKB_CB(skb)->seq;
4048	u32 ack = TCP_SKB_CB(skb)->ack_seq;
4049
4050	return (/* 1. Pure ACK with correct sequence number. */
4051		(th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4052
4053		/* 2. ... and duplicate ACK. */
4054		ack == tp->snd_una &&
4055
4056		/* 3. ... and does not update window. */
4057		!tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4058
4059		/* 4. ... and sits in replay window. */
4060		(s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4061}
4062
4063static inline int tcp_paws_discard(const struct sock *sk,
4064				   const struct sk_buff *skb)
4065{
4066	const struct tcp_sock *tp = tcp_sk(sk);
4067
4068	return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4069	       !tcp_disordered_ack(sk, skb);
4070}
4071
4072/* Check segment sequence number for validity.
4073 *
4074 * Segment controls are considered valid, if the segment
4075 * fits to the window after truncation to the window. Acceptability
4076 * of data (and SYN, FIN, of course) is checked separately.
4077 * See tcp_data_queue(), for example.
4078 *
4079 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4080 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4081 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4082 * (borrowed from freebsd)
4083 */
4084
4085static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4086{
4087	return	!before(end_seq, tp->rcv_wup) &&
4088		!after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4089}
4090
4091/* When we get a reset we do this. */
4092static void tcp_reset(struct sock *sk)
4093{
4094	/* We want the right error as BSD sees it (and indeed as we do). */
4095	switch (sk->sk_state) {
4096	case TCP_SYN_SENT:
4097		sk->sk_err = ECONNREFUSED;
4098		break;
4099	case TCP_CLOSE_WAIT:
4100		sk->sk_err = EPIPE;
4101		break;
4102	case TCP_CLOSE:
4103		return;
4104	default:
4105		sk->sk_err = ECONNRESET;
4106	}
4107	/* This barrier is coupled with smp_rmb() in tcp_poll() */
4108	smp_wmb();
4109
4110	if (!sock_flag(sk, SOCK_DEAD))
4111		sk->sk_error_report(sk);
4112
4113	tcp_done(sk);
4114}
4115
4116/*
4117 * 	Process the FIN bit. This now behaves as it is supposed to work
4118 *	and the FIN takes effect when it is validly part of sequence
4119 *	space. Not before when we get holes.
4120 *
4121 *	If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4122 *	(and thence onto LAST-ACK and finally, CLOSE, we never enter
4123 *	TIME-WAIT)
4124 *
4125 *	If we are in FINWAIT-1, a received FIN indicates simultaneous
4126 *	close and we go into CLOSING (and later onto TIME-WAIT)
4127 *
4128 *	If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4129 */
4130static void tcp_fin(struct sock *sk)
4131{
4132	struct tcp_sock *tp = tcp_sk(sk);
4133
4134	inet_csk_schedule_ack(sk);
4135
4136	sk->sk_shutdown |= RCV_SHUTDOWN;
4137	sock_set_flag(sk, SOCK_DONE);
4138
4139	switch (sk->sk_state) {
4140	case TCP_SYN_RECV:
4141	case TCP_ESTABLISHED:
4142		/* Move to CLOSE_WAIT */
4143		tcp_set_state(sk, TCP_CLOSE_WAIT);
4144		inet_csk(sk)->icsk_ack.pingpong = 1;
4145		break;
4146
4147	case TCP_CLOSE_WAIT:
4148	case TCP_CLOSING:
4149		/* Received a retransmission of the FIN, do
4150		 * nothing.
4151		 */
4152		break;
4153	case TCP_LAST_ACK:
4154		/* RFC793: Remain in the LAST-ACK state. */
4155		break;
4156
4157	case TCP_FIN_WAIT1:
4158		/* This case occurs when a simultaneous close
4159		 * happens, we must ack the received FIN and
4160		 * enter the CLOSING state.
4161		 */
4162		tcp_send_ack(sk);
4163		tcp_set_state(sk, TCP_CLOSING);
4164		break;
4165	case TCP_FIN_WAIT2:
4166		/* Received a FIN -- send ACK and enter TIME_WAIT. */
4167		tcp_send_ack(sk);
4168		tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4169		break;
4170	default:
4171		/* Only TCP_LISTEN and TCP_CLOSE are left, in these
4172		 * cases we should never reach this piece of code.
4173		 */
4174		printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4175		       __func__, sk->sk_state);
4176		break;
4177	}
4178
4179	/* It _is_ possible, that we have something out-of-order _after_ FIN.
4180	 * Probably, we should reset in this case. For now drop them.
4181	 */
4182	__skb_queue_purge(&tp->out_of_order_queue);
4183	if (tcp_is_sack(tp))
4184		tcp_sack_reset(&tp->rx_opt);
4185	sk_mem_reclaim(sk);
4186
4187	if (!sock_flag(sk, SOCK_DEAD)) {
4188		sk->sk_state_change(sk);
4189
4190		/* Do not send POLL_HUP for half duplex close. */
4191		if (sk->sk_shutdown == SHUTDOWN_MASK ||
4192		    sk->sk_state == TCP_CLOSE)
4193			sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4194		else
4195			sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4196	}
4197}
4198
4199static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4200				  u32 end_seq)
4201{
4202	if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4203		if (before(seq, sp->start_seq))
4204			sp->start_seq = seq;
4205		if (after(end_seq, sp->end_seq))
4206			sp->end_seq = end_seq;
4207		return 1;
4208	}
4209	return 0;
4210}
4211
4212static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4213{
4214	struct tcp_sock *tp = tcp_sk(sk);
4215
4216	if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4217		int mib_idx;
4218
4219		if (before(seq, tp->rcv_nxt))
4220			mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4221		else
4222			mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4223
4224		NET_INC_STATS_BH(sock_net(sk), mib_idx);
4225
4226		tp->rx_opt.dsack = 1;
4227		tp->duplicate_sack[0].start_seq = seq;
4228		tp->duplicate_sack[0].end_seq = end_seq;
4229	}
4230}
4231
4232static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4233{
4234	struct tcp_sock *tp = tcp_sk(sk);
4235
4236	if (!tp->rx_opt.dsack)
4237		tcp_dsack_set(sk, seq, end_seq);
4238	else
4239		tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4240}
4241
4242static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4243{
4244	struct tcp_sock *tp = tcp_sk(sk);
4245
4246	if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4247	    before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4248		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4249		tcp_enter_quickack_mode(sk);
4250
4251		if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4252			u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4253
4254			if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4255				end_seq = tp->rcv_nxt;
4256			tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4257		}
4258	}
4259
4260	tcp_send_ack(sk);
4261}
4262
4263/* These routines update the SACK block as out-of-order packets arrive or
4264 * in-order packets close up the sequence space.
4265 */
4266static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4267{
4268	int this_sack;
4269	struct tcp_sack_block *sp = &tp->selective_acks[0];
4270	struct tcp_sack_block *swalk = sp + 1;
4271
4272	/* See if the recent change to the first SACK eats into
4273	 * or hits the sequence space of other SACK blocks, if so coalesce.
4274	 */
4275	for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4276		if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4277			int i;
4278
4279			/* Zap SWALK, by moving every further SACK up by one slot.
4280			 * Decrease num_sacks.
4281			 */
4282			tp->rx_opt.num_sacks--;
4283			for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4284				sp[i] = sp[i + 1];
4285			continue;
4286		}
4287		this_sack++, swalk++;
4288	}
4289}
4290
4291static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4292{
4293	struct tcp_sock *tp = tcp_sk(sk);
4294	struct tcp_sack_block *sp = &tp->selective_acks[0];
4295	int cur_sacks = tp->rx_opt.num_sacks;
4296	int this_sack;
4297
4298	if (!cur_sacks)
4299		goto new_sack;
4300
4301	for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4302		if (tcp_sack_extend(sp, seq, end_seq)) {
4303			/* Rotate this_sack to the first one. */
4304			for (; this_sack > 0; this_sack--, sp--)
4305				swap(*sp, *(sp - 1));
4306			if (cur_sacks > 1)
4307				tcp_sack_maybe_coalesce(tp);
4308			return;
4309		}
4310	}
4311
4312	/* Could not find an adjacent existing SACK, build a new one,
4313	 * put it at the front, and shift everyone else down.  We
4314	 * always know there is at least one SACK present already here.
4315	 *
4316	 * If the sack array is full, forget about the last one.
4317	 */
4318	if (this_sack >= TCP_NUM_SACKS) {
4319		this_sack--;
4320		tp->rx_opt.num_sacks--;
4321		sp--;
4322	}
4323	for (; this_sack > 0; this_sack--, sp--)
4324		*sp = *(sp - 1);
4325
4326new_sack:
4327	/* Build the new head SACK, and we're done. */
4328	sp->start_seq = seq;
4329	sp->end_seq = end_seq;
4330	tp->rx_opt.num_sacks++;
4331}
4332
4333/* RCV.NXT advances, some SACKs should be eaten. */
4334
4335static void tcp_sack_remove(struct tcp_sock *tp)
4336{
4337	struct tcp_sack_block *sp = &tp->selective_acks[0];
4338	int num_sacks = tp->rx_opt.num_sacks;
4339	int this_sack;
4340
4341	/* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4342	if (skb_queue_empty(&tp->out_of_order_queue)) {
4343		tp->rx_opt.num_sacks = 0;
4344		return;
4345	}
4346
4347	for (this_sack = 0; this_sack < num_sacks;) {
4348		/* Check if the start of the sack is covered by RCV.NXT. */
4349		if (!before(tp->rcv_nxt, sp->start_seq)) {
4350			int i;
4351
4352			/* RCV.NXT must cover all the block! */
4353			WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4354
4355			/* Zap this SACK, by moving forward any other SACKS. */
4356			for (i=this_sack+1; i < num_sacks; i++)
4357				tp->selective_acks[i-1] = tp->selective_acks[i];
4358			num_sacks--;
4359			continue;
4360		}
4361		this_sack++;
4362		sp++;
4363	}
4364	tp->rx_opt.num_sacks = num_sacks;
4365}
4366
4367/* This one checks to see if we can put data from the
4368 * out_of_order queue into the receive_queue.
4369 */
4370static void tcp_ofo_queue(struct sock *sk)
4371{
4372	struct tcp_sock *tp = tcp_sk(sk);
4373	__u32 dsack_high = tp->rcv_nxt;
4374	struct sk_buff *skb;
4375
4376	while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4377		if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4378			break;
4379
4380		if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4381			__u32 dsack = dsack_high;
4382			if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4383				dsack_high = TCP_SKB_CB(skb)->end_seq;
4384			tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4385		}
4386
4387		if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4388			SOCK_DEBUG(sk, "ofo packet was already received\n");
4389			__skb_unlink(skb, &tp->out_of_order_queue);
4390			__kfree_skb(skb);
4391			continue;
4392		}
4393		SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4394			   tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4395			   TCP_SKB_CB(skb)->end_seq);
4396
4397		__skb_unlink(skb, &tp->out_of_order_queue);
4398		__skb_queue_tail(&sk->sk_receive_queue, skb);
4399		tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4400		if (tcp_hdr(skb)->fin)
4401			tcp_fin(sk);
4402	}
4403}
4404
4405static int tcp_prune_ofo_queue(struct sock *sk);
4406static int tcp_prune_queue(struct sock *sk);
4407
4408static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4409{
4410	if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4411	    !sk_rmem_schedule(sk, size)) {
4412
4413		if (tcp_prune_queue(sk) < 0)
4414			return -1;
4415
4416		if (!sk_rmem_schedule(sk, size)) {
4417			if (!tcp_prune_ofo_queue(sk))
4418				return -1;
4419
4420			if (!sk_rmem_schedule(sk, size))
4421				return -1;
4422		}
4423	}
4424	return 0;
4425}
4426
4427static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4428{
4429	const struct tcphdr *th = tcp_hdr(skb);
4430	struct tcp_sock *tp = tcp_sk(sk);
4431	int eaten = -1;
4432
4433	if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4434		goto drop;
4435
4436	skb_dst_drop(skb);
4437	__skb_pull(skb, th->doff * 4);
4438
4439	TCP_ECN_accept_cwr(tp, skb);
4440
4441	tp->rx_opt.dsack = 0;
4442
4443	/*  Queue data for delivery to the user.
4444	 *  Packets in sequence go to the receive queue.
4445	 *  Out of sequence packets to the out_of_order_queue.
4446	 */
4447	if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4448		if (tcp_receive_window(tp) == 0)
4449			goto out_of_window;
4450
4451		/* Ok. In sequence. In window. */
4452		if (tp->ucopy.task == current &&
4453		    tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4454		    sock_owned_by_user(sk) && !tp->urg_data) {
4455			int chunk = min_t(unsigned int, skb->len,
4456					  tp->ucopy.len);
4457
4458			__set_current_state(TASK_RUNNING);
4459
4460			local_bh_enable();
4461			if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4462				tp->ucopy.len -= chunk;
4463				tp->copied_seq += chunk;
4464				eaten = (chunk == skb->len);
4465				tcp_rcv_space_adjust(sk);
4466			}
4467			local_bh_disable();
4468		}
4469
4470		if (eaten <= 0) {
4471queue_and_out:
4472			if (eaten < 0 &&
4473			    tcp_try_rmem_schedule(sk, skb->truesize))
4474				goto drop;
4475
4476			skb_set_owner_r(skb, sk);
4477			__skb_queue_tail(&sk->sk_receive_queue, skb);
4478		}
4479		tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4480		if (skb->len)
4481			tcp_event_data_recv(sk, skb);
4482		if (th->fin)
4483			tcp_fin(sk);
4484
4485		if (!skb_queue_empty(&tp->out_of_order_queue)) {
4486			tcp_ofo_queue(sk);
4487
4488			/* RFC2581. 4.2. SHOULD send immediate ACK, when
4489			 * gap in queue is filled.
4490			 */
4491			if (skb_queue_empty(&tp->out_of_order_queue))
4492				inet_csk(sk)->icsk_ack.pingpong = 0;
4493		}
4494
4495		if (tp->rx_opt.num_sacks)
4496			tcp_sack_remove(tp);
4497
4498		tcp_fast_path_check(sk);
4499
4500		if (eaten > 0)
4501			__kfree_skb(skb);
4502		else if (!sock_flag(sk, SOCK_DEAD))
4503			sk->sk_data_ready(sk, 0);
4504		return;
4505	}
4506
4507	if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4508		/* A retransmit, 2nd most common case.  Force an immediate ack. */
4509		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4510		tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4511
4512out_of_window:
4513		tcp_enter_quickack_mode(sk);
4514		inet_csk_schedule_ack(sk);
4515drop:
4516		__kfree_skb(skb);
4517		return;
4518	}
4519
4520	/* Out of window. F.e. zero window probe. */
4521	if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4522		goto out_of_window;
4523
4524	tcp_enter_quickack_mode(sk);
4525
4526	if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4527		/* Partial packet, seq < rcv_next < end_seq */
4528		SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4529			   tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4530			   TCP_SKB_CB(skb)->end_seq);
4531
4532		tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4533
4534		/* If window is closed, drop tail of packet. But after
4535		 * remembering D-SACK for its head made in previous line.
4536		 */
4537		if (!tcp_receive_window(tp))
4538			goto out_of_window;
4539		goto queue_and_out;
4540	}
4541
4542	TCP_ECN_check_ce(tp, skb);
4543
4544	if (tcp_try_rmem_schedule(sk, skb->truesize))
4545		goto drop;
4546
4547	/* Disable header prediction. */
4548	tp->pred_flags = 0;
4549	inet_csk_schedule_ack(sk);
4550
4551	SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4552		   tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4553
4554	skb_set_owner_r(skb, sk);
4555
4556	if (!skb_peek(&tp->out_of_order_queue)) {
4557		/* Initial out of order segment, build 1 SACK. */
4558		if (tcp_is_sack(tp)) {
4559			tp->rx_opt.num_sacks = 1;
4560			tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4561			tp->selective_acks[0].end_seq =
4562						TCP_SKB_CB(skb)->end_seq;
4563		}
4564		__skb_queue_head(&tp->out_of_order_queue, skb);
4565	} else {
4566		struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue);
4567		u32 seq = TCP_SKB_CB(skb)->seq;
4568		u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4569
4570		if (seq == TCP_SKB_CB(skb1)->end_seq) {
4571			__skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4572
4573			if (!tp->rx_opt.num_sacks ||
4574			    tp->selective_acks[0].end_seq != seq)
4575				goto add_sack;
4576
4577			/* Common case: data arrive in order after hole. */
4578			tp->selective_acks[0].end_seq = end_seq;
4579			return;
4580		}
4581
4582		/* Find place to insert this segment. */
4583		while (1) {
4584			if (!after(TCP_SKB_CB(skb1)->seq, seq))
4585				break;
4586			if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4587				skb1 = NULL;
4588				break;
4589			}
4590			skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4591		}
4592
4593		/* Do skb overlap to previous one? */
4594		if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4595			if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4596				/* All the bits are present. Drop. */
4597				__kfree_skb(skb);
4598				tcp_dsack_set(sk, seq, end_seq);
4599				goto add_sack;
4600			}
4601			if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4602				/* Partial overlap. */
4603				tcp_dsack_set(sk, seq,
4604					      TCP_SKB_CB(skb1)->end_seq);
4605			} else {
4606				if (skb_queue_is_first(&tp->out_of_order_queue,
4607						       skb1))
4608					skb1 = NULL;
4609				else
4610					skb1 = skb_queue_prev(
4611						&tp->out_of_order_queue,
4612						skb1);
4613			}
4614		}
4615		if (!skb1)
4616			__skb_queue_head(&tp->out_of_order_queue, skb);
4617		else
4618			__skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4619
4620		/* And clean segments covered by new one as whole. */
4621		while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4622			skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4623
4624			if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4625				break;
4626			if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4627				tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4628						 end_seq);
4629				break;
4630			}
4631			__skb_unlink(skb1, &tp->out_of_order_queue);
4632			tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4633					 TCP_SKB_CB(skb1)->end_seq);
4634			__kfree_skb(skb1);
4635		}
4636
4637add_sack:
4638		if (tcp_is_sack(tp))
4639			tcp_sack_new_ofo_skb(sk, seq, end_seq);
4640	}
4641}
4642
4643static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4644					struct sk_buff_head *list)
4645{
4646	struct sk_buff *next = NULL;
4647
4648	if (!skb_queue_is_last(list, skb))
4649		next = skb_queue_next(list, skb);
4650
4651	__skb_unlink(skb, list);
4652	__kfree_skb(skb);
4653	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4654
4655	return next;
4656}
4657
4658/* Collapse contiguous sequence of skbs head..tail with
4659 * sequence numbers start..end.
4660 *
4661 * If tail is NULL, this means until the end of the list.
4662 *
4663 * Segments with FIN/SYN are not collapsed (only because this
4664 * simplifies code)
4665 */
4666static void
4667tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4668	     struct sk_buff *head, struct sk_buff *tail,
4669	     u32 start, u32 end)
4670{
4671	struct sk_buff *skb, *n;
4672	bool end_of_skbs;
4673
4674	/* First, check that queue is collapsible and find
4675	 * the point where collapsing can be useful. */
4676	skb = head;
4677restart:
4678	end_of_skbs = true;
4679	skb_queue_walk_from_safe(list, skb, n) {
4680		if (skb == tail)
4681			break;
4682		/* No new bits? It is possible on ofo queue. */
4683		if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4684			skb = tcp_collapse_one(sk, skb, list);
4685			if (!skb)
4686				break;
4687			goto restart;
4688		}
4689
4690		/* The first skb to collapse is:
4691		 * - not SYN/FIN and
4692		 * - bloated or contains data before "start" or
4693		 *   overlaps to the next one.
4694		 */
4695		if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4696		    (tcp_win_from_space(skb->truesize) > skb->len ||
4697		     before(TCP_SKB_CB(skb)->seq, start))) {
4698			end_of_skbs = false;
4699			break;
4700		}
4701
4702		if (!skb_queue_is_last(list, skb)) {
4703			struct sk_buff *next = skb_queue_next(list, skb);
4704			if (next != tail &&
4705			    TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4706				end_of_skbs = false;
4707				break;
4708			}
4709		}
4710
4711		/* Decided to skip this, advance start seq. */
4712		start = TCP_SKB_CB(skb)->end_seq;
4713	}
4714	if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4715		return;
4716
4717	while (before(start, end)) {
4718		struct sk_buff *nskb;
4719		unsigned int header = skb_headroom(skb);
4720		int copy = SKB_MAX_ORDER(header, 0);
4721
4722		/* Too big header? This can happen with IPv6. */
4723		if (copy < 0)
4724			return;
4725		if (end - start < copy)
4726			copy = end - start;
4727		nskb = alloc_skb(copy + header, GFP_ATOMIC);
4728		if (!nskb)
4729			return;
4730
4731		skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4732		skb_set_network_header(nskb, (skb_network_header(skb) -
4733					      skb->head));
4734		skb_set_transport_header(nskb, (skb_transport_header(skb) -
4735						skb->head));
4736		skb_reserve(nskb, header);
4737		memcpy(nskb->head, skb->head, header);
4738		memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4739		TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4740		__skb_queue_before(list, skb, nskb);
4741		skb_set_owner_r(nskb, sk);
4742
4743		/* Copy data, releasing collapsed skbs. */
4744		while (copy > 0) {
4745			int offset = start - TCP_SKB_CB(skb)->seq;
4746			int size = TCP_SKB_CB(skb)->end_seq - start;
4747
4748			BUG_ON(offset < 0);
4749			if (size > 0) {
4750				size = min(copy, size);
4751				if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4752					BUG();
4753				TCP_SKB_CB(nskb)->end_seq += size;
4754				copy -= size;
4755				start += size;
4756			}
4757			if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4758				skb = tcp_collapse_one(sk, skb, list);
4759				if (!skb ||
4760				    skb == tail ||
4761				    tcp_hdr(skb)->syn ||
4762				    tcp_hdr(skb)->fin)
4763					return;
4764			}
4765		}
4766	}
4767}
4768
4769/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4770 * and tcp_collapse() them until all the queue is collapsed.
4771 */
4772static void tcp_collapse_ofo_queue(struct sock *sk)
4773{
4774	struct tcp_sock *tp = tcp_sk(sk);
4775	struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4776	struct sk_buff *head;
4777	u32 start, end;
4778
4779	if (skb == NULL)
4780		return;
4781
4782	start = TCP_SKB_CB(skb)->seq;
4783	end = TCP_SKB_CB(skb)->end_seq;
4784	head = skb;
4785
4786	for (;;) {
4787		struct sk_buff *next = NULL;
4788
4789		if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4790			next = skb_queue_next(&tp->out_of_order_queue, skb);
4791		skb = next;
4792
4793		/* Segment is terminated when we see gap or when
4794		 * we are at the end of all the queue. */
4795		if (!skb ||
4796		    after(TCP_SKB_CB(skb)->seq, end) ||
4797		    before(TCP_SKB_CB(skb)->end_seq, start)) {
4798			tcp_collapse(sk, &tp->out_of_order_queue,
4799				     head, skb, start, end);
4800			head = skb;
4801			if (!skb)
4802				break;
4803			/* Start new segment */
4804			start = TCP_SKB_CB(skb)->seq;
4805			end = TCP_SKB_CB(skb)->end_seq;
4806		} else {
4807			if (before(TCP_SKB_CB(skb)->seq, start))
4808				start = TCP_SKB_CB(skb)->seq;
4809			if (after(TCP_SKB_CB(skb)->end_seq, end))
4810				end = TCP_SKB_CB(skb)->end_seq;
4811		}
4812	}
4813}
4814
4815/*
4816 * Purge the out-of-order queue.
4817 * Return true if queue was pruned.
4818 */
4819static int tcp_prune_ofo_queue(struct sock *sk)
4820{
4821	struct tcp_sock *tp = tcp_sk(sk);
4822	int res = 0;
4823
4824	if (!skb_queue_empty(&tp->out_of_order_queue)) {
4825		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4826		__skb_queue_purge(&tp->out_of_order_queue);
4827
4828		/* Reset SACK state.  A conforming SACK implementation will
4829		 * do the same at a timeout based retransmit.  When a connection
4830		 * is in a sad state like this, we care only about integrity
4831		 * of the connection not performance.
4832		 */
4833		if (tp->rx_opt.sack_ok)
4834			tcp_sack_reset(&tp->rx_opt);
4835		sk_mem_reclaim(sk);
4836		res = 1;
4837	}
4838	return res;
4839}
4840
4841/* Reduce allocated memory if we can, trying to get
4842 * the socket within its memory limits again.
4843 *
4844 * Return less than zero if we should start dropping frames
4845 * until the socket owning process reads some of the data
4846 * to stabilize the situation.
4847 */
4848static int tcp_prune_queue(struct sock *sk)
4849{
4850	struct tcp_sock *tp = tcp_sk(sk);
4851
4852	SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4853
4854	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4855
4856	if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4857		tcp_clamp_window(sk);
4858	else if (sk_under_memory_pressure(sk))
4859		tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4860
4861	tcp_collapse_ofo_queue(sk);
4862	if (!skb_queue_empty(&sk->sk_receive_queue))
4863		tcp_collapse(sk, &sk->sk_receive_queue,
4864			     skb_peek(&sk->sk_receive_queue),
4865			     NULL,
4866			     tp->copied_seq, tp->rcv_nxt);
4867	sk_mem_reclaim(sk);
4868
4869	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4870		return 0;
4871
4872	/* Collapsing did not help, destructive actions follow.
4873	 * This must not ever occur. */
4874
4875	tcp_prune_ofo_queue(sk);
4876
4877	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4878		return 0;
4879
4880	/* If we are really being abused, tell the caller to silently
4881	 * drop receive data on the floor.  It will get retransmitted
4882	 * and hopefully then we'll have sufficient space.
4883	 */
4884	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4885
4886	/* Massive buffer overcommit. */
4887	tp->pred_flags = 0;
4888	return -1;
4889}
4890
4891/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4892 * As additional protections, we do not touch cwnd in retransmission phases,
4893 * and if application hit its sndbuf limit recently.
4894 */
4895void tcp_cwnd_application_limited(struct sock *sk)
4896{
4897	struct tcp_sock *tp = tcp_sk(sk);
4898
4899	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4900	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4901		/* Limited by application or receiver window. */
4902		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4903		u32 win_used = max(tp->snd_cwnd_used, init_win);
4904		if (win_used < tp->snd_cwnd) {
4905			tp->snd_ssthresh = tcp_current_ssthresh(sk);
4906			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4907		}
4908		tp->snd_cwnd_used = 0;
4909	}
4910	tp->snd_cwnd_stamp = tcp_time_stamp;
4911}
4912
4913static int tcp_should_expand_sndbuf(const struct sock *sk)
4914{
4915	const struct tcp_sock *tp = tcp_sk(sk);
4916
4917	/* If the user specified a specific send buffer setting, do
4918	 * not modify it.
4919	 */
4920	if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4921		return 0;
4922
4923	/* If we are under global TCP memory pressure, do not expand.  */
4924	if (sk_under_memory_pressure(sk))
4925		return 0;
4926
4927	/* If we are under soft global TCP memory pressure, do not expand.  */
4928	if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4929		return 0;
4930
4931	/* If we filled the congestion window, do not expand.  */
4932	if (tp->packets_out >= tp->snd_cwnd)
4933		return 0;
4934
4935	return 1;
4936}
4937
4938/* When incoming ACK allowed to free some skb from write_queue,
4939 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4940 * on the exit from tcp input handler.
4941 *
4942 * PROBLEM: sndbuf expansion does not work well with largesend.
4943 */
4944static void tcp_new_space(struct sock *sk)
4945{
4946	struct tcp_sock *tp = tcp_sk(sk);
4947
4948	if (tcp_should_expand_sndbuf(sk)) {
4949		int sndmem = SKB_TRUESIZE(max_t(u32,
4950						tp->rx_opt.mss_clamp,
4951						tp->mss_cache) +
4952					  MAX_TCP_HEADER);
4953		int demanded = max_t(unsigned int, tp->snd_cwnd,
4954				     tp->reordering + 1);
4955		sndmem *= 2 * demanded;
4956		if (sndmem > sk->sk_sndbuf)
4957			sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4958		tp->snd_cwnd_stamp = tcp_time_stamp;
4959	}
4960
4961	sk->sk_write_space(sk);
4962}
4963
4964static void tcp_check_space(struct sock *sk)
4965{
4966	if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4967		sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4968		if (sk->sk_socket &&
4969		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4970			tcp_new_space(sk);
4971	}
4972}
4973
4974static inline void tcp_data_snd_check(struct sock *sk)
4975{
4976	tcp_push_pending_frames(sk);
4977	tcp_check_space(sk);
4978}
4979
4980/*
4981 * Check if sending an ack is needed.
4982 */
4983static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4984{
4985	struct tcp_sock *tp = tcp_sk(sk);
4986
4987	    /* More than one full frame received... */
4988	if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4989	     /* ... and right edge of window advances far enough.
4990	      * (tcp_recvmsg() will send ACK otherwise). Or...
4991	      */
4992	     __tcp_select_window(sk) >= tp->rcv_wnd) ||
4993	    /* We ACK each frame or... */
4994	    tcp_in_quickack_mode(sk) ||
4995	    /* We have out of order data. */
4996	    (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4997		/* Then ack it now */
4998		tcp_send_ack(sk);
4999	} else {
5000		/* Else, send delayed ack. */
5001		tcp_send_delayed_ack(sk);
5002	}
5003}
5004
5005static inline void tcp_ack_snd_check(struct sock *sk)
5006{
5007	if (!inet_csk_ack_scheduled(sk)) {
5008		/* We sent a data segment already. */
5009		return;
5010	}
5011	__tcp_ack_snd_check(sk, 1);
5012}
5013
5014/*
5015 *	This routine is only called when we have urgent data
5016 *	signaled. Its the 'slow' part of tcp_urg. It could be
5017 *	moved inline now as tcp_urg is only called from one
5018 *	place. We handle URGent data wrong. We have to - as
5019 *	BSD still doesn't use the correction from RFC961.
5020 *	For 1003.1g we should support a new option TCP_STDURG to permit
5021 *	either form (or just set the sysctl tcp_stdurg).
5022 */
5023
5024static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5025{
5026	struct tcp_sock *tp = tcp_sk(sk);
5027	u32 ptr = ntohs(th->urg_ptr);
5028
5029	if (ptr && !sysctl_tcp_stdurg)
5030		ptr--;
5031	ptr += ntohl(th->seq);
5032
5033	/* Ignore urgent data that we've already seen and read. */
5034	if (after(tp->copied_seq, ptr))
5035		return;
5036
5037	/* Do not replay urg ptr.
5038	 *
5039	 * NOTE: interesting situation not covered by specs.
5040	 * Misbehaving sender may send urg ptr, pointing to segment,
5041	 * which we already have in ofo queue. We are not able to fetch
5042	 * such data and will stay in TCP_URG_NOTYET until will be eaten
5043	 * by recvmsg(). Seems, we are not obliged to handle such wicked
5044	 * situations. But it is worth to think about possibility of some
5045	 * DoSes using some hypothetical application level deadlock.
5046	 */
5047	if (before(ptr, tp->rcv_nxt))
5048		return;
5049
5050	/* Do we already have a newer (or duplicate) urgent pointer? */
5051	if (tp->urg_data && !after(ptr, tp->urg_seq))
5052		return;
5053
5054	/* Tell the world about our new urgent pointer. */
5055	sk_send_sigurg(sk);
5056
5057	/* We may be adding urgent data when the last byte read was
5058	 * urgent. To do this requires some care. We cannot just ignore
5059	 * tp->copied_seq since we would read the last urgent byte again
5060	 * as data, nor can we alter copied_seq until this data arrives
5061	 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5062	 *
5063	 * NOTE. Double Dutch. Rendering to plain English: author of comment
5064	 * above did something sort of 	send("A", MSG_OOB); send("B", MSG_OOB);
5065	 * and expect that both A and B disappear from stream. This is _wrong_.
5066	 * Though this happens in BSD with high probability, this is occasional.
5067	 * Any application relying on this is buggy. Note also, that fix "works"
5068	 * only in this artificial test. Insert some normal data between A and B and we will
5069	 * decline of BSD again. Verdict: it is better to remove to trap
5070	 * buggy users.
5071	 */
5072	if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5073	    !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5074		struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5075		tp->copied_seq++;
5076		if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5077			__skb_unlink(skb, &sk->sk_receive_queue);
5078			__kfree_skb(skb);
5079		}
5080	}
5081
5082	tp->urg_data = TCP_URG_NOTYET;
5083	tp->urg_seq = ptr;
5084
5085	/* Disable header prediction. */
5086	tp->pred_flags = 0;
5087}
5088
5089/* This is the 'fast' part of urgent handling. */
5090static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5091{
5092	struct tcp_sock *tp = tcp_sk(sk);
5093
5094	/* Check if we get a new urgent pointer - normally not. */
5095	if (th->urg)
5096		tcp_check_urg(sk, th);
5097
5098	/* Do we wait for any urgent data? - normally not... */
5099	if (tp->urg_data == TCP_URG_NOTYET) {
5100		u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5101			  th->syn;
5102
5103		/* Is the urgent pointer pointing into this packet? */
5104		if (ptr < skb->len) {
5105			u8 tmp;
5106			if (skb_copy_bits(skb, ptr, &tmp, 1))
5107				BUG();
5108			tp->urg_data = TCP_URG_VALID | tmp;
5109			if (!sock_flag(sk, SOCK_DEAD))
5110				sk->sk_data_ready(sk, 0);
5111		}
5112	}
5113}
5114
5115static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5116{
5117	struct tcp_sock *tp = tcp_sk(sk);
5118	int chunk = skb->len - hlen;
5119	int err;
5120
5121	local_bh_enable();
5122	if (skb_csum_unnecessary(skb))
5123		err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5124	else
5125		err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5126						       tp->ucopy.iov);
5127
5128	if (!err) {
5129		tp->ucopy.len -= chunk;
5130		tp->copied_seq += chunk;
5131		tcp_rcv_space_adjust(sk);
5132	}
5133
5134	local_bh_disable();
5135	return err;
5136}
5137
5138static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5139					    struct sk_buff *skb)
5140{
5141	__sum16 result;
5142
5143	if (sock_owned_by_user(sk)) {
5144		local_bh_enable();
5145		result = __tcp_checksum_complete(skb);
5146		local_bh_disable();
5147	} else {
5148		result = __tcp_checksum_complete(skb);
5149	}
5150	return result;
5151}
5152
5153static inline int tcp_checksum_complete_user(struct sock *sk,
5154					     struct sk_buff *skb)
5155{
5156	return !skb_csum_unnecessary(skb) &&
5157	       __tcp_checksum_complete_user(sk, skb);
5158}
5159
5160#ifdef CONFIG_NET_DMA
5161static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5162				  int hlen)
5163{
5164	struct tcp_sock *tp = tcp_sk(sk);
5165	int chunk = skb->len - hlen;
5166	int dma_cookie;
5167	int copied_early = 0;
5168
5169	if (tp->ucopy.wakeup)
5170		return 0;
5171
5172	if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5173		tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5174
5175	if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5176
5177		dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5178							 skb, hlen,
5179							 tp->ucopy.iov, chunk,
5180							 tp->ucopy.pinned_list);
5181
5182		if (dma_cookie < 0)
5183			goto out;
5184
5185		tp->ucopy.dma_cookie = dma_cookie;
5186		copied_early = 1;
5187
5188		tp->ucopy.len -= chunk;
5189		tp->copied_seq += chunk;
5190		tcp_rcv_space_adjust(sk);
5191
5192		if ((tp->ucopy.len == 0) ||
5193		    (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5194		    (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5195			tp->ucopy.wakeup = 1;
5196			sk->sk_data_ready(sk, 0);
5197		}
5198	} else if (chunk > 0) {
5199		tp->ucopy.wakeup = 1;
5200		sk->sk_data_ready(sk, 0);
5201	}
5202out:
5203	return copied_early;
5204}
5205#endif /* CONFIG_NET_DMA */
5206
5207/* Does PAWS and seqno based validation of an incoming segment, flags will
5208 * play significant role here.
5209 */
5210static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5211			      const struct tcphdr *th, int syn_inerr)
5212{
5213	const u8 *hash_location;
5214	struct tcp_sock *tp = tcp_sk(sk);
5215
5216	/* RFC1323: H1. Apply PAWS check first. */
5217	if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5218	    tp->rx_opt.saw_tstamp &&
5219	    tcp_paws_discard(sk, skb)) {
5220		if (!th->rst) {
5221			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5222			tcp_send_dupack(sk, skb);
5223			goto discard;
5224		}
5225		/* Reset is accepted even if it did not pass PAWS. */
5226	}
5227
5228	/* Step 1: check sequence number */
5229	if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5230		/* RFC793, page 37: "In all states except SYN-SENT, all reset
5231		 * (RST) segments are validated by checking their SEQ-fields."
5232		 * And page 69: "If an incoming segment is not acceptable,
5233		 * an acknowledgment should be sent in reply (unless the RST
5234		 * bit is set, if so drop the segment and return)".
5235		 */
5236		if (!th->rst)
5237			tcp_send_dupack(sk, skb);
5238		goto discard;
5239	}
5240
5241	/* Step 2: check RST bit */
5242	if (th->rst) {
5243		tcp_reset(sk);
5244		goto discard;
5245	}
5246
5247	/* ts_recent update must be made after we are sure that the packet
5248	 * is in window.
5249	 */
5250	tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5251
5252	/* step 3: check security and precedence [ignored] */
5253
5254	/* step 4: Check for a SYN in window. */
5255	if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5256		if (syn_inerr)
5257			TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5258		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5259		tcp_reset(sk);
5260		return -1;
5261	}
5262
5263	return 1;
5264
5265discard:
5266	__kfree_skb(skb);
5267	return 0;
5268}
5269
5270/*
5271 *	TCP receive function for the ESTABLISHED state.
5272 *
5273 *	It is split into a fast path and a slow path. The fast path is
5274 * 	disabled when:
5275 *	- A zero window was announced from us - zero window probing
5276 *        is only handled properly in the slow path.
5277 *	- Out of order segments arrived.
5278 *	- Urgent data is expected.
5279 *	- There is no buffer space left
5280 *	- Unexpected TCP flags/window values/header lengths are received
5281 *	  (detected by checking the TCP header against pred_flags)
5282 *	- Data is sent in both directions. Fast path only supports pure senders
5283 *	  or pure receivers (this means either the sequence number or the ack
5284 *	  value must stay constant)
5285 *	- Unexpected TCP option.
5286 *
5287 *	When these conditions are not satisfied it drops into a standard
5288 *	receive procedure patterned after RFC793 to handle all cases.
5289 *	The first three cases are guaranteed by proper pred_flags setting,
5290 *	the rest is checked inline. Fast processing is turned on in
5291 *	tcp_data_queue when everything is OK.
5292 */
5293int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5294			const struct tcphdr *th, unsigned int len)
5295{
5296	struct tcp_sock *tp = tcp_sk(sk);
5297	int res;
5298
5299	/*
5300	 *	Header prediction.
5301	 *	The code loosely follows the one in the famous
5302	 *	"30 instruction TCP receive" Van Jacobson mail.
5303	 *
5304	 *	Van's trick is to deposit buffers into socket queue
5305	 *	on a device interrupt, to call tcp_recv function
5306	 *	on the receive process context and checksum and copy
5307	 *	the buffer to user space. smart...
5308	 *
5309	 *	Our current scheme is not silly either but we take the
5310	 *	extra cost of the net_bh soft interrupt processing...
5311	 *	We do checksum and copy also but from device to kernel.
5312	 */
5313
5314	tp->rx_opt.saw_tstamp = 0;
5315
5316	/*	pred_flags is 0xS?10 << 16 + snd_wnd
5317	 *	if header_prediction is to be made
5318	 *	'S' will always be tp->tcp_header_len >> 2
5319	 *	'?' will be 0 for the fast path, otherwise pred_flags is 0 to
5320	 *  turn it off	(when there are holes in the receive
5321	 *	 space for instance)
5322	 *	PSH flag is ignored.
5323	 */
5324
5325	if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5326	    TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5327	    !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5328		int tcp_header_len = tp->tcp_header_len;
5329
5330		/* Timestamp header prediction: tcp_header_len
5331		 * is automatically equal to th->doff*4 due to pred_flags
5332		 * match.
5333		 */
5334
5335		/* Check timestamp */
5336		if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5337			/* No? Slow path! */
5338			if (!tcp_parse_aligned_timestamp(tp, th))
5339				goto slow_path;
5340
5341			/* If PAWS failed, check it more carefully in slow path */
5342			if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5343				goto slow_path;
5344
5345			/* DO NOT update ts_recent here, if checksum fails
5346			 * and timestamp was corrupted part, it will result
5347			 * in a hung connection since we will drop all
5348			 * future packets due to the PAWS test.
5349			 */
5350		}
5351
5352		if (len <= tcp_header_len) {
5353			/* Bulk data transfer: sender */
5354			if (len == tcp_header_len) {
5355				/* Predicted packet is in window by definition.
5356				 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5357				 * Hence, check seq<=rcv_wup reduces to:
5358				 */
5359				if (tcp_header_len ==
5360				    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5361				    tp->rcv_nxt == tp->rcv_wup)
5362					tcp_store_ts_recent(tp);
5363
5364				/* We know that such packets are checksummed
5365				 * on entry.
5366				 */
5367				tcp_ack(sk, skb, 0);
5368				__kfree_skb(skb);
5369				tcp_data_snd_check(sk);
5370				return 0;
5371			} else { /* Header too small */
5372				TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5373				goto discard;
5374			}
5375		} else {
5376			int eaten = 0;
5377			int copied_early = 0;
5378
5379			if (tp->copied_seq == tp->rcv_nxt &&
5380			    len - tcp_header_len <= tp->ucopy.len) {
5381#ifdef CONFIG_NET_DMA
5382				if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5383					copied_early = 1;
5384					eaten = 1;
5385				}
5386#endif
5387				if (tp->ucopy.task == current &&
5388				    sock_owned_by_user(sk) && !copied_early) {
5389					__set_current_state(TASK_RUNNING);
5390
5391					if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5392						eaten = 1;
5393				}
5394				if (eaten) {
5395					/* Predicted packet is in window by definition.
5396					 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5397					 * Hence, check seq<=rcv_wup reduces to:
5398					 */
5399					if (tcp_header_len ==
5400					    (sizeof(struct tcphdr) +
5401					     TCPOLEN_TSTAMP_ALIGNED) &&
5402					    tp->rcv_nxt == tp->rcv_wup)
5403						tcp_store_ts_recent(tp);
5404
5405					tcp_rcv_rtt_measure_ts(sk, skb);
5406
5407					__skb_pull(skb, tcp_header_len);
5408					tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5409					NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5410				}
5411				if (copied_early)
5412					tcp_cleanup_rbuf(sk, skb->len);
5413			}
5414			if (!eaten) {
5415				if (tcp_checksum_complete_user(sk, skb))
5416					goto csum_error;
5417
5418				/* Predicted packet is in window by definition.
5419				 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5420				 * Hence, check seq<=rcv_wup reduces to:
5421				 */
5422				if (tcp_header_len ==
5423				    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5424				    tp->rcv_nxt == tp->rcv_wup)
5425					tcp_store_ts_recent(tp);
5426
5427				tcp_rcv_rtt_measure_ts(sk, skb);
5428
5429				if ((int)skb->truesize > sk->sk_forward_alloc)
5430					goto step5;
5431
5432				NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5433
5434				/* Bulk data transfer: receiver */
5435				__skb_pull(skb, tcp_header_len);
5436				__skb_queue_tail(&sk->sk_receive_queue, skb);
5437				skb_set_owner_r(skb, sk);
5438				tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5439			}
5440
5441			tcp_event_data_recv(sk, skb);
5442
5443			if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5444				/* Well, only one small jumplet in fast path... */
5445				tcp_ack(sk, skb, FLAG_DATA);
5446				tcp_data_snd_check(sk);
5447				if (!inet_csk_ack_scheduled(sk))
5448					goto no_ack;
5449			}
5450
5451			if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5452				__tcp_ack_snd_check(sk, 0);
5453no_ack:
5454#ifdef CONFIG_NET_DMA
5455			if (copied_early)
5456				__skb_queue_tail(&sk->sk_async_wait_queue, skb);
5457			else
5458#endif
5459			if (eaten)
5460				__kfree_skb(skb);
5461			else
5462				sk->sk_data_ready(sk, 0);
5463			return 0;
5464		}
5465	}
5466
5467slow_path:
5468	if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5469		goto csum_error;
5470
5471	/*
5472	 *	Standard slow path.
5473	 */
5474
5475	res = tcp_validate_incoming(sk, skb, th, 1);
5476	if (res <= 0)
5477		return -res;
5478
5479step5:
5480	if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5481		goto discard;
5482
5483	tcp_rcv_rtt_measure_ts(sk, skb);
5484
5485	/* Process urgent data. */
5486	tcp_urg(sk, skb, th);
5487
5488	/* step 7: process the segment text */
5489	tcp_data_queue(sk, skb);
5490
5491	tcp_data_snd_check(sk);
5492	tcp_ack_snd_check(sk);
5493	return 0;
5494
5495csum_error:
5496	TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5497
5498discard:
5499	__kfree_skb(skb);
5500	return 0;
5501}
5502EXPORT_SYMBOL(tcp_rcv_established);
5503
5504static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5505					 const struct tcphdr *th, unsigned int len)
5506{
5507	const u8 *hash_location;
5508	struct inet_connection_sock *icsk = inet_csk(sk);
5509	struct tcp_sock *tp = tcp_sk(sk);
5510	struct tcp_cookie_values *cvp = tp->cookie_values;
5511	int saved_clamp = tp->rx_opt.mss_clamp;
5512
5513	tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5514
5515	if (th->ack) {
5516		/* rfc793:
5517		 * "If the state is SYN-SENT then
5518		 *    first check the ACK bit
5519		 *      If the ACK bit is set
5520		 *	  If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5521		 *        a reset (unless the RST bit is set, if so drop
5522		 *        the segment and return)"
5523		 *
5524		 *  We do not send data with SYN, so that RFC-correct
5525		 *  test reduces to:
5526		 */
5527		if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5528			goto reset_and_undo;
5529
5530		if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5531		    !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5532			     tcp_time_stamp)) {
5533			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5534			goto reset_and_undo;
5535		}
5536
5537		/* Now ACK is acceptable.
5538		 *
5539		 * "If the RST bit is set
5540		 *    If the ACK was acceptable then signal the user "error:
5541		 *    connection reset", drop the segment, enter CLOSED state,
5542		 *    delete TCB, and return."
5543		 */
5544
5545		if (th->rst) {
5546			tcp_reset(sk);
5547			goto discard;
5548		}
5549
5550		/* rfc793:
5551		 *   "fifth, if neither of the SYN or RST bits is set then
5552		 *    drop the segment and return."
5553		 *
5554		 *    See note below!
5555		 *                                        --ANK(990513)
5556		 */
5557		if (!th->syn)
5558			goto discard_and_undo;
5559
5560		/* rfc793:
5561		 *   "If the SYN bit is on ...
5562		 *    are acceptable then ...
5563		 *    (our SYN has been ACKed), change the connection
5564		 *    state to ESTABLISHED..."
5565		 */
5566
5567		TCP_ECN_rcv_synack(tp, th);
5568
5569		tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5570		tcp_ack(sk, skb, FLAG_SLOWPATH);
5571
5572		/* Ok.. it's good. Set up sequence numbers and
5573		 * move to established.
5574		 */
5575		tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5576		tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5577
5578		/* RFC1323: The window in SYN & SYN/ACK segments is
5579		 * never scaled.
5580		 */
5581		tp->snd_wnd = ntohs(th->window);
5582		tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5583
5584		if (!tp->rx_opt.wscale_ok) {
5585			tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5586			tp->window_clamp = min(tp->window_clamp, 65535U);
5587		}
5588
5589		if (tp->rx_opt.saw_tstamp) {
5590			tp->rx_opt.tstamp_ok	   = 1;
5591			tp->tcp_header_len =
5592				sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5593			tp->advmss	    -= TCPOLEN_TSTAMP_ALIGNED;
5594			tcp_store_ts_recent(tp);
5595		} else {
5596			tp->tcp_header_len = sizeof(struct tcphdr);
5597		}
5598
5599		if (tcp_is_sack(tp) && sysctl_tcp_fack)
5600			tcp_enable_fack(tp);
5601
5602		tcp_mtup_init(sk);
5603		tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5604		tcp_initialize_rcv_mss(sk);
5605
5606		/* Remember, tcp_poll() does not lock socket!
5607		 * Change state from SYN-SENT only after copied_seq
5608		 * is initialized. */
5609		tp->copied_seq = tp->rcv_nxt;
5610
5611		if (cvp != NULL &&
5612		    cvp->cookie_pair_size > 0 &&
5613		    tp->rx_opt.cookie_plus > 0) {
5614			int cookie_size = tp->rx_opt.cookie_plus
5615					- TCPOLEN_COOKIE_BASE;
5616			int cookie_pair_size = cookie_size
5617					     + cvp->cookie_desired;
5618
5619			/* A cookie extension option was sent and returned.
5620			 * Note that each incoming SYNACK replaces the
5621			 * Responder cookie.  The initial exchange is most
5622			 * fragile, as protection against spoofing relies
5623			 * entirely upon the sequence and timestamp (above).
5624			 * This replacement strategy allows the correct pair to
5625			 * pass through, while any others will be filtered via
5626			 * Responder verification later.
5627			 */
5628			if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5629				memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5630				       hash_location, cookie_size);
5631				cvp->cookie_pair_size = cookie_pair_size;
5632			}
5633		}
5634
5635		smp_mb();
5636		tcp_set_state(sk, TCP_ESTABLISHED);
5637
5638		security_inet_conn_established(sk, skb);
5639
5640		/* Make sure socket is routed, for correct metrics.  */
5641		icsk->icsk_af_ops->rebuild_header(sk);
5642
5643		tcp_init_metrics(sk);
5644
5645		tcp_init_congestion_control(sk);
5646
5647		/* Prevent spurious tcp_cwnd_restart() on first data
5648		 * packet.
5649		 */
5650		tp->lsndtime = tcp_time_stamp;
5651
5652		tcp_init_buffer_space(sk);
5653
5654		if (sock_flag(sk, SOCK_KEEPOPEN))
5655			inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5656
5657		if (!tp->rx_opt.snd_wscale)
5658			__tcp_fast_path_on(tp, tp->snd_wnd);
5659		else
5660			tp->pred_flags = 0;
5661
5662		if (!sock_flag(sk, SOCK_DEAD)) {
5663			sk->sk_state_change(sk);
5664			sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5665		}
5666
5667		if (sk->sk_write_pending ||
5668		    icsk->icsk_accept_queue.rskq_defer_accept ||
5669		    icsk->icsk_ack.pingpong) {
5670			/* Save one ACK. Data will be ready after
5671			 * several ticks, if write_pending is set.
5672			 *
5673			 * It may be deleted, but with this feature tcpdumps
5674			 * look so _wonderfully_ clever, that I was not able
5675			 * to stand against the temptation 8)     --ANK
5676			 */
5677			inet_csk_schedule_ack(sk);
5678			icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5679			icsk->icsk_ack.ato	 = TCP_ATO_MIN;
5680			tcp_incr_quickack(sk);
5681			tcp_enter_quickack_mode(sk);
5682			inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5683						  TCP_DELACK_MAX, TCP_RTO_MAX);
5684
5685discard:
5686			__kfree_skb(skb);
5687			return 0;
5688		} else {
5689			tcp_send_ack(sk);
5690		}
5691		return -1;
5692	}
5693
5694	/* No ACK in the segment */
5695
5696	if (th->rst) {
5697		/* rfc793:
5698		 * "If the RST bit is set
5699		 *
5700		 *      Otherwise (no ACK) drop the segment and return."
5701		 */
5702
5703		goto discard_and_undo;
5704	}
5705
5706	/* PAWS check. */
5707	if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5708	    tcp_paws_reject(&tp->rx_opt, 0))
5709		goto discard_and_undo;
5710
5711	if (th->syn) {
5712		/* We see SYN without ACK. It is attempt of
5713		 * simultaneous connect with crossed SYNs.
5714		 * Particularly, it can be connect to self.
5715		 */
5716		tcp_set_state(sk, TCP_SYN_RECV);
5717
5718		if (tp->rx_opt.saw_tstamp) {
5719			tp->rx_opt.tstamp_ok = 1;
5720			tcp_store_ts_recent(tp);
5721			tp->tcp_header_len =
5722				sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5723		} else {
5724			tp->tcp_header_len = sizeof(struct tcphdr);
5725		}
5726
5727		tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5728		tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5729
5730		/* RFC1323: The window in SYN & SYN/ACK segments is
5731		 * never scaled.
5732		 */
5733		tp->snd_wnd    = ntohs(th->window);
5734		tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5735		tp->max_window = tp->snd_wnd;
5736
5737		TCP_ECN_rcv_syn(tp, th);
5738
5739		tcp_mtup_init(sk);
5740		tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5741		tcp_initialize_rcv_mss(sk);
5742
5743		tcp_send_synack(sk);
5744#if 0
5745		/* Note, we could accept data and URG from this segment.
5746		 * There are no obstacles to make this.
5747		 *
5748		 * However, if we ignore data in ACKless segments sometimes,
5749		 * we have no reasons to accept it sometimes.
5750		 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5751		 * is not flawless. So, discard packet for sanity.
5752		 * Uncomment this return to process the data.
5753		 */
5754		return -1;
5755#else
5756		goto discard;
5757#endif
5758	}
5759	/* "fifth, if neither of the SYN or RST bits is set then
5760	 * drop the segment and return."
5761	 */
5762
5763discard_and_undo:
5764	tcp_clear_options(&tp->rx_opt);
5765	tp->rx_opt.mss_clamp = saved_clamp;
5766	goto discard;
5767
5768reset_and_undo:
5769	tcp_clear_options(&tp->rx_opt);
5770	tp->rx_opt.mss_clamp = saved_clamp;
5771	return 1;
5772}
5773
5774/*
5775 *	This function implements the receiving procedure of RFC 793 for
5776 *	all states except ESTABLISHED and TIME_WAIT.
5777 *	It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5778 *	address independent.
5779 */
5780
5781int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5782			  const struct tcphdr *th, unsigned int len)
5783{
5784	struct tcp_sock *tp = tcp_sk(sk);
5785	struct inet_connection_sock *icsk = inet_csk(sk);
5786	int queued = 0;
5787	int res;
5788
5789	tp->rx_opt.saw_tstamp = 0;
5790
5791	switch (sk->sk_state) {
5792	case TCP_CLOSE:
5793		goto discard;
5794
5795	case TCP_LISTEN:
5796		if (th->ack)
5797			return 1;
5798
5799		if (th->rst)
5800			goto discard;
5801
5802		if (th->syn) {
5803			if (th->fin)
5804				goto discard;
5805			if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5806				return 1;
5807
5808			/* Now we have several options: In theory there is
5809			 * nothing else in the frame. KA9Q has an option to
5810			 * send data with the syn, BSD accepts data with the
5811			 * syn up to the [to be] advertised window and
5812			 * Solaris 2.1 gives you a protocol error. For now
5813			 * we just ignore it, that fits the spec precisely
5814			 * and avoids incompatibilities. It would be nice in
5815			 * future to drop through and process the data.
5816			 *
5817			 * Now that TTCP is starting to be used we ought to
5818			 * queue this data.
5819			 * But, this leaves one open to an easy denial of
5820			 * service attack, and SYN cookies can't defend
5821			 * against this problem. So, we drop the data
5822			 * in the interest of security over speed unless
5823			 * it's still in use.
5824			 */
5825			kfree_skb(skb);
5826			return 0;
5827		}
5828		goto discard;
5829
5830	case TCP_SYN_SENT:
5831		queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5832		if (queued >= 0)
5833			return queued;
5834
5835		/* Do step6 onward by hand. */
5836		tcp_urg(sk, skb, th);
5837		__kfree_skb(skb);
5838		tcp_data_snd_check(sk);
5839		return 0;
5840	}
5841
5842	res = tcp_validate_incoming(sk, skb, th, 0);
5843	if (res <= 0)
5844		return -res;
5845
5846	/* step 5: check the ACK field */
5847	if (th->ack) {
5848		int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5849
5850		switch (sk->sk_state) {
5851		case TCP_SYN_RECV:
5852			if (acceptable) {
5853				tp->copied_seq = tp->rcv_nxt;
5854				smp_mb();
5855				tcp_set_state(sk, TCP_ESTABLISHED);
5856				sk->sk_state_change(sk);
5857
5858				/* Note, that this wakeup is only for marginal
5859				 * crossed SYN case. Passively open sockets
5860				 * are not waked up, because sk->sk_sleep ==
5861				 * NULL and sk->sk_socket == NULL.
5862				 */
5863				if (sk->sk_socket)
5864					sk_wake_async(sk,
5865						      SOCK_WAKE_IO, POLL_OUT);
5866
5867				tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5868				tp->snd_wnd = ntohs(th->window) <<
5869					      tp->rx_opt.snd_wscale;
5870				tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5871
5872				if (tp->rx_opt.tstamp_ok)
5873					tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5874
5875				/* Make sure socket is routed, for
5876				 * correct metrics.
5877				 */
5878				icsk->icsk_af_ops->rebuild_header(sk);
5879
5880				tcp_init_metrics(sk);
5881
5882				tcp_init_congestion_control(sk);
5883
5884				/* Prevent spurious tcp_cwnd_restart() on
5885				 * first data packet.
5886				 */
5887				tp->lsndtime = tcp_time_stamp;
5888
5889				tcp_mtup_init(sk);
5890				tcp_initialize_rcv_mss(sk);
5891				tcp_init_buffer_space(sk);
5892				tcp_fast_path_on(tp);
5893			} else {
5894				return 1;
5895			}
5896			break;
5897
5898		case TCP_FIN_WAIT1:
5899			if (tp->snd_una == tp->write_seq) {
5900				tcp_set_state(sk, TCP_FIN_WAIT2);
5901				sk->sk_shutdown |= SEND_SHUTDOWN;
5902				dst_confirm(__sk_dst_get(sk));
5903
5904				if (!sock_flag(sk, SOCK_DEAD))
5905					/* Wake up lingering close() */
5906					sk->sk_state_change(sk);
5907				else {
5908					int tmo;
5909
5910					if (tp->linger2 < 0 ||
5911					    (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5912					     after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5913						tcp_done(sk);
5914						NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5915						return 1;
5916					}
5917
5918					tmo = tcp_fin_time(sk);
5919					if (tmo > TCP_TIMEWAIT_LEN) {
5920						inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5921					} else if (th->fin || sock_owned_by_user(sk)) {
5922						/* Bad case. We could lose such FIN otherwise.
5923						 * It is not a big problem, but it looks confusing
5924						 * and not so rare event. We still can lose it now,
5925						 * if it spins in bh_lock_sock(), but it is really
5926						 * marginal case.
5927						 */
5928						inet_csk_reset_keepalive_timer(sk, tmo);
5929					} else {
5930						tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5931						goto discard;
5932					}
5933				}
5934			}
5935			break;
5936
5937		case TCP_CLOSING:
5938			if (tp->snd_una == tp->write_seq) {
5939				tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5940				goto discard;
5941			}
5942			break;
5943
5944		case TCP_LAST_ACK:
5945			if (tp->snd_una == tp->write_seq) {
5946				tcp_update_metrics(sk);
5947				tcp_done(sk);
5948				goto discard;
5949			}
5950			break;
5951		}
5952	} else
5953		goto discard;
5954
5955	/* step 6: check the URG bit */
5956	tcp_urg(sk, skb, th);
5957
5958	/* step 7: process the segment text */
5959	switch (sk->sk_state) {
5960	case TCP_CLOSE_WAIT:
5961	case TCP_CLOSING:
5962	case TCP_LAST_ACK:
5963		if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5964			break;
5965	case TCP_FIN_WAIT1:
5966	case TCP_FIN_WAIT2:
5967		/* RFC 793 says to queue data in these states,
5968		 * RFC 1122 says we MUST send a reset.
5969		 * BSD 4.4 also does reset.
5970		 */
5971		if (sk->sk_shutdown & RCV_SHUTDOWN) {
5972			if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5973			    after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5974				NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5975				tcp_reset(sk);
5976				return 1;
5977			}
5978		}
5979		/* Fall through */
5980	case TCP_ESTABLISHED:
5981		tcp_data_queue(sk, skb);
5982		queued = 1;
5983		break;
5984	}
5985
5986	/* tcp_data could move socket to TIME-WAIT */
5987	if (sk->sk_state != TCP_CLOSE) {
5988		tcp_data_snd_check(sk);
5989		tcp_ack_snd_check(sk);
5990	}
5991
5992	if (!queued) {
5993discard:
5994		__kfree_skb(skb);
5995	}
5996	return 0;
5997}
5998EXPORT_SYMBOL(tcp_rcv_state_process);
5999