xref: /linux/net/ipv4/tcp_output.c (revision 40d3057ac036f2501c1930728a6179be4fca577b)
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 #define OPTION_SACK_ADVERTISE	(1 << 0)
349 #define OPTION_TS		(1 << 1)
350 #define OPTION_MD5		(1 << 2)
351 
352 struct tcp_out_options {
353 	u8 options;		/* bit field of OPTION_* */
354 	u8 ws;			/* window scale, 0 to disable */
355 	u8 num_sack_blocks;	/* number of SACK blocks to include */
356 	u16 mss;		/* 0 to disable */
357 	__u32 tsval, tsecr;	/* need to include OPTION_TS */
358 };
359 
360 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
361 			      const struct tcp_out_options *opts,
362 			      __u8 **md5_hash) {
363 	if (unlikely(OPTION_MD5 & opts->options)) {
364 		*ptr++ = htonl((TCPOPT_NOP << 24) |
365 			       (TCPOPT_NOP << 16) |
366 			       (TCPOPT_MD5SIG << 8) |
367 			       TCPOLEN_MD5SIG);
368 		*md5_hash = (__u8 *)ptr;
369 		ptr += 4;
370 	} else {
371 		*md5_hash = NULL;
372 	}
373 
374 	if (likely(OPTION_TS & opts->options)) {
375 		if (unlikely(OPTION_SACK_ADVERTISE & opts->options)) {
376 			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
377 				       (TCPOLEN_SACK_PERM << 16) |
378 				       (TCPOPT_TIMESTAMP << 8) |
379 				       TCPOLEN_TIMESTAMP);
380 		} else {
381 			*ptr++ = htonl((TCPOPT_NOP << 24) |
382 				       (TCPOPT_NOP << 16) |
383 				       (TCPOPT_TIMESTAMP << 8) |
384 				       TCPOLEN_TIMESTAMP);
385 		}
386 		*ptr++ = htonl(opts->tsval);
387 		*ptr++ = htonl(opts->tsecr);
388 	}
389 
390 	if (unlikely(opts->mss)) {
391 		*ptr++ = htonl((TCPOPT_MSS << 24) |
392 			       (TCPOLEN_MSS << 16) |
393 			       opts->mss);
394 	}
395 
396 	if (unlikely(OPTION_SACK_ADVERTISE & opts->options &&
397 		     !(OPTION_TS & opts->options))) {
398 		*ptr++ = htonl((TCPOPT_NOP << 24) |
399 			       (TCPOPT_NOP << 16) |
400 			       (TCPOPT_SACK_PERM << 8) |
401 			       TCPOLEN_SACK_PERM);
402 	}
403 
404 	if (unlikely(opts->ws)) {
405 		*ptr++ = htonl((TCPOPT_NOP << 24) |
406 			       (TCPOPT_WINDOW << 16) |
407 			       (TCPOLEN_WINDOW << 8) |
408 			       opts->ws);
409 	}
410 
411 	if (unlikely(opts->num_sack_blocks)) {
412 		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
413 			tp->duplicate_sack : tp->selective_acks;
414 		int this_sack;
415 
416 		*ptr++ = htonl((TCPOPT_NOP  << 24) |
417 			       (TCPOPT_NOP  << 16) |
418 			       (TCPOPT_SACK <<  8) |
419 			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
420 						     TCPOLEN_SACK_PERBLOCK)));
421 
422 		for (this_sack = 0; this_sack < opts->num_sack_blocks;
423 		     ++this_sack) {
424 			*ptr++ = htonl(sp[this_sack].start_seq);
425 			*ptr++ = htonl(sp[this_sack].end_seq);
426 		}
427 
428 		if (tp->rx_opt.dsack) {
429 			tp->rx_opt.dsack = 0;
430 			tp->rx_opt.eff_sacks--;
431 		}
432 	}
433 }
434 
435 static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
436 				struct tcp_out_options *opts,
437 				struct tcp_md5sig_key **md5) {
438 	struct tcp_sock *tp = tcp_sk(sk);
439 	unsigned size = 0;
440 
441 #ifdef CONFIG_TCP_MD5SIG
442 	*md5 = tp->af_specific->md5_lookup(sk, sk);
443 	if (*md5) {
444 		opts->options |= OPTION_MD5;
445 		size += TCPOLEN_MD5SIG_ALIGNED;
446 	}
447 #else
448 	*md5 = NULL;
449 #endif
450 
451 	/* We always get an MSS option.  The option bytes which will be seen in
452 	 * normal data packets should timestamps be used, must be in the MSS
453 	 * advertised.  But we subtract them from tp->mss_cache so that
454 	 * calculations in tcp_sendmsg are simpler etc.  So account for this
455 	 * fact here if necessary.  If we don't do this correctly, as a
456 	 * receiver we won't recognize data packets as being full sized when we
457 	 * should, and thus we won't abide by the delayed ACK rules correctly.
458 	 * SACKs don't matter, we never delay an ACK when we have any of those
459 	 * going out.  */
460 	opts->mss = tcp_advertise_mss(sk);
461 	size += TCPOLEN_MSS_ALIGNED;
462 
463 	if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
464 		opts->options |= OPTION_TS;
465 		opts->tsval = TCP_SKB_CB(skb)->when;
466 		opts->tsecr = tp->rx_opt.ts_recent;
467 		size += TCPOLEN_TSTAMP_ALIGNED;
468 	}
469 	if (likely(sysctl_tcp_window_scaling)) {
470 		opts->ws = tp->rx_opt.rcv_wscale;
471 		if(likely(opts->ws))
472 			size += TCPOLEN_WSCALE_ALIGNED;
473 	}
474 	if (likely(sysctl_tcp_sack)) {
475 		opts->options |= OPTION_SACK_ADVERTISE;
476 		if (unlikely(!(OPTION_TS & opts->options)))
477 			size += TCPOLEN_SACKPERM_ALIGNED;
478 	}
479 
480 	return size;
481 }
482 
483 static unsigned tcp_synack_options(struct sock *sk,
484 				   struct request_sock *req,
485 				   unsigned mss, struct sk_buff *skb,
486 				   struct tcp_out_options *opts,
487 				   struct tcp_md5sig_key **md5) {
488 	unsigned size = 0;
489 	struct inet_request_sock *ireq = inet_rsk(req);
490 	char doing_ts;
491 
492 #ifdef CONFIG_TCP_MD5SIG
493 	*md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
494 	if (*md5) {
495 		opts->options |= OPTION_MD5;
496 		size += TCPOLEN_MD5SIG_ALIGNED;
497 	}
498 #else
499 	*md5 = NULL;
500 #endif
501 
502 	/* we can't fit any SACK blocks in a packet with MD5 + TS
503 	   options. There was discussion about disabling SACK rather than TS in
504 	   order to fit in better with old, buggy kernels, but that was deemed
505 	   to be unnecessary. */
506 	doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok);
507 
508 	opts->mss = mss;
509 	size += TCPOLEN_MSS_ALIGNED;
510 
511 	if (likely(ireq->wscale_ok)) {
512 		opts->ws = ireq->rcv_wscale;
513 		if(likely(opts->ws))
514 			size += TCPOLEN_WSCALE_ALIGNED;
515 	}
516 	if (likely(doing_ts)) {
517 		opts->options |= OPTION_TS;
518 		opts->tsval = TCP_SKB_CB(skb)->when;
519 		opts->tsecr = req->ts_recent;
520 		size += TCPOLEN_TSTAMP_ALIGNED;
521 	}
522 	if (likely(ireq->sack_ok)) {
523 		opts->options |= OPTION_SACK_ADVERTISE;
524 		if (unlikely(!doing_ts))
525 			size += TCPOLEN_SACKPERM_ALIGNED;
526 	}
527 
528 	return size;
529 }
530 
531 static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
532 					struct tcp_out_options *opts,
533 					struct tcp_md5sig_key **md5) {
534 	struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
535 	struct tcp_sock *tp = tcp_sk(sk);
536 	unsigned size = 0;
537 
538 #ifdef CONFIG_TCP_MD5SIG
539 	*md5 = tp->af_specific->md5_lookup(sk, sk);
540 	if (unlikely(*md5)) {
541 		opts->options |= OPTION_MD5;
542 		size += TCPOLEN_MD5SIG_ALIGNED;
543 	}
544 #else
545 	*md5 = NULL;
546 #endif
547 
548 	if (likely(tp->rx_opt.tstamp_ok)) {
549 		opts->options |= OPTION_TS;
550 		opts->tsval = tcb ? tcb->when : 0;
551 		opts->tsecr = tp->rx_opt.ts_recent;
552 		size += TCPOLEN_TSTAMP_ALIGNED;
553 	}
554 
555 	if (unlikely(tp->rx_opt.eff_sacks)) {
556 		const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
557 		opts->num_sack_blocks =
558 			min_t(unsigned, tp->rx_opt.eff_sacks,
559 			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
560 			      TCPOLEN_SACK_PERBLOCK);
561 		size += TCPOLEN_SACK_BASE_ALIGNED +
562 			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
563 	}
564 
565 	return size;
566 }
567 
568 /* This routine actually transmits TCP packets queued in by
569  * tcp_do_sendmsg().  This is used by both the initial
570  * transmission and possible later retransmissions.
571  * All SKB's seen here are completely headerless.  It is our
572  * job to build the TCP header, and pass the packet down to
573  * IP so it can do the same plus pass the packet off to the
574  * device.
575  *
576  * We are working here with either a clone of the original
577  * SKB, or a fresh unique copy made by the retransmit engine.
578  */
579 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
580 			    gfp_t gfp_mask)
581 {
582 	const struct inet_connection_sock *icsk = inet_csk(sk);
583 	struct inet_sock *inet;
584 	struct tcp_sock *tp;
585 	struct tcp_skb_cb *tcb;
586 	struct tcp_out_options opts;
587 	unsigned tcp_options_size, tcp_header_size;
588 	struct tcp_md5sig_key *md5;
589 	__u8 *md5_hash_location;
590 	struct tcphdr *th;
591 	int err;
592 
593 	BUG_ON(!skb || !tcp_skb_pcount(skb));
594 
595 	/* If congestion control is doing timestamping, we must
596 	 * take such a timestamp before we potentially clone/copy.
597 	 */
598 	if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
599 		__net_timestamp(skb);
600 
601 	if (likely(clone_it)) {
602 		if (unlikely(skb_cloned(skb)))
603 			skb = pskb_copy(skb, gfp_mask);
604 		else
605 			skb = skb_clone(skb, gfp_mask);
606 		if (unlikely(!skb))
607 			return -ENOBUFS;
608 	}
609 
610 	inet = inet_sk(sk);
611 	tp = tcp_sk(sk);
612 	tcb = TCP_SKB_CB(skb);
613 	memset(&opts, 0, sizeof(opts));
614 
615 	if (unlikely(tcb->flags & TCPCB_FLAG_SYN))
616 		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
617 	else
618 		tcp_options_size = tcp_established_options(sk, skb, &opts,
619 							   &md5);
620 	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
621 
622 	if (tcp_packets_in_flight(tp) == 0)
623 		tcp_ca_event(sk, CA_EVENT_TX_START);
624 
625 	skb_push(skb, tcp_header_size);
626 	skb_reset_transport_header(skb);
627 	skb_set_owner_w(skb, sk);
628 
629 	/* Build TCP header and checksum it. */
630 	th = tcp_hdr(skb);
631 	th->source		= inet->sport;
632 	th->dest		= inet->dport;
633 	th->seq			= htonl(tcb->seq);
634 	th->ack_seq		= htonl(tp->rcv_nxt);
635 	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
636 					tcb->flags);
637 
638 	if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
639 		/* RFC1323: The window in SYN & SYN/ACK segments
640 		 * is never scaled.
641 		 */
642 		th->window	= htons(min(tp->rcv_wnd, 65535U));
643 	} else {
644 		th->window	= htons(tcp_select_window(sk));
645 	}
646 	th->check		= 0;
647 	th->urg_ptr		= 0;
648 
649 	if (unlikely(tp->urg_mode &&
650 		     between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) {
651 		th->urg_ptr		= htons(tp->snd_up - tcb->seq);
652 		th->urg			= 1;
653 	}
654 
655 	tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
656 	if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0))
657 		TCP_ECN_send(sk, skb, tcp_header_size);
658 
659 #ifdef CONFIG_TCP_MD5SIG
660 	/* Calculate the MD5 hash, as we have all we need now */
661 	if (md5) {
662 		sk->sk_route_caps &= ~NETIF_F_GSO_MASK;
663 		tp->af_specific->calc_md5_hash(md5_hash_location,
664 					       md5, sk, NULL, skb);
665 	}
666 #endif
667 
668 	icsk->icsk_af_ops->send_check(sk, skb->len, skb);
669 
670 	if (likely(tcb->flags & TCPCB_FLAG_ACK))
671 		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
672 
673 	if (skb->len != tcp_header_size)
674 		tcp_event_data_sent(tp, skb, sk);
675 
676 	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
677 		TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
678 
679 	err = icsk->icsk_af_ops->queue_xmit(skb, 0);
680 	if (likely(err <= 0))
681 		return err;
682 
683 	tcp_enter_cwr(sk, 1);
684 
685 	return net_xmit_eval(err);
686 }
687 
688 /* This routine just queue's the buffer
689  *
690  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
691  * otherwise socket can stall.
692  */
693 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
694 {
695 	struct tcp_sock *tp = tcp_sk(sk);
696 
697 	/* Advance write_seq and place onto the write_queue. */
698 	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
699 	skb_header_release(skb);
700 	tcp_add_write_queue_tail(sk, skb);
701 	sk->sk_wmem_queued += skb->truesize;
702 	sk_mem_charge(sk, skb->truesize);
703 }
704 
705 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
706 				 unsigned int mss_now)
707 {
708 	if (skb->len <= mss_now || !sk_can_gso(sk)) {
709 		/* Avoid the costly divide in the normal
710 		 * non-TSO case.
711 		 */
712 		skb_shinfo(skb)->gso_segs = 1;
713 		skb_shinfo(skb)->gso_size = 0;
714 		skb_shinfo(skb)->gso_type = 0;
715 	} else {
716 		skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
717 		skb_shinfo(skb)->gso_size = mss_now;
718 		skb_shinfo(skb)->gso_type = sk->sk_gso_type;
719 	}
720 }
721 
722 /* When a modification to fackets out becomes necessary, we need to check
723  * skb is counted to fackets_out or not.
724  */
725 static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
726 				   int decr)
727 {
728 	struct tcp_sock *tp = tcp_sk(sk);
729 
730 	if (!tp->sacked_out || tcp_is_reno(tp))
731 		return;
732 
733 	if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
734 		tp->fackets_out -= decr;
735 }
736 
737 /* Function to create two new TCP segments.  Shrinks the given segment
738  * to the specified size and appends a new segment with the rest of the
739  * packet to the list.  This won't be called frequently, I hope.
740  * Remember, these are still headerless SKBs at this point.
741  */
742 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
743 		 unsigned int mss_now)
744 {
745 	struct tcp_sock *tp = tcp_sk(sk);
746 	struct sk_buff *buff;
747 	int nsize, old_factor;
748 	int nlen;
749 	u16 flags;
750 
751 	BUG_ON(len > skb->len);
752 
753 	tcp_clear_retrans_hints_partial(tp);
754 	nsize = skb_headlen(skb) - len;
755 	if (nsize < 0)
756 		nsize = 0;
757 
758 	if (skb_cloned(skb) &&
759 	    skb_is_nonlinear(skb) &&
760 	    pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
761 		return -ENOMEM;
762 
763 	/* Get a new skb... force flag on. */
764 	buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
765 	if (buff == NULL)
766 		return -ENOMEM; /* We'll just try again later. */
767 
768 	sk->sk_wmem_queued += buff->truesize;
769 	sk_mem_charge(sk, buff->truesize);
770 	nlen = skb->len - len - nsize;
771 	buff->truesize += nlen;
772 	skb->truesize -= nlen;
773 
774 	/* Correct the sequence numbers. */
775 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
776 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
777 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
778 
779 	/* PSH and FIN should only be set in the second packet. */
780 	flags = TCP_SKB_CB(skb)->flags;
781 	TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
782 	TCP_SKB_CB(buff)->flags = flags;
783 	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
784 
785 	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
786 		/* Copy and checksum data tail into the new buffer. */
787 		buff->csum = csum_partial_copy_nocheck(skb->data + len,
788 						       skb_put(buff, nsize),
789 						       nsize, 0);
790 
791 		skb_trim(skb, len);
792 
793 		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
794 	} else {
795 		skb->ip_summed = CHECKSUM_PARTIAL;
796 		skb_split(skb, buff, len);
797 	}
798 
799 	buff->ip_summed = skb->ip_summed;
800 
801 	/* Looks stupid, but our code really uses when of
802 	 * skbs, which it never sent before. --ANK
803 	 */
804 	TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
805 	buff->tstamp = skb->tstamp;
806 
807 	old_factor = tcp_skb_pcount(skb);
808 
809 	/* Fix up tso_factor for both original and new SKB.  */
810 	tcp_set_skb_tso_segs(sk, skb, mss_now);
811 	tcp_set_skb_tso_segs(sk, buff, mss_now);
812 
813 	/* If this packet has been sent out already, we must
814 	 * adjust the various packet counters.
815 	 */
816 	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
817 		int diff = old_factor - tcp_skb_pcount(skb) -
818 			tcp_skb_pcount(buff);
819 
820 		tp->packets_out -= diff;
821 
822 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
823 			tp->sacked_out -= diff;
824 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
825 			tp->retrans_out -= diff;
826 
827 		if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
828 			tp->lost_out -= diff;
829 
830 		/* Adjust Reno SACK estimate. */
831 		if (tcp_is_reno(tp) && diff > 0) {
832 			tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
833 			tcp_verify_left_out(tp);
834 		}
835 		tcp_adjust_fackets_out(sk, skb, diff);
836 	}
837 
838 	/* Link BUFF into the send queue. */
839 	skb_header_release(buff);
840 	tcp_insert_write_queue_after(skb, buff, sk);
841 
842 	return 0;
843 }
844 
845 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
846  * eventually). The difference is that pulled data not copied, but
847  * immediately discarded.
848  */
849 static void __pskb_trim_head(struct sk_buff *skb, int len)
850 {
851 	int i, k, eat;
852 
853 	eat = len;
854 	k = 0;
855 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
856 		if (skb_shinfo(skb)->frags[i].size <= eat) {
857 			put_page(skb_shinfo(skb)->frags[i].page);
858 			eat -= skb_shinfo(skb)->frags[i].size;
859 		} else {
860 			skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
861 			if (eat) {
862 				skb_shinfo(skb)->frags[k].page_offset += eat;
863 				skb_shinfo(skb)->frags[k].size -= eat;
864 				eat = 0;
865 			}
866 			k++;
867 		}
868 	}
869 	skb_shinfo(skb)->nr_frags = k;
870 
871 	skb_reset_tail_pointer(skb);
872 	skb->data_len -= len;
873 	skb->len = skb->data_len;
874 }
875 
876 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
877 {
878 	if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
879 		return -ENOMEM;
880 
881 	/* If len == headlen, we avoid __skb_pull to preserve alignment. */
882 	if (unlikely(len < skb_headlen(skb)))
883 		__skb_pull(skb, len);
884 	else
885 		__pskb_trim_head(skb, len - skb_headlen(skb));
886 
887 	TCP_SKB_CB(skb)->seq += len;
888 	skb->ip_summed = CHECKSUM_PARTIAL;
889 
890 	skb->truesize	     -= len;
891 	sk->sk_wmem_queued   -= len;
892 	sk_mem_uncharge(sk, len);
893 	sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
894 
895 	/* Any change of skb->len requires recalculation of tso
896 	 * factor and mss.
897 	 */
898 	if (tcp_skb_pcount(skb) > 1)
899 		tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
900 
901 	return 0;
902 }
903 
904 /* Not accounting for SACKs here. */
905 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
906 {
907 	struct tcp_sock *tp = tcp_sk(sk);
908 	struct inet_connection_sock *icsk = inet_csk(sk);
909 	int mss_now;
910 
911 	/* Calculate base mss without TCP options:
912 	   It is MMS_S - sizeof(tcphdr) of rfc1122
913 	 */
914 	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
915 
916 	/* Clamp it (mss_clamp does not include tcp options) */
917 	if (mss_now > tp->rx_opt.mss_clamp)
918 		mss_now = tp->rx_opt.mss_clamp;
919 
920 	/* Now subtract optional transport overhead */
921 	mss_now -= icsk->icsk_ext_hdr_len;
922 
923 	/* Then reserve room for full set of TCP options and 8 bytes of data */
924 	if (mss_now < 48)
925 		mss_now = 48;
926 
927 	/* Now subtract TCP options size, not including SACKs */
928 	mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
929 
930 	return mss_now;
931 }
932 
933 /* Inverse of above */
934 int tcp_mss_to_mtu(struct sock *sk, int mss)
935 {
936 	struct tcp_sock *tp = tcp_sk(sk);
937 	struct inet_connection_sock *icsk = inet_csk(sk);
938 	int mtu;
939 
940 	mtu = mss +
941 	      tp->tcp_header_len +
942 	      icsk->icsk_ext_hdr_len +
943 	      icsk->icsk_af_ops->net_header_len;
944 
945 	return mtu;
946 }
947 
948 void tcp_mtup_init(struct sock *sk)
949 {
950 	struct tcp_sock *tp = tcp_sk(sk);
951 	struct inet_connection_sock *icsk = inet_csk(sk);
952 
953 	icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
954 	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
955 			       icsk->icsk_af_ops->net_header_len;
956 	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
957 	icsk->icsk_mtup.probe_size = 0;
958 }
959 
960 /* Bound MSS / TSO packet size with the half of the window */
961 static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
962 {
963 	if (tp->max_window && pktsize > (tp->max_window >> 1))
964 		return max(tp->max_window >> 1, 68U - tp->tcp_header_len);
965 	else
966 		return pktsize;
967 }
968 
969 /* This function synchronize snd mss to current pmtu/exthdr set.
970 
971    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
972    for TCP options, but includes only bare TCP header.
973 
974    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
975    It is minimum of user_mss and mss received with SYN.
976    It also does not include TCP options.
977 
978    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
979 
980    tp->mss_cache is current effective sending mss, including
981    all tcp options except for SACKs. It is evaluated,
982    taking into account current pmtu, but never exceeds
983    tp->rx_opt.mss_clamp.
984 
985    NOTE1. rfc1122 clearly states that advertised MSS
986    DOES NOT include either tcp or ip options.
987 
988    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
989    are READ ONLY outside this function.		--ANK (980731)
990  */
991 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
992 {
993 	struct tcp_sock *tp = tcp_sk(sk);
994 	struct inet_connection_sock *icsk = inet_csk(sk);
995 	int mss_now;
996 
997 	if (icsk->icsk_mtup.search_high > pmtu)
998 		icsk->icsk_mtup.search_high = pmtu;
999 
1000 	mss_now = tcp_mtu_to_mss(sk, pmtu);
1001 	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1002 
1003 	/* And store cached results */
1004 	icsk->icsk_pmtu_cookie = pmtu;
1005 	if (icsk->icsk_mtup.enabled)
1006 		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1007 	tp->mss_cache = mss_now;
1008 
1009 	return mss_now;
1010 }
1011 
1012 /* Compute the current effective MSS, taking SACKs and IP options,
1013  * and even PMTU discovery events into account.
1014  *
1015  * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
1016  * cannot be large. However, taking into account rare use of URG, this
1017  * is not a big flaw.
1018  */
1019 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
1020 {
1021 	struct tcp_sock *tp = tcp_sk(sk);
1022 	struct dst_entry *dst = __sk_dst_get(sk);
1023 	u32 mss_now;
1024 	u16 xmit_size_goal;
1025 	int doing_tso = 0;
1026 	unsigned header_len;
1027 	struct tcp_out_options opts;
1028 	struct tcp_md5sig_key *md5;
1029 
1030 	mss_now = tp->mss_cache;
1031 
1032 	if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
1033 		doing_tso = 1;
1034 
1035 	if (dst) {
1036 		u32 mtu = dst_mtu(dst);
1037 		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1038 			mss_now = tcp_sync_mss(sk, mtu);
1039 	}
1040 
1041 	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1042 		     sizeof(struct tcphdr);
1043 	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1044 	 * some common options. If this is an odd packet (because we have SACK
1045 	 * blocks etc) then our calculated header_len will be different, and
1046 	 * we have to adjust mss_now correspondingly */
1047 	if (header_len != tp->tcp_header_len) {
1048 		int delta = (int) header_len - tp->tcp_header_len;
1049 		mss_now -= delta;
1050 	}
1051 
1052 	xmit_size_goal = mss_now;
1053 
1054 	if (doing_tso) {
1055 		xmit_size_goal = ((sk->sk_gso_max_size - 1) -
1056 				  inet_csk(sk)->icsk_af_ops->net_header_len -
1057 				  inet_csk(sk)->icsk_ext_hdr_len -
1058 				  tp->tcp_header_len);
1059 
1060 		xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
1061 		xmit_size_goal -= (xmit_size_goal % mss_now);
1062 	}
1063 	tp->xmit_size_goal = xmit_size_goal;
1064 
1065 	return mss_now;
1066 }
1067 
1068 /* Congestion window validation. (RFC2861) */
1069 static void tcp_cwnd_validate(struct sock *sk)
1070 {
1071 	struct tcp_sock *tp = tcp_sk(sk);
1072 
1073 	if (tp->packets_out >= tp->snd_cwnd) {
1074 		/* Network is feed fully. */
1075 		tp->snd_cwnd_used = 0;
1076 		tp->snd_cwnd_stamp = tcp_time_stamp;
1077 	} else {
1078 		/* Network starves. */
1079 		if (tp->packets_out > tp->snd_cwnd_used)
1080 			tp->snd_cwnd_used = tp->packets_out;
1081 
1082 		if (sysctl_tcp_slow_start_after_idle &&
1083 		    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1084 			tcp_cwnd_application_limited(sk);
1085 	}
1086 }
1087 
1088 /* Returns the portion of skb which can be sent right away without
1089  * introducing MSS oddities to segment boundaries. In rare cases where
1090  * mss_now != mss_cache, we will request caller to create a small skb
1091  * per input skb which could be mostly avoided here (if desired).
1092  *
1093  * We explicitly want to create a request for splitting write queue tail
1094  * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1095  * thus all the complexity (cwnd_len is always MSS multiple which we
1096  * return whenever allowed by the other factors). Basically we need the
1097  * modulo only when the receiver window alone is the limiting factor or
1098  * when we would be allowed to send the split-due-to-Nagle skb fully.
1099  */
1100 static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1101 					unsigned int mss_now, unsigned int cwnd)
1102 {
1103 	struct tcp_sock *tp = tcp_sk(sk);
1104 	u32 needed, window, cwnd_len;
1105 
1106 	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1107 	cwnd_len = mss_now * cwnd;
1108 
1109 	if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1110 		return cwnd_len;
1111 
1112 	needed = min(skb->len, window);
1113 
1114 	if (cwnd_len <= needed)
1115 		return cwnd_len;
1116 
1117 	return needed - needed % mss_now;
1118 }
1119 
1120 /* Can at least one segment of SKB be sent right now, according to the
1121  * congestion window rules?  If so, return how many segments are allowed.
1122  */
1123 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1124 					 struct sk_buff *skb)
1125 {
1126 	u32 in_flight, cwnd;
1127 
1128 	/* Don't be strict about the congestion window for the final FIN.  */
1129 	if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1130 	    tcp_skb_pcount(skb) == 1)
1131 		return 1;
1132 
1133 	in_flight = tcp_packets_in_flight(tp);
1134 	cwnd = tp->snd_cwnd;
1135 	if (in_flight < cwnd)
1136 		return (cwnd - in_flight);
1137 
1138 	return 0;
1139 }
1140 
1141 /* This must be invoked the first time we consider transmitting
1142  * SKB onto the wire.
1143  */
1144 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1145 			     unsigned int mss_now)
1146 {
1147 	int tso_segs = tcp_skb_pcount(skb);
1148 
1149 	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1150 		tcp_set_skb_tso_segs(sk, skb, mss_now);
1151 		tso_segs = tcp_skb_pcount(skb);
1152 	}
1153 	return tso_segs;
1154 }
1155 
1156 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1157 {
1158 	return after(tp->snd_sml,tp->snd_una) &&
1159 		!after(tp->snd_sml, tp->snd_nxt);
1160 }
1161 
1162 /* Return 0, if packet can be sent now without violation Nagle's rules:
1163  * 1. It is full sized.
1164  * 2. Or it contains FIN. (already checked by caller)
1165  * 3. Or TCP_NODELAY was set.
1166  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1167  *    With Minshall's modification: all sent small packets are ACKed.
1168  */
1169 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1170 				  const struct sk_buff *skb,
1171 				  unsigned mss_now, int nonagle)
1172 {
1173 	return (skb->len < mss_now &&
1174 		((nonagle & TCP_NAGLE_CORK) ||
1175 		 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
1176 }
1177 
1178 /* Return non-zero if the Nagle test allows this packet to be
1179  * sent now.
1180  */
1181 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1182 				 unsigned int cur_mss, int nonagle)
1183 {
1184 	/* Nagle rule does not apply to frames, which sit in the middle of the
1185 	 * write_queue (they have no chances to get new data).
1186 	 *
1187 	 * This is implemented in the callers, where they modify the 'nonagle'
1188 	 * argument based upon the location of SKB in the send queue.
1189 	 */
1190 	if (nonagle & TCP_NAGLE_PUSH)
1191 		return 1;
1192 
1193 	/* Don't use the nagle rule for urgent data (or for the final FIN).
1194 	 * Nagle can be ignored during F-RTO too (see RFC4138).
1195 	 */
1196 	if (tp->urg_mode || (tp->frto_counter == 2) ||
1197 	    (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1198 		return 1;
1199 
1200 	if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1201 		return 1;
1202 
1203 	return 0;
1204 }
1205 
1206 /* Does at least the first segment of SKB fit into the send window? */
1207 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1208 				   unsigned int cur_mss)
1209 {
1210 	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1211 
1212 	if (skb->len > cur_mss)
1213 		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1214 
1215 	return !after(end_seq, tcp_wnd_end(tp));
1216 }
1217 
1218 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1219  * should be put on the wire right now.  If so, it returns the number of
1220  * packets allowed by the congestion window.
1221  */
1222 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1223 				 unsigned int cur_mss, int nonagle)
1224 {
1225 	struct tcp_sock *tp = tcp_sk(sk);
1226 	unsigned int cwnd_quota;
1227 
1228 	tcp_init_tso_segs(sk, skb, cur_mss);
1229 
1230 	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1231 		return 0;
1232 
1233 	cwnd_quota = tcp_cwnd_test(tp, skb);
1234 	if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1235 		cwnd_quota = 0;
1236 
1237 	return cwnd_quota;
1238 }
1239 
1240 int tcp_may_send_now(struct sock *sk)
1241 {
1242 	struct tcp_sock *tp = tcp_sk(sk);
1243 	struct sk_buff *skb = tcp_send_head(sk);
1244 
1245 	return (skb &&
1246 		tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1247 			     (tcp_skb_is_last(sk, skb) ?
1248 			      tp->nonagle : TCP_NAGLE_PUSH)));
1249 }
1250 
1251 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1252  * which is put after SKB on the list.  It is very much like
1253  * tcp_fragment() except that it may make several kinds of assumptions
1254  * in order to speed up the splitting operation.  In particular, we
1255  * know that all the data is in scatter-gather pages, and that the
1256  * packet has never been sent out before (and thus is not cloned).
1257  */
1258 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1259 			unsigned int mss_now)
1260 {
1261 	struct sk_buff *buff;
1262 	int nlen = skb->len - len;
1263 	u16 flags;
1264 
1265 	/* All of a TSO frame must be composed of paged data.  */
1266 	if (skb->len != skb->data_len)
1267 		return tcp_fragment(sk, skb, len, mss_now);
1268 
1269 	buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
1270 	if (unlikely(buff == NULL))
1271 		return -ENOMEM;
1272 
1273 	sk->sk_wmem_queued += buff->truesize;
1274 	sk_mem_charge(sk, buff->truesize);
1275 	buff->truesize += nlen;
1276 	skb->truesize -= nlen;
1277 
1278 	/* Correct the sequence numbers. */
1279 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1280 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1281 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1282 
1283 	/* PSH and FIN should only be set in the second packet. */
1284 	flags = TCP_SKB_CB(skb)->flags;
1285 	TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
1286 	TCP_SKB_CB(buff)->flags = flags;
1287 
1288 	/* This packet was never sent out yet, so no SACK bits. */
1289 	TCP_SKB_CB(buff)->sacked = 0;
1290 
1291 	buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1292 	skb_split(skb, buff, len);
1293 
1294 	/* Fix up tso_factor for both original and new SKB.  */
1295 	tcp_set_skb_tso_segs(sk, skb, mss_now);
1296 	tcp_set_skb_tso_segs(sk, buff, mss_now);
1297 
1298 	/* Link BUFF into the send queue. */
1299 	skb_header_release(buff);
1300 	tcp_insert_write_queue_after(skb, buff, sk);
1301 
1302 	return 0;
1303 }
1304 
1305 /* Try to defer sending, if possible, in order to minimize the amount
1306  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1307  *
1308  * This algorithm is from John Heffner.
1309  */
1310 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1311 {
1312 	struct tcp_sock *tp = tcp_sk(sk);
1313 	const struct inet_connection_sock *icsk = inet_csk(sk);
1314 	u32 send_win, cong_win, limit, in_flight;
1315 
1316 	if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1317 		goto send_now;
1318 
1319 	if (icsk->icsk_ca_state != TCP_CA_Open)
1320 		goto send_now;
1321 
1322 	/* Defer for less than two clock ticks. */
1323 	if (tp->tso_deferred &&
1324 	    ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1325 		goto send_now;
1326 
1327 	in_flight = tcp_packets_in_flight(tp);
1328 
1329 	BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1330 
1331 	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1332 
1333 	/* From in_flight test above, we know that cwnd > in_flight.  */
1334 	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1335 
1336 	limit = min(send_win, cong_win);
1337 
1338 	/* If a full-sized TSO skb can be sent, do it. */
1339 	if (limit >= sk->sk_gso_max_size)
1340 		goto send_now;
1341 
1342 	if (sysctl_tcp_tso_win_divisor) {
1343 		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1344 
1345 		/* If at least some fraction of a window is available,
1346 		 * just use it.
1347 		 */
1348 		chunk /= sysctl_tcp_tso_win_divisor;
1349 		if (limit >= chunk)
1350 			goto send_now;
1351 	} else {
1352 		/* Different approach, try not to defer past a single
1353 		 * ACK.  Receiver should ACK every other full sized
1354 		 * frame, so if we have space for more than 3 frames
1355 		 * then send now.
1356 		 */
1357 		if (limit > tcp_max_burst(tp) * tp->mss_cache)
1358 			goto send_now;
1359 	}
1360 
1361 	/* Ok, it looks like it is advisable to defer.  */
1362 	tp->tso_deferred = 1 | (jiffies << 1);
1363 
1364 	return 1;
1365 
1366 send_now:
1367 	tp->tso_deferred = 0;
1368 	return 0;
1369 }
1370 
1371 /* Create a new MTU probe if we are ready.
1372  * Returns 0 if we should wait to probe (no cwnd available),
1373  *         1 if a probe was sent,
1374  *         -1 otherwise
1375  */
1376 static int tcp_mtu_probe(struct sock *sk)
1377 {
1378 	struct tcp_sock *tp = tcp_sk(sk);
1379 	struct inet_connection_sock *icsk = inet_csk(sk);
1380 	struct sk_buff *skb, *nskb, *next;
1381 	int len;
1382 	int probe_size;
1383 	int size_needed;
1384 	int copy;
1385 	int mss_now;
1386 
1387 	/* Not currently probing/verifying,
1388 	 * not in recovery,
1389 	 * have enough cwnd, and
1390 	 * not SACKing (the variable headers throw things off) */
1391 	if (!icsk->icsk_mtup.enabled ||
1392 	    icsk->icsk_mtup.probe_size ||
1393 	    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1394 	    tp->snd_cwnd < 11 ||
1395 	    tp->rx_opt.eff_sacks)
1396 		return -1;
1397 
1398 	/* Very simple search strategy: just double the MSS. */
1399 	mss_now = tcp_current_mss(sk, 0);
1400 	probe_size = 2 * tp->mss_cache;
1401 	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1402 	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1403 		/* TODO: set timer for probe_converge_event */
1404 		return -1;
1405 	}
1406 
1407 	/* Have enough data in the send queue to probe? */
1408 	if (tp->write_seq - tp->snd_nxt < size_needed)
1409 		return -1;
1410 
1411 	if (tp->snd_wnd < size_needed)
1412 		return -1;
1413 	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1414 		return 0;
1415 
1416 	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
1417 	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1418 		if (!tcp_packets_in_flight(tp))
1419 			return -1;
1420 		else
1421 			return 0;
1422 	}
1423 
1424 	/* We're allowed to probe.  Build it now. */
1425 	if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1426 		return -1;
1427 	sk->sk_wmem_queued += nskb->truesize;
1428 	sk_mem_charge(sk, nskb->truesize);
1429 
1430 	skb = tcp_send_head(sk);
1431 
1432 	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1433 	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1434 	TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1435 	TCP_SKB_CB(nskb)->sacked = 0;
1436 	nskb->csum = 0;
1437 	nskb->ip_summed = skb->ip_summed;
1438 
1439 	tcp_insert_write_queue_before(nskb, skb, sk);
1440 
1441 	len = 0;
1442 	tcp_for_write_queue_from_safe(skb, next, sk) {
1443 		copy = min_t(int, skb->len, probe_size - len);
1444 		if (nskb->ip_summed)
1445 			skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1446 		else
1447 			nskb->csum = skb_copy_and_csum_bits(skb, 0,
1448 							    skb_put(nskb, copy),
1449 							    copy, nskb->csum);
1450 
1451 		if (skb->len <= copy) {
1452 			/* We've eaten all the data from this skb.
1453 			 * Throw it away. */
1454 			TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1455 			tcp_unlink_write_queue(skb, sk);
1456 			sk_wmem_free_skb(sk, skb);
1457 		} else {
1458 			TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1459 						   ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1460 			if (!skb_shinfo(skb)->nr_frags) {
1461 				skb_pull(skb, copy);
1462 				if (skb->ip_summed != CHECKSUM_PARTIAL)
1463 					skb->csum = csum_partial(skb->data,
1464 								 skb->len, 0);
1465 			} else {
1466 				__pskb_trim_head(skb, copy);
1467 				tcp_set_skb_tso_segs(sk, skb, mss_now);
1468 			}
1469 			TCP_SKB_CB(skb)->seq += copy;
1470 		}
1471 
1472 		len += copy;
1473 
1474 		if (len >= probe_size)
1475 			break;
1476 	}
1477 	tcp_init_tso_segs(sk, nskb, nskb->len);
1478 
1479 	/* We're ready to send.  If this fails, the probe will
1480 	 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1481 	TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1482 	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1483 		/* Decrement cwnd here because we are sending
1484 		 * effectively two packets. */
1485 		tp->snd_cwnd--;
1486 		tcp_event_new_data_sent(sk, nskb);
1487 
1488 		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1489 		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1490 		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1491 
1492 		return 1;
1493 	}
1494 
1495 	return -1;
1496 }
1497 
1498 /* This routine writes packets to the network.  It advances the
1499  * send_head.  This happens as incoming acks open up the remote
1500  * window for us.
1501  *
1502  * Returns 1, if no segments are in flight and we have queued segments, but
1503  * cannot send anything now because of SWS or another problem.
1504  */
1505 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1506 {
1507 	struct tcp_sock *tp = tcp_sk(sk);
1508 	struct sk_buff *skb;
1509 	unsigned int tso_segs, sent_pkts;
1510 	int cwnd_quota;
1511 	int result;
1512 
1513 	/* If we are closed, the bytes will have to remain here.
1514 	 * In time closedown will finish, we empty the write queue and all
1515 	 * will be happy.
1516 	 */
1517 	if (unlikely(sk->sk_state == TCP_CLOSE))
1518 		return 0;
1519 
1520 	sent_pkts = 0;
1521 
1522 	/* Do MTU probing. */
1523 	if ((result = tcp_mtu_probe(sk)) == 0) {
1524 		return 0;
1525 	} else if (result > 0) {
1526 		sent_pkts = 1;
1527 	}
1528 
1529 	while ((skb = tcp_send_head(sk))) {
1530 		unsigned int limit;
1531 
1532 		tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1533 		BUG_ON(!tso_segs);
1534 
1535 		cwnd_quota = tcp_cwnd_test(tp, skb);
1536 		if (!cwnd_quota)
1537 			break;
1538 
1539 		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1540 			break;
1541 
1542 		if (tso_segs == 1) {
1543 			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1544 						     (tcp_skb_is_last(sk, skb) ?
1545 						      nonagle : TCP_NAGLE_PUSH))))
1546 				break;
1547 		} else {
1548 			if (tcp_tso_should_defer(sk, skb))
1549 				break;
1550 		}
1551 
1552 		limit = mss_now;
1553 		if (tso_segs > 1)
1554 			limit = tcp_mss_split_point(sk, skb, mss_now,
1555 						    cwnd_quota);
1556 
1557 		if (skb->len > limit &&
1558 		    unlikely(tso_fragment(sk, skb, limit, mss_now)))
1559 			break;
1560 
1561 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
1562 
1563 		if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1564 			break;
1565 
1566 		/* Advance the send_head.  This one is sent out.
1567 		 * This call will increment packets_out.
1568 		 */
1569 		tcp_event_new_data_sent(sk, skb);
1570 
1571 		tcp_minshall_update(tp, mss_now, skb);
1572 		sent_pkts++;
1573 	}
1574 
1575 	if (likely(sent_pkts)) {
1576 		tcp_cwnd_validate(sk);
1577 		return 0;
1578 	}
1579 	return !tp->packets_out && tcp_send_head(sk);
1580 }
1581 
1582 /* Push out any pending frames which were held back due to
1583  * TCP_CORK or attempt at coalescing tiny packets.
1584  * The socket must be locked by the caller.
1585  */
1586 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1587 			       int nonagle)
1588 {
1589 	struct sk_buff *skb = tcp_send_head(sk);
1590 
1591 	if (skb) {
1592 		if (tcp_write_xmit(sk, cur_mss, nonagle))
1593 			tcp_check_probe_timer(sk);
1594 	}
1595 }
1596 
1597 /* Send _single_ skb sitting at the send head. This function requires
1598  * true push pending frames to setup probe timer etc.
1599  */
1600 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1601 {
1602 	struct sk_buff *skb = tcp_send_head(sk);
1603 	unsigned int tso_segs, cwnd_quota;
1604 
1605 	BUG_ON(!skb || skb->len < mss_now);
1606 
1607 	tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1608 	cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1609 
1610 	if (likely(cwnd_quota)) {
1611 		unsigned int limit;
1612 
1613 		BUG_ON(!tso_segs);
1614 
1615 		limit = mss_now;
1616 		if (tso_segs > 1)
1617 			limit = tcp_mss_split_point(sk, skb, mss_now,
1618 						    cwnd_quota);
1619 
1620 		if (skb->len > limit &&
1621 		    unlikely(tso_fragment(sk, skb, limit, mss_now)))
1622 			return;
1623 
1624 		/* Send it out now. */
1625 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
1626 
1627 		if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1628 			tcp_event_new_data_sent(sk, skb);
1629 			tcp_cwnd_validate(sk);
1630 			return;
1631 		}
1632 	}
1633 }
1634 
1635 /* This function returns the amount that we can raise the
1636  * usable window based on the following constraints
1637  *
1638  * 1. The window can never be shrunk once it is offered (RFC 793)
1639  * 2. We limit memory per socket
1640  *
1641  * RFC 1122:
1642  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1643  *  RECV.NEXT + RCV.WIN fixed until:
1644  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1645  *
1646  * i.e. don't raise the right edge of the window until you can raise
1647  * it at least MSS bytes.
1648  *
1649  * Unfortunately, the recommended algorithm breaks header prediction,
1650  * since header prediction assumes th->window stays fixed.
1651  *
1652  * Strictly speaking, keeping th->window fixed violates the receiver
1653  * side SWS prevention criteria. The problem is that under this rule
1654  * a stream of single byte packets will cause the right side of the
1655  * window to always advance by a single byte.
1656  *
1657  * Of course, if the sender implements sender side SWS prevention
1658  * then this will not be a problem.
1659  *
1660  * BSD seems to make the following compromise:
1661  *
1662  *	If the free space is less than the 1/4 of the maximum
1663  *	space available and the free space is less than 1/2 mss,
1664  *	then set the window to 0.
1665  *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1666  *	Otherwise, just prevent the window from shrinking
1667  *	and from being larger than the largest representable value.
1668  *
1669  * This prevents incremental opening of the window in the regime
1670  * where TCP is limited by the speed of the reader side taking
1671  * data out of the TCP receive queue. It does nothing about
1672  * those cases where the window is constrained on the sender side
1673  * because the pipeline is full.
1674  *
1675  * BSD also seems to "accidentally" limit itself to windows that are a
1676  * multiple of MSS, at least until the free space gets quite small.
1677  * This would appear to be a side effect of the mbuf implementation.
1678  * Combining these two algorithms results in the observed behavior
1679  * of having a fixed window size at almost all times.
1680  *
1681  * Below we obtain similar behavior by forcing the offered window to
1682  * a multiple of the mss when it is feasible to do so.
1683  *
1684  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1685  * Regular options like TIMESTAMP are taken into account.
1686  */
1687 u32 __tcp_select_window(struct sock *sk)
1688 {
1689 	struct inet_connection_sock *icsk = inet_csk(sk);
1690 	struct tcp_sock *tp = tcp_sk(sk);
1691 	/* MSS for the peer's data.  Previous versions used mss_clamp
1692 	 * here.  I don't know if the value based on our guesses
1693 	 * of peer's MSS is better for the performance.  It's more correct
1694 	 * but may be worse for the performance because of rcv_mss
1695 	 * fluctuations.  --SAW  1998/11/1
1696 	 */
1697 	int mss = icsk->icsk_ack.rcv_mss;
1698 	int free_space = tcp_space(sk);
1699 	int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1700 	int window;
1701 
1702 	if (mss > full_space)
1703 		mss = full_space;
1704 
1705 	if (free_space < (full_space >> 1)) {
1706 		icsk->icsk_ack.quick = 0;
1707 
1708 		if (tcp_memory_pressure)
1709 			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1710 					       4U * tp->advmss);
1711 
1712 		if (free_space < mss)
1713 			return 0;
1714 	}
1715 
1716 	if (free_space > tp->rcv_ssthresh)
1717 		free_space = tp->rcv_ssthresh;
1718 
1719 	/* Don't do rounding if we are using window scaling, since the
1720 	 * scaled window will not line up with the MSS boundary anyway.
1721 	 */
1722 	window = tp->rcv_wnd;
1723 	if (tp->rx_opt.rcv_wscale) {
1724 		window = free_space;
1725 
1726 		/* Advertise enough space so that it won't get scaled away.
1727 		 * Import case: prevent zero window announcement if
1728 		 * 1<<rcv_wscale > mss.
1729 		 */
1730 		if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1731 			window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1732 				  << tp->rx_opt.rcv_wscale);
1733 	} else {
1734 		/* Get the largest window that is a nice multiple of mss.
1735 		 * Window clamp already applied above.
1736 		 * If our current window offering is within 1 mss of the
1737 		 * free space we just keep it. This prevents the divide
1738 		 * and multiply from happening most of the time.
1739 		 * We also don't do any window rounding when the free space
1740 		 * is too small.
1741 		 */
1742 		if (window <= free_space - mss || window > free_space)
1743 			window = (free_space / mss) * mss;
1744 		else if (mss == full_space &&
1745 			 free_space > window + (full_space >> 1))
1746 			window = free_space;
1747 	}
1748 
1749 	return window;
1750 }
1751 
1752 /* Attempt to collapse two adjacent SKB's during retransmission. */
1753 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb,
1754 				     int mss_now)
1755 {
1756 	struct tcp_sock *tp = tcp_sk(sk);
1757 	struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1758 	int skb_size, next_skb_size;
1759 	u16 flags;
1760 
1761 	/* The first test we must make is that neither of these two
1762 	 * SKB's are still referenced by someone else.
1763 	 */
1764 	if (skb_cloned(skb) || skb_cloned(next_skb))
1765 		return;
1766 
1767 	skb_size = skb->len;
1768 	next_skb_size = next_skb->len;
1769 	flags = TCP_SKB_CB(skb)->flags;
1770 
1771 	/* Also punt if next skb has been SACK'd. */
1772 	if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1773 		return;
1774 
1775 	/* Next skb is out of window. */
1776 	if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp)))
1777 		return;
1778 
1779 	/* Punt if not enough space exists in the first SKB for
1780 	 * the data in the second, or the total combined payload
1781 	 * would exceed the MSS.
1782 	 */
1783 	if ((next_skb_size > skb_tailroom(skb)) ||
1784 	    ((skb_size + next_skb_size) > mss_now))
1785 		return;
1786 
1787 	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1788 
1789 	tcp_highest_sack_combine(sk, next_skb, skb);
1790 
1791 	/* Ok.	We will be able to collapse the packet. */
1792 	tcp_unlink_write_queue(next_skb, sk);
1793 
1794 	skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1795 				  next_skb_size);
1796 
1797 	if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1798 		skb->ip_summed = CHECKSUM_PARTIAL;
1799 
1800 	if (skb->ip_summed != CHECKSUM_PARTIAL)
1801 		skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1802 
1803 	/* Update sequence range on original skb. */
1804 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1805 
1806 	/* Merge over control information. */
1807 	flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1808 	TCP_SKB_CB(skb)->flags = flags;
1809 
1810 	/* All done, get rid of second SKB and account for it so
1811 	 * packet counting does not break.
1812 	 */
1813 	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1814 	if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS)
1815 		tp->retrans_out -= tcp_skb_pcount(next_skb);
1816 	if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST)
1817 		tp->lost_out -= tcp_skb_pcount(next_skb);
1818 	/* Reno case is special. Sigh... */
1819 	if (tcp_is_reno(tp) && tp->sacked_out)
1820 		tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1821 
1822 	tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
1823 	tp->packets_out -= tcp_skb_pcount(next_skb);
1824 
1825 	/* changed transmit queue under us so clear hints */
1826 	tcp_clear_retrans_hints_partial(tp);
1827 
1828 	sk_wmem_free_skb(sk, next_skb);
1829 }
1830 
1831 /* Do a simple retransmit without using the backoff mechanisms in
1832  * tcp_timer. This is used for path mtu discovery.
1833  * The socket is already locked here.
1834  */
1835 void tcp_simple_retransmit(struct sock *sk)
1836 {
1837 	const struct inet_connection_sock *icsk = inet_csk(sk);
1838 	struct tcp_sock *tp = tcp_sk(sk);
1839 	struct sk_buff *skb;
1840 	unsigned int mss = tcp_current_mss(sk, 0);
1841 	int lost = 0;
1842 
1843 	tcp_for_write_queue(skb, sk) {
1844 		if (skb == tcp_send_head(sk))
1845 			break;
1846 		if (skb->len > mss &&
1847 		    !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1848 			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1849 				TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1850 				tp->retrans_out -= tcp_skb_pcount(skb);
1851 			}
1852 			if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) {
1853 				TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1854 				tp->lost_out += tcp_skb_pcount(skb);
1855 				lost = 1;
1856 			}
1857 		}
1858 	}
1859 
1860 	tcp_clear_all_retrans_hints(tp);
1861 
1862 	if (!lost)
1863 		return;
1864 
1865 	if (tcp_is_reno(tp))
1866 		tcp_limit_reno_sacked(tp);
1867 
1868 	tcp_verify_left_out(tp);
1869 
1870 	/* Don't muck with the congestion window here.
1871 	 * Reason is that we do not increase amount of _data_
1872 	 * in network, but units changed and effective
1873 	 * cwnd/ssthresh really reduced now.
1874 	 */
1875 	if (icsk->icsk_ca_state != TCP_CA_Loss) {
1876 		tp->high_seq = tp->snd_nxt;
1877 		tp->snd_ssthresh = tcp_current_ssthresh(sk);
1878 		tp->prior_ssthresh = 0;
1879 		tp->undo_marker = 0;
1880 		tcp_set_ca_state(sk, TCP_CA_Loss);
1881 	}
1882 	tcp_xmit_retransmit_queue(sk);
1883 }
1884 
1885 /* This retransmits one SKB.  Policy decisions and retransmit queue
1886  * state updates are done by the caller.  Returns non-zero if an
1887  * error occurred which prevented the send.
1888  */
1889 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1890 {
1891 	struct tcp_sock *tp = tcp_sk(sk);
1892 	struct inet_connection_sock *icsk = inet_csk(sk);
1893 	unsigned int cur_mss;
1894 	int err;
1895 
1896 	/* Inconslusive MTU probe */
1897 	if (icsk->icsk_mtup.probe_size) {
1898 		icsk->icsk_mtup.probe_size = 0;
1899 	}
1900 
1901 	/* Do not sent more than we queued. 1/4 is reserved for possible
1902 	 * copying overhead: fragmentation, tunneling, mangling etc.
1903 	 */
1904 	if (atomic_read(&sk->sk_wmem_alloc) >
1905 	    min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1906 		return -EAGAIN;
1907 
1908 	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1909 		if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1910 			BUG();
1911 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1912 			return -ENOMEM;
1913 	}
1914 
1915 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1916 		return -EHOSTUNREACH; /* Routing failure or similar. */
1917 
1918 	cur_mss = tcp_current_mss(sk, 0);
1919 
1920 	/* If receiver has shrunk his window, and skb is out of
1921 	 * new window, do not retransmit it. The exception is the
1922 	 * case, when window is shrunk to zero. In this case
1923 	 * our retransmit serves as a zero window probe.
1924 	 */
1925 	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
1926 	    && TCP_SKB_CB(skb)->seq != tp->snd_una)
1927 		return -EAGAIN;
1928 
1929 	if (skb->len > cur_mss) {
1930 		if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1931 			return -ENOMEM; /* We'll try again later. */
1932 	}
1933 
1934 	/* Collapse two adjacent packets if worthwhile and we can. */
1935 	if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1936 	    (skb->len < (cur_mss >> 1)) &&
1937 	    (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1938 	    (!tcp_skb_is_last(sk, skb)) &&
1939 	    (skb_shinfo(skb)->nr_frags == 0 &&
1940 	     skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1941 	    (tcp_skb_pcount(skb) == 1 &&
1942 	     tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
1943 	    (sysctl_tcp_retrans_collapse != 0))
1944 		tcp_retrans_try_collapse(sk, skb, cur_mss);
1945 
1946 	/* Some Solaris stacks overoptimize and ignore the FIN on a
1947 	 * retransmit when old data is attached.  So strip it off
1948 	 * since it is cheap to do so and saves bytes on the network.
1949 	 */
1950 	if (skb->len > 0 &&
1951 	    (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1952 	    tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1953 		if (!pskb_trim(skb, 0)) {
1954 			/* Reuse, even though it does some unnecessary work */
1955 			tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
1956 					     TCP_SKB_CB(skb)->flags);
1957 			skb->ip_summed = CHECKSUM_NONE;
1958 		}
1959 	}
1960 
1961 	/* Make a copy, if the first transmission SKB clone we made
1962 	 * is still in somebody's hands, else make a clone.
1963 	 */
1964 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
1965 
1966 	err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1967 
1968 	if (err == 0) {
1969 		/* Update global TCP statistics. */
1970 		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
1971 
1972 		tp->total_retrans++;
1973 
1974 #if FASTRETRANS_DEBUG > 0
1975 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1976 			if (net_ratelimit())
1977 				printk(KERN_DEBUG "retrans_out leaked.\n");
1978 		}
1979 #endif
1980 		if (!tp->retrans_out)
1981 			tp->lost_retrans_low = tp->snd_nxt;
1982 		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1983 		tp->retrans_out += tcp_skb_pcount(skb);
1984 
1985 		/* Save stamp of the first retransmit. */
1986 		if (!tp->retrans_stamp)
1987 			tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1988 
1989 		tp->undo_retrans++;
1990 
1991 		/* snd_nxt is stored to detect loss of retransmitted segment,
1992 		 * see tcp_input.c tcp_sacktag_write_queue().
1993 		 */
1994 		TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1995 	}
1996 	return err;
1997 }
1998 
1999 /* This gets called after a retransmit timeout, and the initially
2000  * retransmitted data is acknowledged.  It tries to continue
2001  * resending the rest of the retransmit queue, until either
2002  * we've sent it all or the congestion window limit is reached.
2003  * If doing SACK, the first ACK which comes back for a timeout
2004  * based retransmit packet might feed us FACK information again.
2005  * If so, we use it to avoid unnecessarily retransmissions.
2006  */
2007 void tcp_xmit_retransmit_queue(struct sock *sk)
2008 {
2009 	const struct inet_connection_sock *icsk = inet_csk(sk);
2010 	struct tcp_sock *tp = tcp_sk(sk);
2011 	struct sk_buff *skb;
2012 	int packet_cnt;
2013 
2014 	if (tp->retransmit_skb_hint) {
2015 		skb = tp->retransmit_skb_hint;
2016 		packet_cnt = tp->retransmit_cnt_hint;
2017 	} else {
2018 		skb = tcp_write_queue_head(sk);
2019 		packet_cnt = 0;
2020 	}
2021 
2022 	/* First pass: retransmit lost packets. */
2023 	if (tp->lost_out) {
2024 		tcp_for_write_queue_from(skb, sk) {
2025 			__u8 sacked = TCP_SKB_CB(skb)->sacked;
2026 
2027 			if (skb == tcp_send_head(sk))
2028 				break;
2029 			/* we could do better than to assign each time */
2030 			tp->retransmit_skb_hint = skb;
2031 			tp->retransmit_cnt_hint = packet_cnt;
2032 
2033 			/* Assume this retransmit will generate
2034 			 * only one packet for congestion window
2035 			 * calculation purposes.  This works because
2036 			 * tcp_retransmit_skb() will chop up the
2037 			 * packet to be MSS sized and all the
2038 			 * packet counting works out.
2039 			 */
2040 			if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2041 				return;
2042 
2043 			if (sacked & TCPCB_LOST) {
2044 				if (!(sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
2045 					int mib_idx;
2046 
2047 					if (tcp_retransmit_skb(sk, skb)) {
2048 						tp->retransmit_skb_hint = NULL;
2049 						return;
2050 					}
2051 					if (icsk->icsk_ca_state != TCP_CA_Loss)
2052 						mib_idx = LINUX_MIB_TCPFASTRETRANS;
2053 					else
2054 						mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2055 					NET_INC_STATS_BH(sock_net(sk), mib_idx);
2056 
2057 					if (skb == tcp_write_queue_head(sk))
2058 						inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2059 									  inet_csk(sk)->icsk_rto,
2060 									  TCP_RTO_MAX);
2061 				}
2062 
2063 				packet_cnt += tcp_skb_pcount(skb);
2064 				if (packet_cnt >= tp->lost_out)
2065 					break;
2066 			}
2067 		}
2068 	}
2069 
2070 	/* OK, demanded retransmission is finished. */
2071 
2072 	/* Forward retransmissions are possible only during Recovery. */
2073 	if (icsk->icsk_ca_state != TCP_CA_Recovery)
2074 		return;
2075 
2076 	/* No forward retransmissions in Reno are possible. */
2077 	if (tcp_is_reno(tp))
2078 		return;
2079 
2080 	/* Yeah, we have to make difficult choice between forward transmission
2081 	 * and retransmission... Both ways have their merits...
2082 	 *
2083 	 * For now we do not retransmit anything, while we have some new
2084 	 * segments to send. In the other cases, follow rule 3 for
2085 	 * NextSeg() specified in RFC3517.
2086 	 */
2087 
2088 	if (tcp_may_send_now(sk))
2089 		return;
2090 
2091 	/* If nothing is SACKed, highest_sack in the loop won't be valid */
2092 	if (!tp->sacked_out)
2093 		return;
2094 
2095 	if (tp->forward_skb_hint)
2096 		skb = tp->forward_skb_hint;
2097 	else
2098 		skb = tcp_write_queue_head(sk);
2099 
2100 	tcp_for_write_queue_from(skb, sk) {
2101 		if (skb == tcp_send_head(sk))
2102 			break;
2103 		tp->forward_skb_hint = skb;
2104 
2105 		if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2106 			break;
2107 
2108 		if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2109 			break;
2110 
2111 		if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
2112 			continue;
2113 
2114 		/* Ok, retransmit it. */
2115 		if (tcp_retransmit_skb(sk, skb)) {
2116 			tp->forward_skb_hint = NULL;
2117 			break;
2118 		}
2119 
2120 		if (skb == tcp_write_queue_head(sk))
2121 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2122 						  inet_csk(sk)->icsk_rto,
2123 						  TCP_RTO_MAX);
2124 
2125 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFORWARDRETRANS);
2126 	}
2127 }
2128 
2129 /* Send a fin.  The caller locks the socket for us.  This cannot be
2130  * allowed to fail queueing a FIN frame under any circumstances.
2131  */
2132 void tcp_send_fin(struct sock *sk)
2133 {
2134 	struct tcp_sock *tp = tcp_sk(sk);
2135 	struct sk_buff *skb = tcp_write_queue_tail(sk);
2136 	int mss_now;
2137 
2138 	/* Optimization, tack on the FIN if we have a queue of
2139 	 * unsent frames.  But be careful about outgoing SACKS
2140 	 * and IP options.
2141 	 */
2142 	mss_now = tcp_current_mss(sk, 1);
2143 
2144 	if (tcp_send_head(sk) != NULL) {
2145 		TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2146 		TCP_SKB_CB(skb)->end_seq++;
2147 		tp->write_seq++;
2148 	} else {
2149 		/* Socket is locked, keep trying until memory is available. */
2150 		for (;;) {
2151 			skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
2152 			if (skb)
2153 				break;
2154 			yield();
2155 		}
2156 
2157 		/* Reserve space for headers and prepare control bits. */
2158 		skb_reserve(skb, MAX_TCP_HEADER);
2159 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2160 		tcp_init_nondata_skb(skb, tp->write_seq,
2161 				     TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2162 		tcp_queue_skb(sk, skb);
2163 	}
2164 	__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2165 }
2166 
2167 /* We get here when a process closes a file descriptor (either due to
2168  * an explicit close() or as a byproduct of exit()'ing) and there
2169  * was unread data in the receive queue.  This behavior is recommended
2170  * by RFC 2525, section 2.17.  -DaveM
2171  */
2172 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2173 {
2174 	struct sk_buff *skb;
2175 
2176 	/* NOTE: No TCP options attached and we never retransmit this. */
2177 	skb = alloc_skb(MAX_TCP_HEADER, priority);
2178 	if (!skb) {
2179 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2180 		return;
2181 	}
2182 
2183 	/* Reserve space for headers and prepare control bits. */
2184 	skb_reserve(skb, MAX_TCP_HEADER);
2185 	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2186 			     TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2187 	/* Send it off. */
2188 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2189 	if (tcp_transmit_skb(sk, skb, 0, priority))
2190 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2191 
2192 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2193 }
2194 
2195 /* WARNING: This routine must only be called when we have already sent
2196  * a SYN packet that crossed the incoming SYN that caused this routine
2197  * to get called. If this assumption fails then the initial rcv_wnd
2198  * and rcv_wscale values will not be correct.
2199  */
2200 int tcp_send_synack(struct sock *sk)
2201 {
2202 	struct sk_buff *skb;
2203 
2204 	skb = tcp_write_queue_head(sk);
2205 	if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
2206 		printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2207 		return -EFAULT;
2208 	}
2209 	if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
2210 		if (skb_cloned(skb)) {
2211 			struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2212 			if (nskb == NULL)
2213 				return -ENOMEM;
2214 			tcp_unlink_write_queue(skb, sk);
2215 			skb_header_release(nskb);
2216 			__tcp_add_write_queue_head(sk, nskb);
2217 			sk_wmem_free_skb(sk, skb);
2218 			sk->sk_wmem_queued += nskb->truesize;
2219 			sk_mem_charge(sk, nskb->truesize);
2220 			skb = nskb;
2221 		}
2222 
2223 		TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2224 		TCP_ECN_send_synack(tcp_sk(sk), skb);
2225 	}
2226 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2227 	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2228 }
2229 
2230 /*
2231  * Prepare a SYN-ACK.
2232  */
2233 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2234 				struct request_sock *req)
2235 {
2236 	struct inet_request_sock *ireq = inet_rsk(req);
2237 	struct tcp_sock *tp = tcp_sk(sk);
2238 	struct tcphdr *th;
2239 	int tcp_header_size;
2240 	struct tcp_out_options opts;
2241 	struct sk_buff *skb;
2242 	struct tcp_md5sig_key *md5;
2243 	__u8 *md5_hash_location;
2244 
2245 	skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2246 	if (skb == NULL)
2247 		return NULL;
2248 
2249 	/* Reserve space for headers. */
2250 	skb_reserve(skb, MAX_TCP_HEADER);
2251 
2252 	skb->dst = dst_clone(dst);
2253 
2254 	if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2255 		__u8 rcv_wscale;
2256 		/* Set this up on the first call only */
2257 		req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2258 		/* tcp_full_space because it is guaranteed to be the first packet */
2259 		tcp_select_initial_window(tcp_full_space(sk),
2260 			dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2261 			&req->rcv_wnd,
2262 			&req->window_clamp,
2263 			ireq->wscale_ok,
2264 			&rcv_wscale);
2265 		ireq->rcv_wscale = rcv_wscale;
2266 	}
2267 
2268 	memset(&opts, 0, sizeof(opts));
2269 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2270 	tcp_header_size = tcp_synack_options(sk, req,
2271 					     dst_metric(dst, RTAX_ADVMSS),
2272 					     skb, &opts, &md5) +
2273 			  sizeof(struct tcphdr);
2274 
2275 	skb_push(skb, tcp_header_size);
2276 	skb_reset_transport_header(skb);
2277 
2278 	th = tcp_hdr(skb);
2279 	memset(th, 0, sizeof(struct tcphdr));
2280 	th->syn = 1;
2281 	th->ack = 1;
2282 	TCP_ECN_make_synack(req, th);
2283 	th->source = inet_sk(sk)->sport;
2284 	th->dest = ireq->rmt_port;
2285 	/* Setting of flags are superfluous here for callers (and ECE is
2286 	 * not even correctly set)
2287 	 */
2288 	tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2289 			     TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
2290 	th->seq = htonl(TCP_SKB_CB(skb)->seq);
2291 	th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2292 
2293 	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2294 	th->window = htons(min(req->rcv_wnd, 65535U));
2295 #ifdef CONFIG_SYN_COOKIES
2296 	if (unlikely(req->cookie_ts))
2297 		TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2298 	else
2299 #endif
2300 	tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
2301 	th->doff = (tcp_header_size >> 2);
2302 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
2303 
2304 #ifdef CONFIG_TCP_MD5SIG
2305 	/* Okay, we have all we need - do the md5 hash if needed */
2306 	if (md5) {
2307 		tp->af_specific->calc_md5_hash(md5_hash_location,
2308 					       md5, NULL, req, skb);
2309 	}
2310 #endif
2311 
2312 	return skb;
2313 }
2314 
2315 /*
2316  * Do all connect socket setups that can be done AF independent.
2317  */
2318 static void tcp_connect_init(struct sock *sk)
2319 {
2320 	struct dst_entry *dst = __sk_dst_get(sk);
2321 	struct tcp_sock *tp = tcp_sk(sk);
2322 	__u8 rcv_wscale;
2323 
2324 	/* We'll fix this up when we get a response from the other end.
2325 	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2326 	 */
2327 	tp->tcp_header_len = sizeof(struct tcphdr) +
2328 		(sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2329 
2330 #ifdef CONFIG_TCP_MD5SIG
2331 	if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2332 		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2333 #endif
2334 
2335 	/* If user gave his TCP_MAXSEG, record it to clamp */
2336 	if (tp->rx_opt.user_mss)
2337 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2338 	tp->max_window = 0;
2339 	tcp_mtup_init(sk);
2340 	tcp_sync_mss(sk, dst_mtu(dst));
2341 
2342 	if (!tp->window_clamp)
2343 		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2344 	tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2345 	tcp_initialize_rcv_mss(sk);
2346 
2347 	tcp_select_initial_window(tcp_full_space(sk),
2348 				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2349 				  &tp->rcv_wnd,
2350 				  &tp->window_clamp,
2351 				  sysctl_tcp_window_scaling,
2352 				  &rcv_wscale);
2353 
2354 	tp->rx_opt.rcv_wscale = rcv_wscale;
2355 	tp->rcv_ssthresh = tp->rcv_wnd;
2356 
2357 	sk->sk_err = 0;
2358 	sock_reset_flag(sk, SOCK_DONE);
2359 	tp->snd_wnd = 0;
2360 	tcp_init_wl(tp, tp->write_seq, 0);
2361 	tp->snd_una = tp->write_seq;
2362 	tp->snd_sml = tp->write_seq;
2363 	tp->rcv_nxt = 0;
2364 	tp->rcv_wup = 0;
2365 	tp->copied_seq = 0;
2366 
2367 	inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2368 	inet_csk(sk)->icsk_retransmits = 0;
2369 	tcp_clear_retrans(tp);
2370 }
2371 
2372 /*
2373  * Build a SYN and send it off.
2374  */
2375 int tcp_connect(struct sock *sk)
2376 {
2377 	struct tcp_sock *tp = tcp_sk(sk);
2378 	struct sk_buff *buff;
2379 
2380 	tcp_connect_init(sk);
2381 
2382 	buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2383 	if (unlikely(buff == NULL))
2384 		return -ENOBUFS;
2385 
2386 	/* Reserve space for headers. */
2387 	skb_reserve(buff, MAX_TCP_HEADER);
2388 
2389 	tp->snd_nxt = tp->write_seq;
2390 	tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
2391 	TCP_ECN_send_syn(sk, buff);
2392 
2393 	/* Send it off. */
2394 	TCP_SKB_CB(buff)->when = tcp_time_stamp;
2395 	tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2396 	skb_header_release(buff);
2397 	__tcp_add_write_queue_tail(sk, buff);
2398 	sk->sk_wmem_queued += buff->truesize;
2399 	sk_mem_charge(sk, buff->truesize);
2400 	tp->packets_out += tcp_skb_pcount(buff);
2401 	tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2402 
2403 	/* We change tp->snd_nxt after the tcp_transmit_skb() call
2404 	 * in order to make this packet get counted in tcpOutSegs.
2405 	 */
2406 	tp->snd_nxt = tp->write_seq;
2407 	tp->pushed_seq = tp->write_seq;
2408 	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2409 
2410 	/* Timer for repeating the SYN until an answer. */
2411 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2412 				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2413 	return 0;
2414 }
2415 
2416 /* Send out a delayed ack, the caller does the policy checking
2417  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
2418  * for details.
2419  */
2420 void tcp_send_delayed_ack(struct sock *sk)
2421 {
2422 	struct inet_connection_sock *icsk = inet_csk(sk);
2423 	int ato = icsk->icsk_ack.ato;
2424 	unsigned long timeout;
2425 
2426 	if (ato > TCP_DELACK_MIN) {
2427 		const struct tcp_sock *tp = tcp_sk(sk);
2428 		int max_ato = HZ / 2;
2429 
2430 		if (icsk->icsk_ack.pingpong ||
2431 		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2432 			max_ato = TCP_DELACK_MAX;
2433 
2434 		/* Slow path, intersegment interval is "high". */
2435 
2436 		/* If some rtt estimate is known, use it to bound delayed ack.
2437 		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2438 		 * directly.
2439 		 */
2440 		if (tp->srtt) {
2441 			int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2442 
2443 			if (rtt < max_ato)
2444 				max_ato = rtt;
2445 		}
2446 
2447 		ato = min(ato, max_ato);
2448 	}
2449 
2450 	/* Stay within the limit we were given */
2451 	timeout = jiffies + ato;
2452 
2453 	/* Use new timeout only if there wasn't a older one earlier. */
2454 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2455 		/* If delack timer was blocked or is about to expire,
2456 		 * send ACK now.
2457 		 */
2458 		if (icsk->icsk_ack.blocked ||
2459 		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2460 			tcp_send_ack(sk);
2461 			return;
2462 		}
2463 
2464 		if (!time_before(timeout, icsk->icsk_ack.timeout))
2465 			timeout = icsk->icsk_ack.timeout;
2466 	}
2467 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2468 	icsk->icsk_ack.timeout = timeout;
2469 	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2470 }
2471 
2472 /* This routine sends an ack and also updates the window. */
2473 void tcp_send_ack(struct sock *sk)
2474 {
2475 	struct sk_buff *buff;
2476 
2477 	/* If we have been reset, we may not send again. */
2478 	if (sk->sk_state == TCP_CLOSE)
2479 		return;
2480 
2481 	/* We are not putting this on the write queue, so
2482 	 * tcp_transmit_skb() will set the ownership to this
2483 	 * sock.
2484 	 */
2485 	buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2486 	if (buff == NULL) {
2487 		inet_csk_schedule_ack(sk);
2488 		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2489 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2490 					  TCP_DELACK_MAX, TCP_RTO_MAX);
2491 		return;
2492 	}
2493 
2494 	/* Reserve space for headers and prepare control bits. */
2495 	skb_reserve(buff, MAX_TCP_HEADER);
2496 	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
2497 
2498 	/* Send it off, this clears delayed acks for us. */
2499 	TCP_SKB_CB(buff)->when = tcp_time_stamp;
2500 	tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2501 }
2502 
2503 /* This routine sends a packet with an out of date sequence
2504  * number. It assumes the other end will try to ack it.
2505  *
2506  * Question: what should we make while urgent mode?
2507  * 4.4BSD forces sending single byte of data. We cannot send
2508  * out of window data, because we have SND.NXT==SND.MAX...
2509  *
2510  * Current solution: to send TWO zero-length segments in urgent mode:
2511  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2512  * out-of-date with SND.UNA-1 to probe window.
2513  */
2514 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2515 {
2516 	struct tcp_sock *tp = tcp_sk(sk);
2517 	struct sk_buff *skb;
2518 
2519 	/* We don't queue it, tcp_transmit_skb() sets ownership. */
2520 	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2521 	if (skb == NULL)
2522 		return -1;
2523 
2524 	/* Reserve space for headers and set control bits. */
2525 	skb_reserve(skb, MAX_TCP_HEADER);
2526 	/* Use a previous sequence.  This should cause the other
2527 	 * end to send an ack.  Don't queue or clone SKB, just
2528 	 * send it.
2529 	 */
2530 	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
2531 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2532 	return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2533 }
2534 
2535 int tcp_write_wakeup(struct sock *sk)
2536 {
2537 	struct tcp_sock *tp = tcp_sk(sk);
2538 	struct sk_buff *skb;
2539 
2540 	if (sk->sk_state == TCP_CLOSE)
2541 		return -1;
2542 
2543 	if ((skb = tcp_send_head(sk)) != NULL &&
2544 	    before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2545 		int err;
2546 		unsigned int mss = tcp_current_mss(sk, 0);
2547 		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2548 
2549 		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2550 			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2551 
2552 		/* We are probing the opening of a window
2553 		 * but the window size is != 0
2554 		 * must have been a result SWS avoidance ( sender )
2555 		 */
2556 		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2557 		    skb->len > mss) {
2558 			seg_size = min(seg_size, mss);
2559 			TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2560 			if (tcp_fragment(sk, skb, seg_size, mss))
2561 				return -1;
2562 		} else if (!tcp_skb_pcount(skb))
2563 			tcp_set_skb_tso_segs(sk, skb, mss);
2564 
2565 		TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2566 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
2567 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2568 		if (!err)
2569 			tcp_event_new_data_sent(sk, skb);
2570 		return err;
2571 	} else {
2572 		if (tp->urg_mode &&
2573 		    between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2574 			tcp_xmit_probe_skb(sk, 1);
2575 		return tcp_xmit_probe_skb(sk, 0);
2576 	}
2577 }
2578 
2579 /* A window probe timeout has occurred.  If window is not closed send
2580  * a partial packet else a zero probe.
2581  */
2582 void tcp_send_probe0(struct sock *sk)
2583 {
2584 	struct inet_connection_sock *icsk = inet_csk(sk);
2585 	struct tcp_sock *tp = tcp_sk(sk);
2586 	int err;
2587 
2588 	err = tcp_write_wakeup(sk);
2589 
2590 	if (tp->packets_out || !tcp_send_head(sk)) {
2591 		/* Cancel probe timer, if it is not required. */
2592 		icsk->icsk_probes_out = 0;
2593 		icsk->icsk_backoff = 0;
2594 		return;
2595 	}
2596 
2597 	if (err <= 0) {
2598 		if (icsk->icsk_backoff < sysctl_tcp_retries2)
2599 			icsk->icsk_backoff++;
2600 		icsk->icsk_probes_out++;
2601 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2602 					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2603 					  TCP_RTO_MAX);
2604 	} else {
2605 		/* If packet was not sent due to local congestion,
2606 		 * do not backoff and do not remember icsk_probes_out.
2607 		 * Let local senders to fight for local resources.
2608 		 *
2609 		 * Use accumulated backoff yet.
2610 		 */
2611 		if (!icsk->icsk_probes_out)
2612 			icsk->icsk_probes_out = 1;
2613 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2614 					  min(icsk->icsk_rto << icsk->icsk_backoff,
2615 					      TCP_RESOURCE_PROBE_INTERVAL),
2616 					  TCP_RTO_MAX);
2617 	}
2618 }
2619 
2620 EXPORT_SYMBOL(tcp_select_initial_window);
2621 EXPORT_SYMBOL(tcp_connect);
2622 EXPORT_SYMBOL(tcp_make_synack);
2623 EXPORT_SYMBOL(tcp_simple_retransmit);
2624 EXPORT_SYMBOL(tcp_sync_mss);
2625 EXPORT_SYMBOL(tcp_mtup_init);
2626