xref: /openbmc/linux/net/ipv4/tcp_output.c (revision 9b9c2cd4)
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 four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
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  */
142 void tcp_cwnd_restart(struct sock *sk, s32 delta)
143 {
144 	struct tcp_sock *tp = tcp_sk(sk);
145 	u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
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 
167 	if (tcp_packets_in_flight(tp) == 0)
168 		tcp_ca_event(sk, CA_EVENT_TX_START);
169 
170 	tp->lsndtime = now;
171 
172 	/* If it is a reply for ato after last received
173 	 * packet, enter pingpong mode.
174 	 */
175 	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 		icsk->icsk_ack.pingpong = 1;
177 }
178 
179 /* Account for an ACK we sent. */
180 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
181 {
182 	tcp_dec_quickack_mode(sk, pkts);
183 	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
184 }
185 
186 
187 u32 tcp_default_init_rwnd(u32 mss)
188 {
189 	/* Initial receive window should be twice of TCP_INIT_CWND to
190 	 * enable proper sending of new unsent data during fast recovery
191 	 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
192 	 * limit when mss is larger than 1460.
193 	 */
194 	u32 init_rwnd = TCP_INIT_CWND * 2;
195 
196 	if (mss > 1460)
197 		init_rwnd = max((1460 * init_rwnd) / mss, 2U);
198 	return init_rwnd;
199 }
200 
201 /* Determine a window scaling and initial window to offer.
202  * Based on the assumption that the given amount of space
203  * will be offered. Store the results in the tp structure.
204  * NOTE: for smooth operation initial space offering should
205  * be a multiple of mss if possible. We assume here that mss >= 1.
206  * This MUST be enforced by all callers.
207  */
208 void tcp_select_initial_window(int __space, __u32 mss,
209 			       __u32 *rcv_wnd, __u32 *window_clamp,
210 			       int wscale_ok, __u8 *rcv_wscale,
211 			       __u32 init_rcv_wnd)
212 {
213 	unsigned int space = (__space < 0 ? 0 : __space);
214 
215 	/* If no clamp set the clamp to the max possible scaled window */
216 	if (*window_clamp == 0)
217 		(*window_clamp) = (65535 << 14);
218 	space = min(*window_clamp, space);
219 
220 	/* Quantize space offering to a multiple of mss if possible. */
221 	if (space > mss)
222 		space = (space / mss) * mss;
223 
224 	/* NOTE: offering an initial window larger than 32767
225 	 * will break some buggy TCP stacks. If the admin tells us
226 	 * it is likely we could be speaking with such a buggy stack
227 	 * we will truncate our initial window offering to 32K-1
228 	 * unless the remote has sent us a window scaling option,
229 	 * which we interpret as a sign the remote TCP is not
230 	 * misinterpreting the window field as a signed quantity.
231 	 */
232 	if (sysctl_tcp_workaround_signed_windows)
233 		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
234 	else
235 		(*rcv_wnd) = space;
236 
237 	(*rcv_wscale) = 0;
238 	if (wscale_ok) {
239 		/* Set window scaling on max possible window
240 		 * See RFC1323 for an explanation of the limit to 14
241 		 */
242 		space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
243 		space = min_t(u32, space, *window_clamp);
244 		while (space > 65535 && (*rcv_wscale) < 14) {
245 			space >>= 1;
246 			(*rcv_wscale)++;
247 		}
248 	}
249 
250 	if (mss > (1 << *rcv_wscale)) {
251 		if (!init_rcv_wnd) /* Use default unless specified otherwise */
252 			init_rcv_wnd = tcp_default_init_rwnd(mss);
253 		*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
254 	}
255 
256 	/* Set the clamp no higher than max representable value */
257 	(*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
258 }
259 EXPORT_SYMBOL(tcp_select_initial_window);
260 
261 /* Chose a new window to advertise, update state in tcp_sock for the
262  * socket, and return result with RFC1323 scaling applied.  The return
263  * value can be stuffed directly into th->window for an outgoing
264  * frame.
265  */
266 static u16 tcp_select_window(struct sock *sk)
267 {
268 	struct tcp_sock *tp = tcp_sk(sk);
269 	u32 old_win = tp->rcv_wnd;
270 	u32 cur_win = tcp_receive_window(tp);
271 	u32 new_win = __tcp_select_window(sk);
272 
273 	/* Never shrink the offered window */
274 	if (new_win < cur_win) {
275 		/* Danger Will Robinson!
276 		 * Don't update rcv_wup/rcv_wnd here or else
277 		 * we will not be able to advertise a zero
278 		 * window in time.  --DaveM
279 		 *
280 		 * Relax Will Robinson.
281 		 */
282 		if (new_win == 0)
283 			NET_INC_STATS(sock_net(sk),
284 				      LINUX_MIB_TCPWANTZEROWINDOWADV);
285 		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
286 	}
287 	tp->rcv_wnd = new_win;
288 	tp->rcv_wup = tp->rcv_nxt;
289 
290 	/* Make sure we do not exceed the maximum possible
291 	 * scaled window.
292 	 */
293 	if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
294 		new_win = min(new_win, MAX_TCP_WINDOW);
295 	else
296 		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
297 
298 	/* RFC1323 scaling applied */
299 	new_win >>= tp->rx_opt.rcv_wscale;
300 
301 	/* If we advertise zero window, disable fast path. */
302 	if (new_win == 0) {
303 		tp->pred_flags = 0;
304 		if (old_win)
305 			NET_INC_STATS(sock_net(sk),
306 				      LINUX_MIB_TCPTOZEROWINDOWADV);
307 	} else if (old_win == 0) {
308 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
309 	}
310 
311 	return new_win;
312 }
313 
314 /* Packet ECN state for a SYN-ACK */
315 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
316 {
317 	const struct tcp_sock *tp = tcp_sk(sk);
318 
319 	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
320 	if (!(tp->ecn_flags & TCP_ECN_OK))
321 		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
322 	else if (tcp_ca_needs_ecn(sk))
323 		INET_ECN_xmit(sk);
324 }
325 
326 /* Packet ECN state for a SYN.  */
327 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
328 {
329 	struct tcp_sock *tp = tcp_sk(sk);
330 	bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
331 		       tcp_ca_needs_ecn(sk);
332 
333 	if (!use_ecn) {
334 		const struct dst_entry *dst = __sk_dst_get(sk);
335 
336 		if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
337 			use_ecn = true;
338 	}
339 
340 	tp->ecn_flags = 0;
341 
342 	if (use_ecn) {
343 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
344 		tp->ecn_flags = TCP_ECN_OK;
345 		if (tcp_ca_needs_ecn(sk))
346 			INET_ECN_xmit(sk);
347 	}
348 }
349 
350 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
351 {
352 	if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
353 		/* tp->ecn_flags are cleared at a later point in time when
354 		 * SYN ACK is ultimatively being received.
355 		 */
356 		TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
357 }
358 
359 static void
360 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
361 {
362 	if (inet_rsk(req)->ecn_ok)
363 		th->ece = 1;
364 }
365 
366 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
367  * be sent.
368  */
369 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
370 				int tcp_header_len)
371 {
372 	struct tcp_sock *tp = tcp_sk(sk);
373 
374 	if (tp->ecn_flags & TCP_ECN_OK) {
375 		/* Not-retransmitted data segment: set ECT and inject CWR. */
376 		if (skb->len != tcp_header_len &&
377 		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
378 			INET_ECN_xmit(sk);
379 			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
380 				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
381 				tcp_hdr(skb)->cwr = 1;
382 				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
383 			}
384 		} else if (!tcp_ca_needs_ecn(sk)) {
385 			/* ACK or retransmitted segment: clear ECT|CE */
386 			INET_ECN_dontxmit(sk);
387 		}
388 		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
389 			tcp_hdr(skb)->ece = 1;
390 	}
391 }
392 
393 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
394  * auto increment end seqno.
395  */
396 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
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 
406 	TCP_SKB_CB(skb)->seq = seq;
407 	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
408 		seq++;
409 	TCP_SKB_CB(skb)->end_seq = seq;
410 }
411 
412 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
413 {
414 	return tp->snd_una != tp->snd_up;
415 }
416 
417 #define OPTION_SACK_ADVERTISE	(1 << 0)
418 #define OPTION_TS		(1 << 1)
419 #define OPTION_MD5		(1 << 2)
420 #define OPTION_WSCALE		(1 << 3)
421 #define OPTION_FAST_OPEN_COOKIE	(1 << 8)
422 
423 struct tcp_out_options {
424 	u16 options;		/* bit field of OPTION_* */
425 	u16 mss;		/* 0 to disable */
426 	u8 ws;			/* window scale, 0 to disable */
427 	u8 num_sack_blocks;	/* number of SACK blocks to include */
428 	u8 hash_size;		/* bytes in hash_location */
429 	__u8 *hash_location;	/* temporary pointer, overloaded */
430 	__u32 tsval, tsecr;	/* need to include OPTION_TS */
431 	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
432 };
433 
434 /* Write previously computed TCP options to the packet.
435  *
436  * Beware: Something in the Internet is very sensitive to the ordering of
437  * TCP options, we learned this through the hard way, so be careful here.
438  * Luckily we can at least blame others for their non-compliance but from
439  * inter-operability perspective it seems that we're somewhat stuck with
440  * the ordering which we have been using if we want to keep working with
441  * those broken things (not that it currently hurts anybody as there isn't
442  * particular reason why the ordering would need to be changed).
443  *
444  * At least SACK_PERM as the first option is known to lead to a disaster
445  * (but it may well be that other scenarios fail similarly).
446  */
447 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
448 			      struct tcp_out_options *opts)
449 {
450 	u16 options = opts->options;	/* mungable copy */
451 
452 	if (unlikely(OPTION_MD5 & options)) {
453 		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
454 			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
455 		/* overload cookie hash location */
456 		opts->hash_location = (__u8 *)ptr;
457 		ptr += 4;
458 	}
459 
460 	if (unlikely(opts->mss)) {
461 		*ptr++ = htonl((TCPOPT_MSS << 24) |
462 			       (TCPOLEN_MSS << 16) |
463 			       opts->mss);
464 	}
465 
466 	if (likely(OPTION_TS & options)) {
467 		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
468 			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
469 				       (TCPOLEN_SACK_PERM << 16) |
470 				       (TCPOPT_TIMESTAMP << 8) |
471 				       TCPOLEN_TIMESTAMP);
472 			options &= ~OPTION_SACK_ADVERTISE;
473 		} else {
474 			*ptr++ = htonl((TCPOPT_NOP << 24) |
475 				       (TCPOPT_NOP << 16) |
476 				       (TCPOPT_TIMESTAMP << 8) |
477 				       TCPOLEN_TIMESTAMP);
478 		}
479 		*ptr++ = htonl(opts->tsval);
480 		*ptr++ = htonl(opts->tsecr);
481 	}
482 
483 	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
484 		*ptr++ = htonl((TCPOPT_NOP << 24) |
485 			       (TCPOPT_NOP << 16) |
486 			       (TCPOPT_SACK_PERM << 8) |
487 			       TCPOLEN_SACK_PERM);
488 	}
489 
490 	if (unlikely(OPTION_WSCALE & options)) {
491 		*ptr++ = htonl((TCPOPT_NOP << 24) |
492 			       (TCPOPT_WINDOW << 16) |
493 			       (TCPOLEN_WINDOW << 8) |
494 			       opts->ws);
495 	}
496 
497 	if (unlikely(opts->num_sack_blocks)) {
498 		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
499 			tp->duplicate_sack : tp->selective_acks;
500 		int this_sack;
501 
502 		*ptr++ = htonl((TCPOPT_NOP  << 24) |
503 			       (TCPOPT_NOP  << 16) |
504 			       (TCPOPT_SACK <<  8) |
505 			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
506 						     TCPOLEN_SACK_PERBLOCK)));
507 
508 		for (this_sack = 0; this_sack < opts->num_sack_blocks;
509 		     ++this_sack) {
510 			*ptr++ = htonl(sp[this_sack].start_seq);
511 			*ptr++ = htonl(sp[this_sack].end_seq);
512 		}
513 
514 		tp->rx_opt.dsack = 0;
515 	}
516 
517 	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
518 		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
519 		u8 *p = (u8 *)ptr;
520 		u32 len; /* Fast Open option length */
521 
522 		if (foc->exp) {
523 			len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
524 			*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
525 				     TCPOPT_FASTOPEN_MAGIC);
526 			p += TCPOLEN_EXP_FASTOPEN_BASE;
527 		} else {
528 			len = TCPOLEN_FASTOPEN_BASE + foc->len;
529 			*p++ = TCPOPT_FASTOPEN;
530 			*p++ = len;
531 		}
532 
533 		memcpy(p, foc->val, foc->len);
534 		if ((len & 3) == 2) {
535 			p[foc->len] = TCPOPT_NOP;
536 			p[foc->len + 1] = TCPOPT_NOP;
537 		}
538 		ptr += (len + 3) >> 2;
539 	}
540 }
541 
542 /* Compute TCP options for SYN packets. This is not the final
543  * network wire format yet.
544  */
545 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
546 				struct tcp_out_options *opts,
547 				struct tcp_md5sig_key **md5)
548 {
549 	struct tcp_sock *tp = tcp_sk(sk);
550 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
551 	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
552 
553 #ifdef CONFIG_TCP_MD5SIG
554 	*md5 = tp->af_specific->md5_lookup(sk, sk);
555 	if (*md5) {
556 		opts->options |= OPTION_MD5;
557 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
558 	}
559 #else
560 	*md5 = NULL;
561 #endif
562 
563 	/* We always get an MSS option.  The option bytes which will be seen in
564 	 * normal data packets should timestamps be used, must be in the MSS
565 	 * advertised.  But we subtract them from tp->mss_cache so that
566 	 * calculations in tcp_sendmsg are simpler etc.  So account for this
567 	 * fact here if necessary.  If we don't do this correctly, as a
568 	 * receiver we won't recognize data packets as being full sized when we
569 	 * should, and thus we won't abide by the delayed ACK rules correctly.
570 	 * SACKs don't matter, we never delay an ACK when we have any of those
571 	 * going out.  */
572 	opts->mss = tcp_advertise_mss(sk);
573 	remaining -= TCPOLEN_MSS_ALIGNED;
574 
575 	if (likely(sysctl_tcp_timestamps && !*md5)) {
576 		opts->options |= OPTION_TS;
577 		opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
578 		opts->tsecr = tp->rx_opt.ts_recent;
579 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
580 	}
581 	if (likely(sysctl_tcp_window_scaling)) {
582 		opts->ws = tp->rx_opt.rcv_wscale;
583 		opts->options |= OPTION_WSCALE;
584 		remaining -= TCPOLEN_WSCALE_ALIGNED;
585 	}
586 	if (likely(sysctl_tcp_sack)) {
587 		opts->options |= OPTION_SACK_ADVERTISE;
588 		if (unlikely(!(OPTION_TS & opts->options)))
589 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
590 	}
591 
592 	if (fastopen && fastopen->cookie.len >= 0) {
593 		u32 need = fastopen->cookie.len;
594 
595 		need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
596 					       TCPOLEN_FASTOPEN_BASE;
597 		need = (need + 3) & ~3U;  /* Align to 32 bits */
598 		if (remaining >= need) {
599 			opts->options |= OPTION_FAST_OPEN_COOKIE;
600 			opts->fastopen_cookie = &fastopen->cookie;
601 			remaining -= need;
602 			tp->syn_fastopen = 1;
603 			tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
604 		}
605 	}
606 
607 	return MAX_TCP_OPTION_SPACE - remaining;
608 }
609 
610 /* Set up TCP options for SYN-ACKs. */
611 static unsigned int tcp_synack_options(struct request_sock *req,
612 				       unsigned int mss, struct sk_buff *skb,
613 				       struct tcp_out_options *opts,
614 				       const struct tcp_md5sig_key *md5,
615 				       struct tcp_fastopen_cookie *foc)
616 {
617 	struct inet_request_sock *ireq = inet_rsk(req);
618 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
619 
620 #ifdef CONFIG_TCP_MD5SIG
621 	if (md5) {
622 		opts->options |= OPTION_MD5;
623 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
624 
625 		/* We can't fit any SACK blocks in a packet with MD5 + TS
626 		 * options. There was discussion about disabling SACK
627 		 * rather than TS in order to fit in better with old,
628 		 * buggy kernels, but that was deemed to be unnecessary.
629 		 */
630 		ireq->tstamp_ok &= !ireq->sack_ok;
631 	}
632 #endif
633 
634 	/* We always send an MSS option. */
635 	opts->mss = mss;
636 	remaining -= TCPOLEN_MSS_ALIGNED;
637 
638 	if (likely(ireq->wscale_ok)) {
639 		opts->ws = ireq->rcv_wscale;
640 		opts->options |= OPTION_WSCALE;
641 		remaining -= TCPOLEN_WSCALE_ALIGNED;
642 	}
643 	if (likely(ireq->tstamp_ok)) {
644 		opts->options |= OPTION_TS;
645 		opts->tsval = tcp_skb_timestamp(skb);
646 		opts->tsecr = req->ts_recent;
647 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
648 	}
649 	if (likely(ireq->sack_ok)) {
650 		opts->options |= OPTION_SACK_ADVERTISE;
651 		if (unlikely(!ireq->tstamp_ok))
652 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
653 	}
654 	if (foc != NULL && foc->len >= 0) {
655 		u32 need = foc->len;
656 
657 		need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
658 				   TCPOLEN_FASTOPEN_BASE;
659 		need = (need + 3) & ~3U;  /* Align to 32 bits */
660 		if (remaining >= need) {
661 			opts->options |= OPTION_FAST_OPEN_COOKIE;
662 			opts->fastopen_cookie = foc;
663 			remaining -= need;
664 		}
665 	}
666 
667 	return MAX_TCP_OPTION_SPACE - remaining;
668 }
669 
670 /* Compute TCP options for ESTABLISHED sockets. This is not the
671  * final wire format yet.
672  */
673 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
674 					struct tcp_out_options *opts,
675 					struct tcp_md5sig_key **md5)
676 {
677 	struct tcp_sock *tp = tcp_sk(sk);
678 	unsigned int size = 0;
679 	unsigned int eff_sacks;
680 
681 	opts->options = 0;
682 
683 #ifdef CONFIG_TCP_MD5SIG
684 	*md5 = tp->af_specific->md5_lookup(sk, sk);
685 	if (unlikely(*md5)) {
686 		opts->options |= OPTION_MD5;
687 		size += TCPOLEN_MD5SIG_ALIGNED;
688 	}
689 #else
690 	*md5 = NULL;
691 #endif
692 
693 	if (likely(tp->rx_opt.tstamp_ok)) {
694 		opts->options |= OPTION_TS;
695 		opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
696 		opts->tsecr = tp->rx_opt.ts_recent;
697 		size += TCPOLEN_TSTAMP_ALIGNED;
698 	}
699 
700 	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
701 	if (unlikely(eff_sacks)) {
702 		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
703 		opts->num_sack_blocks =
704 			min_t(unsigned int, eff_sacks,
705 			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
706 			      TCPOLEN_SACK_PERBLOCK);
707 		size += TCPOLEN_SACK_BASE_ALIGNED +
708 			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
709 	}
710 
711 	return size;
712 }
713 
714 
715 /* TCP SMALL QUEUES (TSQ)
716  *
717  * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
718  * to reduce RTT and bufferbloat.
719  * We do this using a special skb destructor (tcp_wfree).
720  *
721  * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
722  * needs to be reallocated in a driver.
723  * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
724  *
725  * Since transmit from skb destructor is forbidden, we use a tasklet
726  * to process all sockets that eventually need to send more skbs.
727  * We use one tasklet per cpu, with its own queue of sockets.
728  */
729 struct tsq_tasklet {
730 	struct tasklet_struct	tasklet;
731 	struct list_head	head; /* queue of tcp sockets */
732 };
733 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
734 
735 static void tcp_tsq_handler(struct sock *sk)
736 {
737 	if ((1 << sk->sk_state) &
738 	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
739 	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK))
740 		tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
741 			       0, GFP_ATOMIC);
742 }
743 /*
744  * One tasklet per cpu tries to send more skbs.
745  * We run in tasklet context but need to disable irqs when
746  * transferring tsq->head because tcp_wfree() might
747  * interrupt us (non NAPI drivers)
748  */
749 static void tcp_tasklet_func(unsigned long data)
750 {
751 	struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
752 	LIST_HEAD(list);
753 	unsigned long flags;
754 	struct list_head *q, *n;
755 	struct tcp_sock *tp;
756 	struct sock *sk;
757 
758 	local_irq_save(flags);
759 	list_splice_init(&tsq->head, &list);
760 	local_irq_restore(flags);
761 
762 	list_for_each_safe(q, n, &list) {
763 		tp = list_entry(q, struct tcp_sock, tsq_node);
764 		list_del(&tp->tsq_node);
765 
766 		sk = (struct sock *)tp;
767 		bh_lock_sock(sk);
768 
769 		if (!sock_owned_by_user(sk)) {
770 			tcp_tsq_handler(sk);
771 		} else {
772 			/* defer the work to tcp_release_cb() */
773 			set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
774 		}
775 		bh_unlock_sock(sk);
776 
777 		clear_bit(TSQ_QUEUED, &tp->tsq_flags);
778 		sk_free(sk);
779 	}
780 }
781 
782 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) |		\
783 			  (1UL << TCP_WRITE_TIMER_DEFERRED) |	\
784 			  (1UL << TCP_DELACK_TIMER_DEFERRED) |	\
785 			  (1UL << TCP_MTU_REDUCED_DEFERRED))
786 /**
787  * tcp_release_cb - tcp release_sock() callback
788  * @sk: socket
789  *
790  * called from release_sock() to perform protocol dependent
791  * actions before socket release.
792  */
793 void tcp_release_cb(struct sock *sk)
794 {
795 	struct tcp_sock *tp = tcp_sk(sk);
796 	unsigned long flags, nflags;
797 
798 	/* perform an atomic operation only if at least one flag is set */
799 	do {
800 		flags = tp->tsq_flags;
801 		if (!(flags & TCP_DEFERRED_ALL))
802 			return;
803 		nflags = flags & ~TCP_DEFERRED_ALL;
804 	} while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
805 
806 	if (flags & (1UL << TCP_TSQ_DEFERRED))
807 		tcp_tsq_handler(sk);
808 
809 	/* Here begins the tricky part :
810 	 * We are called from release_sock() with :
811 	 * 1) BH disabled
812 	 * 2) sk_lock.slock spinlock held
813 	 * 3) socket owned by us (sk->sk_lock.owned == 1)
814 	 *
815 	 * But following code is meant to be called from BH handlers,
816 	 * so we should keep BH disabled, but early release socket ownership
817 	 */
818 	sock_release_ownership(sk);
819 
820 	if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
821 		tcp_write_timer_handler(sk);
822 		__sock_put(sk);
823 	}
824 	if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
825 		tcp_delack_timer_handler(sk);
826 		__sock_put(sk);
827 	}
828 	if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
829 		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
830 		__sock_put(sk);
831 	}
832 }
833 EXPORT_SYMBOL(tcp_release_cb);
834 
835 void __init tcp_tasklet_init(void)
836 {
837 	int i;
838 
839 	for_each_possible_cpu(i) {
840 		struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
841 
842 		INIT_LIST_HEAD(&tsq->head);
843 		tasklet_init(&tsq->tasklet,
844 			     tcp_tasklet_func,
845 			     (unsigned long)tsq);
846 	}
847 }
848 
849 /*
850  * Write buffer destructor automatically called from kfree_skb.
851  * We can't xmit new skbs from this context, as we might already
852  * hold qdisc lock.
853  */
854 void tcp_wfree(struct sk_buff *skb)
855 {
856 	struct sock *sk = skb->sk;
857 	struct tcp_sock *tp = tcp_sk(sk);
858 	int wmem;
859 
860 	/* Keep one reference on sk_wmem_alloc.
861 	 * Will be released by sk_free() from here or tcp_tasklet_func()
862 	 */
863 	wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
864 
865 	/* If this softirq is serviced by ksoftirqd, we are likely under stress.
866 	 * Wait until our queues (qdisc + devices) are drained.
867 	 * This gives :
868 	 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
869 	 * - chance for incoming ACK (processed by another cpu maybe)
870 	 *   to migrate this flow (skb->ooo_okay will be eventually set)
871 	 */
872 	if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
873 		goto out;
874 
875 	if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
876 	    !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
877 		unsigned long flags;
878 		struct tsq_tasklet *tsq;
879 
880 		/* queue this socket to tasklet queue */
881 		local_irq_save(flags);
882 		tsq = this_cpu_ptr(&tsq_tasklet);
883 		list_add(&tp->tsq_node, &tsq->head);
884 		tasklet_schedule(&tsq->tasklet);
885 		local_irq_restore(flags);
886 		return;
887 	}
888 out:
889 	sk_free(sk);
890 }
891 
892 /* This routine actually transmits TCP packets queued in by
893  * tcp_do_sendmsg().  This is used by both the initial
894  * transmission and possible later retransmissions.
895  * All SKB's seen here are completely headerless.  It is our
896  * job to build the TCP header, and pass the packet down to
897  * IP so it can do the same plus pass the packet off to the
898  * device.
899  *
900  * We are working here with either a clone of the original
901  * SKB, or a fresh unique copy made by the retransmit engine.
902  */
903 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
904 			    gfp_t gfp_mask)
905 {
906 	const struct inet_connection_sock *icsk = inet_csk(sk);
907 	struct inet_sock *inet;
908 	struct tcp_sock *tp;
909 	struct tcp_skb_cb *tcb;
910 	struct tcp_out_options opts;
911 	unsigned int tcp_options_size, tcp_header_size;
912 	struct tcp_md5sig_key *md5;
913 	struct tcphdr *th;
914 	int err;
915 
916 	BUG_ON(!skb || !tcp_skb_pcount(skb));
917 
918 	if (clone_it) {
919 		skb_mstamp_get(&skb->skb_mstamp);
920 
921 		if (unlikely(skb_cloned(skb)))
922 			skb = pskb_copy(skb, gfp_mask);
923 		else
924 			skb = skb_clone(skb, gfp_mask);
925 		if (unlikely(!skb))
926 			return -ENOBUFS;
927 	}
928 
929 	inet = inet_sk(sk);
930 	tp = tcp_sk(sk);
931 	tcb = TCP_SKB_CB(skb);
932 	memset(&opts, 0, sizeof(opts));
933 
934 	if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
935 		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
936 	else
937 		tcp_options_size = tcp_established_options(sk, skb, &opts,
938 							   &md5);
939 	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
940 
941 	/* if no packet is in qdisc/device queue, then allow XPS to select
942 	 * another queue. We can be called from tcp_tsq_handler()
943 	 * which holds one reference to sk_wmem_alloc.
944 	 *
945 	 * TODO: Ideally, in-flight pure ACK packets should not matter here.
946 	 * One way to get this would be to set skb->truesize = 2 on them.
947 	 */
948 	skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
949 
950 	skb_push(skb, tcp_header_size);
951 	skb_reset_transport_header(skb);
952 
953 	skb_orphan(skb);
954 	skb->sk = sk;
955 	skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree;
956 	skb_set_hash_from_sk(skb, sk);
957 	atomic_add(skb->truesize, &sk->sk_wmem_alloc);
958 
959 	/* Build TCP header and checksum it. */
960 	th = tcp_hdr(skb);
961 	th->source		= inet->inet_sport;
962 	th->dest		= inet->inet_dport;
963 	th->seq			= htonl(tcb->seq);
964 	th->ack_seq		= htonl(tp->rcv_nxt);
965 	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
966 					tcb->tcp_flags);
967 
968 	if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
969 		/* RFC1323: The window in SYN & SYN/ACK segments
970 		 * is never scaled.
971 		 */
972 		th->window	= htons(min(tp->rcv_wnd, 65535U));
973 	} else {
974 		th->window	= htons(tcp_select_window(sk));
975 	}
976 	th->check		= 0;
977 	th->urg_ptr		= 0;
978 
979 	/* The urg_mode check is necessary during a below snd_una win probe */
980 	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
981 		if (before(tp->snd_up, tcb->seq + 0x10000)) {
982 			th->urg_ptr = htons(tp->snd_up - tcb->seq);
983 			th->urg = 1;
984 		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
985 			th->urg_ptr = htons(0xFFFF);
986 			th->urg = 1;
987 		}
988 	}
989 
990 	tcp_options_write((__be32 *)(th + 1), tp, &opts);
991 	skb_shinfo(skb)->gso_type = sk->sk_gso_type;
992 	if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
993 		tcp_ecn_send(sk, skb, tcp_header_size);
994 
995 #ifdef CONFIG_TCP_MD5SIG
996 	/* Calculate the MD5 hash, as we have all we need now */
997 	if (md5) {
998 		sk_nocaps_add(sk, NETIF_F_GSO_MASK);
999 		tp->af_specific->calc_md5_hash(opts.hash_location,
1000 					       md5, sk, skb);
1001 	}
1002 #endif
1003 
1004 	icsk->icsk_af_ops->send_check(sk, skb);
1005 
1006 	if (likely(tcb->tcp_flags & TCPHDR_ACK))
1007 		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1008 
1009 	if (skb->len != tcp_header_size)
1010 		tcp_event_data_sent(tp, sk);
1011 
1012 	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1013 		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1014 			      tcp_skb_pcount(skb));
1015 
1016 	tp->segs_out += tcp_skb_pcount(skb);
1017 	/* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1018 	skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1019 	skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1020 
1021 	/* Our usage of tstamp should remain private */
1022 	skb->tstamp.tv64 = 0;
1023 
1024 	/* Cleanup our debris for IP stacks */
1025 	memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1026 			       sizeof(struct inet6_skb_parm)));
1027 
1028 	err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1029 
1030 	if (likely(err <= 0))
1031 		return err;
1032 
1033 	tcp_enter_cwr(sk);
1034 
1035 	return net_xmit_eval(err);
1036 }
1037 
1038 /* This routine just queues the buffer for sending.
1039  *
1040  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1041  * otherwise socket can stall.
1042  */
1043 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1044 {
1045 	struct tcp_sock *tp = tcp_sk(sk);
1046 
1047 	/* Advance write_seq and place onto the write_queue. */
1048 	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1049 	__skb_header_release(skb);
1050 	tcp_add_write_queue_tail(sk, skb);
1051 	sk->sk_wmem_queued += skb->truesize;
1052 	sk_mem_charge(sk, skb->truesize);
1053 }
1054 
1055 /* Initialize TSO segments for a packet. */
1056 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1057 {
1058 	if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1059 		/* Avoid the costly divide in the normal
1060 		 * non-TSO case.
1061 		 */
1062 		tcp_skb_pcount_set(skb, 1);
1063 		TCP_SKB_CB(skb)->tcp_gso_size = 0;
1064 	} else {
1065 		tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1066 		TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1067 	}
1068 }
1069 
1070 /* When a modification to fackets out becomes necessary, we need to check
1071  * skb is counted to fackets_out or not.
1072  */
1073 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1074 				   int decr)
1075 {
1076 	struct tcp_sock *tp = tcp_sk(sk);
1077 
1078 	if (!tp->sacked_out || tcp_is_reno(tp))
1079 		return;
1080 
1081 	if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1082 		tp->fackets_out -= decr;
1083 }
1084 
1085 /* Pcount in the middle of the write queue got changed, we need to do various
1086  * tweaks to fix counters
1087  */
1088 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1089 {
1090 	struct tcp_sock *tp = tcp_sk(sk);
1091 
1092 	tp->packets_out -= decr;
1093 
1094 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1095 		tp->sacked_out -= decr;
1096 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1097 		tp->retrans_out -= decr;
1098 	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1099 		tp->lost_out -= decr;
1100 
1101 	/* Reno case is special. Sigh... */
1102 	if (tcp_is_reno(tp) && decr > 0)
1103 		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1104 
1105 	tcp_adjust_fackets_out(sk, skb, decr);
1106 
1107 	if (tp->lost_skb_hint &&
1108 	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1109 	    (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1110 		tp->lost_cnt_hint -= decr;
1111 
1112 	tcp_verify_left_out(tp);
1113 }
1114 
1115 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1116 {
1117 	struct skb_shared_info *shinfo = skb_shinfo(skb);
1118 
1119 	if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1120 	    !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1121 		struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1122 		u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1123 
1124 		shinfo->tx_flags &= ~tsflags;
1125 		shinfo2->tx_flags |= tsflags;
1126 		swap(shinfo->tskey, shinfo2->tskey);
1127 	}
1128 }
1129 
1130 /* Function to create two new TCP segments.  Shrinks the given segment
1131  * to the specified size and appends a new segment with the rest of the
1132  * packet to the list.  This won't be called frequently, I hope.
1133  * Remember, these are still headerless SKBs at this point.
1134  */
1135 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1136 		 unsigned int mss_now, gfp_t gfp)
1137 {
1138 	struct tcp_sock *tp = tcp_sk(sk);
1139 	struct sk_buff *buff;
1140 	int nsize, old_factor;
1141 	int nlen;
1142 	u8 flags;
1143 
1144 	if (WARN_ON(len > skb->len))
1145 		return -EINVAL;
1146 
1147 	nsize = skb_headlen(skb) - len;
1148 	if (nsize < 0)
1149 		nsize = 0;
1150 
1151 	if (skb_unclone(skb, gfp))
1152 		return -ENOMEM;
1153 
1154 	/* Get a new skb... force flag on. */
1155 	buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1156 	if (!buff)
1157 		return -ENOMEM; /* We'll just try again later. */
1158 
1159 	sk->sk_wmem_queued += buff->truesize;
1160 	sk_mem_charge(sk, buff->truesize);
1161 	nlen = skb->len - len - nsize;
1162 	buff->truesize += nlen;
1163 	skb->truesize -= nlen;
1164 
1165 	/* Correct the sequence numbers. */
1166 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1167 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1168 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1169 
1170 	/* PSH and FIN should only be set in the second packet. */
1171 	flags = TCP_SKB_CB(skb)->tcp_flags;
1172 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1173 	TCP_SKB_CB(buff)->tcp_flags = flags;
1174 	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1175 
1176 	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1177 		/* Copy and checksum data tail into the new buffer. */
1178 		buff->csum = csum_partial_copy_nocheck(skb->data + len,
1179 						       skb_put(buff, nsize),
1180 						       nsize, 0);
1181 
1182 		skb_trim(skb, len);
1183 
1184 		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1185 	} else {
1186 		skb->ip_summed = CHECKSUM_PARTIAL;
1187 		skb_split(skb, buff, len);
1188 	}
1189 
1190 	buff->ip_summed = skb->ip_summed;
1191 
1192 	buff->tstamp = skb->tstamp;
1193 	tcp_fragment_tstamp(skb, buff);
1194 
1195 	old_factor = tcp_skb_pcount(skb);
1196 
1197 	/* Fix up tso_factor for both original and new SKB.  */
1198 	tcp_set_skb_tso_segs(skb, mss_now);
1199 	tcp_set_skb_tso_segs(buff, mss_now);
1200 
1201 	/* If this packet has been sent out already, we must
1202 	 * adjust the various packet counters.
1203 	 */
1204 	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1205 		int diff = old_factor - tcp_skb_pcount(skb) -
1206 			tcp_skb_pcount(buff);
1207 
1208 		if (diff)
1209 			tcp_adjust_pcount(sk, skb, diff);
1210 	}
1211 
1212 	/* Link BUFF into the send queue. */
1213 	__skb_header_release(buff);
1214 	tcp_insert_write_queue_after(skb, buff, sk);
1215 
1216 	return 0;
1217 }
1218 
1219 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1220  * eventually). The difference is that pulled data not copied, but
1221  * immediately discarded.
1222  */
1223 static void __pskb_trim_head(struct sk_buff *skb, int len)
1224 {
1225 	struct skb_shared_info *shinfo;
1226 	int i, k, eat;
1227 
1228 	eat = min_t(int, len, skb_headlen(skb));
1229 	if (eat) {
1230 		__skb_pull(skb, eat);
1231 		len -= eat;
1232 		if (!len)
1233 			return;
1234 	}
1235 	eat = len;
1236 	k = 0;
1237 	shinfo = skb_shinfo(skb);
1238 	for (i = 0; i < shinfo->nr_frags; i++) {
1239 		int size = skb_frag_size(&shinfo->frags[i]);
1240 
1241 		if (size <= eat) {
1242 			skb_frag_unref(skb, i);
1243 			eat -= size;
1244 		} else {
1245 			shinfo->frags[k] = shinfo->frags[i];
1246 			if (eat) {
1247 				shinfo->frags[k].page_offset += eat;
1248 				skb_frag_size_sub(&shinfo->frags[k], eat);
1249 				eat = 0;
1250 			}
1251 			k++;
1252 		}
1253 	}
1254 	shinfo->nr_frags = k;
1255 
1256 	skb_reset_tail_pointer(skb);
1257 	skb->data_len -= len;
1258 	skb->len = skb->data_len;
1259 }
1260 
1261 /* Remove acked data from a packet in the transmit queue. */
1262 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1263 {
1264 	if (skb_unclone(skb, GFP_ATOMIC))
1265 		return -ENOMEM;
1266 
1267 	__pskb_trim_head(skb, len);
1268 
1269 	TCP_SKB_CB(skb)->seq += len;
1270 	skb->ip_summed = CHECKSUM_PARTIAL;
1271 
1272 	skb->truesize	     -= len;
1273 	sk->sk_wmem_queued   -= len;
1274 	sk_mem_uncharge(sk, len);
1275 	sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1276 
1277 	/* Any change of skb->len requires recalculation of tso factor. */
1278 	if (tcp_skb_pcount(skb) > 1)
1279 		tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1280 
1281 	return 0;
1282 }
1283 
1284 /* Calculate MSS not accounting any TCP options.  */
1285 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1286 {
1287 	const struct tcp_sock *tp = tcp_sk(sk);
1288 	const struct inet_connection_sock *icsk = inet_csk(sk);
1289 	int mss_now;
1290 
1291 	/* Calculate base mss without TCP options:
1292 	   It is MMS_S - sizeof(tcphdr) of rfc1122
1293 	 */
1294 	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1295 
1296 	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1297 	if (icsk->icsk_af_ops->net_frag_header_len) {
1298 		const struct dst_entry *dst = __sk_dst_get(sk);
1299 
1300 		if (dst && dst_allfrag(dst))
1301 			mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1302 	}
1303 
1304 	/* Clamp it (mss_clamp does not include tcp options) */
1305 	if (mss_now > tp->rx_opt.mss_clamp)
1306 		mss_now = tp->rx_opt.mss_clamp;
1307 
1308 	/* Now subtract optional transport overhead */
1309 	mss_now -= icsk->icsk_ext_hdr_len;
1310 
1311 	/* Then reserve room for full set of TCP options and 8 bytes of data */
1312 	if (mss_now < 48)
1313 		mss_now = 48;
1314 	return mss_now;
1315 }
1316 
1317 /* Calculate MSS. Not accounting for SACKs here.  */
1318 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1319 {
1320 	/* Subtract TCP options size, not including SACKs */
1321 	return __tcp_mtu_to_mss(sk, pmtu) -
1322 	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1323 }
1324 
1325 /* Inverse of above */
1326 int tcp_mss_to_mtu(struct sock *sk, int mss)
1327 {
1328 	const struct tcp_sock *tp = tcp_sk(sk);
1329 	const struct inet_connection_sock *icsk = inet_csk(sk);
1330 	int mtu;
1331 
1332 	mtu = mss +
1333 	      tp->tcp_header_len +
1334 	      icsk->icsk_ext_hdr_len +
1335 	      icsk->icsk_af_ops->net_header_len;
1336 
1337 	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1338 	if (icsk->icsk_af_ops->net_frag_header_len) {
1339 		const struct dst_entry *dst = __sk_dst_get(sk);
1340 
1341 		if (dst && dst_allfrag(dst))
1342 			mtu += icsk->icsk_af_ops->net_frag_header_len;
1343 	}
1344 	return mtu;
1345 }
1346 
1347 /* MTU probing init per socket */
1348 void tcp_mtup_init(struct sock *sk)
1349 {
1350 	struct tcp_sock *tp = tcp_sk(sk);
1351 	struct inet_connection_sock *icsk = inet_csk(sk);
1352 	struct net *net = sock_net(sk);
1353 
1354 	icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1355 	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1356 			       icsk->icsk_af_ops->net_header_len;
1357 	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1358 	icsk->icsk_mtup.probe_size = 0;
1359 	if (icsk->icsk_mtup.enabled)
1360 		icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1361 }
1362 EXPORT_SYMBOL(tcp_mtup_init);
1363 
1364 /* This function synchronize snd mss to current pmtu/exthdr set.
1365 
1366    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1367    for TCP options, but includes only bare TCP header.
1368 
1369    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1370    It is minimum of user_mss and mss received with SYN.
1371    It also does not include TCP options.
1372 
1373    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1374 
1375    tp->mss_cache is current effective sending mss, including
1376    all tcp options except for SACKs. It is evaluated,
1377    taking into account current pmtu, but never exceeds
1378    tp->rx_opt.mss_clamp.
1379 
1380    NOTE1. rfc1122 clearly states that advertised MSS
1381    DOES NOT include either tcp or ip options.
1382 
1383    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1384    are READ ONLY outside this function.		--ANK (980731)
1385  */
1386 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1387 {
1388 	struct tcp_sock *tp = tcp_sk(sk);
1389 	struct inet_connection_sock *icsk = inet_csk(sk);
1390 	int mss_now;
1391 
1392 	if (icsk->icsk_mtup.search_high > pmtu)
1393 		icsk->icsk_mtup.search_high = pmtu;
1394 
1395 	mss_now = tcp_mtu_to_mss(sk, pmtu);
1396 	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1397 
1398 	/* And store cached results */
1399 	icsk->icsk_pmtu_cookie = pmtu;
1400 	if (icsk->icsk_mtup.enabled)
1401 		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1402 	tp->mss_cache = mss_now;
1403 
1404 	return mss_now;
1405 }
1406 EXPORT_SYMBOL(tcp_sync_mss);
1407 
1408 /* Compute the current effective MSS, taking SACKs and IP options,
1409  * and even PMTU discovery events into account.
1410  */
1411 unsigned int tcp_current_mss(struct sock *sk)
1412 {
1413 	const struct tcp_sock *tp = tcp_sk(sk);
1414 	const struct dst_entry *dst = __sk_dst_get(sk);
1415 	u32 mss_now;
1416 	unsigned int header_len;
1417 	struct tcp_out_options opts;
1418 	struct tcp_md5sig_key *md5;
1419 
1420 	mss_now = tp->mss_cache;
1421 
1422 	if (dst) {
1423 		u32 mtu = dst_mtu(dst);
1424 		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1425 			mss_now = tcp_sync_mss(sk, mtu);
1426 	}
1427 
1428 	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1429 		     sizeof(struct tcphdr);
1430 	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1431 	 * some common options. If this is an odd packet (because we have SACK
1432 	 * blocks etc) then our calculated header_len will be different, and
1433 	 * we have to adjust mss_now correspondingly */
1434 	if (header_len != tp->tcp_header_len) {
1435 		int delta = (int) header_len - tp->tcp_header_len;
1436 		mss_now -= delta;
1437 	}
1438 
1439 	return mss_now;
1440 }
1441 
1442 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1443  * As additional protections, we do not touch cwnd in retransmission phases,
1444  * and if application hit its sndbuf limit recently.
1445  */
1446 static void tcp_cwnd_application_limited(struct sock *sk)
1447 {
1448 	struct tcp_sock *tp = tcp_sk(sk);
1449 
1450 	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1451 	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1452 		/* Limited by application or receiver window. */
1453 		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1454 		u32 win_used = max(tp->snd_cwnd_used, init_win);
1455 		if (win_used < tp->snd_cwnd) {
1456 			tp->snd_ssthresh = tcp_current_ssthresh(sk);
1457 			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1458 		}
1459 		tp->snd_cwnd_used = 0;
1460 	}
1461 	tp->snd_cwnd_stamp = tcp_time_stamp;
1462 }
1463 
1464 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1465 {
1466 	struct tcp_sock *tp = tcp_sk(sk);
1467 
1468 	/* Track the maximum number of outstanding packets in each
1469 	 * window, and remember whether we were cwnd-limited then.
1470 	 */
1471 	if (!before(tp->snd_una, tp->max_packets_seq) ||
1472 	    tp->packets_out > tp->max_packets_out) {
1473 		tp->max_packets_out = tp->packets_out;
1474 		tp->max_packets_seq = tp->snd_nxt;
1475 		tp->is_cwnd_limited = is_cwnd_limited;
1476 	}
1477 
1478 	if (tcp_is_cwnd_limited(sk)) {
1479 		/* Network is feed fully. */
1480 		tp->snd_cwnd_used = 0;
1481 		tp->snd_cwnd_stamp = tcp_time_stamp;
1482 	} else {
1483 		/* Network starves. */
1484 		if (tp->packets_out > tp->snd_cwnd_used)
1485 			tp->snd_cwnd_used = tp->packets_out;
1486 
1487 		if (sysctl_tcp_slow_start_after_idle &&
1488 		    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1489 			tcp_cwnd_application_limited(sk);
1490 	}
1491 }
1492 
1493 /* Minshall's variant of the Nagle send check. */
1494 static bool tcp_minshall_check(const struct tcp_sock *tp)
1495 {
1496 	return after(tp->snd_sml, tp->snd_una) &&
1497 		!after(tp->snd_sml, tp->snd_nxt);
1498 }
1499 
1500 /* Update snd_sml if this skb is under mss
1501  * Note that a TSO packet might end with a sub-mss segment
1502  * The test is really :
1503  * if ((skb->len % mss) != 0)
1504  *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1505  * But we can avoid doing the divide again given we already have
1506  *  skb_pcount = skb->len / mss_now
1507  */
1508 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1509 				const struct sk_buff *skb)
1510 {
1511 	if (skb->len < tcp_skb_pcount(skb) * mss_now)
1512 		tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1513 }
1514 
1515 /* Return false, if packet can be sent now without violation Nagle's rules:
1516  * 1. It is full sized. (provided by caller in %partial bool)
1517  * 2. Or it contains FIN. (already checked by caller)
1518  * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1519  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1520  *    With Minshall's modification: all sent small packets are ACKed.
1521  */
1522 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1523 			    int nonagle)
1524 {
1525 	return partial &&
1526 		((nonagle & TCP_NAGLE_CORK) ||
1527 		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1528 }
1529 
1530 /* Return how many segs we'd like on a TSO packet,
1531  * to send one TSO packet per ms
1532  */
1533 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1534 {
1535 	u32 bytes, segs;
1536 
1537 	bytes = min(sk->sk_pacing_rate >> 10,
1538 		    sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1539 
1540 	/* Goal is to send at least one packet per ms,
1541 	 * not one big TSO packet every 100 ms.
1542 	 * This preserves ACK clocking and is consistent
1543 	 * with tcp_tso_should_defer() heuristic.
1544 	 */
1545 	segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1546 
1547 	return min_t(u32, segs, sk->sk_gso_max_segs);
1548 }
1549 
1550 /* Returns the portion of skb which can be sent right away */
1551 static unsigned int tcp_mss_split_point(const struct sock *sk,
1552 					const struct sk_buff *skb,
1553 					unsigned int mss_now,
1554 					unsigned int max_segs,
1555 					int nonagle)
1556 {
1557 	const struct tcp_sock *tp = tcp_sk(sk);
1558 	u32 partial, needed, window, max_len;
1559 
1560 	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1561 	max_len = mss_now * max_segs;
1562 
1563 	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1564 		return max_len;
1565 
1566 	needed = min(skb->len, window);
1567 
1568 	if (max_len <= needed)
1569 		return max_len;
1570 
1571 	partial = needed % mss_now;
1572 	/* If last segment is not a full MSS, check if Nagle rules allow us
1573 	 * to include this last segment in this skb.
1574 	 * Otherwise, we'll split the skb at last MSS boundary
1575 	 */
1576 	if (tcp_nagle_check(partial != 0, tp, nonagle))
1577 		return needed - partial;
1578 
1579 	return needed;
1580 }
1581 
1582 /* Can at least one segment of SKB be sent right now, according to the
1583  * congestion window rules?  If so, return how many segments are allowed.
1584  */
1585 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1586 					 const struct sk_buff *skb)
1587 {
1588 	u32 in_flight, cwnd, halfcwnd;
1589 
1590 	/* Don't be strict about the congestion window for the final FIN.  */
1591 	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1592 	    tcp_skb_pcount(skb) == 1)
1593 		return 1;
1594 
1595 	in_flight = tcp_packets_in_flight(tp);
1596 	cwnd = tp->snd_cwnd;
1597 	if (in_flight >= cwnd)
1598 		return 0;
1599 
1600 	/* For better scheduling, ensure we have at least
1601 	 * 2 GSO packets in flight.
1602 	 */
1603 	halfcwnd = max(cwnd >> 1, 1U);
1604 	return min(halfcwnd, cwnd - in_flight);
1605 }
1606 
1607 /* Initialize TSO state of a skb.
1608  * This must be invoked the first time we consider transmitting
1609  * SKB onto the wire.
1610  */
1611 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1612 {
1613 	int tso_segs = tcp_skb_pcount(skb);
1614 
1615 	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1616 		tcp_set_skb_tso_segs(skb, mss_now);
1617 		tso_segs = tcp_skb_pcount(skb);
1618 	}
1619 	return tso_segs;
1620 }
1621 
1622 
1623 /* Return true if the Nagle test allows this packet to be
1624  * sent now.
1625  */
1626 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1627 				  unsigned int cur_mss, int nonagle)
1628 {
1629 	/* Nagle rule does not apply to frames, which sit in the middle of the
1630 	 * write_queue (they have no chances to get new data).
1631 	 *
1632 	 * This is implemented in the callers, where they modify the 'nonagle'
1633 	 * argument based upon the location of SKB in the send queue.
1634 	 */
1635 	if (nonagle & TCP_NAGLE_PUSH)
1636 		return true;
1637 
1638 	/* Don't use the nagle rule for urgent data (or for the final FIN). */
1639 	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1640 		return true;
1641 
1642 	if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1643 		return true;
1644 
1645 	return false;
1646 }
1647 
1648 /* Does at least the first segment of SKB fit into the send window? */
1649 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1650 			     const struct sk_buff *skb,
1651 			     unsigned int cur_mss)
1652 {
1653 	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1654 
1655 	if (skb->len > cur_mss)
1656 		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1657 
1658 	return !after(end_seq, tcp_wnd_end(tp));
1659 }
1660 
1661 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1662  * should be put on the wire right now.  If so, it returns the number of
1663  * packets allowed by the congestion window.
1664  */
1665 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1666 				 unsigned int cur_mss, int nonagle)
1667 {
1668 	const struct tcp_sock *tp = tcp_sk(sk);
1669 	unsigned int cwnd_quota;
1670 
1671 	tcp_init_tso_segs(skb, cur_mss);
1672 
1673 	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1674 		return 0;
1675 
1676 	cwnd_quota = tcp_cwnd_test(tp, skb);
1677 	if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1678 		cwnd_quota = 0;
1679 
1680 	return cwnd_quota;
1681 }
1682 
1683 /* Test if sending is allowed right now. */
1684 bool tcp_may_send_now(struct sock *sk)
1685 {
1686 	const struct tcp_sock *tp = tcp_sk(sk);
1687 	struct sk_buff *skb = tcp_send_head(sk);
1688 
1689 	return skb &&
1690 		tcp_snd_test(sk, skb, tcp_current_mss(sk),
1691 			     (tcp_skb_is_last(sk, skb) ?
1692 			      tp->nonagle : TCP_NAGLE_PUSH));
1693 }
1694 
1695 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1696  * which is put after SKB on the list.  It is very much like
1697  * tcp_fragment() except that it may make several kinds of assumptions
1698  * in order to speed up the splitting operation.  In particular, we
1699  * know that all the data is in scatter-gather pages, and that the
1700  * packet has never been sent out before (and thus is not cloned).
1701  */
1702 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1703 			unsigned int mss_now, gfp_t gfp)
1704 {
1705 	struct sk_buff *buff;
1706 	int nlen = skb->len - len;
1707 	u8 flags;
1708 
1709 	/* All of a TSO frame must be composed of paged data.  */
1710 	if (skb->len != skb->data_len)
1711 		return tcp_fragment(sk, skb, len, mss_now, gfp);
1712 
1713 	buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1714 	if (unlikely(!buff))
1715 		return -ENOMEM;
1716 
1717 	sk->sk_wmem_queued += buff->truesize;
1718 	sk_mem_charge(sk, buff->truesize);
1719 	buff->truesize += nlen;
1720 	skb->truesize -= nlen;
1721 
1722 	/* Correct the sequence numbers. */
1723 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1724 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1725 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1726 
1727 	/* PSH and FIN should only be set in the second packet. */
1728 	flags = TCP_SKB_CB(skb)->tcp_flags;
1729 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1730 	TCP_SKB_CB(buff)->tcp_flags = flags;
1731 
1732 	/* This packet was never sent out yet, so no SACK bits. */
1733 	TCP_SKB_CB(buff)->sacked = 0;
1734 
1735 	buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1736 	skb_split(skb, buff, len);
1737 	tcp_fragment_tstamp(skb, buff);
1738 
1739 	/* Fix up tso_factor for both original and new SKB.  */
1740 	tcp_set_skb_tso_segs(skb, mss_now);
1741 	tcp_set_skb_tso_segs(buff, mss_now);
1742 
1743 	/* Link BUFF into the send queue. */
1744 	__skb_header_release(buff);
1745 	tcp_insert_write_queue_after(skb, buff, sk);
1746 
1747 	return 0;
1748 }
1749 
1750 /* Try to defer sending, if possible, in order to minimize the amount
1751  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1752  *
1753  * This algorithm is from John Heffner.
1754  */
1755 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1756 				 bool *is_cwnd_limited, u32 max_segs)
1757 {
1758 	const struct inet_connection_sock *icsk = inet_csk(sk);
1759 	u32 age, send_win, cong_win, limit, in_flight;
1760 	struct tcp_sock *tp = tcp_sk(sk);
1761 	struct skb_mstamp now;
1762 	struct sk_buff *head;
1763 	int win_divisor;
1764 
1765 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1766 		goto send_now;
1767 
1768 	if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1769 		goto send_now;
1770 
1771 	/* Avoid bursty behavior by allowing defer
1772 	 * only if the last write was recent.
1773 	 */
1774 	if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1775 		goto send_now;
1776 
1777 	in_flight = tcp_packets_in_flight(tp);
1778 
1779 	BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1780 
1781 	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1782 
1783 	/* From in_flight test above, we know that cwnd > in_flight.  */
1784 	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1785 
1786 	limit = min(send_win, cong_win);
1787 
1788 	/* If a full-sized TSO skb can be sent, do it. */
1789 	if (limit >= max_segs * tp->mss_cache)
1790 		goto send_now;
1791 
1792 	/* Middle in queue won't get any more data, full sendable already? */
1793 	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1794 		goto send_now;
1795 
1796 	win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1797 	if (win_divisor) {
1798 		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1799 
1800 		/* If at least some fraction of a window is available,
1801 		 * just use it.
1802 		 */
1803 		chunk /= win_divisor;
1804 		if (limit >= chunk)
1805 			goto send_now;
1806 	} else {
1807 		/* Different approach, try not to defer past a single
1808 		 * ACK.  Receiver should ACK every other full sized
1809 		 * frame, so if we have space for more than 3 frames
1810 		 * then send now.
1811 		 */
1812 		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1813 			goto send_now;
1814 	}
1815 
1816 	head = tcp_write_queue_head(sk);
1817 	skb_mstamp_get(&now);
1818 	age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1819 	/* If next ACK is likely to come too late (half srtt), do not defer */
1820 	if (age < (tp->srtt_us >> 4))
1821 		goto send_now;
1822 
1823 	/* Ok, it looks like it is advisable to defer. */
1824 
1825 	if (cong_win < send_win && cong_win <= skb->len)
1826 		*is_cwnd_limited = true;
1827 
1828 	return true;
1829 
1830 send_now:
1831 	return false;
1832 }
1833 
1834 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1835 {
1836 	struct inet_connection_sock *icsk = inet_csk(sk);
1837 	struct tcp_sock *tp = tcp_sk(sk);
1838 	struct net *net = sock_net(sk);
1839 	u32 interval;
1840 	s32 delta;
1841 
1842 	interval = net->ipv4.sysctl_tcp_probe_interval;
1843 	delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1844 	if (unlikely(delta >= interval * HZ)) {
1845 		int mss = tcp_current_mss(sk);
1846 
1847 		/* Update current search range */
1848 		icsk->icsk_mtup.probe_size = 0;
1849 		icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1850 			sizeof(struct tcphdr) +
1851 			icsk->icsk_af_ops->net_header_len;
1852 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1853 
1854 		/* Update probe time stamp */
1855 		icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1856 	}
1857 }
1858 
1859 /* Create a new MTU probe if we are ready.
1860  * MTU probe is regularly attempting to increase the path MTU by
1861  * deliberately sending larger packets.  This discovers routing
1862  * changes resulting in larger path MTUs.
1863  *
1864  * Returns 0 if we should wait to probe (no cwnd available),
1865  *         1 if a probe was sent,
1866  *         -1 otherwise
1867  */
1868 static int tcp_mtu_probe(struct sock *sk)
1869 {
1870 	struct tcp_sock *tp = tcp_sk(sk);
1871 	struct inet_connection_sock *icsk = inet_csk(sk);
1872 	struct sk_buff *skb, *nskb, *next;
1873 	struct net *net = sock_net(sk);
1874 	int len;
1875 	int probe_size;
1876 	int size_needed;
1877 	int copy;
1878 	int mss_now;
1879 	int interval;
1880 
1881 	/* Not currently probing/verifying,
1882 	 * not in recovery,
1883 	 * have enough cwnd, and
1884 	 * not SACKing (the variable headers throw things off) */
1885 	if (!icsk->icsk_mtup.enabled ||
1886 	    icsk->icsk_mtup.probe_size ||
1887 	    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1888 	    tp->snd_cwnd < 11 ||
1889 	    tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1890 		return -1;
1891 
1892 	/* Use binary search for probe_size between tcp_mss_base,
1893 	 * and current mss_clamp. if (search_high - search_low)
1894 	 * smaller than a threshold, backoff from probing.
1895 	 */
1896 	mss_now = tcp_current_mss(sk);
1897 	probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1898 				    icsk->icsk_mtup.search_low) >> 1);
1899 	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1900 	interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1901 	/* When misfortune happens, we are reprobing actively,
1902 	 * and then reprobe timer has expired. We stick with current
1903 	 * probing process by not resetting search range to its orignal.
1904 	 */
1905 	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1906 		interval < net->ipv4.sysctl_tcp_probe_threshold) {
1907 		/* Check whether enough time has elaplased for
1908 		 * another round of probing.
1909 		 */
1910 		tcp_mtu_check_reprobe(sk);
1911 		return -1;
1912 	}
1913 
1914 	/* Have enough data in the send queue to probe? */
1915 	if (tp->write_seq - tp->snd_nxt < size_needed)
1916 		return -1;
1917 
1918 	if (tp->snd_wnd < size_needed)
1919 		return -1;
1920 	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1921 		return 0;
1922 
1923 	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
1924 	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1925 		if (!tcp_packets_in_flight(tp))
1926 			return -1;
1927 		else
1928 			return 0;
1929 	}
1930 
1931 	/* We're allowed to probe.  Build it now. */
1932 	nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
1933 	if (!nskb)
1934 		return -1;
1935 	sk->sk_wmem_queued += nskb->truesize;
1936 	sk_mem_charge(sk, nskb->truesize);
1937 
1938 	skb = tcp_send_head(sk);
1939 
1940 	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1941 	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1942 	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1943 	TCP_SKB_CB(nskb)->sacked = 0;
1944 	nskb->csum = 0;
1945 	nskb->ip_summed = skb->ip_summed;
1946 
1947 	tcp_insert_write_queue_before(nskb, skb, sk);
1948 
1949 	len = 0;
1950 	tcp_for_write_queue_from_safe(skb, next, sk) {
1951 		copy = min_t(int, skb->len, probe_size - len);
1952 		if (nskb->ip_summed)
1953 			skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1954 		else
1955 			nskb->csum = skb_copy_and_csum_bits(skb, 0,
1956 							    skb_put(nskb, copy),
1957 							    copy, nskb->csum);
1958 
1959 		if (skb->len <= copy) {
1960 			/* We've eaten all the data from this skb.
1961 			 * Throw it away. */
1962 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1963 			tcp_unlink_write_queue(skb, sk);
1964 			sk_wmem_free_skb(sk, skb);
1965 		} else {
1966 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1967 						   ~(TCPHDR_FIN|TCPHDR_PSH);
1968 			if (!skb_shinfo(skb)->nr_frags) {
1969 				skb_pull(skb, copy);
1970 				if (skb->ip_summed != CHECKSUM_PARTIAL)
1971 					skb->csum = csum_partial(skb->data,
1972 								 skb->len, 0);
1973 			} else {
1974 				__pskb_trim_head(skb, copy);
1975 				tcp_set_skb_tso_segs(skb, mss_now);
1976 			}
1977 			TCP_SKB_CB(skb)->seq += copy;
1978 		}
1979 
1980 		len += copy;
1981 
1982 		if (len >= probe_size)
1983 			break;
1984 	}
1985 	tcp_init_tso_segs(nskb, nskb->len);
1986 
1987 	/* We're ready to send.  If this fails, the probe will
1988 	 * be resegmented into mss-sized pieces by tcp_write_xmit().
1989 	 */
1990 	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1991 		/* Decrement cwnd here because we are sending
1992 		 * effectively two packets. */
1993 		tp->snd_cwnd--;
1994 		tcp_event_new_data_sent(sk, nskb);
1995 
1996 		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1997 		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1998 		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1999 
2000 		return 1;
2001 	}
2002 
2003 	return -1;
2004 }
2005 
2006 /* This routine writes packets to the network.  It advances the
2007  * send_head.  This happens as incoming acks open up the remote
2008  * window for us.
2009  *
2010  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2011  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2012  * account rare use of URG, this is not a big flaw.
2013  *
2014  * Send at most one packet when push_one > 0. Temporarily ignore
2015  * cwnd limit to force at most one packet out when push_one == 2.
2016 
2017  * Returns true, if no segments are in flight and we have queued segments,
2018  * but cannot send anything now because of SWS or another problem.
2019  */
2020 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2021 			   int push_one, gfp_t gfp)
2022 {
2023 	struct tcp_sock *tp = tcp_sk(sk);
2024 	struct sk_buff *skb;
2025 	unsigned int tso_segs, sent_pkts;
2026 	int cwnd_quota;
2027 	int result;
2028 	bool is_cwnd_limited = false;
2029 	u32 max_segs;
2030 
2031 	sent_pkts = 0;
2032 
2033 	if (!push_one) {
2034 		/* Do MTU probing. */
2035 		result = tcp_mtu_probe(sk);
2036 		if (!result) {
2037 			return false;
2038 		} else if (result > 0) {
2039 			sent_pkts = 1;
2040 		}
2041 	}
2042 
2043 	max_segs = tcp_tso_autosize(sk, mss_now);
2044 	while ((skb = tcp_send_head(sk))) {
2045 		unsigned int limit;
2046 
2047 		tso_segs = tcp_init_tso_segs(skb, mss_now);
2048 		BUG_ON(!tso_segs);
2049 
2050 		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2051 			/* "skb_mstamp" is used as a start point for the retransmit timer */
2052 			skb_mstamp_get(&skb->skb_mstamp);
2053 			goto repair; /* Skip network transmission */
2054 		}
2055 
2056 		cwnd_quota = tcp_cwnd_test(tp, skb);
2057 		if (!cwnd_quota) {
2058 			if (push_one == 2)
2059 				/* Force out a loss probe pkt. */
2060 				cwnd_quota = 1;
2061 			else
2062 				break;
2063 		}
2064 
2065 		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2066 			break;
2067 
2068 		if (tso_segs == 1) {
2069 			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2070 						     (tcp_skb_is_last(sk, skb) ?
2071 						      nonagle : TCP_NAGLE_PUSH))))
2072 				break;
2073 		} else {
2074 			if (!push_one &&
2075 			    tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2076 						 max_segs))
2077 				break;
2078 		}
2079 
2080 		limit = mss_now;
2081 		if (tso_segs > 1 && !tcp_urg_mode(tp))
2082 			limit = tcp_mss_split_point(sk, skb, mss_now,
2083 						    min_t(unsigned int,
2084 							  cwnd_quota,
2085 							  max_segs),
2086 						    nonagle);
2087 
2088 		if (skb->len > limit &&
2089 		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2090 			break;
2091 
2092 		/* TCP Small Queues :
2093 		 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2094 		 * This allows for :
2095 		 *  - better RTT estimation and ACK scheduling
2096 		 *  - faster recovery
2097 		 *  - high rates
2098 		 * Alas, some drivers / subsystems require a fair amount
2099 		 * of queued bytes to ensure line rate.
2100 		 * One example is wifi aggregation (802.11 AMPDU)
2101 		 */
2102 		limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2103 		limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2104 
2105 		if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2106 			set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2107 			/* It is possible TX completion already happened
2108 			 * before we set TSQ_THROTTLED, so we must
2109 			 * test again the condition.
2110 			 */
2111 			smp_mb__after_atomic();
2112 			if (atomic_read(&sk->sk_wmem_alloc) > limit)
2113 				break;
2114 		}
2115 
2116 		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2117 			break;
2118 
2119 repair:
2120 		/* Advance the send_head.  This one is sent out.
2121 		 * This call will increment packets_out.
2122 		 */
2123 		tcp_event_new_data_sent(sk, skb);
2124 
2125 		tcp_minshall_update(tp, mss_now, skb);
2126 		sent_pkts += tcp_skb_pcount(skb);
2127 
2128 		if (push_one)
2129 			break;
2130 	}
2131 
2132 	if (likely(sent_pkts)) {
2133 		if (tcp_in_cwnd_reduction(sk))
2134 			tp->prr_out += sent_pkts;
2135 
2136 		/* Send one loss probe per tail loss episode. */
2137 		if (push_one != 2)
2138 			tcp_schedule_loss_probe(sk);
2139 		is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2140 		tcp_cwnd_validate(sk, is_cwnd_limited);
2141 		return false;
2142 	}
2143 	return !tp->packets_out && tcp_send_head(sk);
2144 }
2145 
2146 bool tcp_schedule_loss_probe(struct sock *sk)
2147 {
2148 	struct inet_connection_sock *icsk = inet_csk(sk);
2149 	struct tcp_sock *tp = tcp_sk(sk);
2150 	u32 timeout, tlp_time_stamp, rto_time_stamp;
2151 	u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2152 
2153 	if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2154 		return false;
2155 	/* No consecutive loss probes. */
2156 	if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2157 		tcp_rearm_rto(sk);
2158 		return false;
2159 	}
2160 	/* Don't do any loss probe on a Fast Open connection before 3WHS
2161 	 * finishes.
2162 	 */
2163 	if (tp->fastopen_rsk)
2164 		return false;
2165 
2166 	/* TLP is only scheduled when next timer event is RTO. */
2167 	if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2168 		return false;
2169 
2170 	/* Schedule a loss probe in 2*RTT for SACK capable connections
2171 	 * in Open state, that are either limited by cwnd or application.
2172 	 */
2173 	if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
2174 	    !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2175 		return false;
2176 
2177 	if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2178 	     tcp_send_head(sk))
2179 		return false;
2180 
2181 	/* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2182 	 * for delayed ack when there's one outstanding packet. If no RTT
2183 	 * sample is available then probe after TCP_TIMEOUT_INIT.
2184 	 */
2185 	timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2186 	if (tp->packets_out == 1)
2187 		timeout = max_t(u32, timeout,
2188 				(rtt + (rtt >> 1) + TCP_DELACK_MAX));
2189 	timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2190 
2191 	/* If RTO is shorter, just schedule TLP in its place. */
2192 	tlp_time_stamp = tcp_time_stamp + timeout;
2193 	rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2194 	if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2195 		s32 delta = rto_time_stamp - tcp_time_stamp;
2196 		if (delta > 0)
2197 			timeout = delta;
2198 	}
2199 
2200 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2201 				  TCP_RTO_MAX);
2202 	return true;
2203 }
2204 
2205 /* Thanks to skb fast clones, we can detect if a prior transmit of
2206  * a packet is still in a qdisc or driver queue.
2207  * In this case, there is very little point doing a retransmit !
2208  * Note: This is called from BH context only.
2209  */
2210 static bool skb_still_in_host_queue(const struct sock *sk,
2211 				    const struct sk_buff *skb)
2212 {
2213 	if (unlikely(skb_fclone_busy(sk, skb))) {
2214 		NET_INC_STATS_BH(sock_net(sk),
2215 				 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2216 		return true;
2217 	}
2218 	return false;
2219 }
2220 
2221 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2222  * retransmit the last segment.
2223  */
2224 void tcp_send_loss_probe(struct sock *sk)
2225 {
2226 	struct tcp_sock *tp = tcp_sk(sk);
2227 	struct sk_buff *skb;
2228 	int pcount;
2229 	int mss = tcp_current_mss(sk);
2230 
2231 	skb = tcp_send_head(sk);
2232 	if (skb) {
2233 		if (tcp_snd_wnd_test(tp, skb, mss)) {
2234 			pcount = tp->packets_out;
2235 			tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2236 			if (tp->packets_out > pcount)
2237 				goto probe_sent;
2238 			goto rearm_timer;
2239 		}
2240 		skb = tcp_write_queue_prev(sk, skb);
2241 	} else {
2242 		skb = tcp_write_queue_tail(sk);
2243 	}
2244 
2245 	/* At most one outstanding TLP retransmission. */
2246 	if (tp->tlp_high_seq)
2247 		goto rearm_timer;
2248 
2249 	/* Retransmit last segment. */
2250 	if (WARN_ON(!skb))
2251 		goto rearm_timer;
2252 
2253 	if (skb_still_in_host_queue(sk, skb))
2254 		goto rearm_timer;
2255 
2256 	pcount = tcp_skb_pcount(skb);
2257 	if (WARN_ON(!pcount))
2258 		goto rearm_timer;
2259 
2260 	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2261 		if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2262 					  GFP_ATOMIC)))
2263 			goto rearm_timer;
2264 		skb = tcp_write_queue_next(sk, skb);
2265 	}
2266 
2267 	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2268 		goto rearm_timer;
2269 
2270 	if (__tcp_retransmit_skb(sk, skb))
2271 		goto rearm_timer;
2272 
2273 	/* Record snd_nxt for loss detection. */
2274 	tp->tlp_high_seq = tp->snd_nxt;
2275 
2276 probe_sent:
2277 	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2278 	/* Reset s.t. tcp_rearm_rto will restart timer from now */
2279 	inet_csk(sk)->icsk_pending = 0;
2280 rearm_timer:
2281 	tcp_rearm_rto(sk);
2282 }
2283 
2284 /* Push out any pending frames which were held back due to
2285  * TCP_CORK or attempt at coalescing tiny packets.
2286  * The socket must be locked by the caller.
2287  */
2288 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2289 			       int nonagle)
2290 {
2291 	/* If we are closed, the bytes will have to remain here.
2292 	 * In time closedown will finish, we empty the write queue and
2293 	 * all will be happy.
2294 	 */
2295 	if (unlikely(sk->sk_state == TCP_CLOSE))
2296 		return;
2297 
2298 	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2299 			   sk_gfp_atomic(sk, GFP_ATOMIC)))
2300 		tcp_check_probe_timer(sk);
2301 }
2302 
2303 /* Send _single_ skb sitting at the send head. This function requires
2304  * true push pending frames to setup probe timer etc.
2305  */
2306 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2307 {
2308 	struct sk_buff *skb = tcp_send_head(sk);
2309 
2310 	BUG_ON(!skb || skb->len < mss_now);
2311 
2312 	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2313 }
2314 
2315 /* This function returns the amount that we can raise the
2316  * usable window based on the following constraints
2317  *
2318  * 1. The window can never be shrunk once it is offered (RFC 793)
2319  * 2. We limit memory per socket
2320  *
2321  * RFC 1122:
2322  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2323  *  RECV.NEXT + RCV.WIN fixed until:
2324  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2325  *
2326  * i.e. don't raise the right edge of the window until you can raise
2327  * it at least MSS bytes.
2328  *
2329  * Unfortunately, the recommended algorithm breaks header prediction,
2330  * since header prediction assumes th->window stays fixed.
2331  *
2332  * Strictly speaking, keeping th->window fixed violates the receiver
2333  * side SWS prevention criteria. The problem is that under this rule
2334  * a stream of single byte packets will cause the right side of the
2335  * window to always advance by a single byte.
2336  *
2337  * Of course, if the sender implements sender side SWS prevention
2338  * then this will not be a problem.
2339  *
2340  * BSD seems to make the following compromise:
2341  *
2342  *	If the free space is less than the 1/4 of the maximum
2343  *	space available and the free space is less than 1/2 mss,
2344  *	then set the window to 0.
2345  *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2346  *	Otherwise, just prevent the window from shrinking
2347  *	and from being larger than the largest representable value.
2348  *
2349  * This prevents incremental opening of the window in the regime
2350  * where TCP is limited by the speed of the reader side taking
2351  * data out of the TCP receive queue. It does nothing about
2352  * those cases where the window is constrained on the sender side
2353  * because the pipeline is full.
2354  *
2355  * BSD also seems to "accidentally" limit itself to windows that are a
2356  * multiple of MSS, at least until the free space gets quite small.
2357  * This would appear to be a side effect of the mbuf implementation.
2358  * Combining these two algorithms results in the observed behavior
2359  * of having a fixed window size at almost all times.
2360  *
2361  * Below we obtain similar behavior by forcing the offered window to
2362  * a multiple of the mss when it is feasible to do so.
2363  *
2364  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2365  * Regular options like TIMESTAMP are taken into account.
2366  */
2367 u32 __tcp_select_window(struct sock *sk)
2368 {
2369 	struct inet_connection_sock *icsk = inet_csk(sk);
2370 	struct tcp_sock *tp = tcp_sk(sk);
2371 	/* MSS for the peer's data.  Previous versions used mss_clamp
2372 	 * here.  I don't know if the value based on our guesses
2373 	 * of peer's MSS is better for the performance.  It's more correct
2374 	 * but may be worse for the performance because of rcv_mss
2375 	 * fluctuations.  --SAW  1998/11/1
2376 	 */
2377 	int mss = icsk->icsk_ack.rcv_mss;
2378 	int free_space = tcp_space(sk);
2379 	int allowed_space = tcp_full_space(sk);
2380 	int full_space = min_t(int, tp->window_clamp, allowed_space);
2381 	int window;
2382 
2383 	if (mss > full_space)
2384 		mss = full_space;
2385 
2386 	if (free_space < (full_space >> 1)) {
2387 		icsk->icsk_ack.quick = 0;
2388 
2389 		if (tcp_under_memory_pressure(sk))
2390 			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2391 					       4U * tp->advmss);
2392 
2393 		/* free_space might become our new window, make sure we don't
2394 		 * increase it due to wscale.
2395 		 */
2396 		free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2397 
2398 		/* if free space is less than mss estimate, or is below 1/16th
2399 		 * of the maximum allowed, try to move to zero-window, else
2400 		 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2401 		 * new incoming data is dropped due to memory limits.
2402 		 * With large window, mss test triggers way too late in order
2403 		 * to announce zero window in time before rmem limit kicks in.
2404 		 */
2405 		if (free_space < (allowed_space >> 4) || free_space < mss)
2406 			return 0;
2407 	}
2408 
2409 	if (free_space > tp->rcv_ssthresh)
2410 		free_space = tp->rcv_ssthresh;
2411 
2412 	/* Don't do rounding if we are using window scaling, since the
2413 	 * scaled window will not line up with the MSS boundary anyway.
2414 	 */
2415 	window = tp->rcv_wnd;
2416 	if (tp->rx_opt.rcv_wscale) {
2417 		window = free_space;
2418 
2419 		/* Advertise enough space so that it won't get scaled away.
2420 		 * Import case: prevent zero window announcement if
2421 		 * 1<<rcv_wscale > mss.
2422 		 */
2423 		if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2424 			window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2425 				  << tp->rx_opt.rcv_wscale);
2426 	} else {
2427 		/* Get the largest window that is a nice multiple of mss.
2428 		 * Window clamp already applied above.
2429 		 * If our current window offering is within 1 mss of the
2430 		 * free space we just keep it. This prevents the divide
2431 		 * and multiply from happening most of the time.
2432 		 * We also don't do any window rounding when the free space
2433 		 * is too small.
2434 		 */
2435 		if (window <= free_space - mss || window > free_space)
2436 			window = (free_space / mss) * mss;
2437 		else if (mss == full_space &&
2438 			 free_space > window + (full_space >> 1))
2439 			window = free_space;
2440 	}
2441 
2442 	return window;
2443 }
2444 
2445 /* Collapses two adjacent SKB's during retransmission. */
2446 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2447 {
2448 	struct tcp_sock *tp = tcp_sk(sk);
2449 	struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2450 	int skb_size, next_skb_size;
2451 
2452 	skb_size = skb->len;
2453 	next_skb_size = next_skb->len;
2454 
2455 	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2456 
2457 	tcp_highest_sack_combine(sk, next_skb, skb);
2458 
2459 	tcp_unlink_write_queue(next_skb, sk);
2460 
2461 	skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2462 				  next_skb_size);
2463 
2464 	if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2465 		skb->ip_summed = CHECKSUM_PARTIAL;
2466 
2467 	if (skb->ip_summed != CHECKSUM_PARTIAL)
2468 		skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2469 
2470 	/* Update sequence range on original skb. */
2471 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2472 
2473 	/* Merge over control information. This moves PSH/FIN etc. over */
2474 	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2475 
2476 	/* All done, get rid of second SKB and account for it so
2477 	 * packet counting does not break.
2478 	 */
2479 	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2480 
2481 	/* changed transmit queue under us so clear hints */
2482 	tcp_clear_retrans_hints_partial(tp);
2483 	if (next_skb == tp->retransmit_skb_hint)
2484 		tp->retransmit_skb_hint = skb;
2485 
2486 	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2487 
2488 	sk_wmem_free_skb(sk, next_skb);
2489 }
2490 
2491 /* Check if coalescing SKBs is legal. */
2492 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2493 {
2494 	if (tcp_skb_pcount(skb) > 1)
2495 		return false;
2496 	/* TODO: SACK collapsing could be used to remove this condition */
2497 	if (skb_shinfo(skb)->nr_frags != 0)
2498 		return false;
2499 	if (skb_cloned(skb))
2500 		return false;
2501 	if (skb == tcp_send_head(sk))
2502 		return false;
2503 	/* Some heurestics for collapsing over SACK'd could be invented */
2504 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2505 		return false;
2506 
2507 	return true;
2508 }
2509 
2510 /* Collapse packets in the retransmit queue to make to create
2511  * less packets on the wire. This is only done on retransmission.
2512  */
2513 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2514 				     int space)
2515 {
2516 	struct tcp_sock *tp = tcp_sk(sk);
2517 	struct sk_buff *skb = to, *tmp;
2518 	bool first = true;
2519 
2520 	if (!sysctl_tcp_retrans_collapse)
2521 		return;
2522 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2523 		return;
2524 
2525 	tcp_for_write_queue_from_safe(skb, tmp, sk) {
2526 		if (!tcp_can_collapse(sk, skb))
2527 			break;
2528 
2529 		space -= skb->len;
2530 
2531 		if (first) {
2532 			first = false;
2533 			continue;
2534 		}
2535 
2536 		if (space < 0)
2537 			break;
2538 		/* Punt if not enough space exists in the first SKB for
2539 		 * the data in the second
2540 		 */
2541 		if (skb->len > skb_availroom(to))
2542 			break;
2543 
2544 		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2545 			break;
2546 
2547 		tcp_collapse_retrans(sk, to);
2548 	}
2549 }
2550 
2551 /* This retransmits one SKB.  Policy decisions and retransmit queue
2552  * state updates are done by the caller.  Returns non-zero if an
2553  * error occurred which prevented the send.
2554  */
2555 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2556 {
2557 	struct tcp_sock *tp = tcp_sk(sk);
2558 	struct inet_connection_sock *icsk = inet_csk(sk);
2559 	unsigned int cur_mss;
2560 	int err;
2561 
2562 	/* Inconslusive MTU probe */
2563 	if (icsk->icsk_mtup.probe_size) {
2564 		icsk->icsk_mtup.probe_size = 0;
2565 	}
2566 
2567 	/* Do not sent more than we queued. 1/4 is reserved for possible
2568 	 * copying overhead: fragmentation, tunneling, mangling etc.
2569 	 */
2570 	if (atomic_read(&sk->sk_wmem_alloc) >
2571 	    min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2572 		return -EAGAIN;
2573 
2574 	if (skb_still_in_host_queue(sk, skb))
2575 		return -EBUSY;
2576 
2577 	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2578 		if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2579 			BUG();
2580 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2581 			return -ENOMEM;
2582 	}
2583 
2584 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2585 		return -EHOSTUNREACH; /* Routing failure or similar. */
2586 
2587 	cur_mss = tcp_current_mss(sk);
2588 
2589 	/* If receiver has shrunk his window, and skb is out of
2590 	 * new window, do not retransmit it. The exception is the
2591 	 * case, when window is shrunk to zero. In this case
2592 	 * our retransmit serves as a zero window probe.
2593 	 */
2594 	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2595 	    TCP_SKB_CB(skb)->seq != tp->snd_una)
2596 		return -EAGAIN;
2597 
2598 	if (skb->len > cur_mss) {
2599 		if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
2600 			return -ENOMEM; /* We'll try again later. */
2601 	} else {
2602 		int oldpcount = tcp_skb_pcount(skb);
2603 
2604 		if (unlikely(oldpcount > 1)) {
2605 			if (skb_unclone(skb, GFP_ATOMIC))
2606 				return -ENOMEM;
2607 			tcp_init_tso_segs(skb, cur_mss);
2608 			tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2609 		}
2610 	}
2611 
2612 	/* RFC3168, section 6.1.1.1. ECN fallback */
2613 	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2614 		tcp_ecn_clear_syn(sk, skb);
2615 
2616 	tcp_retrans_try_collapse(sk, skb, cur_mss);
2617 
2618 	/* Make a copy, if the first transmission SKB clone we made
2619 	 * is still in somebody's hands, else make a clone.
2620 	 */
2621 
2622 	/* make sure skb->data is aligned on arches that require it
2623 	 * and check if ack-trimming & collapsing extended the headroom
2624 	 * beyond what csum_start can cover.
2625 	 */
2626 	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2627 		     skb_headroom(skb) >= 0xFFFF)) {
2628 		struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2629 						   GFP_ATOMIC);
2630 		err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2631 			     -ENOBUFS;
2632 	} else {
2633 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2634 	}
2635 
2636 	if (likely(!err)) {
2637 		TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2638 		/* Update global TCP statistics. */
2639 		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2640 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2641 			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2642 		tp->total_retrans++;
2643 	}
2644 	return err;
2645 }
2646 
2647 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2648 {
2649 	struct tcp_sock *tp = tcp_sk(sk);
2650 	int err = __tcp_retransmit_skb(sk, skb);
2651 
2652 	if (err == 0) {
2653 #if FASTRETRANS_DEBUG > 0
2654 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2655 			net_dbg_ratelimited("retrans_out leaked\n");
2656 		}
2657 #endif
2658 		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2659 		tp->retrans_out += tcp_skb_pcount(skb);
2660 
2661 		/* Save stamp of the first retransmit. */
2662 		if (!tp->retrans_stamp)
2663 			tp->retrans_stamp = tcp_skb_timestamp(skb);
2664 
2665 	} else if (err != -EBUSY) {
2666 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2667 	}
2668 
2669 	if (tp->undo_retrans < 0)
2670 		tp->undo_retrans = 0;
2671 	tp->undo_retrans += tcp_skb_pcount(skb);
2672 	return err;
2673 }
2674 
2675 /* Check if we forward retransmits are possible in the current
2676  * window/congestion state.
2677  */
2678 static bool tcp_can_forward_retransmit(struct sock *sk)
2679 {
2680 	const struct inet_connection_sock *icsk = inet_csk(sk);
2681 	const struct tcp_sock *tp = tcp_sk(sk);
2682 
2683 	/* Forward retransmissions are possible only during Recovery. */
2684 	if (icsk->icsk_ca_state != TCP_CA_Recovery)
2685 		return false;
2686 
2687 	/* No forward retransmissions in Reno are possible. */
2688 	if (tcp_is_reno(tp))
2689 		return false;
2690 
2691 	/* Yeah, we have to make difficult choice between forward transmission
2692 	 * and retransmission... Both ways have their merits...
2693 	 *
2694 	 * For now we do not retransmit anything, while we have some new
2695 	 * segments to send. In the other cases, follow rule 3 for
2696 	 * NextSeg() specified in RFC3517.
2697 	 */
2698 
2699 	if (tcp_may_send_now(sk))
2700 		return false;
2701 
2702 	return true;
2703 }
2704 
2705 /* This gets called after a retransmit timeout, and the initially
2706  * retransmitted data is acknowledged.  It tries to continue
2707  * resending the rest of the retransmit queue, until either
2708  * we've sent it all or the congestion window limit is reached.
2709  * If doing SACK, the first ACK which comes back for a timeout
2710  * based retransmit packet might feed us FACK information again.
2711  * If so, we use it to avoid unnecessarily retransmissions.
2712  */
2713 void tcp_xmit_retransmit_queue(struct sock *sk)
2714 {
2715 	const struct inet_connection_sock *icsk = inet_csk(sk);
2716 	struct tcp_sock *tp = tcp_sk(sk);
2717 	struct sk_buff *skb;
2718 	struct sk_buff *hole = NULL;
2719 	u32 last_lost;
2720 	int mib_idx;
2721 	int fwd_rexmitting = 0;
2722 
2723 	if (!tp->packets_out)
2724 		return;
2725 
2726 	if (!tp->lost_out)
2727 		tp->retransmit_high = tp->snd_una;
2728 
2729 	if (tp->retransmit_skb_hint) {
2730 		skb = tp->retransmit_skb_hint;
2731 		last_lost = TCP_SKB_CB(skb)->end_seq;
2732 		if (after(last_lost, tp->retransmit_high))
2733 			last_lost = tp->retransmit_high;
2734 	} else {
2735 		skb = tcp_write_queue_head(sk);
2736 		last_lost = tp->snd_una;
2737 	}
2738 
2739 	tcp_for_write_queue_from(skb, sk) {
2740 		__u8 sacked = TCP_SKB_CB(skb)->sacked;
2741 
2742 		if (skb == tcp_send_head(sk))
2743 			break;
2744 		/* we could do better than to assign each time */
2745 		if (!hole)
2746 			tp->retransmit_skb_hint = skb;
2747 
2748 		/* Assume this retransmit will generate
2749 		 * only one packet for congestion window
2750 		 * calculation purposes.  This works because
2751 		 * tcp_retransmit_skb() will chop up the
2752 		 * packet to be MSS sized and all the
2753 		 * packet counting works out.
2754 		 */
2755 		if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2756 			return;
2757 
2758 		if (fwd_rexmitting) {
2759 begin_fwd:
2760 			if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2761 				break;
2762 			mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2763 
2764 		} else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2765 			tp->retransmit_high = last_lost;
2766 			if (!tcp_can_forward_retransmit(sk))
2767 				break;
2768 			/* Backtrack if necessary to non-L'ed skb */
2769 			if (hole) {
2770 				skb = hole;
2771 				hole = NULL;
2772 			}
2773 			fwd_rexmitting = 1;
2774 			goto begin_fwd;
2775 
2776 		} else if (!(sacked & TCPCB_LOST)) {
2777 			if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2778 				hole = skb;
2779 			continue;
2780 
2781 		} else {
2782 			last_lost = TCP_SKB_CB(skb)->end_seq;
2783 			if (icsk->icsk_ca_state != TCP_CA_Loss)
2784 				mib_idx = LINUX_MIB_TCPFASTRETRANS;
2785 			else
2786 				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2787 		}
2788 
2789 		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2790 			continue;
2791 
2792 		if (tcp_retransmit_skb(sk, skb))
2793 			return;
2794 
2795 		NET_INC_STATS_BH(sock_net(sk), mib_idx);
2796 
2797 		if (tcp_in_cwnd_reduction(sk))
2798 			tp->prr_out += tcp_skb_pcount(skb);
2799 
2800 		if (skb == tcp_write_queue_head(sk))
2801 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2802 						  inet_csk(sk)->icsk_rto,
2803 						  TCP_RTO_MAX);
2804 	}
2805 }
2806 
2807 /* We allow to exceed memory limits for FIN packets to expedite
2808  * connection tear down and (memory) recovery.
2809  * Otherwise tcp_send_fin() could be tempted to either delay FIN
2810  * or even be forced to close flow without any FIN.
2811  * In general, we want to allow one skb per socket to avoid hangs
2812  * with edge trigger epoll()
2813  */
2814 void sk_forced_mem_schedule(struct sock *sk, int size)
2815 {
2816 	int amt, status;
2817 
2818 	if (size <= sk->sk_forward_alloc)
2819 		return;
2820 	amt = sk_mem_pages(size);
2821 	sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2822 	sk_memory_allocated_add(sk, amt, &status);
2823 }
2824 
2825 /* Send a FIN. The caller locks the socket for us.
2826  * We should try to send a FIN packet really hard, but eventually give up.
2827  */
2828 void tcp_send_fin(struct sock *sk)
2829 {
2830 	struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2831 	struct tcp_sock *tp = tcp_sk(sk);
2832 
2833 	/* Optimization, tack on the FIN if we have one skb in write queue and
2834 	 * this skb was not yet sent, or we are under memory pressure.
2835 	 * Note: in the latter case, FIN packet will be sent after a timeout,
2836 	 * as TCP stack thinks it has already been transmitted.
2837 	 */
2838 	if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
2839 coalesce:
2840 		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2841 		TCP_SKB_CB(tskb)->end_seq++;
2842 		tp->write_seq++;
2843 		if (!tcp_send_head(sk)) {
2844 			/* This means tskb was already sent.
2845 			 * Pretend we included the FIN on previous transmit.
2846 			 * We need to set tp->snd_nxt to the value it would have
2847 			 * if FIN had been sent. This is because retransmit path
2848 			 * does not change tp->snd_nxt.
2849 			 */
2850 			tp->snd_nxt++;
2851 			return;
2852 		}
2853 	} else {
2854 		skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2855 		if (unlikely(!skb)) {
2856 			if (tskb)
2857 				goto coalesce;
2858 			return;
2859 		}
2860 		skb_reserve(skb, MAX_TCP_HEADER);
2861 		sk_forced_mem_schedule(sk, skb->truesize);
2862 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2863 		tcp_init_nondata_skb(skb, tp->write_seq,
2864 				     TCPHDR_ACK | TCPHDR_FIN);
2865 		tcp_queue_skb(sk, skb);
2866 	}
2867 	__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2868 }
2869 
2870 /* We get here when a process closes a file descriptor (either due to
2871  * an explicit close() or as a byproduct of exit()'ing) and there
2872  * was unread data in the receive queue.  This behavior is recommended
2873  * by RFC 2525, section 2.17.  -DaveM
2874  */
2875 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2876 {
2877 	struct sk_buff *skb;
2878 
2879 	/* NOTE: No TCP options attached and we never retransmit this. */
2880 	skb = alloc_skb(MAX_TCP_HEADER, priority);
2881 	if (!skb) {
2882 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2883 		return;
2884 	}
2885 
2886 	/* Reserve space for headers and prepare control bits. */
2887 	skb_reserve(skb, MAX_TCP_HEADER);
2888 	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2889 			     TCPHDR_ACK | TCPHDR_RST);
2890 	skb_mstamp_get(&skb->skb_mstamp);
2891 	/* Send it off. */
2892 	if (tcp_transmit_skb(sk, skb, 0, priority))
2893 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2894 
2895 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2896 }
2897 
2898 /* Send a crossed SYN-ACK during socket establishment.
2899  * WARNING: This routine must only be called when we have already sent
2900  * a SYN packet that crossed the incoming SYN that caused this routine
2901  * to get called. If this assumption fails then the initial rcv_wnd
2902  * and rcv_wscale values will not be correct.
2903  */
2904 int tcp_send_synack(struct sock *sk)
2905 {
2906 	struct sk_buff *skb;
2907 
2908 	skb = tcp_write_queue_head(sk);
2909 	if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2910 		pr_debug("%s: wrong queue state\n", __func__);
2911 		return -EFAULT;
2912 	}
2913 	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2914 		if (skb_cloned(skb)) {
2915 			struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2916 			if (!nskb)
2917 				return -ENOMEM;
2918 			tcp_unlink_write_queue(skb, sk);
2919 			__skb_header_release(nskb);
2920 			__tcp_add_write_queue_head(sk, nskb);
2921 			sk_wmem_free_skb(sk, skb);
2922 			sk->sk_wmem_queued += nskb->truesize;
2923 			sk_mem_charge(sk, nskb->truesize);
2924 			skb = nskb;
2925 		}
2926 
2927 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2928 		tcp_ecn_send_synack(sk, skb);
2929 	}
2930 	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2931 }
2932 
2933 /**
2934  * tcp_make_synack - Prepare a SYN-ACK.
2935  * sk: listener socket
2936  * dst: dst entry attached to the SYNACK
2937  * req: request_sock pointer
2938  *
2939  * Allocate one skb and build a SYNACK packet.
2940  * @dst is consumed : Caller should not use it again.
2941  */
2942 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
2943 				struct request_sock *req,
2944 				struct tcp_fastopen_cookie *foc,
2945 				bool attach_req)
2946 {
2947 	struct inet_request_sock *ireq = inet_rsk(req);
2948 	const struct tcp_sock *tp = tcp_sk(sk);
2949 	struct tcp_md5sig_key *md5 = NULL;
2950 	struct tcp_out_options opts;
2951 	struct sk_buff *skb;
2952 	int tcp_header_size;
2953 	struct tcphdr *th;
2954 	u16 user_mss;
2955 	int mss;
2956 
2957 	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2958 	if (unlikely(!skb)) {
2959 		dst_release(dst);
2960 		return NULL;
2961 	}
2962 	/* Reserve space for headers. */
2963 	skb_reserve(skb, MAX_TCP_HEADER);
2964 
2965 	if (attach_req) {
2966 		skb_set_owner_w(skb, req_to_sk(req));
2967 	} else {
2968 		/* sk is a const pointer, because we want to express multiple
2969 		 * cpu might call us concurrently.
2970 		 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
2971 		 */
2972 		skb_set_owner_w(skb, (struct sock *)sk);
2973 	}
2974 	skb_dst_set(skb, dst);
2975 
2976 	mss = dst_metric_advmss(dst);
2977 	user_mss = READ_ONCE(tp->rx_opt.user_mss);
2978 	if (user_mss && user_mss < mss)
2979 		mss = user_mss;
2980 
2981 	memset(&opts, 0, sizeof(opts));
2982 #ifdef CONFIG_SYN_COOKIES
2983 	if (unlikely(req->cookie_ts))
2984 		skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
2985 	else
2986 #endif
2987 	skb_mstamp_get(&skb->skb_mstamp);
2988 
2989 #ifdef CONFIG_TCP_MD5SIG
2990 	rcu_read_lock();
2991 	md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
2992 #endif
2993 	skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
2994 	tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
2995 			  sizeof(*th);
2996 
2997 	skb_push(skb, tcp_header_size);
2998 	skb_reset_transport_header(skb);
2999 
3000 	th = tcp_hdr(skb);
3001 	memset(th, 0, sizeof(struct tcphdr));
3002 	th->syn = 1;
3003 	th->ack = 1;
3004 	tcp_ecn_make_synack(req, th);
3005 	th->source = htons(ireq->ir_num);
3006 	th->dest = ireq->ir_rmt_port;
3007 	/* Setting of flags are superfluous here for callers (and ECE is
3008 	 * not even correctly set)
3009 	 */
3010 	tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3011 			     TCPHDR_SYN | TCPHDR_ACK);
3012 
3013 	th->seq = htonl(TCP_SKB_CB(skb)->seq);
3014 	/* XXX data is queued and acked as is. No buffer/window check */
3015 	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3016 
3017 	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3018 	th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3019 	tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3020 	th->doff = (tcp_header_size >> 2);
3021 	TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
3022 
3023 #ifdef CONFIG_TCP_MD5SIG
3024 	/* Okay, we have all we need - do the md5 hash if needed */
3025 	if (md5)
3026 		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3027 					       md5, req_to_sk(req), skb);
3028 	rcu_read_unlock();
3029 #endif
3030 
3031 	/* Do not fool tcpdump (if any), clean our debris */
3032 	skb->tstamp.tv64 = 0;
3033 	return skb;
3034 }
3035 EXPORT_SYMBOL(tcp_make_synack);
3036 
3037 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3038 {
3039 	struct inet_connection_sock *icsk = inet_csk(sk);
3040 	const struct tcp_congestion_ops *ca;
3041 	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3042 
3043 	if (ca_key == TCP_CA_UNSPEC)
3044 		return;
3045 
3046 	rcu_read_lock();
3047 	ca = tcp_ca_find_key(ca_key);
3048 	if (likely(ca && try_module_get(ca->owner))) {
3049 		module_put(icsk->icsk_ca_ops->owner);
3050 		icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3051 		icsk->icsk_ca_ops = ca;
3052 	}
3053 	rcu_read_unlock();
3054 }
3055 
3056 /* Do all connect socket setups that can be done AF independent. */
3057 static void tcp_connect_init(struct sock *sk)
3058 {
3059 	const struct dst_entry *dst = __sk_dst_get(sk);
3060 	struct tcp_sock *tp = tcp_sk(sk);
3061 	__u8 rcv_wscale;
3062 
3063 	/* We'll fix this up when we get a response from the other end.
3064 	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3065 	 */
3066 	tp->tcp_header_len = sizeof(struct tcphdr) +
3067 		(sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3068 
3069 #ifdef CONFIG_TCP_MD5SIG
3070 	if (tp->af_specific->md5_lookup(sk, sk))
3071 		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3072 #endif
3073 
3074 	/* If user gave his TCP_MAXSEG, record it to clamp */
3075 	if (tp->rx_opt.user_mss)
3076 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3077 	tp->max_window = 0;
3078 	tcp_mtup_init(sk);
3079 	tcp_sync_mss(sk, dst_mtu(dst));
3080 
3081 	tcp_ca_dst_init(sk, dst);
3082 
3083 	if (!tp->window_clamp)
3084 		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3085 	tp->advmss = dst_metric_advmss(dst);
3086 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3087 		tp->advmss = tp->rx_opt.user_mss;
3088 
3089 	tcp_initialize_rcv_mss(sk);
3090 
3091 	/* limit the window selection if the user enforce a smaller rx buffer */
3092 	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3093 	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3094 		tp->window_clamp = tcp_full_space(sk);
3095 
3096 	tcp_select_initial_window(tcp_full_space(sk),
3097 				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3098 				  &tp->rcv_wnd,
3099 				  &tp->window_clamp,
3100 				  sysctl_tcp_window_scaling,
3101 				  &rcv_wscale,
3102 				  dst_metric(dst, RTAX_INITRWND));
3103 
3104 	tp->rx_opt.rcv_wscale = rcv_wscale;
3105 	tp->rcv_ssthresh = tp->rcv_wnd;
3106 
3107 	sk->sk_err = 0;
3108 	sock_reset_flag(sk, SOCK_DONE);
3109 	tp->snd_wnd = 0;
3110 	tcp_init_wl(tp, 0);
3111 	tp->snd_una = tp->write_seq;
3112 	tp->snd_sml = tp->write_seq;
3113 	tp->snd_up = tp->write_seq;
3114 	tp->snd_nxt = tp->write_seq;
3115 
3116 	if (likely(!tp->repair))
3117 		tp->rcv_nxt = 0;
3118 	else
3119 		tp->rcv_tstamp = tcp_time_stamp;
3120 	tp->rcv_wup = tp->rcv_nxt;
3121 	tp->copied_seq = tp->rcv_nxt;
3122 
3123 	inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3124 	inet_csk(sk)->icsk_retransmits = 0;
3125 	tcp_clear_retrans(tp);
3126 }
3127 
3128 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3129 {
3130 	struct tcp_sock *tp = tcp_sk(sk);
3131 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3132 
3133 	tcb->end_seq += skb->len;
3134 	__skb_header_release(skb);
3135 	__tcp_add_write_queue_tail(sk, skb);
3136 	sk->sk_wmem_queued += skb->truesize;
3137 	sk_mem_charge(sk, skb->truesize);
3138 	tp->write_seq = tcb->end_seq;
3139 	tp->packets_out += tcp_skb_pcount(skb);
3140 }
3141 
3142 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3143  * queue a data-only packet after the regular SYN, such that regular SYNs
3144  * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3145  * only the SYN sequence, the data are retransmitted in the first ACK.
3146  * If cookie is not cached or other error occurs, falls back to send a
3147  * regular SYN with Fast Open cookie request option.
3148  */
3149 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3150 {
3151 	struct tcp_sock *tp = tcp_sk(sk);
3152 	struct tcp_fastopen_request *fo = tp->fastopen_req;
3153 	int syn_loss = 0, space, err = 0;
3154 	unsigned long last_syn_loss = 0;
3155 	struct sk_buff *syn_data;
3156 
3157 	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3158 	tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3159 			       &syn_loss, &last_syn_loss);
3160 	/* Recurring FO SYN losses: revert to regular handshake temporarily */
3161 	if (syn_loss > 1 &&
3162 	    time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3163 		fo->cookie.len = -1;
3164 		goto fallback;
3165 	}
3166 
3167 	if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3168 		fo->cookie.len = -1;
3169 	else if (fo->cookie.len <= 0)
3170 		goto fallback;
3171 
3172 	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3173 	 * user-MSS. Reserve maximum option space for middleboxes that add
3174 	 * private TCP options. The cost is reduced data space in SYN :(
3175 	 */
3176 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3177 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3178 	space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3179 		MAX_TCP_OPTION_SPACE;
3180 
3181 	space = min_t(size_t, space, fo->size);
3182 
3183 	/* limit to order-0 allocations */
3184 	space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3185 
3186 	syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3187 	if (!syn_data)
3188 		goto fallback;
3189 	syn_data->ip_summed = CHECKSUM_PARTIAL;
3190 	memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3191 	if (space) {
3192 		int copied = copy_from_iter(skb_put(syn_data, space), space,
3193 					    &fo->data->msg_iter);
3194 		if (unlikely(!copied)) {
3195 			kfree_skb(syn_data);
3196 			goto fallback;
3197 		}
3198 		if (copied != space) {
3199 			skb_trim(syn_data, copied);
3200 			space = copied;
3201 		}
3202 	}
3203 	/* No more data pending in inet_wait_for_connect() */
3204 	if (space == fo->size)
3205 		fo->data = NULL;
3206 	fo->copied = space;
3207 
3208 	tcp_connect_queue_skb(sk, syn_data);
3209 
3210 	err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3211 
3212 	syn->skb_mstamp = syn_data->skb_mstamp;
3213 
3214 	/* Now full SYN+DATA was cloned and sent (or not),
3215 	 * remove the SYN from the original skb (syn_data)
3216 	 * we keep in write queue in case of a retransmit, as we
3217 	 * also have the SYN packet (with no data) in the same queue.
3218 	 */
3219 	TCP_SKB_CB(syn_data)->seq++;
3220 	TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3221 	if (!err) {
3222 		tp->syn_data = (fo->copied > 0);
3223 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3224 		goto done;
3225 	}
3226 
3227 fallback:
3228 	/* Send a regular SYN with Fast Open cookie request option */
3229 	if (fo->cookie.len > 0)
3230 		fo->cookie.len = 0;
3231 	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3232 	if (err)
3233 		tp->syn_fastopen = 0;
3234 done:
3235 	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3236 	return err;
3237 }
3238 
3239 /* Build a SYN and send it off. */
3240 int tcp_connect(struct sock *sk)
3241 {
3242 	struct tcp_sock *tp = tcp_sk(sk);
3243 	struct sk_buff *buff;
3244 	int err;
3245 
3246 	tcp_connect_init(sk);
3247 
3248 	if (unlikely(tp->repair)) {
3249 		tcp_finish_connect(sk, NULL);
3250 		return 0;
3251 	}
3252 
3253 	buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3254 	if (unlikely(!buff))
3255 		return -ENOBUFS;
3256 
3257 	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3258 	tp->retrans_stamp = tcp_time_stamp;
3259 	tcp_connect_queue_skb(sk, buff);
3260 	tcp_ecn_send_syn(sk, buff);
3261 
3262 	/* Send off SYN; include data in Fast Open. */
3263 	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3264 	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3265 	if (err == -ECONNREFUSED)
3266 		return err;
3267 
3268 	/* We change tp->snd_nxt after the tcp_transmit_skb() call
3269 	 * in order to make this packet get counted in tcpOutSegs.
3270 	 */
3271 	tp->snd_nxt = tp->write_seq;
3272 	tp->pushed_seq = tp->write_seq;
3273 	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3274 
3275 	/* Timer for repeating the SYN until an answer. */
3276 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3277 				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3278 	return 0;
3279 }
3280 EXPORT_SYMBOL(tcp_connect);
3281 
3282 /* Send out a delayed ack, the caller does the policy checking
3283  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3284  * for details.
3285  */
3286 void tcp_send_delayed_ack(struct sock *sk)
3287 {
3288 	struct inet_connection_sock *icsk = inet_csk(sk);
3289 	int ato = icsk->icsk_ack.ato;
3290 	unsigned long timeout;
3291 
3292 	tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3293 
3294 	if (ato > TCP_DELACK_MIN) {
3295 		const struct tcp_sock *tp = tcp_sk(sk);
3296 		int max_ato = HZ / 2;
3297 
3298 		if (icsk->icsk_ack.pingpong ||
3299 		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3300 			max_ato = TCP_DELACK_MAX;
3301 
3302 		/* Slow path, intersegment interval is "high". */
3303 
3304 		/* If some rtt estimate is known, use it to bound delayed ack.
3305 		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3306 		 * directly.
3307 		 */
3308 		if (tp->srtt_us) {
3309 			int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3310 					TCP_DELACK_MIN);
3311 
3312 			if (rtt < max_ato)
3313 				max_ato = rtt;
3314 		}
3315 
3316 		ato = min(ato, max_ato);
3317 	}
3318 
3319 	/* Stay within the limit we were given */
3320 	timeout = jiffies + ato;
3321 
3322 	/* Use new timeout only if there wasn't a older one earlier. */
3323 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3324 		/* If delack timer was blocked or is about to expire,
3325 		 * send ACK now.
3326 		 */
3327 		if (icsk->icsk_ack.blocked ||
3328 		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3329 			tcp_send_ack(sk);
3330 			return;
3331 		}
3332 
3333 		if (!time_before(timeout, icsk->icsk_ack.timeout))
3334 			timeout = icsk->icsk_ack.timeout;
3335 	}
3336 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3337 	icsk->icsk_ack.timeout = timeout;
3338 	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3339 }
3340 
3341 /* This routine sends an ack and also updates the window. */
3342 void tcp_send_ack(struct sock *sk)
3343 {
3344 	struct sk_buff *buff;
3345 
3346 	/* If we have been reset, we may not send again. */
3347 	if (sk->sk_state == TCP_CLOSE)
3348 		return;
3349 
3350 	tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3351 
3352 	/* We are not putting this on the write queue, so
3353 	 * tcp_transmit_skb() will set the ownership to this
3354 	 * sock.
3355 	 */
3356 	buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3357 	if (!buff) {
3358 		inet_csk_schedule_ack(sk);
3359 		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3360 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3361 					  TCP_DELACK_MAX, TCP_RTO_MAX);
3362 		return;
3363 	}
3364 
3365 	/* Reserve space for headers and prepare control bits. */
3366 	skb_reserve(buff, MAX_TCP_HEADER);
3367 	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3368 
3369 	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
3370 	 * too much.
3371 	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3372 	 * We also avoid tcp_wfree() overhead (cache line miss accessing
3373 	 * tp->tsq_flags) by using regular sock_wfree()
3374 	 */
3375 	skb_set_tcp_pure_ack(buff);
3376 
3377 	/* Send it off, this clears delayed acks for us. */
3378 	skb_mstamp_get(&buff->skb_mstamp);
3379 	tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3380 }
3381 EXPORT_SYMBOL_GPL(tcp_send_ack);
3382 
3383 /* This routine sends a packet with an out of date sequence
3384  * number. It assumes the other end will try to ack it.
3385  *
3386  * Question: what should we make while urgent mode?
3387  * 4.4BSD forces sending single byte of data. We cannot send
3388  * out of window data, because we have SND.NXT==SND.MAX...
3389  *
3390  * Current solution: to send TWO zero-length segments in urgent mode:
3391  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3392  * out-of-date with SND.UNA-1 to probe window.
3393  */
3394 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3395 {
3396 	struct tcp_sock *tp = tcp_sk(sk);
3397 	struct sk_buff *skb;
3398 
3399 	/* We don't queue it, tcp_transmit_skb() sets ownership. */
3400 	skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3401 	if (!skb)
3402 		return -1;
3403 
3404 	/* Reserve space for headers and set control bits. */
3405 	skb_reserve(skb, MAX_TCP_HEADER);
3406 	/* Use a previous sequence.  This should cause the other
3407 	 * end to send an ack.  Don't queue or clone SKB, just
3408 	 * send it.
3409 	 */
3410 	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3411 	skb_mstamp_get(&skb->skb_mstamp);
3412 	NET_INC_STATS(sock_net(sk), mib);
3413 	return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3414 }
3415 
3416 void tcp_send_window_probe(struct sock *sk)
3417 {
3418 	if (sk->sk_state == TCP_ESTABLISHED) {
3419 		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3420 		tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3421 	}
3422 }
3423 
3424 /* Initiate keepalive or window probe from timer. */
3425 int tcp_write_wakeup(struct sock *sk, int mib)
3426 {
3427 	struct tcp_sock *tp = tcp_sk(sk);
3428 	struct sk_buff *skb;
3429 
3430 	if (sk->sk_state == TCP_CLOSE)
3431 		return -1;
3432 
3433 	skb = tcp_send_head(sk);
3434 	if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3435 		int err;
3436 		unsigned int mss = tcp_current_mss(sk);
3437 		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3438 
3439 		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3440 			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3441 
3442 		/* We are probing the opening of a window
3443 		 * but the window size is != 0
3444 		 * must have been a result SWS avoidance ( sender )
3445 		 */
3446 		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3447 		    skb->len > mss) {
3448 			seg_size = min(seg_size, mss);
3449 			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3450 			if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3451 				return -1;
3452 		} else if (!tcp_skb_pcount(skb))
3453 			tcp_set_skb_tso_segs(skb, mss);
3454 
3455 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3456 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3457 		if (!err)
3458 			tcp_event_new_data_sent(sk, skb);
3459 		return err;
3460 	} else {
3461 		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3462 			tcp_xmit_probe_skb(sk, 1, mib);
3463 		return tcp_xmit_probe_skb(sk, 0, mib);
3464 	}
3465 }
3466 
3467 /* A window probe timeout has occurred.  If window is not closed send
3468  * a partial packet else a zero probe.
3469  */
3470 void tcp_send_probe0(struct sock *sk)
3471 {
3472 	struct inet_connection_sock *icsk = inet_csk(sk);
3473 	struct tcp_sock *tp = tcp_sk(sk);
3474 	unsigned long probe_max;
3475 	int err;
3476 
3477 	err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3478 
3479 	if (tp->packets_out || !tcp_send_head(sk)) {
3480 		/* Cancel probe timer, if it is not required. */
3481 		icsk->icsk_probes_out = 0;
3482 		icsk->icsk_backoff = 0;
3483 		return;
3484 	}
3485 
3486 	if (err <= 0) {
3487 		if (icsk->icsk_backoff < sysctl_tcp_retries2)
3488 			icsk->icsk_backoff++;
3489 		icsk->icsk_probes_out++;
3490 		probe_max = TCP_RTO_MAX;
3491 	} else {
3492 		/* If packet was not sent due to local congestion,
3493 		 * do not backoff and do not remember icsk_probes_out.
3494 		 * Let local senders to fight for local resources.
3495 		 *
3496 		 * Use accumulated backoff yet.
3497 		 */
3498 		if (!icsk->icsk_probes_out)
3499 			icsk->icsk_probes_out = 1;
3500 		probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3501 	}
3502 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3503 				  tcp_probe0_when(sk, probe_max),
3504 				  TCP_RTO_MAX);
3505 }
3506 
3507 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3508 {
3509 	const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3510 	struct flowi fl;
3511 	int res;
3512 
3513 	tcp_rsk(req)->txhash = net_tx_rndhash();
3514 	res = af_ops->send_synack(sk, NULL, &fl, req, NULL, true);
3515 	if (!res) {
3516 		TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3517 		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3518 	}
3519 	return res;
3520 }
3521 EXPORT_SYMBOL(tcp_rtx_synack);
3522