xref: /openbmc/linux/net/ipv4/tcp_output.c (revision 82003e04)
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 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 			   int push_one, gfp_t gfp);
67 
68 /* Account for new data that has been sent to the network. */
69 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
70 {
71 	struct inet_connection_sock *icsk = inet_csk(sk);
72 	struct tcp_sock *tp = tcp_sk(sk);
73 	unsigned int prior_packets = tp->packets_out;
74 
75 	tcp_advance_send_head(sk, skb);
76 	tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
77 
78 	tp->packets_out += tcp_skb_pcount(skb);
79 	if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
80 	    icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
81 		tcp_rearm_rto(sk);
82 	}
83 
84 	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
85 		      tcp_skb_pcount(skb));
86 }
87 
88 /* SND.NXT, if window was not shrunk.
89  * If window has been shrunk, what should we make? It is not clear at all.
90  * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91  * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92  * invalid. OK, let's make this for now:
93  */
94 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
95 {
96 	const struct tcp_sock *tp = tcp_sk(sk);
97 
98 	if (!before(tcp_wnd_end(tp), tp->snd_nxt))
99 		return tp->snd_nxt;
100 	else
101 		return tcp_wnd_end(tp);
102 }
103 
104 /* Calculate mss to advertise in SYN segment.
105  * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
106  *
107  * 1. It is independent of path mtu.
108  * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109  * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110  *    attached devices, because some buggy hosts are confused by
111  *    large MSS.
112  * 4. We do not make 3, we advertise MSS, calculated from first
113  *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
114  *    This may be overridden via information stored in routing table.
115  * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116  *    probably even Jumbo".
117  */
118 static __u16 tcp_advertise_mss(struct sock *sk)
119 {
120 	struct tcp_sock *tp = tcp_sk(sk);
121 	const struct dst_entry *dst = __sk_dst_get(sk);
122 	int mss = tp->advmss;
123 
124 	if (dst) {
125 		unsigned int metric = dst_metric_advmss(dst);
126 
127 		if (metric < mss) {
128 			mss = metric;
129 			tp->advmss = mss;
130 		}
131 	}
132 
133 	return (__u16)mss;
134 }
135 
136 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
137  * This is the first part of cwnd validation mechanism.
138  */
139 void tcp_cwnd_restart(struct sock *sk, s32 delta)
140 {
141 	struct tcp_sock *tp = tcp_sk(sk);
142 	u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
143 	u32 cwnd = tp->snd_cwnd;
144 
145 	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
146 
147 	tp->snd_ssthresh = tcp_current_ssthresh(sk);
148 	restart_cwnd = min(restart_cwnd, cwnd);
149 
150 	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
151 		cwnd >>= 1;
152 	tp->snd_cwnd = max(cwnd, restart_cwnd);
153 	tp->snd_cwnd_stamp = tcp_time_stamp;
154 	tp->snd_cwnd_used = 0;
155 }
156 
157 /* Congestion state accounting after a packet has been sent. */
158 static void tcp_event_data_sent(struct tcp_sock *tp,
159 				struct sock *sk)
160 {
161 	struct inet_connection_sock *icsk = inet_csk(sk);
162 	const u32 now = tcp_time_stamp;
163 
164 	if (tcp_packets_in_flight(tp) == 0)
165 		tcp_ca_event(sk, CA_EVENT_TX_START);
166 
167 	tp->lsndtime = now;
168 
169 	/* If it is a reply for ato after last received
170 	 * packet, enter pingpong mode.
171 	 */
172 	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
173 		icsk->icsk_ack.pingpong = 1;
174 }
175 
176 /* Account for an ACK we sent. */
177 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
178 {
179 	tcp_dec_quickack_mode(sk, pkts);
180 	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
181 }
182 
183 
184 u32 tcp_default_init_rwnd(u32 mss)
185 {
186 	/* Initial receive window should be twice of TCP_INIT_CWND to
187 	 * enable proper sending of new unsent data during fast recovery
188 	 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
189 	 * limit when mss is larger than 1460.
190 	 */
191 	u32 init_rwnd = TCP_INIT_CWND * 2;
192 
193 	if (mss > 1460)
194 		init_rwnd = max((1460 * init_rwnd) / mss, 2U);
195 	return init_rwnd;
196 }
197 
198 /* Determine a window scaling and initial window to offer.
199  * Based on the assumption that the given amount of space
200  * will be offered. Store the results in the tp structure.
201  * NOTE: for smooth operation initial space offering should
202  * be a multiple of mss if possible. We assume here that mss >= 1.
203  * This MUST be enforced by all callers.
204  */
205 void tcp_select_initial_window(int __space, __u32 mss,
206 			       __u32 *rcv_wnd, __u32 *window_clamp,
207 			       int wscale_ok, __u8 *rcv_wscale,
208 			       __u32 init_rcv_wnd)
209 {
210 	unsigned int space = (__space < 0 ? 0 : __space);
211 
212 	/* If no clamp set the clamp to the max possible scaled window */
213 	if (*window_clamp == 0)
214 		(*window_clamp) = (65535 << 14);
215 	space = min(*window_clamp, space);
216 
217 	/* Quantize space offering to a multiple of mss if possible. */
218 	if (space > mss)
219 		space = (space / mss) * mss;
220 
221 	/* NOTE: offering an initial window larger than 32767
222 	 * will break some buggy TCP stacks. If the admin tells us
223 	 * it is likely we could be speaking with such a buggy stack
224 	 * we will truncate our initial window offering to 32K-1
225 	 * unless the remote has sent us a window scaling option,
226 	 * which we interpret as a sign the remote TCP is not
227 	 * misinterpreting the window field as a signed quantity.
228 	 */
229 	if (sysctl_tcp_workaround_signed_windows)
230 		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
231 	else
232 		(*rcv_wnd) = space;
233 
234 	(*rcv_wscale) = 0;
235 	if (wscale_ok) {
236 		/* Set window scaling on max possible window
237 		 * See RFC1323 for an explanation of the limit to 14
238 		 */
239 		space = max_t(u32, space, sysctl_tcp_rmem[2]);
240 		space = max_t(u32, space, sysctl_rmem_max);
241 		space = min_t(u32, space, *window_clamp);
242 		while (space > 65535 && (*rcv_wscale) < 14) {
243 			space >>= 1;
244 			(*rcv_wscale)++;
245 		}
246 	}
247 
248 	if (mss > (1 << *rcv_wscale)) {
249 		if (!init_rcv_wnd) /* Use default unless specified otherwise */
250 			init_rcv_wnd = tcp_default_init_rwnd(mss);
251 		*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
252 	}
253 
254 	/* Set the clamp no higher than max representable value */
255 	(*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
256 }
257 EXPORT_SYMBOL(tcp_select_initial_window);
258 
259 /* Chose a new window to advertise, update state in tcp_sock for the
260  * socket, and return result with RFC1323 scaling applied.  The return
261  * value can be stuffed directly into th->window for an outgoing
262  * frame.
263  */
264 static u16 tcp_select_window(struct sock *sk)
265 {
266 	struct tcp_sock *tp = tcp_sk(sk);
267 	u32 old_win = tp->rcv_wnd;
268 	u32 cur_win = tcp_receive_window(tp);
269 	u32 new_win = __tcp_select_window(sk);
270 
271 	/* Never shrink the offered window */
272 	if (new_win < cur_win) {
273 		/* Danger Will Robinson!
274 		 * Don't update rcv_wup/rcv_wnd here or else
275 		 * we will not be able to advertise a zero
276 		 * window in time.  --DaveM
277 		 *
278 		 * Relax Will Robinson.
279 		 */
280 		if (new_win == 0)
281 			NET_INC_STATS(sock_net(sk),
282 				      LINUX_MIB_TCPWANTZEROWINDOWADV);
283 		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
284 	}
285 	tp->rcv_wnd = new_win;
286 	tp->rcv_wup = tp->rcv_nxt;
287 
288 	/* Make sure we do not exceed the maximum possible
289 	 * scaled window.
290 	 */
291 	if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
292 		new_win = min(new_win, MAX_TCP_WINDOW);
293 	else
294 		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
295 
296 	/* RFC1323 scaling applied */
297 	new_win >>= tp->rx_opt.rcv_wscale;
298 
299 	/* If we advertise zero window, disable fast path. */
300 	if (new_win == 0) {
301 		tp->pred_flags = 0;
302 		if (old_win)
303 			NET_INC_STATS(sock_net(sk),
304 				      LINUX_MIB_TCPTOZEROWINDOWADV);
305 	} else if (old_win == 0) {
306 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
307 	}
308 
309 	return new_win;
310 }
311 
312 /* Packet ECN state for a SYN-ACK */
313 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
314 {
315 	const struct tcp_sock *tp = tcp_sk(sk);
316 
317 	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
318 	if (!(tp->ecn_flags & TCP_ECN_OK))
319 		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
320 	else if (tcp_ca_needs_ecn(sk))
321 		INET_ECN_xmit(sk);
322 }
323 
324 /* Packet ECN state for a SYN.  */
325 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
326 {
327 	struct tcp_sock *tp = tcp_sk(sk);
328 	bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
329 		       tcp_ca_needs_ecn(sk);
330 
331 	if (!use_ecn) {
332 		const struct dst_entry *dst = __sk_dst_get(sk);
333 
334 		if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
335 			use_ecn = true;
336 	}
337 
338 	tp->ecn_flags = 0;
339 
340 	if (use_ecn) {
341 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
342 		tp->ecn_flags = TCP_ECN_OK;
343 		if (tcp_ca_needs_ecn(sk))
344 			INET_ECN_xmit(sk);
345 	}
346 }
347 
348 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
349 {
350 	if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
351 		/* tp->ecn_flags are cleared at a later point in time when
352 		 * SYN ACK is ultimatively being received.
353 		 */
354 		TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
355 }
356 
357 static void
358 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
359 {
360 	if (inet_rsk(req)->ecn_ok)
361 		th->ece = 1;
362 }
363 
364 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
365  * be sent.
366  */
367 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
368 			 struct tcphdr *th, int tcp_header_len)
369 {
370 	struct tcp_sock *tp = tcp_sk(sk);
371 
372 	if (tp->ecn_flags & TCP_ECN_OK) {
373 		/* Not-retransmitted data segment: set ECT and inject CWR. */
374 		if (skb->len != tcp_header_len &&
375 		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
376 			INET_ECN_xmit(sk);
377 			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
378 				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
379 				th->cwr = 1;
380 				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
381 			}
382 		} else if (!tcp_ca_needs_ecn(sk)) {
383 			/* ACK or retransmitted segment: clear ECT|CE */
384 			INET_ECN_dontxmit(sk);
385 		}
386 		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
387 			th->ece = 1;
388 	}
389 }
390 
391 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
392  * auto increment end seqno.
393  */
394 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
395 {
396 	skb->ip_summed = CHECKSUM_PARTIAL;
397 	skb->csum = 0;
398 
399 	TCP_SKB_CB(skb)->tcp_flags = flags;
400 	TCP_SKB_CB(skb)->sacked = 0;
401 
402 	tcp_skb_pcount_set(skb, 1);
403 
404 	TCP_SKB_CB(skb)->seq = seq;
405 	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
406 		seq++;
407 	TCP_SKB_CB(skb)->end_seq = seq;
408 }
409 
410 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
411 {
412 	return tp->snd_una != tp->snd_up;
413 }
414 
415 #define OPTION_SACK_ADVERTISE	(1 << 0)
416 #define OPTION_TS		(1 << 1)
417 #define OPTION_MD5		(1 << 2)
418 #define OPTION_WSCALE		(1 << 3)
419 #define OPTION_FAST_OPEN_COOKIE	(1 << 8)
420 
421 struct tcp_out_options {
422 	u16 options;		/* bit field of OPTION_* */
423 	u16 mss;		/* 0 to disable */
424 	u8 ws;			/* window scale, 0 to disable */
425 	u8 num_sack_blocks;	/* number of SACK blocks to include */
426 	u8 hash_size;		/* bytes in hash_location */
427 	__u8 *hash_location;	/* temporary pointer, overloaded */
428 	__u32 tsval, tsecr;	/* need to include OPTION_TS */
429 	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
430 };
431 
432 /* Write previously computed TCP options to the packet.
433  *
434  * Beware: Something in the Internet is very sensitive to the ordering of
435  * TCP options, we learned this through the hard way, so be careful here.
436  * Luckily we can at least blame others for their non-compliance but from
437  * inter-operability perspective it seems that we're somewhat stuck with
438  * the ordering which we have been using if we want to keep working with
439  * those broken things (not that it currently hurts anybody as there isn't
440  * particular reason why the ordering would need to be changed).
441  *
442  * At least SACK_PERM as the first option is known to lead to a disaster
443  * (but it may well be that other scenarios fail similarly).
444  */
445 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
446 			      struct tcp_out_options *opts)
447 {
448 	u16 options = opts->options;	/* mungable copy */
449 
450 	if (unlikely(OPTION_MD5 & options)) {
451 		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
452 			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
453 		/* overload cookie hash location */
454 		opts->hash_location = (__u8 *)ptr;
455 		ptr += 4;
456 	}
457 
458 	if (unlikely(opts->mss)) {
459 		*ptr++ = htonl((TCPOPT_MSS << 24) |
460 			       (TCPOLEN_MSS << 16) |
461 			       opts->mss);
462 	}
463 
464 	if (likely(OPTION_TS & options)) {
465 		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
466 			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
467 				       (TCPOLEN_SACK_PERM << 16) |
468 				       (TCPOPT_TIMESTAMP << 8) |
469 				       TCPOLEN_TIMESTAMP);
470 			options &= ~OPTION_SACK_ADVERTISE;
471 		} else {
472 			*ptr++ = htonl((TCPOPT_NOP << 24) |
473 				       (TCPOPT_NOP << 16) |
474 				       (TCPOPT_TIMESTAMP << 8) |
475 				       TCPOLEN_TIMESTAMP);
476 		}
477 		*ptr++ = htonl(opts->tsval);
478 		*ptr++ = htonl(opts->tsecr);
479 	}
480 
481 	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
482 		*ptr++ = htonl((TCPOPT_NOP << 24) |
483 			       (TCPOPT_NOP << 16) |
484 			       (TCPOPT_SACK_PERM << 8) |
485 			       TCPOLEN_SACK_PERM);
486 	}
487 
488 	if (unlikely(OPTION_WSCALE & options)) {
489 		*ptr++ = htonl((TCPOPT_NOP << 24) |
490 			       (TCPOPT_WINDOW << 16) |
491 			       (TCPOLEN_WINDOW << 8) |
492 			       opts->ws);
493 	}
494 
495 	if (unlikely(opts->num_sack_blocks)) {
496 		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
497 			tp->duplicate_sack : tp->selective_acks;
498 		int this_sack;
499 
500 		*ptr++ = htonl((TCPOPT_NOP  << 24) |
501 			       (TCPOPT_NOP  << 16) |
502 			       (TCPOPT_SACK <<  8) |
503 			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
504 						     TCPOLEN_SACK_PERBLOCK)));
505 
506 		for (this_sack = 0; this_sack < opts->num_sack_blocks;
507 		     ++this_sack) {
508 			*ptr++ = htonl(sp[this_sack].start_seq);
509 			*ptr++ = htonl(sp[this_sack].end_seq);
510 		}
511 
512 		tp->rx_opt.dsack = 0;
513 	}
514 
515 	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
516 		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
517 		u8 *p = (u8 *)ptr;
518 		u32 len; /* Fast Open option length */
519 
520 		if (foc->exp) {
521 			len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
522 			*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
523 				     TCPOPT_FASTOPEN_MAGIC);
524 			p += TCPOLEN_EXP_FASTOPEN_BASE;
525 		} else {
526 			len = TCPOLEN_FASTOPEN_BASE + foc->len;
527 			*p++ = TCPOPT_FASTOPEN;
528 			*p++ = len;
529 		}
530 
531 		memcpy(p, foc->val, foc->len);
532 		if ((len & 3) == 2) {
533 			p[foc->len] = TCPOPT_NOP;
534 			p[foc->len + 1] = TCPOPT_NOP;
535 		}
536 		ptr += (len + 3) >> 2;
537 	}
538 }
539 
540 /* Compute TCP options for SYN packets. This is not the final
541  * network wire format yet.
542  */
543 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
544 				struct tcp_out_options *opts,
545 				struct tcp_md5sig_key **md5)
546 {
547 	struct tcp_sock *tp = tcp_sk(sk);
548 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
549 	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
550 
551 #ifdef CONFIG_TCP_MD5SIG
552 	*md5 = tp->af_specific->md5_lookup(sk, sk);
553 	if (*md5) {
554 		opts->options |= OPTION_MD5;
555 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
556 	}
557 #else
558 	*md5 = NULL;
559 #endif
560 
561 	/* We always get an MSS option.  The option bytes which will be seen in
562 	 * normal data packets should timestamps be used, must be in the MSS
563 	 * advertised.  But we subtract them from tp->mss_cache so that
564 	 * calculations in tcp_sendmsg are simpler etc.  So account for this
565 	 * fact here if necessary.  If we don't do this correctly, as a
566 	 * receiver we won't recognize data packets as being full sized when we
567 	 * should, and thus we won't abide by the delayed ACK rules correctly.
568 	 * SACKs don't matter, we never delay an ACK when we have any of those
569 	 * going out.  */
570 	opts->mss = tcp_advertise_mss(sk);
571 	remaining -= TCPOLEN_MSS_ALIGNED;
572 
573 	if (likely(sysctl_tcp_timestamps && !*md5)) {
574 		opts->options |= OPTION_TS;
575 		opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
576 		opts->tsecr = tp->rx_opt.ts_recent;
577 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
578 	}
579 	if (likely(sysctl_tcp_window_scaling)) {
580 		opts->ws = tp->rx_opt.rcv_wscale;
581 		opts->options |= OPTION_WSCALE;
582 		remaining -= TCPOLEN_WSCALE_ALIGNED;
583 	}
584 	if (likely(sysctl_tcp_sack)) {
585 		opts->options |= OPTION_SACK_ADVERTISE;
586 		if (unlikely(!(OPTION_TS & opts->options)))
587 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
588 	}
589 
590 	if (fastopen && fastopen->cookie.len >= 0) {
591 		u32 need = fastopen->cookie.len;
592 
593 		need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
594 					       TCPOLEN_FASTOPEN_BASE;
595 		need = (need + 3) & ~3U;  /* Align to 32 bits */
596 		if (remaining >= need) {
597 			opts->options |= OPTION_FAST_OPEN_COOKIE;
598 			opts->fastopen_cookie = &fastopen->cookie;
599 			remaining -= need;
600 			tp->syn_fastopen = 1;
601 			tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
602 		}
603 	}
604 
605 	return MAX_TCP_OPTION_SPACE - remaining;
606 }
607 
608 /* Set up TCP options for SYN-ACKs. */
609 static unsigned int tcp_synack_options(struct request_sock *req,
610 				       unsigned int mss, struct sk_buff *skb,
611 				       struct tcp_out_options *opts,
612 				       const struct tcp_md5sig_key *md5,
613 				       struct tcp_fastopen_cookie *foc)
614 {
615 	struct inet_request_sock *ireq = inet_rsk(req);
616 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
617 
618 #ifdef CONFIG_TCP_MD5SIG
619 	if (md5) {
620 		opts->options |= OPTION_MD5;
621 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
622 
623 		/* We can't fit any SACK blocks in a packet with MD5 + TS
624 		 * options. There was discussion about disabling SACK
625 		 * rather than TS in order to fit in better with old,
626 		 * buggy kernels, but that was deemed to be unnecessary.
627 		 */
628 		ireq->tstamp_ok &= !ireq->sack_ok;
629 	}
630 #endif
631 
632 	/* We always send an MSS option. */
633 	opts->mss = mss;
634 	remaining -= TCPOLEN_MSS_ALIGNED;
635 
636 	if (likely(ireq->wscale_ok)) {
637 		opts->ws = ireq->rcv_wscale;
638 		opts->options |= OPTION_WSCALE;
639 		remaining -= TCPOLEN_WSCALE_ALIGNED;
640 	}
641 	if (likely(ireq->tstamp_ok)) {
642 		opts->options |= OPTION_TS;
643 		opts->tsval = tcp_skb_timestamp(skb);
644 		opts->tsecr = req->ts_recent;
645 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
646 	}
647 	if (likely(ireq->sack_ok)) {
648 		opts->options |= OPTION_SACK_ADVERTISE;
649 		if (unlikely(!ireq->tstamp_ok))
650 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
651 	}
652 	if (foc != NULL && foc->len >= 0) {
653 		u32 need = foc->len;
654 
655 		need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
656 				   TCPOLEN_FASTOPEN_BASE;
657 		need = (need + 3) & ~3U;  /* Align to 32 bits */
658 		if (remaining >= need) {
659 			opts->options |= OPTION_FAST_OPEN_COOKIE;
660 			opts->fastopen_cookie = foc;
661 			remaining -= need;
662 		}
663 	}
664 
665 	return MAX_TCP_OPTION_SPACE - remaining;
666 }
667 
668 /* Compute TCP options for ESTABLISHED sockets. This is not the
669  * final wire format yet.
670  */
671 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
672 					struct tcp_out_options *opts,
673 					struct tcp_md5sig_key **md5)
674 {
675 	struct tcp_sock *tp = tcp_sk(sk);
676 	unsigned int size = 0;
677 	unsigned int eff_sacks;
678 
679 	opts->options = 0;
680 
681 #ifdef CONFIG_TCP_MD5SIG
682 	*md5 = tp->af_specific->md5_lookup(sk, sk);
683 	if (unlikely(*md5)) {
684 		opts->options |= OPTION_MD5;
685 		size += TCPOLEN_MD5SIG_ALIGNED;
686 	}
687 #else
688 	*md5 = NULL;
689 #endif
690 
691 	if (likely(tp->rx_opt.tstamp_ok)) {
692 		opts->options |= OPTION_TS;
693 		opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
694 		opts->tsecr = tp->rx_opt.ts_recent;
695 		size += TCPOLEN_TSTAMP_ALIGNED;
696 	}
697 
698 	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
699 	if (unlikely(eff_sacks)) {
700 		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
701 		opts->num_sack_blocks =
702 			min_t(unsigned int, eff_sacks,
703 			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
704 			      TCPOLEN_SACK_PERBLOCK);
705 		size += TCPOLEN_SACK_BASE_ALIGNED +
706 			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
707 	}
708 
709 	return size;
710 }
711 
712 
713 /* TCP SMALL QUEUES (TSQ)
714  *
715  * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
716  * to reduce RTT and bufferbloat.
717  * We do this using a special skb destructor (tcp_wfree).
718  *
719  * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
720  * needs to be reallocated in a driver.
721  * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
722  *
723  * Since transmit from skb destructor is forbidden, we use a tasklet
724  * to process all sockets that eventually need to send more skbs.
725  * We use one tasklet per cpu, with its own queue of sockets.
726  */
727 struct tsq_tasklet {
728 	struct tasklet_struct	tasklet;
729 	struct list_head	head; /* queue of tcp sockets */
730 };
731 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
732 
733 static void tcp_tsq_handler(struct sock *sk)
734 {
735 	if ((1 << sk->sk_state) &
736 	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
737 	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK)) {
738 		struct tcp_sock *tp = tcp_sk(sk);
739 
740 		if (tp->lost_out > tp->retrans_out &&
741 		    tp->snd_cwnd > tcp_packets_in_flight(tp))
742 			tcp_xmit_retransmit_queue(sk);
743 
744 		tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
745 			       0, GFP_ATOMIC);
746 	}
747 }
748 /*
749  * One tasklet per cpu tries to send more skbs.
750  * We run in tasklet context but need to disable irqs when
751  * transferring tsq->head because tcp_wfree() might
752  * interrupt us (non NAPI drivers)
753  */
754 static void tcp_tasklet_func(unsigned long data)
755 {
756 	struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
757 	LIST_HEAD(list);
758 	unsigned long flags;
759 	struct list_head *q, *n;
760 	struct tcp_sock *tp;
761 	struct sock *sk;
762 
763 	local_irq_save(flags);
764 	list_splice_init(&tsq->head, &list);
765 	local_irq_restore(flags);
766 
767 	list_for_each_safe(q, n, &list) {
768 		tp = list_entry(q, struct tcp_sock, tsq_node);
769 		list_del(&tp->tsq_node);
770 
771 		sk = (struct sock *)tp;
772 		bh_lock_sock(sk);
773 
774 		if (!sock_owned_by_user(sk)) {
775 			tcp_tsq_handler(sk);
776 		} else {
777 			/* defer the work to tcp_release_cb() */
778 			set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
779 		}
780 		bh_unlock_sock(sk);
781 
782 		clear_bit(TSQ_QUEUED, &tp->tsq_flags);
783 		sk_free(sk);
784 	}
785 }
786 
787 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) |		\
788 			  (1UL << TCP_WRITE_TIMER_DEFERRED) |	\
789 			  (1UL << TCP_DELACK_TIMER_DEFERRED) |	\
790 			  (1UL << TCP_MTU_REDUCED_DEFERRED))
791 /**
792  * tcp_release_cb - tcp release_sock() callback
793  * @sk: socket
794  *
795  * called from release_sock() to perform protocol dependent
796  * actions before socket release.
797  */
798 void tcp_release_cb(struct sock *sk)
799 {
800 	struct tcp_sock *tp = tcp_sk(sk);
801 	unsigned long flags, nflags;
802 
803 	/* perform an atomic operation only if at least one flag is set */
804 	do {
805 		flags = tp->tsq_flags;
806 		if (!(flags & TCP_DEFERRED_ALL))
807 			return;
808 		nflags = flags & ~TCP_DEFERRED_ALL;
809 	} while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
810 
811 	if (flags & (1UL << TCP_TSQ_DEFERRED))
812 		tcp_tsq_handler(sk);
813 
814 	/* Here begins the tricky part :
815 	 * We are called from release_sock() with :
816 	 * 1) BH disabled
817 	 * 2) sk_lock.slock spinlock held
818 	 * 3) socket owned by us (sk->sk_lock.owned == 1)
819 	 *
820 	 * But following code is meant to be called from BH handlers,
821 	 * so we should keep BH disabled, but early release socket ownership
822 	 */
823 	sock_release_ownership(sk);
824 
825 	if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
826 		tcp_write_timer_handler(sk);
827 		__sock_put(sk);
828 	}
829 	if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
830 		tcp_delack_timer_handler(sk);
831 		__sock_put(sk);
832 	}
833 	if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
834 		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
835 		__sock_put(sk);
836 	}
837 }
838 EXPORT_SYMBOL(tcp_release_cb);
839 
840 void __init tcp_tasklet_init(void)
841 {
842 	int i;
843 
844 	for_each_possible_cpu(i) {
845 		struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
846 
847 		INIT_LIST_HEAD(&tsq->head);
848 		tasklet_init(&tsq->tasklet,
849 			     tcp_tasklet_func,
850 			     (unsigned long)tsq);
851 	}
852 }
853 
854 /*
855  * Write buffer destructor automatically called from kfree_skb.
856  * We can't xmit new skbs from this context, as we might already
857  * hold qdisc lock.
858  */
859 void tcp_wfree(struct sk_buff *skb)
860 {
861 	struct sock *sk = skb->sk;
862 	struct tcp_sock *tp = tcp_sk(sk);
863 	int wmem;
864 
865 	/* Keep one reference on sk_wmem_alloc.
866 	 * Will be released by sk_free() from here or tcp_tasklet_func()
867 	 */
868 	wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
869 
870 	/* If this softirq is serviced by ksoftirqd, we are likely under stress.
871 	 * Wait until our queues (qdisc + devices) are drained.
872 	 * This gives :
873 	 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
874 	 * - chance for incoming ACK (processed by another cpu maybe)
875 	 *   to migrate this flow (skb->ooo_okay will be eventually set)
876 	 */
877 	if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
878 		goto out;
879 
880 	if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
881 	    !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
882 		unsigned long flags;
883 		struct tsq_tasklet *tsq;
884 
885 		/* queue this socket to tasklet queue */
886 		local_irq_save(flags);
887 		tsq = this_cpu_ptr(&tsq_tasklet);
888 		list_add(&tp->tsq_node, &tsq->head);
889 		tasklet_schedule(&tsq->tasklet);
890 		local_irq_restore(flags);
891 		return;
892 	}
893 out:
894 	sk_free(sk);
895 }
896 
897 /* This routine actually transmits TCP packets queued in by
898  * tcp_do_sendmsg().  This is used by both the initial
899  * transmission and possible later retransmissions.
900  * All SKB's seen here are completely headerless.  It is our
901  * job to build the TCP header, and pass the packet down to
902  * IP so it can do the same plus pass the packet off to the
903  * device.
904  *
905  * We are working here with either a clone of the original
906  * SKB, or a fresh unique copy made by the retransmit engine.
907  */
908 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
909 			    gfp_t gfp_mask)
910 {
911 	const struct inet_connection_sock *icsk = inet_csk(sk);
912 	struct inet_sock *inet;
913 	struct tcp_sock *tp;
914 	struct tcp_skb_cb *tcb;
915 	struct tcp_out_options opts;
916 	unsigned int tcp_options_size, tcp_header_size;
917 	struct tcp_md5sig_key *md5;
918 	struct tcphdr *th;
919 	int err;
920 
921 	BUG_ON(!skb || !tcp_skb_pcount(skb));
922 	tp = tcp_sk(sk);
923 
924 	if (clone_it) {
925 		skb_mstamp_get(&skb->skb_mstamp);
926 		TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
927 			- tp->snd_una;
928 		tcp_rate_skb_sent(sk, skb);
929 
930 		if (unlikely(skb_cloned(skb)))
931 			skb = pskb_copy(skb, gfp_mask);
932 		else
933 			skb = skb_clone(skb, gfp_mask);
934 		if (unlikely(!skb))
935 			return -ENOBUFS;
936 	}
937 
938 	inet = inet_sk(sk);
939 	tcb = TCP_SKB_CB(skb);
940 	memset(&opts, 0, sizeof(opts));
941 
942 	if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
943 		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
944 	else
945 		tcp_options_size = tcp_established_options(sk, skb, &opts,
946 							   &md5);
947 	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
948 
949 	/* if no packet is in qdisc/device queue, then allow XPS to select
950 	 * another queue. We can be called from tcp_tsq_handler()
951 	 * which holds one reference to sk_wmem_alloc.
952 	 *
953 	 * TODO: Ideally, in-flight pure ACK packets should not matter here.
954 	 * One way to get this would be to set skb->truesize = 2 on them.
955 	 */
956 	skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
957 
958 	skb_push(skb, tcp_header_size);
959 	skb_reset_transport_header(skb);
960 
961 	skb_orphan(skb);
962 	skb->sk = sk;
963 	skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
964 	skb_set_hash_from_sk(skb, sk);
965 	atomic_add(skb->truesize, &sk->sk_wmem_alloc);
966 
967 	/* Build TCP header and checksum it. */
968 	th = (struct tcphdr *)skb->data;
969 	th->source		= inet->inet_sport;
970 	th->dest		= inet->inet_dport;
971 	th->seq			= htonl(tcb->seq);
972 	th->ack_seq		= htonl(tp->rcv_nxt);
973 	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
974 					tcb->tcp_flags);
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))) {
993 		th->window      = htons(tcp_select_window(sk));
994 		tcp_ecn_send(sk, skb, th, tcp_header_size);
995 	} else {
996 		/* RFC1323: The window in SYN & SYN/ACK segments
997 		 * is never scaled.
998 		 */
999 		th->window	= htons(min(tp->rcv_wnd, 65535U));
1000 	}
1001 #ifdef CONFIG_TCP_MD5SIG
1002 	/* Calculate the MD5 hash, as we have all we need now */
1003 	if (md5) {
1004 		sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1005 		tp->af_specific->calc_md5_hash(opts.hash_location,
1006 					       md5, sk, skb);
1007 	}
1008 #endif
1009 
1010 	icsk->icsk_af_ops->send_check(sk, skb);
1011 
1012 	if (likely(tcb->tcp_flags & TCPHDR_ACK))
1013 		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1014 
1015 	if (skb->len != tcp_header_size) {
1016 		tcp_event_data_sent(tp, sk);
1017 		tp->data_segs_out += tcp_skb_pcount(skb);
1018 	}
1019 
1020 	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1021 		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1022 			      tcp_skb_pcount(skb));
1023 
1024 	tp->segs_out += tcp_skb_pcount(skb);
1025 	/* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1026 	skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1027 	skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1028 
1029 	/* Our usage of tstamp should remain private */
1030 	skb->tstamp.tv64 = 0;
1031 
1032 	/* Cleanup our debris for IP stacks */
1033 	memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1034 			       sizeof(struct inet6_skb_parm)));
1035 
1036 	err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1037 
1038 	if (likely(err <= 0))
1039 		return err;
1040 
1041 	tcp_enter_cwr(sk);
1042 
1043 	return net_xmit_eval(err);
1044 }
1045 
1046 /* This routine just queues the buffer for sending.
1047  *
1048  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1049  * otherwise socket can stall.
1050  */
1051 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1052 {
1053 	struct tcp_sock *tp = tcp_sk(sk);
1054 
1055 	/* Advance write_seq and place onto the write_queue. */
1056 	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1057 	__skb_header_release(skb);
1058 	tcp_add_write_queue_tail(sk, skb);
1059 	sk->sk_wmem_queued += skb->truesize;
1060 	sk_mem_charge(sk, skb->truesize);
1061 }
1062 
1063 /* Initialize TSO segments for a packet. */
1064 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1065 {
1066 	if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1067 		/* Avoid the costly divide in the normal
1068 		 * non-TSO case.
1069 		 */
1070 		tcp_skb_pcount_set(skb, 1);
1071 		TCP_SKB_CB(skb)->tcp_gso_size = 0;
1072 	} else {
1073 		tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1074 		TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1075 	}
1076 }
1077 
1078 /* When a modification to fackets out becomes necessary, we need to check
1079  * skb is counted to fackets_out or not.
1080  */
1081 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1082 				   int decr)
1083 {
1084 	struct tcp_sock *tp = tcp_sk(sk);
1085 
1086 	if (!tp->sacked_out || tcp_is_reno(tp))
1087 		return;
1088 
1089 	if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1090 		tp->fackets_out -= decr;
1091 }
1092 
1093 /* Pcount in the middle of the write queue got changed, we need to do various
1094  * tweaks to fix counters
1095  */
1096 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1097 {
1098 	struct tcp_sock *tp = tcp_sk(sk);
1099 
1100 	tp->packets_out -= decr;
1101 
1102 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1103 		tp->sacked_out -= decr;
1104 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1105 		tp->retrans_out -= decr;
1106 	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1107 		tp->lost_out -= decr;
1108 
1109 	/* Reno case is special. Sigh... */
1110 	if (tcp_is_reno(tp) && decr > 0)
1111 		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1112 
1113 	tcp_adjust_fackets_out(sk, skb, decr);
1114 
1115 	if (tp->lost_skb_hint &&
1116 	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1117 	    (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1118 		tp->lost_cnt_hint -= decr;
1119 
1120 	tcp_verify_left_out(tp);
1121 }
1122 
1123 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1124 {
1125 	return TCP_SKB_CB(skb)->txstamp_ack ||
1126 		(skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1127 }
1128 
1129 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1130 {
1131 	struct skb_shared_info *shinfo = skb_shinfo(skb);
1132 
1133 	if (unlikely(tcp_has_tx_tstamp(skb)) &&
1134 	    !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1135 		struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1136 		u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1137 
1138 		shinfo->tx_flags &= ~tsflags;
1139 		shinfo2->tx_flags |= tsflags;
1140 		swap(shinfo->tskey, shinfo2->tskey);
1141 		TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1142 		TCP_SKB_CB(skb)->txstamp_ack = 0;
1143 	}
1144 }
1145 
1146 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1147 {
1148 	TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1149 	TCP_SKB_CB(skb)->eor = 0;
1150 }
1151 
1152 /* Function to create two new TCP segments.  Shrinks the given segment
1153  * to the specified size and appends a new segment with the rest of the
1154  * packet to the list.  This won't be called frequently, I hope.
1155  * Remember, these are still headerless SKBs at this point.
1156  */
1157 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1158 		 unsigned int mss_now, gfp_t gfp)
1159 {
1160 	struct tcp_sock *tp = tcp_sk(sk);
1161 	struct sk_buff *buff;
1162 	int nsize, old_factor;
1163 	int nlen;
1164 	u8 flags;
1165 
1166 	if (WARN_ON(len > skb->len))
1167 		return -EINVAL;
1168 
1169 	nsize = skb_headlen(skb) - len;
1170 	if (nsize < 0)
1171 		nsize = 0;
1172 
1173 	if (skb_unclone(skb, gfp))
1174 		return -ENOMEM;
1175 
1176 	/* Get a new skb... force flag on. */
1177 	buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1178 	if (!buff)
1179 		return -ENOMEM; /* We'll just try again later. */
1180 
1181 	sk->sk_wmem_queued += buff->truesize;
1182 	sk_mem_charge(sk, buff->truesize);
1183 	nlen = skb->len - len - nsize;
1184 	buff->truesize += nlen;
1185 	skb->truesize -= nlen;
1186 
1187 	/* Correct the sequence numbers. */
1188 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1189 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1190 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1191 
1192 	/* PSH and FIN should only be set in the second packet. */
1193 	flags = TCP_SKB_CB(skb)->tcp_flags;
1194 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1195 	TCP_SKB_CB(buff)->tcp_flags = flags;
1196 	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1197 	tcp_skb_fragment_eor(skb, buff);
1198 
1199 	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1200 		/* Copy and checksum data tail into the new buffer. */
1201 		buff->csum = csum_partial_copy_nocheck(skb->data + len,
1202 						       skb_put(buff, nsize),
1203 						       nsize, 0);
1204 
1205 		skb_trim(skb, len);
1206 
1207 		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1208 	} else {
1209 		skb->ip_summed = CHECKSUM_PARTIAL;
1210 		skb_split(skb, buff, len);
1211 	}
1212 
1213 	buff->ip_summed = skb->ip_summed;
1214 
1215 	buff->tstamp = skb->tstamp;
1216 	tcp_fragment_tstamp(skb, buff);
1217 
1218 	old_factor = tcp_skb_pcount(skb);
1219 
1220 	/* Fix up tso_factor for both original and new SKB.  */
1221 	tcp_set_skb_tso_segs(skb, mss_now);
1222 	tcp_set_skb_tso_segs(buff, mss_now);
1223 
1224 	/* Update delivered info for the new segment */
1225 	TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1226 
1227 	/* If this packet has been sent out already, we must
1228 	 * adjust the various packet counters.
1229 	 */
1230 	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1231 		int diff = old_factor - tcp_skb_pcount(skb) -
1232 			tcp_skb_pcount(buff);
1233 
1234 		if (diff)
1235 			tcp_adjust_pcount(sk, skb, diff);
1236 	}
1237 
1238 	/* Link BUFF into the send queue. */
1239 	__skb_header_release(buff);
1240 	tcp_insert_write_queue_after(skb, buff, sk);
1241 
1242 	return 0;
1243 }
1244 
1245 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1246  * eventually). The difference is that pulled data not copied, but
1247  * immediately discarded.
1248  */
1249 static void __pskb_trim_head(struct sk_buff *skb, int len)
1250 {
1251 	struct skb_shared_info *shinfo;
1252 	int i, k, eat;
1253 
1254 	eat = min_t(int, len, skb_headlen(skb));
1255 	if (eat) {
1256 		__skb_pull(skb, eat);
1257 		len -= eat;
1258 		if (!len)
1259 			return;
1260 	}
1261 	eat = len;
1262 	k = 0;
1263 	shinfo = skb_shinfo(skb);
1264 	for (i = 0; i < shinfo->nr_frags; i++) {
1265 		int size = skb_frag_size(&shinfo->frags[i]);
1266 
1267 		if (size <= eat) {
1268 			skb_frag_unref(skb, i);
1269 			eat -= size;
1270 		} else {
1271 			shinfo->frags[k] = shinfo->frags[i];
1272 			if (eat) {
1273 				shinfo->frags[k].page_offset += eat;
1274 				skb_frag_size_sub(&shinfo->frags[k], eat);
1275 				eat = 0;
1276 			}
1277 			k++;
1278 		}
1279 	}
1280 	shinfo->nr_frags = k;
1281 
1282 	skb_reset_tail_pointer(skb);
1283 	skb->data_len -= len;
1284 	skb->len = skb->data_len;
1285 }
1286 
1287 /* Remove acked data from a packet in the transmit queue. */
1288 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1289 {
1290 	if (skb_unclone(skb, GFP_ATOMIC))
1291 		return -ENOMEM;
1292 
1293 	__pskb_trim_head(skb, len);
1294 
1295 	TCP_SKB_CB(skb)->seq += len;
1296 	skb->ip_summed = CHECKSUM_PARTIAL;
1297 
1298 	skb->truesize	     -= len;
1299 	sk->sk_wmem_queued   -= len;
1300 	sk_mem_uncharge(sk, len);
1301 	sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1302 
1303 	/* Any change of skb->len requires recalculation of tso factor. */
1304 	if (tcp_skb_pcount(skb) > 1)
1305 		tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1306 
1307 	return 0;
1308 }
1309 
1310 /* Calculate MSS not accounting any TCP options.  */
1311 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1312 {
1313 	const struct tcp_sock *tp = tcp_sk(sk);
1314 	const struct inet_connection_sock *icsk = inet_csk(sk);
1315 	int mss_now;
1316 
1317 	/* Calculate base mss without TCP options:
1318 	   It is MMS_S - sizeof(tcphdr) of rfc1122
1319 	 */
1320 	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1321 
1322 	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1323 	if (icsk->icsk_af_ops->net_frag_header_len) {
1324 		const struct dst_entry *dst = __sk_dst_get(sk);
1325 
1326 		if (dst && dst_allfrag(dst))
1327 			mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1328 	}
1329 
1330 	/* Clamp it (mss_clamp does not include tcp options) */
1331 	if (mss_now > tp->rx_opt.mss_clamp)
1332 		mss_now = tp->rx_opt.mss_clamp;
1333 
1334 	/* Now subtract optional transport overhead */
1335 	mss_now -= icsk->icsk_ext_hdr_len;
1336 
1337 	/* Then reserve room for full set of TCP options and 8 bytes of data */
1338 	if (mss_now < 48)
1339 		mss_now = 48;
1340 	return mss_now;
1341 }
1342 
1343 /* Calculate MSS. Not accounting for SACKs here.  */
1344 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1345 {
1346 	/* Subtract TCP options size, not including SACKs */
1347 	return __tcp_mtu_to_mss(sk, pmtu) -
1348 	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1349 }
1350 
1351 /* Inverse of above */
1352 int tcp_mss_to_mtu(struct sock *sk, int mss)
1353 {
1354 	const struct tcp_sock *tp = tcp_sk(sk);
1355 	const struct inet_connection_sock *icsk = inet_csk(sk);
1356 	int mtu;
1357 
1358 	mtu = mss +
1359 	      tp->tcp_header_len +
1360 	      icsk->icsk_ext_hdr_len +
1361 	      icsk->icsk_af_ops->net_header_len;
1362 
1363 	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1364 	if (icsk->icsk_af_ops->net_frag_header_len) {
1365 		const struct dst_entry *dst = __sk_dst_get(sk);
1366 
1367 		if (dst && dst_allfrag(dst))
1368 			mtu += icsk->icsk_af_ops->net_frag_header_len;
1369 	}
1370 	return mtu;
1371 }
1372 EXPORT_SYMBOL(tcp_mss_to_mtu);
1373 
1374 /* MTU probing init per socket */
1375 void tcp_mtup_init(struct sock *sk)
1376 {
1377 	struct tcp_sock *tp = tcp_sk(sk);
1378 	struct inet_connection_sock *icsk = inet_csk(sk);
1379 	struct net *net = sock_net(sk);
1380 
1381 	icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1382 	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1383 			       icsk->icsk_af_ops->net_header_len;
1384 	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1385 	icsk->icsk_mtup.probe_size = 0;
1386 	if (icsk->icsk_mtup.enabled)
1387 		icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1388 }
1389 EXPORT_SYMBOL(tcp_mtup_init);
1390 
1391 /* This function synchronize snd mss to current pmtu/exthdr set.
1392 
1393    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1394    for TCP options, but includes only bare TCP header.
1395 
1396    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1397    It is minimum of user_mss and mss received with SYN.
1398    It also does not include TCP options.
1399 
1400    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1401 
1402    tp->mss_cache is current effective sending mss, including
1403    all tcp options except for SACKs. It is evaluated,
1404    taking into account current pmtu, but never exceeds
1405    tp->rx_opt.mss_clamp.
1406 
1407    NOTE1. rfc1122 clearly states that advertised MSS
1408    DOES NOT include either tcp or ip options.
1409 
1410    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1411    are READ ONLY outside this function.		--ANK (980731)
1412  */
1413 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1414 {
1415 	struct tcp_sock *tp = tcp_sk(sk);
1416 	struct inet_connection_sock *icsk = inet_csk(sk);
1417 	int mss_now;
1418 
1419 	if (icsk->icsk_mtup.search_high > pmtu)
1420 		icsk->icsk_mtup.search_high = pmtu;
1421 
1422 	mss_now = tcp_mtu_to_mss(sk, pmtu);
1423 	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1424 
1425 	/* And store cached results */
1426 	icsk->icsk_pmtu_cookie = pmtu;
1427 	if (icsk->icsk_mtup.enabled)
1428 		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1429 	tp->mss_cache = mss_now;
1430 
1431 	return mss_now;
1432 }
1433 EXPORT_SYMBOL(tcp_sync_mss);
1434 
1435 /* Compute the current effective MSS, taking SACKs and IP options,
1436  * and even PMTU discovery events into account.
1437  */
1438 unsigned int tcp_current_mss(struct sock *sk)
1439 {
1440 	const struct tcp_sock *tp = tcp_sk(sk);
1441 	const struct dst_entry *dst = __sk_dst_get(sk);
1442 	u32 mss_now;
1443 	unsigned int header_len;
1444 	struct tcp_out_options opts;
1445 	struct tcp_md5sig_key *md5;
1446 
1447 	mss_now = tp->mss_cache;
1448 
1449 	if (dst) {
1450 		u32 mtu = dst_mtu(dst);
1451 		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1452 			mss_now = tcp_sync_mss(sk, mtu);
1453 	}
1454 
1455 	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1456 		     sizeof(struct tcphdr);
1457 	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1458 	 * some common options. If this is an odd packet (because we have SACK
1459 	 * blocks etc) then our calculated header_len will be different, and
1460 	 * we have to adjust mss_now correspondingly */
1461 	if (header_len != tp->tcp_header_len) {
1462 		int delta = (int) header_len - tp->tcp_header_len;
1463 		mss_now -= delta;
1464 	}
1465 
1466 	return mss_now;
1467 }
1468 
1469 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1470  * As additional protections, we do not touch cwnd in retransmission phases,
1471  * and if application hit its sndbuf limit recently.
1472  */
1473 static void tcp_cwnd_application_limited(struct sock *sk)
1474 {
1475 	struct tcp_sock *tp = tcp_sk(sk);
1476 
1477 	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1478 	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1479 		/* Limited by application or receiver window. */
1480 		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1481 		u32 win_used = max(tp->snd_cwnd_used, init_win);
1482 		if (win_used < tp->snd_cwnd) {
1483 			tp->snd_ssthresh = tcp_current_ssthresh(sk);
1484 			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1485 		}
1486 		tp->snd_cwnd_used = 0;
1487 	}
1488 	tp->snd_cwnd_stamp = tcp_time_stamp;
1489 }
1490 
1491 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1492 {
1493 	struct tcp_sock *tp = tcp_sk(sk);
1494 
1495 	/* Track the maximum number of outstanding packets in each
1496 	 * window, and remember whether we were cwnd-limited then.
1497 	 */
1498 	if (!before(tp->snd_una, tp->max_packets_seq) ||
1499 	    tp->packets_out > tp->max_packets_out) {
1500 		tp->max_packets_out = tp->packets_out;
1501 		tp->max_packets_seq = tp->snd_nxt;
1502 		tp->is_cwnd_limited = is_cwnd_limited;
1503 	}
1504 
1505 	if (tcp_is_cwnd_limited(sk)) {
1506 		/* Network is feed fully. */
1507 		tp->snd_cwnd_used = 0;
1508 		tp->snd_cwnd_stamp = tcp_time_stamp;
1509 	} else {
1510 		/* Network starves. */
1511 		if (tp->packets_out > tp->snd_cwnd_used)
1512 			tp->snd_cwnd_used = tp->packets_out;
1513 
1514 		if (sysctl_tcp_slow_start_after_idle &&
1515 		    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1516 			tcp_cwnd_application_limited(sk);
1517 	}
1518 }
1519 
1520 /* Minshall's variant of the Nagle send check. */
1521 static bool tcp_minshall_check(const struct tcp_sock *tp)
1522 {
1523 	return after(tp->snd_sml, tp->snd_una) &&
1524 		!after(tp->snd_sml, tp->snd_nxt);
1525 }
1526 
1527 /* Update snd_sml if this skb is under mss
1528  * Note that a TSO packet might end with a sub-mss segment
1529  * The test is really :
1530  * if ((skb->len % mss) != 0)
1531  *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1532  * But we can avoid doing the divide again given we already have
1533  *  skb_pcount = skb->len / mss_now
1534  */
1535 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1536 				const struct sk_buff *skb)
1537 {
1538 	if (skb->len < tcp_skb_pcount(skb) * mss_now)
1539 		tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1540 }
1541 
1542 /* Return false, if packet can be sent now without violation Nagle's rules:
1543  * 1. It is full sized. (provided by caller in %partial bool)
1544  * 2. Or it contains FIN. (already checked by caller)
1545  * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1546  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1547  *    With Minshall's modification: all sent small packets are ACKed.
1548  */
1549 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1550 			    int nonagle)
1551 {
1552 	return partial &&
1553 		((nonagle & TCP_NAGLE_CORK) ||
1554 		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1555 }
1556 
1557 /* Return how many segs we'd like on a TSO packet,
1558  * to send one TSO packet per ms
1559  */
1560 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1561 		     int min_tso_segs)
1562 {
1563 	u32 bytes, segs;
1564 
1565 	bytes = min(sk->sk_pacing_rate >> 10,
1566 		    sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1567 
1568 	/* Goal is to send at least one packet per ms,
1569 	 * not one big TSO packet every 100 ms.
1570 	 * This preserves ACK clocking and is consistent
1571 	 * with tcp_tso_should_defer() heuristic.
1572 	 */
1573 	segs = max_t(u32, bytes / mss_now, min_tso_segs);
1574 
1575 	return min_t(u32, segs, sk->sk_gso_max_segs);
1576 }
1577 EXPORT_SYMBOL(tcp_tso_autosize);
1578 
1579 /* Return the number of segments we want in the skb we are transmitting.
1580  * See if congestion control module wants to decide; otherwise, autosize.
1581  */
1582 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1583 {
1584 	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1585 	u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1586 
1587 	return tso_segs ? :
1588 		tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs);
1589 }
1590 
1591 /* Returns the portion of skb which can be sent right away */
1592 static unsigned int tcp_mss_split_point(const struct sock *sk,
1593 					const struct sk_buff *skb,
1594 					unsigned int mss_now,
1595 					unsigned int max_segs,
1596 					int nonagle)
1597 {
1598 	const struct tcp_sock *tp = tcp_sk(sk);
1599 	u32 partial, needed, window, max_len;
1600 
1601 	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1602 	max_len = mss_now * max_segs;
1603 
1604 	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1605 		return max_len;
1606 
1607 	needed = min(skb->len, window);
1608 
1609 	if (max_len <= needed)
1610 		return max_len;
1611 
1612 	partial = needed % mss_now;
1613 	/* If last segment is not a full MSS, check if Nagle rules allow us
1614 	 * to include this last segment in this skb.
1615 	 * Otherwise, we'll split the skb at last MSS boundary
1616 	 */
1617 	if (tcp_nagle_check(partial != 0, tp, nonagle))
1618 		return needed - partial;
1619 
1620 	return needed;
1621 }
1622 
1623 /* Can at least one segment of SKB be sent right now, according to the
1624  * congestion window rules?  If so, return how many segments are allowed.
1625  */
1626 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1627 					 const struct sk_buff *skb)
1628 {
1629 	u32 in_flight, cwnd, halfcwnd;
1630 
1631 	/* Don't be strict about the congestion window for the final FIN.  */
1632 	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1633 	    tcp_skb_pcount(skb) == 1)
1634 		return 1;
1635 
1636 	in_flight = tcp_packets_in_flight(tp);
1637 	cwnd = tp->snd_cwnd;
1638 	if (in_flight >= cwnd)
1639 		return 0;
1640 
1641 	/* For better scheduling, ensure we have at least
1642 	 * 2 GSO packets in flight.
1643 	 */
1644 	halfcwnd = max(cwnd >> 1, 1U);
1645 	return min(halfcwnd, cwnd - in_flight);
1646 }
1647 
1648 /* Initialize TSO state of a skb.
1649  * This must be invoked the first time we consider transmitting
1650  * SKB onto the wire.
1651  */
1652 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1653 {
1654 	int tso_segs = tcp_skb_pcount(skb);
1655 
1656 	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1657 		tcp_set_skb_tso_segs(skb, mss_now);
1658 		tso_segs = tcp_skb_pcount(skb);
1659 	}
1660 	return tso_segs;
1661 }
1662 
1663 
1664 /* Return true if the Nagle test allows this packet to be
1665  * sent now.
1666  */
1667 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1668 				  unsigned int cur_mss, int nonagle)
1669 {
1670 	/* Nagle rule does not apply to frames, which sit in the middle of the
1671 	 * write_queue (they have no chances to get new data).
1672 	 *
1673 	 * This is implemented in the callers, where they modify the 'nonagle'
1674 	 * argument based upon the location of SKB in the send queue.
1675 	 */
1676 	if (nonagle & TCP_NAGLE_PUSH)
1677 		return true;
1678 
1679 	/* Don't use the nagle rule for urgent data (or for the final FIN). */
1680 	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1681 		return true;
1682 
1683 	if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1684 		return true;
1685 
1686 	return false;
1687 }
1688 
1689 /* Does at least the first segment of SKB fit into the send window? */
1690 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1691 			     const struct sk_buff *skb,
1692 			     unsigned int cur_mss)
1693 {
1694 	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1695 
1696 	if (skb->len > cur_mss)
1697 		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1698 
1699 	return !after(end_seq, tcp_wnd_end(tp));
1700 }
1701 
1702 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1703  * should be put on the wire right now.  If so, it returns the number of
1704  * packets allowed by the congestion window.
1705  */
1706 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1707 				 unsigned int cur_mss, int nonagle)
1708 {
1709 	const struct tcp_sock *tp = tcp_sk(sk);
1710 	unsigned int cwnd_quota;
1711 
1712 	tcp_init_tso_segs(skb, cur_mss);
1713 
1714 	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1715 		return 0;
1716 
1717 	cwnd_quota = tcp_cwnd_test(tp, skb);
1718 	if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1719 		cwnd_quota = 0;
1720 
1721 	return cwnd_quota;
1722 }
1723 
1724 /* Test if sending is allowed right now. */
1725 bool tcp_may_send_now(struct sock *sk)
1726 {
1727 	const struct tcp_sock *tp = tcp_sk(sk);
1728 	struct sk_buff *skb = tcp_send_head(sk);
1729 
1730 	return skb &&
1731 		tcp_snd_test(sk, skb, tcp_current_mss(sk),
1732 			     (tcp_skb_is_last(sk, skb) ?
1733 			      tp->nonagle : TCP_NAGLE_PUSH));
1734 }
1735 
1736 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1737  * which is put after SKB on the list.  It is very much like
1738  * tcp_fragment() except that it may make several kinds of assumptions
1739  * in order to speed up the splitting operation.  In particular, we
1740  * know that all the data is in scatter-gather pages, and that the
1741  * packet has never been sent out before (and thus is not cloned).
1742  */
1743 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1744 			unsigned int mss_now, gfp_t gfp)
1745 {
1746 	struct sk_buff *buff;
1747 	int nlen = skb->len - len;
1748 	u8 flags;
1749 
1750 	/* All of a TSO frame must be composed of paged data.  */
1751 	if (skb->len != skb->data_len)
1752 		return tcp_fragment(sk, skb, len, mss_now, gfp);
1753 
1754 	buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1755 	if (unlikely(!buff))
1756 		return -ENOMEM;
1757 
1758 	sk->sk_wmem_queued += buff->truesize;
1759 	sk_mem_charge(sk, buff->truesize);
1760 	buff->truesize += nlen;
1761 	skb->truesize -= nlen;
1762 
1763 	/* Correct the sequence numbers. */
1764 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1765 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1766 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1767 
1768 	/* PSH and FIN should only be set in the second packet. */
1769 	flags = TCP_SKB_CB(skb)->tcp_flags;
1770 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1771 	TCP_SKB_CB(buff)->tcp_flags = flags;
1772 
1773 	/* This packet was never sent out yet, so no SACK bits. */
1774 	TCP_SKB_CB(buff)->sacked = 0;
1775 
1776 	tcp_skb_fragment_eor(skb, buff);
1777 
1778 	buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1779 	skb_split(skb, buff, len);
1780 	tcp_fragment_tstamp(skb, buff);
1781 
1782 	/* Fix up tso_factor for both original and new SKB.  */
1783 	tcp_set_skb_tso_segs(skb, mss_now);
1784 	tcp_set_skb_tso_segs(buff, mss_now);
1785 
1786 	/* Link BUFF into the send queue. */
1787 	__skb_header_release(buff);
1788 	tcp_insert_write_queue_after(skb, buff, sk);
1789 
1790 	return 0;
1791 }
1792 
1793 /* Try to defer sending, if possible, in order to minimize the amount
1794  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1795  *
1796  * This algorithm is from John Heffner.
1797  */
1798 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1799 				 bool *is_cwnd_limited, u32 max_segs)
1800 {
1801 	const struct inet_connection_sock *icsk = inet_csk(sk);
1802 	u32 age, send_win, cong_win, limit, in_flight;
1803 	struct tcp_sock *tp = tcp_sk(sk);
1804 	struct skb_mstamp now;
1805 	struct sk_buff *head;
1806 	int win_divisor;
1807 
1808 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1809 		goto send_now;
1810 
1811 	if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1812 		goto send_now;
1813 
1814 	/* Avoid bursty behavior by allowing defer
1815 	 * only if the last write was recent.
1816 	 */
1817 	if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1818 		goto send_now;
1819 
1820 	in_flight = tcp_packets_in_flight(tp);
1821 
1822 	BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1823 
1824 	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1825 
1826 	/* From in_flight test above, we know that cwnd > in_flight.  */
1827 	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1828 
1829 	limit = min(send_win, cong_win);
1830 
1831 	/* If a full-sized TSO skb can be sent, do it. */
1832 	if (limit >= max_segs * tp->mss_cache)
1833 		goto send_now;
1834 
1835 	/* Middle in queue won't get any more data, full sendable already? */
1836 	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1837 		goto send_now;
1838 
1839 	win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1840 	if (win_divisor) {
1841 		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1842 
1843 		/* If at least some fraction of a window is available,
1844 		 * just use it.
1845 		 */
1846 		chunk /= win_divisor;
1847 		if (limit >= chunk)
1848 			goto send_now;
1849 	} else {
1850 		/* Different approach, try not to defer past a single
1851 		 * ACK.  Receiver should ACK every other full sized
1852 		 * frame, so if we have space for more than 3 frames
1853 		 * then send now.
1854 		 */
1855 		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1856 			goto send_now;
1857 	}
1858 
1859 	head = tcp_write_queue_head(sk);
1860 	skb_mstamp_get(&now);
1861 	age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1862 	/* If next ACK is likely to come too late (half srtt), do not defer */
1863 	if (age < (tp->srtt_us >> 4))
1864 		goto send_now;
1865 
1866 	/* Ok, it looks like it is advisable to defer. */
1867 
1868 	if (cong_win < send_win && cong_win <= skb->len)
1869 		*is_cwnd_limited = true;
1870 
1871 	return true;
1872 
1873 send_now:
1874 	return false;
1875 }
1876 
1877 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1878 {
1879 	struct inet_connection_sock *icsk = inet_csk(sk);
1880 	struct tcp_sock *tp = tcp_sk(sk);
1881 	struct net *net = sock_net(sk);
1882 	u32 interval;
1883 	s32 delta;
1884 
1885 	interval = net->ipv4.sysctl_tcp_probe_interval;
1886 	delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1887 	if (unlikely(delta >= interval * HZ)) {
1888 		int mss = tcp_current_mss(sk);
1889 
1890 		/* Update current search range */
1891 		icsk->icsk_mtup.probe_size = 0;
1892 		icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1893 			sizeof(struct tcphdr) +
1894 			icsk->icsk_af_ops->net_header_len;
1895 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1896 
1897 		/* Update probe time stamp */
1898 		icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1899 	}
1900 }
1901 
1902 /* Create a new MTU probe if we are ready.
1903  * MTU probe is regularly attempting to increase the path MTU by
1904  * deliberately sending larger packets.  This discovers routing
1905  * changes resulting in larger path MTUs.
1906  *
1907  * Returns 0 if we should wait to probe (no cwnd available),
1908  *         1 if a probe was sent,
1909  *         -1 otherwise
1910  */
1911 static int tcp_mtu_probe(struct sock *sk)
1912 {
1913 	struct tcp_sock *tp = tcp_sk(sk);
1914 	struct inet_connection_sock *icsk = inet_csk(sk);
1915 	struct sk_buff *skb, *nskb, *next;
1916 	struct net *net = sock_net(sk);
1917 	int len;
1918 	int probe_size;
1919 	int size_needed;
1920 	int copy;
1921 	int mss_now;
1922 	int interval;
1923 
1924 	/* Not currently probing/verifying,
1925 	 * not in recovery,
1926 	 * have enough cwnd, and
1927 	 * not SACKing (the variable headers throw things off) */
1928 	if (!icsk->icsk_mtup.enabled ||
1929 	    icsk->icsk_mtup.probe_size ||
1930 	    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1931 	    tp->snd_cwnd < 11 ||
1932 	    tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1933 		return -1;
1934 
1935 	/* Use binary search for probe_size between tcp_mss_base,
1936 	 * and current mss_clamp. if (search_high - search_low)
1937 	 * smaller than a threshold, backoff from probing.
1938 	 */
1939 	mss_now = tcp_current_mss(sk);
1940 	probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1941 				    icsk->icsk_mtup.search_low) >> 1);
1942 	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1943 	interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1944 	/* When misfortune happens, we are reprobing actively,
1945 	 * and then reprobe timer has expired. We stick with current
1946 	 * probing process by not resetting search range to its orignal.
1947 	 */
1948 	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1949 		interval < net->ipv4.sysctl_tcp_probe_threshold) {
1950 		/* Check whether enough time has elaplased for
1951 		 * another round of probing.
1952 		 */
1953 		tcp_mtu_check_reprobe(sk);
1954 		return -1;
1955 	}
1956 
1957 	/* Have enough data in the send queue to probe? */
1958 	if (tp->write_seq - tp->snd_nxt < size_needed)
1959 		return -1;
1960 
1961 	if (tp->snd_wnd < size_needed)
1962 		return -1;
1963 	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1964 		return 0;
1965 
1966 	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
1967 	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1968 		if (!tcp_packets_in_flight(tp))
1969 			return -1;
1970 		else
1971 			return 0;
1972 	}
1973 
1974 	/* We're allowed to probe.  Build it now. */
1975 	nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
1976 	if (!nskb)
1977 		return -1;
1978 	sk->sk_wmem_queued += nskb->truesize;
1979 	sk_mem_charge(sk, nskb->truesize);
1980 
1981 	skb = tcp_send_head(sk);
1982 
1983 	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1984 	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1985 	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1986 	TCP_SKB_CB(nskb)->sacked = 0;
1987 	nskb->csum = 0;
1988 	nskb->ip_summed = skb->ip_summed;
1989 
1990 	tcp_insert_write_queue_before(nskb, skb, sk);
1991 
1992 	len = 0;
1993 	tcp_for_write_queue_from_safe(skb, next, sk) {
1994 		copy = min_t(int, skb->len, probe_size - len);
1995 		if (nskb->ip_summed) {
1996 			skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1997 		} else {
1998 			__wsum csum = skb_copy_and_csum_bits(skb, 0,
1999 							     skb_put(nskb, copy),
2000 							     copy, 0);
2001 			nskb->csum = csum_block_add(nskb->csum, csum, len);
2002 		}
2003 
2004 		if (skb->len <= copy) {
2005 			/* We've eaten all the data from this skb.
2006 			 * Throw it away. */
2007 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2008 			tcp_unlink_write_queue(skb, sk);
2009 			sk_wmem_free_skb(sk, skb);
2010 		} else {
2011 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2012 						   ~(TCPHDR_FIN|TCPHDR_PSH);
2013 			if (!skb_shinfo(skb)->nr_frags) {
2014 				skb_pull(skb, copy);
2015 				if (skb->ip_summed != CHECKSUM_PARTIAL)
2016 					skb->csum = csum_partial(skb->data,
2017 								 skb->len, 0);
2018 			} else {
2019 				__pskb_trim_head(skb, copy);
2020 				tcp_set_skb_tso_segs(skb, mss_now);
2021 			}
2022 			TCP_SKB_CB(skb)->seq += copy;
2023 		}
2024 
2025 		len += copy;
2026 
2027 		if (len >= probe_size)
2028 			break;
2029 	}
2030 	tcp_init_tso_segs(nskb, nskb->len);
2031 
2032 	/* We're ready to send.  If this fails, the probe will
2033 	 * be resegmented into mss-sized pieces by tcp_write_xmit().
2034 	 */
2035 	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2036 		/* Decrement cwnd here because we are sending
2037 		 * effectively two packets. */
2038 		tp->snd_cwnd--;
2039 		tcp_event_new_data_sent(sk, nskb);
2040 
2041 		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2042 		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2043 		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2044 
2045 		return 1;
2046 	}
2047 
2048 	return -1;
2049 }
2050 
2051 /* TCP Small Queues :
2052  * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2053  * (These limits are doubled for retransmits)
2054  * This allows for :
2055  *  - better RTT estimation and ACK scheduling
2056  *  - faster recovery
2057  *  - high rates
2058  * Alas, some drivers / subsystems require a fair amount
2059  * of queued bytes to ensure line rate.
2060  * One example is wifi aggregation (802.11 AMPDU)
2061  */
2062 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2063 				  unsigned int factor)
2064 {
2065 	unsigned int limit;
2066 
2067 	limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2068 	limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2069 	limit <<= factor;
2070 
2071 	if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2072 		set_bit(TSQ_THROTTLED, &tcp_sk(sk)->tsq_flags);
2073 		/* It is possible TX completion already happened
2074 		 * before we set TSQ_THROTTLED, so we must
2075 		 * test again the condition.
2076 		 */
2077 		smp_mb__after_atomic();
2078 		if (atomic_read(&sk->sk_wmem_alloc) > limit)
2079 			return true;
2080 	}
2081 	return false;
2082 }
2083 
2084 /* This routine writes packets to the network.  It advances the
2085  * send_head.  This happens as incoming acks open up the remote
2086  * window for us.
2087  *
2088  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2089  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2090  * account rare use of URG, this is not a big flaw.
2091  *
2092  * Send at most one packet when push_one > 0. Temporarily ignore
2093  * cwnd limit to force at most one packet out when push_one == 2.
2094 
2095  * Returns true, if no segments are in flight and we have queued segments,
2096  * but cannot send anything now because of SWS or another problem.
2097  */
2098 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2099 			   int push_one, gfp_t gfp)
2100 {
2101 	struct tcp_sock *tp = tcp_sk(sk);
2102 	struct sk_buff *skb;
2103 	unsigned int tso_segs, sent_pkts;
2104 	int cwnd_quota;
2105 	int result;
2106 	bool is_cwnd_limited = false;
2107 	u32 max_segs;
2108 
2109 	sent_pkts = 0;
2110 
2111 	if (!push_one) {
2112 		/* Do MTU probing. */
2113 		result = tcp_mtu_probe(sk);
2114 		if (!result) {
2115 			return false;
2116 		} else if (result > 0) {
2117 			sent_pkts = 1;
2118 		}
2119 	}
2120 
2121 	max_segs = tcp_tso_segs(sk, mss_now);
2122 	while ((skb = tcp_send_head(sk))) {
2123 		unsigned int limit;
2124 
2125 		tso_segs = tcp_init_tso_segs(skb, mss_now);
2126 		BUG_ON(!tso_segs);
2127 
2128 		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2129 			/* "skb_mstamp" is used as a start point for the retransmit timer */
2130 			skb_mstamp_get(&skb->skb_mstamp);
2131 			goto repair; /* Skip network transmission */
2132 		}
2133 
2134 		cwnd_quota = tcp_cwnd_test(tp, skb);
2135 		if (!cwnd_quota) {
2136 			if (push_one == 2)
2137 				/* Force out a loss probe pkt. */
2138 				cwnd_quota = 1;
2139 			else
2140 				break;
2141 		}
2142 
2143 		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2144 			break;
2145 
2146 		if (tso_segs == 1) {
2147 			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2148 						     (tcp_skb_is_last(sk, skb) ?
2149 						      nonagle : TCP_NAGLE_PUSH))))
2150 				break;
2151 		} else {
2152 			if (!push_one &&
2153 			    tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2154 						 max_segs))
2155 				break;
2156 		}
2157 
2158 		limit = mss_now;
2159 		if (tso_segs > 1 && !tcp_urg_mode(tp))
2160 			limit = tcp_mss_split_point(sk, skb, mss_now,
2161 						    min_t(unsigned int,
2162 							  cwnd_quota,
2163 							  max_segs),
2164 						    nonagle);
2165 
2166 		if (skb->len > limit &&
2167 		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2168 			break;
2169 
2170 		if (tcp_small_queue_check(sk, skb, 0))
2171 			break;
2172 
2173 		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2174 			break;
2175 
2176 repair:
2177 		/* Advance the send_head.  This one is sent out.
2178 		 * This call will increment packets_out.
2179 		 */
2180 		tcp_event_new_data_sent(sk, skb);
2181 
2182 		tcp_minshall_update(tp, mss_now, skb);
2183 		sent_pkts += tcp_skb_pcount(skb);
2184 
2185 		if (push_one)
2186 			break;
2187 	}
2188 
2189 	if (likely(sent_pkts)) {
2190 		if (tcp_in_cwnd_reduction(sk))
2191 			tp->prr_out += sent_pkts;
2192 
2193 		/* Send one loss probe per tail loss episode. */
2194 		if (push_one != 2)
2195 			tcp_schedule_loss_probe(sk);
2196 		is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2197 		tcp_cwnd_validate(sk, is_cwnd_limited);
2198 		return false;
2199 	}
2200 	return !tp->packets_out && tcp_send_head(sk);
2201 }
2202 
2203 bool tcp_schedule_loss_probe(struct sock *sk)
2204 {
2205 	struct inet_connection_sock *icsk = inet_csk(sk);
2206 	struct tcp_sock *tp = tcp_sk(sk);
2207 	u32 timeout, tlp_time_stamp, rto_time_stamp;
2208 	u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2209 
2210 	if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2211 		return false;
2212 	/* No consecutive loss probes. */
2213 	if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2214 		tcp_rearm_rto(sk);
2215 		return false;
2216 	}
2217 	/* Don't do any loss probe on a Fast Open connection before 3WHS
2218 	 * finishes.
2219 	 */
2220 	if (tp->fastopen_rsk)
2221 		return false;
2222 
2223 	/* TLP is only scheduled when next timer event is RTO. */
2224 	if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2225 		return false;
2226 
2227 	/* Schedule a loss probe in 2*RTT for SACK capable connections
2228 	 * in Open state, that are either limited by cwnd or application.
2229 	 */
2230 	if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
2231 	    !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2232 		return false;
2233 
2234 	if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2235 	     tcp_send_head(sk))
2236 		return false;
2237 
2238 	/* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2239 	 * for delayed ack when there's one outstanding packet. If no RTT
2240 	 * sample is available then probe after TCP_TIMEOUT_INIT.
2241 	 */
2242 	timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2243 	if (tp->packets_out == 1)
2244 		timeout = max_t(u32, timeout,
2245 				(rtt + (rtt >> 1) + TCP_DELACK_MAX));
2246 	timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2247 
2248 	/* If RTO is shorter, just schedule TLP in its place. */
2249 	tlp_time_stamp = tcp_time_stamp + timeout;
2250 	rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2251 	if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2252 		s32 delta = rto_time_stamp - tcp_time_stamp;
2253 		if (delta > 0)
2254 			timeout = delta;
2255 	}
2256 
2257 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2258 				  TCP_RTO_MAX);
2259 	return true;
2260 }
2261 
2262 /* Thanks to skb fast clones, we can detect if a prior transmit of
2263  * a packet is still in a qdisc or driver queue.
2264  * In this case, there is very little point doing a retransmit !
2265  */
2266 static bool skb_still_in_host_queue(const struct sock *sk,
2267 				    const struct sk_buff *skb)
2268 {
2269 	if (unlikely(skb_fclone_busy(sk, skb))) {
2270 		NET_INC_STATS(sock_net(sk),
2271 			      LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2272 		return true;
2273 	}
2274 	return false;
2275 }
2276 
2277 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2278  * retransmit the last segment.
2279  */
2280 void tcp_send_loss_probe(struct sock *sk)
2281 {
2282 	struct tcp_sock *tp = tcp_sk(sk);
2283 	struct sk_buff *skb;
2284 	int pcount;
2285 	int mss = tcp_current_mss(sk);
2286 
2287 	skb = tcp_send_head(sk);
2288 	if (skb) {
2289 		if (tcp_snd_wnd_test(tp, skb, mss)) {
2290 			pcount = tp->packets_out;
2291 			tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2292 			if (tp->packets_out > pcount)
2293 				goto probe_sent;
2294 			goto rearm_timer;
2295 		}
2296 		skb = tcp_write_queue_prev(sk, skb);
2297 	} else {
2298 		skb = tcp_write_queue_tail(sk);
2299 	}
2300 
2301 	/* At most one outstanding TLP retransmission. */
2302 	if (tp->tlp_high_seq)
2303 		goto rearm_timer;
2304 
2305 	/* Retransmit last segment. */
2306 	if (WARN_ON(!skb))
2307 		goto rearm_timer;
2308 
2309 	if (skb_still_in_host_queue(sk, skb))
2310 		goto rearm_timer;
2311 
2312 	pcount = tcp_skb_pcount(skb);
2313 	if (WARN_ON(!pcount))
2314 		goto rearm_timer;
2315 
2316 	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2317 		if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2318 					  GFP_ATOMIC)))
2319 			goto rearm_timer;
2320 		skb = tcp_write_queue_next(sk, skb);
2321 	}
2322 
2323 	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2324 		goto rearm_timer;
2325 
2326 	if (__tcp_retransmit_skb(sk, skb, 1))
2327 		goto rearm_timer;
2328 
2329 	/* Record snd_nxt for loss detection. */
2330 	tp->tlp_high_seq = tp->snd_nxt;
2331 
2332 probe_sent:
2333 	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2334 	/* Reset s.t. tcp_rearm_rto will restart timer from now */
2335 	inet_csk(sk)->icsk_pending = 0;
2336 rearm_timer:
2337 	tcp_rearm_rto(sk);
2338 }
2339 
2340 /* Push out any pending frames which were held back due to
2341  * TCP_CORK or attempt at coalescing tiny packets.
2342  * The socket must be locked by the caller.
2343  */
2344 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2345 			       int nonagle)
2346 {
2347 	/* If we are closed, the bytes will have to remain here.
2348 	 * In time closedown will finish, we empty the write queue and
2349 	 * all will be happy.
2350 	 */
2351 	if (unlikely(sk->sk_state == TCP_CLOSE))
2352 		return;
2353 
2354 	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2355 			   sk_gfp_mask(sk, GFP_ATOMIC)))
2356 		tcp_check_probe_timer(sk);
2357 }
2358 
2359 /* Send _single_ skb sitting at the send head. This function requires
2360  * true push pending frames to setup probe timer etc.
2361  */
2362 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2363 {
2364 	struct sk_buff *skb = tcp_send_head(sk);
2365 
2366 	BUG_ON(!skb || skb->len < mss_now);
2367 
2368 	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2369 }
2370 
2371 /* This function returns the amount that we can raise the
2372  * usable window based on the following constraints
2373  *
2374  * 1. The window can never be shrunk once it is offered (RFC 793)
2375  * 2. We limit memory per socket
2376  *
2377  * RFC 1122:
2378  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2379  *  RECV.NEXT + RCV.WIN fixed until:
2380  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2381  *
2382  * i.e. don't raise the right edge of the window until you can raise
2383  * it at least MSS bytes.
2384  *
2385  * Unfortunately, the recommended algorithm breaks header prediction,
2386  * since header prediction assumes th->window stays fixed.
2387  *
2388  * Strictly speaking, keeping th->window fixed violates the receiver
2389  * side SWS prevention criteria. The problem is that under this rule
2390  * a stream of single byte packets will cause the right side of the
2391  * window to always advance by a single byte.
2392  *
2393  * Of course, if the sender implements sender side SWS prevention
2394  * then this will not be a problem.
2395  *
2396  * BSD seems to make the following compromise:
2397  *
2398  *	If the free space is less than the 1/4 of the maximum
2399  *	space available and the free space is less than 1/2 mss,
2400  *	then set the window to 0.
2401  *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2402  *	Otherwise, just prevent the window from shrinking
2403  *	and from being larger than the largest representable value.
2404  *
2405  * This prevents incremental opening of the window in the regime
2406  * where TCP is limited by the speed of the reader side taking
2407  * data out of the TCP receive queue. It does nothing about
2408  * those cases where the window is constrained on the sender side
2409  * because the pipeline is full.
2410  *
2411  * BSD also seems to "accidentally" limit itself to windows that are a
2412  * multiple of MSS, at least until the free space gets quite small.
2413  * This would appear to be a side effect of the mbuf implementation.
2414  * Combining these two algorithms results in the observed behavior
2415  * of having a fixed window size at almost all times.
2416  *
2417  * Below we obtain similar behavior by forcing the offered window to
2418  * a multiple of the mss when it is feasible to do so.
2419  *
2420  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2421  * Regular options like TIMESTAMP are taken into account.
2422  */
2423 u32 __tcp_select_window(struct sock *sk)
2424 {
2425 	struct inet_connection_sock *icsk = inet_csk(sk);
2426 	struct tcp_sock *tp = tcp_sk(sk);
2427 	/* MSS for the peer's data.  Previous versions used mss_clamp
2428 	 * here.  I don't know if the value based on our guesses
2429 	 * of peer's MSS is better for the performance.  It's more correct
2430 	 * but may be worse for the performance because of rcv_mss
2431 	 * fluctuations.  --SAW  1998/11/1
2432 	 */
2433 	int mss = icsk->icsk_ack.rcv_mss;
2434 	int free_space = tcp_space(sk);
2435 	int allowed_space = tcp_full_space(sk);
2436 	int full_space = min_t(int, tp->window_clamp, allowed_space);
2437 	int window;
2438 
2439 	if (mss > full_space)
2440 		mss = full_space;
2441 
2442 	if (free_space < (full_space >> 1)) {
2443 		icsk->icsk_ack.quick = 0;
2444 
2445 		if (tcp_under_memory_pressure(sk))
2446 			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2447 					       4U * tp->advmss);
2448 
2449 		/* free_space might become our new window, make sure we don't
2450 		 * increase it due to wscale.
2451 		 */
2452 		free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2453 
2454 		/* if free space is less than mss estimate, or is below 1/16th
2455 		 * of the maximum allowed, try to move to zero-window, else
2456 		 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2457 		 * new incoming data is dropped due to memory limits.
2458 		 * With large window, mss test triggers way too late in order
2459 		 * to announce zero window in time before rmem limit kicks in.
2460 		 */
2461 		if (free_space < (allowed_space >> 4) || free_space < mss)
2462 			return 0;
2463 	}
2464 
2465 	if (free_space > tp->rcv_ssthresh)
2466 		free_space = tp->rcv_ssthresh;
2467 
2468 	/* Don't do rounding if we are using window scaling, since the
2469 	 * scaled window will not line up with the MSS boundary anyway.
2470 	 */
2471 	window = tp->rcv_wnd;
2472 	if (tp->rx_opt.rcv_wscale) {
2473 		window = free_space;
2474 
2475 		/* Advertise enough space so that it won't get scaled away.
2476 		 * Import case: prevent zero window announcement if
2477 		 * 1<<rcv_wscale > mss.
2478 		 */
2479 		if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2480 			window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2481 				  << tp->rx_opt.rcv_wscale);
2482 	} else {
2483 		/* Get the largest window that is a nice multiple of mss.
2484 		 * Window clamp already applied above.
2485 		 * If our current window offering is within 1 mss of the
2486 		 * free space we just keep it. This prevents the divide
2487 		 * and multiply from happening most of the time.
2488 		 * We also don't do any window rounding when the free space
2489 		 * is too small.
2490 		 */
2491 		if (window <= free_space - mss || window > free_space)
2492 			window = (free_space / mss) * mss;
2493 		else if (mss == full_space &&
2494 			 free_space > window + (full_space >> 1))
2495 			window = free_space;
2496 	}
2497 
2498 	return window;
2499 }
2500 
2501 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2502 			     const struct sk_buff *next_skb)
2503 {
2504 	if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2505 		const struct skb_shared_info *next_shinfo =
2506 			skb_shinfo(next_skb);
2507 		struct skb_shared_info *shinfo = skb_shinfo(skb);
2508 
2509 		shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2510 		shinfo->tskey = next_shinfo->tskey;
2511 		TCP_SKB_CB(skb)->txstamp_ack |=
2512 			TCP_SKB_CB(next_skb)->txstamp_ack;
2513 	}
2514 }
2515 
2516 /* Collapses two adjacent SKB's during retransmission. */
2517 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2518 {
2519 	struct tcp_sock *tp = tcp_sk(sk);
2520 	struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2521 	int skb_size, next_skb_size;
2522 
2523 	skb_size = skb->len;
2524 	next_skb_size = next_skb->len;
2525 
2526 	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2527 
2528 	tcp_highest_sack_combine(sk, next_skb, skb);
2529 
2530 	tcp_unlink_write_queue(next_skb, sk);
2531 
2532 	skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2533 				  next_skb_size);
2534 
2535 	if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2536 		skb->ip_summed = CHECKSUM_PARTIAL;
2537 
2538 	if (skb->ip_summed != CHECKSUM_PARTIAL)
2539 		skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2540 
2541 	/* Update sequence range on original skb. */
2542 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2543 
2544 	/* Merge over control information. This moves PSH/FIN etc. over */
2545 	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2546 
2547 	/* All done, get rid of second SKB and account for it so
2548 	 * packet counting does not break.
2549 	 */
2550 	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2551 	TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2552 
2553 	/* changed transmit queue under us so clear hints */
2554 	tcp_clear_retrans_hints_partial(tp);
2555 	if (next_skb == tp->retransmit_skb_hint)
2556 		tp->retransmit_skb_hint = skb;
2557 
2558 	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2559 
2560 	tcp_skb_collapse_tstamp(skb, next_skb);
2561 
2562 	sk_wmem_free_skb(sk, next_skb);
2563 }
2564 
2565 /* Check if coalescing SKBs is legal. */
2566 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2567 {
2568 	if (tcp_skb_pcount(skb) > 1)
2569 		return false;
2570 	/* TODO: SACK collapsing could be used to remove this condition */
2571 	if (skb_shinfo(skb)->nr_frags != 0)
2572 		return false;
2573 	if (skb_cloned(skb))
2574 		return false;
2575 	if (skb == tcp_send_head(sk))
2576 		return false;
2577 	/* Some heurestics for collapsing over SACK'd could be invented */
2578 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2579 		return false;
2580 
2581 	return true;
2582 }
2583 
2584 /* Collapse packets in the retransmit queue to make to create
2585  * less packets on the wire. This is only done on retransmission.
2586  */
2587 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2588 				     int space)
2589 {
2590 	struct tcp_sock *tp = tcp_sk(sk);
2591 	struct sk_buff *skb = to, *tmp;
2592 	bool first = true;
2593 
2594 	if (!sysctl_tcp_retrans_collapse)
2595 		return;
2596 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2597 		return;
2598 
2599 	tcp_for_write_queue_from_safe(skb, tmp, sk) {
2600 		if (!tcp_can_collapse(sk, skb))
2601 			break;
2602 
2603 		if (!tcp_skb_can_collapse_to(to))
2604 			break;
2605 
2606 		space -= skb->len;
2607 
2608 		if (first) {
2609 			first = false;
2610 			continue;
2611 		}
2612 
2613 		if (space < 0)
2614 			break;
2615 		/* Punt if not enough space exists in the first SKB for
2616 		 * the data in the second
2617 		 */
2618 		if (skb->len > skb_availroom(to))
2619 			break;
2620 
2621 		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2622 			break;
2623 
2624 		tcp_collapse_retrans(sk, to);
2625 	}
2626 }
2627 
2628 /* This retransmits one SKB.  Policy decisions and retransmit queue
2629  * state updates are done by the caller.  Returns non-zero if an
2630  * error occurred which prevented the send.
2631  */
2632 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2633 {
2634 	struct inet_connection_sock *icsk = inet_csk(sk);
2635 	struct tcp_sock *tp = tcp_sk(sk);
2636 	unsigned int cur_mss;
2637 	int diff, len, err;
2638 
2639 
2640 	/* Inconclusive MTU probe */
2641 	if (icsk->icsk_mtup.probe_size)
2642 		icsk->icsk_mtup.probe_size = 0;
2643 
2644 	/* Do not sent more than we queued. 1/4 is reserved for possible
2645 	 * copying overhead: fragmentation, tunneling, mangling etc.
2646 	 */
2647 	if (atomic_read(&sk->sk_wmem_alloc) >
2648 	    min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2649 		  sk->sk_sndbuf))
2650 		return -EAGAIN;
2651 
2652 	if (skb_still_in_host_queue(sk, skb))
2653 		return -EBUSY;
2654 
2655 	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2656 		if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2657 			BUG();
2658 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2659 			return -ENOMEM;
2660 	}
2661 
2662 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2663 		return -EHOSTUNREACH; /* Routing failure or similar. */
2664 
2665 	cur_mss = tcp_current_mss(sk);
2666 
2667 	/* If receiver has shrunk his window, and skb is out of
2668 	 * new window, do not retransmit it. The exception is the
2669 	 * case, when window is shrunk to zero. In this case
2670 	 * our retransmit serves as a zero window probe.
2671 	 */
2672 	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2673 	    TCP_SKB_CB(skb)->seq != tp->snd_una)
2674 		return -EAGAIN;
2675 
2676 	len = cur_mss * segs;
2677 	if (skb->len > len) {
2678 		if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2679 			return -ENOMEM; /* We'll try again later. */
2680 	} else {
2681 		if (skb_unclone(skb, GFP_ATOMIC))
2682 			return -ENOMEM;
2683 
2684 		diff = tcp_skb_pcount(skb);
2685 		tcp_set_skb_tso_segs(skb, cur_mss);
2686 		diff -= tcp_skb_pcount(skb);
2687 		if (diff)
2688 			tcp_adjust_pcount(sk, skb, diff);
2689 		if (skb->len < cur_mss)
2690 			tcp_retrans_try_collapse(sk, skb, cur_mss);
2691 	}
2692 
2693 	/* RFC3168, section 6.1.1.1. ECN fallback */
2694 	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2695 		tcp_ecn_clear_syn(sk, skb);
2696 
2697 	/* make sure skb->data is aligned on arches that require it
2698 	 * and check if ack-trimming & collapsing extended the headroom
2699 	 * beyond what csum_start can cover.
2700 	 */
2701 	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2702 		     skb_headroom(skb) >= 0xFFFF)) {
2703 		struct sk_buff *nskb;
2704 
2705 		skb_mstamp_get(&skb->skb_mstamp);
2706 		nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2707 		err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2708 			     -ENOBUFS;
2709 	} else {
2710 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2711 	}
2712 
2713 	if (likely(!err)) {
2714 		segs = tcp_skb_pcount(skb);
2715 
2716 		TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2717 		/* Update global TCP statistics. */
2718 		TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2719 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2720 			__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2721 		tp->total_retrans += segs;
2722 	}
2723 	return err;
2724 }
2725 
2726 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2727 {
2728 	struct tcp_sock *tp = tcp_sk(sk);
2729 	int err = __tcp_retransmit_skb(sk, skb, segs);
2730 
2731 	if (err == 0) {
2732 #if FASTRETRANS_DEBUG > 0
2733 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2734 			net_dbg_ratelimited("retrans_out leaked\n");
2735 		}
2736 #endif
2737 		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2738 		tp->retrans_out += tcp_skb_pcount(skb);
2739 
2740 		/* Save stamp of the first retransmit. */
2741 		if (!tp->retrans_stamp)
2742 			tp->retrans_stamp = tcp_skb_timestamp(skb);
2743 
2744 	} else if (err != -EBUSY) {
2745 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2746 	}
2747 
2748 	if (tp->undo_retrans < 0)
2749 		tp->undo_retrans = 0;
2750 	tp->undo_retrans += tcp_skb_pcount(skb);
2751 	return err;
2752 }
2753 
2754 /* Check if we forward retransmits are possible in the current
2755  * window/congestion state.
2756  */
2757 static bool tcp_can_forward_retransmit(struct sock *sk)
2758 {
2759 	const struct inet_connection_sock *icsk = inet_csk(sk);
2760 	const struct tcp_sock *tp = tcp_sk(sk);
2761 
2762 	/* Forward retransmissions are possible only during Recovery. */
2763 	if (icsk->icsk_ca_state != TCP_CA_Recovery)
2764 		return false;
2765 
2766 	/* No forward retransmissions in Reno are possible. */
2767 	if (tcp_is_reno(tp))
2768 		return false;
2769 
2770 	/* Yeah, we have to make difficult choice between forward transmission
2771 	 * and retransmission... Both ways have their merits...
2772 	 *
2773 	 * For now we do not retransmit anything, while we have some new
2774 	 * segments to send. In the other cases, follow rule 3 for
2775 	 * NextSeg() specified in RFC3517.
2776 	 */
2777 
2778 	if (tcp_may_send_now(sk))
2779 		return false;
2780 
2781 	return true;
2782 }
2783 
2784 /* This gets called after a retransmit timeout, and the initially
2785  * retransmitted data is acknowledged.  It tries to continue
2786  * resending the rest of the retransmit queue, until either
2787  * we've sent it all or the congestion window limit is reached.
2788  * If doing SACK, the first ACK which comes back for a timeout
2789  * based retransmit packet might feed us FACK information again.
2790  * If so, we use it to avoid unnecessarily retransmissions.
2791  */
2792 void tcp_xmit_retransmit_queue(struct sock *sk)
2793 {
2794 	const struct inet_connection_sock *icsk = inet_csk(sk);
2795 	struct tcp_sock *tp = tcp_sk(sk);
2796 	struct sk_buff *skb;
2797 	struct sk_buff *hole = NULL;
2798 	u32 max_segs, last_lost;
2799 	int mib_idx;
2800 	int fwd_rexmitting = 0;
2801 
2802 	if (!tp->packets_out)
2803 		return;
2804 
2805 	if (!tp->lost_out)
2806 		tp->retransmit_high = tp->snd_una;
2807 
2808 	if (tp->retransmit_skb_hint) {
2809 		skb = tp->retransmit_skb_hint;
2810 		last_lost = TCP_SKB_CB(skb)->end_seq;
2811 		if (after(last_lost, tp->retransmit_high))
2812 			last_lost = tp->retransmit_high;
2813 	} else {
2814 		skb = tcp_write_queue_head(sk);
2815 		last_lost = tp->snd_una;
2816 	}
2817 
2818 	max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2819 	tcp_for_write_queue_from(skb, sk) {
2820 		__u8 sacked;
2821 		int segs;
2822 
2823 		if (skb == tcp_send_head(sk))
2824 			break;
2825 		/* we could do better than to assign each time */
2826 		if (!hole)
2827 			tp->retransmit_skb_hint = skb;
2828 
2829 		segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2830 		if (segs <= 0)
2831 			return;
2832 		sacked = TCP_SKB_CB(skb)->sacked;
2833 		/* In case tcp_shift_skb_data() have aggregated large skbs,
2834 		 * we need to make sure not sending too bigs TSO packets
2835 		 */
2836 		segs = min_t(int, segs, max_segs);
2837 
2838 		if (fwd_rexmitting) {
2839 begin_fwd:
2840 			if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2841 				break;
2842 			mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2843 
2844 		} else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2845 			tp->retransmit_high = last_lost;
2846 			if (!tcp_can_forward_retransmit(sk))
2847 				break;
2848 			/* Backtrack if necessary to non-L'ed skb */
2849 			if (hole) {
2850 				skb = hole;
2851 				hole = NULL;
2852 			}
2853 			fwd_rexmitting = 1;
2854 			goto begin_fwd;
2855 
2856 		} else if (!(sacked & TCPCB_LOST)) {
2857 			if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2858 				hole = skb;
2859 			continue;
2860 
2861 		} else {
2862 			last_lost = TCP_SKB_CB(skb)->end_seq;
2863 			if (icsk->icsk_ca_state != TCP_CA_Loss)
2864 				mib_idx = LINUX_MIB_TCPFASTRETRANS;
2865 			else
2866 				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2867 		}
2868 
2869 		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2870 			continue;
2871 
2872 		if (tcp_small_queue_check(sk, skb, 1))
2873 			return;
2874 
2875 		if (tcp_retransmit_skb(sk, skb, segs))
2876 			return;
2877 
2878 		NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
2879 
2880 		if (tcp_in_cwnd_reduction(sk))
2881 			tp->prr_out += tcp_skb_pcount(skb);
2882 
2883 		if (skb == tcp_write_queue_head(sk))
2884 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2885 						  inet_csk(sk)->icsk_rto,
2886 						  TCP_RTO_MAX);
2887 	}
2888 }
2889 
2890 /* We allow to exceed memory limits for FIN packets to expedite
2891  * connection tear down and (memory) recovery.
2892  * Otherwise tcp_send_fin() could be tempted to either delay FIN
2893  * or even be forced to close flow without any FIN.
2894  * In general, we want to allow one skb per socket to avoid hangs
2895  * with edge trigger epoll()
2896  */
2897 void sk_forced_mem_schedule(struct sock *sk, int size)
2898 {
2899 	int amt;
2900 
2901 	if (size <= sk->sk_forward_alloc)
2902 		return;
2903 	amt = sk_mem_pages(size);
2904 	sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2905 	sk_memory_allocated_add(sk, amt);
2906 
2907 	if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2908 		mem_cgroup_charge_skmem(sk->sk_memcg, amt);
2909 }
2910 
2911 /* Send a FIN. The caller locks the socket for us.
2912  * We should try to send a FIN packet really hard, but eventually give up.
2913  */
2914 void tcp_send_fin(struct sock *sk)
2915 {
2916 	struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2917 	struct tcp_sock *tp = tcp_sk(sk);
2918 
2919 	/* Optimization, tack on the FIN if we have one skb in write queue and
2920 	 * this skb was not yet sent, or we are under memory pressure.
2921 	 * Note: in the latter case, FIN packet will be sent after a timeout,
2922 	 * as TCP stack thinks it has already been transmitted.
2923 	 */
2924 	if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
2925 coalesce:
2926 		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2927 		TCP_SKB_CB(tskb)->end_seq++;
2928 		tp->write_seq++;
2929 		if (!tcp_send_head(sk)) {
2930 			/* This means tskb was already sent.
2931 			 * Pretend we included the FIN on previous transmit.
2932 			 * We need to set tp->snd_nxt to the value it would have
2933 			 * if FIN had been sent. This is because retransmit path
2934 			 * does not change tp->snd_nxt.
2935 			 */
2936 			tp->snd_nxt++;
2937 			return;
2938 		}
2939 	} else {
2940 		skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2941 		if (unlikely(!skb)) {
2942 			if (tskb)
2943 				goto coalesce;
2944 			return;
2945 		}
2946 		skb_reserve(skb, MAX_TCP_HEADER);
2947 		sk_forced_mem_schedule(sk, skb->truesize);
2948 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2949 		tcp_init_nondata_skb(skb, tp->write_seq,
2950 				     TCPHDR_ACK | TCPHDR_FIN);
2951 		tcp_queue_skb(sk, skb);
2952 	}
2953 	__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2954 }
2955 
2956 /* We get here when a process closes a file descriptor (either due to
2957  * an explicit close() or as a byproduct of exit()'ing) and there
2958  * was unread data in the receive queue.  This behavior is recommended
2959  * by RFC 2525, section 2.17.  -DaveM
2960  */
2961 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2962 {
2963 	struct sk_buff *skb;
2964 
2965 	/* NOTE: No TCP options attached and we never retransmit this. */
2966 	skb = alloc_skb(MAX_TCP_HEADER, priority);
2967 	if (!skb) {
2968 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2969 		return;
2970 	}
2971 
2972 	/* Reserve space for headers and prepare control bits. */
2973 	skb_reserve(skb, MAX_TCP_HEADER);
2974 	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2975 			     TCPHDR_ACK | TCPHDR_RST);
2976 	skb_mstamp_get(&skb->skb_mstamp);
2977 	/* Send it off. */
2978 	if (tcp_transmit_skb(sk, skb, 0, priority))
2979 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2980 
2981 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2982 }
2983 
2984 /* Send a crossed SYN-ACK during socket establishment.
2985  * WARNING: This routine must only be called when we have already sent
2986  * a SYN packet that crossed the incoming SYN that caused this routine
2987  * to get called. If this assumption fails then the initial rcv_wnd
2988  * and rcv_wscale values will not be correct.
2989  */
2990 int tcp_send_synack(struct sock *sk)
2991 {
2992 	struct sk_buff *skb;
2993 
2994 	skb = tcp_write_queue_head(sk);
2995 	if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2996 		pr_debug("%s: wrong queue state\n", __func__);
2997 		return -EFAULT;
2998 	}
2999 	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3000 		if (skb_cloned(skb)) {
3001 			struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
3002 			if (!nskb)
3003 				return -ENOMEM;
3004 			tcp_unlink_write_queue(skb, sk);
3005 			__skb_header_release(nskb);
3006 			__tcp_add_write_queue_head(sk, nskb);
3007 			sk_wmem_free_skb(sk, skb);
3008 			sk->sk_wmem_queued += nskb->truesize;
3009 			sk_mem_charge(sk, nskb->truesize);
3010 			skb = nskb;
3011 		}
3012 
3013 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3014 		tcp_ecn_send_synack(sk, skb);
3015 	}
3016 	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3017 }
3018 
3019 /**
3020  * tcp_make_synack - Prepare a SYN-ACK.
3021  * sk: listener socket
3022  * dst: dst entry attached to the SYNACK
3023  * req: request_sock pointer
3024  *
3025  * Allocate one skb and build a SYNACK packet.
3026  * @dst is consumed : Caller should not use it again.
3027  */
3028 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3029 				struct request_sock *req,
3030 				struct tcp_fastopen_cookie *foc,
3031 				enum tcp_synack_type synack_type)
3032 {
3033 	struct inet_request_sock *ireq = inet_rsk(req);
3034 	const struct tcp_sock *tp = tcp_sk(sk);
3035 	struct tcp_md5sig_key *md5 = NULL;
3036 	struct tcp_out_options opts;
3037 	struct sk_buff *skb;
3038 	int tcp_header_size;
3039 	struct tcphdr *th;
3040 	u16 user_mss;
3041 	int mss;
3042 
3043 	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3044 	if (unlikely(!skb)) {
3045 		dst_release(dst);
3046 		return NULL;
3047 	}
3048 	/* Reserve space for headers. */
3049 	skb_reserve(skb, MAX_TCP_HEADER);
3050 
3051 	switch (synack_type) {
3052 	case TCP_SYNACK_NORMAL:
3053 		skb_set_owner_w(skb, req_to_sk(req));
3054 		break;
3055 	case TCP_SYNACK_COOKIE:
3056 		/* Under synflood, we do not attach skb to a socket,
3057 		 * to avoid false sharing.
3058 		 */
3059 		break;
3060 	case TCP_SYNACK_FASTOPEN:
3061 		/* sk is a const pointer, because we want to express multiple
3062 		 * cpu might call us concurrently.
3063 		 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3064 		 */
3065 		skb_set_owner_w(skb, (struct sock *)sk);
3066 		break;
3067 	}
3068 	skb_dst_set(skb, dst);
3069 
3070 	mss = dst_metric_advmss(dst);
3071 	user_mss = READ_ONCE(tp->rx_opt.user_mss);
3072 	if (user_mss && user_mss < mss)
3073 		mss = user_mss;
3074 
3075 	memset(&opts, 0, sizeof(opts));
3076 #ifdef CONFIG_SYN_COOKIES
3077 	if (unlikely(req->cookie_ts))
3078 		skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3079 	else
3080 #endif
3081 	skb_mstamp_get(&skb->skb_mstamp);
3082 
3083 #ifdef CONFIG_TCP_MD5SIG
3084 	rcu_read_lock();
3085 	md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3086 #endif
3087 	skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3088 	tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3089 			  sizeof(*th);
3090 
3091 	skb_push(skb, tcp_header_size);
3092 	skb_reset_transport_header(skb);
3093 
3094 	th = (struct tcphdr *)skb->data;
3095 	memset(th, 0, sizeof(struct tcphdr));
3096 	th->syn = 1;
3097 	th->ack = 1;
3098 	tcp_ecn_make_synack(req, th);
3099 	th->source = htons(ireq->ir_num);
3100 	th->dest = ireq->ir_rmt_port;
3101 	/* Setting of flags are superfluous here for callers (and ECE is
3102 	 * not even correctly set)
3103 	 */
3104 	tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3105 			     TCPHDR_SYN | TCPHDR_ACK);
3106 
3107 	th->seq = htonl(TCP_SKB_CB(skb)->seq);
3108 	/* XXX data is queued and acked as is. No buffer/window check */
3109 	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3110 
3111 	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3112 	th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3113 	tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3114 	th->doff = (tcp_header_size >> 2);
3115 	__TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3116 
3117 #ifdef CONFIG_TCP_MD5SIG
3118 	/* Okay, we have all we need - do the md5 hash if needed */
3119 	if (md5)
3120 		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3121 					       md5, req_to_sk(req), skb);
3122 	rcu_read_unlock();
3123 #endif
3124 
3125 	/* Do not fool tcpdump (if any), clean our debris */
3126 	skb->tstamp.tv64 = 0;
3127 	return skb;
3128 }
3129 EXPORT_SYMBOL(tcp_make_synack);
3130 
3131 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3132 {
3133 	struct inet_connection_sock *icsk = inet_csk(sk);
3134 	const struct tcp_congestion_ops *ca;
3135 	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3136 
3137 	if (ca_key == TCP_CA_UNSPEC)
3138 		return;
3139 
3140 	rcu_read_lock();
3141 	ca = tcp_ca_find_key(ca_key);
3142 	if (likely(ca && try_module_get(ca->owner))) {
3143 		module_put(icsk->icsk_ca_ops->owner);
3144 		icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3145 		icsk->icsk_ca_ops = ca;
3146 	}
3147 	rcu_read_unlock();
3148 }
3149 
3150 /* Do all connect socket setups that can be done AF independent. */
3151 static void tcp_connect_init(struct sock *sk)
3152 {
3153 	const struct dst_entry *dst = __sk_dst_get(sk);
3154 	struct tcp_sock *tp = tcp_sk(sk);
3155 	__u8 rcv_wscale;
3156 
3157 	/* We'll fix this up when we get a response from the other end.
3158 	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3159 	 */
3160 	tp->tcp_header_len = sizeof(struct tcphdr) +
3161 		(sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3162 
3163 #ifdef CONFIG_TCP_MD5SIG
3164 	if (tp->af_specific->md5_lookup(sk, sk))
3165 		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3166 #endif
3167 
3168 	/* If user gave his TCP_MAXSEG, record it to clamp */
3169 	if (tp->rx_opt.user_mss)
3170 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3171 	tp->max_window = 0;
3172 	tcp_mtup_init(sk);
3173 	tcp_sync_mss(sk, dst_mtu(dst));
3174 
3175 	tcp_ca_dst_init(sk, dst);
3176 
3177 	if (!tp->window_clamp)
3178 		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3179 	tp->advmss = dst_metric_advmss(dst);
3180 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3181 		tp->advmss = tp->rx_opt.user_mss;
3182 
3183 	tcp_initialize_rcv_mss(sk);
3184 
3185 	/* limit the window selection if the user enforce a smaller rx buffer */
3186 	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3187 	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3188 		tp->window_clamp = tcp_full_space(sk);
3189 
3190 	tcp_select_initial_window(tcp_full_space(sk),
3191 				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3192 				  &tp->rcv_wnd,
3193 				  &tp->window_clamp,
3194 				  sysctl_tcp_window_scaling,
3195 				  &rcv_wscale,
3196 				  dst_metric(dst, RTAX_INITRWND));
3197 
3198 	tp->rx_opt.rcv_wscale = rcv_wscale;
3199 	tp->rcv_ssthresh = tp->rcv_wnd;
3200 
3201 	sk->sk_err = 0;
3202 	sock_reset_flag(sk, SOCK_DONE);
3203 	tp->snd_wnd = 0;
3204 	tcp_init_wl(tp, 0);
3205 	tp->snd_una = tp->write_seq;
3206 	tp->snd_sml = tp->write_seq;
3207 	tp->snd_up = tp->write_seq;
3208 	tp->snd_nxt = tp->write_seq;
3209 
3210 	if (likely(!tp->repair))
3211 		tp->rcv_nxt = 0;
3212 	else
3213 		tp->rcv_tstamp = tcp_time_stamp;
3214 	tp->rcv_wup = tp->rcv_nxt;
3215 	tp->copied_seq = tp->rcv_nxt;
3216 
3217 	inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3218 	inet_csk(sk)->icsk_retransmits = 0;
3219 	tcp_clear_retrans(tp);
3220 }
3221 
3222 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3223 {
3224 	struct tcp_sock *tp = tcp_sk(sk);
3225 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3226 
3227 	tcb->end_seq += skb->len;
3228 	__skb_header_release(skb);
3229 	__tcp_add_write_queue_tail(sk, skb);
3230 	sk->sk_wmem_queued += skb->truesize;
3231 	sk_mem_charge(sk, skb->truesize);
3232 	tp->write_seq = tcb->end_seq;
3233 	tp->packets_out += tcp_skb_pcount(skb);
3234 }
3235 
3236 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3237  * queue a data-only packet after the regular SYN, such that regular SYNs
3238  * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3239  * only the SYN sequence, the data are retransmitted in the first ACK.
3240  * If cookie is not cached or other error occurs, falls back to send a
3241  * regular SYN with Fast Open cookie request option.
3242  */
3243 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3244 {
3245 	struct tcp_sock *tp = tcp_sk(sk);
3246 	struct tcp_fastopen_request *fo = tp->fastopen_req;
3247 	int syn_loss = 0, space, err = 0;
3248 	unsigned long last_syn_loss = 0;
3249 	struct sk_buff *syn_data;
3250 
3251 	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3252 	tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3253 			       &syn_loss, &last_syn_loss);
3254 	/* Recurring FO SYN losses: revert to regular handshake temporarily */
3255 	if (syn_loss > 1 &&
3256 	    time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3257 		fo->cookie.len = -1;
3258 		goto fallback;
3259 	}
3260 
3261 	if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3262 		fo->cookie.len = -1;
3263 	else if (fo->cookie.len <= 0)
3264 		goto fallback;
3265 
3266 	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3267 	 * user-MSS. Reserve maximum option space for middleboxes that add
3268 	 * private TCP options. The cost is reduced data space in SYN :(
3269 	 */
3270 	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3271 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3272 	space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3273 		MAX_TCP_OPTION_SPACE;
3274 
3275 	space = min_t(size_t, space, fo->size);
3276 
3277 	/* limit to order-0 allocations */
3278 	space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3279 
3280 	syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3281 	if (!syn_data)
3282 		goto fallback;
3283 	syn_data->ip_summed = CHECKSUM_PARTIAL;
3284 	memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3285 	if (space) {
3286 		int copied = copy_from_iter(skb_put(syn_data, space), space,
3287 					    &fo->data->msg_iter);
3288 		if (unlikely(!copied)) {
3289 			kfree_skb(syn_data);
3290 			goto fallback;
3291 		}
3292 		if (copied != space) {
3293 			skb_trim(syn_data, copied);
3294 			space = copied;
3295 		}
3296 	}
3297 	/* No more data pending in inet_wait_for_connect() */
3298 	if (space == fo->size)
3299 		fo->data = NULL;
3300 	fo->copied = space;
3301 
3302 	tcp_connect_queue_skb(sk, syn_data);
3303 
3304 	err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3305 
3306 	syn->skb_mstamp = syn_data->skb_mstamp;
3307 
3308 	/* Now full SYN+DATA was cloned and sent (or not),
3309 	 * remove the SYN from the original skb (syn_data)
3310 	 * we keep in write queue in case of a retransmit, as we
3311 	 * also have the SYN packet (with no data) in the same queue.
3312 	 */
3313 	TCP_SKB_CB(syn_data)->seq++;
3314 	TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3315 	if (!err) {
3316 		tp->syn_data = (fo->copied > 0);
3317 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3318 		goto done;
3319 	}
3320 
3321 fallback:
3322 	/* Send a regular SYN with Fast Open cookie request option */
3323 	if (fo->cookie.len > 0)
3324 		fo->cookie.len = 0;
3325 	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3326 	if (err)
3327 		tp->syn_fastopen = 0;
3328 done:
3329 	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3330 	return err;
3331 }
3332 
3333 /* Build a SYN and send it off. */
3334 int tcp_connect(struct sock *sk)
3335 {
3336 	struct tcp_sock *tp = tcp_sk(sk);
3337 	struct sk_buff *buff;
3338 	int err;
3339 
3340 	tcp_connect_init(sk);
3341 
3342 	if (unlikely(tp->repair)) {
3343 		tcp_finish_connect(sk, NULL);
3344 		return 0;
3345 	}
3346 
3347 	buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3348 	if (unlikely(!buff))
3349 		return -ENOBUFS;
3350 
3351 	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3352 	tp->retrans_stamp = tcp_time_stamp;
3353 	tcp_connect_queue_skb(sk, buff);
3354 	tcp_ecn_send_syn(sk, buff);
3355 
3356 	/* Send off SYN; include data in Fast Open. */
3357 	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3358 	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3359 	if (err == -ECONNREFUSED)
3360 		return err;
3361 
3362 	/* We change tp->snd_nxt after the tcp_transmit_skb() call
3363 	 * in order to make this packet get counted in tcpOutSegs.
3364 	 */
3365 	tp->snd_nxt = tp->write_seq;
3366 	tp->pushed_seq = tp->write_seq;
3367 	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3368 
3369 	/* Timer for repeating the SYN until an answer. */
3370 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3371 				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3372 	return 0;
3373 }
3374 EXPORT_SYMBOL(tcp_connect);
3375 
3376 /* Send out a delayed ack, the caller does the policy checking
3377  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3378  * for details.
3379  */
3380 void tcp_send_delayed_ack(struct sock *sk)
3381 {
3382 	struct inet_connection_sock *icsk = inet_csk(sk);
3383 	int ato = icsk->icsk_ack.ato;
3384 	unsigned long timeout;
3385 
3386 	tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3387 
3388 	if (ato > TCP_DELACK_MIN) {
3389 		const struct tcp_sock *tp = tcp_sk(sk);
3390 		int max_ato = HZ / 2;
3391 
3392 		if (icsk->icsk_ack.pingpong ||
3393 		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3394 			max_ato = TCP_DELACK_MAX;
3395 
3396 		/* Slow path, intersegment interval is "high". */
3397 
3398 		/* If some rtt estimate is known, use it to bound delayed ack.
3399 		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3400 		 * directly.
3401 		 */
3402 		if (tp->srtt_us) {
3403 			int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3404 					TCP_DELACK_MIN);
3405 
3406 			if (rtt < max_ato)
3407 				max_ato = rtt;
3408 		}
3409 
3410 		ato = min(ato, max_ato);
3411 	}
3412 
3413 	/* Stay within the limit we were given */
3414 	timeout = jiffies + ato;
3415 
3416 	/* Use new timeout only if there wasn't a older one earlier. */
3417 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3418 		/* If delack timer was blocked or is about to expire,
3419 		 * send ACK now.
3420 		 */
3421 		if (icsk->icsk_ack.blocked ||
3422 		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3423 			tcp_send_ack(sk);
3424 			return;
3425 		}
3426 
3427 		if (!time_before(timeout, icsk->icsk_ack.timeout))
3428 			timeout = icsk->icsk_ack.timeout;
3429 	}
3430 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3431 	icsk->icsk_ack.timeout = timeout;
3432 	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3433 }
3434 
3435 /* This routine sends an ack and also updates the window. */
3436 void tcp_send_ack(struct sock *sk)
3437 {
3438 	struct sk_buff *buff;
3439 
3440 	/* If we have been reset, we may not send again. */
3441 	if (sk->sk_state == TCP_CLOSE)
3442 		return;
3443 
3444 	tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3445 
3446 	/* We are not putting this on the write queue, so
3447 	 * tcp_transmit_skb() will set the ownership to this
3448 	 * sock.
3449 	 */
3450 	buff = alloc_skb(MAX_TCP_HEADER,
3451 			 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3452 	if (unlikely(!buff)) {
3453 		inet_csk_schedule_ack(sk);
3454 		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3455 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3456 					  TCP_DELACK_MAX, TCP_RTO_MAX);
3457 		return;
3458 	}
3459 
3460 	/* Reserve space for headers and prepare control bits. */
3461 	skb_reserve(buff, MAX_TCP_HEADER);
3462 	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3463 
3464 	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
3465 	 * too much.
3466 	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3467 	 * We also avoid tcp_wfree() overhead (cache line miss accessing
3468 	 * tp->tsq_flags) by using regular sock_wfree()
3469 	 */
3470 	skb_set_tcp_pure_ack(buff);
3471 
3472 	/* Send it off, this clears delayed acks for us. */
3473 	skb_mstamp_get(&buff->skb_mstamp);
3474 	tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3475 }
3476 EXPORT_SYMBOL_GPL(tcp_send_ack);
3477 
3478 /* This routine sends a packet with an out of date sequence
3479  * number. It assumes the other end will try to ack it.
3480  *
3481  * Question: what should we make while urgent mode?
3482  * 4.4BSD forces sending single byte of data. We cannot send
3483  * out of window data, because we have SND.NXT==SND.MAX...
3484  *
3485  * Current solution: to send TWO zero-length segments in urgent mode:
3486  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3487  * out-of-date with SND.UNA-1 to probe window.
3488  */
3489 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3490 {
3491 	struct tcp_sock *tp = tcp_sk(sk);
3492 	struct sk_buff *skb;
3493 
3494 	/* We don't queue it, tcp_transmit_skb() sets ownership. */
3495 	skb = alloc_skb(MAX_TCP_HEADER,
3496 			sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3497 	if (!skb)
3498 		return -1;
3499 
3500 	/* Reserve space for headers and set control bits. */
3501 	skb_reserve(skb, MAX_TCP_HEADER);
3502 	/* Use a previous sequence.  This should cause the other
3503 	 * end to send an ack.  Don't queue or clone SKB, just
3504 	 * send it.
3505 	 */
3506 	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3507 	skb_mstamp_get(&skb->skb_mstamp);
3508 	NET_INC_STATS(sock_net(sk), mib);
3509 	return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3510 }
3511 
3512 void tcp_send_window_probe(struct sock *sk)
3513 {
3514 	if (sk->sk_state == TCP_ESTABLISHED) {
3515 		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3516 		tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3517 	}
3518 }
3519 
3520 /* Initiate keepalive or window probe from timer. */
3521 int tcp_write_wakeup(struct sock *sk, int mib)
3522 {
3523 	struct tcp_sock *tp = tcp_sk(sk);
3524 	struct sk_buff *skb;
3525 
3526 	if (sk->sk_state == TCP_CLOSE)
3527 		return -1;
3528 
3529 	skb = tcp_send_head(sk);
3530 	if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3531 		int err;
3532 		unsigned int mss = tcp_current_mss(sk);
3533 		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3534 
3535 		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3536 			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3537 
3538 		/* We are probing the opening of a window
3539 		 * but the window size is != 0
3540 		 * must have been a result SWS avoidance ( sender )
3541 		 */
3542 		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3543 		    skb->len > mss) {
3544 			seg_size = min(seg_size, mss);
3545 			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3546 			if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3547 				return -1;
3548 		} else if (!tcp_skb_pcount(skb))
3549 			tcp_set_skb_tso_segs(skb, mss);
3550 
3551 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3552 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3553 		if (!err)
3554 			tcp_event_new_data_sent(sk, skb);
3555 		return err;
3556 	} else {
3557 		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3558 			tcp_xmit_probe_skb(sk, 1, mib);
3559 		return tcp_xmit_probe_skb(sk, 0, mib);
3560 	}
3561 }
3562 
3563 /* A window probe timeout has occurred.  If window is not closed send
3564  * a partial packet else a zero probe.
3565  */
3566 void tcp_send_probe0(struct sock *sk)
3567 {
3568 	struct inet_connection_sock *icsk = inet_csk(sk);
3569 	struct tcp_sock *tp = tcp_sk(sk);
3570 	struct net *net = sock_net(sk);
3571 	unsigned long probe_max;
3572 	int err;
3573 
3574 	err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3575 
3576 	if (tp->packets_out || !tcp_send_head(sk)) {
3577 		/* Cancel probe timer, if it is not required. */
3578 		icsk->icsk_probes_out = 0;
3579 		icsk->icsk_backoff = 0;
3580 		return;
3581 	}
3582 
3583 	if (err <= 0) {
3584 		if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3585 			icsk->icsk_backoff++;
3586 		icsk->icsk_probes_out++;
3587 		probe_max = TCP_RTO_MAX;
3588 	} else {
3589 		/* If packet was not sent due to local congestion,
3590 		 * do not backoff and do not remember icsk_probes_out.
3591 		 * Let local senders to fight for local resources.
3592 		 *
3593 		 * Use accumulated backoff yet.
3594 		 */
3595 		if (!icsk->icsk_probes_out)
3596 			icsk->icsk_probes_out = 1;
3597 		probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3598 	}
3599 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3600 				  tcp_probe0_when(sk, probe_max),
3601 				  TCP_RTO_MAX);
3602 }
3603 
3604 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3605 {
3606 	const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3607 	struct flowi fl;
3608 	int res;
3609 
3610 	tcp_rsk(req)->txhash = net_tx_rndhash();
3611 	res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3612 	if (!res) {
3613 		__TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3614 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3615 		if (unlikely(tcp_passive_fastopen(sk)))
3616 			tcp_sk(sk)->total_retrans++;
3617 	}
3618 	return res;
3619 }
3620 EXPORT_SYMBOL(tcp_rtx_synack);
3621