xref: /openbmc/linux/net/ipv4/tcp_output.c (revision 4800cd83)
1 /*
2  * INET		An implementation of the TCP/IP protocol suite for the LINUX
3  *		operating system.  INET is implemented using the  BSD Socket
4  *		interface as the means of communication with the user level.
5  *
6  *		Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Authors:	Ross Biro
9  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *		Mark Evans, <evansmp@uhura.aston.ac.uk>
11  *		Corey Minyard <wf-rch!minyard@relay.EU.net>
12  *		Florian La Roche, <flla@stud.uni-sb.de>
13  *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14  *		Linus Torvalds, <torvalds@cs.helsinki.fi>
15  *		Alan Cox, <gw4pts@gw4pts.ampr.org>
16  *		Matthew Dillon, <dillon@apollo.west.oic.com>
17  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *		Jorge Cwik, <jorge@laser.satlink.net>
19  */
20 
21 /*
22  * Changes:	Pedro Roque	:	Retransmit queue handled by TCP.
23  *				:	Fragmentation on mtu decrease
24  *				:	Segment collapse on retransmit
25  *				:	AF independence
26  *
27  *		Linus Torvalds	:	send_delayed_ack
28  *		David S. Miller	:	Charge memory using the right skb
29  *					during syn/ack processing.
30  *		David S. Miller :	Output engine completely rewritten.
31  *		Andrea Arcangeli:	SYNACK carry ts_recent in tsecr.
32  *		Cacophonix Gaul :	draft-minshall-nagle-01
33  *		J Hadi Salim	:	ECN support
34  *
35  */
36 
37 #include <net/tcp.h>
38 
39 #include <linux/compiler.h>
40 #include <linux/gfp.h>
41 #include <linux/module.h>
42 
43 /* People can turn this off for buggy TCP's found in printers etc. */
44 int sysctl_tcp_retrans_collapse __read_mostly = 1;
45 
46 /* People can turn this on to work with those rare, broken TCPs that
47  * interpret the window field as a signed quantity.
48  */
49 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
50 
51 /* This limits the percentage of the congestion window which we
52  * will allow a single TSO frame to consume.  Building TSO frames
53  * which are too large can cause TCP streams to be bursty.
54  */
55 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
56 
57 int sysctl_tcp_mtu_probing __read_mostly = 0;
58 int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
59 
60 /* By default, RFC2861 behavior.  */
61 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
62 
63 int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
64 EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
65 
66 
67 /* Account for new data that has been sent to the network. */
68 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
69 {
70 	struct tcp_sock *tp = tcp_sk(sk);
71 	unsigned int prior_packets = tp->packets_out;
72 
73 	tcp_advance_send_head(sk, skb);
74 	tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
75 
76 	/* Don't override Nagle indefinately with F-RTO */
77 	if (tp->frto_counter == 2)
78 		tp->frto_counter = 3;
79 
80 	tp->packets_out += tcp_skb_pcount(skb);
81 	if (!prior_packets)
82 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
83 					  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
84 }
85 
86 /* SND.NXT, if window was not shrunk.
87  * If window has been shrunk, what should we make? It is not clear at all.
88  * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
89  * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
90  * invalid. OK, let's make this for now:
91  */
92 static inline __u32 tcp_acceptable_seq(struct sock *sk)
93 {
94 	struct tcp_sock *tp = tcp_sk(sk);
95 
96 	if (!before(tcp_wnd_end(tp), tp->snd_nxt))
97 		return tp->snd_nxt;
98 	else
99 		return tcp_wnd_end(tp);
100 }
101 
102 /* Calculate mss to advertise in SYN segment.
103  * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
104  *
105  * 1. It is independent of path mtu.
106  * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
107  * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
108  *    attached devices, because some buggy hosts are confused by
109  *    large MSS.
110  * 4. We do not make 3, we advertise MSS, calculated from first
111  *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
112  *    This may be overridden via information stored in routing table.
113  * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
114  *    probably even Jumbo".
115  */
116 static __u16 tcp_advertise_mss(struct sock *sk)
117 {
118 	struct tcp_sock *tp = tcp_sk(sk);
119 	struct dst_entry *dst = __sk_dst_get(sk);
120 	int mss = tp->advmss;
121 
122 	if (dst) {
123 		unsigned int metric = dst_metric_advmss(dst);
124 
125 		if (metric < mss) {
126 			mss = metric;
127 			tp->advmss = mss;
128 		}
129 	}
130 
131 	return (__u16)mss;
132 }
133 
134 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
135  * This is the first part of cwnd validation mechanism. */
136 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
137 {
138 	struct tcp_sock *tp = tcp_sk(sk);
139 	s32 delta = tcp_time_stamp - tp->lsndtime;
140 	u32 restart_cwnd = tcp_init_cwnd(tp, dst);
141 	u32 cwnd = tp->snd_cwnd;
142 
143 	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
144 
145 	tp->snd_ssthresh = tcp_current_ssthresh(sk);
146 	restart_cwnd = min(restart_cwnd, cwnd);
147 
148 	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
149 		cwnd >>= 1;
150 	tp->snd_cwnd = max(cwnd, restart_cwnd);
151 	tp->snd_cwnd_stamp = tcp_time_stamp;
152 	tp->snd_cwnd_used = 0;
153 }
154 
155 /* Congestion state accounting after a packet has been sent. */
156 static void tcp_event_data_sent(struct tcp_sock *tp,
157 				struct sk_buff *skb, struct sock *sk)
158 {
159 	struct inet_connection_sock *icsk = inet_csk(sk);
160 	const u32 now = tcp_time_stamp;
161 
162 	if (sysctl_tcp_slow_start_after_idle &&
163 	    (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
164 		tcp_cwnd_restart(sk, __sk_dst_get(sk));
165 
166 	tp->lsndtime = now;
167 
168 	/* If it is a reply for ato after last received
169 	 * packet, enter pingpong mode.
170 	 */
171 	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
172 		icsk->icsk_ack.pingpong = 1;
173 }
174 
175 /* Account for an ACK we sent. */
176 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
177 {
178 	tcp_dec_quickack_mode(sk, pkts);
179 	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
180 }
181 
182 /* Determine a window scaling and initial window to offer.
183  * Based on the assumption that the given amount of space
184  * will be offered. Store the results in the tp structure.
185  * NOTE: for smooth operation initial space offering should
186  * be a multiple of mss if possible. We assume here that mss >= 1.
187  * This MUST be enforced by all callers.
188  */
189 void tcp_select_initial_window(int __space, __u32 mss,
190 			       __u32 *rcv_wnd, __u32 *window_clamp,
191 			       int wscale_ok, __u8 *rcv_wscale,
192 			       __u32 init_rcv_wnd)
193 {
194 	unsigned int space = (__space < 0 ? 0 : __space);
195 
196 	/* If no clamp set the clamp to the max possible scaled window */
197 	if (*window_clamp == 0)
198 		(*window_clamp) = (65535 << 14);
199 	space = min(*window_clamp, space);
200 
201 	/* Quantize space offering to a multiple of mss if possible. */
202 	if (space > mss)
203 		space = (space / mss) * mss;
204 
205 	/* NOTE: offering an initial window larger than 32767
206 	 * will break some buggy TCP stacks. If the admin tells us
207 	 * it is likely we could be speaking with such a buggy stack
208 	 * we will truncate our initial window offering to 32K-1
209 	 * unless the remote has sent us a window scaling option,
210 	 * which we interpret as a sign the remote TCP is not
211 	 * misinterpreting the window field as a signed quantity.
212 	 */
213 	if (sysctl_tcp_workaround_signed_windows)
214 		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
215 	else
216 		(*rcv_wnd) = space;
217 
218 	(*rcv_wscale) = 0;
219 	if (wscale_ok) {
220 		/* Set window scaling on max possible window
221 		 * See RFC1323 for an explanation of the limit to 14
222 		 */
223 		space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
224 		space = min_t(u32, space, *window_clamp);
225 		while (space > 65535 && (*rcv_wscale) < 14) {
226 			space >>= 1;
227 			(*rcv_wscale)++;
228 		}
229 	}
230 
231 	/* Set initial window to a value enough for senders starting with
232 	 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
233 	 * a limit on the initial window when mss is larger than 1460.
234 	 */
235 	if (mss > (1 << *rcv_wscale)) {
236 		int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
237 		if (mss > 1460)
238 			init_cwnd =
239 			max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
240 		/* when initializing use the value from init_rcv_wnd
241 		 * rather than the default from above
242 		 */
243 		if (init_rcv_wnd)
244 			*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
245 		else
246 			*rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
247 	}
248 
249 	/* Set the clamp no higher than max representable value */
250 	(*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
251 }
252 EXPORT_SYMBOL(tcp_select_initial_window);
253 
254 /* Chose a new window to advertise, update state in tcp_sock for the
255  * socket, and return result with RFC1323 scaling applied.  The return
256  * value can be stuffed directly into th->window for an outgoing
257  * frame.
258  */
259 static u16 tcp_select_window(struct sock *sk)
260 {
261 	struct tcp_sock *tp = tcp_sk(sk);
262 	u32 cur_win = tcp_receive_window(tp);
263 	u32 new_win = __tcp_select_window(sk);
264 
265 	/* Never shrink the offered window */
266 	if (new_win < cur_win) {
267 		/* Danger Will Robinson!
268 		 * Don't update rcv_wup/rcv_wnd here or else
269 		 * we will not be able to advertise a zero
270 		 * window in time.  --DaveM
271 		 *
272 		 * Relax Will Robinson.
273 		 */
274 		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
275 	}
276 	tp->rcv_wnd = new_win;
277 	tp->rcv_wup = tp->rcv_nxt;
278 
279 	/* Make sure we do not exceed the maximum possible
280 	 * scaled window.
281 	 */
282 	if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
283 		new_win = min(new_win, MAX_TCP_WINDOW);
284 	else
285 		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
286 
287 	/* RFC1323 scaling applied */
288 	new_win >>= tp->rx_opt.rcv_wscale;
289 
290 	/* If we advertise zero window, disable fast path. */
291 	if (new_win == 0)
292 		tp->pred_flags = 0;
293 
294 	return new_win;
295 }
296 
297 /* Packet ECN state for a SYN-ACK */
298 static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
299 {
300 	TCP_SKB_CB(skb)->flags &= ~TCPHDR_CWR;
301 	if (!(tp->ecn_flags & TCP_ECN_OK))
302 		TCP_SKB_CB(skb)->flags &= ~TCPHDR_ECE;
303 }
304 
305 /* Packet ECN state for a SYN.  */
306 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
307 {
308 	struct tcp_sock *tp = tcp_sk(sk);
309 
310 	tp->ecn_flags = 0;
311 	if (sysctl_tcp_ecn == 1) {
312 		TCP_SKB_CB(skb)->flags |= TCPHDR_ECE | TCPHDR_CWR;
313 		tp->ecn_flags = TCP_ECN_OK;
314 	}
315 }
316 
317 static __inline__ void
318 TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
319 {
320 	if (inet_rsk(req)->ecn_ok)
321 		th->ece = 1;
322 }
323 
324 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
325  * be sent.
326  */
327 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
328 				int tcp_header_len)
329 {
330 	struct tcp_sock *tp = tcp_sk(sk);
331 
332 	if (tp->ecn_flags & TCP_ECN_OK) {
333 		/* Not-retransmitted data segment: set ECT and inject CWR. */
334 		if (skb->len != tcp_header_len &&
335 		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
336 			INET_ECN_xmit(sk);
337 			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
338 				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
339 				tcp_hdr(skb)->cwr = 1;
340 				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
341 			}
342 		} else {
343 			/* ACK or retransmitted segment: clear ECT|CE */
344 			INET_ECN_dontxmit(sk);
345 		}
346 		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
347 			tcp_hdr(skb)->ece = 1;
348 	}
349 }
350 
351 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
352  * auto increment end seqno.
353  */
354 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
355 {
356 	skb->ip_summed = CHECKSUM_PARTIAL;
357 	skb->csum = 0;
358 
359 	TCP_SKB_CB(skb)->flags = flags;
360 	TCP_SKB_CB(skb)->sacked = 0;
361 
362 	skb_shinfo(skb)->gso_segs = 1;
363 	skb_shinfo(skb)->gso_size = 0;
364 	skb_shinfo(skb)->gso_type = 0;
365 
366 	TCP_SKB_CB(skb)->seq = seq;
367 	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
368 		seq++;
369 	TCP_SKB_CB(skb)->end_seq = seq;
370 }
371 
372 static inline int tcp_urg_mode(const struct tcp_sock *tp)
373 {
374 	return tp->snd_una != tp->snd_up;
375 }
376 
377 #define OPTION_SACK_ADVERTISE	(1 << 0)
378 #define OPTION_TS		(1 << 1)
379 #define OPTION_MD5		(1 << 2)
380 #define OPTION_WSCALE		(1 << 3)
381 #define OPTION_COOKIE_EXTENSION	(1 << 4)
382 
383 struct tcp_out_options {
384 	u8 options;		/* bit field of OPTION_* */
385 	u8 ws;			/* window scale, 0 to disable */
386 	u8 num_sack_blocks;	/* number of SACK blocks to include */
387 	u8 hash_size;		/* bytes in hash_location */
388 	u16 mss;		/* 0 to disable */
389 	__u32 tsval, tsecr;	/* need to include OPTION_TS */
390 	__u8 *hash_location;	/* temporary pointer, overloaded */
391 };
392 
393 /* The sysctl int routines are generic, so check consistency here.
394  */
395 static u8 tcp_cookie_size_check(u8 desired)
396 {
397 	int cookie_size;
398 
399 	if (desired > 0)
400 		/* previously specified */
401 		return desired;
402 
403 	cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
404 	if (cookie_size <= 0)
405 		/* no default specified */
406 		return 0;
407 
408 	if (cookie_size <= TCP_COOKIE_MIN)
409 		/* value too small, specify minimum */
410 		return TCP_COOKIE_MIN;
411 
412 	if (cookie_size >= TCP_COOKIE_MAX)
413 		/* value too large, specify maximum */
414 		return TCP_COOKIE_MAX;
415 
416 	if (cookie_size & 1)
417 		/* 8-bit multiple, illegal, fix it */
418 		cookie_size++;
419 
420 	return (u8)cookie_size;
421 }
422 
423 /* Write previously computed TCP options to the packet.
424  *
425  * Beware: Something in the Internet is very sensitive to the ordering of
426  * TCP options, we learned this through the hard way, so be careful here.
427  * Luckily we can at least blame others for their non-compliance but from
428  * inter-operatibility perspective it seems that we're somewhat stuck with
429  * the ordering which we have been using if we want to keep working with
430  * those broken things (not that it currently hurts anybody as there isn't
431  * particular reason why the ordering would need to be changed).
432  *
433  * At least SACK_PERM as the first option is known to lead to a disaster
434  * (but it may well be that other scenarios fail similarly).
435  */
436 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
437 			      struct tcp_out_options *opts)
438 {
439 	u8 options = opts->options;	/* mungable copy */
440 
441 	/* Having both authentication and cookies for security is redundant,
442 	 * and there's certainly not enough room.  Instead, the cookie-less
443 	 * extension variant is proposed.
444 	 *
445 	 * Consider the pessimal case with authentication.  The options
446 	 * could look like:
447 	 *   COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
448 	 */
449 	if (unlikely(OPTION_MD5 & options)) {
450 		if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
451 			*ptr++ = htonl((TCPOPT_COOKIE << 24) |
452 				       (TCPOLEN_COOKIE_BASE << 16) |
453 				       (TCPOPT_MD5SIG << 8) |
454 				       TCPOLEN_MD5SIG);
455 		} else {
456 			*ptr++ = htonl((TCPOPT_NOP << 24) |
457 				       (TCPOPT_NOP << 16) |
458 				       (TCPOPT_MD5SIG << 8) |
459 				       TCPOLEN_MD5SIG);
460 		}
461 		options &= ~OPTION_COOKIE_EXTENSION;
462 		/* overload cookie hash location */
463 		opts->hash_location = (__u8 *)ptr;
464 		ptr += 4;
465 	}
466 
467 	if (unlikely(opts->mss)) {
468 		*ptr++ = htonl((TCPOPT_MSS << 24) |
469 			       (TCPOLEN_MSS << 16) |
470 			       opts->mss);
471 	}
472 
473 	if (likely(OPTION_TS & options)) {
474 		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
475 			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
476 				       (TCPOLEN_SACK_PERM << 16) |
477 				       (TCPOPT_TIMESTAMP << 8) |
478 				       TCPOLEN_TIMESTAMP);
479 			options &= ~OPTION_SACK_ADVERTISE;
480 		} else {
481 			*ptr++ = htonl((TCPOPT_NOP << 24) |
482 				       (TCPOPT_NOP << 16) |
483 				       (TCPOPT_TIMESTAMP << 8) |
484 				       TCPOLEN_TIMESTAMP);
485 		}
486 		*ptr++ = htonl(opts->tsval);
487 		*ptr++ = htonl(opts->tsecr);
488 	}
489 
490 	/* Specification requires after timestamp, so do it now.
491 	 *
492 	 * Consider the pessimal case without authentication.  The options
493 	 * could look like:
494 	 *   MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
495 	 */
496 	if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
497 		__u8 *cookie_copy = opts->hash_location;
498 		u8 cookie_size = opts->hash_size;
499 
500 		/* 8-bit multiple handled in tcp_cookie_size_check() above,
501 		 * and elsewhere.
502 		 */
503 		if (0x2 & cookie_size) {
504 			__u8 *p = (__u8 *)ptr;
505 
506 			/* 16-bit multiple */
507 			*p++ = TCPOPT_COOKIE;
508 			*p++ = TCPOLEN_COOKIE_BASE + cookie_size;
509 			*p++ = *cookie_copy++;
510 			*p++ = *cookie_copy++;
511 			ptr++;
512 			cookie_size -= 2;
513 		} else {
514 			/* 32-bit multiple */
515 			*ptr++ = htonl(((TCPOPT_NOP << 24) |
516 					(TCPOPT_NOP << 16) |
517 					(TCPOPT_COOKIE << 8) |
518 					TCPOLEN_COOKIE_BASE) +
519 				       cookie_size);
520 		}
521 
522 		if (cookie_size > 0) {
523 			memcpy(ptr, cookie_copy, cookie_size);
524 			ptr += (cookie_size / 4);
525 		}
526 	}
527 
528 	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
529 		*ptr++ = htonl((TCPOPT_NOP << 24) |
530 			       (TCPOPT_NOP << 16) |
531 			       (TCPOPT_SACK_PERM << 8) |
532 			       TCPOLEN_SACK_PERM);
533 	}
534 
535 	if (unlikely(OPTION_WSCALE & options)) {
536 		*ptr++ = htonl((TCPOPT_NOP << 24) |
537 			       (TCPOPT_WINDOW << 16) |
538 			       (TCPOLEN_WINDOW << 8) |
539 			       opts->ws);
540 	}
541 
542 	if (unlikely(opts->num_sack_blocks)) {
543 		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
544 			tp->duplicate_sack : tp->selective_acks;
545 		int this_sack;
546 
547 		*ptr++ = htonl((TCPOPT_NOP  << 24) |
548 			       (TCPOPT_NOP  << 16) |
549 			       (TCPOPT_SACK <<  8) |
550 			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
551 						     TCPOLEN_SACK_PERBLOCK)));
552 
553 		for (this_sack = 0; this_sack < opts->num_sack_blocks;
554 		     ++this_sack) {
555 			*ptr++ = htonl(sp[this_sack].start_seq);
556 			*ptr++ = htonl(sp[this_sack].end_seq);
557 		}
558 
559 		tp->rx_opt.dsack = 0;
560 	}
561 }
562 
563 /* Compute TCP options for SYN packets. This is not the final
564  * network wire format yet.
565  */
566 static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
567 				struct tcp_out_options *opts,
568 				struct tcp_md5sig_key **md5) {
569 	struct tcp_sock *tp = tcp_sk(sk);
570 	struct tcp_cookie_values *cvp = tp->cookie_values;
571 	unsigned remaining = MAX_TCP_OPTION_SPACE;
572 	u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
573 			 tcp_cookie_size_check(cvp->cookie_desired) :
574 			 0;
575 
576 #ifdef CONFIG_TCP_MD5SIG
577 	*md5 = tp->af_specific->md5_lookup(sk, sk);
578 	if (*md5) {
579 		opts->options |= OPTION_MD5;
580 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
581 	}
582 #else
583 	*md5 = NULL;
584 #endif
585 
586 	/* We always get an MSS option.  The option bytes which will be seen in
587 	 * normal data packets should timestamps be used, must be in the MSS
588 	 * advertised.  But we subtract them from tp->mss_cache so that
589 	 * calculations in tcp_sendmsg are simpler etc.  So account for this
590 	 * fact here if necessary.  If we don't do this correctly, as a
591 	 * receiver we won't recognize data packets as being full sized when we
592 	 * should, and thus we won't abide by the delayed ACK rules correctly.
593 	 * SACKs don't matter, we never delay an ACK when we have any of those
594 	 * going out.  */
595 	opts->mss = tcp_advertise_mss(sk);
596 	remaining -= TCPOLEN_MSS_ALIGNED;
597 
598 	if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
599 		opts->options |= OPTION_TS;
600 		opts->tsval = TCP_SKB_CB(skb)->when;
601 		opts->tsecr = tp->rx_opt.ts_recent;
602 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
603 	}
604 	if (likely(sysctl_tcp_window_scaling)) {
605 		opts->ws = tp->rx_opt.rcv_wscale;
606 		opts->options |= OPTION_WSCALE;
607 		remaining -= TCPOLEN_WSCALE_ALIGNED;
608 	}
609 	if (likely(sysctl_tcp_sack)) {
610 		opts->options |= OPTION_SACK_ADVERTISE;
611 		if (unlikely(!(OPTION_TS & opts->options)))
612 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
613 	}
614 
615 	/* Note that timestamps are required by the specification.
616 	 *
617 	 * Odd numbers of bytes are prohibited by the specification, ensuring
618 	 * that the cookie is 16-bit aligned, and the resulting cookie pair is
619 	 * 32-bit aligned.
620 	 */
621 	if (*md5 == NULL &&
622 	    (OPTION_TS & opts->options) &&
623 	    cookie_size > 0) {
624 		int need = TCPOLEN_COOKIE_BASE + cookie_size;
625 
626 		if (0x2 & need) {
627 			/* 32-bit multiple */
628 			need += 2; /* NOPs */
629 
630 			if (need > remaining) {
631 				/* try shrinking cookie to fit */
632 				cookie_size -= 2;
633 				need -= 4;
634 			}
635 		}
636 		while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
637 			cookie_size -= 4;
638 			need -= 4;
639 		}
640 		if (TCP_COOKIE_MIN <= cookie_size) {
641 			opts->options |= OPTION_COOKIE_EXTENSION;
642 			opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
643 			opts->hash_size = cookie_size;
644 
645 			/* Remember for future incarnations. */
646 			cvp->cookie_desired = cookie_size;
647 
648 			if (cvp->cookie_desired != cvp->cookie_pair_size) {
649 				/* Currently use random bytes as a nonce,
650 				 * assuming these are completely unpredictable
651 				 * by hostile users of the same system.
652 				 */
653 				get_random_bytes(&cvp->cookie_pair[0],
654 						 cookie_size);
655 				cvp->cookie_pair_size = cookie_size;
656 			}
657 
658 			remaining -= need;
659 		}
660 	}
661 	return MAX_TCP_OPTION_SPACE - remaining;
662 }
663 
664 /* Set up TCP options for SYN-ACKs. */
665 static unsigned tcp_synack_options(struct sock *sk,
666 				   struct request_sock *req,
667 				   unsigned mss, struct sk_buff *skb,
668 				   struct tcp_out_options *opts,
669 				   struct tcp_md5sig_key **md5,
670 				   struct tcp_extend_values *xvp)
671 {
672 	struct inet_request_sock *ireq = inet_rsk(req);
673 	unsigned remaining = MAX_TCP_OPTION_SPACE;
674 	u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
675 			 xvp->cookie_plus :
676 			 0;
677 
678 #ifdef CONFIG_TCP_MD5SIG
679 	*md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
680 	if (*md5) {
681 		opts->options |= OPTION_MD5;
682 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
683 
684 		/* We can't fit any SACK blocks in a packet with MD5 + TS
685 		 * options. There was discussion about disabling SACK
686 		 * rather than TS in order to fit in better with old,
687 		 * buggy kernels, but that was deemed to be unnecessary.
688 		 */
689 		ireq->tstamp_ok &= !ireq->sack_ok;
690 	}
691 #else
692 	*md5 = NULL;
693 #endif
694 
695 	/* We always send an MSS option. */
696 	opts->mss = mss;
697 	remaining -= TCPOLEN_MSS_ALIGNED;
698 
699 	if (likely(ireq->wscale_ok)) {
700 		opts->ws = ireq->rcv_wscale;
701 		opts->options |= OPTION_WSCALE;
702 		remaining -= TCPOLEN_WSCALE_ALIGNED;
703 	}
704 	if (likely(ireq->tstamp_ok)) {
705 		opts->options |= OPTION_TS;
706 		opts->tsval = TCP_SKB_CB(skb)->when;
707 		opts->tsecr = req->ts_recent;
708 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
709 	}
710 	if (likely(ireq->sack_ok)) {
711 		opts->options |= OPTION_SACK_ADVERTISE;
712 		if (unlikely(!ireq->tstamp_ok))
713 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
714 	}
715 
716 	/* Similar rationale to tcp_syn_options() applies here, too.
717 	 * If the <SYN> options fit, the same options should fit now!
718 	 */
719 	if (*md5 == NULL &&
720 	    ireq->tstamp_ok &&
721 	    cookie_plus > TCPOLEN_COOKIE_BASE) {
722 		int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
723 
724 		if (0x2 & need) {
725 			/* 32-bit multiple */
726 			need += 2; /* NOPs */
727 		}
728 		if (need <= remaining) {
729 			opts->options |= OPTION_COOKIE_EXTENSION;
730 			opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
731 			remaining -= need;
732 		} else {
733 			/* There's no error return, so flag it. */
734 			xvp->cookie_out_never = 1; /* true */
735 			opts->hash_size = 0;
736 		}
737 	}
738 	return MAX_TCP_OPTION_SPACE - remaining;
739 }
740 
741 /* Compute TCP options for ESTABLISHED sockets. This is not the
742  * final wire format yet.
743  */
744 static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
745 					struct tcp_out_options *opts,
746 					struct tcp_md5sig_key **md5) {
747 	struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
748 	struct tcp_sock *tp = tcp_sk(sk);
749 	unsigned size = 0;
750 	unsigned int eff_sacks;
751 
752 #ifdef CONFIG_TCP_MD5SIG
753 	*md5 = tp->af_specific->md5_lookup(sk, sk);
754 	if (unlikely(*md5)) {
755 		opts->options |= OPTION_MD5;
756 		size += TCPOLEN_MD5SIG_ALIGNED;
757 	}
758 #else
759 	*md5 = NULL;
760 #endif
761 
762 	if (likely(tp->rx_opt.tstamp_ok)) {
763 		opts->options |= OPTION_TS;
764 		opts->tsval = tcb ? tcb->when : 0;
765 		opts->tsecr = tp->rx_opt.ts_recent;
766 		size += TCPOLEN_TSTAMP_ALIGNED;
767 	}
768 
769 	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
770 	if (unlikely(eff_sacks)) {
771 		const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
772 		opts->num_sack_blocks =
773 			min_t(unsigned, eff_sacks,
774 			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
775 			      TCPOLEN_SACK_PERBLOCK);
776 		size += TCPOLEN_SACK_BASE_ALIGNED +
777 			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
778 	}
779 
780 	return size;
781 }
782 
783 /* This routine actually transmits TCP packets queued in by
784  * tcp_do_sendmsg().  This is used by both the initial
785  * transmission and possible later retransmissions.
786  * All SKB's seen here are completely headerless.  It is our
787  * job to build the TCP header, and pass the packet down to
788  * IP so it can do the same plus pass the packet off to the
789  * device.
790  *
791  * We are working here with either a clone of the original
792  * SKB, or a fresh unique copy made by the retransmit engine.
793  */
794 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
795 			    gfp_t gfp_mask)
796 {
797 	const struct inet_connection_sock *icsk = inet_csk(sk);
798 	struct inet_sock *inet;
799 	struct tcp_sock *tp;
800 	struct tcp_skb_cb *tcb;
801 	struct tcp_out_options opts;
802 	unsigned tcp_options_size, tcp_header_size;
803 	struct tcp_md5sig_key *md5;
804 	struct tcphdr *th;
805 	int err;
806 
807 	BUG_ON(!skb || !tcp_skb_pcount(skb));
808 
809 	/* If congestion control is doing timestamping, we must
810 	 * take such a timestamp before we potentially clone/copy.
811 	 */
812 	if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
813 		__net_timestamp(skb);
814 
815 	if (likely(clone_it)) {
816 		if (unlikely(skb_cloned(skb)))
817 			skb = pskb_copy(skb, gfp_mask);
818 		else
819 			skb = skb_clone(skb, gfp_mask);
820 		if (unlikely(!skb))
821 			return -ENOBUFS;
822 	}
823 
824 	inet = inet_sk(sk);
825 	tp = tcp_sk(sk);
826 	tcb = TCP_SKB_CB(skb);
827 	memset(&opts, 0, sizeof(opts));
828 
829 	if (unlikely(tcb->flags & TCPHDR_SYN))
830 		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
831 	else
832 		tcp_options_size = tcp_established_options(sk, skb, &opts,
833 							   &md5);
834 	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
835 
836 	if (tcp_packets_in_flight(tp) == 0) {
837 		tcp_ca_event(sk, CA_EVENT_TX_START);
838 		skb->ooo_okay = 1;
839 	} else
840 		skb->ooo_okay = 0;
841 
842 	skb_push(skb, tcp_header_size);
843 	skb_reset_transport_header(skb);
844 	skb_set_owner_w(skb, sk);
845 
846 	/* Build TCP header and checksum it. */
847 	th = tcp_hdr(skb);
848 	th->source		= inet->inet_sport;
849 	th->dest		= inet->inet_dport;
850 	th->seq			= htonl(tcb->seq);
851 	th->ack_seq		= htonl(tp->rcv_nxt);
852 	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
853 					tcb->flags);
854 
855 	if (unlikely(tcb->flags & TCPHDR_SYN)) {
856 		/* RFC1323: The window in SYN & SYN/ACK segments
857 		 * is never scaled.
858 		 */
859 		th->window	= htons(min(tp->rcv_wnd, 65535U));
860 	} else {
861 		th->window	= htons(tcp_select_window(sk));
862 	}
863 	th->check		= 0;
864 	th->urg_ptr		= 0;
865 
866 	/* The urg_mode check is necessary during a below snd_una win probe */
867 	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
868 		if (before(tp->snd_up, tcb->seq + 0x10000)) {
869 			th->urg_ptr = htons(tp->snd_up - tcb->seq);
870 			th->urg = 1;
871 		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
872 			th->urg_ptr = htons(0xFFFF);
873 			th->urg = 1;
874 		}
875 	}
876 
877 	tcp_options_write((__be32 *)(th + 1), tp, &opts);
878 	if (likely((tcb->flags & TCPHDR_SYN) == 0))
879 		TCP_ECN_send(sk, skb, tcp_header_size);
880 
881 #ifdef CONFIG_TCP_MD5SIG
882 	/* Calculate the MD5 hash, as we have all we need now */
883 	if (md5) {
884 		sk_nocaps_add(sk, NETIF_F_GSO_MASK);
885 		tp->af_specific->calc_md5_hash(opts.hash_location,
886 					       md5, sk, NULL, skb);
887 	}
888 #endif
889 
890 	icsk->icsk_af_ops->send_check(sk, skb);
891 
892 	if (likely(tcb->flags & TCPHDR_ACK))
893 		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
894 
895 	if (skb->len != tcp_header_size)
896 		tcp_event_data_sent(tp, skb, sk);
897 
898 	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
899 		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
900 			      tcp_skb_pcount(skb));
901 
902 	err = icsk->icsk_af_ops->queue_xmit(skb);
903 	if (likely(err <= 0))
904 		return err;
905 
906 	tcp_enter_cwr(sk, 1);
907 
908 	return net_xmit_eval(err);
909 }
910 
911 /* This routine just queues the buffer for sending.
912  *
913  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
914  * otherwise socket can stall.
915  */
916 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
917 {
918 	struct tcp_sock *tp = tcp_sk(sk);
919 
920 	/* Advance write_seq and place onto the write_queue. */
921 	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
922 	skb_header_release(skb);
923 	tcp_add_write_queue_tail(sk, skb);
924 	sk->sk_wmem_queued += skb->truesize;
925 	sk_mem_charge(sk, skb->truesize);
926 }
927 
928 /* Initialize TSO segments for a packet. */
929 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
930 				 unsigned int mss_now)
931 {
932 	if (skb->len <= mss_now || !sk_can_gso(sk) ||
933 	    skb->ip_summed == CHECKSUM_NONE) {
934 		/* Avoid the costly divide in the normal
935 		 * non-TSO case.
936 		 */
937 		skb_shinfo(skb)->gso_segs = 1;
938 		skb_shinfo(skb)->gso_size = 0;
939 		skb_shinfo(skb)->gso_type = 0;
940 	} else {
941 		skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
942 		skb_shinfo(skb)->gso_size = mss_now;
943 		skb_shinfo(skb)->gso_type = sk->sk_gso_type;
944 	}
945 }
946 
947 /* When a modification to fackets out becomes necessary, we need to check
948  * skb is counted to fackets_out or not.
949  */
950 static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
951 				   int decr)
952 {
953 	struct tcp_sock *tp = tcp_sk(sk);
954 
955 	if (!tp->sacked_out || tcp_is_reno(tp))
956 		return;
957 
958 	if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
959 		tp->fackets_out -= decr;
960 }
961 
962 /* Pcount in the middle of the write queue got changed, we need to do various
963  * tweaks to fix counters
964  */
965 static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr)
966 {
967 	struct tcp_sock *tp = tcp_sk(sk);
968 
969 	tp->packets_out -= decr;
970 
971 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
972 		tp->sacked_out -= decr;
973 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
974 		tp->retrans_out -= decr;
975 	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
976 		tp->lost_out -= decr;
977 
978 	/* Reno case is special. Sigh... */
979 	if (tcp_is_reno(tp) && decr > 0)
980 		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
981 
982 	tcp_adjust_fackets_out(sk, skb, decr);
983 
984 	if (tp->lost_skb_hint &&
985 	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
986 	    (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
987 		tp->lost_cnt_hint -= decr;
988 
989 	tcp_verify_left_out(tp);
990 }
991 
992 /* Function to create two new TCP segments.  Shrinks the given segment
993  * to the specified size and appends a new segment with the rest of the
994  * packet to the list.  This won't be called frequently, I hope.
995  * Remember, these are still headerless SKBs at this point.
996  */
997 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
998 		 unsigned int mss_now)
999 {
1000 	struct tcp_sock *tp = tcp_sk(sk);
1001 	struct sk_buff *buff;
1002 	int nsize, old_factor;
1003 	int nlen;
1004 	u8 flags;
1005 
1006 	BUG_ON(len > skb->len);
1007 
1008 	nsize = skb_headlen(skb) - len;
1009 	if (nsize < 0)
1010 		nsize = 0;
1011 
1012 	if (skb_cloned(skb) &&
1013 	    skb_is_nonlinear(skb) &&
1014 	    pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1015 		return -ENOMEM;
1016 
1017 	/* Get a new skb... force flag on. */
1018 	buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1019 	if (buff == NULL)
1020 		return -ENOMEM; /* We'll just try again later. */
1021 
1022 	sk->sk_wmem_queued += buff->truesize;
1023 	sk_mem_charge(sk, buff->truesize);
1024 	nlen = skb->len - len - nsize;
1025 	buff->truesize += nlen;
1026 	skb->truesize -= nlen;
1027 
1028 	/* Correct the sequence numbers. */
1029 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1030 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1031 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1032 
1033 	/* PSH and FIN should only be set in the second packet. */
1034 	flags = TCP_SKB_CB(skb)->flags;
1035 	TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1036 	TCP_SKB_CB(buff)->flags = flags;
1037 	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1038 
1039 	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1040 		/* Copy and checksum data tail into the new buffer. */
1041 		buff->csum = csum_partial_copy_nocheck(skb->data + len,
1042 						       skb_put(buff, nsize),
1043 						       nsize, 0);
1044 
1045 		skb_trim(skb, len);
1046 
1047 		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1048 	} else {
1049 		skb->ip_summed = CHECKSUM_PARTIAL;
1050 		skb_split(skb, buff, len);
1051 	}
1052 
1053 	buff->ip_summed = skb->ip_summed;
1054 
1055 	/* Looks stupid, but our code really uses when of
1056 	 * skbs, which it never sent before. --ANK
1057 	 */
1058 	TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1059 	buff->tstamp = skb->tstamp;
1060 
1061 	old_factor = tcp_skb_pcount(skb);
1062 
1063 	/* Fix up tso_factor for both original and new SKB.  */
1064 	tcp_set_skb_tso_segs(sk, skb, mss_now);
1065 	tcp_set_skb_tso_segs(sk, buff, mss_now);
1066 
1067 	/* If this packet has been sent out already, we must
1068 	 * adjust the various packet counters.
1069 	 */
1070 	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1071 		int diff = old_factor - tcp_skb_pcount(skb) -
1072 			tcp_skb_pcount(buff);
1073 
1074 		if (diff)
1075 			tcp_adjust_pcount(sk, skb, diff);
1076 	}
1077 
1078 	/* Link BUFF into the send queue. */
1079 	skb_header_release(buff);
1080 	tcp_insert_write_queue_after(skb, buff, sk);
1081 
1082 	return 0;
1083 }
1084 
1085 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1086  * eventually). The difference is that pulled data not copied, but
1087  * immediately discarded.
1088  */
1089 static void __pskb_trim_head(struct sk_buff *skb, int len)
1090 {
1091 	int i, k, eat;
1092 
1093 	eat = len;
1094 	k = 0;
1095 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1096 		if (skb_shinfo(skb)->frags[i].size <= eat) {
1097 			put_page(skb_shinfo(skb)->frags[i].page);
1098 			eat -= skb_shinfo(skb)->frags[i].size;
1099 		} else {
1100 			skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1101 			if (eat) {
1102 				skb_shinfo(skb)->frags[k].page_offset += eat;
1103 				skb_shinfo(skb)->frags[k].size -= eat;
1104 				eat = 0;
1105 			}
1106 			k++;
1107 		}
1108 	}
1109 	skb_shinfo(skb)->nr_frags = k;
1110 
1111 	skb_reset_tail_pointer(skb);
1112 	skb->data_len -= len;
1113 	skb->len = skb->data_len;
1114 }
1115 
1116 /* Remove acked data from a packet in the transmit queue. */
1117 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1118 {
1119 	if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1120 		return -ENOMEM;
1121 
1122 	/* If len == headlen, we avoid __skb_pull to preserve alignment. */
1123 	if (unlikely(len < skb_headlen(skb)))
1124 		__skb_pull(skb, len);
1125 	else
1126 		__pskb_trim_head(skb, len - skb_headlen(skb));
1127 
1128 	TCP_SKB_CB(skb)->seq += len;
1129 	skb->ip_summed = CHECKSUM_PARTIAL;
1130 
1131 	skb->truesize	     -= len;
1132 	sk->sk_wmem_queued   -= len;
1133 	sk_mem_uncharge(sk, len);
1134 	sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1135 
1136 	/* Any change of skb->len requires recalculation of tso
1137 	 * factor and mss.
1138 	 */
1139 	if (tcp_skb_pcount(skb) > 1)
1140 		tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk));
1141 
1142 	return 0;
1143 }
1144 
1145 /* Calculate MSS. Not accounting for SACKs here.  */
1146 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1147 {
1148 	struct tcp_sock *tp = tcp_sk(sk);
1149 	struct inet_connection_sock *icsk = inet_csk(sk);
1150 	int mss_now;
1151 
1152 	/* Calculate base mss without TCP options:
1153 	   It is MMS_S - sizeof(tcphdr) of rfc1122
1154 	 */
1155 	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1156 
1157 	/* Clamp it (mss_clamp does not include tcp options) */
1158 	if (mss_now > tp->rx_opt.mss_clamp)
1159 		mss_now = tp->rx_opt.mss_clamp;
1160 
1161 	/* Now subtract optional transport overhead */
1162 	mss_now -= icsk->icsk_ext_hdr_len;
1163 
1164 	/* Then reserve room for full set of TCP options and 8 bytes of data */
1165 	if (mss_now < 48)
1166 		mss_now = 48;
1167 
1168 	/* Now subtract TCP options size, not including SACKs */
1169 	mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1170 
1171 	return mss_now;
1172 }
1173 
1174 /* Inverse of above */
1175 int tcp_mss_to_mtu(struct sock *sk, int mss)
1176 {
1177 	struct tcp_sock *tp = tcp_sk(sk);
1178 	struct inet_connection_sock *icsk = inet_csk(sk);
1179 	int mtu;
1180 
1181 	mtu = mss +
1182 	      tp->tcp_header_len +
1183 	      icsk->icsk_ext_hdr_len +
1184 	      icsk->icsk_af_ops->net_header_len;
1185 
1186 	return mtu;
1187 }
1188 
1189 /* MTU probing init per socket */
1190 void tcp_mtup_init(struct sock *sk)
1191 {
1192 	struct tcp_sock *tp = tcp_sk(sk);
1193 	struct inet_connection_sock *icsk = inet_csk(sk);
1194 
1195 	icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1196 	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1197 			       icsk->icsk_af_ops->net_header_len;
1198 	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1199 	icsk->icsk_mtup.probe_size = 0;
1200 }
1201 EXPORT_SYMBOL(tcp_mtup_init);
1202 
1203 /* This function synchronize snd mss to current pmtu/exthdr set.
1204 
1205    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1206    for TCP options, but includes only bare TCP header.
1207 
1208    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1209    It is minimum of user_mss and mss received with SYN.
1210    It also does not include TCP options.
1211 
1212    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1213 
1214    tp->mss_cache is current effective sending mss, including
1215    all tcp options except for SACKs. It is evaluated,
1216    taking into account current pmtu, but never exceeds
1217    tp->rx_opt.mss_clamp.
1218 
1219    NOTE1. rfc1122 clearly states that advertised MSS
1220    DOES NOT include either tcp or ip options.
1221 
1222    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1223    are READ ONLY outside this function.		--ANK (980731)
1224  */
1225 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1226 {
1227 	struct tcp_sock *tp = tcp_sk(sk);
1228 	struct inet_connection_sock *icsk = inet_csk(sk);
1229 	int mss_now;
1230 
1231 	if (icsk->icsk_mtup.search_high > pmtu)
1232 		icsk->icsk_mtup.search_high = pmtu;
1233 
1234 	mss_now = tcp_mtu_to_mss(sk, pmtu);
1235 	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1236 
1237 	/* And store cached results */
1238 	icsk->icsk_pmtu_cookie = pmtu;
1239 	if (icsk->icsk_mtup.enabled)
1240 		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1241 	tp->mss_cache = mss_now;
1242 
1243 	return mss_now;
1244 }
1245 EXPORT_SYMBOL(tcp_sync_mss);
1246 
1247 /* Compute the current effective MSS, taking SACKs and IP options,
1248  * and even PMTU discovery events into account.
1249  */
1250 unsigned int tcp_current_mss(struct sock *sk)
1251 {
1252 	struct tcp_sock *tp = tcp_sk(sk);
1253 	struct dst_entry *dst = __sk_dst_get(sk);
1254 	u32 mss_now;
1255 	unsigned header_len;
1256 	struct tcp_out_options opts;
1257 	struct tcp_md5sig_key *md5;
1258 
1259 	mss_now = tp->mss_cache;
1260 
1261 	if (dst) {
1262 		u32 mtu = dst_mtu(dst);
1263 		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1264 			mss_now = tcp_sync_mss(sk, mtu);
1265 	}
1266 
1267 	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1268 		     sizeof(struct tcphdr);
1269 	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1270 	 * some common options. If this is an odd packet (because we have SACK
1271 	 * blocks etc) then our calculated header_len will be different, and
1272 	 * we have to adjust mss_now correspondingly */
1273 	if (header_len != tp->tcp_header_len) {
1274 		int delta = (int) header_len - tp->tcp_header_len;
1275 		mss_now -= delta;
1276 	}
1277 
1278 	return mss_now;
1279 }
1280 
1281 /* Congestion window validation. (RFC2861) */
1282 static void tcp_cwnd_validate(struct sock *sk)
1283 {
1284 	struct tcp_sock *tp = tcp_sk(sk);
1285 
1286 	if (tp->packets_out >= tp->snd_cwnd) {
1287 		/* Network is feed fully. */
1288 		tp->snd_cwnd_used = 0;
1289 		tp->snd_cwnd_stamp = tcp_time_stamp;
1290 	} else {
1291 		/* Network starves. */
1292 		if (tp->packets_out > tp->snd_cwnd_used)
1293 			tp->snd_cwnd_used = tp->packets_out;
1294 
1295 		if (sysctl_tcp_slow_start_after_idle &&
1296 		    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1297 			tcp_cwnd_application_limited(sk);
1298 	}
1299 }
1300 
1301 /* Returns the portion of skb which can be sent right away without
1302  * introducing MSS oddities to segment boundaries. In rare cases where
1303  * mss_now != mss_cache, we will request caller to create a small skb
1304  * per input skb which could be mostly avoided here (if desired).
1305  *
1306  * We explicitly want to create a request for splitting write queue tail
1307  * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1308  * thus all the complexity (cwnd_len is always MSS multiple which we
1309  * return whenever allowed by the other factors). Basically we need the
1310  * modulo only when the receiver window alone is the limiting factor or
1311  * when we would be allowed to send the split-due-to-Nagle skb fully.
1312  */
1313 static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1314 					unsigned int mss_now, unsigned int cwnd)
1315 {
1316 	struct tcp_sock *tp = tcp_sk(sk);
1317 	u32 needed, window, cwnd_len;
1318 
1319 	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1320 	cwnd_len = mss_now * cwnd;
1321 
1322 	if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1323 		return cwnd_len;
1324 
1325 	needed = min(skb->len, window);
1326 
1327 	if (cwnd_len <= needed)
1328 		return cwnd_len;
1329 
1330 	return needed - needed % mss_now;
1331 }
1332 
1333 /* Can at least one segment of SKB be sent right now, according to the
1334  * congestion window rules?  If so, return how many segments are allowed.
1335  */
1336 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1337 					 struct sk_buff *skb)
1338 {
1339 	u32 in_flight, cwnd;
1340 
1341 	/* Don't be strict about the congestion window for the final FIN.  */
1342 	if ((TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && tcp_skb_pcount(skb) == 1)
1343 		return 1;
1344 
1345 	in_flight = tcp_packets_in_flight(tp);
1346 	cwnd = tp->snd_cwnd;
1347 	if (in_flight < cwnd)
1348 		return (cwnd - in_flight);
1349 
1350 	return 0;
1351 }
1352 
1353 /* Initialize TSO state of a skb.
1354  * This must be invoked the first time we consider transmitting
1355  * SKB onto the wire.
1356  */
1357 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1358 			     unsigned int mss_now)
1359 {
1360 	int tso_segs = tcp_skb_pcount(skb);
1361 
1362 	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1363 		tcp_set_skb_tso_segs(sk, skb, mss_now);
1364 		tso_segs = tcp_skb_pcount(skb);
1365 	}
1366 	return tso_segs;
1367 }
1368 
1369 /* Minshall's variant of the Nagle send check. */
1370 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1371 {
1372 	return after(tp->snd_sml, tp->snd_una) &&
1373 		!after(tp->snd_sml, tp->snd_nxt);
1374 }
1375 
1376 /* Return 0, if packet can be sent now without violation Nagle's rules:
1377  * 1. It is full sized.
1378  * 2. Or it contains FIN. (already checked by caller)
1379  * 3. Or TCP_NODELAY was set.
1380  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1381  *    With Minshall's modification: all sent small packets are ACKed.
1382  */
1383 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1384 				  const struct sk_buff *skb,
1385 				  unsigned mss_now, int nonagle)
1386 {
1387 	return skb->len < mss_now &&
1388 		((nonagle & TCP_NAGLE_CORK) ||
1389 		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1390 }
1391 
1392 /* Return non-zero if the Nagle test allows this packet to be
1393  * sent now.
1394  */
1395 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1396 				 unsigned int cur_mss, int nonagle)
1397 {
1398 	/* Nagle rule does not apply to frames, which sit in the middle of the
1399 	 * write_queue (they have no chances to get new data).
1400 	 *
1401 	 * This is implemented in the callers, where they modify the 'nonagle'
1402 	 * argument based upon the location of SKB in the send queue.
1403 	 */
1404 	if (nonagle & TCP_NAGLE_PUSH)
1405 		return 1;
1406 
1407 	/* Don't use the nagle rule for urgent data (or for the final FIN).
1408 	 * Nagle can be ignored during F-RTO too (see RFC4138).
1409 	 */
1410 	if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1411 	    (TCP_SKB_CB(skb)->flags & TCPHDR_FIN))
1412 		return 1;
1413 
1414 	if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1415 		return 1;
1416 
1417 	return 0;
1418 }
1419 
1420 /* Does at least the first segment of SKB fit into the send window? */
1421 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1422 				   unsigned int cur_mss)
1423 {
1424 	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1425 
1426 	if (skb->len > cur_mss)
1427 		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1428 
1429 	return !after(end_seq, tcp_wnd_end(tp));
1430 }
1431 
1432 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1433  * should be put on the wire right now.  If so, it returns the number of
1434  * packets allowed by the congestion window.
1435  */
1436 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1437 				 unsigned int cur_mss, int nonagle)
1438 {
1439 	struct tcp_sock *tp = tcp_sk(sk);
1440 	unsigned int cwnd_quota;
1441 
1442 	tcp_init_tso_segs(sk, skb, cur_mss);
1443 
1444 	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1445 		return 0;
1446 
1447 	cwnd_quota = tcp_cwnd_test(tp, skb);
1448 	if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1449 		cwnd_quota = 0;
1450 
1451 	return cwnd_quota;
1452 }
1453 
1454 /* Test if sending is allowed right now. */
1455 int tcp_may_send_now(struct sock *sk)
1456 {
1457 	struct tcp_sock *tp = tcp_sk(sk);
1458 	struct sk_buff *skb = tcp_send_head(sk);
1459 
1460 	return skb &&
1461 		tcp_snd_test(sk, skb, tcp_current_mss(sk),
1462 			     (tcp_skb_is_last(sk, skb) ?
1463 			      tp->nonagle : TCP_NAGLE_PUSH));
1464 }
1465 
1466 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1467  * which is put after SKB on the list.  It is very much like
1468  * tcp_fragment() except that it may make several kinds of assumptions
1469  * in order to speed up the splitting operation.  In particular, we
1470  * know that all the data is in scatter-gather pages, and that the
1471  * packet has never been sent out before (and thus is not cloned).
1472  */
1473 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1474 			unsigned int mss_now, gfp_t gfp)
1475 {
1476 	struct sk_buff *buff;
1477 	int nlen = skb->len - len;
1478 	u8 flags;
1479 
1480 	/* All of a TSO frame must be composed of paged data.  */
1481 	if (skb->len != skb->data_len)
1482 		return tcp_fragment(sk, skb, len, mss_now);
1483 
1484 	buff = sk_stream_alloc_skb(sk, 0, gfp);
1485 	if (unlikely(buff == NULL))
1486 		return -ENOMEM;
1487 
1488 	sk->sk_wmem_queued += buff->truesize;
1489 	sk_mem_charge(sk, buff->truesize);
1490 	buff->truesize += nlen;
1491 	skb->truesize -= nlen;
1492 
1493 	/* Correct the sequence numbers. */
1494 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1495 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1496 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1497 
1498 	/* PSH and FIN should only be set in the second packet. */
1499 	flags = TCP_SKB_CB(skb)->flags;
1500 	TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1501 	TCP_SKB_CB(buff)->flags = flags;
1502 
1503 	/* This packet was never sent out yet, so no SACK bits. */
1504 	TCP_SKB_CB(buff)->sacked = 0;
1505 
1506 	buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1507 	skb_split(skb, buff, len);
1508 
1509 	/* Fix up tso_factor for both original and new SKB.  */
1510 	tcp_set_skb_tso_segs(sk, skb, mss_now);
1511 	tcp_set_skb_tso_segs(sk, buff, mss_now);
1512 
1513 	/* Link BUFF into the send queue. */
1514 	skb_header_release(buff);
1515 	tcp_insert_write_queue_after(skb, buff, sk);
1516 
1517 	return 0;
1518 }
1519 
1520 /* Try to defer sending, if possible, in order to minimize the amount
1521  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1522  *
1523  * This algorithm is from John Heffner.
1524  */
1525 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1526 {
1527 	struct tcp_sock *tp = tcp_sk(sk);
1528 	const struct inet_connection_sock *icsk = inet_csk(sk);
1529 	u32 send_win, cong_win, limit, in_flight;
1530 	int win_divisor;
1531 
1532 	if (TCP_SKB_CB(skb)->flags & TCPHDR_FIN)
1533 		goto send_now;
1534 
1535 	if (icsk->icsk_ca_state != TCP_CA_Open)
1536 		goto send_now;
1537 
1538 	/* Defer for less than two clock ticks. */
1539 	if (tp->tso_deferred &&
1540 	    (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1541 		goto send_now;
1542 
1543 	in_flight = tcp_packets_in_flight(tp);
1544 
1545 	BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1546 
1547 	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1548 
1549 	/* From in_flight test above, we know that cwnd > in_flight.  */
1550 	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1551 
1552 	limit = min(send_win, cong_win);
1553 
1554 	/* If a full-sized TSO skb can be sent, do it. */
1555 	if (limit >= sk->sk_gso_max_size)
1556 		goto send_now;
1557 
1558 	/* Middle in queue won't get any more data, full sendable already? */
1559 	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1560 		goto send_now;
1561 
1562 	win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1563 	if (win_divisor) {
1564 		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1565 
1566 		/* If at least some fraction of a window is available,
1567 		 * just use it.
1568 		 */
1569 		chunk /= win_divisor;
1570 		if (limit >= chunk)
1571 			goto send_now;
1572 	} else {
1573 		/* Different approach, try not to defer past a single
1574 		 * ACK.  Receiver should ACK every other full sized
1575 		 * frame, so if we have space for more than 3 frames
1576 		 * then send now.
1577 		 */
1578 		if (limit > tcp_max_burst(tp) * tp->mss_cache)
1579 			goto send_now;
1580 	}
1581 
1582 	/* Ok, it looks like it is advisable to defer.  */
1583 	tp->tso_deferred = 1 | (jiffies << 1);
1584 
1585 	return 1;
1586 
1587 send_now:
1588 	tp->tso_deferred = 0;
1589 	return 0;
1590 }
1591 
1592 /* Create a new MTU probe if we are ready.
1593  * MTU probe is regularly attempting to increase the path MTU by
1594  * deliberately sending larger packets.  This discovers routing
1595  * changes resulting in larger path MTUs.
1596  *
1597  * Returns 0 if we should wait to probe (no cwnd available),
1598  *         1 if a probe was sent,
1599  *         -1 otherwise
1600  */
1601 static int tcp_mtu_probe(struct sock *sk)
1602 {
1603 	struct tcp_sock *tp = tcp_sk(sk);
1604 	struct inet_connection_sock *icsk = inet_csk(sk);
1605 	struct sk_buff *skb, *nskb, *next;
1606 	int len;
1607 	int probe_size;
1608 	int size_needed;
1609 	int copy;
1610 	int mss_now;
1611 
1612 	/* Not currently probing/verifying,
1613 	 * not in recovery,
1614 	 * have enough cwnd, and
1615 	 * not SACKing (the variable headers throw things off) */
1616 	if (!icsk->icsk_mtup.enabled ||
1617 	    icsk->icsk_mtup.probe_size ||
1618 	    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1619 	    tp->snd_cwnd < 11 ||
1620 	    tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1621 		return -1;
1622 
1623 	/* Very simple search strategy: just double the MSS. */
1624 	mss_now = tcp_current_mss(sk);
1625 	probe_size = 2 * tp->mss_cache;
1626 	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1627 	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1628 		/* TODO: set timer for probe_converge_event */
1629 		return -1;
1630 	}
1631 
1632 	/* Have enough data in the send queue to probe? */
1633 	if (tp->write_seq - tp->snd_nxt < size_needed)
1634 		return -1;
1635 
1636 	if (tp->snd_wnd < size_needed)
1637 		return -1;
1638 	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1639 		return 0;
1640 
1641 	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
1642 	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1643 		if (!tcp_packets_in_flight(tp))
1644 			return -1;
1645 		else
1646 			return 0;
1647 	}
1648 
1649 	/* We're allowed to probe.  Build it now. */
1650 	if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1651 		return -1;
1652 	sk->sk_wmem_queued += nskb->truesize;
1653 	sk_mem_charge(sk, nskb->truesize);
1654 
1655 	skb = tcp_send_head(sk);
1656 
1657 	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1658 	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1659 	TCP_SKB_CB(nskb)->flags = TCPHDR_ACK;
1660 	TCP_SKB_CB(nskb)->sacked = 0;
1661 	nskb->csum = 0;
1662 	nskb->ip_summed = skb->ip_summed;
1663 
1664 	tcp_insert_write_queue_before(nskb, skb, sk);
1665 
1666 	len = 0;
1667 	tcp_for_write_queue_from_safe(skb, next, sk) {
1668 		copy = min_t(int, skb->len, probe_size - len);
1669 		if (nskb->ip_summed)
1670 			skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1671 		else
1672 			nskb->csum = skb_copy_and_csum_bits(skb, 0,
1673 							    skb_put(nskb, copy),
1674 							    copy, nskb->csum);
1675 
1676 		if (skb->len <= copy) {
1677 			/* We've eaten all the data from this skb.
1678 			 * Throw it away. */
1679 			TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1680 			tcp_unlink_write_queue(skb, sk);
1681 			sk_wmem_free_skb(sk, skb);
1682 		} else {
1683 			TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1684 						   ~(TCPHDR_FIN|TCPHDR_PSH);
1685 			if (!skb_shinfo(skb)->nr_frags) {
1686 				skb_pull(skb, copy);
1687 				if (skb->ip_summed != CHECKSUM_PARTIAL)
1688 					skb->csum = csum_partial(skb->data,
1689 								 skb->len, 0);
1690 			} else {
1691 				__pskb_trim_head(skb, copy);
1692 				tcp_set_skb_tso_segs(sk, skb, mss_now);
1693 			}
1694 			TCP_SKB_CB(skb)->seq += copy;
1695 		}
1696 
1697 		len += copy;
1698 
1699 		if (len >= probe_size)
1700 			break;
1701 	}
1702 	tcp_init_tso_segs(sk, nskb, nskb->len);
1703 
1704 	/* We're ready to send.  If this fails, the probe will
1705 	 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1706 	TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1707 	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1708 		/* Decrement cwnd here because we are sending
1709 		 * effectively two packets. */
1710 		tp->snd_cwnd--;
1711 		tcp_event_new_data_sent(sk, nskb);
1712 
1713 		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1714 		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1715 		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1716 
1717 		return 1;
1718 	}
1719 
1720 	return -1;
1721 }
1722 
1723 /* This routine writes packets to the network.  It advances the
1724  * send_head.  This happens as incoming acks open up the remote
1725  * window for us.
1726  *
1727  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1728  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1729  * account rare use of URG, this is not a big flaw.
1730  *
1731  * Returns 1, if no segments are in flight and we have queued segments, but
1732  * cannot send anything now because of SWS or another problem.
1733  */
1734 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1735 			  int push_one, gfp_t gfp)
1736 {
1737 	struct tcp_sock *tp = tcp_sk(sk);
1738 	struct sk_buff *skb;
1739 	unsigned int tso_segs, sent_pkts;
1740 	int cwnd_quota;
1741 	int result;
1742 
1743 	sent_pkts = 0;
1744 
1745 	if (!push_one) {
1746 		/* Do MTU probing. */
1747 		result = tcp_mtu_probe(sk);
1748 		if (!result) {
1749 			return 0;
1750 		} else if (result > 0) {
1751 			sent_pkts = 1;
1752 		}
1753 	}
1754 
1755 	while ((skb = tcp_send_head(sk))) {
1756 		unsigned int limit;
1757 
1758 		tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1759 		BUG_ON(!tso_segs);
1760 
1761 		cwnd_quota = tcp_cwnd_test(tp, skb);
1762 		if (!cwnd_quota)
1763 			break;
1764 
1765 		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1766 			break;
1767 
1768 		if (tso_segs == 1) {
1769 			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1770 						     (tcp_skb_is_last(sk, skb) ?
1771 						      nonagle : TCP_NAGLE_PUSH))))
1772 				break;
1773 		} else {
1774 			if (!push_one && tcp_tso_should_defer(sk, skb))
1775 				break;
1776 		}
1777 
1778 		limit = mss_now;
1779 		if (tso_segs > 1 && !tcp_urg_mode(tp))
1780 			limit = tcp_mss_split_point(sk, skb, mss_now,
1781 						    cwnd_quota);
1782 
1783 		if (skb->len > limit &&
1784 		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1785 			break;
1786 
1787 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
1788 
1789 		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1790 			break;
1791 
1792 		/* Advance the send_head.  This one is sent out.
1793 		 * This call will increment packets_out.
1794 		 */
1795 		tcp_event_new_data_sent(sk, skb);
1796 
1797 		tcp_minshall_update(tp, mss_now, skb);
1798 		sent_pkts++;
1799 
1800 		if (push_one)
1801 			break;
1802 	}
1803 
1804 	if (likely(sent_pkts)) {
1805 		tcp_cwnd_validate(sk);
1806 		return 0;
1807 	}
1808 	return !tp->packets_out && tcp_send_head(sk);
1809 }
1810 
1811 /* Push out any pending frames which were held back due to
1812  * TCP_CORK or attempt at coalescing tiny packets.
1813  * The socket must be locked by the caller.
1814  */
1815 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1816 			       int nonagle)
1817 {
1818 	/* If we are closed, the bytes will have to remain here.
1819 	 * In time closedown will finish, we empty the write queue and
1820 	 * all will be happy.
1821 	 */
1822 	if (unlikely(sk->sk_state == TCP_CLOSE))
1823 		return;
1824 
1825 	if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
1826 		tcp_check_probe_timer(sk);
1827 }
1828 
1829 /* Send _single_ skb sitting at the send head. This function requires
1830  * true push pending frames to setup probe timer etc.
1831  */
1832 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1833 {
1834 	struct sk_buff *skb = tcp_send_head(sk);
1835 
1836 	BUG_ON(!skb || skb->len < mss_now);
1837 
1838 	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
1839 }
1840 
1841 /* This function returns the amount that we can raise the
1842  * usable window based on the following constraints
1843  *
1844  * 1. The window can never be shrunk once it is offered (RFC 793)
1845  * 2. We limit memory per socket
1846  *
1847  * RFC 1122:
1848  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1849  *  RECV.NEXT + RCV.WIN fixed until:
1850  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1851  *
1852  * i.e. don't raise the right edge of the window until you can raise
1853  * it at least MSS bytes.
1854  *
1855  * Unfortunately, the recommended algorithm breaks header prediction,
1856  * since header prediction assumes th->window stays fixed.
1857  *
1858  * Strictly speaking, keeping th->window fixed violates the receiver
1859  * side SWS prevention criteria. The problem is that under this rule
1860  * a stream of single byte packets will cause the right side of the
1861  * window to always advance by a single byte.
1862  *
1863  * Of course, if the sender implements sender side SWS prevention
1864  * then this will not be a problem.
1865  *
1866  * BSD seems to make the following compromise:
1867  *
1868  *	If the free space is less than the 1/4 of the maximum
1869  *	space available and the free space is less than 1/2 mss,
1870  *	then set the window to 0.
1871  *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1872  *	Otherwise, just prevent the window from shrinking
1873  *	and from being larger than the largest representable value.
1874  *
1875  * This prevents incremental opening of the window in the regime
1876  * where TCP is limited by the speed of the reader side taking
1877  * data out of the TCP receive queue. It does nothing about
1878  * those cases where the window is constrained on the sender side
1879  * because the pipeline is full.
1880  *
1881  * BSD also seems to "accidentally" limit itself to windows that are a
1882  * multiple of MSS, at least until the free space gets quite small.
1883  * This would appear to be a side effect of the mbuf implementation.
1884  * Combining these two algorithms results in the observed behavior
1885  * of having a fixed window size at almost all times.
1886  *
1887  * Below we obtain similar behavior by forcing the offered window to
1888  * a multiple of the mss when it is feasible to do so.
1889  *
1890  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1891  * Regular options like TIMESTAMP are taken into account.
1892  */
1893 u32 __tcp_select_window(struct sock *sk)
1894 {
1895 	struct inet_connection_sock *icsk = inet_csk(sk);
1896 	struct tcp_sock *tp = tcp_sk(sk);
1897 	/* MSS for the peer's data.  Previous versions used mss_clamp
1898 	 * here.  I don't know if the value based on our guesses
1899 	 * of peer's MSS is better for the performance.  It's more correct
1900 	 * but may be worse for the performance because of rcv_mss
1901 	 * fluctuations.  --SAW  1998/11/1
1902 	 */
1903 	int mss = icsk->icsk_ack.rcv_mss;
1904 	int free_space = tcp_space(sk);
1905 	int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1906 	int window;
1907 
1908 	if (mss > full_space)
1909 		mss = full_space;
1910 
1911 	if (free_space < (full_space >> 1)) {
1912 		icsk->icsk_ack.quick = 0;
1913 
1914 		if (tcp_memory_pressure)
1915 			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1916 					       4U * tp->advmss);
1917 
1918 		if (free_space < mss)
1919 			return 0;
1920 	}
1921 
1922 	if (free_space > tp->rcv_ssthresh)
1923 		free_space = tp->rcv_ssthresh;
1924 
1925 	/* Don't do rounding if we are using window scaling, since the
1926 	 * scaled window will not line up with the MSS boundary anyway.
1927 	 */
1928 	window = tp->rcv_wnd;
1929 	if (tp->rx_opt.rcv_wscale) {
1930 		window = free_space;
1931 
1932 		/* Advertise enough space so that it won't get scaled away.
1933 		 * Import case: prevent zero window announcement if
1934 		 * 1<<rcv_wscale > mss.
1935 		 */
1936 		if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1937 			window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1938 				  << tp->rx_opt.rcv_wscale);
1939 	} else {
1940 		/* Get the largest window that is a nice multiple of mss.
1941 		 * Window clamp already applied above.
1942 		 * If our current window offering is within 1 mss of the
1943 		 * free space we just keep it. This prevents the divide
1944 		 * and multiply from happening most of the time.
1945 		 * We also don't do any window rounding when the free space
1946 		 * is too small.
1947 		 */
1948 		if (window <= free_space - mss || window > free_space)
1949 			window = (free_space / mss) * mss;
1950 		else if (mss == full_space &&
1951 			 free_space > window + (full_space >> 1))
1952 			window = free_space;
1953 	}
1954 
1955 	return window;
1956 }
1957 
1958 /* Collapses two adjacent SKB's during retransmission. */
1959 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1960 {
1961 	struct tcp_sock *tp = tcp_sk(sk);
1962 	struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1963 	int skb_size, next_skb_size;
1964 
1965 	skb_size = skb->len;
1966 	next_skb_size = next_skb->len;
1967 
1968 	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1969 
1970 	tcp_highest_sack_combine(sk, next_skb, skb);
1971 
1972 	tcp_unlink_write_queue(next_skb, sk);
1973 
1974 	skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1975 				  next_skb_size);
1976 
1977 	if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1978 		skb->ip_summed = CHECKSUM_PARTIAL;
1979 
1980 	if (skb->ip_summed != CHECKSUM_PARTIAL)
1981 		skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1982 
1983 	/* Update sequence range on original skb. */
1984 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1985 
1986 	/* Merge over control information. This moves PSH/FIN etc. over */
1987 	TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags;
1988 
1989 	/* All done, get rid of second SKB and account for it so
1990 	 * packet counting does not break.
1991 	 */
1992 	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1993 
1994 	/* changed transmit queue under us so clear hints */
1995 	tcp_clear_retrans_hints_partial(tp);
1996 	if (next_skb == tp->retransmit_skb_hint)
1997 		tp->retransmit_skb_hint = skb;
1998 
1999 	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2000 
2001 	sk_wmem_free_skb(sk, next_skb);
2002 }
2003 
2004 /* Check if coalescing SKBs is legal. */
2005 static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
2006 {
2007 	if (tcp_skb_pcount(skb) > 1)
2008 		return 0;
2009 	/* TODO: SACK collapsing could be used to remove this condition */
2010 	if (skb_shinfo(skb)->nr_frags != 0)
2011 		return 0;
2012 	if (skb_cloned(skb))
2013 		return 0;
2014 	if (skb == tcp_send_head(sk))
2015 		return 0;
2016 	/* Some heurestics for collapsing over SACK'd could be invented */
2017 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2018 		return 0;
2019 
2020 	return 1;
2021 }
2022 
2023 /* Collapse packets in the retransmit queue to make to create
2024  * less packets on the wire. This is only done on retransmission.
2025  */
2026 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2027 				     int space)
2028 {
2029 	struct tcp_sock *tp = tcp_sk(sk);
2030 	struct sk_buff *skb = to, *tmp;
2031 	int first = 1;
2032 
2033 	if (!sysctl_tcp_retrans_collapse)
2034 		return;
2035 	if (TCP_SKB_CB(skb)->flags & TCPHDR_SYN)
2036 		return;
2037 
2038 	tcp_for_write_queue_from_safe(skb, tmp, sk) {
2039 		if (!tcp_can_collapse(sk, skb))
2040 			break;
2041 
2042 		space -= skb->len;
2043 
2044 		if (first) {
2045 			first = 0;
2046 			continue;
2047 		}
2048 
2049 		if (space < 0)
2050 			break;
2051 		/* Punt if not enough space exists in the first SKB for
2052 		 * the data in the second
2053 		 */
2054 		if (skb->len > skb_tailroom(to))
2055 			break;
2056 
2057 		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2058 			break;
2059 
2060 		tcp_collapse_retrans(sk, to);
2061 	}
2062 }
2063 
2064 /* This retransmits one SKB.  Policy decisions and retransmit queue
2065  * state updates are done by the caller.  Returns non-zero if an
2066  * error occurred which prevented the send.
2067  */
2068 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2069 {
2070 	struct tcp_sock *tp = tcp_sk(sk);
2071 	struct inet_connection_sock *icsk = inet_csk(sk);
2072 	unsigned int cur_mss;
2073 	int err;
2074 
2075 	/* Inconslusive MTU probe */
2076 	if (icsk->icsk_mtup.probe_size) {
2077 		icsk->icsk_mtup.probe_size = 0;
2078 	}
2079 
2080 	/* Do not sent more than we queued. 1/4 is reserved for possible
2081 	 * copying overhead: fragmentation, tunneling, mangling etc.
2082 	 */
2083 	if (atomic_read(&sk->sk_wmem_alloc) >
2084 	    min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2085 		return -EAGAIN;
2086 
2087 	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2088 		if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2089 			BUG();
2090 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2091 			return -ENOMEM;
2092 	}
2093 
2094 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2095 		return -EHOSTUNREACH; /* Routing failure or similar. */
2096 
2097 	cur_mss = tcp_current_mss(sk);
2098 
2099 	/* If receiver has shrunk his window, and skb is out of
2100 	 * new window, do not retransmit it. The exception is the
2101 	 * case, when window is shrunk to zero. In this case
2102 	 * our retransmit serves as a zero window probe.
2103 	 */
2104 	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2105 	    TCP_SKB_CB(skb)->seq != tp->snd_una)
2106 		return -EAGAIN;
2107 
2108 	if (skb->len > cur_mss) {
2109 		if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2110 			return -ENOMEM; /* We'll try again later. */
2111 	} else {
2112 		int oldpcount = tcp_skb_pcount(skb);
2113 
2114 		if (unlikely(oldpcount > 1)) {
2115 			tcp_init_tso_segs(sk, skb, cur_mss);
2116 			tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2117 		}
2118 	}
2119 
2120 	tcp_retrans_try_collapse(sk, skb, cur_mss);
2121 
2122 	/* Some Solaris stacks overoptimize and ignore the FIN on a
2123 	 * retransmit when old data is attached.  So strip it off
2124 	 * since it is cheap to do so and saves bytes on the network.
2125 	 */
2126 	if (skb->len > 0 &&
2127 	    (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) &&
2128 	    tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2129 		if (!pskb_trim(skb, 0)) {
2130 			/* Reuse, even though it does some unnecessary work */
2131 			tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2132 					     TCP_SKB_CB(skb)->flags);
2133 			skb->ip_summed = CHECKSUM_NONE;
2134 		}
2135 	}
2136 
2137 	/* Make a copy, if the first transmission SKB clone we made
2138 	 * is still in somebody's hands, else make a clone.
2139 	 */
2140 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2141 
2142 	err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2143 
2144 	if (err == 0) {
2145 		/* Update global TCP statistics. */
2146 		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2147 
2148 		tp->total_retrans++;
2149 
2150 #if FASTRETRANS_DEBUG > 0
2151 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2152 			if (net_ratelimit())
2153 				printk(KERN_DEBUG "retrans_out leaked.\n");
2154 		}
2155 #endif
2156 		if (!tp->retrans_out)
2157 			tp->lost_retrans_low = tp->snd_nxt;
2158 		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2159 		tp->retrans_out += tcp_skb_pcount(skb);
2160 
2161 		/* Save stamp of the first retransmit. */
2162 		if (!tp->retrans_stamp)
2163 			tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2164 
2165 		tp->undo_retrans += tcp_skb_pcount(skb);
2166 
2167 		/* snd_nxt is stored to detect loss of retransmitted segment,
2168 		 * see tcp_input.c tcp_sacktag_write_queue().
2169 		 */
2170 		TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2171 	}
2172 	return err;
2173 }
2174 
2175 /* Check if we forward retransmits are possible in the current
2176  * window/congestion state.
2177  */
2178 static int tcp_can_forward_retransmit(struct sock *sk)
2179 {
2180 	const struct inet_connection_sock *icsk = inet_csk(sk);
2181 	struct tcp_sock *tp = tcp_sk(sk);
2182 
2183 	/* Forward retransmissions are possible only during Recovery. */
2184 	if (icsk->icsk_ca_state != TCP_CA_Recovery)
2185 		return 0;
2186 
2187 	/* No forward retransmissions in Reno are possible. */
2188 	if (tcp_is_reno(tp))
2189 		return 0;
2190 
2191 	/* Yeah, we have to make difficult choice between forward transmission
2192 	 * and retransmission... Both ways have their merits...
2193 	 *
2194 	 * For now we do not retransmit anything, while we have some new
2195 	 * segments to send. In the other cases, follow rule 3 for
2196 	 * NextSeg() specified in RFC3517.
2197 	 */
2198 
2199 	if (tcp_may_send_now(sk))
2200 		return 0;
2201 
2202 	return 1;
2203 }
2204 
2205 /* This gets called after a retransmit timeout, and the initially
2206  * retransmitted data is acknowledged.  It tries to continue
2207  * resending the rest of the retransmit queue, until either
2208  * we've sent it all or the congestion window limit is reached.
2209  * If doing SACK, the first ACK which comes back for a timeout
2210  * based retransmit packet might feed us FACK information again.
2211  * If so, we use it to avoid unnecessarily retransmissions.
2212  */
2213 void tcp_xmit_retransmit_queue(struct sock *sk)
2214 {
2215 	const struct inet_connection_sock *icsk = inet_csk(sk);
2216 	struct tcp_sock *tp = tcp_sk(sk);
2217 	struct sk_buff *skb;
2218 	struct sk_buff *hole = NULL;
2219 	u32 last_lost;
2220 	int mib_idx;
2221 	int fwd_rexmitting = 0;
2222 
2223 	if (!tp->packets_out)
2224 		return;
2225 
2226 	if (!tp->lost_out)
2227 		tp->retransmit_high = tp->snd_una;
2228 
2229 	if (tp->retransmit_skb_hint) {
2230 		skb = tp->retransmit_skb_hint;
2231 		last_lost = TCP_SKB_CB(skb)->end_seq;
2232 		if (after(last_lost, tp->retransmit_high))
2233 			last_lost = tp->retransmit_high;
2234 	} else {
2235 		skb = tcp_write_queue_head(sk);
2236 		last_lost = tp->snd_una;
2237 	}
2238 
2239 	tcp_for_write_queue_from(skb, sk) {
2240 		__u8 sacked = TCP_SKB_CB(skb)->sacked;
2241 
2242 		if (skb == tcp_send_head(sk))
2243 			break;
2244 		/* we could do better than to assign each time */
2245 		if (hole == NULL)
2246 			tp->retransmit_skb_hint = skb;
2247 
2248 		/* Assume this retransmit will generate
2249 		 * only one packet for congestion window
2250 		 * calculation purposes.  This works because
2251 		 * tcp_retransmit_skb() will chop up the
2252 		 * packet to be MSS sized and all the
2253 		 * packet counting works out.
2254 		 */
2255 		if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2256 			return;
2257 
2258 		if (fwd_rexmitting) {
2259 begin_fwd:
2260 			if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2261 				break;
2262 			mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2263 
2264 		} else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2265 			tp->retransmit_high = last_lost;
2266 			if (!tcp_can_forward_retransmit(sk))
2267 				break;
2268 			/* Backtrack if necessary to non-L'ed skb */
2269 			if (hole != NULL) {
2270 				skb = hole;
2271 				hole = NULL;
2272 			}
2273 			fwd_rexmitting = 1;
2274 			goto begin_fwd;
2275 
2276 		} else if (!(sacked & TCPCB_LOST)) {
2277 			if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2278 				hole = skb;
2279 			continue;
2280 
2281 		} else {
2282 			last_lost = TCP_SKB_CB(skb)->end_seq;
2283 			if (icsk->icsk_ca_state != TCP_CA_Loss)
2284 				mib_idx = LINUX_MIB_TCPFASTRETRANS;
2285 			else
2286 				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2287 		}
2288 
2289 		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2290 			continue;
2291 
2292 		if (tcp_retransmit_skb(sk, skb))
2293 			return;
2294 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
2295 
2296 		if (skb == tcp_write_queue_head(sk))
2297 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2298 						  inet_csk(sk)->icsk_rto,
2299 						  TCP_RTO_MAX);
2300 	}
2301 }
2302 
2303 /* Send a fin.  The caller locks the socket for us.  This cannot be
2304  * allowed to fail queueing a FIN frame under any circumstances.
2305  */
2306 void tcp_send_fin(struct sock *sk)
2307 {
2308 	struct tcp_sock *tp = tcp_sk(sk);
2309 	struct sk_buff *skb = tcp_write_queue_tail(sk);
2310 	int mss_now;
2311 
2312 	/* Optimization, tack on the FIN if we have a queue of
2313 	 * unsent frames.  But be careful about outgoing SACKS
2314 	 * and IP options.
2315 	 */
2316 	mss_now = tcp_current_mss(sk);
2317 
2318 	if (tcp_send_head(sk) != NULL) {
2319 		TCP_SKB_CB(skb)->flags |= TCPHDR_FIN;
2320 		TCP_SKB_CB(skb)->end_seq++;
2321 		tp->write_seq++;
2322 	} else {
2323 		/* Socket is locked, keep trying until memory is available. */
2324 		for (;;) {
2325 			skb = alloc_skb_fclone(MAX_TCP_HEADER,
2326 					       sk->sk_allocation);
2327 			if (skb)
2328 				break;
2329 			yield();
2330 		}
2331 
2332 		/* Reserve space for headers and prepare control bits. */
2333 		skb_reserve(skb, MAX_TCP_HEADER);
2334 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2335 		tcp_init_nondata_skb(skb, tp->write_seq,
2336 				     TCPHDR_ACK | TCPHDR_FIN);
2337 		tcp_queue_skb(sk, skb);
2338 	}
2339 	__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2340 }
2341 
2342 /* We get here when a process closes a file descriptor (either due to
2343  * an explicit close() or as a byproduct of exit()'ing) and there
2344  * was unread data in the receive queue.  This behavior is recommended
2345  * by RFC 2525, section 2.17.  -DaveM
2346  */
2347 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2348 {
2349 	struct sk_buff *skb;
2350 
2351 	/* NOTE: No TCP options attached and we never retransmit this. */
2352 	skb = alloc_skb(MAX_TCP_HEADER, priority);
2353 	if (!skb) {
2354 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2355 		return;
2356 	}
2357 
2358 	/* Reserve space for headers and prepare control bits. */
2359 	skb_reserve(skb, MAX_TCP_HEADER);
2360 	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2361 			     TCPHDR_ACK | TCPHDR_RST);
2362 	/* Send it off. */
2363 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2364 	if (tcp_transmit_skb(sk, skb, 0, priority))
2365 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2366 
2367 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2368 }
2369 
2370 /* Send a crossed SYN-ACK during socket establishment.
2371  * WARNING: This routine must only be called when we have already sent
2372  * a SYN packet that crossed the incoming SYN that caused this routine
2373  * to get called. If this assumption fails then the initial rcv_wnd
2374  * and rcv_wscale values will not be correct.
2375  */
2376 int tcp_send_synack(struct sock *sk)
2377 {
2378 	struct sk_buff *skb;
2379 
2380 	skb = tcp_write_queue_head(sk);
2381 	if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPHDR_SYN)) {
2382 		printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2383 		return -EFAULT;
2384 	}
2385 	if (!(TCP_SKB_CB(skb)->flags & TCPHDR_ACK)) {
2386 		if (skb_cloned(skb)) {
2387 			struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2388 			if (nskb == NULL)
2389 				return -ENOMEM;
2390 			tcp_unlink_write_queue(skb, sk);
2391 			skb_header_release(nskb);
2392 			__tcp_add_write_queue_head(sk, nskb);
2393 			sk_wmem_free_skb(sk, skb);
2394 			sk->sk_wmem_queued += nskb->truesize;
2395 			sk_mem_charge(sk, nskb->truesize);
2396 			skb = nskb;
2397 		}
2398 
2399 		TCP_SKB_CB(skb)->flags |= TCPHDR_ACK;
2400 		TCP_ECN_send_synack(tcp_sk(sk), skb);
2401 	}
2402 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2403 	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2404 }
2405 
2406 /* Prepare a SYN-ACK. */
2407 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2408 				struct request_sock *req,
2409 				struct request_values *rvp)
2410 {
2411 	struct tcp_out_options opts;
2412 	struct tcp_extend_values *xvp = tcp_xv(rvp);
2413 	struct inet_request_sock *ireq = inet_rsk(req);
2414 	struct tcp_sock *tp = tcp_sk(sk);
2415 	const struct tcp_cookie_values *cvp = tp->cookie_values;
2416 	struct tcphdr *th;
2417 	struct sk_buff *skb;
2418 	struct tcp_md5sig_key *md5;
2419 	int tcp_header_size;
2420 	int mss;
2421 	int s_data_desired = 0;
2422 
2423 	if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2424 		s_data_desired = cvp->s_data_desired;
2425 	skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
2426 	if (skb == NULL)
2427 		return NULL;
2428 
2429 	/* Reserve space for headers. */
2430 	skb_reserve(skb, MAX_TCP_HEADER);
2431 
2432 	skb_dst_set(skb, dst_clone(dst));
2433 
2434 	mss = dst_metric_advmss(dst);
2435 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2436 		mss = tp->rx_opt.user_mss;
2437 
2438 	if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2439 		__u8 rcv_wscale;
2440 		/* Set this up on the first call only */
2441 		req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2442 
2443 		/* limit the window selection if the user enforce a smaller rx buffer */
2444 		if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2445 		    (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2446 			req->window_clamp = tcp_full_space(sk);
2447 
2448 		/* tcp_full_space because it is guaranteed to be the first packet */
2449 		tcp_select_initial_window(tcp_full_space(sk),
2450 			mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2451 			&req->rcv_wnd,
2452 			&req->window_clamp,
2453 			ireq->wscale_ok,
2454 			&rcv_wscale,
2455 			dst_metric(dst, RTAX_INITRWND));
2456 		ireq->rcv_wscale = rcv_wscale;
2457 	}
2458 
2459 	memset(&opts, 0, sizeof(opts));
2460 #ifdef CONFIG_SYN_COOKIES
2461 	if (unlikely(req->cookie_ts))
2462 		TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2463 	else
2464 #endif
2465 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2466 	tcp_header_size = tcp_synack_options(sk, req, mss,
2467 					     skb, &opts, &md5, xvp)
2468 			+ sizeof(*th);
2469 
2470 	skb_push(skb, tcp_header_size);
2471 	skb_reset_transport_header(skb);
2472 
2473 	th = tcp_hdr(skb);
2474 	memset(th, 0, sizeof(struct tcphdr));
2475 	th->syn = 1;
2476 	th->ack = 1;
2477 	TCP_ECN_make_synack(req, th);
2478 	th->source = ireq->loc_port;
2479 	th->dest = ireq->rmt_port;
2480 	/* Setting of flags are superfluous here for callers (and ECE is
2481 	 * not even correctly set)
2482 	 */
2483 	tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2484 			     TCPHDR_SYN | TCPHDR_ACK);
2485 
2486 	if (OPTION_COOKIE_EXTENSION & opts.options) {
2487 		if (s_data_desired) {
2488 			u8 *buf = skb_put(skb, s_data_desired);
2489 
2490 			/* copy data directly from the listening socket. */
2491 			memcpy(buf, cvp->s_data_payload, s_data_desired);
2492 			TCP_SKB_CB(skb)->end_seq += s_data_desired;
2493 		}
2494 
2495 		if (opts.hash_size > 0) {
2496 			__u32 workspace[SHA_WORKSPACE_WORDS];
2497 			u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2498 			u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2499 
2500 			/* Secret recipe depends on the Timestamp, (future)
2501 			 * Sequence and Acknowledgment Numbers, Initiator
2502 			 * Cookie, and others handled by IP variant caller.
2503 			 */
2504 			*tail-- ^= opts.tsval;
2505 			*tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2506 			*tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2507 
2508 			/* recommended */
2509 			*tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2510 			*tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2511 
2512 			sha_transform((__u32 *)&xvp->cookie_bakery[0],
2513 				      (char *)mess,
2514 				      &workspace[0]);
2515 			opts.hash_location =
2516 				(__u8 *)&xvp->cookie_bakery[0];
2517 		}
2518 	}
2519 
2520 	th->seq = htonl(TCP_SKB_CB(skb)->seq);
2521 	th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2522 
2523 	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2524 	th->window = htons(min(req->rcv_wnd, 65535U));
2525 	tcp_options_write((__be32 *)(th + 1), tp, &opts);
2526 	th->doff = (tcp_header_size >> 2);
2527 	TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2528 
2529 #ifdef CONFIG_TCP_MD5SIG
2530 	/* Okay, we have all we need - do the md5 hash if needed */
2531 	if (md5) {
2532 		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2533 					       md5, NULL, req, skb);
2534 	}
2535 #endif
2536 
2537 	return skb;
2538 }
2539 EXPORT_SYMBOL(tcp_make_synack);
2540 
2541 /* Do all connect socket setups that can be done AF independent. */
2542 static void tcp_connect_init(struct sock *sk)
2543 {
2544 	struct dst_entry *dst = __sk_dst_get(sk);
2545 	struct tcp_sock *tp = tcp_sk(sk);
2546 	__u8 rcv_wscale;
2547 
2548 	/* We'll fix this up when we get a response from the other end.
2549 	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2550 	 */
2551 	tp->tcp_header_len = sizeof(struct tcphdr) +
2552 		(sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2553 
2554 #ifdef CONFIG_TCP_MD5SIG
2555 	if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2556 		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2557 #endif
2558 
2559 	/* If user gave his TCP_MAXSEG, record it to clamp */
2560 	if (tp->rx_opt.user_mss)
2561 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2562 	tp->max_window = 0;
2563 	tcp_mtup_init(sk);
2564 	tcp_sync_mss(sk, dst_mtu(dst));
2565 
2566 	if (!tp->window_clamp)
2567 		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2568 	tp->advmss = dst_metric_advmss(dst);
2569 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2570 		tp->advmss = tp->rx_opt.user_mss;
2571 
2572 	tcp_initialize_rcv_mss(sk);
2573 
2574 	/* limit the window selection if the user enforce a smaller rx buffer */
2575 	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2576 	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2577 		tp->window_clamp = tcp_full_space(sk);
2578 
2579 	tcp_select_initial_window(tcp_full_space(sk),
2580 				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2581 				  &tp->rcv_wnd,
2582 				  &tp->window_clamp,
2583 				  sysctl_tcp_window_scaling,
2584 				  &rcv_wscale,
2585 				  dst_metric(dst, RTAX_INITRWND));
2586 
2587 	tp->rx_opt.rcv_wscale = rcv_wscale;
2588 	tp->rcv_ssthresh = tp->rcv_wnd;
2589 
2590 	sk->sk_err = 0;
2591 	sock_reset_flag(sk, SOCK_DONE);
2592 	tp->snd_wnd = 0;
2593 	tcp_init_wl(tp, 0);
2594 	tp->snd_una = tp->write_seq;
2595 	tp->snd_sml = tp->write_seq;
2596 	tp->snd_up = tp->write_seq;
2597 	tp->rcv_nxt = 0;
2598 	tp->rcv_wup = 0;
2599 	tp->copied_seq = 0;
2600 
2601 	inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2602 	inet_csk(sk)->icsk_retransmits = 0;
2603 	tcp_clear_retrans(tp);
2604 }
2605 
2606 /* Build a SYN and send it off. */
2607 int tcp_connect(struct sock *sk)
2608 {
2609 	struct tcp_sock *tp = tcp_sk(sk);
2610 	struct sk_buff *buff;
2611 	int err;
2612 
2613 	tcp_connect_init(sk);
2614 
2615 	buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2616 	if (unlikely(buff == NULL))
2617 		return -ENOBUFS;
2618 
2619 	/* Reserve space for headers. */
2620 	skb_reserve(buff, MAX_TCP_HEADER);
2621 
2622 	tp->snd_nxt = tp->write_seq;
2623 	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2624 	TCP_ECN_send_syn(sk, buff);
2625 
2626 	/* Send it off. */
2627 	TCP_SKB_CB(buff)->when = tcp_time_stamp;
2628 	tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2629 	skb_header_release(buff);
2630 	__tcp_add_write_queue_tail(sk, buff);
2631 	sk->sk_wmem_queued += buff->truesize;
2632 	sk_mem_charge(sk, buff->truesize);
2633 	tp->packets_out += tcp_skb_pcount(buff);
2634 	err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2635 	if (err == -ECONNREFUSED)
2636 		return err;
2637 
2638 	/* We change tp->snd_nxt after the tcp_transmit_skb() call
2639 	 * in order to make this packet get counted in tcpOutSegs.
2640 	 */
2641 	tp->snd_nxt = tp->write_seq;
2642 	tp->pushed_seq = tp->write_seq;
2643 	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2644 
2645 	/* Timer for repeating the SYN until an answer. */
2646 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2647 				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2648 	return 0;
2649 }
2650 EXPORT_SYMBOL(tcp_connect);
2651 
2652 /* Send out a delayed ack, the caller does the policy checking
2653  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
2654  * for details.
2655  */
2656 void tcp_send_delayed_ack(struct sock *sk)
2657 {
2658 	struct inet_connection_sock *icsk = inet_csk(sk);
2659 	int ato = icsk->icsk_ack.ato;
2660 	unsigned long timeout;
2661 
2662 	if (ato > TCP_DELACK_MIN) {
2663 		const struct tcp_sock *tp = tcp_sk(sk);
2664 		int max_ato = HZ / 2;
2665 
2666 		if (icsk->icsk_ack.pingpong ||
2667 		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2668 			max_ato = TCP_DELACK_MAX;
2669 
2670 		/* Slow path, intersegment interval is "high". */
2671 
2672 		/* If some rtt estimate is known, use it to bound delayed ack.
2673 		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2674 		 * directly.
2675 		 */
2676 		if (tp->srtt) {
2677 			int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2678 
2679 			if (rtt < max_ato)
2680 				max_ato = rtt;
2681 		}
2682 
2683 		ato = min(ato, max_ato);
2684 	}
2685 
2686 	/* Stay within the limit we were given */
2687 	timeout = jiffies + ato;
2688 
2689 	/* Use new timeout only if there wasn't a older one earlier. */
2690 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2691 		/* If delack timer was blocked or is about to expire,
2692 		 * send ACK now.
2693 		 */
2694 		if (icsk->icsk_ack.blocked ||
2695 		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2696 			tcp_send_ack(sk);
2697 			return;
2698 		}
2699 
2700 		if (!time_before(timeout, icsk->icsk_ack.timeout))
2701 			timeout = icsk->icsk_ack.timeout;
2702 	}
2703 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2704 	icsk->icsk_ack.timeout = timeout;
2705 	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2706 }
2707 
2708 /* This routine sends an ack and also updates the window. */
2709 void tcp_send_ack(struct sock *sk)
2710 {
2711 	struct sk_buff *buff;
2712 
2713 	/* If we have been reset, we may not send again. */
2714 	if (sk->sk_state == TCP_CLOSE)
2715 		return;
2716 
2717 	/* We are not putting this on the write queue, so
2718 	 * tcp_transmit_skb() will set the ownership to this
2719 	 * sock.
2720 	 */
2721 	buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2722 	if (buff == NULL) {
2723 		inet_csk_schedule_ack(sk);
2724 		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2725 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2726 					  TCP_DELACK_MAX, TCP_RTO_MAX);
2727 		return;
2728 	}
2729 
2730 	/* Reserve space for headers and prepare control bits. */
2731 	skb_reserve(buff, MAX_TCP_HEADER);
2732 	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
2733 
2734 	/* Send it off, this clears delayed acks for us. */
2735 	TCP_SKB_CB(buff)->when = tcp_time_stamp;
2736 	tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2737 }
2738 
2739 /* This routine sends a packet with an out of date sequence
2740  * number. It assumes the other end will try to ack it.
2741  *
2742  * Question: what should we make while urgent mode?
2743  * 4.4BSD forces sending single byte of data. We cannot send
2744  * out of window data, because we have SND.NXT==SND.MAX...
2745  *
2746  * Current solution: to send TWO zero-length segments in urgent mode:
2747  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2748  * out-of-date with SND.UNA-1 to probe window.
2749  */
2750 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2751 {
2752 	struct tcp_sock *tp = tcp_sk(sk);
2753 	struct sk_buff *skb;
2754 
2755 	/* We don't queue it, tcp_transmit_skb() sets ownership. */
2756 	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2757 	if (skb == NULL)
2758 		return -1;
2759 
2760 	/* Reserve space for headers and set control bits. */
2761 	skb_reserve(skb, MAX_TCP_HEADER);
2762 	/* Use a previous sequence.  This should cause the other
2763 	 * end to send an ack.  Don't queue or clone SKB, just
2764 	 * send it.
2765 	 */
2766 	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
2767 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2768 	return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2769 }
2770 
2771 /* Initiate keepalive or window probe from timer. */
2772 int tcp_write_wakeup(struct sock *sk)
2773 {
2774 	struct tcp_sock *tp = tcp_sk(sk);
2775 	struct sk_buff *skb;
2776 
2777 	if (sk->sk_state == TCP_CLOSE)
2778 		return -1;
2779 
2780 	if ((skb = tcp_send_head(sk)) != NULL &&
2781 	    before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2782 		int err;
2783 		unsigned int mss = tcp_current_mss(sk);
2784 		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2785 
2786 		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2787 			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2788 
2789 		/* We are probing the opening of a window
2790 		 * but the window size is != 0
2791 		 * must have been a result SWS avoidance ( sender )
2792 		 */
2793 		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2794 		    skb->len > mss) {
2795 			seg_size = min(seg_size, mss);
2796 			TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
2797 			if (tcp_fragment(sk, skb, seg_size, mss))
2798 				return -1;
2799 		} else if (!tcp_skb_pcount(skb))
2800 			tcp_set_skb_tso_segs(sk, skb, mss);
2801 
2802 		TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
2803 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
2804 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2805 		if (!err)
2806 			tcp_event_new_data_sent(sk, skb);
2807 		return err;
2808 	} else {
2809 		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2810 			tcp_xmit_probe_skb(sk, 1);
2811 		return tcp_xmit_probe_skb(sk, 0);
2812 	}
2813 }
2814 
2815 /* A window probe timeout has occurred.  If window is not closed send
2816  * a partial packet else a zero probe.
2817  */
2818 void tcp_send_probe0(struct sock *sk)
2819 {
2820 	struct inet_connection_sock *icsk = inet_csk(sk);
2821 	struct tcp_sock *tp = tcp_sk(sk);
2822 	int err;
2823 
2824 	err = tcp_write_wakeup(sk);
2825 
2826 	if (tp->packets_out || !tcp_send_head(sk)) {
2827 		/* Cancel probe timer, if it is not required. */
2828 		icsk->icsk_probes_out = 0;
2829 		icsk->icsk_backoff = 0;
2830 		return;
2831 	}
2832 
2833 	if (err <= 0) {
2834 		if (icsk->icsk_backoff < sysctl_tcp_retries2)
2835 			icsk->icsk_backoff++;
2836 		icsk->icsk_probes_out++;
2837 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2838 					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2839 					  TCP_RTO_MAX);
2840 	} else {
2841 		/* If packet was not sent due to local congestion,
2842 		 * do not backoff and do not remember icsk_probes_out.
2843 		 * Let local senders to fight for local resources.
2844 		 *
2845 		 * Use accumulated backoff yet.
2846 		 */
2847 		if (!icsk->icsk_probes_out)
2848 			icsk->icsk_probes_out = 1;
2849 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2850 					  min(icsk->icsk_rto << icsk->icsk_backoff,
2851 					      TCP_RESOURCE_PROBE_INTERVAL),
2852 					  TCP_RTO_MAX);
2853 	}
2854 }
2855