xref: /openbmc/linux/net/ipv4/tcp_input.c (revision ce932d0c5589e9766e089c22c66890dfc48fbd94)
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 #define pr_fmt(fmt) "TCP: " fmt
65 
66 #include <linux/mm.h>
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <net/dst.h>
72 #include <net/tcp.h>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
77 
78 int sysctl_tcp_timestamps __read_mostly = 1;
79 int sysctl_tcp_window_scaling __read_mostly = 1;
80 int sysctl_tcp_sack __read_mostly = 1;
81 int sysctl_tcp_fack __read_mostly = 1;
82 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
83 EXPORT_SYMBOL(sysctl_tcp_reordering);
84 int sysctl_tcp_ecn __read_mostly = 2;
85 EXPORT_SYMBOL(sysctl_tcp_ecn);
86 int sysctl_tcp_dsack __read_mostly = 1;
87 int sysctl_tcp_app_win __read_mostly = 31;
88 int sysctl_tcp_adv_win_scale __read_mostly = 1;
89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
90 
91 int sysctl_tcp_stdurg __read_mostly;
92 int sysctl_tcp_rfc1337 __read_mostly;
93 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
94 int sysctl_tcp_frto __read_mostly = 2;
95 int sysctl_tcp_frto_response __read_mostly;
96 int sysctl_tcp_nometrics_save __read_mostly;
97 
98 int sysctl_tcp_thin_dupack __read_mostly;
99 
100 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
101 int sysctl_tcp_abc __read_mostly;
102 
103 #define FLAG_DATA		0x01 /* Incoming frame contained data.		*/
104 #define FLAG_WIN_UPDATE		0x02 /* Incoming ACK was a window update.	*/
105 #define FLAG_DATA_ACKED		0x04 /* This ACK acknowledged new data.		*/
106 #define FLAG_RETRANS_DATA_ACKED	0x08 /* "" "" some of which was retransmitted.	*/
107 #define FLAG_SYN_ACKED		0x10 /* This ACK acknowledged SYN.		*/
108 #define FLAG_DATA_SACKED	0x20 /* New SACK.				*/
109 #define FLAG_ECE		0x40 /* ECE in this ACK				*/
110 #define FLAG_SLOWPATH		0x100 /* Do not skip RFC checks for window update.*/
111 #define FLAG_ONLY_ORIG_SACKED	0x200 /* SACKs only non-rexmit sent before RTO */
112 #define FLAG_SND_UNA_ADVANCED	0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
113 #define FLAG_DSACKING_ACK	0x800 /* SACK blocks contained D-SACK info */
114 #define FLAG_NONHEAD_RETRANS_ACKED	0x1000 /* Non-head rexmitted data was ACKed */
115 #define FLAG_SACK_RENEGING	0x2000 /* snd_una advanced to a sacked seq */
116 
117 #define FLAG_ACKED		(FLAG_DATA_ACKED|FLAG_SYN_ACKED)
118 #define FLAG_NOT_DUP		(FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
119 #define FLAG_CA_ALERT		(FLAG_DATA_SACKED|FLAG_ECE)
120 #define FLAG_FORWARD_PROGRESS	(FLAG_ACKED|FLAG_DATA_SACKED)
121 #define FLAG_ANY_PROGRESS	(FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
122 
123 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
124 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 
126 /* Adapt the MSS value used to make delayed ack decision to the
127  * real world.
128  */
129 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
130 {
131 	struct inet_connection_sock *icsk = inet_csk(sk);
132 	const unsigned int lss = icsk->icsk_ack.last_seg_size;
133 	unsigned int len;
134 
135 	icsk->icsk_ack.last_seg_size = 0;
136 
137 	/* skb->len may jitter because of SACKs, even if peer
138 	 * sends good full-sized frames.
139 	 */
140 	len = skb_shinfo(skb)->gso_size ? : skb->len;
141 	if (len >= icsk->icsk_ack.rcv_mss) {
142 		icsk->icsk_ack.rcv_mss = len;
143 	} else {
144 		/* Otherwise, we make more careful check taking into account,
145 		 * that SACKs block is variable.
146 		 *
147 		 * "len" is invariant segment length, including TCP header.
148 		 */
149 		len += skb->data - skb_transport_header(skb);
150 		if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
151 		    /* If PSH is not set, packet should be
152 		     * full sized, provided peer TCP is not badly broken.
153 		     * This observation (if it is correct 8)) allows
154 		     * to handle super-low mtu links fairly.
155 		     */
156 		    (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
157 		     !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
158 			/* Subtract also invariant (if peer is RFC compliant),
159 			 * tcp header plus fixed timestamp option length.
160 			 * Resulting "len" is MSS free of SACK jitter.
161 			 */
162 			len -= tcp_sk(sk)->tcp_header_len;
163 			icsk->icsk_ack.last_seg_size = len;
164 			if (len == lss) {
165 				icsk->icsk_ack.rcv_mss = len;
166 				return;
167 			}
168 		}
169 		if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
170 			icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
171 		icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
172 	}
173 }
174 
175 static void tcp_incr_quickack(struct sock *sk)
176 {
177 	struct inet_connection_sock *icsk = inet_csk(sk);
178 	unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
179 
180 	if (quickacks == 0)
181 		quickacks = 2;
182 	if (quickacks > icsk->icsk_ack.quick)
183 		icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
184 }
185 
186 static void tcp_enter_quickack_mode(struct sock *sk)
187 {
188 	struct inet_connection_sock *icsk = inet_csk(sk);
189 	tcp_incr_quickack(sk);
190 	icsk->icsk_ack.pingpong = 0;
191 	icsk->icsk_ack.ato = TCP_ATO_MIN;
192 }
193 
194 /* Send ACKs quickly, if "quick" count is not exhausted
195  * and the session is not interactive.
196  */
197 
198 static inline int tcp_in_quickack_mode(const struct sock *sk)
199 {
200 	const struct inet_connection_sock *icsk = inet_csk(sk);
201 	return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
202 }
203 
204 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
205 {
206 	if (tp->ecn_flags & TCP_ECN_OK)
207 		tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
208 }
209 
210 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
211 {
212 	if (tcp_hdr(skb)->cwr)
213 		tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
214 }
215 
216 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
217 {
218 	tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
219 }
220 
221 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
222 {
223 	if (!(tp->ecn_flags & TCP_ECN_OK))
224 		return;
225 
226 	switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
227 	case INET_ECN_NOT_ECT:
228 		/* Funny extension: if ECT is not set on a segment,
229 		 * and we already seen ECT on a previous segment,
230 		 * it is probably a retransmit.
231 		 */
232 		if (tp->ecn_flags & TCP_ECN_SEEN)
233 			tcp_enter_quickack_mode((struct sock *)tp);
234 		break;
235 	case INET_ECN_CE:
236 		tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
237 		/* fallinto */
238 	default:
239 		tp->ecn_flags |= TCP_ECN_SEEN;
240 	}
241 }
242 
243 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
244 {
245 	if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
246 		tp->ecn_flags &= ~TCP_ECN_OK;
247 }
248 
249 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
250 {
251 	if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
252 		tp->ecn_flags &= ~TCP_ECN_OK;
253 }
254 
255 static inline int TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
256 {
257 	if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
258 		return 1;
259 	return 0;
260 }
261 
262 /* Buffer size and advertised window tuning.
263  *
264  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
265  */
266 
267 static void tcp_fixup_sndbuf(struct sock *sk)
268 {
269 	int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
270 
271 	sndmem *= TCP_INIT_CWND;
272 	if (sk->sk_sndbuf < sndmem)
273 		sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
274 }
275 
276 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
277  *
278  * All tcp_full_space() is split to two parts: "network" buffer, allocated
279  * forward and advertised in receiver window (tp->rcv_wnd) and
280  * "application buffer", required to isolate scheduling/application
281  * latencies from network.
282  * window_clamp is maximal advertised window. It can be less than
283  * tcp_full_space(), in this case tcp_full_space() - window_clamp
284  * is reserved for "application" buffer. The less window_clamp is
285  * the smoother our behaviour from viewpoint of network, but the lower
286  * throughput and the higher sensitivity of the connection to losses. 8)
287  *
288  * rcv_ssthresh is more strict window_clamp used at "slow start"
289  * phase to predict further behaviour of this connection.
290  * It is used for two goals:
291  * - to enforce header prediction at sender, even when application
292  *   requires some significant "application buffer". It is check #1.
293  * - to prevent pruning of receive queue because of misprediction
294  *   of receiver window. Check #2.
295  *
296  * The scheme does not work when sender sends good segments opening
297  * window and then starts to feed us spaghetti. But it should work
298  * in common situations. Otherwise, we have to rely on queue collapsing.
299  */
300 
301 /* Slow part of check#2. */
302 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
303 {
304 	struct tcp_sock *tp = tcp_sk(sk);
305 	/* Optimize this! */
306 	int truesize = tcp_win_from_space(skb->truesize) >> 1;
307 	int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
308 
309 	while (tp->rcv_ssthresh <= window) {
310 		if (truesize <= skb->len)
311 			return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
312 
313 		truesize >>= 1;
314 		window >>= 1;
315 	}
316 	return 0;
317 }
318 
319 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
320 {
321 	struct tcp_sock *tp = tcp_sk(sk);
322 
323 	/* Check #1 */
324 	if (tp->rcv_ssthresh < tp->window_clamp &&
325 	    (int)tp->rcv_ssthresh < tcp_space(sk) &&
326 	    !sk_under_memory_pressure(sk)) {
327 		int incr;
328 
329 		/* Check #2. Increase window, if skb with such overhead
330 		 * will fit to rcvbuf in future.
331 		 */
332 		if (tcp_win_from_space(skb->truesize) <= skb->len)
333 			incr = 2 * tp->advmss;
334 		else
335 			incr = __tcp_grow_window(sk, skb);
336 
337 		if (incr) {
338 			incr = max_t(int, incr, 2 * skb->len);
339 			tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
340 					       tp->window_clamp);
341 			inet_csk(sk)->icsk_ack.quick |= 1;
342 		}
343 	}
344 }
345 
346 /* 3. Tuning rcvbuf, when connection enters established state. */
347 
348 static void tcp_fixup_rcvbuf(struct sock *sk)
349 {
350 	u32 mss = tcp_sk(sk)->advmss;
351 	u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
352 	int rcvmem;
353 
354 	/* Limit to 10 segments if mss <= 1460,
355 	 * or 14600/mss segments, with a minimum of two segments.
356 	 */
357 	if (mss > 1460)
358 		icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
359 
360 	rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
361 	while (tcp_win_from_space(rcvmem) < mss)
362 		rcvmem += 128;
363 
364 	rcvmem *= icwnd;
365 
366 	if (sk->sk_rcvbuf < rcvmem)
367 		sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
368 }
369 
370 /* 4. Try to fixup all. It is made immediately after connection enters
371  *    established state.
372  */
373 static void tcp_init_buffer_space(struct sock *sk)
374 {
375 	struct tcp_sock *tp = tcp_sk(sk);
376 	int maxwin;
377 
378 	if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
379 		tcp_fixup_rcvbuf(sk);
380 	if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
381 		tcp_fixup_sndbuf(sk);
382 
383 	tp->rcvq_space.space = tp->rcv_wnd;
384 
385 	maxwin = tcp_full_space(sk);
386 
387 	if (tp->window_clamp >= maxwin) {
388 		tp->window_clamp = maxwin;
389 
390 		if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
391 			tp->window_clamp = max(maxwin -
392 					       (maxwin >> sysctl_tcp_app_win),
393 					       4 * tp->advmss);
394 	}
395 
396 	/* Force reservation of one segment. */
397 	if (sysctl_tcp_app_win &&
398 	    tp->window_clamp > 2 * tp->advmss &&
399 	    tp->window_clamp + tp->advmss > maxwin)
400 		tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
401 
402 	tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
403 	tp->snd_cwnd_stamp = tcp_time_stamp;
404 }
405 
406 /* 5. Recalculate window clamp after socket hit its memory bounds. */
407 static void tcp_clamp_window(struct sock *sk)
408 {
409 	struct tcp_sock *tp = tcp_sk(sk);
410 	struct inet_connection_sock *icsk = inet_csk(sk);
411 
412 	icsk->icsk_ack.quick = 0;
413 
414 	if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
415 	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
416 	    !sk_under_memory_pressure(sk) &&
417 	    sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
418 		sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
419 				    sysctl_tcp_rmem[2]);
420 	}
421 	if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
422 		tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
423 }
424 
425 /* Initialize RCV_MSS value.
426  * RCV_MSS is an our guess about MSS used by the peer.
427  * We haven't any direct information about the MSS.
428  * It's better to underestimate the RCV_MSS rather than overestimate.
429  * Overestimations make us ACKing less frequently than needed.
430  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
431  */
432 void tcp_initialize_rcv_mss(struct sock *sk)
433 {
434 	const struct tcp_sock *tp = tcp_sk(sk);
435 	unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
436 
437 	hint = min(hint, tp->rcv_wnd / 2);
438 	hint = min(hint, TCP_MSS_DEFAULT);
439 	hint = max(hint, TCP_MIN_MSS);
440 
441 	inet_csk(sk)->icsk_ack.rcv_mss = hint;
442 }
443 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
444 
445 /* Receiver "autotuning" code.
446  *
447  * The algorithm for RTT estimation w/o timestamps is based on
448  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
449  * <http://public.lanl.gov/radiant/pubs.html#DRS>
450  *
451  * More detail on this code can be found at
452  * <http://staff.psc.edu/jheffner/>,
453  * though this reference is out of date.  A new paper
454  * is pending.
455  */
456 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
457 {
458 	u32 new_sample = tp->rcv_rtt_est.rtt;
459 	long m = sample;
460 
461 	if (m == 0)
462 		m = 1;
463 
464 	if (new_sample != 0) {
465 		/* If we sample in larger samples in the non-timestamp
466 		 * case, we could grossly overestimate the RTT especially
467 		 * with chatty applications or bulk transfer apps which
468 		 * are stalled on filesystem I/O.
469 		 *
470 		 * Also, since we are only going for a minimum in the
471 		 * non-timestamp case, we do not smooth things out
472 		 * else with timestamps disabled convergence takes too
473 		 * long.
474 		 */
475 		if (!win_dep) {
476 			m -= (new_sample >> 3);
477 			new_sample += m;
478 		} else {
479 			m <<= 3;
480 			if (m < new_sample)
481 				new_sample = m;
482 		}
483 	} else {
484 		/* No previous measure. */
485 		new_sample = m << 3;
486 	}
487 
488 	if (tp->rcv_rtt_est.rtt != new_sample)
489 		tp->rcv_rtt_est.rtt = new_sample;
490 }
491 
492 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
493 {
494 	if (tp->rcv_rtt_est.time == 0)
495 		goto new_measure;
496 	if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
497 		return;
498 	tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
499 
500 new_measure:
501 	tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
502 	tp->rcv_rtt_est.time = tcp_time_stamp;
503 }
504 
505 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
506 					  const struct sk_buff *skb)
507 {
508 	struct tcp_sock *tp = tcp_sk(sk);
509 	if (tp->rx_opt.rcv_tsecr &&
510 	    (TCP_SKB_CB(skb)->end_seq -
511 	     TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
512 		tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
513 }
514 
515 /*
516  * This function should be called every time data is copied to user space.
517  * It calculates the appropriate TCP receive buffer space.
518  */
519 void tcp_rcv_space_adjust(struct sock *sk)
520 {
521 	struct tcp_sock *tp = tcp_sk(sk);
522 	int time;
523 	int space;
524 
525 	if (tp->rcvq_space.time == 0)
526 		goto new_measure;
527 
528 	time = tcp_time_stamp - tp->rcvq_space.time;
529 	if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
530 		return;
531 
532 	space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
533 
534 	space = max(tp->rcvq_space.space, space);
535 
536 	if (tp->rcvq_space.space != space) {
537 		int rcvmem;
538 
539 		tp->rcvq_space.space = space;
540 
541 		if (sysctl_tcp_moderate_rcvbuf &&
542 		    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
543 			int new_clamp = space;
544 
545 			/* Receive space grows, normalize in order to
546 			 * take into account packet headers and sk_buff
547 			 * structure overhead.
548 			 */
549 			space /= tp->advmss;
550 			if (!space)
551 				space = 1;
552 			rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
553 			while (tcp_win_from_space(rcvmem) < tp->advmss)
554 				rcvmem += 128;
555 			space *= rcvmem;
556 			space = min(space, sysctl_tcp_rmem[2]);
557 			if (space > sk->sk_rcvbuf) {
558 				sk->sk_rcvbuf = space;
559 
560 				/* Make the window clamp follow along.  */
561 				tp->window_clamp = new_clamp;
562 			}
563 		}
564 	}
565 
566 new_measure:
567 	tp->rcvq_space.seq = tp->copied_seq;
568 	tp->rcvq_space.time = tcp_time_stamp;
569 }
570 
571 /* There is something which you must keep in mind when you analyze the
572  * behavior of the tp->ato delayed ack timeout interval.  When a
573  * connection starts up, we want to ack as quickly as possible.  The
574  * problem is that "good" TCP's do slow start at the beginning of data
575  * transmission.  The means that until we send the first few ACK's the
576  * sender will sit on his end and only queue most of his data, because
577  * he can only send snd_cwnd unacked packets at any given time.  For
578  * each ACK we send, he increments snd_cwnd and transmits more of his
579  * queue.  -DaveM
580  */
581 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
582 {
583 	struct tcp_sock *tp = tcp_sk(sk);
584 	struct inet_connection_sock *icsk = inet_csk(sk);
585 	u32 now;
586 
587 	inet_csk_schedule_ack(sk);
588 
589 	tcp_measure_rcv_mss(sk, skb);
590 
591 	tcp_rcv_rtt_measure(tp);
592 
593 	now = tcp_time_stamp;
594 
595 	if (!icsk->icsk_ack.ato) {
596 		/* The _first_ data packet received, initialize
597 		 * delayed ACK engine.
598 		 */
599 		tcp_incr_quickack(sk);
600 		icsk->icsk_ack.ato = TCP_ATO_MIN;
601 	} else {
602 		int m = now - icsk->icsk_ack.lrcvtime;
603 
604 		if (m <= TCP_ATO_MIN / 2) {
605 			/* The fastest case is the first. */
606 			icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
607 		} else if (m < icsk->icsk_ack.ato) {
608 			icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
609 			if (icsk->icsk_ack.ato > icsk->icsk_rto)
610 				icsk->icsk_ack.ato = icsk->icsk_rto;
611 		} else if (m > icsk->icsk_rto) {
612 			/* Too long gap. Apparently sender failed to
613 			 * restart window, so that we send ACKs quickly.
614 			 */
615 			tcp_incr_quickack(sk);
616 			sk_mem_reclaim(sk);
617 		}
618 	}
619 	icsk->icsk_ack.lrcvtime = now;
620 
621 	TCP_ECN_check_ce(tp, skb);
622 
623 	if (skb->len >= 128)
624 		tcp_grow_window(sk, skb);
625 }
626 
627 /* Called to compute a smoothed rtt estimate. The data fed to this
628  * routine either comes from timestamps, or from segments that were
629  * known _not_ to have been retransmitted [see Karn/Partridge
630  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
631  * piece by Van Jacobson.
632  * NOTE: the next three routines used to be one big routine.
633  * To save cycles in the RFC 1323 implementation it was better to break
634  * it up into three procedures. -- erics
635  */
636 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
637 {
638 	struct tcp_sock *tp = tcp_sk(sk);
639 	long m = mrtt; /* RTT */
640 
641 	/*	The following amusing code comes from Jacobson's
642 	 *	article in SIGCOMM '88.  Note that rtt and mdev
643 	 *	are scaled versions of rtt and mean deviation.
644 	 *	This is designed to be as fast as possible
645 	 *	m stands for "measurement".
646 	 *
647 	 *	On a 1990 paper the rto value is changed to:
648 	 *	RTO = rtt + 4 * mdev
649 	 *
650 	 * Funny. This algorithm seems to be very broken.
651 	 * These formulae increase RTO, when it should be decreased, increase
652 	 * too slowly, when it should be increased quickly, decrease too quickly
653 	 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
654 	 * does not matter how to _calculate_ it. Seems, it was trap
655 	 * that VJ failed to avoid. 8)
656 	 */
657 	if (m == 0)
658 		m = 1;
659 	if (tp->srtt != 0) {
660 		m -= (tp->srtt >> 3);	/* m is now error in rtt est */
661 		tp->srtt += m;		/* rtt = 7/8 rtt + 1/8 new */
662 		if (m < 0) {
663 			m = -m;		/* m is now abs(error) */
664 			m -= (tp->mdev >> 2);   /* similar update on mdev */
665 			/* This is similar to one of Eifel findings.
666 			 * Eifel blocks mdev updates when rtt decreases.
667 			 * This solution is a bit different: we use finer gain
668 			 * for mdev in this case (alpha*beta).
669 			 * Like Eifel it also prevents growth of rto,
670 			 * but also it limits too fast rto decreases,
671 			 * happening in pure Eifel.
672 			 */
673 			if (m > 0)
674 				m >>= 3;
675 		} else {
676 			m -= (tp->mdev >> 2);   /* similar update on mdev */
677 		}
678 		tp->mdev += m;	    	/* mdev = 3/4 mdev + 1/4 new */
679 		if (tp->mdev > tp->mdev_max) {
680 			tp->mdev_max = tp->mdev;
681 			if (tp->mdev_max > tp->rttvar)
682 				tp->rttvar = tp->mdev_max;
683 		}
684 		if (after(tp->snd_una, tp->rtt_seq)) {
685 			if (tp->mdev_max < tp->rttvar)
686 				tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
687 			tp->rtt_seq = tp->snd_nxt;
688 			tp->mdev_max = tcp_rto_min(sk);
689 		}
690 	} else {
691 		/* no previous measure. */
692 		tp->srtt = m << 3;	/* take the measured time to be rtt */
693 		tp->mdev = m << 1;	/* make sure rto = 3*rtt */
694 		tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
695 		tp->rtt_seq = tp->snd_nxt;
696 	}
697 }
698 
699 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
700  * routine referred to above.
701  */
702 static inline void tcp_set_rto(struct sock *sk)
703 {
704 	const struct tcp_sock *tp = tcp_sk(sk);
705 	/* Old crap is replaced with new one. 8)
706 	 *
707 	 * More seriously:
708 	 * 1. If rtt variance happened to be less 50msec, it is hallucination.
709 	 *    It cannot be less due to utterly erratic ACK generation made
710 	 *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
711 	 *    to do with delayed acks, because at cwnd>2 true delack timeout
712 	 *    is invisible. Actually, Linux-2.4 also generates erratic
713 	 *    ACKs in some circumstances.
714 	 */
715 	inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
716 
717 	/* 2. Fixups made earlier cannot be right.
718 	 *    If we do not estimate RTO correctly without them,
719 	 *    all the algo is pure shit and should be replaced
720 	 *    with correct one. It is exactly, which we pretend to do.
721 	 */
722 
723 	/* NOTE: clamping at TCP_RTO_MIN is not required, current algo
724 	 * guarantees that rto is higher.
725 	 */
726 	tcp_bound_rto(sk);
727 }
728 
729 /* Save metrics learned by this TCP session.
730    This function is called only, when TCP finishes successfully
731    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
732  */
733 void tcp_update_metrics(struct sock *sk)
734 {
735 	struct tcp_sock *tp = tcp_sk(sk);
736 	struct dst_entry *dst = __sk_dst_get(sk);
737 
738 	if (sysctl_tcp_nometrics_save)
739 		return;
740 
741 	dst_confirm(dst);
742 
743 	if (dst && (dst->flags & DST_HOST)) {
744 		const struct inet_connection_sock *icsk = inet_csk(sk);
745 		int m;
746 		unsigned long rtt;
747 
748 		if (icsk->icsk_backoff || !tp->srtt) {
749 			/* This session failed to estimate rtt. Why?
750 			 * Probably, no packets returned in time.
751 			 * Reset our results.
752 			 */
753 			if (!(dst_metric_locked(dst, RTAX_RTT)))
754 				dst_metric_set(dst, RTAX_RTT, 0);
755 			return;
756 		}
757 
758 		rtt = dst_metric_rtt(dst, RTAX_RTT);
759 		m = rtt - tp->srtt;
760 
761 		/* If newly calculated rtt larger than stored one,
762 		 * store new one. Otherwise, use EWMA. Remember,
763 		 * rtt overestimation is always better than underestimation.
764 		 */
765 		if (!(dst_metric_locked(dst, RTAX_RTT))) {
766 			if (m <= 0)
767 				set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
768 			else
769 				set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
770 		}
771 
772 		if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
773 			unsigned long var;
774 			if (m < 0)
775 				m = -m;
776 
777 			/* Scale deviation to rttvar fixed point */
778 			m >>= 1;
779 			if (m < tp->mdev)
780 				m = tp->mdev;
781 
782 			var = dst_metric_rtt(dst, RTAX_RTTVAR);
783 			if (m >= var)
784 				var = m;
785 			else
786 				var -= (var - m) >> 2;
787 
788 			set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
789 		}
790 
791 		if (tcp_in_initial_slowstart(tp)) {
792 			/* Slow start still did not finish. */
793 			if (dst_metric(dst, RTAX_SSTHRESH) &&
794 			    !dst_metric_locked(dst, RTAX_SSTHRESH) &&
795 			    (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
796 				dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
797 			if (!dst_metric_locked(dst, RTAX_CWND) &&
798 			    tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
799 				dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
800 		} else if (tp->snd_cwnd > tp->snd_ssthresh &&
801 			   icsk->icsk_ca_state == TCP_CA_Open) {
802 			/* Cong. avoidance phase, cwnd is reliable. */
803 			if (!dst_metric_locked(dst, RTAX_SSTHRESH))
804 				dst_metric_set(dst, RTAX_SSTHRESH,
805 					       max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
806 			if (!dst_metric_locked(dst, RTAX_CWND))
807 				dst_metric_set(dst, RTAX_CWND,
808 					       (dst_metric(dst, RTAX_CWND) +
809 						tp->snd_cwnd) >> 1);
810 		} else {
811 			/* Else slow start did not finish, cwnd is non-sense,
812 			   ssthresh may be also invalid.
813 			 */
814 			if (!dst_metric_locked(dst, RTAX_CWND))
815 				dst_metric_set(dst, RTAX_CWND,
816 					       (dst_metric(dst, RTAX_CWND) +
817 						tp->snd_ssthresh) >> 1);
818 			if (dst_metric(dst, RTAX_SSTHRESH) &&
819 			    !dst_metric_locked(dst, RTAX_SSTHRESH) &&
820 			    tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
821 				dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
822 		}
823 
824 		if (!dst_metric_locked(dst, RTAX_REORDERING)) {
825 			if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
826 			    tp->reordering != sysctl_tcp_reordering)
827 				dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
828 		}
829 	}
830 }
831 
832 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
833 {
834 	__u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
835 
836 	if (!cwnd)
837 		cwnd = TCP_INIT_CWND;
838 	return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
839 }
840 
841 /* Set slow start threshold and cwnd not falling to slow start */
842 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
843 {
844 	struct tcp_sock *tp = tcp_sk(sk);
845 	const struct inet_connection_sock *icsk = inet_csk(sk);
846 
847 	tp->prior_ssthresh = 0;
848 	tp->bytes_acked = 0;
849 	if (icsk->icsk_ca_state < TCP_CA_CWR) {
850 		tp->undo_marker = 0;
851 		if (set_ssthresh)
852 			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
853 		tp->snd_cwnd = min(tp->snd_cwnd,
854 				   tcp_packets_in_flight(tp) + 1U);
855 		tp->snd_cwnd_cnt = 0;
856 		tp->high_seq = tp->snd_nxt;
857 		tp->snd_cwnd_stamp = tcp_time_stamp;
858 		TCP_ECN_queue_cwr(tp);
859 
860 		tcp_set_ca_state(sk, TCP_CA_CWR);
861 	}
862 }
863 
864 /*
865  * Packet counting of FACK is based on in-order assumptions, therefore TCP
866  * disables it when reordering is detected
867  */
868 static void tcp_disable_fack(struct tcp_sock *tp)
869 {
870 	/* RFC3517 uses different metric in lost marker => reset on change */
871 	if (tcp_is_fack(tp))
872 		tp->lost_skb_hint = NULL;
873 	tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
874 }
875 
876 /* Take a notice that peer is sending D-SACKs */
877 static void tcp_dsack_seen(struct tcp_sock *tp)
878 {
879 	tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
880 }
881 
882 /* Initialize metrics on socket. */
883 
884 static void tcp_init_metrics(struct sock *sk)
885 {
886 	struct tcp_sock *tp = tcp_sk(sk);
887 	struct dst_entry *dst = __sk_dst_get(sk);
888 
889 	if (dst == NULL)
890 		goto reset;
891 
892 	dst_confirm(dst);
893 
894 	if (dst_metric_locked(dst, RTAX_CWND))
895 		tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
896 	if (dst_metric(dst, RTAX_SSTHRESH)) {
897 		tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
898 		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
899 			tp->snd_ssthresh = tp->snd_cwnd_clamp;
900 	} else {
901 		/* ssthresh may have been reduced unnecessarily during.
902 		 * 3WHS. Restore it back to its initial default.
903 		 */
904 		tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
905 	}
906 	if (dst_metric(dst, RTAX_REORDERING) &&
907 	    tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
908 		tcp_disable_fack(tp);
909 		tp->reordering = dst_metric(dst, RTAX_REORDERING);
910 	}
911 
912 	if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
913 		goto reset;
914 
915 	/* Initial rtt is determined from SYN,SYN-ACK.
916 	 * The segment is small and rtt may appear much
917 	 * less than real one. Use per-dst memory
918 	 * to make it more realistic.
919 	 *
920 	 * A bit of theory. RTT is time passed after "normal" sized packet
921 	 * is sent until it is ACKed. In normal circumstances sending small
922 	 * packets force peer to delay ACKs and calculation is correct too.
923 	 * The algorithm is adaptive and, provided we follow specs, it
924 	 * NEVER underestimate RTT. BUT! If peer tries to make some clever
925 	 * tricks sort of "quick acks" for time long enough to decrease RTT
926 	 * to low value, and then abruptly stops to do it and starts to delay
927 	 * ACKs, wait for troubles.
928 	 */
929 	if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
930 		tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
931 		tp->rtt_seq = tp->snd_nxt;
932 	}
933 	if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
934 		tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
935 		tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
936 	}
937 	tcp_set_rto(sk);
938 reset:
939 	if (tp->srtt == 0) {
940 		/* RFC2988bis: We've failed to get a valid RTT sample from
941 		 * 3WHS. This is most likely due to retransmission,
942 		 * including spurious one. Reset the RTO back to 3secs
943 		 * from the more aggressive 1sec to avoid more spurious
944 		 * retransmission.
945 		 */
946 		tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
947 		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
948 	}
949 	/* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
950 	 * retransmitted. In light of RFC2988bis' more aggressive 1sec
951 	 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
952 	 * retransmission has occurred.
953 	 */
954 	if (tp->total_retrans > 1)
955 		tp->snd_cwnd = 1;
956 	else
957 		tp->snd_cwnd = tcp_init_cwnd(tp, dst);
958 	tp->snd_cwnd_stamp = tcp_time_stamp;
959 }
960 
961 static void tcp_update_reordering(struct sock *sk, const int metric,
962 				  const int ts)
963 {
964 	struct tcp_sock *tp = tcp_sk(sk);
965 	if (metric > tp->reordering) {
966 		int mib_idx;
967 
968 		tp->reordering = min(TCP_MAX_REORDERING, metric);
969 
970 		/* This exciting event is worth to be remembered. 8) */
971 		if (ts)
972 			mib_idx = LINUX_MIB_TCPTSREORDER;
973 		else if (tcp_is_reno(tp))
974 			mib_idx = LINUX_MIB_TCPRENOREORDER;
975 		else if (tcp_is_fack(tp))
976 			mib_idx = LINUX_MIB_TCPFACKREORDER;
977 		else
978 			mib_idx = LINUX_MIB_TCPSACKREORDER;
979 
980 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
981 #if FASTRETRANS_DEBUG > 1
982 		printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
983 		       tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
984 		       tp->reordering,
985 		       tp->fackets_out,
986 		       tp->sacked_out,
987 		       tp->undo_marker ? tp->undo_retrans : 0);
988 #endif
989 		tcp_disable_fack(tp);
990 	}
991 }
992 
993 /* This must be called before lost_out is incremented */
994 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
995 {
996 	if ((tp->retransmit_skb_hint == NULL) ||
997 	    before(TCP_SKB_CB(skb)->seq,
998 		   TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
999 		tp->retransmit_skb_hint = skb;
1000 
1001 	if (!tp->lost_out ||
1002 	    after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
1003 		tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1004 }
1005 
1006 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
1007 {
1008 	if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1009 		tcp_verify_retransmit_hint(tp, skb);
1010 
1011 		tp->lost_out += tcp_skb_pcount(skb);
1012 		TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1013 	}
1014 }
1015 
1016 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1017 					    struct sk_buff *skb)
1018 {
1019 	tcp_verify_retransmit_hint(tp, skb);
1020 
1021 	if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1022 		tp->lost_out += tcp_skb_pcount(skb);
1023 		TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1024 	}
1025 }
1026 
1027 /* This procedure tags the retransmission queue when SACKs arrive.
1028  *
1029  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1030  * Packets in queue with these bits set are counted in variables
1031  * sacked_out, retrans_out and lost_out, correspondingly.
1032  *
1033  * Valid combinations are:
1034  * Tag  InFlight	Description
1035  * 0	1		- orig segment is in flight.
1036  * S	0		- nothing flies, orig reached receiver.
1037  * L	0		- nothing flies, orig lost by net.
1038  * R	2		- both orig and retransmit are in flight.
1039  * L|R	1		- orig is lost, retransmit is in flight.
1040  * S|R  1		- orig reached receiver, retrans is still in flight.
1041  * (L|S|R is logically valid, it could occur when L|R is sacked,
1042  *  but it is equivalent to plain S and code short-curcuits it to S.
1043  *  L|S is logically invalid, it would mean -1 packet in flight 8))
1044  *
1045  * These 6 states form finite state machine, controlled by the following events:
1046  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1047  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1048  * 3. Loss detection event of two flavors:
1049  *	A. Scoreboard estimator decided the packet is lost.
1050  *	   A'. Reno "three dupacks" marks head of queue lost.
1051  *	   A''. Its FACK modification, head until snd.fack is lost.
1052  *	B. SACK arrives sacking SND.NXT at the moment, when the
1053  *	   segment was retransmitted.
1054  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1055  *
1056  * It is pleasant to note, that state diagram turns out to be commutative,
1057  * so that we are allowed not to be bothered by order of our actions,
1058  * when multiple events arrive simultaneously. (see the function below).
1059  *
1060  * Reordering detection.
1061  * --------------------
1062  * Reordering metric is maximal distance, which a packet can be displaced
1063  * in packet stream. With SACKs we can estimate it:
1064  *
1065  * 1. SACK fills old hole and the corresponding segment was not
1066  *    ever retransmitted -> reordering. Alas, we cannot use it
1067  *    when segment was retransmitted.
1068  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1069  *    for retransmitted and already SACKed segment -> reordering..
1070  * Both of these heuristics are not used in Loss state, when we cannot
1071  * account for retransmits accurately.
1072  *
1073  * SACK block validation.
1074  * ----------------------
1075  *
1076  * SACK block range validation checks that the received SACK block fits to
1077  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1078  * Note that SND.UNA is not included to the range though being valid because
1079  * it means that the receiver is rather inconsistent with itself reporting
1080  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1081  * perfectly valid, however, in light of RFC2018 which explicitly states
1082  * that "SACK block MUST reflect the newest segment.  Even if the newest
1083  * segment is going to be discarded ...", not that it looks very clever
1084  * in case of head skb. Due to potentional receiver driven attacks, we
1085  * choose to avoid immediate execution of a walk in write queue due to
1086  * reneging and defer head skb's loss recovery to standard loss recovery
1087  * procedure that will eventually trigger (nothing forbids us doing this).
1088  *
1089  * Implements also blockage to start_seq wrap-around. Problem lies in the
1090  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1091  * there's no guarantee that it will be before snd_nxt (n). The problem
1092  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1093  * wrap (s_w):
1094  *
1095  *         <- outs wnd ->                          <- wrapzone ->
1096  *         u     e      n                         u_w   e_w  s n_w
1097  *         |     |      |                          |     |   |  |
1098  * |<------------+------+----- TCP seqno space --------------+---------->|
1099  * ...-- <2^31 ->|                                           |<--------...
1100  * ...---- >2^31 ------>|                                    |<--------...
1101  *
1102  * Current code wouldn't be vulnerable but it's better still to discard such
1103  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1104  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1105  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1106  * equal to the ideal case (infinite seqno space without wrap caused issues).
1107  *
1108  * With D-SACK the lower bound is extended to cover sequence space below
1109  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1110  * again, D-SACK block must not to go across snd_una (for the same reason as
1111  * for the normal SACK blocks, explained above). But there all simplicity
1112  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1113  * fully below undo_marker they do not affect behavior in anyway and can
1114  * therefore be safely ignored. In rare cases (which are more or less
1115  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1116  * fragmentation and packet reordering past skb's retransmission. To consider
1117  * them correctly, the acceptable range must be extended even more though
1118  * the exact amount is rather hard to quantify. However, tp->max_window can
1119  * be used as an exaggerated estimate.
1120  */
1121 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1122 				  u32 start_seq, u32 end_seq)
1123 {
1124 	/* Too far in future, or reversed (interpretation is ambiguous) */
1125 	if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1126 		return 0;
1127 
1128 	/* Nasty start_seq wrap-around check (see comments above) */
1129 	if (!before(start_seq, tp->snd_nxt))
1130 		return 0;
1131 
1132 	/* In outstanding window? ...This is valid exit for D-SACKs too.
1133 	 * start_seq == snd_una is non-sensical (see comments above)
1134 	 */
1135 	if (after(start_seq, tp->snd_una))
1136 		return 1;
1137 
1138 	if (!is_dsack || !tp->undo_marker)
1139 		return 0;
1140 
1141 	/* ...Then it's D-SACK, and must reside below snd_una completely */
1142 	if (after(end_seq, tp->snd_una))
1143 		return 0;
1144 
1145 	if (!before(start_seq, tp->undo_marker))
1146 		return 1;
1147 
1148 	/* Too old */
1149 	if (!after(end_seq, tp->undo_marker))
1150 		return 0;
1151 
1152 	/* Undo_marker boundary crossing (overestimates a lot). Known already:
1153 	 *   start_seq < undo_marker and end_seq >= undo_marker.
1154 	 */
1155 	return !before(start_seq, end_seq - tp->max_window);
1156 }
1157 
1158 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1159  * Event "B". Later note: FACK people cheated me again 8), we have to account
1160  * for reordering! Ugly, but should help.
1161  *
1162  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1163  * less than what is now known to be received by the other end (derived from
1164  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1165  * retransmitted skbs to avoid some costly processing per ACKs.
1166  */
1167 static void tcp_mark_lost_retrans(struct sock *sk)
1168 {
1169 	const struct inet_connection_sock *icsk = inet_csk(sk);
1170 	struct tcp_sock *tp = tcp_sk(sk);
1171 	struct sk_buff *skb;
1172 	int cnt = 0;
1173 	u32 new_low_seq = tp->snd_nxt;
1174 	u32 received_upto = tcp_highest_sack_seq(tp);
1175 
1176 	if (!tcp_is_fack(tp) || !tp->retrans_out ||
1177 	    !after(received_upto, tp->lost_retrans_low) ||
1178 	    icsk->icsk_ca_state != TCP_CA_Recovery)
1179 		return;
1180 
1181 	tcp_for_write_queue(skb, sk) {
1182 		u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1183 
1184 		if (skb == tcp_send_head(sk))
1185 			break;
1186 		if (cnt == tp->retrans_out)
1187 			break;
1188 		if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1189 			continue;
1190 
1191 		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1192 			continue;
1193 
1194 		/* TODO: We would like to get rid of tcp_is_fack(tp) only
1195 		 * constraint here (see above) but figuring out that at
1196 		 * least tp->reordering SACK blocks reside between ack_seq
1197 		 * and received_upto is not easy task to do cheaply with
1198 		 * the available datastructures.
1199 		 *
1200 		 * Whether FACK should check here for tp->reordering segs
1201 		 * in-between one could argue for either way (it would be
1202 		 * rather simple to implement as we could count fack_count
1203 		 * during the walk and do tp->fackets_out - fack_count).
1204 		 */
1205 		if (after(received_upto, ack_seq)) {
1206 			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1207 			tp->retrans_out -= tcp_skb_pcount(skb);
1208 
1209 			tcp_skb_mark_lost_uncond_verify(tp, skb);
1210 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1211 		} else {
1212 			if (before(ack_seq, new_low_seq))
1213 				new_low_seq = ack_seq;
1214 			cnt += tcp_skb_pcount(skb);
1215 		}
1216 	}
1217 
1218 	if (tp->retrans_out)
1219 		tp->lost_retrans_low = new_low_seq;
1220 }
1221 
1222 static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1223 			   struct tcp_sack_block_wire *sp, int num_sacks,
1224 			   u32 prior_snd_una)
1225 {
1226 	struct tcp_sock *tp = tcp_sk(sk);
1227 	u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1228 	u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1229 	int dup_sack = 0;
1230 
1231 	if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1232 		dup_sack = 1;
1233 		tcp_dsack_seen(tp);
1234 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1235 	} else if (num_sacks > 1) {
1236 		u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1237 		u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1238 
1239 		if (!after(end_seq_0, end_seq_1) &&
1240 		    !before(start_seq_0, start_seq_1)) {
1241 			dup_sack = 1;
1242 			tcp_dsack_seen(tp);
1243 			NET_INC_STATS_BH(sock_net(sk),
1244 					LINUX_MIB_TCPDSACKOFORECV);
1245 		}
1246 	}
1247 
1248 	/* D-SACK for already forgotten data... Do dumb counting. */
1249 	if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1250 	    !after(end_seq_0, prior_snd_una) &&
1251 	    after(end_seq_0, tp->undo_marker))
1252 		tp->undo_retrans--;
1253 
1254 	return dup_sack;
1255 }
1256 
1257 struct tcp_sacktag_state {
1258 	int reord;
1259 	int fack_count;
1260 	int flag;
1261 };
1262 
1263 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1264  * the incoming SACK may not exactly match but we can find smaller MSS
1265  * aligned portion of it that matches. Therefore we might need to fragment
1266  * which may fail and creates some hassle (caller must handle error case
1267  * returns).
1268  *
1269  * FIXME: this could be merged to shift decision code
1270  */
1271 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1272 				 u32 start_seq, u32 end_seq)
1273 {
1274 	int in_sack, err;
1275 	unsigned int pkt_len;
1276 	unsigned int mss;
1277 
1278 	in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1279 		  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1280 
1281 	if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1282 	    after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1283 		mss = tcp_skb_mss(skb);
1284 		in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1285 
1286 		if (!in_sack) {
1287 			pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1288 			if (pkt_len < mss)
1289 				pkt_len = mss;
1290 		} else {
1291 			pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1292 			if (pkt_len < mss)
1293 				return -EINVAL;
1294 		}
1295 
1296 		/* Round if necessary so that SACKs cover only full MSSes
1297 		 * and/or the remaining small portion (if present)
1298 		 */
1299 		if (pkt_len > mss) {
1300 			unsigned int new_len = (pkt_len / mss) * mss;
1301 			if (!in_sack && new_len < pkt_len) {
1302 				new_len += mss;
1303 				if (new_len > skb->len)
1304 					return 0;
1305 			}
1306 			pkt_len = new_len;
1307 		}
1308 		err = tcp_fragment(sk, skb, pkt_len, mss);
1309 		if (err < 0)
1310 			return err;
1311 	}
1312 
1313 	return in_sack;
1314 }
1315 
1316 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1317 static u8 tcp_sacktag_one(struct sock *sk,
1318 			  struct tcp_sacktag_state *state, u8 sacked,
1319 			  u32 start_seq, u32 end_seq,
1320 			  int dup_sack, int pcount)
1321 {
1322 	struct tcp_sock *tp = tcp_sk(sk);
1323 	int fack_count = state->fack_count;
1324 
1325 	/* Account D-SACK for retransmitted packet. */
1326 	if (dup_sack && (sacked & TCPCB_RETRANS)) {
1327 		if (tp->undo_marker && tp->undo_retrans &&
1328 		    after(end_seq, tp->undo_marker))
1329 			tp->undo_retrans--;
1330 		if (sacked & TCPCB_SACKED_ACKED)
1331 			state->reord = min(fack_count, state->reord);
1332 	}
1333 
1334 	/* Nothing to do; acked frame is about to be dropped (was ACKed). */
1335 	if (!after(end_seq, tp->snd_una))
1336 		return sacked;
1337 
1338 	if (!(sacked & TCPCB_SACKED_ACKED)) {
1339 		if (sacked & TCPCB_SACKED_RETRANS) {
1340 			/* If the segment is not tagged as lost,
1341 			 * we do not clear RETRANS, believing
1342 			 * that retransmission is still in flight.
1343 			 */
1344 			if (sacked & TCPCB_LOST) {
1345 				sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1346 				tp->lost_out -= pcount;
1347 				tp->retrans_out -= pcount;
1348 			}
1349 		} else {
1350 			if (!(sacked & TCPCB_RETRANS)) {
1351 				/* New sack for not retransmitted frame,
1352 				 * which was in hole. It is reordering.
1353 				 */
1354 				if (before(start_seq,
1355 					   tcp_highest_sack_seq(tp)))
1356 					state->reord = min(fack_count,
1357 							   state->reord);
1358 
1359 				/* SACK enhanced F-RTO (RFC4138; Appendix B) */
1360 				if (!after(end_seq, tp->frto_highmark))
1361 					state->flag |= FLAG_ONLY_ORIG_SACKED;
1362 			}
1363 
1364 			if (sacked & TCPCB_LOST) {
1365 				sacked &= ~TCPCB_LOST;
1366 				tp->lost_out -= pcount;
1367 			}
1368 		}
1369 
1370 		sacked |= TCPCB_SACKED_ACKED;
1371 		state->flag |= FLAG_DATA_SACKED;
1372 		tp->sacked_out += pcount;
1373 
1374 		fack_count += pcount;
1375 
1376 		/* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1377 		if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1378 		    before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1379 			tp->lost_cnt_hint += pcount;
1380 
1381 		if (fack_count > tp->fackets_out)
1382 			tp->fackets_out = fack_count;
1383 	}
1384 
1385 	/* D-SACK. We can detect redundant retransmission in S|R and plain R
1386 	 * frames and clear it. undo_retrans is decreased above, L|R frames
1387 	 * are accounted above as well.
1388 	 */
1389 	if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1390 		sacked &= ~TCPCB_SACKED_RETRANS;
1391 		tp->retrans_out -= pcount;
1392 	}
1393 
1394 	return sacked;
1395 }
1396 
1397 /* Shift newly-SACKed bytes from this skb to the immediately previous
1398  * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1399  */
1400 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1401 			   struct tcp_sacktag_state *state,
1402 			   unsigned int pcount, int shifted, int mss,
1403 			   int dup_sack)
1404 {
1405 	struct tcp_sock *tp = tcp_sk(sk);
1406 	struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1407 	u32 start_seq = TCP_SKB_CB(skb)->seq;	/* start of newly-SACKed */
1408 	u32 end_seq = start_seq + shifted;	/* end of newly-SACKed */
1409 
1410 	BUG_ON(!pcount);
1411 
1412 	/* Adjust counters and hints for the newly sacked sequence
1413 	 * range but discard the return value since prev is already
1414 	 * marked. We must tag the range first because the seq
1415 	 * advancement below implicitly advances
1416 	 * tcp_highest_sack_seq() when skb is highest_sack.
1417 	 */
1418 	tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1419 			start_seq, end_seq, dup_sack, pcount);
1420 
1421 	if (skb == tp->lost_skb_hint)
1422 		tp->lost_cnt_hint += pcount;
1423 
1424 	TCP_SKB_CB(prev)->end_seq += shifted;
1425 	TCP_SKB_CB(skb)->seq += shifted;
1426 
1427 	skb_shinfo(prev)->gso_segs += pcount;
1428 	BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1429 	skb_shinfo(skb)->gso_segs -= pcount;
1430 
1431 	/* When we're adding to gso_segs == 1, gso_size will be zero,
1432 	 * in theory this shouldn't be necessary but as long as DSACK
1433 	 * code can come after this skb later on it's better to keep
1434 	 * setting gso_size to something.
1435 	 */
1436 	if (!skb_shinfo(prev)->gso_size) {
1437 		skb_shinfo(prev)->gso_size = mss;
1438 		skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1439 	}
1440 
1441 	/* CHECKME: To clear or not to clear? Mimics normal skb currently */
1442 	if (skb_shinfo(skb)->gso_segs <= 1) {
1443 		skb_shinfo(skb)->gso_size = 0;
1444 		skb_shinfo(skb)->gso_type = 0;
1445 	}
1446 
1447 	/* Difference in this won't matter, both ACKed by the same cumul. ACK */
1448 	TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1449 
1450 	if (skb->len > 0) {
1451 		BUG_ON(!tcp_skb_pcount(skb));
1452 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1453 		return 0;
1454 	}
1455 
1456 	/* Whole SKB was eaten :-) */
1457 
1458 	if (skb == tp->retransmit_skb_hint)
1459 		tp->retransmit_skb_hint = prev;
1460 	if (skb == tp->scoreboard_skb_hint)
1461 		tp->scoreboard_skb_hint = prev;
1462 	if (skb == tp->lost_skb_hint) {
1463 		tp->lost_skb_hint = prev;
1464 		tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1465 	}
1466 
1467 	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1468 	if (skb == tcp_highest_sack(sk))
1469 		tcp_advance_highest_sack(sk, skb);
1470 
1471 	tcp_unlink_write_queue(skb, sk);
1472 	sk_wmem_free_skb(sk, skb);
1473 
1474 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1475 
1476 	return 1;
1477 }
1478 
1479 /* I wish gso_size would have a bit more sane initialization than
1480  * something-or-zero which complicates things
1481  */
1482 static int tcp_skb_seglen(const struct sk_buff *skb)
1483 {
1484 	return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1485 }
1486 
1487 /* Shifting pages past head area doesn't work */
1488 static int skb_can_shift(const struct sk_buff *skb)
1489 {
1490 	return !skb_headlen(skb) && skb_is_nonlinear(skb);
1491 }
1492 
1493 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1494  * skb.
1495  */
1496 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1497 					  struct tcp_sacktag_state *state,
1498 					  u32 start_seq, u32 end_seq,
1499 					  int dup_sack)
1500 {
1501 	struct tcp_sock *tp = tcp_sk(sk);
1502 	struct sk_buff *prev;
1503 	int mss;
1504 	int pcount = 0;
1505 	int len;
1506 	int in_sack;
1507 
1508 	if (!sk_can_gso(sk))
1509 		goto fallback;
1510 
1511 	/* Normally R but no L won't result in plain S */
1512 	if (!dup_sack &&
1513 	    (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1514 		goto fallback;
1515 	if (!skb_can_shift(skb))
1516 		goto fallback;
1517 	/* This frame is about to be dropped (was ACKed). */
1518 	if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1519 		goto fallback;
1520 
1521 	/* Can only happen with delayed DSACK + discard craziness */
1522 	if (unlikely(skb == tcp_write_queue_head(sk)))
1523 		goto fallback;
1524 	prev = tcp_write_queue_prev(sk, skb);
1525 
1526 	if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1527 		goto fallback;
1528 
1529 	in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1530 		  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1531 
1532 	if (in_sack) {
1533 		len = skb->len;
1534 		pcount = tcp_skb_pcount(skb);
1535 		mss = tcp_skb_seglen(skb);
1536 
1537 		/* TODO: Fix DSACKs to not fragment already SACKed and we can
1538 		 * drop this restriction as unnecessary
1539 		 */
1540 		if (mss != tcp_skb_seglen(prev))
1541 			goto fallback;
1542 	} else {
1543 		if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1544 			goto noop;
1545 		/* CHECKME: This is non-MSS split case only?, this will
1546 		 * cause skipped skbs due to advancing loop btw, original
1547 		 * has that feature too
1548 		 */
1549 		if (tcp_skb_pcount(skb) <= 1)
1550 			goto noop;
1551 
1552 		in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1553 		if (!in_sack) {
1554 			/* TODO: head merge to next could be attempted here
1555 			 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1556 			 * though it might not be worth of the additional hassle
1557 			 *
1558 			 * ...we can probably just fallback to what was done
1559 			 * previously. We could try merging non-SACKed ones
1560 			 * as well but it probably isn't going to buy off
1561 			 * because later SACKs might again split them, and
1562 			 * it would make skb timestamp tracking considerably
1563 			 * harder problem.
1564 			 */
1565 			goto fallback;
1566 		}
1567 
1568 		len = end_seq - TCP_SKB_CB(skb)->seq;
1569 		BUG_ON(len < 0);
1570 		BUG_ON(len > skb->len);
1571 
1572 		/* MSS boundaries should be honoured or else pcount will
1573 		 * severely break even though it makes things bit trickier.
1574 		 * Optimize common case to avoid most of the divides
1575 		 */
1576 		mss = tcp_skb_mss(skb);
1577 
1578 		/* TODO: Fix DSACKs to not fragment already SACKed and we can
1579 		 * drop this restriction as unnecessary
1580 		 */
1581 		if (mss != tcp_skb_seglen(prev))
1582 			goto fallback;
1583 
1584 		if (len == mss) {
1585 			pcount = 1;
1586 		} else if (len < mss) {
1587 			goto noop;
1588 		} else {
1589 			pcount = len / mss;
1590 			len = pcount * mss;
1591 		}
1592 	}
1593 
1594 	/* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1595 	if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1596 		goto fallback;
1597 
1598 	if (!skb_shift(prev, skb, len))
1599 		goto fallback;
1600 	if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1601 		goto out;
1602 
1603 	/* Hole filled allows collapsing with the next as well, this is very
1604 	 * useful when hole on every nth skb pattern happens
1605 	 */
1606 	if (prev == tcp_write_queue_tail(sk))
1607 		goto out;
1608 	skb = tcp_write_queue_next(sk, prev);
1609 
1610 	if (!skb_can_shift(skb) ||
1611 	    (skb == tcp_send_head(sk)) ||
1612 	    ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1613 	    (mss != tcp_skb_seglen(skb)))
1614 		goto out;
1615 
1616 	len = skb->len;
1617 	if (skb_shift(prev, skb, len)) {
1618 		pcount += tcp_skb_pcount(skb);
1619 		tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1620 	}
1621 
1622 out:
1623 	state->fack_count += pcount;
1624 	return prev;
1625 
1626 noop:
1627 	return skb;
1628 
1629 fallback:
1630 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1631 	return NULL;
1632 }
1633 
1634 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1635 					struct tcp_sack_block *next_dup,
1636 					struct tcp_sacktag_state *state,
1637 					u32 start_seq, u32 end_seq,
1638 					int dup_sack_in)
1639 {
1640 	struct tcp_sock *tp = tcp_sk(sk);
1641 	struct sk_buff *tmp;
1642 
1643 	tcp_for_write_queue_from(skb, sk) {
1644 		int in_sack = 0;
1645 		int dup_sack = dup_sack_in;
1646 
1647 		if (skb == tcp_send_head(sk))
1648 			break;
1649 
1650 		/* queue is in-order => we can short-circuit the walk early */
1651 		if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1652 			break;
1653 
1654 		if ((next_dup != NULL) &&
1655 		    before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1656 			in_sack = tcp_match_skb_to_sack(sk, skb,
1657 							next_dup->start_seq,
1658 							next_dup->end_seq);
1659 			if (in_sack > 0)
1660 				dup_sack = 1;
1661 		}
1662 
1663 		/* skb reference here is a bit tricky to get right, since
1664 		 * shifting can eat and free both this skb and the next,
1665 		 * so not even _safe variant of the loop is enough.
1666 		 */
1667 		if (in_sack <= 0) {
1668 			tmp = tcp_shift_skb_data(sk, skb, state,
1669 						 start_seq, end_seq, dup_sack);
1670 			if (tmp != NULL) {
1671 				if (tmp != skb) {
1672 					skb = tmp;
1673 					continue;
1674 				}
1675 
1676 				in_sack = 0;
1677 			} else {
1678 				in_sack = tcp_match_skb_to_sack(sk, skb,
1679 								start_seq,
1680 								end_seq);
1681 			}
1682 		}
1683 
1684 		if (unlikely(in_sack < 0))
1685 			break;
1686 
1687 		if (in_sack) {
1688 			TCP_SKB_CB(skb)->sacked =
1689 				tcp_sacktag_one(sk,
1690 						state,
1691 						TCP_SKB_CB(skb)->sacked,
1692 						TCP_SKB_CB(skb)->seq,
1693 						TCP_SKB_CB(skb)->end_seq,
1694 						dup_sack,
1695 						tcp_skb_pcount(skb));
1696 
1697 			if (!before(TCP_SKB_CB(skb)->seq,
1698 				    tcp_highest_sack_seq(tp)))
1699 				tcp_advance_highest_sack(sk, skb);
1700 		}
1701 
1702 		state->fack_count += tcp_skb_pcount(skb);
1703 	}
1704 	return skb;
1705 }
1706 
1707 /* Avoid all extra work that is being done by sacktag while walking in
1708  * a normal way
1709  */
1710 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1711 					struct tcp_sacktag_state *state,
1712 					u32 skip_to_seq)
1713 {
1714 	tcp_for_write_queue_from(skb, sk) {
1715 		if (skb == tcp_send_head(sk))
1716 			break;
1717 
1718 		if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1719 			break;
1720 
1721 		state->fack_count += tcp_skb_pcount(skb);
1722 	}
1723 	return skb;
1724 }
1725 
1726 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1727 						struct sock *sk,
1728 						struct tcp_sack_block *next_dup,
1729 						struct tcp_sacktag_state *state,
1730 						u32 skip_to_seq)
1731 {
1732 	if (next_dup == NULL)
1733 		return skb;
1734 
1735 	if (before(next_dup->start_seq, skip_to_seq)) {
1736 		skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1737 		skb = tcp_sacktag_walk(skb, sk, NULL, state,
1738 				       next_dup->start_seq, next_dup->end_seq,
1739 				       1);
1740 	}
1741 
1742 	return skb;
1743 }
1744 
1745 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1746 {
1747 	return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1748 }
1749 
1750 static int
1751 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1752 			u32 prior_snd_una)
1753 {
1754 	const struct inet_connection_sock *icsk = inet_csk(sk);
1755 	struct tcp_sock *tp = tcp_sk(sk);
1756 	const unsigned char *ptr = (skb_transport_header(ack_skb) +
1757 				    TCP_SKB_CB(ack_skb)->sacked);
1758 	struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1759 	struct tcp_sack_block sp[TCP_NUM_SACKS];
1760 	struct tcp_sack_block *cache;
1761 	struct tcp_sacktag_state state;
1762 	struct sk_buff *skb;
1763 	int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1764 	int used_sacks;
1765 	int found_dup_sack = 0;
1766 	int i, j;
1767 	int first_sack_index;
1768 
1769 	state.flag = 0;
1770 	state.reord = tp->packets_out;
1771 
1772 	if (!tp->sacked_out) {
1773 		if (WARN_ON(tp->fackets_out))
1774 			tp->fackets_out = 0;
1775 		tcp_highest_sack_reset(sk);
1776 	}
1777 
1778 	found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1779 					 num_sacks, prior_snd_una);
1780 	if (found_dup_sack)
1781 		state.flag |= FLAG_DSACKING_ACK;
1782 
1783 	/* Eliminate too old ACKs, but take into
1784 	 * account more or less fresh ones, they can
1785 	 * contain valid SACK info.
1786 	 */
1787 	if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1788 		return 0;
1789 
1790 	if (!tp->packets_out)
1791 		goto out;
1792 
1793 	used_sacks = 0;
1794 	first_sack_index = 0;
1795 	for (i = 0; i < num_sacks; i++) {
1796 		int dup_sack = !i && found_dup_sack;
1797 
1798 		sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1799 		sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1800 
1801 		if (!tcp_is_sackblock_valid(tp, dup_sack,
1802 					    sp[used_sacks].start_seq,
1803 					    sp[used_sacks].end_seq)) {
1804 			int mib_idx;
1805 
1806 			if (dup_sack) {
1807 				if (!tp->undo_marker)
1808 					mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1809 				else
1810 					mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1811 			} else {
1812 				/* Don't count olds caused by ACK reordering */
1813 				if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1814 				    !after(sp[used_sacks].end_seq, tp->snd_una))
1815 					continue;
1816 				mib_idx = LINUX_MIB_TCPSACKDISCARD;
1817 			}
1818 
1819 			NET_INC_STATS_BH(sock_net(sk), mib_idx);
1820 			if (i == 0)
1821 				first_sack_index = -1;
1822 			continue;
1823 		}
1824 
1825 		/* Ignore very old stuff early */
1826 		if (!after(sp[used_sacks].end_seq, prior_snd_una))
1827 			continue;
1828 
1829 		used_sacks++;
1830 	}
1831 
1832 	/* order SACK blocks to allow in order walk of the retrans queue */
1833 	for (i = used_sacks - 1; i > 0; i--) {
1834 		for (j = 0; j < i; j++) {
1835 			if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1836 				swap(sp[j], sp[j + 1]);
1837 
1838 				/* Track where the first SACK block goes to */
1839 				if (j == first_sack_index)
1840 					first_sack_index = j + 1;
1841 			}
1842 		}
1843 	}
1844 
1845 	skb = tcp_write_queue_head(sk);
1846 	state.fack_count = 0;
1847 	i = 0;
1848 
1849 	if (!tp->sacked_out) {
1850 		/* It's already past, so skip checking against it */
1851 		cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1852 	} else {
1853 		cache = tp->recv_sack_cache;
1854 		/* Skip empty blocks in at head of the cache */
1855 		while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1856 		       !cache->end_seq)
1857 			cache++;
1858 	}
1859 
1860 	while (i < used_sacks) {
1861 		u32 start_seq = sp[i].start_seq;
1862 		u32 end_seq = sp[i].end_seq;
1863 		int dup_sack = (found_dup_sack && (i == first_sack_index));
1864 		struct tcp_sack_block *next_dup = NULL;
1865 
1866 		if (found_dup_sack && ((i + 1) == first_sack_index))
1867 			next_dup = &sp[i + 1];
1868 
1869 		/* Skip too early cached blocks */
1870 		while (tcp_sack_cache_ok(tp, cache) &&
1871 		       !before(start_seq, cache->end_seq))
1872 			cache++;
1873 
1874 		/* Can skip some work by looking recv_sack_cache? */
1875 		if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1876 		    after(end_seq, cache->start_seq)) {
1877 
1878 			/* Head todo? */
1879 			if (before(start_seq, cache->start_seq)) {
1880 				skb = tcp_sacktag_skip(skb, sk, &state,
1881 						       start_seq);
1882 				skb = tcp_sacktag_walk(skb, sk, next_dup,
1883 						       &state,
1884 						       start_seq,
1885 						       cache->start_seq,
1886 						       dup_sack);
1887 			}
1888 
1889 			/* Rest of the block already fully processed? */
1890 			if (!after(end_seq, cache->end_seq))
1891 				goto advance_sp;
1892 
1893 			skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1894 						       &state,
1895 						       cache->end_seq);
1896 
1897 			/* ...tail remains todo... */
1898 			if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1899 				/* ...but better entrypoint exists! */
1900 				skb = tcp_highest_sack(sk);
1901 				if (skb == NULL)
1902 					break;
1903 				state.fack_count = tp->fackets_out;
1904 				cache++;
1905 				goto walk;
1906 			}
1907 
1908 			skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1909 			/* Check overlap against next cached too (past this one already) */
1910 			cache++;
1911 			continue;
1912 		}
1913 
1914 		if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1915 			skb = tcp_highest_sack(sk);
1916 			if (skb == NULL)
1917 				break;
1918 			state.fack_count = tp->fackets_out;
1919 		}
1920 		skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1921 
1922 walk:
1923 		skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1924 				       start_seq, end_seq, dup_sack);
1925 
1926 advance_sp:
1927 		/* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1928 		 * due to in-order walk
1929 		 */
1930 		if (after(end_seq, tp->frto_highmark))
1931 			state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1932 
1933 		i++;
1934 	}
1935 
1936 	/* Clear the head of the cache sack blocks so we can skip it next time */
1937 	for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1938 		tp->recv_sack_cache[i].start_seq = 0;
1939 		tp->recv_sack_cache[i].end_seq = 0;
1940 	}
1941 	for (j = 0; j < used_sacks; j++)
1942 		tp->recv_sack_cache[i++] = sp[j];
1943 
1944 	tcp_mark_lost_retrans(sk);
1945 
1946 	tcp_verify_left_out(tp);
1947 
1948 	if ((state.reord < tp->fackets_out) &&
1949 	    ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1950 	    (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1951 		tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1952 
1953 out:
1954 
1955 #if FASTRETRANS_DEBUG > 0
1956 	WARN_ON((int)tp->sacked_out < 0);
1957 	WARN_ON((int)tp->lost_out < 0);
1958 	WARN_ON((int)tp->retrans_out < 0);
1959 	WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1960 #endif
1961 	return state.flag;
1962 }
1963 
1964 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1965  * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1966  */
1967 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1968 {
1969 	u32 holes;
1970 
1971 	holes = max(tp->lost_out, 1U);
1972 	holes = min(holes, tp->packets_out);
1973 
1974 	if ((tp->sacked_out + holes) > tp->packets_out) {
1975 		tp->sacked_out = tp->packets_out - holes;
1976 		return 1;
1977 	}
1978 	return 0;
1979 }
1980 
1981 /* If we receive more dupacks than we expected counting segments
1982  * in assumption of absent reordering, interpret this as reordering.
1983  * The only another reason could be bug in receiver TCP.
1984  */
1985 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1986 {
1987 	struct tcp_sock *tp = tcp_sk(sk);
1988 	if (tcp_limit_reno_sacked(tp))
1989 		tcp_update_reordering(sk, tp->packets_out + addend, 0);
1990 }
1991 
1992 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1993 
1994 static void tcp_add_reno_sack(struct sock *sk)
1995 {
1996 	struct tcp_sock *tp = tcp_sk(sk);
1997 	tp->sacked_out++;
1998 	tcp_check_reno_reordering(sk, 0);
1999 	tcp_verify_left_out(tp);
2000 }
2001 
2002 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2003 
2004 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
2005 {
2006 	struct tcp_sock *tp = tcp_sk(sk);
2007 
2008 	if (acked > 0) {
2009 		/* One ACK acked hole. The rest eat duplicate ACKs. */
2010 		if (acked - 1 >= tp->sacked_out)
2011 			tp->sacked_out = 0;
2012 		else
2013 			tp->sacked_out -= acked - 1;
2014 	}
2015 	tcp_check_reno_reordering(sk, acked);
2016 	tcp_verify_left_out(tp);
2017 }
2018 
2019 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2020 {
2021 	tp->sacked_out = 0;
2022 }
2023 
2024 static int tcp_is_sackfrto(const struct tcp_sock *tp)
2025 {
2026 	return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2027 }
2028 
2029 /* F-RTO can only be used if TCP has never retransmitted anything other than
2030  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2031  */
2032 int tcp_use_frto(struct sock *sk)
2033 {
2034 	const struct tcp_sock *tp = tcp_sk(sk);
2035 	const struct inet_connection_sock *icsk = inet_csk(sk);
2036 	struct sk_buff *skb;
2037 
2038 	if (!sysctl_tcp_frto)
2039 		return 0;
2040 
2041 	/* MTU probe and F-RTO won't really play nicely along currently */
2042 	if (icsk->icsk_mtup.probe_size)
2043 		return 0;
2044 
2045 	if (tcp_is_sackfrto(tp))
2046 		return 1;
2047 
2048 	/* Avoid expensive walking of rexmit queue if possible */
2049 	if (tp->retrans_out > 1)
2050 		return 0;
2051 
2052 	skb = tcp_write_queue_head(sk);
2053 	if (tcp_skb_is_last(sk, skb))
2054 		return 1;
2055 	skb = tcp_write_queue_next(sk, skb);	/* Skips head */
2056 	tcp_for_write_queue_from(skb, sk) {
2057 		if (skb == tcp_send_head(sk))
2058 			break;
2059 		if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2060 			return 0;
2061 		/* Short-circuit when first non-SACKed skb has been checked */
2062 		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2063 			break;
2064 	}
2065 	return 1;
2066 }
2067 
2068 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2069  * recovery a bit and use heuristics in tcp_process_frto() to detect if
2070  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2071  * keep retrans_out counting accurate (with SACK F-RTO, other than head
2072  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2073  * bits are handled if the Loss state is really to be entered (in
2074  * tcp_enter_frto_loss).
2075  *
2076  * Do like tcp_enter_loss() would; when RTO expires the second time it
2077  * does:
2078  *  "Reduce ssthresh if it has not yet been made inside this window."
2079  */
2080 void tcp_enter_frto(struct sock *sk)
2081 {
2082 	const struct inet_connection_sock *icsk = inet_csk(sk);
2083 	struct tcp_sock *tp = tcp_sk(sk);
2084 	struct sk_buff *skb;
2085 
2086 	if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2087 	    tp->snd_una == tp->high_seq ||
2088 	    ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2089 	     !icsk->icsk_retransmits)) {
2090 		tp->prior_ssthresh = tcp_current_ssthresh(sk);
2091 		/* Our state is too optimistic in ssthresh() call because cwnd
2092 		 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2093 		 * recovery has not yet completed. Pattern would be this: RTO,
2094 		 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2095 		 * up here twice).
2096 		 * RFC4138 should be more specific on what to do, even though
2097 		 * RTO is quite unlikely to occur after the first Cumulative ACK
2098 		 * due to back-off and complexity of triggering events ...
2099 		 */
2100 		if (tp->frto_counter) {
2101 			u32 stored_cwnd;
2102 			stored_cwnd = tp->snd_cwnd;
2103 			tp->snd_cwnd = 2;
2104 			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2105 			tp->snd_cwnd = stored_cwnd;
2106 		} else {
2107 			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2108 		}
2109 		/* ... in theory, cong.control module could do "any tricks" in
2110 		 * ssthresh(), which means that ca_state, lost bits and lost_out
2111 		 * counter would have to be faked before the call occurs. We
2112 		 * consider that too expensive, unlikely and hacky, so modules
2113 		 * using these in ssthresh() must deal these incompatibility
2114 		 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2115 		 */
2116 		tcp_ca_event(sk, CA_EVENT_FRTO);
2117 	}
2118 
2119 	tp->undo_marker = tp->snd_una;
2120 	tp->undo_retrans = 0;
2121 
2122 	skb = tcp_write_queue_head(sk);
2123 	if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2124 		tp->undo_marker = 0;
2125 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2126 		TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2127 		tp->retrans_out -= tcp_skb_pcount(skb);
2128 	}
2129 	tcp_verify_left_out(tp);
2130 
2131 	/* Too bad if TCP was application limited */
2132 	tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2133 
2134 	/* Earlier loss recovery underway (see RFC4138; Appendix B).
2135 	 * The last condition is necessary at least in tp->frto_counter case.
2136 	 */
2137 	if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2138 	    ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2139 	    after(tp->high_seq, tp->snd_una)) {
2140 		tp->frto_highmark = tp->high_seq;
2141 	} else {
2142 		tp->frto_highmark = tp->snd_nxt;
2143 	}
2144 	tcp_set_ca_state(sk, TCP_CA_Disorder);
2145 	tp->high_seq = tp->snd_nxt;
2146 	tp->frto_counter = 1;
2147 }
2148 
2149 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2150  * which indicates that we should follow the traditional RTO recovery,
2151  * i.e. mark everything lost and do go-back-N retransmission.
2152  */
2153 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2154 {
2155 	struct tcp_sock *tp = tcp_sk(sk);
2156 	struct sk_buff *skb;
2157 
2158 	tp->lost_out = 0;
2159 	tp->retrans_out = 0;
2160 	if (tcp_is_reno(tp))
2161 		tcp_reset_reno_sack(tp);
2162 
2163 	tcp_for_write_queue(skb, sk) {
2164 		if (skb == tcp_send_head(sk))
2165 			break;
2166 
2167 		TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2168 		/*
2169 		 * Count the retransmission made on RTO correctly (only when
2170 		 * waiting for the first ACK and did not get it)...
2171 		 */
2172 		if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2173 			/* For some reason this R-bit might get cleared? */
2174 			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2175 				tp->retrans_out += tcp_skb_pcount(skb);
2176 			/* ...enter this if branch just for the first segment */
2177 			flag |= FLAG_DATA_ACKED;
2178 		} else {
2179 			if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2180 				tp->undo_marker = 0;
2181 			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2182 		}
2183 
2184 		/* Marking forward transmissions that were made after RTO lost
2185 		 * can cause unnecessary retransmissions in some scenarios,
2186 		 * SACK blocks will mitigate that in some but not in all cases.
2187 		 * We used to not mark them but it was causing break-ups with
2188 		 * receivers that do only in-order receival.
2189 		 *
2190 		 * TODO: we could detect presence of such receiver and select
2191 		 * different behavior per flow.
2192 		 */
2193 		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2194 			TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2195 			tp->lost_out += tcp_skb_pcount(skb);
2196 			tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2197 		}
2198 	}
2199 	tcp_verify_left_out(tp);
2200 
2201 	tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2202 	tp->snd_cwnd_cnt = 0;
2203 	tp->snd_cwnd_stamp = tcp_time_stamp;
2204 	tp->frto_counter = 0;
2205 	tp->bytes_acked = 0;
2206 
2207 	tp->reordering = min_t(unsigned int, tp->reordering,
2208 			       sysctl_tcp_reordering);
2209 	tcp_set_ca_state(sk, TCP_CA_Loss);
2210 	tp->high_seq = tp->snd_nxt;
2211 	TCP_ECN_queue_cwr(tp);
2212 
2213 	tcp_clear_all_retrans_hints(tp);
2214 }
2215 
2216 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2217 {
2218 	tp->retrans_out = 0;
2219 	tp->lost_out = 0;
2220 
2221 	tp->undo_marker = 0;
2222 	tp->undo_retrans = 0;
2223 }
2224 
2225 void tcp_clear_retrans(struct tcp_sock *tp)
2226 {
2227 	tcp_clear_retrans_partial(tp);
2228 
2229 	tp->fackets_out = 0;
2230 	tp->sacked_out = 0;
2231 }
2232 
2233 /* Enter Loss state. If "how" is not zero, forget all SACK information
2234  * and reset tags completely, otherwise preserve SACKs. If receiver
2235  * dropped its ofo queue, we will know this due to reneging detection.
2236  */
2237 void tcp_enter_loss(struct sock *sk, int how)
2238 {
2239 	const struct inet_connection_sock *icsk = inet_csk(sk);
2240 	struct tcp_sock *tp = tcp_sk(sk);
2241 	struct sk_buff *skb;
2242 
2243 	/* Reduce ssthresh if it has not yet been made inside this window. */
2244 	if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2245 	    (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2246 		tp->prior_ssthresh = tcp_current_ssthresh(sk);
2247 		tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2248 		tcp_ca_event(sk, CA_EVENT_LOSS);
2249 	}
2250 	tp->snd_cwnd	   = 1;
2251 	tp->snd_cwnd_cnt   = 0;
2252 	tp->snd_cwnd_stamp = tcp_time_stamp;
2253 
2254 	tp->bytes_acked = 0;
2255 	tcp_clear_retrans_partial(tp);
2256 
2257 	if (tcp_is_reno(tp))
2258 		tcp_reset_reno_sack(tp);
2259 
2260 	if (!how) {
2261 		/* Push undo marker, if it was plain RTO and nothing
2262 		 * was retransmitted. */
2263 		tp->undo_marker = tp->snd_una;
2264 	} else {
2265 		tp->sacked_out = 0;
2266 		tp->fackets_out = 0;
2267 	}
2268 	tcp_clear_all_retrans_hints(tp);
2269 
2270 	tcp_for_write_queue(skb, sk) {
2271 		if (skb == tcp_send_head(sk))
2272 			break;
2273 
2274 		if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2275 			tp->undo_marker = 0;
2276 		TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2277 		if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2278 			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2279 			TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2280 			tp->lost_out += tcp_skb_pcount(skb);
2281 			tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2282 		}
2283 	}
2284 	tcp_verify_left_out(tp);
2285 
2286 	tp->reordering = min_t(unsigned int, tp->reordering,
2287 			       sysctl_tcp_reordering);
2288 	tcp_set_ca_state(sk, TCP_CA_Loss);
2289 	tp->high_seq = tp->snd_nxt;
2290 	TCP_ECN_queue_cwr(tp);
2291 	/* Abort F-RTO algorithm if one is in progress */
2292 	tp->frto_counter = 0;
2293 }
2294 
2295 /* If ACK arrived pointing to a remembered SACK, it means that our
2296  * remembered SACKs do not reflect real state of receiver i.e.
2297  * receiver _host_ is heavily congested (or buggy).
2298  *
2299  * Do processing similar to RTO timeout.
2300  */
2301 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2302 {
2303 	if (flag & FLAG_SACK_RENEGING) {
2304 		struct inet_connection_sock *icsk = inet_csk(sk);
2305 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2306 
2307 		tcp_enter_loss(sk, 1);
2308 		icsk->icsk_retransmits++;
2309 		tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2310 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2311 					  icsk->icsk_rto, TCP_RTO_MAX);
2312 		return 1;
2313 	}
2314 	return 0;
2315 }
2316 
2317 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2318 {
2319 	return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2320 }
2321 
2322 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2323  * counter when SACK is enabled (without SACK, sacked_out is used for
2324  * that purpose).
2325  *
2326  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2327  * segments up to the highest received SACK block so far and holes in
2328  * between them.
2329  *
2330  * With reordering, holes may still be in flight, so RFC3517 recovery
2331  * uses pure sacked_out (total number of SACKed segments) even though
2332  * it violates the RFC that uses duplicate ACKs, often these are equal
2333  * but when e.g. out-of-window ACKs or packet duplication occurs,
2334  * they differ. Since neither occurs due to loss, TCP should really
2335  * ignore them.
2336  */
2337 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2338 {
2339 	return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2340 }
2341 
2342 static inline int tcp_skb_timedout(const struct sock *sk,
2343 				   const struct sk_buff *skb)
2344 {
2345 	return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2346 }
2347 
2348 static inline int tcp_head_timedout(const struct sock *sk)
2349 {
2350 	const struct tcp_sock *tp = tcp_sk(sk);
2351 
2352 	return tp->packets_out &&
2353 	       tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2354 }
2355 
2356 /* Linux NewReno/SACK/FACK/ECN state machine.
2357  * --------------------------------------
2358  *
2359  * "Open"	Normal state, no dubious events, fast path.
2360  * "Disorder"   In all the respects it is "Open",
2361  *		but requires a bit more attention. It is entered when
2362  *		we see some SACKs or dupacks. It is split of "Open"
2363  *		mainly to move some processing from fast path to slow one.
2364  * "CWR"	CWND was reduced due to some Congestion Notification event.
2365  *		It can be ECN, ICMP source quench, local device congestion.
2366  * "Recovery"	CWND was reduced, we are fast-retransmitting.
2367  * "Loss"	CWND was reduced due to RTO timeout or SACK reneging.
2368  *
2369  * tcp_fastretrans_alert() is entered:
2370  * - each incoming ACK, if state is not "Open"
2371  * - when arrived ACK is unusual, namely:
2372  *	* SACK
2373  *	* Duplicate ACK.
2374  *	* ECN ECE.
2375  *
2376  * Counting packets in flight is pretty simple.
2377  *
2378  *	in_flight = packets_out - left_out + retrans_out
2379  *
2380  *	packets_out is SND.NXT-SND.UNA counted in packets.
2381  *
2382  *	retrans_out is number of retransmitted segments.
2383  *
2384  *	left_out is number of segments left network, but not ACKed yet.
2385  *
2386  *		left_out = sacked_out + lost_out
2387  *
2388  *     sacked_out: Packets, which arrived to receiver out of order
2389  *		   and hence not ACKed. With SACKs this number is simply
2390  *		   amount of SACKed data. Even without SACKs
2391  *		   it is easy to give pretty reliable estimate of this number,
2392  *		   counting duplicate ACKs.
2393  *
2394  *       lost_out: Packets lost by network. TCP has no explicit
2395  *		   "loss notification" feedback from network (for now).
2396  *		   It means that this number can be only _guessed_.
2397  *		   Actually, it is the heuristics to predict lossage that
2398  *		   distinguishes different algorithms.
2399  *
2400  *	F.e. after RTO, when all the queue is considered as lost,
2401  *	lost_out = packets_out and in_flight = retrans_out.
2402  *
2403  *		Essentially, we have now two algorithms counting
2404  *		lost packets.
2405  *
2406  *		FACK: It is the simplest heuristics. As soon as we decided
2407  *		that something is lost, we decide that _all_ not SACKed
2408  *		packets until the most forward SACK are lost. I.e.
2409  *		lost_out = fackets_out - sacked_out and left_out = fackets_out.
2410  *		It is absolutely correct estimate, if network does not reorder
2411  *		packets. And it loses any connection to reality when reordering
2412  *		takes place. We use FACK by default until reordering
2413  *		is suspected on the path to this destination.
2414  *
2415  *		NewReno: when Recovery is entered, we assume that one segment
2416  *		is lost (classic Reno). While we are in Recovery and
2417  *		a partial ACK arrives, we assume that one more packet
2418  *		is lost (NewReno). This heuristics are the same in NewReno
2419  *		and SACK.
2420  *
2421  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2422  *  deflation etc. CWND is real congestion window, never inflated, changes
2423  *  only according to classic VJ rules.
2424  *
2425  * Really tricky (and requiring careful tuning) part of algorithm
2426  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2427  * The first determines the moment _when_ we should reduce CWND and,
2428  * hence, slow down forward transmission. In fact, it determines the moment
2429  * when we decide that hole is caused by loss, rather than by a reorder.
2430  *
2431  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2432  * holes, caused by lost packets.
2433  *
2434  * And the most logically complicated part of algorithm is undo
2435  * heuristics. We detect false retransmits due to both too early
2436  * fast retransmit (reordering) and underestimated RTO, analyzing
2437  * timestamps and D-SACKs. When we detect that some segments were
2438  * retransmitted by mistake and CWND reduction was wrong, we undo
2439  * window reduction and abort recovery phase. This logic is hidden
2440  * inside several functions named tcp_try_undo_<something>.
2441  */
2442 
2443 /* This function decides, when we should leave Disordered state
2444  * and enter Recovery phase, reducing congestion window.
2445  *
2446  * Main question: may we further continue forward transmission
2447  * with the same cwnd?
2448  */
2449 static int tcp_time_to_recover(struct sock *sk)
2450 {
2451 	struct tcp_sock *tp = tcp_sk(sk);
2452 	__u32 packets_out;
2453 
2454 	/* Do not perform any recovery during F-RTO algorithm */
2455 	if (tp->frto_counter)
2456 		return 0;
2457 
2458 	/* Trick#1: The loss is proven. */
2459 	if (tp->lost_out)
2460 		return 1;
2461 
2462 	/* Not-A-Trick#2 : Classic rule... */
2463 	if (tcp_dupack_heuristics(tp) > tp->reordering)
2464 		return 1;
2465 
2466 	/* Trick#3 : when we use RFC2988 timer restart, fast
2467 	 * retransmit can be triggered by timeout of queue head.
2468 	 */
2469 	if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2470 		return 1;
2471 
2472 	/* Trick#4: It is still not OK... But will it be useful to delay
2473 	 * recovery more?
2474 	 */
2475 	packets_out = tp->packets_out;
2476 	if (packets_out <= tp->reordering &&
2477 	    tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2478 	    !tcp_may_send_now(sk)) {
2479 		/* We have nothing to send. This connection is limited
2480 		 * either by receiver window or by application.
2481 		 */
2482 		return 1;
2483 	}
2484 
2485 	/* If a thin stream is detected, retransmit after first
2486 	 * received dupack. Employ only if SACK is supported in order
2487 	 * to avoid possible corner-case series of spurious retransmissions
2488 	 * Use only if there are no unsent data.
2489 	 */
2490 	if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2491 	    tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2492 	    tcp_is_sack(tp) && !tcp_send_head(sk))
2493 		return 1;
2494 
2495 	return 0;
2496 }
2497 
2498 /* New heuristics: it is possible only after we switched to restart timer
2499  * each time when something is ACKed. Hence, we can detect timed out packets
2500  * during fast retransmit without falling to slow start.
2501  *
2502  * Usefulness of this as is very questionable, since we should know which of
2503  * the segments is the next to timeout which is relatively expensive to find
2504  * in general case unless we add some data structure just for that. The
2505  * current approach certainly won't find the right one too often and when it
2506  * finally does find _something_ it usually marks large part of the window
2507  * right away (because a retransmission with a larger timestamp blocks the
2508  * loop from advancing). -ij
2509  */
2510 static void tcp_timeout_skbs(struct sock *sk)
2511 {
2512 	struct tcp_sock *tp = tcp_sk(sk);
2513 	struct sk_buff *skb;
2514 
2515 	if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2516 		return;
2517 
2518 	skb = tp->scoreboard_skb_hint;
2519 	if (tp->scoreboard_skb_hint == NULL)
2520 		skb = tcp_write_queue_head(sk);
2521 
2522 	tcp_for_write_queue_from(skb, sk) {
2523 		if (skb == tcp_send_head(sk))
2524 			break;
2525 		if (!tcp_skb_timedout(sk, skb))
2526 			break;
2527 
2528 		tcp_skb_mark_lost(tp, skb);
2529 	}
2530 
2531 	tp->scoreboard_skb_hint = skb;
2532 
2533 	tcp_verify_left_out(tp);
2534 }
2535 
2536 /* Detect loss in event "A" above by marking head of queue up as lost.
2537  * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2538  * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2539  * has at least tp->reordering SACKed seqments above it; "packets" refers to
2540  * the maximum SACKed segments to pass before reaching this limit.
2541  */
2542 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2543 {
2544 	struct tcp_sock *tp = tcp_sk(sk);
2545 	struct sk_buff *skb;
2546 	int cnt, oldcnt;
2547 	int err;
2548 	unsigned int mss;
2549 	/* Use SACK to deduce losses of new sequences sent during recovery */
2550 	const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2551 
2552 	WARN_ON(packets > tp->packets_out);
2553 	if (tp->lost_skb_hint) {
2554 		skb = tp->lost_skb_hint;
2555 		cnt = tp->lost_cnt_hint;
2556 		/* Head already handled? */
2557 		if (mark_head && skb != tcp_write_queue_head(sk))
2558 			return;
2559 	} else {
2560 		skb = tcp_write_queue_head(sk);
2561 		cnt = 0;
2562 	}
2563 
2564 	tcp_for_write_queue_from(skb, sk) {
2565 		if (skb == tcp_send_head(sk))
2566 			break;
2567 		/* TODO: do this better */
2568 		/* this is not the most efficient way to do this... */
2569 		tp->lost_skb_hint = skb;
2570 		tp->lost_cnt_hint = cnt;
2571 
2572 		if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2573 			break;
2574 
2575 		oldcnt = cnt;
2576 		if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2577 		    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2578 			cnt += tcp_skb_pcount(skb);
2579 
2580 		if (cnt > packets) {
2581 			if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2582 			    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2583 			    (oldcnt >= packets))
2584 				break;
2585 
2586 			mss = skb_shinfo(skb)->gso_size;
2587 			err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2588 			if (err < 0)
2589 				break;
2590 			cnt = packets;
2591 		}
2592 
2593 		tcp_skb_mark_lost(tp, skb);
2594 
2595 		if (mark_head)
2596 			break;
2597 	}
2598 	tcp_verify_left_out(tp);
2599 }
2600 
2601 /* Account newly detected lost packet(s) */
2602 
2603 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2604 {
2605 	struct tcp_sock *tp = tcp_sk(sk);
2606 
2607 	if (tcp_is_reno(tp)) {
2608 		tcp_mark_head_lost(sk, 1, 1);
2609 	} else if (tcp_is_fack(tp)) {
2610 		int lost = tp->fackets_out - tp->reordering;
2611 		if (lost <= 0)
2612 			lost = 1;
2613 		tcp_mark_head_lost(sk, lost, 0);
2614 	} else {
2615 		int sacked_upto = tp->sacked_out - tp->reordering;
2616 		if (sacked_upto >= 0)
2617 			tcp_mark_head_lost(sk, sacked_upto, 0);
2618 		else if (fast_rexmit)
2619 			tcp_mark_head_lost(sk, 1, 1);
2620 	}
2621 
2622 	tcp_timeout_skbs(sk);
2623 }
2624 
2625 /* CWND moderation, preventing bursts due to too big ACKs
2626  * in dubious situations.
2627  */
2628 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2629 {
2630 	tp->snd_cwnd = min(tp->snd_cwnd,
2631 			   tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2632 	tp->snd_cwnd_stamp = tcp_time_stamp;
2633 }
2634 
2635 /* Lower bound on congestion window is slow start threshold
2636  * unless congestion avoidance choice decides to overide it.
2637  */
2638 static inline u32 tcp_cwnd_min(const struct sock *sk)
2639 {
2640 	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2641 
2642 	return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2643 }
2644 
2645 /* Decrease cwnd each second ack. */
2646 static void tcp_cwnd_down(struct sock *sk, int flag)
2647 {
2648 	struct tcp_sock *tp = tcp_sk(sk);
2649 	int decr = tp->snd_cwnd_cnt + 1;
2650 
2651 	if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2652 	    (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2653 		tp->snd_cwnd_cnt = decr & 1;
2654 		decr >>= 1;
2655 
2656 		if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2657 			tp->snd_cwnd -= decr;
2658 
2659 		tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2660 		tp->snd_cwnd_stamp = tcp_time_stamp;
2661 	}
2662 }
2663 
2664 /* Nothing was retransmitted or returned timestamp is less
2665  * than timestamp of the first retransmission.
2666  */
2667 static inline int tcp_packet_delayed(const struct tcp_sock *tp)
2668 {
2669 	return !tp->retrans_stamp ||
2670 		(tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2671 		 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2672 }
2673 
2674 /* Undo procedures. */
2675 
2676 #if FASTRETRANS_DEBUG > 1
2677 static void DBGUNDO(struct sock *sk, const char *msg)
2678 {
2679 	struct tcp_sock *tp = tcp_sk(sk);
2680 	struct inet_sock *inet = inet_sk(sk);
2681 
2682 	if (sk->sk_family == AF_INET) {
2683 		printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2684 		       msg,
2685 		       &inet->inet_daddr, ntohs(inet->inet_dport),
2686 		       tp->snd_cwnd, tcp_left_out(tp),
2687 		       tp->snd_ssthresh, tp->prior_ssthresh,
2688 		       tp->packets_out);
2689 	}
2690 #if IS_ENABLED(CONFIG_IPV6)
2691 	else if (sk->sk_family == AF_INET6) {
2692 		struct ipv6_pinfo *np = inet6_sk(sk);
2693 		printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2694 		       msg,
2695 		       &np->daddr, ntohs(inet->inet_dport),
2696 		       tp->snd_cwnd, tcp_left_out(tp),
2697 		       tp->snd_ssthresh, tp->prior_ssthresh,
2698 		       tp->packets_out);
2699 	}
2700 #endif
2701 }
2702 #else
2703 #define DBGUNDO(x...) do { } while (0)
2704 #endif
2705 
2706 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2707 {
2708 	struct tcp_sock *tp = tcp_sk(sk);
2709 
2710 	if (tp->prior_ssthresh) {
2711 		const struct inet_connection_sock *icsk = inet_csk(sk);
2712 
2713 		if (icsk->icsk_ca_ops->undo_cwnd)
2714 			tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2715 		else
2716 			tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2717 
2718 		if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2719 			tp->snd_ssthresh = tp->prior_ssthresh;
2720 			TCP_ECN_withdraw_cwr(tp);
2721 		}
2722 	} else {
2723 		tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2724 	}
2725 	tp->snd_cwnd_stamp = tcp_time_stamp;
2726 }
2727 
2728 static inline int tcp_may_undo(const struct tcp_sock *tp)
2729 {
2730 	return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2731 }
2732 
2733 /* People celebrate: "We love our President!" */
2734 static int tcp_try_undo_recovery(struct sock *sk)
2735 {
2736 	struct tcp_sock *tp = tcp_sk(sk);
2737 
2738 	if (tcp_may_undo(tp)) {
2739 		int mib_idx;
2740 
2741 		/* Happy end! We did not retransmit anything
2742 		 * or our original transmission succeeded.
2743 		 */
2744 		DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2745 		tcp_undo_cwr(sk, true);
2746 		if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2747 			mib_idx = LINUX_MIB_TCPLOSSUNDO;
2748 		else
2749 			mib_idx = LINUX_MIB_TCPFULLUNDO;
2750 
2751 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
2752 		tp->undo_marker = 0;
2753 	}
2754 	if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2755 		/* Hold old state until something *above* high_seq
2756 		 * is ACKed. For Reno it is MUST to prevent false
2757 		 * fast retransmits (RFC2582). SACK TCP is safe. */
2758 		tcp_moderate_cwnd(tp);
2759 		return 1;
2760 	}
2761 	tcp_set_ca_state(sk, TCP_CA_Open);
2762 	return 0;
2763 }
2764 
2765 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2766 static void tcp_try_undo_dsack(struct sock *sk)
2767 {
2768 	struct tcp_sock *tp = tcp_sk(sk);
2769 
2770 	if (tp->undo_marker && !tp->undo_retrans) {
2771 		DBGUNDO(sk, "D-SACK");
2772 		tcp_undo_cwr(sk, true);
2773 		tp->undo_marker = 0;
2774 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2775 	}
2776 }
2777 
2778 /* We can clear retrans_stamp when there are no retransmissions in the
2779  * window. It would seem that it is trivially available for us in
2780  * tp->retrans_out, however, that kind of assumptions doesn't consider
2781  * what will happen if errors occur when sending retransmission for the
2782  * second time. ...It could the that such segment has only
2783  * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2784  * the head skb is enough except for some reneging corner cases that
2785  * are not worth the effort.
2786  *
2787  * Main reason for all this complexity is the fact that connection dying
2788  * time now depends on the validity of the retrans_stamp, in particular,
2789  * that successive retransmissions of a segment must not advance
2790  * retrans_stamp under any conditions.
2791  */
2792 static int tcp_any_retrans_done(const struct sock *sk)
2793 {
2794 	const struct tcp_sock *tp = tcp_sk(sk);
2795 	struct sk_buff *skb;
2796 
2797 	if (tp->retrans_out)
2798 		return 1;
2799 
2800 	skb = tcp_write_queue_head(sk);
2801 	if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2802 		return 1;
2803 
2804 	return 0;
2805 }
2806 
2807 /* Undo during fast recovery after partial ACK. */
2808 
2809 static int tcp_try_undo_partial(struct sock *sk, int acked)
2810 {
2811 	struct tcp_sock *tp = tcp_sk(sk);
2812 	/* Partial ACK arrived. Force Hoe's retransmit. */
2813 	int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2814 
2815 	if (tcp_may_undo(tp)) {
2816 		/* Plain luck! Hole if filled with delayed
2817 		 * packet, rather than with a retransmit.
2818 		 */
2819 		if (!tcp_any_retrans_done(sk))
2820 			tp->retrans_stamp = 0;
2821 
2822 		tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2823 
2824 		DBGUNDO(sk, "Hoe");
2825 		tcp_undo_cwr(sk, false);
2826 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2827 
2828 		/* So... Do not make Hoe's retransmit yet.
2829 		 * If the first packet was delayed, the rest
2830 		 * ones are most probably delayed as well.
2831 		 */
2832 		failed = 0;
2833 	}
2834 	return failed;
2835 }
2836 
2837 /* Undo during loss recovery after partial ACK. */
2838 static int tcp_try_undo_loss(struct sock *sk)
2839 {
2840 	struct tcp_sock *tp = tcp_sk(sk);
2841 
2842 	if (tcp_may_undo(tp)) {
2843 		struct sk_buff *skb;
2844 		tcp_for_write_queue(skb, sk) {
2845 			if (skb == tcp_send_head(sk))
2846 				break;
2847 			TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2848 		}
2849 
2850 		tcp_clear_all_retrans_hints(tp);
2851 
2852 		DBGUNDO(sk, "partial loss");
2853 		tp->lost_out = 0;
2854 		tcp_undo_cwr(sk, true);
2855 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2856 		inet_csk(sk)->icsk_retransmits = 0;
2857 		tp->undo_marker = 0;
2858 		if (tcp_is_sack(tp))
2859 			tcp_set_ca_state(sk, TCP_CA_Open);
2860 		return 1;
2861 	}
2862 	return 0;
2863 }
2864 
2865 static inline void tcp_complete_cwr(struct sock *sk)
2866 {
2867 	struct tcp_sock *tp = tcp_sk(sk);
2868 
2869 	/* Do not moderate cwnd if it's already undone in cwr or recovery. */
2870 	if (tp->undo_marker) {
2871 		if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) {
2872 			tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2873 			tp->snd_cwnd_stamp = tcp_time_stamp;
2874 		} else if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH) {
2875 			/* PRR algorithm. */
2876 			tp->snd_cwnd = tp->snd_ssthresh;
2877 			tp->snd_cwnd_stamp = tcp_time_stamp;
2878 		}
2879 	}
2880 	tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2881 }
2882 
2883 static void tcp_try_keep_open(struct sock *sk)
2884 {
2885 	struct tcp_sock *tp = tcp_sk(sk);
2886 	int state = TCP_CA_Open;
2887 
2888 	if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2889 		state = TCP_CA_Disorder;
2890 
2891 	if (inet_csk(sk)->icsk_ca_state != state) {
2892 		tcp_set_ca_state(sk, state);
2893 		tp->high_seq = tp->snd_nxt;
2894 	}
2895 }
2896 
2897 static void tcp_try_to_open(struct sock *sk, int flag)
2898 {
2899 	struct tcp_sock *tp = tcp_sk(sk);
2900 
2901 	tcp_verify_left_out(tp);
2902 
2903 	if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2904 		tp->retrans_stamp = 0;
2905 
2906 	if (flag & FLAG_ECE)
2907 		tcp_enter_cwr(sk, 1);
2908 
2909 	if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2910 		tcp_try_keep_open(sk);
2911 		if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2912 			tcp_moderate_cwnd(tp);
2913 	} else {
2914 		tcp_cwnd_down(sk, flag);
2915 	}
2916 }
2917 
2918 static void tcp_mtup_probe_failed(struct sock *sk)
2919 {
2920 	struct inet_connection_sock *icsk = inet_csk(sk);
2921 
2922 	icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2923 	icsk->icsk_mtup.probe_size = 0;
2924 }
2925 
2926 static void tcp_mtup_probe_success(struct sock *sk)
2927 {
2928 	struct tcp_sock *tp = tcp_sk(sk);
2929 	struct inet_connection_sock *icsk = inet_csk(sk);
2930 
2931 	/* FIXME: breaks with very large cwnd */
2932 	tp->prior_ssthresh = tcp_current_ssthresh(sk);
2933 	tp->snd_cwnd = tp->snd_cwnd *
2934 		       tcp_mss_to_mtu(sk, tp->mss_cache) /
2935 		       icsk->icsk_mtup.probe_size;
2936 	tp->snd_cwnd_cnt = 0;
2937 	tp->snd_cwnd_stamp = tcp_time_stamp;
2938 	tp->snd_ssthresh = tcp_current_ssthresh(sk);
2939 
2940 	icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2941 	icsk->icsk_mtup.probe_size = 0;
2942 	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2943 }
2944 
2945 /* Do a simple retransmit without using the backoff mechanisms in
2946  * tcp_timer. This is used for path mtu discovery.
2947  * The socket is already locked here.
2948  */
2949 void tcp_simple_retransmit(struct sock *sk)
2950 {
2951 	const struct inet_connection_sock *icsk = inet_csk(sk);
2952 	struct tcp_sock *tp = tcp_sk(sk);
2953 	struct sk_buff *skb;
2954 	unsigned int mss = tcp_current_mss(sk);
2955 	u32 prior_lost = tp->lost_out;
2956 
2957 	tcp_for_write_queue(skb, sk) {
2958 		if (skb == tcp_send_head(sk))
2959 			break;
2960 		if (tcp_skb_seglen(skb) > mss &&
2961 		    !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2962 			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2963 				TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2964 				tp->retrans_out -= tcp_skb_pcount(skb);
2965 			}
2966 			tcp_skb_mark_lost_uncond_verify(tp, skb);
2967 		}
2968 	}
2969 
2970 	tcp_clear_retrans_hints_partial(tp);
2971 
2972 	if (prior_lost == tp->lost_out)
2973 		return;
2974 
2975 	if (tcp_is_reno(tp))
2976 		tcp_limit_reno_sacked(tp);
2977 
2978 	tcp_verify_left_out(tp);
2979 
2980 	/* Don't muck with the congestion window here.
2981 	 * Reason is that we do not increase amount of _data_
2982 	 * in network, but units changed and effective
2983 	 * cwnd/ssthresh really reduced now.
2984 	 */
2985 	if (icsk->icsk_ca_state != TCP_CA_Loss) {
2986 		tp->high_seq = tp->snd_nxt;
2987 		tp->snd_ssthresh = tcp_current_ssthresh(sk);
2988 		tp->prior_ssthresh = 0;
2989 		tp->undo_marker = 0;
2990 		tcp_set_ca_state(sk, TCP_CA_Loss);
2991 	}
2992 	tcp_xmit_retransmit_queue(sk);
2993 }
2994 EXPORT_SYMBOL(tcp_simple_retransmit);
2995 
2996 /* This function implements the PRR algorithm, specifcally the PRR-SSRB
2997  * (proportional rate reduction with slow start reduction bound) as described in
2998  * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2999  * It computes the number of packets to send (sndcnt) based on packets newly
3000  * delivered:
3001  *   1) If the packets in flight is larger than ssthresh, PRR spreads the
3002  *	cwnd reductions across a full RTT.
3003  *   2) If packets in flight is lower than ssthresh (such as due to excess
3004  *	losses and/or application stalls), do not perform any further cwnd
3005  *	reductions, but instead slow start up to ssthresh.
3006  */
3007 static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
3008 					int fast_rexmit, int flag)
3009 {
3010 	struct tcp_sock *tp = tcp_sk(sk);
3011 	int sndcnt = 0;
3012 	int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
3013 
3014 	if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
3015 		u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
3016 			       tp->prior_cwnd - 1;
3017 		sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
3018 	} else {
3019 		sndcnt = min_t(int, delta,
3020 			       max_t(int, tp->prr_delivered - tp->prr_out,
3021 				     newly_acked_sacked) + 1);
3022 	}
3023 
3024 	sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
3025 	tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
3026 }
3027 
3028 /* Process an event, which can update packets-in-flight not trivially.
3029  * Main goal of this function is to calculate new estimate for left_out,
3030  * taking into account both packets sitting in receiver's buffer and
3031  * packets lost by network.
3032  *
3033  * Besides that it does CWND reduction, when packet loss is detected
3034  * and changes state of machine.
3035  *
3036  * It does _not_ decide what to send, it is made in function
3037  * tcp_xmit_retransmit_queue().
3038  */
3039 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
3040 				  int newly_acked_sacked, bool is_dupack,
3041 				  int flag)
3042 {
3043 	struct inet_connection_sock *icsk = inet_csk(sk);
3044 	struct tcp_sock *tp = tcp_sk(sk);
3045 	int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
3046 				    (tcp_fackets_out(tp) > tp->reordering));
3047 	int fast_rexmit = 0, mib_idx;
3048 
3049 	if (WARN_ON(!tp->packets_out && tp->sacked_out))
3050 		tp->sacked_out = 0;
3051 	if (WARN_ON(!tp->sacked_out && tp->fackets_out))
3052 		tp->fackets_out = 0;
3053 
3054 	/* Now state machine starts.
3055 	 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3056 	if (flag & FLAG_ECE)
3057 		tp->prior_ssthresh = 0;
3058 
3059 	/* B. In all the states check for reneging SACKs. */
3060 	if (tcp_check_sack_reneging(sk, flag))
3061 		return;
3062 
3063 	/* C. Check consistency of the current state. */
3064 	tcp_verify_left_out(tp);
3065 
3066 	/* D. Check state exit conditions. State can be terminated
3067 	 *    when high_seq is ACKed. */
3068 	if (icsk->icsk_ca_state == TCP_CA_Open) {
3069 		WARN_ON(tp->retrans_out != 0);
3070 		tp->retrans_stamp = 0;
3071 	} else if (!before(tp->snd_una, tp->high_seq)) {
3072 		switch (icsk->icsk_ca_state) {
3073 		case TCP_CA_Loss:
3074 			icsk->icsk_retransmits = 0;
3075 			if (tcp_try_undo_recovery(sk))
3076 				return;
3077 			break;
3078 
3079 		case TCP_CA_CWR:
3080 			/* CWR is to be held something *above* high_seq
3081 			 * is ACKed for CWR bit to reach receiver. */
3082 			if (tp->snd_una != tp->high_seq) {
3083 				tcp_complete_cwr(sk);
3084 				tcp_set_ca_state(sk, TCP_CA_Open);
3085 			}
3086 			break;
3087 
3088 		case TCP_CA_Recovery:
3089 			if (tcp_is_reno(tp))
3090 				tcp_reset_reno_sack(tp);
3091 			if (tcp_try_undo_recovery(sk))
3092 				return;
3093 			tcp_complete_cwr(sk);
3094 			break;
3095 		}
3096 	}
3097 
3098 	/* E. Process state. */
3099 	switch (icsk->icsk_ca_state) {
3100 	case TCP_CA_Recovery:
3101 		if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3102 			if (tcp_is_reno(tp) && is_dupack)
3103 				tcp_add_reno_sack(sk);
3104 		} else
3105 			do_lost = tcp_try_undo_partial(sk, pkts_acked);
3106 		break;
3107 	case TCP_CA_Loss:
3108 		if (flag & FLAG_DATA_ACKED)
3109 			icsk->icsk_retransmits = 0;
3110 		if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3111 			tcp_reset_reno_sack(tp);
3112 		if (!tcp_try_undo_loss(sk)) {
3113 			tcp_moderate_cwnd(tp);
3114 			tcp_xmit_retransmit_queue(sk);
3115 			return;
3116 		}
3117 		if (icsk->icsk_ca_state != TCP_CA_Open)
3118 			return;
3119 		/* Loss is undone; fall through to processing in Open state. */
3120 	default:
3121 		if (tcp_is_reno(tp)) {
3122 			if (flag & FLAG_SND_UNA_ADVANCED)
3123 				tcp_reset_reno_sack(tp);
3124 			if (is_dupack)
3125 				tcp_add_reno_sack(sk);
3126 		}
3127 
3128 		if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3129 			tcp_try_undo_dsack(sk);
3130 
3131 		if (!tcp_time_to_recover(sk)) {
3132 			tcp_try_to_open(sk, flag);
3133 			return;
3134 		}
3135 
3136 		/* MTU probe failure: don't reduce cwnd */
3137 		if (icsk->icsk_ca_state < TCP_CA_CWR &&
3138 		    icsk->icsk_mtup.probe_size &&
3139 		    tp->snd_una == tp->mtu_probe.probe_seq_start) {
3140 			tcp_mtup_probe_failed(sk);
3141 			/* Restores the reduction we did in tcp_mtup_probe() */
3142 			tp->snd_cwnd++;
3143 			tcp_simple_retransmit(sk);
3144 			return;
3145 		}
3146 
3147 		/* Otherwise enter Recovery state */
3148 
3149 		if (tcp_is_reno(tp))
3150 			mib_idx = LINUX_MIB_TCPRENORECOVERY;
3151 		else
3152 			mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3153 
3154 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
3155 
3156 		tp->high_seq = tp->snd_nxt;
3157 		tp->prior_ssthresh = 0;
3158 		tp->undo_marker = tp->snd_una;
3159 		tp->undo_retrans = tp->retrans_out;
3160 
3161 		if (icsk->icsk_ca_state < TCP_CA_CWR) {
3162 			if (!(flag & FLAG_ECE))
3163 				tp->prior_ssthresh = tcp_current_ssthresh(sk);
3164 			tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3165 			TCP_ECN_queue_cwr(tp);
3166 		}
3167 
3168 		tp->bytes_acked = 0;
3169 		tp->snd_cwnd_cnt = 0;
3170 		tp->prior_cwnd = tp->snd_cwnd;
3171 		tp->prr_delivered = 0;
3172 		tp->prr_out = 0;
3173 		tcp_set_ca_state(sk, TCP_CA_Recovery);
3174 		fast_rexmit = 1;
3175 	}
3176 
3177 	if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3178 		tcp_update_scoreboard(sk, fast_rexmit);
3179 	tp->prr_delivered += newly_acked_sacked;
3180 	tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3181 	tcp_xmit_retransmit_queue(sk);
3182 }
3183 
3184 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3185 {
3186 	tcp_rtt_estimator(sk, seq_rtt);
3187 	tcp_set_rto(sk);
3188 	inet_csk(sk)->icsk_backoff = 0;
3189 }
3190 EXPORT_SYMBOL(tcp_valid_rtt_meas);
3191 
3192 /* Read draft-ietf-tcplw-high-performance before mucking
3193  * with this code. (Supersedes RFC1323)
3194  */
3195 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3196 {
3197 	/* RTTM Rule: A TSecr value received in a segment is used to
3198 	 * update the averaged RTT measurement only if the segment
3199 	 * acknowledges some new data, i.e., only if it advances the
3200 	 * left edge of the send window.
3201 	 *
3202 	 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3203 	 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3204 	 *
3205 	 * Changed: reset backoff as soon as we see the first valid sample.
3206 	 * If we do not, we get strongly overestimated rto. With timestamps
3207 	 * samples are accepted even from very old segments: f.e., when rtt=1
3208 	 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3209 	 * answer arrives rto becomes 120 seconds! If at least one of segments
3210 	 * in window is lost... Voila.	 			--ANK (010210)
3211 	 */
3212 	struct tcp_sock *tp = tcp_sk(sk);
3213 
3214 	tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3215 }
3216 
3217 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3218 {
3219 	/* We don't have a timestamp. Can only use
3220 	 * packets that are not retransmitted to determine
3221 	 * rtt estimates. Also, we must not reset the
3222 	 * backoff for rto until we get a non-retransmitted
3223 	 * packet. This allows us to deal with a situation
3224 	 * where the network delay has increased suddenly.
3225 	 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3226 	 */
3227 
3228 	if (flag & FLAG_RETRANS_DATA_ACKED)
3229 		return;
3230 
3231 	tcp_valid_rtt_meas(sk, seq_rtt);
3232 }
3233 
3234 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3235 				      const s32 seq_rtt)
3236 {
3237 	const struct tcp_sock *tp = tcp_sk(sk);
3238 	/* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3239 	if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3240 		tcp_ack_saw_tstamp(sk, flag);
3241 	else if (seq_rtt >= 0)
3242 		tcp_ack_no_tstamp(sk, seq_rtt, flag);
3243 }
3244 
3245 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3246 {
3247 	const struct inet_connection_sock *icsk = inet_csk(sk);
3248 	icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3249 	tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3250 }
3251 
3252 /* Restart timer after forward progress on connection.
3253  * RFC2988 recommends to restart timer to now+rto.
3254  */
3255 static void tcp_rearm_rto(struct sock *sk)
3256 {
3257 	const struct tcp_sock *tp = tcp_sk(sk);
3258 
3259 	if (!tp->packets_out) {
3260 		inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3261 	} else {
3262 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3263 					  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3264 	}
3265 }
3266 
3267 /* If we get here, the whole TSO packet has not been acked. */
3268 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3269 {
3270 	struct tcp_sock *tp = tcp_sk(sk);
3271 	u32 packets_acked;
3272 
3273 	BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3274 
3275 	packets_acked = tcp_skb_pcount(skb);
3276 	if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3277 		return 0;
3278 	packets_acked -= tcp_skb_pcount(skb);
3279 
3280 	if (packets_acked) {
3281 		BUG_ON(tcp_skb_pcount(skb) == 0);
3282 		BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3283 	}
3284 
3285 	return packets_acked;
3286 }
3287 
3288 /* Remove acknowledged frames from the retransmission queue. If our packet
3289  * is before the ack sequence we can discard it as it's confirmed to have
3290  * arrived at the other end.
3291  */
3292 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3293 			       u32 prior_snd_una)
3294 {
3295 	struct tcp_sock *tp = tcp_sk(sk);
3296 	const struct inet_connection_sock *icsk = inet_csk(sk);
3297 	struct sk_buff *skb;
3298 	u32 now = tcp_time_stamp;
3299 	int fully_acked = 1;
3300 	int flag = 0;
3301 	u32 pkts_acked = 0;
3302 	u32 reord = tp->packets_out;
3303 	u32 prior_sacked = tp->sacked_out;
3304 	s32 seq_rtt = -1;
3305 	s32 ca_seq_rtt = -1;
3306 	ktime_t last_ackt = net_invalid_timestamp();
3307 
3308 	while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3309 		struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3310 		u32 acked_pcount;
3311 		u8 sacked = scb->sacked;
3312 
3313 		/* Determine how many packets and what bytes were acked, tso and else */
3314 		if (after(scb->end_seq, tp->snd_una)) {
3315 			if (tcp_skb_pcount(skb) == 1 ||
3316 			    !after(tp->snd_una, scb->seq))
3317 				break;
3318 
3319 			acked_pcount = tcp_tso_acked(sk, skb);
3320 			if (!acked_pcount)
3321 				break;
3322 
3323 			fully_acked = 0;
3324 		} else {
3325 			acked_pcount = tcp_skb_pcount(skb);
3326 		}
3327 
3328 		if (sacked & TCPCB_RETRANS) {
3329 			if (sacked & TCPCB_SACKED_RETRANS)
3330 				tp->retrans_out -= acked_pcount;
3331 			flag |= FLAG_RETRANS_DATA_ACKED;
3332 			ca_seq_rtt = -1;
3333 			seq_rtt = -1;
3334 			if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3335 				flag |= FLAG_NONHEAD_RETRANS_ACKED;
3336 		} else {
3337 			ca_seq_rtt = now - scb->when;
3338 			last_ackt = skb->tstamp;
3339 			if (seq_rtt < 0) {
3340 				seq_rtt = ca_seq_rtt;
3341 			}
3342 			if (!(sacked & TCPCB_SACKED_ACKED))
3343 				reord = min(pkts_acked, reord);
3344 		}
3345 
3346 		if (sacked & TCPCB_SACKED_ACKED)
3347 			tp->sacked_out -= acked_pcount;
3348 		if (sacked & TCPCB_LOST)
3349 			tp->lost_out -= acked_pcount;
3350 
3351 		tp->packets_out -= acked_pcount;
3352 		pkts_acked += acked_pcount;
3353 
3354 		/* Initial outgoing SYN's get put onto the write_queue
3355 		 * just like anything else we transmit.  It is not
3356 		 * true data, and if we misinform our callers that
3357 		 * this ACK acks real data, we will erroneously exit
3358 		 * connection startup slow start one packet too
3359 		 * quickly.  This is severely frowned upon behavior.
3360 		 */
3361 		if (!(scb->tcp_flags & TCPHDR_SYN)) {
3362 			flag |= FLAG_DATA_ACKED;
3363 		} else {
3364 			flag |= FLAG_SYN_ACKED;
3365 			tp->retrans_stamp = 0;
3366 		}
3367 
3368 		if (!fully_acked)
3369 			break;
3370 
3371 		tcp_unlink_write_queue(skb, sk);
3372 		sk_wmem_free_skb(sk, skb);
3373 		tp->scoreboard_skb_hint = NULL;
3374 		if (skb == tp->retransmit_skb_hint)
3375 			tp->retransmit_skb_hint = NULL;
3376 		if (skb == tp->lost_skb_hint)
3377 			tp->lost_skb_hint = NULL;
3378 	}
3379 
3380 	if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3381 		tp->snd_up = tp->snd_una;
3382 
3383 	if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3384 		flag |= FLAG_SACK_RENEGING;
3385 
3386 	if (flag & FLAG_ACKED) {
3387 		const struct tcp_congestion_ops *ca_ops
3388 			= inet_csk(sk)->icsk_ca_ops;
3389 
3390 		if (unlikely(icsk->icsk_mtup.probe_size &&
3391 			     !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3392 			tcp_mtup_probe_success(sk);
3393 		}
3394 
3395 		tcp_ack_update_rtt(sk, flag, seq_rtt);
3396 		tcp_rearm_rto(sk);
3397 
3398 		if (tcp_is_reno(tp)) {
3399 			tcp_remove_reno_sacks(sk, pkts_acked);
3400 		} else {
3401 			int delta;
3402 
3403 			/* Non-retransmitted hole got filled? That's reordering */
3404 			if (reord < prior_fackets)
3405 				tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3406 
3407 			delta = tcp_is_fack(tp) ? pkts_acked :
3408 						  prior_sacked - tp->sacked_out;
3409 			tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3410 		}
3411 
3412 		tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3413 
3414 		if (ca_ops->pkts_acked) {
3415 			s32 rtt_us = -1;
3416 
3417 			/* Is the ACK triggering packet unambiguous? */
3418 			if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3419 				/* High resolution needed and available? */
3420 				if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3421 				    !ktime_equal(last_ackt,
3422 						 net_invalid_timestamp()))
3423 					rtt_us = ktime_us_delta(ktime_get_real(),
3424 								last_ackt);
3425 				else if (ca_seq_rtt >= 0)
3426 					rtt_us = jiffies_to_usecs(ca_seq_rtt);
3427 			}
3428 
3429 			ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3430 		}
3431 	}
3432 
3433 #if FASTRETRANS_DEBUG > 0
3434 	WARN_ON((int)tp->sacked_out < 0);
3435 	WARN_ON((int)tp->lost_out < 0);
3436 	WARN_ON((int)tp->retrans_out < 0);
3437 	if (!tp->packets_out && tcp_is_sack(tp)) {
3438 		icsk = inet_csk(sk);
3439 		if (tp->lost_out) {
3440 			printk(KERN_DEBUG "Leak l=%u %d\n",
3441 			       tp->lost_out, icsk->icsk_ca_state);
3442 			tp->lost_out = 0;
3443 		}
3444 		if (tp->sacked_out) {
3445 			printk(KERN_DEBUG "Leak s=%u %d\n",
3446 			       tp->sacked_out, icsk->icsk_ca_state);
3447 			tp->sacked_out = 0;
3448 		}
3449 		if (tp->retrans_out) {
3450 			printk(KERN_DEBUG "Leak r=%u %d\n",
3451 			       tp->retrans_out, icsk->icsk_ca_state);
3452 			tp->retrans_out = 0;
3453 		}
3454 	}
3455 #endif
3456 	return flag;
3457 }
3458 
3459 static void tcp_ack_probe(struct sock *sk)
3460 {
3461 	const struct tcp_sock *tp = tcp_sk(sk);
3462 	struct inet_connection_sock *icsk = inet_csk(sk);
3463 
3464 	/* Was it a usable window open? */
3465 
3466 	if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3467 		icsk->icsk_backoff = 0;
3468 		inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3469 		/* Socket must be waked up by subsequent tcp_data_snd_check().
3470 		 * This function is not for random using!
3471 		 */
3472 	} else {
3473 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3474 					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3475 					  TCP_RTO_MAX);
3476 	}
3477 }
3478 
3479 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3480 {
3481 	return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3482 		inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3483 }
3484 
3485 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3486 {
3487 	const struct tcp_sock *tp = tcp_sk(sk);
3488 	return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3489 		!((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3490 }
3491 
3492 /* Check that window update is acceptable.
3493  * The function assumes that snd_una<=ack<=snd_next.
3494  */
3495 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3496 					const u32 ack, const u32 ack_seq,
3497 					const u32 nwin)
3498 {
3499 	return	after(ack, tp->snd_una) ||
3500 		after(ack_seq, tp->snd_wl1) ||
3501 		(ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3502 }
3503 
3504 /* Update our send window.
3505  *
3506  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3507  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3508  */
3509 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3510 				 u32 ack_seq)
3511 {
3512 	struct tcp_sock *tp = tcp_sk(sk);
3513 	int flag = 0;
3514 	u32 nwin = ntohs(tcp_hdr(skb)->window);
3515 
3516 	if (likely(!tcp_hdr(skb)->syn))
3517 		nwin <<= tp->rx_opt.snd_wscale;
3518 
3519 	if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3520 		flag |= FLAG_WIN_UPDATE;
3521 		tcp_update_wl(tp, ack_seq);
3522 
3523 		if (tp->snd_wnd != nwin) {
3524 			tp->snd_wnd = nwin;
3525 
3526 			/* Note, it is the only place, where
3527 			 * fast path is recovered for sending TCP.
3528 			 */
3529 			tp->pred_flags = 0;
3530 			tcp_fast_path_check(sk);
3531 
3532 			if (nwin > tp->max_window) {
3533 				tp->max_window = nwin;
3534 				tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3535 			}
3536 		}
3537 	}
3538 
3539 	tp->snd_una = ack;
3540 
3541 	return flag;
3542 }
3543 
3544 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3545  * continue in congestion avoidance.
3546  */
3547 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3548 {
3549 	tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3550 	tp->snd_cwnd_cnt = 0;
3551 	tp->bytes_acked = 0;
3552 	TCP_ECN_queue_cwr(tp);
3553 	tcp_moderate_cwnd(tp);
3554 }
3555 
3556 /* A conservative spurious RTO response algorithm: reduce cwnd using
3557  * rate halving and continue in congestion avoidance.
3558  */
3559 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3560 {
3561 	tcp_enter_cwr(sk, 0);
3562 }
3563 
3564 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3565 {
3566 	if (flag & FLAG_ECE)
3567 		tcp_ratehalving_spur_to_response(sk);
3568 	else
3569 		tcp_undo_cwr(sk, true);
3570 }
3571 
3572 /* F-RTO spurious RTO detection algorithm (RFC4138)
3573  *
3574  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3575  * comments). State (ACK number) is kept in frto_counter. When ACK advances
3576  * window (but not to or beyond highest sequence sent before RTO):
3577  *   On First ACK,  send two new segments out.
3578  *   On Second ACK, RTO was likely spurious. Do spurious response (response
3579  *                  algorithm is not part of the F-RTO detection algorithm
3580  *                  given in RFC4138 but can be selected separately).
3581  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3582  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3583  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3584  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3585  *
3586  * Rationale: if the RTO was spurious, new ACKs should arrive from the
3587  * original window even after we transmit two new data segments.
3588  *
3589  * SACK version:
3590  *   on first step, wait until first cumulative ACK arrives, then move to
3591  *   the second step. In second step, the next ACK decides.
3592  *
3593  * F-RTO is implemented (mainly) in four functions:
3594  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3595  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3596  *     called when tcp_use_frto() showed green light
3597  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3598  *   - tcp_enter_frto_loss() is called if there is not enough evidence
3599  *     to prove that the RTO is indeed spurious. It transfers the control
3600  *     from F-RTO to the conventional RTO recovery
3601  */
3602 static int tcp_process_frto(struct sock *sk, int flag)
3603 {
3604 	struct tcp_sock *tp = tcp_sk(sk);
3605 
3606 	tcp_verify_left_out(tp);
3607 
3608 	/* Duplicate the behavior from Loss state (fastretrans_alert) */
3609 	if (flag & FLAG_DATA_ACKED)
3610 		inet_csk(sk)->icsk_retransmits = 0;
3611 
3612 	if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3613 	    ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3614 		tp->undo_marker = 0;
3615 
3616 	if (!before(tp->snd_una, tp->frto_highmark)) {
3617 		tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3618 		return 1;
3619 	}
3620 
3621 	if (!tcp_is_sackfrto(tp)) {
3622 		/* RFC4138 shortcoming in step 2; should also have case c):
3623 		 * ACK isn't duplicate nor advances window, e.g., opposite dir
3624 		 * data, winupdate
3625 		 */
3626 		if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3627 			return 1;
3628 
3629 		if (!(flag & FLAG_DATA_ACKED)) {
3630 			tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3631 					    flag);
3632 			return 1;
3633 		}
3634 	} else {
3635 		if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3636 			/* Prevent sending of new data. */
3637 			tp->snd_cwnd = min(tp->snd_cwnd,
3638 					   tcp_packets_in_flight(tp));
3639 			return 1;
3640 		}
3641 
3642 		if ((tp->frto_counter >= 2) &&
3643 		    (!(flag & FLAG_FORWARD_PROGRESS) ||
3644 		     ((flag & FLAG_DATA_SACKED) &&
3645 		      !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3646 			/* RFC4138 shortcoming (see comment above) */
3647 			if (!(flag & FLAG_FORWARD_PROGRESS) &&
3648 			    (flag & FLAG_NOT_DUP))
3649 				return 1;
3650 
3651 			tcp_enter_frto_loss(sk, 3, flag);
3652 			return 1;
3653 		}
3654 	}
3655 
3656 	if (tp->frto_counter == 1) {
3657 		/* tcp_may_send_now needs to see updated state */
3658 		tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3659 		tp->frto_counter = 2;
3660 
3661 		if (!tcp_may_send_now(sk))
3662 			tcp_enter_frto_loss(sk, 2, flag);
3663 
3664 		return 1;
3665 	} else {
3666 		switch (sysctl_tcp_frto_response) {
3667 		case 2:
3668 			tcp_undo_spur_to_response(sk, flag);
3669 			break;
3670 		case 1:
3671 			tcp_conservative_spur_to_response(tp);
3672 			break;
3673 		default:
3674 			tcp_ratehalving_spur_to_response(sk);
3675 			break;
3676 		}
3677 		tp->frto_counter = 0;
3678 		tp->undo_marker = 0;
3679 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3680 	}
3681 	return 0;
3682 }
3683 
3684 /* This routine deals with incoming acks, but not outgoing ones. */
3685 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3686 {
3687 	struct inet_connection_sock *icsk = inet_csk(sk);
3688 	struct tcp_sock *tp = tcp_sk(sk);
3689 	u32 prior_snd_una = tp->snd_una;
3690 	u32 ack_seq = TCP_SKB_CB(skb)->seq;
3691 	u32 ack = TCP_SKB_CB(skb)->ack_seq;
3692 	bool is_dupack = false;
3693 	u32 prior_in_flight;
3694 	u32 prior_fackets;
3695 	int prior_packets;
3696 	int prior_sacked = tp->sacked_out;
3697 	int pkts_acked = 0;
3698 	int newly_acked_sacked = 0;
3699 	int frto_cwnd = 0;
3700 
3701 	/* If the ack is older than previous acks
3702 	 * then we can probably ignore it.
3703 	 */
3704 	if (before(ack, prior_snd_una))
3705 		goto old_ack;
3706 
3707 	/* If the ack includes data we haven't sent yet, discard
3708 	 * this segment (RFC793 Section 3.9).
3709 	 */
3710 	if (after(ack, tp->snd_nxt))
3711 		goto invalid_ack;
3712 
3713 	if (after(ack, prior_snd_una))
3714 		flag |= FLAG_SND_UNA_ADVANCED;
3715 
3716 	if (sysctl_tcp_abc) {
3717 		if (icsk->icsk_ca_state < TCP_CA_CWR)
3718 			tp->bytes_acked += ack - prior_snd_una;
3719 		else if (icsk->icsk_ca_state == TCP_CA_Loss)
3720 			/* we assume just one segment left network */
3721 			tp->bytes_acked += min(ack - prior_snd_una,
3722 					       tp->mss_cache);
3723 	}
3724 
3725 	prior_fackets = tp->fackets_out;
3726 	prior_in_flight = tcp_packets_in_flight(tp);
3727 
3728 	if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3729 		/* Window is constant, pure forward advance.
3730 		 * No more checks are required.
3731 		 * Note, we use the fact that SND.UNA>=SND.WL2.
3732 		 */
3733 		tcp_update_wl(tp, ack_seq);
3734 		tp->snd_una = ack;
3735 		flag |= FLAG_WIN_UPDATE;
3736 
3737 		tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3738 
3739 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3740 	} else {
3741 		if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3742 			flag |= FLAG_DATA;
3743 		else
3744 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3745 
3746 		flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3747 
3748 		if (TCP_SKB_CB(skb)->sacked)
3749 			flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3750 
3751 		if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3752 			flag |= FLAG_ECE;
3753 
3754 		tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3755 	}
3756 
3757 	/* We passed data and got it acked, remove any soft error
3758 	 * log. Something worked...
3759 	 */
3760 	sk->sk_err_soft = 0;
3761 	icsk->icsk_probes_out = 0;
3762 	tp->rcv_tstamp = tcp_time_stamp;
3763 	prior_packets = tp->packets_out;
3764 	if (!prior_packets)
3765 		goto no_queue;
3766 
3767 	/* See if we can take anything off of the retransmit queue. */
3768 	flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3769 
3770 	pkts_acked = prior_packets - tp->packets_out;
3771 	newly_acked_sacked = (prior_packets - prior_sacked) -
3772 			     (tp->packets_out - tp->sacked_out);
3773 
3774 	if (tp->frto_counter)
3775 		frto_cwnd = tcp_process_frto(sk, flag);
3776 	/* Guarantee sacktag reordering detection against wrap-arounds */
3777 	if (before(tp->frto_highmark, tp->snd_una))
3778 		tp->frto_highmark = 0;
3779 
3780 	if (tcp_ack_is_dubious(sk, flag)) {
3781 		/* Advance CWND, if state allows this. */
3782 		if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3783 		    tcp_may_raise_cwnd(sk, flag))
3784 			tcp_cong_avoid(sk, ack, prior_in_flight);
3785 		is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3786 		tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3787 				      is_dupack, flag);
3788 	} else {
3789 		if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3790 			tcp_cong_avoid(sk, ack, prior_in_flight);
3791 	}
3792 
3793 	if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3794 		dst_confirm(__sk_dst_get(sk));
3795 
3796 	return 1;
3797 
3798 no_queue:
3799 	/* If data was DSACKed, see if we can undo a cwnd reduction. */
3800 	if (flag & FLAG_DSACKING_ACK)
3801 		tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3802 				      is_dupack, flag);
3803 	/* If this ack opens up a zero window, clear backoff.  It was
3804 	 * being used to time the probes, and is probably far higher than
3805 	 * it needs to be for normal retransmission.
3806 	 */
3807 	if (tcp_send_head(sk))
3808 		tcp_ack_probe(sk);
3809 	return 1;
3810 
3811 invalid_ack:
3812 	SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3813 	return -1;
3814 
3815 old_ack:
3816 	/* If data was SACKed, tag it and see if we should send more data.
3817 	 * If data was DSACKed, see if we can undo a cwnd reduction.
3818 	 */
3819 	if (TCP_SKB_CB(skb)->sacked) {
3820 		flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3821 		newly_acked_sacked = tp->sacked_out - prior_sacked;
3822 		tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3823 				      is_dupack, flag);
3824 	}
3825 
3826 	SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3827 	return 0;
3828 }
3829 
3830 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3831  * But, this can also be called on packets in the established flow when
3832  * the fast version below fails.
3833  */
3834 void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3835 		       const u8 **hvpp, int estab)
3836 {
3837 	const unsigned char *ptr;
3838 	const struct tcphdr *th = tcp_hdr(skb);
3839 	int length = (th->doff * 4) - sizeof(struct tcphdr);
3840 
3841 	ptr = (const unsigned char *)(th + 1);
3842 	opt_rx->saw_tstamp = 0;
3843 
3844 	while (length > 0) {
3845 		int opcode = *ptr++;
3846 		int opsize;
3847 
3848 		switch (opcode) {
3849 		case TCPOPT_EOL:
3850 			return;
3851 		case TCPOPT_NOP:	/* Ref: RFC 793 section 3.1 */
3852 			length--;
3853 			continue;
3854 		default:
3855 			opsize = *ptr++;
3856 			if (opsize < 2) /* "silly options" */
3857 				return;
3858 			if (opsize > length)
3859 				return;	/* don't parse partial options */
3860 			switch (opcode) {
3861 			case TCPOPT_MSS:
3862 				if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3863 					u16 in_mss = get_unaligned_be16(ptr);
3864 					if (in_mss) {
3865 						if (opt_rx->user_mss &&
3866 						    opt_rx->user_mss < in_mss)
3867 							in_mss = opt_rx->user_mss;
3868 						opt_rx->mss_clamp = in_mss;
3869 					}
3870 				}
3871 				break;
3872 			case TCPOPT_WINDOW:
3873 				if (opsize == TCPOLEN_WINDOW && th->syn &&
3874 				    !estab && sysctl_tcp_window_scaling) {
3875 					__u8 snd_wscale = *(__u8 *)ptr;
3876 					opt_rx->wscale_ok = 1;
3877 					if (snd_wscale > 14) {
3878 						if (net_ratelimit())
3879 							pr_info("%s: Illegal window scaling value %d >14 received\n",
3880 								__func__,
3881 								snd_wscale);
3882 						snd_wscale = 14;
3883 					}
3884 					opt_rx->snd_wscale = snd_wscale;
3885 				}
3886 				break;
3887 			case TCPOPT_TIMESTAMP:
3888 				if ((opsize == TCPOLEN_TIMESTAMP) &&
3889 				    ((estab && opt_rx->tstamp_ok) ||
3890 				     (!estab && sysctl_tcp_timestamps))) {
3891 					opt_rx->saw_tstamp = 1;
3892 					opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3893 					opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3894 				}
3895 				break;
3896 			case TCPOPT_SACK_PERM:
3897 				if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3898 				    !estab && sysctl_tcp_sack) {
3899 					opt_rx->sack_ok = TCP_SACK_SEEN;
3900 					tcp_sack_reset(opt_rx);
3901 				}
3902 				break;
3903 
3904 			case TCPOPT_SACK:
3905 				if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3906 				   !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3907 				   opt_rx->sack_ok) {
3908 					TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3909 				}
3910 				break;
3911 #ifdef CONFIG_TCP_MD5SIG
3912 			case TCPOPT_MD5SIG:
3913 				/*
3914 				 * The MD5 Hash has already been
3915 				 * checked (see tcp_v{4,6}_do_rcv()).
3916 				 */
3917 				break;
3918 #endif
3919 			case TCPOPT_COOKIE:
3920 				/* This option is variable length.
3921 				 */
3922 				switch (opsize) {
3923 				case TCPOLEN_COOKIE_BASE:
3924 					/* not yet implemented */
3925 					break;
3926 				case TCPOLEN_COOKIE_PAIR:
3927 					/* not yet implemented */
3928 					break;
3929 				case TCPOLEN_COOKIE_MIN+0:
3930 				case TCPOLEN_COOKIE_MIN+2:
3931 				case TCPOLEN_COOKIE_MIN+4:
3932 				case TCPOLEN_COOKIE_MIN+6:
3933 				case TCPOLEN_COOKIE_MAX:
3934 					/* 16-bit multiple */
3935 					opt_rx->cookie_plus = opsize;
3936 					*hvpp = ptr;
3937 					break;
3938 				default:
3939 					/* ignore option */
3940 					break;
3941 				}
3942 				break;
3943 			}
3944 
3945 			ptr += opsize-2;
3946 			length -= opsize;
3947 		}
3948 	}
3949 }
3950 EXPORT_SYMBOL(tcp_parse_options);
3951 
3952 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3953 {
3954 	const __be32 *ptr = (const __be32 *)(th + 1);
3955 
3956 	if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3957 			  | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3958 		tp->rx_opt.saw_tstamp = 1;
3959 		++ptr;
3960 		tp->rx_opt.rcv_tsval = ntohl(*ptr);
3961 		++ptr;
3962 		tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3963 		return 1;
3964 	}
3965 	return 0;
3966 }
3967 
3968 /* Fast parse options. This hopes to only see timestamps.
3969  * If it is wrong it falls back on tcp_parse_options().
3970  */
3971 static int tcp_fast_parse_options(const struct sk_buff *skb,
3972 				  const struct tcphdr *th,
3973 				  struct tcp_sock *tp, const u8 **hvpp)
3974 {
3975 	/* In the spirit of fast parsing, compare doff directly to constant
3976 	 * values.  Because equality is used, short doff can be ignored here.
3977 	 */
3978 	if (th->doff == (sizeof(*th) / 4)) {
3979 		tp->rx_opt.saw_tstamp = 0;
3980 		return 0;
3981 	} else if (tp->rx_opt.tstamp_ok &&
3982 		   th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3983 		if (tcp_parse_aligned_timestamp(tp, th))
3984 			return 1;
3985 	}
3986 	tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3987 	return 1;
3988 }
3989 
3990 #ifdef CONFIG_TCP_MD5SIG
3991 /*
3992  * Parse MD5 Signature option
3993  */
3994 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3995 {
3996 	int length = (th->doff << 2) - sizeof(*th);
3997 	const u8 *ptr = (const u8 *)(th + 1);
3998 
3999 	/* If the TCP option is too short, we can short cut */
4000 	if (length < TCPOLEN_MD5SIG)
4001 		return NULL;
4002 
4003 	while (length > 0) {
4004 		int opcode = *ptr++;
4005 		int opsize;
4006 
4007 		switch(opcode) {
4008 		case TCPOPT_EOL:
4009 			return NULL;
4010 		case TCPOPT_NOP:
4011 			length--;
4012 			continue;
4013 		default:
4014 			opsize = *ptr++;
4015 			if (opsize < 2 || opsize > length)
4016 				return NULL;
4017 			if (opcode == TCPOPT_MD5SIG)
4018 				return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4019 		}
4020 		ptr += opsize - 2;
4021 		length -= opsize;
4022 	}
4023 	return NULL;
4024 }
4025 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4026 #endif
4027 
4028 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
4029 {
4030 	tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
4031 	tp->rx_opt.ts_recent_stamp = get_seconds();
4032 }
4033 
4034 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
4035 {
4036 	if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
4037 		/* PAWS bug workaround wrt. ACK frames, the PAWS discard
4038 		 * extra check below makes sure this can only happen
4039 		 * for pure ACK frames.  -DaveM
4040 		 *
4041 		 * Not only, also it occurs for expired timestamps.
4042 		 */
4043 
4044 		if (tcp_paws_check(&tp->rx_opt, 0))
4045 			tcp_store_ts_recent(tp);
4046 	}
4047 }
4048 
4049 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4050  *
4051  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4052  * it can pass through stack. So, the following predicate verifies that
4053  * this segment is not used for anything but congestion avoidance or
4054  * fast retransmit. Moreover, we even are able to eliminate most of such
4055  * second order effects, if we apply some small "replay" window (~RTO)
4056  * to timestamp space.
4057  *
4058  * All these measures still do not guarantee that we reject wrapped ACKs
4059  * on networks with high bandwidth, when sequence space is recycled fastly,
4060  * but it guarantees that such events will be very rare and do not affect
4061  * connection seriously. This doesn't look nice, but alas, PAWS is really
4062  * buggy extension.
4063  *
4064  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4065  * states that events when retransmit arrives after original data are rare.
4066  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4067  * the biggest problem on large power networks even with minor reordering.
4068  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4069  * up to bandwidth of 18Gigabit/sec. 8) ]
4070  */
4071 
4072 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4073 {
4074 	const struct tcp_sock *tp = tcp_sk(sk);
4075 	const struct tcphdr *th = tcp_hdr(skb);
4076 	u32 seq = TCP_SKB_CB(skb)->seq;
4077 	u32 ack = TCP_SKB_CB(skb)->ack_seq;
4078 
4079 	return (/* 1. Pure ACK with correct sequence number. */
4080 		(th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4081 
4082 		/* 2. ... and duplicate ACK. */
4083 		ack == tp->snd_una &&
4084 
4085 		/* 3. ... and does not update window. */
4086 		!tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4087 
4088 		/* 4. ... and sits in replay window. */
4089 		(s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4090 }
4091 
4092 static inline int tcp_paws_discard(const struct sock *sk,
4093 				   const struct sk_buff *skb)
4094 {
4095 	const struct tcp_sock *tp = tcp_sk(sk);
4096 
4097 	return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4098 	       !tcp_disordered_ack(sk, skb);
4099 }
4100 
4101 /* Check segment sequence number for validity.
4102  *
4103  * Segment controls are considered valid, if the segment
4104  * fits to the window after truncation to the window. Acceptability
4105  * of data (and SYN, FIN, of course) is checked separately.
4106  * See tcp_data_queue(), for example.
4107  *
4108  * Also, controls (RST is main one) are accepted using RCV.WUP instead
4109  * of RCV.NXT. Peer still did not advance his SND.UNA when we
4110  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4111  * (borrowed from freebsd)
4112  */
4113 
4114 static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4115 {
4116 	return	!before(end_seq, tp->rcv_wup) &&
4117 		!after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4118 }
4119 
4120 /* When we get a reset we do this. */
4121 static void tcp_reset(struct sock *sk)
4122 {
4123 	/* We want the right error as BSD sees it (and indeed as we do). */
4124 	switch (sk->sk_state) {
4125 	case TCP_SYN_SENT:
4126 		sk->sk_err = ECONNREFUSED;
4127 		break;
4128 	case TCP_CLOSE_WAIT:
4129 		sk->sk_err = EPIPE;
4130 		break;
4131 	case TCP_CLOSE:
4132 		return;
4133 	default:
4134 		sk->sk_err = ECONNRESET;
4135 	}
4136 	/* This barrier is coupled with smp_rmb() in tcp_poll() */
4137 	smp_wmb();
4138 
4139 	if (!sock_flag(sk, SOCK_DEAD))
4140 		sk->sk_error_report(sk);
4141 
4142 	tcp_done(sk);
4143 }
4144 
4145 /*
4146  * 	Process the FIN bit. This now behaves as it is supposed to work
4147  *	and the FIN takes effect when it is validly part of sequence
4148  *	space. Not before when we get holes.
4149  *
4150  *	If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4151  *	(and thence onto LAST-ACK and finally, CLOSE, we never enter
4152  *	TIME-WAIT)
4153  *
4154  *	If we are in FINWAIT-1, a received FIN indicates simultaneous
4155  *	close and we go into CLOSING (and later onto TIME-WAIT)
4156  *
4157  *	If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4158  */
4159 static void tcp_fin(struct sock *sk)
4160 {
4161 	struct tcp_sock *tp = tcp_sk(sk);
4162 
4163 	inet_csk_schedule_ack(sk);
4164 
4165 	sk->sk_shutdown |= RCV_SHUTDOWN;
4166 	sock_set_flag(sk, SOCK_DONE);
4167 
4168 	switch (sk->sk_state) {
4169 	case TCP_SYN_RECV:
4170 	case TCP_ESTABLISHED:
4171 		/* Move to CLOSE_WAIT */
4172 		tcp_set_state(sk, TCP_CLOSE_WAIT);
4173 		inet_csk(sk)->icsk_ack.pingpong = 1;
4174 		break;
4175 
4176 	case TCP_CLOSE_WAIT:
4177 	case TCP_CLOSING:
4178 		/* Received a retransmission of the FIN, do
4179 		 * nothing.
4180 		 */
4181 		break;
4182 	case TCP_LAST_ACK:
4183 		/* RFC793: Remain in the LAST-ACK state. */
4184 		break;
4185 
4186 	case TCP_FIN_WAIT1:
4187 		/* This case occurs when a simultaneous close
4188 		 * happens, we must ack the received FIN and
4189 		 * enter the CLOSING state.
4190 		 */
4191 		tcp_send_ack(sk);
4192 		tcp_set_state(sk, TCP_CLOSING);
4193 		break;
4194 	case TCP_FIN_WAIT2:
4195 		/* Received a FIN -- send ACK and enter TIME_WAIT. */
4196 		tcp_send_ack(sk);
4197 		tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4198 		break;
4199 	default:
4200 		/* Only TCP_LISTEN and TCP_CLOSE are left, in these
4201 		 * cases we should never reach this piece of code.
4202 		 */
4203 		pr_err("%s: Impossible, sk->sk_state=%d\n",
4204 		       __func__, sk->sk_state);
4205 		break;
4206 	}
4207 
4208 	/* It _is_ possible, that we have something out-of-order _after_ FIN.
4209 	 * Probably, we should reset in this case. For now drop them.
4210 	 */
4211 	__skb_queue_purge(&tp->out_of_order_queue);
4212 	if (tcp_is_sack(tp))
4213 		tcp_sack_reset(&tp->rx_opt);
4214 	sk_mem_reclaim(sk);
4215 
4216 	if (!sock_flag(sk, SOCK_DEAD)) {
4217 		sk->sk_state_change(sk);
4218 
4219 		/* Do not send POLL_HUP for half duplex close. */
4220 		if (sk->sk_shutdown == SHUTDOWN_MASK ||
4221 		    sk->sk_state == TCP_CLOSE)
4222 			sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4223 		else
4224 			sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4225 	}
4226 }
4227 
4228 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4229 				  u32 end_seq)
4230 {
4231 	if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4232 		if (before(seq, sp->start_seq))
4233 			sp->start_seq = seq;
4234 		if (after(end_seq, sp->end_seq))
4235 			sp->end_seq = end_seq;
4236 		return 1;
4237 	}
4238 	return 0;
4239 }
4240 
4241 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4242 {
4243 	struct tcp_sock *tp = tcp_sk(sk);
4244 
4245 	if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4246 		int mib_idx;
4247 
4248 		if (before(seq, tp->rcv_nxt))
4249 			mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4250 		else
4251 			mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4252 
4253 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
4254 
4255 		tp->rx_opt.dsack = 1;
4256 		tp->duplicate_sack[0].start_seq = seq;
4257 		tp->duplicate_sack[0].end_seq = end_seq;
4258 	}
4259 }
4260 
4261 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4262 {
4263 	struct tcp_sock *tp = tcp_sk(sk);
4264 
4265 	if (!tp->rx_opt.dsack)
4266 		tcp_dsack_set(sk, seq, end_seq);
4267 	else
4268 		tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4269 }
4270 
4271 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4272 {
4273 	struct tcp_sock *tp = tcp_sk(sk);
4274 
4275 	if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4276 	    before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4277 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4278 		tcp_enter_quickack_mode(sk);
4279 
4280 		if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4281 			u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4282 
4283 			if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4284 				end_seq = tp->rcv_nxt;
4285 			tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4286 		}
4287 	}
4288 
4289 	tcp_send_ack(sk);
4290 }
4291 
4292 /* These routines update the SACK block as out-of-order packets arrive or
4293  * in-order packets close up the sequence space.
4294  */
4295 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4296 {
4297 	int this_sack;
4298 	struct tcp_sack_block *sp = &tp->selective_acks[0];
4299 	struct tcp_sack_block *swalk = sp + 1;
4300 
4301 	/* See if the recent change to the first SACK eats into
4302 	 * or hits the sequence space of other SACK blocks, if so coalesce.
4303 	 */
4304 	for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4305 		if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4306 			int i;
4307 
4308 			/* Zap SWALK, by moving every further SACK up by one slot.
4309 			 * Decrease num_sacks.
4310 			 */
4311 			tp->rx_opt.num_sacks--;
4312 			for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4313 				sp[i] = sp[i + 1];
4314 			continue;
4315 		}
4316 		this_sack++, swalk++;
4317 	}
4318 }
4319 
4320 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4321 {
4322 	struct tcp_sock *tp = tcp_sk(sk);
4323 	struct tcp_sack_block *sp = &tp->selective_acks[0];
4324 	int cur_sacks = tp->rx_opt.num_sacks;
4325 	int this_sack;
4326 
4327 	if (!cur_sacks)
4328 		goto new_sack;
4329 
4330 	for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4331 		if (tcp_sack_extend(sp, seq, end_seq)) {
4332 			/* Rotate this_sack to the first one. */
4333 			for (; this_sack > 0; this_sack--, sp--)
4334 				swap(*sp, *(sp - 1));
4335 			if (cur_sacks > 1)
4336 				tcp_sack_maybe_coalesce(tp);
4337 			return;
4338 		}
4339 	}
4340 
4341 	/* Could not find an adjacent existing SACK, build a new one,
4342 	 * put it at the front, and shift everyone else down.  We
4343 	 * always know there is at least one SACK present already here.
4344 	 *
4345 	 * If the sack array is full, forget about the last one.
4346 	 */
4347 	if (this_sack >= TCP_NUM_SACKS) {
4348 		this_sack--;
4349 		tp->rx_opt.num_sacks--;
4350 		sp--;
4351 	}
4352 	for (; this_sack > 0; this_sack--, sp--)
4353 		*sp = *(sp - 1);
4354 
4355 new_sack:
4356 	/* Build the new head SACK, and we're done. */
4357 	sp->start_seq = seq;
4358 	sp->end_seq = end_seq;
4359 	tp->rx_opt.num_sacks++;
4360 }
4361 
4362 /* RCV.NXT advances, some SACKs should be eaten. */
4363 
4364 static void tcp_sack_remove(struct tcp_sock *tp)
4365 {
4366 	struct tcp_sack_block *sp = &tp->selective_acks[0];
4367 	int num_sacks = tp->rx_opt.num_sacks;
4368 	int this_sack;
4369 
4370 	/* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4371 	if (skb_queue_empty(&tp->out_of_order_queue)) {
4372 		tp->rx_opt.num_sacks = 0;
4373 		return;
4374 	}
4375 
4376 	for (this_sack = 0; this_sack < num_sacks;) {
4377 		/* Check if the start of the sack is covered by RCV.NXT. */
4378 		if (!before(tp->rcv_nxt, sp->start_seq)) {
4379 			int i;
4380 
4381 			/* RCV.NXT must cover all the block! */
4382 			WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4383 
4384 			/* Zap this SACK, by moving forward any other SACKS. */
4385 			for (i=this_sack+1; i < num_sacks; i++)
4386 				tp->selective_acks[i-1] = tp->selective_acks[i];
4387 			num_sacks--;
4388 			continue;
4389 		}
4390 		this_sack++;
4391 		sp++;
4392 	}
4393 	tp->rx_opt.num_sacks = num_sacks;
4394 }
4395 
4396 /* This one checks to see if we can put data from the
4397  * out_of_order queue into the receive_queue.
4398  */
4399 static void tcp_ofo_queue(struct sock *sk)
4400 {
4401 	struct tcp_sock *tp = tcp_sk(sk);
4402 	__u32 dsack_high = tp->rcv_nxt;
4403 	struct sk_buff *skb;
4404 
4405 	while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4406 		if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4407 			break;
4408 
4409 		if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4410 			__u32 dsack = dsack_high;
4411 			if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4412 				dsack_high = TCP_SKB_CB(skb)->end_seq;
4413 			tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4414 		}
4415 
4416 		if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4417 			SOCK_DEBUG(sk, "ofo packet was already received\n");
4418 			__skb_unlink(skb, &tp->out_of_order_queue);
4419 			__kfree_skb(skb);
4420 			continue;
4421 		}
4422 		SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4423 			   tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4424 			   TCP_SKB_CB(skb)->end_seq);
4425 
4426 		__skb_unlink(skb, &tp->out_of_order_queue);
4427 		__skb_queue_tail(&sk->sk_receive_queue, skb);
4428 		tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4429 		if (tcp_hdr(skb)->fin)
4430 			tcp_fin(sk);
4431 	}
4432 }
4433 
4434 static int tcp_prune_ofo_queue(struct sock *sk);
4435 static int tcp_prune_queue(struct sock *sk);
4436 
4437 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4438 {
4439 	if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4440 	    !sk_rmem_schedule(sk, size)) {
4441 
4442 		if (tcp_prune_queue(sk) < 0)
4443 			return -1;
4444 
4445 		if (!sk_rmem_schedule(sk, size)) {
4446 			if (!tcp_prune_ofo_queue(sk))
4447 				return -1;
4448 
4449 			if (!sk_rmem_schedule(sk, size))
4450 				return -1;
4451 		}
4452 	}
4453 	return 0;
4454 }
4455 
4456 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4457 {
4458 	struct tcp_sock *tp = tcp_sk(sk);
4459 	struct sk_buff *skb1;
4460 	u32 seq, end_seq;
4461 
4462 	TCP_ECN_check_ce(tp, skb);
4463 
4464 	if (tcp_try_rmem_schedule(sk, skb->truesize)) {
4465 		/* TODO: should increment a counter */
4466 		__kfree_skb(skb);
4467 		return;
4468 	}
4469 
4470 	/* Disable header prediction. */
4471 	tp->pred_flags = 0;
4472 	inet_csk_schedule_ack(sk);
4473 
4474 	SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4475 		   tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4476 
4477 	skb1 = skb_peek_tail(&tp->out_of_order_queue);
4478 	if (!skb1) {
4479 		/* Initial out of order segment, build 1 SACK. */
4480 		if (tcp_is_sack(tp)) {
4481 			tp->rx_opt.num_sacks = 1;
4482 			tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4483 			tp->selective_acks[0].end_seq =
4484 						TCP_SKB_CB(skb)->end_seq;
4485 		}
4486 		__skb_queue_head(&tp->out_of_order_queue, skb);
4487 		goto end;
4488 	}
4489 
4490 	seq = TCP_SKB_CB(skb)->seq;
4491 	end_seq = TCP_SKB_CB(skb)->end_seq;
4492 
4493 	if (seq == TCP_SKB_CB(skb1)->end_seq) {
4494 		/* Packets in ofo can stay in queue a long time.
4495 		 * Better try to coalesce them right now
4496 		 * to avoid future tcp_collapse_ofo_queue(),
4497 		 * probably the most expensive function in tcp stack.
4498 		 */
4499 		if (skb->len <= skb_tailroom(skb1) && !tcp_hdr(skb)->fin) {
4500 			NET_INC_STATS_BH(sock_net(sk),
4501 					 LINUX_MIB_TCPRCVCOALESCE);
4502 			BUG_ON(skb_copy_bits(skb, 0,
4503 					     skb_put(skb1, skb->len),
4504 					     skb->len));
4505 			TCP_SKB_CB(skb1)->end_seq = end_seq;
4506 			TCP_SKB_CB(skb1)->ack_seq = TCP_SKB_CB(skb)->ack_seq;
4507 			__kfree_skb(skb);
4508 			skb = NULL;
4509 		} else {
4510 			__skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4511 		}
4512 
4513 		if (!tp->rx_opt.num_sacks ||
4514 		    tp->selective_acks[0].end_seq != seq)
4515 			goto add_sack;
4516 
4517 		/* Common case: data arrive in order after hole. */
4518 		tp->selective_acks[0].end_seq = end_seq;
4519 		goto end;
4520 	}
4521 
4522 	/* Find place to insert this segment. */
4523 	while (1) {
4524 		if (!after(TCP_SKB_CB(skb1)->seq, seq))
4525 			break;
4526 		if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4527 			skb1 = NULL;
4528 			break;
4529 		}
4530 		skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4531 	}
4532 
4533 	/* Do skb overlap to previous one? */
4534 	if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4535 		if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4536 			/* All the bits are present. Drop. */
4537 			__kfree_skb(skb);
4538 			skb = NULL;
4539 			tcp_dsack_set(sk, seq, end_seq);
4540 			goto add_sack;
4541 		}
4542 		if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4543 			/* Partial overlap. */
4544 			tcp_dsack_set(sk, seq,
4545 				      TCP_SKB_CB(skb1)->end_seq);
4546 		} else {
4547 			if (skb_queue_is_first(&tp->out_of_order_queue,
4548 					       skb1))
4549 				skb1 = NULL;
4550 			else
4551 				skb1 = skb_queue_prev(
4552 					&tp->out_of_order_queue,
4553 					skb1);
4554 		}
4555 	}
4556 	if (!skb1)
4557 		__skb_queue_head(&tp->out_of_order_queue, skb);
4558 	else
4559 		__skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4560 
4561 	/* And clean segments covered by new one as whole. */
4562 	while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4563 		skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4564 
4565 		if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4566 			break;
4567 		if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4568 			tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4569 					 end_seq);
4570 			break;
4571 		}
4572 		__skb_unlink(skb1, &tp->out_of_order_queue);
4573 		tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4574 				 TCP_SKB_CB(skb1)->end_seq);
4575 		__kfree_skb(skb1);
4576 	}
4577 
4578 add_sack:
4579 	if (tcp_is_sack(tp))
4580 		tcp_sack_new_ofo_skb(sk, seq, end_seq);
4581 end:
4582 	if (skb)
4583 		skb_set_owner_r(skb, sk);
4584 }
4585 
4586 
4587 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4588 {
4589 	const struct tcphdr *th = tcp_hdr(skb);
4590 	struct tcp_sock *tp = tcp_sk(sk);
4591 	int eaten = -1;
4592 
4593 	if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4594 		goto drop;
4595 
4596 	skb_dst_drop(skb);
4597 	__skb_pull(skb, th->doff * 4);
4598 
4599 	TCP_ECN_accept_cwr(tp, skb);
4600 
4601 	tp->rx_opt.dsack = 0;
4602 
4603 	/*  Queue data for delivery to the user.
4604 	 *  Packets in sequence go to the receive queue.
4605 	 *  Out of sequence packets to the out_of_order_queue.
4606 	 */
4607 	if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4608 		if (tcp_receive_window(tp) == 0)
4609 			goto out_of_window;
4610 
4611 		/* Ok. In sequence. In window. */
4612 		if (tp->ucopy.task == current &&
4613 		    tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4614 		    sock_owned_by_user(sk) && !tp->urg_data) {
4615 			int chunk = min_t(unsigned int, skb->len,
4616 					  tp->ucopy.len);
4617 
4618 			__set_current_state(TASK_RUNNING);
4619 
4620 			local_bh_enable();
4621 			if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4622 				tp->ucopy.len -= chunk;
4623 				tp->copied_seq += chunk;
4624 				eaten = (chunk == skb->len);
4625 				tcp_rcv_space_adjust(sk);
4626 			}
4627 			local_bh_disable();
4628 		}
4629 
4630 		if (eaten <= 0) {
4631 queue_and_out:
4632 			if (eaten < 0 &&
4633 			    tcp_try_rmem_schedule(sk, skb->truesize))
4634 				goto drop;
4635 
4636 			skb_set_owner_r(skb, sk);
4637 			__skb_queue_tail(&sk->sk_receive_queue, skb);
4638 		}
4639 		tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4640 		if (skb->len)
4641 			tcp_event_data_recv(sk, skb);
4642 		if (th->fin)
4643 			tcp_fin(sk);
4644 
4645 		if (!skb_queue_empty(&tp->out_of_order_queue)) {
4646 			tcp_ofo_queue(sk);
4647 
4648 			/* RFC2581. 4.2. SHOULD send immediate ACK, when
4649 			 * gap in queue is filled.
4650 			 */
4651 			if (skb_queue_empty(&tp->out_of_order_queue))
4652 				inet_csk(sk)->icsk_ack.pingpong = 0;
4653 		}
4654 
4655 		if (tp->rx_opt.num_sacks)
4656 			tcp_sack_remove(tp);
4657 
4658 		tcp_fast_path_check(sk);
4659 
4660 		if (eaten > 0)
4661 			__kfree_skb(skb);
4662 		else if (!sock_flag(sk, SOCK_DEAD))
4663 			sk->sk_data_ready(sk, 0);
4664 		return;
4665 	}
4666 
4667 	if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4668 		/* A retransmit, 2nd most common case.  Force an immediate ack. */
4669 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4670 		tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4671 
4672 out_of_window:
4673 		tcp_enter_quickack_mode(sk);
4674 		inet_csk_schedule_ack(sk);
4675 drop:
4676 		__kfree_skb(skb);
4677 		return;
4678 	}
4679 
4680 	/* Out of window. F.e. zero window probe. */
4681 	if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4682 		goto out_of_window;
4683 
4684 	tcp_enter_quickack_mode(sk);
4685 
4686 	if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4687 		/* Partial packet, seq < rcv_next < end_seq */
4688 		SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4689 			   tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4690 			   TCP_SKB_CB(skb)->end_seq);
4691 
4692 		tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4693 
4694 		/* If window is closed, drop tail of packet. But after
4695 		 * remembering D-SACK for its head made in previous line.
4696 		 */
4697 		if (!tcp_receive_window(tp))
4698 			goto out_of_window;
4699 		goto queue_and_out;
4700 	}
4701 
4702 	tcp_data_queue_ofo(sk, skb);
4703 }
4704 
4705 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4706 					struct sk_buff_head *list)
4707 {
4708 	struct sk_buff *next = NULL;
4709 
4710 	if (!skb_queue_is_last(list, skb))
4711 		next = skb_queue_next(list, skb);
4712 
4713 	__skb_unlink(skb, list);
4714 	__kfree_skb(skb);
4715 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4716 
4717 	return next;
4718 }
4719 
4720 /* Collapse contiguous sequence of skbs head..tail with
4721  * sequence numbers start..end.
4722  *
4723  * If tail is NULL, this means until the end of the list.
4724  *
4725  * Segments with FIN/SYN are not collapsed (only because this
4726  * simplifies code)
4727  */
4728 static void
4729 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4730 	     struct sk_buff *head, struct sk_buff *tail,
4731 	     u32 start, u32 end)
4732 {
4733 	struct sk_buff *skb, *n;
4734 	bool end_of_skbs;
4735 
4736 	/* First, check that queue is collapsible and find
4737 	 * the point where collapsing can be useful. */
4738 	skb = head;
4739 restart:
4740 	end_of_skbs = true;
4741 	skb_queue_walk_from_safe(list, skb, n) {
4742 		if (skb == tail)
4743 			break;
4744 		/* No new bits? It is possible on ofo queue. */
4745 		if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4746 			skb = tcp_collapse_one(sk, skb, list);
4747 			if (!skb)
4748 				break;
4749 			goto restart;
4750 		}
4751 
4752 		/* The first skb to collapse is:
4753 		 * - not SYN/FIN and
4754 		 * - bloated or contains data before "start" or
4755 		 *   overlaps to the next one.
4756 		 */
4757 		if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4758 		    (tcp_win_from_space(skb->truesize) > skb->len ||
4759 		     before(TCP_SKB_CB(skb)->seq, start))) {
4760 			end_of_skbs = false;
4761 			break;
4762 		}
4763 
4764 		if (!skb_queue_is_last(list, skb)) {
4765 			struct sk_buff *next = skb_queue_next(list, skb);
4766 			if (next != tail &&
4767 			    TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4768 				end_of_skbs = false;
4769 				break;
4770 			}
4771 		}
4772 
4773 		/* Decided to skip this, advance start seq. */
4774 		start = TCP_SKB_CB(skb)->end_seq;
4775 	}
4776 	if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4777 		return;
4778 
4779 	while (before(start, end)) {
4780 		struct sk_buff *nskb;
4781 		unsigned int header = skb_headroom(skb);
4782 		int copy = SKB_MAX_ORDER(header, 0);
4783 
4784 		/* Too big header? This can happen with IPv6. */
4785 		if (copy < 0)
4786 			return;
4787 		if (end - start < copy)
4788 			copy = end - start;
4789 		nskb = alloc_skb(copy + header, GFP_ATOMIC);
4790 		if (!nskb)
4791 			return;
4792 
4793 		skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4794 		skb_set_network_header(nskb, (skb_network_header(skb) -
4795 					      skb->head));
4796 		skb_set_transport_header(nskb, (skb_transport_header(skb) -
4797 						skb->head));
4798 		skb_reserve(nskb, header);
4799 		memcpy(nskb->head, skb->head, header);
4800 		memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4801 		TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4802 		__skb_queue_before(list, skb, nskb);
4803 		skb_set_owner_r(nskb, sk);
4804 
4805 		/* Copy data, releasing collapsed skbs. */
4806 		while (copy > 0) {
4807 			int offset = start - TCP_SKB_CB(skb)->seq;
4808 			int size = TCP_SKB_CB(skb)->end_seq - start;
4809 
4810 			BUG_ON(offset < 0);
4811 			if (size > 0) {
4812 				size = min(copy, size);
4813 				if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4814 					BUG();
4815 				TCP_SKB_CB(nskb)->end_seq += size;
4816 				copy -= size;
4817 				start += size;
4818 			}
4819 			if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4820 				skb = tcp_collapse_one(sk, skb, list);
4821 				if (!skb ||
4822 				    skb == tail ||
4823 				    tcp_hdr(skb)->syn ||
4824 				    tcp_hdr(skb)->fin)
4825 					return;
4826 			}
4827 		}
4828 	}
4829 }
4830 
4831 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4832  * and tcp_collapse() them until all the queue is collapsed.
4833  */
4834 static void tcp_collapse_ofo_queue(struct sock *sk)
4835 {
4836 	struct tcp_sock *tp = tcp_sk(sk);
4837 	struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4838 	struct sk_buff *head;
4839 	u32 start, end;
4840 
4841 	if (skb == NULL)
4842 		return;
4843 
4844 	start = TCP_SKB_CB(skb)->seq;
4845 	end = TCP_SKB_CB(skb)->end_seq;
4846 	head = skb;
4847 
4848 	for (;;) {
4849 		struct sk_buff *next = NULL;
4850 
4851 		if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4852 			next = skb_queue_next(&tp->out_of_order_queue, skb);
4853 		skb = next;
4854 
4855 		/* Segment is terminated when we see gap or when
4856 		 * we are at the end of all the queue. */
4857 		if (!skb ||
4858 		    after(TCP_SKB_CB(skb)->seq, end) ||
4859 		    before(TCP_SKB_CB(skb)->end_seq, start)) {
4860 			tcp_collapse(sk, &tp->out_of_order_queue,
4861 				     head, skb, start, end);
4862 			head = skb;
4863 			if (!skb)
4864 				break;
4865 			/* Start new segment */
4866 			start = TCP_SKB_CB(skb)->seq;
4867 			end = TCP_SKB_CB(skb)->end_seq;
4868 		} else {
4869 			if (before(TCP_SKB_CB(skb)->seq, start))
4870 				start = TCP_SKB_CB(skb)->seq;
4871 			if (after(TCP_SKB_CB(skb)->end_seq, end))
4872 				end = TCP_SKB_CB(skb)->end_seq;
4873 		}
4874 	}
4875 }
4876 
4877 /*
4878  * Purge the out-of-order queue.
4879  * Return true if queue was pruned.
4880  */
4881 static int tcp_prune_ofo_queue(struct sock *sk)
4882 {
4883 	struct tcp_sock *tp = tcp_sk(sk);
4884 	int res = 0;
4885 
4886 	if (!skb_queue_empty(&tp->out_of_order_queue)) {
4887 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4888 		__skb_queue_purge(&tp->out_of_order_queue);
4889 
4890 		/* Reset SACK state.  A conforming SACK implementation will
4891 		 * do the same at a timeout based retransmit.  When a connection
4892 		 * is in a sad state like this, we care only about integrity
4893 		 * of the connection not performance.
4894 		 */
4895 		if (tp->rx_opt.sack_ok)
4896 			tcp_sack_reset(&tp->rx_opt);
4897 		sk_mem_reclaim(sk);
4898 		res = 1;
4899 	}
4900 	return res;
4901 }
4902 
4903 /* Reduce allocated memory if we can, trying to get
4904  * the socket within its memory limits again.
4905  *
4906  * Return less than zero if we should start dropping frames
4907  * until the socket owning process reads some of the data
4908  * to stabilize the situation.
4909  */
4910 static int tcp_prune_queue(struct sock *sk)
4911 {
4912 	struct tcp_sock *tp = tcp_sk(sk);
4913 
4914 	SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4915 
4916 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4917 
4918 	if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4919 		tcp_clamp_window(sk);
4920 	else if (sk_under_memory_pressure(sk))
4921 		tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4922 
4923 	tcp_collapse_ofo_queue(sk);
4924 	if (!skb_queue_empty(&sk->sk_receive_queue))
4925 		tcp_collapse(sk, &sk->sk_receive_queue,
4926 			     skb_peek(&sk->sk_receive_queue),
4927 			     NULL,
4928 			     tp->copied_seq, tp->rcv_nxt);
4929 	sk_mem_reclaim(sk);
4930 
4931 	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4932 		return 0;
4933 
4934 	/* Collapsing did not help, destructive actions follow.
4935 	 * This must not ever occur. */
4936 
4937 	tcp_prune_ofo_queue(sk);
4938 
4939 	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4940 		return 0;
4941 
4942 	/* If we are really being abused, tell the caller to silently
4943 	 * drop receive data on the floor.  It will get retransmitted
4944 	 * and hopefully then we'll have sufficient space.
4945 	 */
4946 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4947 
4948 	/* Massive buffer overcommit. */
4949 	tp->pred_flags = 0;
4950 	return -1;
4951 }
4952 
4953 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4954  * As additional protections, we do not touch cwnd in retransmission phases,
4955  * and if application hit its sndbuf limit recently.
4956  */
4957 void tcp_cwnd_application_limited(struct sock *sk)
4958 {
4959 	struct tcp_sock *tp = tcp_sk(sk);
4960 
4961 	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4962 	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4963 		/* Limited by application or receiver window. */
4964 		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4965 		u32 win_used = max(tp->snd_cwnd_used, init_win);
4966 		if (win_used < tp->snd_cwnd) {
4967 			tp->snd_ssthresh = tcp_current_ssthresh(sk);
4968 			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4969 		}
4970 		tp->snd_cwnd_used = 0;
4971 	}
4972 	tp->snd_cwnd_stamp = tcp_time_stamp;
4973 }
4974 
4975 static int tcp_should_expand_sndbuf(const struct sock *sk)
4976 {
4977 	const struct tcp_sock *tp = tcp_sk(sk);
4978 
4979 	/* If the user specified a specific send buffer setting, do
4980 	 * not modify it.
4981 	 */
4982 	if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4983 		return 0;
4984 
4985 	/* If we are under global TCP memory pressure, do not expand.  */
4986 	if (sk_under_memory_pressure(sk))
4987 		return 0;
4988 
4989 	/* If we are under soft global TCP memory pressure, do not expand.  */
4990 	if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4991 		return 0;
4992 
4993 	/* If we filled the congestion window, do not expand.  */
4994 	if (tp->packets_out >= tp->snd_cwnd)
4995 		return 0;
4996 
4997 	return 1;
4998 }
4999 
5000 /* When incoming ACK allowed to free some skb from write_queue,
5001  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5002  * on the exit from tcp input handler.
5003  *
5004  * PROBLEM: sndbuf expansion does not work well with largesend.
5005  */
5006 static void tcp_new_space(struct sock *sk)
5007 {
5008 	struct tcp_sock *tp = tcp_sk(sk);
5009 
5010 	if (tcp_should_expand_sndbuf(sk)) {
5011 		int sndmem = SKB_TRUESIZE(max_t(u32,
5012 						tp->rx_opt.mss_clamp,
5013 						tp->mss_cache) +
5014 					  MAX_TCP_HEADER);
5015 		int demanded = max_t(unsigned int, tp->snd_cwnd,
5016 				     tp->reordering + 1);
5017 		sndmem *= 2 * demanded;
5018 		if (sndmem > sk->sk_sndbuf)
5019 			sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
5020 		tp->snd_cwnd_stamp = tcp_time_stamp;
5021 	}
5022 
5023 	sk->sk_write_space(sk);
5024 }
5025 
5026 static void tcp_check_space(struct sock *sk)
5027 {
5028 	if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5029 		sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5030 		if (sk->sk_socket &&
5031 		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5032 			tcp_new_space(sk);
5033 	}
5034 }
5035 
5036 static inline void tcp_data_snd_check(struct sock *sk)
5037 {
5038 	tcp_push_pending_frames(sk);
5039 	tcp_check_space(sk);
5040 }
5041 
5042 /*
5043  * Check if sending an ack is needed.
5044  */
5045 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5046 {
5047 	struct tcp_sock *tp = tcp_sk(sk);
5048 
5049 	    /* More than one full frame received... */
5050 	if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5051 	     /* ... and right edge of window advances far enough.
5052 	      * (tcp_recvmsg() will send ACK otherwise). Or...
5053 	      */
5054 	     __tcp_select_window(sk) >= tp->rcv_wnd) ||
5055 	    /* We ACK each frame or... */
5056 	    tcp_in_quickack_mode(sk) ||
5057 	    /* We have out of order data. */
5058 	    (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5059 		/* Then ack it now */
5060 		tcp_send_ack(sk);
5061 	} else {
5062 		/* Else, send delayed ack. */
5063 		tcp_send_delayed_ack(sk);
5064 	}
5065 }
5066 
5067 static inline void tcp_ack_snd_check(struct sock *sk)
5068 {
5069 	if (!inet_csk_ack_scheduled(sk)) {
5070 		/* We sent a data segment already. */
5071 		return;
5072 	}
5073 	__tcp_ack_snd_check(sk, 1);
5074 }
5075 
5076 /*
5077  *	This routine is only called when we have urgent data
5078  *	signaled. Its the 'slow' part of tcp_urg. It could be
5079  *	moved inline now as tcp_urg is only called from one
5080  *	place. We handle URGent data wrong. We have to - as
5081  *	BSD still doesn't use the correction from RFC961.
5082  *	For 1003.1g we should support a new option TCP_STDURG to permit
5083  *	either form (or just set the sysctl tcp_stdurg).
5084  */
5085 
5086 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5087 {
5088 	struct tcp_sock *tp = tcp_sk(sk);
5089 	u32 ptr = ntohs(th->urg_ptr);
5090 
5091 	if (ptr && !sysctl_tcp_stdurg)
5092 		ptr--;
5093 	ptr += ntohl(th->seq);
5094 
5095 	/* Ignore urgent data that we've already seen and read. */
5096 	if (after(tp->copied_seq, ptr))
5097 		return;
5098 
5099 	/* Do not replay urg ptr.
5100 	 *
5101 	 * NOTE: interesting situation not covered by specs.
5102 	 * Misbehaving sender may send urg ptr, pointing to segment,
5103 	 * which we already have in ofo queue. We are not able to fetch
5104 	 * such data and will stay in TCP_URG_NOTYET until will be eaten
5105 	 * by recvmsg(). Seems, we are not obliged to handle such wicked
5106 	 * situations. But it is worth to think about possibility of some
5107 	 * DoSes using some hypothetical application level deadlock.
5108 	 */
5109 	if (before(ptr, tp->rcv_nxt))
5110 		return;
5111 
5112 	/* Do we already have a newer (or duplicate) urgent pointer? */
5113 	if (tp->urg_data && !after(ptr, tp->urg_seq))
5114 		return;
5115 
5116 	/* Tell the world about our new urgent pointer. */
5117 	sk_send_sigurg(sk);
5118 
5119 	/* We may be adding urgent data when the last byte read was
5120 	 * urgent. To do this requires some care. We cannot just ignore
5121 	 * tp->copied_seq since we would read the last urgent byte again
5122 	 * as data, nor can we alter copied_seq until this data arrives
5123 	 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5124 	 *
5125 	 * NOTE. Double Dutch. Rendering to plain English: author of comment
5126 	 * above did something sort of 	send("A", MSG_OOB); send("B", MSG_OOB);
5127 	 * and expect that both A and B disappear from stream. This is _wrong_.
5128 	 * Though this happens in BSD with high probability, this is occasional.
5129 	 * Any application relying on this is buggy. Note also, that fix "works"
5130 	 * only in this artificial test. Insert some normal data between A and B and we will
5131 	 * decline of BSD again. Verdict: it is better to remove to trap
5132 	 * buggy users.
5133 	 */
5134 	if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5135 	    !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5136 		struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5137 		tp->copied_seq++;
5138 		if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5139 			__skb_unlink(skb, &sk->sk_receive_queue);
5140 			__kfree_skb(skb);
5141 		}
5142 	}
5143 
5144 	tp->urg_data = TCP_URG_NOTYET;
5145 	tp->urg_seq = ptr;
5146 
5147 	/* Disable header prediction. */
5148 	tp->pred_flags = 0;
5149 }
5150 
5151 /* This is the 'fast' part of urgent handling. */
5152 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5153 {
5154 	struct tcp_sock *tp = tcp_sk(sk);
5155 
5156 	/* Check if we get a new urgent pointer - normally not. */
5157 	if (th->urg)
5158 		tcp_check_urg(sk, th);
5159 
5160 	/* Do we wait for any urgent data? - normally not... */
5161 	if (tp->urg_data == TCP_URG_NOTYET) {
5162 		u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5163 			  th->syn;
5164 
5165 		/* Is the urgent pointer pointing into this packet? */
5166 		if (ptr < skb->len) {
5167 			u8 tmp;
5168 			if (skb_copy_bits(skb, ptr, &tmp, 1))
5169 				BUG();
5170 			tp->urg_data = TCP_URG_VALID | tmp;
5171 			if (!sock_flag(sk, SOCK_DEAD))
5172 				sk->sk_data_ready(sk, 0);
5173 		}
5174 	}
5175 }
5176 
5177 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5178 {
5179 	struct tcp_sock *tp = tcp_sk(sk);
5180 	int chunk = skb->len - hlen;
5181 	int err;
5182 
5183 	local_bh_enable();
5184 	if (skb_csum_unnecessary(skb))
5185 		err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5186 	else
5187 		err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5188 						       tp->ucopy.iov);
5189 
5190 	if (!err) {
5191 		tp->ucopy.len -= chunk;
5192 		tp->copied_seq += chunk;
5193 		tcp_rcv_space_adjust(sk);
5194 	}
5195 
5196 	local_bh_disable();
5197 	return err;
5198 }
5199 
5200 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5201 					    struct sk_buff *skb)
5202 {
5203 	__sum16 result;
5204 
5205 	if (sock_owned_by_user(sk)) {
5206 		local_bh_enable();
5207 		result = __tcp_checksum_complete(skb);
5208 		local_bh_disable();
5209 	} else {
5210 		result = __tcp_checksum_complete(skb);
5211 	}
5212 	return result;
5213 }
5214 
5215 static inline int tcp_checksum_complete_user(struct sock *sk,
5216 					     struct sk_buff *skb)
5217 {
5218 	return !skb_csum_unnecessary(skb) &&
5219 	       __tcp_checksum_complete_user(sk, skb);
5220 }
5221 
5222 #ifdef CONFIG_NET_DMA
5223 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5224 				  int hlen)
5225 {
5226 	struct tcp_sock *tp = tcp_sk(sk);
5227 	int chunk = skb->len - hlen;
5228 	int dma_cookie;
5229 	int copied_early = 0;
5230 
5231 	if (tp->ucopy.wakeup)
5232 		return 0;
5233 
5234 	if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5235 		tp->ucopy.dma_chan = net_dma_find_channel();
5236 
5237 	if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5238 
5239 		dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5240 							 skb, hlen,
5241 							 tp->ucopy.iov, chunk,
5242 							 tp->ucopy.pinned_list);
5243 
5244 		if (dma_cookie < 0)
5245 			goto out;
5246 
5247 		tp->ucopy.dma_cookie = dma_cookie;
5248 		copied_early = 1;
5249 
5250 		tp->ucopy.len -= chunk;
5251 		tp->copied_seq += chunk;
5252 		tcp_rcv_space_adjust(sk);
5253 
5254 		if ((tp->ucopy.len == 0) ||
5255 		    (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5256 		    (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5257 			tp->ucopy.wakeup = 1;
5258 			sk->sk_data_ready(sk, 0);
5259 		}
5260 	} else if (chunk > 0) {
5261 		tp->ucopy.wakeup = 1;
5262 		sk->sk_data_ready(sk, 0);
5263 	}
5264 out:
5265 	return copied_early;
5266 }
5267 #endif /* CONFIG_NET_DMA */
5268 
5269 /* Does PAWS and seqno based validation of an incoming segment, flags will
5270  * play significant role here.
5271  */
5272 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5273 			      const struct tcphdr *th, int syn_inerr)
5274 {
5275 	const u8 *hash_location;
5276 	struct tcp_sock *tp = tcp_sk(sk);
5277 
5278 	/* RFC1323: H1. Apply PAWS check first. */
5279 	if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5280 	    tp->rx_opt.saw_tstamp &&
5281 	    tcp_paws_discard(sk, skb)) {
5282 		if (!th->rst) {
5283 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5284 			tcp_send_dupack(sk, skb);
5285 			goto discard;
5286 		}
5287 		/* Reset is accepted even if it did not pass PAWS. */
5288 	}
5289 
5290 	/* Step 1: check sequence number */
5291 	if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5292 		/* RFC793, page 37: "In all states except SYN-SENT, all reset
5293 		 * (RST) segments are validated by checking their SEQ-fields."
5294 		 * And page 69: "If an incoming segment is not acceptable,
5295 		 * an acknowledgment should be sent in reply (unless the RST
5296 		 * bit is set, if so drop the segment and return)".
5297 		 */
5298 		if (!th->rst)
5299 			tcp_send_dupack(sk, skb);
5300 		goto discard;
5301 	}
5302 
5303 	/* Step 2: check RST bit */
5304 	if (th->rst) {
5305 		tcp_reset(sk);
5306 		goto discard;
5307 	}
5308 
5309 	/* ts_recent update must be made after we are sure that the packet
5310 	 * is in window.
5311 	 */
5312 	tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5313 
5314 	/* step 3: check security and precedence [ignored] */
5315 
5316 	/* step 4: Check for a SYN in window. */
5317 	if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5318 		if (syn_inerr)
5319 			TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5320 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5321 		tcp_reset(sk);
5322 		return -1;
5323 	}
5324 
5325 	return 1;
5326 
5327 discard:
5328 	__kfree_skb(skb);
5329 	return 0;
5330 }
5331 
5332 /*
5333  *	TCP receive function for the ESTABLISHED state.
5334  *
5335  *	It is split into a fast path and a slow path. The fast path is
5336  * 	disabled when:
5337  *	- A zero window was announced from us - zero window probing
5338  *        is only handled properly in the slow path.
5339  *	- Out of order segments arrived.
5340  *	- Urgent data is expected.
5341  *	- There is no buffer space left
5342  *	- Unexpected TCP flags/window values/header lengths are received
5343  *	  (detected by checking the TCP header against pred_flags)
5344  *	- Data is sent in both directions. Fast path only supports pure senders
5345  *	  or pure receivers (this means either the sequence number or the ack
5346  *	  value must stay constant)
5347  *	- Unexpected TCP option.
5348  *
5349  *	When these conditions are not satisfied it drops into a standard
5350  *	receive procedure patterned after RFC793 to handle all cases.
5351  *	The first three cases are guaranteed by proper pred_flags setting,
5352  *	the rest is checked inline. Fast processing is turned on in
5353  *	tcp_data_queue when everything is OK.
5354  */
5355 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5356 			const struct tcphdr *th, unsigned int len)
5357 {
5358 	struct tcp_sock *tp = tcp_sk(sk);
5359 	int res;
5360 
5361 	/*
5362 	 *	Header prediction.
5363 	 *	The code loosely follows the one in the famous
5364 	 *	"30 instruction TCP receive" Van Jacobson mail.
5365 	 *
5366 	 *	Van's trick is to deposit buffers into socket queue
5367 	 *	on a device interrupt, to call tcp_recv function
5368 	 *	on the receive process context and checksum and copy
5369 	 *	the buffer to user space. smart...
5370 	 *
5371 	 *	Our current scheme is not silly either but we take the
5372 	 *	extra cost of the net_bh soft interrupt processing...
5373 	 *	We do checksum and copy also but from device to kernel.
5374 	 */
5375 
5376 	tp->rx_opt.saw_tstamp = 0;
5377 
5378 	/*	pred_flags is 0xS?10 << 16 + snd_wnd
5379 	 *	if header_prediction is to be made
5380 	 *	'S' will always be tp->tcp_header_len >> 2
5381 	 *	'?' will be 0 for the fast path, otherwise pred_flags is 0 to
5382 	 *  turn it off	(when there are holes in the receive
5383 	 *	 space for instance)
5384 	 *	PSH flag is ignored.
5385 	 */
5386 
5387 	if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5388 	    TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5389 	    !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5390 		int tcp_header_len = tp->tcp_header_len;
5391 
5392 		/* Timestamp header prediction: tcp_header_len
5393 		 * is automatically equal to th->doff*4 due to pred_flags
5394 		 * match.
5395 		 */
5396 
5397 		/* Check timestamp */
5398 		if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5399 			/* No? Slow path! */
5400 			if (!tcp_parse_aligned_timestamp(tp, th))
5401 				goto slow_path;
5402 
5403 			/* If PAWS failed, check it more carefully in slow path */
5404 			if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5405 				goto slow_path;
5406 
5407 			/* DO NOT update ts_recent here, if checksum fails
5408 			 * and timestamp was corrupted part, it will result
5409 			 * in a hung connection since we will drop all
5410 			 * future packets due to the PAWS test.
5411 			 */
5412 		}
5413 
5414 		if (len <= tcp_header_len) {
5415 			/* Bulk data transfer: sender */
5416 			if (len == tcp_header_len) {
5417 				/* Predicted packet is in window by definition.
5418 				 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5419 				 * Hence, check seq<=rcv_wup reduces to:
5420 				 */
5421 				if (tcp_header_len ==
5422 				    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5423 				    tp->rcv_nxt == tp->rcv_wup)
5424 					tcp_store_ts_recent(tp);
5425 
5426 				/* We know that such packets are checksummed
5427 				 * on entry.
5428 				 */
5429 				tcp_ack(sk, skb, 0);
5430 				__kfree_skb(skb);
5431 				tcp_data_snd_check(sk);
5432 				return 0;
5433 			} else { /* Header too small */
5434 				TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5435 				goto discard;
5436 			}
5437 		} else {
5438 			int eaten = 0;
5439 			int copied_early = 0;
5440 
5441 			if (tp->copied_seq == tp->rcv_nxt &&
5442 			    len - tcp_header_len <= tp->ucopy.len) {
5443 #ifdef CONFIG_NET_DMA
5444 				if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5445 					copied_early = 1;
5446 					eaten = 1;
5447 				}
5448 #endif
5449 				if (tp->ucopy.task == current &&
5450 				    sock_owned_by_user(sk) && !copied_early) {
5451 					__set_current_state(TASK_RUNNING);
5452 
5453 					if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5454 						eaten = 1;
5455 				}
5456 				if (eaten) {
5457 					/* Predicted packet is in window by definition.
5458 					 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5459 					 * Hence, check seq<=rcv_wup reduces to:
5460 					 */
5461 					if (tcp_header_len ==
5462 					    (sizeof(struct tcphdr) +
5463 					     TCPOLEN_TSTAMP_ALIGNED) &&
5464 					    tp->rcv_nxt == tp->rcv_wup)
5465 						tcp_store_ts_recent(tp);
5466 
5467 					tcp_rcv_rtt_measure_ts(sk, skb);
5468 
5469 					__skb_pull(skb, tcp_header_len);
5470 					tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5471 					NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5472 				}
5473 				if (copied_early)
5474 					tcp_cleanup_rbuf(sk, skb->len);
5475 			}
5476 			if (!eaten) {
5477 				if (tcp_checksum_complete_user(sk, skb))
5478 					goto csum_error;
5479 
5480 				/* Predicted packet is in window by definition.
5481 				 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5482 				 * Hence, check seq<=rcv_wup reduces to:
5483 				 */
5484 				if (tcp_header_len ==
5485 				    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5486 				    tp->rcv_nxt == tp->rcv_wup)
5487 					tcp_store_ts_recent(tp);
5488 
5489 				tcp_rcv_rtt_measure_ts(sk, skb);
5490 
5491 				if ((int)skb->truesize > sk->sk_forward_alloc)
5492 					goto step5;
5493 
5494 				NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5495 
5496 				/* Bulk data transfer: receiver */
5497 				__skb_pull(skb, tcp_header_len);
5498 				__skb_queue_tail(&sk->sk_receive_queue, skb);
5499 				skb_set_owner_r(skb, sk);
5500 				tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5501 			}
5502 
5503 			tcp_event_data_recv(sk, skb);
5504 
5505 			if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5506 				/* Well, only one small jumplet in fast path... */
5507 				tcp_ack(sk, skb, FLAG_DATA);
5508 				tcp_data_snd_check(sk);
5509 				if (!inet_csk_ack_scheduled(sk))
5510 					goto no_ack;
5511 			}
5512 
5513 			if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5514 				__tcp_ack_snd_check(sk, 0);
5515 no_ack:
5516 #ifdef CONFIG_NET_DMA
5517 			if (copied_early)
5518 				__skb_queue_tail(&sk->sk_async_wait_queue, skb);
5519 			else
5520 #endif
5521 			if (eaten)
5522 				__kfree_skb(skb);
5523 			else
5524 				sk->sk_data_ready(sk, 0);
5525 			return 0;
5526 		}
5527 	}
5528 
5529 slow_path:
5530 	if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5531 		goto csum_error;
5532 
5533 	/*
5534 	 *	Standard slow path.
5535 	 */
5536 
5537 	res = tcp_validate_incoming(sk, skb, th, 1);
5538 	if (res <= 0)
5539 		return -res;
5540 
5541 step5:
5542 	if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5543 		goto discard;
5544 
5545 	tcp_rcv_rtt_measure_ts(sk, skb);
5546 
5547 	/* Process urgent data. */
5548 	tcp_urg(sk, skb, th);
5549 
5550 	/* step 7: process the segment text */
5551 	tcp_data_queue(sk, skb);
5552 
5553 	tcp_data_snd_check(sk);
5554 	tcp_ack_snd_check(sk);
5555 	return 0;
5556 
5557 csum_error:
5558 	TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5559 
5560 discard:
5561 	__kfree_skb(skb);
5562 	return 0;
5563 }
5564 EXPORT_SYMBOL(tcp_rcv_established);
5565 
5566 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5567 					 const struct tcphdr *th, unsigned int len)
5568 {
5569 	const u8 *hash_location;
5570 	struct inet_connection_sock *icsk = inet_csk(sk);
5571 	struct tcp_sock *tp = tcp_sk(sk);
5572 	struct tcp_cookie_values *cvp = tp->cookie_values;
5573 	int saved_clamp = tp->rx_opt.mss_clamp;
5574 
5575 	tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5576 
5577 	if (th->ack) {
5578 		/* rfc793:
5579 		 * "If the state is SYN-SENT then
5580 		 *    first check the ACK bit
5581 		 *      If the ACK bit is set
5582 		 *	  If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5583 		 *        a reset (unless the RST bit is set, if so drop
5584 		 *        the segment and return)"
5585 		 *
5586 		 *  We do not send data with SYN, so that RFC-correct
5587 		 *  test reduces to:
5588 		 */
5589 		if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5590 			goto reset_and_undo;
5591 
5592 		if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5593 		    !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5594 			     tcp_time_stamp)) {
5595 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5596 			goto reset_and_undo;
5597 		}
5598 
5599 		/* Now ACK is acceptable.
5600 		 *
5601 		 * "If the RST bit is set
5602 		 *    If the ACK was acceptable then signal the user "error:
5603 		 *    connection reset", drop the segment, enter CLOSED state,
5604 		 *    delete TCB, and return."
5605 		 */
5606 
5607 		if (th->rst) {
5608 			tcp_reset(sk);
5609 			goto discard;
5610 		}
5611 
5612 		/* rfc793:
5613 		 *   "fifth, if neither of the SYN or RST bits is set then
5614 		 *    drop the segment and return."
5615 		 *
5616 		 *    See note below!
5617 		 *                                        --ANK(990513)
5618 		 */
5619 		if (!th->syn)
5620 			goto discard_and_undo;
5621 
5622 		/* rfc793:
5623 		 *   "If the SYN bit is on ...
5624 		 *    are acceptable then ...
5625 		 *    (our SYN has been ACKed), change the connection
5626 		 *    state to ESTABLISHED..."
5627 		 */
5628 
5629 		TCP_ECN_rcv_synack(tp, th);
5630 
5631 		tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5632 		tcp_ack(sk, skb, FLAG_SLOWPATH);
5633 
5634 		/* Ok.. it's good. Set up sequence numbers and
5635 		 * move to established.
5636 		 */
5637 		tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5638 		tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5639 
5640 		/* RFC1323: The window in SYN & SYN/ACK segments is
5641 		 * never scaled.
5642 		 */
5643 		tp->snd_wnd = ntohs(th->window);
5644 		tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5645 
5646 		if (!tp->rx_opt.wscale_ok) {
5647 			tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5648 			tp->window_clamp = min(tp->window_clamp, 65535U);
5649 		}
5650 
5651 		if (tp->rx_opt.saw_tstamp) {
5652 			tp->rx_opt.tstamp_ok	   = 1;
5653 			tp->tcp_header_len =
5654 				sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5655 			tp->advmss	    -= TCPOLEN_TSTAMP_ALIGNED;
5656 			tcp_store_ts_recent(tp);
5657 		} else {
5658 			tp->tcp_header_len = sizeof(struct tcphdr);
5659 		}
5660 
5661 		if (tcp_is_sack(tp) && sysctl_tcp_fack)
5662 			tcp_enable_fack(tp);
5663 
5664 		tcp_mtup_init(sk);
5665 		tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5666 		tcp_initialize_rcv_mss(sk);
5667 
5668 		/* Remember, tcp_poll() does not lock socket!
5669 		 * Change state from SYN-SENT only after copied_seq
5670 		 * is initialized. */
5671 		tp->copied_seq = tp->rcv_nxt;
5672 
5673 		if (cvp != NULL &&
5674 		    cvp->cookie_pair_size > 0 &&
5675 		    tp->rx_opt.cookie_plus > 0) {
5676 			int cookie_size = tp->rx_opt.cookie_plus
5677 					- TCPOLEN_COOKIE_BASE;
5678 			int cookie_pair_size = cookie_size
5679 					     + cvp->cookie_desired;
5680 
5681 			/* A cookie extension option was sent and returned.
5682 			 * Note that each incoming SYNACK replaces the
5683 			 * Responder cookie.  The initial exchange is most
5684 			 * fragile, as protection against spoofing relies
5685 			 * entirely upon the sequence and timestamp (above).
5686 			 * This replacement strategy allows the correct pair to
5687 			 * pass through, while any others will be filtered via
5688 			 * Responder verification later.
5689 			 */
5690 			if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5691 				memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5692 				       hash_location, cookie_size);
5693 				cvp->cookie_pair_size = cookie_pair_size;
5694 			}
5695 		}
5696 
5697 		smp_mb();
5698 		tcp_set_state(sk, TCP_ESTABLISHED);
5699 
5700 		security_inet_conn_established(sk, skb);
5701 
5702 		/* Make sure socket is routed, for correct metrics.  */
5703 		icsk->icsk_af_ops->rebuild_header(sk);
5704 
5705 		tcp_init_metrics(sk);
5706 
5707 		tcp_init_congestion_control(sk);
5708 
5709 		/* Prevent spurious tcp_cwnd_restart() on first data
5710 		 * packet.
5711 		 */
5712 		tp->lsndtime = tcp_time_stamp;
5713 
5714 		tcp_init_buffer_space(sk);
5715 
5716 		if (sock_flag(sk, SOCK_KEEPOPEN))
5717 			inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5718 
5719 		if (!tp->rx_opt.snd_wscale)
5720 			__tcp_fast_path_on(tp, tp->snd_wnd);
5721 		else
5722 			tp->pred_flags = 0;
5723 
5724 		if (!sock_flag(sk, SOCK_DEAD)) {
5725 			sk->sk_state_change(sk);
5726 			sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5727 		}
5728 
5729 		if (sk->sk_write_pending ||
5730 		    icsk->icsk_accept_queue.rskq_defer_accept ||
5731 		    icsk->icsk_ack.pingpong) {
5732 			/* Save one ACK. Data will be ready after
5733 			 * several ticks, if write_pending is set.
5734 			 *
5735 			 * It may be deleted, but with this feature tcpdumps
5736 			 * look so _wonderfully_ clever, that I was not able
5737 			 * to stand against the temptation 8)     --ANK
5738 			 */
5739 			inet_csk_schedule_ack(sk);
5740 			icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5741 			icsk->icsk_ack.ato	 = TCP_ATO_MIN;
5742 			tcp_incr_quickack(sk);
5743 			tcp_enter_quickack_mode(sk);
5744 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5745 						  TCP_DELACK_MAX, TCP_RTO_MAX);
5746 
5747 discard:
5748 			__kfree_skb(skb);
5749 			return 0;
5750 		} else {
5751 			tcp_send_ack(sk);
5752 		}
5753 		return -1;
5754 	}
5755 
5756 	/* No ACK in the segment */
5757 
5758 	if (th->rst) {
5759 		/* rfc793:
5760 		 * "If the RST bit is set
5761 		 *
5762 		 *      Otherwise (no ACK) drop the segment and return."
5763 		 */
5764 
5765 		goto discard_and_undo;
5766 	}
5767 
5768 	/* PAWS check. */
5769 	if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5770 	    tcp_paws_reject(&tp->rx_opt, 0))
5771 		goto discard_and_undo;
5772 
5773 	if (th->syn) {
5774 		/* We see SYN without ACK. It is attempt of
5775 		 * simultaneous connect with crossed SYNs.
5776 		 * Particularly, it can be connect to self.
5777 		 */
5778 		tcp_set_state(sk, TCP_SYN_RECV);
5779 
5780 		if (tp->rx_opt.saw_tstamp) {
5781 			tp->rx_opt.tstamp_ok = 1;
5782 			tcp_store_ts_recent(tp);
5783 			tp->tcp_header_len =
5784 				sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5785 		} else {
5786 			tp->tcp_header_len = sizeof(struct tcphdr);
5787 		}
5788 
5789 		tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5790 		tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5791 
5792 		/* RFC1323: The window in SYN & SYN/ACK segments is
5793 		 * never scaled.
5794 		 */
5795 		tp->snd_wnd    = ntohs(th->window);
5796 		tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5797 		tp->max_window = tp->snd_wnd;
5798 
5799 		TCP_ECN_rcv_syn(tp, th);
5800 
5801 		tcp_mtup_init(sk);
5802 		tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5803 		tcp_initialize_rcv_mss(sk);
5804 
5805 		tcp_send_synack(sk);
5806 #if 0
5807 		/* Note, we could accept data and URG from this segment.
5808 		 * There are no obstacles to make this.
5809 		 *
5810 		 * However, if we ignore data in ACKless segments sometimes,
5811 		 * we have no reasons to accept it sometimes.
5812 		 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5813 		 * is not flawless. So, discard packet for sanity.
5814 		 * Uncomment this return to process the data.
5815 		 */
5816 		return -1;
5817 #else
5818 		goto discard;
5819 #endif
5820 	}
5821 	/* "fifth, if neither of the SYN or RST bits is set then
5822 	 * drop the segment and return."
5823 	 */
5824 
5825 discard_and_undo:
5826 	tcp_clear_options(&tp->rx_opt);
5827 	tp->rx_opt.mss_clamp = saved_clamp;
5828 	goto discard;
5829 
5830 reset_and_undo:
5831 	tcp_clear_options(&tp->rx_opt);
5832 	tp->rx_opt.mss_clamp = saved_clamp;
5833 	return 1;
5834 }
5835 
5836 /*
5837  *	This function implements the receiving procedure of RFC 793 for
5838  *	all states except ESTABLISHED and TIME_WAIT.
5839  *	It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5840  *	address independent.
5841  */
5842 
5843 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5844 			  const struct tcphdr *th, unsigned int len)
5845 {
5846 	struct tcp_sock *tp = tcp_sk(sk);
5847 	struct inet_connection_sock *icsk = inet_csk(sk);
5848 	int queued = 0;
5849 	int res;
5850 
5851 	tp->rx_opt.saw_tstamp = 0;
5852 
5853 	switch (sk->sk_state) {
5854 	case TCP_CLOSE:
5855 		goto discard;
5856 
5857 	case TCP_LISTEN:
5858 		if (th->ack)
5859 			return 1;
5860 
5861 		if (th->rst)
5862 			goto discard;
5863 
5864 		if (th->syn) {
5865 			if (th->fin)
5866 				goto discard;
5867 			if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5868 				return 1;
5869 
5870 			/* Now we have several options: In theory there is
5871 			 * nothing else in the frame. KA9Q has an option to
5872 			 * send data with the syn, BSD accepts data with the
5873 			 * syn up to the [to be] advertised window and
5874 			 * Solaris 2.1 gives you a protocol error. For now
5875 			 * we just ignore it, that fits the spec precisely
5876 			 * and avoids incompatibilities. It would be nice in
5877 			 * future to drop through and process the data.
5878 			 *
5879 			 * Now that TTCP is starting to be used we ought to
5880 			 * queue this data.
5881 			 * But, this leaves one open to an easy denial of
5882 			 * service attack, and SYN cookies can't defend
5883 			 * against this problem. So, we drop the data
5884 			 * in the interest of security over speed unless
5885 			 * it's still in use.
5886 			 */
5887 			kfree_skb(skb);
5888 			return 0;
5889 		}
5890 		goto discard;
5891 
5892 	case TCP_SYN_SENT:
5893 		queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5894 		if (queued >= 0)
5895 			return queued;
5896 
5897 		/* Do step6 onward by hand. */
5898 		tcp_urg(sk, skb, th);
5899 		__kfree_skb(skb);
5900 		tcp_data_snd_check(sk);
5901 		return 0;
5902 	}
5903 
5904 	res = tcp_validate_incoming(sk, skb, th, 0);
5905 	if (res <= 0)
5906 		return -res;
5907 
5908 	/* step 5: check the ACK field */
5909 	if (th->ack) {
5910 		int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5911 
5912 		switch (sk->sk_state) {
5913 		case TCP_SYN_RECV:
5914 			if (acceptable) {
5915 				tp->copied_seq = tp->rcv_nxt;
5916 				smp_mb();
5917 				tcp_set_state(sk, TCP_ESTABLISHED);
5918 				sk->sk_state_change(sk);
5919 
5920 				/* Note, that this wakeup is only for marginal
5921 				 * crossed SYN case. Passively open sockets
5922 				 * are not waked up, because sk->sk_sleep ==
5923 				 * NULL and sk->sk_socket == NULL.
5924 				 */
5925 				if (sk->sk_socket)
5926 					sk_wake_async(sk,
5927 						      SOCK_WAKE_IO, POLL_OUT);
5928 
5929 				tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5930 				tp->snd_wnd = ntohs(th->window) <<
5931 					      tp->rx_opt.snd_wscale;
5932 				tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5933 
5934 				if (tp->rx_opt.tstamp_ok)
5935 					tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5936 
5937 				/* Make sure socket is routed, for
5938 				 * correct metrics.
5939 				 */
5940 				icsk->icsk_af_ops->rebuild_header(sk);
5941 
5942 				tcp_init_metrics(sk);
5943 
5944 				tcp_init_congestion_control(sk);
5945 
5946 				/* Prevent spurious tcp_cwnd_restart() on
5947 				 * first data packet.
5948 				 */
5949 				tp->lsndtime = tcp_time_stamp;
5950 
5951 				tcp_mtup_init(sk);
5952 				tcp_initialize_rcv_mss(sk);
5953 				tcp_init_buffer_space(sk);
5954 				tcp_fast_path_on(tp);
5955 			} else {
5956 				return 1;
5957 			}
5958 			break;
5959 
5960 		case TCP_FIN_WAIT1:
5961 			if (tp->snd_una == tp->write_seq) {
5962 				tcp_set_state(sk, TCP_FIN_WAIT2);
5963 				sk->sk_shutdown |= SEND_SHUTDOWN;
5964 				dst_confirm(__sk_dst_get(sk));
5965 
5966 				if (!sock_flag(sk, SOCK_DEAD))
5967 					/* Wake up lingering close() */
5968 					sk->sk_state_change(sk);
5969 				else {
5970 					int tmo;
5971 
5972 					if (tp->linger2 < 0 ||
5973 					    (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5974 					     after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5975 						tcp_done(sk);
5976 						NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5977 						return 1;
5978 					}
5979 
5980 					tmo = tcp_fin_time(sk);
5981 					if (tmo > TCP_TIMEWAIT_LEN) {
5982 						inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5983 					} else if (th->fin || sock_owned_by_user(sk)) {
5984 						/* Bad case. We could lose such FIN otherwise.
5985 						 * It is not a big problem, but it looks confusing
5986 						 * and not so rare event. We still can lose it now,
5987 						 * if it spins in bh_lock_sock(), but it is really
5988 						 * marginal case.
5989 						 */
5990 						inet_csk_reset_keepalive_timer(sk, tmo);
5991 					} else {
5992 						tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5993 						goto discard;
5994 					}
5995 				}
5996 			}
5997 			break;
5998 
5999 		case TCP_CLOSING:
6000 			if (tp->snd_una == tp->write_seq) {
6001 				tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6002 				goto discard;
6003 			}
6004 			break;
6005 
6006 		case TCP_LAST_ACK:
6007 			if (tp->snd_una == tp->write_seq) {
6008 				tcp_update_metrics(sk);
6009 				tcp_done(sk);
6010 				goto discard;
6011 			}
6012 			break;
6013 		}
6014 	} else
6015 		goto discard;
6016 
6017 	/* step 6: check the URG bit */
6018 	tcp_urg(sk, skb, th);
6019 
6020 	/* step 7: process the segment text */
6021 	switch (sk->sk_state) {
6022 	case TCP_CLOSE_WAIT:
6023 	case TCP_CLOSING:
6024 	case TCP_LAST_ACK:
6025 		if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6026 			break;
6027 	case TCP_FIN_WAIT1:
6028 	case TCP_FIN_WAIT2:
6029 		/* RFC 793 says to queue data in these states,
6030 		 * RFC 1122 says we MUST send a reset.
6031 		 * BSD 4.4 also does reset.
6032 		 */
6033 		if (sk->sk_shutdown & RCV_SHUTDOWN) {
6034 			if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6035 			    after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6036 				NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6037 				tcp_reset(sk);
6038 				return 1;
6039 			}
6040 		}
6041 		/* Fall through */
6042 	case TCP_ESTABLISHED:
6043 		tcp_data_queue(sk, skb);
6044 		queued = 1;
6045 		break;
6046 	}
6047 
6048 	/* tcp_data could move socket to TIME-WAIT */
6049 	if (sk->sk_state != TCP_CLOSE) {
6050 		tcp_data_snd_check(sk);
6051 		tcp_ack_snd_check(sk);
6052 	}
6053 
6054 	if (!queued) {
6055 discard:
6056 		__kfree_skb(skb);
6057 	}
6058 	return 0;
6059 }
6060 EXPORT_SYMBOL(tcp_rcv_state_process);
6061