xref: /linux/net/ipv4/tcp_input.c (revision f2ee442115c9b6219083c019939a9cc0c9abb2f8)
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 
76 int sysctl_tcp_timestamps __read_mostly = 1;
77 int sysctl_tcp_window_scaling __read_mostly = 1;
78 int sysctl_tcp_sack __read_mostly = 1;
79 int sysctl_tcp_fack __read_mostly = 1;
80 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
81 EXPORT_SYMBOL(sysctl_tcp_reordering);
82 int sysctl_tcp_ecn __read_mostly = 2;
83 EXPORT_SYMBOL(sysctl_tcp_ecn);
84 int sysctl_tcp_dsack __read_mostly = 1;
85 int sysctl_tcp_app_win __read_mostly = 31;
86 int sysctl_tcp_adv_win_scale __read_mostly = 2;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
88 
89 int sysctl_tcp_stdurg __read_mostly;
90 int sysctl_tcp_rfc1337 __read_mostly;
91 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
92 int sysctl_tcp_frto __read_mostly = 2;
93 int sysctl_tcp_frto_response __read_mostly;
94 int sysctl_tcp_nometrics_save __read_mostly;
95 
96 int sysctl_tcp_thin_dupack __read_mostly;
97 
98 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
99 int 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_DATA_LOST		0x80 /* SACK detected data lossage.		*/
109 #define FLAG_SLOWPATH		0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ONLY_ORIG_SACKED	0x200 /* SACKs only non-rexmit sent before RTO */
111 #define FLAG_SND_UNA_ADVANCED	0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK	0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_NONHEAD_RETRANS_ACKED	0x1000 /* Non-head rexmitted data was ACKed */
114 #define FLAG_SACK_RENEGING	0x2000 /* snd_una advanced to a sacked seq */
115 
116 #define FLAG_ACKED		(FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP		(FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT		(FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS	(FLAG_ACKED|FLAG_DATA_SACKED)
120 #define FLAG_ANY_PROGRESS	(FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
121 
122 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
123 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124 
125 /* Adapt the MSS value used to make delayed ack decision to the
126  * real world.
127  */
128 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
129 {
130 	struct inet_connection_sock *icsk = inet_csk(sk);
131 	const unsigned int lss = icsk->icsk_ack.last_seg_size;
132 	unsigned int len;
133 
134 	icsk->icsk_ack.last_seg_size = 0;
135 
136 	/* skb->len may jitter because of SACKs, even if peer
137 	 * sends good full-sized frames.
138 	 */
139 	len = skb_shinfo(skb)->gso_size ? : skb->len;
140 	if (len >= icsk->icsk_ack.rcv_mss) {
141 		icsk->icsk_ack.rcv_mss = len;
142 	} else {
143 		/* Otherwise, we make more careful check taking into account,
144 		 * that SACKs block is variable.
145 		 *
146 		 * "len" is invariant segment length, including TCP header.
147 		 */
148 		len += skb->data - skb_transport_header(skb);
149 		if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
150 		    /* If PSH is not set, packet should be
151 		     * full sized, provided peer TCP is not badly broken.
152 		     * This observation (if it is correct 8)) allows
153 		     * to handle super-low mtu links fairly.
154 		     */
155 		    (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
156 		     !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
157 			/* Subtract also invariant (if peer is RFC compliant),
158 			 * tcp header plus fixed timestamp option length.
159 			 * Resulting "len" is MSS free of SACK jitter.
160 			 */
161 			len -= tcp_sk(sk)->tcp_header_len;
162 			icsk->icsk_ack.last_seg_size = len;
163 			if (len == lss) {
164 				icsk->icsk_ack.rcv_mss = len;
165 				return;
166 			}
167 		}
168 		if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
169 			icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
170 		icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
171 	}
172 }
173 
174 static void tcp_incr_quickack(struct sock *sk)
175 {
176 	struct inet_connection_sock *icsk = inet_csk(sk);
177 	unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
178 
179 	if (quickacks == 0)
180 		quickacks = 2;
181 	if (quickacks > icsk->icsk_ack.quick)
182 		icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
183 }
184 
185 static void tcp_enter_quickack_mode(struct sock *sk)
186 {
187 	struct inet_connection_sock *icsk = inet_csk(sk);
188 	tcp_incr_quickack(sk);
189 	icsk->icsk_ack.pingpong = 0;
190 	icsk->icsk_ack.ato = TCP_ATO_MIN;
191 }
192 
193 /* Send ACKs quickly, if "quick" count is not exhausted
194  * and the session is not interactive.
195  */
196 
197 static inline int tcp_in_quickack_mode(const struct sock *sk)
198 {
199 	const struct inet_connection_sock *icsk = inet_csk(sk);
200 	return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
201 }
202 
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
204 {
205 	if (tp->ecn_flags & TCP_ECN_OK)
206 		tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
207 }
208 
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
210 {
211 	if (tcp_hdr(skb)->cwr)
212 		tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
213 }
214 
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
216 {
217 	tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
218 }
219 
220 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
221 {
222 	if (!(tp->ecn_flags & TCP_ECN_OK))
223 		return;
224 
225 	switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
226 	case INET_ECN_NOT_ECT:
227 		/* Funny extension: if ECT is not set on a segment,
228 		 * and we already seen ECT on a previous segment,
229 		 * it is probably a retransmit.
230 		 */
231 		if (tp->ecn_flags & TCP_ECN_SEEN)
232 			tcp_enter_quickack_mode((struct sock *)tp);
233 		break;
234 	case INET_ECN_CE:
235 		tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
236 		/* fallinto */
237 	default:
238 		tp->ecn_flags |= TCP_ECN_SEEN;
239 	}
240 }
241 
242 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
243 {
244 	if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
245 		tp->ecn_flags &= ~TCP_ECN_OK;
246 }
247 
248 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
249 {
250 	if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
251 		tp->ecn_flags &= ~TCP_ECN_OK;
252 }
253 
254 static inline int TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
255 {
256 	if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
257 		return 1;
258 	return 0;
259 }
260 
261 /* Buffer size and advertised window tuning.
262  *
263  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
264  */
265 
266 static void tcp_fixup_sndbuf(struct sock *sk)
267 {
268 	int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
269 
270 	sndmem *= TCP_INIT_CWND;
271 	if (sk->sk_sndbuf < sndmem)
272 		sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
273 }
274 
275 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
276  *
277  * All tcp_full_space() is split to two parts: "network" buffer, allocated
278  * forward and advertised in receiver window (tp->rcv_wnd) and
279  * "application buffer", required to isolate scheduling/application
280  * latencies from network.
281  * window_clamp is maximal advertised window. It can be less than
282  * tcp_full_space(), in this case tcp_full_space() - window_clamp
283  * is reserved for "application" buffer. The less window_clamp is
284  * the smoother our behaviour from viewpoint of network, but the lower
285  * throughput and the higher sensitivity of the connection to losses. 8)
286  *
287  * rcv_ssthresh is more strict window_clamp used at "slow start"
288  * phase to predict further behaviour of this connection.
289  * It is used for two goals:
290  * - to enforce header prediction at sender, even when application
291  *   requires some significant "application buffer". It is check #1.
292  * - to prevent pruning of receive queue because of misprediction
293  *   of receiver window. Check #2.
294  *
295  * The scheme does not work when sender sends good segments opening
296  * window and then starts to feed us spaghetti. But it should work
297  * in common situations. Otherwise, we have to rely on queue collapsing.
298  */
299 
300 /* Slow part of check#2. */
301 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
302 {
303 	struct tcp_sock *tp = tcp_sk(sk);
304 	/* Optimize this! */
305 	int truesize = tcp_win_from_space(skb->truesize) >> 1;
306 	int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
307 
308 	while (tp->rcv_ssthresh <= window) {
309 		if (truesize <= skb->len)
310 			return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
311 
312 		truesize >>= 1;
313 		window >>= 1;
314 	}
315 	return 0;
316 }
317 
318 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
319 {
320 	struct tcp_sock *tp = tcp_sk(sk);
321 
322 	/* Check #1 */
323 	if (tp->rcv_ssthresh < tp->window_clamp &&
324 	    (int)tp->rcv_ssthresh < tcp_space(sk) &&
325 	    !tcp_memory_pressure) {
326 		int incr;
327 
328 		/* Check #2. Increase window, if skb with such overhead
329 		 * will fit to rcvbuf in future.
330 		 */
331 		if (tcp_win_from_space(skb->truesize) <= skb->len)
332 			incr = 2 * tp->advmss;
333 		else
334 			incr = __tcp_grow_window(sk, skb);
335 
336 		if (incr) {
337 			tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
338 					       tp->window_clamp);
339 			inet_csk(sk)->icsk_ack.quick |= 1;
340 		}
341 	}
342 }
343 
344 /* 3. Tuning rcvbuf, when connection enters established state. */
345 
346 static void tcp_fixup_rcvbuf(struct sock *sk)
347 {
348 	u32 mss = tcp_sk(sk)->advmss;
349 	u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
350 	int rcvmem;
351 
352 	/* Limit to 10 segments if mss <= 1460,
353 	 * or 14600/mss segments, with a minimum of two segments.
354 	 */
355 	if (mss > 1460)
356 		icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
357 
358 	rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
359 	while (tcp_win_from_space(rcvmem) < mss)
360 		rcvmem += 128;
361 
362 	rcvmem *= icwnd;
363 
364 	if (sk->sk_rcvbuf < rcvmem)
365 		sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
366 }
367 
368 /* 4. Try to fixup all. It is made immediately after connection enters
369  *    established state.
370  */
371 static void tcp_init_buffer_space(struct sock *sk)
372 {
373 	struct tcp_sock *tp = tcp_sk(sk);
374 	int maxwin;
375 
376 	if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
377 		tcp_fixup_rcvbuf(sk);
378 	if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
379 		tcp_fixup_sndbuf(sk);
380 
381 	tp->rcvq_space.space = tp->rcv_wnd;
382 
383 	maxwin = tcp_full_space(sk);
384 
385 	if (tp->window_clamp >= maxwin) {
386 		tp->window_clamp = maxwin;
387 
388 		if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
389 			tp->window_clamp = max(maxwin -
390 					       (maxwin >> sysctl_tcp_app_win),
391 					       4 * tp->advmss);
392 	}
393 
394 	/* Force reservation of one segment. */
395 	if (sysctl_tcp_app_win &&
396 	    tp->window_clamp > 2 * tp->advmss &&
397 	    tp->window_clamp + tp->advmss > maxwin)
398 		tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
399 
400 	tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
401 	tp->snd_cwnd_stamp = tcp_time_stamp;
402 }
403 
404 /* 5. Recalculate window clamp after socket hit its memory bounds. */
405 static void tcp_clamp_window(struct sock *sk)
406 {
407 	struct tcp_sock *tp = tcp_sk(sk);
408 	struct inet_connection_sock *icsk = inet_csk(sk);
409 
410 	icsk->icsk_ack.quick = 0;
411 
412 	if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
413 	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
414 	    !tcp_memory_pressure &&
415 	    atomic_long_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
416 		sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
417 				    sysctl_tcp_rmem[2]);
418 	}
419 	if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
420 		tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
421 }
422 
423 /* Initialize RCV_MSS value.
424  * RCV_MSS is an our guess about MSS used by the peer.
425  * We haven't any direct information about the MSS.
426  * It's better to underestimate the RCV_MSS rather than overestimate.
427  * Overestimations make us ACKing less frequently than needed.
428  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
429  */
430 void tcp_initialize_rcv_mss(struct sock *sk)
431 {
432 	const struct tcp_sock *tp = tcp_sk(sk);
433 	unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
434 
435 	hint = min(hint, tp->rcv_wnd / 2);
436 	hint = min(hint, TCP_MSS_DEFAULT);
437 	hint = max(hint, TCP_MIN_MSS);
438 
439 	inet_csk(sk)->icsk_ack.rcv_mss = hint;
440 }
441 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
442 
443 /* Receiver "autotuning" code.
444  *
445  * The algorithm for RTT estimation w/o timestamps is based on
446  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
447  * <http://public.lanl.gov/radiant/pubs.html#DRS>
448  *
449  * More detail on this code can be found at
450  * <http://staff.psc.edu/jheffner/>,
451  * though this reference is out of date.  A new paper
452  * is pending.
453  */
454 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
455 {
456 	u32 new_sample = tp->rcv_rtt_est.rtt;
457 	long m = sample;
458 
459 	if (m == 0)
460 		m = 1;
461 
462 	if (new_sample != 0) {
463 		/* If we sample in larger samples in the non-timestamp
464 		 * case, we could grossly overestimate the RTT especially
465 		 * with chatty applications or bulk transfer apps which
466 		 * are stalled on filesystem I/O.
467 		 *
468 		 * Also, since we are only going for a minimum in the
469 		 * non-timestamp case, we do not smooth things out
470 		 * else with timestamps disabled convergence takes too
471 		 * long.
472 		 */
473 		if (!win_dep) {
474 			m -= (new_sample >> 3);
475 			new_sample += m;
476 		} else if (m < new_sample)
477 			new_sample = m << 3;
478 	} else {
479 		/* No previous measure. */
480 		new_sample = m << 3;
481 	}
482 
483 	if (tp->rcv_rtt_est.rtt != new_sample)
484 		tp->rcv_rtt_est.rtt = new_sample;
485 }
486 
487 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
488 {
489 	if (tp->rcv_rtt_est.time == 0)
490 		goto new_measure;
491 	if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
492 		return;
493 	tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
494 
495 new_measure:
496 	tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
497 	tp->rcv_rtt_est.time = tcp_time_stamp;
498 }
499 
500 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
501 					  const struct sk_buff *skb)
502 {
503 	struct tcp_sock *tp = tcp_sk(sk);
504 	if (tp->rx_opt.rcv_tsecr &&
505 	    (TCP_SKB_CB(skb)->end_seq -
506 	     TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
507 		tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
508 }
509 
510 /*
511  * This function should be called every time data is copied to user space.
512  * It calculates the appropriate TCP receive buffer space.
513  */
514 void tcp_rcv_space_adjust(struct sock *sk)
515 {
516 	struct tcp_sock *tp = tcp_sk(sk);
517 	int time;
518 	int space;
519 
520 	if (tp->rcvq_space.time == 0)
521 		goto new_measure;
522 
523 	time = tcp_time_stamp - tp->rcvq_space.time;
524 	if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
525 		return;
526 
527 	space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
528 
529 	space = max(tp->rcvq_space.space, space);
530 
531 	if (tp->rcvq_space.space != space) {
532 		int rcvmem;
533 
534 		tp->rcvq_space.space = space;
535 
536 		if (sysctl_tcp_moderate_rcvbuf &&
537 		    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
538 			int new_clamp = space;
539 
540 			/* Receive space grows, normalize in order to
541 			 * take into account packet headers and sk_buff
542 			 * structure overhead.
543 			 */
544 			space /= tp->advmss;
545 			if (!space)
546 				space = 1;
547 			rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
548 			while (tcp_win_from_space(rcvmem) < tp->advmss)
549 				rcvmem += 128;
550 			space *= rcvmem;
551 			space = min(space, sysctl_tcp_rmem[2]);
552 			if (space > sk->sk_rcvbuf) {
553 				sk->sk_rcvbuf = space;
554 
555 				/* Make the window clamp follow along.  */
556 				tp->window_clamp = new_clamp;
557 			}
558 		}
559 	}
560 
561 new_measure:
562 	tp->rcvq_space.seq = tp->copied_seq;
563 	tp->rcvq_space.time = tcp_time_stamp;
564 }
565 
566 /* There is something which you must keep in mind when you analyze the
567  * behavior of the tp->ato delayed ack timeout interval.  When a
568  * connection starts up, we want to ack as quickly as possible.  The
569  * problem is that "good" TCP's do slow start at the beginning of data
570  * transmission.  The means that until we send the first few ACK's the
571  * sender will sit on his end and only queue most of his data, because
572  * he can only send snd_cwnd unacked packets at any given time.  For
573  * each ACK we send, he increments snd_cwnd and transmits more of his
574  * queue.  -DaveM
575  */
576 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
577 {
578 	struct tcp_sock *tp = tcp_sk(sk);
579 	struct inet_connection_sock *icsk = inet_csk(sk);
580 	u32 now;
581 
582 	inet_csk_schedule_ack(sk);
583 
584 	tcp_measure_rcv_mss(sk, skb);
585 
586 	tcp_rcv_rtt_measure(tp);
587 
588 	now = tcp_time_stamp;
589 
590 	if (!icsk->icsk_ack.ato) {
591 		/* The _first_ data packet received, initialize
592 		 * delayed ACK engine.
593 		 */
594 		tcp_incr_quickack(sk);
595 		icsk->icsk_ack.ato = TCP_ATO_MIN;
596 	} else {
597 		int m = now - icsk->icsk_ack.lrcvtime;
598 
599 		if (m <= TCP_ATO_MIN / 2) {
600 			/* The fastest case is the first. */
601 			icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
602 		} else if (m < icsk->icsk_ack.ato) {
603 			icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
604 			if (icsk->icsk_ack.ato > icsk->icsk_rto)
605 				icsk->icsk_ack.ato = icsk->icsk_rto;
606 		} else if (m > icsk->icsk_rto) {
607 			/* Too long gap. Apparently sender failed to
608 			 * restart window, so that we send ACKs quickly.
609 			 */
610 			tcp_incr_quickack(sk);
611 			sk_mem_reclaim(sk);
612 		}
613 	}
614 	icsk->icsk_ack.lrcvtime = now;
615 
616 	TCP_ECN_check_ce(tp, skb);
617 
618 	if (skb->len >= 128)
619 		tcp_grow_window(sk, skb);
620 }
621 
622 /* Called to compute a smoothed rtt estimate. The data fed to this
623  * routine either comes from timestamps, or from segments that were
624  * known _not_ to have been retransmitted [see Karn/Partridge
625  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
626  * piece by Van Jacobson.
627  * NOTE: the next three routines used to be one big routine.
628  * To save cycles in the RFC 1323 implementation it was better to break
629  * it up into three procedures. -- erics
630  */
631 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
632 {
633 	struct tcp_sock *tp = tcp_sk(sk);
634 	long m = mrtt; /* RTT */
635 
636 	/*	The following amusing code comes from Jacobson's
637 	 *	article in SIGCOMM '88.  Note that rtt and mdev
638 	 *	are scaled versions of rtt and mean deviation.
639 	 *	This is designed to be as fast as possible
640 	 *	m stands for "measurement".
641 	 *
642 	 *	On a 1990 paper the rto value is changed to:
643 	 *	RTO = rtt + 4 * mdev
644 	 *
645 	 * Funny. This algorithm seems to be very broken.
646 	 * These formulae increase RTO, when it should be decreased, increase
647 	 * too slowly, when it should be increased quickly, decrease too quickly
648 	 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
649 	 * does not matter how to _calculate_ it. Seems, it was trap
650 	 * that VJ failed to avoid. 8)
651 	 */
652 	if (m == 0)
653 		m = 1;
654 	if (tp->srtt != 0) {
655 		m -= (tp->srtt >> 3);	/* m is now error in rtt est */
656 		tp->srtt += m;		/* rtt = 7/8 rtt + 1/8 new */
657 		if (m < 0) {
658 			m = -m;		/* m is now abs(error) */
659 			m -= (tp->mdev >> 2);   /* similar update on mdev */
660 			/* This is similar to one of Eifel findings.
661 			 * Eifel blocks mdev updates when rtt decreases.
662 			 * This solution is a bit different: we use finer gain
663 			 * for mdev in this case (alpha*beta).
664 			 * Like Eifel it also prevents growth of rto,
665 			 * but also it limits too fast rto decreases,
666 			 * happening in pure Eifel.
667 			 */
668 			if (m > 0)
669 				m >>= 3;
670 		} else {
671 			m -= (tp->mdev >> 2);   /* similar update on mdev */
672 		}
673 		tp->mdev += m;	    	/* mdev = 3/4 mdev + 1/4 new */
674 		if (tp->mdev > tp->mdev_max) {
675 			tp->mdev_max = tp->mdev;
676 			if (tp->mdev_max > tp->rttvar)
677 				tp->rttvar = tp->mdev_max;
678 		}
679 		if (after(tp->snd_una, tp->rtt_seq)) {
680 			if (tp->mdev_max < tp->rttvar)
681 				tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
682 			tp->rtt_seq = tp->snd_nxt;
683 			tp->mdev_max = tcp_rto_min(sk);
684 		}
685 	} else {
686 		/* no previous measure. */
687 		tp->srtt = m << 3;	/* take the measured time to be rtt */
688 		tp->mdev = m << 1;	/* make sure rto = 3*rtt */
689 		tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
690 		tp->rtt_seq = tp->snd_nxt;
691 	}
692 }
693 
694 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
695  * routine referred to above.
696  */
697 static inline void tcp_set_rto(struct sock *sk)
698 {
699 	const struct tcp_sock *tp = tcp_sk(sk);
700 	/* Old crap is replaced with new one. 8)
701 	 *
702 	 * More seriously:
703 	 * 1. If rtt variance happened to be less 50msec, it is hallucination.
704 	 *    It cannot be less due to utterly erratic ACK generation made
705 	 *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
706 	 *    to do with delayed acks, because at cwnd>2 true delack timeout
707 	 *    is invisible. Actually, Linux-2.4 also generates erratic
708 	 *    ACKs in some circumstances.
709 	 */
710 	inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
711 
712 	/* 2. Fixups made earlier cannot be right.
713 	 *    If we do not estimate RTO correctly without them,
714 	 *    all the algo is pure shit and should be replaced
715 	 *    with correct one. It is exactly, which we pretend to do.
716 	 */
717 
718 	/* NOTE: clamping at TCP_RTO_MIN is not required, current algo
719 	 * guarantees that rto is higher.
720 	 */
721 	tcp_bound_rto(sk);
722 }
723 
724 /* Save metrics learned by this TCP session.
725    This function is called only, when TCP finishes successfully
726    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
727  */
728 void tcp_update_metrics(struct sock *sk)
729 {
730 	struct tcp_sock *tp = tcp_sk(sk);
731 	struct dst_entry *dst = __sk_dst_get(sk);
732 
733 	if (sysctl_tcp_nometrics_save)
734 		return;
735 
736 	dst_confirm(dst);
737 
738 	if (dst && (dst->flags & DST_HOST)) {
739 		const struct inet_connection_sock *icsk = inet_csk(sk);
740 		int m;
741 		unsigned long rtt;
742 
743 		if (icsk->icsk_backoff || !tp->srtt) {
744 			/* This session failed to estimate rtt. Why?
745 			 * Probably, no packets returned in time.
746 			 * Reset our results.
747 			 */
748 			if (!(dst_metric_locked(dst, RTAX_RTT)))
749 				dst_metric_set(dst, RTAX_RTT, 0);
750 			return;
751 		}
752 
753 		rtt = dst_metric_rtt(dst, RTAX_RTT);
754 		m = rtt - tp->srtt;
755 
756 		/* If newly calculated rtt larger than stored one,
757 		 * store new one. Otherwise, use EWMA. Remember,
758 		 * rtt overestimation is always better than underestimation.
759 		 */
760 		if (!(dst_metric_locked(dst, RTAX_RTT))) {
761 			if (m <= 0)
762 				set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
763 			else
764 				set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
765 		}
766 
767 		if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
768 			unsigned long var;
769 			if (m < 0)
770 				m = -m;
771 
772 			/* Scale deviation to rttvar fixed point */
773 			m >>= 1;
774 			if (m < tp->mdev)
775 				m = tp->mdev;
776 
777 			var = dst_metric_rtt(dst, RTAX_RTTVAR);
778 			if (m >= var)
779 				var = m;
780 			else
781 				var -= (var - m) >> 2;
782 
783 			set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
784 		}
785 
786 		if (tcp_in_initial_slowstart(tp)) {
787 			/* Slow start still did not finish. */
788 			if (dst_metric(dst, RTAX_SSTHRESH) &&
789 			    !dst_metric_locked(dst, RTAX_SSTHRESH) &&
790 			    (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
791 				dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
792 			if (!dst_metric_locked(dst, RTAX_CWND) &&
793 			    tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
794 				dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
795 		} else if (tp->snd_cwnd > tp->snd_ssthresh &&
796 			   icsk->icsk_ca_state == TCP_CA_Open) {
797 			/* Cong. avoidance phase, cwnd is reliable. */
798 			if (!dst_metric_locked(dst, RTAX_SSTHRESH))
799 				dst_metric_set(dst, RTAX_SSTHRESH,
800 					       max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
801 			if (!dst_metric_locked(dst, RTAX_CWND))
802 				dst_metric_set(dst, RTAX_CWND,
803 					       (dst_metric(dst, RTAX_CWND) +
804 						tp->snd_cwnd) >> 1);
805 		} else {
806 			/* Else slow start did not finish, cwnd is non-sense,
807 			   ssthresh may be also invalid.
808 			 */
809 			if (!dst_metric_locked(dst, RTAX_CWND))
810 				dst_metric_set(dst, RTAX_CWND,
811 					       (dst_metric(dst, RTAX_CWND) +
812 						tp->snd_ssthresh) >> 1);
813 			if (dst_metric(dst, RTAX_SSTHRESH) &&
814 			    !dst_metric_locked(dst, RTAX_SSTHRESH) &&
815 			    tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
816 				dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
817 		}
818 
819 		if (!dst_metric_locked(dst, RTAX_REORDERING)) {
820 			if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
821 			    tp->reordering != sysctl_tcp_reordering)
822 				dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
823 		}
824 	}
825 }
826 
827 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
828 {
829 	__u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
830 
831 	if (!cwnd)
832 		cwnd = TCP_INIT_CWND;
833 	return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
834 }
835 
836 /* Set slow start threshold and cwnd not falling to slow start */
837 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
838 {
839 	struct tcp_sock *tp = tcp_sk(sk);
840 	const struct inet_connection_sock *icsk = inet_csk(sk);
841 
842 	tp->prior_ssthresh = 0;
843 	tp->bytes_acked = 0;
844 	if (icsk->icsk_ca_state < TCP_CA_CWR) {
845 		tp->undo_marker = 0;
846 		if (set_ssthresh)
847 			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
848 		tp->snd_cwnd = min(tp->snd_cwnd,
849 				   tcp_packets_in_flight(tp) + 1U);
850 		tp->snd_cwnd_cnt = 0;
851 		tp->high_seq = tp->snd_nxt;
852 		tp->snd_cwnd_stamp = tcp_time_stamp;
853 		TCP_ECN_queue_cwr(tp);
854 
855 		tcp_set_ca_state(sk, TCP_CA_CWR);
856 	}
857 }
858 
859 /*
860  * Packet counting of FACK is based on in-order assumptions, therefore TCP
861  * disables it when reordering is detected
862  */
863 static void tcp_disable_fack(struct tcp_sock *tp)
864 {
865 	/* RFC3517 uses different metric in lost marker => reset on change */
866 	if (tcp_is_fack(tp))
867 		tp->lost_skb_hint = NULL;
868 	tp->rx_opt.sack_ok &= ~2;
869 }
870 
871 /* Take a notice that peer is sending D-SACKs */
872 static void tcp_dsack_seen(struct tcp_sock *tp)
873 {
874 	tp->rx_opt.sack_ok |= 4;
875 }
876 
877 /* Initialize metrics on socket. */
878 
879 static void tcp_init_metrics(struct sock *sk)
880 {
881 	struct tcp_sock *tp = tcp_sk(sk);
882 	struct dst_entry *dst = __sk_dst_get(sk);
883 
884 	if (dst == NULL)
885 		goto reset;
886 
887 	dst_confirm(dst);
888 
889 	if (dst_metric_locked(dst, RTAX_CWND))
890 		tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
891 	if (dst_metric(dst, RTAX_SSTHRESH)) {
892 		tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
893 		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
894 			tp->snd_ssthresh = tp->snd_cwnd_clamp;
895 	} else {
896 		/* ssthresh may have been reduced unnecessarily during.
897 		 * 3WHS. Restore it back to its initial default.
898 		 */
899 		tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
900 	}
901 	if (dst_metric(dst, RTAX_REORDERING) &&
902 	    tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
903 		tcp_disable_fack(tp);
904 		tp->reordering = dst_metric(dst, RTAX_REORDERING);
905 	}
906 
907 	if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
908 		goto reset;
909 
910 	/* Initial rtt is determined from SYN,SYN-ACK.
911 	 * The segment is small and rtt may appear much
912 	 * less than real one. Use per-dst memory
913 	 * to make it more realistic.
914 	 *
915 	 * A bit of theory. RTT is time passed after "normal" sized packet
916 	 * is sent until it is ACKed. In normal circumstances sending small
917 	 * packets force peer to delay ACKs and calculation is correct too.
918 	 * The algorithm is adaptive and, provided we follow specs, it
919 	 * NEVER underestimate RTT. BUT! If peer tries to make some clever
920 	 * tricks sort of "quick acks" for time long enough to decrease RTT
921 	 * to low value, and then abruptly stops to do it and starts to delay
922 	 * ACKs, wait for troubles.
923 	 */
924 	if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
925 		tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
926 		tp->rtt_seq = tp->snd_nxt;
927 	}
928 	if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
929 		tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
930 		tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
931 	}
932 	tcp_set_rto(sk);
933 reset:
934 	if (tp->srtt == 0) {
935 		/* RFC2988bis: We've failed to get a valid RTT sample from
936 		 * 3WHS. This is most likely due to retransmission,
937 		 * including spurious one. Reset the RTO back to 3secs
938 		 * from the more aggressive 1sec to avoid more spurious
939 		 * retransmission.
940 		 */
941 		tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
942 		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
943 	}
944 	/* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
945 	 * retransmitted. In light of RFC2988bis' more aggressive 1sec
946 	 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
947 	 * retransmission has occurred.
948 	 */
949 	if (tp->total_retrans > 1)
950 		tp->snd_cwnd = 1;
951 	else
952 		tp->snd_cwnd = tcp_init_cwnd(tp, dst);
953 	tp->snd_cwnd_stamp = tcp_time_stamp;
954 }
955 
956 static void tcp_update_reordering(struct sock *sk, const int metric,
957 				  const int ts)
958 {
959 	struct tcp_sock *tp = tcp_sk(sk);
960 	if (metric > tp->reordering) {
961 		int mib_idx;
962 
963 		tp->reordering = min(TCP_MAX_REORDERING, metric);
964 
965 		/* This exciting event is worth to be remembered. 8) */
966 		if (ts)
967 			mib_idx = LINUX_MIB_TCPTSREORDER;
968 		else if (tcp_is_reno(tp))
969 			mib_idx = LINUX_MIB_TCPRENOREORDER;
970 		else if (tcp_is_fack(tp))
971 			mib_idx = LINUX_MIB_TCPFACKREORDER;
972 		else
973 			mib_idx = LINUX_MIB_TCPSACKREORDER;
974 
975 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
976 #if FASTRETRANS_DEBUG > 1
977 		printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
978 		       tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
979 		       tp->reordering,
980 		       tp->fackets_out,
981 		       tp->sacked_out,
982 		       tp->undo_marker ? tp->undo_retrans : 0);
983 #endif
984 		tcp_disable_fack(tp);
985 	}
986 }
987 
988 /* This must be called before lost_out is incremented */
989 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
990 {
991 	if ((tp->retransmit_skb_hint == NULL) ||
992 	    before(TCP_SKB_CB(skb)->seq,
993 		   TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
994 		tp->retransmit_skb_hint = skb;
995 
996 	if (!tp->lost_out ||
997 	    after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
998 		tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
999 }
1000 
1001 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
1002 {
1003 	if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1004 		tcp_verify_retransmit_hint(tp, skb);
1005 
1006 		tp->lost_out += tcp_skb_pcount(skb);
1007 		TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1008 	}
1009 }
1010 
1011 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1012 					    struct sk_buff *skb)
1013 {
1014 	tcp_verify_retransmit_hint(tp, skb);
1015 
1016 	if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1017 		tp->lost_out += tcp_skb_pcount(skb);
1018 		TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1019 	}
1020 }
1021 
1022 /* This procedure tags the retransmission queue when SACKs arrive.
1023  *
1024  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1025  * Packets in queue with these bits set are counted in variables
1026  * sacked_out, retrans_out and lost_out, correspondingly.
1027  *
1028  * Valid combinations are:
1029  * Tag  InFlight	Description
1030  * 0	1		- orig segment is in flight.
1031  * S	0		- nothing flies, orig reached receiver.
1032  * L	0		- nothing flies, orig lost by net.
1033  * R	2		- both orig and retransmit are in flight.
1034  * L|R	1		- orig is lost, retransmit is in flight.
1035  * S|R  1		- orig reached receiver, retrans is still in flight.
1036  * (L|S|R is logically valid, it could occur when L|R is sacked,
1037  *  but it is equivalent to plain S and code short-curcuits it to S.
1038  *  L|S is logically invalid, it would mean -1 packet in flight 8))
1039  *
1040  * These 6 states form finite state machine, controlled by the following events:
1041  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1042  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1043  * 3. Loss detection event of one of three flavors:
1044  *	A. Scoreboard estimator decided the packet is lost.
1045  *	   A'. Reno "three dupacks" marks head of queue lost.
1046  *	   A''. Its FACK modfication, head until snd.fack is lost.
1047  *	B. SACK arrives sacking data transmitted after never retransmitted
1048  *	   hole was sent out.
1049  *	C. SACK arrives sacking SND.NXT at the moment, when the
1050  *	   segment was retransmitted.
1051  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1052  *
1053  * It is pleasant to note, that state diagram turns out to be commutative,
1054  * so that we are allowed not to be bothered by order of our actions,
1055  * when multiple events arrive simultaneously. (see the function below).
1056  *
1057  * Reordering detection.
1058  * --------------------
1059  * Reordering metric is maximal distance, which a packet can be displaced
1060  * in packet stream. With SACKs we can estimate it:
1061  *
1062  * 1. SACK fills old hole and the corresponding segment was not
1063  *    ever retransmitted -> reordering. Alas, we cannot use it
1064  *    when segment was retransmitted.
1065  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1066  *    for retransmitted and already SACKed segment -> reordering..
1067  * Both of these heuristics are not used in Loss state, when we cannot
1068  * account for retransmits accurately.
1069  *
1070  * SACK block validation.
1071  * ----------------------
1072  *
1073  * SACK block range validation checks that the received SACK block fits to
1074  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1075  * Note that SND.UNA is not included to the range though being valid because
1076  * it means that the receiver is rather inconsistent with itself reporting
1077  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1078  * perfectly valid, however, in light of RFC2018 which explicitly states
1079  * that "SACK block MUST reflect the newest segment.  Even if the newest
1080  * segment is going to be discarded ...", not that it looks very clever
1081  * in case of head skb. Due to potentional receiver driven attacks, we
1082  * choose to avoid immediate execution of a walk in write queue due to
1083  * reneging and defer head skb's loss recovery to standard loss recovery
1084  * procedure that will eventually trigger (nothing forbids us doing this).
1085  *
1086  * Implements also blockage to start_seq wrap-around. Problem lies in the
1087  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1088  * there's no guarantee that it will be before snd_nxt (n). The problem
1089  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1090  * wrap (s_w):
1091  *
1092  *         <- outs wnd ->                          <- wrapzone ->
1093  *         u     e      n                         u_w   e_w  s n_w
1094  *         |     |      |                          |     |   |  |
1095  * |<------------+------+----- TCP seqno space --------------+---------->|
1096  * ...-- <2^31 ->|                                           |<--------...
1097  * ...---- >2^31 ------>|                                    |<--------...
1098  *
1099  * Current code wouldn't be vulnerable but it's better still to discard such
1100  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1101  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1102  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1103  * equal to the ideal case (infinite seqno space without wrap caused issues).
1104  *
1105  * With D-SACK the lower bound is extended to cover sequence space below
1106  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1107  * again, D-SACK block must not to go across snd_una (for the same reason as
1108  * for the normal SACK blocks, explained above). But there all simplicity
1109  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1110  * fully below undo_marker they do not affect behavior in anyway and can
1111  * therefore be safely ignored. In rare cases (which are more or less
1112  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1113  * fragmentation and packet reordering past skb's retransmission. To consider
1114  * them correctly, the acceptable range must be extended even more though
1115  * the exact amount is rather hard to quantify. However, tp->max_window can
1116  * be used as an exaggerated estimate.
1117  */
1118 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1119 				  u32 start_seq, u32 end_seq)
1120 {
1121 	/* Too far in future, or reversed (interpretation is ambiguous) */
1122 	if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1123 		return 0;
1124 
1125 	/* Nasty start_seq wrap-around check (see comments above) */
1126 	if (!before(start_seq, tp->snd_nxt))
1127 		return 0;
1128 
1129 	/* In outstanding window? ...This is valid exit for D-SACKs too.
1130 	 * start_seq == snd_una is non-sensical (see comments above)
1131 	 */
1132 	if (after(start_seq, tp->snd_una))
1133 		return 1;
1134 
1135 	if (!is_dsack || !tp->undo_marker)
1136 		return 0;
1137 
1138 	/* ...Then it's D-SACK, and must reside below snd_una completely */
1139 	if (after(end_seq, tp->snd_una))
1140 		return 0;
1141 
1142 	if (!before(start_seq, tp->undo_marker))
1143 		return 1;
1144 
1145 	/* Too old */
1146 	if (!after(end_seq, tp->undo_marker))
1147 		return 0;
1148 
1149 	/* Undo_marker boundary crossing (overestimates a lot). Known already:
1150 	 *   start_seq < undo_marker and end_seq >= undo_marker.
1151 	 */
1152 	return !before(start_seq, end_seq - tp->max_window);
1153 }
1154 
1155 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1156  * Event "C". Later note: FACK people cheated me again 8), we have to account
1157  * for reordering! Ugly, but should help.
1158  *
1159  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1160  * less than what is now known to be received by the other end (derived from
1161  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1162  * retransmitted skbs to avoid some costly processing per ACKs.
1163  */
1164 static void tcp_mark_lost_retrans(struct sock *sk)
1165 {
1166 	const struct inet_connection_sock *icsk = inet_csk(sk);
1167 	struct tcp_sock *tp = tcp_sk(sk);
1168 	struct sk_buff *skb;
1169 	int cnt = 0;
1170 	u32 new_low_seq = tp->snd_nxt;
1171 	u32 received_upto = tcp_highest_sack_seq(tp);
1172 
1173 	if (!tcp_is_fack(tp) || !tp->retrans_out ||
1174 	    !after(received_upto, tp->lost_retrans_low) ||
1175 	    icsk->icsk_ca_state != TCP_CA_Recovery)
1176 		return;
1177 
1178 	tcp_for_write_queue(skb, sk) {
1179 		u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1180 
1181 		if (skb == tcp_send_head(sk))
1182 			break;
1183 		if (cnt == tp->retrans_out)
1184 			break;
1185 		if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1186 			continue;
1187 
1188 		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1189 			continue;
1190 
1191 		/* TODO: We would like to get rid of tcp_is_fack(tp) only
1192 		 * constraint here (see above) but figuring out that at
1193 		 * least tp->reordering SACK blocks reside between ack_seq
1194 		 * and received_upto is not easy task to do cheaply with
1195 		 * the available datastructures.
1196 		 *
1197 		 * Whether FACK should check here for tp->reordering segs
1198 		 * in-between one could argue for either way (it would be
1199 		 * rather simple to implement as we could count fack_count
1200 		 * during the walk and do tp->fackets_out - fack_count).
1201 		 */
1202 		if (after(received_upto, ack_seq)) {
1203 			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1204 			tp->retrans_out -= tcp_skb_pcount(skb);
1205 
1206 			tcp_skb_mark_lost_uncond_verify(tp, skb);
1207 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1208 		} else {
1209 			if (before(ack_seq, new_low_seq))
1210 				new_low_seq = ack_seq;
1211 			cnt += tcp_skb_pcount(skb);
1212 		}
1213 	}
1214 
1215 	if (tp->retrans_out)
1216 		tp->lost_retrans_low = new_low_seq;
1217 }
1218 
1219 static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1220 			   struct tcp_sack_block_wire *sp, int num_sacks,
1221 			   u32 prior_snd_una)
1222 {
1223 	struct tcp_sock *tp = tcp_sk(sk);
1224 	u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1225 	u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1226 	int dup_sack = 0;
1227 
1228 	if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1229 		dup_sack = 1;
1230 		tcp_dsack_seen(tp);
1231 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1232 	} else if (num_sacks > 1) {
1233 		u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1234 		u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1235 
1236 		if (!after(end_seq_0, end_seq_1) &&
1237 		    !before(start_seq_0, start_seq_1)) {
1238 			dup_sack = 1;
1239 			tcp_dsack_seen(tp);
1240 			NET_INC_STATS_BH(sock_net(sk),
1241 					LINUX_MIB_TCPDSACKOFORECV);
1242 		}
1243 	}
1244 
1245 	/* D-SACK for already forgotten data... Do dumb counting. */
1246 	if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1247 	    !after(end_seq_0, prior_snd_una) &&
1248 	    after(end_seq_0, tp->undo_marker))
1249 		tp->undo_retrans--;
1250 
1251 	return dup_sack;
1252 }
1253 
1254 struct tcp_sacktag_state {
1255 	int reord;
1256 	int fack_count;
1257 	int flag;
1258 };
1259 
1260 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1261  * the incoming SACK may not exactly match but we can find smaller MSS
1262  * aligned portion of it that matches. Therefore we might need to fragment
1263  * which may fail and creates some hassle (caller must handle error case
1264  * returns).
1265  *
1266  * FIXME: this could be merged to shift decision code
1267  */
1268 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1269 				 u32 start_seq, u32 end_seq)
1270 {
1271 	int in_sack, err;
1272 	unsigned int pkt_len;
1273 	unsigned int mss;
1274 
1275 	in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1276 		  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1277 
1278 	if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1279 	    after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1280 		mss = tcp_skb_mss(skb);
1281 		in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1282 
1283 		if (!in_sack) {
1284 			pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1285 			if (pkt_len < mss)
1286 				pkt_len = mss;
1287 		} else {
1288 			pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1289 			if (pkt_len < mss)
1290 				return -EINVAL;
1291 		}
1292 
1293 		/* Round if necessary so that SACKs cover only full MSSes
1294 		 * and/or the remaining small portion (if present)
1295 		 */
1296 		if (pkt_len > mss) {
1297 			unsigned int new_len = (pkt_len / mss) * mss;
1298 			if (!in_sack && new_len < pkt_len) {
1299 				new_len += mss;
1300 				if (new_len > skb->len)
1301 					return 0;
1302 			}
1303 			pkt_len = new_len;
1304 		}
1305 		err = tcp_fragment(sk, skb, pkt_len, mss);
1306 		if (err < 0)
1307 			return err;
1308 	}
1309 
1310 	return in_sack;
1311 }
1312 
1313 static u8 tcp_sacktag_one(const struct sk_buff *skb, struct sock *sk,
1314 			  struct tcp_sacktag_state *state,
1315 			  int dup_sack, int pcount)
1316 {
1317 	struct tcp_sock *tp = tcp_sk(sk);
1318 	u8 sacked = TCP_SKB_CB(skb)->sacked;
1319 	int fack_count = state->fack_count;
1320 
1321 	/* Account D-SACK for retransmitted packet. */
1322 	if (dup_sack && (sacked & TCPCB_RETRANS)) {
1323 		if (tp->undo_marker && tp->undo_retrans &&
1324 		    after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1325 			tp->undo_retrans--;
1326 		if (sacked & TCPCB_SACKED_ACKED)
1327 			state->reord = min(fack_count, state->reord);
1328 	}
1329 
1330 	/* Nothing to do; acked frame is about to be dropped (was ACKed). */
1331 	if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1332 		return sacked;
1333 
1334 	if (!(sacked & TCPCB_SACKED_ACKED)) {
1335 		if (sacked & TCPCB_SACKED_RETRANS) {
1336 			/* If the segment is not tagged as lost,
1337 			 * we do not clear RETRANS, believing
1338 			 * that retransmission is still in flight.
1339 			 */
1340 			if (sacked & TCPCB_LOST) {
1341 				sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1342 				tp->lost_out -= pcount;
1343 				tp->retrans_out -= pcount;
1344 			}
1345 		} else {
1346 			if (!(sacked & TCPCB_RETRANS)) {
1347 				/* New sack for not retransmitted frame,
1348 				 * which was in hole. It is reordering.
1349 				 */
1350 				if (before(TCP_SKB_CB(skb)->seq,
1351 					   tcp_highest_sack_seq(tp)))
1352 					state->reord = min(fack_count,
1353 							   state->reord);
1354 
1355 				/* SACK enhanced F-RTO (RFC4138; Appendix B) */
1356 				if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1357 					state->flag |= FLAG_ONLY_ORIG_SACKED;
1358 			}
1359 
1360 			if (sacked & TCPCB_LOST) {
1361 				sacked &= ~TCPCB_LOST;
1362 				tp->lost_out -= pcount;
1363 			}
1364 		}
1365 
1366 		sacked |= TCPCB_SACKED_ACKED;
1367 		state->flag |= FLAG_DATA_SACKED;
1368 		tp->sacked_out += pcount;
1369 
1370 		fack_count += pcount;
1371 
1372 		/* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1373 		if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1374 		    before(TCP_SKB_CB(skb)->seq,
1375 			   TCP_SKB_CB(tp->lost_skb_hint)->seq))
1376 			tp->lost_cnt_hint += pcount;
1377 
1378 		if (fack_count > tp->fackets_out)
1379 			tp->fackets_out = fack_count;
1380 	}
1381 
1382 	/* D-SACK. We can detect redundant retransmission in S|R and plain R
1383 	 * frames and clear it. undo_retrans is decreased above, L|R frames
1384 	 * are accounted above as well.
1385 	 */
1386 	if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1387 		sacked &= ~TCPCB_SACKED_RETRANS;
1388 		tp->retrans_out -= pcount;
1389 	}
1390 
1391 	return sacked;
1392 }
1393 
1394 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1395 			   struct tcp_sacktag_state *state,
1396 			   unsigned int pcount, int shifted, int mss,
1397 			   int dup_sack)
1398 {
1399 	struct tcp_sock *tp = tcp_sk(sk);
1400 	struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1401 
1402 	BUG_ON(!pcount);
1403 
1404 	if (skb == tp->lost_skb_hint)
1405 		tp->lost_cnt_hint += pcount;
1406 
1407 	TCP_SKB_CB(prev)->end_seq += shifted;
1408 	TCP_SKB_CB(skb)->seq += shifted;
1409 
1410 	skb_shinfo(prev)->gso_segs += pcount;
1411 	BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1412 	skb_shinfo(skb)->gso_segs -= pcount;
1413 
1414 	/* When we're adding to gso_segs == 1, gso_size will be zero,
1415 	 * in theory this shouldn't be necessary but as long as DSACK
1416 	 * code can come after this skb later on it's better to keep
1417 	 * setting gso_size to something.
1418 	 */
1419 	if (!skb_shinfo(prev)->gso_size) {
1420 		skb_shinfo(prev)->gso_size = mss;
1421 		skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1422 	}
1423 
1424 	/* CHECKME: To clear or not to clear? Mimics normal skb currently */
1425 	if (skb_shinfo(skb)->gso_segs <= 1) {
1426 		skb_shinfo(skb)->gso_size = 0;
1427 		skb_shinfo(skb)->gso_type = 0;
1428 	}
1429 
1430 	/* We discard results */
1431 	tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
1432 
1433 	/* Difference in this won't matter, both ACKed by the same cumul. ACK */
1434 	TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1435 
1436 	if (skb->len > 0) {
1437 		BUG_ON(!tcp_skb_pcount(skb));
1438 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1439 		return 0;
1440 	}
1441 
1442 	/* Whole SKB was eaten :-) */
1443 
1444 	if (skb == tp->retransmit_skb_hint)
1445 		tp->retransmit_skb_hint = prev;
1446 	if (skb == tp->scoreboard_skb_hint)
1447 		tp->scoreboard_skb_hint = prev;
1448 	if (skb == tp->lost_skb_hint) {
1449 		tp->lost_skb_hint = prev;
1450 		tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1451 	}
1452 
1453 	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1454 	if (skb == tcp_highest_sack(sk))
1455 		tcp_advance_highest_sack(sk, skb);
1456 
1457 	tcp_unlink_write_queue(skb, sk);
1458 	sk_wmem_free_skb(sk, skb);
1459 
1460 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1461 
1462 	return 1;
1463 }
1464 
1465 /* I wish gso_size would have a bit more sane initialization than
1466  * something-or-zero which complicates things
1467  */
1468 static int tcp_skb_seglen(const struct sk_buff *skb)
1469 {
1470 	return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1471 }
1472 
1473 /* Shifting pages past head area doesn't work */
1474 static int skb_can_shift(const struct sk_buff *skb)
1475 {
1476 	return !skb_headlen(skb) && skb_is_nonlinear(skb);
1477 }
1478 
1479 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1480  * skb.
1481  */
1482 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1483 					  struct tcp_sacktag_state *state,
1484 					  u32 start_seq, u32 end_seq,
1485 					  int dup_sack)
1486 {
1487 	struct tcp_sock *tp = tcp_sk(sk);
1488 	struct sk_buff *prev;
1489 	int mss;
1490 	int pcount = 0;
1491 	int len;
1492 	int in_sack;
1493 
1494 	if (!sk_can_gso(sk))
1495 		goto fallback;
1496 
1497 	/* Normally R but no L won't result in plain S */
1498 	if (!dup_sack &&
1499 	    (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1500 		goto fallback;
1501 	if (!skb_can_shift(skb))
1502 		goto fallback;
1503 	/* This frame is about to be dropped (was ACKed). */
1504 	if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1505 		goto fallback;
1506 
1507 	/* Can only happen with delayed DSACK + discard craziness */
1508 	if (unlikely(skb == tcp_write_queue_head(sk)))
1509 		goto fallback;
1510 	prev = tcp_write_queue_prev(sk, skb);
1511 
1512 	if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1513 		goto fallback;
1514 
1515 	in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1516 		  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1517 
1518 	if (in_sack) {
1519 		len = skb->len;
1520 		pcount = tcp_skb_pcount(skb);
1521 		mss = tcp_skb_seglen(skb);
1522 
1523 		/* TODO: Fix DSACKs to not fragment already SACKed and we can
1524 		 * drop this restriction as unnecessary
1525 		 */
1526 		if (mss != tcp_skb_seglen(prev))
1527 			goto fallback;
1528 	} else {
1529 		if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1530 			goto noop;
1531 		/* CHECKME: This is non-MSS split case only?, this will
1532 		 * cause skipped skbs due to advancing loop btw, original
1533 		 * has that feature too
1534 		 */
1535 		if (tcp_skb_pcount(skb) <= 1)
1536 			goto noop;
1537 
1538 		in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1539 		if (!in_sack) {
1540 			/* TODO: head merge to next could be attempted here
1541 			 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1542 			 * though it might not be worth of the additional hassle
1543 			 *
1544 			 * ...we can probably just fallback to what was done
1545 			 * previously. We could try merging non-SACKed ones
1546 			 * as well but it probably isn't going to buy off
1547 			 * because later SACKs might again split them, and
1548 			 * it would make skb timestamp tracking considerably
1549 			 * harder problem.
1550 			 */
1551 			goto fallback;
1552 		}
1553 
1554 		len = end_seq - TCP_SKB_CB(skb)->seq;
1555 		BUG_ON(len < 0);
1556 		BUG_ON(len > skb->len);
1557 
1558 		/* MSS boundaries should be honoured or else pcount will
1559 		 * severely break even though it makes things bit trickier.
1560 		 * Optimize common case to avoid most of the divides
1561 		 */
1562 		mss = tcp_skb_mss(skb);
1563 
1564 		/* TODO: Fix DSACKs to not fragment already SACKed and we can
1565 		 * drop this restriction as unnecessary
1566 		 */
1567 		if (mss != tcp_skb_seglen(prev))
1568 			goto fallback;
1569 
1570 		if (len == mss) {
1571 			pcount = 1;
1572 		} else if (len < mss) {
1573 			goto noop;
1574 		} else {
1575 			pcount = len / mss;
1576 			len = pcount * mss;
1577 		}
1578 	}
1579 
1580 	if (!skb_shift(prev, skb, len))
1581 		goto fallback;
1582 	if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1583 		goto out;
1584 
1585 	/* Hole filled allows collapsing with the next as well, this is very
1586 	 * useful when hole on every nth skb pattern happens
1587 	 */
1588 	if (prev == tcp_write_queue_tail(sk))
1589 		goto out;
1590 	skb = tcp_write_queue_next(sk, prev);
1591 
1592 	if (!skb_can_shift(skb) ||
1593 	    (skb == tcp_send_head(sk)) ||
1594 	    ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1595 	    (mss != tcp_skb_seglen(skb)))
1596 		goto out;
1597 
1598 	len = skb->len;
1599 	if (skb_shift(prev, skb, len)) {
1600 		pcount += tcp_skb_pcount(skb);
1601 		tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1602 	}
1603 
1604 out:
1605 	state->fack_count += pcount;
1606 	return prev;
1607 
1608 noop:
1609 	return skb;
1610 
1611 fallback:
1612 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1613 	return NULL;
1614 }
1615 
1616 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1617 					struct tcp_sack_block *next_dup,
1618 					struct tcp_sacktag_state *state,
1619 					u32 start_seq, u32 end_seq,
1620 					int dup_sack_in)
1621 {
1622 	struct tcp_sock *tp = tcp_sk(sk);
1623 	struct sk_buff *tmp;
1624 
1625 	tcp_for_write_queue_from(skb, sk) {
1626 		int in_sack = 0;
1627 		int dup_sack = dup_sack_in;
1628 
1629 		if (skb == tcp_send_head(sk))
1630 			break;
1631 
1632 		/* queue is in-order => we can short-circuit the walk early */
1633 		if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1634 			break;
1635 
1636 		if ((next_dup != NULL) &&
1637 		    before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1638 			in_sack = tcp_match_skb_to_sack(sk, skb,
1639 							next_dup->start_seq,
1640 							next_dup->end_seq);
1641 			if (in_sack > 0)
1642 				dup_sack = 1;
1643 		}
1644 
1645 		/* skb reference here is a bit tricky to get right, since
1646 		 * shifting can eat and free both this skb and the next,
1647 		 * so not even _safe variant of the loop is enough.
1648 		 */
1649 		if (in_sack <= 0) {
1650 			tmp = tcp_shift_skb_data(sk, skb, state,
1651 						 start_seq, end_seq, dup_sack);
1652 			if (tmp != NULL) {
1653 				if (tmp != skb) {
1654 					skb = tmp;
1655 					continue;
1656 				}
1657 
1658 				in_sack = 0;
1659 			} else {
1660 				in_sack = tcp_match_skb_to_sack(sk, skb,
1661 								start_seq,
1662 								end_seq);
1663 			}
1664 		}
1665 
1666 		if (unlikely(in_sack < 0))
1667 			break;
1668 
1669 		if (in_sack) {
1670 			TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1671 								  state,
1672 								  dup_sack,
1673 								  tcp_skb_pcount(skb));
1674 
1675 			if (!before(TCP_SKB_CB(skb)->seq,
1676 				    tcp_highest_sack_seq(tp)))
1677 				tcp_advance_highest_sack(sk, skb);
1678 		}
1679 
1680 		state->fack_count += tcp_skb_pcount(skb);
1681 	}
1682 	return skb;
1683 }
1684 
1685 /* Avoid all extra work that is being done by sacktag while walking in
1686  * a normal way
1687  */
1688 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1689 					struct tcp_sacktag_state *state,
1690 					u32 skip_to_seq)
1691 {
1692 	tcp_for_write_queue_from(skb, sk) {
1693 		if (skb == tcp_send_head(sk))
1694 			break;
1695 
1696 		if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1697 			break;
1698 
1699 		state->fack_count += tcp_skb_pcount(skb);
1700 	}
1701 	return skb;
1702 }
1703 
1704 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1705 						struct sock *sk,
1706 						struct tcp_sack_block *next_dup,
1707 						struct tcp_sacktag_state *state,
1708 						u32 skip_to_seq)
1709 {
1710 	if (next_dup == NULL)
1711 		return skb;
1712 
1713 	if (before(next_dup->start_seq, skip_to_seq)) {
1714 		skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1715 		skb = tcp_sacktag_walk(skb, sk, NULL, state,
1716 				       next_dup->start_seq, next_dup->end_seq,
1717 				       1);
1718 	}
1719 
1720 	return skb;
1721 }
1722 
1723 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1724 {
1725 	return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1726 }
1727 
1728 static int
1729 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1730 			u32 prior_snd_una)
1731 {
1732 	const struct inet_connection_sock *icsk = inet_csk(sk);
1733 	struct tcp_sock *tp = tcp_sk(sk);
1734 	const unsigned char *ptr = (skb_transport_header(ack_skb) +
1735 				    TCP_SKB_CB(ack_skb)->sacked);
1736 	struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1737 	struct tcp_sack_block sp[TCP_NUM_SACKS];
1738 	struct tcp_sack_block *cache;
1739 	struct tcp_sacktag_state state;
1740 	struct sk_buff *skb;
1741 	int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1742 	int used_sacks;
1743 	int found_dup_sack = 0;
1744 	int i, j;
1745 	int first_sack_index;
1746 
1747 	state.flag = 0;
1748 	state.reord = tp->packets_out;
1749 
1750 	if (!tp->sacked_out) {
1751 		if (WARN_ON(tp->fackets_out))
1752 			tp->fackets_out = 0;
1753 		tcp_highest_sack_reset(sk);
1754 	}
1755 
1756 	found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1757 					 num_sacks, prior_snd_una);
1758 	if (found_dup_sack)
1759 		state.flag |= FLAG_DSACKING_ACK;
1760 
1761 	/* Eliminate too old ACKs, but take into
1762 	 * account more or less fresh ones, they can
1763 	 * contain valid SACK info.
1764 	 */
1765 	if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1766 		return 0;
1767 
1768 	if (!tp->packets_out)
1769 		goto out;
1770 
1771 	used_sacks = 0;
1772 	first_sack_index = 0;
1773 	for (i = 0; i < num_sacks; i++) {
1774 		int dup_sack = !i && found_dup_sack;
1775 
1776 		sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1777 		sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1778 
1779 		if (!tcp_is_sackblock_valid(tp, dup_sack,
1780 					    sp[used_sacks].start_seq,
1781 					    sp[used_sacks].end_seq)) {
1782 			int mib_idx;
1783 
1784 			if (dup_sack) {
1785 				if (!tp->undo_marker)
1786 					mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1787 				else
1788 					mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1789 			} else {
1790 				/* Don't count olds caused by ACK reordering */
1791 				if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1792 				    !after(sp[used_sacks].end_seq, tp->snd_una))
1793 					continue;
1794 				mib_idx = LINUX_MIB_TCPSACKDISCARD;
1795 			}
1796 
1797 			NET_INC_STATS_BH(sock_net(sk), mib_idx);
1798 			if (i == 0)
1799 				first_sack_index = -1;
1800 			continue;
1801 		}
1802 
1803 		/* Ignore very old stuff early */
1804 		if (!after(sp[used_sacks].end_seq, prior_snd_una))
1805 			continue;
1806 
1807 		used_sacks++;
1808 	}
1809 
1810 	/* order SACK blocks to allow in order walk of the retrans queue */
1811 	for (i = used_sacks - 1; i > 0; i--) {
1812 		for (j = 0; j < i; j++) {
1813 			if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1814 				swap(sp[j], sp[j + 1]);
1815 
1816 				/* Track where the first SACK block goes to */
1817 				if (j == first_sack_index)
1818 					first_sack_index = j + 1;
1819 			}
1820 		}
1821 	}
1822 
1823 	skb = tcp_write_queue_head(sk);
1824 	state.fack_count = 0;
1825 	i = 0;
1826 
1827 	if (!tp->sacked_out) {
1828 		/* It's already past, so skip checking against it */
1829 		cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1830 	} else {
1831 		cache = tp->recv_sack_cache;
1832 		/* Skip empty blocks in at head of the cache */
1833 		while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1834 		       !cache->end_seq)
1835 			cache++;
1836 	}
1837 
1838 	while (i < used_sacks) {
1839 		u32 start_seq = sp[i].start_seq;
1840 		u32 end_seq = sp[i].end_seq;
1841 		int dup_sack = (found_dup_sack && (i == first_sack_index));
1842 		struct tcp_sack_block *next_dup = NULL;
1843 
1844 		if (found_dup_sack && ((i + 1) == first_sack_index))
1845 			next_dup = &sp[i + 1];
1846 
1847 		/* Event "B" in the comment above. */
1848 		if (after(end_seq, tp->high_seq))
1849 			state.flag |= FLAG_DATA_LOST;
1850 
1851 		/* Skip too early cached blocks */
1852 		while (tcp_sack_cache_ok(tp, cache) &&
1853 		       !before(start_seq, cache->end_seq))
1854 			cache++;
1855 
1856 		/* Can skip some work by looking recv_sack_cache? */
1857 		if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1858 		    after(end_seq, cache->start_seq)) {
1859 
1860 			/* Head todo? */
1861 			if (before(start_seq, cache->start_seq)) {
1862 				skb = tcp_sacktag_skip(skb, sk, &state,
1863 						       start_seq);
1864 				skb = tcp_sacktag_walk(skb, sk, next_dup,
1865 						       &state,
1866 						       start_seq,
1867 						       cache->start_seq,
1868 						       dup_sack);
1869 			}
1870 
1871 			/* Rest of the block already fully processed? */
1872 			if (!after(end_seq, cache->end_seq))
1873 				goto advance_sp;
1874 
1875 			skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1876 						       &state,
1877 						       cache->end_seq);
1878 
1879 			/* ...tail remains todo... */
1880 			if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1881 				/* ...but better entrypoint exists! */
1882 				skb = tcp_highest_sack(sk);
1883 				if (skb == NULL)
1884 					break;
1885 				state.fack_count = tp->fackets_out;
1886 				cache++;
1887 				goto walk;
1888 			}
1889 
1890 			skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1891 			/* Check overlap against next cached too (past this one already) */
1892 			cache++;
1893 			continue;
1894 		}
1895 
1896 		if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1897 			skb = tcp_highest_sack(sk);
1898 			if (skb == NULL)
1899 				break;
1900 			state.fack_count = tp->fackets_out;
1901 		}
1902 		skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1903 
1904 walk:
1905 		skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1906 				       start_seq, end_seq, dup_sack);
1907 
1908 advance_sp:
1909 		/* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1910 		 * due to in-order walk
1911 		 */
1912 		if (after(end_seq, tp->frto_highmark))
1913 			state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1914 
1915 		i++;
1916 	}
1917 
1918 	/* Clear the head of the cache sack blocks so we can skip it next time */
1919 	for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1920 		tp->recv_sack_cache[i].start_seq = 0;
1921 		tp->recv_sack_cache[i].end_seq = 0;
1922 	}
1923 	for (j = 0; j < used_sacks; j++)
1924 		tp->recv_sack_cache[i++] = sp[j];
1925 
1926 	tcp_mark_lost_retrans(sk);
1927 
1928 	tcp_verify_left_out(tp);
1929 
1930 	if ((state.reord < tp->fackets_out) &&
1931 	    ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1932 	    (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1933 		tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1934 
1935 out:
1936 
1937 #if FASTRETRANS_DEBUG > 0
1938 	WARN_ON((int)tp->sacked_out < 0);
1939 	WARN_ON((int)tp->lost_out < 0);
1940 	WARN_ON((int)tp->retrans_out < 0);
1941 	WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1942 #endif
1943 	return state.flag;
1944 }
1945 
1946 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1947  * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1948  */
1949 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1950 {
1951 	u32 holes;
1952 
1953 	holes = max(tp->lost_out, 1U);
1954 	holes = min(holes, tp->packets_out);
1955 
1956 	if ((tp->sacked_out + holes) > tp->packets_out) {
1957 		tp->sacked_out = tp->packets_out - holes;
1958 		return 1;
1959 	}
1960 	return 0;
1961 }
1962 
1963 /* If we receive more dupacks than we expected counting segments
1964  * in assumption of absent reordering, interpret this as reordering.
1965  * The only another reason could be bug in receiver TCP.
1966  */
1967 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1968 {
1969 	struct tcp_sock *tp = tcp_sk(sk);
1970 	if (tcp_limit_reno_sacked(tp))
1971 		tcp_update_reordering(sk, tp->packets_out + addend, 0);
1972 }
1973 
1974 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1975 
1976 static void tcp_add_reno_sack(struct sock *sk)
1977 {
1978 	struct tcp_sock *tp = tcp_sk(sk);
1979 	tp->sacked_out++;
1980 	tcp_check_reno_reordering(sk, 0);
1981 	tcp_verify_left_out(tp);
1982 }
1983 
1984 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1985 
1986 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1987 {
1988 	struct tcp_sock *tp = tcp_sk(sk);
1989 
1990 	if (acked > 0) {
1991 		/* One ACK acked hole. The rest eat duplicate ACKs. */
1992 		if (acked - 1 >= tp->sacked_out)
1993 			tp->sacked_out = 0;
1994 		else
1995 			tp->sacked_out -= acked - 1;
1996 	}
1997 	tcp_check_reno_reordering(sk, acked);
1998 	tcp_verify_left_out(tp);
1999 }
2000 
2001 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2002 {
2003 	tp->sacked_out = 0;
2004 }
2005 
2006 static int tcp_is_sackfrto(const struct tcp_sock *tp)
2007 {
2008 	return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2009 }
2010 
2011 /* F-RTO can only be used if TCP has never retransmitted anything other than
2012  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2013  */
2014 int tcp_use_frto(struct sock *sk)
2015 {
2016 	const struct tcp_sock *tp = tcp_sk(sk);
2017 	const struct inet_connection_sock *icsk = inet_csk(sk);
2018 	struct sk_buff *skb;
2019 
2020 	if (!sysctl_tcp_frto)
2021 		return 0;
2022 
2023 	/* MTU probe and F-RTO won't really play nicely along currently */
2024 	if (icsk->icsk_mtup.probe_size)
2025 		return 0;
2026 
2027 	if (tcp_is_sackfrto(tp))
2028 		return 1;
2029 
2030 	/* Avoid expensive walking of rexmit queue if possible */
2031 	if (tp->retrans_out > 1)
2032 		return 0;
2033 
2034 	skb = tcp_write_queue_head(sk);
2035 	if (tcp_skb_is_last(sk, skb))
2036 		return 1;
2037 	skb = tcp_write_queue_next(sk, skb);	/* Skips head */
2038 	tcp_for_write_queue_from(skb, sk) {
2039 		if (skb == tcp_send_head(sk))
2040 			break;
2041 		if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2042 			return 0;
2043 		/* Short-circuit when first non-SACKed skb has been checked */
2044 		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2045 			break;
2046 	}
2047 	return 1;
2048 }
2049 
2050 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2051  * recovery a bit and use heuristics in tcp_process_frto() to detect if
2052  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2053  * keep retrans_out counting accurate (with SACK F-RTO, other than head
2054  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2055  * bits are handled if the Loss state is really to be entered (in
2056  * tcp_enter_frto_loss).
2057  *
2058  * Do like tcp_enter_loss() would; when RTO expires the second time it
2059  * does:
2060  *  "Reduce ssthresh if it has not yet been made inside this window."
2061  */
2062 void tcp_enter_frto(struct sock *sk)
2063 {
2064 	const struct inet_connection_sock *icsk = inet_csk(sk);
2065 	struct tcp_sock *tp = tcp_sk(sk);
2066 	struct sk_buff *skb;
2067 
2068 	if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2069 	    tp->snd_una == tp->high_seq ||
2070 	    ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2071 	     !icsk->icsk_retransmits)) {
2072 		tp->prior_ssthresh = tcp_current_ssthresh(sk);
2073 		/* Our state is too optimistic in ssthresh() call because cwnd
2074 		 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2075 		 * recovery has not yet completed. Pattern would be this: RTO,
2076 		 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2077 		 * up here twice).
2078 		 * RFC4138 should be more specific on what to do, even though
2079 		 * RTO is quite unlikely to occur after the first Cumulative ACK
2080 		 * due to back-off and complexity of triggering events ...
2081 		 */
2082 		if (tp->frto_counter) {
2083 			u32 stored_cwnd;
2084 			stored_cwnd = tp->snd_cwnd;
2085 			tp->snd_cwnd = 2;
2086 			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2087 			tp->snd_cwnd = stored_cwnd;
2088 		} else {
2089 			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2090 		}
2091 		/* ... in theory, cong.control module could do "any tricks" in
2092 		 * ssthresh(), which means that ca_state, lost bits and lost_out
2093 		 * counter would have to be faked before the call occurs. We
2094 		 * consider that too expensive, unlikely and hacky, so modules
2095 		 * using these in ssthresh() must deal these incompatibility
2096 		 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2097 		 */
2098 		tcp_ca_event(sk, CA_EVENT_FRTO);
2099 	}
2100 
2101 	tp->undo_marker = tp->snd_una;
2102 	tp->undo_retrans = 0;
2103 
2104 	skb = tcp_write_queue_head(sk);
2105 	if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2106 		tp->undo_marker = 0;
2107 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2108 		TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2109 		tp->retrans_out -= tcp_skb_pcount(skb);
2110 	}
2111 	tcp_verify_left_out(tp);
2112 
2113 	/* Too bad if TCP was application limited */
2114 	tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2115 
2116 	/* Earlier loss recovery underway (see RFC4138; Appendix B).
2117 	 * The last condition is necessary at least in tp->frto_counter case.
2118 	 */
2119 	if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2120 	    ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2121 	    after(tp->high_seq, tp->snd_una)) {
2122 		tp->frto_highmark = tp->high_seq;
2123 	} else {
2124 		tp->frto_highmark = tp->snd_nxt;
2125 	}
2126 	tcp_set_ca_state(sk, TCP_CA_Disorder);
2127 	tp->high_seq = tp->snd_nxt;
2128 	tp->frto_counter = 1;
2129 }
2130 
2131 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2132  * which indicates that we should follow the traditional RTO recovery,
2133  * i.e. mark everything lost and do go-back-N retransmission.
2134  */
2135 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2136 {
2137 	struct tcp_sock *tp = tcp_sk(sk);
2138 	struct sk_buff *skb;
2139 
2140 	tp->lost_out = 0;
2141 	tp->retrans_out = 0;
2142 	if (tcp_is_reno(tp))
2143 		tcp_reset_reno_sack(tp);
2144 
2145 	tcp_for_write_queue(skb, sk) {
2146 		if (skb == tcp_send_head(sk))
2147 			break;
2148 
2149 		TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2150 		/*
2151 		 * Count the retransmission made on RTO correctly (only when
2152 		 * waiting for the first ACK and did not get it)...
2153 		 */
2154 		if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2155 			/* For some reason this R-bit might get cleared? */
2156 			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2157 				tp->retrans_out += tcp_skb_pcount(skb);
2158 			/* ...enter this if branch just for the first segment */
2159 			flag |= FLAG_DATA_ACKED;
2160 		} else {
2161 			if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2162 				tp->undo_marker = 0;
2163 			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2164 		}
2165 
2166 		/* Marking forward transmissions that were made after RTO lost
2167 		 * can cause unnecessary retransmissions in some scenarios,
2168 		 * SACK blocks will mitigate that in some but not in all cases.
2169 		 * We used to not mark them but it was causing break-ups with
2170 		 * receivers that do only in-order receival.
2171 		 *
2172 		 * TODO: we could detect presence of such receiver and select
2173 		 * different behavior per flow.
2174 		 */
2175 		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2176 			TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2177 			tp->lost_out += tcp_skb_pcount(skb);
2178 			tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2179 		}
2180 	}
2181 	tcp_verify_left_out(tp);
2182 
2183 	tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2184 	tp->snd_cwnd_cnt = 0;
2185 	tp->snd_cwnd_stamp = tcp_time_stamp;
2186 	tp->frto_counter = 0;
2187 	tp->bytes_acked = 0;
2188 
2189 	tp->reordering = min_t(unsigned int, tp->reordering,
2190 			       sysctl_tcp_reordering);
2191 	tcp_set_ca_state(sk, TCP_CA_Loss);
2192 	tp->high_seq = tp->snd_nxt;
2193 	TCP_ECN_queue_cwr(tp);
2194 
2195 	tcp_clear_all_retrans_hints(tp);
2196 }
2197 
2198 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2199 {
2200 	tp->retrans_out = 0;
2201 	tp->lost_out = 0;
2202 
2203 	tp->undo_marker = 0;
2204 	tp->undo_retrans = 0;
2205 }
2206 
2207 void tcp_clear_retrans(struct tcp_sock *tp)
2208 {
2209 	tcp_clear_retrans_partial(tp);
2210 
2211 	tp->fackets_out = 0;
2212 	tp->sacked_out = 0;
2213 }
2214 
2215 /* Enter Loss state. If "how" is not zero, forget all SACK information
2216  * and reset tags completely, otherwise preserve SACKs. If receiver
2217  * dropped its ofo queue, we will know this due to reneging detection.
2218  */
2219 void tcp_enter_loss(struct sock *sk, int how)
2220 {
2221 	const struct inet_connection_sock *icsk = inet_csk(sk);
2222 	struct tcp_sock *tp = tcp_sk(sk);
2223 	struct sk_buff *skb;
2224 
2225 	/* Reduce ssthresh if it has not yet been made inside this window. */
2226 	if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2227 	    (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2228 		tp->prior_ssthresh = tcp_current_ssthresh(sk);
2229 		tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2230 		tcp_ca_event(sk, CA_EVENT_LOSS);
2231 	}
2232 	tp->snd_cwnd	   = 1;
2233 	tp->snd_cwnd_cnt   = 0;
2234 	tp->snd_cwnd_stamp = tcp_time_stamp;
2235 
2236 	tp->bytes_acked = 0;
2237 	tcp_clear_retrans_partial(tp);
2238 
2239 	if (tcp_is_reno(tp))
2240 		tcp_reset_reno_sack(tp);
2241 
2242 	if (!how) {
2243 		/* Push undo marker, if it was plain RTO and nothing
2244 		 * was retransmitted. */
2245 		tp->undo_marker = tp->snd_una;
2246 	} else {
2247 		tp->sacked_out = 0;
2248 		tp->fackets_out = 0;
2249 	}
2250 	tcp_clear_all_retrans_hints(tp);
2251 
2252 	tcp_for_write_queue(skb, sk) {
2253 		if (skb == tcp_send_head(sk))
2254 			break;
2255 
2256 		if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2257 			tp->undo_marker = 0;
2258 		TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2259 		if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2260 			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2261 			TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2262 			tp->lost_out += tcp_skb_pcount(skb);
2263 			tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2264 		}
2265 	}
2266 	tcp_verify_left_out(tp);
2267 
2268 	tp->reordering = min_t(unsigned int, tp->reordering,
2269 			       sysctl_tcp_reordering);
2270 	tcp_set_ca_state(sk, TCP_CA_Loss);
2271 	tp->high_seq = tp->snd_nxt;
2272 	TCP_ECN_queue_cwr(tp);
2273 	/* Abort F-RTO algorithm if one is in progress */
2274 	tp->frto_counter = 0;
2275 }
2276 
2277 /* If ACK arrived pointing to a remembered SACK, it means that our
2278  * remembered SACKs do not reflect real state of receiver i.e.
2279  * receiver _host_ is heavily congested (or buggy).
2280  *
2281  * Do processing similar to RTO timeout.
2282  */
2283 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2284 {
2285 	if (flag & FLAG_SACK_RENEGING) {
2286 		struct inet_connection_sock *icsk = inet_csk(sk);
2287 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2288 
2289 		tcp_enter_loss(sk, 1);
2290 		icsk->icsk_retransmits++;
2291 		tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2292 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2293 					  icsk->icsk_rto, TCP_RTO_MAX);
2294 		return 1;
2295 	}
2296 	return 0;
2297 }
2298 
2299 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2300 {
2301 	return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2302 }
2303 
2304 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2305  * counter when SACK is enabled (without SACK, sacked_out is used for
2306  * that purpose).
2307  *
2308  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2309  * segments up to the highest received SACK block so far and holes in
2310  * between them.
2311  *
2312  * With reordering, holes may still be in flight, so RFC3517 recovery
2313  * uses pure sacked_out (total number of SACKed segments) even though
2314  * it violates the RFC that uses duplicate ACKs, often these are equal
2315  * but when e.g. out-of-window ACKs or packet duplication occurs,
2316  * they differ. Since neither occurs due to loss, TCP should really
2317  * ignore them.
2318  */
2319 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2320 {
2321 	return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2322 }
2323 
2324 static inline int tcp_skb_timedout(const struct sock *sk,
2325 				   const struct sk_buff *skb)
2326 {
2327 	return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2328 }
2329 
2330 static inline int tcp_head_timedout(const struct sock *sk)
2331 {
2332 	const struct tcp_sock *tp = tcp_sk(sk);
2333 
2334 	return tp->packets_out &&
2335 	       tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2336 }
2337 
2338 /* Linux NewReno/SACK/FACK/ECN state machine.
2339  * --------------------------------------
2340  *
2341  * "Open"	Normal state, no dubious events, fast path.
2342  * "Disorder"   In all the respects it is "Open",
2343  *		but requires a bit more attention. It is entered when
2344  *		we see some SACKs or dupacks. It is split of "Open"
2345  *		mainly to move some processing from fast path to slow one.
2346  * "CWR"	CWND was reduced due to some Congestion Notification event.
2347  *		It can be ECN, ICMP source quench, local device congestion.
2348  * "Recovery"	CWND was reduced, we are fast-retransmitting.
2349  * "Loss"	CWND was reduced due to RTO timeout or SACK reneging.
2350  *
2351  * tcp_fastretrans_alert() is entered:
2352  * - each incoming ACK, if state is not "Open"
2353  * - when arrived ACK is unusual, namely:
2354  *	* SACK
2355  *	* Duplicate ACK.
2356  *	* ECN ECE.
2357  *
2358  * Counting packets in flight is pretty simple.
2359  *
2360  *	in_flight = packets_out - left_out + retrans_out
2361  *
2362  *	packets_out is SND.NXT-SND.UNA counted in packets.
2363  *
2364  *	retrans_out is number of retransmitted segments.
2365  *
2366  *	left_out is number of segments left network, but not ACKed yet.
2367  *
2368  *		left_out = sacked_out + lost_out
2369  *
2370  *     sacked_out: Packets, which arrived to receiver out of order
2371  *		   and hence not ACKed. With SACKs this number is simply
2372  *		   amount of SACKed data. Even without SACKs
2373  *		   it is easy to give pretty reliable estimate of this number,
2374  *		   counting duplicate ACKs.
2375  *
2376  *       lost_out: Packets lost by network. TCP has no explicit
2377  *		   "loss notification" feedback from network (for now).
2378  *		   It means that this number can be only _guessed_.
2379  *		   Actually, it is the heuristics to predict lossage that
2380  *		   distinguishes different algorithms.
2381  *
2382  *	F.e. after RTO, when all the queue is considered as lost,
2383  *	lost_out = packets_out and in_flight = retrans_out.
2384  *
2385  *		Essentially, we have now two algorithms counting
2386  *		lost packets.
2387  *
2388  *		FACK: It is the simplest heuristics. As soon as we decided
2389  *		that something is lost, we decide that _all_ not SACKed
2390  *		packets until the most forward SACK are lost. I.e.
2391  *		lost_out = fackets_out - sacked_out and left_out = fackets_out.
2392  *		It is absolutely correct estimate, if network does not reorder
2393  *		packets. And it loses any connection to reality when reordering
2394  *		takes place. We use FACK by default until reordering
2395  *		is suspected on the path to this destination.
2396  *
2397  *		NewReno: when Recovery is entered, we assume that one segment
2398  *		is lost (classic Reno). While we are in Recovery and
2399  *		a partial ACK arrives, we assume that one more packet
2400  *		is lost (NewReno). This heuristics are the same in NewReno
2401  *		and SACK.
2402  *
2403  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2404  *  deflation etc. CWND is real congestion window, never inflated, changes
2405  *  only according to classic VJ rules.
2406  *
2407  * Really tricky (and requiring careful tuning) part of algorithm
2408  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2409  * The first determines the moment _when_ we should reduce CWND and,
2410  * hence, slow down forward transmission. In fact, it determines the moment
2411  * when we decide that hole is caused by loss, rather than by a reorder.
2412  *
2413  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2414  * holes, caused by lost packets.
2415  *
2416  * And the most logically complicated part of algorithm is undo
2417  * heuristics. We detect false retransmits due to both too early
2418  * fast retransmit (reordering) and underestimated RTO, analyzing
2419  * timestamps and D-SACKs. When we detect that some segments were
2420  * retransmitted by mistake and CWND reduction was wrong, we undo
2421  * window reduction and abort recovery phase. This logic is hidden
2422  * inside several functions named tcp_try_undo_<something>.
2423  */
2424 
2425 /* This function decides, when we should leave Disordered state
2426  * and enter Recovery phase, reducing congestion window.
2427  *
2428  * Main question: may we further continue forward transmission
2429  * with the same cwnd?
2430  */
2431 static int tcp_time_to_recover(struct sock *sk)
2432 {
2433 	struct tcp_sock *tp = tcp_sk(sk);
2434 	__u32 packets_out;
2435 
2436 	/* Do not perform any recovery during F-RTO algorithm */
2437 	if (tp->frto_counter)
2438 		return 0;
2439 
2440 	/* Trick#1: The loss is proven. */
2441 	if (tp->lost_out)
2442 		return 1;
2443 
2444 	/* Not-A-Trick#2 : Classic rule... */
2445 	if (tcp_dupack_heuristics(tp) > tp->reordering)
2446 		return 1;
2447 
2448 	/* Trick#3 : when we use RFC2988 timer restart, fast
2449 	 * retransmit can be triggered by timeout of queue head.
2450 	 */
2451 	if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2452 		return 1;
2453 
2454 	/* Trick#4: It is still not OK... But will it be useful to delay
2455 	 * recovery more?
2456 	 */
2457 	packets_out = tp->packets_out;
2458 	if (packets_out <= tp->reordering &&
2459 	    tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2460 	    !tcp_may_send_now(sk)) {
2461 		/* We have nothing to send. This connection is limited
2462 		 * either by receiver window or by application.
2463 		 */
2464 		return 1;
2465 	}
2466 
2467 	/* If a thin stream is detected, retransmit after first
2468 	 * received dupack. Employ only if SACK is supported in order
2469 	 * to avoid possible corner-case series of spurious retransmissions
2470 	 * Use only if there are no unsent data.
2471 	 */
2472 	if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2473 	    tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2474 	    tcp_is_sack(tp) && !tcp_send_head(sk))
2475 		return 1;
2476 
2477 	return 0;
2478 }
2479 
2480 /* New heuristics: it is possible only after we switched to restart timer
2481  * each time when something is ACKed. Hence, we can detect timed out packets
2482  * during fast retransmit without falling to slow start.
2483  *
2484  * Usefulness of this as is very questionable, since we should know which of
2485  * the segments is the next to timeout which is relatively expensive to find
2486  * in general case unless we add some data structure just for that. The
2487  * current approach certainly won't find the right one too often and when it
2488  * finally does find _something_ it usually marks large part of the window
2489  * right away (because a retransmission with a larger timestamp blocks the
2490  * loop from advancing). -ij
2491  */
2492 static void tcp_timeout_skbs(struct sock *sk)
2493 {
2494 	struct tcp_sock *tp = tcp_sk(sk);
2495 	struct sk_buff *skb;
2496 
2497 	if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2498 		return;
2499 
2500 	skb = tp->scoreboard_skb_hint;
2501 	if (tp->scoreboard_skb_hint == NULL)
2502 		skb = tcp_write_queue_head(sk);
2503 
2504 	tcp_for_write_queue_from(skb, sk) {
2505 		if (skb == tcp_send_head(sk))
2506 			break;
2507 		if (!tcp_skb_timedout(sk, skb))
2508 			break;
2509 
2510 		tcp_skb_mark_lost(tp, skb);
2511 	}
2512 
2513 	tp->scoreboard_skb_hint = skb;
2514 
2515 	tcp_verify_left_out(tp);
2516 }
2517 
2518 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2519  * is against sacked "cnt", otherwise it's against facked "cnt"
2520  */
2521 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2522 {
2523 	struct tcp_sock *tp = tcp_sk(sk);
2524 	struct sk_buff *skb;
2525 	int cnt, oldcnt;
2526 	int err;
2527 	unsigned int mss;
2528 
2529 	WARN_ON(packets > tp->packets_out);
2530 	if (tp->lost_skb_hint) {
2531 		skb = tp->lost_skb_hint;
2532 		cnt = tp->lost_cnt_hint;
2533 		/* Head already handled? */
2534 		if (mark_head && skb != tcp_write_queue_head(sk))
2535 			return;
2536 	} else {
2537 		skb = tcp_write_queue_head(sk);
2538 		cnt = 0;
2539 	}
2540 
2541 	tcp_for_write_queue_from(skb, sk) {
2542 		if (skb == tcp_send_head(sk))
2543 			break;
2544 		/* TODO: do this better */
2545 		/* this is not the most efficient way to do this... */
2546 		tp->lost_skb_hint = skb;
2547 		tp->lost_cnt_hint = cnt;
2548 
2549 		if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2550 			break;
2551 
2552 		oldcnt = cnt;
2553 		if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2554 		    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2555 			cnt += tcp_skb_pcount(skb);
2556 
2557 		if (cnt > packets) {
2558 			if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2559 			    (oldcnt >= packets))
2560 				break;
2561 
2562 			mss = skb_shinfo(skb)->gso_size;
2563 			err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2564 			if (err < 0)
2565 				break;
2566 			cnt = packets;
2567 		}
2568 
2569 		tcp_skb_mark_lost(tp, skb);
2570 
2571 		if (mark_head)
2572 			break;
2573 	}
2574 	tcp_verify_left_out(tp);
2575 }
2576 
2577 /* Account newly detected lost packet(s) */
2578 
2579 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2580 {
2581 	struct tcp_sock *tp = tcp_sk(sk);
2582 
2583 	if (tcp_is_reno(tp)) {
2584 		tcp_mark_head_lost(sk, 1, 1);
2585 	} else if (tcp_is_fack(tp)) {
2586 		int lost = tp->fackets_out - tp->reordering;
2587 		if (lost <= 0)
2588 			lost = 1;
2589 		tcp_mark_head_lost(sk, lost, 0);
2590 	} else {
2591 		int sacked_upto = tp->sacked_out - tp->reordering;
2592 		if (sacked_upto >= 0)
2593 			tcp_mark_head_lost(sk, sacked_upto, 0);
2594 		else if (fast_rexmit)
2595 			tcp_mark_head_lost(sk, 1, 1);
2596 	}
2597 
2598 	tcp_timeout_skbs(sk);
2599 }
2600 
2601 /* CWND moderation, preventing bursts due to too big ACKs
2602  * in dubious situations.
2603  */
2604 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2605 {
2606 	tp->snd_cwnd = min(tp->snd_cwnd,
2607 			   tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2608 	tp->snd_cwnd_stamp = tcp_time_stamp;
2609 }
2610 
2611 /* Lower bound on congestion window is slow start threshold
2612  * unless congestion avoidance choice decides to overide it.
2613  */
2614 static inline u32 tcp_cwnd_min(const struct sock *sk)
2615 {
2616 	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2617 
2618 	return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2619 }
2620 
2621 /* Decrease cwnd each second ack. */
2622 static void tcp_cwnd_down(struct sock *sk, int flag)
2623 {
2624 	struct tcp_sock *tp = tcp_sk(sk);
2625 	int decr = tp->snd_cwnd_cnt + 1;
2626 
2627 	if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2628 	    (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2629 		tp->snd_cwnd_cnt = decr & 1;
2630 		decr >>= 1;
2631 
2632 		if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2633 			tp->snd_cwnd -= decr;
2634 
2635 		tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2636 		tp->snd_cwnd_stamp = tcp_time_stamp;
2637 	}
2638 }
2639 
2640 /* Nothing was retransmitted or returned timestamp is less
2641  * than timestamp of the first retransmission.
2642  */
2643 static inline int tcp_packet_delayed(const struct tcp_sock *tp)
2644 {
2645 	return !tp->retrans_stamp ||
2646 		(tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2647 		 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2648 }
2649 
2650 /* Undo procedures. */
2651 
2652 #if FASTRETRANS_DEBUG > 1
2653 static void DBGUNDO(struct sock *sk, const char *msg)
2654 {
2655 	struct tcp_sock *tp = tcp_sk(sk);
2656 	struct inet_sock *inet = inet_sk(sk);
2657 
2658 	if (sk->sk_family == AF_INET) {
2659 		printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2660 		       msg,
2661 		       &inet->inet_daddr, ntohs(inet->inet_dport),
2662 		       tp->snd_cwnd, tcp_left_out(tp),
2663 		       tp->snd_ssthresh, tp->prior_ssthresh,
2664 		       tp->packets_out);
2665 	}
2666 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2667 	else if (sk->sk_family == AF_INET6) {
2668 		struct ipv6_pinfo *np = inet6_sk(sk);
2669 		printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2670 		       msg,
2671 		       &np->daddr, ntohs(inet->inet_dport),
2672 		       tp->snd_cwnd, tcp_left_out(tp),
2673 		       tp->snd_ssthresh, tp->prior_ssthresh,
2674 		       tp->packets_out);
2675 	}
2676 #endif
2677 }
2678 #else
2679 #define DBGUNDO(x...) do { } while (0)
2680 #endif
2681 
2682 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2683 {
2684 	struct tcp_sock *tp = tcp_sk(sk);
2685 
2686 	if (tp->prior_ssthresh) {
2687 		const struct inet_connection_sock *icsk = inet_csk(sk);
2688 
2689 		if (icsk->icsk_ca_ops->undo_cwnd)
2690 			tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2691 		else
2692 			tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2693 
2694 		if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2695 			tp->snd_ssthresh = tp->prior_ssthresh;
2696 			TCP_ECN_withdraw_cwr(tp);
2697 		}
2698 	} else {
2699 		tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2700 	}
2701 	tp->snd_cwnd_stamp = tcp_time_stamp;
2702 }
2703 
2704 static inline int tcp_may_undo(const struct tcp_sock *tp)
2705 {
2706 	return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2707 }
2708 
2709 /* People celebrate: "We love our President!" */
2710 static int tcp_try_undo_recovery(struct sock *sk)
2711 {
2712 	struct tcp_sock *tp = tcp_sk(sk);
2713 
2714 	if (tcp_may_undo(tp)) {
2715 		int mib_idx;
2716 
2717 		/* Happy end! We did not retransmit anything
2718 		 * or our original transmission succeeded.
2719 		 */
2720 		DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2721 		tcp_undo_cwr(sk, true);
2722 		if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2723 			mib_idx = LINUX_MIB_TCPLOSSUNDO;
2724 		else
2725 			mib_idx = LINUX_MIB_TCPFULLUNDO;
2726 
2727 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
2728 		tp->undo_marker = 0;
2729 	}
2730 	if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2731 		/* Hold old state until something *above* high_seq
2732 		 * is ACKed. For Reno it is MUST to prevent false
2733 		 * fast retransmits (RFC2582). SACK TCP is safe. */
2734 		tcp_moderate_cwnd(tp);
2735 		return 1;
2736 	}
2737 	tcp_set_ca_state(sk, TCP_CA_Open);
2738 	return 0;
2739 }
2740 
2741 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2742 static void tcp_try_undo_dsack(struct sock *sk)
2743 {
2744 	struct tcp_sock *tp = tcp_sk(sk);
2745 
2746 	if (tp->undo_marker && !tp->undo_retrans) {
2747 		DBGUNDO(sk, "D-SACK");
2748 		tcp_undo_cwr(sk, true);
2749 		tp->undo_marker = 0;
2750 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2751 	}
2752 }
2753 
2754 /* We can clear retrans_stamp when there are no retransmissions in the
2755  * window. It would seem that it is trivially available for us in
2756  * tp->retrans_out, however, that kind of assumptions doesn't consider
2757  * what will happen if errors occur when sending retransmission for the
2758  * second time. ...It could the that such segment has only
2759  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2760  * the head skb is enough except for some reneging corner cases that
2761  * are not worth the effort.
2762  *
2763  * Main reason for all this complexity is the fact that connection dying
2764  * time now depends on the validity of the retrans_stamp, in particular,
2765  * that successive retransmissions of a segment must not advance
2766  * retrans_stamp under any conditions.
2767  */
2768 static int tcp_any_retrans_done(const struct sock *sk)
2769 {
2770 	const struct tcp_sock *tp = tcp_sk(sk);
2771 	struct sk_buff *skb;
2772 
2773 	if (tp->retrans_out)
2774 		return 1;
2775 
2776 	skb = tcp_write_queue_head(sk);
2777 	if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2778 		return 1;
2779 
2780 	return 0;
2781 }
2782 
2783 /* Undo during fast recovery after partial ACK. */
2784 
2785 static int tcp_try_undo_partial(struct sock *sk, int acked)
2786 {
2787 	struct tcp_sock *tp = tcp_sk(sk);
2788 	/* Partial ACK arrived. Force Hoe's retransmit. */
2789 	int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2790 
2791 	if (tcp_may_undo(tp)) {
2792 		/* Plain luck! Hole if filled with delayed
2793 		 * packet, rather than with a retransmit.
2794 		 */
2795 		if (!tcp_any_retrans_done(sk))
2796 			tp->retrans_stamp = 0;
2797 
2798 		tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2799 
2800 		DBGUNDO(sk, "Hoe");
2801 		tcp_undo_cwr(sk, false);
2802 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2803 
2804 		/* So... Do not make Hoe's retransmit yet.
2805 		 * If the first packet was delayed, the rest
2806 		 * ones are most probably delayed as well.
2807 		 */
2808 		failed = 0;
2809 	}
2810 	return failed;
2811 }
2812 
2813 /* Undo during loss recovery after partial ACK. */
2814 static int tcp_try_undo_loss(struct sock *sk)
2815 {
2816 	struct tcp_sock *tp = tcp_sk(sk);
2817 
2818 	if (tcp_may_undo(tp)) {
2819 		struct sk_buff *skb;
2820 		tcp_for_write_queue(skb, sk) {
2821 			if (skb == tcp_send_head(sk))
2822 				break;
2823 			TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2824 		}
2825 
2826 		tcp_clear_all_retrans_hints(tp);
2827 
2828 		DBGUNDO(sk, "partial loss");
2829 		tp->lost_out = 0;
2830 		tcp_undo_cwr(sk, true);
2831 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2832 		inet_csk(sk)->icsk_retransmits = 0;
2833 		tp->undo_marker = 0;
2834 		if (tcp_is_sack(tp))
2835 			tcp_set_ca_state(sk, TCP_CA_Open);
2836 		return 1;
2837 	}
2838 	return 0;
2839 }
2840 
2841 static inline void tcp_complete_cwr(struct sock *sk)
2842 {
2843 	struct tcp_sock *tp = tcp_sk(sk);
2844 
2845 	/* Do not moderate cwnd if it's already undone in cwr or recovery. */
2846 	if (tp->undo_marker) {
2847 		if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR)
2848 			tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2849 		else /* PRR */
2850 			tp->snd_cwnd = tp->snd_ssthresh;
2851 		tp->snd_cwnd_stamp = tcp_time_stamp;
2852 	}
2853 	tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2854 }
2855 
2856 static void tcp_try_keep_open(struct sock *sk)
2857 {
2858 	struct tcp_sock *tp = tcp_sk(sk);
2859 	int state = TCP_CA_Open;
2860 
2861 	if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker)
2862 		state = TCP_CA_Disorder;
2863 
2864 	if (inet_csk(sk)->icsk_ca_state != state) {
2865 		tcp_set_ca_state(sk, state);
2866 		tp->high_seq = tp->snd_nxt;
2867 	}
2868 }
2869 
2870 static void tcp_try_to_open(struct sock *sk, int flag)
2871 {
2872 	struct tcp_sock *tp = tcp_sk(sk);
2873 
2874 	tcp_verify_left_out(tp);
2875 
2876 	if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2877 		tp->retrans_stamp = 0;
2878 
2879 	if (flag & FLAG_ECE)
2880 		tcp_enter_cwr(sk, 1);
2881 
2882 	if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2883 		tcp_try_keep_open(sk);
2884 		tcp_moderate_cwnd(tp);
2885 	} else {
2886 		tcp_cwnd_down(sk, flag);
2887 	}
2888 }
2889 
2890 static void tcp_mtup_probe_failed(struct sock *sk)
2891 {
2892 	struct inet_connection_sock *icsk = inet_csk(sk);
2893 
2894 	icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2895 	icsk->icsk_mtup.probe_size = 0;
2896 }
2897 
2898 static void tcp_mtup_probe_success(struct sock *sk)
2899 {
2900 	struct tcp_sock *tp = tcp_sk(sk);
2901 	struct inet_connection_sock *icsk = inet_csk(sk);
2902 
2903 	/* FIXME: breaks with very large cwnd */
2904 	tp->prior_ssthresh = tcp_current_ssthresh(sk);
2905 	tp->snd_cwnd = tp->snd_cwnd *
2906 		       tcp_mss_to_mtu(sk, tp->mss_cache) /
2907 		       icsk->icsk_mtup.probe_size;
2908 	tp->snd_cwnd_cnt = 0;
2909 	tp->snd_cwnd_stamp = tcp_time_stamp;
2910 	tp->snd_ssthresh = tcp_current_ssthresh(sk);
2911 
2912 	icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2913 	icsk->icsk_mtup.probe_size = 0;
2914 	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2915 }
2916 
2917 /* Do a simple retransmit without using the backoff mechanisms in
2918  * tcp_timer. This is used for path mtu discovery.
2919  * The socket is already locked here.
2920  */
2921 void tcp_simple_retransmit(struct sock *sk)
2922 {
2923 	const struct inet_connection_sock *icsk = inet_csk(sk);
2924 	struct tcp_sock *tp = tcp_sk(sk);
2925 	struct sk_buff *skb;
2926 	unsigned int mss = tcp_current_mss(sk);
2927 	u32 prior_lost = tp->lost_out;
2928 
2929 	tcp_for_write_queue(skb, sk) {
2930 		if (skb == tcp_send_head(sk))
2931 			break;
2932 		if (tcp_skb_seglen(skb) > mss &&
2933 		    !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2934 			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2935 				TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2936 				tp->retrans_out -= tcp_skb_pcount(skb);
2937 			}
2938 			tcp_skb_mark_lost_uncond_verify(tp, skb);
2939 		}
2940 	}
2941 
2942 	tcp_clear_retrans_hints_partial(tp);
2943 
2944 	if (prior_lost == tp->lost_out)
2945 		return;
2946 
2947 	if (tcp_is_reno(tp))
2948 		tcp_limit_reno_sacked(tp);
2949 
2950 	tcp_verify_left_out(tp);
2951 
2952 	/* Don't muck with the congestion window here.
2953 	 * Reason is that we do not increase amount of _data_
2954 	 * in network, but units changed and effective
2955 	 * cwnd/ssthresh really reduced now.
2956 	 */
2957 	if (icsk->icsk_ca_state != TCP_CA_Loss) {
2958 		tp->high_seq = tp->snd_nxt;
2959 		tp->snd_ssthresh = tcp_current_ssthresh(sk);
2960 		tp->prior_ssthresh = 0;
2961 		tp->undo_marker = 0;
2962 		tcp_set_ca_state(sk, TCP_CA_Loss);
2963 	}
2964 	tcp_xmit_retransmit_queue(sk);
2965 }
2966 EXPORT_SYMBOL(tcp_simple_retransmit);
2967 
2968 /* This function implements the PRR algorithm, specifcally the PRR-SSRB
2969  * (proportional rate reduction with slow start reduction bound) as described in
2970  * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2971  * It computes the number of packets to send (sndcnt) based on packets newly
2972  * delivered:
2973  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2974  *	cwnd reductions across a full RTT.
2975  *   2) If packets in flight is lower than ssthresh (such as due to excess
2976  *	losses and/or application stalls), do not perform any further cwnd
2977  *	reductions, but instead slow start up to ssthresh.
2978  */
2979 static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
2980 					int fast_rexmit, int flag)
2981 {
2982 	struct tcp_sock *tp = tcp_sk(sk);
2983 	int sndcnt = 0;
2984 	int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2985 
2986 	if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2987 		u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2988 			       tp->prior_cwnd - 1;
2989 		sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2990 	} else {
2991 		sndcnt = min_t(int, delta,
2992 			       max_t(int, tp->prr_delivered - tp->prr_out,
2993 				     newly_acked_sacked) + 1);
2994 	}
2995 
2996 	sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2997 	tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2998 }
2999 
3000 /* Process an event, which can update packets-in-flight not trivially.
3001  * Main goal of this function is to calculate new estimate for left_out,
3002  * taking into account both packets sitting in receiver's buffer and
3003  * packets lost by network.
3004  *
3005  * Besides that it does CWND reduction, when packet loss is detected
3006  * and changes state of machine.
3007  *
3008  * It does _not_ decide what to send, it is made in function
3009  * tcp_xmit_retransmit_queue().
3010  */
3011 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
3012 				  int newly_acked_sacked, int flag)
3013 {
3014 	struct inet_connection_sock *icsk = inet_csk(sk);
3015 	struct tcp_sock *tp = tcp_sk(sk);
3016 	int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3017 	int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
3018 				    (tcp_fackets_out(tp) > tp->reordering));
3019 	int fast_rexmit = 0, mib_idx;
3020 
3021 	if (WARN_ON(!tp->packets_out && tp->sacked_out))
3022 		tp->sacked_out = 0;
3023 	if (WARN_ON(!tp->sacked_out && tp->fackets_out))
3024 		tp->fackets_out = 0;
3025 
3026 	/* Now state machine starts.
3027 	 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3028 	if (flag & FLAG_ECE)
3029 		tp->prior_ssthresh = 0;
3030 
3031 	/* B. In all the states check for reneging SACKs. */
3032 	if (tcp_check_sack_reneging(sk, flag))
3033 		return;
3034 
3035 	/* C. Process data loss notification, provided it is valid. */
3036 	if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
3037 	    before(tp->snd_una, tp->high_seq) &&
3038 	    icsk->icsk_ca_state != TCP_CA_Open &&
3039 	    tp->fackets_out > tp->reordering) {
3040 		tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0);
3041 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
3042 	}
3043 
3044 	/* D. Check consistency of the current state. */
3045 	tcp_verify_left_out(tp);
3046 
3047 	/* E. Check state exit conditions. State can be terminated
3048 	 *    when high_seq is ACKed. */
3049 	if (icsk->icsk_ca_state == TCP_CA_Open) {
3050 		WARN_ON(tp->retrans_out != 0);
3051 		tp->retrans_stamp = 0;
3052 	} else if (!before(tp->snd_una, tp->high_seq)) {
3053 		switch (icsk->icsk_ca_state) {
3054 		case TCP_CA_Loss:
3055 			icsk->icsk_retransmits = 0;
3056 			if (tcp_try_undo_recovery(sk))
3057 				return;
3058 			break;
3059 
3060 		case TCP_CA_CWR:
3061 			/* CWR is to be held something *above* high_seq
3062 			 * is ACKed for CWR bit to reach receiver. */
3063 			if (tp->snd_una != tp->high_seq) {
3064 				tcp_complete_cwr(sk);
3065 				tcp_set_ca_state(sk, TCP_CA_Open);
3066 			}
3067 			break;
3068 
3069 		case TCP_CA_Disorder:
3070 			tcp_try_undo_dsack(sk);
3071 			if (!tp->undo_marker ||
3072 			    /* For SACK case do not Open to allow to undo
3073 			     * catching for all duplicate ACKs. */
3074 			    tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
3075 				tp->undo_marker = 0;
3076 				tcp_set_ca_state(sk, TCP_CA_Open);
3077 			}
3078 			break;
3079 
3080 		case TCP_CA_Recovery:
3081 			if (tcp_is_reno(tp))
3082 				tcp_reset_reno_sack(tp);
3083 			if (tcp_try_undo_recovery(sk))
3084 				return;
3085 			tcp_complete_cwr(sk);
3086 			break;
3087 		}
3088 	}
3089 
3090 	/* F. Process state. */
3091 	switch (icsk->icsk_ca_state) {
3092 	case TCP_CA_Recovery:
3093 		if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3094 			if (tcp_is_reno(tp) && is_dupack)
3095 				tcp_add_reno_sack(sk);
3096 		} else
3097 			do_lost = tcp_try_undo_partial(sk, pkts_acked);
3098 		break;
3099 	case TCP_CA_Loss:
3100 		if (flag & FLAG_DATA_ACKED)
3101 			icsk->icsk_retransmits = 0;
3102 		if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3103 			tcp_reset_reno_sack(tp);
3104 		if (!tcp_try_undo_loss(sk)) {
3105 			tcp_moderate_cwnd(tp);
3106 			tcp_xmit_retransmit_queue(sk);
3107 			return;
3108 		}
3109 		if (icsk->icsk_ca_state != TCP_CA_Open)
3110 			return;
3111 		/* Loss is undone; fall through to processing in Open state. */
3112 	default:
3113 		if (tcp_is_reno(tp)) {
3114 			if (flag & FLAG_SND_UNA_ADVANCED)
3115 				tcp_reset_reno_sack(tp);
3116 			if (is_dupack)
3117 				tcp_add_reno_sack(sk);
3118 		}
3119 
3120 		if (icsk->icsk_ca_state == TCP_CA_Disorder)
3121 			tcp_try_undo_dsack(sk);
3122 
3123 		if (!tcp_time_to_recover(sk)) {
3124 			tcp_try_to_open(sk, flag);
3125 			return;
3126 		}
3127 
3128 		/* MTU probe failure: don't reduce cwnd */
3129 		if (icsk->icsk_ca_state < TCP_CA_CWR &&
3130 		    icsk->icsk_mtup.probe_size &&
3131 		    tp->snd_una == tp->mtu_probe.probe_seq_start) {
3132 			tcp_mtup_probe_failed(sk);
3133 			/* Restores the reduction we did in tcp_mtup_probe() */
3134 			tp->snd_cwnd++;
3135 			tcp_simple_retransmit(sk);
3136 			return;
3137 		}
3138 
3139 		/* Otherwise enter Recovery state */
3140 
3141 		if (tcp_is_reno(tp))
3142 			mib_idx = LINUX_MIB_TCPRENORECOVERY;
3143 		else
3144 			mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3145 
3146 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
3147 
3148 		tp->high_seq = tp->snd_nxt;
3149 		tp->prior_ssthresh = 0;
3150 		tp->undo_marker = tp->snd_una;
3151 		tp->undo_retrans = tp->retrans_out;
3152 
3153 		if (icsk->icsk_ca_state < TCP_CA_CWR) {
3154 			if (!(flag & FLAG_ECE))
3155 				tp->prior_ssthresh = tcp_current_ssthresh(sk);
3156 			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3157 			TCP_ECN_queue_cwr(tp);
3158 		}
3159 
3160 		tp->bytes_acked = 0;
3161 		tp->snd_cwnd_cnt = 0;
3162 		tp->prior_cwnd = tp->snd_cwnd;
3163 		tp->prr_delivered = 0;
3164 		tp->prr_out = 0;
3165 		tcp_set_ca_state(sk, TCP_CA_Recovery);
3166 		fast_rexmit = 1;
3167 	}
3168 
3169 	if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3170 		tcp_update_scoreboard(sk, fast_rexmit);
3171 	tp->prr_delivered += newly_acked_sacked;
3172 	tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3173 	tcp_xmit_retransmit_queue(sk);
3174 }
3175 
3176 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3177 {
3178 	tcp_rtt_estimator(sk, seq_rtt);
3179 	tcp_set_rto(sk);
3180 	inet_csk(sk)->icsk_backoff = 0;
3181 }
3182 EXPORT_SYMBOL(tcp_valid_rtt_meas);
3183 
3184 /* Read draft-ietf-tcplw-high-performance before mucking
3185  * with this code. (Supersedes RFC1323)
3186  */
3187 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3188 {
3189 	/* RTTM Rule: A TSecr value received in a segment is used to
3190 	 * update the averaged RTT measurement only if the segment
3191 	 * acknowledges some new data, i.e., only if it advances the
3192 	 * left edge of the send window.
3193 	 *
3194 	 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3195 	 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3196 	 *
3197 	 * Changed: reset backoff as soon as we see the first valid sample.
3198 	 * If we do not, we get strongly overestimated rto. With timestamps
3199 	 * samples are accepted even from very old segments: f.e., when rtt=1
3200 	 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3201 	 * answer arrives rto becomes 120 seconds! If at least one of segments
3202 	 * in window is lost... Voila.	 			--ANK (010210)
3203 	 */
3204 	struct tcp_sock *tp = tcp_sk(sk);
3205 
3206 	tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3207 }
3208 
3209 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3210 {
3211 	/* We don't have a timestamp. Can only use
3212 	 * packets that are not retransmitted to determine
3213 	 * rtt estimates. Also, we must not reset the
3214 	 * backoff for rto until we get a non-retransmitted
3215 	 * packet. This allows us to deal with a situation
3216 	 * where the network delay has increased suddenly.
3217 	 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3218 	 */
3219 
3220 	if (flag & FLAG_RETRANS_DATA_ACKED)
3221 		return;
3222 
3223 	tcp_valid_rtt_meas(sk, seq_rtt);
3224 }
3225 
3226 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3227 				      const s32 seq_rtt)
3228 {
3229 	const struct tcp_sock *tp = tcp_sk(sk);
3230 	/* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3231 	if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3232 		tcp_ack_saw_tstamp(sk, flag);
3233 	else if (seq_rtt >= 0)
3234 		tcp_ack_no_tstamp(sk, seq_rtt, flag);
3235 }
3236 
3237 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3238 {
3239 	const struct inet_connection_sock *icsk = inet_csk(sk);
3240 	icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3241 	tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3242 }
3243 
3244 /* Restart timer after forward progress on connection.
3245  * RFC2988 recommends to restart timer to now+rto.
3246  */
3247 static void tcp_rearm_rto(struct sock *sk)
3248 {
3249 	const struct tcp_sock *tp = tcp_sk(sk);
3250 
3251 	if (!tp->packets_out) {
3252 		inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3253 	} else {
3254 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3255 					  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3256 	}
3257 }
3258 
3259 /* If we get here, the whole TSO packet has not been acked. */
3260 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3261 {
3262 	struct tcp_sock *tp = tcp_sk(sk);
3263 	u32 packets_acked;
3264 
3265 	BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3266 
3267 	packets_acked = tcp_skb_pcount(skb);
3268 	if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3269 		return 0;
3270 	packets_acked -= tcp_skb_pcount(skb);
3271 
3272 	if (packets_acked) {
3273 		BUG_ON(tcp_skb_pcount(skb) == 0);
3274 		BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3275 	}
3276 
3277 	return packets_acked;
3278 }
3279 
3280 /* Remove acknowledged frames from the retransmission queue. If our packet
3281  * is before the ack sequence we can discard it as it's confirmed to have
3282  * arrived at the other end.
3283  */
3284 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3285 			       u32 prior_snd_una)
3286 {
3287 	struct tcp_sock *tp = tcp_sk(sk);
3288 	const struct inet_connection_sock *icsk = inet_csk(sk);
3289 	struct sk_buff *skb;
3290 	u32 now = tcp_time_stamp;
3291 	int fully_acked = 1;
3292 	int flag = 0;
3293 	u32 pkts_acked = 0;
3294 	u32 reord = tp->packets_out;
3295 	u32 prior_sacked = tp->sacked_out;
3296 	s32 seq_rtt = -1;
3297 	s32 ca_seq_rtt = -1;
3298 	ktime_t last_ackt = net_invalid_timestamp();
3299 
3300 	while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3301 		struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3302 		u32 acked_pcount;
3303 		u8 sacked = scb->sacked;
3304 
3305 		/* Determine how many packets and what bytes were acked, tso and else */
3306 		if (after(scb->end_seq, tp->snd_una)) {
3307 			if (tcp_skb_pcount(skb) == 1 ||
3308 			    !after(tp->snd_una, scb->seq))
3309 				break;
3310 
3311 			acked_pcount = tcp_tso_acked(sk, skb);
3312 			if (!acked_pcount)
3313 				break;
3314 
3315 			fully_acked = 0;
3316 		} else {
3317 			acked_pcount = tcp_skb_pcount(skb);
3318 		}
3319 
3320 		if (sacked & TCPCB_RETRANS) {
3321 			if (sacked & TCPCB_SACKED_RETRANS)
3322 				tp->retrans_out -= acked_pcount;
3323 			flag |= FLAG_RETRANS_DATA_ACKED;
3324 			ca_seq_rtt = -1;
3325 			seq_rtt = -1;
3326 			if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3327 				flag |= FLAG_NONHEAD_RETRANS_ACKED;
3328 		} else {
3329 			ca_seq_rtt = now - scb->when;
3330 			last_ackt = skb->tstamp;
3331 			if (seq_rtt < 0) {
3332 				seq_rtt = ca_seq_rtt;
3333 			}
3334 			if (!(sacked & TCPCB_SACKED_ACKED))
3335 				reord = min(pkts_acked, reord);
3336 		}
3337 
3338 		if (sacked & TCPCB_SACKED_ACKED)
3339 			tp->sacked_out -= acked_pcount;
3340 		if (sacked & TCPCB_LOST)
3341 			tp->lost_out -= acked_pcount;
3342 
3343 		tp->packets_out -= acked_pcount;
3344 		pkts_acked += acked_pcount;
3345 
3346 		/* Initial outgoing SYN's get put onto the write_queue
3347 		 * just like anything else we transmit.  It is not
3348 		 * true data, and if we misinform our callers that
3349 		 * this ACK acks real data, we will erroneously exit
3350 		 * connection startup slow start one packet too
3351 		 * quickly.  This is severely frowned upon behavior.
3352 		 */
3353 		if (!(scb->tcp_flags & TCPHDR_SYN)) {
3354 			flag |= FLAG_DATA_ACKED;
3355 		} else {
3356 			flag |= FLAG_SYN_ACKED;
3357 			tp->retrans_stamp = 0;
3358 		}
3359 
3360 		if (!fully_acked)
3361 			break;
3362 
3363 		tcp_unlink_write_queue(skb, sk);
3364 		sk_wmem_free_skb(sk, skb);
3365 		tp->scoreboard_skb_hint = NULL;
3366 		if (skb == tp->retransmit_skb_hint)
3367 			tp->retransmit_skb_hint = NULL;
3368 		if (skb == tp->lost_skb_hint)
3369 			tp->lost_skb_hint = NULL;
3370 	}
3371 
3372 	if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3373 		tp->snd_up = tp->snd_una;
3374 
3375 	if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3376 		flag |= FLAG_SACK_RENEGING;
3377 
3378 	if (flag & FLAG_ACKED) {
3379 		const struct tcp_congestion_ops *ca_ops
3380 			= inet_csk(sk)->icsk_ca_ops;
3381 
3382 		if (unlikely(icsk->icsk_mtup.probe_size &&
3383 			     !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3384 			tcp_mtup_probe_success(sk);
3385 		}
3386 
3387 		tcp_ack_update_rtt(sk, flag, seq_rtt);
3388 		tcp_rearm_rto(sk);
3389 
3390 		if (tcp_is_reno(tp)) {
3391 			tcp_remove_reno_sacks(sk, pkts_acked);
3392 		} else {
3393 			int delta;
3394 
3395 			/* Non-retransmitted hole got filled? That's reordering */
3396 			if (reord < prior_fackets)
3397 				tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3398 
3399 			delta = tcp_is_fack(tp) ? pkts_acked :
3400 						  prior_sacked - tp->sacked_out;
3401 			tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3402 		}
3403 
3404 		tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3405 
3406 		if (ca_ops->pkts_acked) {
3407 			s32 rtt_us = -1;
3408 
3409 			/* Is the ACK triggering packet unambiguous? */
3410 			if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3411 				/* High resolution needed and available? */
3412 				if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3413 				    !ktime_equal(last_ackt,
3414 						 net_invalid_timestamp()))
3415 					rtt_us = ktime_us_delta(ktime_get_real(),
3416 								last_ackt);
3417 				else if (ca_seq_rtt >= 0)
3418 					rtt_us = jiffies_to_usecs(ca_seq_rtt);
3419 			}
3420 
3421 			ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3422 		}
3423 	}
3424 
3425 #if FASTRETRANS_DEBUG > 0
3426 	WARN_ON((int)tp->sacked_out < 0);
3427 	WARN_ON((int)tp->lost_out < 0);
3428 	WARN_ON((int)tp->retrans_out < 0);
3429 	if (!tp->packets_out && tcp_is_sack(tp)) {
3430 		icsk = inet_csk(sk);
3431 		if (tp->lost_out) {
3432 			printk(KERN_DEBUG "Leak l=%u %d\n",
3433 			       tp->lost_out, icsk->icsk_ca_state);
3434 			tp->lost_out = 0;
3435 		}
3436 		if (tp->sacked_out) {
3437 			printk(KERN_DEBUG "Leak s=%u %d\n",
3438 			       tp->sacked_out, icsk->icsk_ca_state);
3439 			tp->sacked_out = 0;
3440 		}
3441 		if (tp->retrans_out) {
3442 			printk(KERN_DEBUG "Leak r=%u %d\n",
3443 			       tp->retrans_out, icsk->icsk_ca_state);
3444 			tp->retrans_out = 0;
3445 		}
3446 	}
3447 #endif
3448 	return flag;
3449 }
3450 
3451 static void tcp_ack_probe(struct sock *sk)
3452 {
3453 	const struct tcp_sock *tp = tcp_sk(sk);
3454 	struct inet_connection_sock *icsk = inet_csk(sk);
3455 
3456 	/* Was it a usable window open? */
3457 
3458 	if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3459 		icsk->icsk_backoff = 0;
3460 		inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3461 		/* Socket must be waked up by subsequent tcp_data_snd_check().
3462 		 * This function is not for random using!
3463 		 */
3464 	} else {
3465 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3466 					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3467 					  TCP_RTO_MAX);
3468 	}
3469 }
3470 
3471 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3472 {
3473 	return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3474 		inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3475 }
3476 
3477 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3478 {
3479 	const struct tcp_sock *tp = tcp_sk(sk);
3480 	return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3481 		!((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3482 }
3483 
3484 /* Check that window update is acceptable.
3485  * The function assumes that snd_una<=ack<=snd_next.
3486  */
3487 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3488 					const u32 ack, const u32 ack_seq,
3489 					const u32 nwin)
3490 {
3491 	return	after(ack, tp->snd_una) ||
3492 		after(ack_seq, tp->snd_wl1) ||
3493 		(ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3494 }
3495 
3496 /* Update our send window.
3497  *
3498  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3499  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3500  */
3501 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3502 				 u32 ack_seq)
3503 {
3504 	struct tcp_sock *tp = tcp_sk(sk);
3505 	int flag = 0;
3506 	u32 nwin = ntohs(tcp_hdr(skb)->window);
3507 
3508 	if (likely(!tcp_hdr(skb)->syn))
3509 		nwin <<= tp->rx_opt.snd_wscale;
3510 
3511 	if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3512 		flag |= FLAG_WIN_UPDATE;
3513 		tcp_update_wl(tp, ack_seq);
3514 
3515 		if (tp->snd_wnd != nwin) {
3516 			tp->snd_wnd = nwin;
3517 
3518 			/* Note, it is the only place, where
3519 			 * fast path is recovered for sending TCP.
3520 			 */
3521 			tp->pred_flags = 0;
3522 			tcp_fast_path_check(sk);
3523 
3524 			if (nwin > tp->max_window) {
3525 				tp->max_window = nwin;
3526 				tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3527 			}
3528 		}
3529 	}
3530 
3531 	tp->snd_una = ack;
3532 
3533 	return flag;
3534 }
3535 
3536 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3537  * continue in congestion avoidance.
3538  */
3539 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3540 {
3541 	tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3542 	tp->snd_cwnd_cnt = 0;
3543 	tp->bytes_acked = 0;
3544 	TCP_ECN_queue_cwr(tp);
3545 	tcp_moderate_cwnd(tp);
3546 }
3547 
3548 /* A conservative spurious RTO response algorithm: reduce cwnd using
3549  * rate halving and continue in congestion avoidance.
3550  */
3551 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3552 {
3553 	tcp_enter_cwr(sk, 0);
3554 }
3555 
3556 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3557 {
3558 	if (flag & FLAG_ECE)
3559 		tcp_ratehalving_spur_to_response(sk);
3560 	else
3561 		tcp_undo_cwr(sk, true);
3562 }
3563 
3564 /* F-RTO spurious RTO detection algorithm (RFC4138)
3565  *
3566  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3567  * comments). State (ACK number) is kept in frto_counter. When ACK advances
3568  * window (but not to or beyond highest sequence sent before RTO):
3569  *   On First ACK,  send two new segments out.
3570  *   On Second ACK, RTO was likely spurious. Do spurious response (response
3571  *                  algorithm is not part of the F-RTO detection algorithm
3572  *                  given in RFC4138 but can be selected separately).
3573  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3574  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3575  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3576  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3577  *
3578  * Rationale: if the RTO was spurious, new ACKs should arrive from the
3579  * original window even after we transmit two new data segments.
3580  *
3581  * SACK version:
3582  *   on first step, wait until first cumulative ACK arrives, then move to
3583  *   the second step. In second step, the next ACK decides.
3584  *
3585  * F-RTO is implemented (mainly) in four functions:
3586  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3587  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3588  *     called when tcp_use_frto() showed green light
3589  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3590  *   - tcp_enter_frto_loss() is called if there is not enough evidence
3591  *     to prove that the RTO is indeed spurious. It transfers the control
3592  *     from F-RTO to the conventional RTO recovery
3593  */
3594 static int tcp_process_frto(struct sock *sk, int flag)
3595 {
3596 	struct tcp_sock *tp = tcp_sk(sk);
3597 
3598 	tcp_verify_left_out(tp);
3599 
3600 	/* Duplicate the behavior from Loss state (fastretrans_alert) */
3601 	if (flag & FLAG_DATA_ACKED)
3602 		inet_csk(sk)->icsk_retransmits = 0;
3603 
3604 	if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3605 	    ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3606 		tp->undo_marker = 0;
3607 
3608 	if (!before(tp->snd_una, tp->frto_highmark)) {
3609 		tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3610 		return 1;
3611 	}
3612 
3613 	if (!tcp_is_sackfrto(tp)) {
3614 		/* RFC4138 shortcoming in step 2; should also have case c):
3615 		 * ACK isn't duplicate nor advances window, e.g., opposite dir
3616 		 * data, winupdate
3617 		 */
3618 		if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3619 			return 1;
3620 
3621 		if (!(flag & FLAG_DATA_ACKED)) {
3622 			tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3623 					    flag);
3624 			return 1;
3625 		}
3626 	} else {
3627 		if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3628 			/* Prevent sending of new data. */
3629 			tp->snd_cwnd = min(tp->snd_cwnd,
3630 					   tcp_packets_in_flight(tp));
3631 			return 1;
3632 		}
3633 
3634 		if ((tp->frto_counter >= 2) &&
3635 		    (!(flag & FLAG_FORWARD_PROGRESS) ||
3636 		     ((flag & FLAG_DATA_SACKED) &&
3637 		      !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3638 			/* RFC4138 shortcoming (see comment above) */
3639 			if (!(flag & FLAG_FORWARD_PROGRESS) &&
3640 			    (flag & FLAG_NOT_DUP))
3641 				return 1;
3642 
3643 			tcp_enter_frto_loss(sk, 3, flag);
3644 			return 1;
3645 		}
3646 	}
3647 
3648 	if (tp->frto_counter == 1) {
3649 		/* tcp_may_send_now needs to see updated state */
3650 		tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3651 		tp->frto_counter = 2;
3652 
3653 		if (!tcp_may_send_now(sk))
3654 			tcp_enter_frto_loss(sk, 2, flag);
3655 
3656 		return 1;
3657 	} else {
3658 		switch (sysctl_tcp_frto_response) {
3659 		case 2:
3660 			tcp_undo_spur_to_response(sk, flag);
3661 			break;
3662 		case 1:
3663 			tcp_conservative_spur_to_response(tp);
3664 			break;
3665 		default:
3666 			tcp_ratehalving_spur_to_response(sk);
3667 			break;
3668 		}
3669 		tp->frto_counter = 0;
3670 		tp->undo_marker = 0;
3671 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3672 	}
3673 	return 0;
3674 }
3675 
3676 /* This routine deals with incoming acks, but not outgoing ones. */
3677 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3678 {
3679 	struct inet_connection_sock *icsk = inet_csk(sk);
3680 	struct tcp_sock *tp = tcp_sk(sk);
3681 	u32 prior_snd_una = tp->snd_una;
3682 	u32 ack_seq = TCP_SKB_CB(skb)->seq;
3683 	u32 ack = TCP_SKB_CB(skb)->ack_seq;
3684 	u32 prior_in_flight;
3685 	u32 prior_fackets;
3686 	int prior_packets;
3687 	int prior_sacked = tp->sacked_out;
3688 	int newly_acked_sacked = 0;
3689 	int frto_cwnd = 0;
3690 
3691 	/* If the ack is older than previous acks
3692 	 * then we can probably ignore it.
3693 	 */
3694 	if (before(ack, prior_snd_una))
3695 		goto old_ack;
3696 
3697 	/* If the ack includes data we haven't sent yet, discard
3698 	 * this segment (RFC793 Section 3.9).
3699 	 */
3700 	if (after(ack, tp->snd_nxt))
3701 		goto invalid_ack;
3702 
3703 	if (after(ack, prior_snd_una))
3704 		flag |= FLAG_SND_UNA_ADVANCED;
3705 
3706 	if (sysctl_tcp_abc) {
3707 		if (icsk->icsk_ca_state < TCP_CA_CWR)
3708 			tp->bytes_acked += ack - prior_snd_una;
3709 		else if (icsk->icsk_ca_state == TCP_CA_Loss)
3710 			/* we assume just one segment left network */
3711 			tp->bytes_acked += min(ack - prior_snd_una,
3712 					       tp->mss_cache);
3713 	}
3714 
3715 	prior_fackets = tp->fackets_out;
3716 	prior_in_flight = tcp_packets_in_flight(tp);
3717 
3718 	if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3719 		/* Window is constant, pure forward advance.
3720 		 * No more checks are required.
3721 		 * Note, we use the fact that SND.UNA>=SND.WL2.
3722 		 */
3723 		tcp_update_wl(tp, ack_seq);
3724 		tp->snd_una = ack;
3725 		flag |= FLAG_WIN_UPDATE;
3726 
3727 		tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3728 
3729 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3730 	} else {
3731 		if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3732 			flag |= FLAG_DATA;
3733 		else
3734 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3735 
3736 		flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3737 
3738 		if (TCP_SKB_CB(skb)->sacked)
3739 			flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3740 
3741 		if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3742 			flag |= FLAG_ECE;
3743 
3744 		tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3745 	}
3746 
3747 	/* We passed data and got it acked, remove any soft error
3748 	 * log. Something worked...
3749 	 */
3750 	sk->sk_err_soft = 0;
3751 	icsk->icsk_probes_out = 0;
3752 	tp->rcv_tstamp = tcp_time_stamp;
3753 	prior_packets = tp->packets_out;
3754 	if (!prior_packets)
3755 		goto no_queue;
3756 
3757 	/* See if we can take anything off of the retransmit queue. */
3758 	flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3759 
3760 	newly_acked_sacked = (prior_packets - prior_sacked) -
3761 			     (tp->packets_out - tp->sacked_out);
3762 
3763 	if (tp->frto_counter)
3764 		frto_cwnd = tcp_process_frto(sk, flag);
3765 	/* Guarantee sacktag reordering detection against wrap-arounds */
3766 	if (before(tp->frto_highmark, tp->snd_una))
3767 		tp->frto_highmark = 0;
3768 
3769 	if (tcp_ack_is_dubious(sk, flag)) {
3770 		/* Advance CWND, if state allows this. */
3771 		if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3772 		    tcp_may_raise_cwnd(sk, flag))
3773 			tcp_cong_avoid(sk, ack, prior_in_flight);
3774 		tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3775 				      newly_acked_sacked, flag);
3776 	} else {
3777 		if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3778 			tcp_cong_avoid(sk, ack, prior_in_flight);
3779 	}
3780 
3781 	if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3782 		dst_confirm(__sk_dst_get(sk));
3783 
3784 	return 1;
3785 
3786 no_queue:
3787 	/* If this ack opens up a zero window, clear backoff.  It was
3788 	 * being used to time the probes, and is probably far higher than
3789 	 * it needs to be for normal retransmission.
3790 	 */
3791 	if (tcp_send_head(sk))
3792 		tcp_ack_probe(sk);
3793 	return 1;
3794 
3795 invalid_ack:
3796 	SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3797 	return -1;
3798 
3799 old_ack:
3800 	if (TCP_SKB_CB(skb)->sacked) {
3801 		tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3802 		if (icsk->icsk_ca_state == TCP_CA_Open)
3803 			tcp_try_keep_open(sk);
3804 	}
3805 
3806 	SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3807 	return 0;
3808 }
3809 
3810 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3811  * But, this can also be called on packets in the established flow when
3812  * the fast version below fails.
3813  */
3814 void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3815 		       const u8 **hvpp, int estab)
3816 {
3817 	const unsigned char *ptr;
3818 	const struct tcphdr *th = tcp_hdr(skb);
3819 	int length = (th->doff * 4) - sizeof(struct tcphdr);
3820 
3821 	ptr = (const unsigned char *)(th + 1);
3822 	opt_rx->saw_tstamp = 0;
3823 
3824 	while (length > 0) {
3825 		int opcode = *ptr++;
3826 		int opsize;
3827 
3828 		switch (opcode) {
3829 		case TCPOPT_EOL:
3830 			return;
3831 		case TCPOPT_NOP:	/* Ref: RFC 793 section 3.1 */
3832 			length--;
3833 			continue;
3834 		default:
3835 			opsize = *ptr++;
3836 			if (opsize < 2) /* "silly options" */
3837 				return;
3838 			if (opsize > length)
3839 				return;	/* don't parse partial options */
3840 			switch (opcode) {
3841 			case TCPOPT_MSS:
3842 				if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3843 					u16 in_mss = get_unaligned_be16(ptr);
3844 					if (in_mss) {
3845 						if (opt_rx->user_mss &&
3846 						    opt_rx->user_mss < in_mss)
3847 							in_mss = opt_rx->user_mss;
3848 						opt_rx->mss_clamp = in_mss;
3849 					}
3850 				}
3851 				break;
3852 			case TCPOPT_WINDOW:
3853 				if (opsize == TCPOLEN_WINDOW && th->syn &&
3854 				    !estab && sysctl_tcp_window_scaling) {
3855 					__u8 snd_wscale = *(__u8 *)ptr;
3856 					opt_rx->wscale_ok = 1;
3857 					if (snd_wscale > 14) {
3858 						if (net_ratelimit())
3859 							printk(KERN_INFO "tcp_parse_options: Illegal window "
3860 							       "scaling value %d >14 received.\n",
3861 							       snd_wscale);
3862 						snd_wscale = 14;
3863 					}
3864 					opt_rx->snd_wscale = snd_wscale;
3865 				}
3866 				break;
3867 			case TCPOPT_TIMESTAMP:
3868 				if ((opsize == TCPOLEN_TIMESTAMP) &&
3869 				    ((estab && opt_rx->tstamp_ok) ||
3870 				     (!estab && sysctl_tcp_timestamps))) {
3871 					opt_rx->saw_tstamp = 1;
3872 					opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3873 					opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3874 				}
3875 				break;
3876 			case TCPOPT_SACK_PERM:
3877 				if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3878 				    !estab && sysctl_tcp_sack) {
3879 					opt_rx->sack_ok = 1;
3880 					tcp_sack_reset(opt_rx);
3881 				}
3882 				break;
3883 
3884 			case TCPOPT_SACK:
3885 				if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3886 				   !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3887 				   opt_rx->sack_ok) {
3888 					TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3889 				}
3890 				break;
3891 #ifdef CONFIG_TCP_MD5SIG
3892 			case TCPOPT_MD5SIG:
3893 				/*
3894 				 * The MD5 Hash has already been
3895 				 * checked (see tcp_v{4,6}_do_rcv()).
3896 				 */
3897 				break;
3898 #endif
3899 			case TCPOPT_COOKIE:
3900 				/* This option is variable length.
3901 				 */
3902 				switch (opsize) {
3903 				case TCPOLEN_COOKIE_BASE:
3904 					/* not yet implemented */
3905 					break;
3906 				case TCPOLEN_COOKIE_PAIR:
3907 					/* not yet implemented */
3908 					break;
3909 				case TCPOLEN_COOKIE_MIN+0:
3910 				case TCPOLEN_COOKIE_MIN+2:
3911 				case TCPOLEN_COOKIE_MIN+4:
3912 				case TCPOLEN_COOKIE_MIN+6:
3913 				case TCPOLEN_COOKIE_MAX:
3914 					/* 16-bit multiple */
3915 					opt_rx->cookie_plus = opsize;
3916 					*hvpp = ptr;
3917 					break;
3918 				default:
3919 					/* ignore option */
3920 					break;
3921 				}
3922 				break;
3923 			}
3924 
3925 			ptr += opsize-2;
3926 			length -= opsize;
3927 		}
3928 	}
3929 }
3930 EXPORT_SYMBOL(tcp_parse_options);
3931 
3932 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3933 {
3934 	const __be32 *ptr = (const __be32 *)(th + 1);
3935 
3936 	if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3937 			  | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3938 		tp->rx_opt.saw_tstamp = 1;
3939 		++ptr;
3940 		tp->rx_opt.rcv_tsval = ntohl(*ptr);
3941 		++ptr;
3942 		tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3943 		return 1;
3944 	}
3945 	return 0;
3946 }
3947 
3948 /* Fast parse options. This hopes to only see timestamps.
3949  * If it is wrong it falls back on tcp_parse_options().
3950  */
3951 static int tcp_fast_parse_options(const struct sk_buff *skb,
3952 				  const struct tcphdr *th,
3953 				  struct tcp_sock *tp, const u8 **hvpp)
3954 {
3955 	/* In the spirit of fast parsing, compare doff directly to constant
3956 	 * values.  Because equality is used, short doff can be ignored here.
3957 	 */
3958 	if (th->doff == (sizeof(*th) / 4)) {
3959 		tp->rx_opt.saw_tstamp = 0;
3960 		return 0;
3961 	} else if (tp->rx_opt.tstamp_ok &&
3962 		   th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3963 		if (tcp_parse_aligned_timestamp(tp, th))
3964 			return 1;
3965 	}
3966 	tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3967 	return 1;
3968 }
3969 
3970 #ifdef CONFIG_TCP_MD5SIG
3971 /*
3972  * Parse MD5 Signature option
3973  */
3974 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3975 {
3976 	int length = (th->doff << 2) - sizeof(*th);
3977 	const u8 *ptr = (const u8 *)(th + 1);
3978 
3979 	/* If the TCP option is too short, we can short cut */
3980 	if (length < TCPOLEN_MD5SIG)
3981 		return NULL;
3982 
3983 	while (length > 0) {
3984 		int opcode = *ptr++;
3985 		int opsize;
3986 
3987 		switch(opcode) {
3988 		case TCPOPT_EOL:
3989 			return NULL;
3990 		case TCPOPT_NOP:
3991 			length--;
3992 			continue;
3993 		default:
3994 			opsize = *ptr++;
3995 			if (opsize < 2 || opsize > length)
3996 				return NULL;
3997 			if (opcode == TCPOPT_MD5SIG)
3998 				return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3999 		}
4000 		ptr += opsize - 2;
4001 		length -= opsize;
4002 	}
4003 	return NULL;
4004 }
4005 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4006 #endif
4007 
4008 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
4009 {
4010 	tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
4011 	tp->rx_opt.ts_recent_stamp = get_seconds();
4012 }
4013 
4014 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
4015 {
4016 	if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
4017 		/* PAWS bug workaround wrt. ACK frames, the PAWS discard
4018 		 * extra check below makes sure this can only happen
4019 		 * for pure ACK frames.  -DaveM
4020 		 *
4021 		 * Not only, also it occurs for expired timestamps.
4022 		 */
4023 
4024 		if (tcp_paws_check(&tp->rx_opt, 0))
4025 			tcp_store_ts_recent(tp);
4026 	}
4027 }
4028 
4029 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4030  *
4031  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4032  * it can pass through stack. So, the following predicate verifies that
4033  * this segment is not used for anything but congestion avoidance or
4034  * fast retransmit. Moreover, we even are able to eliminate most of such
4035  * second order effects, if we apply some small "replay" window (~RTO)
4036  * to timestamp space.
4037  *
4038  * All these measures still do not guarantee that we reject wrapped ACKs
4039  * on networks with high bandwidth, when sequence space is recycled fastly,
4040  * but it guarantees that such events will be very rare and do not affect
4041  * connection seriously. This doesn't look nice, but alas, PAWS is really
4042  * buggy extension.
4043  *
4044  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4045  * states that events when retransmit arrives after original data are rare.
4046  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4047  * the biggest problem on large power networks even with minor reordering.
4048  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4049  * up to bandwidth of 18Gigabit/sec. 8) ]
4050  */
4051 
4052 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4053 {
4054 	const struct tcp_sock *tp = tcp_sk(sk);
4055 	const struct tcphdr *th = tcp_hdr(skb);
4056 	u32 seq = TCP_SKB_CB(skb)->seq;
4057 	u32 ack = TCP_SKB_CB(skb)->ack_seq;
4058 
4059 	return (/* 1. Pure ACK with correct sequence number. */
4060 		(th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4061 
4062 		/* 2. ... and duplicate ACK. */
4063 		ack == tp->snd_una &&
4064 
4065 		/* 3. ... and does not update window. */
4066 		!tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4067 
4068 		/* 4. ... and sits in replay window. */
4069 		(s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4070 }
4071 
4072 static inline int tcp_paws_discard(const struct sock *sk,
4073 				   const struct sk_buff *skb)
4074 {
4075 	const struct tcp_sock *tp = tcp_sk(sk);
4076 
4077 	return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4078 	       !tcp_disordered_ack(sk, skb);
4079 }
4080 
4081 /* Check segment sequence number for validity.
4082  *
4083  * Segment controls are considered valid, if the segment
4084  * fits to the window after truncation to the window. Acceptability
4085  * of data (and SYN, FIN, of course) is checked separately.
4086  * See tcp_data_queue(), for example.
4087  *
4088  * Also, controls (RST is main one) are accepted using RCV.WUP instead
4089  * of RCV.NXT. Peer still did not advance his SND.UNA when we
4090  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4091  * (borrowed from freebsd)
4092  */
4093 
4094 static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4095 {
4096 	return	!before(end_seq, tp->rcv_wup) &&
4097 		!after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4098 }
4099 
4100 /* When we get a reset we do this. */
4101 static void tcp_reset(struct sock *sk)
4102 {
4103 	/* We want the right error as BSD sees it (and indeed as we do). */
4104 	switch (sk->sk_state) {
4105 	case TCP_SYN_SENT:
4106 		sk->sk_err = ECONNREFUSED;
4107 		break;
4108 	case TCP_CLOSE_WAIT:
4109 		sk->sk_err = EPIPE;
4110 		break;
4111 	case TCP_CLOSE:
4112 		return;
4113 	default:
4114 		sk->sk_err = ECONNRESET;
4115 	}
4116 	/* This barrier is coupled with smp_rmb() in tcp_poll() */
4117 	smp_wmb();
4118 
4119 	if (!sock_flag(sk, SOCK_DEAD))
4120 		sk->sk_error_report(sk);
4121 
4122 	tcp_done(sk);
4123 }
4124 
4125 /*
4126  * 	Process the FIN bit. This now behaves as it is supposed to work
4127  *	and the FIN takes effect when it is validly part of sequence
4128  *	space. Not before when we get holes.
4129  *
4130  *	If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4131  *	(and thence onto LAST-ACK and finally, CLOSE, we never enter
4132  *	TIME-WAIT)
4133  *
4134  *	If we are in FINWAIT-1, a received FIN indicates simultaneous
4135  *	close and we go into CLOSING (and later onto TIME-WAIT)
4136  *
4137  *	If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4138  */
4139 static void tcp_fin(struct sock *sk)
4140 {
4141 	struct tcp_sock *tp = tcp_sk(sk);
4142 
4143 	inet_csk_schedule_ack(sk);
4144 
4145 	sk->sk_shutdown |= RCV_SHUTDOWN;
4146 	sock_set_flag(sk, SOCK_DONE);
4147 
4148 	switch (sk->sk_state) {
4149 	case TCP_SYN_RECV:
4150 	case TCP_ESTABLISHED:
4151 		/* Move to CLOSE_WAIT */
4152 		tcp_set_state(sk, TCP_CLOSE_WAIT);
4153 		inet_csk(sk)->icsk_ack.pingpong = 1;
4154 		break;
4155 
4156 	case TCP_CLOSE_WAIT:
4157 	case TCP_CLOSING:
4158 		/* Received a retransmission of the FIN, do
4159 		 * nothing.
4160 		 */
4161 		break;
4162 	case TCP_LAST_ACK:
4163 		/* RFC793: Remain in the LAST-ACK state. */
4164 		break;
4165 
4166 	case TCP_FIN_WAIT1:
4167 		/* This case occurs when a simultaneous close
4168 		 * happens, we must ack the received FIN and
4169 		 * enter the CLOSING state.
4170 		 */
4171 		tcp_send_ack(sk);
4172 		tcp_set_state(sk, TCP_CLOSING);
4173 		break;
4174 	case TCP_FIN_WAIT2:
4175 		/* Received a FIN -- send ACK and enter TIME_WAIT. */
4176 		tcp_send_ack(sk);
4177 		tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4178 		break;
4179 	default:
4180 		/* Only TCP_LISTEN and TCP_CLOSE are left, in these
4181 		 * cases we should never reach this piece of code.
4182 		 */
4183 		printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4184 		       __func__, sk->sk_state);
4185 		break;
4186 	}
4187 
4188 	/* It _is_ possible, that we have something out-of-order _after_ FIN.
4189 	 * Probably, we should reset in this case. For now drop them.
4190 	 */
4191 	__skb_queue_purge(&tp->out_of_order_queue);
4192 	if (tcp_is_sack(tp))
4193 		tcp_sack_reset(&tp->rx_opt);
4194 	sk_mem_reclaim(sk);
4195 
4196 	if (!sock_flag(sk, SOCK_DEAD)) {
4197 		sk->sk_state_change(sk);
4198 
4199 		/* Do not send POLL_HUP for half duplex close. */
4200 		if (sk->sk_shutdown == SHUTDOWN_MASK ||
4201 		    sk->sk_state == TCP_CLOSE)
4202 			sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4203 		else
4204 			sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4205 	}
4206 }
4207 
4208 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4209 				  u32 end_seq)
4210 {
4211 	if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4212 		if (before(seq, sp->start_seq))
4213 			sp->start_seq = seq;
4214 		if (after(end_seq, sp->end_seq))
4215 			sp->end_seq = end_seq;
4216 		return 1;
4217 	}
4218 	return 0;
4219 }
4220 
4221 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4222 {
4223 	struct tcp_sock *tp = tcp_sk(sk);
4224 
4225 	if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4226 		int mib_idx;
4227 
4228 		if (before(seq, tp->rcv_nxt))
4229 			mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4230 		else
4231 			mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4232 
4233 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
4234 
4235 		tp->rx_opt.dsack = 1;
4236 		tp->duplicate_sack[0].start_seq = seq;
4237 		tp->duplicate_sack[0].end_seq = end_seq;
4238 	}
4239 }
4240 
4241 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4242 {
4243 	struct tcp_sock *tp = tcp_sk(sk);
4244 
4245 	if (!tp->rx_opt.dsack)
4246 		tcp_dsack_set(sk, seq, end_seq);
4247 	else
4248 		tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4249 }
4250 
4251 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4252 {
4253 	struct tcp_sock *tp = tcp_sk(sk);
4254 
4255 	if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4256 	    before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4257 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4258 		tcp_enter_quickack_mode(sk);
4259 
4260 		if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4261 			u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4262 
4263 			if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4264 				end_seq = tp->rcv_nxt;
4265 			tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4266 		}
4267 	}
4268 
4269 	tcp_send_ack(sk);
4270 }
4271 
4272 /* These routines update the SACK block as out-of-order packets arrive or
4273  * in-order packets close up the sequence space.
4274  */
4275 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4276 {
4277 	int this_sack;
4278 	struct tcp_sack_block *sp = &tp->selective_acks[0];
4279 	struct tcp_sack_block *swalk = sp + 1;
4280 
4281 	/* See if the recent change to the first SACK eats into
4282 	 * or hits the sequence space of other SACK blocks, if so coalesce.
4283 	 */
4284 	for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4285 		if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4286 			int i;
4287 
4288 			/* Zap SWALK, by moving every further SACK up by one slot.
4289 			 * Decrease num_sacks.
4290 			 */
4291 			tp->rx_opt.num_sacks--;
4292 			for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4293 				sp[i] = sp[i + 1];
4294 			continue;
4295 		}
4296 		this_sack++, swalk++;
4297 	}
4298 }
4299 
4300 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4301 {
4302 	struct tcp_sock *tp = tcp_sk(sk);
4303 	struct tcp_sack_block *sp = &tp->selective_acks[0];
4304 	int cur_sacks = tp->rx_opt.num_sacks;
4305 	int this_sack;
4306 
4307 	if (!cur_sacks)
4308 		goto new_sack;
4309 
4310 	for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4311 		if (tcp_sack_extend(sp, seq, end_seq)) {
4312 			/* Rotate this_sack to the first one. */
4313 			for (; this_sack > 0; this_sack--, sp--)
4314 				swap(*sp, *(sp - 1));
4315 			if (cur_sacks > 1)
4316 				tcp_sack_maybe_coalesce(tp);
4317 			return;
4318 		}
4319 	}
4320 
4321 	/* Could not find an adjacent existing SACK, build a new one,
4322 	 * put it at the front, and shift everyone else down.  We
4323 	 * always know there is at least one SACK present already here.
4324 	 *
4325 	 * If the sack array is full, forget about the last one.
4326 	 */
4327 	if (this_sack >= TCP_NUM_SACKS) {
4328 		this_sack--;
4329 		tp->rx_opt.num_sacks--;
4330 		sp--;
4331 	}
4332 	for (; this_sack > 0; this_sack--, sp--)
4333 		*sp = *(sp - 1);
4334 
4335 new_sack:
4336 	/* Build the new head SACK, and we're done. */
4337 	sp->start_seq = seq;
4338 	sp->end_seq = end_seq;
4339 	tp->rx_opt.num_sacks++;
4340 }
4341 
4342 /* RCV.NXT advances, some SACKs should be eaten. */
4343 
4344 static void tcp_sack_remove(struct tcp_sock *tp)
4345 {
4346 	struct tcp_sack_block *sp = &tp->selective_acks[0];
4347 	int num_sacks = tp->rx_opt.num_sacks;
4348 	int this_sack;
4349 
4350 	/* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4351 	if (skb_queue_empty(&tp->out_of_order_queue)) {
4352 		tp->rx_opt.num_sacks = 0;
4353 		return;
4354 	}
4355 
4356 	for (this_sack = 0; this_sack < num_sacks;) {
4357 		/* Check if the start of the sack is covered by RCV.NXT. */
4358 		if (!before(tp->rcv_nxt, sp->start_seq)) {
4359 			int i;
4360 
4361 			/* RCV.NXT must cover all the block! */
4362 			WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4363 
4364 			/* Zap this SACK, by moving forward any other SACKS. */
4365 			for (i=this_sack+1; i < num_sacks; i++)
4366 				tp->selective_acks[i-1] = tp->selective_acks[i];
4367 			num_sacks--;
4368 			continue;
4369 		}
4370 		this_sack++;
4371 		sp++;
4372 	}
4373 	tp->rx_opt.num_sacks = num_sacks;
4374 }
4375 
4376 /* This one checks to see if we can put data from the
4377  * out_of_order queue into the receive_queue.
4378  */
4379 static void tcp_ofo_queue(struct sock *sk)
4380 {
4381 	struct tcp_sock *tp = tcp_sk(sk);
4382 	__u32 dsack_high = tp->rcv_nxt;
4383 	struct sk_buff *skb;
4384 
4385 	while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4386 		if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4387 			break;
4388 
4389 		if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4390 			__u32 dsack = dsack_high;
4391 			if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4392 				dsack_high = TCP_SKB_CB(skb)->end_seq;
4393 			tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4394 		}
4395 
4396 		if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4397 			SOCK_DEBUG(sk, "ofo packet was already received\n");
4398 			__skb_unlink(skb, &tp->out_of_order_queue);
4399 			__kfree_skb(skb);
4400 			continue;
4401 		}
4402 		SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4403 			   tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4404 			   TCP_SKB_CB(skb)->end_seq);
4405 
4406 		__skb_unlink(skb, &tp->out_of_order_queue);
4407 		__skb_queue_tail(&sk->sk_receive_queue, skb);
4408 		tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4409 		if (tcp_hdr(skb)->fin)
4410 			tcp_fin(sk);
4411 	}
4412 }
4413 
4414 static int tcp_prune_ofo_queue(struct sock *sk);
4415 static int tcp_prune_queue(struct sock *sk);
4416 
4417 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4418 {
4419 	if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4420 	    !sk_rmem_schedule(sk, size)) {
4421 
4422 		if (tcp_prune_queue(sk) < 0)
4423 			return -1;
4424 
4425 		if (!sk_rmem_schedule(sk, size)) {
4426 			if (!tcp_prune_ofo_queue(sk))
4427 				return -1;
4428 
4429 			if (!sk_rmem_schedule(sk, size))
4430 				return -1;
4431 		}
4432 	}
4433 	return 0;
4434 }
4435 
4436 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4437 {
4438 	const struct tcphdr *th = tcp_hdr(skb);
4439 	struct tcp_sock *tp = tcp_sk(sk);
4440 	int eaten = -1;
4441 
4442 	if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4443 		goto drop;
4444 
4445 	skb_dst_drop(skb);
4446 	__skb_pull(skb, th->doff * 4);
4447 
4448 	TCP_ECN_accept_cwr(tp, skb);
4449 
4450 	tp->rx_opt.dsack = 0;
4451 
4452 	/*  Queue data for delivery to the user.
4453 	 *  Packets in sequence go to the receive queue.
4454 	 *  Out of sequence packets to the out_of_order_queue.
4455 	 */
4456 	if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4457 		if (tcp_receive_window(tp) == 0)
4458 			goto out_of_window;
4459 
4460 		/* Ok. In sequence. In window. */
4461 		if (tp->ucopy.task == current &&
4462 		    tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4463 		    sock_owned_by_user(sk) && !tp->urg_data) {
4464 			int chunk = min_t(unsigned int, skb->len,
4465 					  tp->ucopy.len);
4466 
4467 			__set_current_state(TASK_RUNNING);
4468 
4469 			local_bh_enable();
4470 			if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4471 				tp->ucopy.len -= chunk;
4472 				tp->copied_seq += chunk;
4473 				eaten = (chunk == skb->len);
4474 				tcp_rcv_space_adjust(sk);
4475 			}
4476 			local_bh_disable();
4477 		}
4478 
4479 		if (eaten <= 0) {
4480 queue_and_out:
4481 			if (eaten < 0 &&
4482 			    tcp_try_rmem_schedule(sk, skb->truesize))
4483 				goto drop;
4484 
4485 			skb_set_owner_r(skb, sk);
4486 			__skb_queue_tail(&sk->sk_receive_queue, skb);
4487 		}
4488 		tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4489 		if (skb->len)
4490 			tcp_event_data_recv(sk, skb);
4491 		if (th->fin)
4492 			tcp_fin(sk);
4493 
4494 		if (!skb_queue_empty(&tp->out_of_order_queue)) {
4495 			tcp_ofo_queue(sk);
4496 
4497 			/* RFC2581. 4.2. SHOULD send immediate ACK, when
4498 			 * gap in queue is filled.
4499 			 */
4500 			if (skb_queue_empty(&tp->out_of_order_queue))
4501 				inet_csk(sk)->icsk_ack.pingpong = 0;
4502 		}
4503 
4504 		if (tp->rx_opt.num_sacks)
4505 			tcp_sack_remove(tp);
4506 
4507 		tcp_fast_path_check(sk);
4508 
4509 		if (eaten > 0)
4510 			__kfree_skb(skb);
4511 		else if (!sock_flag(sk, SOCK_DEAD))
4512 			sk->sk_data_ready(sk, 0);
4513 		return;
4514 	}
4515 
4516 	if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4517 		/* A retransmit, 2nd most common case.  Force an immediate ack. */
4518 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4519 		tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4520 
4521 out_of_window:
4522 		tcp_enter_quickack_mode(sk);
4523 		inet_csk_schedule_ack(sk);
4524 drop:
4525 		__kfree_skb(skb);
4526 		return;
4527 	}
4528 
4529 	/* Out of window. F.e. zero window probe. */
4530 	if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4531 		goto out_of_window;
4532 
4533 	tcp_enter_quickack_mode(sk);
4534 
4535 	if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4536 		/* Partial packet, seq < rcv_next < end_seq */
4537 		SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4538 			   tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4539 			   TCP_SKB_CB(skb)->end_seq);
4540 
4541 		tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4542 
4543 		/* If window is closed, drop tail of packet. But after
4544 		 * remembering D-SACK for its head made in previous line.
4545 		 */
4546 		if (!tcp_receive_window(tp))
4547 			goto out_of_window;
4548 		goto queue_and_out;
4549 	}
4550 
4551 	TCP_ECN_check_ce(tp, skb);
4552 
4553 	if (tcp_try_rmem_schedule(sk, skb->truesize))
4554 		goto drop;
4555 
4556 	/* Disable header prediction. */
4557 	tp->pred_flags = 0;
4558 	inet_csk_schedule_ack(sk);
4559 
4560 	SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4561 		   tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4562 
4563 	skb_set_owner_r(skb, sk);
4564 
4565 	if (!skb_peek(&tp->out_of_order_queue)) {
4566 		/* Initial out of order segment, build 1 SACK. */
4567 		if (tcp_is_sack(tp)) {
4568 			tp->rx_opt.num_sacks = 1;
4569 			tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4570 			tp->selective_acks[0].end_seq =
4571 						TCP_SKB_CB(skb)->end_seq;
4572 		}
4573 		__skb_queue_head(&tp->out_of_order_queue, skb);
4574 	} else {
4575 		struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue);
4576 		u32 seq = TCP_SKB_CB(skb)->seq;
4577 		u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4578 
4579 		if (seq == TCP_SKB_CB(skb1)->end_seq) {
4580 			__skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4581 
4582 			if (!tp->rx_opt.num_sacks ||
4583 			    tp->selective_acks[0].end_seq != seq)
4584 				goto add_sack;
4585 
4586 			/* Common case: data arrive in order after hole. */
4587 			tp->selective_acks[0].end_seq = end_seq;
4588 			return;
4589 		}
4590 
4591 		/* Find place to insert this segment. */
4592 		while (1) {
4593 			if (!after(TCP_SKB_CB(skb1)->seq, seq))
4594 				break;
4595 			if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4596 				skb1 = NULL;
4597 				break;
4598 			}
4599 			skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4600 		}
4601 
4602 		/* Do skb overlap to previous one? */
4603 		if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4604 			if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4605 				/* All the bits are present. Drop. */
4606 				__kfree_skb(skb);
4607 				tcp_dsack_set(sk, seq, end_seq);
4608 				goto add_sack;
4609 			}
4610 			if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4611 				/* Partial overlap. */
4612 				tcp_dsack_set(sk, seq,
4613 					      TCP_SKB_CB(skb1)->end_seq);
4614 			} else {
4615 				if (skb_queue_is_first(&tp->out_of_order_queue,
4616 						       skb1))
4617 					skb1 = NULL;
4618 				else
4619 					skb1 = skb_queue_prev(
4620 						&tp->out_of_order_queue,
4621 						skb1);
4622 			}
4623 		}
4624 		if (!skb1)
4625 			__skb_queue_head(&tp->out_of_order_queue, skb);
4626 		else
4627 			__skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4628 
4629 		/* And clean segments covered by new one as whole. */
4630 		while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4631 			skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4632 
4633 			if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4634 				break;
4635 			if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4636 				tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4637 						 end_seq);
4638 				break;
4639 			}
4640 			__skb_unlink(skb1, &tp->out_of_order_queue);
4641 			tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4642 					 TCP_SKB_CB(skb1)->end_seq);
4643 			__kfree_skb(skb1);
4644 		}
4645 
4646 add_sack:
4647 		if (tcp_is_sack(tp))
4648 			tcp_sack_new_ofo_skb(sk, seq, end_seq);
4649 	}
4650 }
4651 
4652 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4653 					struct sk_buff_head *list)
4654 {
4655 	struct sk_buff *next = NULL;
4656 
4657 	if (!skb_queue_is_last(list, skb))
4658 		next = skb_queue_next(list, skb);
4659 
4660 	__skb_unlink(skb, list);
4661 	__kfree_skb(skb);
4662 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4663 
4664 	return next;
4665 }
4666 
4667 /* Collapse contiguous sequence of skbs head..tail with
4668  * sequence numbers start..end.
4669  *
4670  * If tail is NULL, this means until the end of the list.
4671  *
4672  * Segments with FIN/SYN are not collapsed (only because this
4673  * simplifies code)
4674  */
4675 static void
4676 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4677 	     struct sk_buff *head, struct sk_buff *tail,
4678 	     u32 start, u32 end)
4679 {
4680 	struct sk_buff *skb, *n;
4681 	bool end_of_skbs;
4682 
4683 	/* First, check that queue is collapsible and find
4684 	 * the point where collapsing can be useful. */
4685 	skb = head;
4686 restart:
4687 	end_of_skbs = true;
4688 	skb_queue_walk_from_safe(list, skb, n) {
4689 		if (skb == tail)
4690 			break;
4691 		/* No new bits? It is possible on ofo queue. */
4692 		if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4693 			skb = tcp_collapse_one(sk, skb, list);
4694 			if (!skb)
4695 				break;
4696 			goto restart;
4697 		}
4698 
4699 		/* The first skb to collapse is:
4700 		 * - not SYN/FIN and
4701 		 * - bloated or contains data before "start" or
4702 		 *   overlaps to the next one.
4703 		 */
4704 		if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4705 		    (tcp_win_from_space(skb->truesize) > skb->len ||
4706 		     before(TCP_SKB_CB(skb)->seq, start))) {
4707 			end_of_skbs = false;
4708 			break;
4709 		}
4710 
4711 		if (!skb_queue_is_last(list, skb)) {
4712 			struct sk_buff *next = skb_queue_next(list, skb);
4713 			if (next != tail &&
4714 			    TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4715 				end_of_skbs = false;
4716 				break;
4717 			}
4718 		}
4719 
4720 		/* Decided to skip this, advance start seq. */
4721 		start = TCP_SKB_CB(skb)->end_seq;
4722 	}
4723 	if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4724 		return;
4725 
4726 	while (before(start, end)) {
4727 		struct sk_buff *nskb;
4728 		unsigned int header = skb_headroom(skb);
4729 		int copy = SKB_MAX_ORDER(header, 0);
4730 
4731 		/* Too big header? This can happen with IPv6. */
4732 		if (copy < 0)
4733 			return;
4734 		if (end - start < copy)
4735 			copy = end - start;
4736 		nskb = alloc_skb(copy + header, GFP_ATOMIC);
4737 		if (!nskb)
4738 			return;
4739 
4740 		skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4741 		skb_set_network_header(nskb, (skb_network_header(skb) -
4742 					      skb->head));
4743 		skb_set_transport_header(nskb, (skb_transport_header(skb) -
4744 						skb->head));
4745 		skb_reserve(nskb, header);
4746 		memcpy(nskb->head, skb->head, header);
4747 		memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4748 		TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4749 		__skb_queue_before(list, skb, nskb);
4750 		skb_set_owner_r(nskb, sk);
4751 
4752 		/* Copy data, releasing collapsed skbs. */
4753 		while (copy > 0) {
4754 			int offset = start - TCP_SKB_CB(skb)->seq;
4755 			int size = TCP_SKB_CB(skb)->end_seq - start;
4756 
4757 			BUG_ON(offset < 0);
4758 			if (size > 0) {
4759 				size = min(copy, size);
4760 				if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4761 					BUG();
4762 				TCP_SKB_CB(nskb)->end_seq += size;
4763 				copy -= size;
4764 				start += size;
4765 			}
4766 			if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4767 				skb = tcp_collapse_one(sk, skb, list);
4768 				if (!skb ||
4769 				    skb == tail ||
4770 				    tcp_hdr(skb)->syn ||
4771 				    tcp_hdr(skb)->fin)
4772 					return;
4773 			}
4774 		}
4775 	}
4776 }
4777 
4778 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4779  * and tcp_collapse() them until all the queue is collapsed.
4780  */
4781 static void tcp_collapse_ofo_queue(struct sock *sk)
4782 {
4783 	struct tcp_sock *tp = tcp_sk(sk);
4784 	struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4785 	struct sk_buff *head;
4786 	u32 start, end;
4787 
4788 	if (skb == NULL)
4789 		return;
4790 
4791 	start = TCP_SKB_CB(skb)->seq;
4792 	end = TCP_SKB_CB(skb)->end_seq;
4793 	head = skb;
4794 
4795 	for (;;) {
4796 		struct sk_buff *next = NULL;
4797 
4798 		if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4799 			next = skb_queue_next(&tp->out_of_order_queue, skb);
4800 		skb = next;
4801 
4802 		/* Segment is terminated when we see gap or when
4803 		 * we are at the end of all the queue. */
4804 		if (!skb ||
4805 		    after(TCP_SKB_CB(skb)->seq, end) ||
4806 		    before(TCP_SKB_CB(skb)->end_seq, start)) {
4807 			tcp_collapse(sk, &tp->out_of_order_queue,
4808 				     head, skb, start, end);
4809 			head = skb;
4810 			if (!skb)
4811 				break;
4812 			/* Start new segment */
4813 			start = TCP_SKB_CB(skb)->seq;
4814 			end = TCP_SKB_CB(skb)->end_seq;
4815 		} else {
4816 			if (before(TCP_SKB_CB(skb)->seq, start))
4817 				start = TCP_SKB_CB(skb)->seq;
4818 			if (after(TCP_SKB_CB(skb)->end_seq, end))
4819 				end = TCP_SKB_CB(skb)->end_seq;
4820 		}
4821 	}
4822 }
4823 
4824 /*
4825  * Purge the out-of-order queue.
4826  * Return true if queue was pruned.
4827  */
4828 static int tcp_prune_ofo_queue(struct sock *sk)
4829 {
4830 	struct tcp_sock *tp = tcp_sk(sk);
4831 	int res = 0;
4832 
4833 	if (!skb_queue_empty(&tp->out_of_order_queue)) {
4834 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4835 		__skb_queue_purge(&tp->out_of_order_queue);
4836 
4837 		/* Reset SACK state.  A conforming SACK implementation will
4838 		 * do the same at a timeout based retransmit.  When a connection
4839 		 * is in a sad state like this, we care only about integrity
4840 		 * of the connection not performance.
4841 		 */
4842 		if (tp->rx_opt.sack_ok)
4843 			tcp_sack_reset(&tp->rx_opt);
4844 		sk_mem_reclaim(sk);
4845 		res = 1;
4846 	}
4847 	return res;
4848 }
4849 
4850 /* Reduce allocated memory if we can, trying to get
4851  * the socket within its memory limits again.
4852  *
4853  * Return less than zero if we should start dropping frames
4854  * until the socket owning process reads some of the data
4855  * to stabilize the situation.
4856  */
4857 static int tcp_prune_queue(struct sock *sk)
4858 {
4859 	struct tcp_sock *tp = tcp_sk(sk);
4860 
4861 	SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4862 
4863 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4864 
4865 	if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4866 		tcp_clamp_window(sk);
4867 	else if (tcp_memory_pressure)
4868 		tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4869 
4870 	tcp_collapse_ofo_queue(sk);
4871 	if (!skb_queue_empty(&sk->sk_receive_queue))
4872 		tcp_collapse(sk, &sk->sk_receive_queue,
4873 			     skb_peek(&sk->sk_receive_queue),
4874 			     NULL,
4875 			     tp->copied_seq, tp->rcv_nxt);
4876 	sk_mem_reclaim(sk);
4877 
4878 	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4879 		return 0;
4880 
4881 	/* Collapsing did not help, destructive actions follow.
4882 	 * This must not ever occur. */
4883 
4884 	tcp_prune_ofo_queue(sk);
4885 
4886 	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4887 		return 0;
4888 
4889 	/* If we are really being abused, tell the caller to silently
4890 	 * drop receive data on the floor.  It will get retransmitted
4891 	 * and hopefully then we'll have sufficient space.
4892 	 */
4893 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4894 
4895 	/* Massive buffer overcommit. */
4896 	tp->pred_flags = 0;
4897 	return -1;
4898 }
4899 
4900 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4901  * As additional protections, we do not touch cwnd in retransmission phases,
4902  * and if application hit its sndbuf limit recently.
4903  */
4904 void tcp_cwnd_application_limited(struct sock *sk)
4905 {
4906 	struct tcp_sock *tp = tcp_sk(sk);
4907 
4908 	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4909 	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4910 		/* Limited by application or receiver window. */
4911 		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4912 		u32 win_used = max(tp->snd_cwnd_used, init_win);
4913 		if (win_used < tp->snd_cwnd) {
4914 			tp->snd_ssthresh = tcp_current_ssthresh(sk);
4915 			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4916 		}
4917 		tp->snd_cwnd_used = 0;
4918 	}
4919 	tp->snd_cwnd_stamp = tcp_time_stamp;
4920 }
4921 
4922 static int tcp_should_expand_sndbuf(const struct sock *sk)
4923 {
4924 	const struct tcp_sock *tp = tcp_sk(sk);
4925 
4926 	/* If the user specified a specific send buffer setting, do
4927 	 * not modify it.
4928 	 */
4929 	if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4930 		return 0;
4931 
4932 	/* If we are under global TCP memory pressure, do not expand.  */
4933 	if (tcp_memory_pressure)
4934 		return 0;
4935 
4936 	/* If we are under soft global TCP memory pressure, do not expand.  */
4937 	if (atomic_long_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4938 		return 0;
4939 
4940 	/* If we filled the congestion window, do not expand.  */
4941 	if (tp->packets_out >= tp->snd_cwnd)
4942 		return 0;
4943 
4944 	return 1;
4945 }
4946 
4947 /* When incoming ACK allowed to free some skb from write_queue,
4948  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4949  * on the exit from tcp input handler.
4950  *
4951  * PROBLEM: sndbuf expansion does not work well with largesend.
4952  */
4953 static void tcp_new_space(struct sock *sk)
4954 {
4955 	struct tcp_sock *tp = tcp_sk(sk);
4956 
4957 	if (tcp_should_expand_sndbuf(sk)) {
4958 		int sndmem = SKB_TRUESIZE(max_t(u32,
4959 						tp->rx_opt.mss_clamp,
4960 						tp->mss_cache) +
4961 					  MAX_TCP_HEADER);
4962 		int demanded = max_t(unsigned int, tp->snd_cwnd,
4963 				     tp->reordering + 1);
4964 		sndmem *= 2 * demanded;
4965 		if (sndmem > sk->sk_sndbuf)
4966 			sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4967 		tp->snd_cwnd_stamp = tcp_time_stamp;
4968 	}
4969 
4970 	sk->sk_write_space(sk);
4971 }
4972 
4973 static void tcp_check_space(struct sock *sk)
4974 {
4975 	if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4976 		sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4977 		if (sk->sk_socket &&
4978 		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4979 			tcp_new_space(sk);
4980 	}
4981 }
4982 
4983 static inline void tcp_data_snd_check(struct sock *sk)
4984 {
4985 	tcp_push_pending_frames(sk);
4986 	tcp_check_space(sk);
4987 }
4988 
4989 /*
4990  * Check if sending an ack is needed.
4991  */
4992 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4993 {
4994 	struct tcp_sock *tp = tcp_sk(sk);
4995 
4996 	    /* More than one full frame received... */
4997 	if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4998 	     /* ... and right edge of window advances far enough.
4999 	      * (tcp_recvmsg() will send ACK otherwise). Or...
5000 	      */
5001 	     __tcp_select_window(sk) >= tp->rcv_wnd) ||
5002 	    /* We ACK each frame or... */
5003 	    tcp_in_quickack_mode(sk) ||
5004 	    /* We have out of order data. */
5005 	    (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5006 		/* Then ack it now */
5007 		tcp_send_ack(sk);
5008 	} else {
5009 		/* Else, send delayed ack. */
5010 		tcp_send_delayed_ack(sk);
5011 	}
5012 }
5013 
5014 static inline void tcp_ack_snd_check(struct sock *sk)
5015 {
5016 	if (!inet_csk_ack_scheduled(sk)) {
5017 		/* We sent a data segment already. */
5018 		return;
5019 	}
5020 	__tcp_ack_snd_check(sk, 1);
5021 }
5022 
5023 /*
5024  *	This routine is only called when we have urgent data
5025  *	signaled. Its the 'slow' part of tcp_urg. It could be
5026  *	moved inline now as tcp_urg is only called from one
5027  *	place. We handle URGent data wrong. We have to - as
5028  *	BSD still doesn't use the correction from RFC961.
5029  *	For 1003.1g we should support a new option TCP_STDURG to permit
5030  *	either form (or just set the sysctl tcp_stdurg).
5031  */
5032 
5033 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5034 {
5035 	struct tcp_sock *tp = tcp_sk(sk);
5036 	u32 ptr = ntohs(th->urg_ptr);
5037 
5038 	if (ptr && !sysctl_tcp_stdurg)
5039 		ptr--;
5040 	ptr += ntohl(th->seq);
5041 
5042 	/* Ignore urgent data that we've already seen and read. */
5043 	if (after(tp->copied_seq, ptr))
5044 		return;
5045 
5046 	/* Do not replay urg ptr.
5047 	 *
5048 	 * NOTE: interesting situation not covered by specs.
5049 	 * Misbehaving sender may send urg ptr, pointing to segment,
5050 	 * which we already have in ofo queue. We are not able to fetch
5051 	 * such data and will stay in TCP_URG_NOTYET until will be eaten
5052 	 * by recvmsg(). Seems, we are not obliged to handle such wicked
5053 	 * situations. But it is worth to think about possibility of some
5054 	 * DoSes using some hypothetical application level deadlock.
5055 	 */
5056 	if (before(ptr, tp->rcv_nxt))
5057 		return;
5058 
5059 	/* Do we already have a newer (or duplicate) urgent pointer? */
5060 	if (tp->urg_data && !after(ptr, tp->urg_seq))
5061 		return;
5062 
5063 	/* Tell the world about our new urgent pointer. */
5064 	sk_send_sigurg(sk);
5065 
5066 	/* We may be adding urgent data when the last byte read was
5067 	 * urgent. To do this requires some care. We cannot just ignore
5068 	 * tp->copied_seq since we would read the last urgent byte again
5069 	 * as data, nor can we alter copied_seq until this data arrives
5070 	 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5071 	 *
5072 	 * NOTE. Double Dutch. Rendering to plain English: author of comment
5073 	 * above did something sort of 	send("A", MSG_OOB); send("B", MSG_OOB);
5074 	 * and expect that both A and B disappear from stream. This is _wrong_.
5075 	 * Though this happens in BSD with high probability, this is occasional.
5076 	 * Any application relying on this is buggy. Note also, that fix "works"
5077 	 * only in this artificial test. Insert some normal data between A and B and we will
5078 	 * decline of BSD again. Verdict: it is better to remove to trap
5079 	 * buggy users.
5080 	 */
5081 	if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5082 	    !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5083 		struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5084 		tp->copied_seq++;
5085 		if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5086 			__skb_unlink(skb, &sk->sk_receive_queue);
5087 			__kfree_skb(skb);
5088 		}
5089 	}
5090 
5091 	tp->urg_data = TCP_URG_NOTYET;
5092 	tp->urg_seq = ptr;
5093 
5094 	/* Disable header prediction. */
5095 	tp->pred_flags = 0;
5096 }
5097 
5098 /* This is the 'fast' part of urgent handling. */
5099 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5100 {
5101 	struct tcp_sock *tp = tcp_sk(sk);
5102 
5103 	/* Check if we get a new urgent pointer - normally not. */
5104 	if (th->urg)
5105 		tcp_check_urg(sk, th);
5106 
5107 	/* Do we wait for any urgent data? - normally not... */
5108 	if (tp->urg_data == TCP_URG_NOTYET) {
5109 		u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5110 			  th->syn;
5111 
5112 		/* Is the urgent pointer pointing into this packet? */
5113 		if (ptr < skb->len) {
5114 			u8 tmp;
5115 			if (skb_copy_bits(skb, ptr, &tmp, 1))
5116 				BUG();
5117 			tp->urg_data = TCP_URG_VALID | tmp;
5118 			if (!sock_flag(sk, SOCK_DEAD))
5119 				sk->sk_data_ready(sk, 0);
5120 		}
5121 	}
5122 }
5123 
5124 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5125 {
5126 	struct tcp_sock *tp = tcp_sk(sk);
5127 	int chunk = skb->len - hlen;
5128 	int err;
5129 
5130 	local_bh_enable();
5131 	if (skb_csum_unnecessary(skb))
5132 		err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5133 	else
5134 		err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5135 						       tp->ucopy.iov);
5136 
5137 	if (!err) {
5138 		tp->ucopy.len -= chunk;
5139 		tp->copied_seq += chunk;
5140 		tcp_rcv_space_adjust(sk);
5141 	}
5142 
5143 	local_bh_disable();
5144 	return err;
5145 }
5146 
5147 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5148 					    struct sk_buff *skb)
5149 {
5150 	__sum16 result;
5151 
5152 	if (sock_owned_by_user(sk)) {
5153 		local_bh_enable();
5154 		result = __tcp_checksum_complete(skb);
5155 		local_bh_disable();
5156 	} else {
5157 		result = __tcp_checksum_complete(skb);
5158 	}
5159 	return result;
5160 }
5161 
5162 static inline int tcp_checksum_complete_user(struct sock *sk,
5163 					     struct sk_buff *skb)
5164 {
5165 	return !skb_csum_unnecessary(skb) &&
5166 	       __tcp_checksum_complete_user(sk, skb);
5167 }
5168 
5169 #ifdef CONFIG_NET_DMA
5170 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5171 				  int hlen)
5172 {
5173 	struct tcp_sock *tp = tcp_sk(sk);
5174 	int chunk = skb->len - hlen;
5175 	int dma_cookie;
5176 	int copied_early = 0;
5177 
5178 	if (tp->ucopy.wakeup)
5179 		return 0;
5180 
5181 	if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5182 		tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5183 
5184 	if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5185 
5186 		dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5187 							 skb, hlen,
5188 							 tp->ucopy.iov, chunk,
5189 							 tp->ucopy.pinned_list);
5190 
5191 		if (dma_cookie < 0)
5192 			goto out;
5193 
5194 		tp->ucopy.dma_cookie = dma_cookie;
5195 		copied_early = 1;
5196 
5197 		tp->ucopy.len -= chunk;
5198 		tp->copied_seq += chunk;
5199 		tcp_rcv_space_adjust(sk);
5200 
5201 		if ((tp->ucopy.len == 0) ||
5202 		    (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5203 		    (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5204 			tp->ucopy.wakeup = 1;
5205 			sk->sk_data_ready(sk, 0);
5206 		}
5207 	} else if (chunk > 0) {
5208 		tp->ucopy.wakeup = 1;
5209 		sk->sk_data_ready(sk, 0);
5210 	}
5211 out:
5212 	return copied_early;
5213 }
5214 #endif /* CONFIG_NET_DMA */
5215 
5216 /* Does PAWS and seqno based validation of an incoming segment, flags will
5217  * play significant role here.
5218  */
5219 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5220 			      const struct tcphdr *th, int syn_inerr)
5221 {
5222 	const u8 *hash_location;
5223 	struct tcp_sock *tp = tcp_sk(sk);
5224 
5225 	/* RFC1323: H1. Apply PAWS check first. */
5226 	if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5227 	    tp->rx_opt.saw_tstamp &&
5228 	    tcp_paws_discard(sk, skb)) {
5229 		if (!th->rst) {
5230 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5231 			tcp_send_dupack(sk, skb);
5232 			goto discard;
5233 		}
5234 		/* Reset is accepted even if it did not pass PAWS. */
5235 	}
5236 
5237 	/* Step 1: check sequence number */
5238 	if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5239 		/* RFC793, page 37: "In all states except SYN-SENT, all reset
5240 		 * (RST) segments are validated by checking their SEQ-fields."
5241 		 * And page 69: "If an incoming segment is not acceptable,
5242 		 * an acknowledgment should be sent in reply (unless the RST
5243 		 * bit is set, if so drop the segment and return)".
5244 		 */
5245 		if (!th->rst)
5246 			tcp_send_dupack(sk, skb);
5247 		goto discard;
5248 	}
5249 
5250 	/* Step 2: check RST bit */
5251 	if (th->rst) {
5252 		tcp_reset(sk);
5253 		goto discard;
5254 	}
5255 
5256 	/* ts_recent update must be made after we are sure that the packet
5257 	 * is in window.
5258 	 */
5259 	tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5260 
5261 	/* step 3: check security and precedence [ignored] */
5262 
5263 	/* step 4: Check for a SYN in window. */
5264 	if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5265 		if (syn_inerr)
5266 			TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5267 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5268 		tcp_reset(sk);
5269 		return -1;
5270 	}
5271 
5272 	return 1;
5273 
5274 discard:
5275 	__kfree_skb(skb);
5276 	return 0;
5277 }
5278 
5279 /*
5280  *	TCP receive function for the ESTABLISHED state.
5281  *
5282  *	It is split into a fast path and a slow path. The fast path is
5283  * 	disabled when:
5284  *	- A zero window was announced from us - zero window probing
5285  *        is only handled properly in the slow path.
5286  *	- Out of order segments arrived.
5287  *	- Urgent data is expected.
5288  *	- There is no buffer space left
5289  *	- Unexpected TCP flags/window values/header lengths are received
5290  *	  (detected by checking the TCP header against pred_flags)
5291  *	- Data is sent in both directions. Fast path only supports pure senders
5292  *	  or pure receivers (this means either the sequence number or the ack
5293  *	  value must stay constant)
5294  *	- Unexpected TCP option.
5295  *
5296  *	When these conditions are not satisfied it drops into a standard
5297  *	receive procedure patterned after RFC793 to handle all cases.
5298  *	The first three cases are guaranteed by proper pred_flags setting,
5299  *	the rest is checked inline. Fast processing is turned on in
5300  *	tcp_data_queue when everything is OK.
5301  */
5302 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5303 			const struct tcphdr *th, unsigned int len)
5304 {
5305 	struct tcp_sock *tp = tcp_sk(sk);
5306 	int res;
5307 
5308 	/*
5309 	 *	Header prediction.
5310 	 *	The code loosely follows the one in the famous
5311 	 *	"30 instruction TCP receive" Van Jacobson mail.
5312 	 *
5313 	 *	Van's trick is to deposit buffers into socket queue
5314 	 *	on a device interrupt, to call tcp_recv function
5315 	 *	on the receive process context and checksum and copy
5316 	 *	the buffer to user space. smart...
5317 	 *
5318 	 *	Our current scheme is not silly either but we take the
5319 	 *	extra cost of the net_bh soft interrupt processing...
5320 	 *	We do checksum and copy also but from device to kernel.
5321 	 */
5322 
5323 	tp->rx_opt.saw_tstamp = 0;
5324 
5325 	/*	pred_flags is 0xS?10 << 16 + snd_wnd
5326 	 *	if header_prediction is to be made
5327 	 *	'S' will always be tp->tcp_header_len >> 2
5328 	 *	'?' will be 0 for the fast path, otherwise pred_flags is 0 to
5329 	 *  turn it off	(when there are holes in the receive
5330 	 *	 space for instance)
5331 	 *	PSH flag is ignored.
5332 	 */
5333 
5334 	if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5335 	    TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5336 	    !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5337 		int tcp_header_len = tp->tcp_header_len;
5338 
5339 		/* Timestamp header prediction: tcp_header_len
5340 		 * is automatically equal to th->doff*4 due to pred_flags
5341 		 * match.
5342 		 */
5343 
5344 		/* Check timestamp */
5345 		if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5346 			/* No? Slow path! */
5347 			if (!tcp_parse_aligned_timestamp(tp, th))
5348 				goto slow_path;
5349 
5350 			/* If PAWS failed, check it more carefully in slow path */
5351 			if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5352 				goto slow_path;
5353 
5354 			/* DO NOT update ts_recent here, if checksum fails
5355 			 * and timestamp was corrupted part, it will result
5356 			 * in a hung connection since we will drop all
5357 			 * future packets due to the PAWS test.
5358 			 */
5359 		}
5360 
5361 		if (len <= tcp_header_len) {
5362 			/* Bulk data transfer: sender */
5363 			if (len == tcp_header_len) {
5364 				/* Predicted packet is in window by definition.
5365 				 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5366 				 * Hence, check seq<=rcv_wup reduces to:
5367 				 */
5368 				if (tcp_header_len ==
5369 				    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5370 				    tp->rcv_nxt == tp->rcv_wup)
5371 					tcp_store_ts_recent(tp);
5372 
5373 				/* We know that such packets are checksummed
5374 				 * on entry.
5375 				 */
5376 				tcp_ack(sk, skb, 0);
5377 				__kfree_skb(skb);
5378 				tcp_data_snd_check(sk);
5379 				return 0;
5380 			} else { /* Header too small */
5381 				TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5382 				goto discard;
5383 			}
5384 		} else {
5385 			int eaten = 0;
5386 			int copied_early = 0;
5387 
5388 			if (tp->copied_seq == tp->rcv_nxt &&
5389 			    len - tcp_header_len <= tp->ucopy.len) {
5390 #ifdef CONFIG_NET_DMA
5391 				if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5392 					copied_early = 1;
5393 					eaten = 1;
5394 				}
5395 #endif
5396 				if (tp->ucopy.task == current &&
5397 				    sock_owned_by_user(sk) && !copied_early) {
5398 					__set_current_state(TASK_RUNNING);
5399 
5400 					if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5401 						eaten = 1;
5402 				}
5403 				if (eaten) {
5404 					/* Predicted packet is in window by definition.
5405 					 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5406 					 * Hence, check seq<=rcv_wup reduces to:
5407 					 */
5408 					if (tcp_header_len ==
5409 					    (sizeof(struct tcphdr) +
5410 					     TCPOLEN_TSTAMP_ALIGNED) &&
5411 					    tp->rcv_nxt == tp->rcv_wup)
5412 						tcp_store_ts_recent(tp);
5413 
5414 					tcp_rcv_rtt_measure_ts(sk, skb);
5415 
5416 					__skb_pull(skb, tcp_header_len);
5417 					tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5418 					NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5419 				}
5420 				if (copied_early)
5421 					tcp_cleanup_rbuf(sk, skb->len);
5422 			}
5423 			if (!eaten) {
5424 				if (tcp_checksum_complete_user(sk, skb))
5425 					goto csum_error;
5426 
5427 				/* Predicted packet is in window by definition.
5428 				 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5429 				 * Hence, check seq<=rcv_wup reduces to:
5430 				 */
5431 				if (tcp_header_len ==
5432 				    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5433 				    tp->rcv_nxt == tp->rcv_wup)
5434 					tcp_store_ts_recent(tp);
5435 
5436 				tcp_rcv_rtt_measure_ts(sk, skb);
5437 
5438 				if ((int)skb->truesize > sk->sk_forward_alloc)
5439 					goto step5;
5440 
5441 				NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5442 
5443 				/* Bulk data transfer: receiver */
5444 				__skb_pull(skb, tcp_header_len);
5445 				__skb_queue_tail(&sk->sk_receive_queue, skb);
5446 				skb_set_owner_r(skb, sk);
5447 				tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5448 			}
5449 
5450 			tcp_event_data_recv(sk, skb);
5451 
5452 			if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5453 				/* Well, only one small jumplet in fast path... */
5454 				tcp_ack(sk, skb, FLAG_DATA);
5455 				tcp_data_snd_check(sk);
5456 				if (!inet_csk_ack_scheduled(sk))
5457 					goto no_ack;
5458 			}
5459 
5460 			if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5461 				__tcp_ack_snd_check(sk, 0);
5462 no_ack:
5463 #ifdef CONFIG_NET_DMA
5464 			if (copied_early)
5465 				__skb_queue_tail(&sk->sk_async_wait_queue, skb);
5466 			else
5467 #endif
5468 			if (eaten)
5469 				__kfree_skb(skb);
5470 			else
5471 				sk->sk_data_ready(sk, 0);
5472 			return 0;
5473 		}
5474 	}
5475 
5476 slow_path:
5477 	if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5478 		goto csum_error;
5479 
5480 	/*
5481 	 *	Standard slow path.
5482 	 */
5483 
5484 	res = tcp_validate_incoming(sk, skb, th, 1);
5485 	if (res <= 0)
5486 		return -res;
5487 
5488 step5:
5489 	if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5490 		goto discard;
5491 
5492 	tcp_rcv_rtt_measure_ts(sk, skb);
5493 
5494 	/* Process urgent data. */
5495 	tcp_urg(sk, skb, th);
5496 
5497 	/* step 7: process the segment text */
5498 	tcp_data_queue(sk, skb);
5499 
5500 	tcp_data_snd_check(sk);
5501 	tcp_ack_snd_check(sk);
5502 	return 0;
5503 
5504 csum_error:
5505 	TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5506 
5507 discard:
5508 	__kfree_skb(skb);
5509 	return 0;
5510 }
5511 EXPORT_SYMBOL(tcp_rcv_established);
5512 
5513 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5514 					 const struct tcphdr *th, unsigned int len)
5515 {
5516 	const u8 *hash_location;
5517 	struct inet_connection_sock *icsk = inet_csk(sk);
5518 	struct tcp_sock *tp = tcp_sk(sk);
5519 	struct tcp_cookie_values *cvp = tp->cookie_values;
5520 	int saved_clamp = tp->rx_opt.mss_clamp;
5521 
5522 	tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5523 
5524 	if (th->ack) {
5525 		/* rfc793:
5526 		 * "If the state is SYN-SENT then
5527 		 *    first check the ACK bit
5528 		 *      If the ACK bit is set
5529 		 *	  If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5530 		 *        a reset (unless the RST bit is set, if so drop
5531 		 *        the segment and return)"
5532 		 *
5533 		 *  We do not send data with SYN, so that RFC-correct
5534 		 *  test reduces to:
5535 		 */
5536 		if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5537 			goto reset_and_undo;
5538 
5539 		if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5540 		    !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5541 			     tcp_time_stamp)) {
5542 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5543 			goto reset_and_undo;
5544 		}
5545 
5546 		/* Now ACK is acceptable.
5547 		 *
5548 		 * "If the RST bit is set
5549 		 *    If the ACK was acceptable then signal the user "error:
5550 		 *    connection reset", drop the segment, enter CLOSED state,
5551 		 *    delete TCB, and return."
5552 		 */
5553 
5554 		if (th->rst) {
5555 			tcp_reset(sk);
5556 			goto discard;
5557 		}
5558 
5559 		/* rfc793:
5560 		 *   "fifth, if neither of the SYN or RST bits is set then
5561 		 *    drop the segment and return."
5562 		 *
5563 		 *    See note below!
5564 		 *                                        --ANK(990513)
5565 		 */
5566 		if (!th->syn)
5567 			goto discard_and_undo;
5568 
5569 		/* rfc793:
5570 		 *   "If the SYN bit is on ...
5571 		 *    are acceptable then ...
5572 		 *    (our SYN has been ACKed), change the connection
5573 		 *    state to ESTABLISHED..."
5574 		 */
5575 
5576 		TCP_ECN_rcv_synack(tp, th);
5577 
5578 		tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5579 		tcp_ack(sk, skb, FLAG_SLOWPATH);
5580 
5581 		/* Ok.. it's good. Set up sequence numbers and
5582 		 * move to established.
5583 		 */
5584 		tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5585 		tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5586 
5587 		/* RFC1323: The window in SYN & SYN/ACK segments is
5588 		 * never scaled.
5589 		 */
5590 		tp->snd_wnd = ntohs(th->window);
5591 		tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5592 
5593 		if (!tp->rx_opt.wscale_ok) {
5594 			tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5595 			tp->window_clamp = min(tp->window_clamp, 65535U);
5596 		}
5597 
5598 		if (tp->rx_opt.saw_tstamp) {
5599 			tp->rx_opt.tstamp_ok	   = 1;
5600 			tp->tcp_header_len =
5601 				sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5602 			tp->advmss	    -= TCPOLEN_TSTAMP_ALIGNED;
5603 			tcp_store_ts_recent(tp);
5604 		} else {
5605 			tp->tcp_header_len = sizeof(struct tcphdr);
5606 		}
5607 
5608 		if (tcp_is_sack(tp) && sysctl_tcp_fack)
5609 			tcp_enable_fack(tp);
5610 
5611 		tcp_mtup_init(sk);
5612 		tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5613 		tcp_initialize_rcv_mss(sk);
5614 
5615 		/* Remember, tcp_poll() does not lock socket!
5616 		 * Change state from SYN-SENT only after copied_seq
5617 		 * is initialized. */
5618 		tp->copied_seq = tp->rcv_nxt;
5619 
5620 		if (cvp != NULL &&
5621 		    cvp->cookie_pair_size > 0 &&
5622 		    tp->rx_opt.cookie_plus > 0) {
5623 			int cookie_size = tp->rx_opt.cookie_plus
5624 					- TCPOLEN_COOKIE_BASE;
5625 			int cookie_pair_size = cookie_size
5626 					     + cvp->cookie_desired;
5627 
5628 			/* A cookie extension option was sent and returned.
5629 			 * Note that each incoming SYNACK replaces the
5630 			 * Responder cookie.  The initial exchange is most
5631 			 * fragile, as protection against spoofing relies
5632 			 * entirely upon the sequence and timestamp (above).
5633 			 * This replacement strategy allows the correct pair to
5634 			 * pass through, while any others will be filtered via
5635 			 * Responder verification later.
5636 			 */
5637 			if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5638 				memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5639 				       hash_location, cookie_size);
5640 				cvp->cookie_pair_size = cookie_pair_size;
5641 			}
5642 		}
5643 
5644 		smp_mb();
5645 		tcp_set_state(sk, TCP_ESTABLISHED);
5646 
5647 		security_inet_conn_established(sk, skb);
5648 
5649 		/* Make sure socket is routed, for correct metrics.  */
5650 		icsk->icsk_af_ops->rebuild_header(sk);
5651 
5652 		tcp_init_metrics(sk);
5653 
5654 		tcp_init_congestion_control(sk);
5655 
5656 		/* Prevent spurious tcp_cwnd_restart() on first data
5657 		 * packet.
5658 		 */
5659 		tp->lsndtime = tcp_time_stamp;
5660 
5661 		tcp_init_buffer_space(sk);
5662 
5663 		if (sock_flag(sk, SOCK_KEEPOPEN))
5664 			inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5665 
5666 		if (!tp->rx_opt.snd_wscale)
5667 			__tcp_fast_path_on(tp, tp->snd_wnd);
5668 		else
5669 			tp->pred_flags = 0;
5670 
5671 		if (!sock_flag(sk, SOCK_DEAD)) {
5672 			sk->sk_state_change(sk);
5673 			sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5674 		}
5675 
5676 		if (sk->sk_write_pending ||
5677 		    icsk->icsk_accept_queue.rskq_defer_accept ||
5678 		    icsk->icsk_ack.pingpong) {
5679 			/* Save one ACK. Data will be ready after
5680 			 * several ticks, if write_pending is set.
5681 			 *
5682 			 * It may be deleted, but with this feature tcpdumps
5683 			 * look so _wonderfully_ clever, that I was not able
5684 			 * to stand against the temptation 8)     --ANK
5685 			 */
5686 			inet_csk_schedule_ack(sk);
5687 			icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5688 			icsk->icsk_ack.ato	 = TCP_ATO_MIN;
5689 			tcp_incr_quickack(sk);
5690 			tcp_enter_quickack_mode(sk);
5691 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5692 						  TCP_DELACK_MAX, TCP_RTO_MAX);
5693 
5694 discard:
5695 			__kfree_skb(skb);
5696 			return 0;
5697 		} else {
5698 			tcp_send_ack(sk);
5699 		}
5700 		return -1;
5701 	}
5702 
5703 	/* No ACK in the segment */
5704 
5705 	if (th->rst) {
5706 		/* rfc793:
5707 		 * "If the RST bit is set
5708 		 *
5709 		 *      Otherwise (no ACK) drop the segment and return."
5710 		 */
5711 
5712 		goto discard_and_undo;
5713 	}
5714 
5715 	/* PAWS check. */
5716 	if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5717 	    tcp_paws_reject(&tp->rx_opt, 0))
5718 		goto discard_and_undo;
5719 
5720 	if (th->syn) {
5721 		/* We see SYN without ACK. It is attempt of
5722 		 * simultaneous connect with crossed SYNs.
5723 		 * Particularly, it can be connect to self.
5724 		 */
5725 		tcp_set_state(sk, TCP_SYN_RECV);
5726 
5727 		if (tp->rx_opt.saw_tstamp) {
5728 			tp->rx_opt.tstamp_ok = 1;
5729 			tcp_store_ts_recent(tp);
5730 			tp->tcp_header_len =
5731 				sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5732 		} else {
5733 			tp->tcp_header_len = sizeof(struct tcphdr);
5734 		}
5735 
5736 		tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5737 		tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5738 
5739 		/* RFC1323: The window in SYN & SYN/ACK segments is
5740 		 * never scaled.
5741 		 */
5742 		tp->snd_wnd    = ntohs(th->window);
5743 		tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5744 		tp->max_window = tp->snd_wnd;
5745 
5746 		TCP_ECN_rcv_syn(tp, th);
5747 
5748 		tcp_mtup_init(sk);
5749 		tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5750 		tcp_initialize_rcv_mss(sk);
5751 
5752 		tcp_send_synack(sk);
5753 #if 0
5754 		/* Note, we could accept data and URG from this segment.
5755 		 * There are no obstacles to make this.
5756 		 *
5757 		 * However, if we ignore data in ACKless segments sometimes,
5758 		 * we have no reasons to accept it sometimes.
5759 		 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5760 		 * is not flawless. So, discard packet for sanity.
5761 		 * Uncomment this return to process the data.
5762 		 */
5763 		return -1;
5764 #else
5765 		goto discard;
5766 #endif
5767 	}
5768 	/* "fifth, if neither of the SYN or RST bits is set then
5769 	 * drop the segment and return."
5770 	 */
5771 
5772 discard_and_undo:
5773 	tcp_clear_options(&tp->rx_opt);
5774 	tp->rx_opt.mss_clamp = saved_clamp;
5775 	goto discard;
5776 
5777 reset_and_undo:
5778 	tcp_clear_options(&tp->rx_opt);
5779 	tp->rx_opt.mss_clamp = saved_clamp;
5780 	return 1;
5781 }
5782 
5783 /*
5784  *	This function implements the receiving procedure of RFC 793 for
5785  *	all states except ESTABLISHED and TIME_WAIT.
5786  *	It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5787  *	address independent.
5788  */
5789 
5790 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5791 			  const struct tcphdr *th, unsigned int len)
5792 {
5793 	struct tcp_sock *tp = tcp_sk(sk);
5794 	struct inet_connection_sock *icsk = inet_csk(sk);
5795 	int queued = 0;
5796 	int res;
5797 
5798 	tp->rx_opt.saw_tstamp = 0;
5799 
5800 	switch (sk->sk_state) {
5801 	case TCP_CLOSE:
5802 		goto discard;
5803 
5804 	case TCP_LISTEN:
5805 		if (th->ack)
5806 			return 1;
5807 
5808 		if (th->rst)
5809 			goto discard;
5810 
5811 		if (th->syn) {
5812 			if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5813 				return 1;
5814 
5815 			/* Now we have several options: In theory there is
5816 			 * nothing else in the frame. KA9Q has an option to
5817 			 * send data with the syn, BSD accepts data with the
5818 			 * syn up to the [to be] advertised window and
5819 			 * Solaris 2.1 gives you a protocol error. For now
5820 			 * we just ignore it, that fits the spec precisely
5821 			 * and avoids incompatibilities. It would be nice in
5822 			 * future to drop through and process the data.
5823 			 *
5824 			 * Now that TTCP is starting to be used we ought to
5825 			 * queue this data.
5826 			 * But, this leaves one open to an easy denial of
5827 			 * service attack, and SYN cookies can't defend
5828 			 * against this problem. So, we drop the data
5829 			 * in the interest of security over speed unless
5830 			 * it's still in use.
5831 			 */
5832 			kfree_skb(skb);
5833 			return 0;
5834 		}
5835 		goto discard;
5836 
5837 	case TCP_SYN_SENT:
5838 		queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5839 		if (queued >= 0)
5840 			return queued;
5841 
5842 		/* Do step6 onward by hand. */
5843 		tcp_urg(sk, skb, th);
5844 		__kfree_skb(skb);
5845 		tcp_data_snd_check(sk);
5846 		return 0;
5847 	}
5848 
5849 	res = tcp_validate_incoming(sk, skb, th, 0);
5850 	if (res <= 0)
5851 		return -res;
5852 
5853 	/* step 5: check the ACK field */
5854 	if (th->ack) {
5855 		int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5856 
5857 		switch (sk->sk_state) {
5858 		case TCP_SYN_RECV:
5859 			if (acceptable) {
5860 				tp->copied_seq = tp->rcv_nxt;
5861 				smp_mb();
5862 				tcp_set_state(sk, TCP_ESTABLISHED);
5863 				sk->sk_state_change(sk);
5864 
5865 				/* Note, that this wakeup is only for marginal
5866 				 * crossed SYN case. Passively open sockets
5867 				 * are not waked up, because sk->sk_sleep ==
5868 				 * NULL and sk->sk_socket == NULL.
5869 				 */
5870 				if (sk->sk_socket)
5871 					sk_wake_async(sk,
5872 						      SOCK_WAKE_IO, POLL_OUT);
5873 
5874 				tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5875 				tp->snd_wnd = ntohs(th->window) <<
5876 					      tp->rx_opt.snd_wscale;
5877 				tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5878 
5879 				if (tp->rx_opt.tstamp_ok)
5880 					tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5881 
5882 				/* Make sure socket is routed, for
5883 				 * correct metrics.
5884 				 */
5885 				icsk->icsk_af_ops->rebuild_header(sk);
5886 
5887 				tcp_init_metrics(sk);
5888 
5889 				tcp_init_congestion_control(sk);
5890 
5891 				/* Prevent spurious tcp_cwnd_restart() on
5892 				 * first data packet.
5893 				 */
5894 				tp->lsndtime = tcp_time_stamp;
5895 
5896 				tcp_mtup_init(sk);
5897 				tcp_initialize_rcv_mss(sk);
5898 				tcp_init_buffer_space(sk);
5899 				tcp_fast_path_on(tp);
5900 			} else {
5901 				return 1;
5902 			}
5903 			break;
5904 
5905 		case TCP_FIN_WAIT1:
5906 			if (tp->snd_una == tp->write_seq) {
5907 				tcp_set_state(sk, TCP_FIN_WAIT2);
5908 				sk->sk_shutdown |= SEND_SHUTDOWN;
5909 				dst_confirm(__sk_dst_get(sk));
5910 
5911 				if (!sock_flag(sk, SOCK_DEAD))
5912 					/* Wake up lingering close() */
5913 					sk->sk_state_change(sk);
5914 				else {
5915 					int tmo;
5916 
5917 					if (tp->linger2 < 0 ||
5918 					    (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5919 					     after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5920 						tcp_done(sk);
5921 						NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5922 						return 1;
5923 					}
5924 
5925 					tmo = tcp_fin_time(sk);
5926 					if (tmo > TCP_TIMEWAIT_LEN) {
5927 						inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5928 					} else if (th->fin || sock_owned_by_user(sk)) {
5929 						/* Bad case. We could lose such FIN otherwise.
5930 						 * It is not a big problem, but it looks confusing
5931 						 * and not so rare event. We still can lose it now,
5932 						 * if it spins in bh_lock_sock(), but it is really
5933 						 * marginal case.
5934 						 */
5935 						inet_csk_reset_keepalive_timer(sk, tmo);
5936 					} else {
5937 						tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5938 						goto discard;
5939 					}
5940 				}
5941 			}
5942 			break;
5943 
5944 		case TCP_CLOSING:
5945 			if (tp->snd_una == tp->write_seq) {
5946 				tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5947 				goto discard;
5948 			}
5949 			break;
5950 
5951 		case TCP_LAST_ACK:
5952 			if (tp->snd_una == tp->write_seq) {
5953 				tcp_update_metrics(sk);
5954 				tcp_done(sk);
5955 				goto discard;
5956 			}
5957 			break;
5958 		}
5959 	} else
5960 		goto discard;
5961 
5962 	/* step 6: check the URG bit */
5963 	tcp_urg(sk, skb, th);
5964 
5965 	/* step 7: process the segment text */
5966 	switch (sk->sk_state) {
5967 	case TCP_CLOSE_WAIT:
5968 	case TCP_CLOSING:
5969 	case TCP_LAST_ACK:
5970 		if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5971 			break;
5972 	case TCP_FIN_WAIT1:
5973 	case TCP_FIN_WAIT2:
5974 		/* RFC 793 says to queue data in these states,
5975 		 * RFC 1122 says we MUST send a reset.
5976 		 * BSD 4.4 also does reset.
5977 		 */
5978 		if (sk->sk_shutdown & RCV_SHUTDOWN) {
5979 			if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5980 			    after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5981 				NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5982 				tcp_reset(sk);
5983 				return 1;
5984 			}
5985 		}
5986 		/* Fall through */
5987 	case TCP_ESTABLISHED:
5988 		tcp_data_queue(sk, skb);
5989 		queued = 1;
5990 		break;
5991 	}
5992 
5993 	/* tcp_data could move socket to TIME-WAIT */
5994 	if (sk->sk_state != TCP_CLOSE) {
5995 		tcp_data_snd_check(sk);
5996 		tcp_ack_snd_check(sk);
5997 	}
5998 
5999 	if (!queued) {
6000 discard:
6001 		__kfree_skb(skb);
6002 	}
6003 	return 0;
6004 }
6005 EXPORT_SYMBOL(tcp_rcv_state_process);
6006