xref: /linux/net/ipv4/tcp_output.c (revision 87c2ce3b9305b9b723faeedf6e32ef703ec9b33a)
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  * Version:	$Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:	Ross Biro
11  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *		Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *		Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *		Florian La Roche, <flla@stud.uni-sb.de>
15  *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *		Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *		Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *		Matthew Dillon, <dillon@apollo.west.oic.com>
19  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *		Jorge Cwik, <jorge@laser.satlink.net>
21  */
22 
23 /*
24  * Changes:	Pedro Roque	:	Retransmit queue handled by TCP.
25  *				:	Fragmentation on mtu decrease
26  *				:	Segment collapse on retransmit
27  *				:	AF independence
28  *
29  *		Linus Torvalds	:	send_delayed_ack
30  *		David S. Miller	:	Charge memory using the right skb
31  *					during syn/ack processing.
32  *		David S. Miller :	Output engine completely rewritten.
33  *		Andrea Arcangeli:	SYNACK carry ts_recent in tsecr.
34  *		Cacophonix Gaul :	draft-minshall-nagle-01
35  *		J Hadi Salim	:	ECN support
36  *
37  */
38 
39 #include <net/tcp.h>
40 
41 #include <linux/compiler.h>
42 #include <linux/module.h>
43 #include <linux/smp_lock.h>
44 
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse = 1;
47 
48 /* This limits the percentage of the congestion window which we
49  * will allow a single TSO frame to consume.  Building TSO frames
50  * which are too large can cause TCP streams to be bursty.
51  */
52 int sysctl_tcp_tso_win_divisor = 3;
53 
54 static void update_send_head(struct sock *sk, struct tcp_sock *tp,
55 			     struct sk_buff *skb)
56 {
57 	sk->sk_send_head = skb->next;
58 	if (sk->sk_send_head == (struct sk_buff *)&sk->sk_write_queue)
59 		sk->sk_send_head = NULL;
60 	tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
61 	tcp_packets_out_inc(sk, tp, skb);
62 }
63 
64 /* SND.NXT, if window was not shrunk.
65  * If window has been shrunk, what should we make? It is not clear at all.
66  * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
67  * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
68  * invalid. OK, let's make this for now:
69  */
70 static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp)
71 {
72 	if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
73 		return tp->snd_nxt;
74 	else
75 		return tp->snd_una+tp->snd_wnd;
76 }
77 
78 /* Calculate mss to advertise in SYN segment.
79  * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
80  *
81  * 1. It is independent of path mtu.
82  * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
83  * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
84  *    attached devices, because some buggy hosts are confused by
85  *    large MSS.
86  * 4. We do not make 3, we advertise MSS, calculated from first
87  *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
88  *    This may be overridden via information stored in routing table.
89  * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
90  *    probably even Jumbo".
91  */
92 static __u16 tcp_advertise_mss(struct sock *sk)
93 {
94 	struct tcp_sock *tp = tcp_sk(sk);
95 	struct dst_entry *dst = __sk_dst_get(sk);
96 	int mss = tp->advmss;
97 
98 	if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
99 		mss = dst_metric(dst, RTAX_ADVMSS);
100 		tp->advmss = mss;
101 	}
102 
103 	return (__u16)mss;
104 }
105 
106 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
107  * This is the first part of cwnd validation mechanism. */
108 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
109 {
110 	struct tcp_sock *tp = tcp_sk(sk);
111 	s32 delta = tcp_time_stamp - tp->lsndtime;
112 	u32 restart_cwnd = tcp_init_cwnd(tp, dst);
113 	u32 cwnd = tp->snd_cwnd;
114 
115 	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
116 
117 	tp->snd_ssthresh = tcp_current_ssthresh(sk);
118 	restart_cwnd = min(restart_cwnd, cwnd);
119 
120 	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
121 		cwnd >>= 1;
122 	tp->snd_cwnd = max(cwnd, restart_cwnd);
123 	tp->snd_cwnd_stamp = tcp_time_stamp;
124 	tp->snd_cwnd_used = 0;
125 }
126 
127 static void tcp_event_data_sent(struct tcp_sock *tp,
128 				struct sk_buff *skb, struct sock *sk)
129 {
130 	struct inet_connection_sock *icsk = inet_csk(sk);
131 	const u32 now = tcp_time_stamp;
132 
133 	if (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)
134 		tcp_cwnd_restart(sk, __sk_dst_get(sk));
135 
136 	tp->lsndtime = now;
137 
138 	/* If it is a reply for ato after last received
139 	 * packet, enter pingpong mode.
140 	 */
141 	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
142 		icsk->icsk_ack.pingpong = 1;
143 }
144 
145 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
146 {
147 	tcp_dec_quickack_mode(sk, pkts);
148 	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
149 }
150 
151 /* Determine a window scaling and initial window to offer.
152  * Based on the assumption that the given amount of space
153  * will be offered. Store the results in the tp structure.
154  * NOTE: for smooth operation initial space offering should
155  * be a multiple of mss if possible. We assume here that mss >= 1.
156  * This MUST be enforced by all callers.
157  */
158 void tcp_select_initial_window(int __space, __u32 mss,
159 			       __u32 *rcv_wnd, __u32 *window_clamp,
160 			       int wscale_ok, __u8 *rcv_wscale)
161 {
162 	unsigned int space = (__space < 0 ? 0 : __space);
163 
164 	/* If no clamp set the clamp to the max possible scaled window */
165 	if (*window_clamp == 0)
166 		(*window_clamp) = (65535 << 14);
167 	space = min(*window_clamp, space);
168 
169 	/* Quantize space offering to a multiple of mss if possible. */
170 	if (space > mss)
171 		space = (space / mss) * mss;
172 
173 	/* NOTE: offering an initial window larger than 32767
174 	 * will break some buggy TCP stacks. We try to be nice.
175 	 * If we are not window scaling, then this truncates
176 	 * our initial window offering to 32k. There should also
177 	 * be a sysctl option to stop being nice.
178 	 */
179 	(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
180 	(*rcv_wscale) = 0;
181 	if (wscale_ok) {
182 		/* Set window scaling on max possible window
183 		 * See RFC1323 for an explanation of the limit to 14
184 		 */
185 		space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
186 		while (space > 65535 && (*rcv_wscale) < 14) {
187 			space >>= 1;
188 			(*rcv_wscale)++;
189 		}
190 	}
191 
192 	/* Set initial window to value enough for senders,
193 	 * following RFC2414. Senders, not following this RFC,
194 	 * will be satisfied with 2.
195 	 */
196 	if (mss > (1<<*rcv_wscale)) {
197 		int init_cwnd = 4;
198 		if (mss > 1460*3)
199 			init_cwnd = 2;
200 		else if (mss > 1460)
201 			init_cwnd = 3;
202 		if (*rcv_wnd > init_cwnd*mss)
203 			*rcv_wnd = init_cwnd*mss;
204 	}
205 
206 	/* Set the clamp no higher than max representable value */
207 	(*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
208 }
209 
210 /* Chose a new window to advertise, update state in tcp_sock for the
211  * socket, and return result with RFC1323 scaling applied.  The return
212  * value can be stuffed directly into th->window for an outgoing
213  * frame.
214  */
215 static u16 tcp_select_window(struct sock *sk)
216 {
217 	struct tcp_sock *tp = tcp_sk(sk);
218 	u32 cur_win = tcp_receive_window(tp);
219 	u32 new_win = __tcp_select_window(sk);
220 
221 	/* Never shrink the offered window */
222 	if(new_win < cur_win) {
223 		/* Danger Will Robinson!
224 		 * Don't update rcv_wup/rcv_wnd here or else
225 		 * we will not be able to advertise a zero
226 		 * window in time.  --DaveM
227 		 *
228 		 * Relax Will Robinson.
229 		 */
230 		new_win = cur_win;
231 	}
232 	tp->rcv_wnd = new_win;
233 	tp->rcv_wup = tp->rcv_nxt;
234 
235 	/* Make sure we do not exceed the maximum possible
236 	 * scaled window.
237 	 */
238 	if (!tp->rx_opt.rcv_wscale)
239 		new_win = min(new_win, MAX_TCP_WINDOW);
240 	else
241 		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
242 
243 	/* RFC1323 scaling applied */
244 	new_win >>= tp->rx_opt.rcv_wscale;
245 
246 	/* If we advertise zero window, disable fast path. */
247 	if (new_win == 0)
248 		tp->pred_flags = 0;
249 
250 	return new_win;
251 }
252 
253 static void tcp_build_and_update_options(__u32 *ptr, struct tcp_sock *tp,
254 					 __u32 tstamp)
255 {
256 	if (tp->rx_opt.tstamp_ok) {
257 		*ptr++ = __constant_htonl((TCPOPT_NOP << 24) |
258 					  (TCPOPT_NOP << 16) |
259 					  (TCPOPT_TIMESTAMP << 8) |
260 					  TCPOLEN_TIMESTAMP);
261 		*ptr++ = htonl(tstamp);
262 		*ptr++ = htonl(tp->rx_opt.ts_recent);
263 	}
264 	if (tp->rx_opt.eff_sacks) {
265 		struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
266 		int this_sack;
267 
268 		*ptr++ = htonl((TCPOPT_NOP  << 24) |
269 			       (TCPOPT_NOP  << 16) |
270 			       (TCPOPT_SACK <<  8) |
271 			       (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
272 						     TCPOLEN_SACK_PERBLOCK)));
273 		for(this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
274 			*ptr++ = htonl(sp[this_sack].start_seq);
275 			*ptr++ = htonl(sp[this_sack].end_seq);
276 		}
277 		if (tp->rx_opt.dsack) {
278 			tp->rx_opt.dsack = 0;
279 			tp->rx_opt.eff_sacks--;
280 		}
281 	}
282 }
283 
284 /* Construct a tcp options header for a SYN or SYN_ACK packet.
285  * If this is every changed make sure to change the definition of
286  * MAX_SYN_SIZE to match the new maximum number of options that you
287  * can generate.
288  */
289 static void tcp_syn_build_options(__u32 *ptr, int mss, int ts, int sack,
290 				  int offer_wscale, int wscale, __u32 tstamp,
291 				  __u32 ts_recent)
292 {
293 	/* We always get an MSS option.
294 	 * The option bytes which will be seen in normal data
295 	 * packets should timestamps be used, must be in the MSS
296 	 * advertised.  But we subtract them from tp->mss_cache so
297 	 * that calculations in tcp_sendmsg are simpler etc.
298 	 * So account for this fact here if necessary.  If we
299 	 * don't do this correctly, as a receiver we won't
300 	 * recognize data packets as being full sized when we
301 	 * should, and thus we won't abide by the delayed ACK
302 	 * rules correctly.
303 	 * SACKs don't matter, we never delay an ACK when we
304 	 * have any of those going out.
305 	 */
306 	*ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
307 	if (ts) {
308 		if(sack)
309 			*ptr++ = __constant_htonl((TCPOPT_SACK_PERM << 24) | (TCPOLEN_SACK_PERM << 16) |
310 						  (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
311 		else
312 			*ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
313 						  (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
314 		*ptr++ = htonl(tstamp);		/* TSVAL */
315 		*ptr++ = htonl(ts_recent);	/* TSECR */
316 	} else if(sack)
317 		*ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
318 					  (TCPOPT_SACK_PERM << 8) | TCPOLEN_SACK_PERM);
319 	if (offer_wscale)
320 		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_WINDOW << 16) | (TCPOLEN_WINDOW << 8) | (wscale));
321 }
322 
323 /* This routine actually transmits TCP packets queued in by
324  * tcp_do_sendmsg().  This is used by both the initial
325  * transmission and possible later retransmissions.
326  * All SKB's seen here are completely headerless.  It is our
327  * job to build the TCP header, and pass the packet down to
328  * IP so it can do the same plus pass the packet off to the
329  * device.
330  *
331  * We are working here with either a clone of the original
332  * SKB, or a fresh unique copy made by the retransmit engine.
333  */
334 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask)
335 {
336 	const struct inet_connection_sock *icsk = inet_csk(sk);
337 	struct inet_sock *inet;
338 	struct tcp_sock *tp;
339 	struct tcp_skb_cb *tcb;
340 	int tcp_header_size;
341 	struct tcphdr *th;
342 	int sysctl_flags;
343 	int err;
344 
345 	BUG_ON(!skb || !tcp_skb_pcount(skb));
346 
347 	/* If congestion control is doing timestamping, we must
348 	 * take such a timestamp before we potentially clone/copy.
349 	 */
350 	if (icsk->icsk_ca_ops->rtt_sample)
351 		__net_timestamp(skb);
352 
353 	if (likely(clone_it)) {
354 		if (unlikely(skb_cloned(skb)))
355 			skb = pskb_copy(skb, gfp_mask);
356 		else
357 			skb = skb_clone(skb, gfp_mask);
358 		if (unlikely(!skb))
359 			return -ENOBUFS;
360 	}
361 
362 	inet = inet_sk(sk);
363 	tp = tcp_sk(sk);
364 	tcb = TCP_SKB_CB(skb);
365 	tcp_header_size = tp->tcp_header_len;
366 
367 #define SYSCTL_FLAG_TSTAMPS	0x1
368 #define SYSCTL_FLAG_WSCALE	0x2
369 #define SYSCTL_FLAG_SACK	0x4
370 
371 	sysctl_flags = 0;
372 	if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
373 		tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
374 		if(sysctl_tcp_timestamps) {
375 			tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
376 			sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
377 		}
378 		if (sysctl_tcp_window_scaling) {
379 			tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
380 			sysctl_flags |= SYSCTL_FLAG_WSCALE;
381 		}
382 		if (sysctl_tcp_sack) {
383 			sysctl_flags |= SYSCTL_FLAG_SACK;
384 			if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
385 				tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
386 		}
387 	} else if (unlikely(tp->rx_opt.eff_sacks)) {
388 		/* A SACK is 2 pad bytes, a 2 byte header, plus
389 		 * 2 32-bit sequence numbers for each SACK block.
390 		 */
391 		tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
392 				    (tp->rx_opt.eff_sacks *
393 				     TCPOLEN_SACK_PERBLOCK));
394 	}
395 
396 	if (tcp_packets_in_flight(tp) == 0)
397 		tcp_ca_event(sk, CA_EVENT_TX_START);
398 
399 	th = (struct tcphdr *) skb_push(skb, tcp_header_size);
400 	skb->h.th = th;
401 	skb_set_owner_w(skb, sk);
402 
403 	/* Build TCP header and checksum it. */
404 	th->source		= inet->sport;
405 	th->dest		= inet->dport;
406 	th->seq			= htonl(tcb->seq);
407 	th->ack_seq		= htonl(tp->rcv_nxt);
408 	*(((__u16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
409 					tcb->flags);
410 
411 	if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
412 		/* RFC1323: The window in SYN & SYN/ACK segments
413 		 * is never scaled.
414 		 */
415 		th->window	= htons(tp->rcv_wnd);
416 	} else {
417 		th->window	= htons(tcp_select_window(sk));
418 	}
419 	th->check		= 0;
420 	th->urg_ptr		= 0;
421 
422 	if (unlikely(tp->urg_mode &&
423 		     between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
424 		th->urg_ptr		= htons(tp->snd_up-tcb->seq);
425 		th->urg			= 1;
426 	}
427 
428 	if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
429 		tcp_syn_build_options((__u32 *)(th + 1),
430 				      tcp_advertise_mss(sk),
431 				      (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
432 				      (sysctl_flags & SYSCTL_FLAG_SACK),
433 				      (sysctl_flags & SYSCTL_FLAG_WSCALE),
434 				      tp->rx_opt.rcv_wscale,
435 				      tcb->when,
436 				      tp->rx_opt.ts_recent);
437 	} else {
438 		tcp_build_and_update_options((__u32 *)(th + 1),
439 					     tp, tcb->when);
440 		TCP_ECN_send(sk, tp, skb, tcp_header_size);
441 	}
442 
443 	icsk->icsk_af_ops->send_check(sk, skb->len, skb);
444 
445 	if (likely(tcb->flags & TCPCB_FLAG_ACK))
446 		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
447 
448 	if (skb->len != tcp_header_size)
449 		tcp_event_data_sent(tp, skb, sk);
450 
451 	TCP_INC_STATS(TCP_MIB_OUTSEGS);
452 
453 	err = icsk->icsk_af_ops->queue_xmit(skb, 0);
454 	if (unlikely(err <= 0))
455 		return err;
456 
457 	tcp_enter_cwr(sk);
458 
459 	/* NET_XMIT_CN is special. It does not guarantee,
460 	 * that this packet is lost. It tells that device
461 	 * is about to start to drop packets or already
462 	 * drops some packets of the same priority and
463 	 * invokes us to send less aggressively.
464 	 */
465 	return err == NET_XMIT_CN ? 0 : err;
466 
467 #undef SYSCTL_FLAG_TSTAMPS
468 #undef SYSCTL_FLAG_WSCALE
469 #undef SYSCTL_FLAG_SACK
470 }
471 
472 
473 /* This routine just queue's the buffer
474  *
475  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
476  * otherwise socket can stall.
477  */
478 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
479 {
480 	struct tcp_sock *tp = tcp_sk(sk);
481 
482 	/* Advance write_seq and place onto the write_queue. */
483 	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
484 	skb_header_release(skb);
485 	__skb_queue_tail(&sk->sk_write_queue, skb);
486 	sk_charge_skb(sk, skb);
487 
488 	/* Queue it, remembering where we must start sending. */
489 	if (sk->sk_send_head == NULL)
490 		sk->sk_send_head = skb;
491 }
492 
493 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
494 {
495 	if (skb->len <= mss_now ||
496 	    !(sk->sk_route_caps & NETIF_F_TSO)) {
497 		/* Avoid the costly divide in the normal
498 		 * non-TSO case.
499 		 */
500 		skb_shinfo(skb)->tso_segs = 1;
501 		skb_shinfo(skb)->tso_size = 0;
502 	} else {
503 		unsigned int factor;
504 
505 		factor = skb->len + (mss_now - 1);
506 		factor /= mss_now;
507 		skb_shinfo(skb)->tso_segs = factor;
508 		skb_shinfo(skb)->tso_size = mss_now;
509 	}
510 }
511 
512 /* Function to create two new TCP segments.  Shrinks the given segment
513  * to the specified size and appends a new segment with the rest of the
514  * packet to the list.  This won't be called frequently, I hope.
515  * Remember, these are still headerless SKBs at this point.
516  */
517 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
518 {
519 	struct tcp_sock *tp = tcp_sk(sk);
520 	struct sk_buff *buff;
521 	int nsize, old_factor;
522 	u16 flags;
523 
524 	BUG_ON(len > skb->len);
525 
526  	clear_all_retrans_hints(tp);
527 	nsize = skb_headlen(skb) - len;
528 	if (nsize < 0)
529 		nsize = 0;
530 
531 	if (skb_cloned(skb) &&
532 	    skb_is_nonlinear(skb) &&
533 	    pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
534 		return -ENOMEM;
535 
536 	/* Get a new skb... force flag on. */
537 	buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
538 	if (buff == NULL)
539 		return -ENOMEM; /* We'll just try again later. */
540 	sk_charge_skb(sk, buff);
541 
542 	/* Correct the sequence numbers. */
543 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
544 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
545 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
546 
547 	/* PSH and FIN should only be set in the second packet. */
548 	flags = TCP_SKB_CB(skb)->flags;
549 	TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
550 	TCP_SKB_CB(buff)->flags = flags;
551 	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
552 	TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
553 
554 	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) {
555 		/* Copy and checksum data tail into the new buffer. */
556 		buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
557 						       nsize, 0);
558 
559 		skb_trim(skb, len);
560 
561 		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
562 	} else {
563 		skb->ip_summed = CHECKSUM_HW;
564 		skb_split(skb, buff, len);
565 	}
566 
567 	buff->ip_summed = skb->ip_summed;
568 
569 	/* Looks stupid, but our code really uses when of
570 	 * skbs, which it never sent before. --ANK
571 	 */
572 	TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
573 	buff->tstamp = skb->tstamp;
574 
575 	old_factor = tcp_skb_pcount(skb);
576 
577 	/* Fix up tso_factor for both original and new SKB.  */
578 	tcp_set_skb_tso_segs(sk, skb, mss_now);
579 	tcp_set_skb_tso_segs(sk, buff, mss_now);
580 
581 	/* If this packet has been sent out already, we must
582 	 * adjust the various packet counters.
583 	 */
584 	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
585 		int diff = old_factor - tcp_skb_pcount(skb) -
586 			tcp_skb_pcount(buff);
587 
588 		tp->packets_out -= diff;
589 
590 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
591 			tp->sacked_out -= diff;
592 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
593 			tp->retrans_out -= diff;
594 
595 		if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
596 			tp->lost_out -= diff;
597 			tp->left_out -= diff;
598 		}
599 
600 		if (diff > 0) {
601 			/* Adjust Reno SACK estimate. */
602 			if (!tp->rx_opt.sack_ok) {
603 				tp->sacked_out -= diff;
604 				if ((int)tp->sacked_out < 0)
605 					tp->sacked_out = 0;
606 				tcp_sync_left_out(tp);
607 			}
608 
609 			tp->fackets_out -= diff;
610 			if ((int)tp->fackets_out < 0)
611 				tp->fackets_out = 0;
612 		}
613 	}
614 
615 	/* Link BUFF into the send queue. */
616 	skb_header_release(buff);
617 	__skb_append(skb, buff, &sk->sk_write_queue);
618 
619 	return 0;
620 }
621 
622 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
623  * eventually). The difference is that pulled data not copied, but
624  * immediately discarded.
625  */
626 static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len)
627 {
628 	int i, k, eat;
629 
630 	eat = len;
631 	k = 0;
632 	for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
633 		if (skb_shinfo(skb)->frags[i].size <= eat) {
634 			put_page(skb_shinfo(skb)->frags[i].page);
635 			eat -= skb_shinfo(skb)->frags[i].size;
636 		} else {
637 			skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
638 			if (eat) {
639 				skb_shinfo(skb)->frags[k].page_offset += eat;
640 				skb_shinfo(skb)->frags[k].size -= eat;
641 				eat = 0;
642 			}
643 			k++;
644 		}
645 	}
646 	skb_shinfo(skb)->nr_frags = k;
647 
648 	skb->tail = skb->data;
649 	skb->data_len -= len;
650 	skb->len = skb->data_len;
651 	return skb->tail;
652 }
653 
654 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
655 {
656 	if (skb_cloned(skb) &&
657 	    pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
658 		return -ENOMEM;
659 
660 	if (len <= skb_headlen(skb)) {
661 		__skb_pull(skb, len);
662 	} else {
663 		if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL)
664 			return -ENOMEM;
665 	}
666 
667 	TCP_SKB_CB(skb)->seq += len;
668 	skb->ip_summed = CHECKSUM_HW;
669 
670 	skb->truesize	     -= len;
671 	sk->sk_wmem_queued   -= len;
672 	sk->sk_forward_alloc += len;
673 	sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
674 
675 	/* Any change of skb->len requires recalculation of tso
676 	 * factor and mss.
677 	 */
678 	if (tcp_skb_pcount(skb) > 1)
679 		tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
680 
681 	return 0;
682 }
683 
684 /* This function synchronize snd mss to current pmtu/exthdr set.
685 
686    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
687    for TCP options, but includes only bare TCP header.
688 
689    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
690    It is minimum of user_mss and mss received with SYN.
691    It also does not include TCP options.
692 
693    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
694 
695    tp->mss_cache is current effective sending mss, including
696    all tcp options except for SACKs. It is evaluated,
697    taking into account current pmtu, but never exceeds
698    tp->rx_opt.mss_clamp.
699 
700    NOTE1. rfc1122 clearly states that advertised MSS
701    DOES NOT include either tcp or ip options.
702 
703    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
704    are READ ONLY outside this function.		--ANK (980731)
705  */
706 
707 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
708 {
709 	struct tcp_sock *tp = tcp_sk(sk);
710 	struct inet_connection_sock *icsk = inet_csk(sk);
711 	/* Calculate base mss without TCP options:
712 	   It is MMS_S - sizeof(tcphdr) of rfc1122
713 	 */
714 	int mss_now = (pmtu - icsk->icsk_af_ops->net_header_len -
715 		       sizeof(struct tcphdr));
716 
717 	/* Clamp it (mss_clamp does not include tcp options) */
718 	if (mss_now > tp->rx_opt.mss_clamp)
719 		mss_now = tp->rx_opt.mss_clamp;
720 
721 	/* Now subtract optional transport overhead */
722 	mss_now -= icsk->icsk_ext_hdr_len;
723 
724 	/* Then reserve room for full set of TCP options and 8 bytes of data */
725 	if (mss_now < 48)
726 		mss_now = 48;
727 
728 	/* Now subtract TCP options size, not including SACKs */
729 	mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
730 
731 	/* Bound mss with half of window */
732 	if (tp->max_window && mss_now > (tp->max_window>>1))
733 		mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
734 
735 	/* And store cached results */
736 	icsk->icsk_pmtu_cookie = pmtu;
737 	tp->mss_cache = mss_now;
738 
739 	return mss_now;
740 }
741 
742 /* Compute the current effective MSS, taking SACKs and IP options,
743  * and even PMTU discovery events into account.
744  *
745  * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
746  * cannot be large. However, taking into account rare use of URG, this
747  * is not a big flaw.
748  */
749 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
750 {
751 	struct tcp_sock *tp = tcp_sk(sk);
752 	struct dst_entry *dst = __sk_dst_get(sk);
753 	u32 mss_now;
754 	u16 xmit_size_goal;
755 	int doing_tso = 0;
756 
757 	mss_now = tp->mss_cache;
758 
759 	if (large_allowed &&
760 	    (sk->sk_route_caps & NETIF_F_TSO) &&
761 	    !tp->urg_mode)
762 		doing_tso = 1;
763 
764 	if (dst) {
765 		u32 mtu = dst_mtu(dst);
766 		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
767 			mss_now = tcp_sync_mss(sk, mtu);
768 	}
769 
770 	if (tp->rx_opt.eff_sacks)
771 		mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
772 			    (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
773 
774 	xmit_size_goal = mss_now;
775 
776 	if (doing_tso) {
777 		xmit_size_goal = (65535 -
778 				  inet_csk(sk)->icsk_af_ops->net_header_len -
779 				  inet_csk(sk)->icsk_ext_hdr_len -
780 				  tp->tcp_header_len);
781 
782 		if (tp->max_window &&
783 		    (xmit_size_goal > (tp->max_window >> 1)))
784 			xmit_size_goal = max((tp->max_window >> 1),
785 					     68U - tp->tcp_header_len);
786 
787 		xmit_size_goal -= (xmit_size_goal % mss_now);
788 	}
789 	tp->xmit_size_goal = xmit_size_goal;
790 
791 	return mss_now;
792 }
793 
794 /* Congestion window validation. (RFC2861) */
795 
796 static void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
797 {
798 	__u32 packets_out = tp->packets_out;
799 
800 	if (packets_out >= tp->snd_cwnd) {
801 		/* Network is feed fully. */
802 		tp->snd_cwnd_used = 0;
803 		tp->snd_cwnd_stamp = tcp_time_stamp;
804 	} else {
805 		/* Network starves. */
806 		if (tp->packets_out > tp->snd_cwnd_used)
807 			tp->snd_cwnd_used = tp->packets_out;
808 
809 		if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
810 			tcp_cwnd_application_limited(sk);
811 	}
812 }
813 
814 static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
815 {
816 	u32 window, cwnd_len;
817 
818 	window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
819 	cwnd_len = mss_now * cwnd;
820 	return min(window, cwnd_len);
821 }
822 
823 /* Can at least one segment of SKB be sent right now, according to the
824  * congestion window rules?  If so, return how many segments are allowed.
825  */
826 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
827 {
828 	u32 in_flight, cwnd;
829 
830 	/* Don't be strict about the congestion window for the final FIN.  */
831 	if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
832 		return 1;
833 
834 	in_flight = tcp_packets_in_flight(tp);
835 	cwnd = tp->snd_cwnd;
836 	if (in_flight < cwnd)
837 		return (cwnd - in_flight);
838 
839 	return 0;
840 }
841 
842 /* This must be invoked the first time we consider transmitting
843  * SKB onto the wire.
844  */
845 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
846 {
847 	int tso_segs = tcp_skb_pcount(skb);
848 
849 	if (!tso_segs ||
850 	    (tso_segs > 1 &&
851 	     skb_shinfo(skb)->tso_size != mss_now)) {
852 		tcp_set_skb_tso_segs(sk, skb, mss_now);
853 		tso_segs = tcp_skb_pcount(skb);
854 	}
855 	return tso_segs;
856 }
857 
858 static inline int tcp_minshall_check(const struct tcp_sock *tp)
859 {
860 	return after(tp->snd_sml,tp->snd_una) &&
861 		!after(tp->snd_sml, tp->snd_nxt);
862 }
863 
864 /* Return 0, if packet can be sent now without violation Nagle's rules:
865  * 1. It is full sized.
866  * 2. Or it contains FIN. (already checked by caller)
867  * 3. Or TCP_NODELAY was set.
868  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
869  *    With Minshall's modification: all sent small packets are ACKed.
870  */
871 
872 static inline int tcp_nagle_check(const struct tcp_sock *tp,
873 				  const struct sk_buff *skb,
874 				  unsigned mss_now, int nonagle)
875 {
876 	return (skb->len < mss_now &&
877 		((nonagle&TCP_NAGLE_CORK) ||
878 		 (!nonagle &&
879 		  tp->packets_out &&
880 		  tcp_minshall_check(tp))));
881 }
882 
883 /* Return non-zero if the Nagle test allows this packet to be
884  * sent now.
885  */
886 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
887 				 unsigned int cur_mss, int nonagle)
888 {
889 	/* Nagle rule does not apply to frames, which sit in the middle of the
890 	 * write_queue (they have no chances to get new data).
891 	 *
892 	 * This is implemented in the callers, where they modify the 'nonagle'
893 	 * argument based upon the location of SKB in the send queue.
894 	 */
895 	if (nonagle & TCP_NAGLE_PUSH)
896 		return 1;
897 
898 	/* Don't use the nagle rule for urgent data (or for the final FIN).  */
899 	if (tp->urg_mode ||
900 	    (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
901 		return 1;
902 
903 	if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
904 		return 1;
905 
906 	return 0;
907 }
908 
909 /* Does at least the first segment of SKB fit into the send window? */
910 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
911 {
912 	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
913 
914 	if (skb->len > cur_mss)
915 		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
916 
917 	return !after(end_seq, tp->snd_una + tp->snd_wnd);
918 }
919 
920 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
921  * should be put on the wire right now.  If so, it returns the number of
922  * packets allowed by the congestion window.
923  */
924 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
925 				 unsigned int cur_mss, int nonagle)
926 {
927 	struct tcp_sock *tp = tcp_sk(sk);
928 	unsigned int cwnd_quota;
929 
930 	tcp_init_tso_segs(sk, skb, cur_mss);
931 
932 	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
933 		return 0;
934 
935 	cwnd_quota = tcp_cwnd_test(tp, skb);
936 	if (cwnd_quota &&
937 	    !tcp_snd_wnd_test(tp, skb, cur_mss))
938 		cwnd_quota = 0;
939 
940 	return cwnd_quota;
941 }
942 
943 static inline int tcp_skb_is_last(const struct sock *sk,
944 				  const struct sk_buff *skb)
945 {
946 	return skb->next == (struct sk_buff *)&sk->sk_write_queue;
947 }
948 
949 int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
950 {
951 	struct sk_buff *skb = sk->sk_send_head;
952 
953 	return (skb &&
954 		tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
955 			     (tcp_skb_is_last(sk, skb) ?
956 			      TCP_NAGLE_PUSH :
957 			      tp->nonagle)));
958 }
959 
960 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
961  * which is put after SKB on the list.  It is very much like
962  * tcp_fragment() except that it may make several kinds of assumptions
963  * in order to speed up the splitting operation.  In particular, we
964  * know that all the data is in scatter-gather pages, and that the
965  * packet has never been sent out before (and thus is not cloned).
966  */
967 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
968 {
969 	struct sk_buff *buff;
970 	int nlen = skb->len - len;
971 	u16 flags;
972 
973 	/* All of a TSO frame must be composed of paged data.  */
974 	if (skb->len != skb->data_len)
975 		return tcp_fragment(sk, skb, len, mss_now);
976 
977 	buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
978 	if (unlikely(buff == NULL))
979 		return -ENOMEM;
980 
981 	buff->truesize = nlen;
982 	skb->truesize -= nlen;
983 
984 	/* Correct the sequence numbers. */
985 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
986 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
987 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
988 
989 	/* PSH and FIN should only be set in the second packet. */
990 	flags = TCP_SKB_CB(skb)->flags;
991 	TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
992 	TCP_SKB_CB(buff)->flags = flags;
993 
994 	/* This packet was never sent out yet, so no SACK bits. */
995 	TCP_SKB_CB(buff)->sacked = 0;
996 
997 	buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
998 	skb_split(skb, buff, len);
999 
1000 	/* Fix up tso_factor for both original and new SKB.  */
1001 	tcp_set_skb_tso_segs(sk, skb, mss_now);
1002 	tcp_set_skb_tso_segs(sk, buff, mss_now);
1003 
1004 	/* Link BUFF into the send queue. */
1005 	skb_header_release(buff);
1006 	__skb_append(skb, buff, &sk->sk_write_queue);
1007 
1008 	return 0;
1009 }
1010 
1011 /* Try to defer sending, if possible, in order to minimize the amount
1012  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1013  *
1014  * This algorithm is from John Heffner.
1015  */
1016 static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
1017 {
1018 	const struct inet_connection_sock *icsk = inet_csk(sk);
1019 	u32 send_win, cong_win, limit, in_flight;
1020 
1021 	if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1022 		return 0;
1023 
1024 	if (icsk->icsk_ca_state != TCP_CA_Open)
1025 		return 0;
1026 
1027 	in_flight = tcp_packets_in_flight(tp);
1028 
1029 	BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1030 	       (tp->snd_cwnd <= in_flight));
1031 
1032 	send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
1033 
1034 	/* From in_flight test above, we know that cwnd > in_flight.  */
1035 	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1036 
1037 	limit = min(send_win, cong_win);
1038 
1039 	if (sysctl_tcp_tso_win_divisor) {
1040 		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1041 
1042 		/* If at least some fraction of a window is available,
1043 		 * just use it.
1044 		 */
1045 		chunk /= sysctl_tcp_tso_win_divisor;
1046 		if (limit >= chunk)
1047 			return 0;
1048 	} else {
1049 		/* Different approach, try not to defer past a single
1050 		 * ACK.  Receiver should ACK every other full sized
1051 		 * frame, so if we have space for more than 3 frames
1052 		 * then send now.
1053 		 */
1054 		if (limit > tcp_max_burst(tp) * tp->mss_cache)
1055 			return 0;
1056 	}
1057 
1058 	/* Ok, it looks like it is advisable to defer.  */
1059 	return 1;
1060 }
1061 
1062 /* This routine writes packets to the network.  It advances the
1063  * send_head.  This happens as incoming acks open up the remote
1064  * window for us.
1065  *
1066  * Returns 1, if no segments are in flight and we have queued segments, but
1067  * cannot send anything now because of SWS or another problem.
1068  */
1069 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1070 {
1071 	struct tcp_sock *tp = tcp_sk(sk);
1072 	struct sk_buff *skb;
1073 	unsigned int tso_segs, sent_pkts;
1074 	int cwnd_quota;
1075 
1076 	/* If we are closed, the bytes will have to remain here.
1077 	 * In time closedown will finish, we empty the write queue and all
1078 	 * will be happy.
1079 	 */
1080 	if (unlikely(sk->sk_state == TCP_CLOSE))
1081 		return 0;
1082 
1083 	sent_pkts = 0;
1084 	while ((skb = sk->sk_send_head)) {
1085 		unsigned int limit;
1086 
1087 		tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1088 		BUG_ON(!tso_segs);
1089 
1090 		cwnd_quota = tcp_cwnd_test(tp, skb);
1091 		if (!cwnd_quota)
1092 			break;
1093 
1094 		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1095 			break;
1096 
1097 		if (tso_segs == 1) {
1098 			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1099 						     (tcp_skb_is_last(sk, skb) ?
1100 						      nonagle : TCP_NAGLE_PUSH))))
1101 				break;
1102 		} else {
1103 			if (tcp_tso_should_defer(sk, tp, skb))
1104 				break;
1105 		}
1106 
1107 		limit = mss_now;
1108 		if (tso_segs > 1) {
1109 			limit = tcp_window_allows(tp, skb,
1110 						  mss_now, cwnd_quota);
1111 
1112 			if (skb->len < limit) {
1113 				unsigned int trim = skb->len % mss_now;
1114 
1115 				if (trim)
1116 					limit = skb->len - trim;
1117 			}
1118 		}
1119 
1120 		if (skb->len > limit &&
1121 		    unlikely(tso_fragment(sk, skb, limit, mss_now)))
1122 			break;
1123 
1124 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
1125 
1126 		if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1127 			break;
1128 
1129 		/* Advance the send_head.  This one is sent out.
1130 		 * This call will increment packets_out.
1131 		 */
1132 		update_send_head(sk, tp, skb);
1133 
1134 		tcp_minshall_update(tp, mss_now, skb);
1135 		sent_pkts++;
1136 	}
1137 
1138 	if (likely(sent_pkts)) {
1139 		tcp_cwnd_validate(sk, tp);
1140 		return 0;
1141 	}
1142 	return !tp->packets_out && sk->sk_send_head;
1143 }
1144 
1145 /* Push out any pending frames which were held back due to
1146  * TCP_CORK or attempt at coalescing tiny packets.
1147  * The socket must be locked by the caller.
1148  */
1149 void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
1150 			       unsigned int cur_mss, int nonagle)
1151 {
1152 	struct sk_buff *skb = sk->sk_send_head;
1153 
1154 	if (skb) {
1155 		if (tcp_write_xmit(sk, cur_mss, nonagle))
1156 			tcp_check_probe_timer(sk, tp);
1157 	}
1158 }
1159 
1160 /* Send _single_ skb sitting at the send head. This function requires
1161  * true push pending frames to setup probe timer etc.
1162  */
1163 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1164 {
1165 	struct tcp_sock *tp = tcp_sk(sk);
1166 	struct sk_buff *skb = sk->sk_send_head;
1167 	unsigned int tso_segs, cwnd_quota;
1168 
1169 	BUG_ON(!skb || skb->len < mss_now);
1170 
1171 	tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1172 	cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1173 
1174 	if (likely(cwnd_quota)) {
1175 		unsigned int limit;
1176 
1177 		BUG_ON(!tso_segs);
1178 
1179 		limit = mss_now;
1180 		if (tso_segs > 1) {
1181 			limit = tcp_window_allows(tp, skb,
1182 						  mss_now, cwnd_quota);
1183 
1184 			if (skb->len < limit) {
1185 				unsigned int trim = skb->len % mss_now;
1186 
1187 				if (trim)
1188 					limit = skb->len - trim;
1189 			}
1190 		}
1191 
1192 		if (skb->len > limit &&
1193 		    unlikely(tso_fragment(sk, skb, limit, mss_now)))
1194 			return;
1195 
1196 		/* Send it out now. */
1197 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
1198 
1199 		if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1200 			update_send_head(sk, tp, skb);
1201 			tcp_cwnd_validate(sk, tp);
1202 			return;
1203 		}
1204 	}
1205 }
1206 
1207 /* This function returns the amount that we can raise the
1208  * usable window based on the following constraints
1209  *
1210  * 1. The window can never be shrunk once it is offered (RFC 793)
1211  * 2. We limit memory per socket
1212  *
1213  * RFC 1122:
1214  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1215  *  RECV.NEXT + RCV.WIN fixed until:
1216  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1217  *
1218  * i.e. don't raise the right edge of the window until you can raise
1219  * it at least MSS bytes.
1220  *
1221  * Unfortunately, the recommended algorithm breaks header prediction,
1222  * since header prediction assumes th->window stays fixed.
1223  *
1224  * Strictly speaking, keeping th->window fixed violates the receiver
1225  * side SWS prevention criteria. The problem is that under this rule
1226  * a stream of single byte packets will cause the right side of the
1227  * window to always advance by a single byte.
1228  *
1229  * Of course, if the sender implements sender side SWS prevention
1230  * then this will not be a problem.
1231  *
1232  * BSD seems to make the following compromise:
1233  *
1234  *	If the free space is less than the 1/4 of the maximum
1235  *	space available and the free space is less than 1/2 mss,
1236  *	then set the window to 0.
1237  *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1238  *	Otherwise, just prevent the window from shrinking
1239  *	and from being larger than the largest representable value.
1240  *
1241  * This prevents incremental opening of the window in the regime
1242  * where TCP is limited by the speed of the reader side taking
1243  * data out of the TCP receive queue. It does nothing about
1244  * those cases where the window is constrained on the sender side
1245  * because the pipeline is full.
1246  *
1247  * BSD also seems to "accidentally" limit itself to windows that are a
1248  * multiple of MSS, at least until the free space gets quite small.
1249  * This would appear to be a side effect of the mbuf implementation.
1250  * Combining these two algorithms results in the observed behavior
1251  * of having a fixed window size at almost all times.
1252  *
1253  * Below we obtain similar behavior by forcing the offered window to
1254  * a multiple of the mss when it is feasible to do so.
1255  *
1256  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1257  * Regular options like TIMESTAMP are taken into account.
1258  */
1259 u32 __tcp_select_window(struct sock *sk)
1260 {
1261 	struct inet_connection_sock *icsk = inet_csk(sk);
1262 	struct tcp_sock *tp = tcp_sk(sk);
1263 	/* MSS for the peer's data.  Previous versions used mss_clamp
1264 	 * here.  I don't know if the value based on our guesses
1265 	 * of peer's MSS is better for the performance.  It's more correct
1266 	 * but may be worse for the performance because of rcv_mss
1267 	 * fluctuations.  --SAW  1998/11/1
1268 	 */
1269 	int mss = icsk->icsk_ack.rcv_mss;
1270 	int free_space = tcp_space(sk);
1271 	int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1272 	int window;
1273 
1274 	if (mss > full_space)
1275 		mss = full_space;
1276 
1277 	if (free_space < full_space/2) {
1278 		icsk->icsk_ack.quick = 0;
1279 
1280 		if (tcp_memory_pressure)
1281 			tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1282 
1283 		if (free_space < mss)
1284 			return 0;
1285 	}
1286 
1287 	if (free_space > tp->rcv_ssthresh)
1288 		free_space = tp->rcv_ssthresh;
1289 
1290 	/* Don't do rounding if we are using window scaling, since the
1291 	 * scaled window will not line up with the MSS boundary anyway.
1292 	 */
1293 	window = tp->rcv_wnd;
1294 	if (tp->rx_opt.rcv_wscale) {
1295 		window = free_space;
1296 
1297 		/* Advertise enough space so that it won't get scaled away.
1298 		 * Import case: prevent zero window announcement if
1299 		 * 1<<rcv_wscale > mss.
1300 		 */
1301 		if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1302 			window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1303 				  << tp->rx_opt.rcv_wscale);
1304 	} else {
1305 		/* Get the largest window that is a nice multiple of mss.
1306 		 * Window clamp already applied above.
1307 		 * If our current window offering is within 1 mss of the
1308 		 * free space we just keep it. This prevents the divide
1309 		 * and multiply from happening most of the time.
1310 		 * We also don't do any window rounding when the free space
1311 		 * is too small.
1312 		 */
1313 		if (window <= free_space - mss || window > free_space)
1314 			window = (free_space/mss)*mss;
1315 	}
1316 
1317 	return window;
1318 }
1319 
1320 /* Attempt to collapse two adjacent SKB's during retransmission. */
1321 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1322 {
1323 	struct tcp_sock *tp = tcp_sk(sk);
1324 	struct sk_buff *next_skb = skb->next;
1325 
1326 	/* The first test we must make is that neither of these two
1327 	 * SKB's are still referenced by someone else.
1328 	 */
1329 	if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1330 		int skb_size = skb->len, next_skb_size = next_skb->len;
1331 		u16 flags = TCP_SKB_CB(skb)->flags;
1332 
1333 		/* Also punt if next skb has been SACK'd. */
1334 		if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1335 			return;
1336 
1337 		/* Next skb is out of window. */
1338 		if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1339 			return;
1340 
1341 		/* Punt if not enough space exists in the first SKB for
1342 		 * the data in the second, or the total combined payload
1343 		 * would exceed the MSS.
1344 		 */
1345 		if ((next_skb_size > skb_tailroom(skb)) ||
1346 		    ((skb_size + next_skb_size) > mss_now))
1347 			return;
1348 
1349 		BUG_ON(tcp_skb_pcount(skb) != 1 ||
1350 		       tcp_skb_pcount(next_skb) != 1);
1351 
1352 		/* changing transmit queue under us so clear hints */
1353 		clear_all_retrans_hints(tp);
1354 
1355 		/* Ok.	We will be able to collapse the packet. */
1356 		__skb_unlink(next_skb, &sk->sk_write_queue);
1357 
1358 		memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1359 
1360 		if (next_skb->ip_summed == CHECKSUM_HW)
1361 			skb->ip_summed = CHECKSUM_HW;
1362 
1363 		if (skb->ip_summed != CHECKSUM_HW)
1364 			skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1365 
1366 		/* Update sequence range on original skb. */
1367 		TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1368 
1369 		/* Merge over control information. */
1370 		flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1371 		TCP_SKB_CB(skb)->flags = flags;
1372 
1373 		/* All done, get rid of second SKB and account for it so
1374 		 * packet counting does not break.
1375 		 */
1376 		TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1377 		if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1378 			tp->retrans_out -= tcp_skb_pcount(next_skb);
1379 		if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1380 			tp->lost_out -= tcp_skb_pcount(next_skb);
1381 			tp->left_out -= tcp_skb_pcount(next_skb);
1382 		}
1383 		/* Reno case is special. Sigh... */
1384 		if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1385 			tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1386 			tp->left_out -= tcp_skb_pcount(next_skb);
1387 		}
1388 
1389 		/* Not quite right: it can be > snd.fack, but
1390 		 * it is better to underestimate fackets.
1391 		 */
1392 		tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1393 		tcp_packets_out_dec(tp, next_skb);
1394 		sk_stream_free_skb(sk, next_skb);
1395 	}
1396 }
1397 
1398 /* Do a simple retransmit without using the backoff mechanisms in
1399  * tcp_timer. This is used for path mtu discovery.
1400  * The socket is already locked here.
1401  */
1402 void tcp_simple_retransmit(struct sock *sk)
1403 {
1404 	const struct inet_connection_sock *icsk = inet_csk(sk);
1405 	struct tcp_sock *tp = tcp_sk(sk);
1406 	struct sk_buff *skb;
1407 	unsigned int mss = tcp_current_mss(sk, 0);
1408 	int lost = 0;
1409 
1410 	sk_stream_for_retrans_queue(skb, sk) {
1411 		if (skb->len > mss &&
1412 		    !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1413 			if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1414 				TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1415 				tp->retrans_out -= tcp_skb_pcount(skb);
1416 			}
1417 			if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1418 				TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1419 				tp->lost_out += tcp_skb_pcount(skb);
1420 				lost = 1;
1421 			}
1422 		}
1423 	}
1424 
1425 	clear_all_retrans_hints(tp);
1426 
1427 	if (!lost)
1428 		return;
1429 
1430 	tcp_sync_left_out(tp);
1431 
1432  	/* Don't muck with the congestion window here.
1433 	 * Reason is that we do not increase amount of _data_
1434 	 * in network, but units changed and effective
1435 	 * cwnd/ssthresh really reduced now.
1436 	 */
1437 	if (icsk->icsk_ca_state != TCP_CA_Loss) {
1438 		tp->high_seq = tp->snd_nxt;
1439 		tp->snd_ssthresh = tcp_current_ssthresh(sk);
1440 		tp->prior_ssthresh = 0;
1441 		tp->undo_marker = 0;
1442 		tcp_set_ca_state(sk, TCP_CA_Loss);
1443 	}
1444 	tcp_xmit_retransmit_queue(sk);
1445 }
1446 
1447 /* This retransmits one SKB.  Policy decisions and retransmit queue
1448  * state updates are done by the caller.  Returns non-zero if an
1449  * error occurred which prevented the send.
1450  */
1451 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1452 {
1453 	struct tcp_sock *tp = tcp_sk(sk);
1454  	unsigned int cur_mss = tcp_current_mss(sk, 0);
1455 	int err;
1456 
1457 	/* Do not sent more than we queued. 1/4 is reserved for possible
1458 	 * copying overhead: fragmentation, tunneling, mangling etc.
1459 	 */
1460 	if (atomic_read(&sk->sk_wmem_alloc) >
1461 	    min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1462 		return -EAGAIN;
1463 
1464 	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1465 		if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1466 			BUG();
1467 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1468 			return -ENOMEM;
1469 	}
1470 
1471 	/* If receiver has shrunk his window, and skb is out of
1472 	 * new window, do not retransmit it. The exception is the
1473 	 * case, when window is shrunk to zero. In this case
1474 	 * our retransmit serves as a zero window probe.
1475 	 */
1476 	if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1477 	    && TCP_SKB_CB(skb)->seq != tp->snd_una)
1478 		return -EAGAIN;
1479 
1480 	if (skb->len > cur_mss) {
1481 		if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1482 			return -ENOMEM; /* We'll try again later. */
1483 	}
1484 
1485 	/* Collapse two adjacent packets if worthwhile and we can. */
1486 	if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1487 	   (skb->len < (cur_mss >> 1)) &&
1488 	   (skb->next != sk->sk_send_head) &&
1489 	   (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1490 	   (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1491 	   (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1492 	   (sysctl_tcp_retrans_collapse != 0))
1493 		tcp_retrans_try_collapse(sk, skb, cur_mss);
1494 
1495 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1496 		return -EHOSTUNREACH; /* Routing failure or similar. */
1497 
1498 	/* Some Solaris stacks overoptimize and ignore the FIN on a
1499 	 * retransmit when old data is attached.  So strip it off
1500 	 * since it is cheap to do so and saves bytes on the network.
1501 	 */
1502 	if(skb->len > 0 &&
1503 	   (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1504 	   tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1505 		if (!pskb_trim(skb, 0)) {
1506 			TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1507 			skb_shinfo(skb)->tso_segs = 1;
1508 			skb_shinfo(skb)->tso_size = 0;
1509 			skb->ip_summed = CHECKSUM_NONE;
1510 			skb->csum = 0;
1511 		}
1512 	}
1513 
1514 	/* Make a copy, if the first transmission SKB clone we made
1515 	 * is still in somebody's hands, else make a clone.
1516 	 */
1517 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
1518 
1519 	err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1520 
1521 	if (err == 0) {
1522 		/* Update global TCP statistics. */
1523 		TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1524 
1525 		tp->total_retrans++;
1526 
1527 #if FASTRETRANS_DEBUG > 0
1528 		if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1529 			if (net_ratelimit())
1530 				printk(KERN_DEBUG "retrans_out leaked.\n");
1531 		}
1532 #endif
1533 		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1534 		tp->retrans_out += tcp_skb_pcount(skb);
1535 
1536 		/* Save stamp of the first retransmit. */
1537 		if (!tp->retrans_stamp)
1538 			tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1539 
1540 		tp->undo_retrans++;
1541 
1542 		/* snd_nxt is stored to detect loss of retransmitted segment,
1543 		 * see tcp_input.c tcp_sacktag_write_queue().
1544 		 */
1545 		TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1546 	}
1547 	return err;
1548 }
1549 
1550 /* This gets called after a retransmit timeout, and the initially
1551  * retransmitted data is acknowledged.  It tries to continue
1552  * resending the rest of the retransmit queue, until either
1553  * we've sent it all or the congestion window limit is reached.
1554  * If doing SACK, the first ACK which comes back for a timeout
1555  * based retransmit packet might feed us FACK information again.
1556  * If so, we use it to avoid unnecessarily retransmissions.
1557  */
1558 void tcp_xmit_retransmit_queue(struct sock *sk)
1559 {
1560 	const struct inet_connection_sock *icsk = inet_csk(sk);
1561 	struct tcp_sock *tp = tcp_sk(sk);
1562 	struct sk_buff *skb;
1563 	int packet_cnt;
1564 
1565 	if (tp->retransmit_skb_hint) {
1566 		skb = tp->retransmit_skb_hint;
1567 		packet_cnt = tp->retransmit_cnt_hint;
1568 	}else{
1569 		skb = sk->sk_write_queue.next;
1570 		packet_cnt = 0;
1571 	}
1572 
1573 	/* First pass: retransmit lost packets. */
1574 	if (tp->lost_out) {
1575 		sk_stream_for_retrans_queue_from(skb, sk) {
1576 			__u8 sacked = TCP_SKB_CB(skb)->sacked;
1577 
1578 			/* we could do better than to assign each time */
1579 			tp->retransmit_skb_hint = skb;
1580 			tp->retransmit_cnt_hint = packet_cnt;
1581 
1582 			/* Assume this retransmit will generate
1583 			 * only one packet for congestion window
1584 			 * calculation purposes.  This works because
1585 			 * tcp_retransmit_skb() will chop up the
1586 			 * packet to be MSS sized and all the
1587 			 * packet counting works out.
1588 			 */
1589 			if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1590 				return;
1591 
1592 			if (sacked & TCPCB_LOST) {
1593 				if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1594 					if (tcp_retransmit_skb(sk, skb)) {
1595 						tp->retransmit_skb_hint = NULL;
1596 						return;
1597 					}
1598 					if (icsk->icsk_ca_state != TCP_CA_Loss)
1599 						NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1600 					else
1601 						NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1602 
1603 					if (skb ==
1604 					    skb_peek(&sk->sk_write_queue))
1605 						inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1606 									  inet_csk(sk)->icsk_rto,
1607 									  TCP_RTO_MAX);
1608 				}
1609 
1610 				packet_cnt += tcp_skb_pcount(skb);
1611 				if (packet_cnt >= tp->lost_out)
1612 					break;
1613 			}
1614 		}
1615 	}
1616 
1617 	/* OK, demanded retransmission is finished. */
1618 
1619 	/* Forward retransmissions are possible only during Recovery. */
1620 	if (icsk->icsk_ca_state != TCP_CA_Recovery)
1621 		return;
1622 
1623 	/* No forward retransmissions in Reno are possible. */
1624 	if (!tp->rx_opt.sack_ok)
1625 		return;
1626 
1627 	/* Yeah, we have to make difficult choice between forward transmission
1628 	 * and retransmission... Both ways have their merits...
1629 	 *
1630 	 * For now we do not retransmit anything, while we have some new
1631 	 * segments to send.
1632 	 */
1633 
1634 	if (tcp_may_send_now(sk, tp))
1635 		return;
1636 
1637 	if (tp->forward_skb_hint) {
1638 		skb = tp->forward_skb_hint;
1639 		packet_cnt = tp->forward_cnt_hint;
1640 	} else{
1641 		skb = sk->sk_write_queue.next;
1642 		packet_cnt = 0;
1643 	}
1644 
1645 	sk_stream_for_retrans_queue_from(skb, sk) {
1646 		tp->forward_cnt_hint = packet_cnt;
1647 		tp->forward_skb_hint = skb;
1648 
1649 		/* Similar to the retransmit loop above we
1650 		 * can pretend that the retransmitted SKB
1651 		 * we send out here will be composed of one
1652 		 * real MSS sized packet because tcp_retransmit_skb()
1653 		 * will fragment it if necessary.
1654 		 */
1655 		if (++packet_cnt > tp->fackets_out)
1656 			break;
1657 
1658 		if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1659 			break;
1660 
1661 		if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1662 			continue;
1663 
1664 		/* Ok, retransmit it. */
1665 		if (tcp_retransmit_skb(sk, skb)) {
1666 			tp->forward_skb_hint = NULL;
1667 			break;
1668 		}
1669 
1670 		if (skb == skb_peek(&sk->sk_write_queue))
1671 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1672 						  inet_csk(sk)->icsk_rto,
1673 						  TCP_RTO_MAX);
1674 
1675 		NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1676 	}
1677 }
1678 
1679 
1680 /* Send a fin.  The caller locks the socket for us.  This cannot be
1681  * allowed to fail queueing a FIN frame under any circumstances.
1682  */
1683 void tcp_send_fin(struct sock *sk)
1684 {
1685 	struct tcp_sock *tp = tcp_sk(sk);
1686 	struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1687 	int mss_now;
1688 
1689 	/* Optimization, tack on the FIN if we have a queue of
1690 	 * unsent frames.  But be careful about outgoing SACKS
1691 	 * and IP options.
1692 	 */
1693 	mss_now = tcp_current_mss(sk, 1);
1694 
1695 	if (sk->sk_send_head != NULL) {
1696 		TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1697 		TCP_SKB_CB(skb)->end_seq++;
1698 		tp->write_seq++;
1699 	} else {
1700 		/* Socket is locked, keep trying until memory is available. */
1701 		for (;;) {
1702 			skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
1703 			if (skb)
1704 				break;
1705 			yield();
1706 		}
1707 
1708 		/* Reserve space for headers and prepare control bits. */
1709 		skb_reserve(skb, MAX_TCP_HEADER);
1710 		skb->csum = 0;
1711 		TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1712 		TCP_SKB_CB(skb)->sacked = 0;
1713 		skb_shinfo(skb)->tso_segs = 1;
1714 		skb_shinfo(skb)->tso_size = 0;
1715 
1716 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1717 		TCP_SKB_CB(skb)->seq = tp->write_seq;
1718 		TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1719 		tcp_queue_skb(sk, skb);
1720 	}
1721 	__tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1722 }
1723 
1724 /* We get here when a process closes a file descriptor (either due to
1725  * an explicit close() or as a byproduct of exit()'ing) and there
1726  * was unread data in the receive queue.  This behavior is recommended
1727  * by draft-ietf-tcpimpl-prob-03.txt section 3.10.  -DaveM
1728  */
1729 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
1730 {
1731 	struct tcp_sock *tp = tcp_sk(sk);
1732 	struct sk_buff *skb;
1733 
1734 	/* NOTE: No TCP options attached and we never retransmit this. */
1735 	skb = alloc_skb(MAX_TCP_HEADER, priority);
1736 	if (!skb) {
1737 		NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1738 		return;
1739 	}
1740 
1741 	/* Reserve space for headers and prepare control bits. */
1742 	skb_reserve(skb, MAX_TCP_HEADER);
1743 	skb->csum = 0;
1744 	TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1745 	TCP_SKB_CB(skb)->sacked = 0;
1746 	skb_shinfo(skb)->tso_segs = 1;
1747 	skb_shinfo(skb)->tso_size = 0;
1748 
1749 	/* Send it off. */
1750 	TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1751 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1752 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
1753 	if (tcp_transmit_skb(sk, skb, 0, priority))
1754 		NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1755 }
1756 
1757 /* WARNING: This routine must only be called when we have already sent
1758  * a SYN packet that crossed the incoming SYN that caused this routine
1759  * to get called. If this assumption fails then the initial rcv_wnd
1760  * and rcv_wscale values will not be correct.
1761  */
1762 int tcp_send_synack(struct sock *sk)
1763 {
1764 	struct sk_buff* skb;
1765 
1766 	skb = skb_peek(&sk->sk_write_queue);
1767 	if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1768 		printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1769 		return -EFAULT;
1770 	}
1771 	if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1772 		if (skb_cloned(skb)) {
1773 			struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
1774 			if (nskb == NULL)
1775 				return -ENOMEM;
1776 			__skb_unlink(skb, &sk->sk_write_queue);
1777 			skb_header_release(nskb);
1778 			__skb_queue_head(&sk->sk_write_queue, nskb);
1779 			sk_stream_free_skb(sk, skb);
1780 			sk_charge_skb(sk, nskb);
1781 			skb = nskb;
1782 		}
1783 
1784 		TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
1785 		TCP_ECN_send_synack(tcp_sk(sk), skb);
1786 	}
1787 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
1788 	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1789 }
1790 
1791 /*
1792  * Prepare a SYN-ACK.
1793  */
1794 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
1795 				 struct request_sock *req)
1796 {
1797 	struct inet_request_sock *ireq = inet_rsk(req);
1798 	struct tcp_sock *tp = tcp_sk(sk);
1799 	struct tcphdr *th;
1800 	int tcp_header_size;
1801 	struct sk_buff *skb;
1802 
1803 	skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
1804 	if (skb == NULL)
1805 		return NULL;
1806 
1807 	/* Reserve space for headers. */
1808 	skb_reserve(skb, MAX_TCP_HEADER);
1809 
1810 	skb->dst = dst_clone(dst);
1811 
1812 	tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
1813 			   (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
1814 			   (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
1815 			   /* SACK_PERM is in the place of NOP NOP of TS */
1816 			   ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
1817 	skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
1818 
1819 	memset(th, 0, sizeof(struct tcphdr));
1820 	th->syn = 1;
1821 	th->ack = 1;
1822 	if (dst->dev->features&NETIF_F_TSO)
1823 		ireq->ecn_ok = 0;
1824 	TCP_ECN_make_synack(req, th);
1825 	th->source = inet_sk(sk)->sport;
1826 	th->dest = ireq->rmt_port;
1827 	TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
1828 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1829 	TCP_SKB_CB(skb)->sacked = 0;
1830 	skb_shinfo(skb)->tso_segs = 1;
1831 	skb_shinfo(skb)->tso_size = 0;
1832 	th->seq = htonl(TCP_SKB_CB(skb)->seq);
1833 	th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
1834 	if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
1835 		__u8 rcv_wscale;
1836 		/* Set this up on the first call only */
1837 		req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
1838 		/* tcp_full_space because it is guaranteed to be the first packet */
1839 		tcp_select_initial_window(tcp_full_space(sk),
1840 			dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
1841 			&req->rcv_wnd,
1842 			&req->window_clamp,
1843 			ireq->wscale_ok,
1844 			&rcv_wscale);
1845 		ireq->rcv_wscale = rcv_wscale;
1846 	}
1847 
1848 	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1849 	th->window = htons(req->rcv_wnd);
1850 
1851 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
1852 	tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
1853 			      ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
1854 			      TCP_SKB_CB(skb)->when,
1855 			      req->ts_recent);
1856 
1857 	skb->csum = 0;
1858 	th->doff = (tcp_header_size >> 2);
1859 	TCP_INC_STATS(TCP_MIB_OUTSEGS);
1860 	return skb;
1861 }
1862 
1863 /*
1864  * Do all connect socket setups that can be done AF independent.
1865  */
1866 static void tcp_connect_init(struct sock *sk)
1867 {
1868 	struct dst_entry *dst = __sk_dst_get(sk);
1869 	struct tcp_sock *tp = tcp_sk(sk);
1870 	__u8 rcv_wscale;
1871 
1872 	/* We'll fix this up when we get a response from the other end.
1873 	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1874 	 */
1875 	tp->tcp_header_len = sizeof(struct tcphdr) +
1876 		(sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
1877 
1878 	/* If user gave his TCP_MAXSEG, record it to clamp */
1879 	if (tp->rx_opt.user_mss)
1880 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
1881 	tp->max_window = 0;
1882 	tcp_sync_mss(sk, dst_mtu(dst));
1883 
1884 	if (!tp->window_clamp)
1885 		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
1886 	tp->advmss = dst_metric(dst, RTAX_ADVMSS);
1887 	tcp_initialize_rcv_mss(sk);
1888 
1889 	tcp_select_initial_window(tcp_full_space(sk),
1890 				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
1891 				  &tp->rcv_wnd,
1892 				  &tp->window_clamp,
1893 				  sysctl_tcp_window_scaling,
1894 				  &rcv_wscale);
1895 
1896 	tp->rx_opt.rcv_wscale = rcv_wscale;
1897 	tp->rcv_ssthresh = tp->rcv_wnd;
1898 
1899 	sk->sk_err = 0;
1900 	sock_reset_flag(sk, SOCK_DONE);
1901 	tp->snd_wnd = 0;
1902 	tcp_init_wl(tp, tp->write_seq, 0);
1903 	tp->snd_una = tp->write_seq;
1904 	tp->snd_sml = tp->write_seq;
1905 	tp->rcv_nxt = 0;
1906 	tp->rcv_wup = 0;
1907 	tp->copied_seq = 0;
1908 
1909 	inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
1910 	inet_csk(sk)->icsk_retransmits = 0;
1911 	tcp_clear_retrans(tp);
1912 }
1913 
1914 /*
1915  * Build a SYN and send it off.
1916  */
1917 int tcp_connect(struct sock *sk)
1918 {
1919 	struct tcp_sock *tp = tcp_sk(sk);
1920 	struct sk_buff *buff;
1921 
1922 	tcp_connect_init(sk);
1923 
1924 	buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
1925 	if (unlikely(buff == NULL))
1926 		return -ENOBUFS;
1927 
1928 	/* Reserve space for headers. */
1929 	skb_reserve(buff, MAX_TCP_HEADER);
1930 
1931 	TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
1932 	TCP_ECN_send_syn(sk, tp, buff);
1933 	TCP_SKB_CB(buff)->sacked = 0;
1934 	skb_shinfo(buff)->tso_segs = 1;
1935 	skb_shinfo(buff)->tso_size = 0;
1936 	buff->csum = 0;
1937 	TCP_SKB_CB(buff)->seq = tp->write_seq++;
1938 	TCP_SKB_CB(buff)->end_seq = tp->write_seq;
1939 	tp->snd_nxt = tp->write_seq;
1940 	tp->pushed_seq = tp->write_seq;
1941 
1942 	/* Send it off. */
1943 	TCP_SKB_CB(buff)->when = tcp_time_stamp;
1944 	tp->retrans_stamp = TCP_SKB_CB(buff)->when;
1945 	skb_header_release(buff);
1946 	__skb_queue_tail(&sk->sk_write_queue, buff);
1947 	sk_charge_skb(sk, buff);
1948 	tp->packets_out += tcp_skb_pcount(buff);
1949 	tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
1950 	TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
1951 
1952 	/* Timer for repeating the SYN until an answer. */
1953 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1954 				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
1955 	return 0;
1956 }
1957 
1958 /* Send out a delayed ack, the caller does the policy checking
1959  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
1960  * for details.
1961  */
1962 void tcp_send_delayed_ack(struct sock *sk)
1963 {
1964 	struct inet_connection_sock *icsk = inet_csk(sk);
1965 	int ato = icsk->icsk_ack.ato;
1966 	unsigned long timeout;
1967 
1968 	if (ato > TCP_DELACK_MIN) {
1969 		const struct tcp_sock *tp = tcp_sk(sk);
1970 		int max_ato = HZ/2;
1971 
1972 		if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
1973 			max_ato = TCP_DELACK_MAX;
1974 
1975 		/* Slow path, intersegment interval is "high". */
1976 
1977 		/* If some rtt estimate is known, use it to bound delayed ack.
1978 		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
1979 		 * directly.
1980 		 */
1981 		if (tp->srtt) {
1982 			int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
1983 
1984 			if (rtt < max_ato)
1985 				max_ato = rtt;
1986 		}
1987 
1988 		ato = min(ato, max_ato);
1989 	}
1990 
1991 	/* Stay within the limit we were given */
1992 	timeout = jiffies + ato;
1993 
1994 	/* Use new timeout only if there wasn't a older one earlier. */
1995 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
1996 		/* If delack timer was blocked or is about to expire,
1997 		 * send ACK now.
1998 		 */
1999 		if (icsk->icsk_ack.blocked ||
2000 		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2001 			tcp_send_ack(sk);
2002 			return;
2003 		}
2004 
2005 		if (!time_before(timeout, icsk->icsk_ack.timeout))
2006 			timeout = icsk->icsk_ack.timeout;
2007 	}
2008 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2009 	icsk->icsk_ack.timeout = timeout;
2010 	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2011 }
2012 
2013 /* This routine sends an ack and also updates the window. */
2014 void tcp_send_ack(struct sock *sk)
2015 {
2016 	/* If we have been reset, we may not send again. */
2017 	if (sk->sk_state != TCP_CLOSE) {
2018 		struct tcp_sock *tp = tcp_sk(sk);
2019 		struct sk_buff *buff;
2020 
2021 		/* We are not putting this on the write queue, so
2022 		 * tcp_transmit_skb() will set the ownership to this
2023 		 * sock.
2024 		 */
2025 		buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2026 		if (buff == NULL) {
2027 			inet_csk_schedule_ack(sk);
2028 			inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2029 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2030 						  TCP_DELACK_MAX, TCP_RTO_MAX);
2031 			return;
2032 		}
2033 
2034 		/* Reserve space for headers and prepare control bits. */
2035 		skb_reserve(buff, MAX_TCP_HEADER);
2036 		buff->csum = 0;
2037 		TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2038 		TCP_SKB_CB(buff)->sacked = 0;
2039 		skb_shinfo(buff)->tso_segs = 1;
2040 		skb_shinfo(buff)->tso_size = 0;
2041 
2042 		/* Send it off, this clears delayed acks for us. */
2043 		TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
2044 		TCP_SKB_CB(buff)->when = tcp_time_stamp;
2045 		tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2046 	}
2047 }
2048 
2049 /* This routine sends a packet with an out of date sequence
2050  * number. It assumes the other end will try to ack it.
2051  *
2052  * Question: what should we make while urgent mode?
2053  * 4.4BSD forces sending single byte of data. We cannot send
2054  * out of window data, because we have SND.NXT==SND.MAX...
2055  *
2056  * Current solution: to send TWO zero-length segments in urgent mode:
2057  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2058  * out-of-date with SND.UNA-1 to probe window.
2059  */
2060 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2061 {
2062 	struct tcp_sock *tp = tcp_sk(sk);
2063 	struct sk_buff *skb;
2064 
2065 	/* We don't queue it, tcp_transmit_skb() sets ownership. */
2066 	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2067 	if (skb == NULL)
2068 		return -1;
2069 
2070 	/* Reserve space for headers and set control bits. */
2071 	skb_reserve(skb, MAX_TCP_HEADER);
2072 	skb->csum = 0;
2073 	TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2074 	TCP_SKB_CB(skb)->sacked = urgent;
2075 	skb_shinfo(skb)->tso_segs = 1;
2076 	skb_shinfo(skb)->tso_size = 0;
2077 
2078 	/* Use a previous sequence.  This should cause the other
2079 	 * end to send an ack.  Don't queue or clone SKB, just
2080 	 * send it.
2081 	 */
2082 	TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2083 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2084 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2085 	return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2086 }
2087 
2088 int tcp_write_wakeup(struct sock *sk)
2089 {
2090 	if (sk->sk_state != TCP_CLOSE) {
2091 		struct tcp_sock *tp = tcp_sk(sk);
2092 		struct sk_buff *skb;
2093 
2094 		if ((skb = sk->sk_send_head) != NULL &&
2095 		    before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
2096 			int err;
2097 			unsigned int mss = tcp_current_mss(sk, 0);
2098 			unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
2099 
2100 			if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2101 				tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2102 
2103 			/* We are probing the opening of a window
2104 			 * but the window size is != 0
2105 			 * must have been a result SWS avoidance ( sender )
2106 			 */
2107 			if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2108 			    skb->len > mss) {
2109 				seg_size = min(seg_size, mss);
2110 				TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2111 				if (tcp_fragment(sk, skb, seg_size, mss))
2112 					return -1;
2113 			} else if (!tcp_skb_pcount(skb))
2114 				tcp_set_skb_tso_segs(sk, skb, mss);
2115 
2116 			TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2117 			TCP_SKB_CB(skb)->when = tcp_time_stamp;
2118 			err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2119 			if (!err) {
2120 				update_send_head(sk, tp, skb);
2121 			}
2122 			return err;
2123 		} else {
2124 			if (tp->urg_mode &&
2125 			    between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2126 				tcp_xmit_probe_skb(sk, TCPCB_URG);
2127 			return tcp_xmit_probe_skb(sk, 0);
2128 		}
2129 	}
2130 	return -1;
2131 }
2132 
2133 /* A window probe timeout has occurred.  If window is not closed send
2134  * a partial packet else a zero probe.
2135  */
2136 void tcp_send_probe0(struct sock *sk)
2137 {
2138 	struct inet_connection_sock *icsk = inet_csk(sk);
2139 	struct tcp_sock *tp = tcp_sk(sk);
2140 	int err;
2141 
2142 	err = tcp_write_wakeup(sk);
2143 
2144 	if (tp->packets_out || !sk->sk_send_head) {
2145 		/* Cancel probe timer, if it is not required. */
2146 		icsk->icsk_probes_out = 0;
2147 		icsk->icsk_backoff = 0;
2148 		return;
2149 	}
2150 
2151 	if (err <= 0) {
2152 		if (icsk->icsk_backoff < sysctl_tcp_retries2)
2153 			icsk->icsk_backoff++;
2154 		icsk->icsk_probes_out++;
2155 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2156 					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2157 					  TCP_RTO_MAX);
2158 	} else {
2159 		/* If packet was not sent due to local congestion,
2160 		 * do not backoff and do not remember icsk_probes_out.
2161 		 * Let local senders to fight for local resources.
2162 		 *
2163 		 * Use accumulated backoff yet.
2164 		 */
2165 		if (!icsk->icsk_probes_out)
2166 			icsk->icsk_probes_out = 1;
2167 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2168 					  min(icsk->icsk_rto << icsk->icsk_backoff,
2169 					      TCP_RESOURCE_PROBE_INTERVAL),
2170 					  TCP_RTO_MAX);
2171 	}
2172 }
2173 
2174 EXPORT_SYMBOL(tcp_connect);
2175 EXPORT_SYMBOL(tcp_make_synack);
2176 EXPORT_SYMBOL(tcp_simple_retransmit);
2177 EXPORT_SYMBOL(tcp_sync_mss);
2178 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
2179