xref: /linux/net/ipv4/tcp_input.c (revision a115bc070b1fc57ab23f3972401425927b5b465c)
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/module.h>
66 #include <linux/sysctl.h>
67 #include <linux/kernel.h>
68 #include <net/dst.h>
69 #include <net/tcp.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
74 
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly = 2;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84 
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly = 2;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
91 
92 int sysctl_tcp_thin_dupack __read_mostly;
93 
94 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
95 int sysctl_tcp_abc __read_mostly;
96 
97 #define FLAG_DATA		0x01 /* Incoming frame contained data.		*/
98 #define FLAG_WIN_UPDATE		0x02 /* Incoming ACK was a window update.	*/
99 #define FLAG_DATA_ACKED		0x04 /* This ACK acknowledged new data.		*/
100 #define FLAG_RETRANS_DATA_ACKED	0x08 /* "" "" some of which was retransmitted.	*/
101 #define FLAG_SYN_ACKED		0x10 /* This ACK acknowledged SYN.		*/
102 #define FLAG_DATA_SACKED	0x20 /* New SACK.				*/
103 #define FLAG_ECE		0x40 /* ECE in this ACK				*/
104 #define FLAG_DATA_LOST		0x80 /* SACK detected data lossage.		*/
105 #define FLAG_SLOWPATH		0x100 /* Do not skip RFC checks for window update.*/
106 #define FLAG_ONLY_ORIG_SACKED	0x200 /* SACKs only non-rexmit sent before RTO */
107 #define FLAG_SND_UNA_ADVANCED	0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
108 #define FLAG_DSACKING_ACK	0x800 /* SACK blocks contained D-SACK info */
109 #define FLAG_NONHEAD_RETRANS_ACKED	0x1000 /* Non-head rexmitted data was ACKed */
110 #define FLAG_SACK_RENEGING	0x2000 /* snd_una advanced to a sacked seq */
111 
112 #define FLAG_ACKED		(FLAG_DATA_ACKED|FLAG_SYN_ACKED)
113 #define FLAG_NOT_DUP		(FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
114 #define FLAG_CA_ALERT		(FLAG_DATA_SACKED|FLAG_ECE)
115 #define FLAG_FORWARD_PROGRESS	(FLAG_ACKED|FLAG_DATA_SACKED)
116 #define FLAG_ANY_PROGRESS	(FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
117 
118 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
119 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
120 
121 /* Adapt the MSS value used to make delayed ack decision to the
122  * real world.
123  */
124 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
125 {
126 	struct inet_connection_sock *icsk = inet_csk(sk);
127 	const unsigned int lss = icsk->icsk_ack.last_seg_size;
128 	unsigned int len;
129 
130 	icsk->icsk_ack.last_seg_size = 0;
131 
132 	/* skb->len may jitter because of SACKs, even if peer
133 	 * sends good full-sized frames.
134 	 */
135 	len = skb_shinfo(skb)->gso_size ? : skb->len;
136 	if (len >= icsk->icsk_ack.rcv_mss) {
137 		icsk->icsk_ack.rcv_mss = len;
138 	} else {
139 		/* Otherwise, we make more careful check taking into account,
140 		 * that SACKs block is variable.
141 		 *
142 		 * "len" is invariant segment length, including TCP header.
143 		 */
144 		len += skb->data - skb_transport_header(skb);
145 		if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
146 		    /* If PSH is not set, packet should be
147 		     * full sized, provided peer TCP is not badly broken.
148 		     * This observation (if it is correct 8)) allows
149 		     * to handle super-low mtu links fairly.
150 		     */
151 		    (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
152 		     !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
153 			/* Subtract also invariant (if peer is RFC compliant),
154 			 * tcp header plus fixed timestamp option length.
155 			 * Resulting "len" is MSS free of SACK jitter.
156 			 */
157 			len -= tcp_sk(sk)->tcp_header_len;
158 			icsk->icsk_ack.last_seg_size = len;
159 			if (len == lss) {
160 				icsk->icsk_ack.rcv_mss = len;
161 				return;
162 			}
163 		}
164 		if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
165 			icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
166 		icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
167 	}
168 }
169 
170 static void tcp_incr_quickack(struct sock *sk)
171 {
172 	struct inet_connection_sock *icsk = inet_csk(sk);
173 	unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
174 
175 	if (quickacks == 0)
176 		quickacks = 2;
177 	if (quickacks > icsk->icsk_ack.quick)
178 		icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
179 }
180 
181 void tcp_enter_quickack_mode(struct sock *sk)
182 {
183 	struct inet_connection_sock *icsk = inet_csk(sk);
184 	tcp_incr_quickack(sk);
185 	icsk->icsk_ack.pingpong = 0;
186 	icsk->icsk_ack.ato = TCP_ATO_MIN;
187 }
188 
189 /* Send ACKs quickly, if "quick" count is not exhausted
190  * and the session is not interactive.
191  */
192 
193 static inline int tcp_in_quickack_mode(const struct sock *sk)
194 {
195 	const struct inet_connection_sock *icsk = inet_csk(sk);
196 	return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
197 }
198 
199 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
200 {
201 	if (tp->ecn_flags & TCP_ECN_OK)
202 		tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
203 }
204 
205 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
206 {
207 	if (tcp_hdr(skb)->cwr)
208 		tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
209 }
210 
211 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
212 {
213 	tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
214 }
215 
216 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
217 {
218 	if (tp->ecn_flags & TCP_ECN_OK) {
219 		if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
220 			tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
221 		/* Funny extension: if ECT is not set on a segment,
222 		 * it is surely retransmit. It is not in ECN RFC,
223 		 * but Linux follows this rule. */
224 		else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
225 			tcp_enter_quickack_mode((struct sock *)tp);
226 	}
227 }
228 
229 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
230 {
231 	if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
232 		tp->ecn_flags &= ~TCP_ECN_OK;
233 }
234 
235 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
236 {
237 	if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
238 		tp->ecn_flags &= ~TCP_ECN_OK;
239 }
240 
241 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
242 {
243 	if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
244 		return 1;
245 	return 0;
246 }
247 
248 /* Buffer size and advertised window tuning.
249  *
250  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
251  */
252 
253 static void tcp_fixup_sndbuf(struct sock *sk)
254 {
255 	int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
256 		     sizeof(struct sk_buff);
257 
258 	if (sk->sk_sndbuf < 3 * sndmem)
259 		sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
260 }
261 
262 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
263  *
264  * All tcp_full_space() is split to two parts: "network" buffer, allocated
265  * forward and advertised in receiver window (tp->rcv_wnd) and
266  * "application buffer", required to isolate scheduling/application
267  * latencies from network.
268  * window_clamp is maximal advertised window. It can be less than
269  * tcp_full_space(), in this case tcp_full_space() - window_clamp
270  * is reserved for "application" buffer. The less window_clamp is
271  * the smoother our behaviour from viewpoint of network, but the lower
272  * throughput and the higher sensitivity of the connection to losses. 8)
273  *
274  * rcv_ssthresh is more strict window_clamp used at "slow start"
275  * phase to predict further behaviour of this connection.
276  * It is used for two goals:
277  * - to enforce header prediction at sender, even when application
278  *   requires some significant "application buffer". It is check #1.
279  * - to prevent pruning of receive queue because of misprediction
280  *   of receiver window. Check #2.
281  *
282  * The scheme does not work when sender sends good segments opening
283  * window and then starts to feed us spaghetti. But it should work
284  * in common situations. Otherwise, we have to rely on queue collapsing.
285  */
286 
287 /* Slow part of check#2. */
288 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
289 {
290 	struct tcp_sock *tp = tcp_sk(sk);
291 	/* Optimize this! */
292 	int truesize = tcp_win_from_space(skb->truesize) >> 1;
293 	int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
294 
295 	while (tp->rcv_ssthresh <= window) {
296 		if (truesize <= skb->len)
297 			return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
298 
299 		truesize >>= 1;
300 		window >>= 1;
301 	}
302 	return 0;
303 }
304 
305 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
306 {
307 	struct tcp_sock *tp = tcp_sk(sk);
308 
309 	/* Check #1 */
310 	if (tp->rcv_ssthresh < tp->window_clamp &&
311 	    (int)tp->rcv_ssthresh < tcp_space(sk) &&
312 	    !tcp_memory_pressure) {
313 		int incr;
314 
315 		/* Check #2. Increase window, if skb with such overhead
316 		 * will fit to rcvbuf in future.
317 		 */
318 		if (tcp_win_from_space(skb->truesize) <= skb->len)
319 			incr = 2 * tp->advmss;
320 		else
321 			incr = __tcp_grow_window(sk, skb);
322 
323 		if (incr) {
324 			tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
325 					       tp->window_clamp);
326 			inet_csk(sk)->icsk_ack.quick |= 1;
327 		}
328 	}
329 }
330 
331 /* 3. Tuning rcvbuf, when connection enters established state. */
332 
333 static void tcp_fixup_rcvbuf(struct sock *sk)
334 {
335 	struct tcp_sock *tp = tcp_sk(sk);
336 	int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
337 
338 	/* Try to select rcvbuf so that 4 mss-sized segments
339 	 * will fit to window and corresponding skbs will fit to our rcvbuf.
340 	 * (was 3; 4 is minimum to allow fast retransmit to work.)
341 	 */
342 	while (tcp_win_from_space(rcvmem) < tp->advmss)
343 		rcvmem += 128;
344 	if (sk->sk_rcvbuf < 4 * rcvmem)
345 		sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
346 }
347 
348 /* 4. Try to fixup all. It is made immediately after connection enters
349  *    established state.
350  */
351 static void tcp_init_buffer_space(struct sock *sk)
352 {
353 	struct tcp_sock *tp = tcp_sk(sk);
354 	int maxwin;
355 
356 	if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
357 		tcp_fixup_rcvbuf(sk);
358 	if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
359 		tcp_fixup_sndbuf(sk);
360 
361 	tp->rcvq_space.space = tp->rcv_wnd;
362 
363 	maxwin = tcp_full_space(sk);
364 
365 	if (tp->window_clamp >= maxwin) {
366 		tp->window_clamp = maxwin;
367 
368 		if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
369 			tp->window_clamp = max(maxwin -
370 					       (maxwin >> sysctl_tcp_app_win),
371 					       4 * tp->advmss);
372 	}
373 
374 	/* Force reservation of one segment. */
375 	if (sysctl_tcp_app_win &&
376 	    tp->window_clamp > 2 * tp->advmss &&
377 	    tp->window_clamp + tp->advmss > maxwin)
378 		tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
379 
380 	tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
381 	tp->snd_cwnd_stamp = tcp_time_stamp;
382 }
383 
384 /* 5. Recalculate window clamp after socket hit its memory bounds. */
385 static void tcp_clamp_window(struct sock *sk)
386 {
387 	struct tcp_sock *tp = tcp_sk(sk);
388 	struct inet_connection_sock *icsk = inet_csk(sk);
389 
390 	icsk->icsk_ack.quick = 0;
391 
392 	if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
393 	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
394 	    !tcp_memory_pressure &&
395 	    atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
396 		sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
397 				    sysctl_tcp_rmem[2]);
398 	}
399 	if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
400 		tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
401 }
402 
403 /* Initialize RCV_MSS value.
404  * RCV_MSS is an our guess about MSS used by the peer.
405  * We haven't any direct information about the MSS.
406  * It's better to underestimate the RCV_MSS rather than overestimate.
407  * Overestimations make us ACKing less frequently than needed.
408  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
409  */
410 void tcp_initialize_rcv_mss(struct sock *sk)
411 {
412 	struct tcp_sock *tp = tcp_sk(sk);
413 	unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
414 
415 	hint = min(hint, tp->rcv_wnd / 2);
416 	hint = min(hint, TCP_MSS_DEFAULT);
417 	hint = max(hint, TCP_MIN_MSS);
418 
419 	inet_csk(sk)->icsk_ack.rcv_mss = hint;
420 }
421 
422 /* Receiver "autotuning" code.
423  *
424  * The algorithm for RTT estimation w/o timestamps is based on
425  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
426  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
427  *
428  * More detail on this code can be found at
429  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
430  * though this reference is out of date.  A new paper
431  * is pending.
432  */
433 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
434 {
435 	u32 new_sample = tp->rcv_rtt_est.rtt;
436 	long m = sample;
437 
438 	if (m == 0)
439 		m = 1;
440 
441 	if (new_sample != 0) {
442 		/* If we sample in larger samples in the non-timestamp
443 		 * case, we could grossly overestimate the RTT especially
444 		 * with chatty applications or bulk transfer apps which
445 		 * are stalled on filesystem I/O.
446 		 *
447 		 * Also, since we are only going for a minimum in the
448 		 * non-timestamp case, we do not smooth things out
449 		 * else with timestamps disabled convergence takes too
450 		 * long.
451 		 */
452 		if (!win_dep) {
453 			m -= (new_sample >> 3);
454 			new_sample += m;
455 		} else if (m < new_sample)
456 			new_sample = m << 3;
457 	} else {
458 		/* No previous measure. */
459 		new_sample = m << 3;
460 	}
461 
462 	if (tp->rcv_rtt_est.rtt != new_sample)
463 		tp->rcv_rtt_est.rtt = new_sample;
464 }
465 
466 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
467 {
468 	if (tp->rcv_rtt_est.time == 0)
469 		goto new_measure;
470 	if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
471 		return;
472 	tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
473 
474 new_measure:
475 	tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
476 	tp->rcv_rtt_est.time = tcp_time_stamp;
477 }
478 
479 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
480 					  const struct sk_buff *skb)
481 {
482 	struct tcp_sock *tp = tcp_sk(sk);
483 	if (tp->rx_opt.rcv_tsecr &&
484 	    (TCP_SKB_CB(skb)->end_seq -
485 	     TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
486 		tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
487 }
488 
489 /*
490  * This function should be called every time data is copied to user space.
491  * It calculates the appropriate TCP receive buffer space.
492  */
493 void tcp_rcv_space_adjust(struct sock *sk)
494 {
495 	struct tcp_sock *tp = tcp_sk(sk);
496 	int time;
497 	int space;
498 
499 	if (tp->rcvq_space.time == 0)
500 		goto new_measure;
501 
502 	time = tcp_time_stamp - tp->rcvq_space.time;
503 	if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
504 		return;
505 
506 	space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
507 
508 	space = max(tp->rcvq_space.space, space);
509 
510 	if (tp->rcvq_space.space != space) {
511 		int rcvmem;
512 
513 		tp->rcvq_space.space = space;
514 
515 		if (sysctl_tcp_moderate_rcvbuf &&
516 		    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
517 			int new_clamp = space;
518 
519 			/* Receive space grows, normalize in order to
520 			 * take into account packet headers and sk_buff
521 			 * structure overhead.
522 			 */
523 			space /= tp->advmss;
524 			if (!space)
525 				space = 1;
526 			rcvmem = (tp->advmss + MAX_TCP_HEADER +
527 				  16 + sizeof(struct sk_buff));
528 			while (tcp_win_from_space(rcvmem) < tp->advmss)
529 				rcvmem += 128;
530 			space *= rcvmem;
531 			space = min(space, sysctl_tcp_rmem[2]);
532 			if (space > sk->sk_rcvbuf) {
533 				sk->sk_rcvbuf = space;
534 
535 				/* Make the window clamp follow along.  */
536 				tp->window_clamp = new_clamp;
537 			}
538 		}
539 	}
540 
541 new_measure:
542 	tp->rcvq_space.seq = tp->copied_seq;
543 	tp->rcvq_space.time = tcp_time_stamp;
544 }
545 
546 /* There is something which you must keep in mind when you analyze the
547  * behavior of the tp->ato delayed ack timeout interval.  When a
548  * connection starts up, we want to ack as quickly as possible.  The
549  * problem is that "good" TCP's do slow start at the beginning of data
550  * transmission.  The means that until we send the first few ACK's the
551  * sender will sit on his end and only queue most of his data, because
552  * he can only send snd_cwnd unacked packets at any given time.  For
553  * each ACK we send, he increments snd_cwnd and transmits more of his
554  * queue.  -DaveM
555  */
556 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
557 {
558 	struct tcp_sock *tp = tcp_sk(sk);
559 	struct inet_connection_sock *icsk = inet_csk(sk);
560 	u32 now;
561 
562 	inet_csk_schedule_ack(sk);
563 
564 	tcp_measure_rcv_mss(sk, skb);
565 
566 	tcp_rcv_rtt_measure(tp);
567 
568 	now = tcp_time_stamp;
569 
570 	if (!icsk->icsk_ack.ato) {
571 		/* The _first_ data packet received, initialize
572 		 * delayed ACK engine.
573 		 */
574 		tcp_incr_quickack(sk);
575 		icsk->icsk_ack.ato = TCP_ATO_MIN;
576 	} else {
577 		int m = now - icsk->icsk_ack.lrcvtime;
578 
579 		if (m <= TCP_ATO_MIN / 2) {
580 			/* The fastest case is the first. */
581 			icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
582 		} else if (m < icsk->icsk_ack.ato) {
583 			icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
584 			if (icsk->icsk_ack.ato > icsk->icsk_rto)
585 				icsk->icsk_ack.ato = icsk->icsk_rto;
586 		} else if (m > icsk->icsk_rto) {
587 			/* Too long gap. Apparently sender failed to
588 			 * restart window, so that we send ACKs quickly.
589 			 */
590 			tcp_incr_quickack(sk);
591 			sk_mem_reclaim(sk);
592 		}
593 	}
594 	icsk->icsk_ack.lrcvtime = now;
595 
596 	TCP_ECN_check_ce(tp, skb);
597 
598 	if (skb->len >= 128)
599 		tcp_grow_window(sk, skb);
600 }
601 
602 /* Called to compute a smoothed rtt estimate. The data fed to this
603  * routine either comes from timestamps, or from segments that were
604  * known _not_ to have been retransmitted [see Karn/Partridge
605  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
606  * piece by Van Jacobson.
607  * NOTE: the next three routines used to be one big routine.
608  * To save cycles in the RFC 1323 implementation it was better to break
609  * it up into three procedures. -- erics
610  */
611 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
612 {
613 	struct tcp_sock *tp = tcp_sk(sk);
614 	long m = mrtt; /* RTT */
615 
616 	/*	The following amusing code comes from Jacobson's
617 	 *	article in SIGCOMM '88.  Note that rtt and mdev
618 	 *	are scaled versions of rtt and mean deviation.
619 	 *	This is designed to be as fast as possible
620 	 *	m stands for "measurement".
621 	 *
622 	 *	On a 1990 paper the rto value is changed to:
623 	 *	RTO = rtt + 4 * mdev
624 	 *
625 	 * Funny. This algorithm seems to be very broken.
626 	 * These formulae increase RTO, when it should be decreased, increase
627 	 * too slowly, when it should be increased quickly, decrease too quickly
628 	 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
629 	 * does not matter how to _calculate_ it. Seems, it was trap
630 	 * that VJ failed to avoid. 8)
631 	 */
632 	if (m == 0)
633 		m = 1;
634 	if (tp->srtt != 0) {
635 		m -= (tp->srtt >> 3);	/* m is now error in rtt est */
636 		tp->srtt += m;		/* rtt = 7/8 rtt + 1/8 new */
637 		if (m < 0) {
638 			m = -m;		/* m is now abs(error) */
639 			m -= (tp->mdev >> 2);   /* similar update on mdev */
640 			/* This is similar to one of Eifel findings.
641 			 * Eifel blocks mdev updates when rtt decreases.
642 			 * This solution is a bit different: we use finer gain
643 			 * for mdev in this case (alpha*beta).
644 			 * Like Eifel it also prevents growth of rto,
645 			 * but also it limits too fast rto decreases,
646 			 * happening in pure Eifel.
647 			 */
648 			if (m > 0)
649 				m >>= 3;
650 		} else {
651 			m -= (tp->mdev >> 2);   /* similar update on mdev */
652 		}
653 		tp->mdev += m;	    	/* mdev = 3/4 mdev + 1/4 new */
654 		if (tp->mdev > tp->mdev_max) {
655 			tp->mdev_max = tp->mdev;
656 			if (tp->mdev_max > tp->rttvar)
657 				tp->rttvar = tp->mdev_max;
658 		}
659 		if (after(tp->snd_una, tp->rtt_seq)) {
660 			if (tp->mdev_max < tp->rttvar)
661 				tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
662 			tp->rtt_seq = tp->snd_nxt;
663 			tp->mdev_max = tcp_rto_min(sk);
664 		}
665 	} else {
666 		/* no previous measure. */
667 		tp->srtt = m << 3;	/* take the measured time to be rtt */
668 		tp->mdev = m << 1;	/* make sure rto = 3*rtt */
669 		tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
670 		tp->rtt_seq = tp->snd_nxt;
671 	}
672 }
673 
674 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
675  * routine referred to above.
676  */
677 static inline void tcp_set_rto(struct sock *sk)
678 {
679 	const struct tcp_sock *tp = tcp_sk(sk);
680 	/* Old crap is replaced with new one. 8)
681 	 *
682 	 * More seriously:
683 	 * 1. If rtt variance happened to be less 50msec, it is hallucination.
684 	 *    It cannot be less due to utterly erratic ACK generation made
685 	 *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
686 	 *    to do with delayed acks, because at cwnd>2 true delack timeout
687 	 *    is invisible. Actually, Linux-2.4 also generates erratic
688 	 *    ACKs in some circumstances.
689 	 */
690 	inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
691 
692 	/* 2. Fixups made earlier cannot be right.
693 	 *    If we do not estimate RTO correctly without them,
694 	 *    all the algo is pure shit and should be replaced
695 	 *    with correct one. It is exactly, which we pretend to do.
696 	 */
697 
698 	/* NOTE: clamping at TCP_RTO_MIN is not required, current algo
699 	 * guarantees that rto is higher.
700 	 */
701 	tcp_bound_rto(sk);
702 }
703 
704 /* Save metrics learned by this TCP session.
705    This function is called only, when TCP finishes successfully
706    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
707  */
708 void tcp_update_metrics(struct sock *sk)
709 {
710 	struct tcp_sock *tp = tcp_sk(sk);
711 	struct dst_entry *dst = __sk_dst_get(sk);
712 
713 	if (sysctl_tcp_nometrics_save)
714 		return;
715 
716 	dst_confirm(dst);
717 
718 	if (dst && (dst->flags & DST_HOST)) {
719 		const struct inet_connection_sock *icsk = inet_csk(sk);
720 		int m;
721 		unsigned long rtt;
722 
723 		if (icsk->icsk_backoff || !tp->srtt) {
724 			/* This session failed to estimate rtt. Why?
725 			 * Probably, no packets returned in time.
726 			 * Reset our results.
727 			 */
728 			if (!(dst_metric_locked(dst, RTAX_RTT)))
729 				dst->metrics[RTAX_RTT - 1] = 0;
730 			return;
731 		}
732 
733 		rtt = dst_metric_rtt(dst, RTAX_RTT);
734 		m = rtt - tp->srtt;
735 
736 		/* If newly calculated rtt larger than stored one,
737 		 * store new one. Otherwise, use EWMA. Remember,
738 		 * rtt overestimation is always better than underestimation.
739 		 */
740 		if (!(dst_metric_locked(dst, RTAX_RTT))) {
741 			if (m <= 0)
742 				set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
743 			else
744 				set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
745 		}
746 
747 		if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
748 			unsigned long var;
749 			if (m < 0)
750 				m = -m;
751 
752 			/* Scale deviation to rttvar fixed point */
753 			m >>= 1;
754 			if (m < tp->mdev)
755 				m = tp->mdev;
756 
757 			var = dst_metric_rtt(dst, RTAX_RTTVAR);
758 			if (m >= var)
759 				var = m;
760 			else
761 				var -= (var - m) >> 2;
762 
763 			set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
764 		}
765 
766 		if (tcp_in_initial_slowstart(tp)) {
767 			/* Slow start still did not finish. */
768 			if (dst_metric(dst, RTAX_SSTHRESH) &&
769 			    !dst_metric_locked(dst, RTAX_SSTHRESH) &&
770 			    (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
771 				dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
772 			if (!dst_metric_locked(dst, RTAX_CWND) &&
773 			    tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
774 				dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
775 		} else if (tp->snd_cwnd > tp->snd_ssthresh &&
776 			   icsk->icsk_ca_state == TCP_CA_Open) {
777 			/* Cong. avoidance phase, cwnd is reliable. */
778 			if (!dst_metric_locked(dst, RTAX_SSTHRESH))
779 				dst->metrics[RTAX_SSTHRESH-1] =
780 					max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
781 			if (!dst_metric_locked(dst, RTAX_CWND))
782 				dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1;
783 		} else {
784 			/* Else slow start did not finish, cwnd is non-sense,
785 			   ssthresh may be also invalid.
786 			 */
787 			if (!dst_metric_locked(dst, RTAX_CWND))
788 				dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1;
789 			if (dst_metric(dst, RTAX_SSTHRESH) &&
790 			    !dst_metric_locked(dst, RTAX_SSTHRESH) &&
791 			    tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
792 				dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
793 		}
794 
795 		if (!dst_metric_locked(dst, RTAX_REORDERING)) {
796 			if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
797 			    tp->reordering != sysctl_tcp_reordering)
798 				dst->metrics[RTAX_REORDERING-1] = tp->reordering;
799 		}
800 	}
801 }
802 
803 /* Numbers are taken from RFC3390.
804  *
805  * John Heffner states:
806  *
807  *	The RFC specifies a window of no more than 4380 bytes
808  *	unless 2*MSS > 4380.  Reading the pseudocode in the RFC
809  *	is a bit misleading because they use a clamp at 4380 bytes
810  *	rather than use a multiplier in the relevant range.
811  */
812 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
813 {
814 	__u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
815 
816 	if (!cwnd) {
817 		if (tp->mss_cache > 1460)
818 			cwnd = 2;
819 		else
820 			cwnd = (tp->mss_cache > 1095) ? 3 : 4;
821 	}
822 	return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
823 }
824 
825 /* Set slow start threshold and cwnd not falling to slow start */
826 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
827 {
828 	struct tcp_sock *tp = tcp_sk(sk);
829 	const struct inet_connection_sock *icsk = inet_csk(sk);
830 
831 	tp->prior_ssthresh = 0;
832 	tp->bytes_acked = 0;
833 	if (icsk->icsk_ca_state < TCP_CA_CWR) {
834 		tp->undo_marker = 0;
835 		if (set_ssthresh)
836 			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
837 		tp->snd_cwnd = min(tp->snd_cwnd,
838 				   tcp_packets_in_flight(tp) + 1U);
839 		tp->snd_cwnd_cnt = 0;
840 		tp->high_seq = tp->snd_nxt;
841 		tp->snd_cwnd_stamp = tcp_time_stamp;
842 		TCP_ECN_queue_cwr(tp);
843 
844 		tcp_set_ca_state(sk, TCP_CA_CWR);
845 	}
846 }
847 
848 /*
849  * Packet counting of FACK is based on in-order assumptions, therefore TCP
850  * disables it when reordering is detected
851  */
852 static void tcp_disable_fack(struct tcp_sock *tp)
853 {
854 	/* RFC3517 uses different metric in lost marker => reset on change */
855 	if (tcp_is_fack(tp))
856 		tp->lost_skb_hint = NULL;
857 	tp->rx_opt.sack_ok &= ~2;
858 }
859 
860 /* Take a notice that peer is sending D-SACKs */
861 static void tcp_dsack_seen(struct tcp_sock *tp)
862 {
863 	tp->rx_opt.sack_ok |= 4;
864 }
865 
866 /* Initialize metrics on socket. */
867 
868 static void tcp_init_metrics(struct sock *sk)
869 {
870 	struct tcp_sock *tp = tcp_sk(sk);
871 	struct dst_entry *dst = __sk_dst_get(sk);
872 
873 	if (dst == NULL)
874 		goto reset;
875 
876 	dst_confirm(dst);
877 
878 	if (dst_metric_locked(dst, RTAX_CWND))
879 		tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
880 	if (dst_metric(dst, RTAX_SSTHRESH)) {
881 		tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
882 		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
883 			tp->snd_ssthresh = tp->snd_cwnd_clamp;
884 	}
885 	if (dst_metric(dst, RTAX_REORDERING) &&
886 	    tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
887 		tcp_disable_fack(tp);
888 		tp->reordering = dst_metric(dst, RTAX_REORDERING);
889 	}
890 
891 	if (dst_metric(dst, RTAX_RTT) == 0)
892 		goto reset;
893 
894 	if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
895 		goto reset;
896 
897 	/* Initial rtt is determined from SYN,SYN-ACK.
898 	 * The segment is small and rtt may appear much
899 	 * less than real one. Use per-dst memory
900 	 * to make it more realistic.
901 	 *
902 	 * A bit of theory. RTT is time passed after "normal" sized packet
903 	 * is sent until it is ACKed. In normal circumstances sending small
904 	 * packets force peer to delay ACKs and calculation is correct too.
905 	 * The algorithm is adaptive and, provided we follow specs, it
906 	 * NEVER underestimate RTT. BUT! If peer tries to make some clever
907 	 * tricks sort of "quick acks" for time long enough to decrease RTT
908 	 * to low value, and then abruptly stops to do it and starts to delay
909 	 * ACKs, wait for troubles.
910 	 */
911 	if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
912 		tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
913 		tp->rtt_seq = tp->snd_nxt;
914 	}
915 	if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
916 		tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
917 		tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
918 	}
919 	tcp_set_rto(sk);
920 	if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
921 		goto reset;
922 
923 cwnd:
924 	tp->snd_cwnd = tcp_init_cwnd(tp, dst);
925 	tp->snd_cwnd_stamp = tcp_time_stamp;
926 	return;
927 
928 reset:
929 	/* Play conservative. If timestamps are not
930 	 * supported, TCP will fail to recalculate correct
931 	 * rtt, if initial rto is too small. FORGET ALL AND RESET!
932 	 */
933 	if (!tp->rx_opt.saw_tstamp && tp->srtt) {
934 		tp->srtt = 0;
935 		tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
936 		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
937 	}
938 	goto cwnd;
939 }
940 
941 static void tcp_update_reordering(struct sock *sk, const int metric,
942 				  const int ts)
943 {
944 	struct tcp_sock *tp = tcp_sk(sk);
945 	if (metric > tp->reordering) {
946 		int mib_idx;
947 
948 		tp->reordering = min(TCP_MAX_REORDERING, metric);
949 
950 		/* This exciting event is worth to be remembered. 8) */
951 		if (ts)
952 			mib_idx = LINUX_MIB_TCPTSREORDER;
953 		else if (tcp_is_reno(tp))
954 			mib_idx = LINUX_MIB_TCPRENOREORDER;
955 		else if (tcp_is_fack(tp))
956 			mib_idx = LINUX_MIB_TCPFACKREORDER;
957 		else
958 			mib_idx = LINUX_MIB_TCPSACKREORDER;
959 
960 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
961 #if FASTRETRANS_DEBUG > 1
962 		printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
963 		       tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
964 		       tp->reordering,
965 		       tp->fackets_out,
966 		       tp->sacked_out,
967 		       tp->undo_marker ? tp->undo_retrans : 0);
968 #endif
969 		tcp_disable_fack(tp);
970 	}
971 }
972 
973 /* This must be called before lost_out is incremented */
974 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
975 {
976 	if ((tp->retransmit_skb_hint == NULL) ||
977 	    before(TCP_SKB_CB(skb)->seq,
978 		   TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
979 		tp->retransmit_skb_hint = skb;
980 
981 	if (!tp->lost_out ||
982 	    after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
983 		tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
984 }
985 
986 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
987 {
988 	if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
989 		tcp_verify_retransmit_hint(tp, skb);
990 
991 		tp->lost_out += tcp_skb_pcount(skb);
992 		TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
993 	}
994 }
995 
996 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
997 					    struct sk_buff *skb)
998 {
999 	tcp_verify_retransmit_hint(tp, skb);
1000 
1001 	if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1002 		tp->lost_out += tcp_skb_pcount(skb);
1003 		TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1004 	}
1005 }
1006 
1007 /* This procedure tags the retransmission queue when SACKs arrive.
1008  *
1009  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1010  * Packets in queue with these bits set are counted in variables
1011  * sacked_out, retrans_out and lost_out, correspondingly.
1012  *
1013  * Valid combinations are:
1014  * Tag  InFlight	Description
1015  * 0	1		- orig segment is in flight.
1016  * S	0		- nothing flies, orig reached receiver.
1017  * L	0		- nothing flies, orig lost by net.
1018  * R	2		- both orig and retransmit are in flight.
1019  * L|R	1		- orig is lost, retransmit is in flight.
1020  * S|R  1		- orig reached receiver, retrans is still in flight.
1021  * (L|S|R is logically valid, it could occur when L|R is sacked,
1022  *  but it is equivalent to plain S and code short-curcuits it to S.
1023  *  L|S is logically invalid, it would mean -1 packet in flight 8))
1024  *
1025  * These 6 states form finite state machine, controlled by the following events:
1026  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1027  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1028  * 3. Loss detection event of one of three flavors:
1029  *	A. Scoreboard estimator decided the packet is lost.
1030  *	   A'. Reno "three dupacks" marks head of queue lost.
1031  *	   A''. Its FACK modfication, head until snd.fack is lost.
1032  *	B. SACK arrives sacking data transmitted after never retransmitted
1033  *	   hole was sent out.
1034  *	C. SACK arrives sacking SND.NXT at the moment, when the
1035  *	   segment was retransmitted.
1036  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1037  *
1038  * It is pleasant to note, that state diagram turns out to be commutative,
1039  * so that we are allowed not to be bothered by order of our actions,
1040  * when multiple events arrive simultaneously. (see the function below).
1041  *
1042  * Reordering detection.
1043  * --------------------
1044  * Reordering metric is maximal distance, which a packet can be displaced
1045  * in packet stream. With SACKs we can estimate it:
1046  *
1047  * 1. SACK fills old hole and the corresponding segment was not
1048  *    ever retransmitted -> reordering. Alas, we cannot use it
1049  *    when segment was retransmitted.
1050  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1051  *    for retransmitted and already SACKed segment -> reordering..
1052  * Both of these heuristics are not used in Loss state, when we cannot
1053  * account for retransmits accurately.
1054  *
1055  * SACK block validation.
1056  * ----------------------
1057  *
1058  * SACK block range validation checks that the received SACK block fits to
1059  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1060  * Note that SND.UNA is not included to the range though being valid because
1061  * it means that the receiver is rather inconsistent with itself reporting
1062  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1063  * perfectly valid, however, in light of RFC2018 which explicitly states
1064  * that "SACK block MUST reflect the newest segment.  Even if the newest
1065  * segment is going to be discarded ...", not that it looks very clever
1066  * in case of head skb. Due to potentional receiver driven attacks, we
1067  * choose to avoid immediate execution of a walk in write queue due to
1068  * reneging and defer head skb's loss recovery to standard loss recovery
1069  * procedure that will eventually trigger (nothing forbids us doing this).
1070  *
1071  * Implements also blockage to start_seq wrap-around. Problem lies in the
1072  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1073  * there's no guarantee that it will be before snd_nxt (n). The problem
1074  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1075  * wrap (s_w):
1076  *
1077  *         <- outs wnd ->                          <- wrapzone ->
1078  *         u     e      n                         u_w   e_w  s n_w
1079  *         |     |      |                          |     |   |  |
1080  * |<------------+------+----- TCP seqno space --------------+---------->|
1081  * ...-- <2^31 ->|                                           |<--------...
1082  * ...---- >2^31 ------>|                                    |<--------...
1083  *
1084  * Current code wouldn't be vulnerable but it's better still to discard such
1085  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1086  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1087  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1088  * equal to the ideal case (infinite seqno space without wrap caused issues).
1089  *
1090  * With D-SACK the lower bound is extended to cover sequence space below
1091  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1092  * again, D-SACK block must not to go across snd_una (for the same reason as
1093  * for the normal SACK blocks, explained above). But there all simplicity
1094  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1095  * fully below undo_marker they do not affect behavior in anyway and can
1096  * therefore be safely ignored. In rare cases (which are more or less
1097  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1098  * fragmentation and packet reordering past skb's retransmission. To consider
1099  * them correctly, the acceptable range must be extended even more though
1100  * the exact amount is rather hard to quantify. However, tp->max_window can
1101  * be used as an exaggerated estimate.
1102  */
1103 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1104 				  u32 start_seq, u32 end_seq)
1105 {
1106 	/* Too far in future, or reversed (interpretation is ambiguous) */
1107 	if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1108 		return 0;
1109 
1110 	/* Nasty start_seq wrap-around check (see comments above) */
1111 	if (!before(start_seq, tp->snd_nxt))
1112 		return 0;
1113 
1114 	/* In outstanding window? ...This is valid exit for D-SACKs too.
1115 	 * start_seq == snd_una is non-sensical (see comments above)
1116 	 */
1117 	if (after(start_seq, tp->snd_una))
1118 		return 1;
1119 
1120 	if (!is_dsack || !tp->undo_marker)
1121 		return 0;
1122 
1123 	/* ...Then it's D-SACK, and must reside below snd_una completely */
1124 	if (!after(end_seq, tp->snd_una))
1125 		return 0;
1126 
1127 	if (!before(start_seq, tp->undo_marker))
1128 		return 1;
1129 
1130 	/* Too old */
1131 	if (!after(end_seq, tp->undo_marker))
1132 		return 0;
1133 
1134 	/* Undo_marker boundary crossing (overestimates a lot). Known already:
1135 	 *   start_seq < undo_marker and end_seq >= undo_marker.
1136 	 */
1137 	return !before(start_seq, end_seq - tp->max_window);
1138 }
1139 
1140 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1141  * Event "C". Later note: FACK people cheated me again 8), we have to account
1142  * for reordering! Ugly, but should help.
1143  *
1144  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1145  * less than what is now known to be received by the other end (derived from
1146  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1147  * retransmitted skbs to avoid some costly processing per ACKs.
1148  */
1149 static void tcp_mark_lost_retrans(struct sock *sk)
1150 {
1151 	const struct inet_connection_sock *icsk = inet_csk(sk);
1152 	struct tcp_sock *tp = tcp_sk(sk);
1153 	struct sk_buff *skb;
1154 	int cnt = 0;
1155 	u32 new_low_seq = tp->snd_nxt;
1156 	u32 received_upto = tcp_highest_sack_seq(tp);
1157 
1158 	if (!tcp_is_fack(tp) || !tp->retrans_out ||
1159 	    !after(received_upto, tp->lost_retrans_low) ||
1160 	    icsk->icsk_ca_state != TCP_CA_Recovery)
1161 		return;
1162 
1163 	tcp_for_write_queue(skb, sk) {
1164 		u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1165 
1166 		if (skb == tcp_send_head(sk))
1167 			break;
1168 		if (cnt == tp->retrans_out)
1169 			break;
1170 		if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1171 			continue;
1172 
1173 		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1174 			continue;
1175 
1176 		/* TODO: We would like to get rid of tcp_is_fack(tp) only
1177 		 * constraint here (see above) but figuring out that at
1178 		 * least tp->reordering SACK blocks reside between ack_seq
1179 		 * and received_upto is not easy task to do cheaply with
1180 		 * the available datastructures.
1181 		 *
1182 		 * Whether FACK should check here for tp->reordering segs
1183 		 * in-between one could argue for either way (it would be
1184 		 * rather simple to implement as we could count fack_count
1185 		 * during the walk and do tp->fackets_out - fack_count).
1186 		 */
1187 		if (after(received_upto, ack_seq)) {
1188 			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1189 			tp->retrans_out -= tcp_skb_pcount(skb);
1190 
1191 			tcp_skb_mark_lost_uncond_verify(tp, skb);
1192 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1193 		} else {
1194 			if (before(ack_seq, new_low_seq))
1195 				new_low_seq = ack_seq;
1196 			cnt += tcp_skb_pcount(skb);
1197 		}
1198 	}
1199 
1200 	if (tp->retrans_out)
1201 		tp->lost_retrans_low = new_low_seq;
1202 }
1203 
1204 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1205 			   struct tcp_sack_block_wire *sp, int num_sacks,
1206 			   u32 prior_snd_una)
1207 {
1208 	struct tcp_sock *tp = tcp_sk(sk);
1209 	u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1210 	u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1211 	int dup_sack = 0;
1212 
1213 	if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1214 		dup_sack = 1;
1215 		tcp_dsack_seen(tp);
1216 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1217 	} else if (num_sacks > 1) {
1218 		u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1219 		u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1220 
1221 		if (!after(end_seq_0, end_seq_1) &&
1222 		    !before(start_seq_0, start_seq_1)) {
1223 			dup_sack = 1;
1224 			tcp_dsack_seen(tp);
1225 			NET_INC_STATS_BH(sock_net(sk),
1226 					LINUX_MIB_TCPDSACKOFORECV);
1227 		}
1228 	}
1229 
1230 	/* D-SACK for already forgotten data... Do dumb counting. */
1231 	if (dup_sack &&
1232 	    !after(end_seq_0, prior_snd_una) &&
1233 	    after(end_seq_0, tp->undo_marker))
1234 		tp->undo_retrans--;
1235 
1236 	return dup_sack;
1237 }
1238 
1239 struct tcp_sacktag_state {
1240 	int reord;
1241 	int fack_count;
1242 	int flag;
1243 };
1244 
1245 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1246  * the incoming SACK may not exactly match but we can find smaller MSS
1247  * aligned portion of it that matches. Therefore we might need to fragment
1248  * which may fail and creates some hassle (caller must handle error case
1249  * returns).
1250  *
1251  * FIXME: this could be merged to shift decision code
1252  */
1253 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1254 				 u32 start_seq, u32 end_seq)
1255 {
1256 	int in_sack, err;
1257 	unsigned int pkt_len;
1258 	unsigned int mss;
1259 
1260 	in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1261 		  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1262 
1263 	if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1264 	    after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1265 		mss = tcp_skb_mss(skb);
1266 		in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1267 
1268 		if (!in_sack) {
1269 			pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1270 			if (pkt_len < mss)
1271 				pkt_len = mss;
1272 		} else {
1273 			pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1274 			if (pkt_len < mss)
1275 				return -EINVAL;
1276 		}
1277 
1278 		/* Round if necessary so that SACKs cover only full MSSes
1279 		 * and/or the remaining small portion (if present)
1280 		 */
1281 		if (pkt_len > mss) {
1282 			unsigned int new_len = (pkt_len / mss) * mss;
1283 			if (!in_sack && new_len < pkt_len) {
1284 				new_len += mss;
1285 				if (new_len > skb->len)
1286 					return 0;
1287 			}
1288 			pkt_len = new_len;
1289 		}
1290 		err = tcp_fragment(sk, skb, pkt_len, mss);
1291 		if (err < 0)
1292 			return err;
1293 	}
1294 
1295 	return in_sack;
1296 }
1297 
1298 static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1299 			  struct tcp_sacktag_state *state,
1300 			  int dup_sack, int pcount)
1301 {
1302 	struct tcp_sock *tp = tcp_sk(sk);
1303 	u8 sacked = TCP_SKB_CB(skb)->sacked;
1304 	int fack_count = state->fack_count;
1305 
1306 	/* Account D-SACK for retransmitted packet. */
1307 	if (dup_sack && (sacked & TCPCB_RETRANS)) {
1308 		if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1309 			tp->undo_retrans--;
1310 		if (sacked & TCPCB_SACKED_ACKED)
1311 			state->reord = min(fack_count, state->reord);
1312 	}
1313 
1314 	/* Nothing to do; acked frame is about to be dropped (was ACKed). */
1315 	if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1316 		return sacked;
1317 
1318 	if (!(sacked & TCPCB_SACKED_ACKED)) {
1319 		if (sacked & TCPCB_SACKED_RETRANS) {
1320 			/* If the segment is not tagged as lost,
1321 			 * we do not clear RETRANS, believing
1322 			 * that retransmission is still in flight.
1323 			 */
1324 			if (sacked & TCPCB_LOST) {
1325 				sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1326 				tp->lost_out -= pcount;
1327 				tp->retrans_out -= pcount;
1328 			}
1329 		} else {
1330 			if (!(sacked & TCPCB_RETRANS)) {
1331 				/* New sack for not retransmitted frame,
1332 				 * which was in hole. It is reordering.
1333 				 */
1334 				if (before(TCP_SKB_CB(skb)->seq,
1335 					   tcp_highest_sack_seq(tp)))
1336 					state->reord = min(fack_count,
1337 							   state->reord);
1338 
1339 				/* SACK enhanced F-RTO (RFC4138; Appendix B) */
1340 				if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1341 					state->flag |= FLAG_ONLY_ORIG_SACKED;
1342 			}
1343 
1344 			if (sacked & TCPCB_LOST) {
1345 				sacked &= ~TCPCB_LOST;
1346 				tp->lost_out -= pcount;
1347 			}
1348 		}
1349 
1350 		sacked |= TCPCB_SACKED_ACKED;
1351 		state->flag |= FLAG_DATA_SACKED;
1352 		tp->sacked_out += pcount;
1353 
1354 		fack_count += pcount;
1355 
1356 		/* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1357 		if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1358 		    before(TCP_SKB_CB(skb)->seq,
1359 			   TCP_SKB_CB(tp->lost_skb_hint)->seq))
1360 			tp->lost_cnt_hint += pcount;
1361 
1362 		if (fack_count > tp->fackets_out)
1363 			tp->fackets_out = fack_count;
1364 	}
1365 
1366 	/* D-SACK. We can detect redundant retransmission in S|R and plain R
1367 	 * frames and clear it. undo_retrans is decreased above, L|R frames
1368 	 * are accounted above as well.
1369 	 */
1370 	if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1371 		sacked &= ~TCPCB_SACKED_RETRANS;
1372 		tp->retrans_out -= pcount;
1373 	}
1374 
1375 	return sacked;
1376 }
1377 
1378 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1379 			   struct tcp_sacktag_state *state,
1380 			   unsigned int pcount, int shifted, int mss,
1381 			   int dup_sack)
1382 {
1383 	struct tcp_sock *tp = tcp_sk(sk);
1384 	struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1385 
1386 	BUG_ON(!pcount);
1387 
1388 	/* Tweak before seqno plays */
1389 	if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint &&
1390 	    !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
1391 		tp->lost_cnt_hint += pcount;
1392 
1393 	TCP_SKB_CB(prev)->end_seq += shifted;
1394 	TCP_SKB_CB(skb)->seq += shifted;
1395 
1396 	skb_shinfo(prev)->gso_segs += pcount;
1397 	BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1398 	skb_shinfo(skb)->gso_segs -= pcount;
1399 
1400 	/* When we're adding to gso_segs == 1, gso_size will be zero,
1401 	 * in theory this shouldn't be necessary but as long as DSACK
1402 	 * code can come after this skb later on it's better to keep
1403 	 * setting gso_size to something.
1404 	 */
1405 	if (!skb_shinfo(prev)->gso_size) {
1406 		skb_shinfo(prev)->gso_size = mss;
1407 		skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1408 	}
1409 
1410 	/* CHECKME: To clear or not to clear? Mimics normal skb currently */
1411 	if (skb_shinfo(skb)->gso_segs <= 1) {
1412 		skb_shinfo(skb)->gso_size = 0;
1413 		skb_shinfo(skb)->gso_type = 0;
1414 	}
1415 
1416 	/* We discard results */
1417 	tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
1418 
1419 	/* Difference in this won't matter, both ACKed by the same cumul. ACK */
1420 	TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1421 
1422 	if (skb->len > 0) {
1423 		BUG_ON(!tcp_skb_pcount(skb));
1424 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1425 		return 0;
1426 	}
1427 
1428 	/* Whole SKB was eaten :-) */
1429 
1430 	if (skb == tp->retransmit_skb_hint)
1431 		tp->retransmit_skb_hint = prev;
1432 	if (skb == tp->scoreboard_skb_hint)
1433 		tp->scoreboard_skb_hint = prev;
1434 	if (skb == tp->lost_skb_hint) {
1435 		tp->lost_skb_hint = prev;
1436 		tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1437 	}
1438 
1439 	TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1440 	if (skb == tcp_highest_sack(sk))
1441 		tcp_advance_highest_sack(sk, skb);
1442 
1443 	tcp_unlink_write_queue(skb, sk);
1444 	sk_wmem_free_skb(sk, skb);
1445 
1446 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1447 
1448 	return 1;
1449 }
1450 
1451 /* I wish gso_size would have a bit more sane initialization than
1452  * something-or-zero which complicates things
1453  */
1454 static int tcp_skb_seglen(struct sk_buff *skb)
1455 {
1456 	return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1457 }
1458 
1459 /* Shifting pages past head area doesn't work */
1460 static int skb_can_shift(struct sk_buff *skb)
1461 {
1462 	return !skb_headlen(skb) && skb_is_nonlinear(skb);
1463 }
1464 
1465 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1466  * skb.
1467  */
1468 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1469 					  struct tcp_sacktag_state *state,
1470 					  u32 start_seq, u32 end_seq,
1471 					  int dup_sack)
1472 {
1473 	struct tcp_sock *tp = tcp_sk(sk);
1474 	struct sk_buff *prev;
1475 	int mss;
1476 	int pcount = 0;
1477 	int len;
1478 	int in_sack;
1479 
1480 	if (!sk_can_gso(sk))
1481 		goto fallback;
1482 
1483 	/* Normally R but no L won't result in plain S */
1484 	if (!dup_sack &&
1485 	    (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1486 		goto fallback;
1487 	if (!skb_can_shift(skb))
1488 		goto fallback;
1489 	/* This frame is about to be dropped (was ACKed). */
1490 	if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1491 		goto fallback;
1492 
1493 	/* Can only happen with delayed DSACK + discard craziness */
1494 	if (unlikely(skb == tcp_write_queue_head(sk)))
1495 		goto fallback;
1496 	prev = tcp_write_queue_prev(sk, skb);
1497 
1498 	if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1499 		goto fallback;
1500 
1501 	in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1502 		  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1503 
1504 	if (in_sack) {
1505 		len = skb->len;
1506 		pcount = tcp_skb_pcount(skb);
1507 		mss = tcp_skb_seglen(skb);
1508 
1509 		/* TODO: Fix DSACKs to not fragment already SACKed and we can
1510 		 * drop this restriction as unnecessary
1511 		 */
1512 		if (mss != tcp_skb_seglen(prev))
1513 			goto fallback;
1514 	} else {
1515 		if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1516 			goto noop;
1517 		/* CHECKME: This is non-MSS split case only?, this will
1518 		 * cause skipped skbs due to advancing loop btw, original
1519 		 * has that feature too
1520 		 */
1521 		if (tcp_skb_pcount(skb) <= 1)
1522 			goto noop;
1523 
1524 		in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1525 		if (!in_sack) {
1526 			/* TODO: head merge to next could be attempted here
1527 			 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1528 			 * though it might not be worth of the additional hassle
1529 			 *
1530 			 * ...we can probably just fallback to what was done
1531 			 * previously. We could try merging non-SACKed ones
1532 			 * as well but it probably isn't going to buy off
1533 			 * because later SACKs might again split them, and
1534 			 * it would make skb timestamp tracking considerably
1535 			 * harder problem.
1536 			 */
1537 			goto fallback;
1538 		}
1539 
1540 		len = end_seq - TCP_SKB_CB(skb)->seq;
1541 		BUG_ON(len < 0);
1542 		BUG_ON(len > skb->len);
1543 
1544 		/* MSS boundaries should be honoured or else pcount will
1545 		 * severely break even though it makes things bit trickier.
1546 		 * Optimize common case to avoid most of the divides
1547 		 */
1548 		mss = tcp_skb_mss(skb);
1549 
1550 		/* TODO: Fix DSACKs to not fragment already SACKed and we can
1551 		 * drop this restriction as unnecessary
1552 		 */
1553 		if (mss != tcp_skb_seglen(prev))
1554 			goto fallback;
1555 
1556 		if (len == mss) {
1557 			pcount = 1;
1558 		} else if (len < mss) {
1559 			goto noop;
1560 		} else {
1561 			pcount = len / mss;
1562 			len = pcount * mss;
1563 		}
1564 	}
1565 
1566 	if (!skb_shift(prev, skb, len))
1567 		goto fallback;
1568 	if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1569 		goto out;
1570 
1571 	/* Hole filled allows collapsing with the next as well, this is very
1572 	 * useful when hole on every nth skb pattern happens
1573 	 */
1574 	if (prev == tcp_write_queue_tail(sk))
1575 		goto out;
1576 	skb = tcp_write_queue_next(sk, prev);
1577 
1578 	if (!skb_can_shift(skb) ||
1579 	    (skb == tcp_send_head(sk)) ||
1580 	    ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1581 	    (mss != tcp_skb_seglen(skb)))
1582 		goto out;
1583 
1584 	len = skb->len;
1585 	if (skb_shift(prev, skb, len)) {
1586 		pcount += tcp_skb_pcount(skb);
1587 		tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1588 	}
1589 
1590 out:
1591 	state->fack_count += pcount;
1592 	return prev;
1593 
1594 noop:
1595 	return skb;
1596 
1597 fallback:
1598 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1599 	return NULL;
1600 }
1601 
1602 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1603 					struct tcp_sack_block *next_dup,
1604 					struct tcp_sacktag_state *state,
1605 					u32 start_seq, u32 end_seq,
1606 					int dup_sack_in)
1607 {
1608 	struct tcp_sock *tp = tcp_sk(sk);
1609 	struct sk_buff *tmp;
1610 
1611 	tcp_for_write_queue_from(skb, sk) {
1612 		int in_sack = 0;
1613 		int dup_sack = dup_sack_in;
1614 
1615 		if (skb == tcp_send_head(sk))
1616 			break;
1617 
1618 		/* queue is in-order => we can short-circuit the walk early */
1619 		if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1620 			break;
1621 
1622 		if ((next_dup != NULL) &&
1623 		    before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1624 			in_sack = tcp_match_skb_to_sack(sk, skb,
1625 							next_dup->start_seq,
1626 							next_dup->end_seq);
1627 			if (in_sack > 0)
1628 				dup_sack = 1;
1629 		}
1630 
1631 		/* skb reference here is a bit tricky to get right, since
1632 		 * shifting can eat and free both this skb and the next,
1633 		 * so not even _safe variant of the loop is enough.
1634 		 */
1635 		if (in_sack <= 0) {
1636 			tmp = tcp_shift_skb_data(sk, skb, state,
1637 						 start_seq, end_seq, dup_sack);
1638 			if (tmp != NULL) {
1639 				if (tmp != skb) {
1640 					skb = tmp;
1641 					continue;
1642 				}
1643 
1644 				in_sack = 0;
1645 			} else {
1646 				in_sack = tcp_match_skb_to_sack(sk, skb,
1647 								start_seq,
1648 								end_seq);
1649 			}
1650 		}
1651 
1652 		if (unlikely(in_sack < 0))
1653 			break;
1654 
1655 		if (in_sack) {
1656 			TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1657 								  state,
1658 								  dup_sack,
1659 								  tcp_skb_pcount(skb));
1660 
1661 			if (!before(TCP_SKB_CB(skb)->seq,
1662 				    tcp_highest_sack_seq(tp)))
1663 				tcp_advance_highest_sack(sk, skb);
1664 		}
1665 
1666 		state->fack_count += tcp_skb_pcount(skb);
1667 	}
1668 	return skb;
1669 }
1670 
1671 /* Avoid all extra work that is being done by sacktag while walking in
1672  * a normal way
1673  */
1674 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1675 					struct tcp_sacktag_state *state,
1676 					u32 skip_to_seq)
1677 {
1678 	tcp_for_write_queue_from(skb, sk) {
1679 		if (skb == tcp_send_head(sk))
1680 			break;
1681 
1682 		if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1683 			break;
1684 
1685 		state->fack_count += tcp_skb_pcount(skb);
1686 	}
1687 	return skb;
1688 }
1689 
1690 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1691 						struct sock *sk,
1692 						struct tcp_sack_block *next_dup,
1693 						struct tcp_sacktag_state *state,
1694 						u32 skip_to_seq)
1695 {
1696 	if (next_dup == NULL)
1697 		return skb;
1698 
1699 	if (before(next_dup->start_seq, skip_to_seq)) {
1700 		skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1701 		skb = tcp_sacktag_walk(skb, sk, NULL, state,
1702 				       next_dup->start_seq, next_dup->end_seq,
1703 				       1);
1704 	}
1705 
1706 	return skb;
1707 }
1708 
1709 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1710 {
1711 	return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1712 }
1713 
1714 static int
1715 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1716 			u32 prior_snd_una)
1717 {
1718 	const struct inet_connection_sock *icsk = inet_csk(sk);
1719 	struct tcp_sock *tp = tcp_sk(sk);
1720 	unsigned char *ptr = (skb_transport_header(ack_skb) +
1721 			      TCP_SKB_CB(ack_skb)->sacked);
1722 	struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1723 	struct tcp_sack_block sp[TCP_NUM_SACKS];
1724 	struct tcp_sack_block *cache;
1725 	struct tcp_sacktag_state state;
1726 	struct sk_buff *skb;
1727 	int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1728 	int used_sacks;
1729 	int found_dup_sack = 0;
1730 	int i, j;
1731 	int first_sack_index;
1732 
1733 	state.flag = 0;
1734 	state.reord = tp->packets_out;
1735 
1736 	if (!tp->sacked_out) {
1737 		if (WARN_ON(tp->fackets_out))
1738 			tp->fackets_out = 0;
1739 		tcp_highest_sack_reset(sk);
1740 	}
1741 
1742 	found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1743 					 num_sacks, prior_snd_una);
1744 	if (found_dup_sack)
1745 		state.flag |= FLAG_DSACKING_ACK;
1746 
1747 	/* Eliminate too old ACKs, but take into
1748 	 * account more or less fresh ones, they can
1749 	 * contain valid SACK info.
1750 	 */
1751 	if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1752 		return 0;
1753 
1754 	if (!tp->packets_out)
1755 		goto out;
1756 
1757 	used_sacks = 0;
1758 	first_sack_index = 0;
1759 	for (i = 0; i < num_sacks; i++) {
1760 		int dup_sack = !i && found_dup_sack;
1761 
1762 		sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1763 		sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1764 
1765 		if (!tcp_is_sackblock_valid(tp, dup_sack,
1766 					    sp[used_sacks].start_seq,
1767 					    sp[used_sacks].end_seq)) {
1768 			int mib_idx;
1769 
1770 			if (dup_sack) {
1771 				if (!tp->undo_marker)
1772 					mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1773 				else
1774 					mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1775 			} else {
1776 				/* Don't count olds caused by ACK reordering */
1777 				if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1778 				    !after(sp[used_sacks].end_seq, tp->snd_una))
1779 					continue;
1780 				mib_idx = LINUX_MIB_TCPSACKDISCARD;
1781 			}
1782 
1783 			NET_INC_STATS_BH(sock_net(sk), mib_idx);
1784 			if (i == 0)
1785 				first_sack_index = -1;
1786 			continue;
1787 		}
1788 
1789 		/* Ignore very old stuff early */
1790 		if (!after(sp[used_sacks].end_seq, prior_snd_una))
1791 			continue;
1792 
1793 		used_sacks++;
1794 	}
1795 
1796 	/* order SACK blocks to allow in order walk of the retrans queue */
1797 	for (i = used_sacks - 1; i > 0; i--) {
1798 		for (j = 0; j < i; j++) {
1799 			if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1800 				swap(sp[j], sp[j + 1]);
1801 
1802 				/* Track where the first SACK block goes to */
1803 				if (j == first_sack_index)
1804 					first_sack_index = j + 1;
1805 			}
1806 		}
1807 	}
1808 
1809 	skb = tcp_write_queue_head(sk);
1810 	state.fack_count = 0;
1811 	i = 0;
1812 
1813 	if (!tp->sacked_out) {
1814 		/* It's already past, so skip checking against it */
1815 		cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1816 	} else {
1817 		cache = tp->recv_sack_cache;
1818 		/* Skip empty blocks in at head of the cache */
1819 		while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1820 		       !cache->end_seq)
1821 			cache++;
1822 	}
1823 
1824 	while (i < used_sacks) {
1825 		u32 start_seq = sp[i].start_seq;
1826 		u32 end_seq = sp[i].end_seq;
1827 		int dup_sack = (found_dup_sack && (i == first_sack_index));
1828 		struct tcp_sack_block *next_dup = NULL;
1829 
1830 		if (found_dup_sack && ((i + 1) == first_sack_index))
1831 			next_dup = &sp[i + 1];
1832 
1833 		/* Event "B" in the comment above. */
1834 		if (after(end_seq, tp->high_seq))
1835 			state.flag |= FLAG_DATA_LOST;
1836 
1837 		/* Skip too early cached blocks */
1838 		while (tcp_sack_cache_ok(tp, cache) &&
1839 		       !before(start_seq, cache->end_seq))
1840 			cache++;
1841 
1842 		/* Can skip some work by looking recv_sack_cache? */
1843 		if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1844 		    after(end_seq, cache->start_seq)) {
1845 
1846 			/* Head todo? */
1847 			if (before(start_seq, cache->start_seq)) {
1848 				skb = tcp_sacktag_skip(skb, sk, &state,
1849 						       start_seq);
1850 				skb = tcp_sacktag_walk(skb, sk, next_dup,
1851 						       &state,
1852 						       start_seq,
1853 						       cache->start_seq,
1854 						       dup_sack);
1855 			}
1856 
1857 			/* Rest of the block already fully processed? */
1858 			if (!after(end_seq, cache->end_seq))
1859 				goto advance_sp;
1860 
1861 			skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1862 						       &state,
1863 						       cache->end_seq);
1864 
1865 			/* ...tail remains todo... */
1866 			if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1867 				/* ...but better entrypoint exists! */
1868 				skb = tcp_highest_sack(sk);
1869 				if (skb == NULL)
1870 					break;
1871 				state.fack_count = tp->fackets_out;
1872 				cache++;
1873 				goto walk;
1874 			}
1875 
1876 			skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1877 			/* Check overlap against next cached too (past this one already) */
1878 			cache++;
1879 			continue;
1880 		}
1881 
1882 		if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1883 			skb = tcp_highest_sack(sk);
1884 			if (skb == NULL)
1885 				break;
1886 			state.fack_count = tp->fackets_out;
1887 		}
1888 		skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1889 
1890 walk:
1891 		skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1892 				       start_seq, end_seq, dup_sack);
1893 
1894 advance_sp:
1895 		/* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1896 		 * due to in-order walk
1897 		 */
1898 		if (after(end_seq, tp->frto_highmark))
1899 			state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1900 
1901 		i++;
1902 	}
1903 
1904 	/* Clear the head of the cache sack blocks so we can skip it next time */
1905 	for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1906 		tp->recv_sack_cache[i].start_seq = 0;
1907 		tp->recv_sack_cache[i].end_seq = 0;
1908 	}
1909 	for (j = 0; j < used_sacks; j++)
1910 		tp->recv_sack_cache[i++] = sp[j];
1911 
1912 	tcp_mark_lost_retrans(sk);
1913 
1914 	tcp_verify_left_out(tp);
1915 
1916 	if ((state.reord < tp->fackets_out) &&
1917 	    ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1918 	    (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1919 		tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1920 
1921 out:
1922 
1923 #if FASTRETRANS_DEBUG > 0
1924 	WARN_ON((int)tp->sacked_out < 0);
1925 	WARN_ON((int)tp->lost_out < 0);
1926 	WARN_ON((int)tp->retrans_out < 0);
1927 	WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1928 #endif
1929 	return state.flag;
1930 }
1931 
1932 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1933  * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1934  */
1935 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1936 {
1937 	u32 holes;
1938 
1939 	holes = max(tp->lost_out, 1U);
1940 	holes = min(holes, tp->packets_out);
1941 
1942 	if ((tp->sacked_out + holes) > tp->packets_out) {
1943 		tp->sacked_out = tp->packets_out - holes;
1944 		return 1;
1945 	}
1946 	return 0;
1947 }
1948 
1949 /* If we receive more dupacks than we expected counting segments
1950  * in assumption of absent reordering, interpret this as reordering.
1951  * The only another reason could be bug in receiver TCP.
1952  */
1953 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1954 {
1955 	struct tcp_sock *tp = tcp_sk(sk);
1956 	if (tcp_limit_reno_sacked(tp))
1957 		tcp_update_reordering(sk, tp->packets_out + addend, 0);
1958 }
1959 
1960 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1961 
1962 static void tcp_add_reno_sack(struct sock *sk)
1963 {
1964 	struct tcp_sock *tp = tcp_sk(sk);
1965 	tp->sacked_out++;
1966 	tcp_check_reno_reordering(sk, 0);
1967 	tcp_verify_left_out(tp);
1968 }
1969 
1970 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1971 
1972 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1973 {
1974 	struct tcp_sock *tp = tcp_sk(sk);
1975 
1976 	if (acked > 0) {
1977 		/* One ACK acked hole. The rest eat duplicate ACKs. */
1978 		if (acked - 1 >= tp->sacked_out)
1979 			tp->sacked_out = 0;
1980 		else
1981 			tp->sacked_out -= acked - 1;
1982 	}
1983 	tcp_check_reno_reordering(sk, acked);
1984 	tcp_verify_left_out(tp);
1985 }
1986 
1987 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1988 {
1989 	tp->sacked_out = 0;
1990 }
1991 
1992 static int tcp_is_sackfrto(const struct tcp_sock *tp)
1993 {
1994 	return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1995 }
1996 
1997 /* F-RTO can only be used if TCP has never retransmitted anything other than
1998  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1999  */
2000 int tcp_use_frto(struct sock *sk)
2001 {
2002 	const struct tcp_sock *tp = tcp_sk(sk);
2003 	const struct inet_connection_sock *icsk = inet_csk(sk);
2004 	struct sk_buff *skb;
2005 
2006 	if (!sysctl_tcp_frto)
2007 		return 0;
2008 
2009 	/* MTU probe and F-RTO won't really play nicely along currently */
2010 	if (icsk->icsk_mtup.probe_size)
2011 		return 0;
2012 
2013 	if (tcp_is_sackfrto(tp))
2014 		return 1;
2015 
2016 	/* Avoid expensive walking of rexmit queue if possible */
2017 	if (tp->retrans_out > 1)
2018 		return 0;
2019 
2020 	skb = tcp_write_queue_head(sk);
2021 	if (tcp_skb_is_last(sk, skb))
2022 		return 1;
2023 	skb = tcp_write_queue_next(sk, skb);	/* Skips head */
2024 	tcp_for_write_queue_from(skb, sk) {
2025 		if (skb == tcp_send_head(sk))
2026 			break;
2027 		if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2028 			return 0;
2029 		/* Short-circuit when first non-SACKed skb has been checked */
2030 		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2031 			break;
2032 	}
2033 	return 1;
2034 }
2035 
2036 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2037  * recovery a bit and use heuristics in tcp_process_frto() to detect if
2038  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2039  * keep retrans_out counting accurate (with SACK F-RTO, other than head
2040  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2041  * bits are handled if the Loss state is really to be entered (in
2042  * tcp_enter_frto_loss).
2043  *
2044  * Do like tcp_enter_loss() would; when RTO expires the second time it
2045  * does:
2046  *  "Reduce ssthresh if it has not yet been made inside this window."
2047  */
2048 void tcp_enter_frto(struct sock *sk)
2049 {
2050 	const struct inet_connection_sock *icsk = inet_csk(sk);
2051 	struct tcp_sock *tp = tcp_sk(sk);
2052 	struct sk_buff *skb;
2053 
2054 	if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2055 	    tp->snd_una == tp->high_seq ||
2056 	    ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2057 	     !icsk->icsk_retransmits)) {
2058 		tp->prior_ssthresh = tcp_current_ssthresh(sk);
2059 		/* Our state is too optimistic in ssthresh() call because cwnd
2060 		 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2061 		 * recovery has not yet completed. Pattern would be this: RTO,
2062 		 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2063 		 * up here twice).
2064 		 * RFC4138 should be more specific on what to do, even though
2065 		 * RTO is quite unlikely to occur after the first Cumulative ACK
2066 		 * due to back-off and complexity of triggering events ...
2067 		 */
2068 		if (tp->frto_counter) {
2069 			u32 stored_cwnd;
2070 			stored_cwnd = tp->snd_cwnd;
2071 			tp->snd_cwnd = 2;
2072 			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2073 			tp->snd_cwnd = stored_cwnd;
2074 		} else {
2075 			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2076 		}
2077 		/* ... in theory, cong.control module could do "any tricks" in
2078 		 * ssthresh(), which means that ca_state, lost bits and lost_out
2079 		 * counter would have to be faked before the call occurs. We
2080 		 * consider that too expensive, unlikely and hacky, so modules
2081 		 * using these in ssthresh() must deal these incompatibility
2082 		 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2083 		 */
2084 		tcp_ca_event(sk, CA_EVENT_FRTO);
2085 	}
2086 
2087 	tp->undo_marker = tp->snd_una;
2088 	tp->undo_retrans = 0;
2089 
2090 	skb = tcp_write_queue_head(sk);
2091 	if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2092 		tp->undo_marker = 0;
2093 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2094 		TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2095 		tp->retrans_out -= tcp_skb_pcount(skb);
2096 	}
2097 	tcp_verify_left_out(tp);
2098 
2099 	/* Too bad if TCP was application limited */
2100 	tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2101 
2102 	/* Earlier loss recovery underway (see RFC4138; Appendix B).
2103 	 * The last condition is necessary at least in tp->frto_counter case.
2104 	 */
2105 	if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2106 	    ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2107 	    after(tp->high_seq, tp->snd_una)) {
2108 		tp->frto_highmark = tp->high_seq;
2109 	} else {
2110 		tp->frto_highmark = tp->snd_nxt;
2111 	}
2112 	tcp_set_ca_state(sk, TCP_CA_Disorder);
2113 	tp->high_seq = tp->snd_nxt;
2114 	tp->frto_counter = 1;
2115 }
2116 
2117 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2118  * which indicates that we should follow the traditional RTO recovery,
2119  * i.e. mark everything lost and do go-back-N retransmission.
2120  */
2121 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2122 {
2123 	struct tcp_sock *tp = tcp_sk(sk);
2124 	struct sk_buff *skb;
2125 
2126 	tp->lost_out = 0;
2127 	tp->retrans_out = 0;
2128 	if (tcp_is_reno(tp))
2129 		tcp_reset_reno_sack(tp);
2130 
2131 	tcp_for_write_queue(skb, sk) {
2132 		if (skb == tcp_send_head(sk))
2133 			break;
2134 
2135 		TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2136 		/*
2137 		 * Count the retransmission made on RTO correctly (only when
2138 		 * waiting for the first ACK and did not get it)...
2139 		 */
2140 		if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2141 			/* For some reason this R-bit might get cleared? */
2142 			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2143 				tp->retrans_out += tcp_skb_pcount(skb);
2144 			/* ...enter this if branch just for the first segment */
2145 			flag |= FLAG_DATA_ACKED;
2146 		} else {
2147 			if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2148 				tp->undo_marker = 0;
2149 			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2150 		}
2151 
2152 		/* Marking forward transmissions that were made after RTO lost
2153 		 * can cause unnecessary retransmissions in some scenarios,
2154 		 * SACK blocks will mitigate that in some but not in all cases.
2155 		 * We used to not mark them but it was causing break-ups with
2156 		 * receivers that do only in-order receival.
2157 		 *
2158 		 * TODO: we could detect presence of such receiver and select
2159 		 * different behavior per flow.
2160 		 */
2161 		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2162 			TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2163 			tp->lost_out += tcp_skb_pcount(skb);
2164 			tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2165 		}
2166 	}
2167 	tcp_verify_left_out(tp);
2168 
2169 	tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2170 	tp->snd_cwnd_cnt = 0;
2171 	tp->snd_cwnd_stamp = tcp_time_stamp;
2172 	tp->frto_counter = 0;
2173 	tp->bytes_acked = 0;
2174 
2175 	tp->reordering = min_t(unsigned int, tp->reordering,
2176 			       sysctl_tcp_reordering);
2177 	tcp_set_ca_state(sk, TCP_CA_Loss);
2178 	tp->high_seq = tp->snd_nxt;
2179 	TCP_ECN_queue_cwr(tp);
2180 
2181 	tcp_clear_all_retrans_hints(tp);
2182 }
2183 
2184 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2185 {
2186 	tp->retrans_out = 0;
2187 	tp->lost_out = 0;
2188 
2189 	tp->undo_marker = 0;
2190 	tp->undo_retrans = 0;
2191 }
2192 
2193 void tcp_clear_retrans(struct tcp_sock *tp)
2194 {
2195 	tcp_clear_retrans_partial(tp);
2196 
2197 	tp->fackets_out = 0;
2198 	tp->sacked_out = 0;
2199 }
2200 
2201 /* Enter Loss state. If "how" is not zero, forget all SACK information
2202  * and reset tags completely, otherwise preserve SACKs. If receiver
2203  * dropped its ofo queue, we will know this due to reneging detection.
2204  */
2205 void tcp_enter_loss(struct sock *sk, int how)
2206 {
2207 	const struct inet_connection_sock *icsk = inet_csk(sk);
2208 	struct tcp_sock *tp = tcp_sk(sk);
2209 	struct sk_buff *skb;
2210 
2211 	/* Reduce ssthresh if it has not yet been made inside this window. */
2212 	if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2213 	    (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2214 		tp->prior_ssthresh = tcp_current_ssthresh(sk);
2215 		tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2216 		tcp_ca_event(sk, CA_EVENT_LOSS);
2217 	}
2218 	tp->snd_cwnd	   = 1;
2219 	tp->snd_cwnd_cnt   = 0;
2220 	tp->snd_cwnd_stamp = tcp_time_stamp;
2221 
2222 	tp->bytes_acked = 0;
2223 	tcp_clear_retrans_partial(tp);
2224 
2225 	if (tcp_is_reno(tp))
2226 		tcp_reset_reno_sack(tp);
2227 
2228 	if (!how) {
2229 		/* Push undo marker, if it was plain RTO and nothing
2230 		 * was retransmitted. */
2231 		tp->undo_marker = tp->snd_una;
2232 	} else {
2233 		tp->sacked_out = 0;
2234 		tp->fackets_out = 0;
2235 	}
2236 	tcp_clear_all_retrans_hints(tp);
2237 
2238 	tcp_for_write_queue(skb, sk) {
2239 		if (skb == tcp_send_head(sk))
2240 			break;
2241 
2242 		if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2243 			tp->undo_marker = 0;
2244 		TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2245 		if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2246 			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2247 			TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2248 			tp->lost_out += tcp_skb_pcount(skb);
2249 			tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2250 		}
2251 	}
2252 	tcp_verify_left_out(tp);
2253 
2254 	tp->reordering = min_t(unsigned int, tp->reordering,
2255 			       sysctl_tcp_reordering);
2256 	tcp_set_ca_state(sk, TCP_CA_Loss);
2257 	tp->high_seq = tp->snd_nxt;
2258 	TCP_ECN_queue_cwr(tp);
2259 	/* Abort F-RTO algorithm if one is in progress */
2260 	tp->frto_counter = 0;
2261 }
2262 
2263 /* If ACK arrived pointing to a remembered SACK, it means that our
2264  * remembered SACKs do not reflect real state of receiver i.e.
2265  * receiver _host_ is heavily congested (or buggy).
2266  *
2267  * Do processing similar to RTO timeout.
2268  */
2269 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2270 {
2271 	if (flag & FLAG_SACK_RENEGING) {
2272 		struct inet_connection_sock *icsk = inet_csk(sk);
2273 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2274 
2275 		tcp_enter_loss(sk, 1);
2276 		icsk->icsk_retransmits++;
2277 		tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2278 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2279 					  icsk->icsk_rto, TCP_RTO_MAX);
2280 		return 1;
2281 	}
2282 	return 0;
2283 }
2284 
2285 static inline int tcp_fackets_out(struct tcp_sock *tp)
2286 {
2287 	return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2288 }
2289 
2290 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2291  * counter when SACK is enabled (without SACK, sacked_out is used for
2292  * that purpose).
2293  *
2294  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2295  * segments up to the highest received SACK block so far and holes in
2296  * between them.
2297  *
2298  * With reordering, holes may still be in flight, so RFC3517 recovery
2299  * uses pure sacked_out (total number of SACKed segments) even though
2300  * it violates the RFC that uses duplicate ACKs, often these are equal
2301  * but when e.g. out-of-window ACKs or packet duplication occurs,
2302  * they differ. Since neither occurs due to loss, TCP should really
2303  * ignore them.
2304  */
2305 static inline int tcp_dupack_heuristics(struct tcp_sock *tp)
2306 {
2307 	return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2308 }
2309 
2310 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2311 {
2312 	return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
2313 }
2314 
2315 static inline int tcp_head_timedout(struct sock *sk)
2316 {
2317 	struct tcp_sock *tp = tcp_sk(sk);
2318 
2319 	return tp->packets_out &&
2320 	       tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2321 }
2322 
2323 /* Linux NewReno/SACK/FACK/ECN state machine.
2324  * --------------------------------------
2325  *
2326  * "Open"	Normal state, no dubious events, fast path.
2327  * "Disorder"   In all the respects it is "Open",
2328  *		but requires a bit more attention. It is entered when
2329  *		we see some SACKs or dupacks. It is split of "Open"
2330  *		mainly to move some processing from fast path to slow one.
2331  * "CWR"	CWND was reduced due to some Congestion Notification event.
2332  *		It can be ECN, ICMP source quench, local device congestion.
2333  * "Recovery"	CWND was reduced, we are fast-retransmitting.
2334  * "Loss"	CWND was reduced due to RTO timeout or SACK reneging.
2335  *
2336  * tcp_fastretrans_alert() is entered:
2337  * - each incoming ACK, if state is not "Open"
2338  * - when arrived ACK is unusual, namely:
2339  *	* SACK
2340  *	* Duplicate ACK.
2341  *	* ECN ECE.
2342  *
2343  * Counting packets in flight is pretty simple.
2344  *
2345  *	in_flight = packets_out - left_out + retrans_out
2346  *
2347  *	packets_out is SND.NXT-SND.UNA counted in packets.
2348  *
2349  *	retrans_out is number of retransmitted segments.
2350  *
2351  *	left_out is number of segments left network, but not ACKed yet.
2352  *
2353  *		left_out = sacked_out + lost_out
2354  *
2355  *     sacked_out: Packets, which arrived to receiver out of order
2356  *		   and hence not ACKed. With SACKs this number is simply
2357  *		   amount of SACKed data. Even without SACKs
2358  *		   it is easy to give pretty reliable estimate of this number,
2359  *		   counting duplicate ACKs.
2360  *
2361  *       lost_out: Packets lost by network. TCP has no explicit
2362  *		   "loss notification" feedback from network (for now).
2363  *		   It means that this number can be only _guessed_.
2364  *		   Actually, it is the heuristics to predict lossage that
2365  *		   distinguishes different algorithms.
2366  *
2367  *	F.e. after RTO, when all the queue is considered as lost,
2368  *	lost_out = packets_out and in_flight = retrans_out.
2369  *
2370  *		Essentially, we have now two algorithms counting
2371  *		lost packets.
2372  *
2373  *		FACK: It is the simplest heuristics. As soon as we decided
2374  *		that something is lost, we decide that _all_ not SACKed
2375  *		packets until the most forward SACK are lost. I.e.
2376  *		lost_out = fackets_out - sacked_out and left_out = fackets_out.
2377  *		It is absolutely correct estimate, if network does not reorder
2378  *		packets. And it loses any connection to reality when reordering
2379  *		takes place. We use FACK by default until reordering
2380  *		is suspected on the path to this destination.
2381  *
2382  *		NewReno: when Recovery is entered, we assume that one segment
2383  *		is lost (classic Reno). While we are in Recovery and
2384  *		a partial ACK arrives, we assume that one more packet
2385  *		is lost (NewReno). This heuristics are the same in NewReno
2386  *		and SACK.
2387  *
2388  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2389  *  deflation etc. CWND is real congestion window, never inflated, changes
2390  *  only according to classic VJ rules.
2391  *
2392  * Really tricky (and requiring careful tuning) part of algorithm
2393  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2394  * The first determines the moment _when_ we should reduce CWND and,
2395  * hence, slow down forward transmission. In fact, it determines the moment
2396  * when we decide that hole is caused by loss, rather than by a reorder.
2397  *
2398  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2399  * holes, caused by lost packets.
2400  *
2401  * And the most logically complicated part of algorithm is undo
2402  * heuristics. We detect false retransmits due to both too early
2403  * fast retransmit (reordering) and underestimated RTO, analyzing
2404  * timestamps and D-SACKs. When we detect that some segments were
2405  * retransmitted by mistake and CWND reduction was wrong, we undo
2406  * window reduction and abort recovery phase. This logic is hidden
2407  * inside several functions named tcp_try_undo_<something>.
2408  */
2409 
2410 /* This function decides, when we should leave Disordered state
2411  * and enter Recovery phase, reducing congestion window.
2412  *
2413  * Main question: may we further continue forward transmission
2414  * with the same cwnd?
2415  */
2416 static int tcp_time_to_recover(struct sock *sk)
2417 {
2418 	struct tcp_sock *tp = tcp_sk(sk);
2419 	__u32 packets_out;
2420 
2421 	/* Do not perform any recovery during F-RTO algorithm */
2422 	if (tp->frto_counter)
2423 		return 0;
2424 
2425 	/* Trick#1: The loss is proven. */
2426 	if (tp->lost_out)
2427 		return 1;
2428 
2429 	/* Not-A-Trick#2 : Classic rule... */
2430 	if (tcp_dupack_heuristics(tp) > tp->reordering)
2431 		return 1;
2432 
2433 	/* Trick#3 : when we use RFC2988 timer restart, fast
2434 	 * retransmit can be triggered by timeout of queue head.
2435 	 */
2436 	if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2437 		return 1;
2438 
2439 	/* Trick#4: It is still not OK... But will it be useful to delay
2440 	 * recovery more?
2441 	 */
2442 	packets_out = tp->packets_out;
2443 	if (packets_out <= tp->reordering &&
2444 	    tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2445 	    !tcp_may_send_now(sk)) {
2446 		/* We have nothing to send. This connection is limited
2447 		 * either by receiver window or by application.
2448 		 */
2449 		return 1;
2450 	}
2451 
2452 	/* If a thin stream is detected, retransmit after first
2453 	 * received dupack. Employ only if SACK is supported in order
2454 	 * to avoid possible corner-case series of spurious retransmissions
2455 	 * Use only if there are no unsent data.
2456 	 */
2457 	if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2458 	    tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2459 	    tcp_is_sack(tp) && !tcp_send_head(sk))
2460 		return 1;
2461 
2462 	return 0;
2463 }
2464 
2465 /* New heuristics: it is possible only after we switched to restart timer
2466  * each time when something is ACKed. Hence, we can detect timed out packets
2467  * during fast retransmit without falling to slow start.
2468  *
2469  * Usefulness of this as is very questionable, since we should know which of
2470  * the segments is the next to timeout which is relatively expensive to find
2471  * in general case unless we add some data structure just for that. The
2472  * current approach certainly won't find the right one too often and when it
2473  * finally does find _something_ it usually marks large part of the window
2474  * right away (because a retransmission with a larger timestamp blocks the
2475  * loop from advancing). -ij
2476  */
2477 static void tcp_timeout_skbs(struct sock *sk)
2478 {
2479 	struct tcp_sock *tp = tcp_sk(sk);
2480 	struct sk_buff *skb;
2481 
2482 	if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2483 		return;
2484 
2485 	skb = tp->scoreboard_skb_hint;
2486 	if (tp->scoreboard_skb_hint == NULL)
2487 		skb = tcp_write_queue_head(sk);
2488 
2489 	tcp_for_write_queue_from(skb, sk) {
2490 		if (skb == tcp_send_head(sk))
2491 			break;
2492 		if (!tcp_skb_timedout(sk, skb))
2493 			break;
2494 
2495 		tcp_skb_mark_lost(tp, skb);
2496 	}
2497 
2498 	tp->scoreboard_skb_hint = skb;
2499 
2500 	tcp_verify_left_out(tp);
2501 }
2502 
2503 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2504  * is against sacked "cnt", otherwise it's against facked "cnt"
2505  */
2506 static void tcp_mark_head_lost(struct sock *sk, int packets)
2507 {
2508 	struct tcp_sock *tp = tcp_sk(sk);
2509 	struct sk_buff *skb;
2510 	int cnt, oldcnt;
2511 	int err;
2512 	unsigned int mss;
2513 
2514 	WARN_ON(packets > tp->packets_out);
2515 	if (tp->lost_skb_hint) {
2516 		skb = tp->lost_skb_hint;
2517 		cnt = tp->lost_cnt_hint;
2518 	} else {
2519 		skb = tcp_write_queue_head(sk);
2520 		cnt = 0;
2521 	}
2522 
2523 	tcp_for_write_queue_from(skb, sk) {
2524 		if (skb == tcp_send_head(sk))
2525 			break;
2526 		/* TODO: do this better */
2527 		/* this is not the most efficient way to do this... */
2528 		tp->lost_skb_hint = skb;
2529 		tp->lost_cnt_hint = cnt;
2530 
2531 		if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2532 			break;
2533 
2534 		oldcnt = cnt;
2535 		if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2536 		    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2537 			cnt += tcp_skb_pcount(skb);
2538 
2539 		if (cnt > packets) {
2540 			if (tcp_is_sack(tp) || (oldcnt >= packets))
2541 				break;
2542 
2543 			mss = skb_shinfo(skb)->gso_size;
2544 			err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2545 			if (err < 0)
2546 				break;
2547 			cnt = packets;
2548 		}
2549 
2550 		tcp_skb_mark_lost(tp, skb);
2551 	}
2552 	tcp_verify_left_out(tp);
2553 }
2554 
2555 /* Account newly detected lost packet(s) */
2556 
2557 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2558 {
2559 	struct tcp_sock *tp = tcp_sk(sk);
2560 
2561 	if (tcp_is_reno(tp)) {
2562 		tcp_mark_head_lost(sk, 1);
2563 	} else if (tcp_is_fack(tp)) {
2564 		int lost = tp->fackets_out - tp->reordering;
2565 		if (lost <= 0)
2566 			lost = 1;
2567 		tcp_mark_head_lost(sk, lost);
2568 	} else {
2569 		int sacked_upto = tp->sacked_out - tp->reordering;
2570 		if (sacked_upto < fast_rexmit)
2571 			sacked_upto = fast_rexmit;
2572 		tcp_mark_head_lost(sk, sacked_upto);
2573 	}
2574 
2575 	tcp_timeout_skbs(sk);
2576 }
2577 
2578 /* CWND moderation, preventing bursts due to too big ACKs
2579  * in dubious situations.
2580  */
2581 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2582 {
2583 	tp->snd_cwnd = min(tp->snd_cwnd,
2584 			   tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2585 	tp->snd_cwnd_stamp = tcp_time_stamp;
2586 }
2587 
2588 /* Lower bound on congestion window is slow start threshold
2589  * unless congestion avoidance choice decides to overide it.
2590  */
2591 static inline u32 tcp_cwnd_min(const struct sock *sk)
2592 {
2593 	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2594 
2595 	return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2596 }
2597 
2598 /* Decrease cwnd each second ack. */
2599 static void tcp_cwnd_down(struct sock *sk, int flag)
2600 {
2601 	struct tcp_sock *tp = tcp_sk(sk);
2602 	int decr = tp->snd_cwnd_cnt + 1;
2603 
2604 	if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2605 	    (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2606 		tp->snd_cwnd_cnt = decr & 1;
2607 		decr >>= 1;
2608 
2609 		if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2610 			tp->snd_cwnd -= decr;
2611 
2612 		tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2613 		tp->snd_cwnd_stamp = tcp_time_stamp;
2614 	}
2615 }
2616 
2617 /* Nothing was retransmitted or returned timestamp is less
2618  * than timestamp of the first retransmission.
2619  */
2620 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2621 {
2622 	return !tp->retrans_stamp ||
2623 		(tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2624 		 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2625 }
2626 
2627 /* Undo procedures. */
2628 
2629 #if FASTRETRANS_DEBUG > 1
2630 static void DBGUNDO(struct sock *sk, const char *msg)
2631 {
2632 	struct tcp_sock *tp = tcp_sk(sk);
2633 	struct inet_sock *inet = inet_sk(sk);
2634 
2635 	if (sk->sk_family == AF_INET) {
2636 		printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2637 		       msg,
2638 		       &inet->daddr, ntohs(inet->dport),
2639 		       tp->snd_cwnd, tcp_left_out(tp),
2640 		       tp->snd_ssthresh, tp->prior_ssthresh,
2641 		       tp->packets_out);
2642 	}
2643 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2644 	else if (sk->sk_family == AF_INET6) {
2645 		struct ipv6_pinfo *np = inet6_sk(sk);
2646 		printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2647 		       msg,
2648 		       &np->daddr, ntohs(inet->dport),
2649 		       tp->snd_cwnd, tcp_left_out(tp),
2650 		       tp->snd_ssthresh, tp->prior_ssthresh,
2651 		       tp->packets_out);
2652 	}
2653 #endif
2654 }
2655 #else
2656 #define DBGUNDO(x...) do { } while (0)
2657 #endif
2658 
2659 static void tcp_undo_cwr(struct sock *sk, const int undo)
2660 {
2661 	struct tcp_sock *tp = tcp_sk(sk);
2662 
2663 	if (tp->prior_ssthresh) {
2664 		const struct inet_connection_sock *icsk = inet_csk(sk);
2665 
2666 		if (icsk->icsk_ca_ops->undo_cwnd)
2667 			tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2668 		else
2669 			tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2670 
2671 		if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2672 			tp->snd_ssthresh = tp->prior_ssthresh;
2673 			TCP_ECN_withdraw_cwr(tp);
2674 		}
2675 	} else {
2676 		tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2677 	}
2678 	tcp_moderate_cwnd(tp);
2679 	tp->snd_cwnd_stamp = tcp_time_stamp;
2680 }
2681 
2682 static inline int tcp_may_undo(struct tcp_sock *tp)
2683 {
2684 	return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2685 }
2686 
2687 /* People celebrate: "We love our President!" */
2688 static int tcp_try_undo_recovery(struct sock *sk)
2689 {
2690 	struct tcp_sock *tp = tcp_sk(sk);
2691 
2692 	if (tcp_may_undo(tp)) {
2693 		int mib_idx;
2694 
2695 		/* Happy end! We did not retransmit anything
2696 		 * or our original transmission succeeded.
2697 		 */
2698 		DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2699 		tcp_undo_cwr(sk, 1);
2700 		if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2701 			mib_idx = LINUX_MIB_TCPLOSSUNDO;
2702 		else
2703 			mib_idx = LINUX_MIB_TCPFULLUNDO;
2704 
2705 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
2706 		tp->undo_marker = 0;
2707 	}
2708 	if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2709 		/* Hold old state until something *above* high_seq
2710 		 * is ACKed. For Reno it is MUST to prevent false
2711 		 * fast retransmits (RFC2582). SACK TCP is safe. */
2712 		tcp_moderate_cwnd(tp);
2713 		return 1;
2714 	}
2715 	tcp_set_ca_state(sk, TCP_CA_Open);
2716 	return 0;
2717 }
2718 
2719 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2720 static void tcp_try_undo_dsack(struct sock *sk)
2721 {
2722 	struct tcp_sock *tp = tcp_sk(sk);
2723 
2724 	if (tp->undo_marker && !tp->undo_retrans) {
2725 		DBGUNDO(sk, "D-SACK");
2726 		tcp_undo_cwr(sk, 1);
2727 		tp->undo_marker = 0;
2728 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2729 	}
2730 }
2731 
2732 /* We can clear retrans_stamp when there are no retransmissions in the
2733  * window. It would seem that it is trivially available for us in
2734  * tp->retrans_out, however, that kind of assumptions doesn't consider
2735  * what will happen if errors occur when sending retransmission for the
2736  * second time. ...It could the that such segment has only
2737  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2738  * the head skb is enough except for some reneging corner cases that
2739  * are not worth the effort.
2740  *
2741  * Main reason for all this complexity is the fact that connection dying
2742  * time now depends on the validity of the retrans_stamp, in particular,
2743  * that successive retransmissions of a segment must not advance
2744  * retrans_stamp under any conditions.
2745  */
2746 static int tcp_any_retrans_done(struct sock *sk)
2747 {
2748 	struct tcp_sock *tp = tcp_sk(sk);
2749 	struct sk_buff *skb;
2750 
2751 	if (tp->retrans_out)
2752 		return 1;
2753 
2754 	skb = tcp_write_queue_head(sk);
2755 	if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2756 		return 1;
2757 
2758 	return 0;
2759 }
2760 
2761 /* Undo during fast recovery after partial ACK. */
2762 
2763 static int tcp_try_undo_partial(struct sock *sk, int acked)
2764 {
2765 	struct tcp_sock *tp = tcp_sk(sk);
2766 	/* Partial ACK arrived. Force Hoe's retransmit. */
2767 	int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2768 
2769 	if (tcp_may_undo(tp)) {
2770 		/* Plain luck! Hole if filled with delayed
2771 		 * packet, rather than with a retransmit.
2772 		 */
2773 		if (!tcp_any_retrans_done(sk))
2774 			tp->retrans_stamp = 0;
2775 
2776 		tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2777 
2778 		DBGUNDO(sk, "Hoe");
2779 		tcp_undo_cwr(sk, 0);
2780 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2781 
2782 		/* So... Do not make Hoe's retransmit yet.
2783 		 * If the first packet was delayed, the rest
2784 		 * ones are most probably delayed as well.
2785 		 */
2786 		failed = 0;
2787 	}
2788 	return failed;
2789 }
2790 
2791 /* Undo during loss recovery after partial ACK. */
2792 static int tcp_try_undo_loss(struct sock *sk)
2793 {
2794 	struct tcp_sock *tp = tcp_sk(sk);
2795 
2796 	if (tcp_may_undo(tp)) {
2797 		struct sk_buff *skb;
2798 		tcp_for_write_queue(skb, sk) {
2799 			if (skb == tcp_send_head(sk))
2800 				break;
2801 			TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2802 		}
2803 
2804 		tcp_clear_all_retrans_hints(tp);
2805 
2806 		DBGUNDO(sk, "partial loss");
2807 		tp->lost_out = 0;
2808 		tcp_undo_cwr(sk, 1);
2809 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2810 		inet_csk(sk)->icsk_retransmits = 0;
2811 		tp->undo_marker = 0;
2812 		if (tcp_is_sack(tp))
2813 			tcp_set_ca_state(sk, TCP_CA_Open);
2814 		return 1;
2815 	}
2816 	return 0;
2817 }
2818 
2819 static inline void tcp_complete_cwr(struct sock *sk)
2820 {
2821 	struct tcp_sock *tp = tcp_sk(sk);
2822 	tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2823 	tp->snd_cwnd_stamp = tcp_time_stamp;
2824 	tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2825 }
2826 
2827 static void tcp_try_keep_open(struct sock *sk)
2828 {
2829 	struct tcp_sock *tp = tcp_sk(sk);
2830 	int state = TCP_CA_Open;
2831 
2832 	if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker)
2833 		state = TCP_CA_Disorder;
2834 
2835 	if (inet_csk(sk)->icsk_ca_state != state) {
2836 		tcp_set_ca_state(sk, state);
2837 		tp->high_seq = tp->snd_nxt;
2838 	}
2839 }
2840 
2841 static void tcp_try_to_open(struct sock *sk, int flag)
2842 {
2843 	struct tcp_sock *tp = tcp_sk(sk);
2844 
2845 	tcp_verify_left_out(tp);
2846 
2847 	if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2848 		tp->retrans_stamp = 0;
2849 
2850 	if (flag & FLAG_ECE)
2851 		tcp_enter_cwr(sk, 1);
2852 
2853 	if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2854 		tcp_try_keep_open(sk);
2855 		tcp_moderate_cwnd(tp);
2856 	} else {
2857 		tcp_cwnd_down(sk, flag);
2858 	}
2859 }
2860 
2861 static void tcp_mtup_probe_failed(struct sock *sk)
2862 {
2863 	struct inet_connection_sock *icsk = inet_csk(sk);
2864 
2865 	icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2866 	icsk->icsk_mtup.probe_size = 0;
2867 }
2868 
2869 static void tcp_mtup_probe_success(struct sock *sk)
2870 {
2871 	struct tcp_sock *tp = tcp_sk(sk);
2872 	struct inet_connection_sock *icsk = inet_csk(sk);
2873 
2874 	/* FIXME: breaks with very large cwnd */
2875 	tp->prior_ssthresh = tcp_current_ssthresh(sk);
2876 	tp->snd_cwnd = tp->snd_cwnd *
2877 		       tcp_mss_to_mtu(sk, tp->mss_cache) /
2878 		       icsk->icsk_mtup.probe_size;
2879 	tp->snd_cwnd_cnt = 0;
2880 	tp->snd_cwnd_stamp = tcp_time_stamp;
2881 	tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2882 
2883 	icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2884 	icsk->icsk_mtup.probe_size = 0;
2885 	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2886 }
2887 
2888 /* Do a simple retransmit without using the backoff mechanisms in
2889  * tcp_timer. This is used for path mtu discovery.
2890  * The socket is already locked here.
2891  */
2892 void tcp_simple_retransmit(struct sock *sk)
2893 {
2894 	const struct inet_connection_sock *icsk = inet_csk(sk);
2895 	struct tcp_sock *tp = tcp_sk(sk);
2896 	struct sk_buff *skb;
2897 	unsigned int mss = tcp_current_mss(sk);
2898 	u32 prior_lost = tp->lost_out;
2899 
2900 	tcp_for_write_queue(skb, sk) {
2901 		if (skb == tcp_send_head(sk))
2902 			break;
2903 		if (tcp_skb_seglen(skb) > mss &&
2904 		    !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2905 			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2906 				TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2907 				tp->retrans_out -= tcp_skb_pcount(skb);
2908 			}
2909 			tcp_skb_mark_lost_uncond_verify(tp, skb);
2910 		}
2911 	}
2912 
2913 	tcp_clear_retrans_hints_partial(tp);
2914 
2915 	if (prior_lost == tp->lost_out)
2916 		return;
2917 
2918 	if (tcp_is_reno(tp))
2919 		tcp_limit_reno_sacked(tp);
2920 
2921 	tcp_verify_left_out(tp);
2922 
2923 	/* Don't muck with the congestion window here.
2924 	 * Reason is that we do not increase amount of _data_
2925 	 * in network, but units changed and effective
2926 	 * cwnd/ssthresh really reduced now.
2927 	 */
2928 	if (icsk->icsk_ca_state != TCP_CA_Loss) {
2929 		tp->high_seq = tp->snd_nxt;
2930 		tp->snd_ssthresh = tcp_current_ssthresh(sk);
2931 		tp->prior_ssthresh = 0;
2932 		tp->undo_marker = 0;
2933 		tcp_set_ca_state(sk, TCP_CA_Loss);
2934 	}
2935 	tcp_xmit_retransmit_queue(sk);
2936 }
2937 
2938 /* Process an event, which can update packets-in-flight not trivially.
2939  * Main goal of this function is to calculate new estimate for left_out,
2940  * taking into account both packets sitting in receiver's buffer and
2941  * packets lost by network.
2942  *
2943  * Besides that it does CWND reduction, when packet loss is detected
2944  * and changes state of machine.
2945  *
2946  * It does _not_ decide what to send, it is made in function
2947  * tcp_xmit_retransmit_queue().
2948  */
2949 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2950 {
2951 	struct inet_connection_sock *icsk = inet_csk(sk);
2952 	struct tcp_sock *tp = tcp_sk(sk);
2953 	int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2954 	int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2955 				    (tcp_fackets_out(tp) > tp->reordering));
2956 	int fast_rexmit = 0, mib_idx;
2957 
2958 	if (WARN_ON(!tp->packets_out && tp->sacked_out))
2959 		tp->sacked_out = 0;
2960 	if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2961 		tp->fackets_out = 0;
2962 
2963 	/* Now state machine starts.
2964 	 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2965 	if (flag & FLAG_ECE)
2966 		tp->prior_ssthresh = 0;
2967 
2968 	/* B. In all the states check for reneging SACKs. */
2969 	if (tcp_check_sack_reneging(sk, flag))
2970 		return;
2971 
2972 	/* C. Process data loss notification, provided it is valid. */
2973 	if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2974 	    before(tp->snd_una, tp->high_seq) &&
2975 	    icsk->icsk_ca_state != TCP_CA_Open &&
2976 	    tp->fackets_out > tp->reordering) {
2977 		tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2978 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
2979 	}
2980 
2981 	/* D. Check consistency of the current state. */
2982 	tcp_verify_left_out(tp);
2983 
2984 	/* E. Check state exit conditions. State can be terminated
2985 	 *    when high_seq is ACKed. */
2986 	if (icsk->icsk_ca_state == TCP_CA_Open) {
2987 		WARN_ON(tp->retrans_out != 0);
2988 		tp->retrans_stamp = 0;
2989 	} else if (!before(tp->snd_una, tp->high_seq)) {
2990 		switch (icsk->icsk_ca_state) {
2991 		case TCP_CA_Loss:
2992 			icsk->icsk_retransmits = 0;
2993 			if (tcp_try_undo_recovery(sk))
2994 				return;
2995 			break;
2996 
2997 		case TCP_CA_CWR:
2998 			/* CWR is to be held something *above* high_seq
2999 			 * is ACKed for CWR bit to reach receiver. */
3000 			if (tp->snd_una != tp->high_seq) {
3001 				tcp_complete_cwr(sk);
3002 				tcp_set_ca_state(sk, TCP_CA_Open);
3003 			}
3004 			break;
3005 
3006 		case TCP_CA_Disorder:
3007 			tcp_try_undo_dsack(sk);
3008 			if (!tp->undo_marker ||
3009 			    /* For SACK case do not Open to allow to undo
3010 			     * catching for all duplicate ACKs. */
3011 			    tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
3012 				tp->undo_marker = 0;
3013 				tcp_set_ca_state(sk, TCP_CA_Open);
3014 			}
3015 			break;
3016 
3017 		case TCP_CA_Recovery:
3018 			if (tcp_is_reno(tp))
3019 				tcp_reset_reno_sack(tp);
3020 			if (tcp_try_undo_recovery(sk))
3021 				return;
3022 			tcp_complete_cwr(sk);
3023 			break;
3024 		}
3025 	}
3026 
3027 	/* F. Process state. */
3028 	switch (icsk->icsk_ca_state) {
3029 	case TCP_CA_Recovery:
3030 		if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3031 			if (tcp_is_reno(tp) && is_dupack)
3032 				tcp_add_reno_sack(sk);
3033 		} else
3034 			do_lost = tcp_try_undo_partial(sk, pkts_acked);
3035 		break;
3036 	case TCP_CA_Loss:
3037 		if (flag & FLAG_DATA_ACKED)
3038 			icsk->icsk_retransmits = 0;
3039 		if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3040 			tcp_reset_reno_sack(tp);
3041 		if (!tcp_try_undo_loss(sk)) {
3042 			tcp_moderate_cwnd(tp);
3043 			tcp_xmit_retransmit_queue(sk);
3044 			return;
3045 		}
3046 		if (icsk->icsk_ca_state != TCP_CA_Open)
3047 			return;
3048 		/* Loss is undone; fall through to processing in Open state. */
3049 	default:
3050 		if (tcp_is_reno(tp)) {
3051 			if (flag & FLAG_SND_UNA_ADVANCED)
3052 				tcp_reset_reno_sack(tp);
3053 			if (is_dupack)
3054 				tcp_add_reno_sack(sk);
3055 		}
3056 
3057 		if (icsk->icsk_ca_state == TCP_CA_Disorder)
3058 			tcp_try_undo_dsack(sk);
3059 
3060 		if (!tcp_time_to_recover(sk)) {
3061 			tcp_try_to_open(sk, flag);
3062 			return;
3063 		}
3064 
3065 		/* MTU probe failure: don't reduce cwnd */
3066 		if (icsk->icsk_ca_state < TCP_CA_CWR &&
3067 		    icsk->icsk_mtup.probe_size &&
3068 		    tp->snd_una == tp->mtu_probe.probe_seq_start) {
3069 			tcp_mtup_probe_failed(sk);
3070 			/* Restores the reduction we did in tcp_mtup_probe() */
3071 			tp->snd_cwnd++;
3072 			tcp_simple_retransmit(sk);
3073 			return;
3074 		}
3075 
3076 		/* Otherwise enter Recovery state */
3077 
3078 		if (tcp_is_reno(tp))
3079 			mib_idx = LINUX_MIB_TCPRENORECOVERY;
3080 		else
3081 			mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3082 
3083 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
3084 
3085 		tp->high_seq = tp->snd_nxt;
3086 		tp->prior_ssthresh = 0;
3087 		tp->undo_marker = tp->snd_una;
3088 		tp->undo_retrans = tp->retrans_out;
3089 
3090 		if (icsk->icsk_ca_state < TCP_CA_CWR) {
3091 			if (!(flag & FLAG_ECE))
3092 				tp->prior_ssthresh = tcp_current_ssthresh(sk);
3093 			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3094 			TCP_ECN_queue_cwr(tp);
3095 		}
3096 
3097 		tp->bytes_acked = 0;
3098 		tp->snd_cwnd_cnt = 0;
3099 		tcp_set_ca_state(sk, TCP_CA_Recovery);
3100 		fast_rexmit = 1;
3101 	}
3102 
3103 	if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3104 		tcp_update_scoreboard(sk, fast_rexmit);
3105 	tcp_cwnd_down(sk, flag);
3106 	tcp_xmit_retransmit_queue(sk);
3107 }
3108 
3109 static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3110 {
3111 	tcp_rtt_estimator(sk, seq_rtt);
3112 	tcp_set_rto(sk);
3113 	inet_csk(sk)->icsk_backoff = 0;
3114 }
3115 
3116 /* Read draft-ietf-tcplw-high-performance before mucking
3117  * with this code. (Supersedes RFC1323)
3118  */
3119 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3120 {
3121 	/* RTTM Rule: A TSecr value received in a segment is used to
3122 	 * update the averaged RTT measurement only if the segment
3123 	 * acknowledges some new data, i.e., only if it advances the
3124 	 * left edge of the send window.
3125 	 *
3126 	 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3127 	 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3128 	 *
3129 	 * Changed: reset backoff as soon as we see the first valid sample.
3130 	 * If we do not, we get strongly overestimated rto. With timestamps
3131 	 * samples are accepted even from very old segments: f.e., when rtt=1
3132 	 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3133 	 * answer arrives rto becomes 120 seconds! If at least one of segments
3134 	 * in window is lost... Voila.	 			--ANK (010210)
3135 	 */
3136 	struct tcp_sock *tp = tcp_sk(sk);
3137 
3138 	tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3139 }
3140 
3141 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3142 {
3143 	/* We don't have a timestamp. Can only use
3144 	 * packets that are not retransmitted to determine
3145 	 * rtt estimates. Also, we must not reset the
3146 	 * backoff for rto until we get a non-retransmitted
3147 	 * packet. This allows us to deal with a situation
3148 	 * where the network delay has increased suddenly.
3149 	 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3150 	 */
3151 
3152 	if (flag & FLAG_RETRANS_DATA_ACKED)
3153 		return;
3154 
3155 	tcp_valid_rtt_meas(sk, seq_rtt);
3156 }
3157 
3158 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3159 				      const s32 seq_rtt)
3160 {
3161 	const struct tcp_sock *tp = tcp_sk(sk);
3162 	/* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3163 	if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3164 		tcp_ack_saw_tstamp(sk, flag);
3165 	else if (seq_rtt >= 0)
3166 		tcp_ack_no_tstamp(sk, seq_rtt, flag);
3167 }
3168 
3169 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3170 {
3171 	const struct inet_connection_sock *icsk = inet_csk(sk);
3172 	icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3173 	tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3174 }
3175 
3176 /* Restart timer after forward progress on connection.
3177  * RFC2988 recommends to restart timer to now+rto.
3178  */
3179 static void tcp_rearm_rto(struct sock *sk)
3180 {
3181 	struct tcp_sock *tp = tcp_sk(sk);
3182 
3183 	if (!tp->packets_out) {
3184 		inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3185 	} else {
3186 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3187 					  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3188 	}
3189 }
3190 
3191 /* If we get here, the whole TSO packet has not been acked. */
3192 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3193 {
3194 	struct tcp_sock *tp = tcp_sk(sk);
3195 	u32 packets_acked;
3196 
3197 	BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3198 
3199 	packets_acked = tcp_skb_pcount(skb);
3200 	if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3201 		return 0;
3202 	packets_acked -= tcp_skb_pcount(skb);
3203 
3204 	if (packets_acked) {
3205 		BUG_ON(tcp_skb_pcount(skb) == 0);
3206 		BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3207 	}
3208 
3209 	return packets_acked;
3210 }
3211 
3212 /* Remove acknowledged frames from the retransmission queue. If our packet
3213  * is before the ack sequence we can discard it as it's confirmed to have
3214  * arrived at the other end.
3215  */
3216 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3217 			       u32 prior_snd_una)
3218 {
3219 	struct tcp_sock *tp = tcp_sk(sk);
3220 	const struct inet_connection_sock *icsk = inet_csk(sk);
3221 	struct sk_buff *skb;
3222 	u32 now = tcp_time_stamp;
3223 	int fully_acked = 1;
3224 	int flag = 0;
3225 	u32 pkts_acked = 0;
3226 	u32 reord = tp->packets_out;
3227 	u32 prior_sacked = tp->sacked_out;
3228 	s32 seq_rtt = -1;
3229 	s32 ca_seq_rtt = -1;
3230 	ktime_t last_ackt = net_invalid_timestamp();
3231 
3232 	while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3233 		struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3234 		u32 acked_pcount;
3235 		u8 sacked = scb->sacked;
3236 
3237 		/* Determine how many packets and what bytes were acked, tso and else */
3238 		if (after(scb->end_seq, tp->snd_una)) {
3239 			if (tcp_skb_pcount(skb) == 1 ||
3240 			    !after(tp->snd_una, scb->seq))
3241 				break;
3242 
3243 			acked_pcount = tcp_tso_acked(sk, skb);
3244 			if (!acked_pcount)
3245 				break;
3246 
3247 			fully_acked = 0;
3248 		} else {
3249 			acked_pcount = tcp_skb_pcount(skb);
3250 		}
3251 
3252 		if (sacked & TCPCB_RETRANS) {
3253 			if (sacked & TCPCB_SACKED_RETRANS)
3254 				tp->retrans_out -= acked_pcount;
3255 			flag |= FLAG_RETRANS_DATA_ACKED;
3256 			ca_seq_rtt = -1;
3257 			seq_rtt = -1;
3258 			if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3259 				flag |= FLAG_NONHEAD_RETRANS_ACKED;
3260 		} else {
3261 			ca_seq_rtt = now - scb->when;
3262 			last_ackt = skb->tstamp;
3263 			if (seq_rtt < 0) {
3264 				seq_rtt = ca_seq_rtt;
3265 			}
3266 			if (!(sacked & TCPCB_SACKED_ACKED))
3267 				reord = min(pkts_acked, reord);
3268 		}
3269 
3270 		if (sacked & TCPCB_SACKED_ACKED)
3271 			tp->sacked_out -= acked_pcount;
3272 		if (sacked & TCPCB_LOST)
3273 			tp->lost_out -= acked_pcount;
3274 
3275 		tp->packets_out -= acked_pcount;
3276 		pkts_acked += acked_pcount;
3277 
3278 		/* Initial outgoing SYN's get put onto the write_queue
3279 		 * just like anything else we transmit.  It is not
3280 		 * true data, and if we misinform our callers that
3281 		 * this ACK acks real data, we will erroneously exit
3282 		 * connection startup slow start one packet too
3283 		 * quickly.  This is severely frowned upon behavior.
3284 		 */
3285 		if (!(scb->flags & TCPCB_FLAG_SYN)) {
3286 			flag |= FLAG_DATA_ACKED;
3287 		} else {
3288 			flag |= FLAG_SYN_ACKED;
3289 			tp->retrans_stamp = 0;
3290 		}
3291 
3292 		if (!fully_acked)
3293 			break;
3294 
3295 		tcp_unlink_write_queue(skb, sk);
3296 		sk_wmem_free_skb(sk, skb);
3297 		tp->scoreboard_skb_hint = NULL;
3298 		if (skb == tp->retransmit_skb_hint)
3299 			tp->retransmit_skb_hint = NULL;
3300 		if (skb == tp->lost_skb_hint)
3301 			tp->lost_skb_hint = NULL;
3302 	}
3303 
3304 	if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3305 		tp->snd_up = tp->snd_una;
3306 
3307 	if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3308 		flag |= FLAG_SACK_RENEGING;
3309 
3310 	if (flag & FLAG_ACKED) {
3311 		const struct tcp_congestion_ops *ca_ops
3312 			= inet_csk(sk)->icsk_ca_ops;
3313 
3314 		if (unlikely(icsk->icsk_mtup.probe_size &&
3315 			     !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3316 			tcp_mtup_probe_success(sk);
3317 		}
3318 
3319 		tcp_ack_update_rtt(sk, flag, seq_rtt);
3320 		tcp_rearm_rto(sk);
3321 
3322 		if (tcp_is_reno(tp)) {
3323 			tcp_remove_reno_sacks(sk, pkts_acked);
3324 		} else {
3325 			int delta;
3326 
3327 			/* Non-retransmitted hole got filled? That's reordering */
3328 			if (reord < prior_fackets)
3329 				tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3330 
3331 			delta = tcp_is_fack(tp) ? pkts_acked :
3332 						  prior_sacked - tp->sacked_out;
3333 			tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3334 		}
3335 
3336 		tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3337 
3338 		if (ca_ops->pkts_acked) {
3339 			s32 rtt_us = -1;
3340 
3341 			/* Is the ACK triggering packet unambiguous? */
3342 			if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3343 				/* High resolution needed and available? */
3344 				if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3345 				    !ktime_equal(last_ackt,
3346 						 net_invalid_timestamp()))
3347 					rtt_us = ktime_us_delta(ktime_get_real(),
3348 								last_ackt);
3349 				else if (ca_seq_rtt > 0)
3350 					rtt_us = jiffies_to_usecs(ca_seq_rtt);
3351 			}
3352 
3353 			ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3354 		}
3355 	}
3356 
3357 #if FASTRETRANS_DEBUG > 0
3358 	WARN_ON((int)tp->sacked_out < 0);
3359 	WARN_ON((int)tp->lost_out < 0);
3360 	WARN_ON((int)tp->retrans_out < 0);
3361 	if (!tp->packets_out && tcp_is_sack(tp)) {
3362 		icsk = inet_csk(sk);
3363 		if (tp->lost_out) {
3364 			printk(KERN_DEBUG "Leak l=%u %d\n",
3365 			       tp->lost_out, icsk->icsk_ca_state);
3366 			tp->lost_out = 0;
3367 		}
3368 		if (tp->sacked_out) {
3369 			printk(KERN_DEBUG "Leak s=%u %d\n",
3370 			       tp->sacked_out, icsk->icsk_ca_state);
3371 			tp->sacked_out = 0;
3372 		}
3373 		if (tp->retrans_out) {
3374 			printk(KERN_DEBUG "Leak r=%u %d\n",
3375 			       tp->retrans_out, icsk->icsk_ca_state);
3376 			tp->retrans_out = 0;
3377 		}
3378 	}
3379 #endif
3380 	return flag;
3381 }
3382 
3383 static void tcp_ack_probe(struct sock *sk)
3384 {
3385 	const struct tcp_sock *tp = tcp_sk(sk);
3386 	struct inet_connection_sock *icsk = inet_csk(sk);
3387 
3388 	/* Was it a usable window open? */
3389 
3390 	if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3391 		icsk->icsk_backoff = 0;
3392 		inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3393 		/* Socket must be waked up by subsequent tcp_data_snd_check().
3394 		 * This function is not for random using!
3395 		 */
3396 	} else {
3397 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3398 					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3399 					  TCP_RTO_MAX);
3400 	}
3401 }
3402 
3403 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3404 {
3405 	return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3406 		inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
3407 }
3408 
3409 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3410 {
3411 	const struct tcp_sock *tp = tcp_sk(sk);
3412 	return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3413 		!((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3414 }
3415 
3416 /* Check that window update is acceptable.
3417  * The function assumes that snd_una<=ack<=snd_next.
3418  */
3419 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3420 					const u32 ack, const u32 ack_seq,
3421 					const u32 nwin)
3422 {
3423 	return (after(ack, tp->snd_una) ||
3424 		after(ack_seq, tp->snd_wl1) ||
3425 		(ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3426 }
3427 
3428 /* Update our send window.
3429  *
3430  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3431  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3432  */
3433 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3434 				 u32 ack_seq)
3435 {
3436 	struct tcp_sock *tp = tcp_sk(sk);
3437 	int flag = 0;
3438 	u32 nwin = ntohs(tcp_hdr(skb)->window);
3439 
3440 	if (likely(!tcp_hdr(skb)->syn))
3441 		nwin <<= tp->rx_opt.snd_wscale;
3442 
3443 	if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3444 		flag |= FLAG_WIN_UPDATE;
3445 		tcp_update_wl(tp, ack_seq);
3446 
3447 		if (tp->snd_wnd != nwin) {
3448 			tp->snd_wnd = nwin;
3449 
3450 			/* Note, it is the only place, where
3451 			 * fast path is recovered for sending TCP.
3452 			 */
3453 			tp->pred_flags = 0;
3454 			tcp_fast_path_check(sk);
3455 
3456 			if (nwin > tp->max_window) {
3457 				tp->max_window = nwin;
3458 				tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3459 			}
3460 		}
3461 	}
3462 
3463 	tp->snd_una = ack;
3464 
3465 	return flag;
3466 }
3467 
3468 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3469  * continue in congestion avoidance.
3470  */
3471 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3472 {
3473 	tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3474 	tp->snd_cwnd_cnt = 0;
3475 	tp->bytes_acked = 0;
3476 	TCP_ECN_queue_cwr(tp);
3477 	tcp_moderate_cwnd(tp);
3478 }
3479 
3480 /* A conservative spurious RTO response algorithm: reduce cwnd using
3481  * rate halving and continue in congestion avoidance.
3482  */
3483 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3484 {
3485 	tcp_enter_cwr(sk, 0);
3486 }
3487 
3488 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3489 {
3490 	if (flag & FLAG_ECE)
3491 		tcp_ratehalving_spur_to_response(sk);
3492 	else
3493 		tcp_undo_cwr(sk, 1);
3494 }
3495 
3496 /* F-RTO spurious RTO detection algorithm (RFC4138)
3497  *
3498  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3499  * comments). State (ACK number) is kept in frto_counter. When ACK advances
3500  * window (but not to or beyond highest sequence sent before RTO):
3501  *   On First ACK,  send two new segments out.
3502  *   On Second ACK, RTO was likely spurious. Do spurious response (response
3503  *                  algorithm is not part of the F-RTO detection algorithm
3504  *                  given in RFC4138 but can be selected separately).
3505  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3506  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3507  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3508  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3509  *
3510  * Rationale: if the RTO was spurious, new ACKs should arrive from the
3511  * original window even after we transmit two new data segments.
3512  *
3513  * SACK version:
3514  *   on first step, wait until first cumulative ACK arrives, then move to
3515  *   the second step. In second step, the next ACK decides.
3516  *
3517  * F-RTO is implemented (mainly) in four functions:
3518  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3519  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3520  *     called when tcp_use_frto() showed green light
3521  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3522  *   - tcp_enter_frto_loss() is called if there is not enough evidence
3523  *     to prove that the RTO is indeed spurious. It transfers the control
3524  *     from F-RTO to the conventional RTO recovery
3525  */
3526 static int tcp_process_frto(struct sock *sk, int flag)
3527 {
3528 	struct tcp_sock *tp = tcp_sk(sk);
3529 
3530 	tcp_verify_left_out(tp);
3531 
3532 	/* Duplicate the behavior from Loss state (fastretrans_alert) */
3533 	if (flag & FLAG_DATA_ACKED)
3534 		inet_csk(sk)->icsk_retransmits = 0;
3535 
3536 	if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3537 	    ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3538 		tp->undo_marker = 0;
3539 
3540 	if (!before(tp->snd_una, tp->frto_highmark)) {
3541 		tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3542 		return 1;
3543 	}
3544 
3545 	if (!tcp_is_sackfrto(tp)) {
3546 		/* RFC4138 shortcoming in step 2; should also have case c):
3547 		 * ACK isn't duplicate nor advances window, e.g., opposite dir
3548 		 * data, winupdate
3549 		 */
3550 		if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3551 			return 1;
3552 
3553 		if (!(flag & FLAG_DATA_ACKED)) {
3554 			tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3555 					    flag);
3556 			return 1;
3557 		}
3558 	} else {
3559 		if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3560 			/* Prevent sending of new data. */
3561 			tp->snd_cwnd = min(tp->snd_cwnd,
3562 					   tcp_packets_in_flight(tp));
3563 			return 1;
3564 		}
3565 
3566 		if ((tp->frto_counter >= 2) &&
3567 		    (!(flag & FLAG_FORWARD_PROGRESS) ||
3568 		     ((flag & FLAG_DATA_SACKED) &&
3569 		      !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3570 			/* RFC4138 shortcoming (see comment above) */
3571 			if (!(flag & FLAG_FORWARD_PROGRESS) &&
3572 			    (flag & FLAG_NOT_DUP))
3573 				return 1;
3574 
3575 			tcp_enter_frto_loss(sk, 3, flag);
3576 			return 1;
3577 		}
3578 	}
3579 
3580 	if (tp->frto_counter == 1) {
3581 		/* tcp_may_send_now needs to see updated state */
3582 		tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3583 		tp->frto_counter = 2;
3584 
3585 		if (!tcp_may_send_now(sk))
3586 			tcp_enter_frto_loss(sk, 2, flag);
3587 
3588 		return 1;
3589 	} else {
3590 		switch (sysctl_tcp_frto_response) {
3591 		case 2:
3592 			tcp_undo_spur_to_response(sk, flag);
3593 			break;
3594 		case 1:
3595 			tcp_conservative_spur_to_response(tp);
3596 			break;
3597 		default:
3598 			tcp_ratehalving_spur_to_response(sk);
3599 			break;
3600 		}
3601 		tp->frto_counter = 0;
3602 		tp->undo_marker = 0;
3603 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3604 	}
3605 	return 0;
3606 }
3607 
3608 /* This routine deals with incoming acks, but not outgoing ones. */
3609 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3610 {
3611 	struct inet_connection_sock *icsk = inet_csk(sk);
3612 	struct tcp_sock *tp = tcp_sk(sk);
3613 	u32 prior_snd_una = tp->snd_una;
3614 	u32 ack_seq = TCP_SKB_CB(skb)->seq;
3615 	u32 ack = TCP_SKB_CB(skb)->ack_seq;
3616 	u32 prior_in_flight;
3617 	u32 prior_fackets;
3618 	int prior_packets;
3619 	int frto_cwnd = 0;
3620 
3621 	/* If the ack is older than previous acks
3622 	 * then we can probably ignore it.
3623 	 */
3624 	if (before(ack, prior_snd_una))
3625 		goto old_ack;
3626 
3627 	/* If the ack includes data we haven't sent yet, discard
3628 	 * this segment (RFC793 Section 3.9).
3629 	 */
3630 	if (after(ack, tp->snd_nxt))
3631 		goto invalid_ack;
3632 
3633 	if (after(ack, prior_snd_una))
3634 		flag |= FLAG_SND_UNA_ADVANCED;
3635 
3636 	if (sysctl_tcp_abc) {
3637 		if (icsk->icsk_ca_state < TCP_CA_CWR)
3638 			tp->bytes_acked += ack - prior_snd_una;
3639 		else if (icsk->icsk_ca_state == TCP_CA_Loss)
3640 			/* we assume just one segment left network */
3641 			tp->bytes_acked += min(ack - prior_snd_una,
3642 					       tp->mss_cache);
3643 	}
3644 
3645 	prior_fackets = tp->fackets_out;
3646 	prior_in_flight = tcp_packets_in_flight(tp);
3647 
3648 	if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3649 		/* Window is constant, pure forward advance.
3650 		 * No more checks are required.
3651 		 * Note, we use the fact that SND.UNA>=SND.WL2.
3652 		 */
3653 		tcp_update_wl(tp, ack_seq);
3654 		tp->snd_una = ack;
3655 		flag |= FLAG_WIN_UPDATE;
3656 
3657 		tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3658 
3659 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3660 	} else {
3661 		if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3662 			flag |= FLAG_DATA;
3663 		else
3664 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3665 
3666 		flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3667 
3668 		if (TCP_SKB_CB(skb)->sacked)
3669 			flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3670 
3671 		if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3672 			flag |= FLAG_ECE;
3673 
3674 		tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3675 	}
3676 
3677 	/* We passed data and got it acked, remove any soft error
3678 	 * log. Something worked...
3679 	 */
3680 	sk->sk_err_soft = 0;
3681 	icsk->icsk_probes_out = 0;
3682 	tp->rcv_tstamp = tcp_time_stamp;
3683 	prior_packets = tp->packets_out;
3684 	if (!prior_packets)
3685 		goto no_queue;
3686 
3687 	/* See if we can take anything off of the retransmit queue. */
3688 	flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3689 
3690 	if (tp->frto_counter)
3691 		frto_cwnd = tcp_process_frto(sk, flag);
3692 	/* Guarantee sacktag reordering detection against wrap-arounds */
3693 	if (before(tp->frto_highmark, tp->snd_una))
3694 		tp->frto_highmark = 0;
3695 
3696 	if (tcp_ack_is_dubious(sk, flag)) {
3697 		/* Advance CWND, if state allows this. */
3698 		if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3699 		    tcp_may_raise_cwnd(sk, flag))
3700 			tcp_cong_avoid(sk, ack, prior_in_flight);
3701 		tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3702 				      flag);
3703 	} else {
3704 		if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3705 			tcp_cong_avoid(sk, ack, prior_in_flight);
3706 	}
3707 
3708 	if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3709 		dst_confirm(sk->sk_dst_cache);
3710 
3711 	return 1;
3712 
3713 no_queue:
3714 	/* If this ack opens up a zero window, clear backoff.  It was
3715 	 * being used to time the probes, and is probably far higher than
3716 	 * it needs to be for normal retransmission.
3717 	 */
3718 	if (tcp_send_head(sk))
3719 		tcp_ack_probe(sk);
3720 	return 1;
3721 
3722 invalid_ack:
3723 	SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3724 	return -1;
3725 
3726 old_ack:
3727 	if (TCP_SKB_CB(skb)->sacked) {
3728 		tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3729 		if (icsk->icsk_ca_state == TCP_CA_Open)
3730 			tcp_try_keep_open(sk);
3731 	}
3732 
3733 	SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3734 	return 0;
3735 }
3736 
3737 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3738  * But, this can also be called on packets in the established flow when
3739  * the fast version below fails.
3740  */
3741 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3742 		       u8 **hvpp, int estab)
3743 {
3744 	unsigned char *ptr;
3745 	struct tcphdr *th = tcp_hdr(skb);
3746 	int length = (th->doff * 4) - sizeof(struct tcphdr);
3747 
3748 	ptr = (unsigned char *)(th + 1);
3749 	opt_rx->saw_tstamp = 0;
3750 
3751 	while (length > 0) {
3752 		int opcode = *ptr++;
3753 		int opsize;
3754 
3755 		switch (opcode) {
3756 		case TCPOPT_EOL:
3757 			return;
3758 		case TCPOPT_NOP:	/* Ref: RFC 793 section 3.1 */
3759 			length--;
3760 			continue;
3761 		default:
3762 			opsize = *ptr++;
3763 			if (opsize < 2) /* "silly options" */
3764 				return;
3765 			if (opsize > length)
3766 				return;	/* don't parse partial options */
3767 			switch (opcode) {
3768 			case TCPOPT_MSS:
3769 				if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3770 					u16 in_mss = get_unaligned_be16(ptr);
3771 					if (in_mss) {
3772 						if (opt_rx->user_mss &&
3773 						    opt_rx->user_mss < in_mss)
3774 							in_mss = opt_rx->user_mss;
3775 						opt_rx->mss_clamp = in_mss;
3776 					}
3777 				}
3778 				break;
3779 			case TCPOPT_WINDOW:
3780 				if (opsize == TCPOLEN_WINDOW && th->syn &&
3781 				    !estab && sysctl_tcp_window_scaling) {
3782 					__u8 snd_wscale = *(__u8 *)ptr;
3783 					opt_rx->wscale_ok = 1;
3784 					if (snd_wscale > 14) {
3785 						if (net_ratelimit())
3786 							printk(KERN_INFO "tcp_parse_options: Illegal window "
3787 							       "scaling value %d >14 received.\n",
3788 							       snd_wscale);
3789 						snd_wscale = 14;
3790 					}
3791 					opt_rx->snd_wscale = snd_wscale;
3792 				}
3793 				break;
3794 			case TCPOPT_TIMESTAMP:
3795 				if ((opsize == TCPOLEN_TIMESTAMP) &&
3796 				    ((estab && opt_rx->tstamp_ok) ||
3797 				     (!estab && sysctl_tcp_timestamps))) {
3798 					opt_rx->saw_tstamp = 1;
3799 					opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3800 					opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3801 				}
3802 				break;
3803 			case TCPOPT_SACK_PERM:
3804 				if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3805 				    !estab && sysctl_tcp_sack) {
3806 					opt_rx->sack_ok = 1;
3807 					tcp_sack_reset(opt_rx);
3808 				}
3809 				break;
3810 
3811 			case TCPOPT_SACK:
3812 				if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3813 				   !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3814 				   opt_rx->sack_ok) {
3815 					TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3816 				}
3817 				break;
3818 #ifdef CONFIG_TCP_MD5SIG
3819 			case TCPOPT_MD5SIG:
3820 				/*
3821 				 * The MD5 Hash has already been
3822 				 * checked (see tcp_v{4,6}_do_rcv()).
3823 				 */
3824 				break;
3825 #endif
3826 			case TCPOPT_COOKIE:
3827 				/* This option is variable length.
3828 				 */
3829 				switch (opsize) {
3830 				case TCPOLEN_COOKIE_BASE:
3831 					/* not yet implemented */
3832 					break;
3833 				case TCPOLEN_COOKIE_PAIR:
3834 					/* not yet implemented */
3835 					break;
3836 				case TCPOLEN_COOKIE_MIN+0:
3837 				case TCPOLEN_COOKIE_MIN+2:
3838 				case TCPOLEN_COOKIE_MIN+4:
3839 				case TCPOLEN_COOKIE_MIN+6:
3840 				case TCPOLEN_COOKIE_MAX:
3841 					/* 16-bit multiple */
3842 					opt_rx->cookie_plus = opsize;
3843 					*hvpp = ptr;
3844 				default:
3845 					/* ignore option */
3846 					break;
3847 				};
3848 				break;
3849 			};
3850 
3851 			ptr += opsize-2;
3852 			length -= opsize;
3853 		}
3854 	}
3855 }
3856 
3857 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3858 {
3859 	__be32 *ptr = (__be32 *)(th + 1);
3860 
3861 	if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3862 			  | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3863 		tp->rx_opt.saw_tstamp = 1;
3864 		++ptr;
3865 		tp->rx_opt.rcv_tsval = ntohl(*ptr);
3866 		++ptr;
3867 		tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3868 		return 1;
3869 	}
3870 	return 0;
3871 }
3872 
3873 /* Fast parse options. This hopes to only see timestamps.
3874  * If it is wrong it falls back on tcp_parse_options().
3875  */
3876 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3877 				  struct tcp_sock *tp, u8 **hvpp)
3878 {
3879 	/* In the spirit of fast parsing, compare doff directly to constant
3880 	 * values.  Because equality is used, short doff can be ignored here.
3881 	 */
3882 	if (th->doff == (sizeof(*th) / 4)) {
3883 		tp->rx_opt.saw_tstamp = 0;
3884 		return 0;
3885 	} else if (tp->rx_opt.tstamp_ok &&
3886 		   th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3887 		if (tcp_parse_aligned_timestamp(tp, th))
3888 			return 1;
3889 	}
3890 	tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3891 	return 1;
3892 }
3893 
3894 #ifdef CONFIG_TCP_MD5SIG
3895 /*
3896  * Parse MD5 Signature option
3897  */
3898 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3899 {
3900 	int length = (th->doff << 2) - sizeof (*th);
3901 	u8 *ptr = (u8*)(th + 1);
3902 
3903 	/* If the TCP option is too short, we can short cut */
3904 	if (length < TCPOLEN_MD5SIG)
3905 		return NULL;
3906 
3907 	while (length > 0) {
3908 		int opcode = *ptr++;
3909 		int opsize;
3910 
3911 		switch(opcode) {
3912 		case TCPOPT_EOL:
3913 			return NULL;
3914 		case TCPOPT_NOP:
3915 			length--;
3916 			continue;
3917 		default:
3918 			opsize = *ptr++;
3919 			if (opsize < 2 || opsize > length)
3920 				return NULL;
3921 			if (opcode == TCPOPT_MD5SIG)
3922 				return ptr;
3923 		}
3924 		ptr += opsize - 2;
3925 		length -= opsize;
3926 	}
3927 	return NULL;
3928 }
3929 #endif
3930 
3931 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3932 {
3933 	tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3934 	tp->rx_opt.ts_recent_stamp = get_seconds();
3935 }
3936 
3937 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3938 {
3939 	if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3940 		/* PAWS bug workaround wrt. ACK frames, the PAWS discard
3941 		 * extra check below makes sure this can only happen
3942 		 * for pure ACK frames.  -DaveM
3943 		 *
3944 		 * Not only, also it occurs for expired timestamps.
3945 		 */
3946 
3947 		if (tcp_paws_check(&tp->rx_opt, 0))
3948 			tcp_store_ts_recent(tp);
3949 	}
3950 }
3951 
3952 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3953  *
3954  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3955  * it can pass through stack. So, the following predicate verifies that
3956  * this segment is not used for anything but congestion avoidance or
3957  * fast retransmit. Moreover, we even are able to eliminate most of such
3958  * second order effects, if we apply some small "replay" window (~RTO)
3959  * to timestamp space.
3960  *
3961  * All these measures still do not guarantee that we reject wrapped ACKs
3962  * on networks with high bandwidth, when sequence space is recycled fastly,
3963  * but it guarantees that such events will be very rare and do not affect
3964  * connection seriously. This doesn't look nice, but alas, PAWS is really
3965  * buggy extension.
3966  *
3967  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3968  * states that events when retransmit arrives after original data are rare.
3969  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3970  * the biggest problem on large power networks even with minor reordering.
3971  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3972  * up to bandwidth of 18Gigabit/sec. 8) ]
3973  */
3974 
3975 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3976 {
3977 	struct tcp_sock *tp = tcp_sk(sk);
3978 	struct tcphdr *th = tcp_hdr(skb);
3979 	u32 seq = TCP_SKB_CB(skb)->seq;
3980 	u32 ack = TCP_SKB_CB(skb)->ack_seq;
3981 
3982 	return (/* 1. Pure ACK with correct sequence number. */
3983 		(th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3984 
3985 		/* 2. ... and duplicate ACK. */
3986 		ack == tp->snd_una &&
3987 
3988 		/* 3. ... and does not update window. */
3989 		!tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3990 
3991 		/* 4. ... and sits in replay window. */
3992 		(s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3993 }
3994 
3995 static inline int tcp_paws_discard(const struct sock *sk,
3996 				   const struct sk_buff *skb)
3997 {
3998 	const struct tcp_sock *tp = tcp_sk(sk);
3999 
4000 	return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4001 	       !tcp_disordered_ack(sk, skb);
4002 }
4003 
4004 /* Check segment sequence number for validity.
4005  *
4006  * Segment controls are considered valid, if the segment
4007  * fits to the window after truncation to the window. Acceptability
4008  * of data (and SYN, FIN, of course) is checked separately.
4009  * See tcp_data_queue(), for example.
4010  *
4011  * Also, controls (RST is main one) are accepted using RCV.WUP instead
4012  * of RCV.NXT. Peer still did not advance his SND.UNA when we
4013  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4014  * (borrowed from freebsd)
4015  */
4016 
4017 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
4018 {
4019 	return	!before(end_seq, tp->rcv_wup) &&
4020 		!after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4021 }
4022 
4023 /* When we get a reset we do this. */
4024 static void tcp_reset(struct sock *sk)
4025 {
4026 	/* We want the right error as BSD sees it (and indeed as we do). */
4027 	switch (sk->sk_state) {
4028 	case TCP_SYN_SENT:
4029 		sk->sk_err = ECONNREFUSED;
4030 		break;
4031 	case TCP_CLOSE_WAIT:
4032 		sk->sk_err = EPIPE;
4033 		break;
4034 	case TCP_CLOSE:
4035 		return;
4036 	default:
4037 		sk->sk_err = ECONNRESET;
4038 	}
4039 
4040 	if (!sock_flag(sk, SOCK_DEAD))
4041 		sk->sk_error_report(sk);
4042 
4043 	tcp_done(sk);
4044 }
4045 
4046 /*
4047  * 	Process the FIN bit. This now behaves as it is supposed to work
4048  *	and the FIN takes effect when it is validly part of sequence
4049  *	space. Not before when we get holes.
4050  *
4051  *	If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4052  *	(and thence onto LAST-ACK and finally, CLOSE, we never enter
4053  *	TIME-WAIT)
4054  *
4055  *	If we are in FINWAIT-1, a received FIN indicates simultaneous
4056  *	close and we go into CLOSING (and later onto TIME-WAIT)
4057  *
4058  *	If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4059  */
4060 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
4061 {
4062 	struct tcp_sock *tp = tcp_sk(sk);
4063 
4064 	inet_csk_schedule_ack(sk);
4065 
4066 	sk->sk_shutdown |= RCV_SHUTDOWN;
4067 	sock_set_flag(sk, SOCK_DONE);
4068 
4069 	switch (sk->sk_state) {
4070 	case TCP_SYN_RECV:
4071 	case TCP_ESTABLISHED:
4072 		/* Move to CLOSE_WAIT */
4073 		tcp_set_state(sk, TCP_CLOSE_WAIT);
4074 		inet_csk(sk)->icsk_ack.pingpong = 1;
4075 		break;
4076 
4077 	case TCP_CLOSE_WAIT:
4078 	case TCP_CLOSING:
4079 		/* Received a retransmission of the FIN, do
4080 		 * nothing.
4081 		 */
4082 		break;
4083 	case TCP_LAST_ACK:
4084 		/* RFC793: Remain in the LAST-ACK state. */
4085 		break;
4086 
4087 	case TCP_FIN_WAIT1:
4088 		/* This case occurs when a simultaneous close
4089 		 * happens, we must ack the received FIN and
4090 		 * enter the CLOSING state.
4091 		 */
4092 		tcp_send_ack(sk);
4093 		tcp_set_state(sk, TCP_CLOSING);
4094 		break;
4095 	case TCP_FIN_WAIT2:
4096 		/* Received a FIN -- send ACK and enter TIME_WAIT. */
4097 		tcp_send_ack(sk);
4098 		tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4099 		break;
4100 	default:
4101 		/* Only TCP_LISTEN and TCP_CLOSE are left, in these
4102 		 * cases we should never reach this piece of code.
4103 		 */
4104 		printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4105 		       __func__, sk->sk_state);
4106 		break;
4107 	}
4108 
4109 	/* It _is_ possible, that we have something out-of-order _after_ FIN.
4110 	 * Probably, we should reset in this case. For now drop them.
4111 	 */
4112 	__skb_queue_purge(&tp->out_of_order_queue);
4113 	if (tcp_is_sack(tp))
4114 		tcp_sack_reset(&tp->rx_opt);
4115 	sk_mem_reclaim(sk);
4116 
4117 	if (!sock_flag(sk, SOCK_DEAD)) {
4118 		sk->sk_state_change(sk);
4119 
4120 		/* Do not send POLL_HUP for half duplex close. */
4121 		if (sk->sk_shutdown == SHUTDOWN_MASK ||
4122 		    sk->sk_state == TCP_CLOSE)
4123 			sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4124 		else
4125 			sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4126 	}
4127 }
4128 
4129 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4130 				  u32 end_seq)
4131 {
4132 	if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4133 		if (before(seq, sp->start_seq))
4134 			sp->start_seq = seq;
4135 		if (after(end_seq, sp->end_seq))
4136 			sp->end_seq = end_seq;
4137 		return 1;
4138 	}
4139 	return 0;
4140 }
4141 
4142 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4143 {
4144 	struct tcp_sock *tp = tcp_sk(sk);
4145 
4146 	if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4147 		int mib_idx;
4148 
4149 		if (before(seq, tp->rcv_nxt))
4150 			mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4151 		else
4152 			mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4153 
4154 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
4155 
4156 		tp->rx_opt.dsack = 1;
4157 		tp->duplicate_sack[0].start_seq = seq;
4158 		tp->duplicate_sack[0].end_seq = end_seq;
4159 	}
4160 }
4161 
4162 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4163 {
4164 	struct tcp_sock *tp = tcp_sk(sk);
4165 
4166 	if (!tp->rx_opt.dsack)
4167 		tcp_dsack_set(sk, seq, end_seq);
4168 	else
4169 		tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4170 }
4171 
4172 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4173 {
4174 	struct tcp_sock *tp = tcp_sk(sk);
4175 
4176 	if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4177 	    before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4178 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4179 		tcp_enter_quickack_mode(sk);
4180 
4181 		if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4182 			u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4183 
4184 			if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4185 				end_seq = tp->rcv_nxt;
4186 			tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4187 		}
4188 	}
4189 
4190 	tcp_send_ack(sk);
4191 }
4192 
4193 /* These routines update the SACK block as out-of-order packets arrive or
4194  * in-order packets close up the sequence space.
4195  */
4196 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4197 {
4198 	int this_sack;
4199 	struct tcp_sack_block *sp = &tp->selective_acks[0];
4200 	struct tcp_sack_block *swalk = sp + 1;
4201 
4202 	/* See if the recent change to the first SACK eats into
4203 	 * or hits the sequence space of other SACK blocks, if so coalesce.
4204 	 */
4205 	for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4206 		if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4207 			int i;
4208 
4209 			/* Zap SWALK, by moving every further SACK up by one slot.
4210 			 * Decrease num_sacks.
4211 			 */
4212 			tp->rx_opt.num_sacks--;
4213 			for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4214 				sp[i] = sp[i + 1];
4215 			continue;
4216 		}
4217 		this_sack++, swalk++;
4218 	}
4219 }
4220 
4221 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4222 {
4223 	struct tcp_sock *tp = tcp_sk(sk);
4224 	struct tcp_sack_block *sp = &tp->selective_acks[0];
4225 	int cur_sacks = tp->rx_opt.num_sacks;
4226 	int this_sack;
4227 
4228 	if (!cur_sacks)
4229 		goto new_sack;
4230 
4231 	for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4232 		if (tcp_sack_extend(sp, seq, end_seq)) {
4233 			/* Rotate this_sack to the first one. */
4234 			for (; this_sack > 0; this_sack--, sp--)
4235 				swap(*sp, *(sp - 1));
4236 			if (cur_sacks > 1)
4237 				tcp_sack_maybe_coalesce(tp);
4238 			return;
4239 		}
4240 	}
4241 
4242 	/* Could not find an adjacent existing SACK, build a new one,
4243 	 * put it at the front, and shift everyone else down.  We
4244 	 * always know there is at least one SACK present already here.
4245 	 *
4246 	 * If the sack array is full, forget about the last one.
4247 	 */
4248 	if (this_sack >= TCP_NUM_SACKS) {
4249 		this_sack--;
4250 		tp->rx_opt.num_sacks--;
4251 		sp--;
4252 	}
4253 	for (; this_sack > 0; this_sack--, sp--)
4254 		*sp = *(sp - 1);
4255 
4256 new_sack:
4257 	/* Build the new head SACK, and we're done. */
4258 	sp->start_seq = seq;
4259 	sp->end_seq = end_seq;
4260 	tp->rx_opt.num_sacks++;
4261 }
4262 
4263 /* RCV.NXT advances, some SACKs should be eaten. */
4264 
4265 static void tcp_sack_remove(struct tcp_sock *tp)
4266 {
4267 	struct tcp_sack_block *sp = &tp->selective_acks[0];
4268 	int num_sacks = tp->rx_opt.num_sacks;
4269 	int this_sack;
4270 
4271 	/* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4272 	if (skb_queue_empty(&tp->out_of_order_queue)) {
4273 		tp->rx_opt.num_sacks = 0;
4274 		return;
4275 	}
4276 
4277 	for (this_sack = 0; this_sack < num_sacks;) {
4278 		/* Check if the start of the sack is covered by RCV.NXT. */
4279 		if (!before(tp->rcv_nxt, sp->start_seq)) {
4280 			int i;
4281 
4282 			/* RCV.NXT must cover all the block! */
4283 			WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4284 
4285 			/* Zap this SACK, by moving forward any other SACKS. */
4286 			for (i=this_sack+1; i < num_sacks; i++)
4287 				tp->selective_acks[i-1] = tp->selective_acks[i];
4288 			num_sacks--;
4289 			continue;
4290 		}
4291 		this_sack++;
4292 		sp++;
4293 	}
4294 	tp->rx_opt.num_sacks = num_sacks;
4295 }
4296 
4297 /* This one checks to see if we can put data from the
4298  * out_of_order queue into the receive_queue.
4299  */
4300 static void tcp_ofo_queue(struct sock *sk)
4301 {
4302 	struct tcp_sock *tp = tcp_sk(sk);
4303 	__u32 dsack_high = tp->rcv_nxt;
4304 	struct sk_buff *skb;
4305 
4306 	while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4307 		if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4308 			break;
4309 
4310 		if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4311 			__u32 dsack = dsack_high;
4312 			if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4313 				dsack_high = TCP_SKB_CB(skb)->end_seq;
4314 			tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4315 		}
4316 
4317 		if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4318 			SOCK_DEBUG(sk, "ofo packet was already received \n");
4319 			__skb_unlink(skb, &tp->out_of_order_queue);
4320 			__kfree_skb(skb);
4321 			continue;
4322 		}
4323 		SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4324 			   tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4325 			   TCP_SKB_CB(skb)->end_seq);
4326 
4327 		__skb_unlink(skb, &tp->out_of_order_queue);
4328 		__skb_queue_tail(&sk->sk_receive_queue, skb);
4329 		tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4330 		if (tcp_hdr(skb)->fin)
4331 			tcp_fin(skb, sk, tcp_hdr(skb));
4332 	}
4333 }
4334 
4335 static int tcp_prune_ofo_queue(struct sock *sk);
4336 static int tcp_prune_queue(struct sock *sk);
4337 
4338 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4339 {
4340 	if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4341 	    !sk_rmem_schedule(sk, size)) {
4342 
4343 		if (tcp_prune_queue(sk) < 0)
4344 			return -1;
4345 
4346 		if (!sk_rmem_schedule(sk, size)) {
4347 			if (!tcp_prune_ofo_queue(sk))
4348 				return -1;
4349 
4350 			if (!sk_rmem_schedule(sk, size))
4351 				return -1;
4352 		}
4353 	}
4354 	return 0;
4355 }
4356 
4357 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4358 {
4359 	struct tcphdr *th = tcp_hdr(skb);
4360 	struct tcp_sock *tp = tcp_sk(sk);
4361 	int eaten = -1;
4362 
4363 	if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4364 		goto drop;
4365 
4366 	__skb_pull(skb, th->doff * 4);
4367 
4368 	TCP_ECN_accept_cwr(tp, skb);
4369 
4370 	tp->rx_opt.dsack = 0;
4371 
4372 	/*  Queue data for delivery to the user.
4373 	 *  Packets in sequence go to the receive queue.
4374 	 *  Out of sequence packets to the out_of_order_queue.
4375 	 */
4376 	if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4377 		if (tcp_receive_window(tp) == 0)
4378 			goto out_of_window;
4379 
4380 		/* Ok. In sequence. In window. */
4381 		if (tp->ucopy.task == current &&
4382 		    tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4383 		    sock_owned_by_user(sk) && !tp->urg_data) {
4384 			int chunk = min_t(unsigned int, skb->len,
4385 					  tp->ucopy.len);
4386 
4387 			__set_current_state(TASK_RUNNING);
4388 
4389 			local_bh_enable();
4390 			if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4391 				tp->ucopy.len -= chunk;
4392 				tp->copied_seq += chunk;
4393 				eaten = (chunk == skb->len && !th->fin);
4394 				tcp_rcv_space_adjust(sk);
4395 			}
4396 			local_bh_disable();
4397 		}
4398 
4399 		if (eaten <= 0) {
4400 queue_and_out:
4401 			if (eaten < 0 &&
4402 			    tcp_try_rmem_schedule(sk, skb->truesize))
4403 				goto drop;
4404 
4405 			skb_set_owner_r(skb, sk);
4406 			__skb_queue_tail(&sk->sk_receive_queue, skb);
4407 		}
4408 		tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4409 		if (skb->len)
4410 			tcp_event_data_recv(sk, skb);
4411 		if (th->fin)
4412 			tcp_fin(skb, sk, th);
4413 
4414 		if (!skb_queue_empty(&tp->out_of_order_queue)) {
4415 			tcp_ofo_queue(sk);
4416 
4417 			/* RFC2581. 4.2. SHOULD send immediate ACK, when
4418 			 * gap in queue is filled.
4419 			 */
4420 			if (skb_queue_empty(&tp->out_of_order_queue))
4421 				inet_csk(sk)->icsk_ack.pingpong = 0;
4422 		}
4423 
4424 		if (tp->rx_opt.num_sacks)
4425 			tcp_sack_remove(tp);
4426 
4427 		tcp_fast_path_check(sk);
4428 
4429 		if (eaten > 0)
4430 			__kfree_skb(skb);
4431 		else if (!sock_flag(sk, SOCK_DEAD))
4432 			sk->sk_data_ready(sk, 0);
4433 		return;
4434 	}
4435 
4436 	if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4437 		/* A retransmit, 2nd most common case.  Force an immediate ack. */
4438 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4439 		tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4440 
4441 out_of_window:
4442 		tcp_enter_quickack_mode(sk);
4443 		inet_csk_schedule_ack(sk);
4444 drop:
4445 		__kfree_skb(skb);
4446 		return;
4447 	}
4448 
4449 	/* Out of window. F.e. zero window probe. */
4450 	if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4451 		goto out_of_window;
4452 
4453 	tcp_enter_quickack_mode(sk);
4454 
4455 	if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4456 		/* Partial packet, seq < rcv_next < end_seq */
4457 		SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4458 			   tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4459 			   TCP_SKB_CB(skb)->end_seq);
4460 
4461 		tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4462 
4463 		/* If window is closed, drop tail of packet. But after
4464 		 * remembering D-SACK for its head made in previous line.
4465 		 */
4466 		if (!tcp_receive_window(tp))
4467 			goto out_of_window;
4468 		goto queue_and_out;
4469 	}
4470 
4471 	TCP_ECN_check_ce(tp, skb);
4472 
4473 	if (tcp_try_rmem_schedule(sk, skb->truesize))
4474 		goto drop;
4475 
4476 	/* Disable header prediction. */
4477 	tp->pred_flags = 0;
4478 	inet_csk_schedule_ack(sk);
4479 
4480 	SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4481 		   tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4482 
4483 	skb_set_owner_r(skb, sk);
4484 
4485 	if (!skb_peek(&tp->out_of_order_queue)) {
4486 		/* Initial out of order segment, build 1 SACK. */
4487 		if (tcp_is_sack(tp)) {
4488 			tp->rx_opt.num_sacks = 1;
4489 			tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4490 			tp->selective_acks[0].end_seq =
4491 						TCP_SKB_CB(skb)->end_seq;
4492 		}
4493 		__skb_queue_head(&tp->out_of_order_queue, skb);
4494 	} else {
4495 		struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue);
4496 		u32 seq = TCP_SKB_CB(skb)->seq;
4497 		u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4498 
4499 		if (seq == TCP_SKB_CB(skb1)->end_seq) {
4500 			__skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4501 
4502 			if (!tp->rx_opt.num_sacks ||
4503 			    tp->selective_acks[0].end_seq != seq)
4504 				goto add_sack;
4505 
4506 			/* Common case: data arrive in order after hole. */
4507 			tp->selective_acks[0].end_seq = end_seq;
4508 			return;
4509 		}
4510 
4511 		/* Find place to insert this segment. */
4512 		while (1) {
4513 			if (!after(TCP_SKB_CB(skb1)->seq, seq))
4514 				break;
4515 			if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4516 				skb1 = NULL;
4517 				break;
4518 			}
4519 			skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4520 		}
4521 
4522 		/* Do skb overlap to previous one? */
4523 		if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4524 			if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4525 				/* All the bits are present. Drop. */
4526 				__kfree_skb(skb);
4527 				tcp_dsack_set(sk, seq, end_seq);
4528 				goto add_sack;
4529 			}
4530 			if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4531 				/* Partial overlap. */
4532 				tcp_dsack_set(sk, seq,
4533 					      TCP_SKB_CB(skb1)->end_seq);
4534 			} else {
4535 				if (skb_queue_is_first(&tp->out_of_order_queue,
4536 						       skb1))
4537 					skb1 = NULL;
4538 				else
4539 					skb1 = skb_queue_prev(
4540 						&tp->out_of_order_queue,
4541 						skb1);
4542 			}
4543 		}
4544 		if (!skb1)
4545 			__skb_queue_head(&tp->out_of_order_queue, skb);
4546 		else
4547 			__skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4548 
4549 		/* And clean segments covered by new one as whole. */
4550 		while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4551 			skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4552 
4553 			if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4554 				break;
4555 			if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4556 				tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4557 						 end_seq);
4558 				break;
4559 			}
4560 			__skb_unlink(skb1, &tp->out_of_order_queue);
4561 			tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4562 					 TCP_SKB_CB(skb1)->end_seq);
4563 			__kfree_skb(skb1);
4564 		}
4565 
4566 add_sack:
4567 		if (tcp_is_sack(tp))
4568 			tcp_sack_new_ofo_skb(sk, seq, end_seq);
4569 	}
4570 }
4571 
4572 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4573 					struct sk_buff_head *list)
4574 {
4575 	struct sk_buff *next = NULL;
4576 
4577 	if (!skb_queue_is_last(list, skb))
4578 		next = skb_queue_next(list, skb);
4579 
4580 	__skb_unlink(skb, list);
4581 	__kfree_skb(skb);
4582 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4583 
4584 	return next;
4585 }
4586 
4587 /* Collapse contiguous sequence of skbs head..tail with
4588  * sequence numbers start..end.
4589  *
4590  * If tail is NULL, this means until the end of the list.
4591  *
4592  * Segments with FIN/SYN are not collapsed (only because this
4593  * simplifies code)
4594  */
4595 static void
4596 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4597 	     struct sk_buff *head, struct sk_buff *tail,
4598 	     u32 start, u32 end)
4599 {
4600 	struct sk_buff *skb, *n;
4601 	bool end_of_skbs;
4602 
4603 	/* First, check that queue is collapsible and find
4604 	 * the point where collapsing can be useful. */
4605 	skb = head;
4606 restart:
4607 	end_of_skbs = true;
4608 	skb_queue_walk_from_safe(list, skb, n) {
4609 		if (skb == tail)
4610 			break;
4611 		/* No new bits? It is possible on ofo queue. */
4612 		if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4613 			skb = tcp_collapse_one(sk, skb, list);
4614 			if (!skb)
4615 				break;
4616 			goto restart;
4617 		}
4618 
4619 		/* The first skb to collapse is:
4620 		 * - not SYN/FIN and
4621 		 * - bloated or contains data before "start" or
4622 		 *   overlaps to the next one.
4623 		 */
4624 		if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4625 		    (tcp_win_from_space(skb->truesize) > skb->len ||
4626 		     before(TCP_SKB_CB(skb)->seq, start))) {
4627 			end_of_skbs = false;
4628 			break;
4629 		}
4630 
4631 		if (!skb_queue_is_last(list, skb)) {
4632 			struct sk_buff *next = skb_queue_next(list, skb);
4633 			if (next != tail &&
4634 			    TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4635 				end_of_skbs = false;
4636 				break;
4637 			}
4638 		}
4639 
4640 		/* Decided to skip this, advance start seq. */
4641 		start = TCP_SKB_CB(skb)->end_seq;
4642 	}
4643 	if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4644 		return;
4645 
4646 	while (before(start, end)) {
4647 		struct sk_buff *nskb;
4648 		unsigned int header = skb_headroom(skb);
4649 		int copy = SKB_MAX_ORDER(header, 0);
4650 
4651 		/* Too big header? This can happen with IPv6. */
4652 		if (copy < 0)
4653 			return;
4654 		if (end - start < copy)
4655 			copy = end - start;
4656 		nskb = alloc_skb(copy + header, GFP_ATOMIC);
4657 		if (!nskb)
4658 			return;
4659 
4660 		skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4661 		skb_set_network_header(nskb, (skb_network_header(skb) -
4662 					      skb->head));
4663 		skb_set_transport_header(nskb, (skb_transport_header(skb) -
4664 						skb->head));
4665 		skb_reserve(nskb, header);
4666 		memcpy(nskb->head, skb->head, header);
4667 		memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4668 		TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4669 		__skb_queue_before(list, skb, nskb);
4670 		skb_set_owner_r(nskb, sk);
4671 
4672 		/* Copy data, releasing collapsed skbs. */
4673 		while (copy > 0) {
4674 			int offset = start - TCP_SKB_CB(skb)->seq;
4675 			int size = TCP_SKB_CB(skb)->end_seq - start;
4676 
4677 			BUG_ON(offset < 0);
4678 			if (size > 0) {
4679 				size = min(copy, size);
4680 				if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4681 					BUG();
4682 				TCP_SKB_CB(nskb)->end_seq += size;
4683 				copy -= size;
4684 				start += size;
4685 			}
4686 			if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4687 				skb = tcp_collapse_one(sk, skb, list);
4688 				if (!skb ||
4689 				    skb == tail ||
4690 				    tcp_hdr(skb)->syn ||
4691 				    tcp_hdr(skb)->fin)
4692 					return;
4693 			}
4694 		}
4695 	}
4696 }
4697 
4698 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4699  * and tcp_collapse() them until all the queue is collapsed.
4700  */
4701 static void tcp_collapse_ofo_queue(struct sock *sk)
4702 {
4703 	struct tcp_sock *tp = tcp_sk(sk);
4704 	struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4705 	struct sk_buff *head;
4706 	u32 start, end;
4707 
4708 	if (skb == NULL)
4709 		return;
4710 
4711 	start = TCP_SKB_CB(skb)->seq;
4712 	end = TCP_SKB_CB(skb)->end_seq;
4713 	head = skb;
4714 
4715 	for (;;) {
4716 		struct sk_buff *next = NULL;
4717 
4718 		if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4719 			next = skb_queue_next(&tp->out_of_order_queue, skb);
4720 		skb = next;
4721 
4722 		/* Segment is terminated when we see gap or when
4723 		 * we are at the end of all the queue. */
4724 		if (!skb ||
4725 		    after(TCP_SKB_CB(skb)->seq, end) ||
4726 		    before(TCP_SKB_CB(skb)->end_seq, start)) {
4727 			tcp_collapse(sk, &tp->out_of_order_queue,
4728 				     head, skb, start, end);
4729 			head = skb;
4730 			if (!skb)
4731 				break;
4732 			/* Start new segment */
4733 			start = TCP_SKB_CB(skb)->seq;
4734 			end = TCP_SKB_CB(skb)->end_seq;
4735 		} else {
4736 			if (before(TCP_SKB_CB(skb)->seq, start))
4737 				start = TCP_SKB_CB(skb)->seq;
4738 			if (after(TCP_SKB_CB(skb)->end_seq, end))
4739 				end = TCP_SKB_CB(skb)->end_seq;
4740 		}
4741 	}
4742 }
4743 
4744 /*
4745  * Purge the out-of-order queue.
4746  * Return true if queue was pruned.
4747  */
4748 static int tcp_prune_ofo_queue(struct sock *sk)
4749 {
4750 	struct tcp_sock *tp = tcp_sk(sk);
4751 	int res = 0;
4752 
4753 	if (!skb_queue_empty(&tp->out_of_order_queue)) {
4754 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4755 		__skb_queue_purge(&tp->out_of_order_queue);
4756 
4757 		/* Reset SACK state.  A conforming SACK implementation will
4758 		 * do the same at a timeout based retransmit.  When a connection
4759 		 * is in a sad state like this, we care only about integrity
4760 		 * of the connection not performance.
4761 		 */
4762 		if (tp->rx_opt.sack_ok)
4763 			tcp_sack_reset(&tp->rx_opt);
4764 		sk_mem_reclaim(sk);
4765 		res = 1;
4766 	}
4767 	return res;
4768 }
4769 
4770 /* Reduce allocated memory if we can, trying to get
4771  * the socket within its memory limits again.
4772  *
4773  * Return less than zero if we should start dropping frames
4774  * until the socket owning process reads some of the data
4775  * to stabilize the situation.
4776  */
4777 static int tcp_prune_queue(struct sock *sk)
4778 {
4779 	struct tcp_sock *tp = tcp_sk(sk);
4780 
4781 	SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4782 
4783 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4784 
4785 	if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4786 		tcp_clamp_window(sk);
4787 	else if (tcp_memory_pressure)
4788 		tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4789 
4790 	tcp_collapse_ofo_queue(sk);
4791 	if (!skb_queue_empty(&sk->sk_receive_queue))
4792 		tcp_collapse(sk, &sk->sk_receive_queue,
4793 			     skb_peek(&sk->sk_receive_queue),
4794 			     NULL,
4795 			     tp->copied_seq, tp->rcv_nxt);
4796 	sk_mem_reclaim(sk);
4797 
4798 	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4799 		return 0;
4800 
4801 	/* Collapsing did not help, destructive actions follow.
4802 	 * This must not ever occur. */
4803 
4804 	tcp_prune_ofo_queue(sk);
4805 
4806 	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4807 		return 0;
4808 
4809 	/* If we are really being abused, tell the caller to silently
4810 	 * drop receive data on the floor.  It will get retransmitted
4811 	 * and hopefully then we'll have sufficient space.
4812 	 */
4813 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4814 
4815 	/* Massive buffer overcommit. */
4816 	tp->pred_flags = 0;
4817 	return -1;
4818 }
4819 
4820 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4821  * As additional protections, we do not touch cwnd in retransmission phases,
4822  * and if application hit its sndbuf limit recently.
4823  */
4824 void tcp_cwnd_application_limited(struct sock *sk)
4825 {
4826 	struct tcp_sock *tp = tcp_sk(sk);
4827 
4828 	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4829 	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4830 		/* Limited by application or receiver window. */
4831 		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4832 		u32 win_used = max(tp->snd_cwnd_used, init_win);
4833 		if (win_used < tp->snd_cwnd) {
4834 			tp->snd_ssthresh = tcp_current_ssthresh(sk);
4835 			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4836 		}
4837 		tp->snd_cwnd_used = 0;
4838 	}
4839 	tp->snd_cwnd_stamp = tcp_time_stamp;
4840 }
4841 
4842 static int tcp_should_expand_sndbuf(struct sock *sk)
4843 {
4844 	struct tcp_sock *tp = tcp_sk(sk);
4845 
4846 	/* If the user specified a specific send buffer setting, do
4847 	 * not modify it.
4848 	 */
4849 	if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4850 		return 0;
4851 
4852 	/* If we are under global TCP memory pressure, do not expand.  */
4853 	if (tcp_memory_pressure)
4854 		return 0;
4855 
4856 	/* If we are under soft global TCP memory pressure, do not expand.  */
4857 	if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4858 		return 0;
4859 
4860 	/* If we filled the congestion window, do not expand.  */
4861 	if (tp->packets_out >= tp->snd_cwnd)
4862 		return 0;
4863 
4864 	return 1;
4865 }
4866 
4867 /* When incoming ACK allowed to free some skb from write_queue,
4868  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4869  * on the exit from tcp input handler.
4870  *
4871  * PROBLEM: sndbuf expansion does not work well with largesend.
4872  */
4873 static void tcp_new_space(struct sock *sk)
4874 {
4875 	struct tcp_sock *tp = tcp_sk(sk);
4876 
4877 	if (tcp_should_expand_sndbuf(sk)) {
4878 		int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4879 			MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4880 		int demanded = max_t(unsigned int, tp->snd_cwnd,
4881 				     tp->reordering + 1);
4882 		sndmem *= 2 * demanded;
4883 		if (sndmem > sk->sk_sndbuf)
4884 			sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4885 		tp->snd_cwnd_stamp = tcp_time_stamp;
4886 	}
4887 
4888 	sk->sk_write_space(sk);
4889 }
4890 
4891 static void tcp_check_space(struct sock *sk)
4892 {
4893 	if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4894 		sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4895 		if (sk->sk_socket &&
4896 		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4897 			tcp_new_space(sk);
4898 	}
4899 }
4900 
4901 static inline void tcp_data_snd_check(struct sock *sk)
4902 {
4903 	tcp_push_pending_frames(sk);
4904 	tcp_check_space(sk);
4905 }
4906 
4907 /*
4908  * Check if sending an ack is needed.
4909  */
4910 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4911 {
4912 	struct tcp_sock *tp = tcp_sk(sk);
4913 
4914 	    /* More than one full frame received... */
4915 	if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4916 	     /* ... and right edge of window advances far enough.
4917 	      * (tcp_recvmsg() will send ACK otherwise). Or...
4918 	      */
4919 	     __tcp_select_window(sk) >= tp->rcv_wnd) ||
4920 	    /* We ACK each frame or... */
4921 	    tcp_in_quickack_mode(sk) ||
4922 	    /* We have out of order data. */
4923 	    (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4924 		/* Then ack it now */
4925 		tcp_send_ack(sk);
4926 	} else {
4927 		/* Else, send delayed ack. */
4928 		tcp_send_delayed_ack(sk);
4929 	}
4930 }
4931 
4932 static inline void tcp_ack_snd_check(struct sock *sk)
4933 {
4934 	if (!inet_csk_ack_scheduled(sk)) {
4935 		/* We sent a data segment already. */
4936 		return;
4937 	}
4938 	__tcp_ack_snd_check(sk, 1);
4939 }
4940 
4941 /*
4942  *	This routine is only called when we have urgent data
4943  *	signaled. Its the 'slow' part of tcp_urg. It could be
4944  *	moved inline now as tcp_urg is only called from one
4945  *	place. We handle URGent data wrong. We have to - as
4946  *	BSD still doesn't use the correction from RFC961.
4947  *	For 1003.1g we should support a new option TCP_STDURG to permit
4948  *	either form (or just set the sysctl tcp_stdurg).
4949  */
4950 
4951 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4952 {
4953 	struct tcp_sock *tp = tcp_sk(sk);
4954 	u32 ptr = ntohs(th->urg_ptr);
4955 
4956 	if (ptr && !sysctl_tcp_stdurg)
4957 		ptr--;
4958 	ptr += ntohl(th->seq);
4959 
4960 	/* Ignore urgent data that we've already seen and read. */
4961 	if (after(tp->copied_seq, ptr))
4962 		return;
4963 
4964 	/* Do not replay urg ptr.
4965 	 *
4966 	 * NOTE: interesting situation not covered by specs.
4967 	 * Misbehaving sender may send urg ptr, pointing to segment,
4968 	 * which we already have in ofo queue. We are not able to fetch
4969 	 * such data and will stay in TCP_URG_NOTYET until will be eaten
4970 	 * by recvmsg(). Seems, we are not obliged to handle such wicked
4971 	 * situations. But it is worth to think about possibility of some
4972 	 * DoSes using some hypothetical application level deadlock.
4973 	 */
4974 	if (before(ptr, tp->rcv_nxt))
4975 		return;
4976 
4977 	/* Do we already have a newer (or duplicate) urgent pointer? */
4978 	if (tp->urg_data && !after(ptr, tp->urg_seq))
4979 		return;
4980 
4981 	/* Tell the world about our new urgent pointer. */
4982 	sk_send_sigurg(sk);
4983 
4984 	/* We may be adding urgent data when the last byte read was
4985 	 * urgent. To do this requires some care. We cannot just ignore
4986 	 * tp->copied_seq since we would read the last urgent byte again
4987 	 * as data, nor can we alter copied_seq until this data arrives
4988 	 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4989 	 *
4990 	 * NOTE. Double Dutch. Rendering to plain English: author of comment
4991 	 * above did something sort of 	send("A", MSG_OOB); send("B", MSG_OOB);
4992 	 * and expect that both A and B disappear from stream. This is _wrong_.
4993 	 * Though this happens in BSD with high probability, this is occasional.
4994 	 * Any application relying on this is buggy. Note also, that fix "works"
4995 	 * only in this artificial test. Insert some normal data between A and B and we will
4996 	 * decline of BSD again. Verdict: it is better to remove to trap
4997 	 * buggy users.
4998 	 */
4999 	if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5000 	    !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5001 		struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5002 		tp->copied_seq++;
5003 		if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5004 			__skb_unlink(skb, &sk->sk_receive_queue);
5005 			__kfree_skb(skb);
5006 		}
5007 	}
5008 
5009 	tp->urg_data = TCP_URG_NOTYET;
5010 	tp->urg_seq = ptr;
5011 
5012 	/* Disable header prediction. */
5013 	tp->pred_flags = 0;
5014 }
5015 
5016 /* This is the 'fast' part of urgent handling. */
5017 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
5018 {
5019 	struct tcp_sock *tp = tcp_sk(sk);
5020 
5021 	/* Check if we get a new urgent pointer - normally not. */
5022 	if (th->urg)
5023 		tcp_check_urg(sk, th);
5024 
5025 	/* Do we wait for any urgent data? - normally not... */
5026 	if (tp->urg_data == TCP_URG_NOTYET) {
5027 		u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5028 			  th->syn;
5029 
5030 		/* Is the urgent pointer pointing into this packet? */
5031 		if (ptr < skb->len) {
5032 			u8 tmp;
5033 			if (skb_copy_bits(skb, ptr, &tmp, 1))
5034 				BUG();
5035 			tp->urg_data = TCP_URG_VALID | tmp;
5036 			if (!sock_flag(sk, SOCK_DEAD))
5037 				sk->sk_data_ready(sk, 0);
5038 		}
5039 	}
5040 }
5041 
5042 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5043 {
5044 	struct tcp_sock *tp = tcp_sk(sk);
5045 	int chunk = skb->len - hlen;
5046 	int err;
5047 
5048 	local_bh_enable();
5049 	if (skb_csum_unnecessary(skb))
5050 		err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5051 	else
5052 		err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5053 						       tp->ucopy.iov);
5054 
5055 	if (!err) {
5056 		tp->ucopy.len -= chunk;
5057 		tp->copied_seq += chunk;
5058 		tcp_rcv_space_adjust(sk);
5059 	}
5060 
5061 	local_bh_disable();
5062 	return err;
5063 }
5064 
5065 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5066 					    struct sk_buff *skb)
5067 {
5068 	__sum16 result;
5069 
5070 	if (sock_owned_by_user(sk)) {
5071 		local_bh_enable();
5072 		result = __tcp_checksum_complete(skb);
5073 		local_bh_disable();
5074 	} else {
5075 		result = __tcp_checksum_complete(skb);
5076 	}
5077 	return result;
5078 }
5079 
5080 static inline int tcp_checksum_complete_user(struct sock *sk,
5081 					     struct sk_buff *skb)
5082 {
5083 	return !skb_csum_unnecessary(skb) &&
5084 	       __tcp_checksum_complete_user(sk, skb);
5085 }
5086 
5087 #ifdef CONFIG_NET_DMA
5088 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5089 				  int hlen)
5090 {
5091 	struct tcp_sock *tp = tcp_sk(sk);
5092 	int chunk = skb->len - hlen;
5093 	int dma_cookie;
5094 	int copied_early = 0;
5095 
5096 	if (tp->ucopy.wakeup)
5097 		return 0;
5098 
5099 	if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5100 		tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5101 
5102 	if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5103 
5104 		dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5105 							 skb, hlen,
5106 							 tp->ucopy.iov, chunk,
5107 							 tp->ucopy.pinned_list);
5108 
5109 		if (dma_cookie < 0)
5110 			goto out;
5111 
5112 		tp->ucopy.dma_cookie = dma_cookie;
5113 		copied_early = 1;
5114 
5115 		tp->ucopy.len -= chunk;
5116 		tp->copied_seq += chunk;
5117 		tcp_rcv_space_adjust(sk);
5118 
5119 		if ((tp->ucopy.len == 0) ||
5120 		    (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5121 		    (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5122 			tp->ucopy.wakeup = 1;
5123 			sk->sk_data_ready(sk, 0);
5124 		}
5125 	} else if (chunk > 0) {
5126 		tp->ucopy.wakeup = 1;
5127 		sk->sk_data_ready(sk, 0);
5128 	}
5129 out:
5130 	return copied_early;
5131 }
5132 #endif /* CONFIG_NET_DMA */
5133 
5134 /* Does PAWS and seqno based validation of an incoming segment, flags will
5135  * play significant role here.
5136  */
5137 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5138 			      struct tcphdr *th, int syn_inerr)
5139 {
5140 	u8 *hash_location;
5141 	struct tcp_sock *tp = tcp_sk(sk);
5142 
5143 	/* RFC1323: H1. Apply PAWS check first. */
5144 	if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5145 	    tp->rx_opt.saw_tstamp &&
5146 	    tcp_paws_discard(sk, skb)) {
5147 		if (!th->rst) {
5148 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5149 			tcp_send_dupack(sk, skb);
5150 			goto discard;
5151 		}
5152 		/* Reset is accepted even if it did not pass PAWS. */
5153 	}
5154 
5155 	/* Step 1: check sequence number */
5156 	if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5157 		/* RFC793, page 37: "In all states except SYN-SENT, all reset
5158 		 * (RST) segments are validated by checking their SEQ-fields."
5159 		 * And page 69: "If an incoming segment is not acceptable,
5160 		 * an acknowledgment should be sent in reply (unless the RST
5161 		 * bit is set, if so drop the segment and return)".
5162 		 */
5163 		if (!th->rst)
5164 			tcp_send_dupack(sk, skb);
5165 		goto discard;
5166 	}
5167 
5168 	/* Step 2: check RST bit */
5169 	if (th->rst) {
5170 		tcp_reset(sk);
5171 		goto discard;
5172 	}
5173 
5174 	/* ts_recent update must be made after we are sure that the packet
5175 	 * is in window.
5176 	 */
5177 	tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5178 
5179 	/* step 3: check security and precedence [ignored] */
5180 
5181 	/* step 4: Check for a SYN in window. */
5182 	if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5183 		if (syn_inerr)
5184 			TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5185 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5186 		tcp_reset(sk);
5187 		return -1;
5188 	}
5189 
5190 	return 1;
5191 
5192 discard:
5193 	__kfree_skb(skb);
5194 	return 0;
5195 }
5196 
5197 /*
5198  *	TCP receive function for the ESTABLISHED state.
5199  *
5200  *	It is split into a fast path and a slow path. The fast path is
5201  * 	disabled when:
5202  *	- A zero window was announced from us - zero window probing
5203  *        is only handled properly in the slow path.
5204  *	- Out of order segments arrived.
5205  *	- Urgent data is expected.
5206  *	- There is no buffer space left
5207  *	- Unexpected TCP flags/window values/header lengths are received
5208  *	  (detected by checking the TCP header against pred_flags)
5209  *	- Data is sent in both directions. Fast path only supports pure senders
5210  *	  or pure receivers (this means either the sequence number or the ack
5211  *	  value must stay constant)
5212  *	- Unexpected TCP option.
5213  *
5214  *	When these conditions are not satisfied it drops into a standard
5215  *	receive procedure patterned after RFC793 to handle all cases.
5216  *	The first three cases are guaranteed by proper pred_flags setting,
5217  *	the rest is checked inline. Fast processing is turned on in
5218  *	tcp_data_queue when everything is OK.
5219  */
5220 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5221 			struct tcphdr *th, unsigned len)
5222 {
5223 	struct tcp_sock *tp = tcp_sk(sk);
5224 	int res;
5225 
5226 	/*
5227 	 *	Header prediction.
5228 	 *	The code loosely follows the one in the famous
5229 	 *	"30 instruction TCP receive" Van Jacobson mail.
5230 	 *
5231 	 *	Van's trick is to deposit buffers into socket queue
5232 	 *	on a device interrupt, to call tcp_recv function
5233 	 *	on the receive process context and checksum and copy
5234 	 *	the buffer to user space. smart...
5235 	 *
5236 	 *	Our current scheme is not silly either but we take the
5237 	 *	extra cost of the net_bh soft interrupt processing...
5238 	 *	We do checksum and copy also but from device to kernel.
5239 	 */
5240 
5241 	tp->rx_opt.saw_tstamp = 0;
5242 
5243 	/*	pred_flags is 0xS?10 << 16 + snd_wnd
5244 	 *	if header_prediction is to be made
5245 	 *	'S' will always be tp->tcp_header_len >> 2
5246 	 *	'?' will be 0 for the fast path, otherwise pred_flags is 0 to
5247 	 *  turn it off	(when there are holes in the receive
5248 	 *	 space for instance)
5249 	 *	PSH flag is ignored.
5250 	 */
5251 
5252 	if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5253 	    TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5254 	    !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5255 		int tcp_header_len = tp->tcp_header_len;
5256 
5257 		/* Timestamp header prediction: tcp_header_len
5258 		 * is automatically equal to th->doff*4 due to pred_flags
5259 		 * match.
5260 		 */
5261 
5262 		/* Check timestamp */
5263 		if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5264 			/* No? Slow path! */
5265 			if (!tcp_parse_aligned_timestamp(tp, th))
5266 				goto slow_path;
5267 
5268 			/* If PAWS failed, check it more carefully in slow path */
5269 			if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5270 				goto slow_path;
5271 
5272 			/* DO NOT update ts_recent here, if checksum fails
5273 			 * and timestamp was corrupted part, it will result
5274 			 * in a hung connection since we will drop all
5275 			 * future packets due to the PAWS test.
5276 			 */
5277 		}
5278 
5279 		if (len <= tcp_header_len) {
5280 			/* Bulk data transfer: sender */
5281 			if (len == tcp_header_len) {
5282 				/* Predicted packet is in window by definition.
5283 				 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5284 				 * Hence, check seq<=rcv_wup reduces to:
5285 				 */
5286 				if (tcp_header_len ==
5287 				    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5288 				    tp->rcv_nxt == tp->rcv_wup)
5289 					tcp_store_ts_recent(tp);
5290 
5291 				/* We know that such packets are checksummed
5292 				 * on entry.
5293 				 */
5294 				tcp_ack(sk, skb, 0);
5295 				__kfree_skb(skb);
5296 				tcp_data_snd_check(sk);
5297 				return 0;
5298 			} else { /* Header too small */
5299 				TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5300 				goto discard;
5301 			}
5302 		} else {
5303 			int eaten = 0;
5304 			int copied_early = 0;
5305 
5306 			if (tp->copied_seq == tp->rcv_nxt &&
5307 			    len - tcp_header_len <= tp->ucopy.len) {
5308 #ifdef CONFIG_NET_DMA
5309 				if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5310 					copied_early = 1;
5311 					eaten = 1;
5312 				}
5313 #endif
5314 				if (tp->ucopy.task == current &&
5315 				    sock_owned_by_user(sk) && !copied_early) {
5316 					__set_current_state(TASK_RUNNING);
5317 
5318 					if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5319 						eaten = 1;
5320 				}
5321 				if (eaten) {
5322 					/* Predicted packet is in window by definition.
5323 					 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5324 					 * Hence, check seq<=rcv_wup reduces to:
5325 					 */
5326 					if (tcp_header_len ==
5327 					    (sizeof(struct tcphdr) +
5328 					     TCPOLEN_TSTAMP_ALIGNED) &&
5329 					    tp->rcv_nxt == tp->rcv_wup)
5330 						tcp_store_ts_recent(tp);
5331 
5332 					tcp_rcv_rtt_measure_ts(sk, skb);
5333 
5334 					__skb_pull(skb, tcp_header_len);
5335 					tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5336 					NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5337 				}
5338 				if (copied_early)
5339 					tcp_cleanup_rbuf(sk, skb->len);
5340 			}
5341 			if (!eaten) {
5342 				if (tcp_checksum_complete_user(sk, skb))
5343 					goto csum_error;
5344 
5345 				/* Predicted packet is in window by definition.
5346 				 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5347 				 * Hence, check seq<=rcv_wup reduces to:
5348 				 */
5349 				if (tcp_header_len ==
5350 				    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5351 				    tp->rcv_nxt == tp->rcv_wup)
5352 					tcp_store_ts_recent(tp);
5353 
5354 				tcp_rcv_rtt_measure_ts(sk, skb);
5355 
5356 				if ((int)skb->truesize > sk->sk_forward_alloc)
5357 					goto step5;
5358 
5359 				NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5360 
5361 				/* Bulk data transfer: receiver */
5362 				__skb_pull(skb, tcp_header_len);
5363 				__skb_queue_tail(&sk->sk_receive_queue, skb);
5364 				skb_set_owner_r(skb, sk);
5365 				tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5366 			}
5367 
5368 			tcp_event_data_recv(sk, skb);
5369 
5370 			if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5371 				/* Well, only one small jumplet in fast path... */
5372 				tcp_ack(sk, skb, FLAG_DATA);
5373 				tcp_data_snd_check(sk);
5374 				if (!inet_csk_ack_scheduled(sk))
5375 					goto no_ack;
5376 			}
5377 
5378 			if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5379 				__tcp_ack_snd_check(sk, 0);
5380 no_ack:
5381 #ifdef CONFIG_NET_DMA
5382 			if (copied_early)
5383 				__skb_queue_tail(&sk->sk_async_wait_queue, skb);
5384 			else
5385 #endif
5386 			if (eaten)
5387 				__kfree_skb(skb);
5388 			else
5389 				sk->sk_data_ready(sk, 0);
5390 			return 0;
5391 		}
5392 	}
5393 
5394 slow_path:
5395 	if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5396 		goto csum_error;
5397 
5398 	/*
5399 	 *	Standard slow path.
5400 	 */
5401 
5402 	res = tcp_validate_incoming(sk, skb, th, 1);
5403 	if (res <= 0)
5404 		return -res;
5405 
5406 step5:
5407 	if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5408 		goto discard;
5409 
5410 	tcp_rcv_rtt_measure_ts(sk, skb);
5411 
5412 	/* Process urgent data. */
5413 	tcp_urg(sk, skb, th);
5414 
5415 	/* step 7: process the segment text */
5416 	tcp_data_queue(sk, skb);
5417 
5418 	tcp_data_snd_check(sk);
5419 	tcp_ack_snd_check(sk);
5420 	return 0;
5421 
5422 csum_error:
5423 	TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5424 
5425 discard:
5426 	__kfree_skb(skb);
5427 	return 0;
5428 }
5429 
5430 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5431 					 struct tcphdr *th, unsigned len)
5432 {
5433 	u8 *hash_location;
5434 	struct inet_connection_sock *icsk = inet_csk(sk);
5435 	struct tcp_sock *tp = tcp_sk(sk);
5436 	struct tcp_cookie_values *cvp = tp->cookie_values;
5437 	int saved_clamp = tp->rx_opt.mss_clamp;
5438 
5439 	tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5440 
5441 	if (th->ack) {
5442 		/* rfc793:
5443 		 * "If the state is SYN-SENT then
5444 		 *    first check the ACK bit
5445 		 *      If the ACK bit is set
5446 		 *	  If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5447 		 *        a reset (unless the RST bit is set, if so drop
5448 		 *        the segment and return)"
5449 		 *
5450 		 *  We do not send data with SYN, so that RFC-correct
5451 		 *  test reduces to:
5452 		 */
5453 		if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5454 			goto reset_and_undo;
5455 
5456 		if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5457 		    !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5458 			     tcp_time_stamp)) {
5459 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5460 			goto reset_and_undo;
5461 		}
5462 
5463 		/* Now ACK is acceptable.
5464 		 *
5465 		 * "If the RST bit is set
5466 		 *    If the ACK was acceptable then signal the user "error:
5467 		 *    connection reset", drop the segment, enter CLOSED state,
5468 		 *    delete TCB, and return."
5469 		 */
5470 
5471 		if (th->rst) {
5472 			tcp_reset(sk);
5473 			goto discard;
5474 		}
5475 
5476 		/* rfc793:
5477 		 *   "fifth, if neither of the SYN or RST bits is set then
5478 		 *    drop the segment and return."
5479 		 *
5480 		 *    See note below!
5481 		 *                                        --ANK(990513)
5482 		 */
5483 		if (!th->syn)
5484 			goto discard_and_undo;
5485 
5486 		/* rfc793:
5487 		 *   "If the SYN bit is on ...
5488 		 *    are acceptable then ...
5489 		 *    (our SYN has been ACKed), change the connection
5490 		 *    state to ESTABLISHED..."
5491 		 */
5492 
5493 		TCP_ECN_rcv_synack(tp, th);
5494 
5495 		tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5496 		tcp_ack(sk, skb, FLAG_SLOWPATH);
5497 
5498 		/* Ok.. it's good. Set up sequence numbers and
5499 		 * move to established.
5500 		 */
5501 		tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5502 		tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5503 
5504 		/* RFC1323: The window in SYN & SYN/ACK segments is
5505 		 * never scaled.
5506 		 */
5507 		tp->snd_wnd = ntohs(th->window);
5508 		tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5509 
5510 		if (!tp->rx_opt.wscale_ok) {
5511 			tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5512 			tp->window_clamp = min(tp->window_clamp, 65535U);
5513 		}
5514 
5515 		if (tp->rx_opt.saw_tstamp) {
5516 			tp->rx_opt.tstamp_ok	   = 1;
5517 			tp->tcp_header_len =
5518 				sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5519 			tp->advmss	    -= TCPOLEN_TSTAMP_ALIGNED;
5520 			tcp_store_ts_recent(tp);
5521 		} else {
5522 			tp->tcp_header_len = sizeof(struct tcphdr);
5523 		}
5524 
5525 		if (tcp_is_sack(tp) && sysctl_tcp_fack)
5526 			tcp_enable_fack(tp);
5527 
5528 		tcp_mtup_init(sk);
5529 		tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5530 		tcp_initialize_rcv_mss(sk);
5531 
5532 		/* Remember, tcp_poll() does not lock socket!
5533 		 * Change state from SYN-SENT only after copied_seq
5534 		 * is initialized. */
5535 		tp->copied_seq = tp->rcv_nxt;
5536 
5537 		if (cvp != NULL &&
5538 		    cvp->cookie_pair_size > 0 &&
5539 		    tp->rx_opt.cookie_plus > 0) {
5540 			int cookie_size = tp->rx_opt.cookie_plus
5541 					- TCPOLEN_COOKIE_BASE;
5542 			int cookie_pair_size = cookie_size
5543 					     + cvp->cookie_desired;
5544 
5545 			/* A cookie extension option was sent and returned.
5546 			 * Note that each incoming SYNACK replaces the
5547 			 * Responder cookie.  The initial exchange is most
5548 			 * fragile, as protection against spoofing relies
5549 			 * entirely upon the sequence and timestamp (above).
5550 			 * This replacement strategy allows the correct pair to
5551 			 * pass through, while any others will be filtered via
5552 			 * Responder verification later.
5553 			 */
5554 			if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5555 				memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5556 				       hash_location, cookie_size);
5557 				cvp->cookie_pair_size = cookie_pair_size;
5558 			}
5559 		}
5560 
5561 		smp_mb();
5562 		tcp_set_state(sk, TCP_ESTABLISHED);
5563 
5564 		security_inet_conn_established(sk, skb);
5565 
5566 		/* Make sure socket is routed, for correct metrics.  */
5567 		icsk->icsk_af_ops->rebuild_header(sk);
5568 
5569 		tcp_init_metrics(sk);
5570 
5571 		tcp_init_congestion_control(sk);
5572 
5573 		/* Prevent spurious tcp_cwnd_restart() on first data
5574 		 * packet.
5575 		 */
5576 		tp->lsndtime = tcp_time_stamp;
5577 
5578 		tcp_init_buffer_space(sk);
5579 
5580 		if (sock_flag(sk, SOCK_KEEPOPEN))
5581 			inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5582 
5583 		if (!tp->rx_opt.snd_wscale)
5584 			__tcp_fast_path_on(tp, tp->snd_wnd);
5585 		else
5586 			tp->pred_flags = 0;
5587 
5588 		if (!sock_flag(sk, SOCK_DEAD)) {
5589 			sk->sk_state_change(sk);
5590 			sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5591 		}
5592 
5593 		if (sk->sk_write_pending ||
5594 		    icsk->icsk_accept_queue.rskq_defer_accept ||
5595 		    icsk->icsk_ack.pingpong) {
5596 			/* Save one ACK. Data will be ready after
5597 			 * several ticks, if write_pending is set.
5598 			 *
5599 			 * It may be deleted, but with this feature tcpdumps
5600 			 * look so _wonderfully_ clever, that I was not able
5601 			 * to stand against the temptation 8)     --ANK
5602 			 */
5603 			inet_csk_schedule_ack(sk);
5604 			icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5605 			icsk->icsk_ack.ato	 = TCP_ATO_MIN;
5606 			tcp_incr_quickack(sk);
5607 			tcp_enter_quickack_mode(sk);
5608 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5609 						  TCP_DELACK_MAX, TCP_RTO_MAX);
5610 
5611 discard:
5612 			__kfree_skb(skb);
5613 			return 0;
5614 		} else {
5615 			tcp_send_ack(sk);
5616 		}
5617 		return -1;
5618 	}
5619 
5620 	/* No ACK in the segment */
5621 
5622 	if (th->rst) {
5623 		/* rfc793:
5624 		 * "If the RST bit is set
5625 		 *
5626 		 *      Otherwise (no ACK) drop the segment and return."
5627 		 */
5628 
5629 		goto discard_and_undo;
5630 	}
5631 
5632 	/* PAWS check. */
5633 	if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5634 	    tcp_paws_reject(&tp->rx_opt, 0))
5635 		goto discard_and_undo;
5636 
5637 	if (th->syn) {
5638 		/* We see SYN without ACK. It is attempt of
5639 		 * simultaneous connect with crossed SYNs.
5640 		 * Particularly, it can be connect to self.
5641 		 */
5642 		tcp_set_state(sk, TCP_SYN_RECV);
5643 
5644 		if (tp->rx_opt.saw_tstamp) {
5645 			tp->rx_opt.tstamp_ok = 1;
5646 			tcp_store_ts_recent(tp);
5647 			tp->tcp_header_len =
5648 				sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5649 		} else {
5650 			tp->tcp_header_len = sizeof(struct tcphdr);
5651 		}
5652 
5653 		tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5654 		tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5655 
5656 		/* RFC1323: The window in SYN & SYN/ACK segments is
5657 		 * never scaled.
5658 		 */
5659 		tp->snd_wnd    = ntohs(th->window);
5660 		tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5661 		tp->max_window = tp->snd_wnd;
5662 
5663 		TCP_ECN_rcv_syn(tp, th);
5664 
5665 		tcp_mtup_init(sk);
5666 		tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5667 		tcp_initialize_rcv_mss(sk);
5668 
5669 		tcp_send_synack(sk);
5670 #if 0
5671 		/* Note, we could accept data and URG from this segment.
5672 		 * There are no obstacles to make this.
5673 		 *
5674 		 * However, if we ignore data in ACKless segments sometimes,
5675 		 * we have no reasons to accept it sometimes.
5676 		 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5677 		 * is not flawless. So, discard packet for sanity.
5678 		 * Uncomment this return to process the data.
5679 		 */
5680 		return -1;
5681 #else
5682 		goto discard;
5683 #endif
5684 	}
5685 	/* "fifth, if neither of the SYN or RST bits is set then
5686 	 * drop the segment and return."
5687 	 */
5688 
5689 discard_and_undo:
5690 	tcp_clear_options(&tp->rx_opt);
5691 	tp->rx_opt.mss_clamp = saved_clamp;
5692 	goto discard;
5693 
5694 reset_and_undo:
5695 	tcp_clear_options(&tp->rx_opt);
5696 	tp->rx_opt.mss_clamp = saved_clamp;
5697 	return 1;
5698 }
5699 
5700 /*
5701  *	This function implements the receiving procedure of RFC 793 for
5702  *	all states except ESTABLISHED and TIME_WAIT.
5703  *	It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5704  *	address independent.
5705  */
5706 
5707 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5708 			  struct tcphdr *th, unsigned len)
5709 {
5710 	struct tcp_sock *tp = tcp_sk(sk);
5711 	struct inet_connection_sock *icsk = inet_csk(sk);
5712 	int queued = 0;
5713 	int res;
5714 
5715 	tp->rx_opt.saw_tstamp = 0;
5716 
5717 	switch (sk->sk_state) {
5718 	case TCP_CLOSE:
5719 		goto discard;
5720 
5721 	case TCP_LISTEN:
5722 		if (th->ack)
5723 			return 1;
5724 
5725 		if (th->rst)
5726 			goto discard;
5727 
5728 		if (th->syn) {
5729 			if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5730 				return 1;
5731 
5732 			/* Now we have several options: In theory there is
5733 			 * nothing else in the frame. KA9Q has an option to
5734 			 * send data with the syn, BSD accepts data with the
5735 			 * syn up to the [to be] advertised window and
5736 			 * Solaris 2.1 gives you a protocol error. For now
5737 			 * we just ignore it, that fits the spec precisely
5738 			 * and avoids incompatibilities. It would be nice in
5739 			 * future to drop through and process the data.
5740 			 *
5741 			 * Now that TTCP is starting to be used we ought to
5742 			 * queue this data.
5743 			 * But, this leaves one open to an easy denial of
5744 			 * service attack, and SYN cookies can't defend
5745 			 * against this problem. So, we drop the data
5746 			 * in the interest of security over speed unless
5747 			 * it's still in use.
5748 			 */
5749 			kfree_skb(skb);
5750 			return 0;
5751 		}
5752 		goto discard;
5753 
5754 	case TCP_SYN_SENT:
5755 		queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5756 		if (queued >= 0)
5757 			return queued;
5758 
5759 		/* Do step6 onward by hand. */
5760 		tcp_urg(sk, skb, th);
5761 		__kfree_skb(skb);
5762 		tcp_data_snd_check(sk);
5763 		return 0;
5764 	}
5765 
5766 	res = tcp_validate_incoming(sk, skb, th, 0);
5767 	if (res <= 0)
5768 		return -res;
5769 
5770 	/* step 5: check the ACK field */
5771 	if (th->ack) {
5772 		int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5773 
5774 		switch (sk->sk_state) {
5775 		case TCP_SYN_RECV:
5776 			if (acceptable) {
5777 				tp->copied_seq = tp->rcv_nxt;
5778 				smp_mb();
5779 				tcp_set_state(sk, TCP_ESTABLISHED);
5780 				sk->sk_state_change(sk);
5781 
5782 				/* Note, that this wakeup is only for marginal
5783 				 * crossed SYN case. Passively open sockets
5784 				 * are not waked up, because sk->sk_sleep ==
5785 				 * NULL and sk->sk_socket == NULL.
5786 				 */
5787 				if (sk->sk_socket)
5788 					sk_wake_async(sk,
5789 						      SOCK_WAKE_IO, POLL_OUT);
5790 
5791 				tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5792 				tp->snd_wnd = ntohs(th->window) <<
5793 					      tp->rx_opt.snd_wscale;
5794 				tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5795 
5796 				/* tcp_ack considers this ACK as duplicate
5797 				 * and does not calculate rtt.
5798 				 * Force it here.
5799 				 */
5800 				tcp_ack_update_rtt(sk, 0, 0);
5801 
5802 				if (tp->rx_opt.tstamp_ok)
5803 					tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5804 
5805 				/* Make sure socket is routed, for
5806 				 * correct metrics.
5807 				 */
5808 				icsk->icsk_af_ops->rebuild_header(sk);
5809 
5810 				tcp_init_metrics(sk);
5811 
5812 				tcp_init_congestion_control(sk);
5813 
5814 				/* Prevent spurious tcp_cwnd_restart() on
5815 				 * first data packet.
5816 				 */
5817 				tp->lsndtime = tcp_time_stamp;
5818 
5819 				tcp_mtup_init(sk);
5820 				tcp_initialize_rcv_mss(sk);
5821 				tcp_init_buffer_space(sk);
5822 				tcp_fast_path_on(tp);
5823 			} else {
5824 				return 1;
5825 			}
5826 			break;
5827 
5828 		case TCP_FIN_WAIT1:
5829 			if (tp->snd_una == tp->write_seq) {
5830 				tcp_set_state(sk, TCP_FIN_WAIT2);
5831 				sk->sk_shutdown |= SEND_SHUTDOWN;
5832 				dst_confirm(sk->sk_dst_cache);
5833 
5834 				if (!sock_flag(sk, SOCK_DEAD))
5835 					/* Wake up lingering close() */
5836 					sk->sk_state_change(sk);
5837 				else {
5838 					int tmo;
5839 
5840 					if (tp->linger2 < 0 ||
5841 					    (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5842 					     after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5843 						tcp_done(sk);
5844 						NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5845 						return 1;
5846 					}
5847 
5848 					tmo = tcp_fin_time(sk);
5849 					if (tmo > TCP_TIMEWAIT_LEN) {
5850 						inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5851 					} else if (th->fin || sock_owned_by_user(sk)) {
5852 						/* Bad case. We could lose such FIN otherwise.
5853 						 * It is not a big problem, but it looks confusing
5854 						 * and not so rare event. We still can lose it now,
5855 						 * if it spins in bh_lock_sock(), but it is really
5856 						 * marginal case.
5857 						 */
5858 						inet_csk_reset_keepalive_timer(sk, tmo);
5859 					} else {
5860 						tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5861 						goto discard;
5862 					}
5863 				}
5864 			}
5865 			break;
5866 
5867 		case TCP_CLOSING:
5868 			if (tp->snd_una == tp->write_seq) {
5869 				tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5870 				goto discard;
5871 			}
5872 			break;
5873 
5874 		case TCP_LAST_ACK:
5875 			if (tp->snd_una == tp->write_seq) {
5876 				tcp_update_metrics(sk);
5877 				tcp_done(sk);
5878 				goto discard;
5879 			}
5880 			break;
5881 		}
5882 	} else
5883 		goto discard;
5884 
5885 	/* step 6: check the URG bit */
5886 	tcp_urg(sk, skb, th);
5887 
5888 	/* step 7: process the segment text */
5889 	switch (sk->sk_state) {
5890 	case TCP_CLOSE_WAIT:
5891 	case TCP_CLOSING:
5892 	case TCP_LAST_ACK:
5893 		if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5894 			break;
5895 	case TCP_FIN_WAIT1:
5896 	case TCP_FIN_WAIT2:
5897 		/* RFC 793 says to queue data in these states,
5898 		 * RFC 1122 says we MUST send a reset.
5899 		 * BSD 4.4 also does reset.
5900 		 */
5901 		if (sk->sk_shutdown & RCV_SHUTDOWN) {
5902 			if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5903 			    after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5904 				NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5905 				tcp_reset(sk);
5906 				return 1;
5907 			}
5908 		}
5909 		/* Fall through */
5910 	case TCP_ESTABLISHED:
5911 		tcp_data_queue(sk, skb);
5912 		queued = 1;
5913 		break;
5914 	}
5915 
5916 	/* tcp_data could move socket to TIME-WAIT */
5917 	if (sk->sk_state != TCP_CLOSE) {
5918 		tcp_data_snd_check(sk);
5919 		tcp_ack_snd_check(sk);
5920 	}
5921 
5922 	if (!queued) {
5923 discard:
5924 		__kfree_skb(skb);
5925 	}
5926 	return 0;
5927 }
5928 
5929 EXPORT_SYMBOL(sysctl_tcp_ecn);
5930 EXPORT_SYMBOL(sysctl_tcp_reordering);
5931 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
5932 EXPORT_SYMBOL(tcp_parse_options);
5933 #ifdef CONFIG_TCP_MD5SIG
5934 EXPORT_SYMBOL(tcp_parse_md5sig_option);
5935 #endif
5936 EXPORT_SYMBOL(tcp_rcv_established);
5937 EXPORT_SYMBOL(tcp_rcv_state_process);
5938 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
5939