xref: /linux/net/ipv4/tcp_output.c (revision 5f4123be3cdb1dbd77fa9d6d2bb96bb9689a0a19)
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:	Pedro Roque	:	Retransmit queue handled by TCP.
23  *				:	Fragmentation on mtu decrease
24  *				:	Segment collapse on retransmit
25  *				:	AF independence
26  *
27  *		Linus Torvalds	:	send_delayed_ack
28  *		David S. Miller	:	Charge memory using the right skb
29  *					during syn/ack processing.
30  *		David S. Miller :	Output engine completely rewritten.
31  *		Andrea Arcangeli:	SYNACK carry ts_recent in tsecr.
32  *		Cacophonix Gaul :	draft-minshall-nagle-01
33  *		J Hadi Salim	:	ECN support
34  *
35  */
36 
37 #include <net/tcp.h>
38 
39 #include <linux/compiler.h>
40 #include <linux/module.h>
41 
42 /* People can turn this off for buggy TCP's found in printers etc. */
43 int sysctl_tcp_retrans_collapse __read_mostly = 1;
44 
45 /* People can turn this on to  work with those rare, broken TCPs that
46  * interpret the window field as a signed quantity.
47  */
48 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
49 
50 /* This limits the percentage of the congestion window which we
51  * will allow a single TSO frame to consume.  Building TSO frames
52  * which are too large can cause TCP streams to be bursty.
53  */
54 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
55 
56 int sysctl_tcp_mtu_probing __read_mostly = 0;
57 int sysctl_tcp_base_mss __read_mostly = 512;
58 
59 /* By default, RFC2861 behavior.  */
60 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
61 
62 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
63 {
64 	struct tcp_sock *tp = tcp_sk(sk);
65 	unsigned int prior_packets = tp->packets_out;
66 
67 	tcp_advance_send_head(sk, skb);
68 	tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
69 
70 	/* Don't override Nagle indefinately with F-RTO */
71 	if (tp->frto_counter == 2)
72 		tp->frto_counter = 3;
73 
74 	tp->packets_out += tcp_skb_pcount(skb);
75 	if (!prior_packets)
76 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
77 					  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
78 }
79 
80 /* SND.NXT, if window was not shrunk.
81  * If window has been shrunk, what should we make? It is not clear at all.
82  * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
83  * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
84  * invalid. OK, let's make this for now:
85  */
86 static inline __u32 tcp_acceptable_seq(struct sock *sk)
87 {
88 	struct tcp_sock *tp = tcp_sk(sk);
89 
90 	if (!before(tcp_wnd_end(tp), tp->snd_nxt))
91 		return tp->snd_nxt;
92 	else
93 		return tcp_wnd_end(tp);
94 }
95 
96 /* Calculate mss to advertise in SYN segment.
97  * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
98  *
99  * 1. It is independent of path mtu.
100  * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
101  * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
102  *    attached devices, because some buggy hosts are confused by
103  *    large MSS.
104  * 4. We do not make 3, we advertise MSS, calculated from first
105  *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
106  *    This may be overridden via information stored in routing table.
107  * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
108  *    probably even Jumbo".
109  */
110 static __u16 tcp_advertise_mss(struct sock *sk)
111 {
112 	struct tcp_sock *tp = tcp_sk(sk);
113 	struct dst_entry *dst = __sk_dst_get(sk);
114 	int mss = tp->advmss;
115 
116 	if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
117 		mss = dst_metric(dst, RTAX_ADVMSS);
118 		tp->advmss = mss;
119 	}
120 
121 	return (__u16)mss;
122 }
123 
124 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
125  * This is the first part of cwnd validation mechanism. */
126 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
127 {
128 	struct tcp_sock *tp = tcp_sk(sk);
129 	s32 delta = tcp_time_stamp - tp->lsndtime;
130 	u32 restart_cwnd = tcp_init_cwnd(tp, dst);
131 	u32 cwnd = tp->snd_cwnd;
132 
133 	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
134 
135 	tp->snd_ssthresh = tcp_current_ssthresh(sk);
136 	restart_cwnd = min(restart_cwnd, cwnd);
137 
138 	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
139 		cwnd >>= 1;
140 	tp->snd_cwnd = max(cwnd, restart_cwnd);
141 	tp->snd_cwnd_stamp = tcp_time_stamp;
142 	tp->snd_cwnd_used = 0;
143 }
144 
145 static void tcp_event_data_sent(struct tcp_sock *tp,
146 				struct sk_buff *skb, struct sock *sk)
147 {
148 	struct inet_connection_sock *icsk = inet_csk(sk);
149 	const u32 now = tcp_time_stamp;
150 
151 	if (sysctl_tcp_slow_start_after_idle &&
152 	    (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
153 		tcp_cwnd_restart(sk, __sk_dst_get(sk));
154 
155 	tp->lsndtime = now;
156 
157 	/* If it is a reply for ato after last received
158 	 * packet, enter pingpong mode.
159 	 */
160 	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
161 		icsk->icsk_ack.pingpong = 1;
162 }
163 
164 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
165 {
166 	tcp_dec_quickack_mode(sk, pkts);
167 	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
168 }
169 
170 /* Determine a window scaling and initial window to offer.
171  * Based on the assumption that the given amount of space
172  * will be offered. Store the results in the tp structure.
173  * NOTE: for smooth operation initial space offering should
174  * be a multiple of mss if possible. We assume here that mss >= 1.
175  * This MUST be enforced by all callers.
176  */
177 void tcp_select_initial_window(int __space, __u32 mss,
178 			       __u32 *rcv_wnd, __u32 *window_clamp,
179 			       int wscale_ok, __u8 *rcv_wscale)
180 {
181 	unsigned int space = (__space < 0 ? 0 : __space);
182 
183 	/* If no clamp set the clamp to the max possible scaled window */
184 	if (*window_clamp == 0)
185 		(*window_clamp) = (65535 << 14);
186 	space = min(*window_clamp, space);
187 
188 	/* Quantize space offering to a multiple of mss if possible. */
189 	if (space > mss)
190 		space = (space / mss) * mss;
191 
192 	/* NOTE: offering an initial window larger than 32767
193 	 * will break some buggy TCP stacks. If the admin tells us
194 	 * it is likely we could be speaking with such a buggy stack
195 	 * we will truncate our initial window offering to 32K-1
196 	 * unless the remote has sent us a window scaling option,
197 	 * which we interpret as a sign the remote TCP is not
198 	 * misinterpreting the window field as a signed quantity.
199 	 */
200 	if (sysctl_tcp_workaround_signed_windows)
201 		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
202 	else
203 		(*rcv_wnd) = space;
204 
205 	(*rcv_wscale) = 0;
206 	if (wscale_ok) {
207 		/* Set window scaling on max possible window
208 		 * See RFC1323 for an explanation of the limit to 14
209 		 */
210 		space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
211 		space = min_t(u32, space, *window_clamp);
212 		while (space > 65535 && (*rcv_wscale) < 14) {
213 			space >>= 1;
214 			(*rcv_wscale)++;
215 		}
216 	}
217 
218 	/* Set initial window to value enough for senders,
219 	 * following RFC2414. Senders, not following this RFC,
220 	 * will be satisfied with 2.
221 	 */
222 	if (mss > (1 << *rcv_wscale)) {
223 		int init_cwnd = 4;
224 		if (mss > 1460 * 3)
225 			init_cwnd = 2;
226 		else if (mss > 1460)
227 			init_cwnd = 3;
228 		if (*rcv_wnd > init_cwnd * mss)
229 			*rcv_wnd = init_cwnd * mss;
230 	}
231 
232 	/* Set the clamp no higher than max representable value */
233 	(*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
234 }
235 
236 /* Chose a new window to advertise, update state in tcp_sock for the
237  * socket, and return result with RFC1323 scaling applied.  The return
238  * value can be stuffed directly into th->window for an outgoing
239  * frame.
240  */
241 static u16 tcp_select_window(struct sock *sk)
242 {
243 	struct tcp_sock *tp = tcp_sk(sk);
244 	u32 cur_win = tcp_receive_window(tp);
245 	u32 new_win = __tcp_select_window(sk);
246 
247 	/* Never shrink the offered window */
248 	if (new_win < cur_win) {
249 		/* Danger Will Robinson!
250 		 * Don't update rcv_wup/rcv_wnd here or else
251 		 * we will not be able to advertise a zero
252 		 * window in time.  --DaveM
253 		 *
254 		 * Relax Will Robinson.
255 		 */
256 		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
257 	}
258 	tp->rcv_wnd = new_win;
259 	tp->rcv_wup = tp->rcv_nxt;
260 
261 	/* Make sure we do not exceed the maximum possible
262 	 * scaled window.
263 	 */
264 	if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
265 		new_win = min(new_win, MAX_TCP_WINDOW);
266 	else
267 		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
268 
269 	/* RFC1323 scaling applied */
270 	new_win >>= tp->rx_opt.rcv_wscale;
271 
272 	/* If we advertise zero window, disable fast path. */
273 	if (new_win == 0)
274 		tp->pred_flags = 0;
275 
276 	return new_win;
277 }
278 
279 static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
280 {
281 	TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
282 	if (!(tp->ecn_flags & TCP_ECN_OK))
283 		TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
284 }
285 
286 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
287 {
288 	struct tcp_sock *tp = tcp_sk(sk);
289 
290 	tp->ecn_flags = 0;
291 	if (sysctl_tcp_ecn) {
292 		TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR;
293 		tp->ecn_flags = TCP_ECN_OK;
294 	}
295 }
296 
297 static __inline__ void
298 TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
299 {
300 	if (inet_rsk(req)->ecn_ok)
301 		th->ece = 1;
302 }
303 
304 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
305 				int tcp_header_len)
306 {
307 	struct tcp_sock *tp = tcp_sk(sk);
308 
309 	if (tp->ecn_flags & TCP_ECN_OK) {
310 		/* Not-retransmitted data segment: set ECT and inject CWR. */
311 		if (skb->len != tcp_header_len &&
312 		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
313 			INET_ECN_xmit(sk);
314 			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
315 				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
316 				tcp_hdr(skb)->cwr = 1;
317 				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
318 			}
319 		} else {
320 			/* ACK or retransmitted segment: clear ECT|CE */
321 			INET_ECN_dontxmit(sk);
322 		}
323 		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
324 			tcp_hdr(skb)->ece = 1;
325 	}
326 }
327 
328 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
329  * auto increment end seqno.
330  */
331 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
332 {
333 	skb->csum = 0;
334 
335 	TCP_SKB_CB(skb)->flags = flags;
336 	TCP_SKB_CB(skb)->sacked = 0;
337 
338 	skb_shinfo(skb)->gso_segs = 1;
339 	skb_shinfo(skb)->gso_size = 0;
340 	skb_shinfo(skb)->gso_type = 0;
341 
342 	TCP_SKB_CB(skb)->seq = seq;
343 	if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN))
344 		seq++;
345 	TCP_SKB_CB(skb)->end_seq = seq;
346 }
347 
348 static inline int tcp_urg_mode(const struct tcp_sock *tp)
349 {
350 	return tp->snd_una != tp->snd_up;
351 }
352 
353 #define OPTION_SACK_ADVERTISE	(1 << 0)
354 #define OPTION_TS		(1 << 1)
355 #define OPTION_MD5		(1 << 2)
356 
357 struct tcp_out_options {
358 	u8 options;		/* bit field of OPTION_* */
359 	u8 ws;			/* window scale, 0 to disable */
360 	u8 num_sack_blocks;	/* number of SACK blocks to include */
361 	u16 mss;		/* 0 to disable */
362 	__u32 tsval, tsecr;	/* need to include OPTION_TS */
363 };
364 
365 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
366 			      const struct tcp_out_options *opts,
367 			      __u8 **md5_hash) {
368 	if (unlikely(OPTION_MD5 & opts->options)) {
369 		*ptr++ = htonl((TCPOPT_NOP << 24) |
370 			       (TCPOPT_NOP << 16) |
371 			       (TCPOPT_MD5SIG << 8) |
372 			       TCPOLEN_MD5SIG);
373 		*md5_hash = (__u8 *)ptr;
374 		ptr += 4;
375 	} else {
376 		*md5_hash = NULL;
377 	}
378 
379 	if (likely(OPTION_TS & opts->options)) {
380 		if (unlikely(OPTION_SACK_ADVERTISE & opts->options)) {
381 			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
382 				       (TCPOLEN_SACK_PERM << 16) |
383 				       (TCPOPT_TIMESTAMP << 8) |
384 				       TCPOLEN_TIMESTAMP);
385 		} else {
386 			*ptr++ = htonl((TCPOPT_NOP << 24) |
387 				       (TCPOPT_NOP << 16) |
388 				       (TCPOPT_TIMESTAMP << 8) |
389 				       TCPOLEN_TIMESTAMP);
390 		}
391 		*ptr++ = htonl(opts->tsval);
392 		*ptr++ = htonl(opts->tsecr);
393 	}
394 
395 	if (unlikely(opts->mss)) {
396 		*ptr++ = htonl((TCPOPT_MSS << 24) |
397 			       (TCPOLEN_MSS << 16) |
398 			       opts->mss);
399 	}
400 
401 	if (unlikely(OPTION_SACK_ADVERTISE & opts->options &&
402 		     !(OPTION_TS & opts->options))) {
403 		*ptr++ = htonl((TCPOPT_NOP << 24) |
404 			       (TCPOPT_NOP << 16) |
405 			       (TCPOPT_SACK_PERM << 8) |
406 			       TCPOLEN_SACK_PERM);
407 	}
408 
409 	if (unlikely(opts->ws)) {
410 		*ptr++ = htonl((TCPOPT_NOP << 24) |
411 			       (TCPOPT_WINDOW << 16) |
412 			       (TCPOLEN_WINDOW << 8) |
413 			       opts->ws);
414 	}
415 
416 	if (unlikely(opts->num_sack_blocks)) {
417 		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
418 			tp->duplicate_sack : tp->selective_acks;
419 		int this_sack;
420 
421 		*ptr++ = htonl((TCPOPT_NOP  << 24) |
422 			       (TCPOPT_NOP  << 16) |
423 			       (TCPOPT_SACK <<  8) |
424 			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
425 						     TCPOLEN_SACK_PERBLOCK)));
426 
427 		for (this_sack = 0; this_sack < opts->num_sack_blocks;
428 		     ++this_sack) {
429 			*ptr++ = htonl(sp[this_sack].start_seq);
430 			*ptr++ = htonl(sp[this_sack].end_seq);
431 		}
432 
433 		if (tp->rx_opt.dsack) {
434 			tp->rx_opt.dsack = 0;
435 			tp->rx_opt.eff_sacks--;
436 		}
437 	}
438 }
439 
440 static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
441 				struct tcp_out_options *opts,
442 				struct tcp_md5sig_key **md5) {
443 	struct tcp_sock *tp = tcp_sk(sk);
444 	unsigned size = 0;
445 
446 #ifdef CONFIG_TCP_MD5SIG
447 	*md5 = tp->af_specific->md5_lookup(sk, sk);
448 	if (*md5) {
449 		opts->options |= OPTION_MD5;
450 		size += TCPOLEN_MD5SIG_ALIGNED;
451 	}
452 #else
453 	*md5 = NULL;
454 #endif
455 
456 	/* We always get an MSS option.  The option bytes which will be seen in
457 	 * normal data packets should timestamps be used, must be in the MSS
458 	 * advertised.  But we subtract them from tp->mss_cache so that
459 	 * calculations in tcp_sendmsg are simpler etc.  So account for this
460 	 * fact here if necessary.  If we don't do this correctly, as a
461 	 * receiver we won't recognize data packets as being full sized when we
462 	 * should, and thus we won't abide by the delayed ACK rules correctly.
463 	 * SACKs don't matter, we never delay an ACK when we have any of those
464 	 * going out.  */
465 	opts->mss = tcp_advertise_mss(sk);
466 	size += TCPOLEN_MSS_ALIGNED;
467 
468 	if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
469 		opts->options |= OPTION_TS;
470 		opts->tsval = TCP_SKB_CB(skb)->when;
471 		opts->tsecr = tp->rx_opt.ts_recent;
472 		size += TCPOLEN_TSTAMP_ALIGNED;
473 	}
474 	if (likely(sysctl_tcp_window_scaling)) {
475 		opts->ws = tp->rx_opt.rcv_wscale;
476 		if(likely(opts->ws))
477 			size += TCPOLEN_WSCALE_ALIGNED;
478 	}
479 	if (likely(sysctl_tcp_sack)) {
480 		opts->options |= OPTION_SACK_ADVERTISE;
481 		if (unlikely(!(OPTION_TS & opts->options)))
482 			size += TCPOLEN_SACKPERM_ALIGNED;
483 	}
484 
485 	return size;
486 }
487 
488 static unsigned tcp_synack_options(struct sock *sk,
489 				   struct request_sock *req,
490 				   unsigned mss, struct sk_buff *skb,
491 				   struct tcp_out_options *opts,
492 				   struct tcp_md5sig_key **md5) {
493 	unsigned size = 0;
494 	struct inet_request_sock *ireq = inet_rsk(req);
495 	char doing_ts;
496 
497 #ifdef CONFIG_TCP_MD5SIG
498 	*md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
499 	if (*md5) {
500 		opts->options |= OPTION_MD5;
501 		size += TCPOLEN_MD5SIG_ALIGNED;
502 	}
503 #else
504 	*md5 = NULL;
505 #endif
506 
507 	/* we can't fit any SACK blocks in a packet with MD5 + TS
508 	   options. There was discussion about disabling SACK rather than TS in
509 	   order to fit in better with old, buggy kernels, but that was deemed
510 	   to be unnecessary. */
511 	doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok);
512 
513 	opts->mss = mss;
514 	size += TCPOLEN_MSS_ALIGNED;
515 
516 	if (likely(ireq->wscale_ok)) {
517 		opts->ws = ireq->rcv_wscale;
518 		if(likely(opts->ws))
519 			size += TCPOLEN_WSCALE_ALIGNED;
520 	}
521 	if (likely(doing_ts)) {
522 		opts->options |= OPTION_TS;
523 		opts->tsval = TCP_SKB_CB(skb)->when;
524 		opts->tsecr = req->ts_recent;
525 		size += TCPOLEN_TSTAMP_ALIGNED;
526 	}
527 	if (likely(ireq->sack_ok)) {
528 		opts->options |= OPTION_SACK_ADVERTISE;
529 		if (unlikely(!doing_ts))
530 			size += TCPOLEN_SACKPERM_ALIGNED;
531 	}
532 
533 	return size;
534 }
535 
536 static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
537 					struct tcp_out_options *opts,
538 					struct tcp_md5sig_key **md5) {
539 	struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
540 	struct tcp_sock *tp = tcp_sk(sk);
541 	unsigned size = 0;
542 
543 #ifdef CONFIG_TCP_MD5SIG
544 	*md5 = tp->af_specific->md5_lookup(sk, sk);
545 	if (unlikely(*md5)) {
546 		opts->options |= OPTION_MD5;
547 		size += TCPOLEN_MD5SIG_ALIGNED;
548 	}
549 #else
550 	*md5 = NULL;
551 #endif
552 
553 	if (likely(tp->rx_opt.tstamp_ok)) {
554 		opts->options |= OPTION_TS;
555 		opts->tsval = tcb ? tcb->when : 0;
556 		opts->tsecr = tp->rx_opt.ts_recent;
557 		size += TCPOLEN_TSTAMP_ALIGNED;
558 	}
559 
560 	if (unlikely(tp->rx_opt.eff_sacks)) {
561 		const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
562 		opts->num_sack_blocks =
563 			min_t(unsigned, tp->rx_opt.eff_sacks,
564 			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
565 			      TCPOLEN_SACK_PERBLOCK);
566 		size += TCPOLEN_SACK_BASE_ALIGNED +
567 			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
568 	}
569 
570 	return size;
571 }
572 
573 /* This routine actually transmits TCP packets queued in by
574  * tcp_do_sendmsg().  This is used by both the initial
575  * transmission and possible later retransmissions.
576  * All SKB's seen here are completely headerless.  It is our
577  * job to build the TCP header, and pass the packet down to
578  * IP so it can do the same plus pass the packet off to the
579  * device.
580  *
581  * We are working here with either a clone of the original
582  * SKB, or a fresh unique copy made by the retransmit engine.
583  */
584 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
585 			    gfp_t gfp_mask)
586 {
587 	const struct inet_connection_sock *icsk = inet_csk(sk);
588 	struct inet_sock *inet;
589 	struct tcp_sock *tp;
590 	struct tcp_skb_cb *tcb;
591 	struct tcp_out_options opts;
592 	unsigned tcp_options_size, tcp_header_size;
593 	struct tcp_md5sig_key *md5;
594 	__u8 *md5_hash_location;
595 	struct tcphdr *th;
596 	int err;
597 
598 	BUG_ON(!skb || !tcp_skb_pcount(skb));
599 
600 	/* If congestion control is doing timestamping, we must
601 	 * take such a timestamp before we potentially clone/copy.
602 	 */
603 	if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
604 		__net_timestamp(skb);
605 
606 	if (likely(clone_it)) {
607 		if (unlikely(skb_cloned(skb)))
608 			skb = pskb_copy(skb, gfp_mask);
609 		else
610 			skb = skb_clone(skb, gfp_mask);
611 		if (unlikely(!skb))
612 			return -ENOBUFS;
613 	}
614 
615 	inet = inet_sk(sk);
616 	tp = tcp_sk(sk);
617 	tcb = TCP_SKB_CB(skb);
618 	memset(&opts, 0, sizeof(opts));
619 
620 	if (unlikely(tcb->flags & TCPCB_FLAG_SYN))
621 		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
622 	else
623 		tcp_options_size = tcp_established_options(sk, skb, &opts,
624 							   &md5);
625 	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
626 
627 	if (tcp_packets_in_flight(tp) == 0)
628 		tcp_ca_event(sk, CA_EVENT_TX_START);
629 
630 	skb_push(skb, tcp_header_size);
631 	skb_reset_transport_header(skb);
632 	skb_set_owner_w(skb, sk);
633 
634 	/* Build TCP header and checksum it. */
635 	th = tcp_hdr(skb);
636 	th->source		= inet->sport;
637 	th->dest		= inet->dport;
638 	th->seq			= htonl(tcb->seq);
639 	th->ack_seq		= htonl(tp->rcv_nxt);
640 	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
641 					tcb->flags);
642 
643 	if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
644 		/* RFC1323: The window in SYN & SYN/ACK segments
645 		 * is never scaled.
646 		 */
647 		th->window	= htons(min(tp->rcv_wnd, 65535U));
648 	} else {
649 		th->window	= htons(tcp_select_window(sk));
650 	}
651 	th->check		= 0;
652 	th->urg_ptr		= 0;
653 
654 	/* The urg_mode check is necessary during a below snd_una win probe */
655 	if (unlikely(tcp_urg_mode(tp) &&
656 		     between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) {
657 		th->urg_ptr		= htons(tp->snd_up - tcb->seq);
658 		th->urg			= 1;
659 	}
660 
661 	tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
662 	if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0))
663 		TCP_ECN_send(sk, skb, tcp_header_size);
664 
665 #ifdef CONFIG_TCP_MD5SIG
666 	/* Calculate the MD5 hash, as we have all we need now */
667 	if (md5) {
668 		sk->sk_route_caps &= ~NETIF_F_GSO_MASK;
669 		tp->af_specific->calc_md5_hash(md5_hash_location,
670 					       md5, sk, NULL, skb);
671 	}
672 #endif
673 
674 	icsk->icsk_af_ops->send_check(sk, skb->len, skb);
675 
676 	if (likely(tcb->flags & TCPCB_FLAG_ACK))
677 		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
678 
679 	if (skb->len != tcp_header_size)
680 		tcp_event_data_sent(tp, skb, sk);
681 
682 	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
683 		TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
684 
685 	err = icsk->icsk_af_ops->queue_xmit(skb, 0);
686 	if (likely(err <= 0))
687 		return err;
688 
689 	tcp_enter_cwr(sk, 1);
690 
691 	return net_xmit_eval(err);
692 }
693 
694 /* This routine just queue's the buffer
695  *
696  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
697  * otherwise socket can stall.
698  */
699 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
700 {
701 	struct tcp_sock *tp = tcp_sk(sk);
702 
703 	/* Advance write_seq and place onto the write_queue. */
704 	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
705 	skb_header_release(skb);
706 	tcp_add_write_queue_tail(sk, skb);
707 	sk->sk_wmem_queued += skb->truesize;
708 	sk_mem_charge(sk, skb->truesize);
709 }
710 
711 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
712 				 unsigned int mss_now)
713 {
714 	if (skb->len <= mss_now || !sk_can_gso(sk)) {
715 		/* Avoid the costly divide in the normal
716 		 * non-TSO case.
717 		 */
718 		skb_shinfo(skb)->gso_segs = 1;
719 		skb_shinfo(skb)->gso_size = 0;
720 		skb_shinfo(skb)->gso_type = 0;
721 	} else {
722 		skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
723 		skb_shinfo(skb)->gso_size = mss_now;
724 		skb_shinfo(skb)->gso_type = sk->sk_gso_type;
725 	}
726 }
727 
728 /* When a modification to fackets out becomes necessary, we need to check
729  * skb is counted to fackets_out or not.
730  */
731 static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
732 				   int decr)
733 {
734 	struct tcp_sock *tp = tcp_sk(sk);
735 
736 	if (!tp->sacked_out || tcp_is_reno(tp))
737 		return;
738 
739 	if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
740 		tp->fackets_out -= decr;
741 }
742 
743 /* Function to create two new TCP segments.  Shrinks the given segment
744  * to the specified size and appends a new segment with the rest of the
745  * packet to the list.  This won't be called frequently, I hope.
746  * Remember, these are still headerless SKBs at this point.
747  */
748 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
749 		 unsigned int mss_now)
750 {
751 	struct tcp_sock *tp = tcp_sk(sk);
752 	struct sk_buff *buff;
753 	int nsize, old_factor;
754 	int nlen;
755 	u16 flags;
756 
757 	BUG_ON(len > skb->len);
758 
759 	tcp_clear_retrans_hints_partial(tp);
760 	nsize = skb_headlen(skb) - len;
761 	if (nsize < 0)
762 		nsize = 0;
763 
764 	if (skb_cloned(skb) &&
765 	    skb_is_nonlinear(skb) &&
766 	    pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
767 		return -ENOMEM;
768 
769 	/* Get a new skb... force flag on. */
770 	buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
771 	if (buff == NULL)
772 		return -ENOMEM; /* We'll just try again later. */
773 
774 	sk->sk_wmem_queued += buff->truesize;
775 	sk_mem_charge(sk, buff->truesize);
776 	nlen = skb->len - len - nsize;
777 	buff->truesize += nlen;
778 	skb->truesize -= nlen;
779 
780 	/* Correct the sequence numbers. */
781 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
782 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
783 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
784 
785 	/* PSH and FIN should only be set in the second packet. */
786 	flags = TCP_SKB_CB(skb)->flags;
787 	TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
788 	TCP_SKB_CB(buff)->flags = flags;
789 	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
790 
791 	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
792 		/* Copy and checksum data tail into the new buffer. */
793 		buff->csum = csum_partial_copy_nocheck(skb->data + len,
794 						       skb_put(buff, nsize),
795 						       nsize, 0);
796 
797 		skb_trim(skb, len);
798 
799 		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
800 	} else {
801 		skb->ip_summed = CHECKSUM_PARTIAL;
802 		skb_split(skb, buff, len);
803 	}
804 
805 	buff->ip_summed = skb->ip_summed;
806 
807 	/* Looks stupid, but our code really uses when of
808 	 * skbs, which it never sent before. --ANK
809 	 */
810 	TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
811 	buff->tstamp = skb->tstamp;
812 
813 	old_factor = tcp_skb_pcount(skb);
814 
815 	/* Fix up tso_factor for both original and new SKB.  */
816 	tcp_set_skb_tso_segs(sk, skb, mss_now);
817 	tcp_set_skb_tso_segs(sk, buff, mss_now);
818 
819 	/* If this packet has been sent out already, we must
820 	 * adjust the various packet counters.
821 	 */
822 	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
823 		int diff = old_factor - tcp_skb_pcount(skb) -
824 			tcp_skb_pcount(buff);
825 
826 		tp->packets_out -= diff;
827 
828 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
829 			tp->sacked_out -= diff;
830 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
831 			tp->retrans_out -= diff;
832 
833 		if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
834 			tp->lost_out -= diff;
835 
836 		/* Adjust Reno SACK estimate. */
837 		if (tcp_is_reno(tp) && diff > 0) {
838 			tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
839 			tcp_verify_left_out(tp);
840 		}
841 		tcp_adjust_fackets_out(sk, skb, diff);
842 	}
843 
844 	/* Link BUFF into the send queue. */
845 	skb_header_release(buff);
846 	tcp_insert_write_queue_after(skb, buff, sk);
847 
848 	return 0;
849 }
850 
851 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
852  * eventually). The difference is that pulled data not copied, but
853  * immediately discarded.
854  */
855 static void __pskb_trim_head(struct sk_buff *skb, int len)
856 {
857 	int i, k, eat;
858 
859 	eat = len;
860 	k = 0;
861 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
862 		if (skb_shinfo(skb)->frags[i].size <= eat) {
863 			put_page(skb_shinfo(skb)->frags[i].page);
864 			eat -= skb_shinfo(skb)->frags[i].size;
865 		} else {
866 			skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
867 			if (eat) {
868 				skb_shinfo(skb)->frags[k].page_offset += eat;
869 				skb_shinfo(skb)->frags[k].size -= eat;
870 				eat = 0;
871 			}
872 			k++;
873 		}
874 	}
875 	skb_shinfo(skb)->nr_frags = k;
876 
877 	skb_reset_tail_pointer(skb);
878 	skb->data_len -= len;
879 	skb->len = skb->data_len;
880 }
881 
882 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
883 {
884 	if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
885 		return -ENOMEM;
886 
887 	/* If len == headlen, we avoid __skb_pull to preserve alignment. */
888 	if (unlikely(len < skb_headlen(skb)))
889 		__skb_pull(skb, len);
890 	else
891 		__pskb_trim_head(skb, len - skb_headlen(skb));
892 
893 	TCP_SKB_CB(skb)->seq += len;
894 	skb->ip_summed = CHECKSUM_PARTIAL;
895 
896 	skb->truesize	     -= len;
897 	sk->sk_wmem_queued   -= len;
898 	sk_mem_uncharge(sk, len);
899 	sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
900 
901 	/* Any change of skb->len requires recalculation of tso
902 	 * factor and mss.
903 	 */
904 	if (tcp_skb_pcount(skb) > 1)
905 		tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
906 
907 	return 0;
908 }
909 
910 /* Not accounting for SACKs here. */
911 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
912 {
913 	struct tcp_sock *tp = tcp_sk(sk);
914 	struct inet_connection_sock *icsk = inet_csk(sk);
915 	int mss_now;
916 
917 	/* Calculate base mss without TCP options:
918 	   It is MMS_S - sizeof(tcphdr) of rfc1122
919 	 */
920 	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
921 
922 	/* Clamp it (mss_clamp does not include tcp options) */
923 	if (mss_now > tp->rx_opt.mss_clamp)
924 		mss_now = tp->rx_opt.mss_clamp;
925 
926 	/* Now subtract optional transport overhead */
927 	mss_now -= icsk->icsk_ext_hdr_len;
928 
929 	/* Then reserve room for full set of TCP options and 8 bytes of data */
930 	if (mss_now < 48)
931 		mss_now = 48;
932 
933 	/* Now subtract TCP options size, not including SACKs */
934 	mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
935 
936 	return mss_now;
937 }
938 
939 /* Inverse of above */
940 int tcp_mss_to_mtu(struct sock *sk, int mss)
941 {
942 	struct tcp_sock *tp = tcp_sk(sk);
943 	struct inet_connection_sock *icsk = inet_csk(sk);
944 	int mtu;
945 
946 	mtu = mss +
947 	      tp->tcp_header_len +
948 	      icsk->icsk_ext_hdr_len +
949 	      icsk->icsk_af_ops->net_header_len;
950 
951 	return mtu;
952 }
953 
954 void tcp_mtup_init(struct sock *sk)
955 {
956 	struct tcp_sock *tp = tcp_sk(sk);
957 	struct inet_connection_sock *icsk = inet_csk(sk);
958 
959 	icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
960 	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
961 			       icsk->icsk_af_ops->net_header_len;
962 	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
963 	icsk->icsk_mtup.probe_size = 0;
964 }
965 
966 /* Bound MSS / TSO packet size with the half of the window */
967 static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
968 {
969 	if (tp->max_window && pktsize > (tp->max_window >> 1))
970 		return max(tp->max_window >> 1, 68U - tp->tcp_header_len);
971 	else
972 		return pktsize;
973 }
974 
975 /* This function synchronize snd mss to current pmtu/exthdr set.
976 
977    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
978    for TCP options, but includes only bare TCP header.
979 
980    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
981    It is minimum of user_mss and mss received with SYN.
982    It also does not include TCP options.
983 
984    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
985 
986    tp->mss_cache is current effective sending mss, including
987    all tcp options except for SACKs. It is evaluated,
988    taking into account current pmtu, but never exceeds
989    tp->rx_opt.mss_clamp.
990 
991    NOTE1. rfc1122 clearly states that advertised MSS
992    DOES NOT include either tcp or ip options.
993 
994    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
995    are READ ONLY outside this function.		--ANK (980731)
996  */
997 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
998 {
999 	struct tcp_sock *tp = tcp_sk(sk);
1000 	struct inet_connection_sock *icsk = inet_csk(sk);
1001 	int mss_now;
1002 
1003 	if (icsk->icsk_mtup.search_high > pmtu)
1004 		icsk->icsk_mtup.search_high = pmtu;
1005 
1006 	mss_now = tcp_mtu_to_mss(sk, pmtu);
1007 	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1008 
1009 	/* And store cached results */
1010 	icsk->icsk_pmtu_cookie = pmtu;
1011 	if (icsk->icsk_mtup.enabled)
1012 		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1013 	tp->mss_cache = mss_now;
1014 
1015 	return mss_now;
1016 }
1017 
1018 /* Compute the current effective MSS, taking SACKs and IP options,
1019  * and even PMTU discovery events into account.
1020  *
1021  * LARGESEND note: !tcp_urg_mode is overkill, only frames up to snd_up
1022  * cannot be large. However, taking into account rare use of URG, this
1023  * is not a big flaw.
1024  */
1025 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
1026 {
1027 	struct tcp_sock *tp = tcp_sk(sk);
1028 	struct dst_entry *dst = __sk_dst_get(sk);
1029 	u32 mss_now;
1030 	u16 xmit_size_goal;
1031 	int doing_tso = 0;
1032 	unsigned header_len;
1033 	struct tcp_out_options opts;
1034 	struct tcp_md5sig_key *md5;
1035 
1036 	mss_now = tp->mss_cache;
1037 
1038 	if (large_allowed && sk_can_gso(sk) && !tcp_urg_mode(tp))
1039 		doing_tso = 1;
1040 
1041 	if (dst) {
1042 		u32 mtu = dst_mtu(dst);
1043 		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1044 			mss_now = tcp_sync_mss(sk, mtu);
1045 	}
1046 
1047 	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1048 		     sizeof(struct tcphdr);
1049 	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1050 	 * some common options. If this is an odd packet (because we have SACK
1051 	 * blocks etc) then our calculated header_len will be different, and
1052 	 * we have to adjust mss_now correspondingly */
1053 	if (header_len != tp->tcp_header_len) {
1054 		int delta = (int) header_len - tp->tcp_header_len;
1055 		mss_now -= delta;
1056 	}
1057 
1058 	xmit_size_goal = mss_now;
1059 
1060 	if (doing_tso) {
1061 		xmit_size_goal = ((sk->sk_gso_max_size - 1) -
1062 				  inet_csk(sk)->icsk_af_ops->net_header_len -
1063 				  inet_csk(sk)->icsk_ext_hdr_len -
1064 				  tp->tcp_header_len);
1065 
1066 		xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
1067 		xmit_size_goal -= (xmit_size_goal % mss_now);
1068 	}
1069 	tp->xmit_size_goal = xmit_size_goal;
1070 
1071 	return mss_now;
1072 }
1073 
1074 /* Congestion window validation. (RFC2861) */
1075 static void tcp_cwnd_validate(struct sock *sk)
1076 {
1077 	struct tcp_sock *tp = tcp_sk(sk);
1078 
1079 	if (tp->packets_out >= tp->snd_cwnd) {
1080 		/* Network is feed fully. */
1081 		tp->snd_cwnd_used = 0;
1082 		tp->snd_cwnd_stamp = tcp_time_stamp;
1083 	} else {
1084 		/* Network starves. */
1085 		if (tp->packets_out > tp->snd_cwnd_used)
1086 			tp->snd_cwnd_used = tp->packets_out;
1087 
1088 		if (sysctl_tcp_slow_start_after_idle &&
1089 		    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1090 			tcp_cwnd_application_limited(sk);
1091 	}
1092 }
1093 
1094 /* Returns the portion of skb which can be sent right away without
1095  * introducing MSS oddities to segment boundaries. In rare cases where
1096  * mss_now != mss_cache, we will request caller to create a small skb
1097  * per input skb which could be mostly avoided here (if desired).
1098  *
1099  * We explicitly want to create a request for splitting write queue tail
1100  * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1101  * thus all the complexity (cwnd_len is always MSS multiple which we
1102  * return whenever allowed by the other factors). Basically we need the
1103  * modulo only when the receiver window alone is the limiting factor or
1104  * when we would be allowed to send the split-due-to-Nagle skb fully.
1105  */
1106 static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1107 					unsigned int mss_now, unsigned int cwnd)
1108 {
1109 	struct tcp_sock *tp = tcp_sk(sk);
1110 	u32 needed, window, cwnd_len;
1111 
1112 	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1113 	cwnd_len = mss_now * cwnd;
1114 
1115 	if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1116 		return cwnd_len;
1117 
1118 	needed = min(skb->len, window);
1119 
1120 	if (cwnd_len <= needed)
1121 		return cwnd_len;
1122 
1123 	return needed - needed % mss_now;
1124 }
1125 
1126 /* Can at least one segment of SKB be sent right now, according to the
1127  * congestion window rules?  If so, return how many segments are allowed.
1128  */
1129 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1130 					 struct sk_buff *skb)
1131 {
1132 	u32 in_flight, cwnd;
1133 
1134 	/* Don't be strict about the congestion window for the final FIN.  */
1135 	if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1136 	    tcp_skb_pcount(skb) == 1)
1137 		return 1;
1138 
1139 	in_flight = tcp_packets_in_flight(tp);
1140 	cwnd = tp->snd_cwnd;
1141 	if (in_flight < cwnd)
1142 		return (cwnd - in_flight);
1143 
1144 	return 0;
1145 }
1146 
1147 /* This must be invoked the first time we consider transmitting
1148  * SKB onto the wire.
1149  */
1150 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1151 			     unsigned int mss_now)
1152 {
1153 	int tso_segs = tcp_skb_pcount(skb);
1154 
1155 	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1156 		tcp_set_skb_tso_segs(sk, skb, mss_now);
1157 		tso_segs = tcp_skb_pcount(skb);
1158 	}
1159 	return tso_segs;
1160 }
1161 
1162 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1163 {
1164 	return after(tp->snd_sml,tp->snd_una) &&
1165 		!after(tp->snd_sml, tp->snd_nxt);
1166 }
1167 
1168 /* Return 0, if packet can be sent now without violation Nagle's rules:
1169  * 1. It is full sized.
1170  * 2. Or it contains FIN. (already checked by caller)
1171  * 3. Or TCP_NODELAY was set.
1172  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1173  *    With Minshall's modification: all sent small packets are ACKed.
1174  */
1175 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1176 				  const struct sk_buff *skb,
1177 				  unsigned mss_now, int nonagle)
1178 {
1179 	return (skb->len < mss_now &&
1180 		((nonagle & TCP_NAGLE_CORK) ||
1181 		 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
1182 }
1183 
1184 /* Return non-zero if the Nagle test allows this packet to be
1185  * sent now.
1186  */
1187 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1188 				 unsigned int cur_mss, int nonagle)
1189 {
1190 	/* Nagle rule does not apply to frames, which sit in the middle of the
1191 	 * write_queue (they have no chances to get new data).
1192 	 *
1193 	 * This is implemented in the callers, where they modify the 'nonagle'
1194 	 * argument based upon the location of SKB in the send queue.
1195 	 */
1196 	if (nonagle & TCP_NAGLE_PUSH)
1197 		return 1;
1198 
1199 	/* Don't use the nagle rule for urgent data (or for the final FIN).
1200 	 * Nagle can be ignored during F-RTO too (see RFC4138).
1201 	 */
1202 	if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1203 	    (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1204 		return 1;
1205 
1206 	if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1207 		return 1;
1208 
1209 	return 0;
1210 }
1211 
1212 /* Does at least the first segment of SKB fit into the send window? */
1213 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1214 				   unsigned int cur_mss)
1215 {
1216 	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1217 
1218 	if (skb->len > cur_mss)
1219 		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1220 
1221 	return !after(end_seq, tcp_wnd_end(tp));
1222 }
1223 
1224 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1225  * should be put on the wire right now.  If so, it returns the number of
1226  * packets allowed by the congestion window.
1227  */
1228 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1229 				 unsigned int cur_mss, int nonagle)
1230 {
1231 	struct tcp_sock *tp = tcp_sk(sk);
1232 	unsigned int cwnd_quota;
1233 
1234 	tcp_init_tso_segs(sk, skb, cur_mss);
1235 
1236 	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1237 		return 0;
1238 
1239 	cwnd_quota = tcp_cwnd_test(tp, skb);
1240 	if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1241 		cwnd_quota = 0;
1242 
1243 	return cwnd_quota;
1244 }
1245 
1246 int tcp_may_send_now(struct sock *sk)
1247 {
1248 	struct tcp_sock *tp = tcp_sk(sk);
1249 	struct sk_buff *skb = tcp_send_head(sk);
1250 
1251 	return (skb &&
1252 		tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1253 			     (tcp_skb_is_last(sk, skb) ?
1254 			      tp->nonagle : TCP_NAGLE_PUSH)));
1255 }
1256 
1257 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1258  * which is put after SKB on the list.  It is very much like
1259  * tcp_fragment() except that it may make several kinds of assumptions
1260  * in order to speed up the splitting operation.  In particular, we
1261  * know that all the data is in scatter-gather pages, and that the
1262  * packet has never been sent out before (and thus is not cloned).
1263  */
1264 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1265 			unsigned int mss_now)
1266 {
1267 	struct sk_buff *buff;
1268 	int nlen = skb->len - len;
1269 	u16 flags;
1270 
1271 	/* All of a TSO frame must be composed of paged data.  */
1272 	if (skb->len != skb->data_len)
1273 		return tcp_fragment(sk, skb, len, mss_now);
1274 
1275 	buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
1276 	if (unlikely(buff == NULL))
1277 		return -ENOMEM;
1278 
1279 	sk->sk_wmem_queued += buff->truesize;
1280 	sk_mem_charge(sk, buff->truesize);
1281 	buff->truesize += nlen;
1282 	skb->truesize -= nlen;
1283 
1284 	/* Correct the sequence numbers. */
1285 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1286 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1287 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1288 
1289 	/* PSH and FIN should only be set in the second packet. */
1290 	flags = TCP_SKB_CB(skb)->flags;
1291 	TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
1292 	TCP_SKB_CB(buff)->flags = flags;
1293 
1294 	/* This packet was never sent out yet, so no SACK bits. */
1295 	TCP_SKB_CB(buff)->sacked = 0;
1296 
1297 	buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1298 	skb_split(skb, buff, len);
1299 
1300 	/* Fix up tso_factor for both original and new SKB.  */
1301 	tcp_set_skb_tso_segs(sk, skb, mss_now);
1302 	tcp_set_skb_tso_segs(sk, buff, mss_now);
1303 
1304 	/* Link BUFF into the send queue. */
1305 	skb_header_release(buff);
1306 	tcp_insert_write_queue_after(skb, buff, sk);
1307 
1308 	return 0;
1309 }
1310 
1311 /* Try to defer sending, if possible, in order to minimize the amount
1312  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1313  *
1314  * This algorithm is from John Heffner.
1315  */
1316 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1317 {
1318 	struct tcp_sock *tp = tcp_sk(sk);
1319 	const struct inet_connection_sock *icsk = inet_csk(sk);
1320 	u32 send_win, cong_win, limit, in_flight;
1321 
1322 	if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1323 		goto send_now;
1324 
1325 	if (icsk->icsk_ca_state != TCP_CA_Open)
1326 		goto send_now;
1327 
1328 	/* Defer for less than two clock ticks. */
1329 	if (tp->tso_deferred &&
1330 	    ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1331 		goto send_now;
1332 
1333 	in_flight = tcp_packets_in_flight(tp);
1334 
1335 	BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1336 
1337 	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1338 
1339 	/* From in_flight test above, we know that cwnd > in_flight.  */
1340 	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1341 
1342 	limit = min(send_win, cong_win);
1343 
1344 	/* If a full-sized TSO skb can be sent, do it. */
1345 	if (limit >= sk->sk_gso_max_size)
1346 		goto send_now;
1347 
1348 	if (sysctl_tcp_tso_win_divisor) {
1349 		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1350 
1351 		/* If at least some fraction of a window is available,
1352 		 * just use it.
1353 		 */
1354 		chunk /= sysctl_tcp_tso_win_divisor;
1355 		if (limit >= chunk)
1356 			goto send_now;
1357 	} else {
1358 		/* Different approach, try not to defer past a single
1359 		 * ACK.  Receiver should ACK every other full sized
1360 		 * frame, so if we have space for more than 3 frames
1361 		 * then send now.
1362 		 */
1363 		if (limit > tcp_max_burst(tp) * tp->mss_cache)
1364 			goto send_now;
1365 	}
1366 
1367 	/* Ok, it looks like it is advisable to defer.  */
1368 	tp->tso_deferred = 1 | (jiffies << 1);
1369 
1370 	return 1;
1371 
1372 send_now:
1373 	tp->tso_deferred = 0;
1374 	return 0;
1375 }
1376 
1377 /* Create a new MTU probe if we are ready.
1378  * Returns 0 if we should wait to probe (no cwnd available),
1379  *         1 if a probe was sent,
1380  *         -1 otherwise
1381  */
1382 static int tcp_mtu_probe(struct sock *sk)
1383 {
1384 	struct tcp_sock *tp = tcp_sk(sk);
1385 	struct inet_connection_sock *icsk = inet_csk(sk);
1386 	struct sk_buff *skb, *nskb, *next;
1387 	int len;
1388 	int probe_size;
1389 	int size_needed;
1390 	int copy;
1391 	int mss_now;
1392 
1393 	/* Not currently probing/verifying,
1394 	 * not in recovery,
1395 	 * have enough cwnd, and
1396 	 * not SACKing (the variable headers throw things off) */
1397 	if (!icsk->icsk_mtup.enabled ||
1398 	    icsk->icsk_mtup.probe_size ||
1399 	    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1400 	    tp->snd_cwnd < 11 ||
1401 	    tp->rx_opt.eff_sacks)
1402 		return -1;
1403 
1404 	/* Very simple search strategy: just double the MSS. */
1405 	mss_now = tcp_current_mss(sk, 0);
1406 	probe_size = 2 * tp->mss_cache;
1407 	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1408 	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1409 		/* TODO: set timer for probe_converge_event */
1410 		return -1;
1411 	}
1412 
1413 	/* Have enough data in the send queue to probe? */
1414 	if (tp->write_seq - tp->snd_nxt < size_needed)
1415 		return -1;
1416 
1417 	if (tp->snd_wnd < size_needed)
1418 		return -1;
1419 	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1420 		return 0;
1421 
1422 	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
1423 	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1424 		if (!tcp_packets_in_flight(tp))
1425 			return -1;
1426 		else
1427 			return 0;
1428 	}
1429 
1430 	/* We're allowed to probe.  Build it now. */
1431 	if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1432 		return -1;
1433 	sk->sk_wmem_queued += nskb->truesize;
1434 	sk_mem_charge(sk, nskb->truesize);
1435 
1436 	skb = tcp_send_head(sk);
1437 
1438 	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1439 	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1440 	TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1441 	TCP_SKB_CB(nskb)->sacked = 0;
1442 	nskb->csum = 0;
1443 	nskb->ip_summed = skb->ip_summed;
1444 
1445 	tcp_insert_write_queue_before(nskb, skb, sk);
1446 
1447 	len = 0;
1448 	tcp_for_write_queue_from_safe(skb, next, sk) {
1449 		copy = min_t(int, skb->len, probe_size - len);
1450 		if (nskb->ip_summed)
1451 			skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1452 		else
1453 			nskb->csum = skb_copy_and_csum_bits(skb, 0,
1454 							    skb_put(nskb, copy),
1455 							    copy, nskb->csum);
1456 
1457 		if (skb->len <= copy) {
1458 			/* We've eaten all the data from this skb.
1459 			 * Throw it away. */
1460 			TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1461 			tcp_unlink_write_queue(skb, sk);
1462 			sk_wmem_free_skb(sk, skb);
1463 		} else {
1464 			TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1465 						   ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1466 			if (!skb_shinfo(skb)->nr_frags) {
1467 				skb_pull(skb, copy);
1468 				if (skb->ip_summed != CHECKSUM_PARTIAL)
1469 					skb->csum = csum_partial(skb->data,
1470 								 skb->len, 0);
1471 			} else {
1472 				__pskb_trim_head(skb, copy);
1473 				tcp_set_skb_tso_segs(sk, skb, mss_now);
1474 			}
1475 			TCP_SKB_CB(skb)->seq += copy;
1476 		}
1477 
1478 		len += copy;
1479 
1480 		if (len >= probe_size)
1481 			break;
1482 	}
1483 	tcp_init_tso_segs(sk, nskb, nskb->len);
1484 
1485 	/* We're ready to send.  If this fails, the probe will
1486 	 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1487 	TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1488 	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1489 		/* Decrement cwnd here because we are sending
1490 		 * effectively two packets. */
1491 		tp->snd_cwnd--;
1492 		tcp_event_new_data_sent(sk, nskb);
1493 
1494 		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1495 		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1496 		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1497 
1498 		return 1;
1499 	}
1500 
1501 	return -1;
1502 }
1503 
1504 /* This routine writes packets to the network.  It advances the
1505  * send_head.  This happens as incoming acks open up the remote
1506  * window for us.
1507  *
1508  * Returns 1, if no segments are in flight and we have queued segments, but
1509  * cannot send anything now because of SWS or another problem.
1510  */
1511 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1512 {
1513 	struct tcp_sock *tp = tcp_sk(sk);
1514 	struct sk_buff *skb;
1515 	unsigned int tso_segs, sent_pkts;
1516 	int cwnd_quota;
1517 	int result;
1518 
1519 	/* If we are closed, the bytes will have to remain here.
1520 	 * In time closedown will finish, we empty the write queue and all
1521 	 * will be happy.
1522 	 */
1523 	if (unlikely(sk->sk_state == TCP_CLOSE))
1524 		return 0;
1525 
1526 	sent_pkts = 0;
1527 
1528 	/* Do MTU probing. */
1529 	if ((result = tcp_mtu_probe(sk)) == 0) {
1530 		return 0;
1531 	} else if (result > 0) {
1532 		sent_pkts = 1;
1533 	}
1534 
1535 	while ((skb = tcp_send_head(sk))) {
1536 		unsigned int limit;
1537 
1538 		tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1539 		BUG_ON(!tso_segs);
1540 
1541 		cwnd_quota = tcp_cwnd_test(tp, skb);
1542 		if (!cwnd_quota)
1543 			break;
1544 
1545 		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1546 			break;
1547 
1548 		if (tso_segs == 1) {
1549 			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1550 						     (tcp_skb_is_last(sk, skb) ?
1551 						      nonagle : TCP_NAGLE_PUSH))))
1552 				break;
1553 		} else {
1554 			if (tcp_tso_should_defer(sk, skb))
1555 				break;
1556 		}
1557 
1558 		limit = mss_now;
1559 		if (tso_segs > 1)
1560 			limit = tcp_mss_split_point(sk, skb, mss_now,
1561 						    cwnd_quota);
1562 
1563 		if (skb->len > limit &&
1564 		    unlikely(tso_fragment(sk, skb, limit, mss_now)))
1565 			break;
1566 
1567 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
1568 
1569 		if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1570 			break;
1571 
1572 		/* Advance the send_head.  This one is sent out.
1573 		 * This call will increment packets_out.
1574 		 */
1575 		tcp_event_new_data_sent(sk, skb);
1576 
1577 		tcp_minshall_update(tp, mss_now, skb);
1578 		sent_pkts++;
1579 	}
1580 
1581 	if (likely(sent_pkts)) {
1582 		tcp_cwnd_validate(sk);
1583 		return 0;
1584 	}
1585 	return !tp->packets_out && tcp_send_head(sk);
1586 }
1587 
1588 /* Push out any pending frames which were held back due to
1589  * TCP_CORK or attempt at coalescing tiny packets.
1590  * The socket must be locked by the caller.
1591  */
1592 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1593 			       int nonagle)
1594 {
1595 	struct sk_buff *skb = tcp_send_head(sk);
1596 
1597 	if (skb) {
1598 		if (tcp_write_xmit(sk, cur_mss, nonagle))
1599 			tcp_check_probe_timer(sk);
1600 	}
1601 }
1602 
1603 /* Send _single_ skb sitting at the send head. This function requires
1604  * true push pending frames to setup probe timer etc.
1605  */
1606 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1607 {
1608 	struct sk_buff *skb = tcp_send_head(sk);
1609 	unsigned int tso_segs, cwnd_quota;
1610 
1611 	BUG_ON(!skb || skb->len < mss_now);
1612 
1613 	tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1614 	cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1615 
1616 	if (likely(cwnd_quota)) {
1617 		unsigned int limit;
1618 
1619 		BUG_ON(!tso_segs);
1620 
1621 		limit = mss_now;
1622 		if (tso_segs > 1)
1623 			limit = tcp_mss_split_point(sk, skb, mss_now,
1624 						    cwnd_quota);
1625 
1626 		if (skb->len > limit &&
1627 		    unlikely(tso_fragment(sk, skb, limit, mss_now)))
1628 			return;
1629 
1630 		/* Send it out now. */
1631 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
1632 
1633 		if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1634 			tcp_event_new_data_sent(sk, skb);
1635 			tcp_cwnd_validate(sk);
1636 			return;
1637 		}
1638 	}
1639 }
1640 
1641 /* This function returns the amount that we can raise the
1642  * usable window based on the following constraints
1643  *
1644  * 1. The window can never be shrunk once it is offered (RFC 793)
1645  * 2. We limit memory per socket
1646  *
1647  * RFC 1122:
1648  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1649  *  RECV.NEXT + RCV.WIN fixed until:
1650  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1651  *
1652  * i.e. don't raise the right edge of the window until you can raise
1653  * it at least MSS bytes.
1654  *
1655  * Unfortunately, the recommended algorithm breaks header prediction,
1656  * since header prediction assumes th->window stays fixed.
1657  *
1658  * Strictly speaking, keeping th->window fixed violates the receiver
1659  * side SWS prevention criteria. The problem is that under this rule
1660  * a stream of single byte packets will cause the right side of the
1661  * window to always advance by a single byte.
1662  *
1663  * Of course, if the sender implements sender side SWS prevention
1664  * then this will not be a problem.
1665  *
1666  * BSD seems to make the following compromise:
1667  *
1668  *	If the free space is less than the 1/4 of the maximum
1669  *	space available and the free space is less than 1/2 mss,
1670  *	then set the window to 0.
1671  *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1672  *	Otherwise, just prevent the window from shrinking
1673  *	and from being larger than the largest representable value.
1674  *
1675  * This prevents incremental opening of the window in the regime
1676  * where TCP is limited by the speed of the reader side taking
1677  * data out of the TCP receive queue. It does nothing about
1678  * those cases where the window is constrained on the sender side
1679  * because the pipeline is full.
1680  *
1681  * BSD also seems to "accidentally" limit itself to windows that are a
1682  * multiple of MSS, at least until the free space gets quite small.
1683  * This would appear to be a side effect of the mbuf implementation.
1684  * Combining these two algorithms results in the observed behavior
1685  * of having a fixed window size at almost all times.
1686  *
1687  * Below we obtain similar behavior by forcing the offered window to
1688  * a multiple of the mss when it is feasible to do so.
1689  *
1690  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1691  * Regular options like TIMESTAMP are taken into account.
1692  */
1693 u32 __tcp_select_window(struct sock *sk)
1694 {
1695 	struct inet_connection_sock *icsk = inet_csk(sk);
1696 	struct tcp_sock *tp = tcp_sk(sk);
1697 	/* MSS for the peer's data.  Previous versions used mss_clamp
1698 	 * here.  I don't know if the value based on our guesses
1699 	 * of peer's MSS is better for the performance.  It's more correct
1700 	 * but may be worse for the performance because of rcv_mss
1701 	 * fluctuations.  --SAW  1998/11/1
1702 	 */
1703 	int mss = icsk->icsk_ack.rcv_mss;
1704 	int free_space = tcp_space(sk);
1705 	int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1706 	int window;
1707 
1708 	if (mss > full_space)
1709 		mss = full_space;
1710 
1711 	if (free_space < (full_space >> 1)) {
1712 		icsk->icsk_ack.quick = 0;
1713 
1714 		if (tcp_memory_pressure)
1715 			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1716 					       4U * tp->advmss);
1717 
1718 		if (free_space < mss)
1719 			return 0;
1720 	}
1721 
1722 	if (free_space > tp->rcv_ssthresh)
1723 		free_space = tp->rcv_ssthresh;
1724 
1725 	/* Don't do rounding if we are using window scaling, since the
1726 	 * scaled window will not line up with the MSS boundary anyway.
1727 	 */
1728 	window = tp->rcv_wnd;
1729 	if (tp->rx_opt.rcv_wscale) {
1730 		window = free_space;
1731 
1732 		/* Advertise enough space so that it won't get scaled away.
1733 		 * Import case: prevent zero window announcement if
1734 		 * 1<<rcv_wscale > mss.
1735 		 */
1736 		if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1737 			window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1738 				  << tp->rx_opt.rcv_wscale);
1739 	} else {
1740 		/* Get the largest window that is a nice multiple of mss.
1741 		 * Window clamp already applied above.
1742 		 * If our current window offering is within 1 mss of the
1743 		 * free space we just keep it. This prevents the divide
1744 		 * and multiply from happening most of the time.
1745 		 * We also don't do any window rounding when the free space
1746 		 * is too small.
1747 		 */
1748 		if (window <= free_space - mss || window > free_space)
1749 			window = (free_space / mss) * mss;
1750 		else if (mss == full_space &&
1751 			 free_space > window + (full_space >> 1))
1752 			window = free_space;
1753 	}
1754 
1755 	return window;
1756 }
1757 
1758 /* Attempt to collapse two adjacent SKB's during retransmission. */
1759 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb,
1760 				     int mss_now)
1761 {
1762 	struct tcp_sock *tp = tcp_sk(sk);
1763 	struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1764 	int skb_size, next_skb_size;
1765 	u16 flags;
1766 
1767 	/* The first test we must make is that neither of these two
1768 	 * SKB's are still referenced by someone else.
1769 	 */
1770 	if (skb_cloned(skb) || skb_cloned(next_skb))
1771 		return;
1772 
1773 	skb_size = skb->len;
1774 	next_skb_size = next_skb->len;
1775 	flags = TCP_SKB_CB(skb)->flags;
1776 
1777 	/* Also punt if next skb has been SACK'd. */
1778 	if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1779 		return;
1780 
1781 	/* Next skb is out of window. */
1782 	if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp)))
1783 		return;
1784 
1785 	/* Punt if not enough space exists in the first SKB for
1786 	 * the data in the second, or the total combined payload
1787 	 * would exceed the MSS.
1788 	 */
1789 	if ((next_skb_size > skb_tailroom(skb)) ||
1790 	    ((skb_size + next_skb_size) > mss_now))
1791 		return;
1792 
1793 	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1794 
1795 	tcp_highest_sack_combine(sk, next_skb, skb);
1796 
1797 	/* Ok.	We will be able to collapse the packet. */
1798 	tcp_unlink_write_queue(next_skb, sk);
1799 
1800 	skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1801 				  next_skb_size);
1802 
1803 	if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1804 		skb->ip_summed = CHECKSUM_PARTIAL;
1805 
1806 	if (skb->ip_summed != CHECKSUM_PARTIAL)
1807 		skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1808 
1809 	/* Update sequence range on original skb. */
1810 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1811 
1812 	/* Merge over control information. */
1813 	flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1814 	TCP_SKB_CB(skb)->flags = flags;
1815 
1816 	/* All done, get rid of second SKB and account for it so
1817 	 * packet counting does not break.
1818 	 */
1819 	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1820 	if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS)
1821 		tp->retrans_out -= tcp_skb_pcount(next_skb);
1822 	if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST)
1823 		tp->lost_out -= tcp_skb_pcount(next_skb);
1824 	/* Reno case is special. Sigh... */
1825 	if (tcp_is_reno(tp) && tp->sacked_out)
1826 		tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1827 
1828 	tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
1829 	tp->packets_out -= tcp_skb_pcount(next_skb);
1830 
1831 	/* changed transmit queue under us so clear hints */
1832 	tcp_clear_retrans_hints_partial(tp);
1833 	if (next_skb == tp->retransmit_skb_hint)
1834 		tp->retransmit_skb_hint = skb;
1835 
1836 	sk_wmem_free_skb(sk, next_skb);
1837 }
1838 
1839 /* Do a simple retransmit without using the backoff mechanisms in
1840  * tcp_timer. This is used for path mtu discovery.
1841  * The socket is already locked here.
1842  */
1843 void tcp_simple_retransmit(struct sock *sk)
1844 {
1845 	const struct inet_connection_sock *icsk = inet_csk(sk);
1846 	struct tcp_sock *tp = tcp_sk(sk);
1847 	struct sk_buff *skb;
1848 	unsigned int mss = tcp_current_mss(sk, 0);
1849 	u32 prior_lost = tp->lost_out;
1850 
1851 	tcp_for_write_queue(skb, sk) {
1852 		if (skb == tcp_send_head(sk))
1853 			break;
1854 		if (skb->len > mss &&
1855 		    !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1856 			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1857 				TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1858 				tp->retrans_out -= tcp_skb_pcount(skb);
1859 			}
1860 			tcp_skb_mark_lost_uncond_verify(tp, skb);
1861 		}
1862 	}
1863 
1864 	tcp_clear_retrans_hints_partial(tp);
1865 
1866 	if (prior_lost == tp->lost_out)
1867 		return;
1868 
1869 	if (tcp_is_reno(tp))
1870 		tcp_limit_reno_sacked(tp);
1871 
1872 	tcp_verify_left_out(tp);
1873 
1874 	/* Don't muck with the congestion window here.
1875 	 * Reason is that we do not increase amount of _data_
1876 	 * in network, but units changed and effective
1877 	 * cwnd/ssthresh really reduced now.
1878 	 */
1879 	if (icsk->icsk_ca_state != TCP_CA_Loss) {
1880 		tp->high_seq = tp->snd_nxt;
1881 		tp->snd_ssthresh = tcp_current_ssthresh(sk);
1882 		tp->prior_ssthresh = 0;
1883 		tp->undo_marker = 0;
1884 		tcp_set_ca_state(sk, TCP_CA_Loss);
1885 	}
1886 	tcp_xmit_retransmit_queue(sk);
1887 }
1888 
1889 /* This retransmits one SKB.  Policy decisions and retransmit queue
1890  * state updates are done by the caller.  Returns non-zero if an
1891  * error occurred which prevented the send.
1892  */
1893 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1894 {
1895 	struct tcp_sock *tp = tcp_sk(sk);
1896 	struct inet_connection_sock *icsk = inet_csk(sk);
1897 	unsigned int cur_mss;
1898 	int err;
1899 
1900 	/* Inconslusive MTU probe */
1901 	if (icsk->icsk_mtup.probe_size) {
1902 		icsk->icsk_mtup.probe_size = 0;
1903 	}
1904 
1905 	/* Do not sent more than we queued. 1/4 is reserved for possible
1906 	 * copying overhead: fragmentation, tunneling, mangling etc.
1907 	 */
1908 	if (atomic_read(&sk->sk_wmem_alloc) >
1909 	    min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1910 		return -EAGAIN;
1911 
1912 	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1913 		if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1914 			BUG();
1915 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1916 			return -ENOMEM;
1917 	}
1918 
1919 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1920 		return -EHOSTUNREACH; /* Routing failure or similar. */
1921 
1922 	cur_mss = tcp_current_mss(sk, 0);
1923 
1924 	/* If receiver has shrunk his window, and skb is out of
1925 	 * new window, do not retransmit it. The exception is the
1926 	 * case, when window is shrunk to zero. In this case
1927 	 * our retransmit serves as a zero window probe.
1928 	 */
1929 	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
1930 	    && TCP_SKB_CB(skb)->seq != tp->snd_una)
1931 		return -EAGAIN;
1932 
1933 	if (skb->len > cur_mss) {
1934 		if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1935 			return -ENOMEM; /* We'll try again later. */
1936 	}
1937 
1938 	/* Collapse two adjacent packets if worthwhile and we can. */
1939 	if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1940 	    (skb->len < (cur_mss >> 1)) &&
1941 	    (!tcp_skb_is_last(sk, skb)) &&
1942 	    (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1943 	    (skb_shinfo(skb)->nr_frags == 0 &&
1944 	     skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1945 	    (tcp_skb_pcount(skb) == 1 &&
1946 	     tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
1947 	    (sysctl_tcp_retrans_collapse != 0))
1948 		tcp_retrans_try_collapse(sk, skb, cur_mss);
1949 
1950 	/* Some Solaris stacks overoptimize and ignore the FIN on a
1951 	 * retransmit when old data is attached.  So strip it off
1952 	 * since it is cheap to do so and saves bytes on the network.
1953 	 */
1954 	if (skb->len > 0 &&
1955 	    (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1956 	    tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1957 		if (!pskb_trim(skb, 0)) {
1958 			/* Reuse, even though it does some unnecessary work */
1959 			tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
1960 					     TCP_SKB_CB(skb)->flags);
1961 			skb->ip_summed = CHECKSUM_NONE;
1962 		}
1963 	}
1964 
1965 	/* Make a copy, if the first transmission SKB clone we made
1966 	 * is still in somebody's hands, else make a clone.
1967 	 */
1968 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
1969 
1970 	err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1971 
1972 	if (err == 0) {
1973 		/* Update global TCP statistics. */
1974 		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
1975 
1976 		tp->total_retrans++;
1977 
1978 #if FASTRETRANS_DEBUG > 0
1979 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1980 			if (net_ratelimit())
1981 				printk(KERN_DEBUG "retrans_out leaked.\n");
1982 		}
1983 #endif
1984 		if (!tp->retrans_out)
1985 			tp->lost_retrans_low = tp->snd_nxt;
1986 		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1987 		tp->retrans_out += tcp_skb_pcount(skb);
1988 
1989 		/* Save stamp of the first retransmit. */
1990 		if (!tp->retrans_stamp)
1991 			tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1992 
1993 		tp->undo_retrans++;
1994 
1995 		/* snd_nxt is stored to detect loss of retransmitted segment,
1996 		 * see tcp_input.c tcp_sacktag_write_queue().
1997 		 */
1998 		TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1999 	}
2000 	return err;
2001 }
2002 
2003 static int tcp_can_forward_retransmit(struct sock *sk)
2004 {
2005 	const struct inet_connection_sock *icsk = inet_csk(sk);
2006 	struct tcp_sock *tp = tcp_sk(sk);
2007 
2008 	/* Forward retransmissions are possible only during Recovery. */
2009 	if (icsk->icsk_ca_state != TCP_CA_Recovery)
2010 		return 0;
2011 
2012 	/* No forward retransmissions in Reno are possible. */
2013 	if (tcp_is_reno(tp))
2014 		return 0;
2015 
2016 	/* Yeah, we have to make difficult choice between forward transmission
2017 	 * and retransmission... Both ways have their merits...
2018 	 *
2019 	 * For now we do not retransmit anything, while we have some new
2020 	 * segments to send. In the other cases, follow rule 3 for
2021 	 * NextSeg() specified in RFC3517.
2022 	 */
2023 
2024 	if (tcp_may_send_now(sk))
2025 		return 0;
2026 
2027 	return 1;
2028 }
2029 
2030 /* This gets called after a retransmit timeout, and the initially
2031  * retransmitted data is acknowledged.  It tries to continue
2032  * resending the rest of the retransmit queue, until either
2033  * we've sent it all or the congestion window limit is reached.
2034  * If doing SACK, the first ACK which comes back for a timeout
2035  * based retransmit packet might feed us FACK information again.
2036  * If so, we use it to avoid unnecessarily retransmissions.
2037  */
2038 void tcp_xmit_retransmit_queue(struct sock *sk)
2039 {
2040 	const struct inet_connection_sock *icsk = inet_csk(sk);
2041 	struct tcp_sock *tp = tcp_sk(sk);
2042 	struct sk_buff *skb;
2043 	struct sk_buff *hole = NULL;
2044 	u32 last_lost;
2045 	int mib_idx;
2046 	int fwd_rexmitting = 0;
2047 
2048 	if (!tp->lost_out)
2049 		tp->retransmit_high = tp->snd_una;
2050 
2051 	if (tp->retransmit_skb_hint) {
2052 		skb = tp->retransmit_skb_hint;
2053 		last_lost = TCP_SKB_CB(skb)->end_seq;
2054 		if (after(last_lost, tp->retransmit_high))
2055 			last_lost = tp->retransmit_high;
2056 	} else {
2057 		skb = tcp_write_queue_head(sk);
2058 		last_lost = tp->snd_una;
2059 	}
2060 
2061 	/* First pass: retransmit lost packets. */
2062 	tcp_for_write_queue_from(skb, sk) {
2063 		__u8 sacked = TCP_SKB_CB(skb)->sacked;
2064 
2065 		if (skb == tcp_send_head(sk))
2066 			break;
2067 		/* we could do better than to assign each time */
2068 		if (hole == NULL)
2069 			tp->retransmit_skb_hint = skb;
2070 
2071 		/* Assume this retransmit will generate
2072 		 * only one packet for congestion window
2073 		 * calculation purposes.  This works because
2074 		 * tcp_retransmit_skb() will chop up the
2075 		 * packet to be MSS sized and all the
2076 		 * packet counting works out.
2077 		 */
2078 		if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2079 			return;
2080 
2081 		if (fwd_rexmitting) {
2082 begin_fwd:
2083 			if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2084 				break;
2085 			mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2086 
2087 		} else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2088 			tp->retransmit_high = last_lost;
2089 			if (!tcp_can_forward_retransmit(sk))
2090 				break;
2091 			/* Backtrack if necessary to non-L'ed skb */
2092 			if (hole != NULL) {
2093 				skb = hole;
2094 				hole = NULL;
2095 			}
2096 			fwd_rexmitting = 1;
2097 			goto begin_fwd;
2098 
2099 		} else if (!(sacked & TCPCB_LOST)) {
2100 			if (hole == NULL && !(sacked & TCPCB_SACKED_RETRANS))
2101 				hole = skb;
2102 			continue;
2103 
2104 		} else {
2105 			last_lost = TCP_SKB_CB(skb)->end_seq;
2106 			if (icsk->icsk_ca_state != TCP_CA_Loss)
2107 				mib_idx = LINUX_MIB_TCPFASTRETRANS;
2108 			else
2109 				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2110 		}
2111 
2112 		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2113 			continue;
2114 
2115 		if (tcp_retransmit_skb(sk, skb))
2116 			return;
2117 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
2118 
2119 		if (skb == tcp_write_queue_head(sk))
2120 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2121 						  inet_csk(sk)->icsk_rto,
2122 						  TCP_RTO_MAX);
2123 	}
2124 }
2125 
2126 /* Send a fin.  The caller locks the socket for us.  This cannot be
2127  * allowed to fail queueing a FIN frame under any circumstances.
2128  */
2129 void tcp_send_fin(struct sock *sk)
2130 {
2131 	struct tcp_sock *tp = tcp_sk(sk);
2132 	struct sk_buff *skb = tcp_write_queue_tail(sk);
2133 	int mss_now;
2134 
2135 	/* Optimization, tack on the FIN if we have a queue of
2136 	 * unsent frames.  But be careful about outgoing SACKS
2137 	 * and IP options.
2138 	 */
2139 	mss_now = tcp_current_mss(sk, 1);
2140 
2141 	if (tcp_send_head(sk) != NULL) {
2142 		TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2143 		TCP_SKB_CB(skb)->end_seq++;
2144 		tp->write_seq++;
2145 	} else {
2146 		/* Socket is locked, keep trying until memory is available. */
2147 		for (;;) {
2148 			skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
2149 			if (skb)
2150 				break;
2151 			yield();
2152 		}
2153 
2154 		/* Reserve space for headers and prepare control bits. */
2155 		skb_reserve(skb, MAX_TCP_HEADER);
2156 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2157 		tcp_init_nondata_skb(skb, tp->write_seq,
2158 				     TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2159 		tcp_queue_skb(sk, skb);
2160 	}
2161 	__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2162 }
2163 
2164 /* We get here when a process closes a file descriptor (either due to
2165  * an explicit close() or as a byproduct of exit()'ing) and there
2166  * was unread data in the receive queue.  This behavior is recommended
2167  * by RFC 2525, section 2.17.  -DaveM
2168  */
2169 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2170 {
2171 	struct sk_buff *skb;
2172 
2173 	/* NOTE: No TCP options attached and we never retransmit this. */
2174 	skb = alloc_skb(MAX_TCP_HEADER, priority);
2175 	if (!skb) {
2176 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2177 		return;
2178 	}
2179 
2180 	/* Reserve space for headers and prepare control bits. */
2181 	skb_reserve(skb, MAX_TCP_HEADER);
2182 	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2183 			     TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2184 	/* Send it off. */
2185 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2186 	if (tcp_transmit_skb(sk, skb, 0, priority))
2187 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2188 
2189 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2190 }
2191 
2192 /* WARNING: This routine must only be called when we have already sent
2193  * a SYN packet that crossed the incoming SYN that caused this routine
2194  * to get called. If this assumption fails then the initial rcv_wnd
2195  * and rcv_wscale values will not be correct.
2196  */
2197 int tcp_send_synack(struct sock *sk)
2198 {
2199 	struct sk_buff *skb;
2200 
2201 	skb = tcp_write_queue_head(sk);
2202 	if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
2203 		printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2204 		return -EFAULT;
2205 	}
2206 	if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
2207 		if (skb_cloned(skb)) {
2208 			struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2209 			if (nskb == NULL)
2210 				return -ENOMEM;
2211 			tcp_unlink_write_queue(skb, sk);
2212 			skb_header_release(nskb);
2213 			__tcp_add_write_queue_head(sk, nskb);
2214 			sk_wmem_free_skb(sk, skb);
2215 			sk->sk_wmem_queued += nskb->truesize;
2216 			sk_mem_charge(sk, nskb->truesize);
2217 			skb = nskb;
2218 		}
2219 
2220 		TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2221 		TCP_ECN_send_synack(tcp_sk(sk), skb);
2222 	}
2223 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2224 	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2225 }
2226 
2227 /*
2228  * Prepare a SYN-ACK.
2229  */
2230 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2231 				struct request_sock *req)
2232 {
2233 	struct inet_request_sock *ireq = inet_rsk(req);
2234 	struct tcp_sock *tp = tcp_sk(sk);
2235 	struct tcphdr *th;
2236 	int tcp_header_size;
2237 	struct tcp_out_options opts;
2238 	struct sk_buff *skb;
2239 	struct tcp_md5sig_key *md5;
2240 	__u8 *md5_hash_location;
2241 	int mss;
2242 
2243 	skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2244 	if (skb == NULL)
2245 		return NULL;
2246 
2247 	/* Reserve space for headers. */
2248 	skb_reserve(skb, MAX_TCP_HEADER);
2249 
2250 	skb->dst = dst_clone(dst);
2251 
2252 	mss = dst_metric(dst, RTAX_ADVMSS);
2253 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2254 		mss = tp->rx_opt.user_mss;
2255 
2256 	if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2257 		__u8 rcv_wscale;
2258 		/* Set this up on the first call only */
2259 		req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2260 		/* tcp_full_space because it is guaranteed to be the first packet */
2261 		tcp_select_initial_window(tcp_full_space(sk),
2262 			mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2263 			&req->rcv_wnd,
2264 			&req->window_clamp,
2265 			ireq->wscale_ok,
2266 			&rcv_wscale);
2267 		ireq->rcv_wscale = rcv_wscale;
2268 	}
2269 
2270 	memset(&opts, 0, sizeof(opts));
2271 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2272 	tcp_header_size = tcp_synack_options(sk, req, mss,
2273 					     skb, &opts, &md5) +
2274 			  sizeof(struct tcphdr);
2275 
2276 	skb_push(skb, tcp_header_size);
2277 	skb_reset_transport_header(skb);
2278 
2279 	th = tcp_hdr(skb);
2280 	memset(th, 0, sizeof(struct tcphdr));
2281 	th->syn = 1;
2282 	th->ack = 1;
2283 	TCP_ECN_make_synack(req, th);
2284 	th->source = ireq->loc_port;
2285 	th->dest = ireq->rmt_port;
2286 	/* Setting of flags are superfluous here for callers (and ECE is
2287 	 * not even correctly set)
2288 	 */
2289 	tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2290 			     TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
2291 	th->seq = htonl(TCP_SKB_CB(skb)->seq);
2292 	th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2293 
2294 	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2295 	th->window = htons(min(req->rcv_wnd, 65535U));
2296 #ifdef CONFIG_SYN_COOKIES
2297 	if (unlikely(req->cookie_ts))
2298 		TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2299 	else
2300 #endif
2301 	tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
2302 	th->doff = (tcp_header_size >> 2);
2303 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
2304 
2305 #ifdef CONFIG_TCP_MD5SIG
2306 	/* Okay, we have all we need - do the md5 hash if needed */
2307 	if (md5) {
2308 		tp->af_specific->calc_md5_hash(md5_hash_location,
2309 					       md5, NULL, req, skb);
2310 	}
2311 #endif
2312 
2313 	return skb;
2314 }
2315 
2316 /*
2317  * Do all connect socket setups that can be done AF independent.
2318  */
2319 static void tcp_connect_init(struct sock *sk)
2320 {
2321 	struct dst_entry *dst = __sk_dst_get(sk);
2322 	struct tcp_sock *tp = tcp_sk(sk);
2323 	__u8 rcv_wscale;
2324 
2325 	/* We'll fix this up when we get a response from the other end.
2326 	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2327 	 */
2328 	tp->tcp_header_len = sizeof(struct tcphdr) +
2329 		(sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2330 
2331 #ifdef CONFIG_TCP_MD5SIG
2332 	if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2333 		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2334 #endif
2335 
2336 	/* If user gave his TCP_MAXSEG, record it to clamp */
2337 	if (tp->rx_opt.user_mss)
2338 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2339 	tp->max_window = 0;
2340 	tcp_mtup_init(sk);
2341 	tcp_sync_mss(sk, dst_mtu(dst));
2342 
2343 	if (!tp->window_clamp)
2344 		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2345 	tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2346 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2347 		tp->advmss = tp->rx_opt.user_mss;
2348 
2349 	tcp_initialize_rcv_mss(sk);
2350 
2351 	tcp_select_initial_window(tcp_full_space(sk),
2352 				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2353 				  &tp->rcv_wnd,
2354 				  &tp->window_clamp,
2355 				  sysctl_tcp_window_scaling,
2356 				  &rcv_wscale);
2357 
2358 	tp->rx_opt.rcv_wscale = rcv_wscale;
2359 	tp->rcv_ssthresh = tp->rcv_wnd;
2360 
2361 	sk->sk_err = 0;
2362 	sock_reset_flag(sk, SOCK_DONE);
2363 	tp->snd_wnd = 0;
2364 	tcp_init_wl(tp, tp->write_seq, 0);
2365 	tp->snd_una = tp->write_seq;
2366 	tp->snd_sml = tp->write_seq;
2367 	tp->snd_up = tp->write_seq;
2368 	tp->rcv_nxt = 0;
2369 	tp->rcv_wup = 0;
2370 	tp->copied_seq = 0;
2371 
2372 	inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2373 	inet_csk(sk)->icsk_retransmits = 0;
2374 	tcp_clear_retrans(tp);
2375 }
2376 
2377 /*
2378  * Build a SYN and send it off.
2379  */
2380 int tcp_connect(struct sock *sk)
2381 {
2382 	struct tcp_sock *tp = tcp_sk(sk);
2383 	struct sk_buff *buff;
2384 
2385 	tcp_connect_init(sk);
2386 
2387 	buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2388 	if (unlikely(buff == NULL))
2389 		return -ENOBUFS;
2390 
2391 	/* Reserve space for headers. */
2392 	skb_reserve(buff, MAX_TCP_HEADER);
2393 
2394 	tp->snd_nxt = tp->write_seq;
2395 	tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
2396 	TCP_ECN_send_syn(sk, buff);
2397 
2398 	/* Send it off. */
2399 	TCP_SKB_CB(buff)->when = tcp_time_stamp;
2400 	tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2401 	skb_header_release(buff);
2402 	__tcp_add_write_queue_tail(sk, buff);
2403 	sk->sk_wmem_queued += buff->truesize;
2404 	sk_mem_charge(sk, buff->truesize);
2405 	tp->packets_out += tcp_skb_pcount(buff);
2406 	tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2407 
2408 	/* We change tp->snd_nxt after the tcp_transmit_skb() call
2409 	 * in order to make this packet get counted in tcpOutSegs.
2410 	 */
2411 	tp->snd_nxt = tp->write_seq;
2412 	tp->pushed_seq = tp->write_seq;
2413 	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2414 
2415 	/* Timer for repeating the SYN until an answer. */
2416 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2417 				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2418 	return 0;
2419 }
2420 
2421 /* Send out a delayed ack, the caller does the policy checking
2422  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
2423  * for details.
2424  */
2425 void tcp_send_delayed_ack(struct sock *sk)
2426 {
2427 	struct inet_connection_sock *icsk = inet_csk(sk);
2428 	int ato = icsk->icsk_ack.ato;
2429 	unsigned long timeout;
2430 
2431 	if (ato > TCP_DELACK_MIN) {
2432 		const struct tcp_sock *tp = tcp_sk(sk);
2433 		int max_ato = HZ / 2;
2434 
2435 		if (icsk->icsk_ack.pingpong ||
2436 		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2437 			max_ato = TCP_DELACK_MAX;
2438 
2439 		/* Slow path, intersegment interval is "high". */
2440 
2441 		/* If some rtt estimate is known, use it to bound delayed ack.
2442 		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2443 		 * directly.
2444 		 */
2445 		if (tp->srtt) {
2446 			int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2447 
2448 			if (rtt < max_ato)
2449 				max_ato = rtt;
2450 		}
2451 
2452 		ato = min(ato, max_ato);
2453 	}
2454 
2455 	/* Stay within the limit we were given */
2456 	timeout = jiffies + ato;
2457 
2458 	/* Use new timeout only if there wasn't a older one earlier. */
2459 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2460 		/* If delack timer was blocked or is about to expire,
2461 		 * send ACK now.
2462 		 */
2463 		if (icsk->icsk_ack.blocked ||
2464 		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2465 			tcp_send_ack(sk);
2466 			return;
2467 		}
2468 
2469 		if (!time_before(timeout, icsk->icsk_ack.timeout))
2470 			timeout = icsk->icsk_ack.timeout;
2471 	}
2472 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2473 	icsk->icsk_ack.timeout = timeout;
2474 	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2475 }
2476 
2477 /* This routine sends an ack and also updates the window. */
2478 void tcp_send_ack(struct sock *sk)
2479 {
2480 	struct sk_buff *buff;
2481 
2482 	/* If we have been reset, we may not send again. */
2483 	if (sk->sk_state == TCP_CLOSE)
2484 		return;
2485 
2486 	/* We are not putting this on the write queue, so
2487 	 * tcp_transmit_skb() will set the ownership to this
2488 	 * sock.
2489 	 */
2490 	buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2491 	if (buff == NULL) {
2492 		inet_csk_schedule_ack(sk);
2493 		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2494 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2495 					  TCP_DELACK_MAX, TCP_RTO_MAX);
2496 		return;
2497 	}
2498 
2499 	/* Reserve space for headers and prepare control bits. */
2500 	skb_reserve(buff, MAX_TCP_HEADER);
2501 	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
2502 
2503 	/* Send it off, this clears delayed acks for us. */
2504 	TCP_SKB_CB(buff)->when = tcp_time_stamp;
2505 	tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2506 }
2507 
2508 /* This routine sends a packet with an out of date sequence
2509  * number. It assumes the other end will try to ack it.
2510  *
2511  * Question: what should we make while urgent mode?
2512  * 4.4BSD forces sending single byte of data. We cannot send
2513  * out of window data, because we have SND.NXT==SND.MAX...
2514  *
2515  * Current solution: to send TWO zero-length segments in urgent mode:
2516  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2517  * out-of-date with SND.UNA-1 to probe window.
2518  */
2519 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2520 {
2521 	struct tcp_sock *tp = tcp_sk(sk);
2522 	struct sk_buff *skb;
2523 
2524 	/* We don't queue it, tcp_transmit_skb() sets ownership. */
2525 	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2526 	if (skb == NULL)
2527 		return -1;
2528 
2529 	/* Reserve space for headers and set control bits. */
2530 	skb_reserve(skb, MAX_TCP_HEADER);
2531 	/* Use a previous sequence.  This should cause the other
2532 	 * end to send an ack.  Don't queue or clone SKB, just
2533 	 * send it.
2534 	 */
2535 	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
2536 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2537 	return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2538 }
2539 
2540 int tcp_write_wakeup(struct sock *sk)
2541 {
2542 	struct tcp_sock *tp = tcp_sk(sk);
2543 	struct sk_buff *skb;
2544 
2545 	if (sk->sk_state == TCP_CLOSE)
2546 		return -1;
2547 
2548 	if ((skb = tcp_send_head(sk)) != NULL &&
2549 	    before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2550 		int err;
2551 		unsigned int mss = tcp_current_mss(sk, 0);
2552 		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2553 
2554 		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2555 			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2556 
2557 		/* We are probing the opening of a window
2558 		 * but the window size is != 0
2559 		 * must have been a result SWS avoidance ( sender )
2560 		 */
2561 		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2562 		    skb->len > mss) {
2563 			seg_size = min(seg_size, mss);
2564 			TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2565 			if (tcp_fragment(sk, skb, seg_size, mss))
2566 				return -1;
2567 		} else if (!tcp_skb_pcount(skb))
2568 			tcp_set_skb_tso_segs(sk, skb, mss);
2569 
2570 		TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2571 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
2572 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2573 		if (!err)
2574 			tcp_event_new_data_sent(sk, skb);
2575 		return err;
2576 	} else {
2577 		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2578 			tcp_xmit_probe_skb(sk, 1);
2579 		return tcp_xmit_probe_skb(sk, 0);
2580 	}
2581 }
2582 
2583 /* A window probe timeout has occurred.  If window is not closed send
2584  * a partial packet else a zero probe.
2585  */
2586 void tcp_send_probe0(struct sock *sk)
2587 {
2588 	struct inet_connection_sock *icsk = inet_csk(sk);
2589 	struct tcp_sock *tp = tcp_sk(sk);
2590 	int err;
2591 
2592 	err = tcp_write_wakeup(sk);
2593 
2594 	if (tp->packets_out || !tcp_send_head(sk)) {
2595 		/* Cancel probe timer, if it is not required. */
2596 		icsk->icsk_probes_out = 0;
2597 		icsk->icsk_backoff = 0;
2598 		return;
2599 	}
2600 
2601 	if (err <= 0) {
2602 		if (icsk->icsk_backoff < sysctl_tcp_retries2)
2603 			icsk->icsk_backoff++;
2604 		icsk->icsk_probes_out++;
2605 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2606 					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2607 					  TCP_RTO_MAX);
2608 	} else {
2609 		/* If packet was not sent due to local congestion,
2610 		 * do not backoff and do not remember icsk_probes_out.
2611 		 * Let local senders to fight for local resources.
2612 		 *
2613 		 * Use accumulated backoff yet.
2614 		 */
2615 		if (!icsk->icsk_probes_out)
2616 			icsk->icsk_probes_out = 1;
2617 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2618 					  min(icsk->icsk_rto << icsk->icsk_backoff,
2619 					      TCP_RESOURCE_PROBE_INTERVAL),
2620 					  TCP_RTO_MAX);
2621 	}
2622 }
2623 
2624 EXPORT_SYMBOL(tcp_select_initial_window);
2625 EXPORT_SYMBOL(tcp_connect);
2626 EXPORT_SYMBOL(tcp_make_synack);
2627 EXPORT_SYMBOL(tcp_simple_retransmit);
2628 EXPORT_SYMBOL(tcp_sync_mss);
2629 EXPORT_SYMBOL(tcp_mtup_init);
2630