xref: /openbmc/linux/net/ipv4/tcp_output.c (revision 64c70b1c)
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 
44 /* People can turn this off for buggy TCP's found in printers etc. */
45 int sysctl_tcp_retrans_collapse __read_mostly = 1;
46 
47 /* People can turn this on to  work with those rare, broken TCPs that
48  * interpret the window field as a signed quantity.
49  */
50 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
51 
52 /* This limits the percentage of the congestion window which we
53  * will allow a single TSO frame to consume.  Building TSO frames
54  * which are too large can cause TCP streams to be bursty.
55  */
56 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
57 
58 int sysctl_tcp_mtu_probing __read_mostly = 0;
59 int sysctl_tcp_base_mss __read_mostly = 512;
60 
61 /* By default, RFC2861 behavior.  */
62 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
63 
64 static void update_send_head(struct sock *sk, struct sk_buff *skb)
65 {
66 	struct tcp_sock *tp = tcp_sk(sk);
67 
68 	tcp_advance_send_head(sk, skb);
69 	tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
70 	tcp_packets_out_inc(sk, skb);
71 }
72 
73 /* SND.NXT, if window was not shrunk.
74  * If window has been shrunk, what should we make? It is not clear at all.
75  * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
76  * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
77  * invalid. OK, let's make this for now:
78  */
79 static inline __u32 tcp_acceptable_seq(struct sock *sk)
80 {
81 	struct tcp_sock *tp = tcp_sk(sk);
82 
83 	if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
84 		return tp->snd_nxt;
85 	else
86 		return tp->snd_una+tp->snd_wnd;
87 }
88 
89 /* Calculate mss to advertise in SYN segment.
90  * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
91  *
92  * 1. It is independent of path mtu.
93  * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
94  * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
95  *    attached devices, because some buggy hosts are confused by
96  *    large MSS.
97  * 4. We do not make 3, we advertise MSS, calculated from first
98  *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
99  *    This may be overridden via information stored in routing table.
100  * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
101  *    probably even Jumbo".
102  */
103 static __u16 tcp_advertise_mss(struct sock *sk)
104 {
105 	struct tcp_sock *tp = tcp_sk(sk);
106 	struct dst_entry *dst = __sk_dst_get(sk);
107 	int mss = tp->advmss;
108 
109 	if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
110 		mss = dst_metric(dst, RTAX_ADVMSS);
111 		tp->advmss = mss;
112 	}
113 
114 	return (__u16)mss;
115 }
116 
117 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
118  * This is the first part of cwnd validation mechanism. */
119 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
120 {
121 	struct tcp_sock *tp = tcp_sk(sk);
122 	s32 delta = tcp_time_stamp - tp->lsndtime;
123 	u32 restart_cwnd = tcp_init_cwnd(tp, dst);
124 	u32 cwnd = tp->snd_cwnd;
125 
126 	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
127 
128 	tp->snd_ssthresh = tcp_current_ssthresh(sk);
129 	restart_cwnd = min(restart_cwnd, cwnd);
130 
131 	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
132 		cwnd >>= 1;
133 	tp->snd_cwnd = max(cwnd, restart_cwnd);
134 	tp->snd_cwnd_stamp = tcp_time_stamp;
135 	tp->snd_cwnd_used = 0;
136 }
137 
138 static void tcp_event_data_sent(struct tcp_sock *tp,
139 				struct sk_buff *skb, struct sock *sk)
140 {
141 	struct inet_connection_sock *icsk = inet_csk(sk);
142 	const u32 now = tcp_time_stamp;
143 
144 	if (sysctl_tcp_slow_start_after_idle &&
145 	    (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
146 		tcp_cwnd_restart(sk, __sk_dst_get(sk));
147 
148 	tp->lsndtime = now;
149 
150 	/* If it is a reply for ato after last received
151 	 * packet, enter pingpong mode.
152 	 */
153 	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
154 		icsk->icsk_ack.pingpong = 1;
155 }
156 
157 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
158 {
159 	tcp_dec_quickack_mode(sk, pkts);
160 	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
161 }
162 
163 /* Determine a window scaling and initial window to offer.
164  * Based on the assumption that the given amount of space
165  * will be offered. Store the results in the tp structure.
166  * NOTE: for smooth operation initial space offering should
167  * be a multiple of mss if possible. We assume here that mss >= 1.
168  * This MUST be enforced by all callers.
169  */
170 void tcp_select_initial_window(int __space, __u32 mss,
171 			       __u32 *rcv_wnd, __u32 *window_clamp,
172 			       int wscale_ok, __u8 *rcv_wscale)
173 {
174 	unsigned int space = (__space < 0 ? 0 : __space);
175 
176 	/* If no clamp set the clamp to the max possible scaled window */
177 	if (*window_clamp == 0)
178 		(*window_clamp) = (65535 << 14);
179 	space = min(*window_clamp, space);
180 
181 	/* Quantize space offering to a multiple of mss if possible. */
182 	if (space > mss)
183 		space = (space / mss) * mss;
184 
185 	/* NOTE: offering an initial window larger than 32767
186 	 * will break some buggy TCP stacks. If the admin tells us
187 	 * it is likely we could be speaking with such a buggy stack
188 	 * we will truncate our initial window offering to 32K-1
189 	 * unless the remote has sent us a window scaling option,
190 	 * which we interpret as a sign the remote TCP is not
191 	 * misinterpreting the window field as a signed quantity.
192 	 */
193 	if (sysctl_tcp_workaround_signed_windows)
194 		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
195 	else
196 		(*rcv_wnd) = space;
197 
198 	(*rcv_wscale) = 0;
199 	if (wscale_ok) {
200 		/* Set window scaling on max possible window
201 		 * See RFC1323 for an explanation of the limit to 14
202 		 */
203 		space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
204 		space = min_t(u32, space, *window_clamp);
205 		while (space > 65535 && (*rcv_wscale) < 14) {
206 			space >>= 1;
207 			(*rcv_wscale)++;
208 		}
209 	}
210 
211 	/* Set initial window to value enough for senders,
212 	 * following RFC2414. Senders, not following this RFC,
213 	 * will be satisfied with 2.
214 	 */
215 	if (mss > (1<<*rcv_wscale)) {
216 		int init_cwnd = 4;
217 		if (mss > 1460*3)
218 			init_cwnd = 2;
219 		else if (mss > 1460)
220 			init_cwnd = 3;
221 		if (*rcv_wnd > init_cwnd*mss)
222 			*rcv_wnd = init_cwnd*mss;
223 	}
224 
225 	/* Set the clamp no higher than max representable value */
226 	(*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
227 }
228 
229 /* Chose a new window to advertise, update state in tcp_sock for the
230  * socket, and return result with RFC1323 scaling applied.  The return
231  * value can be stuffed directly into th->window for an outgoing
232  * frame.
233  */
234 static u16 tcp_select_window(struct sock *sk)
235 {
236 	struct tcp_sock *tp = tcp_sk(sk);
237 	u32 cur_win = tcp_receive_window(tp);
238 	u32 new_win = __tcp_select_window(sk);
239 
240 	/* Never shrink the offered window */
241 	if (new_win < cur_win) {
242 		/* Danger Will Robinson!
243 		 * Don't update rcv_wup/rcv_wnd here or else
244 		 * we will not be able to advertise a zero
245 		 * window in time.  --DaveM
246 		 *
247 		 * Relax Will Robinson.
248 		 */
249 		new_win = cur_win;
250 	}
251 	tp->rcv_wnd = new_win;
252 	tp->rcv_wup = tp->rcv_nxt;
253 
254 	/* Make sure we do not exceed the maximum possible
255 	 * scaled window.
256 	 */
257 	if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
258 		new_win = min(new_win, MAX_TCP_WINDOW);
259 	else
260 		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
261 
262 	/* RFC1323 scaling applied */
263 	new_win >>= tp->rx_opt.rcv_wscale;
264 
265 	/* If we advertise zero window, disable fast path. */
266 	if (new_win == 0)
267 		tp->pred_flags = 0;
268 
269 	return new_win;
270 }
271 
272 static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp,
273 					 __u32 tstamp, __u8 **md5_hash)
274 {
275 	if (tp->rx_opt.tstamp_ok) {
276 		*ptr++ = htonl((TCPOPT_NOP << 24) |
277 			       (TCPOPT_NOP << 16) |
278 			       (TCPOPT_TIMESTAMP << 8) |
279 			       TCPOLEN_TIMESTAMP);
280 		*ptr++ = htonl(tstamp);
281 		*ptr++ = htonl(tp->rx_opt.ts_recent);
282 	}
283 	if (tp->rx_opt.eff_sacks) {
284 		struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
285 		int this_sack;
286 
287 		*ptr++ = htonl((TCPOPT_NOP  << 24) |
288 			       (TCPOPT_NOP  << 16) |
289 			       (TCPOPT_SACK <<  8) |
290 			       (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
291 						     TCPOLEN_SACK_PERBLOCK)));
292 
293 		for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
294 			*ptr++ = htonl(sp[this_sack].start_seq);
295 			*ptr++ = htonl(sp[this_sack].end_seq);
296 		}
297 
298 		if (tp->rx_opt.dsack) {
299 			tp->rx_opt.dsack = 0;
300 			tp->rx_opt.eff_sacks--;
301 		}
302 	}
303 #ifdef CONFIG_TCP_MD5SIG
304 	if (md5_hash) {
305 		*ptr++ = htonl((TCPOPT_NOP << 24) |
306 			       (TCPOPT_NOP << 16) |
307 			       (TCPOPT_MD5SIG << 8) |
308 			       TCPOLEN_MD5SIG);
309 		*md5_hash = (__u8 *)ptr;
310 	}
311 #endif
312 }
313 
314 /* Construct a tcp options header for a SYN or SYN_ACK packet.
315  * If this is every changed make sure to change the definition of
316  * MAX_SYN_SIZE to match the new maximum number of options that you
317  * can generate.
318  *
319  * Note - that with the RFC2385 TCP option, we make room for the
320  * 16 byte MD5 hash. This will be filled in later, so the pointer for the
321  * location to be filled is passed back up.
322  */
323 static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack,
324 				  int offer_wscale, int wscale, __u32 tstamp,
325 				  __u32 ts_recent, __u8 **md5_hash)
326 {
327 	/* We always get an MSS option.
328 	 * The option bytes which will be seen in normal data
329 	 * packets should timestamps be used, must be in the MSS
330 	 * advertised.  But we subtract them from tp->mss_cache so
331 	 * that calculations in tcp_sendmsg are simpler etc.
332 	 * So account for this fact here if necessary.  If we
333 	 * don't do this correctly, as a receiver we won't
334 	 * recognize data packets as being full sized when we
335 	 * should, and thus we won't abide by the delayed ACK
336 	 * rules correctly.
337 	 * SACKs don't matter, we never delay an ACK when we
338 	 * have any of those going out.
339 	 */
340 	*ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
341 	if (ts) {
342 		if (sack)
343 			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
344 				       (TCPOLEN_SACK_PERM << 16) |
345 				       (TCPOPT_TIMESTAMP << 8) |
346 				       TCPOLEN_TIMESTAMP);
347 		else
348 			*ptr++ = htonl((TCPOPT_NOP << 24) |
349 				       (TCPOPT_NOP << 16) |
350 				       (TCPOPT_TIMESTAMP << 8) |
351 				       TCPOLEN_TIMESTAMP);
352 		*ptr++ = htonl(tstamp);		/* TSVAL */
353 		*ptr++ = htonl(ts_recent);	/* TSECR */
354 	} else if (sack)
355 		*ptr++ = htonl((TCPOPT_NOP << 24) |
356 			       (TCPOPT_NOP << 16) |
357 			       (TCPOPT_SACK_PERM << 8) |
358 			       TCPOLEN_SACK_PERM);
359 	if (offer_wscale)
360 		*ptr++ = htonl((TCPOPT_NOP << 24) |
361 			       (TCPOPT_WINDOW << 16) |
362 			       (TCPOLEN_WINDOW << 8) |
363 			       (wscale));
364 #ifdef CONFIG_TCP_MD5SIG
365 	/*
366 	 * If MD5 is enabled, then we set the option, and include the size
367 	 * (always 18). The actual MD5 hash is added just before the
368 	 * packet is sent.
369 	 */
370 	if (md5_hash) {
371 		*ptr++ = htonl((TCPOPT_NOP << 24) |
372 			       (TCPOPT_NOP << 16) |
373 			       (TCPOPT_MD5SIG << 8) |
374 			       TCPOLEN_MD5SIG);
375 		*md5_hash = (__u8 *) ptr;
376 	}
377 #endif
378 }
379 
380 /* This routine actually transmits TCP packets queued in by
381  * tcp_do_sendmsg().  This is used by both the initial
382  * transmission and possible later retransmissions.
383  * All SKB's seen here are completely headerless.  It is our
384  * job to build the TCP header, and pass the packet down to
385  * IP so it can do the same plus pass the packet off to the
386  * device.
387  *
388  * We are working here with either a clone of the original
389  * SKB, or a fresh unique copy made by the retransmit engine.
390  */
391 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask)
392 {
393 	const struct inet_connection_sock *icsk = inet_csk(sk);
394 	struct inet_sock *inet;
395 	struct tcp_sock *tp;
396 	struct tcp_skb_cb *tcb;
397 	int tcp_header_size;
398 #ifdef CONFIG_TCP_MD5SIG
399 	struct tcp_md5sig_key *md5;
400 	__u8 *md5_hash_location;
401 #endif
402 	struct tcphdr *th;
403 	int sysctl_flags;
404 	int err;
405 
406 	BUG_ON(!skb || !tcp_skb_pcount(skb));
407 
408 	/* If congestion control is doing timestamping, we must
409 	 * take such a timestamp before we potentially clone/copy.
410 	 */
411 	if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
412 		__net_timestamp(skb);
413 
414 	if (likely(clone_it)) {
415 		if (unlikely(skb_cloned(skb)))
416 			skb = pskb_copy(skb, gfp_mask);
417 		else
418 			skb = skb_clone(skb, gfp_mask);
419 		if (unlikely(!skb))
420 			return -ENOBUFS;
421 	}
422 
423 	inet = inet_sk(sk);
424 	tp = tcp_sk(sk);
425 	tcb = TCP_SKB_CB(skb);
426 	tcp_header_size = tp->tcp_header_len;
427 
428 #define SYSCTL_FLAG_TSTAMPS	0x1
429 #define SYSCTL_FLAG_WSCALE	0x2
430 #define SYSCTL_FLAG_SACK	0x4
431 
432 	sysctl_flags = 0;
433 	if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
434 		tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
435 		if (sysctl_tcp_timestamps) {
436 			tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
437 			sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
438 		}
439 		if (sysctl_tcp_window_scaling) {
440 			tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
441 			sysctl_flags |= SYSCTL_FLAG_WSCALE;
442 		}
443 		if (sysctl_tcp_sack) {
444 			sysctl_flags |= SYSCTL_FLAG_SACK;
445 			if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
446 				tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
447 		}
448 	} else if (unlikely(tp->rx_opt.eff_sacks)) {
449 		/* A SACK is 2 pad bytes, a 2 byte header, plus
450 		 * 2 32-bit sequence numbers for each SACK block.
451 		 */
452 		tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
453 				    (tp->rx_opt.eff_sacks *
454 				     TCPOLEN_SACK_PERBLOCK));
455 	}
456 
457 	if (tcp_packets_in_flight(tp) == 0)
458 		tcp_ca_event(sk, CA_EVENT_TX_START);
459 
460 #ifdef CONFIG_TCP_MD5SIG
461 	/*
462 	 * Are we doing MD5 on this segment? If so - make
463 	 * room for it.
464 	 */
465 	md5 = tp->af_specific->md5_lookup(sk, sk);
466 	if (md5)
467 		tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
468 #endif
469 
470 	skb_push(skb, tcp_header_size);
471 	skb_reset_transport_header(skb);
472 	skb_set_owner_w(skb, sk);
473 
474 	/* Build TCP header and checksum it. */
475 	th = tcp_hdr(skb);
476 	th->source		= inet->sport;
477 	th->dest		= inet->dport;
478 	th->seq			= htonl(tcb->seq);
479 	th->ack_seq		= htonl(tp->rcv_nxt);
480 	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
481 					tcb->flags);
482 
483 	if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
484 		/* RFC1323: The window in SYN & SYN/ACK segments
485 		 * is never scaled.
486 		 */
487 		th->window	= htons(min(tp->rcv_wnd, 65535U));
488 	} else {
489 		th->window	= htons(tcp_select_window(sk));
490 	}
491 	th->check		= 0;
492 	th->urg_ptr		= 0;
493 
494 	if (unlikely(tp->urg_mode &&
495 		     between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
496 		th->urg_ptr		= htons(tp->snd_up-tcb->seq);
497 		th->urg			= 1;
498 	}
499 
500 	if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
501 		tcp_syn_build_options((__be32 *)(th + 1),
502 				      tcp_advertise_mss(sk),
503 				      (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
504 				      (sysctl_flags & SYSCTL_FLAG_SACK),
505 				      (sysctl_flags & SYSCTL_FLAG_WSCALE),
506 				      tp->rx_opt.rcv_wscale,
507 				      tcb->when,
508 				      tp->rx_opt.ts_recent,
509 
510 #ifdef CONFIG_TCP_MD5SIG
511 				      md5 ? &md5_hash_location :
512 #endif
513 				      NULL);
514 	} else {
515 		tcp_build_and_update_options((__be32 *)(th + 1),
516 					     tp, tcb->when,
517 #ifdef CONFIG_TCP_MD5SIG
518 					     md5 ? &md5_hash_location :
519 #endif
520 					     NULL);
521 		TCP_ECN_send(sk, skb, tcp_header_size);
522 	}
523 
524 #ifdef CONFIG_TCP_MD5SIG
525 	/* Calculate the MD5 hash, as we have all we need now */
526 	if (md5) {
527 		tp->af_specific->calc_md5_hash(md5_hash_location,
528 					       md5,
529 					       sk, NULL, NULL,
530 					       tcp_hdr(skb),
531 					       sk->sk_protocol,
532 					       skb->len);
533 	}
534 #endif
535 
536 	icsk->icsk_af_ops->send_check(sk, skb->len, skb);
537 
538 	if (likely(tcb->flags & TCPCB_FLAG_ACK))
539 		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
540 
541 	if (skb->len != tcp_header_size)
542 		tcp_event_data_sent(tp, skb, sk);
543 
544 	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
545 		TCP_INC_STATS(TCP_MIB_OUTSEGS);
546 
547 	err = icsk->icsk_af_ops->queue_xmit(skb, 0);
548 	if (likely(err <= 0))
549 		return err;
550 
551 	tcp_enter_cwr(sk, 1);
552 
553 	return net_xmit_eval(err);
554 
555 #undef SYSCTL_FLAG_TSTAMPS
556 #undef SYSCTL_FLAG_WSCALE
557 #undef SYSCTL_FLAG_SACK
558 }
559 
560 
561 /* This routine just queue's the buffer
562  *
563  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
564  * otherwise socket can stall.
565  */
566 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
567 {
568 	struct tcp_sock *tp = tcp_sk(sk);
569 
570 	/* Advance write_seq and place onto the write_queue. */
571 	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
572 	skb_header_release(skb);
573 	tcp_add_write_queue_tail(sk, skb);
574 	sk_charge_skb(sk, skb);
575 }
576 
577 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
578 {
579 	if (skb->len <= mss_now || !sk_can_gso(sk)) {
580 		/* Avoid the costly divide in the normal
581 		 * non-TSO case.
582 		 */
583 		skb_shinfo(skb)->gso_segs = 1;
584 		skb_shinfo(skb)->gso_size = 0;
585 		skb_shinfo(skb)->gso_type = 0;
586 	} else {
587 		unsigned int factor;
588 
589 		factor = skb->len + (mss_now - 1);
590 		factor /= mss_now;
591 		skb_shinfo(skb)->gso_segs = factor;
592 		skb_shinfo(skb)->gso_size = mss_now;
593 		skb_shinfo(skb)->gso_type = sk->sk_gso_type;
594 	}
595 }
596 
597 /* Function to create two new TCP segments.  Shrinks the given segment
598  * to the specified size and appends a new segment with the rest of the
599  * packet to the list.  This won't be called frequently, I hope.
600  * Remember, these are still headerless SKBs at this point.
601  */
602 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
603 {
604 	struct tcp_sock *tp = tcp_sk(sk);
605 	struct sk_buff *buff;
606 	int nsize, old_factor;
607 	int nlen;
608 	u16 flags;
609 
610 	BUG_ON(len > skb->len);
611 
612 	clear_all_retrans_hints(tp);
613 	nsize = skb_headlen(skb) - len;
614 	if (nsize < 0)
615 		nsize = 0;
616 
617 	if (skb_cloned(skb) &&
618 	    skb_is_nonlinear(skb) &&
619 	    pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
620 		return -ENOMEM;
621 
622 	/* Get a new skb... force flag on. */
623 	buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
624 	if (buff == NULL)
625 		return -ENOMEM; /* We'll just try again later. */
626 
627 	sk_charge_skb(sk, buff);
628 	nlen = skb->len - len - nsize;
629 	buff->truesize += nlen;
630 	skb->truesize -= nlen;
631 
632 	/* Correct the sequence numbers. */
633 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
634 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
635 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
636 
637 	/* PSH and FIN should only be set in the second packet. */
638 	flags = TCP_SKB_CB(skb)->flags;
639 	TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
640 	TCP_SKB_CB(buff)->flags = flags;
641 	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
642 	TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
643 
644 	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
645 		/* Copy and checksum data tail into the new buffer. */
646 		buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
647 						       nsize, 0);
648 
649 		skb_trim(skb, len);
650 
651 		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
652 	} else {
653 		skb->ip_summed = CHECKSUM_PARTIAL;
654 		skb_split(skb, buff, len);
655 	}
656 
657 	buff->ip_summed = skb->ip_summed;
658 
659 	/* Looks stupid, but our code really uses when of
660 	 * skbs, which it never sent before. --ANK
661 	 */
662 	TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
663 	buff->tstamp = skb->tstamp;
664 
665 	old_factor = tcp_skb_pcount(skb);
666 
667 	/* Fix up tso_factor for both original and new SKB.  */
668 	tcp_set_skb_tso_segs(sk, skb, mss_now);
669 	tcp_set_skb_tso_segs(sk, buff, mss_now);
670 
671 	/* If this packet has been sent out already, we must
672 	 * adjust the various packet counters.
673 	 */
674 	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
675 		int diff = old_factor - tcp_skb_pcount(skb) -
676 			tcp_skb_pcount(buff);
677 
678 		tp->packets_out -= diff;
679 
680 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
681 			tp->sacked_out -= diff;
682 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
683 			tp->retrans_out -= diff;
684 
685 		if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
686 			tp->lost_out -= diff;
687 			tp->left_out -= diff;
688 		}
689 
690 		if (diff > 0) {
691 			/* Adjust Reno SACK estimate. */
692 			if (!tp->rx_opt.sack_ok) {
693 				tp->sacked_out -= diff;
694 				if ((int)tp->sacked_out < 0)
695 					tp->sacked_out = 0;
696 				tcp_sync_left_out(tp);
697 			}
698 
699 			tp->fackets_out -= diff;
700 			if ((int)tp->fackets_out < 0)
701 				tp->fackets_out = 0;
702 		}
703 	}
704 
705 	/* Link BUFF into the send queue. */
706 	skb_header_release(buff);
707 	tcp_insert_write_queue_after(skb, buff, sk);
708 
709 	return 0;
710 }
711 
712 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
713  * eventually). The difference is that pulled data not copied, but
714  * immediately discarded.
715  */
716 static void __pskb_trim_head(struct sk_buff *skb, int len)
717 {
718 	int i, k, eat;
719 
720 	eat = len;
721 	k = 0;
722 	for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
723 		if (skb_shinfo(skb)->frags[i].size <= eat) {
724 			put_page(skb_shinfo(skb)->frags[i].page);
725 			eat -= skb_shinfo(skb)->frags[i].size;
726 		} else {
727 			skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
728 			if (eat) {
729 				skb_shinfo(skb)->frags[k].page_offset += eat;
730 				skb_shinfo(skb)->frags[k].size -= eat;
731 				eat = 0;
732 			}
733 			k++;
734 		}
735 	}
736 	skb_shinfo(skb)->nr_frags = k;
737 
738 	skb_reset_tail_pointer(skb);
739 	skb->data_len -= len;
740 	skb->len = skb->data_len;
741 }
742 
743 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
744 {
745 	if (skb_cloned(skb) &&
746 	    pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
747 		return -ENOMEM;
748 
749 	/* If len == headlen, we avoid __skb_pull to preserve alignment. */
750 	if (unlikely(len < skb_headlen(skb)))
751 		__skb_pull(skb, len);
752 	else
753 		__pskb_trim_head(skb, len - skb_headlen(skb));
754 
755 	TCP_SKB_CB(skb)->seq += len;
756 	skb->ip_summed = CHECKSUM_PARTIAL;
757 
758 	skb->truesize	     -= len;
759 	sk->sk_wmem_queued   -= len;
760 	sk->sk_forward_alloc += len;
761 	sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
762 
763 	/* Any change of skb->len requires recalculation of tso
764 	 * factor and mss.
765 	 */
766 	if (tcp_skb_pcount(skb) > 1)
767 		tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
768 
769 	return 0;
770 }
771 
772 /* Not accounting for SACKs here. */
773 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
774 {
775 	struct tcp_sock *tp = tcp_sk(sk);
776 	struct inet_connection_sock *icsk = inet_csk(sk);
777 	int mss_now;
778 
779 	/* Calculate base mss without TCP options:
780 	   It is MMS_S - sizeof(tcphdr) of rfc1122
781 	 */
782 	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
783 
784 	/* Clamp it (mss_clamp does not include tcp options) */
785 	if (mss_now > tp->rx_opt.mss_clamp)
786 		mss_now = tp->rx_opt.mss_clamp;
787 
788 	/* Now subtract optional transport overhead */
789 	mss_now -= icsk->icsk_ext_hdr_len;
790 
791 	/* Then reserve room for full set of TCP options and 8 bytes of data */
792 	if (mss_now < 48)
793 		mss_now = 48;
794 
795 	/* Now subtract TCP options size, not including SACKs */
796 	mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
797 
798 	return mss_now;
799 }
800 
801 /* Inverse of above */
802 int tcp_mss_to_mtu(struct sock *sk, int mss)
803 {
804 	struct tcp_sock *tp = tcp_sk(sk);
805 	struct inet_connection_sock *icsk = inet_csk(sk);
806 	int mtu;
807 
808 	mtu = mss +
809 	      tp->tcp_header_len +
810 	      icsk->icsk_ext_hdr_len +
811 	      icsk->icsk_af_ops->net_header_len;
812 
813 	return mtu;
814 }
815 
816 void tcp_mtup_init(struct sock *sk)
817 {
818 	struct tcp_sock *tp = tcp_sk(sk);
819 	struct inet_connection_sock *icsk = inet_csk(sk);
820 
821 	icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
822 	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
823 			       icsk->icsk_af_ops->net_header_len;
824 	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
825 	icsk->icsk_mtup.probe_size = 0;
826 }
827 
828 /* This function synchronize snd mss to current pmtu/exthdr set.
829 
830    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
831    for TCP options, but includes only bare TCP header.
832 
833    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
834    It is minimum of user_mss and mss received with SYN.
835    It also does not include TCP options.
836 
837    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
838 
839    tp->mss_cache is current effective sending mss, including
840    all tcp options except for SACKs. It is evaluated,
841    taking into account current pmtu, but never exceeds
842    tp->rx_opt.mss_clamp.
843 
844    NOTE1. rfc1122 clearly states that advertised MSS
845    DOES NOT include either tcp or ip options.
846 
847    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
848    are READ ONLY outside this function.		--ANK (980731)
849  */
850 
851 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
852 {
853 	struct tcp_sock *tp = tcp_sk(sk);
854 	struct inet_connection_sock *icsk = inet_csk(sk);
855 	int mss_now;
856 
857 	if (icsk->icsk_mtup.search_high > pmtu)
858 		icsk->icsk_mtup.search_high = pmtu;
859 
860 	mss_now = tcp_mtu_to_mss(sk, pmtu);
861 
862 	/* Bound mss with half of window */
863 	if (tp->max_window && mss_now > (tp->max_window>>1))
864 		mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
865 
866 	/* And store cached results */
867 	icsk->icsk_pmtu_cookie = pmtu;
868 	if (icsk->icsk_mtup.enabled)
869 		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
870 	tp->mss_cache = mss_now;
871 
872 	return mss_now;
873 }
874 
875 /* Compute the current effective MSS, taking SACKs and IP options,
876  * and even PMTU discovery events into account.
877  *
878  * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
879  * cannot be large. However, taking into account rare use of URG, this
880  * is not a big flaw.
881  */
882 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
883 {
884 	struct tcp_sock *tp = tcp_sk(sk);
885 	struct dst_entry *dst = __sk_dst_get(sk);
886 	u32 mss_now;
887 	u16 xmit_size_goal;
888 	int doing_tso = 0;
889 
890 	mss_now = tp->mss_cache;
891 
892 	if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
893 		doing_tso = 1;
894 
895 	if (dst) {
896 		u32 mtu = dst_mtu(dst);
897 		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
898 			mss_now = tcp_sync_mss(sk, mtu);
899 	}
900 
901 	if (tp->rx_opt.eff_sacks)
902 		mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
903 			    (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
904 
905 #ifdef CONFIG_TCP_MD5SIG
906 	if (tp->af_specific->md5_lookup(sk, sk))
907 		mss_now -= TCPOLEN_MD5SIG_ALIGNED;
908 #endif
909 
910 	xmit_size_goal = mss_now;
911 
912 	if (doing_tso) {
913 		xmit_size_goal = (65535 -
914 				  inet_csk(sk)->icsk_af_ops->net_header_len -
915 				  inet_csk(sk)->icsk_ext_hdr_len -
916 				  tp->tcp_header_len);
917 
918 		if (tp->max_window &&
919 		    (xmit_size_goal > (tp->max_window >> 1)))
920 			xmit_size_goal = max((tp->max_window >> 1),
921 					     68U - tp->tcp_header_len);
922 
923 		xmit_size_goal -= (xmit_size_goal % mss_now);
924 	}
925 	tp->xmit_size_goal = xmit_size_goal;
926 
927 	return mss_now;
928 }
929 
930 /* Congestion window validation. (RFC2861) */
931 
932 static void tcp_cwnd_validate(struct sock *sk)
933 {
934 	struct tcp_sock *tp = tcp_sk(sk);
935 	__u32 packets_out = tp->packets_out;
936 
937 	if (packets_out >= tp->snd_cwnd) {
938 		/* Network is feed fully. */
939 		tp->snd_cwnd_used = 0;
940 		tp->snd_cwnd_stamp = tcp_time_stamp;
941 	} else {
942 		/* Network starves. */
943 		if (tp->packets_out > tp->snd_cwnd_used)
944 			tp->snd_cwnd_used = tp->packets_out;
945 
946 		if (sysctl_tcp_slow_start_after_idle &&
947 		    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
948 			tcp_cwnd_application_limited(sk);
949 	}
950 }
951 
952 static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
953 {
954 	u32 window, cwnd_len;
955 
956 	window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
957 	cwnd_len = mss_now * cwnd;
958 	return min(window, cwnd_len);
959 }
960 
961 /* Can at least one segment of SKB be sent right now, according to the
962  * congestion window rules?  If so, return how many segments are allowed.
963  */
964 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
965 {
966 	u32 in_flight, cwnd;
967 
968 	/* Don't be strict about the congestion window for the final FIN.  */
969 	if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
970 	    tcp_skb_pcount(skb) == 1)
971 		return 1;
972 
973 	in_flight = tcp_packets_in_flight(tp);
974 	cwnd = tp->snd_cwnd;
975 	if (in_flight < cwnd)
976 		return (cwnd - in_flight);
977 
978 	return 0;
979 }
980 
981 /* This must be invoked the first time we consider transmitting
982  * SKB onto the wire.
983  */
984 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
985 {
986 	int tso_segs = tcp_skb_pcount(skb);
987 
988 	if (!tso_segs ||
989 	    (tso_segs > 1 &&
990 	     tcp_skb_mss(skb) != mss_now)) {
991 		tcp_set_skb_tso_segs(sk, skb, mss_now);
992 		tso_segs = tcp_skb_pcount(skb);
993 	}
994 	return tso_segs;
995 }
996 
997 static inline int tcp_minshall_check(const struct tcp_sock *tp)
998 {
999 	return after(tp->snd_sml,tp->snd_una) &&
1000 		!after(tp->snd_sml, tp->snd_nxt);
1001 }
1002 
1003 /* Return 0, if packet can be sent now without violation Nagle's rules:
1004  * 1. It is full sized.
1005  * 2. Or it contains FIN. (already checked by caller)
1006  * 3. Or TCP_NODELAY was set.
1007  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1008  *    With Minshall's modification: all sent small packets are ACKed.
1009  */
1010 
1011 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1012 				  const struct sk_buff *skb,
1013 				  unsigned mss_now, int nonagle)
1014 {
1015 	return (skb->len < mss_now &&
1016 		((nonagle&TCP_NAGLE_CORK) ||
1017 		 (!nonagle &&
1018 		  tp->packets_out &&
1019 		  tcp_minshall_check(tp))));
1020 }
1021 
1022 /* Return non-zero if the Nagle test allows this packet to be
1023  * sent now.
1024  */
1025 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1026 				 unsigned int cur_mss, int nonagle)
1027 {
1028 	/* Nagle rule does not apply to frames, which sit in the middle of the
1029 	 * write_queue (they have no chances to get new data).
1030 	 *
1031 	 * This is implemented in the callers, where they modify the 'nonagle'
1032 	 * argument based upon the location of SKB in the send queue.
1033 	 */
1034 	if (nonagle & TCP_NAGLE_PUSH)
1035 		return 1;
1036 
1037 	/* Don't use the nagle rule for urgent data (or for the final FIN).
1038 	 * Nagle can be ignored during F-RTO too (see RFC4138).
1039 	 */
1040 	if (tp->urg_mode || (tp->frto_counter == 2) ||
1041 	    (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1042 		return 1;
1043 
1044 	if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1045 		return 1;
1046 
1047 	return 0;
1048 }
1049 
1050 /* Does at least the first segment of SKB fit into the send window? */
1051 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
1052 {
1053 	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1054 
1055 	if (skb->len > cur_mss)
1056 		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1057 
1058 	return !after(end_seq, tp->snd_una + tp->snd_wnd);
1059 }
1060 
1061 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1062  * should be put on the wire right now.  If so, it returns the number of
1063  * packets allowed by the congestion window.
1064  */
1065 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1066 				 unsigned int cur_mss, int nonagle)
1067 {
1068 	struct tcp_sock *tp = tcp_sk(sk);
1069 	unsigned int cwnd_quota;
1070 
1071 	tcp_init_tso_segs(sk, skb, cur_mss);
1072 
1073 	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1074 		return 0;
1075 
1076 	cwnd_quota = tcp_cwnd_test(tp, skb);
1077 	if (cwnd_quota &&
1078 	    !tcp_snd_wnd_test(tp, skb, cur_mss))
1079 		cwnd_quota = 0;
1080 
1081 	return cwnd_quota;
1082 }
1083 
1084 int tcp_may_send_now(struct sock *sk)
1085 {
1086 	struct tcp_sock *tp = tcp_sk(sk);
1087 	struct sk_buff *skb = tcp_send_head(sk);
1088 
1089 	return (skb &&
1090 		tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1091 			     (tcp_skb_is_last(sk, skb) ?
1092 			      TCP_NAGLE_PUSH :
1093 			      tp->nonagle)));
1094 }
1095 
1096 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1097  * which is put after SKB on the list.  It is very much like
1098  * tcp_fragment() except that it may make several kinds of assumptions
1099  * in order to speed up the splitting operation.  In particular, we
1100  * know that all the data is in scatter-gather pages, and that the
1101  * packet has never been sent out before (and thus is not cloned).
1102  */
1103 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
1104 {
1105 	struct sk_buff *buff;
1106 	int nlen = skb->len - len;
1107 	u16 flags;
1108 
1109 	/* All of a TSO frame must be composed of paged data.  */
1110 	if (skb->len != skb->data_len)
1111 		return tcp_fragment(sk, skb, len, mss_now);
1112 
1113 	buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
1114 	if (unlikely(buff == NULL))
1115 		return -ENOMEM;
1116 
1117 	sk_charge_skb(sk, buff);
1118 	buff->truesize += nlen;
1119 	skb->truesize -= nlen;
1120 
1121 	/* Correct the sequence numbers. */
1122 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1123 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1124 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1125 
1126 	/* PSH and FIN should only be set in the second packet. */
1127 	flags = TCP_SKB_CB(skb)->flags;
1128 	TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1129 	TCP_SKB_CB(buff)->flags = flags;
1130 
1131 	/* This packet was never sent out yet, so no SACK bits. */
1132 	TCP_SKB_CB(buff)->sacked = 0;
1133 
1134 	buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1135 	skb_split(skb, buff, len);
1136 
1137 	/* Fix up tso_factor for both original and new SKB.  */
1138 	tcp_set_skb_tso_segs(sk, skb, mss_now);
1139 	tcp_set_skb_tso_segs(sk, buff, mss_now);
1140 
1141 	/* Link BUFF into the send queue. */
1142 	skb_header_release(buff);
1143 	tcp_insert_write_queue_after(skb, buff, sk);
1144 
1145 	return 0;
1146 }
1147 
1148 /* Try to defer sending, if possible, in order to minimize the amount
1149  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1150  *
1151  * This algorithm is from John Heffner.
1152  */
1153 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1154 {
1155 	struct tcp_sock *tp = tcp_sk(sk);
1156 	const struct inet_connection_sock *icsk = inet_csk(sk);
1157 	u32 send_win, cong_win, limit, in_flight;
1158 
1159 	if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1160 		goto send_now;
1161 
1162 	if (icsk->icsk_ca_state != TCP_CA_Open)
1163 		goto send_now;
1164 
1165 	/* Defer for less than two clock ticks. */
1166 	if (!tp->tso_deferred && ((jiffies<<1)>>1) - (tp->tso_deferred>>1) > 1)
1167 		goto send_now;
1168 
1169 	in_flight = tcp_packets_in_flight(tp);
1170 
1171 	BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1172 	       (tp->snd_cwnd <= in_flight));
1173 
1174 	send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
1175 
1176 	/* From in_flight test above, we know that cwnd > in_flight.  */
1177 	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1178 
1179 	limit = min(send_win, cong_win);
1180 
1181 	/* If a full-sized TSO skb can be sent, do it. */
1182 	if (limit >= 65536)
1183 		goto send_now;
1184 
1185 	if (sysctl_tcp_tso_win_divisor) {
1186 		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1187 
1188 		/* If at least some fraction of a window is available,
1189 		 * just use it.
1190 		 */
1191 		chunk /= sysctl_tcp_tso_win_divisor;
1192 		if (limit >= chunk)
1193 			goto send_now;
1194 	} else {
1195 		/* Different approach, try not to defer past a single
1196 		 * ACK.  Receiver should ACK every other full sized
1197 		 * frame, so if we have space for more than 3 frames
1198 		 * then send now.
1199 		 */
1200 		if (limit > tcp_max_burst(tp) * tp->mss_cache)
1201 			goto send_now;
1202 	}
1203 
1204 	/* Ok, it looks like it is advisable to defer.  */
1205 	tp->tso_deferred = 1 | (jiffies<<1);
1206 
1207 	return 1;
1208 
1209 send_now:
1210 	tp->tso_deferred = 0;
1211 	return 0;
1212 }
1213 
1214 /* Create a new MTU probe if we are ready.
1215  * Returns 0 if we should wait to probe (no cwnd available),
1216  *         1 if a probe was sent,
1217  *         -1 otherwise */
1218 static int tcp_mtu_probe(struct sock *sk)
1219 {
1220 	struct tcp_sock *tp = tcp_sk(sk);
1221 	struct inet_connection_sock *icsk = inet_csk(sk);
1222 	struct sk_buff *skb, *nskb, *next;
1223 	int len;
1224 	int probe_size;
1225 	unsigned int pif;
1226 	int copy;
1227 	int mss_now;
1228 
1229 	/* Not currently probing/verifying,
1230 	 * not in recovery,
1231 	 * have enough cwnd, and
1232 	 * not SACKing (the variable headers throw things off) */
1233 	if (!icsk->icsk_mtup.enabled ||
1234 	    icsk->icsk_mtup.probe_size ||
1235 	    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1236 	    tp->snd_cwnd < 11 ||
1237 	    tp->rx_opt.eff_sacks)
1238 		return -1;
1239 
1240 	/* Very simple search strategy: just double the MSS. */
1241 	mss_now = tcp_current_mss(sk, 0);
1242 	probe_size = 2*tp->mss_cache;
1243 	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1244 		/* TODO: set timer for probe_converge_event */
1245 		return -1;
1246 	}
1247 
1248 	/* Have enough data in the send queue to probe? */
1249 	len = 0;
1250 	if ((skb = tcp_send_head(sk)) == NULL)
1251 		return -1;
1252 	while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb))
1253 		skb = tcp_write_queue_next(sk, skb);
1254 	if (len < probe_size)
1255 		return -1;
1256 
1257 	/* Receive window check. */
1258 	if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) {
1259 		if (tp->snd_wnd < probe_size)
1260 			return -1;
1261 		else
1262 			return 0;
1263 	}
1264 
1265 	/* Do we need to wait to drain cwnd? */
1266 	pif = tcp_packets_in_flight(tp);
1267 	if (pif + 2 > tp->snd_cwnd) {
1268 		/* With no packets in flight, don't stall. */
1269 		if (pif == 0)
1270 			return -1;
1271 		else
1272 			return 0;
1273 	}
1274 
1275 	/* We're allowed to probe.  Build it now. */
1276 	if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1277 		return -1;
1278 	sk_charge_skb(sk, nskb);
1279 
1280 	skb = tcp_send_head(sk);
1281 	tcp_insert_write_queue_before(nskb, skb, sk);
1282 	tcp_advance_send_head(sk, skb);
1283 
1284 	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1285 	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1286 	TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1287 	TCP_SKB_CB(nskb)->sacked = 0;
1288 	nskb->csum = 0;
1289 	nskb->ip_summed = skb->ip_summed;
1290 
1291 	len = 0;
1292 	while (len < probe_size) {
1293 		next = tcp_write_queue_next(sk, skb);
1294 
1295 		copy = min_t(int, skb->len, probe_size - len);
1296 		if (nskb->ip_summed)
1297 			skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1298 		else
1299 			nskb->csum = skb_copy_and_csum_bits(skb, 0,
1300 					 skb_put(nskb, copy), copy, nskb->csum);
1301 
1302 		if (skb->len <= copy) {
1303 			/* We've eaten all the data from this skb.
1304 			 * Throw it away. */
1305 			TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1306 			tcp_unlink_write_queue(skb, sk);
1307 			sk_stream_free_skb(sk, skb);
1308 		} else {
1309 			TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1310 						   ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1311 			if (!skb_shinfo(skb)->nr_frags) {
1312 				skb_pull(skb, copy);
1313 				if (skb->ip_summed != CHECKSUM_PARTIAL)
1314 					skb->csum = csum_partial(skb->data, skb->len, 0);
1315 			} else {
1316 				__pskb_trim_head(skb, copy);
1317 				tcp_set_skb_tso_segs(sk, skb, mss_now);
1318 			}
1319 			TCP_SKB_CB(skb)->seq += copy;
1320 		}
1321 
1322 		len += copy;
1323 		skb = next;
1324 	}
1325 	tcp_init_tso_segs(sk, nskb, nskb->len);
1326 
1327 	/* We're ready to send.  If this fails, the probe will
1328 	 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1329 	TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1330 	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1331 		/* Decrement cwnd here because we are sending
1332 		* effectively two packets. */
1333 		tp->snd_cwnd--;
1334 		update_send_head(sk, nskb);
1335 
1336 		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1337 		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1338 		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1339 
1340 		return 1;
1341 	}
1342 
1343 	return -1;
1344 }
1345 
1346 
1347 /* This routine writes packets to the network.  It advances the
1348  * send_head.  This happens as incoming acks open up the remote
1349  * window for us.
1350  *
1351  * Returns 1, if no segments are in flight and we have queued segments, but
1352  * cannot send anything now because of SWS or another problem.
1353  */
1354 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1355 {
1356 	struct tcp_sock *tp = tcp_sk(sk);
1357 	struct sk_buff *skb;
1358 	unsigned int tso_segs, sent_pkts;
1359 	int cwnd_quota;
1360 	int result;
1361 
1362 	/* If we are closed, the bytes will have to remain here.
1363 	 * In time closedown will finish, we empty the write queue and all
1364 	 * will be happy.
1365 	 */
1366 	if (unlikely(sk->sk_state == TCP_CLOSE))
1367 		return 0;
1368 
1369 	sent_pkts = 0;
1370 
1371 	/* Do MTU probing. */
1372 	if ((result = tcp_mtu_probe(sk)) == 0) {
1373 		return 0;
1374 	} else if (result > 0) {
1375 		sent_pkts = 1;
1376 	}
1377 
1378 	while ((skb = tcp_send_head(sk))) {
1379 		unsigned int limit;
1380 
1381 		tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1382 		BUG_ON(!tso_segs);
1383 
1384 		cwnd_quota = tcp_cwnd_test(tp, skb);
1385 		if (!cwnd_quota)
1386 			break;
1387 
1388 		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1389 			break;
1390 
1391 		if (tso_segs == 1) {
1392 			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1393 						     (tcp_skb_is_last(sk, skb) ?
1394 						      nonagle : TCP_NAGLE_PUSH))))
1395 				break;
1396 		} else {
1397 			if (tcp_tso_should_defer(sk, skb))
1398 				break;
1399 		}
1400 
1401 		limit = mss_now;
1402 		if (tso_segs > 1) {
1403 			limit = tcp_window_allows(tp, skb,
1404 						  mss_now, cwnd_quota);
1405 
1406 			if (skb->len < limit) {
1407 				unsigned int trim = skb->len % mss_now;
1408 
1409 				if (trim)
1410 					limit = skb->len - trim;
1411 			}
1412 		}
1413 
1414 		if (skb->len > limit &&
1415 		    unlikely(tso_fragment(sk, skb, limit, mss_now)))
1416 			break;
1417 
1418 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
1419 
1420 		if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1421 			break;
1422 
1423 		/* Advance the send_head.  This one is sent out.
1424 		 * This call will increment packets_out.
1425 		 */
1426 		update_send_head(sk, skb);
1427 
1428 		tcp_minshall_update(tp, mss_now, skb);
1429 		sent_pkts++;
1430 	}
1431 
1432 	if (likely(sent_pkts)) {
1433 		tcp_cwnd_validate(sk);
1434 		return 0;
1435 	}
1436 	return !tp->packets_out && tcp_send_head(sk);
1437 }
1438 
1439 /* Push out any pending frames which were held back due to
1440  * TCP_CORK or attempt at coalescing tiny packets.
1441  * The socket must be locked by the caller.
1442  */
1443 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1444 			       int nonagle)
1445 {
1446 	struct sk_buff *skb = tcp_send_head(sk);
1447 
1448 	if (skb) {
1449 		if (tcp_write_xmit(sk, cur_mss, nonagle))
1450 			tcp_check_probe_timer(sk);
1451 	}
1452 }
1453 
1454 /* Send _single_ skb sitting at the send head. This function requires
1455  * true push pending frames to setup probe timer etc.
1456  */
1457 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1458 {
1459 	struct tcp_sock *tp = tcp_sk(sk);
1460 	struct sk_buff *skb = tcp_send_head(sk);
1461 	unsigned int tso_segs, cwnd_quota;
1462 
1463 	BUG_ON(!skb || skb->len < mss_now);
1464 
1465 	tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1466 	cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1467 
1468 	if (likely(cwnd_quota)) {
1469 		unsigned int limit;
1470 
1471 		BUG_ON(!tso_segs);
1472 
1473 		limit = mss_now;
1474 		if (tso_segs > 1) {
1475 			limit = tcp_window_allows(tp, skb,
1476 						  mss_now, cwnd_quota);
1477 
1478 			if (skb->len < limit) {
1479 				unsigned int trim = skb->len % mss_now;
1480 
1481 				if (trim)
1482 					limit = skb->len - trim;
1483 			}
1484 		}
1485 
1486 		if (skb->len > limit &&
1487 		    unlikely(tso_fragment(sk, skb, limit, mss_now)))
1488 			return;
1489 
1490 		/* Send it out now. */
1491 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
1492 
1493 		if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1494 			update_send_head(sk, skb);
1495 			tcp_cwnd_validate(sk);
1496 			return;
1497 		}
1498 	}
1499 }
1500 
1501 /* This function returns the amount that we can raise the
1502  * usable window based on the following constraints
1503  *
1504  * 1. The window can never be shrunk once it is offered (RFC 793)
1505  * 2. We limit memory per socket
1506  *
1507  * RFC 1122:
1508  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1509  *  RECV.NEXT + RCV.WIN fixed until:
1510  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1511  *
1512  * i.e. don't raise the right edge of the window until you can raise
1513  * it at least MSS bytes.
1514  *
1515  * Unfortunately, the recommended algorithm breaks header prediction,
1516  * since header prediction assumes th->window stays fixed.
1517  *
1518  * Strictly speaking, keeping th->window fixed violates the receiver
1519  * side SWS prevention criteria. The problem is that under this rule
1520  * a stream of single byte packets will cause the right side of the
1521  * window to always advance by a single byte.
1522  *
1523  * Of course, if the sender implements sender side SWS prevention
1524  * then this will not be a problem.
1525  *
1526  * BSD seems to make the following compromise:
1527  *
1528  *	If the free space is less than the 1/4 of the maximum
1529  *	space available and the free space is less than 1/2 mss,
1530  *	then set the window to 0.
1531  *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1532  *	Otherwise, just prevent the window from shrinking
1533  *	and from being larger than the largest representable value.
1534  *
1535  * This prevents incremental opening of the window in the regime
1536  * where TCP is limited by the speed of the reader side taking
1537  * data out of the TCP receive queue. It does nothing about
1538  * those cases where the window is constrained on the sender side
1539  * because the pipeline is full.
1540  *
1541  * BSD also seems to "accidentally" limit itself to windows that are a
1542  * multiple of MSS, at least until the free space gets quite small.
1543  * This would appear to be a side effect of the mbuf implementation.
1544  * Combining these two algorithms results in the observed behavior
1545  * of having a fixed window size at almost all times.
1546  *
1547  * Below we obtain similar behavior by forcing the offered window to
1548  * a multiple of the mss when it is feasible to do so.
1549  *
1550  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1551  * Regular options like TIMESTAMP are taken into account.
1552  */
1553 u32 __tcp_select_window(struct sock *sk)
1554 {
1555 	struct inet_connection_sock *icsk = inet_csk(sk);
1556 	struct tcp_sock *tp = tcp_sk(sk);
1557 	/* MSS for the peer's data.  Previous versions used mss_clamp
1558 	 * here.  I don't know if the value based on our guesses
1559 	 * of peer's MSS is better for the performance.  It's more correct
1560 	 * but may be worse for the performance because of rcv_mss
1561 	 * fluctuations.  --SAW  1998/11/1
1562 	 */
1563 	int mss = icsk->icsk_ack.rcv_mss;
1564 	int free_space = tcp_space(sk);
1565 	int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1566 	int window;
1567 
1568 	if (mss > full_space)
1569 		mss = full_space;
1570 
1571 	if (free_space < full_space/2) {
1572 		icsk->icsk_ack.quick = 0;
1573 
1574 		if (tcp_memory_pressure)
1575 			tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1576 
1577 		if (free_space < mss)
1578 			return 0;
1579 	}
1580 
1581 	if (free_space > tp->rcv_ssthresh)
1582 		free_space = tp->rcv_ssthresh;
1583 
1584 	/* Don't do rounding if we are using window scaling, since the
1585 	 * scaled window will not line up with the MSS boundary anyway.
1586 	 */
1587 	window = tp->rcv_wnd;
1588 	if (tp->rx_opt.rcv_wscale) {
1589 		window = free_space;
1590 
1591 		/* Advertise enough space so that it won't get scaled away.
1592 		 * Import case: prevent zero window announcement if
1593 		 * 1<<rcv_wscale > mss.
1594 		 */
1595 		if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1596 			window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1597 				  << tp->rx_opt.rcv_wscale);
1598 	} else {
1599 		/* Get the largest window that is a nice multiple of mss.
1600 		 * Window clamp already applied above.
1601 		 * If our current window offering is within 1 mss of the
1602 		 * free space we just keep it. This prevents the divide
1603 		 * and multiply from happening most of the time.
1604 		 * We also don't do any window rounding when the free space
1605 		 * is too small.
1606 		 */
1607 		if (window <= free_space - mss || window > free_space)
1608 			window = (free_space/mss)*mss;
1609 		else if (mss == full_space &&
1610 		         free_space > window + full_space/2)
1611 			window = free_space;
1612 	}
1613 
1614 	return window;
1615 }
1616 
1617 /* Attempt to collapse two adjacent SKB's during retransmission. */
1618 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1619 {
1620 	struct tcp_sock *tp = tcp_sk(sk);
1621 	struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1622 
1623 	/* The first test we must make is that neither of these two
1624 	 * SKB's are still referenced by someone else.
1625 	 */
1626 	if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1627 		int skb_size = skb->len, next_skb_size = next_skb->len;
1628 		u16 flags = TCP_SKB_CB(skb)->flags;
1629 
1630 		/* Also punt if next skb has been SACK'd. */
1631 		if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1632 			return;
1633 
1634 		/* Next skb is out of window. */
1635 		if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1636 			return;
1637 
1638 		/* Punt if not enough space exists in the first SKB for
1639 		 * the data in the second, or the total combined payload
1640 		 * would exceed the MSS.
1641 		 */
1642 		if ((next_skb_size > skb_tailroom(skb)) ||
1643 		    ((skb_size + next_skb_size) > mss_now))
1644 			return;
1645 
1646 		BUG_ON(tcp_skb_pcount(skb) != 1 ||
1647 		       tcp_skb_pcount(next_skb) != 1);
1648 
1649 		/* changing transmit queue under us so clear hints */
1650 		clear_all_retrans_hints(tp);
1651 
1652 		/* Ok.	We will be able to collapse the packet. */
1653 		tcp_unlink_write_queue(next_skb, sk);
1654 
1655 		skb_copy_from_linear_data(next_skb,
1656 					  skb_put(skb, next_skb_size),
1657 					  next_skb_size);
1658 
1659 		if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1660 			skb->ip_summed = CHECKSUM_PARTIAL;
1661 
1662 		if (skb->ip_summed != CHECKSUM_PARTIAL)
1663 			skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1664 
1665 		/* Update sequence range on original skb. */
1666 		TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1667 
1668 		/* Merge over control information. */
1669 		flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1670 		TCP_SKB_CB(skb)->flags = flags;
1671 
1672 		/* All done, get rid of second SKB and account for it so
1673 		 * packet counting does not break.
1674 		 */
1675 		TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1676 		if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1677 			tp->retrans_out -= tcp_skb_pcount(next_skb);
1678 		if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1679 			tp->lost_out -= tcp_skb_pcount(next_skb);
1680 			tp->left_out -= tcp_skb_pcount(next_skb);
1681 		}
1682 		/* Reno case is special. Sigh... */
1683 		if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1684 			tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1685 			tp->left_out -= tcp_skb_pcount(next_skb);
1686 		}
1687 
1688 		/* Not quite right: it can be > snd.fack, but
1689 		 * it is better to underestimate fackets.
1690 		 */
1691 		tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1692 		tcp_packets_out_dec(tp, next_skb);
1693 		sk_stream_free_skb(sk, next_skb);
1694 	}
1695 }
1696 
1697 /* Do a simple retransmit without using the backoff mechanisms in
1698  * tcp_timer. This is used for path mtu discovery.
1699  * The socket is already locked here.
1700  */
1701 void tcp_simple_retransmit(struct sock *sk)
1702 {
1703 	const struct inet_connection_sock *icsk = inet_csk(sk);
1704 	struct tcp_sock *tp = tcp_sk(sk);
1705 	struct sk_buff *skb;
1706 	unsigned int mss = tcp_current_mss(sk, 0);
1707 	int lost = 0;
1708 
1709 	tcp_for_write_queue(skb, sk) {
1710 		if (skb == tcp_send_head(sk))
1711 			break;
1712 		if (skb->len > mss &&
1713 		    !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1714 			if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1715 				TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1716 				tp->retrans_out -= tcp_skb_pcount(skb);
1717 			}
1718 			if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1719 				TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1720 				tp->lost_out += tcp_skb_pcount(skb);
1721 				lost = 1;
1722 			}
1723 		}
1724 	}
1725 
1726 	clear_all_retrans_hints(tp);
1727 
1728 	if (!lost)
1729 		return;
1730 
1731 	tcp_sync_left_out(tp);
1732 
1733 	/* Don't muck with the congestion window here.
1734 	 * Reason is that we do not increase amount of _data_
1735 	 * in network, but units changed and effective
1736 	 * cwnd/ssthresh really reduced now.
1737 	 */
1738 	if (icsk->icsk_ca_state != TCP_CA_Loss) {
1739 		tp->high_seq = tp->snd_nxt;
1740 		tp->snd_ssthresh = tcp_current_ssthresh(sk);
1741 		tp->prior_ssthresh = 0;
1742 		tp->undo_marker = 0;
1743 		tcp_set_ca_state(sk, TCP_CA_Loss);
1744 	}
1745 	tcp_xmit_retransmit_queue(sk);
1746 }
1747 
1748 /* This retransmits one SKB.  Policy decisions and retransmit queue
1749  * state updates are done by the caller.  Returns non-zero if an
1750  * error occurred which prevented the send.
1751  */
1752 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1753 {
1754 	struct tcp_sock *tp = tcp_sk(sk);
1755 	struct inet_connection_sock *icsk = inet_csk(sk);
1756 	unsigned int cur_mss = tcp_current_mss(sk, 0);
1757 	int err;
1758 
1759 	/* Inconslusive MTU probe */
1760 	if (icsk->icsk_mtup.probe_size) {
1761 		icsk->icsk_mtup.probe_size = 0;
1762 	}
1763 
1764 	/* Do not sent more than we queued. 1/4 is reserved for possible
1765 	 * copying overhead: fragmentation, tunneling, mangling etc.
1766 	 */
1767 	if (atomic_read(&sk->sk_wmem_alloc) >
1768 	    min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1769 		return -EAGAIN;
1770 
1771 	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1772 		if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1773 			BUG();
1774 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1775 			return -ENOMEM;
1776 	}
1777 
1778 	/* If receiver has shrunk his window, and skb is out of
1779 	 * new window, do not retransmit it. The exception is the
1780 	 * case, when window is shrunk to zero. In this case
1781 	 * our retransmit serves as a zero window probe.
1782 	 */
1783 	if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1784 	    && TCP_SKB_CB(skb)->seq != tp->snd_una)
1785 		return -EAGAIN;
1786 
1787 	if (skb->len > cur_mss) {
1788 		if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1789 			return -ENOMEM; /* We'll try again later. */
1790 	}
1791 
1792 	/* Collapse two adjacent packets if worthwhile and we can. */
1793 	if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1794 	    (skb->len < (cur_mss >> 1)) &&
1795 	    (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1796 	    (!tcp_skb_is_last(sk, skb)) &&
1797 	    (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1798 	    (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
1799 	    (sysctl_tcp_retrans_collapse != 0))
1800 		tcp_retrans_try_collapse(sk, skb, cur_mss);
1801 
1802 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1803 		return -EHOSTUNREACH; /* Routing failure or similar. */
1804 
1805 	/* Some Solaris stacks overoptimize and ignore the FIN on a
1806 	 * retransmit when old data is attached.  So strip it off
1807 	 * since it is cheap to do so and saves bytes on the network.
1808 	 */
1809 	if (skb->len > 0 &&
1810 	    (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1811 	    tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1812 		if (!pskb_trim(skb, 0)) {
1813 			TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1814 			skb_shinfo(skb)->gso_segs = 1;
1815 			skb_shinfo(skb)->gso_size = 0;
1816 			skb_shinfo(skb)->gso_type = 0;
1817 			skb->ip_summed = CHECKSUM_NONE;
1818 			skb->csum = 0;
1819 		}
1820 	}
1821 
1822 	/* Make a copy, if the first transmission SKB clone we made
1823 	 * is still in somebody's hands, else make a clone.
1824 	 */
1825 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
1826 
1827 	err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1828 
1829 	if (err == 0) {
1830 		/* Update global TCP statistics. */
1831 		TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1832 
1833 		tp->total_retrans++;
1834 
1835 #if FASTRETRANS_DEBUG > 0
1836 		if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1837 			if (net_ratelimit())
1838 				printk(KERN_DEBUG "retrans_out leaked.\n");
1839 		}
1840 #endif
1841 		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1842 		tp->retrans_out += tcp_skb_pcount(skb);
1843 
1844 		/* Save stamp of the first retransmit. */
1845 		if (!tp->retrans_stamp)
1846 			tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1847 
1848 		tp->undo_retrans++;
1849 
1850 		/* snd_nxt is stored to detect loss of retransmitted segment,
1851 		 * see tcp_input.c tcp_sacktag_write_queue().
1852 		 */
1853 		TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1854 	}
1855 	return err;
1856 }
1857 
1858 /* This gets called after a retransmit timeout, and the initially
1859  * retransmitted data is acknowledged.  It tries to continue
1860  * resending the rest of the retransmit queue, until either
1861  * we've sent it all or the congestion window limit is reached.
1862  * If doing SACK, the first ACK which comes back for a timeout
1863  * based retransmit packet might feed us FACK information again.
1864  * If so, we use it to avoid unnecessarily retransmissions.
1865  */
1866 void tcp_xmit_retransmit_queue(struct sock *sk)
1867 {
1868 	const struct inet_connection_sock *icsk = inet_csk(sk);
1869 	struct tcp_sock *tp = tcp_sk(sk);
1870 	struct sk_buff *skb;
1871 	int packet_cnt;
1872 
1873 	if (tp->retransmit_skb_hint) {
1874 		skb = tp->retransmit_skb_hint;
1875 		packet_cnt = tp->retransmit_cnt_hint;
1876 	}else{
1877 		skb = tcp_write_queue_head(sk);
1878 		packet_cnt = 0;
1879 	}
1880 
1881 	/* First pass: retransmit lost packets. */
1882 	if (tp->lost_out) {
1883 		tcp_for_write_queue_from(skb, sk) {
1884 			__u8 sacked = TCP_SKB_CB(skb)->sacked;
1885 
1886 			if (skb == tcp_send_head(sk))
1887 				break;
1888 			/* we could do better than to assign each time */
1889 			tp->retransmit_skb_hint = skb;
1890 			tp->retransmit_cnt_hint = packet_cnt;
1891 
1892 			/* Assume this retransmit will generate
1893 			 * only one packet for congestion window
1894 			 * calculation purposes.  This works because
1895 			 * tcp_retransmit_skb() will chop up the
1896 			 * packet to be MSS sized and all the
1897 			 * packet counting works out.
1898 			 */
1899 			if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1900 				return;
1901 
1902 			if (sacked & TCPCB_LOST) {
1903 				if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1904 					if (tcp_retransmit_skb(sk, skb)) {
1905 						tp->retransmit_skb_hint = NULL;
1906 						return;
1907 					}
1908 					if (icsk->icsk_ca_state != TCP_CA_Loss)
1909 						NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1910 					else
1911 						NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1912 
1913 					if (skb == tcp_write_queue_head(sk))
1914 						inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1915 									  inet_csk(sk)->icsk_rto,
1916 									  TCP_RTO_MAX);
1917 				}
1918 
1919 				packet_cnt += tcp_skb_pcount(skb);
1920 				if (packet_cnt >= tp->lost_out)
1921 					break;
1922 			}
1923 		}
1924 	}
1925 
1926 	/* OK, demanded retransmission is finished. */
1927 
1928 	/* Forward retransmissions are possible only during Recovery. */
1929 	if (icsk->icsk_ca_state != TCP_CA_Recovery)
1930 		return;
1931 
1932 	/* No forward retransmissions in Reno are possible. */
1933 	if (!tp->rx_opt.sack_ok)
1934 		return;
1935 
1936 	/* Yeah, we have to make difficult choice between forward transmission
1937 	 * and retransmission... Both ways have their merits...
1938 	 *
1939 	 * For now we do not retransmit anything, while we have some new
1940 	 * segments to send.
1941 	 */
1942 
1943 	if (tcp_may_send_now(sk))
1944 		return;
1945 
1946 	if (tp->forward_skb_hint) {
1947 		skb = tp->forward_skb_hint;
1948 		packet_cnt = tp->forward_cnt_hint;
1949 	} else{
1950 		skb = tcp_write_queue_head(sk);
1951 		packet_cnt = 0;
1952 	}
1953 
1954 	tcp_for_write_queue_from(skb, sk) {
1955 		if (skb == tcp_send_head(sk))
1956 			break;
1957 		tp->forward_cnt_hint = packet_cnt;
1958 		tp->forward_skb_hint = skb;
1959 
1960 		/* Similar to the retransmit loop above we
1961 		 * can pretend that the retransmitted SKB
1962 		 * we send out here will be composed of one
1963 		 * real MSS sized packet because tcp_retransmit_skb()
1964 		 * will fragment it if necessary.
1965 		 */
1966 		if (++packet_cnt > tp->fackets_out)
1967 			break;
1968 
1969 		if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1970 			break;
1971 
1972 		if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1973 			continue;
1974 
1975 		/* Ok, retransmit it. */
1976 		if (tcp_retransmit_skb(sk, skb)) {
1977 			tp->forward_skb_hint = NULL;
1978 			break;
1979 		}
1980 
1981 		if (skb == tcp_write_queue_head(sk))
1982 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1983 						  inet_csk(sk)->icsk_rto,
1984 						  TCP_RTO_MAX);
1985 
1986 		NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1987 	}
1988 }
1989 
1990 
1991 /* Send a fin.  The caller locks the socket for us.  This cannot be
1992  * allowed to fail queueing a FIN frame under any circumstances.
1993  */
1994 void tcp_send_fin(struct sock *sk)
1995 {
1996 	struct tcp_sock *tp = tcp_sk(sk);
1997 	struct sk_buff *skb = tcp_write_queue_tail(sk);
1998 	int mss_now;
1999 
2000 	/* Optimization, tack on the FIN if we have a queue of
2001 	 * unsent frames.  But be careful about outgoing SACKS
2002 	 * and IP options.
2003 	 */
2004 	mss_now = tcp_current_mss(sk, 1);
2005 
2006 	if (tcp_send_head(sk) != NULL) {
2007 		TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2008 		TCP_SKB_CB(skb)->end_seq++;
2009 		tp->write_seq++;
2010 	} else {
2011 		/* Socket is locked, keep trying until memory is available. */
2012 		for (;;) {
2013 			skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
2014 			if (skb)
2015 				break;
2016 			yield();
2017 		}
2018 
2019 		/* Reserve space for headers and prepare control bits. */
2020 		skb_reserve(skb, MAX_TCP_HEADER);
2021 		skb->csum = 0;
2022 		TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2023 		TCP_SKB_CB(skb)->sacked = 0;
2024 		skb_shinfo(skb)->gso_segs = 1;
2025 		skb_shinfo(skb)->gso_size = 0;
2026 		skb_shinfo(skb)->gso_type = 0;
2027 
2028 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2029 		TCP_SKB_CB(skb)->seq = tp->write_seq;
2030 		TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2031 		tcp_queue_skb(sk, skb);
2032 	}
2033 	__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2034 }
2035 
2036 /* We get here when a process closes a file descriptor (either due to
2037  * an explicit close() or as a byproduct of exit()'ing) and there
2038  * was unread data in the receive queue.  This behavior is recommended
2039  * by RFC 2525, section 2.17.  -DaveM
2040  */
2041 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2042 {
2043 	struct sk_buff *skb;
2044 
2045 	/* NOTE: No TCP options attached and we never retransmit this. */
2046 	skb = alloc_skb(MAX_TCP_HEADER, priority);
2047 	if (!skb) {
2048 		NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2049 		return;
2050 	}
2051 
2052 	/* Reserve space for headers and prepare control bits. */
2053 	skb_reserve(skb, MAX_TCP_HEADER);
2054 	skb->csum = 0;
2055 	TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2056 	TCP_SKB_CB(skb)->sacked = 0;
2057 	skb_shinfo(skb)->gso_segs = 1;
2058 	skb_shinfo(skb)->gso_size = 0;
2059 	skb_shinfo(skb)->gso_type = 0;
2060 
2061 	/* Send it off. */
2062 	TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk);
2063 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2064 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2065 	if (tcp_transmit_skb(sk, skb, 0, priority))
2066 		NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2067 }
2068 
2069 /* WARNING: This routine must only be called when we have already sent
2070  * a SYN packet that crossed the incoming SYN that caused this routine
2071  * to get called. If this assumption fails then the initial rcv_wnd
2072  * and rcv_wscale values will not be correct.
2073  */
2074 int tcp_send_synack(struct sock *sk)
2075 {
2076 	struct sk_buff* skb;
2077 
2078 	skb = tcp_write_queue_head(sk);
2079 	if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
2080 		printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2081 		return -EFAULT;
2082 	}
2083 	if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
2084 		if (skb_cloned(skb)) {
2085 			struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2086 			if (nskb == NULL)
2087 				return -ENOMEM;
2088 			tcp_unlink_write_queue(skb, sk);
2089 			skb_header_release(nskb);
2090 			__tcp_add_write_queue_head(sk, nskb);
2091 			sk_stream_free_skb(sk, skb);
2092 			sk_charge_skb(sk, nskb);
2093 			skb = nskb;
2094 		}
2095 
2096 		TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2097 		TCP_ECN_send_synack(tcp_sk(sk), skb);
2098 	}
2099 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2100 	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2101 }
2102 
2103 /*
2104  * Prepare a SYN-ACK.
2105  */
2106 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2107 				 struct request_sock *req)
2108 {
2109 	struct inet_request_sock *ireq = inet_rsk(req);
2110 	struct tcp_sock *tp = tcp_sk(sk);
2111 	struct tcphdr *th;
2112 	int tcp_header_size;
2113 	struct sk_buff *skb;
2114 #ifdef CONFIG_TCP_MD5SIG
2115 	struct tcp_md5sig_key *md5;
2116 	__u8 *md5_hash_location;
2117 #endif
2118 
2119 	skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2120 	if (skb == NULL)
2121 		return NULL;
2122 
2123 	/* Reserve space for headers. */
2124 	skb_reserve(skb, MAX_TCP_HEADER);
2125 
2126 	skb->dst = dst_clone(dst);
2127 
2128 	tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
2129 			   (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2130 			   (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
2131 			   /* SACK_PERM is in the place of NOP NOP of TS */
2132 			   ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
2133 
2134 #ifdef CONFIG_TCP_MD5SIG
2135 	/* Are we doing MD5 on this segment? If so - make room for it */
2136 	md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
2137 	if (md5)
2138 		tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
2139 #endif
2140 	skb_push(skb, tcp_header_size);
2141 	skb_reset_transport_header(skb);
2142 
2143 	th = tcp_hdr(skb);
2144 	memset(th, 0, sizeof(struct tcphdr));
2145 	th->syn = 1;
2146 	th->ack = 1;
2147 	TCP_ECN_make_synack(req, th);
2148 	th->source = inet_sk(sk)->sport;
2149 	th->dest = ireq->rmt_port;
2150 	TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
2151 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2152 	TCP_SKB_CB(skb)->sacked = 0;
2153 	skb_shinfo(skb)->gso_segs = 1;
2154 	skb_shinfo(skb)->gso_size = 0;
2155 	skb_shinfo(skb)->gso_type = 0;
2156 	th->seq = htonl(TCP_SKB_CB(skb)->seq);
2157 	th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2158 	if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2159 		__u8 rcv_wscale;
2160 		/* Set this up on the first call only */
2161 		req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2162 		/* tcp_full_space because it is guaranteed to be the first packet */
2163 		tcp_select_initial_window(tcp_full_space(sk),
2164 			dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2165 			&req->rcv_wnd,
2166 			&req->window_clamp,
2167 			ireq->wscale_ok,
2168 			&rcv_wscale);
2169 		ireq->rcv_wscale = rcv_wscale;
2170 	}
2171 
2172 	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2173 	th->window = htons(min(req->rcv_wnd, 65535U));
2174 
2175 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2176 	tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
2177 			      ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
2178 			      TCP_SKB_CB(skb)->when,
2179 			      req->ts_recent,
2180 			      (
2181 #ifdef CONFIG_TCP_MD5SIG
2182 			       md5 ? &md5_hash_location :
2183 #endif
2184 			       NULL)
2185 			      );
2186 
2187 	skb->csum = 0;
2188 	th->doff = (tcp_header_size >> 2);
2189 	TCP_INC_STATS(TCP_MIB_OUTSEGS);
2190 
2191 #ifdef CONFIG_TCP_MD5SIG
2192 	/* Okay, we have all we need - do the md5 hash if needed */
2193 	if (md5) {
2194 		tp->af_specific->calc_md5_hash(md5_hash_location,
2195 					       md5,
2196 					       NULL, dst, req,
2197 					       tcp_hdr(skb), sk->sk_protocol,
2198 					       skb->len);
2199 	}
2200 #endif
2201 
2202 	return skb;
2203 }
2204 
2205 /*
2206  * Do all connect socket setups that can be done AF independent.
2207  */
2208 static void tcp_connect_init(struct sock *sk)
2209 {
2210 	struct dst_entry *dst = __sk_dst_get(sk);
2211 	struct tcp_sock *tp = tcp_sk(sk);
2212 	__u8 rcv_wscale;
2213 
2214 	/* We'll fix this up when we get a response from the other end.
2215 	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2216 	 */
2217 	tp->tcp_header_len = sizeof(struct tcphdr) +
2218 		(sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2219 
2220 #ifdef CONFIG_TCP_MD5SIG
2221 	if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2222 		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2223 #endif
2224 
2225 	/* If user gave his TCP_MAXSEG, record it to clamp */
2226 	if (tp->rx_opt.user_mss)
2227 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2228 	tp->max_window = 0;
2229 	tcp_mtup_init(sk);
2230 	tcp_sync_mss(sk, dst_mtu(dst));
2231 
2232 	if (!tp->window_clamp)
2233 		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2234 	tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2235 	tcp_initialize_rcv_mss(sk);
2236 
2237 	tcp_select_initial_window(tcp_full_space(sk),
2238 				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2239 				  &tp->rcv_wnd,
2240 				  &tp->window_clamp,
2241 				  sysctl_tcp_window_scaling,
2242 				  &rcv_wscale);
2243 
2244 	tp->rx_opt.rcv_wscale = rcv_wscale;
2245 	tp->rcv_ssthresh = tp->rcv_wnd;
2246 
2247 	sk->sk_err = 0;
2248 	sock_reset_flag(sk, SOCK_DONE);
2249 	tp->snd_wnd = 0;
2250 	tcp_init_wl(tp, tp->write_seq, 0);
2251 	tp->snd_una = tp->write_seq;
2252 	tp->snd_sml = tp->write_seq;
2253 	tp->rcv_nxt = 0;
2254 	tp->rcv_wup = 0;
2255 	tp->copied_seq = 0;
2256 
2257 	inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2258 	inet_csk(sk)->icsk_retransmits = 0;
2259 	tcp_clear_retrans(tp);
2260 }
2261 
2262 /*
2263  * Build a SYN and send it off.
2264  */
2265 int tcp_connect(struct sock *sk)
2266 {
2267 	struct tcp_sock *tp = tcp_sk(sk);
2268 	struct sk_buff *buff;
2269 
2270 	tcp_connect_init(sk);
2271 
2272 	buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2273 	if (unlikely(buff == NULL))
2274 		return -ENOBUFS;
2275 
2276 	/* Reserve space for headers. */
2277 	skb_reserve(buff, MAX_TCP_HEADER);
2278 
2279 	TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
2280 	TCP_ECN_send_syn(sk, buff);
2281 	TCP_SKB_CB(buff)->sacked = 0;
2282 	skb_shinfo(buff)->gso_segs = 1;
2283 	skb_shinfo(buff)->gso_size = 0;
2284 	skb_shinfo(buff)->gso_type = 0;
2285 	buff->csum = 0;
2286 	tp->snd_nxt = tp->write_seq;
2287 	TCP_SKB_CB(buff)->seq = tp->write_seq++;
2288 	TCP_SKB_CB(buff)->end_seq = tp->write_seq;
2289 
2290 	/* Send it off. */
2291 	TCP_SKB_CB(buff)->when = tcp_time_stamp;
2292 	tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2293 	skb_header_release(buff);
2294 	__tcp_add_write_queue_tail(sk, buff);
2295 	sk_charge_skb(sk, buff);
2296 	tp->packets_out += tcp_skb_pcount(buff);
2297 	tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2298 
2299 	/* We change tp->snd_nxt after the tcp_transmit_skb() call
2300 	 * in order to make this packet get counted in tcpOutSegs.
2301 	 */
2302 	tp->snd_nxt = tp->write_seq;
2303 	tp->pushed_seq = tp->write_seq;
2304 	TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2305 
2306 	/* Timer for repeating the SYN until an answer. */
2307 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2308 				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2309 	return 0;
2310 }
2311 
2312 /* Send out a delayed ack, the caller does the policy checking
2313  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
2314  * for details.
2315  */
2316 void tcp_send_delayed_ack(struct sock *sk)
2317 {
2318 	struct inet_connection_sock *icsk = inet_csk(sk);
2319 	int ato = icsk->icsk_ack.ato;
2320 	unsigned long timeout;
2321 
2322 	if (ato > TCP_DELACK_MIN) {
2323 		const struct tcp_sock *tp = tcp_sk(sk);
2324 		int max_ato = HZ/2;
2325 
2326 		if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2327 			max_ato = TCP_DELACK_MAX;
2328 
2329 		/* Slow path, intersegment interval is "high". */
2330 
2331 		/* If some rtt estimate is known, use it to bound delayed ack.
2332 		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2333 		 * directly.
2334 		 */
2335 		if (tp->srtt) {
2336 			int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
2337 
2338 			if (rtt < max_ato)
2339 				max_ato = rtt;
2340 		}
2341 
2342 		ato = min(ato, max_ato);
2343 	}
2344 
2345 	/* Stay within the limit we were given */
2346 	timeout = jiffies + ato;
2347 
2348 	/* Use new timeout only if there wasn't a older one earlier. */
2349 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2350 		/* If delack timer was blocked or is about to expire,
2351 		 * send ACK now.
2352 		 */
2353 		if (icsk->icsk_ack.blocked ||
2354 		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2355 			tcp_send_ack(sk);
2356 			return;
2357 		}
2358 
2359 		if (!time_before(timeout, icsk->icsk_ack.timeout))
2360 			timeout = icsk->icsk_ack.timeout;
2361 	}
2362 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2363 	icsk->icsk_ack.timeout = timeout;
2364 	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2365 }
2366 
2367 /* This routine sends an ack and also updates the window. */
2368 void tcp_send_ack(struct sock *sk)
2369 {
2370 	/* If we have been reset, we may not send again. */
2371 	if (sk->sk_state != TCP_CLOSE) {
2372 		struct sk_buff *buff;
2373 
2374 		/* We are not putting this on the write queue, so
2375 		 * tcp_transmit_skb() will set the ownership to this
2376 		 * sock.
2377 		 */
2378 		buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2379 		if (buff == NULL) {
2380 			inet_csk_schedule_ack(sk);
2381 			inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2382 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2383 						  TCP_DELACK_MAX, TCP_RTO_MAX);
2384 			return;
2385 		}
2386 
2387 		/* Reserve space for headers and prepare control bits. */
2388 		skb_reserve(buff, MAX_TCP_HEADER);
2389 		buff->csum = 0;
2390 		TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2391 		TCP_SKB_CB(buff)->sacked = 0;
2392 		skb_shinfo(buff)->gso_segs = 1;
2393 		skb_shinfo(buff)->gso_size = 0;
2394 		skb_shinfo(buff)->gso_type = 0;
2395 
2396 		/* Send it off, this clears delayed acks for us. */
2397 		TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk);
2398 		TCP_SKB_CB(buff)->when = tcp_time_stamp;
2399 		tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2400 	}
2401 }
2402 
2403 /* This routine sends a packet with an out of date sequence
2404  * number. It assumes the other end will try to ack it.
2405  *
2406  * Question: what should we make while urgent mode?
2407  * 4.4BSD forces sending single byte of data. We cannot send
2408  * out of window data, because we have SND.NXT==SND.MAX...
2409  *
2410  * Current solution: to send TWO zero-length segments in urgent mode:
2411  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2412  * out-of-date with SND.UNA-1 to probe window.
2413  */
2414 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2415 {
2416 	struct tcp_sock *tp = tcp_sk(sk);
2417 	struct sk_buff *skb;
2418 
2419 	/* We don't queue it, tcp_transmit_skb() sets ownership. */
2420 	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2421 	if (skb == NULL)
2422 		return -1;
2423 
2424 	/* Reserve space for headers and set control bits. */
2425 	skb_reserve(skb, MAX_TCP_HEADER);
2426 	skb->csum = 0;
2427 	TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2428 	TCP_SKB_CB(skb)->sacked = urgent;
2429 	skb_shinfo(skb)->gso_segs = 1;
2430 	skb_shinfo(skb)->gso_size = 0;
2431 	skb_shinfo(skb)->gso_type = 0;
2432 
2433 	/* Use a previous sequence.  This should cause the other
2434 	 * end to send an ack.  Don't queue or clone SKB, just
2435 	 * send it.
2436 	 */
2437 	TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2438 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2439 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2440 	return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2441 }
2442 
2443 int tcp_write_wakeup(struct sock *sk)
2444 {
2445 	if (sk->sk_state != TCP_CLOSE) {
2446 		struct tcp_sock *tp = tcp_sk(sk);
2447 		struct sk_buff *skb;
2448 
2449 		if ((skb = tcp_send_head(sk)) != NULL &&
2450 		    before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
2451 			int err;
2452 			unsigned int mss = tcp_current_mss(sk, 0);
2453 			unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
2454 
2455 			if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2456 				tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2457 
2458 			/* We are probing the opening of a window
2459 			 * but the window size is != 0
2460 			 * must have been a result SWS avoidance ( sender )
2461 			 */
2462 			if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2463 			    skb->len > mss) {
2464 				seg_size = min(seg_size, mss);
2465 				TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2466 				if (tcp_fragment(sk, skb, seg_size, mss))
2467 					return -1;
2468 			} else if (!tcp_skb_pcount(skb))
2469 				tcp_set_skb_tso_segs(sk, skb, mss);
2470 
2471 			TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2472 			TCP_SKB_CB(skb)->when = tcp_time_stamp;
2473 			err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2474 			if (!err) {
2475 				update_send_head(sk, skb);
2476 			}
2477 			return err;
2478 		} else {
2479 			if (tp->urg_mode &&
2480 			    between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2481 				tcp_xmit_probe_skb(sk, TCPCB_URG);
2482 			return tcp_xmit_probe_skb(sk, 0);
2483 		}
2484 	}
2485 	return -1;
2486 }
2487 
2488 /* A window probe timeout has occurred.  If window is not closed send
2489  * a partial packet else a zero probe.
2490  */
2491 void tcp_send_probe0(struct sock *sk)
2492 {
2493 	struct inet_connection_sock *icsk = inet_csk(sk);
2494 	struct tcp_sock *tp = tcp_sk(sk);
2495 	int err;
2496 
2497 	err = tcp_write_wakeup(sk);
2498 
2499 	if (tp->packets_out || !tcp_send_head(sk)) {
2500 		/* Cancel probe timer, if it is not required. */
2501 		icsk->icsk_probes_out = 0;
2502 		icsk->icsk_backoff = 0;
2503 		return;
2504 	}
2505 
2506 	if (err <= 0) {
2507 		if (icsk->icsk_backoff < sysctl_tcp_retries2)
2508 			icsk->icsk_backoff++;
2509 		icsk->icsk_probes_out++;
2510 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2511 					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2512 					  TCP_RTO_MAX);
2513 	} else {
2514 		/* If packet was not sent due to local congestion,
2515 		 * do not backoff and do not remember icsk_probes_out.
2516 		 * Let local senders to fight for local resources.
2517 		 *
2518 		 * Use accumulated backoff yet.
2519 		 */
2520 		if (!icsk->icsk_probes_out)
2521 			icsk->icsk_probes_out = 1;
2522 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2523 					  min(icsk->icsk_rto << icsk->icsk_backoff,
2524 					      TCP_RESOURCE_PROBE_INTERVAL),
2525 					  TCP_RTO_MAX);
2526 	}
2527 }
2528 
2529 EXPORT_SYMBOL(tcp_connect);
2530 EXPORT_SYMBOL(tcp_make_synack);
2531 EXPORT_SYMBOL(tcp_simple_retransmit);
2532 EXPORT_SYMBOL(tcp_sync_mss);
2533 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
2534 EXPORT_SYMBOL(tcp_mtup_init);
2535