xref: /openbmc/linux/net/ipv4/tcp_output.c (revision d7a3d85e)
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 #define pr_fmt(fmt) "TCP: " fmt
38 
39 #include <net/tcp.h>
40 
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
44 
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly = 1;
47 
48 /* People can turn this on to work with those rare, broken TCPs that
49  * interpret the window field as a signed quantity.
50  */
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
52 
53 /* Default TSQ limit of two TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 131072;
55 
56 /* This limits the percentage of the congestion window which we
57  * will allow a single TSO frame to consume.  Building TSO frames
58  * which are too large can cause TCP streams to be bursty.
59  */
60 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
61 
62 /* By default, RFC2861 behavior.  */
63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
64 
65 unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX;
66 EXPORT_SYMBOL(sysctl_tcp_notsent_lowat);
67 
68 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
69 			   int push_one, gfp_t gfp);
70 
71 /* Account for new data that has been sent to the network. */
72 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
73 {
74 	struct inet_connection_sock *icsk = inet_csk(sk);
75 	struct tcp_sock *tp = tcp_sk(sk);
76 	unsigned int prior_packets = tp->packets_out;
77 
78 	tcp_advance_send_head(sk, skb);
79 	tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
80 
81 	tp->packets_out += tcp_skb_pcount(skb);
82 	if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
83 	    icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
84 		tcp_rearm_rto(sk);
85 	}
86 
87 	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
88 		      tcp_skb_pcount(skb));
89 }
90 
91 /* SND.NXT, if window was not shrunk.
92  * If window has been shrunk, what should we make? It is not clear at all.
93  * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
94  * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
95  * invalid. OK, let's make this for now:
96  */
97 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
98 {
99 	const struct tcp_sock *tp = tcp_sk(sk);
100 
101 	if (!before(tcp_wnd_end(tp), tp->snd_nxt))
102 		return tp->snd_nxt;
103 	else
104 		return tcp_wnd_end(tp);
105 }
106 
107 /* Calculate mss to advertise in SYN segment.
108  * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109  *
110  * 1. It is independent of path mtu.
111  * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112  * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113  *    attached devices, because some buggy hosts are confused by
114  *    large MSS.
115  * 4. We do not make 3, we advertise MSS, calculated from first
116  *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
117  *    This may be overridden via information stored in routing table.
118  * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119  *    probably even Jumbo".
120  */
121 static __u16 tcp_advertise_mss(struct sock *sk)
122 {
123 	struct tcp_sock *tp = tcp_sk(sk);
124 	const struct dst_entry *dst = __sk_dst_get(sk);
125 	int mss = tp->advmss;
126 
127 	if (dst) {
128 		unsigned int metric = dst_metric_advmss(dst);
129 
130 		if (metric < mss) {
131 			mss = metric;
132 			tp->advmss = mss;
133 		}
134 	}
135 
136 	return (__u16)mss;
137 }
138 
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140  * This is the first part of cwnd validation mechanism. */
141 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
142 {
143 	struct tcp_sock *tp = tcp_sk(sk);
144 	s32 delta = tcp_time_stamp - tp->lsndtime;
145 	u32 restart_cwnd = tcp_init_cwnd(tp, dst);
146 	u32 cwnd = tp->snd_cwnd;
147 
148 	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
149 
150 	tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 	restart_cwnd = min(restart_cwnd, cwnd);
152 
153 	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
154 		cwnd >>= 1;
155 	tp->snd_cwnd = max(cwnd, restart_cwnd);
156 	tp->snd_cwnd_stamp = tcp_time_stamp;
157 	tp->snd_cwnd_used = 0;
158 }
159 
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock *tp,
162 				struct sock *sk)
163 {
164 	struct inet_connection_sock *icsk = inet_csk(sk);
165 	const u32 now = tcp_time_stamp;
166 	const struct dst_entry *dst = __sk_dst_get(sk);
167 
168 	if (sysctl_tcp_slow_start_after_idle &&
169 	    (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
170 		tcp_cwnd_restart(sk, __sk_dst_get(sk));
171 
172 	tp->lsndtime = now;
173 
174 	/* If it is a reply for ato after last received
175 	 * packet, enter pingpong mode.
176 	 */
177 	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
178 	    (!dst || !dst_metric(dst, RTAX_QUICKACK)))
179 			icsk->icsk_ack.pingpong = 1;
180 }
181 
182 /* Account for an ACK we sent. */
183 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
184 {
185 	tcp_dec_quickack_mode(sk, pkts);
186 	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
187 }
188 
189 
190 u32 tcp_default_init_rwnd(u32 mss)
191 {
192 	/* Initial receive window should be twice of TCP_INIT_CWND to
193 	 * enable proper sending of new unsent data during fast recovery
194 	 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
195 	 * limit when mss is larger than 1460.
196 	 */
197 	u32 init_rwnd = TCP_INIT_CWND * 2;
198 
199 	if (mss > 1460)
200 		init_rwnd = max((1460 * init_rwnd) / mss, 2U);
201 	return init_rwnd;
202 }
203 
204 /* Determine a window scaling and initial window to offer.
205  * Based on the assumption that the given amount of space
206  * will be offered. Store the results in the tp structure.
207  * NOTE: for smooth operation initial space offering should
208  * be a multiple of mss if possible. We assume here that mss >= 1.
209  * This MUST be enforced by all callers.
210  */
211 void tcp_select_initial_window(int __space, __u32 mss,
212 			       __u32 *rcv_wnd, __u32 *window_clamp,
213 			       int wscale_ok, __u8 *rcv_wscale,
214 			       __u32 init_rcv_wnd)
215 {
216 	unsigned int space = (__space < 0 ? 0 : __space);
217 
218 	/* If no clamp set the clamp to the max possible scaled window */
219 	if (*window_clamp == 0)
220 		(*window_clamp) = (65535 << 14);
221 	space = min(*window_clamp, space);
222 
223 	/* Quantize space offering to a multiple of mss if possible. */
224 	if (space > mss)
225 		space = (space / mss) * mss;
226 
227 	/* NOTE: offering an initial window larger than 32767
228 	 * will break some buggy TCP stacks. If the admin tells us
229 	 * it is likely we could be speaking with such a buggy stack
230 	 * we will truncate our initial window offering to 32K-1
231 	 * unless the remote has sent us a window scaling option,
232 	 * which we interpret as a sign the remote TCP is not
233 	 * misinterpreting the window field as a signed quantity.
234 	 */
235 	if (sysctl_tcp_workaround_signed_windows)
236 		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
237 	else
238 		(*rcv_wnd) = space;
239 
240 	(*rcv_wscale) = 0;
241 	if (wscale_ok) {
242 		/* Set window scaling on max possible window
243 		 * See RFC1323 for an explanation of the limit to 14
244 		 */
245 		space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
246 		space = min_t(u32, space, *window_clamp);
247 		while (space > 65535 && (*rcv_wscale) < 14) {
248 			space >>= 1;
249 			(*rcv_wscale)++;
250 		}
251 	}
252 
253 	if (mss > (1 << *rcv_wscale)) {
254 		if (!init_rcv_wnd) /* Use default unless specified otherwise */
255 			init_rcv_wnd = tcp_default_init_rwnd(mss);
256 		*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
257 	}
258 
259 	/* Set the clamp no higher than max representable value */
260 	(*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
261 }
262 EXPORT_SYMBOL(tcp_select_initial_window);
263 
264 /* Chose a new window to advertise, update state in tcp_sock for the
265  * socket, and return result with RFC1323 scaling applied.  The return
266  * value can be stuffed directly into th->window for an outgoing
267  * frame.
268  */
269 static u16 tcp_select_window(struct sock *sk)
270 {
271 	struct tcp_sock *tp = tcp_sk(sk);
272 	u32 old_win = tp->rcv_wnd;
273 	u32 cur_win = tcp_receive_window(tp);
274 	u32 new_win = __tcp_select_window(sk);
275 
276 	/* Never shrink the offered window */
277 	if (new_win < cur_win) {
278 		/* Danger Will Robinson!
279 		 * Don't update rcv_wup/rcv_wnd here or else
280 		 * we will not be able to advertise a zero
281 		 * window in time.  --DaveM
282 		 *
283 		 * Relax Will Robinson.
284 		 */
285 		if (new_win == 0)
286 			NET_INC_STATS(sock_net(sk),
287 				      LINUX_MIB_TCPWANTZEROWINDOWADV);
288 		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
289 	}
290 	tp->rcv_wnd = new_win;
291 	tp->rcv_wup = tp->rcv_nxt;
292 
293 	/* Make sure we do not exceed the maximum possible
294 	 * scaled window.
295 	 */
296 	if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
297 		new_win = min(new_win, MAX_TCP_WINDOW);
298 	else
299 		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
300 
301 	/* RFC1323 scaling applied */
302 	new_win >>= tp->rx_opt.rcv_wscale;
303 
304 	/* If we advertise zero window, disable fast path. */
305 	if (new_win == 0) {
306 		tp->pred_flags = 0;
307 		if (old_win)
308 			NET_INC_STATS(sock_net(sk),
309 				      LINUX_MIB_TCPTOZEROWINDOWADV);
310 	} else if (old_win == 0) {
311 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
312 	}
313 
314 	return new_win;
315 }
316 
317 /* Packet ECN state for a SYN-ACK */
318 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
319 {
320 	const struct tcp_sock *tp = tcp_sk(sk);
321 
322 	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
323 	if (!(tp->ecn_flags & TCP_ECN_OK))
324 		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
325 	else if (tcp_ca_needs_ecn(sk))
326 		INET_ECN_xmit(sk);
327 }
328 
329 /* Packet ECN state for a SYN.  */
330 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
331 {
332 	struct tcp_sock *tp = tcp_sk(sk);
333 	bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
334 		       tcp_ca_needs_ecn(sk);
335 
336 	if (!use_ecn) {
337 		const struct dst_entry *dst = __sk_dst_get(sk);
338 
339 		if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
340 			use_ecn = true;
341 	}
342 
343 	tp->ecn_flags = 0;
344 
345 	if (use_ecn) {
346 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
347 		tp->ecn_flags = TCP_ECN_OK;
348 		if (tcp_ca_needs_ecn(sk))
349 			INET_ECN_xmit(sk);
350 	}
351 }
352 
353 static void
354 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th,
355 		    struct sock *sk)
356 {
357 	if (inet_rsk(req)->ecn_ok) {
358 		th->ece = 1;
359 		if (tcp_ca_needs_ecn(sk))
360 			INET_ECN_xmit(sk);
361 	}
362 }
363 
364 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
365  * be sent.
366  */
367 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
368 				int tcp_header_len)
369 {
370 	struct tcp_sock *tp = tcp_sk(sk);
371 
372 	if (tp->ecn_flags & TCP_ECN_OK) {
373 		/* Not-retransmitted data segment: set ECT and inject CWR. */
374 		if (skb->len != tcp_header_len &&
375 		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
376 			INET_ECN_xmit(sk);
377 			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
378 				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
379 				tcp_hdr(skb)->cwr = 1;
380 				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
381 			}
382 		} else if (!tcp_ca_needs_ecn(sk)) {
383 			/* ACK or retransmitted segment: clear ECT|CE */
384 			INET_ECN_dontxmit(sk);
385 		}
386 		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
387 			tcp_hdr(skb)->ece = 1;
388 	}
389 }
390 
391 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
392  * auto increment end seqno.
393  */
394 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
395 {
396 	struct skb_shared_info *shinfo = skb_shinfo(skb);
397 
398 	skb->ip_summed = CHECKSUM_PARTIAL;
399 	skb->csum = 0;
400 
401 	TCP_SKB_CB(skb)->tcp_flags = flags;
402 	TCP_SKB_CB(skb)->sacked = 0;
403 
404 	tcp_skb_pcount_set(skb, 1);
405 	shinfo->gso_size = 0;
406 	shinfo->gso_type = 0;
407 
408 	TCP_SKB_CB(skb)->seq = seq;
409 	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
410 		seq++;
411 	TCP_SKB_CB(skb)->end_seq = seq;
412 }
413 
414 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
415 {
416 	return tp->snd_una != tp->snd_up;
417 }
418 
419 #define OPTION_SACK_ADVERTISE	(1 << 0)
420 #define OPTION_TS		(1 << 1)
421 #define OPTION_MD5		(1 << 2)
422 #define OPTION_WSCALE		(1 << 3)
423 #define OPTION_FAST_OPEN_COOKIE	(1 << 8)
424 
425 struct tcp_out_options {
426 	u16 options;		/* bit field of OPTION_* */
427 	u16 mss;		/* 0 to disable */
428 	u8 ws;			/* window scale, 0 to disable */
429 	u8 num_sack_blocks;	/* number of SACK blocks to include */
430 	u8 hash_size;		/* bytes in hash_location */
431 	__u8 *hash_location;	/* temporary pointer, overloaded */
432 	__u32 tsval, tsecr;	/* need to include OPTION_TS */
433 	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
434 };
435 
436 /* Write previously computed TCP options to the packet.
437  *
438  * Beware: Something in the Internet is very sensitive to the ordering of
439  * TCP options, we learned this through the hard way, so be careful here.
440  * Luckily we can at least blame others for their non-compliance but from
441  * inter-operability perspective it seems that we're somewhat stuck with
442  * the ordering which we have been using if we want to keep working with
443  * those broken things (not that it currently hurts anybody as there isn't
444  * particular reason why the ordering would need to be changed).
445  *
446  * At least SACK_PERM as the first option is known to lead to a disaster
447  * (but it may well be that other scenarios fail similarly).
448  */
449 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
450 			      struct tcp_out_options *opts)
451 {
452 	u16 options = opts->options;	/* mungable copy */
453 
454 	if (unlikely(OPTION_MD5 & options)) {
455 		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
456 			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
457 		/* overload cookie hash location */
458 		opts->hash_location = (__u8 *)ptr;
459 		ptr += 4;
460 	}
461 
462 	if (unlikely(opts->mss)) {
463 		*ptr++ = htonl((TCPOPT_MSS << 24) |
464 			       (TCPOLEN_MSS << 16) |
465 			       opts->mss);
466 	}
467 
468 	if (likely(OPTION_TS & options)) {
469 		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
470 			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
471 				       (TCPOLEN_SACK_PERM << 16) |
472 				       (TCPOPT_TIMESTAMP << 8) |
473 				       TCPOLEN_TIMESTAMP);
474 			options &= ~OPTION_SACK_ADVERTISE;
475 		} else {
476 			*ptr++ = htonl((TCPOPT_NOP << 24) |
477 				       (TCPOPT_NOP << 16) |
478 				       (TCPOPT_TIMESTAMP << 8) |
479 				       TCPOLEN_TIMESTAMP);
480 		}
481 		*ptr++ = htonl(opts->tsval);
482 		*ptr++ = htonl(opts->tsecr);
483 	}
484 
485 	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
486 		*ptr++ = htonl((TCPOPT_NOP << 24) |
487 			       (TCPOPT_NOP << 16) |
488 			       (TCPOPT_SACK_PERM << 8) |
489 			       TCPOLEN_SACK_PERM);
490 	}
491 
492 	if (unlikely(OPTION_WSCALE & options)) {
493 		*ptr++ = htonl((TCPOPT_NOP << 24) |
494 			       (TCPOPT_WINDOW << 16) |
495 			       (TCPOLEN_WINDOW << 8) |
496 			       opts->ws);
497 	}
498 
499 	if (unlikely(opts->num_sack_blocks)) {
500 		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
501 			tp->duplicate_sack : tp->selective_acks;
502 		int this_sack;
503 
504 		*ptr++ = htonl((TCPOPT_NOP  << 24) |
505 			       (TCPOPT_NOP  << 16) |
506 			       (TCPOPT_SACK <<  8) |
507 			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
508 						     TCPOLEN_SACK_PERBLOCK)));
509 
510 		for (this_sack = 0; this_sack < opts->num_sack_blocks;
511 		     ++this_sack) {
512 			*ptr++ = htonl(sp[this_sack].start_seq);
513 			*ptr++ = htonl(sp[this_sack].end_seq);
514 		}
515 
516 		tp->rx_opt.dsack = 0;
517 	}
518 
519 	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
520 		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
521 		u8 *p = (u8 *)ptr;
522 		u32 len; /* Fast Open option length */
523 
524 		if (foc->exp) {
525 			len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
526 			*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
527 				     TCPOPT_FASTOPEN_MAGIC);
528 			p += TCPOLEN_EXP_FASTOPEN_BASE;
529 		} else {
530 			len = TCPOLEN_FASTOPEN_BASE + foc->len;
531 			*p++ = TCPOPT_FASTOPEN;
532 			*p++ = len;
533 		}
534 
535 		memcpy(p, foc->val, foc->len);
536 		if ((len & 3) == 2) {
537 			p[foc->len] = TCPOPT_NOP;
538 			p[foc->len + 1] = TCPOPT_NOP;
539 		}
540 		ptr += (len + 3) >> 2;
541 	}
542 }
543 
544 /* Compute TCP options for SYN packets. This is not the final
545  * network wire format yet.
546  */
547 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
548 				struct tcp_out_options *opts,
549 				struct tcp_md5sig_key **md5)
550 {
551 	struct tcp_sock *tp = tcp_sk(sk);
552 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
553 	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
554 
555 #ifdef CONFIG_TCP_MD5SIG
556 	*md5 = tp->af_specific->md5_lookup(sk, sk);
557 	if (*md5) {
558 		opts->options |= OPTION_MD5;
559 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
560 	}
561 #else
562 	*md5 = NULL;
563 #endif
564 
565 	/* We always get an MSS option.  The option bytes which will be seen in
566 	 * normal data packets should timestamps be used, must be in the MSS
567 	 * advertised.  But we subtract them from tp->mss_cache so that
568 	 * calculations in tcp_sendmsg are simpler etc.  So account for this
569 	 * fact here if necessary.  If we don't do this correctly, as a
570 	 * receiver we won't recognize data packets as being full sized when we
571 	 * should, and thus we won't abide by the delayed ACK rules correctly.
572 	 * SACKs don't matter, we never delay an ACK when we have any of those
573 	 * going out.  */
574 	opts->mss = tcp_advertise_mss(sk);
575 	remaining -= TCPOLEN_MSS_ALIGNED;
576 
577 	if (likely(sysctl_tcp_timestamps && !*md5)) {
578 		opts->options |= OPTION_TS;
579 		opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
580 		opts->tsecr = tp->rx_opt.ts_recent;
581 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
582 	}
583 	if (likely(sysctl_tcp_window_scaling)) {
584 		opts->ws = tp->rx_opt.rcv_wscale;
585 		opts->options |= OPTION_WSCALE;
586 		remaining -= TCPOLEN_WSCALE_ALIGNED;
587 	}
588 	if (likely(sysctl_tcp_sack)) {
589 		opts->options |= OPTION_SACK_ADVERTISE;
590 		if (unlikely(!(OPTION_TS & opts->options)))
591 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
592 	}
593 
594 	if (fastopen && fastopen->cookie.len >= 0) {
595 		u32 need = fastopen->cookie.len;
596 
597 		need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
598 					       TCPOLEN_FASTOPEN_BASE;
599 		need = (need + 3) & ~3U;  /* Align to 32 bits */
600 		if (remaining >= need) {
601 			opts->options |= OPTION_FAST_OPEN_COOKIE;
602 			opts->fastopen_cookie = &fastopen->cookie;
603 			remaining -= need;
604 			tp->syn_fastopen = 1;
605 			tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
606 		}
607 	}
608 
609 	return MAX_TCP_OPTION_SPACE - remaining;
610 }
611 
612 /* Set up TCP options for SYN-ACKs. */
613 static unsigned int tcp_synack_options(struct sock *sk,
614 				   struct request_sock *req,
615 				   unsigned int mss, struct sk_buff *skb,
616 				   struct tcp_out_options *opts,
617 				   const struct tcp_md5sig_key *md5,
618 				   struct tcp_fastopen_cookie *foc)
619 {
620 	struct inet_request_sock *ireq = inet_rsk(req);
621 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
622 
623 #ifdef CONFIG_TCP_MD5SIG
624 	if (md5) {
625 		opts->options |= OPTION_MD5;
626 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
627 
628 		/* We can't fit any SACK blocks in a packet with MD5 + TS
629 		 * options. There was discussion about disabling SACK
630 		 * rather than TS in order to fit in better with old,
631 		 * buggy kernels, but that was deemed to be unnecessary.
632 		 */
633 		ireq->tstamp_ok &= !ireq->sack_ok;
634 	}
635 #endif
636 
637 	/* We always send an MSS option. */
638 	opts->mss = mss;
639 	remaining -= TCPOLEN_MSS_ALIGNED;
640 
641 	if (likely(ireq->wscale_ok)) {
642 		opts->ws = ireq->rcv_wscale;
643 		opts->options |= OPTION_WSCALE;
644 		remaining -= TCPOLEN_WSCALE_ALIGNED;
645 	}
646 	if (likely(ireq->tstamp_ok)) {
647 		opts->options |= OPTION_TS;
648 		opts->tsval = tcp_skb_timestamp(skb);
649 		opts->tsecr = req->ts_recent;
650 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
651 	}
652 	if (likely(ireq->sack_ok)) {
653 		opts->options |= OPTION_SACK_ADVERTISE;
654 		if (unlikely(!ireq->tstamp_ok))
655 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
656 	}
657 	if (foc != NULL && foc->len >= 0) {
658 		u32 need = foc->len;
659 
660 		need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
661 				   TCPOLEN_FASTOPEN_BASE;
662 		need = (need + 3) & ~3U;  /* Align to 32 bits */
663 		if (remaining >= need) {
664 			opts->options |= OPTION_FAST_OPEN_COOKIE;
665 			opts->fastopen_cookie = foc;
666 			remaining -= need;
667 		}
668 	}
669 
670 	return MAX_TCP_OPTION_SPACE - remaining;
671 }
672 
673 /* Compute TCP options for ESTABLISHED sockets. This is not the
674  * final wire format yet.
675  */
676 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
677 					struct tcp_out_options *opts,
678 					struct tcp_md5sig_key **md5)
679 {
680 	struct tcp_sock *tp = tcp_sk(sk);
681 	unsigned int size = 0;
682 	unsigned int eff_sacks;
683 
684 	opts->options = 0;
685 
686 #ifdef CONFIG_TCP_MD5SIG
687 	*md5 = tp->af_specific->md5_lookup(sk, sk);
688 	if (unlikely(*md5)) {
689 		opts->options |= OPTION_MD5;
690 		size += TCPOLEN_MD5SIG_ALIGNED;
691 	}
692 #else
693 	*md5 = NULL;
694 #endif
695 
696 	if (likely(tp->rx_opt.tstamp_ok)) {
697 		opts->options |= OPTION_TS;
698 		opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
699 		opts->tsecr = tp->rx_opt.ts_recent;
700 		size += TCPOLEN_TSTAMP_ALIGNED;
701 	}
702 
703 	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
704 	if (unlikely(eff_sacks)) {
705 		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
706 		opts->num_sack_blocks =
707 			min_t(unsigned int, eff_sacks,
708 			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
709 			      TCPOLEN_SACK_PERBLOCK);
710 		size += TCPOLEN_SACK_BASE_ALIGNED +
711 			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
712 	}
713 
714 	return size;
715 }
716 
717 
718 /* TCP SMALL QUEUES (TSQ)
719  *
720  * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
721  * to reduce RTT and bufferbloat.
722  * We do this using a special skb destructor (tcp_wfree).
723  *
724  * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
725  * needs to be reallocated in a driver.
726  * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
727  *
728  * Since transmit from skb destructor is forbidden, we use a tasklet
729  * to process all sockets that eventually need to send more skbs.
730  * We use one tasklet per cpu, with its own queue of sockets.
731  */
732 struct tsq_tasklet {
733 	struct tasklet_struct	tasklet;
734 	struct list_head	head; /* queue of tcp sockets */
735 };
736 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
737 
738 static void tcp_tsq_handler(struct sock *sk)
739 {
740 	if ((1 << sk->sk_state) &
741 	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
742 	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK))
743 		tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
744 			       0, GFP_ATOMIC);
745 }
746 /*
747  * One tasklet per cpu tries to send more skbs.
748  * We run in tasklet context but need to disable irqs when
749  * transferring tsq->head because tcp_wfree() might
750  * interrupt us (non NAPI drivers)
751  */
752 static void tcp_tasklet_func(unsigned long data)
753 {
754 	struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
755 	LIST_HEAD(list);
756 	unsigned long flags;
757 	struct list_head *q, *n;
758 	struct tcp_sock *tp;
759 	struct sock *sk;
760 
761 	local_irq_save(flags);
762 	list_splice_init(&tsq->head, &list);
763 	local_irq_restore(flags);
764 
765 	list_for_each_safe(q, n, &list) {
766 		tp = list_entry(q, struct tcp_sock, tsq_node);
767 		list_del(&tp->tsq_node);
768 
769 		sk = (struct sock *)tp;
770 		bh_lock_sock(sk);
771 
772 		if (!sock_owned_by_user(sk)) {
773 			tcp_tsq_handler(sk);
774 		} else {
775 			/* defer the work to tcp_release_cb() */
776 			set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
777 		}
778 		bh_unlock_sock(sk);
779 
780 		clear_bit(TSQ_QUEUED, &tp->tsq_flags);
781 		sk_free(sk);
782 	}
783 }
784 
785 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) |		\
786 			  (1UL << TCP_WRITE_TIMER_DEFERRED) |	\
787 			  (1UL << TCP_DELACK_TIMER_DEFERRED) |	\
788 			  (1UL << TCP_MTU_REDUCED_DEFERRED))
789 /**
790  * tcp_release_cb - tcp release_sock() callback
791  * @sk: socket
792  *
793  * called from release_sock() to perform protocol dependent
794  * actions before socket release.
795  */
796 void tcp_release_cb(struct sock *sk)
797 {
798 	struct tcp_sock *tp = tcp_sk(sk);
799 	unsigned long flags, nflags;
800 
801 	/* perform an atomic operation only if at least one flag is set */
802 	do {
803 		flags = tp->tsq_flags;
804 		if (!(flags & TCP_DEFERRED_ALL))
805 			return;
806 		nflags = flags & ~TCP_DEFERRED_ALL;
807 	} while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
808 
809 	if (flags & (1UL << TCP_TSQ_DEFERRED))
810 		tcp_tsq_handler(sk);
811 
812 	/* Here begins the tricky part :
813 	 * We are called from release_sock() with :
814 	 * 1) BH disabled
815 	 * 2) sk_lock.slock spinlock held
816 	 * 3) socket owned by us (sk->sk_lock.owned == 1)
817 	 *
818 	 * But following code is meant to be called from BH handlers,
819 	 * so we should keep BH disabled, but early release socket ownership
820 	 */
821 	sock_release_ownership(sk);
822 
823 	if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
824 		tcp_write_timer_handler(sk);
825 		__sock_put(sk);
826 	}
827 	if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
828 		tcp_delack_timer_handler(sk);
829 		__sock_put(sk);
830 	}
831 	if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
832 		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
833 		__sock_put(sk);
834 	}
835 }
836 EXPORT_SYMBOL(tcp_release_cb);
837 
838 void __init tcp_tasklet_init(void)
839 {
840 	int i;
841 
842 	for_each_possible_cpu(i) {
843 		struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
844 
845 		INIT_LIST_HEAD(&tsq->head);
846 		tasklet_init(&tsq->tasklet,
847 			     tcp_tasklet_func,
848 			     (unsigned long)tsq);
849 	}
850 }
851 
852 /*
853  * Write buffer destructor automatically called from kfree_skb.
854  * We can't xmit new skbs from this context, as we might already
855  * hold qdisc lock.
856  */
857 void tcp_wfree(struct sk_buff *skb)
858 {
859 	struct sock *sk = skb->sk;
860 	struct tcp_sock *tp = tcp_sk(sk);
861 	int wmem;
862 
863 	/* Keep one reference on sk_wmem_alloc.
864 	 * Will be released by sk_free() from here or tcp_tasklet_func()
865 	 */
866 	wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
867 
868 	/* If this softirq is serviced by ksoftirqd, we are likely under stress.
869 	 * Wait until our queues (qdisc + devices) are drained.
870 	 * This gives :
871 	 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
872 	 * - chance for incoming ACK (processed by another cpu maybe)
873 	 *   to migrate this flow (skb->ooo_okay will be eventually set)
874 	 */
875 	if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
876 		goto out;
877 
878 	if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
879 	    !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
880 		unsigned long flags;
881 		struct tsq_tasklet *tsq;
882 
883 		/* queue this socket to tasklet queue */
884 		local_irq_save(flags);
885 		tsq = this_cpu_ptr(&tsq_tasklet);
886 		list_add(&tp->tsq_node, &tsq->head);
887 		tasklet_schedule(&tsq->tasklet);
888 		local_irq_restore(flags);
889 		return;
890 	}
891 out:
892 	sk_free(sk);
893 }
894 
895 /* This routine actually transmits TCP packets queued in by
896  * tcp_do_sendmsg().  This is used by both the initial
897  * transmission and possible later retransmissions.
898  * All SKB's seen here are completely headerless.  It is our
899  * job to build the TCP header, and pass the packet down to
900  * IP so it can do the same plus pass the packet off to the
901  * device.
902  *
903  * We are working here with either a clone of the original
904  * SKB, or a fresh unique copy made by the retransmit engine.
905  */
906 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
907 			    gfp_t gfp_mask)
908 {
909 	const struct inet_connection_sock *icsk = inet_csk(sk);
910 	struct inet_sock *inet;
911 	struct tcp_sock *tp;
912 	struct tcp_skb_cb *tcb;
913 	struct tcp_out_options opts;
914 	unsigned int tcp_options_size, tcp_header_size;
915 	struct tcp_md5sig_key *md5;
916 	struct tcphdr *th;
917 	int err;
918 
919 	BUG_ON(!skb || !tcp_skb_pcount(skb));
920 
921 	if (clone_it) {
922 		skb_mstamp_get(&skb->skb_mstamp);
923 
924 		if (unlikely(skb_cloned(skb)))
925 			skb = pskb_copy(skb, gfp_mask);
926 		else
927 			skb = skb_clone(skb, gfp_mask);
928 		if (unlikely(!skb))
929 			return -ENOBUFS;
930 	}
931 
932 	inet = inet_sk(sk);
933 	tp = tcp_sk(sk);
934 	tcb = TCP_SKB_CB(skb);
935 	memset(&opts, 0, sizeof(opts));
936 
937 	if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
938 		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
939 	else
940 		tcp_options_size = tcp_established_options(sk, skb, &opts,
941 							   &md5);
942 	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
943 
944 	if (tcp_packets_in_flight(tp) == 0)
945 		tcp_ca_event(sk, CA_EVENT_TX_START);
946 
947 	/* if no packet is in qdisc/device queue, then allow XPS to select
948 	 * another queue. We can be called from tcp_tsq_handler()
949 	 * which holds one reference to sk_wmem_alloc.
950 	 *
951 	 * TODO: Ideally, in-flight pure ACK packets should not matter here.
952 	 * One way to get this would be to set skb->truesize = 2 on them.
953 	 */
954 	skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
955 
956 	skb_push(skb, tcp_header_size);
957 	skb_reset_transport_header(skb);
958 
959 	skb_orphan(skb);
960 	skb->sk = sk;
961 	skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree;
962 	skb_set_hash_from_sk(skb, sk);
963 	atomic_add(skb->truesize, &sk->sk_wmem_alloc);
964 
965 	/* Build TCP header and checksum it. */
966 	th = tcp_hdr(skb);
967 	th->source		= inet->inet_sport;
968 	th->dest		= inet->inet_dport;
969 	th->seq			= htonl(tcb->seq);
970 	th->ack_seq		= htonl(tp->rcv_nxt);
971 	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
972 					tcb->tcp_flags);
973 
974 	if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
975 		/* RFC1323: The window in SYN & SYN/ACK segments
976 		 * is never scaled.
977 		 */
978 		th->window	= htons(min(tp->rcv_wnd, 65535U));
979 	} else {
980 		th->window	= htons(tcp_select_window(sk));
981 	}
982 	th->check		= 0;
983 	th->urg_ptr		= 0;
984 
985 	/* The urg_mode check is necessary during a below snd_una win probe */
986 	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
987 		if (before(tp->snd_up, tcb->seq + 0x10000)) {
988 			th->urg_ptr = htons(tp->snd_up - tcb->seq);
989 			th->urg = 1;
990 		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
991 			th->urg_ptr = htons(0xFFFF);
992 			th->urg = 1;
993 		}
994 	}
995 
996 	tcp_options_write((__be32 *)(th + 1), tp, &opts);
997 	if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
998 		tcp_ecn_send(sk, skb, tcp_header_size);
999 
1000 #ifdef CONFIG_TCP_MD5SIG
1001 	/* Calculate the MD5 hash, as we have all we need now */
1002 	if (md5) {
1003 		sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1004 		tp->af_specific->calc_md5_hash(opts.hash_location,
1005 					       md5, sk, skb);
1006 	}
1007 #endif
1008 
1009 	icsk->icsk_af_ops->send_check(sk, skb);
1010 
1011 	if (likely(tcb->tcp_flags & TCPHDR_ACK))
1012 		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1013 
1014 	if (skb->len != tcp_header_size)
1015 		tcp_event_data_sent(tp, sk);
1016 
1017 	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1018 		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1019 			      tcp_skb_pcount(skb));
1020 
1021 	/* OK, its time to fill skb_shinfo(skb)->gso_segs */
1022 	skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1023 
1024 	/* Our usage of tstamp should remain private */
1025 	skb->tstamp.tv64 = 0;
1026 
1027 	/* Cleanup our debris for IP stacks */
1028 	memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1029 			       sizeof(struct inet6_skb_parm)));
1030 
1031 	err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1032 
1033 	if (likely(err <= 0))
1034 		return err;
1035 
1036 	tcp_enter_cwr(sk);
1037 
1038 	return net_xmit_eval(err);
1039 }
1040 
1041 /* This routine just queues the buffer for sending.
1042  *
1043  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1044  * otherwise socket can stall.
1045  */
1046 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1047 {
1048 	struct tcp_sock *tp = tcp_sk(sk);
1049 
1050 	/* Advance write_seq and place onto the write_queue. */
1051 	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1052 	__skb_header_release(skb);
1053 	tcp_add_write_queue_tail(sk, skb);
1054 	sk->sk_wmem_queued += skb->truesize;
1055 	sk_mem_charge(sk, skb->truesize);
1056 }
1057 
1058 /* Initialize TSO segments for a packet. */
1059 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1060 				 unsigned int mss_now)
1061 {
1062 	struct skb_shared_info *shinfo = skb_shinfo(skb);
1063 
1064 	/* Make sure we own this skb before messing gso_size/gso_segs */
1065 	WARN_ON_ONCE(skb_cloned(skb));
1066 
1067 	if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1068 		/* Avoid the costly divide in the normal
1069 		 * non-TSO case.
1070 		 */
1071 		tcp_skb_pcount_set(skb, 1);
1072 		shinfo->gso_size = 0;
1073 		shinfo->gso_type = 0;
1074 	} else {
1075 		tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1076 		shinfo->gso_size = mss_now;
1077 		shinfo->gso_type = sk->sk_gso_type;
1078 	}
1079 }
1080 
1081 /* When a modification to fackets out becomes necessary, we need to check
1082  * skb is counted to fackets_out or not.
1083  */
1084 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1085 				   int decr)
1086 {
1087 	struct tcp_sock *tp = tcp_sk(sk);
1088 
1089 	if (!tp->sacked_out || tcp_is_reno(tp))
1090 		return;
1091 
1092 	if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1093 		tp->fackets_out -= decr;
1094 }
1095 
1096 /* Pcount in the middle of the write queue got changed, we need to do various
1097  * tweaks to fix counters
1098  */
1099 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1100 {
1101 	struct tcp_sock *tp = tcp_sk(sk);
1102 
1103 	tp->packets_out -= decr;
1104 
1105 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1106 		tp->sacked_out -= decr;
1107 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1108 		tp->retrans_out -= decr;
1109 	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1110 		tp->lost_out -= decr;
1111 
1112 	/* Reno case is special. Sigh... */
1113 	if (tcp_is_reno(tp) && decr > 0)
1114 		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1115 
1116 	tcp_adjust_fackets_out(sk, skb, decr);
1117 
1118 	if (tp->lost_skb_hint &&
1119 	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1120 	    (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1121 		tp->lost_cnt_hint -= decr;
1122 
1123 	tcp_verify_left_out(tp);
1124 }
1125 
1126 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1127 {
1128 	struct skb_shared_info *shinfo = skb_shinfo(skb);
1129 
1130 	if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1131 	    !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1132 		struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1133 		u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1134 
1135 		shinfo->tx_flags &= ~tsflags;
1136 		shinfo2->tx_flags |= tsflags;
1137 		swap(shinfo->tskey, shinfo2->tskey);
1138 	}
1139 }
1140 
1141 /* Function to create two new TCP segments.  Shrinks the given segment
1142  * to the specified size and appends a new segment with the rest of the
1143  * packet to the list.  This won't be called frequently, I hope.
1144  * Remember, these are still headerless SKBs at this point.
1145  */
1146 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1147 		 unsigned int mss_now, gfp_t gfp)
1148 {
1149 	struct tcp_sock *tp = tcp_sk(sk);
1150 	struct sk_buff *buff;
1151 	int nsize, old_factor;
1152 	int nlen;
1153 	u8 flags;
1154 
1155 	if (WARN_ON(len > skb->len))
1156 		return -EINVAL;
1157 
1158 	nsize = skb_headlen(skb) - len;
1159 	if (nsize < 0)
1160 		nsize = 0;
1161 
1162 	if (skb_unclone(skb, gfp))
1163 		return -ENOMEM;
1164 
1165 	/* Get a new skb... force flag on. */
1166 	buff = sk_stream_alloc_skb(sk, nsize, gfp);
1167 	if (!buff)
1168 		return -ENOMEM; /* We'll just try again later. */
1169 
1170 	sk->sk_wmem_queued += buff->truesize;
1171 	sk_mem_charge(sk, buff->truesize);
1172 	nlen = skb->len - len - nsize;
1173 	buff->truesize += nlen;
1174 	skb->truesize -= nlen;
1175 
1176 	/* Correct the sequence numbers. */
1177 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1178 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1179 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1180 
1181 	/* PSH and FIN should only be set in the second packet. */
1182 	flags = TCP_SKB_CB(skb)->tcp_flags;
1183 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1184 	TCP_SKB_CB(buff)->tcp_flags = flags;
1185 	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1186 
1187 	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1188 		/* Copy and checksum data tail into the new buffer. */
1189 		buff->csum = csum_partial_copy_nocheck(skb->data + len,
1190 						       skb_put(buff, nsize),
1191 						       nsize, 0);
1192 
1193 		skb_trim(skb, len);
1194 
1195 		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1196 	} else {
1197 		skb->ip_summed = CHECKSUM_PARTIAL;
1198 		skb_split(skb, buff, len);
1199 	}
1200 
1201 	buff->ip_summed = skb->ip_summed;
1202 
1203 	buff->tstamp = skb->tstamp;
1204 	tcp_fragment_tstamp(skb, buff);
1205 
1206 	old_factor = tcp_skb_pcount(skb);
1207 
1208 	/* Fix up tso_factor for both original and new SKB.  */
1209 	tcp_set_skb_tso_segs(sk, skb, mss_now);
1210 	tcp_set_skb_tso_segs(sk, buff, mss_now);
1211 
1212 	/* If this packet has been sent out already, we must
1213 	 * adjust the various packet counters.
1214 	 */
1215 	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1216 		int diff = old_factor - tcp_skb_pcount(skb) -
1217 			tcp_skb_pcount(buff);
1218 
1219 		if (diff)
1220 			tcp_adjust_pcount(sk, skb, diff);
1221 	}
1222 
1223 	/* Link BUFF into the send queue. */
1224 	__skb_header_release(buff);
1225 	tcp_insert_write_queue_after(skb, buff, sk);
1226 
1227 	return 0;
1228 }
1229 
1230 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1231  * eventually). The difference is that pulled data not copied, but
1232  * immediately discarded.
1233  */
1234 static void __pskb_trim_head(struct sk_buff *skb, int len)
1235 {
1236 	struct skb_shared_info *shinfo;
1237 	int i, k, eat;
1238 
1239 	eat = min_t(int, len, skb_headlen(skb));
1240 	if (eat) {
1241 		__skb_pull(skb, eat);
1242 		len -= eat;
1243 		if (!len)
1244 			return;
1245 	}
1246 	eat = len;
1247 	k = 0;
1248 	shinfo = skb_shinfo(skb);
1249 	for (i = 0; i < shinfo->nr_frags; i++) {
1250 		int size = skb_frag_size(&shinfo->frags[i]);
1251 
1252 		if (size <= eat) {
1253 			skb_frag_unref(skb, i);
1254 			eat -= size;
1255 		} else {
1256 			shinfo->frags[k] = shinfo->frags[i];
1257 			if (eat) {
1258 				shinfo->frags[k].page_offset += eat;
1259 				skb_frag_size_sub(&shinfo->frags[k], eat);
1260 				eat = 0;
1261 			}
1262 			k++;
1263 		}
1264 	}
1265 	shinfo->nr_frags = k;
1266 
1267 	skb_reset_tail_pointer(skb);
1268 	skb->data_len -= len;
1269 	skb->len = skb->data_len;
1270 }
1271 
1272 /* Remove acked data from a packet in the transmit queue. */
1273 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1274 {
1275 	if (skb_unclone(skb, GFP_ATOMIC))
1276 		return -ENOMEM;
1277 
1278 	__pskb_trim_head(skb, len);
1279 
1280 	TCP_SKB_CB(skb)->seq += len;
1281 	skb->ip_summed = CHECKSUM_PARTIAL;
1282 
1283 	skb->truesize	     -= len;
1284 	sk->sk_wmem_queued   -= len;
1285 	sk_mem_uncharge(sk, len);
1286 	sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1287 
1288 	/* Any change of skb->len requires recalculation of tso factor. */
1289 	if (tcp_skb_pcount(skb) > 1)
1290 		tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1291 
1292 	return 0;
1293 }
1294 
1295 /* Calculate MSS not accounting any TCP options.  */
1296 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1297 {
1298 	const struct tcp_sock *tp = tcp_sk(sk);
1299 	const struct inet_connection_sock *icsk = inet_csk(sk);
1300 	int mss_now;
1301 
1302 	/* Calculate base mss without TCP options:
1303 	   It is MMS_S - sizeof(tcphdr) of rfc1122
1304 	 */
1305 	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1306 
1307 	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1308 	if (icsk->icsk_af_ops->net_frag_header_len) {
1309 		const struct dst_entry *dst = __sk_dst_get(sk);
1310 
1311 		if (dst && dst_allfrag(dst))
1312 			mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1313 	}
1314 
1315 	/* Clamp it (mss_clamp does not include tcp options) */
1316 	if (mss_now > tp->rx_opt.mss_clamp)
1317 		mss_now = tp->rx_opt.mss_clamp;
1318 
1319 	/* Now subtract optional transport overhead */
1320 	mss_now -= icsk->icsk_ext_hdr_len;
1321 
1322 	/* Then reserve room for full set of TCP options and 8 bytes of data */
1323 	if (mss_now < 48)
1324 		mss_now = 48;
1325 	return mss_now;
1326 }
1327 
1328 /* Calculate MSS. Not accounting for SACKs here.  */
1329 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1330 {
1331 	/* Subtract TCP options size, not including SACKs */
1332 	return __tcp_mtu_to_mss(sk, pmtu) -
1333 	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1334 }
1335 
1336 /* Inverse of above */
1337 int tcp_mss_to_mtu(struct sock *sk, int mss)
1338 {
1339 	const struct tcp_sock *tp = tcp_sk(sk);
1340 	const struct inet_connection_sock *icsk = inet_csk(sk);
1341 	int mtu;
1342 
1343 	mtu = mss +
1344 	      tp->tcp_header_len +
1345 	      icsk->icsk_ext_hdr_len +
1346 	      icsk->icsk_af_ops->net_header_len;
1347 
1348 	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1349 	if (icsk->icsk_af_ops->net_frag_header_len) {
1350 		const struct dst_entry *dst = __sk_dst_get(sk);
1351 
1352 		if (dst && dst_allfrag(dst))
1353 			mtu += icsk->icsk_af_ops->net_frag_header_len;
1354 	}
1355 	return mtu;
1356 }
1357 
1358 /* MTU probing init per socket */
1359 void tcp_mtup_init(struct sock *sk)
1360 {
1361 	struct tcp_sock *tp = tcp_sk(sk);
1362 	struct inet_connection_sock *icsk = inet_csk(sk);
1363 	struct net *net = sock_net(sk);
1364 
1365 	icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1366 	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1367 			       icsk->icsk_af_ops->net_header_len;
1368 	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1369 	icsk->icsk_mtup.probe_size = 0;
1370 	if (icsk->icsk_mtup.enabled)
1371 		icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1372 }
1373 EXPORT_SYMBOL(tcp_mtup_init);
1374 
1375 /* This function synchronize snd mss to current pmtu/exthdr set.
1376 
1377    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1378    for TCP options, but includes only bare TCP header.
1379 
1380    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1381    It is minimum of user_mss and mss received with SYN.
1382    It also does not include TCP options.
1383 
1384    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1385 
1386    tp->mss_cache is current effective sending mss, including
1387    all tcp options except for SACKs. It is evaluated,
1388    taking into account current pmtu, but never exceeds
1389    tp->rx_opt.mss_clamp.
1390 
1391    NOTE1. rfc1122 clearly states that advertised MSS
1392    DOES NOT include either tcp or ip options.
1393 
1394    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1395    are READ ONLY outside this function.		--ANK (980731)
1396  */
1397 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1398 {
1399 	struct tcp_sock *tp = tcp_sk(sk);
1400 	struct inet_connection_sock *icsk = inet_csk(sk);
1401 	int mss_now;
1402 
1403 	if (icsk->icsk_mtup.search_high > pmtu)
1404 		icsk->icsk_mtup.search_high = pmtu;
1405 
1406 	mss_now = tcp_mtu_to_mss(sk, pmtu);
1407 	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1408 
1409 	/* And store cached results */
1410 	icsk->icsk_pmtu_cookie = pmtu;
1411 	if (icsk->icsk_mtup.enabled)
1412 		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1413 	tp->mss_cache = mss_now;
1414 
1415 	return mss_now;
1416 }
1417 EXPORT_SYMBOL(tcp_sync_mss);
1418 
1419 /* Compute the current effective MSS, taking SACKs and IP options,
1420  * and even PMTU discovery events into account.
1421  */
1422 unsigned int tcp_current_mss(struct sock *sk)
1423 {
1424 	const struct tcp_sock *tp = tcp_sk(sk);
1425 	const struct dst_entry *dst = __sk_dst_get(sk);
1426 	u32 mss_now;
1427 	unsigned int header_len;
1428 	struct tcp_out_options opts;
1429 	struct tcp_md5sig_key *md5;
1430 
1431 	mss_now = tp->mss_cache;
1432 
1433 	if (dst) {
1434 		u32 mtu = dst_mtu(dst);
1435 		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1436 			mss_now = tcp_sync_mss(sk, mtu);
1437 	}
1438 
1439 	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1440 		     sizeof(struct tcphdr);
1441 	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1442 	 * some common options. If this is an odd packet (because we have SACK
1443 	 * blocks etc) then our calculated header_len will be different, and
1444 	 * we have to adjust mss_now correspondingly */
1445 	if (header_len != tp->tcp_header_len) {
1446 		int delta = (int) header_len - tp->tcp_header_len;
1447 		mss_now -= delta;
1448 	}
1449 
1450 	return mss_now;
1451 }
1452 
1453 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1454  * As additional protections, we do not touch cwnd in retransmission phases,
1455  * and if application hit its sndbuf limit recently.
1456  */
1457 static void tcp_cwnd_application_limited(struct sock *sk)
1458 {
1459 	struct tcp_sock *tp = tcp_sk(sk);
1460 
1461 	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1462 	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1463 		/* Limited by application or receiver window. */
1464 		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1465 		u32 win_used = max(tp->snd_cwnd_used, init_win);
1466 		if (win_used < tp->snd_cwnd) {
1467 			tp->snd_ssthresh = tcp_current_ssthresh(sk);
1468 			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1469 		}
1470 		tp->snd_cwnd_used = 0;
1471 	}
1472 	tp->snd_cwnd_stamp = tcp_time_stamp;
1473 }
1474 
1475 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1476 {
1477 	struct tcp_sock *tp = tcp_sk(sk);
1478 
1479 	/* Track the maximum number of outstanding packets in each
1480 	 * window, and remember whether we were cwnd-limited then.
1481 	 */
1482 	if (!before(tp->snd_una, tp->max_packets_seq) ||
1483 	    tp->packets_out > tp->max_packets_out) {
1484 		tp->max_packets_out = tp->packets_out;
1485 		tp->max_packets_seq = tp->snd_nxt;
1486 		tp->is_cwnd_limited = is_cwnd_limited;
1487 	}
1488 
1489 	if (tcp_is_cwnd_limited(sk)) {
1490 		/* Network is feed fully. */
1491 		tp->snd_cwnd_used = 0;
1492 		tp->snd_cwnd_stamp = tcp_time_stamp;
1493 	} else {
1494 		/* Network starves. */
1495 		if (tp->packets_out > tp->snd_cwnd_used)
1496 			tp->snd_cwnd_used = tp->packets_out;
1497 
1498 		if (sysctl_tcp_slow_start_after_idle &&
1499 		    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1500 			tcp_cwnd_application_limited(sk);
1501 	}
1502 }
1503 
1504 /* Minshall's variant of the Nagle send check. */
1505 static bool tcp_minshall_check(const struct tcp_sock *tp)
1506 {
1507 	return after(tp->snd_sml, tp->snd_una) &&
1508 		!after(tp->snd_sml, tp->snd_nxt);
1509 }
1510 
1511 /* Update snd_sml if this skb is under mss
1512  * Note that a TSO packet might end with a sub-mss segment
1513  * The test is really :
1514  * if ((skb->len % mss) != 0)
1515  *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1516  * But we can avoid doing the divide again given we already have
1517  *  skb_pcount = skb->len / mss_now
1518  */
1519 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1520 				const struct sk_buff *skb)
1521 {
1522 	if (skb->len < tcp_skb_pcount(skb) * mss_now)
1523 		tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1524 }
1525 
1526 /* Return false, if packet can be sent now without violation Nagle's rules:
1527  * 1. It is full sized. (provided by caller in %partial bool)
1528  * 2. Or it contains FIN. (already checked by caller)
1529  * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1530  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1531  *    With Minshall's modification: all sent small packets are ACKed.
1532  */
1533 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1534 			    int nonagle)
1535 {
1536 	return partial &&
1537 		((nonagle & TCP_NAGLE_CORK) ||
1538 		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1539 }
1540 
1541 /* Return how many segs we'd like on a TSO packet,
1542  * to send one TSO packet per ms
1543  */
1544 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1545 {
1546 	u32 bytes, segs;
1547 
1548 	bytes = min(sk->sk_pacing_rate >> 10,
1549 		    sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1550 
1551 	/* Goal is to send at least one packet per ms,
1552 	 * not one big TSO packet every 100 ms.
1553 	 * This preserves ACK clocking and is consistent
1554 	 * with tcp_tso_should_defer() heuristic.
1555 	 */
1556 	segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1557 
1558 	return min_t(u32, segs, sk->sk_gso_max_segs);
1559 }
1560 
1561 /* Returns the portion of skb which can be sent right away */
1562 static unsigned int tcp_mss_split_point(const struct sock *sk,
1563 					const struct sk_buff *skb,
1564 					unsigned int mss_now,
1565 					unsigned int max_segs,
1566 					int nonagle)
1567 {
1568 	const struct tcp_sock *tp = tcp_sk(sk);
1569 	u32 partial, needed, window, max_len;
1570 
1571 	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1572 	max_len = mss_now * max_segs;
1573 
1574 	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1575 		return max_len;
1576 
1577 	needed = min(skb->len, window);
1578 
1579 	if (max_len <= needed)
1580 		return max_len;
1581 
1582 	partial = needed % mss_now;
1583 	/* If last segment is not a full MSS, check if Nagle rules allow us
1584 	 * to include this last segment in this skb.
1585 	 * Otherwise, we'll split the skb at last MSS boundary
1586 	 */
1587 	if (tcp_nagle_check(partial != 0, tp, nonagle))
1588 		return needed - partial;
1589 
1590 	return needed;
1591 }
1592 
1593 /* Can at least one segment of SKB be sent right now, according to the
1594  * congestion window rules?  If so, return how many segments are allowed.
1595  */
1596 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1597 					 const struct sk_buff *skb)
1598 {
1599 	u32 in_flight, cwnd, halfcwnd;
1600 
1601 	/* Don't be strict about the congestion window for the final FIN.  */
1602 	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1603 	    tcp_skb_pcount(skb) == 1)
1604 		return 1;
1605 
1606 	in_flight = tcp_packets_in_flight(tp);
1607 	cwnd = tp->snd_cwnd;
1608 	if (in_flight >= cwnd)
1609 		return 0;
1610 
1611 	/* For better scheduling, ensure we have at least
1612 	 * 2 GSO packets in flight.
1613 	 */
1614 	halfcwnd = max(cwnd >> 1, 1U);
1615 	return min(halfcwnd, cwnd - in_flight);
1616 }
1617 
1618 /* Initialize TSO state of a skb.
1619  * This must be invoked the first time we consider transmitting
1620  * SKB onto the wire.
1621  */
1622 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1623 			     unsigned int mss_now)
1624 {
1625 	int tso_segs = tcp_skb_pcount(skb);
1626 
1627 	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1628 		tcp_set_skb_tso_segs(sk, skb, mss_now);
1629 		tso_segs = tcp_skb_pcount(skb);
1630 	}
1631 	return tso_segs;
1632 }
1633 
1634 
1635 /* Return true if the Nagle test allows this packet to be
1636  * sent now.
1637  */
1638 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1639 				  unsigned int cur_mss, int nonagle)
1640 {
1641 	/* Nagle rule does not apply to frames, which sit in the middle of the
1642 	 * write_queue (they have no chances to get new data).
1643 	 *
1644 	 * This is implemented in the callers, where they modify the 'nonagle'
1645 	 * argument based upon the location of SKB in the send queue.
1646 	 */
1647 	if (nonagle & TCP_NAGLE_PUSH)
1648 		return true;
1649 
1650 	/* Don't use the nagle rule for urgent data (or for the final FIN). */
1651 	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1652 		return true;
1653 
1654 	if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1655 		return true;
1656 
1657 	return false;
1658 }
1659 
1660 /* Does at least the first segment of SKB fit into the send window? */
1661 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1662 			     const struct sk_buff *skb,
1663 			     unsigned int cur_mss)
1664 {
1665 	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1666 
1667 	if (skb->len > cur_mss)
1668 		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1669 
1670 	return !after(end_seq, tcp_wnd_end(tp));
1671 }
1672 
1673 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1674  * should be put on the wire right now.  If so, it returns the number of
1675  * packets allowed by the congestion window.
1676  */
1677 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1678 				 unsigned int cur_mss, int nonagle)
1679 {
1680 	const struct tcp_sock *tp = tcp_sk(sk);
1681 	unsigned int cwnd_quota;
1682 
1683 	tcp_init_tso_segs(sk, skb, cur_mss);
1684 
1685 	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1686 		return 0;
1687 
1688 	cwnd_quota = tcp_cwnd_test(tp, skb);
1689 	if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1690 		cwnd_quota = 0;
1691 
1692 	return cwnd_quota;
1693 }
1694 
1695 /* Test if sending is allowed right now. */
1696 bool tcp_may_send_now(struct sock *sk)
1697 {
1698 	const struct tcp_sock *tp = tcp_sk(sk);
1699 	struct sk_buff *skb = tcp_send_head(sk);
1700 
1701 	return skb &&
1702 		tcp_snd_test(sk, skb, tcp_current_mss(sk),
1703 			     (tcp_skb_is_last(sk, skb) ?
1704 			      tp->nonagle : TCP_NAGLE_PUSH));
1705 }
1706 
1707 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1708  * which is put after SKB on the list.  It is very much like
1709  * tcp_fragment() except that it may make several kinds of assumptions
1710  * in order to speed up the splitting operation.  In particular, we
1711  * know that all the data is in scatter-gather pages, and that the
1712  * packet has never been sent out before (and thus is not cloned).
1713  */
1714 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1715 			unsigned int mss_now, gfp_t gfp)
1716 {
1717 	struct sk_buff *buff;
1718 	int nlen = skb->len - len;
1719 	u8 flags;
1720 
1721 	/* All of a TSO frame must be composed of paged data.  */
1722 	if (skb->len != skb->data_len)
1723 		return tcp_fragment(sk, skb, len, mss_now, gfp);
1724 
1725 	buff = sk_stream_alloc_skb(sk, 0, gfp);
1726 	if (unlikely(!buff))
1727 		return -ENOMEM;
1728 
1729 	sk->sk_wmem_queued += buff->truesize;
1730 	sk_mem_charge(sk, buff->truesize);
1731 	buff->truesize += nlen;
1732 	skb->truesize -= nlen;
1733 
1734 	/* Correct the sequence numbers. */
1735 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1736 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1737 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1738 
1739 	/* PSH and FIN should only be set in the second packet. */
1740 	flags = TCP_SKB_CB(skb)->tcp_flags;
1741 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1742 	TCP_SKB_CB(buff)->tcp_flags = flags;
1743 
1744 	/* This packet was never sent out yet, so no SACK bits. */
1745 	TCP_SKB_CB(buff)->sacked = 0;
1746 
1747 	buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1748 	skb_split(skb, buff, len);
1749 	tcp_fragment_tstamp(skb, buff);
1750 
1751 	/* Fix up tso_factor for both original and new SKB.  */
1752 	tcp_set_skb_tso_segs(sk, skb, mss_now);
1753 	tcp_set_skb_tso_segs(sk, buff, mss_now);
1754 
1755 	/* Link BUFF into the send queue. */
1756 	__skb_header_release(buff);
1757 	tcp_insert_write_queue_after(skb, buff, sk);
1758 
1759 	return 0;
1760 }
1761 
1762 /* Try to defer sending, if possible, in order to minimize the amount
1763  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1764  *
1765  * This algorithm is from John Heffner.
1766  */
1767 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1768 				 bool *is_cwnd_limited, u32 max_segs)
1769 {
1770 	const struct inet_connection_sock *icsk = inet_csk(sk);
1771 	u32 age, send_win, cong_win, limit, in_flight;
1772 	struct tcp_sock *tp = tcp_sk(sk);
1773 	struct skb_mstamp now;
1774 	struct sk_buff *head;
1775 	int win_divisor;
1776 
1777 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1778 		goto send_now;
1779 
1780 	if (!((1 << icsk->icsk_ca_state) & (TCPF_CA_Open | TCPF_CA_CWR)))
1781 		goto send_now;
1782 
1783 	/* Avoid bursty behavior by allowing defer
1784 	 * only if the last write was recent.
1785 	 */
1786 	if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1787 		goto send_now;
1788 
1789 	in_flight = tcp_packets_in_flight(tp);
1790 
1791 	BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1792 
1793 	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1794 
1795 	/* From in_flight test above, we know that cwnd > in_flight.  */
1796 	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1797 
1798 	limit = min(send_win, cong_win);
1799 
1800 	/* If a full-sized TSO skb can be sent, do it. */
1801 	if (limit >= max_segs * tp->mss_cache)
1802 		goto send_now;
1803 
1804 	/* Middle in queue won't get any more data, full sendable already? */
1805 	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1806 		goto send_now;
1807 
1808 	win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1809 	if (win_divisor) {
1810 		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1811 
1812 		/* If at least some fraction of a window is available,
1813 		 * just use it.
1814 		 */
1815 		chunk /= win_divisor;
1816 		if (limit >= chunk)
1817 			goto send_now;
1818 	} else {
1819 		/* Different approach, try not to defer past a single
1820 		 * ACK.  Receiver should ACK every other full sized
1821 		 * frame, so if we have space for more than 3 frames
1822 		 * then send now.
1823 		 */
1824 		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1825 			goto send_now;
1826 	}
1827 
1828 	head = tcp_write_queue_head(sk);
1829 	skb_mstamp_get(&now);
1830 	age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1831 	/* If next ACK is likely to come too late (half srtt), do not defer */
1832 	if (age < (tp->srtt_us >> 4))
1833 		goto send_now;
1834 
1835 	/* Ok, it looks like it is advisable to defer. */
1836 
1837 	if (cong_win < send_win && cong_win < skb->len)
1838 		*is_cwnd_limited = true;
1839 
1840 	return true;
1841 
1842 send_now:
1843 	return false;
1844 }
1845 
1846 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1847 {
1848 	struct inet_connection_sock *icsk = inet_csk(sk);
1849 	struct tcp_sock *tp = tcp_sk(sk);
1850 	struct net *net = sock_net(sk);
1851 	u32 interval;
1852 	s32 delta;
1853 
1854 	interval = net->ipv4.sysctl_tcp_probe_interval;
1855 	delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1856 	if (unlikely(delta >= interval * HZ)) {
1857 		int mss = tcp_current_mss(sk);
1858 
1859 		/* Update current search range */
1860 		icsk->icsk_mtup.probe_size = 0;
1861 		icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1862 			sizeof(struct tcphdr) +
1863 			icsk->icsk_af_ops->net_header_len;
1864 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1865 
1866 		/* Update probe time stamp */
1867 		icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1868 	}
1869 }
1870 
1871 /* Create a new MTU probe if we are ready.
1872  * MTU probe is regularly attempting to increase the path MTU by
1873  * deliberately sending larger packets.  This discovers routing
1874  * changes resulting in larger path MTUs.
1875  *
1876  * Returns 0 if we should wait to probe (no cwnd available),
1877  *         1 if a probe was sent,
1878  *         -1 otherwise
1879  */
1880 static int tcp_mtu_probe(struct sock *sk)
1881 {
1882 	struct tcp_sock *tp = tcp_sk(sk);
1883 	struct inet_connection_sock *icsk = inet_csk(sk);
1884 	struct sk_buff *skb, *nskb, *next;
1885 	struct net *net = sock_net(sk);
1886 	int len;
1887 	int probe_size;
1888 	int size_needed;
1889 	int copy;
1890 	int mss_now;
1891 	int interval;
1892 
1893 	/* Not currently probing/verifying,
1894 	 * not in recovery,
1895 	 * have enough cwnd, and
1896 	 * not SACKing (the variable headers throw things off) */
1897 	if (!icsk->icsk_mtup.enabled ||
1898 	    icsk->icsk_mtup.probe_size ||
1899 	    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1900 	    tp->snd_cwnd < 11 ||
1901 	    tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1902 		return -1;
1903 
1904 	/* Use binary search for probe_size between tcp_mss_base,
1905 	 * and current mss_clamp. if (search_high - search_low)
1906 	 * smaller than a threshold, backoff from probing.
1907 	 */
1908 	mss_now = tcp_current_mss(sk);
1909 	probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1910 				    icsk->icsk_mtup.search_low) >> 1);
1911 	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1912 	interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1913 	/* When misfortune happens, we are reprobing actively,
1914 	 * and then reprobe timer has expired. We stick with current
1915 	 * probing process by not resetting search range to its orignal.
1916 	 */
1917 	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1918 		interval < net->ipv4.sysctl_tcp_probe_threshold) {
1919 		/* Check whether enough time has elaplased for
1920 		 * another round of probing.
1921 		 */
1922 		tcp_mtu_check_reprobe(sk);
1923 		return -1;
1924 	}
1925 
1926 	/* Have enough data in the send queue to probe? */
1927 	if (tp->write_seq - tp->snd_nxt < size_needed)
1928 		return -1;
1929 
1930 	if (tp->snd_wnd < size_needed)
1931 		return -1;
1932 	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1933 		return 0;
1934 
1935 	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
1936 	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1937 		if (!tcp_packets_in_flight(tp))
1938 			return -1;
1939 		else
1940 			return 0;
1941 	}
1942 
1943 	/* We're allowed to probe.  Build it now. */
1944 	nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC);
1945 	if (!nskb)
1946 		return -1;
1947 	sk->sk_wmem_queued += nskb->truesize;
1948 	sk_mem_charge(sk, nskb->truesize);
1949 
1950 	skb = tcp_send_head(sk);
1951 
1952 	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1953 	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1954 	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1955 	TCP_SKB_CB(nskb)->sacked = 0;
1956 	nskb->csum = 0;
1957 	nskb->ip_summed = skb->ip_summed;
1958 
1959 	tcp_insert_write_queue_before(nskb, skb, sk);
1960 
1961 	len = 0;
1962 	tcp_for_write_queue_from_safe(skb, next, sk) {
1963 		copy = min_t(int, skb->len, probe_size - len);
1964 		if (nskb->ip_summed)
1965 			skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1966 		else
1967 			nskb->csum = skb_copy_and_csum_bits(skb, 0,
1968 							    skb_put(nskb, copy),
1969 							    copy, nskb->csum);
1970 
1971 		if (skb->len <= copy) {
1972 			/* We've eaten all the data from this skb.
1973 			 * Throw it away. */
1974 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1975 			tcp_unlink_write_queue(skb, sk);
1976 			sk_wmem_free_skb(sk, skb);
1977 		} else {
1978 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1979 						   ~(TCPHDR_FIN|TCPHDR_PSH);
1980 			if (!skb_shinfo(skb)->nr_frags) {
1981 				skb_pull(skb, copy);
1982 				if (skb->ip_summed != CHECKSUM_PARTIAL)
1983 					skb->csum = csum_partial(skb->data,
1984 								 skb->len, 0);
1985 			} else {
1986 				__pskb_trim_head(skb, copy);
1987 				tcp_set_skb_tso_segs(sk, skb, mss_now);
1988 			}
1989 			TCP_SKB_CB(skb)->seq += copy;
1990 		}
1991 
1992 		len += copy;
1993 
1994 		if (len >= probe_size)
1995 			break;
1996 	}
1997 	tcp_init_tso_segs(sk, nskb, nskb->len);
1998 
1999 	/* We're ready to send.  If this fails, the probe will
2000 	 * be resegmented into mss-sized pieces by tcp_write_xmit().
2001 	 */
2002 	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2003 		/* Decrement cwnd here because we are sending
2004 		 * effectively two packets. */
2005 		tp->snd_cwnd--;
2006 		tcp_event_new_data_sent(sk, nskb);
2007 
2008 		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2009 		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2010 		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2011 
2012 		return 1;
2013 	}
2014 
2015 	return -1;
2016 }
2017 
2018 /* This routine writes packets to the network.  It advances the
2019  * send_head.  This happens as incoming acks open up the remote
2020  * window for us.
2021  *
2022  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2023  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2024  * account rare use of URG, this is not a big flaw.
2025  *
2026  * Send at most one packet when push_one > 0. Temporarily ignore
2027  * cwnd limit to force at most one packet out when push_one == 2.
2028 
2029  * Returns true, if no segments are in flight and we have queued segments,
2030  * but cannot send anything now because of SWS or another problem.
2031  */
2032 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2033 			   int push_one, gfp_t gfp)
2034 {
2035 	struct tcp_sock *tp = tcp_sk(sk);
2036 	struct sk_buff *skb;
2037 	unsigned int tso_segs, sent_pkts;
2038 	int cwnd_quota;
2039 	int result;
2040 	bool is_cwnd_limited = false;
2041 	u32 max_segs;
2042 
2043 	sent_pkts = 0;
2044 
2045 	if (!push_one) {
2046 		/* Do MTU probing. */
2047 		result = tcp_mtu_probe(sk);
2048 		if (!result) {
2049 			return false;
2050 		} else if (result > 0) {
2051 			sent_pkts = 1;
2052 		}
2053 	}
2054 
2055 	max_segs = tcp_tso_autosize(sk, mss_now);
2056 	while ((skb = tcp_send_head(sk))) {
2057 		unsigned int limit;
2058 
2059 		tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
2060 		BUG_ON(!tso_segs);
2061 
2062 		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2063 			/* "skb_mstamp" is used as a start point for the retransmit timer */
2064 			skb_mstamp_get(&skb->skb_mstamp);
2065 			goto repair; /* Skip network transmission */
2066 		}
2067 
2068 		cwnd_quota = tcp_cwnd_test(tp, skb);
2069 		if (!cwnd_quota) {
2070 			is_cwnd_limited = true;
2071 			if (push_one == 2)
2072 				/* Force out a loss probe pkt. */
2073 				cwnd_quota = 1;
2074 			else
2075 				break;
2076 		}
2077 
2078 		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2079 			break;
2080 
2081 		if (tso_segs == 1 || !max_segs) {
2082 			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2083 						     (tcp_skb_is_last(sk, skb) ?
2084 						      nonagle : TCP_NAGLE_PUSH))))
2085 				break;
2086 		} else {
2087 			if (!push_one &&
2088 			    tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2089 						 max_segs))
2090 				break;
2091 		}
2092 
2093 		limit = mss_now;
2094 		if (tso_segs > 1 && max_segs && !tcp_urg_mode(tp))
2095 			limit = tcp_mss_split_point(sk, skb, mss_now,
2096 						    min_t(unsigned int,
2097 							  cwnd_quota,
2098 							  max_segs),
2099 						    nonagle);
2100 
2101 		if (skb->len > limit &&
2102 		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2103 			break;
2104 
2105 		/* TCP Small Queues :
2106 		 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2107 		 * This allows for :
2108 		 *  - better RTT estimation and ACK scheduling
2109 		 *  - faster recovery
2110 		 *  - high rates
2111 		 * Alas, some drivers / subsystems require a fair amount
2112 		 * of queued bytes to ensure line rate.
2113 		 * One example is wifi aggregation (802.11 AMPDU)
2114 		 */
2115 		limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2116 		limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2117 
2118 		if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2119 			set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2120 			/* It is possible TX completion already happened
2121 			 * before we set TSQ_THROTTLED, so we must
2122 			 * test again the condition.
2123 			 */
2124 			smp_mb__after_atomic();
2125 			if (atomic_read(&sk->sk_wmem_alloc) > limit)
2126 				break;
2127 		}
2128 
2129 		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2130 			break;
2131 
2132 repair:
2133 		/* Advance the send_head.  This one is sent out.
2134 		 * This call will increment packets_out.
2135 		 */
2136 		tcp_event_new_data_sent(sk, skb);
2137 
2138 		tcp_minshall_update(tp, mss_now, skb);
2139 		sent_pkts += tcp_skb_pcount(skb);
2140 
2141 		if (push_one)
2142 			break;
2143 	}
2144 
2145 	if (likely(sent_pkts)) {
2146 		if (tcp_in_cwnd_reduction(sk))
2147 			tp->prr_out += sent_pkts;
2148 
2149 		/* Send one loss probe per tail loss episode. */
2150 		if (push_one != 2)
2151 			tcp_schedule_loss_probe(sk);
2152 		tcp_cwnd_validate(sk, is_cwnd_limited);
2153 		return false;
2154 	}
2155 	return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
2156 }
2157 
2158 bool tcp_schedule_loss_probe(struct sock *sk)
2159 {
2160 	struct inet_connection_sock *icsk = inet_csk(sk);
2161 	struct tcp_sock *tp = tcp_sk(sk);
2162 	u32 timeout, tlp_time_stamp, rto_time_stamp;
2163 	u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2164 
2165 	if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2166 		return false;
2167 	/* No consecutive loss probes. */
2168 	if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2169 		tcp_rearm_rto(sk);
2170 		return false;
2171 	}
2172 	/* Don't do any loss probe on a Fast Open connection before 3WHS
2173 	 * finishes.
2174 	 */
2175 	if (sk->sk_state == TCP_SYN_RECV)
2176 		return false;
2177 
2178 	/* TLP is only scheduled when next timer event is RTO. */
2179 	if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2180 		return false;
2181 
2182 	/* Schedule a loss probe in 2*RTT for SACK capable connections
2183 	 * in Open state, that are either limited by cwnd or application.
2184 	 */
2185 	if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out ||
2186 	    !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2187 		return false;
2188 
2189 	if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2190 	     tcp_send_head(sk))
2191 		return false;
2192 
2193 	/* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2194 	 * for delayed ack when there's one outstanding packet.
2195 	 */
2196 	timeout = rtt << 1;
2197 	if (tp->packets_out == 1)
2198 		timeout = max_t(u32, timeout,
2199 				(rtt + (rtt >> 1) + TCP_DELACK_MAX));
2200 	timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2201 
2202 	/* If RTO is shorter, just schedule TLP in its place. */
2203 	tlp_time_stamp = tcp_time_stamp + timeout;
2204 	rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2205 	if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2206 		s32 delta = rto_time_stamp - tcp_time_stamp;
2207 		if (delta > 0)
2208 			timeout = delta;
2209 	}
2210 
2211 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2212 				  TCP_RTO_MAX);
2213 	return true;
2214 }
2215 
2216 /* Thanks to skb fast clones, we can detect if a prior transmit of
2217  * a packet is still in a qdisc or driver queue.
2218  * In this case, there is very little point doing a retransmit !
2219  * Note: This is called from BH context only.
2220  */
2221 static bool skb_still_in_host_queue(const struct sock *sk,
2222 				    const struct sk_buff *skb)
2223 {
2224 	if (unlikely(skb_fclone_busy(sk, skb))) {
2225 		NET_INC_STATS_BH(sock_net(sk),
2226 				 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2227 		return true;
2228 	}
2229 	return false;
2230 }
2231 
2232 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2233  * retransmit the last segment.
2234  */
2235 void tcp_send_loss_probe(struct sock *sk)
2236 {
2237 	struct tcp_sock *tp = tcp_sk(sk);
2238 	struct sk_buff *skb;
2239 	int pcount;
2240 	int mss = tcp_current_mss(sk);
2241 	int err = -1;
2242 
2243 	if (tcp_send_head(sk)) {
2244 		err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2245 		goto rearm_timer;
2246 	}
2247 
2248 	/* At most one outstanding TLP retransmission. */
2249 	if (tp->tlp_high_seq)
2250 		goto rearm_timer;
2251 
2252 	/* Retransmit last segment. */
2253 	skb = tcp_write_queue_tail(sk);
2254 	if (WARN_ON(!skb))
2255 		goto rearm_timer;
2256 
2257 	if (skb_still_in_host_queue(sk, skb))
2258 		goto rearm_timer;
2259 
2260 	pcount = tcp_skb_pcount(skb);
2261 	if (WARN_ON(!pcount))
2262 		goto rearm_timer;
2263 
2264 	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2265 		if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2266 					  GFP_ATOMIC)))
2267 			goto rearm_timer;
2268 		skb = tcp_write_queue_tail(sk);
2269 	}
2270 
2271 	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2272 		goto rearm_timer;
2273 
2274 	err = __tcp_retransmit_skb(sk, skb);
2275 
2276 	/* Record snd_nxt for loss detection. */
2277 	if (likely(!err))
2278 		tp->tlp_high_seq = tp->snd_nxt;
2279 
2280 rearm_timer:
2281 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2282 				  inet_csk(sk)->icsk_rto,
2283 				  TCP_RTO_MAX);
2284 
2285 	if (likely(!err))
2286 		NET_INC_STATS_BH(sock_net(sk),
2287 				 LINUX_MIB_TCPLOSSPROBES);
2288 }
2289 
2290 /* Push out any pending frames which were held back due to
2291  * TCP_CORK or attempt at coalescing tiny packets.
2292  * The socket must be locked by the caller.
2293  */
2294 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2295 			       int nonagle)
2296 {
2297 	/* If we are closed, the bytes will have to remain here.
2298 	 * In time closedown will finish, we empty the write queue and
2299 	 * all will be happy.
2300 	 */
2301 	if (unlikely(sk->sk_state == TCP_CLOSE))
2302 		return;
2303 
2304 	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2305 			   sk_gfp_atomic(sk, GFP_ATOMIC)))
2306 		tcp_check_probe_timer(sk);
2307 }
2308 
2309 /* Send _single_ skb sitting at the send head. This function requires
2310  * true push pending frames to setup probe timer etc.
2311  */
2312 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2313 {
2314 	struct sk_buff *skb = tcp_send_head(sk);
2315 
2316 	BUG_ON(!skb || skb->len < mss_now);
2317 
2318 	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2319 }
2320 
2321 /* This function returns the amount that we can raise the
2322  * usable window based on the following constraints
2323  *
2324  * 1. The window can never be shrunk once it is offered (RFC 793)
2325  * 2. We limit memory per socket
2326  *
2327  * RFC 1122:
2328  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2329  *  RECV.NEXT + RCV.WIN fixed until:
2330  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2331  *
2332  * i.e. don't raise the right edge of the window until you can raise
2333  * it at least MSS bytes.
2334  *
2335  * Unfortunately, the recommended algorithm breaks header prediction,
2336  * since header prediction assumes th->window stays fixed.
2337  *
2338  * Strictly speaking, keeping th->window fixed violates the receiver
2339  * side SWS prevention criteria. The problem is that under this rule
2340  * a stream of single byte packets will cause the right side of the
2341  * window to always advance by a single byte.
2342  *
2343  * Of course, if the sender implements sender side SWS prevention
2344  * then this will not be a problem.
2345  *
2346  * BSD seems to make the following compromise:
2347  *
2348  *	If the free space is less than the 1/4 of the maximum
2349  *	space available and the free space is less than 1/2 mss,
2350  *	then set the window to 0.
2351  *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2352  *	Otherwise, just prevent the window from shrinking
2353  *	and from being larger than the largest representable value.
2354  *
2355  * This prevents incremental opening of the window in the regime
2356  * where TCP is limited by the speed of the reader side taking
2357  * data out of the TCP receive queue. It does nothing about
2358  * those cases where the window is constrained on the sender side
2359  * because the pipeline is full.
2360  *
2361  * BSD also seems to "accidentally" limit itself to windows that are a
2362  * multiple of MSS, at least until the free space gets quite small.
2363  * This would appear to be a side effect of the mbuf implementation.
2364  * Combining these two algorithms results in the observed behavior
2365  * of having a fixed window size at almost all times.
2366  *
2367  * Below we obtain similar behavior by forcing the offered window to
2368  * a multiple of the mss when it is feasible to do so.
2369  *
2370  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2371  * Regular options like TIMESTAMP are taken into account.
2372  */
2373 u32 __tcp_select_window(struct sock *sk)
2374 {
2375 	struct inet_connection_sock *icsk = inet_csk(sk);
2376 	struct tcp_sock *tp = tcp_sk(sk);
2377 	/* MSS for the peer's data.  Previous versions used mss_clamp
2378 	 * here.  I don't know if the value based on our guesses
2379 	 * of peer's MSS is better for the performance.  It's more correct
2380 	 * but may be worse for the performance because of rcv_mss
2381 	 * fluctuations.  --SAW  1998/11/1
2382 	 */
2383 	int mss = icsk->icsk_ack.rcv_mss;
2384 	int free_space = tcp_space(sk);
2385 	int allowed_space = tcp_full_space(sk);
2386 	int full_space = min_t(int, tp->window_clamp, allowed_space);
2387 	int window;
2388 
2389 	if (mss > full_space)
2390 		mss = full_space;
2391 
2392 	if (free_space < (full_space >> 1)) {
2393 		icsk->icsk_ack.quick = 0;
2394 
2395 		if (sk_under_memory_pressure(sk))
2396 			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2397 					       4U * tp->advmss);
2398 
2399 		/* free_space might become our new window, make sure we don't
2400 		 * increase it due to wscale.
2401 		 */
2402 		free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2403 
2404 		/* if free space is less than mss estimate, or is below 1/16th
2405 		 * of the maximum allowed, try to move to zero-window, else
2406 		 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2407 		 * new incoming data is dropped due to memory limits.
2408 		 * With large window, mss test triggers way too late in order
2409 		 * to announce zero window in time before rmem limit kicks in.
2410 		 */
2411 		if (free_space < (allowed_space >> 4) || free_space < mss)
2412 			return 0;
2413 	}
2414 
2415 	if (free_space > tp->rcv_ssthresh)
2416 		free_space = tp->rcv_ssthresh;
2417 
2418 	/* Don't do rounding if we are using window scaling, since the
2419 	 * scaled window will not line up with the MSS boundary anyway.
2420 	 */
2421 	window = tp->rcv_wnd;
2422 	if (tp->rx_opt.rcv_wscale) {
2423 		window = free_space;
2424 
2425 		/* Advertise enough space so that it won't get scaled away.
2426 		 * Import case: prevent zero window announcement if
2427 		 * 1<<rcv_wscale > mss.
2428 		 */
2429 		if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2430 			window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2431 				  << tp->rx_opt.rcv_wscale);
2432 	} else {
2433 		/* Get the largest window that is a nice multiple of mss.
2434 		 * Window clamp already applied above.
2435 		 * If our current window offering is within 1 mss of the
2436 		 * free space we just keep it. This prevents the divide
2437 		 * and multiply from happening most of the time.
2438 		 * We also don't do any window rounding when the free space
2439 		 * is too small.
2440 		 */
2441 		if (window <= free_space - mss || window > free_space)
2442 			window = (free_space / mss) * mss;
2443 		else if (mss == full_space &&
2444 			 free_space > window + (full_space >> 1))
2445 			window = free_space;
2446 	}
2447 
2448 	return window;
2449 }
2450 
2451 /* Collapses two adjacent SKB's during retransmission. */
2452 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2453 {
2454 	struct tcp_sock *tp = tcp_sk(sk);
2455 	struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2456 	int skb_size, next_skb_size;
2457 
2458 	skb_size = skb->len;
2459 	next_skb_size = next_skb->len;
2460 
2461 	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2462 
2463 	tcp_highest_sack_combine(sk, next_skb, skb);
2464 
2465 	tcp_unlink_write_queue(next_skb, sk);
2466 
2467 	skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2468 				  next_skb_size);
2469 
2470 	if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2471 		skb->ip_summed = CHECKSUM_PARTIAL;
2472 
2473 	if (skb->ip_summed != CHECKSUM_PARTIAL)
2474 		skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2475 
2476 	/* Update sequence range on original skb. */
2477 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2478 
2479 	/* Merge over control information. This moves PSH/FIN etc. over */
2480 	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2481 
2482 	/* All done, get rid of second SKB and account for it so
2483 	 * packet counting does not break.
2484 	 */
2485 	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2486 
2487 	/* changed transmit queue under us so clear hints */
2488 	tcp_clear_retrans_hints_partial(tp);
2489 	if (next_skb == tp->retransmit_skb_hint)
2490 		tp->retransmit_skb_hint = skb;
2491 
2492 	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2493 
2494 	sk_wmem_free_skb(sk, next_skb);
2495 }
2496 
2497 /* Check if coalescing SKBs is legal. */
2498 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2499 {
2500 	if (tcp_skb_pcount(skb) > 1)
2501 		return false;
2502 	/* TODO: SACK collapsing could be used to remove this condition */
2503 	if (skb_shinfo(skb)->nr_frags != 0)
2504 		return false;
2505 	if (skb_cloned(skb))
2506 		return false;
2507 	if (skb == tcp_send_head(sk))
2508 		return false;
2509 	/* Some heurestics for collapsing over SACK'd could be invented */
2510 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2511 		return false;
2512 
2513 	return true;
2514 }
2515 
2516 /* Collapse packets in the retransmit queue to make to create
2517  * less packets on the wire. This is only done on retransmission.
2518  */
2519 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2520 				     int space)
2521 {
2522 	struct tcp_sock *tp = tcp_sk(sk);
2523 	struct sk_buff *skb = to, *tmp;
2524 	bool first = true;
2525 
2526 	if (!sysctl_tcp_retrans_collapse)
2527 		return;
2528 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2529 		return;
2530 
2531 	tcp_for_write_queue_from_safe(skb, tmp, sk) {
2532 		if (!tcp_can_collapse(sk, skb))
2533 			break;
2534 
2535 		space -= skb->len;
2536 
2537 		if (first) {
2538 			first = false;
2539 			continue;
2540 		}
2541 
2542 		if (space < 0)
2543 			break;
2544 		/* Punt if not enough space exists in the first SKB for
2545 		 * the data in the second
2546 		 */
2547 		if (skb->len > skb_availroom(to))
2548 			break;
2549 
2550 		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2551 			break;
2552 
2553 		tcp_collapse_retrans(sk, to);
2554 	}
2555 }
2556 
2557 /* This retransmits one SKB.  Policy decisions and retransmit queue
2558  * state updates are done by the caller.  Returns non-zero if an
2559  * error occurred which prevented the send.
2560  */
2561 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2562 {
2563 	struct tcp_sock *tp = tcp_sk(sk);
2564 	struct inet_connection_sock *icsk = inet_csk(sk);
2565 	unsigned int cur_mss;
2566 	int err;
2567 
2568 	/* Inconslusive MTU probe */
2569 	if (icsk->icsk_mtup.probe_size) {
2570 		icsk->icsk_mtup.probe_size = 0;
2571 	}
2572 
2573 	/* Do not sent more than we queued. 1/4 is reserved for possible
2574 	 * copying overhead: fragmentation, tunneling, mangling etc.
2575 	 */
2576 	if (atomic_read(&sk->sk_wmem_alloc) >
2577 	    min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2578 		return -EAGAIN;
2579 
2580 	if (skb_still_in_host_queue(sk, skb))
2581 		return -EBUSY;
2582 
2583 	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2584 		if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2585 			BUG();
2586 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2587 			return -ENOMEM;
2588 	}
2589 
2590 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2591 		return -EHOSTUNREACH; /* Routing failure or similar. */
2592 
2593 	cur_mss = tcp_current_mss(sk);
2594 
2595 	/* If receiver has shrunk his window, and skb is out of
2596 	 * new window, do not retransmit it. The exception is the
2597 	 * case, when window is shrunk to zero. In this case
2598 	 * our retransmit serves as a zero window probe.
2599 	 */
2600 	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2601 	    TCP_SKB_CB(skb)->seq != tp->snd_una)
2602 		return -EAGAIN;
2603 
2604 	if (skb->len > cur_mss) {
2605 		if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
2606 			return -ENOMEM; /* We'll try again later. */
2607 	} else {
2608 		int oldpcount = tcp_skb_pcount(skb);
2609 
2610 		if (unlikely(oldpcount > 1)) {
2611 			if (skb_unclone(skb, GFP_ATOMIC))
2612 				return -ENOMEM;
2613 			tcp_init_tso_segs(sk, skb, cur_mss);
2614 			tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2615 		}
2616 	}
2617 
2618 	tcp_retrans_try_collapse(sk, skb, cur_mss);
2619 
2620 	/* Make a copy, if the first transmission SKB clone we made
2621 	 * is still in somebody's hands, else make a clone.
2622 	 */
2623 
2624 	/* make sure skb->data is aligned on arches that require it
2625 	 * and check if ack-trimming & collapsing extended the headroom
2626 	 * beyond what csum_start can cover.
2627 	 */
2628 	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2629 		     skb_headroom(skb) >= 0xFFFF)) {
2630 		struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2631 						   GFP_ATOMIC);
2632 		err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2633 			     -ENOBUFS;
2634 	} else {
2635 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2636 	}
2637 
2638 	if (likely(!err)) {
2639 		TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2640 		/* Update global TCP statistics. */
2641 		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2642 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2643 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2644 		tp->total_retrans++;
2645 	}
2646 	return err;
2647 }
2648 
2649 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2650 {
2651 	struct tcp_sock *tp = tcp_sk(sk);
2652 	int err = __tcp_retransmit_skb(sk, skb);
2653 
2654 	if (err == 0) {
2655 #if FASTRETRANS_DEBUG > 0
2656 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2657 			net_dbg_ratelimited("retrans_out leaked\n");
2658 		}
2659 #endif
2660 		if (!tp->retrans_out)
2661 			tp->lost_retrans_low = tp->snd_nxt;
2662 		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2663 		tp->retrans_out += tcp_skb_pcount(skb);
2664 
2665 		/* Save stamp of the first retransmit. */
2666 		if (!tp->retrans_stamp)
2667 			tp->retrans_stamp = tcp_skb_timestamp(skb);
2668 
2669 		/* snd_nxt is stored to detect loss of retransmitted segment,
2670 		 * see tcp_input.c tcp_sacktag_write_queue().
2671 		 */
2672 		TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2673 	} else if (err != -EBUSY) {
2674 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2675 	}
2676 
2677 	if (tp->undo_retrans < 0)
2678 		tp->undo_retrans = 0;
2679 	tp->undo_retrans += tcp_skb_pcount(skb);
2680 	return err;
2681 }
2682 
2683 /* Check if we forward retransmits are possible in the current
2684  * window/congestion state.
2685  */
2686 static bool tcp_can_forward_retransmit(struct sock *sk)
2687 {
2688 	const struct inet_connection_sock *icsk = inet_csk(sk);
2689 	const struct tcp_sock *tp = tcp_sk(sk);
2690 
2691 	/* Forward retransmissions are possible only during Recovery. */
2692 	if (icsk->icsk_ca_state != TCP_CA_Recovery)
2693 		return false;
2694 
2695 	/* No forward retransmissions in Reno are possible. */
2696 	if (tcp_is_reno(tp))
2697 		return false;
2698 
2699 	/* Yeah, we have to make difficult choice between forward transmission
2700 	 * and retransmission... Both ways have their merits...
2701 	 *
2702 	 * For now we do not retransmit anything, while we have some new
2703 	 * segments to send. In the other cases, follow rule 3 for
2704 	 * NextSeg() specified in RFC3517.
2705 	 */
2706 
2707 	if (tcp_may_send_now(sk))
2708 		return false;
2709 
2710 	return true;
2711 }
2712 
2713 /* This gets called after a retransmit timeout, and the initially
2714  * retransmitted data is acknowledged.  It tries to continue
2715  * resending the rest of the retransmit queue, until either
2716  * we've sent it all or the congestion window limit is reached.
2717  * If doing SACK, the first ACK which comes back for a timeout
2718  * based retransmit packet might feed us FACK information again.
2719  * If so, we use it to avoid unnecessarily retransmissions.
2720  */
2721 void tcp_xmit_retransmit_queue(struct sock *sk)
2722 {
2723 	const struct inet_connection_sock *icsk = inet_csk(sk);
2724 	struct tcp_sock *tp = tcp_sk(sk);
2725 	struct sk_buff *skb;
2726 	struct sk_buff *hole = NULL;
2727 	u32 last_lost;
2728 	int mib_idx;
2729 	int fwd_rexmitting = 0;
2730 
2731 	if (!tp->packets_out)
2732 		return;
2733 
2734 	if (!tp->lost_out)
2735 		tp->retransmit_high = tp->snd_una;
2736 
2737 	if (tp->retransmit_skb_hint) {
2738 		skb = tp->retransmit_skb_hint;
2739 		last_lost = TCP_SKB_CB(skb)->end_seq;
2740 		if (after(last_lost, tp->retransmit_high))
2741 			last_lost = tp->retransmit_high;
2742 	} else {
2743 		skb = tcp_write_queue_head(sk);
2744 		last_lost = tp->snd_una;
2745 	}
2746 
2747 	tcp_for_write_queue_from(skb, sk) {
2748 		__u8 sacked = TCP_SKB_CB(skb)->sacked;
2749 
2750 		if (skb == tcp_send_head(sk))
2751 			break;
2752 		/* we could do better than to assign each time */
2753 		if (!hole)
2754 			tp->retransmit_skb_hint = skb;
2755 
2756 		/* Assume this retransmit will generate
2757 		 * only one packet for congestion window
2758 		 * calculation purposes.  This works because
2759 		 * tcp_retransmit_skb() will chop up the
2760 		 * packet to be MSS sized and all the
2761 		 * packet counting works out.
2762 		 */
2763 		if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2764 			return;
2765 
2766 		if (fwd_rexmitting) {
2767 begin_fwd:
2768 			if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2769 				break;
2770 			mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2771 
2772 		} else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2773 			tp->retransmit_high = last_lost;
2774 			if (!tcp_can_forward_retransmit(sk))
2775 				break;
2776 			/* Backtrack if necessary to non-L'ed skb */
2777 			if (hole) {
2778 				skb = hole;
2779 				hole = NULL;
2780 			}
2781 			fwd_rexmitting = 1;
2782 			goto begin_fwd;
2783 
2784 		} else if (!(sacked & TCPCB_LOST)) {
2785 			if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2786 				hole = skb;
2787 			continue;
2788 
2789 		} else {
2790 			last_lost = TCP_SKB_CB(skb)->end_seq;
2791 			if (icsk->icsk_ca_state != TCP_CA_Loss)
2792 				mib_idx = LINUX_MIB_TCPFASTRETRANS;
2793 			else
2794 				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2795 		}
2796 
2797 		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2798 			continue;
2799 
2800 		if (tcp_retransmit_skb(sk, skb))
2801 			return;
2802 
2803 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
2804 
2805 		if (tcp_in_cwnd_reduction(sk))
2806 			tp->prr_out += tcp_skb_pcount(skb);
2807 
2808 		if (skb == tcp_write_queue_head(sk))
2809 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2810 						  inet_csk(sk)->icsk_rto,
2811 						  TCP_RTO_MAX);
2812 	}
2813 }
2814 
2815 /* We allow to exceed memory limits for FIN packets to expedite
2816  * connection tear down and (memory) recovery.
2817  * Otherwise tcp_send_fin() could be tempted to either delay FIN
2818  * or even be forced to close flow without any FIN.
2819  */
2820 static void sk_forced_wmem_schedule(struct sock *sk, int size)
2821 {
2822 	int amt, status;
2823 
2824 	if (size <= sk->sk_forward_alloc)
2825 		return;
2826 	amt = sk_mem_pages(size);
2827 	sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2828 	sk_memory_allocated_add(sk, amt, &status);
2829 }
2830 
2831 /* Send a FIN. The caller locks the socket for us.
2832  * We should try to send a FIN packet really hard, but eventually give up.
2833  */
2834 void tcp_send_fin(struct sock *sk)
2835 {
2836 	struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2837 	struct tcp_sock *tp = tcp_sk(sk);
2838 
2839 	/* Optimization, tack on the FIN if we have one skb in write queue and
2840 	 * this skb was not yet sent, or we are under memory pressure.
2841 	 * Note: in the latter case, FIN packet will be sent after a timeout,
2842 	 * as TCP stack thinks it has already been transmitted.
2843 	 */
2844 	if (tskb && (tcp_send_head(sk) || sk_under_memory_pressure(sk))) {
2845 coalesce:
2846 		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2847 		TCP_SKB_CB(tskb)->end_seq++;
2848 		tp->write_seq++;
2849 		if (!tcp_send_head(sk)) {
2850 			/* This means tskb was already sent.
2851 			 * Pretend we included the FIN on previous transmit.
2852 			 * We need to set tp->snd_nxt to the value it would have
2853 			 * if FIN had been sent. This is because retransmit path
2854 			 * does not change tp->snd_nxt.
2855 			 */
2856 			tp->snd_nxt++;
2857 			return;
2858 		}
2859 	} else {
2860 		skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2861 		if (unlikely(!skb)) {
2862 			if (tskb)
2863 				goto coalesce;
2864 			return;
2865 		}
2866 		skb_reserve(skb, MAX_TCP_HEADER);
2867 		sk_forced_wmem_schedule(sk, skb->truesize);
2868 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2869 		tcp_init_nondata_skb(skb, tp->write_seq,
2870 				     TCPHDR_ACK | TCPHDR_FIN);
2871 		tcp_queue_skb(sk, skb);
2872 	}
2873 	__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2874 }
2875 
2876 /* We get here when a process closes a file descriptor (either due to
2877  * an explicit close() or as a byproduct of exit()'ing) and there
2878  * was unread data in the receive queue.  This behavior is recommended
2879  * by RFC 2525, section 2.17.  -DaveM
2880  */
2881 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2882 {
2883 	struct sk_buff *skb;
2884 
2885 	/* NOTE: No TCP options attached and we never retransmit this. */
2886 	skb = alloc_skb(MAX_TCP_HEADER, priority);
2887 	if (!skb) {
2888 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2889 		return;
2890 	}
2891 
2892 	/* Reserve space for headers and prepare control bits. */
2893 	skb_reserve(skb, MAX_TCP_HEADER);
2894 	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2895 			     TCPHDR_ACK | TCPHDR_RST);
2896 	/* Send it off. */
2897 	if (tcp_transmit_skb(sk, skb, 0, priority))
2898 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2899 
2900 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2901 }
2902 
2903 /* Send a crossed SYN-ACK during socket establishment.
2904  * WARNING: This routine must only be called when we have already sent
2905  * a SYN packet that crossed the incoming SYN that caused this routine
2906  * to get called. If this assumption fails then the initial rcv_wnd
2907  * and rcv_wscale values will not be correct.
2908  */
2909 int tcp_send_synack(struct sock *sk)
2910 {
2911 	struct sk_buff *skb;
2912 
2913 	skb = tcp_write_queue_head(sk);
2914 	if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2915 		pr_debug("%s: wrong queue state\n", __func__);
2916 		return -EFAULT;
2917 	}
2918 	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2919 		if (skb_cloned(skb)) {
2920 			struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2921 			if (!nskb)
2922 				return -ENOMEM;
2923 			tcp_unlink_write_queue(skb, sk);
2924 			__skb_header_release(nskb);
2925 			__tcp_add_write_queue_head(sk, nskb);
2926 			sk_wmem_free_skb(sk, skb);
2927 			sk->sk_wmem_queued += nskb->truesize;
2928 			sk_mem_charge(sk, nskb->truesize);
2929 			skb = nskb;
2930 		}
2931 
2932 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2933 		tcp_ecn_send_synack(sk, skb);
2934 	}
2935 	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2936 }
2937 
2938 /**
2939  * tcp_make_synack - Prepare a SYN-ACK.
2940  * sk: listener socket
2941  * dst: dst entry attached to the SYNACK
2942  * req: request_sock pointer
2943  *
2944  * Allocate one skb and build a SYNACK packet.
2945  * @dst is consumed : Caller should not use it again.
2946  */
2947 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2948 				struct request_sock *req,
2949 				struct tcp_fastopen_cookie *foc)
2950 {
2951 	struct tcp_out_options opts;
2952 	struct inet_request_sock *ireq = inet_rsk(req);
2953 	struct tcp_sock *tp = tcp_sk(sk);
2954 	struct tcphdr *th;
2955 	struct sk_buff *skb;
2956 	struct tcp_md5sig_key *md5 = NULL;
2957 	int tcp_header_size;
2958 	int mss;
2959 
2960 	skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC);
2961 	if (unlikely(!skb)) {
2962 		dst_release(dst);
2963 		return NULL;
2964 	}
2965 	/* Reserve space for headers. */
2966 	skb_reserve(skb, MAX_TCP_HEADER);
2967 
2968 	skb_dst_set(skb, dst);
2969 
2970 	mss = dst_metric_advmss(dst);
2971 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2972 		mss = tp->rx_opt.user_mss;
2973 
2974 	memset(&opts, 0, sizeof(opts));
2975 #ifdef CONFIG_SYN_COOKIES
2976 	if (unlikely(req->cookie_ts))
2977 		skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
2978 	else
2979 #endif
2980 	skb_mstamp_get(&skb->skb_mstamp);
2981 
2982 #ifdef CONFIG_TCP_MD5SIG
2983 	rcu_read_lock();
2984 	md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
2985 #endif
2986 	tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
2987 					     foc) + sizeof(*th);
2988 
2989 	skb_push(skb, tcp_header_size);
2990 	skb_reset_transport_header(skb);
2991 
2992 	th = tcp_hdr(skb);
2993 	memset(th, 0, sizeof(struct tcphdr));
2994 	th->syn = 1;
2995 	th->ack = 1;
2996 	tcp_ecn_make_synack(req, th, sk);
2997 	th->source = htons(ireq->ir_num);
2998 	th->dest = ireq->ir_rmt_port;
2999 	/* Setting of flags are superfluous here for callers (and ECE is
3000 	 * not even correctly set)
3001 	 */
3002 	tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3003 			     TCPHDR_SYN | TCPHDR_ACK);
3004 
3005 	th->seq = htonl(TCP_SKB_CB(skb)->seq);
3006 	/* XXX data is queued and acked as is. No buffer/window check */
3007 	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3008 
3009 	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3010 	th->window = htons(min(req->rcv_wnd, 65535U));
3011 	tcp_options_write((__be32 *)(th + 1), tp, &opts);
3012 	th->doff = (tcp_header_size >> 2);
3013 	TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
3014 
3015 #ifdef CONFIG_TCP_MD5SIG
3016 	/* Okay, we have all we need - do the md5 hash if needed */
3017 	if (md5)
3018 		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3019 					       md5, req_to_sk(req), skb);
3020 	rcu_read_unlock();
3021 #endif
3022 
3023 	/* Do not fool tcpdump (if any), clean our debris */
3024 	skb->tstamp.tv64 = 0;
3025 	return skb;
3026 }
3027 EXPORT_SYMBOL(tcp_make_synack);
3028 
3029 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3030 {
3031 	struct inet_connection_sock *icsk = inet_csk(sk);
3032 	const struct tcp_congestion_ops *ca;
3033 	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3034 
3035 	if (ca_key == TCP_CA_UNSPEC)
3036 		return;
3037 
3038 	rcu_read_lock();
3039 	ca = tcp_ca_find_key(ca_key);
3040 	if (likely(ca && try_module_get(ca->owner))) {
3041 		module_put(icsk->icsk_ca_ops->owner);
3042 		icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3043 		icsk->icsk_ca_ops = ca;
3044 	}
3045 	rcu_read_unlock();
3046 }
3047 
3048 /* Do all connect socket setups that can be done AF independent. */
3049 static void tcp_connect_init(struct sock *sk)
3050 {
3051 	const struct dst_entry *dst = __sk_dst_get(sk);
3052 	struct tcp_sock *tp = tcp_sk(sk);
3053 	__u8 rcv_wscale;
3054 
3055 	/* We'll fix this up when we get a response from the other end.
3056 	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3057 	 */
3058 	tp->tcp_header_len = sizeof(struct tcphdr) +
3059 		(sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3060 
3061 #ifdef CONFIG_TCP_MD5SIG
3062 	if (tp->af_specific->md5_lookup(sk, sk))
3063 		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3064 #endif
3065 
3066 	/* If user gave his TCP_MAXSEG, record it to clamp */
3067 	if (tp->rx_opt.user_mss)
3068 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3069 	tp->max_window = 0;
3070 	tcp_mtup_init(sk);
3071 	tcp_sync_mss(sk, dst_mtu(dst));
3072 
3073 	tcp_ca_dst_init(sk, dst);
3074 
3075 	if (!tp->window_clamp)
3076 		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3077 	tp->advmss = dst_metric_advmss(dst);
3078 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3079 		tp->advmss = tp->rx_opt.user_mss;
3080 
3081 	tcp_initialize_rcv_mss(sk);
3082 
3083 	/* limit the window selection if the user enforce a smaller rx buffer */
3084 	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3085 	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3086 		tp->window_clamp = tcp_full_space(sk);
3087 
3088 	tcp_select_initial_window(tcp_full_space(sk),
3089 				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3090 				  &tp->rcv_wnd,
3091 				  &tp->window_clamp,
3092 				  sysctl_tcp_window_scaling,
3093 				  &rcv_wscale,
3094 				  dst_metric(dst, RTAX_INITRWND));
3095 
3096 	tp->rx_opt.rcv_wscale = rcv_wscale;
3097 	tp->rcv_ssthresh = tp->rcv_wnd;
3098 
3099 	sk->sk_err = 0;
3100 	sock_reset_flag(sk, SOCK_DONE);
3101 	tp->snd_wnd = 0;
3102 	tcp_init_wl(tp, 0);
3103 	tp->snd_una = tp->write_seq;
3104 	tp->snd_sml = tp->write_seq;
3105 	tp->snd_up = tp->write_seq;
3106 	tp->snd_nxt = tp->write_seq;
3107 
3108 	if (likely(!tp->repair))
3109 		tp->rcv_nxt = 0;
3110 	else
3111 		tp->rcv_tstamp = tcp_time_stamp;
3112 	tp->rcv_wup = tp->rcv_nxt;
3113 	tp->copied_seq = tp->rcv_nxt;
3114 
3115 	inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3116 	inet_csk(sk)->icsk_retransmits = 0;
3117 	tcp_clear_retrans(tp);
3118 }
3119 
3120 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3121 {
3122 	struct tcp_sock *tp = tcp_sk(sk);
3123 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3124 
3125 	tcb->end_seq += skb->len;
3126 	__skb_header_release(skb);
3127 	__tcp_add_write_queue_tail(sk, skb);
3128 	sk->sk_wmem_queued += skb->truesize;
3129 	sk_mem_charge(sk, skb->truesize);
3130 	tp->write_seq = tcb->end_seq;
3131 	tp->packets_out += tcp_skb_pcount(skb);
3132 }
3133 
3134 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3135  * queue a data-only packet after the regular SYN, such that regular SYNs
3136  * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3137  * only the SYN sequence, the data are retransmitted in the first ACK.
3138  * If cookie is not cached or other error occurs, falls back to send a
3139  * regular SYN with Fast Open cookie request option.
3140  */
3141 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3142 {
3143 	struct tcp_sock *tp = tcp_sk(sk);
3144 	struct tcp_fastopen_request *fo = tp->fastopen_req;
3145 	int syn_loss = 0, space, err = 0, copied;
3146 	unsigned long last_syn_loss = 0;
3147 	struct sk_buff *syn_data;
3148 
3149 	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3150 	tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3151 			       &syn_loss, &last_syn_loss);
3152 	/* Recurring FO SYN losses: revert to regular handshake temporarily */
3153 	if (syn_loss > 1 &&
3154 	    time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3155 		fo->cookie.len = -1;
3156 		goto fallback;
3157 	}
3158 
3159 	if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3160 		fo->cookie.len = -1;
3161 	else if (fo->cookie.len <= 0)
3162 		goto fallback;
3163 
3164 	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3165 	 * user-MSS. Reserve maximum option space for middleboxes that add
3166 	 * private TCP options. The cost is reduced data space in SYN :(
3167 	 */
3168 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3169 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3170 	space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3171 		MAX_TCP_OPTION_SPACE;
3172 
3173 	space = min_t(size_t, space, fo->size);
3174 
3175 	/* limit to order-0 allocations */
3176 	space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3177 
3178 	syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation);
3179 	if (!syn_data)
3180 		goto fallback;
3181 	syn_data->ip_summed = CHECKSUM_PARTIAL;
3182 	memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3183 	copied = copy_from_iter(skb_put(syn_data, space), space,
3184 				&fo->data->msg_iter);
3185 	if (unlikely(!copied)) {
3186 		kfree_skb(syn_data);
3187 		goto fallback;
3188 	}
3189 	if (copied != space) {
3190 		skb_trim(syn_data, copied);
3191 		space = copied;
3192 	}
3193 
3194 	/* No more data pending in inet_wait_for_connect() */
3195 	if (space == fo->size)
3196 		fo->data = NULL;
3197 	fo->copied = space;
3198 
3199 	tcp_connect_queue_skb(sk, syn_data);
3200 
3201 	err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3202 
3203 	syn->skb_mstamp = syn_data->skb_mstamp;
3204 
3205 	/* Now full SYN+DATA was cloned and sent (or not),
3206 	 * remove the SYN from the original skb (syn_data)
3207 	 * we keep in write queue in case of a retransmit, as we
3208 	 * also have the SYN packet (with no data) in the same queue.
3209 	 */
3210 	TCP_SKB_CB(syn_data)->seq++;
3211 	TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3212 	if (!err) {
3213 		tp->syn_data = (fo->copied > 0);
3214 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3215 		goto done;
3216 	}
3217 
3218 fallback:
3219 	/* Send a regular SYN with Fast Open cookie request option */
3220 	if (fo->cookie.len > 0)
3221 		fo->cookie.len = 0;
3222 	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3223 	if (err)
3224 		tp->syn_fastopen = 0;
3225 done:
3226 	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3227 	return err;
3228 }
3229 
3230 /* Build a SYN and send it off. */
3231 int tcp_connect(struct sock *sk)
3232 {
3233 	struct tcp_sock *tp = tcp_sk(sk);
3234 	struct sk_buff *buff;
3235 	int err;
3236 
3237 	tcp_connect_init(sk);
3238 
3239 	if (unlikely(tp->repair)) {
3240 		tcp_finish_connect(sk, NULL);
3241 		return 0;
3242 	}
3243 
3244 	buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
3245 	if (unlikely(!buff))
3246 		return -ENOBUFS;
3247 
3248 	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3249 	tp->retrans_stamp = tcp_time_stamp;
3250 	tcp_connect_queue_skb(sk, buff);
3251 	tcp_ecn_send_syn(sk, buff);
3252 
3253 	/* Send off SYN; include data in Fast Open. */
3254 	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3255 	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3256 	if (err == -ECONNREFUSED)
3257 		return err;
3258 
3259 	/* We change tp->snd_nxt after the tcp_transmit_skb() call
3260 	 * in order to make this packet get counted in tcpOutSegs.
3261 	 */
3262 	tp->snd_nxt = tp->write_seq;
3263 	tp->pushed_seq = tp->write_seq;
3264 	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3265 
3266 	/* Timer for repeating the SYN until an answer. */
3267 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3268 				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3269 	return 0;
3270 }
3271 EXPORT_SYMBOL(tcp_connect);
3272 
3273 /* Send out a delayed ack, the caller does the policy checking
3274  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3275  * for details.
3276  */
3277 void tcp_send_delayed_ack(struct sock *sk)
3278 {
3279 	struct inet_connection_sock *icsk = inet_csk(sk);
3280 	int ato = icsk->icsk_ack.ato;
3281 	unsigned long timeout;
3282 
3283 	tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3284 
3285 	if (ato > TCP_DELACK_MIN) {
3286 		const struct tcp_sock *tp = tcp_sk(sk);
3287 		int max_ato = HZ / 2;
3288 
3289 		if (icsk->icsk_ack.pingpong ||
3290 		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3291 			max_ato = TCP_DELACK_MAX;
3292 
3293 		/* Slow path, intersegment interval is "high". */
3294 
3295 		/* If some rtt estimate is known, use it to bound delayed ack.
3296 		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3297 		 * directly.
3298 		 */
3299 		if (tp->srtt_us) {
3300 			int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3301 					TCP_DELACK_MIN);
3302 
3303 			if (rtt < max_ato)
3304 				max_ato = rtt;
3305 		}
3306 
3307 		ato = min(ato, max_ato);
3308 	}
3309 
3310 	/* Stay within the limit we were given */
3311 	timeout = jiffies + ato;
3312 
3313 	/* Use new timeout only if there wasn't a older one earlier. */
3314 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3315 		/* If delack timer was blocked or is about to expire,
3316 		 * send ACK now.
3317 		 */
3318 		if (icsk->icsk_ack.blocked ||
3319 		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3320 			tcp_send_ack(sk);
3321 			return;
3322 		}
3323 
3324 		if (!time_before(timeout, icsk->icsk_ack.timeout))
3325 			timeout = icsk->icsk_ack.timeout;
3326 	}
3327 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3328 	icsk->icsk_ack.timeout = timeout;
3329 	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3330 }
3331 
3332 /* This routine sends an ack and also updates the window. */
3333 void tcp_send_ack(struct sock *sk)
3334 {
3335 	struct sk_buff *buff;
3336 
3337 	/* If we have been reset, we may not send again. */
3338 	if (sk->sk_state == TCP_CLOSE)
3339 		return;
3340 
3341 	tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3342 
3343 	/* We are not putting this on the write queue, so
3344 	 * tcp_transmit_skb() will set the ownership to this
3345 	 * sock.
3346 	 */
3347 	buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3348 	if (!buff) {
3349 		inet_csk_schedule_ack(sk);
3350 		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3351 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3352 					  TCP_DELACK_MAX, TCP_RTO_MAX);
3353 		return;
3354 	}
3355 
3356 	/* Reserve space for headers and prepare control bits. */
3357 	skb_reserve(buff, MAX_TCP_HEADER);
3358 	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3359 
3360 	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
3361 	 * too much.
3362 	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3363 	 * We also avoid tcp_wfree() overhead (cache line miss accessing
3364 	 * tp->tsq_flags) by using regular sock_wfree()
3365 	 */
3366 	skb_set_tcp_pure_ack(buff);
3367 
3368 	/* Send it off, this clears delayed acks for us. */
3369 	skb_mstamp_get(&buff->skb_mstamp);
3370 	tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3371 }
3372 EXPORT_SYMBOL_GPL(tcp_send_ack);
3373 
3374 /* This routine sends a packet with an out of date sequence
3375  * number. It assumes the other end will try to ack it.
3376  *
3377  * Question: what should we make while urgent mode?
3378  * 4.4BSD forces sending single byte of data. We cannot send
3379  * out of window data, because we have SND.NXT==SND.MAX...
3380  *
3381  * Current solution: to send TWO zero-length segments in urgent mode:
3382  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3383  * out-of-date with SND.UNA-1 to probe window.
3384  */
3385 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3386 {
3387 	struct tcp_sock *tp = tcp_sk(sk);
3388 	struct sk_buff *skb;
3389 
3390 	/* We don't queue it, tcp_transmit_skb() sets ownership. */
3391 	skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3392 	if (!skb)
3393 		return -1;
3394 
3395 	/* Reserve space for headers and set control bits. */
3396 	skb_reserve(skb, MAX_TCP_HEADER);
3397 	/* Use a previous sequence.  This should cause the other
3398 	 * end to send an ack.  Don't queue or clone SKB, just
3399 	 * send it.
3400 	 */
3401 	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3402 	skb_mstamp_get(&skb->skb_mstamp);
3403 	return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3404 }
3405 
3406 void tcp_send_window_probe(struct sock *sk)
3407 {
3408 	if (sk->sk_state == TCP_ESTABLISHED) {
3409 		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3410 		tcp_xmit_probe_skb(sk, 0);
3411 	}
3412 }
3413 
3414 /* Initiate keepalive or window probe from timer. */
3415 int tcp_write_wakeup(struct sock *sk)
3416 {
3417 	struct tcp_sock *tp = tcp_sk(sk);
3418 	struct sk_buff *skb;
3419 
3420 	if (sk->sk_state == TCP_CLOSE)
3421 		return -1;
3422 
3423 	skb = tcp_send_head(sk);
3424 	if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3425 		int err;
3426 		unsigned int mss = tcp_current_mss(sk);
3427 		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3428 
3429 		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3430 			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3431 
3432 		/* We are probing the opening of a window
3433 		 * but the window size is != 0
3434 		 * must have been a result SWS avoidance ( sender )
3435 		 */
3436 		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3437 		    skb->len > mss) {
3438 			seg_size = min(seg_size, mss);
3439 			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3440 			if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3441 				return -1;
3442 		} else if (!tcp_skb_pcount(skb))
3443 			tcp_set_skb_tso_segs(sk, skb, mss);
3444 
3445 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3446 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3447 		if (!err)
3448 			tcp_event_new_data_sent(sk, skb);
3449 		return err;
3450 	} else {
3451 		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3452 			tcp_xmit_probe_skb(sk, 1);
3453 		return tcp_xmit_probe_skb(sk, 0);
3454 	}
3455 }
3456 
3457 /* A window probe timeout has occurred.  If window is not closed send
3458  * a partial packet else a zero probe.
3459  */
3460 void tcp_send_probe0(struct sock *sk)
3461 {
3462 	struct inet_connection_sock *icsk = inet_csk(sk);
3463 	struct tcp_sock *tp = tcp_sk(sk);
3464 	unsigned long probe_max;
3465 	int err;
3466 
3467 	err = tcp_write_wakeup(sk);
3468 
3469 	if (tp->packets_out || !tcp_send_head(sk)) {
3470 		/* Cancel probe timer, if it is not required. */
3471 		icsk->icsk_probes_out = 0;
3472 		icsk->icsk_backoff = 0;
3473 		return;
3474 	}
3475 
3476 	if (err <= 0) {
3477 		if (icsk->icsk_backoff < sysctl_tcp_retries2)
3478 			icsk->icsk_backoff++;
3479 		icsk->icsk_probes_out++;
3480 		probe_max = TCP_RTO_MAX;
3481 	} else {
3482 		/* If packet was not sent due to local congestion,
3483 		 * do not backoff and do not remember icsk_probes_out.
3484 		 * Let local senders to fight for local resources.
3485 		 *
3486 		 * Use accumulated backoff yet.
3487 		 */
3488 		if (!icsk->icsk_probes_out)
3489 			icsk->icsk_probes_out = 1;
3490 		probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3491 	}
3492 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3493 				  inet_csk_rto_backoff(icsk, probe_max),
3494 				  TCP_RTO_MAX);
3495 }
3496 
3497 int tcp_rtx_synack(struct sock *sk, struct request_sock *req)
3498 {
3499 	const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3500 	struct flowi fl;
3501 	int res;
3502 
3503 	res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL);
3504 	if (!res) {
3505 		TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3506 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3507 	}
3508 	return res;
3509 }
3510 EXPORT_SYMBOL(tcp_rtx_synack);
3511