xref: /linux/net/ipv4/tcp_output.c (revision fada1935590f66dc6784981e0d557ca09013c847)
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 #include <linux/static_key.h>
45 
46 #include <trace/events/tcp.h>
47 
48 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
49 			   int push_one, gfp_t gfp);
50 
51 /* Account for new data that has been sent to the network. */
52 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
53 {
54 	struct inet_connection_sock *icsk = inet_csk(sk);
55 	struct tcp_sock *tp = tcp_sk(sk);
56 	unsigned int prior_packets = tp->packets_out;
57 
58 	tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
59 
60 	__skb_unlink(skb, &sk->sk_write_queue);
61 	tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
62 
63 	tp->packets_out += tcp_skb_pcount(skb);
64 	if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
65 		tcp_rearm_rto(sk);
66 
67 	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
68 		      tcp_skb_pcount(skb));
69 }
70 
71 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
72  * window scaling factor due to loss of precision.
73  * If window has been shrunk, what should we make? It is not clear at all.
74  * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
75  * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
76  * invalid. OK, let's make this for now:
77  */
78 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
79 {
80 	const struct tcp_sock *tp = tcp_sk(sk);
81 
82 	if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
83 	    (tp->rx_opt.wscale_ok &&
84 	     ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
85 		return tp->snd_nxt;
86 	else
87 		return tcp_wnd_end(tp);
88 }
89 
90 /* Calculate mss to advertise in SYN segment.
91  * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
92  *
93  * 1. It is independent of path mtu.
94  * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
95  * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
96  *    attached devices, because some buggy hosts are confused by
97  *    large MSS.
98  * 4. We do not make 3, we advertise MSS, calculated from first
99  *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
100  *    This may be overridden via information stored in routing table.
101  * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
102  *    probably even Jumbo".
103  */
104 static __u16 tcp_advertise_mss(struct sock *sk)
105 {
106 	struct tcp_sock *tp = tcp_sk(sk);
107 	const struct dst_entry *dst = __sk_dst_get(sk);
108 	int mss = tp->advmss;
109 
110 	if (dst) {
111 		unsigned int metric = dst_metric_advmss(dst);
112 
113 		if (metric < mss) {
114 			mss = metric;
115 			tp->advmss = mss;
116 		}
117 	}
118 
119 	return (__u16)mss;
120 }
121 
122 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
123  * This is the first part of cwnd validation mechanism.
124  */
125 void tcp_cwnd_restart(struct sock *sk, s32 delta)
126 {
127 	struct tcp_sock *tp = tcp_sk(sk);
128 	u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
129 	u32 cwnd = tp->snd_cwnd;
130 
131 	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
132 
133 	tp->snd_ssthresh = tcp_current_ssthresh(sk);
134 	restart_cwnd = min(restart_cwnd, cwnd);
135 
136 	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
137 		cwnd >>= 1;
138 	tp->snd_cwnd = max(cwnd, restart_cwnd);
139 	tp->snd_cwnd_stamp = tcp_jiffies32;
140 	tp->snd_cwnd_used = 0;
141 }
142 
143 /* Congestion state accounting after a packet has been sent. */
144 static void tcp_event_data_sent(struct tcp_sock *tp,
145 				struct sock *sk)
146 {
147 	struct inet_connection_sock *icsk = inet_csk(sk);
148 	const u32 now = tcp_jiffies32;
149 
150 	if (tcp_packets_in_flight(tp) == 0)
151 		tcp_ca_event(sk, CA_EVENT_TX_START);
152 
153 	tp->lsndtime = now;
154 
155 	/* If it is a reply for ato after last received
156 	 * packet, enter pingpong mode.
157 	 */
158 	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
159 		icsk->icsk_ack.pingpong = 1;
160 }
161 
162 /* Account for an ACK we sent. */
163 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
164 {
165 	tcp_dec_quickack_mode(sk, pkts);
166 	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
167 }
168 
169 
170 u32 tcp_default_init_rwnd(u32 mss)
171 {
172 	/* Initial receive window should be twice of TCP_INIT_CWND to
173 	 * enable proper sending of new unsent data during fast recovery
174 	 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
175 	 * limit when mss is larger than 1460.
176 	 */
177 	u32 init_rwnd = TCP_INIT_CWND * 2;
178 
179 	if (mss > 1460)
180 		init_rwnd = max((1460 * init_rwnd) / mss, 2U);
181 	return init_rwnd;
182 }
183 
184 /* Determine a window scaling and initial window to offer.
185  * Based on the assumption that the given amount of space
186  * will be offered. Store the results in the tp structure.
187  * NOTE: for smooth operation initial space offering should
188  * be a multiple of mss if possible. We assume here that mss >= 1.
189  * This MUST be enforced by all callers.
190  */
191 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
192 			       __u32 *rcv_wnd, __u32 *window_clamp,
193 			       int wscale_ok, __u8 *rcv_wscale,
194 			       __u32 init_rcv_wnd)
195 {
196 	unsigned int space = (__space < 0 ? 0 : __space);
197 
198 	/* If no clamp set the clamp to the max possible scaled window */
199 	if (*window_clamp == 0)
200 		(*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
201 	space = min(*window_clamp, space);
202 
203 	/* Quantize space offering to a multiple of mss if possible. */
204 	if (space > mss)
205 		space = rounddown(space, mss);
206 
207 	/* NOTE: offering an initial window larger than 32767
208 	 * will break some buggy TCP stacks. If the admin tells us
209 	 * it is likely we could be speaking with such a buggy stack
210 	 * we will truncate our initial window offering to 32K-1
211 	 * unless the remote has sent us a window scaling option,
212 	 * which we interpret as a sign the remote TCP is not
213 	 * misinterpreting the window field as a signed quantity.
214 	 */
215 	if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
216 		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
217 	else
218 		(*rcv_wnd) = space;
219 
220 	(*rcv_wscale) = 0;
221 	if (wscale_ok) {
222 		/* Set window scaling on max possible window */
223 		space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
224 		space = max_t(u32, space, sysctl_rmem_max);
225 		space = min_t(u32, space, *window_clamp);
226 		while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
227 			space >>= 1;
228 			(*rcv_wscale)++;
229 		}
230 	}
231 
232 	if (mss > (1 << *rcv_wscale)) {
233 		if (!init_rcv_wnd) /* Use default unless specified otherwise */
234 			init_rcv_wnd = tcp_default_init_rwnd(mss);
235 		*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
236 	}
237 
238 	/* Set the clamp no higher than max representable value */
239 	(*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
240 }
241 EXPORT_SYMBOL(tcp_select_initial_window);
242 
243 /* Chose a new window to advertise, update state in tcp_sock for the
244  * socket, and return result with RFC1323 scaling applied.  The return
245  * value can be stuffed directly into th->window for an outgoing
246  * frame.
247  */
248 static u16 tcp_select_window(struct sock *sk)
249 {
250 	struct tcp_sock *tp = tcp_sk(sk);
251 	u32 old_win = tp->rcv_wnd;
252 	u32 cur_win = tcp_receive_window(tp);
253 	u32 new_win = __tcp_select_window(sk);
254 
255 	/* Never shrink the offered window */
256 	if (new_win < cur_win) {
257 		/* Danger Will Robinson!
258 		 * Don't update rcv_wup/rcv_wnd here or else
259 		 * we will not be able to advertise a zero
260 		 * window in time.  --DaveM
261 		 *
262 		 * Relax Will Robinson.
263 		 */
264 		if (new_win == 0)
265 			NET_INC_STATS(sock_net(sk),
266 				      LINUX_MIB_TCPWANTZEROWINDOWADV);
267 		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
268 	}
269 	tp->rcv_wnd = new_win;
270 	tp->rcv_wup = tp->rcv_nxt;
271 
272 	/* Make sure we do not exceed the maximum possible
273 	 * scaled window.
274 	 */
275 	if (!tp->rx_opt.rcv_wscale &&
276 	    sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
277 		new_win = min(new_win, MAX_TCP_WINDOW);
278 	else
279 		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
280 
281 	/* RFC1323 scaling applied */
282 	new_win >>= tp->rx_opt.rcv_wscale;
283 
284 	/* If we advertise zero window, disable fast path. */
285 	if (new_win == 0) {
286 		tp->pred_flags = 0;
287 		if (old_win)
288 			NET_INC_STATS(sock_net(sk),
289 				      LINUX_MIB_TCPTOZEROWINDOWADV);
290 	} else if (old_win == 0) {
291 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
292 	}
293 
294 	return new_win;
295 }
296 
297 /* Packet ECN state for a SYN-ACK */
298 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
299 {
300 	const struct tcp_sock *tp = tcp_sk(sk);
301 
302 	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
303 	if (!(tp->ecn_flags & TCP_ECN_OK))
304 		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
305 	else if (tcp_ca_needs_ecn(sk) ||
306 		 tcp_bpf_ca_needs_ecn(sk))
307 		INET_ECN_xmit(sk);
308 }
309 
310 /* Packet ECN state for a SYN.  */
311 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
312 {
313 	struct tcp_sock *tp = tcp_sk(sk);
314 	bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
315 	bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
316 		tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
317 
318 	if (!use_ecn) {
319 		const struct dst_entry *dst = __sk_dst_get(sk);
320 
321 		if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
322 			use_ecn = true;
323 	}
324 
325 	tp->ecn_flags = 0;
326 
327 	if (use_ecn) {
328 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
329 		tp->ecn_flags = TCP_ECN_OK;
330 		if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
331 			INET_ECN_xmit(sk);
332 	}
333 }
334 
335 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
336 {
337 	if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
338 		/* tp->ecn_flags are cleared at a later point in time when
339 		 * SYN ACK is ultimatively being received.
340 		 */
341 		TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
342 }
343 
344 static void
345 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
346 {
347 	if (inet_rsk(req)->ecn_ok)
348 		th->ece = 1;
349 }
350 
351 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
352  * be sent.
353  */
354 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
355 			 struct tcphdr *th, int tcp_header_len)
356 {
357 	struct tcp_sock *tp = tcp_sk(sk);
358 
359 	if (tp->ecn_flags & TCP_ECN_OK) {
360 		/* Not-retransmitted data segment: set ECT and inject CWR. */
361 		if (skb->len != tcp_header_len &&
362 		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
363 			INET_ECN_xmit(sk);
364 			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
365 				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
366 				th->cwr = 1;
367 				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
368 			}
369 		} else if (!tcp_ca_needs_ecn(sk)) {
370 			/* ACK or retransmitted segment: clear ECT|CE */
371 			INET_ECN_dontxmit(sk);
372 		}
373 		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
374 			th->ece = 1;
375 	}
376 }
377 
378 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
379  * auto increment end seqno.
380  */
381 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
382 {
383 	skb->ip_summed = CHECKSUM_PARTIAL;
384 
385 	TCP_SKB_CB(skb)->tcp_flags = flags;
386 	TCP_SKB_CB(skb)->sacked = 0;
387 
388 	tcp_skb_pcount_set(skb, 1);
389 
390 	TCP_SKB_CB(skb)->seq = seq;
391 	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
392 		seq++;
393 	TCP_SKB_CB(skb)->end_seq = seq;
394 }
395 
396 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
397 {
398 	return tp->snd_una != tp->snd_up;
399 }
400 
401 #define OPTION_SACK_ADVERTISE	(1 << 0)
402 #define OPTION_TS		(1 << 1)
403 #define OPTION_MD5		(1 << 2)
404 #define OPTION_WSCALE		(1 << 3)
405 #define OPTION_FAST_OPEN_COOKIE	(1 << 8)
406 #define OPTION_SMC		(1 << 9)
407 
408 static void smc_options_write(__be32 *ptr, u16 *options)
409 {
410 #if IS_ENABLED(CONFIG_SMC)
411 	if (static_branch_unlikely(&tcp_have_smc)) {
412 		if (unlikely(OPTION_SMC & *options)) {
413 			*ptr++ = htonl((TCPOPT_NOP  << 24) |
414 				       (TCPOPT_NOP  << 16) |
415 				       (TCPOPT_EXP <<  8) |
416 				       (TCPOLEN_EXP_SMC_BASE));
417 			*ptr++ = htonl(TCPOPT_SMC_MAGIC);
418 		}
419 	}
420 #endif
421 }
422 
423 struct tcp_out_options {
424 	u16 options;		/* bit field of OPTION_* */
425 	u16 mss;		/* 0 to disable */
426 	u8 ws;			/* window scale, 0 to disable */
427 	u8 num_sack_blocks;	/* number of SACK blocks to include */
428 	u8 hash_size;		/* bytes in hash_location */
429 	__u8 *hash_location;	/* temporary pointer, overloaded */
430 	__u32 tsval, tsecr;	/* need to include OPTION_TS */
431 	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
432 };
433 
434 /* Write previously computed TCP options to the packet.
435  *
436  * Beware: Something in the Internet is very sensitive to the ordering of
437  * TCP options, we learned this through the hard way, so be careful here.
438  * Luckily we can at least blame others for their non-compliance but from
439  * inter-operability perspective it seems that we're somewhat stuck with
440  * the ordering which we have been using if we want to keep working with
441  * those broken things (not that it currently hurts anybody as there isn't
442  * particular reason why the ordering would need to be changed).
443  *
444  * At least SACK_PERM as the first option is known to lead to a disaster
445  * (but it may well be that other scenarios fail similarly).
446  */
447 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
448 			      struct tcp_out_options *opts)
449 {
450 	u16 options = opts->options;	/* mungable copy */
451 
452 	if (unlikely(OPTION_MD5 & options)) {
453 		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
454 			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
455 		/* overload cookie hash location */
456 		opts->hash_location = (__u8 *)ptr;
457 		ptr += 4;
458 	}
459 
460 	if (unlikely(opts->mss)) {
461 		*ptr++ = htonl((TCPOPT_MSS << 24) |
462 			       (TCPOLEN_MSS << 16) |
463 			       opts->mss);
464 	}
465 
466 	if (likely(OPTION_TS & options)) {
467 		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
468 			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
469 				       (TCPOLEN_SACK_PERM << 16) |
470 				       (TCPOPT_TIMESTAMP << 8) |
471 				       TCPOLEN_TIMESTAMP);
472 			options &= ~OPTION_SACK_ADVERTISE;
473 		} else {
474 			*ptr++ = htonl((TCPOPT_NOP << 24) |
475 				       (TCPOPT_NOP << 16) |
476 				       (TCPOPT_TIMESTAMP << 8) |
477 				       TCPOLEN_TIMESTAMP);
478 		}
479 		*ptr++ = htonl(opts->tsval);
480 		*ptr++ = htonl(opts->tsecr);
481 	}
482 
483 	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
484 		*ptr++ = htonl((TCPOPT_NOP << 24) |
485 			       (TCPOPT_NOP << 16) |
486 			       (TCPOPT_SACK_PERM << 8) |
487 			       TCPOLEN_SACK_PERM);
488 	}
489 
490 	if (unlikely(OPTION_WSCALE & options)) {
491 		*ptr++ = htonl((TCPOPT_NOP << 24) |
492 			       (TCPOPT_WINDOW << 16) |
493 			       (TCPOLEN_WINDOW << 8) |
494 			       opts->ws);
495 	}
496 
497 	if (unlikely(opts->num_sack_blocks)) {
498 		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
499 			tp->duplicate_sack : tp->selective_acks;
500 		int this_sack;
501 
502 		*ptr++ = htonl((TCPOPT_NOP  << 24) |
503 			       (TCPOPT_NOP  << 16) |
504 			       (TCPOPT_SACK <<  8) |
505 			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
506 						     TCPOLEN_SACK_PERBLOCK)));
507 
508 		for (this_sack = 0; this_sack < opts->num_sack_blocks;
509 		     ++this_sack) {
510 			*ptr++ = htonl(sp[this_sack].start_seq);
511 			*ptr++ = htonl(sp[this_sack].end_seq);
512 		}
513 
514 		tp->rx_opt.dsack = 0;
515 	}
516 
517 	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
518 		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
519 		u8 *p = (u8 *)ptr;
520 		u32 len; /* Fast Open option length */
521 
522 		if (foc->exp) {
523 			len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
524 			*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
525 				     TCPOPT_FASTOPEN_MAGIC);
526 			p += TCPOLEN_EXP_FASTOPEN_BASE;
527 		} else {
528 			len = TCPOLEN_FASTOPEN_BASE + foc->len;
529 			*p++ = TCPOPT_FASTOPEN;
530 			*p++ = len;
531 		}
532 
533 		memcpy(p, foc->val, foc->len);
534 		if ((len & 3) == 2) {
535 			p[foc->len] = TCPOPT_NOP;
536 			p[foc->len + 1] = TCPOPT_NOP;
537 		}
538 		ptr += (len + 3) >> 2;
539 	}
540 
541 	smc_options_write(ptr, &options);
542 }
543 
544 static void smc_set_option(const struct tcp_sock *tp,
545 			   struct tcp_out_options *opts,
546 			   unsigned int *remaining)
547 {
548 #if IS_ENABLED(CONFIG_SMC)
549 	if (static_branch_unlikely(&tcp_have_smc)) {
550 		if (tp->syn_smc) {
551 			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
552 				opts->options |= OPTION_SMC;
553 				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
554 			}
555 		}
556 	}
557 #endif
558 }
559 
560 static void smc_set_option_cond(const struct tcp_sock *tp,
561 				const struct inet_request_sock *ireq,
562 				struct tcp_out_options *opts,
563 				unsigned int *remaining)
564 {
565 #if IS_ENABLED(CONFIG_SMC)
566 	if (static_branch_unlikely(&tcp_have_smc)) {
567 		if (tp->syn_smc && ireq->smc_ok) {
568 			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
569 				opts->options |= OPTION_SMC;
570 				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
571 			}
572 		}
573 	}
574 #endif
575 }
576 
577 /* Compute TCP options for SYN packets. This is not the final
578  * network wire format yet.
579  */
580 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
581 				struct tcp_out_options *opts,
582 				struct tcp_md5sig_key **md5)
583 {
584 	struct tcp_sock *tp = tcp_sk(sk);
585 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
586 	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
587 
588 #ifdef CONFIG_TCP_MD5SIG
589 	*md5 = tp->af_specific->md5_lookup(sk, sk);
590 	if (*md5) {
591 		opts->options |= OPTION_MD5;
592 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
593 	}
594 #else
595 	*md5 = NULL;
596 #endif
597 
598 	/* We always get an MSS option.  The option bytes which will be seen in
599 	 * normal data packets should timestamps be used, must be in the MSS
600 	 * advertised.  But we subtract them from tp->mss_cache so that
601 	 * calculations in tcp_sendmsg are simpler etc.  So account for this
602 	 * fact here if necessary.  If we don't do this correctly, as a
603 	 * receiver we won't recognize data packets as being full sized when we
604 	 * should, and thus we won't abide by the delayed ACK rules correctly.
605 	 * SACKs don't matter, we never delay an ACK when we have any of those
606 	 * going out.  */
607 	opts->mss = tcp_advertise_mss(sk);
608 	remaining -= TCPOLEN_MSS_ALIGNED;
609 
610 	if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
611 		opts->options |= OPTION_TS;
612 		opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
613 		opts->tsecr = tp->rx_opt.ts_recent;
614 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
615 	}
616 	if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
617 		opts->ws = tp->rx_opt.rcv_wscale;
618 		opts->options |= OPTION_WSCALE;
619 		remaining -= TCPOLEN_WSCALE_ALIGNED;
620 	}
621 	if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
622 		opts->options |= OPTION_SACK_ADVERTISE;
623 		if (unlikely(!(OPTION_TS & opts->options)))
624 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
625 	}
626 
627 	if (fastopen && fastopen->cookie.len >= 0) {
628 		u32 need = fastopen->cookie.len;
629 
630 		need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
631 					       TCPOLEN_FASTOPEN_BASE;
632 		need = (need + 3) & ~3U;  /* Align to 32 bits */
633 		if (remaining >= need) {
634 			opts->options |= OPTION_FAST_OPEN_COOKIE;
635 			opts->fastopen_cookie = &fastopen->cookie;
636 			remaining -= need;
637 			tp->syn_fastopen = 1;
638 			tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
639 		}
640 	}
641 
642 	smc_set_option(tp, opts, &remaining);
643 
644 	return MAX_TCP_OPTION_SPACE - remaining;
645 }
646 
647 /* Set up TCP options for SYN-ACKs. */
648 static unsigned int tcp_synack_options(const struct sock *sk,
649 				       struct request_sock *req,
650 				       unsigned int mss, struct sk_buff *skb,
651 				       struct tcp_out_options *opts,
652 				       const struct tcp_md5sig_key *md5,
653 				       struct tcp_fastopen_cookie *foc)
654 {
655 	struct inet_request_sock *ireq = inet_rsk(req);
656 	unsigned int remaining = MAX_TCP_OPTION_SPACE;
657 
658 #ifdef CONFIG_TCP_MD5SIG
659 	if (md5) {
660 		opts->options |= OPTION_MD5;
661 		remaining -= TCPOLEN_MD5SIG_ALIGNED;
662 
663 		/* We can't fit any SACK blocks in a packet with MD5 + TS
664 		 * options. There was discussion about disabling SACK
665 		 * rather than TS in order to fit in better with old,
666 		 * buggy kernels, but that was deemed to be unnecessary.
667 		 */
668 		ireq->tstamp_ok &= !ireq->sack_ok;
669 	}
670 #endif
671 
672 	/* We always send an MSS option. */
673 	opts->mss = mss;
674 	remaining -= TCPOLEN_MSS_ALIGNED;
675 
676 	if (likely(ireq->wscale_ok)) {
677 		opts->ws = ireq->rcv_wscale;
678 		opts->options |= OPTION_WSCALE;
679 		remaining -= TCPOLEN_WSCALE_ALIGNED;
680 	}
681 	if (likely(ireq->tstamp_ok)) {
682 		opts->options |= OPTION_TS;
683 		opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
684 		opts->tsecr = req->ts_recent;
685 		remaining -= TCPOLEN_TSTAMP_ALIGNED;
686 	}
687 	if (likely(ireq->sack_ok)) {
688 		opts->options |= OPTION_SACK_ADVERTISE;
689 		if (unlikely(!ireq->tstamp_ok))
690 			remaining -= TCPOLEN_SACKPERM_ALIGNED;
691 	}
692 	if (foc != NULL && foc->len >= 0) {
693 		u32 need = foc->len;
694 
695 		need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
696 				   TCPOLEN_FASTOPEN_BASE;
697 		need = (need + 3) & ~3U;  /* Align to 32 bits */
698 		if (remaining >= need) {
699 			opts->options |= OPTION_FAST_OPEN_COOKIE;
700 			opts->fastopen_cookie = foc;
701 			remaining -= need;
702 		}
703 	}
704 
705 	smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
706 
707 	return MAX_TCP_OPTION_SPACE - remaining;
708 }
709 
710 /* Compute TCP options for ESTABLISHED sockets. This is not the
711  * final wire format yet.
712  */
713 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
714 					struct tcp_out_options *opts,
715 					struct tcp_md5sig_key **md5)
716 {
717 	struct tcp_sock *tp = tcp_sk(sk);
718 	unsigned int size = 0;
719 	unsigned int eff_sacks;
720 
721 	opts->options = 0;
722 
723 #ifdef CONFIG_TCP_MD5SIG
724 	*md5 = tp->af_specific->md5_lookup(sk, sk);
725 	if (unlikely(*md5)) {
726 		opts->options |= OPTION_MD5;
727 		size += TCPOLEN_MD5SIG_ALIGNED;
728 	}
729 #else
730 	*md5 = NULL;
731 #endif
732 
733 	if (likely(tp->rx_opt.tstamp_ok)) {
734 		opts->options |= OPTION_TS;
735 		opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
736 		opts->tsecr = tp->rx_opt.ts_recent;
737 		size += TCPOLEN_TSTAMP_ALIGNED;
738 	}
739 
740 	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
741 	if (unlikely(eff_sacks)) {
742 		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
743 		opts->num_sack_blocks =
744 			min_t(unsigned int, eff_sacks,
745 			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
746 			      TCPOLEN_SACK_PERBLOCK);
747 		size += TCPOLEN_SACK_BASE_ALIGNED +
748 			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
749 	}
750 
751 	return size;
752 }
753 
754 
755 /* TCP SMALL QUEUES (TSQ)
756  *
757  * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
758  * to reduce RTT and bufferbloat.
759  * We do this using a special skb destructor (tcp_wfree).
760  *
761  * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
762  * needs to be reallocated in a driver.
763  * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
764  *
765  * Since transmit from skb destructor is forbidden, we use a tasklet
766  * to process all sockets that eventually need to send more skbs.
767  * We use one tasklet per cpu, with its own queue of sockets.
768  */
769 struct tsq_tasklet {
770 	struct tasklet_struct	tasklet;
771 	struct list_head	head; /* queue of tcp sockets */
772 };
773 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
774 
775 static void tcp_tsq_handler(struct sock *sk)
776 {
777 	if ((1 << sk->sk_state) &
778 	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
779 	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK)) {
780 		struct tcp_sock *tp = tcp_sk(sk);
781 
782 		if (tp->lost_out > tp->retrans_out &&
783 		    tp->snd_cwnd > tcp_packets_in_flight(tp)) {
784 			tcp_mstamp_refresh(tp);
785 			tcp_xmit_retransmit_queue(sk);
786 		}
787 
788 		tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
789 			       0, GFP_ATOMIC);
790 	}
791 }
792 /*
793  * One tasklet per cpu tries to send more skbs.
794  * We run in tasklet context but need to disable irqs when
795  * transferring tsq->head because tcp_wfree() might
796  * interrupt us (non NAPI drivers)
797  */
798 static void tcp_tasklet_func(unsigned long data)
799 {
800 	struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
801 	LIST_HEAD(list);
802 	unsigned long flags;
803 	struct list_head *q, *n;
804 	struct tcp_sock *tp;
805 	struct sock *sk;
806 
807 	local_irq_save(flags);
808 	list_splice_init(&tsq->head, &list);
809 	local_irq_restore(flags);
810 
811 	list_for_each_safe(q, n, &list) {
812 		tp = list_entry(q, struct tcp_sock, tsq_node);
813 		list_del(&tp->tsq_node);
814 
815 		sk = (struct sock *)tp;
816 		smp_mb__before_atomic();
817 		clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
818 
819 		if (!sk->sk_lock.owned &&
820 		    test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
821 			bh_lock_sock(sk);
822 			if (!sock_owned_by_user(sk)) {
823 				clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
824 				tcp_tsq_handler(sk);
825 			}
826 			bh_unlock_sock(sk);
827 		}
828 
829 		sk_free(sk);
830 	}
831 }
832 
833 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED |		\
834 			  TCPF_WRITE_TIMER_DEFERRED |	\
835 			  TCPF_DELACK_TIMER_DEFERRED |	\
836 			  TCPF_MTU_REDUCED_DEFERRED)
837 /**
838  * tcp_release_cb - tcp release_sock() callback
839  * @sk: socket
840  *
841  * called from release_sock() to perform protocol dependent
842  * actions before socket release.
843  */
844 void tcp_release_cb(struct sock *sk)
845 {
846 	unsigned long flags, nflags;
847 
848 	/* perform an atomic operation only if at least one flag is set */
849 	do {
850 		flags = sk->sk_tsq_flags;
851 		if (!(flags & TCP_DEFERRED_ALL))
852 			return;
853 		nflags = flags & ~TCP_DEFERRED_ALL;
854 	} while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
855 
856 	if (flags & TCPF_TSQ_DEFERRED)
857 		tcp_tsq_handler(sk);
858 
859 	/* Here begins the tricky part :
860 	 * We are called from release_sock() with :
861 	 * 1) BH disabled
862 	 * 2) sk_lock.slock spinlock held
863 	 * 3) socket owned by us (sk->sk_lock.owned == 1)
864 	 *
865 	 * But following code is meant to be called from BH handlers,
866 	 * so we should keep BH disabled, but early release socket ownership
867 	 */
868 	sock_release_ownership(sk);
869 
870 	if (flags & TCPF_WRITE_TIMER_DEFERRED) {
871 		tcp_write_timer_handler(sk);
872 		__sock_put(sk);
873 	}
874 	if (flags & TCPF_DELACK_TIMER_DEFERRED) {
875 		tcp_delack_timer_handler(sk);
876 		__sock_put(sk);
877 	}
878 	if (flags & TCPF_MTU_REDUCED_DEFERRED) {
879 		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
880 		__sock_put(sk);
881 	}
882 }
883 EXPORT_SYMBOL(tcp_release_cb);
884 
885 void __init tcp_tasklet_init(void)
886 {
887 	int i;
888 
889 	for_each_possible_cpu(i) {
890 		struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
891 
892 		INIT_LIST_HEAD(&tsq->head);
893 		tasklet_init(&tsq->tasklet,
894 			     tcp_tasklet_func,
895 			     (unsigned long)tsq);
896 	}
897 }
898 
899 /*
900  * Write buffer destructor automatically called from kfree_skb.
901  * We can't xmit new skbs from this context, as we might already
902  * hold qdisc lock.
903  */
904 void tcp_wfree(struct sk_buff *skb)
905 {
906 	struct sock *sk = skb->sk;
907 	struct tcp_sock *tp = tcp_sk(sk);
908 	unsigned long flags, nval, oval;
909 
910 	/* Keep one reference on sk_wmem_alloc.
911 	 * Will be released by sk_free() from here or tcp_tasklet_func()
912 	 */
913 	WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
914 
915 	/* If this softirq is serviced by ksoftirqd, we are likely under stress.
916 	 * Wait until our queues (qdisc + devices) are drained.
917 	 * This gives :
918 	 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
919 	 * - chance for incoming ACK (processed by another cpu maybe)
920 	 *   to migrate this flow (skb->ooo_okay will be eventually set)
921 	 */
922 	if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
923 		goto out;
924 
925 	for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
926 		struct tsq_tasklet *tsq;
927 		bool empty;
928 
929 		if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
930 			goto out;
931 
932 		nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
933 		nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
934 		if (nval != oval)
935 			continue;
936 
937 		/* queue this socket to tasklet queue */
938 		local_irq_save(flags);
939 		tsq = this_cpu_ptr(&tsq_tasklet);
940 		empty = list_empty(&tsq->head);
941 		list_add(&tp->tsq_node, &tsq->head);
942 		if (empty)
943 			tasklet_schedule(&tsq->tasklet);
944 		local_irq_restore(flags);
945 		return;
946 	}
947 out:
948 	sk_free(sk);
949 }
950 
951 /* Note: Called under hard irq.
952  * We can not call TCP stack right away.
953  */
954 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
955 {
956 	struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
957 	struct sock *sk = (struct sock *)tp;
958 	unsigned long nval, oval;
959 
960 	for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
961 		struct tsq_tasklet *tsq;
962 		bool empty;
963 
964 		if (oval & TSQF_QUEUED)
965 			break;
966 
967 		nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
968 		nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
969 		if (nval != oval)
970 			continue;
971 
972 		if (!refcount_inc_not_zero(&sk->sk_wmem_alloc))
973 			break;
974 		/* queue this socket to tasklet queue */
975 		tsq = this_cpu_ptr(&tsq_tasklet);
976 		empty = list_empty(&tsq->head);
977 		list_add(&tp->tsq_node, &tsq->head);
978 		if (empty)
979 			tasklet_schedule(&tsq->tasklet);
980 		break;
981 	}
982 	return HRTIMER_NORESTART;
983 }
984 
985 /* BBR congestion control needs pacing.
986  * Same remark for SO_MAX_PACING_RATE.
987  * sch_fq packet scheduler is efficiently handling pacing,
988  * but is not always installed/used.
989  * Return true if TCP stack should pace packets itself.
990  */
991 static bool tcp_needs_internal_pacing(const struct sock *sk)
992 {
993 	return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
994 }
995 
996 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
997 {
998 	u64 len_ns;
999 	u32 rate;
1000 
1001 	if (!tcp_needs_internal_pacing(sk))
1002 		return;
1003 	rate = sk->sk_pacing_rate;
1004 	if (!rate || rate == ~0U)
1005 		return;
1006 
1007 	/* Should account for header sizes as sch_fq does,
1008 	 * but lets make things simple.
1009 	 */
1010 	len_ns = (u64)skb->len * NSEC_PER_SEC;
1011 	do_div(len_ns, rate);
1012 	hrtimer_start(&tcp_sk(sk)->pacing_timer,
1013 		      ktime_add_ns(ktime_get(), len_ns),
1014 		      HRTIMER_MODE_ABS_PINNED);
1015 }
1016 
1017 static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb)
1018 {
1019 	skb->skb_mstamp = tp->tcp_mstamp;
1020 	list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1021 }
1022 
1023 /* This routine actually transmits TCP packets queued in by
1024  * tcp_do_sendmsg().  This is used by both the initial
1025  * transmission and possible later retransmissions.
1026  * All SKB's seen here are completely headerless.  It is our
1027  * job to build the TCP header, and pass the packet down to
1028  * IP so it can do the same plus pass the packet off to the
1029  * device.
1030  *
1031  * We are working here with either a clone of the original
1032  * SKB, or a fresh unique copy made by the retransmit engine.
1033  */
1034 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1035 			    gfp_t gfp_mask)
1036 {
1037 	const struct inet_connection_sock *icsk = inet_csk(sk);
1038 	struct inet_sock *inet;
1039 	struct tcp_sock *tp;
1040 	struct tcp_skb_cb *tcb;
1041 	struct tcp_out_options opts;
1042 	unsigned int tcp_options_size, tcp_header_size;
1043 	struct sk_buff *oskb = NULL;
1044 	struct tcp_md5sig_key *md5;
1045 	struct tcphdr *th;
1046 	int err;
1047 
1048 	BUG_ON(!skb || !tcp_skb_pcount(skb));
1049 	tp = tcp_sk(sk);
1050 
1051 	if (clone_it) {
1052 		TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1053 			- tp->snd_una;
1054 		oskb = skb;
1055 
1056 		tcp_skb_tsorted_save(oskb) {
1057 			if (unlikely(skb_cloned(oskb)))
1058 				skb = pskb_copy(oskb, gfp_mask);
1059 			else
1060 				skb = skb_clone(oskb, gfp_mask);
1061 		} tcp_skb_tsorted_restore(oskb);
1062 
1063 		if (unlikely(!skb))
1064 			return -ENOBUFS;
1065 	}
1066 	skb->skb_mstamp = tp->tcp_mstamp;
1067 
1068 	inet = inet_sk(sk);
1069 	tcb = TCP_SKB_CB(skb);
1070 	memset(&opts, 0, sizeof(opts));
1071 
1072 	if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1073 		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1074 	else
1075 		tcp_options_size = tcp_established_options(sk, skb, &opts,
1076 							   &md5);
1077 	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1078 
1079 	/* if no packet is in qdisc/device queue, then allow XPS to select
1080 	 * another queue. We can be called from tcp_tsq_handler()
1081 	 * which holds one reference to sk_wmem_alloc.
1082 	 *
1083 	 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1084 	 * One way to get this would be to set skb->truesize = 2 on them.
1085 	 */
1086 	skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1087 
1088 	/* If we had to use memory reserve to allocate this skb,
1089 	 * this might cause drops if packet is looped back :
1090 	 * Other socket might not have SOCK_MEMALLOC.
1091 	 * Packets not looped back do not care about pfmemalloc.
1092 	 */
1093 	skb->pfmemalloc = 0;
1094 
1095 	skb_push(skb, tcp_header_size);
1096 	skb_reset_transport_header(skb);
1097 
1098 	skb_orphan(skb);
1099 	skb->sk = sk;
1100 	skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1101 	skb_set_hash_from_sk(skb, sk);
1102 	refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1103 
1104 	skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1105 
1106 	/* Build TCP header and checksum it. */
1107 	th = (struct tcphdr *)skb->data;
1108 	th->source		= inet->inet_sport;
1109 	th->dest		= inet->inet_dport;
1110 	th->seq			= htonl(tcb->seq);
1111 	th->ack_seq		= htonl(tp->rcv_nxt);
1112 	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
1113 					tcb->tcp_flags);
1114 
1115 	th->check		= 0;
1116 	th->urg_ptr		= 0;
1117 
1118 	/* The urg_mode check is necessary during a below snd_una win probe */
1119 	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1120 		if (before(tp->snd_up, tcb->seq + 0x10000)) {
1121 			th->urg_ptr = htons(tp->snd_up - tcb->seq);
1122 			th->urg = 1;
1123 		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1124 			th->urg_ptr = htons(0xFFFF);
1125 			th->urg = 1;
1126 		}
1127 	}
1128 
1129 	tcp_options_write((__be32 *)(th + 1), tp, &opts);
1130 	skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1131 	if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1132 		th->window      = htons(tcp_select_window(sk));
1133 		tcp_ecn_send(sk, skb, th, tcp_header_size);
1134 	} else {
1135 		/* RFC1323: The window in SYN & SYN/ACK segments
1136 		 * is never scaled.
1137 		 */
1138 		th->window	= htons(min(tp->rcv_wnd, 65535U));
1139 	}
1140 #ifdef CONFIG_TCP_MD5SIG
1141 	/* Calculate the MD5 hash, as we have all we need now */
1142 	if (md5) {
1143 		sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1144 		tp->af_specific->calc_md5_hash(opts.hash_location,
1145 					       md5, sk, skb);
1146 	}
1147 #endif
1148 
1149 	icsk->icsk_af_ops->send_check(sk, skb);
1150 
1151 	if (likely(tcb->tcp_flags & TCPHDR_ACK))
1152 		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1153 
1154 	if (skb->len != tcp_header_size) {
1155 		tcp_event_data_sent(tp, sk);
1156 		tp->data_segs_out += tcp_skb_pcount(skb);
1157 		tcp_internal_pacing(sk, skb);
1158 	}
1159 
1160 	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1161 		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1162 			      tcp_skb_pcount(skb));
1163 
1164 	tp->segs_out += tcp_skb_pcount(skb);
1165 	/* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1166 	skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1167 	skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1168 
1169 	/* Our usage of tstamp should remain private */
1170 	skb->tstamp = 0;
1171 
1172 	/* Cleanup our debris for IP stacks */
1173 	memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1174 			       sizeof(struct inet6_skb_parm)));
1175 
1176 	err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1177 
1178 	if (unlikely(err > 0)) {
1179 		tcp_enter_cwr(sk);
1180 		err = net_xmit_eval(err);
1181 	}
1182 	if (!err && oskb) {
1183 		tcp_update_skb_after_send(tp, oskb);
1184 		tcp_rate_skb_sent(sk, oskb);
1185 	}
1186 	return err;
1187 }
1188 
1189 /* This routine just queues the buffer for sending.
1190  *
1191  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1192  * otherwise socket can stall.
1193  */
1194 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1195 {
1196 	struct tcp_sock *tp = tcp_sk(sk);
1197 
1198 	/* Advance write_seq and place onto the write_queue. */
1199 	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1200 	__skb_header_release(skb);
1201 	tcp_add_write_queue_tail(sk, skb);
1202 	sk->sk_wmem_queued += skb->truesize;
1203 	sk_mem_charge(sk, skb->truesize);
1204 }
1205 
1206 /* Initialize TSO segments for a packet. */
1207 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1208 {
1209 	if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1210 		/* Avoid the costly divide in the normal
1211 		 * non-TSO case.
1212 		 */
1213 		tcp_skb_pcount_set(skb, 1);
1214 		TCP_SKB_CB(skb)->tcp_gso_size = 0;
1215 	} else {
1216 		tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1217 		TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1218 	}
1219 }
1220 
1221 /* Pcount in the middle of the write queue got changed, we need to do various
1222  * tweaks to fix counters
1223  */
1224 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1225 {
1226 	struct tcp_sock *tp = tcp_sk(sk);
1227 
1228 	tp->packets_out -= decr;
1229 
1230 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1231 		tp->sacked_out -= decr;
1232 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1233 		tp->retrans_out -= decr;
1234 	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1235 		tp->lost_out -= decr;
1236 
1237 	/* Reno case is special. Sigh... */
1238 	if (tcp_is_reno(tp) && decr > 0)
1239 		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1240 
1241 	if (tp->lost_skb_hint &&
1242 	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1243 	    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1244 		tp->lost_cnt_hint -= decr;
1245 
1246 	tcp_verify_left_out(tp);
1247 }
1248 
1249 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1250 {
1251 	return TCP_SKB_CB(skb)->txstamp_ack ||
1252 		(skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1253 }
1254 
1255 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1256 {
1257 	struct skb_shared_info *shinfo = skb_shinfo(skb);
1258 
1259 	if (unlikely(tcp_has_tx_tstamp(skb)) &&
1260 	    !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1261 		struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1262 		u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1263 
1264 		shinfo->tx_flags &= ~tsflags;
1265 		shinfo2->tx_flags |= tsflags;
1266 		swap(shinfo->tskey, shinfo2->tskey);
1267 		TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1268 		TCP_SKB_CB(skb)->txstamp_ack = 0;
1269 	}
1270 }
1271 
1272 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1273 {
1274 	TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1275 	TCP_SKB_CB(skb)->eor = 0;
1276 }
1277 
1278 /* Insert buff after skb on the write or rtx queue of sk.  */
1279 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1280 					 struct sk_buff *buff,
1281 					 struct sock *sk,
1282 					 enum tcp_queue tcp_queue)
1283 {
1284 	if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1285 		__skb_queue_after(&sk->sk_write_queue, skb, buff);
1286 	else
1287 		tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1288 }
1289 
1290 /* Function to create two new TCP segments.  Shrinks the given segment
1291  * to the specified size and appends a new segment with the rest of the
1292  * packet to the list.  This won't be called frequently, I hope.
1293  * Remember, these are still headerless SKBs at this point.
1294  */
1295 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1296 		 struct sk_buff *skb, u32 len,
1297 		 unsigned int mss_now, gfp_t gfp)
1298 {
1299 	struct tcp_sock *tp = tcp_sk(sk);
1300 	struct sk_buff *buff;
1301 	int nsize, old_factor;
1302 	int nlen;
1303 	u8 flags;
1304 
1305 	if (WARN_ON(len > skb->len))
1306 		return -EINVAL;
1307 
1308 	nsize = skb_headlen(skb) - len;
1309 	if (nsize < 0)
1310 		nsize = 0;
1311 
1312 	if (skb_unclone(skb, gfp))
1313 		return -ENOMEM;
1314 
1315 	/* Get a new skb... force flag on. */
1316 	buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1317 	if (!buff)
1318 		return -ENOMEM; /* We'll just try again later. */
1319 
1320 	sk->sk_wmem_queued += buff->truesize;
1321 	sk_mem_charge(sk, buff->truesize);
1322 	nlen = skb->len - len - nsize;
1323 	buff->truesize += nlen;
1324 	skb->truesize -= nlen;
1325 
1326 	/* Correct the sequence numbers. */
1327 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1328 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1329 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1330 
1331 	/* PSH and FIN should only be set in the second packet. */
1332 	flags = TCP_SKB_CB(skb)->tcp_flags;
1333 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1334 	TCP_SKB_CB(buff)->tcp_flags = flags;
1335 	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1336 	tcp_skb_fragment_eor(skb, buff);
1337 
1338 	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1339 		/* Copy and checksum data tail into the new buffer. */
1340 		buff->csum = csum_partial_copy_nocheck(skb->data + len,
1341 						       skb_put(buff, nsize),
1342 						       nsize, 0);
1343 
1344 		skb_trim(skb, len);
1345 
1346 		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1347 	} else {
1348 		skb->ip_summed = CHECKSUM_PARTIAL;
1349 		skb_split(skb, buff, len);
1350 	}
1351 
1352 	buff->ip_summed = skb->ip_summed;
1353 
1354 	buff->tstamp = skb->tstamp;
1355 	tcp_fragment_tstamp(skb, buff);
1356 
1357 	old_factor = tcp_skb_pcount(skb);
1358 
1359 	/* Fix up tso_factor for both original and new SKB.  */
1360 	tcp_set_skb_tso_segs(skb, mss_now);
1361 	tcp_set_skb_tso_segs(buff, mss_now);
1362 
1363 	/* Update delivered info for the new segment */
1364 	TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1365 
1366 	/* If this packet has been sent out already, we must
1367 	 * adjust the various packet counters.
1368 	 */
1369 	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1370 		int diff = old_factor - tcp_skb_pcount(skb) -
1371 			tcp_skb_pcount(buff);
1372 
1373 		if (diff)
1374 			tcp_adjust_pcount(sk, skb, diff);
1375 	}
1376 
1377 	/* Link BUFF into the send queue. */
1378 	__skb_header_release(buff);
1379 	tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1380 	if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1381 		list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1382 
1383 	return 0;
1384 }
1385 
1386 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1387  * data is not copied, but immediately discarded.
1388  */
1389 static int __pskb_trim_head(struct sk_buff *skb, int len)
1390 {
1391 	struct skb_shared_info *shinfo;
1392 	int i, k, eat;
1393 
1394 	eat = min_t(int, len, skb_headlen(skb));
1395 	if (eat) {
1396 		__skb_pull(skb, eat);
1397 		len -= eat;
1398 		if (!len)
1399 			return 0;
1400 	}
1401 	eat = len;
1402 	k = 0;
1403 	shinfo = skb_shinfo(skb);
1404 	for (i = 0; i < shinfo->nr_frags; i++) {
1405 		int size = skb_frag_size(&shinfo->frags[i]);
1406 
1407 		if (size <= eat) {
1408 			skb_frag_unref(skb, i);
1409 			eat -= size;
1410 		} else {
1411 			shinfo->frags[k] = shinfo->frags[i];
1412 			if (eat) {
1413 				shinfo->frags[k].page_offset += eat;
1414 				skb_frag_size_sub(&shinfo->frags[k], eat);
1415 				eat = 0;
1416 			}
1417 			k++;
1418 		}
1419 	}
1420 	shinfo->nr_frags = k;
1421 
1422 	skb->data_len -= len;
1423 	skb->len = skb->data_len;
1424 	return len;
1425 }
1426 
1427 /* Remove acked data from a packet in the transmit queue. */
1428 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1429 {
1430 	u32 delta_truesize;
1431 
1432 	if (skb_unclone(skb, GFP_ATOMIC))
1433 		return -ENOMEM;
1434 
1435 	delta_truesize = __pskb_trim_head(skb, len);
1436 
1437 	TCP_SKB_CB(skb)->seq += len;
1438 	skb->ip_summed = CHECKSUM_PARTIAL;
1439 
1440 	if (delta_truesize) {
1441 		skb->truesize	   -= delta_truesize;
1442 		sk->sk_wmem_queued -= delta_truesize;
1443 		sk_mem_uncharge(sk, delta_truesize);
1444 		sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1445 	}
1446 
1447 	/* Any change of skb->len requires recalculation of tso factor. */
1448 	if (tcp_skb_pcount(skb) > 1)
1449 		tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1450 
1451 	return 0;
1452 }
1453 
1454 /* Calculate MSS not accounting any TCP options.  */
1455 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1456 {
1457 	const struct tcp_sock *tp = tcp_sk(sk);
1458 	const struct inet_connection_sock *icsk = inet_csk(sk);
1459 	int mss_now;
1460 
1461 	/* Calculate base mss without TCP options:
1462 	   It is MMS_S - sizeof(tcphdr) of rfc1122
1463 	 */
1464 	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1465 
1466 	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1467 	if (icsk->icsk_af_ops->net_frag_header_len) {
1468 		const struct dst_entry *dst = __sk_dst_get(sk);
1469 
1470 		if (dst && dst_allfrag(dst))
1471 			mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1472 	}
1473 
1474 	/* Clamp it (mss_clamp does not include tcp options) */
1475 	if (mss_now > tp->rx_opt.mss_clamp)
1476 		mss_now = tp->rx_opt.mss_clamp;
1477 
1478 	/* Now subtract optional transport overhead */
1479 	mss_now -= icsk->icsk_ext_hdr_len;
1480 
1481 	/* Then reserve room for full set of TCP options and 8 bytes of data */
1482 	if (mss_now < 48)
1483 		mss_now = 48;
1484 	return mss_now;
1485 }
1486 
1487 /* Calculate MSS. Not accounting for SACKs here.  */
1488 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1489 {
1490 	/* Subtract TCP options size, not including SACKs */
1491 	return __tcp_mtu_to_mss(sk, pmtu) -
1492 	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1493 }
1494 
1495 /* Inverse of above */
1496 int tcp_mss_to_mtu(struct sock *sk, int mss)
1497 {
1498 	const struct tcp_sock *tp = tcp_sk(sk);
1499 	const struct inet_connection_sock *icsk = inet_csk(sk);
1500 	int mtu;
1501 
1502 	mtu = mss +
1503 	      tp->tcp_header_len +
1504 	      icsk->icsk_ext_hdr_len +
1505 	      icsk->icsk_af_ops->net_header_len;
1506 
1507 	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1508 	if (icsk->icsk_af_ops->net_frag_header_len) {
1509 		const struct dst_entry *dst = __sk_dst_get(sk);
1510 
1511 		if (dst && dst_allfrag(dst))
1512 			mtu += icsk->icsk_af_ops->net_frag_header_len;
1513 	}
1514 	return mtu;
1515 }
1516 EXPORT_SYMBOL(tcp_mss_to_mtu);
1517 
1518 /* MTU probing init per socket */
1519 void tcp_mtup_init(struct sock *sk)
1520 {
1521 	struct tcp_sock *tp = tcp_sk(sk);
1522 	struct inet_connection_sock *icsk = inet_csk(sk);
1523 	struct net *net = sock_net(sk);
1524 
1525 	icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1526 	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1527 			       icsk->icsk_af_ops->net_header_len;
1528 	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1529 	icsk->icsk_mtup.probe_size = 0;
1530 	if (icsk->icsk_mtup.enabled)
1531 		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1532 }
1533 EXPORT_SYMBOL(tcp_mtup_init);
1534 
1535 /* This function synchronize snd mss to current pmtu/exthdr set.
1536 
1537    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1538    for TCP options, but includes only bare TCP header.
1539 
1540    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1541    It is minimum of user_mss and mss received with SYN.
1542    It also does not include TCP options.
1543 
1544    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1545 
1546    tp->mss_cache is current effective sending mss, including
1547    all tcp options except for SACKs. It is evaluated,
1548    taking into account current pmtu, but never exceeds
1549    tp->rx_opt.mss_clamp.
1550 
1551    NOTE1. rfc1122 clearly states that advertised MSS
1552    DOES NOT include either tcp or ip options.
1553 
1554    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1555    are READ ONLY outside this function.		--ANK (980731)
1556  */
1557 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1558 {
1559 	struct tcp_sock *tp = tcp_sk(sk);
1560 	struct inet_connection_sock *icsk = inet_csk(sk);
1561 	int mss_now;
1562 
1563 	if (icsk->icsk_mtup.search_high > pmtu)
1564 		icsk->icsk_mtup.search_high = pmtu;
1565 
1566 	mss_now = tcp_mtu_to_mss(sk, pmtu);
1567 	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1568 
1569 	/* And store cached results */
1570 	icsk->icsk_pmtu_cookie = pmtu;
1571 	if (icsk->icsk_mtup.enabled)
1572 		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1573 	tp->mss_cache = mss_now;
1574 
1575 	return mss_now;
1576 }
1577 EXPORT_SYMBOL(tcp_sync_mss);
1578 
1579 /* Compute the current effective MSS, taking SACKs and IP options,
1580  * and even PMTU discovery events into account.
1581  */
1582 unsigned int tcp_current_mss(struct sock *sk)
1583 {
1584 	const struct tcp_sock *tp = tcp_sk(sk);
1585 	const struct dst_entry *dst = __sk_dst_get(sk);
1586 	u32 mss_now;
1587 	unsigned int header_len;
1588 	struct tcp_out_options opts;
1589 	struct tcp_md5sig_key *md5;
1590 
1591 	mss_now = tp->mss_cache;
1592 
1593 	if (dst) {
1594 		u32 mtu = dst_mtu(dst);
1595 		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1596 			mss_now = tcp_sync_mss(sk, mtu);
1597 	}
1598 
1599 	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1600 		     sizeof(struct tcphdr);
1601 	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1602 	 * some common options. If this is an odd packet (because we have SACK
1603 	 * blocks etc) then our calculated header_len will be different, and
1604 	 * we have to adjust mss_now correspondingly */
1605 	if (header_len != tp->tcp_header_len) {
1606 		int delta = (int) header_len - tp->tcp_header_len;
1607 		mss_now -= delta;
1608 	}
1609 
1610 	return mss_now;
1611 }
1612 
1613 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1614  * As additional protections, we do not touch cwnd in retransmission phases,
1615  * and if application hit its sndbuf limit recently.
1616  */
1617 static void tcp_cwnd_application_limited(struct sock *sk)
1618 {
1619 	struct tcp_sock *tp = tcp_sk(sk);
1620 
1621 	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1622 	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1623 		/* Limited by application or receiver window. */
1624 		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1625 		u32 win_used = max(tp->snd_cwnd_used, init_win);
1626 		if (win_used < tp->snd_cwnd) {
1627 			tp->snd_ssthresh = tcp_current_ssthresh(sk);
1628 			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1629 		}
1630 		tp->snd_cwnd_used = 0;
1631 	}
1632 	tp->snd_cwnd_stamp = tcp_jiffies32;
1633 }
1634 
1635 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1636 {
1637 	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1638 	struct tcp_sock *tp = tcp_sk(sk);
1639 
1640 	/* Track the maximum number of outstanding packets in each
1641 	 * window, and remember whether we were cwnd-limited then.
1642 	 */
1643 	if (!before(tp->snd_una, tp->max_packets_seq) ||
1644 	    tp->packets_out > tp->max_packets_out) {
1645 		tp->max_packets_out = tp->packets_out;
1646 		tp->max_packets_seq = tp->snd_nxt;
1647 		tp->is_cwnd_limited = is_cwnd_limited;
1648 	}
1649 
1650 	if (tcp_is_cwnd_limited(sk)) {
1651 		/* Network is feed fully. */
1652 		tp->snd_cwnd_used = 0;
1653 		tp->snd_cwnd_stamp = tcp_jiffies32;
1654 	} else {
1655 		/* Network starves. */
1656 		if (tp->packets_out > tp->snd_cwnd_used)
1657 			tp->snd_cwnd_used = tp->packets_out;
1658 
1659 		if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1660 		    (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1661 		    !ca_ops->cong_control)
1662 			tcp_cwnd_application_limited(sk);
1663 
1664 		/* The following conditions together indicate the starvation
1665 		 * is caused by insufficient sender buffer:
1666 		 * 1) just sent some data (see tcp_write_xmit)
1667 		 * 2) not cwnd limited (this else condition)
1668 		 * 3) no more data to send (tcp_write_queue_empty())
1669 		 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1670 		 */
1671 		if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1672 		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1673 		    (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1674 			tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1675 	}
1676 }
1677 
1678 /* Minshall's variant of the Nagle send check. */
1679 static bool tcp_minshall_check(const struct tcp_sock *tp)
1680 {
1681 	return after(tp->snd_sml, tp->snd_una) &&
1682 		!after(tp->snd_sml, tp->snd_nxt);
1683 }
1684 
1685 /* Update snd_sml if this skb is under mss
1686  * Note that a TSO packet might end with a sub-mss segment
1687  * The test is really :
1688  * if ((skb->len % mss) != 0)
1689  *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1690  * But we can avoid doing the divide again given we already have
1691  *  skb_pcount = skb->len / mss_now
1692  */
1693 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1694 				const struct sk_buff *skb)
1695 {
1696 	if (skb->len < tcp_skb_pcount(skb) * mss_now)
1697 		tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1698 }
1699 
1700 /* Return false, if packet can be sent now without violation Nagle's rules:
1701  * 1. It is full sized. (provided by caller in %partial bool)
1702  * 2. Or it contains FIN. (already checked by caller)
1703  * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1704  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1705  *    With Minshall's modification: all sent small packets are ACKed.
1706  */
1707 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1708 			    int nonagle)
1709 {
1710 	return partial &&
1711 		((nonagle & TCP_NAGLE_CORK) ||
1712 		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1713 }
1714 
1715 /* Return how many segs we'd like on a TSO packet,
1716  * to send one TSO packet per ms
1717  */
1718 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1719 		     int min_tso_segs)
1720 {
1721 	u32 bytes, segs;
1722 
1723 	bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift,
1724 		    sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1725 
1726 	/* Goal is to send at least one packet per ms,
1727 	 * not one big TSO packet every 100 ms.
1728 	 * This preserves ACK clocking and is consistent
1729 	 * with tcp_tso_should_defer() heuristic.
1730 	 */
1731 	segs = max_t(u32, bytes / mss_now, min_tso_segs);
1732 
1733 	return min_t(u32, segs, sk->sk_gso_max_segs);
1734 }
1735 EXPORT_SYMBOL(tcp_tso_autosize);
1736 
1737 /* Return the number of segments we want in the skb we are transmitting.
1738  * See if congestion control module wants to decide; otherwise, autosize.
1739  */
1740 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1741 {
1742 	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1743 	u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1744 
1745 	return tso_segs ? :
1746 		tcp_tso_autosize(sk, mss_now,
1747 				 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
1748 }
1749 
1750 /* Returns the portion of skb which can be sent right away */
1751 static unsigned int tcp_mss_split_point(const struct sock *sk,
1752 					const struct sk_buff *skb,
1753 					unsigned int mss_now,
1754 					unsigned int max_segs,
1755 					int nonagle)
1756 {
1757 	const struct tcp_sock *tp = tcp_sk(sk);
1758 	u32 partial, needed, window, max_len;
1759 
1760 	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1761 	max_len = mss_now * max_segs;
1762 
1763 	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1764 		return max_len;
1765 
1766 	needed = min(skb->len, window);
1767 
1768 	if (max_len <= needed)
1769 		return max_len;
1770 
1771 	partial = needed % mss_now;
1772 	/* If last segment is not a full MSS, check if Nagle rules allow us
1773 	 * to include this last segment in this skb.
1774 	 * Otherwise, we'll split the skb at last MSS boundary
1775 	 */
1776 	if (tcp_nagle_check(partial != 0, tp, nonagle))
1777 		return needed - partial;
1778 
1779 	return needed;
1780 }
1781 
1782 /* Can at least one segment of SKB be sent right now, according to the
1783  * congestion window rules?  If so, return how many segments are allowed.
1784  */
1785 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1786 					 const struct sk_buff *skb)
1787 {
1788 	u32 in_flight, cwnd, halfcwnd;
1789 
1790 	/* Don't be strict about the congestion window for the final FIN.  */
1791 	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1792 	    tcp_skb_pcount(skb) == 1)
1793 		return 1;
1794 
1795 	in_flight = tcp_packets_in_flight(tp);
1796 	cwnd = tp->snd_cwnd;
1797 	if (in_flight >= cwnd)
1798 		return 0;
1799 
1800 	/* For better scheduling, ensure we have at least
1801 	 * 2 GSO packets in flight.
1802 	 */
1803 	halfcwnd = max(cwnd >> 1, 1U);
1804 	return min(halfcwnd, cwnd - in_flight);
1805 }
1806 
1807 /* Initialize TSO state of a skb.
1808  * This must be invoked the first time we consider transmitting
1809  * SKB onto the wire.
1810  */
1811 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1812 {
1813 	int tso_segs = tcp_skb_pcount(skb);
1814 
1815 	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1816 		tcp_set_skb_tso_segs(skb, mss_now);
1817 		tso_segs = tcp_skb_pcount(skb);
1818 	}
1819 	return tso_segs;
1820 }
1821 
1822 
1823 /* Return true if the Nagle test allows this packet to be
1824  * sent now.
1825  */
1826 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1827 				  unsigned int cur_mss, int nonagle)
1828 {
1829 	/* Nagle rule does not apply to frames, which sit in the middle of the
1830 	 * write_queue (they have no chances to get new data).
1831 	 *
1832 	 * This is implemented in the callers, where they modify the 'nonagle'
1833 	 * argument based upon the location of SKB in the send queue.
1834 	 */
1835 	if (nonagle & TCP_NAGLE_PUSH)
1836 		return true;
1837 
1838 	/* Don't use the nagle rule for urgent data (or for the final FIN). */
1839 	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1840 		return true;
1841 
1842 	if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1843 		return true;
1844 
1845 	return false;
1846 }
1847 
1848 /* Does at least the first segment of SKB fit into the send window? */
1849 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1850 			     const struct sk_buff *skb,
1851 			     unsigned int cur_mss)
1852 {
1853 	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1854 
1855 	if (skb->len > cur_mss)
1856 		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1857 
1858 	return !after(end_seq, tcp_wnd_end(tp));
1859 }
1860 
1861 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1862  * which is put after SKB on the list.  It is very much like
1863  * tcp_fragment() except that it may make several kinds of assumptions
1864  * in order to speed up the splitting operation.  In particular, we
1865  * know that all the data is in scatter-gather pages, and that the
1866  * packet has never been sent out before (and thus is not cloned).
1867  */
1868 static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1869 			struct sk_buff *skb, unsigned int len,
1870 			unsigned int mss_now, gfp_t gfp)
1871 {
1872 	struct sk_buff *buff;
1873 	int nlen = skb->len - len;
1874 	u8 flags;
1875 
1876 	/* All of a TSO frame must be composed of paged data.  */
1877 	if (skb->len != skb->data_len)
1878 		return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
1879 
1880 	buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1881 	if (unlikely(!buff))
1882 		return -ENOMEM;
1883 
1884 	sk->sk_wmem_queued += buff->truesize;
1885 	sk_mem_charge(sk, buff->truesize);
1886 	buff->truesize += nlen;
1887 	skb->truesize -= nlen;
1888 
1889 	/* Correct the sequence numbers. */
1890 	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1891 	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1892 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1893 
1894 	/* PSH and FIN should only be set in the second packet. */
1895 	flags = TCP_SKB_CB(skb)->tcp_flags;
1896 	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1897 	TCP_SKB_CB(buff)->tcp_flags = flags;
1898 
1899 	/* This packet was never sent out yet, so no SACK bits. */
1900 	TCP_SKB_CB(buff)->sacked = 0;
1901 
1902 	tcp_skb_fragment_eor(skb, buff);
1903 
1904 	buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1905 	skb_split(skb, buff, len);
1906 	tcp_fragment_tstamp(skb, buff);
1907 
1908 	/* Fix up tso_factor for both original and new SKB.  */
1909 	tcp_set_skb_tso_segs(skb, mss_now);
1910 	tcp_set_skb_tso_segs(buff, mss_now);
1911 
1912 	/* Link BUFF into the send queue. */
1913 	__skb_header_release(buff);
1914 	tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1915 
1916 	return 0;
1917 }
1918 
1919 /* Try to defer sending, if possible, in order to minimize the amount
1920  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1921  *
1922  * This algorithm is from John Heffner.
1923  */
1924 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1925 				 bool *is_cwnd_limited, u32 max_segs)
1926 {
1927 	const struct inet_connection_sock *icsk = inet_csk(sk);
1928 	u32 age, send_win, cong_win, limit, in_flight;
1929 	struct tcp_sock *tp = tcp_sk(sk);
1930 	struct sk_buff *head;
1931 	int win_divisor;
1932 
1933 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1934 		goto send_now;
1935 
1936 	if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1937 		goto send_now;
1938 
1939 	/* Avoid bursty behavior by allowing defer
1940 	 * only if the last write was recent.
1941 	 */
1942 	if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1943 		goto send_now;
1944 
1945 	in_flight = tcp_packets_in_flight(tp);
1946 
1947 	BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1948 
1949 	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1950 
1951 	/* From in_flight test above, we know that cwnd > in_flight.  */
1952 	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1953 
1954 	limit = min(send_win, cong_win);
1955 
1956 	/* If a full-sized TSO skb can be sent, do it. */
1957 	if (limit >= max_segs * tp->mss_cache)
1958 		goto send_now;
1959 
1960 	/* Middle in queue won't get any more data, full sendable already? */
1961 	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1962 		goto send_now;
1963 
1964 	win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1965 	if (win_divisor) {
1966 		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1967 
1968 		/* If at least some fraction of a window is available,
1969 		 * just use it.
1970 		 */
1971 		chunk /= win_divisor;
1972 		if (limit >= chunk)
1973 			goto send_now;
1974 	} else {
1975 		/* Different approach, try not to defer past a single
1976 		 * ACK.  Receiver should ACK every other full sized
1977 		 * frame, so if we have space for more than 3 frames
1978 		 * then send now.
1979 		 */
1980 		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1981 			goto send_now;
1982 	}
1983 
1984 	/* TODO : use tsorted_sent_queue ? */
1985 	head = tcp_rtx_queue_head(sk);
1986 	if (!head)
1987 		goto send_now;
1988 	age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1989 	/* If next ACK is likely to come too late (half srtt), do not defer */
1990 	if (age < (tp->srtt_us >> 4))
1991 		goto send_now;
1992 
1993 	/* Ok, it looks like it is advisable to defer. */
1994 
1995 	if (cong_win < send_win && cong_win <= skb->len)
1996 		*is_cwnd_limited = true;
1997 
1998 	return true;
1999 
2000 send_now:
2001 	return false;
2002 }
2003 
2004 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2005 {
2006 	struct inet_connection_sock *icsk = inet_csk(sk);
2007 	struct tcp_sock *tp = tcp_sk(sk);
2008 	struct net *net = sock_net(sk);
2009 	u32 interval;
2010 	s32 delta;
2011 
2012 	interval = net->ipv4.sysctl_tcp_probe_interval;
2013 	delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2014 	if (unlikely(delta >= interval * HZ)) {
2015 		int mss = tcp_current_mss(sk);
2016 
2017 		/* Update current search range */
2018 		icsk->icsk_mtup.probe_size = 0;
2019 		icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2020 			sizeof(struct tcphdr) +
2021 			icsk->icsk_af_ops->net_header_len;
2022 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2023 
2024 		/* Update probe time stamp */
2025 		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2026 	}
2027 }
2028 
2029 /* Create a new MTU probe if we are ready.
2030  * MTU probe is regularly attempting to increase the path MTU by
2031  * deliberately sending larger packets.  This discovers routing
2032  * changes resulting in larger path MTUs.
2033  *
2034  * Returns 0 if we should wait to probe (no cwnd available),
2035  *         1 if a probe was sent,
2036  *         -1 otherwise
2037  */
2038 static int tcp_mtu_probe(struct sock *sk)
2039 {
2040 	struct inet_connection_sock *icsk = inet_csk(sk);
2041 	struct tcp_sock *tp = tcp_sk(sk);
2042 	struct sk_buff *skb, *nskb, *next;
2043 	struct net *net = sock_net(sk);
2044 	int probe_size;
2045 	int size_needed;
2046 	int copy, len;
2047 	int mss_now;
2048 	int interval;
2049 
2050 	/* Not currently probing/verifying,
2051 	 * not in recovery,
2052 	 * have enough cwnd, and
2053 	 * not SACKing (the variable headers throw things off)
2054 	 */
2055 	if (likely(!icsk->icsk_mtup.enabled ||
2056 		   icsk->icsk_mtup.probe_size ||
2057 		   inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2058 		   tp->snd_cwnd < 11 ||
2059 		   tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2060 		return -1;
2061 
2062 	/* Use binary search for probe_size between tcp_mss_base,
2063 	 * and current mss_clamp. if (search_high - search_low)
2064 	 * smaller than a threshold, backoff from probing.
2065 	 */
2066 	mss_now = tcp_current_mss(sk);
2067 	probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2068 				    icsk->icsk_mtup.search_low) >> 1);
2069 	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2070 	interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2071 	/* When misfortune happens, we are reprobing actively,
2072 	 * and then reprobe timer has expired. We stick with current
2073 	 * probing process by not resetting search range to its orignal.
2074 	 */
2075 	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2076 		interval < net->ipv4.sysctl_tcp_probe_threshold) {
2077 		/* Check whether enough time has elaplased for
2078 		 * another round of probing.
2079 		 */
2080 		tcp_mtu_check_reprobe(sk);
2081 		return -1;
2082 	}
2083 
2084 	/* Have enough data in the send queue to probe? */
2085 	if (tp->write_seq - tp->snd_nxt < size_needed)
2086 		return -1;
2087 
2088 	if (tp->snd_wnd < size_needed)
2089 		return -1;
2090 	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2091 		return 0;
2092 
2093 	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
2094 	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2095 		if (!tcp_packets_in_flight(tp))
2096 			return -1;
2097 		else
2098 			return 0;
2099 	}
2100 
2101 	/* We're allowed to probe.  Build it now. */
2102 	nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2103 	if (!nskb)
2104 		return -1;
2105 	sk->sk_wmem_queued += nskb->truesize;
2106 	sk_mem_charge(sk, nskb->truesize);
2107 
2108 	skb = tcp_send_head(sk);
2109 
2110 	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2111 	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2112 	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2113 	TCP_SKB_CB(nskb)->sacked = 0;
2114 	nskb->csum = 0;
2115 	nskb->ip_summed = skb->ip_summed;
2116 
2117 	tcp_insert_write_queue_before(nskb, skb, sk);
2118 	tcp_highest_sack_replace(sk, skb, nskb);
2119 
2120 	len = 0;
2121 	tcp_for_write_queue_from_safe(skb, next, sk) {
2122 		copy = min_t(int, skb->len, probe_size - len);
2123 		if (nskb->ip_summed) {
2124 			skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2125 		} else {
2126 			__wsum csum = skb_copy_and_csum_bits(skb, 0,
2127 							     skb_put(nskb, copy),
2128 							     copy, 0);
2129 			nskb->csum = csum_block_add(nskb->csum, csum, len);
2130 		}
2131 
2132 		if (skb->len <= copy) {
2133 			/* We've eaten all the data from this skb.
2134 			 * Throw it away. */
2135 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2136 			tcp_unlink_write_queue(skb, sk);
2137 			sk_wmem_free_skb(sk, skb);
2138 		} else {
2139 			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2140 						   ~(TCPHDR_FIN|TCPHDR_PSH);
2141 			if (!skb_shinfo(skb)->nr_frags) {
2142 				skb_pull(skb, copy);
2143 				if (skb->ip_summed != CHECKSUM_PARTIAL)
2144 					skb->csum = csum_partial(skb->data,
2145 								 skb->len, 0);
2146 			} else {
2147 				__pskb_trim_head(skb, copy);
2148 				tcp_set_skb_tso_segs(skb, mss_now);
2149 			}
2150 			TCP_SKB_CB(skb)->seq += copy;
2151 		}
2152 
2153 		len += copy;
2154 
2155 		if (len >= probe_size)
2156 			break;
2157 	}
2158 	tcp_init_tso_segs(nskb, nskb->len);
2159 
2160 	/* We're ready to send.  If this fails, the probe will
2161 	 * be resegmented into mss-sized pieces by tcp_write_xmit().
2162 	 */
2163 	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2164 		/* Decrement cwnd here because we are sending
2165 		 * effectively two packets. */
2166 		tp->snd_cwnd--;
2167 		tcp_event_new_data_sent(sk, nskb);
2168 
2169 		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2170 		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2171 		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2172 
2173 		return 1;
2174 	}
2175 
2176 	return -1;
2177 }
2178 
2179 static bool tcp_pacing_check(const struct sock *sk)
2180 {
2181 	return tcp_needs_internal_pacing(sk) &&
2182 	       hrtimer_active(&tcp_sk(sk)->pacing_timer);
2183 }
2184 
2185 /* TCP Small Queues :
2186  * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2187  * (These limits are doubled for retransmits)
2188  * This allows for :
2189  *  - better RTT estimation and ACK scheduling
2190  *  - faster recovery
2191  *  - high rates
2192  * Alas, some drivers / subsystems require a fair amount
2193  * of queued bytes to ensure line rate.
2194  * One example is wifi aggregation (802.11 AMPDU)
2195  */
2196 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2197 				  unsigned int factor)
2198 {
2199 	unsigned int limit;
2200 
2201 	limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift);
2202 	limit = min_t(u32, limit,
2203 		      sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2204 	limit <<= factor;
2205 
2206 	if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2207 		/* Always send skb if rtx queue is empty.
2208 		 * No need to wait for TX completion to call us back,
2209 		 * after softirq/tasklet schedule.
2210 		 * This helps when TX completions are delayed too much.
2211 		 */
2212 		if (tcp_rtx_queue_empty(sk))
2213 			return false;
2214 
2215 		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2216 		/* It is possible TX completion already happened
2217 		 * before we set TSQ_THROTTLED, so we must
2218 		 * test again the condition.
2219 		 */
2220 		smp_mb__after_atomic();
2221 		if (refcount_read(&sk->sk_wmem_alloc) > limit)
2222 			return true;
2223 	}
2224 	return false;
2225 }
2226 
2227 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2228 {
2229 	const u32 now = tcp_jiffies32;
2230 	enum tcp_chrono old = tp->chrono_type;
2231 
2232 	if (old > TCP_CHRONO_UNSPEC)
2233 		tp->chrono_stat[old - 1] += now - tp->chrono_start;
2234 	tp->chrono_start = now;
2235 	tp->chrono_type = new;
2236 }
2237 
2238 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2239 {
2240 	struct tcp_sock *tp = tcp_sk(sk);
2241 
2242 	/* If there are multiple conditions worthy of tracking in a
2243 	 * chronograph then the highest priority enum takes precedence
2244 	 * over the other conditions. So that if something "more interesting"
2245 	 * starts happening, stop the previous chrono and start a new one.
2246 	 */
2247 	if (type > tp->chrono_type)
2248 		tcp_chrono_set(tp, type);
2249 }
2250 
2251 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2252 {
2253 	struct tcp_sock *tp = tcp_sk(sk);
2254 
2255 
2256 	/* There are multiple conditions worthy of tracking in a
2257 	 * chronograph, so that the highest priority enum takes
2258 	 * precedence over the other conditions (see tcp_chrono_start).
2259 	 * If a condition stops, we only stop chrono tracking if
2260 	 * it's the "most interesting" or current chrono we are
2261 	 * tracking and starts busy chrono if we have pending data.
2262 	 */
2263 	if (tcp_rtx_and_write_queues_empty(sk))
2264 		tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2265 	else if (type == tp->chrono_type)
2266 		tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2267 }
2268 
2269 /* This routine writes packets to the network.  It advances the
2270  * send_head.  This happens as incoming acks open up the remote
2271  * window for us.
2272  *
2273  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2274  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2275  * account rare use of URG, this is not a big flaw.
2276  *
2277  * Send at most one packet when push_one > 0. Temporarily ignore
2278  * cwnd limit to force at most one packet out when push_one == 2.
2279 
2280  * Returns true, if no segments are in flight and we have queued segments,
2281  * but cannot send anything now because of SWS or another problem.
2282  */
2283 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2284 			   int push_one, gfp_t gfp)
2285 {
2286 	struct tcp_sock *tp = tcp_sk(sk);
2287 	struct sk_buff *skb;
2288 	unsigned int tso_segs, sent_pkts;
2289 	int cwnd_quota;
2290 	int result;
2291 	bool is_cwnd_limited = false, is_rwnd_limited = false;
2292 	u32 max_segs;
2293 
2294 	sent_pkts = 0;
2295 
2296 	tcp_mstamp_refresh(tp);
2297 	if (!push_one) {
2298 		/* Do MTU probing. */
2299 		result = tcp_mtu_probe(sk);
2300 		if (!result) {
2301 			return false;
2302 		} else if (result > 0) {
2303 			sent_pkts = 1;
2304 		}
2305 	}
2306 
2307 	max_segs = tcp_tso_segs(sk, mss_now);
2308 	while ((skb = tcp_send_head(sk))) {
2309 		unsigned int limit;
2310 
2311 		if (tcp_pacing_check(sk))
2312 			break;
2313 
2314 		tso_segs = tcp_init_tso_segs(skb, mss_now);
2315 		BUG_ON(!tso_segs);
2316 
2317 		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2318 			/* "skb_mstamp" is used as a start point for the retransmit timer */
2319 			tcp_update_skb_after_send(tp, skb);
2320 			goto repair; /* Skip network transmission */
2321 		}
2322 
2323 		cwnd_quota = tcp_cwnd_test(tp, skb);
2324 		if (!cwnd_quota) {
2325 			if (push_one == 2)
2326 				/* Force out a loss probe pkt. */
2327 				cwnd_quota = 1;
2328 			else
2329 				break;
2330 		}
2331 
2332 		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2333 			is_rwnd_limited = true;
2334 			break;
2335 		}
2336 
2337 		if (tso_segs == 1) {
2338 			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2339 						     (tcp_skb_is_last(sk, skb) ?
2340 						      nonagle : TCP_NAGLE_PUSH))))
2341 				break;
2342 		} else {
2343 			if (!push_one &&
2344 			    tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2345 						 max_segs))
2346 				break;
2347 		}
2348 
2349 		limit = mss_now;
2350 		if (tso_segs > 1 && !tcp_urg_mode(tp))
2351 			limit = tcp_mss_split_point(sk, skb, mss_now,
2352 						    min_t(unsigned int,
2353 							  cwnd_quota,
2354 							  max_segs),
2355 						    nonagle);
2356 
2357 		if (skb->len > limit &&
2358 		    unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2359 					  skb, limit, mss_now, gfp)))
2360 			break;
2361 
2362 		if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
2363 			clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
2364 		if (tcp_small_queue_check(sk, skb, 0))
2365 			break;
2366 
2367 		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2368 			break;
2369 
2370 repair:
2371 		/* Advance the send_head.  This one is sent out.
2372 		 * This call will increment packets_out.
2373 		 */
2374 		tcp_event_new_data_sent(sk, skb);
2375 
2376 		tcp_minshall_update(tp, mss_now, skb);
2377 		sent_pkts += tcp_skb_pcount(skb);
2378 
2379 		if (push_one)
2380 			break;
2381 	}
2382 
2383 	if (is_rwnd_limited)
2384 		tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2385 	else
2386 		tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2387 
2388 	if (likely(sent_pkts)) {
2389 		if (tcp_in_cwnd_reduction(sk))
2390 			tp->prr_out += sent_pkts;
2391 
2392 		/* Send one loss probe per tail loss episode. */
2393 		if (push_one != 2)
2394 			tcp_schedule_loss_probe(sk, false);
2395 		is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2396 		tcp_cwnd_validate(sk, is_cwnd_limited);
2397 		return false;
2398 	}
2399 	return !tp->packets_out && !tcp_write_queue_empty(sk);
2400 }
2401 
2402 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2403 {
2404 	struct inet_connection_sock *icsk = inet_csk(sk);
2405 	struct tcp_sock *tp = tcp_sk(sk);
2406 	u32 timeout, rto_delta_us;
2407 	int early_retrans;
2408 
2409 	/* Don't do any loss probe on a Fast Open connection before 3WHS
2410 	 * finishes.
2411 	 */
2412 	if (tp->fastopen_rsk)
2413 		return false;
2414 
2415 	early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2416 	/* Schedule a loss probe in 2*RTT for SACK capable connections
2417 	 * in Open state, that are either limited by cwnd or application.
2418 	 */
2419 	if ((early_retrans != 3 && early_retrans != 4) ||
2420 	    !tp->packets_out || !tcp_is_sack(tp) ||
2421 	    icsk->icsk_ca_state != TCP_CA_Open)
2422 		return false;
2423 
2424 	if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2425 	     !tcp_write_queue_empty(sk))
2426 		return false;
2427 
2428 	/* Probe timeout is 2*rtt. Add minimum RTO to account
2429 	 * for delayed ack when there's one outstanding packet. If no RTT
2430 	 * sample is available then probe after TCP_TIMEOUT_INIT.
2431 	 */
2432 	if (tp->srtt_us) {
2433 		timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2434 		if (tp->packets_out == 1)
2435 			timeout += TCP_RTO_MIN;
2436 		else
2437 			timeout += TCP_TIMEOUT_MIN;
2438 	} else {
2439 		timeout = TCP_TIMEOUT_INIT;
2440 	}
2441 
2442 	/* If the RTO formula yields an earlier time, then use that time. */
2443 	rto_delta_us = advancing_rto ?
2444 			jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2445 			tcp_rto_delta_us(sk);  /* How far in future is RTO? */
2446 	if (rto_delta_us > 0)
2447 		timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2448 
2449 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2450 				  TCP_RTO_MAX);
2451 	return true;
2452 }
2453 
2454 /* Thanks to skb fast clones, we can detect if a prior transmit of
2455  * a packet is still in a qdisc or driver queue.
2456  * In this case, there is very little point doing a retransmit !
2457  */
2458 static bool skb_still_in_host_queue(const struct sock *sk,
2459 				    const struct sk_buff *skb)
2460 {
2461 	if (unlikely(skb_fclone_busy(sk, skb))) {
2462 		NET_INC_STATS(sock_net(sk),
2463 			      LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2464 		return true;
2465 	}
2466 	return false;
2467 }
2468 
2469 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2470  * retransmit the last segment.
2471  */
2472 void tcp_send_loss_probe(struct sock *sk)
2473 {
2474 	struct tcp_sock *tp = tcp_sk(sk);
2475 	struct sk_buff *skb;
2476 	int pcount;
2477 	int mss = tcp_current_mss(sk);
2478 
2479 	skb = tcp_send_head(sk);
2480 	if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2481 		pcount = tp->packets_out;
2482 		tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2483 		if (tp->packets_out > pcount)
2484 			goto probe_sent;
2485 		goto rearm_timer;
2486 	}
2487 	skb = skb_rb_last(&sk->tcp_rtx_queue);
2488 
2489 	/* At most one outstanding TLP retransmission. */
2490 	if (tp->tlp_high_seq)
2491 		goto rearm_timer;
2492 
2493 	/* Retransmit last segment. */
2494 	if (WARN_ON(!skb))
2495 		goto rearm_timer;
2496 
2497 	if (skb_still_in_host_queue(sk, skb))
2498 		goto rearm_timer;
2499 
2500 	pcount = tcp_skb_pcount(skb);
2501 	if (WARN_ON(!pcount))
2502 		goto rearm_timer;
2503 
2504 	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2505 		if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2506 					  (pcount - 1) * mss, mss,
2507 					  GFP_ATOMIC)))
2508 			goto rearm_timer;
2509 		skb = skb_rb_next(skb);
2510 	}
2511 
2512 	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2513 		goto rearm_timer;
2514 
2515 	if (__tcp_retransmit_skb(sk, skb, 1))
2516 		goto rearm_timer;
2517 
2518 	/* Record snd_nxt for loss detection. */
2519 	tp->tlp_high_seq = tp->snd_nxt;
2520 
2521 probe_sent:
2522 	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2523 	/* Reset s.t. tcp_rearm_rto will restart timer from now */
2524 	inet_csk(sk)->icsk_pending = 0;
2525 rearm_timer:
2526 	tcp_rearm_rto(sk);
2527 }
2528 
2529 /* Push out any pending frames which were held back due to
2530  * TCP_CORK or attempt at coalescing tiny packets.
2531  * The socket must be locked by the caller.
2532  */
2533 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2534 			       int nonagle)
2535 {
2536 	/* If we are closed, the bytes will have to remain here.
2537 	 * In time closedown will finish, we empty the write queue and
2538 	 * all will be happy.
2539 	 */
2540 	if (unlikely(sk->sk_state == TCP_CLOSE))
2541 		return;
2542 
2543 	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2544 			   sk_gfp_mask(sk, GFP_ATOMIC)))
2545 		tcp_check_probe_timer(sk);
2546 }
2547 
2548 /* Send _single_ skb sitting at the send head. This function requires
2549  * true push pending frames to setup probe timer etc.
2550  */
2551 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2552 {
2553 	struct sk_buff *skb = tcp_send_head(sk);
2554 
2555 	BUG_ON(!skb || skb->len < mss_now);
2556 
2557 	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2558 }
2559 
2560 /* This function returns the amount that we can raise the
2561  * usable window based on the following constraints
2562  *
2563  * 1. The window can never be shrunk once it is offered (RFC 793)
2564  * 2. We limit memory per socket
2565  *
2566  * RFC 1122:
2567  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2568  *  RECV.NEXT + RCV.WIN fixed until:
2569  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2570  *
2571  * i.e. don't raise the right edge of the window until you can raise
2572  * it at least MSS bytes.
2573  *
2574  * Unfortunately, the recommended algorithm breaks header prediction,
2575  * since header prediction assumes th->window stays fixed.
2576  *
2577  * Strictly speaking, keeping th->window fixed violates the receiver
2578  * side SWS prevention criteria. The problem is that under this rule
2579  * a stream of single byte packets will cause the right side of the
2580  * window to always advance by a single byte.
2581  *
2582  * Of course, if the sender implements sender side SWS prevention
2583  * then this will not be a problem.
2584  *
2585  * BSD seems to make the following compromise:
2586  *
2587  *	If the free space is less than the 1/4 of the maximum
2588  *	space available and the free space is less than 1/2 mss,
2589  *	then set the window to 0.
2590  *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2591  *	Otherwise, just prevent the window from shrinking
2592  *	and from being larger than the largest representable value.
2593  *
2594  * This prevents incremental opening of the window in the regime
2595  * where TCP is limited by the speed of the reader side taking
2596  * data out of the TCP receive queue. It does nothing about
2597  * those cases where the window is constrained on the sender side
2598  * because the pipeline is full.
2599  *
2600  * BSD also seems to "accidentally" limit itself to windows that are a
2601  * multiple of MSS, at least until the free space gets quite small.
2602  * This would appear to be a side effect of the mbuf implementation.
2603  * Combining these two algorithms results in the observed behavior
2604  * of having a fixed window size at almost all times.
2605  *
2606  * Below we obtain similar behavior by forcing the offered window to
2607  * a multiple of the mss when it is feasible to do so.
2608  *
2609  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2610  * Regular options like TIMESTAMP are taken into account.
2611  */
2612 u32 __tcp_select_window(struct sock *sk)
2613 {
2614 	struct inet_connection_sock *icsk = inet_csk(sk);
2615 	struct tcp_sock *tp = tcp_sk(sk);
2616 	/* MSS for the peer's data.  Previous versions used mss_clamp
2617 	 * here.  I don't know if the value based on our guesses
2618 	 * of peer's MSS is better for the performance.  It's more correct
2619 	 * but may be worse for the performance because of rcv_mss
2620 	 * fluctuations.  --SAW  1998/11/1
2621 	 */
2622 	int mss = icsk->icsk_ack.rcv_mss;
2623 	int free_space = tcp_space(sk);
2624 	int allowed_space = tcp_full_space(sk);
2625 	int full_space = min_t(int, tp->window_clamp, allowed_space);
2626 	int window;
2627 
2628 	if (unlikely(mss > full_space)) {
2629 		mss = full_space;
2630 		if (mss <= 0)
2631 			return 0;
2632 	}
2633 	if (free_space < (full_space >> 1)) {
2634 		icsk->icsk_ack.quick = 0;
2635 
2636 		if (tcp_under_memory_pressure(sk))
2637 			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2638 					       4U * tp->advmss);
2639 
2640 		/* free_space might become our new window, make sure we don't
2641 		 * increase it due to wscale.
2642 		 */
2643 		free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2644 
2645 		/* if free space is less than mss estimate, or is below 1/16th
2646 		 * of the maximum allowed, try to move to zero-window, else
2647 		 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2648 		 * new incoming data is dropped due to memory limits.
2649 		 * With large window, mss test triggers way too late in order
2650 		 * to announce zero window in time before rmem limit kicks in.
2651 		 */
2652 		if (free_space < (allowed_space >> 4) || free_space < mss)
2653 			return 0;
2654 	}
2655 
2656 	if (free_space > tp->rcv_ssthresh)
2657 		free_space = tp->rcv_ssthresh;
2658 
2659 	/* Don't do rounding if we are using window scaling, since the
2660 	 * scaled window will not line up with the MSS boundary anyway.
2661 	 */
2662 	if (tp->rx_opt.rcv_wscale) {
2663 		window = free_space;
2664 
2665 		/* Advertise enough space so that it won't get scaled away.
2666 		 * Import case: prevent zero window announcement if
2667 		 * 1<<rcv_wscale > mss.
2668 		 */
2669 		window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2670 	} else {
2671 		window = tp->rcv_wnd;
2672 		/* Get the largest window that is a nice multiple of mss.
2673 		 * Window clamp already applied above.
2674 		 * If our current window offering is within 1 mss of the
2675 		 * free space we just keep it. This prevents the divide
2676 		 * and multiply from happening most of the time.
2677 		 * We also don't do any window rounding when the free space
2678 		 * is too small.
2679 		 */
2680 		if (window <= free_space - mss || window > free_space)
2681 			window = rounddown(free_space, mss);
2682 		else if (mss == full_space &&
2683 			 free_space > window + (full_space >> 1))
2684 			window = free_space;
2685 	}
2686 
2687 	return window;
2688 }
2689 
2690 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2691 			     const struct sk_buff *next_skb)
2692 {
2693 	if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2694 		const struct skb_shared_info *next_shinfo =
2695 			skb_shinfo(next_skb);
2696 		struct skb_shared_info *shinfo = skb_shinfo(skb);
2697 
2698 		shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2699 		shinfo->tskey = next_shinfo->tskey;
2700 		TCP_SKB_CB(skb)->txstamp_ack |=
2701 			TCP_SKB_CB(next_skb)->txstamp_ack;
2702 	}
2703 }
2704 
2705 /* Collapses two adjacent SKB's during retransmission. */
2706 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2707 {
2708 	struct tcp_sock *tp = tcp_sk(sk);
2709 	struct sk_buff *next_skb = skb_rb_next(skb);
2710 	int skb_size, next_skb_size;
2711 
2712 	skb_size = skb->len;
2713 	next_skb_size = next_skb->len;
2714 
2715 	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2716 
2717 	if (next_skb_size) {
2718 		if (next_skb_size <= skb_availroom(skb))
2719 			skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2720 				      next_skb_size);
2721 		else if (!skb_shift(skb, next_skb, next_skb_size))
2722 			return false;
2723 	}
2724 	tcp_highest_sack_replace(sk, next_skb, skb);
2725 
2726 	if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2727 		skb->ip_summed = CHECKSUM_PARTIAL;
2728 
2729 	if (skb->ip_summed != CHECKSUM_PARTIAL)
2730 		skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2731 
2732 	/* Update sequence range on original skb. */
2733 	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2734 
2735 	/* Merge over control information. This moves PSH/FIN etc. over */
2736 	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2737 
2738 	/* All done, get rid of second SKB and account for it so
2739 	 * packet counting does not break.
2740 	 */
2741 	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2742 	TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2743 
2744 	/* changed transmit queue under us so clear hints */
2745 	tcp_clear_retrans_hints_partial(tp);
2746 	if (next_skb == tp->retransmit_skb_hint)
2747 		tp->retransmit_skb_hint = skb;
2748 
2749 	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2750 
2751 	tcp_skb_collapse_tstamp(skb, next_skb);
2752 
2753 	tcp_rtx_queue_unlink_and_free(next_skb, sk);
2754 	return true;
2755 }
2756 
2757 /* Check if coalescing SKBs is legal. */
2758 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2759 {
2760 	if (tcp_skb_pcount(skb) > 1)
2761 		return false;
2762 	if (skb_cloned(skb))
2763 		return false;
2764 	/* Some heuristics for collapsing over SACK'd could be invented */
2765 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2766 		return false;
2767 
2768 	return true;
2769 }
2770 
2771 /* Collapse packets in the retransmit queue to make to create
2772  * less packets on the wire. This is only done on retransmission.
2773  */
2774 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2775 				     int space)
2776 {
2777 	struct tcp_sock *tp = tcp_sk(sk);
2778 	struct sk_buff *skb = to, *tmp;
2779 	bool first = true;
2780 
2781 	if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2782 		return;
2783 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2784 		return;
2785 
2786 	skb_rbtree_walk_from_safe(skb, tmp) {
2787 		if (!tcp_can_collapse(sk, skb))
2788 			break;
2789 
2790 		if (!tcp_skb_can_collapse_to(to))
2791 			break;
2792 
2793 		space -= skb->len;
2794 
2795 		if (first) {
2796 			first = false;
2797 			continue;
2798 		}
2799 
2800 		if (space < 0)
2801 			break;
2802 
2803 		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2804 			break;
2805 
2806 		if (!tcp_collapse_retrans(sk, to))
2807 			break;
2808 	}
2809 }
2810 
2811 /* This retransmits one SKB.  Policy decisions and retransmit queue
2812  * state updates are done by the caller.  Returns non-zero if an
2813  * error occurred which prevented the send.
2814  */
2815 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2816 {
2817 	struct inet_connection_sock *icsk = inet_csk(sk);
2818 	struct tcp_sock *tp = tcp_sk(sk);
2819 	unsigned int cur_mss;
2820 	int diff, len, err;
2821 
2822 
2823 	/* Inconclusive MTU probe */
2824 	if (icsk->icsk_mtup.probe_size)
2825 		icsk->icsk_mtup.probe_size = 0;
2826 
2827 	/* Do not sent more than we queued. 1/4 is reserved for possible
2828 	 * copying overhead: fragmentation, tunneling, mangling etc.
2829 	 */
2830 	if (refcount_read(&sk->sk_wmem_alloc) >
2831 	    min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2832 		  sk->sk_sndbuf))
2833 		return -EAGAIN;
2834 
2835 	if (skb_still_in_host_queue(sk, skb))
2836 		return -EBUSY;
2837 
2838 	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2839 		if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2840 			BUG();
2841 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2842 			return -ENOMEM;
2843 	}
2844 
2845 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2846 		return -EHOSTUNREACH; /* Routing failure or similar. */
2847 
2848 	cur_mss = tcp_current_mss(sk);
2849 
2850 	/* If receiver has shrunk his window, and skb is out of
2851 	 * new window, do not retransmit it. The exception is the
2852 	 * case, when window is shrunk to zero. In this case
2853 	 * our retransmit serves as a zero window probe.
2854 	 */
2855 	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2856 	    TCP_SKB_CB(skb)->seq != tp->snd_una)
2857 		return -EAGAIN;
2858 
2859 	len = cur_mss * segs;
2860 	if (skb->len > len) {
2861 		if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2862 				 cur_mss, GFP_ATOMIC))
2863 			return -ENOMEM; /* We'll try again later. */
2864 	} else {
2865 		if (skb_unclone(skb, GFP_ATOMIC))
2866 			return -ENOMEM;
2867 
2868 		diff = tcp_skb_pcount(skb);
2869 		tcp_set_skb_tso_segs(skb, cur_mss);
2870 		diff -= tcp_skb_pcount(skb);
2871 		if (diff)
2872 			tcp_adjust_pcount(sk, skb, diff);
2873 		if (skb->len < cur_mss)
2874 			tcp_retrans_try_collapse(sk, skb, cur_mss);
2875 	}
2876 
2877 	/* RFC3168, section 6.1.1.1. ECN fallback */
2878 	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2879 		tcp_ecn_clear_syn(sk, skb);
2880 
2881 	/* Update global and local TCP statistics. */
2882 	segs = tcp_skb_pcount(skb);
2883 	TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2884 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2885 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2886 	tp->total_retrans += segs;
2887 
2888 	/* make sure skb->data is aligned on arches that require it
2889 	 * and check if ack-trimming & collapsing extended the headroom
2890 	 * beyond what csum_start can cover.
2891 	 */
2892 	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2893 		     skb_headroom(skb) >= 0xFFFF)) {
2894 		struct sk_buff *nskb;
2895 
2896 		tcp_skb_tsorted_save(skb) {
2897 			nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2898 			err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2899 				     -ENOBUFS;
2900 		} tcp_skb_tsorted_restore(skb);
2901 
2902 		if (!err) {
2903 			tcp_update_skb_after_send(tp, skb);
2904 			tcp_rate_skb_sent(sk, skb);
2905 		}
2906 	} else {
2907 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2908 	}
2909 
2910 	if (likely(!err)) {
2911 		TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2912 		trace_tcp_retransmit_skb(sk, skb);
2913 	} else if (err != -EBUSY) {
2914 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2915 	}
2916 	return err;
2917 }
2918 
2919 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2920 {
2921 	struct tcp_sock *tp = tcp_sk(sk);
2922 	int err = __tcp_retransmit_skb(sk, skb, segs);
2923 
2924 	if (err == 0) {
2925 #if FASTRETRANS_DEBUG > 0
2926 		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2927 			net_dbg_ratelimited("retrans_out leaked\n");
2928 		}
2929 #endif
2930 		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2931 		tp->retrans_out += tcp_skb_pcount(skb);
2932 
2933 		/* Save stamp of the first retransmit. */
2934 		if (!tp->retrans_stamp)
2935 			tp->retrans_stamp = tcp_skb_timestamp(skb);
2936 
2937 	}
2938 
2939 	if (tp->undo_retrans < 0)
2940 		tp->undo_retrans = 0;
2941 	tp->undo_retrans += tcp_skb_pcount(skb);
2942 	return err;
2943 }
2944 
2945 /* This gets called after a retransmit timeout, and the initially
2946  * retransmitted data is acknowledged.  It tries to continue
2947  * resending the rest of the retransmit queue, until either
2948  * we've sent it all or the congestion window limit is reached.
2949  */
2950 void tcp_xmit_retransmit_queue(struct sock *sk)
2951 {
2952 	const struct inet_connection_sock *icsk = inet_csk(sk);
2953 	struct sk_buff *skb, *rtx_head, *hole = NULL;
2954 	struct tcp_sock *tp = tcp_sk(sk);
2955 	u32 max_segs;
2956 	int mib_idx;
2957 
2958 	if (!tp->packets_out)
2959 		return;
2960 
2961 	rtx_head = tcp_rtx_queue_head(sk);
2962 	skb = tp->retransmit_skb_hint ?: rtx_head;
2963 	max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2964 	skb_rbtree_walk_from(skb) {
2965 		__u8 sacked;
2966 		int segs;
2967 
2968 		if (tcp_pacing_check(sk))
2969 			break;
2970 
2971 		/* we could do better than to assign each time */
2972 		if (!hole)
2973 			tp->retransmit_skb_hint = skb;
2974 
2975 		segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2976 		if (segs <= 0)
2977 			return;
2978 		sacked = TCP_SKB_CB(skb)->sacked;
2979 		/* In case tcp_shift_skb_data() have aggregated large skbs,
2980 		 * we need to make sure not sending too bigs TSO packets
2981 		 */
2982 		segs = min_t(int, segs, max_segs);
2983 
2984 		if (tp->retrans_out >= tp->lost_out) {
2985 			break;
2986 		} else if (!(sacked & TCPCB_LOST)) {
2987 			if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2988 				hole = skb;
2989 			continue;
2990 
2991 		} else {
2992 			if (icsk->icsk_ca_state != TCP_CA_Loss)
2993 				mib_idx = LINUX_MIB_TCPFASTRETRANS;
2994 			else
2995 				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2996 		}
2997 
2998 		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2999 			continue;
3000 
3001 		if (tcp_small_queue_check(sk, skb, 1))
3002 			return;
3003 
3004 		if (tcp_retransmit_skb(sk, skb, segs))
3005 			return;
3006 
3007 		NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3008 
3009 		if (tcp_in_cwnd_reduction(sk))
3010 			tp->prr_out += tcp_skb_pcount(skb);
3011 
3012 		if (skb == rtx_head &&
3013 		    icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3014 			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3015 						  inet_csk(sk)->icsk_rto,
3016 						  TCP_RTO_MAX);
3017 	}
3018 }
3019 
3020 /* We allow to exceed memory limits for FIN packets to expedite
3021  * connection tear down and (memory) recovery.
3022  * Otherwise tcp_send_fin() could be tempted to either delay FIN
3023  * or even be forced to close flow without any FIN.
3024  * In general, we want to allow one skb per socket to avoid hangs
3025  * with edge trigger epoll()
3026  */
3027 void sk_forced_mem_schedule(struct sock *sk, int size)
3028 {
3029 	int amt;
3030 
3031 	if (size <= sk->sk_forward_alloc)
3032 		return;
3033 	amt = sk_mem_pages(size);
3034 	sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3035 	sk_memory_allocated_add(sk, amt);
3036 
3037 	if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3038 		mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3039 }
3040 
3041 /* Send a FIN. The caller locks the socket for us.
3042  * We should try to send a FIN packet really hard, but eventually give up.
3043  */
3044 void tcp_send_fin(struct sock *sk)
3045 {
3046 	struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3047 	struct tcp_sock *tp = tcp_sk(sk);
3048 
3049 	/* Optimization, tack on the FIN if we have one skb in write queue and
3050 	 * this skb was not yet sent, or we are under memory pressure.
3051 	 * Note: in the latter case, FIN packet will be sent after a timeout,
3052 	 * as TCP stack thinks it has already been transmitted.
3053 	 */
3054 	if (!tskb && tcp_under_memory_pressure(sk))
3055 		tskb = skb_rb_last(&sk->tcp_rtx_queue);
3056 
3057 	if (tskb) {
3058 coalesce:
3059 		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3060 		TCP_SKB_CB(tskb)->end_seq++;
3061 		tp->write_seq++;
3062 		if (tcp_write_queue_empty(sk)) {
3063 			/* This means tskb was already sent.
3064 			 * Pretend we included the FIN on previous transmit.
3065 			 * We need to set tp->snd_nxt to the value it would have
3066 			 * if FIN had been sent. This is because retransmit path
3067 			 * does not change tp->snd_nxt.
3068 			 */
3069 			tp->snd_nxt++;
3070 			return;
3071 		}
3072 	} else {
3073 		skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3074 		if (unlikely(!skb)) {
3075 			if (tskb)
3076 				goto coalesce;
3077 			return;
3078 		}
3079 		INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3080 		skb_reserve(skb, MAX_TCP_HEADER);
3081 		sk_forced_mem_schedule(sk, skb->truesize);
3082 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3083 		tcp_init_nondata_skb(skb, tp->write_seq,
3084 				     TCPHDR_ACK | TCPHDR_FIN);
3085 		tcp_queue_skb(sk, skb);
3086 	}
3087 	__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3088 }
3089 
3090 /* We get here when a process closes a file descriptor (either due to
3091  * an explicit close() or as a byproduct of exit()'ing) and there
3092  * was unread data in the receive queue.  This behavior is recommended
3093  * by RFC 2525, section 2.17.  -DaveM
3094  */
3095 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3096 {
3097 	struct sk_buff *skb;
3098 
3099 	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3100 
3101 	/* NOTE: No TCP options attached and we never retransmit this. */
3102 	skb = alloc_skb(MAX_TCP_HEADER, priority);
3103 	if (!skb) {
3104 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3105 		return;
3106 	}
3107 
3108 	/* Reserve space for headers and prepare control bits. */
3109 	skb_reserve(skb, MAX_TCP_HEADER);
3110 	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3111 			     TCPHDR_ACK | TCPHDR_RST);
3112 	tcp_mstamp_refresh(tcp_sk(sk));
3113 	/* Send it off. */
3114 	if (tcp_transmit_skb(sk, skb, 0, priority))
3115 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3116 
3117 	/* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3118 	 * skb here is different to the troublesome skb, so use NULL
3119 	 */
3120 	trace_tcp_send_reset(sk, NULL);
3121 }
3122 
3123 /* Send a crossed SYN-ACK during socket establishment.
3124  * WARNING: This routine must only be called when we have already sent
3125  * a SYN packet that crossed the incoming SYN that caused this routine
3126  * to get called. If this assumption fails then the initial rcv_wnd
3127  * and rcv_wscale values will not be correct.
3128  */
3129 int tcp_send_synack(struct sock *sk)
3130 {
3131 	struct sk_buff *skb;
3132 
3133 	skb = tcp_rtx_queue_head(sk);
3134 	if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3135 		pr_err("%s: wrong queue state\n", __func__);
3136 		return -EFAULT;
3137 	}
3138 	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3139 		if (skb_cloned(skb)) {
3140 			struct sk_buff *nskb;
3141 
3142 			tcp_skb_tsorted_save(skb) {
3143 				nskb = skb_copy(skb, GFP_ATOMIC);
3144 			} tcp_skb_tsorted_restore(skb);
3145 			if (!nskb)
3146 				return -ENOMEM;
3147 			INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3148 			tcp_rtx_queue_unlink_and_free(skb, sk);
3149 			__skb_header_release(nskb);
3150 			tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3151 			sk->sk_wmem_queued += nskb->truesize;
3152 			sk_mem_charge(sk, nskb->truesize);
3153 			skb = nskb;
3154 		}
3155 
3156 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3157 		tcp_ecn_send_synack(sk, skb);
3158 	}
3159 	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3160 }
3161 
3162 /**
3163  * tcp_make_synack - Prepare a SYN-ACK.
3164  * sk: listener socket
3165  * dst: dst entry attached to the SYNACK
3166  * req: request_sock pointer
3167  *
3168  * Allocate one skb and build a SYNACK packet.
3169  * @dst is consumed : Caller should not use it again.
3170  */
3171 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3172 				struct request_sock *req,
3173 				struct tcp_fastopen_cookie *foc,
3174 				enum tcp_synack_type synack_type)
3175 {
3176 	struct inet_request_sock *ireq = inet_rsk(req);
3177 	const struct tcp_sock *tp = tcp_sk(sk);
3178 	struct tcp_md5sig_key *md5 = NULL;
3179 	struct tcp_out_options opts;
3180 	struct sk_buff *skb;
3181 	int tcp_header_size;
3182 	struct tcphdr *th;
3183 	int mss;
3184 
3185 	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3186 	if (unlikely(!skb)) {
3187 		dst_release(dst);
3188 		return NULL;
3189 	}
3190 	/* Reserve space for headers. */
3191 	skb_reserve(skb, MAX_TCP_HEADER);
3192 
3193 	switch (synack_type) {
3194 	case TCP_SYNACK_NORMAL:
3195 		skb_set_owner_w(skb, req_to_sk(req));
3196 		break;
3197 	case TCP_SYNACK_COOKIE:
3198 		/* Under synflood, we do not attach skb to a socket,
3199 		 * to avoid false sharing.
3200 		 */
3201 		break;
3202 	case TCP_SYNACK_FASTOPEN:
3203 		/* sk is a const pointer, because we want to express multiple
3204 		 * cpu might call us concurrently.
3205 		 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3206 		 */
3207 		skb_set_owner_w(skb, (struct sock *)sk);
3208 		break;
3209 	}
3210 	skb_dst_set(skb, dst);
3211 
3212 	mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3213 
3214 	memset(&opts, 0, sizeof(opts));
3215 #ifdef CONFIG_SYN_COOKIES
3216 	if (unlikely(req->cookie_ts))
3217 		skb->skb_mstamp = cookie_init_timestamp(req);
3218 	else
3219 #endif
3220 		skb->skb_mstamp = tcp_clock_us();
3221 
3222 #ifdef CONFIG_TCP_MD5SIG
3223 	rcu_read_lock();
3224 	md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3225 #endif
3226 	skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3227 	tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3228 					     foc) + sizeof(*th);
3229 
3230 	skb_push(skb, tcp_header_size);
3231 	skb_reset_transport_header(skb);
3232 
3233 	th = (struct tcphdr *)skb->data;
3234 	memset(th, 0, sizeof(struct tcphdr));
3235 	th->syn = 1;
3236 	th->ack = 1;
3237 	tcp_ecn_make_synack(req, th);
3238 	th->source = htons(ireq->ir_num);
3239 	th->dest = ireq->ir_rmt_port;
3240 	skb->mark = ireq->ir_mark;
3241 	skb->ip_summed = CHECKSUM_PARTIAL;
3242 	th->seq = htonl(tcp_rsk(req)->snt_isn);
3243 	/* XXX data is queued and acked as is. No buffer/window check */
3244 	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3245 
3246 	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3247 	th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3248 	tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3249 	th->doff = (tcp_header_size >> 2);
3250 	__TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3251 
3252 #ifdef CONFIG_TCP_MD5SIG
3253 	/* Okay, we have all we need - do the md5 hash if needed */
3254 	if (md5)
3255 		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3256 					       md5, req_to_sk(req), skb);
3257 	rcu_read_unlock();
3258 #endif
3259 
3260 	/* Do not fool tcpdump (if any), clean our debris */
3261 	skb->tstamp = 0;
3262 	return skb;
3263 }
3264 EXPORT_SYMBOL(tcp_make_synack);
3265 
3266 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3267 {
3268 	struct inet_connection_sock *icsk = inet_csk(sk);
3269 	const struct tcp_congestion_ops *ca;
3270 	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3271 
3272 	if (ca_key == TCP_CA_UNSPEC)
3273 		return;
3274 
3275 	rcu_read_lock();
3276 	ca = tcp_ca_find_key(ca_key);
3277 	if (likely(ca && try_module_get(ca->owner))) {
3278 		module_put(icsk->icsk_ca_ops->owner);
3279 		icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3280 		icsk->icsk_ca_ops = ca;
3281 	}
3282 	rcu_read_unlock();
3283 }
3284 
3285 /* Do all connect socket setups that can be done AF independent. */
3286 static void tcp_connect_init(struct sock *sk)
3287 {
3288 	const struct dst_entry *dst = __sk_dst_get(sk);
3289 	struct tcp_sock *tp = tcp_sk(sk);
3290 	__u8 rcv_wscale;
3291 	u32 rcv_wnd;
3292 
3293 	/* We'll fix this up when we get a response from the other end.
3294 	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3295 	 */
3296 	tp->tcp_header_len = sizeof(struct tcphdr);
3297 	if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3298 		tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3299 
3300 #ifdef CONFIG_TCP_MD5SIG
3301 	if (tp->af_specific->md5_lookup(sk, sk))
3302 		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3303 #endif
3304 
3305 	/* If user gave his TCP_MAXSEG, record it to clamp */
3306 	if (tp->rx_opt.user_mss)
3307 		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3308 	tp->max_window = 0;
3309 	tcp_mtup_init(sk);
3310 	tcp_sync_mss(sk, dst_mtu(dst));
3311 
3312 	tcp_ca_dst_init(sk, dst);
3313 
3314 	if (!tp->window_clamp)
3315 		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3316 	tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3317 
3318 	tcp_initialize_rcv_mss(sk);
3319 
3320 	/* limit the window selection if the user enforce a smaller rx buffer */
3321 	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3322 	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3323 		tp->window_clamp = tcp_full_space(sk);
3324 
3325 	rcv_wnd = tcp_rwnd_init_bpf(sk);
3326 	if (rcv_wnd == 0)
3327 		rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3328 
3329 	tcp_select_initial_window(sk, tcp_full_space(sk),
3330 				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3331 				  &tp->rcv_wnd,
3332 				  &tp->window_clamp,
3333 				  sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3334 				  &rcv_wscale,
3335 				  rcv_wnd);
3336 
3337 	tp->rx_opt.rcv_wscale = rcv_wscale;
3338 	tp->rcv_ssthresh = tp->rcv_wnd;
3339 
3340 	sk->sk_err = 0;
3341 	sock_reset_flag(sk, SOCK_DONE);
3342 	tp->snd_wnd = 0;
3343 	tcp_init_wl(tp, 0);
3344 	tp->snd_una = tp->write_seq;
3345 	tp->snd_sml = tp->write_seq;
3346 	tp->snd_up = tp->write_seq;
3347 	tp->snd_nxt = tp->write_seq;
3348 
3349 	if (likely(!tp->repair))
3350 		tp->rcv_nxt = 0;
3351 	else
3352 		tp->rcv_tstamp = tcp_jiffies32;
3353 	tp->rcv_wup = tp->rcv_nxt;
3354 	tp->copied_seq = tp->rcv_nxt;
3355 
3356 	inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3357 	inet_csk(sk)->icsk_retransmits = 0;
3358 	tcp_clear_retrans(tp);
3359 }
3360 
3361 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3362 {
3363 	struct tcp_sock *tp = tcp_sk(sk);
3364 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3365 
3366 	tcb->end_seq += skb->len;
3367 	__skb_header_release(skb);
3368 	sk->sk_wmem_queued += skb->truesize;
3369 	sk_mem_charge(sk, skb->truesize);
3370 	tp->write_seq = tcb->end_seq;
3371 	tp->packets_out += tcp_skb_pcount(skb);
3372 }
3373 
3374 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3375  * queue a data-only packet after the regular SYN, such that regular SYNs
3376  * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3377  * only the SYN sequence, the data are retransmitted in the first ACK.
3378  * If cookie is not cached or other error occurs, falls back to send a
3379  * regular SYN with Fast Open cookie request option.
3380  */
3381 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3382 {
3383 	struct tcp_sock *tp = tcp_sk(sk);
3384 	struct tcp_fastopen_request *fo = tp->fastopen_req;
3385 	int space, err = 0;
3386 	struct sk_buff *syn_data;
3387 
3388 	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3389 	if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3390 		goto fallback;
3391 
3392 	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3393 	 * user-MSS. Reserve maximum option space for middleboxes that add
3394 	 * private TCP options. The cost is reduced data space in SYN :(
3395 	 */
3396 	tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3397 
3398 	space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3399 		MAX_TCP_OPTION_SPACE;
3400 
3401 	space = min_t(size_t, space, fo->size);
3402 
3403 	/* limit to order-0 allocations */
3404 	space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3405 
3406 	syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3407 	if (!syn_data)
3408 		goto fallback;
3409 	syn_data->ip_summed = CHECKSUM_PARTIAL;
3410 	memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3411 	if (space) {
3412 		int copied = copy_from_iter(skb_put(syn_data, space), space,
3413 					    &fo->data->msg_iter);
3414 		if (unlikely(!copied)) {
3415 			tcp_skb_tsorted_anchor_cleanup(syn_data);
3416 			kfree_skb(syn_data);
3417 			goto fallback;
3418 		}
3419 		if (copied != space) {
3420 			skb_trim(syn_data, copied);
3421 			space = copied;
3422 		}
3423 	}
3424 	/* No more data pending in inet_wait_for_connect() */
3425 	if (space == fo->size)
3426 		fo->data = NULL;
3427 	fo->copied = space;
3428 
3429 	tcp_connect_queue_skb(sk, syn_data);
3430 	if (syn_data->len)
3431 		tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3432 
3433 	err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3434 
3435 	syn->skb_mstamp = syn_data->skb_mstamp;
3436 
3437 	/* Now full SYN+DATA was cloned and sent (or not),
3438 	 * remove the SYN from the original skb (syn_data)
3439 	 * we keep in write queue in case of a retransmit, as we
3440 	 * also have the SYN packet (with no data) in the same queue.
3441 	 */
3442 	TCP_SKB_CB(syn_data)->seq++;
3443 	TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3444 	if (!err) {
3445 		tp->syn_data = (fo->copied > 0);
3446 		tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3447 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3448 		goto done;
3449 	}
3450 
3451 	/* data was not sent, put it in write_queue */
3452 	__skb_queue_tail(&sk->sk_write_queue, syn_data);
3453 	tp->packets_out -= tcp_skb_pcount(syn_data);
3454 
3455 fallback:
3456 	/* Send a regular SYN with Fast Open cookie request option */
3457 	if (fo->cookie.len > 0)
3458 		fo->cookie.len = 0;
3459 	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3460 	if (err)
3461 		tp->syn_fastopen = 0;
3462 done:
3463 	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3464 	return err;
3465 }
3466 
3467 /* Build a SYN and send it off. */
3468 int tcp_connect(struct sock *sk)
3469 {
3470 	struct tcp_sock *tp = tcp_sk(sk);
3471 	struct sk_buff *buff;
3472 	int err;
3473 
3474 	tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB);
3475 
3476 	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3477 		return -EHOSTUNREACH; /* Routing failure or similar. */
3478 
3479 	tcp_connect_init(sk);
3480 
3481 	if (unlikely(tp->repair)) {
3482 		tcp_finish_connect(sk, NULL);
3483 		return 0;
3484 	}
3485 
3486 	buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3487 	if (unlikely(!buff))
3488 		return -ENOBUFS;
3489 
3490 	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3491 	tcp_mstamp_refresh(tp);
3492 	tp->retrans_stamp = tcp_time_stamp(tp);
3493 	tcp_connect_queue_skb(sk, buff);
3494 	tcp_ecn_send_syn(sk, buff);
3495 	tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3496 
3497 	/* Send off SYN; include data in Fast Open. */
3498 	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3499 	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3500 	if (err == -ECONNREFUSED)
3501 		return err;
3502 
3503 	/* We change tp->snd_nxt after the tcp_transmit_skb() call
3504 	 * in order to make this packet get counted in tcpOutSegs.
3505 	 */
3506 	tp->snd_nxt = tp->write_seq;
3507 	tp->pushed_seq = tp->write_seq;
3508 	buff = tcp_send_head(sk);
3509 	if (unlikely(buff)) {
3510 		tp->snd_nxt	= TCP_SKB_CB(buff)->seq;
3511 		tp->pushed_seq	= TCP_SKB_CB(buff)->seq;
3512 	}
3513 	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3514 
3515 	/* Timer for repeating the SYN until an answer. */
3516 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3517 				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3518 	return 0;
3519 }
3520 EXPORT_SYMBOL(tcp_connect);
3521 
3522 /* Send out a delayed ack, the caller does the policy checking
3523  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3524  * for details.
3525  */
3526 void tcp_send_delayed_ack(struct sock *sk)
3527 {
3528 	struct inet_connection_sock *icsk = inet_csk(sk);
3529 	int ato = icsk->icsk_ack.ato;
3530 	unsigned long timeout;
3531 
3532 	tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3533 
3534 	if (ato > TCP_DELACK_MIN) {
3535 		const struct tcp_sock *tp = tcp_sk(sk);
3536 		int max_ato = HZ / 2;
3537 
3538 		if (icsk->icsk_ack.pingpong ||
3539 		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3540 			max_ato = TCP_DELACK_MAX;
3541 
3542 		/* Slow path, intersegment interval is "high". */
3543 
3544 		/* If some rtt estimate is known, use it to bound delayed ack.
3545 		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3546 		 * directly.
3547 		 */
3548 		if (tp->srtt_us) {
3549 			int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3550 					TCP_DELACK_MIN);
3551 
3552 			if (rtt < max_ato)
3553 				max_ato = rtt;
3554 		}
3555 
3556 		ato = min(ato, max_ato);
3557 	}
3558 
3559 	/* Stay within the limit we were given */
3560 	timeout = jiffies + ato;
3561 
3562 	/* Use new timeout only if there wasn't a older one earlier. */
3563 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3564 		/* If delack timer was blocked or is about to expire,
3565 		 * send ACK now.
3566 		 */
3567 		if (icsk->icsk_ack.blocked ||
3568 		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3569 			tcp_send_ack(sk);
3570 			return;
3571 		}
3572 
3573 		if (!time_before(timeout, icsk->icsk_ack.timeout))
3574 			timeout = icsk->icsk_ack.timeout;
3575 	}
3576 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3577 	icsk->icsk_ack.timeout = timeout;
3578 	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3579 }
3580 
3581 /* This routine sends an ack and also updates the window. */
3582 void tcp_send_ack(struct sock *sk)
3583 {
3584 	struct sk_buff *buff;
3585 
3586 	/* If we have been reset, we may not send again. */
3587 	if (sk->sk_state == TCP_CLOSE)
3588 		return;
3589 
3590 	tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3591 
3592 	/* We are not putting this on the write queue, so
3593 	 * tcp_transmit_skb() will set the ownership to this
3594 	 * sock.
3595 	 */
3596 	buff = alloc_skb(MAX_TCP_HEADER,
3597 			 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3598 	if (unlikely(!buff)) {
3599 		inet_csk_schedule_ack(sk);
3600 		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3601 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3602 					  TCP_DELACK_MAX, TCP_RTO_MAX);
3603 		return;
3604 	}
3605 
3606 	/* Reserve space for headers and prepare control bits. */
3607 	skb_reserve(buff, MAX_TCP_HEADER);
3608 	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3609 
3610 	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
3611 	 * too much.
3612 	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3613 	 */
3614 	skb_set_tcp_pure_ack(buff);
3615 
3616 	/* Send it off, this clears delayed acks for us. */
3617 	tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3618 }
3619 EXPORT_SYMBOL_GPL(tcp_send_ack);
3620 
3621 /* This routine sends a packet with an out of date sequence
3622  * number. It assumes the other end will try to ack it.
3623  *
3624  * Question: what should we make while urgent mode?
3625  * 4.4BSD forces sending single byte of data. We cannot send
3626  * out of window data, because we have SND.NXT==SND.MAX...
3627  *
3628  * Current solution: to send TWO zero-length segments in urgent mode:
3629  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3630  * out-of-date with SND.UNA-1 to probe window.
3631  */
3632 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3633 {
3634 	struct tcp_sock *tp = tcp_sk(sk);
3635 	struct sk_buff *skb;
3636 
3637 	/* We don't queue it, tcp_transmit_skb() sets ownership. */
3638 	skb = alloc_skb(MAX_TCP_HEADER,
3639 			sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3640 	if (!skb)
3641 		return -1;
3642 
3643 	/* Reserve space for headers and set control bits. */
3644 	skb_reserve(skb, MAX_TCP_HEADER);
3645 	/* Use a previous sequence.  This should cause the other
3646 	 * end to send an ack.  Don't queue or clone SKB, just
3647 	 * send it.
3648 	 */
3649 	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3650 	NET_INC_STATS(sock_net(sk), mib);
3651 	return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3652 }
3653 
3654 /* Called from setsockopt( ... TCP_REPAIR ) */
3655 void tcp_send_window_probe(struct sock *sk)
3656 {
3657 	if (sk->sk_state == TCP_ESTABLISHED) {
3658 		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3659 		tcp_mstamp_refresh(tcp_sk(sk));
3660 		tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3661 	}
3662 }
3663 
3664 /* Initiate keepalive or window probe from timer. */
3665 int tcp_write_wakeup(struct sock *sk, int mib)
3666 {
3667 	struct tcp_sock *tp = tcp_sk(sk);
3668 	struct sk_buff *skb;
3669 
3670 	if (sk->sk_state == TCP_CLOSE)
3671 		return -1;
3672 
3673 	skb = tcp_send_head(sk);
3674 	if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3675 		int err;
3676 		unsigned int mss = tcp_current_mss(sk);
3677 		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3678 
3679 		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3680 			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3681 
3682 		/* We are probing the opening of a window
3683 		 * but the window size is != 0
3684 		 * must have been a result SWS avoidance ( sender )
3685 		 */
3686 		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3687 		    skb->len > mss) {
3688 			seg_size = min(seg_size, mss);
3689 			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3690 			if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3691 					 skb, seg_size, mss, GFP_ATOMIC))
3692 				return -1;
3693 		} else if (!tcp_skb_pcount(skb))
3694 			tcp_set_skb_tso_segs(skb, mss);
3695 
3696 		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3697 		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3698 		if (!err)
3699 			tcp_event_new_data_sent(sk, skb);
3700 		return err;
3701 	} else {
3702 		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3703 			tcp_xmit_probe_skb(sk, 1, mib);
3704 		return tcp_xmit_probe_skb(sk, 0, mib);
3705 	}
3706 }
3707 
3708 /* A window probe timeout has occurred.  If window is not closed send
3709  * a partial packet else a zero probe.
3710  */
3711 void tcp_send_probe0(struct sock *sk)
3712 {
3713 	struct inet_connection_sock *icsk = inet_csk(sk);
3714 	struct tcp_sock *tp = tcp_sk(sk);
3715 	struct net *net = sock_net(sk);
3716 	unsigned long probe_max;
3717 	int err;
3718 
3719 	err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3720 
3721 	if (tp->packets_out || tcp_write_queue_empty(sk)) {
3722 		/* Cancel probe timer, if it is not required. */
3723 		icsk->icsk_probes_out = 0;
3724 		icsk->icsk_backoff = 0;
3725 		return;
3726 	}
3727 
3728 	if (err <= 0) {
3729 		if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3730 			icsk->icsk_backoff++;
3731 		icsk->icsk_probes_out++;
3732 		probe_max = TCP_RTO_MAX;
3733 	} else {
3734 		/* If packet was not sent due to local congestion,
3735 		 * do not backoff and do not remember icsk_probes_out.
3736 		 * Let local senders to fight for local resources.
3737 		 *
3738 		 * Use accumulated backoff yet.
3739 		 */
3740 		if (!icsk->icsk_probes_out)
3741 			icsk->icsk_probes_out = 1;
3742 		probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3743 	}
3744 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3745 				  tcp_probe0_when(sk, probe_max),
3746 				  TCP_RTO_MAX);
3747 }
3748 
3749 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3750 {
3751 	const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3752 	struct flowi fl;
3753 	int res;
3754 
3755 	tcp_rsk(req)->txhash = net_tx_rndhash();
3756 	res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3757 	if (!res) {
3758 		__TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3759 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3760 		if (unlikely(tcp_passive_fastopen(sk)))
3761 			tcp_sk(sk)->total_retrans++;
3762 		trace_tcp_retransmit_synack(sk, req);
3763 	}
3764 	return res;
3765 }
3766 EXPORT_SYMBOL(tcp_rtx_synack);
3767