xref: /linux/net/ipv4/tcp.c (revision f7308991bfeea3f6a4c6281c64fc1ba9dc6e56b3)
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  * Fixes:
21  *		Alan Cox	:	Numerous verify_area() calls
22  *		Alan Cox	:	Set the ACK bit on a reset
23  *		Alan Cox	:	Stopped it crashing if it closed while
24  *					sk->inuse=1 and was trying to connect
25  *					(tcp_err()).
26  *		Alan Cox	:	All icmp error handling was broken
27  *					pointers passed where wrong and the
28  *					socket was looked up backwards. Nobody
29  *					tested any icmp error code obviously.
30  *		Alan Cox	:	tcp_err() now handled properly. It
31  *					wakes people on errors. poll
32  *					behaves and the icmp error race
33  *					has gone by moving it into sock.c
34  *		Alan Cox	:	tcp_send_reset() fixed to work for
35  *					everything not just packets for
36  *					unknown sockets.
37  *		Alan Cox	:	tcp option processing.
38  *		Alan Cox	:	Reset tweaked (still not 100%) [Had
39  *					syn rule wrong]
40  *		Herp Rosmanith  :	More reset fixes
41  *		Alan Cox	:	No longer acks invalid rst frames.
42  *					Acking any kind of RST is right out.
43  *		Alan Cox	:	Sets an ignore me flag on an rst
44  *					receive otherwise odd bits of prattle
45  *					escape still
46  *		Alan Cox	:	Fixed another acking RST frame bug.
47  *					Should stop LAN workplace lockups.
48  *		Alan Cox	: 	Some tidyups using the new skb list
49  *					facilities
50  *		Alan Cox	:	sk->keepopen now seems to work
51  *		Alan Cox	:	Pulls options out correctly on accepts
52  *		Alan Cox	:	Fixed assorted sk->rqueue->next errors
53  *		Alan Cox	:	PSH doesn't end a TCP read. Switched a
54  *					bit to skb ops.
55  *		Alan Cox	:	Tidied tcp_data to avoid a potential
56  *					nasty.
57  *		Alan Cox	:	Added some better commenting, as the
58  *					tcp is hard to follow
59  *		Alan Cox	:	Removed incorrect check for 20 * psh
60  *	Michael O'Reilly	:	ack < copied bug fix.
61  *	Johannes Stille		:	Misc tcp fixes (not all in yet).
62  *		Alan Cox	:	FIN with no memory -> CRASH
63  *		Alan Cox	:	Added socket option proto entries.
64  *					Also added awareness of them to accept.
65  *		Alan Cox	:	Added TCP options (SOL_TCP)
66  *		Alan Cox	:	Switched wakeup calls to callbacks,
67  *					so the kernel can layer network
68  *					sockets.
69  *		Alan Cox	:	Use ip_tos/ip_ttl settings.
70  *		Alan Cox	:	Handle FIN (more) properly (we hope).
71  *		Alan Cox	:	RST frames sent on unsynchronised
72  *					state ack error.
73  *		Alan Cox	:	Put in missing check for SYN bit.
74  *		Alan Cox	:	Added tcp_select_window() aka NET2E
75  *					window non shrink trick.
76  *		Alan Cox	:	Added a couple of small NET2E timer
77  *					fixes
78  *		Charles Hedrick :	TCP fixes
79  *		Toomas Tamm	:	TCP window fixes
80  *		Alan Cox	:	Small URG fix to rlogin ^C ack fight
81  *		Charles Hedrick	:	Rewrote most of it to actually work
82  *		Linus		:	Rewrote tcp_read() and URG handling
83  *					completely
84  *		Gerhard Koerting:	Fixed some missing timer handling
85  *		Matthew Dillon  :	Reworked TCP machine states as per RFC
86  *		Gerhard Koerting:	PC/TCP workarounds
87  *		Adam Caldwell	:	Assorted timer/timing errors
88  *		Matthew Dillon	:	Fixed another RST bug
89  *		Alan Cox	:	Move to kernel side addressing changes.
90  *		Alan Cox	:	Beginning work on TCP fastpathing
91  *					(not yet usable)
92  *		Arnt Gulbrandsen:	Turbocharged tcp_check() routine.
93  *		Alan Cox	:	TCP fast path debugging
94  *		Alan Cox	:	Window clamping
95  *		Michael Riepe	:	Bug in tcp_check()
96  *		Matt Dillon	:	More TCP improvements and RST bug fixes
97  *		Matt Dillon	:	Yet more small nasties remove from the
98  *					TCP code (Be very nice to this man if
99  *					tcp finally works 100%) 8)
100  *		Alan Cox	:	BSD accept semantics.
101  *		Alan Cox	:	Reset on closedown bug.
102  *	Peter De Schrijver	:	ENOTCONN check missing in tcp_sendto().
103  *		Michael Pall	:	Handle poll() after URG properly in
104  *					all cases.
105  *		Michael Pall	:	Undo the last fix in tcp_read_urg()
106  *					(multi URG PUSH broke rlogin).
107  *		Michael Pall	:	Fix the multi URG PUSH problem in
108  *					tcp_readable(), poll() after URG
109  *					works now.
110  *		Michael Pall	:	recv(...,MSG_OOB) never blocks in the
111  *					BSD api.
112  *		Alan Cox	:	Changed the semantics of sk->socket to
113  *					fix a race and a signal problem with
114  *					accept() and async I/O.
115  *		Alan Cox	:	Relaxed the rules on tcp_sendto().
116  *		Yury Shevchuk	:	Really fixed accept() blocking problem.
117  *		Craig I. Hagan  :	Allow for BSD compatible TIME_WAIT for
118  *					clients/servers which listen in on
119  *					fixed ports.
120  *		Alan Cox	:	Cleaned the above up and shrank it to
121  *					a sensible code size.
122  *		Alan Cox	:	Self connect lockup fix.
123  *		Alan Cox	:	No connect to multicast.
124  *		Ross Biro	:	Close unaccepted children on master
125  *					socket close.
126  *		Alan Cox	:	Reset tracing code.
127  *		Alan Cox	:	Spurious resets on shutdown.
128  *		Alan Cox	:	Giant 15 minute/60 second timer error
129  *		Alan Cox	:	Small whoops in polling before an
130  *					accept.
131  *		Alan Cox	:	Kept the state trace facility since
132  *					it's handy for debugging.
133  *		Alan Cox	:	More reset handler fixes.
134  *		Alan Cox	:	Started rewriting the code based on
135  *					the RFC's for other useful protocol
136  *					references see: Comer, KA9Q NOS, and
137  *					for a reference on the difference
138  *					between specifications and how BSD
139  *					works see the 4.4lite source.
140  *		A.N.Kuznetsov	:	Don't time wait on completion of tidy
141  *					close.
142  *		Linus Torvalds	:	Fin/Shutdown & copied_seq changes.
143  *		Linus Torvalds	:	Fixed BSD port reuse to work first syn
144  *		Alan Cox	:	Reimplemented timers as per the RFC
145  *					and using multiple timers for sanity.
146  *		Alan Cox	:	Small bug fixes, and a lot of new
147  *					comments.
148  *		Alan Cox	:	Fixed dual reader crash by locking
149  *					the buffers (much like datagram.c)
150  *		Alan Cox	:	Fixed stuck sockets in probe. A probe
151  *					now gets fed up of retrying without
152  *					(even a no space) answer.
153  *		Alan Cox	:	Extracted closing code better
154  *		Alan Cox	:	Fixed the closing state machine to
155  *					resemble the RFC.
156  *		Alan Cox	:	More 'per spec' fixes.
157  *		Jorge Cwik	:	Even faster checksumming.
158  *		Alan Cox	:	tcp_data() doesn't ack illegal PSH
159  *					only frames. At least one pc tcp stack
160  *					generates them.
161  *		Alan Cox	:	Cache last socket.
162  *		Alan Cox	:	Per route irtt.
163  *		Matt Day	:	poll()->select() match BSD precisely on error
164  *		Alan Cox	:	New buffers
165  *		Marc Tamsky	:	Various sk->prot->retransmits and
166  *					sk->retransmits misupdating fixed.
167  *					Fixed tcp_write_timeout: stuck close,
168  *					and TCP syn retries gets used now.
169  *		Mark Yarvis	:	In tcp_read_wakeup(), don't send an
170  *					ack if state is TCP_CLOSED.
171  *		Alan Cox	:	Look up device on a retransmit - routes may
172  *					change. Doesn't yet cope with MSS shrink right
173  *					but it's a start!
174  *		Marc Tamsky	:	Closing in closing fixes.
175  *		Mike Shaver	:	RFC1122 verifications.
176  *		Alan Cox	:	rcv_saddr errors.
177  *		Alan Cox	:	Block double connect().
178  *		Alan Cox	:	Small hooks for enSKIP.
179  *		Alexey Kuznetsov:	Path MTU discovery.
180  *		Alan Cox	:	Support soft errors.
181  *		Alan Cox	:	Fix MTU discovery pathological case
182  *					when the remote claims no mtu!
183  *		Marc Tamsky	:	TCP_CLOSE fix.
184  *		Colin (G3TNE)	:	Send a reset on syn ack replies in
185  *					window but wrong (fixes NT lpd problems)
186  *		Pedro Roque	:	Better TCP window handling, delayed ack.
187  *		Joerg Reuter	:	No modification of locked buffers in
188  *					tcp_do_retransmit()
189  *		Eric Schenk	:	Changed receiver side silly window
190  *					avoidance algorithm to BSD style
191  *					algorithm. This doubles throughput
192  *					against machines running Solaris,
193  *					and seems to result in general
194  *					improvement.
195  *	Stefan Magdalinski	:	adjusted tcp_readable() to fix FIONREAD
196  *	Willy Konynenberg	:	Transparent proxying support.
197  *	Mike McLagan		:	Routing by source
198  *		Keith Owens	:	Do proper merging with partial SKB's in
199  *					tcp_do_sendmsg to avoid burstiness.
200  *		Eric Schenk	:	Fix fast close down bug with
201  *					shutdown() followed by close().
202  *		Andi Kleen 	:	Make poll agree with SIGIO
203  *	Salvatore Sanfilippo	:	Support SO_LINGER with linger == 1 and
204  *					lingertime == 0 (RFC 793 ABORT Call)
205  *	Hirokazu Takahashi	:	Use copy_from_user() instead of
206  *					csum_and_copy_from_user() if possible.
207  *
208  *		This program is free software; you can redistribute it and/or
209  *		modify it under the terms of the GNU General Public License
210  *		as published by the Free Software Foundation; either version
211  *		2 of the License, or(at your option) any later version.
212  *
213  * Description of States:
214  *
215  *	TCP_SYN_SENT		sent a connection request, waiting for ack
216  *
217  *	TCP_SYN_RECV		received a connection request, sent ack,
218  *				waiting for final ack in three-way handshake.
219  *
220  *	TCP_ESTABLISHED		connection established
221  *
222  *	TCP_FIN_WAIT1		our side has shutdown, waiting to complete
223  *				transmission of remaining buffered data
224  *
225  *	TCP_FIN_WAIT2		all buffered data sent, waiting for remote
226  *				to shutdown
227  *
228  *	TCP_CLOSING		both sides have shutdown but we still have
229  *				data we have to finish sending
230  *
231  *	TCP_TIME_WAIT		timeout to catch resent junk before entering
232  *				closed, can only be entered from FIN_WAIT2
233  *				or CLOSING.  Required because the other end
234  *				may not have gotten our last ACK causing it
235  *				to retransmit the data packet (which we ignore)
236  *
237  *	TCP_CLOSE_WAIT		remote side has shutdown and is waiting for
238  *				us to finish writing our data and to shutdown
239  *				(we have to close() to move on to LAST_ACK)
240  *
241  *	TCP_LAST_ACK		out side has shutdown after remote has
242  *				shutdown.  There may still be data in our
243  *				buffer that we have to finish sending
244  *
245  *	TCP_CLOSE		socket is finished
246  */
247 
248 #define pr_fmt(fmt) "TCP: " fmt
249 
250 #include <crypto/hash.h>
251 #include <linux/kernel.h>
252 #include <linux/module.h>
253 #include <linux/types.h>
254 #include <linux/fcntl.h>
255 #include <linux/poll.h>
256 #include <linux/inet_diag.h>
257 #include <linux/init.h>
258 #include <linux/fs.h>
259 #include <linux/skbuff.h>
260 #include <linux/scatterlist.h>
261 #include <linux/splice.h>
262 #include <linux/net.h>
263 #include <linux/socket.h>
264 #include <linux/random.h>
265 #include <linux/bootmem.h>
266 #include <linux/highmem.h>
267 #include <linux/swap.h>
268 #include <linux/cache.h>
269 #include <linux/err.h>
270 #include <linux/time.h>
271 #include <linux/slab.h>
272 #include <linux/errqueue.h>
273 #include <linux/static_key.h>
274 
275 #include <net/icmp.h>
276 #include <net/inet_common.h>
277 #include <net/tcp.h>
278 #include <net/xfrm.h>
279 #include <net/ip.h>
280 #include <net/sock.h>
281 
282 #include <linux/uaccess.h>
283 #include <asm/ioctls.h>
284 #include <net/busy_poll.h>
285 
286 struct percpu_counter tcp_orphan_count;
287 EXPORT_SYMBOL_GPL(tcp_orphan_count);
288 
289 long sysctl_tcp_mem[3] __read_mostly;
290 EXPORT_SYMBOL(sysctl_tcp_mem);
291 
292 atomic_long_t tcp_memory_allocated;	/* Current allocated memory. */
293 EXPORT_SYMBOL(tcp_memory_allocated);
294 
295 #if IS_ENABLED(CONFIG_SMC)
296 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
297 EXPORT_SYMBOL(tcp_have_smc);
298 #endif
299 
300 /*
301  * Current number of TCP sockets.
302  */
303 struct percpu_counter tcp_sockets_allocated;
304 EXPORT_SYMBOL(tcp_sockets_allocated);
305 
306 /*
307  * TCP splice context
308  */
309 struct tcp_splice_state {
310 	struct pipe_inode_info *pipe;
311 	size_t len;
312 	unsigned int flags;
313 };
314 
315 /*
316  * Pressure flag: try to collapse.
317  * Technical note: it is used by multiple contexts non atomically.
318  * All the __sk_mem_schedule() is of this nature: accounting
319  * is strict, actions are advisory and have some latency.
320  */
321 unsigned long tcp_memory_pressure __read_mostly;
322 EXPORT_SYMBOL_GPL(tcp_memory_pressure);
323 
324 void tcp_enter_memory_pressure(struct sock *sk)
325 {
326 	unsigned long val;
327 
328 	if (tcp_memory_pressure)
329 		return;
330 	val = jiffies;
331 
332 	if (!val)
333 		val--;
334 	if (!cmpxchg(&tcp_memory_pressure, 0, val))
335 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
336 }
337 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
338 
339 void tcp_leave_memory_pressure(struct sock *sk)
340 {
341 	unsigned long val;
342 
343 	if (!tcp_memory_pressure)
344 		return;
345 	val = xchg(&tcp_memory_pressure, 0);
346 	if (val)
347 		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
348 			      jiffies_to_msecs(jiffies - val));
349 }
350 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
351 
352 /* Convert seconds to retransmits based on initial and max timeout */
353 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
354 {
355 	u8 res = 0;
356 
357 	if (seconds > 0) {
358 		int period = timeout;
359 
360 		res = 1;
361 		while (seconds > period && res < 255) {
362 			res++;
363 			timeout <<= 1;
364 			if (timeout > rto_max)
365 				timeout = rto_max;
366 			period += timeout;
367 		}
368 	}
369 	return res;
370 }
371 
372 /* Convert retransmits to seconds based on initial and max timeout */
373 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
374 {
375 	int period = 0;
376 
377 	if (retrans > 0) {
378 		period = timeout;
379 		while (--retrans) {
380 			timeout <<= 1;
381 			if (timeout > rto_max)
382 				timeout = rto_max;
383 			period += timeout;
384 		}
385 	}
386 	return period;
387 }
388 
389 static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
390 {
391 	u32 rate = READ_ONCE(tp->rate_delivered);
392 	u32 intv = READ_ONCE(tp->rate_interval_us);
393 	u64 rate64 = 0;
394 
395 	if (rate && intv) {
396 		rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
397 		do_div(rate64, intv);
398 	}
399 	return rate64;
400 }
401 
402 /* Address-family independent initialization for a tcp_sock.
403  *
404  * NOTE: A lot of things set to zero explicitly by call to
405  *       sk_alloc() so need not be done here.
406  */
407 void tcp_init_sock(struct sock *sk)
408 {
409 	struct inet_connection_sock *icsk = inet_csk(sk);
410 	struct tcp_sock *tp = tcp_sk(sk);
411 
412 	tp->out_of_order_queue = RB_ROOT;
413 	sk->tcp_rtx_queue = RB_ROOT;
414 	tcp_init_xmit_timers(sk);
415 	INIT_LIST_HEAD(&tp->tsq_node);
416 	INIT_LIST_HEAD(&tp->tsorted_sent_queue);
417 
418 	icsk->icsk_rto = TCP_TIMEOUT_INIT;
419 	tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
420 	minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
421 
422 	/* So many TCP implementations out there (incorrectly) count the
423 	 * initial SYN frame in their delayed-ACK and congestion control
424 	 * algorithms that we must have the following bandaid to talk
425 	 * efficiently to them.  -DaveM
426 	 */
427 	tp->snd_cwnd = TCP_INIT_CWND;
428 
429 	/* There's a bubble in the pipe until at least the first ACK. */
430 	tp->app_limited = ~0U;
431 
432 	/* See draft-stevens-tcpca-spec-01 for discussion of the
433 	 * initialization of these values.
434 	 */
435 	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
436 	tp->snd_cwnd_clamp = ~0;
437 	tp->mss_cache = TCP_MSS_DEFAULT;
438 
439 	tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
440 	tcp_assign_congestion_control(sk);
441 
442 	tp->tsoffset = 0;
443 	tp->rack.reo_wnd_steps = 1;
444 
445 	sk->sk_state = TCP_CLOSE;
446 
447 	sk->sk_write_space = sk_stream_write_space;
448 	sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
449 
450 	icsk->icsk_sync_mss = tcp_sync_mss;
451 
452 	sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
453 	sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
454 
455 	sk_sockets_allocated_inc(sk);
456 	sk->sk_route_forced_caps = NETIF_F_GSO;
457 }
458 EXPORT_SYMBOL(tcp_init_sock);
459 
460 void tcp_init_transfer(struct sock *sk, int bpf_op)
461 {
462 	struct inet_connection_sock *icsk = inet_csk(sk);
463 
464 	tcp_mtup_init(sk);
465 	icsk->icsk_af_ops->rebuild_header(sk);
466 	tcp_init_metrics(sk);
467 	tcp_call_bpf(sk, bpf_op, 0, NULL);
468 	tcp_init_congestion_control(sk);
469 	tcp_init_buffer_space(sk);
470 }
471 
472 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
473 {
474 	struct sk_buff *skb = tcp_write_queue_tail(sk);
475 
476 	if (tsflags && skb) {
477 		struct skb_shared_info *shinfo = skb_shinfo(skb);
478 		struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
479 
480 		sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
481 		if (tsflags & SOF_TIMESTAMPING_TX_ACK)
482 			tcb->txstamp_ack = 1;
483 		if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
484 			shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
485 	}
486 }
487 
488 /*
489  *	Wait for a TCP event.
490  *
491  *	Note that we don't need to lock the socket, as the upper poll layers
492  *	take care of normal races (between the test and the event) and we don't
493  *	go look at any of the socket buffers directly.
494  */
495 __poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
496 {
497 	__poll_t mask;
498 	struct sock *sk = sock->sk;
499 	const struct tcp_sock *tp = tcp_sk(sk);
500 	int state;
501 
502 	sock_poll_wait(file, sk_sleep(sk), wait);
503 
504 	state = inet_sk_state_load(sk);
505 	if (state == TCP_LISTEN)
506 		return inet_csk_listen_poll(sk);
507 
508 	/* Socket is not locked. We are protected from async events
509 	 * by poll logic and correct handling of state changes
510 	 * made by other threads is impossible in any case.
511 	 */
512 
513 	mask = 0;
514 
515 	/*
516 	 * EPOLLHUP is certainly not done right. But poll() doesn't
517 	 * have a notion of HUP in just one direction, and for a
518 	 * socket the read side is more interesting.
519 	 *
520 	 * Some poll() documentation says that EPOLLHUP is incompatible
521 	 * with the EPOLLOUT/POLLWR flags, so somebody should check this
522 	 * all. But careful, it tends to be safer to return too many
523 	 * bits than too few, and you can easily break real applications
524 	 * if you don't tell them that something has hung up!
525 	 *
526 	 * Check-me.
527 	 *
528 	 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
529 	 * our fs/select.c). It means that after we received EOF,
530 	 * poll always returns immediately, making impossible poll() on write()
531 	 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
532 	 * if and only if shutdown has been made in both directions.
533 	 * Actually, it is interesting to look how Solaris and DUX
534 	 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
535 	 * then we could set it on SND_SHUTDOWN. BTW examples given
536 	 * in Stevens' books assume exactly this behaviour, it explains
537 	 * why EPOLLHUP is incompatible with EPOLLOUT.	--ANK
538 	 *
539 	 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
540 	 * blocking on fresh not-connected or disconnected socket. --ANK
541 	 */
542 	if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
543 		mask |= EPOLLHUP;
544 	if (sk->sk_shutdown & RCV_SHUTDOWN)
545 		mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
546 
547 	/* Connected or passive Fast Open socket? */
548 	if (state != TCP_SYN_SENT &&
549 	    (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
550 		int target = sock_rcvlowat(sk, 0, INT_MAX);
551 
552 		if (tp->urg_seq == tp->copied_seq &&
553 		    !sock_flag(sk, SOCK_URGINLINE) &&
554 		    tp->urg_data)
555 			target++;
556 
557 		if (tp->rcv_nxt - tp->copied_seq >= target)
558 			mask |= EPOLLIN | EPOLLRDNORM;
559 
560 		if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
561 			if (sk_stream_is_writeable(sk)) {
562 				mask |= EPOLLOUT | EPOLLWRNORM;
563 			} else {  /* send SIGIO later */
564 				sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
565 				set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
566 
567 				/* Race breaker. If space is freed after
568 				 * wspace test but before the flags are set,
569 				 * IO signal will be lost. Memory barrier
570 				 * pairs with the input side.
571 				 */
572 				smp_mb__after_atomic();
573 				if (sk_stream_is_writeable(sk))
574 					mask |= EPOLLOUT | EPOLLWRNORM;
575 			}
576 		} else
577 			mask |= EPOLLOUT | EPOLLWRNORM;
578 
579 		if (tp->urg_data & TCP_URG_VALID)
580 			mask |= EPOLLPRI;
581 	} else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
582 		/* Active TCP fastopen socket with defer_connect
583 		 * Return EPOLLOUT so application can call write()
584 		 * in order for kernel to generate SYN+data
585 		 */
586 		mask |= EPOLLOUT | EPOLLWRNORM;
587 	}
588 	/* This barrier is coupled with smp_wmb() in tcp_reset() */
589 	smp_rmb();
590 	if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
591 		mask |= EPOLLERR;
592 
593 	return mask;
594 }
595 EXPORT_SYMBOL(tcp_poll);
596 
597 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
598 {
599 	struct tcp_sock *tp = tcp_sk(sk);
600 	int answ;
601 	bool slow;
602 
603 	switch (cmd) {
604 	case SIOCINQ:
605 		if (sk->sk_state == TCP_LISTEN)
606 			return -EINVAL;
607 
608 		slow = lock_sock_fast(sk);
609 		answ = tcp_inq(sk);
610 		unlock_sock_fast(sk, slow);
611 		break;
612 	case SIOCATMARK:
613 		answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
614 		break;
615 	case SIOCOUTQ:
616 		if (sk->sk_state == TCP_LISTEN)
617 			return -EINVAL;
618 
619 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
620 			answ = 0;
621 		else
622 			answ = tp->write_seq - tp->snd_una;
623 		break;
624 	case SIOCOUTQNSD:
625 		if (sk->sk_state == TCP_LISTEN)
626 			return -EINVAL;
627 
628 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
629 			answ = 0;
630 		else
631 			answ = tp->write_seq - tp->snd_nxt;
632 		break;
633 	default:
634 		return -ENOIOCTLCMD;
635 	}
636 
637 	return put_user(answ, (int __user *)arg);
638 }
639 EXPORT_SYMBOL(tcp_ioctl);
640 
641 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
642 {
643 	TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
644 	tp->pushed_seq = tp->write_seq;
645 }
646 
647 static inline bool forced_push(const struct tcp_sock *tp)
648 {
649 	return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
650 }
651 
652 static void skb_entail(struct sock *sk, struct sk_buff *skb)
653 {
654 	struct tcp_sock *tp = tcp_sk(sk);
655 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
656 
657 	skb->csum    = 0;
658 	tcb->seq     = tcb->end_seq = tp->write_seq;
659 	tcb->tcp_flags = TCPHDR_ACK;
660 	tcb->sacked  = 0;
661 	__skb_header_release(skb);
662 	tcp_add_write_queue_tail(sk, skb);
663 	sk->sk_wmem_queued += skb->truesize;
664 	sk_mem_charge(sk, skb->truesize);
665 	if (tp->nonagle & TCP_NAGLE_PUSH)
666 		tp->nonagle &= ~TCP_NAGLE_PUSH;
667 
668 	tcp_slow_start_after_idle_check(sk);
669 }
670 
671 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
672 {
673 	if (flags & MSG_OOB)
674 		tp->snd_up = tp->write_seq;
675 }
676 
677 /* If a not yet filled skb is pushed, do not send it if
678  * we have data packets in Qdisc or NIC queues :
679  * Because TX completion will happen shortly, it gives a chance
680  * to coalesce future sendmsg() payload into this skb, without
681  * need for a timer, and with no latency trade off.
682  * As packets containing data payload have a bigger truesize
683  * than pure acks (dataless) packets, the last checks prevent
684  * autocorking if we only have an ACK in Qdisc/NIC queues,
685  * or if TX completion was delayed after we processed ACK packet.
686  */
687 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
688 				int size_goal)
689 {
690 	return skb->len < size_goal &&
691 	       sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
692 	       skb != tcp_write_queue_head(sk) &&
693 	       refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
694 }
695 
696 static void tcp_push(struct sock *sk, int flags, int mss_now,
697 		     int nonagle, int size_goal)
698 {
699 	struct tcp_sock *tp = tcp_sk(sk);
700 	struct sk_buff *skb;
701 
702 	skb = tcp_write_queue_tail(sk);
703 	if (!skb)
704 		return;
705 	if (!(flags & MSG_MORE) || forced_push(tp))
706 		tcp_mark_push(tp, skb);
707 
708 	tcp_mark_urg(tp, flags);
709 
710 	if (tcp_should_autocork(sk, skb, size_goal)) {
711 
712 		/* avoid atomic op if TSQ_THROTTLED bit is already set */
713 		if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
714 			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
715 			set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
716 		}
717 		/* It is possible TX completion already happened
718 		 * before we set TSQ_THROTTLED.
719 		 */
720 		if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
721 			return;
722 	}
723 
724 	if (flags & MSG_MORE)
725 		nonagle = TCP_NAGLE_CORK;
726 
727 	__tcp_push_pending_frames(sk, mss_now, nonagle);
728 }
729 
730 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
731 				unsigned int offset, size_t len)
732 {
733 	struct tcp_splice_state *tss = rd_desc->arg.data;
734 	int ret;
735 
736 	ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
737 			      min(rd_desc->count, len), tss->flags);
738 	if (ret > 0)
739 		rd_desc->count -= ret;
740 	return ret;
741 }
742 
743 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
744 {
745 	/* Store TCP splice context information in read_descriptor_t. */
746 	read_descriptor_t rd_desc = {
747 		.arg.data = tss,
748 		.count	  = tss->len,
749 	};
750 
751 	return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
752 }
753 
754 /**
755  *  tcp_splice_read - splice data from TCP socket to a pipe
756  * @sock:	socket to splice from
757  * @ppos:	position (not valid)
758  * @pipe:	pipe to splice to
759  * @len:	number of bytes to splice
760  * @flags:	splice modifier flags
761  *
762  * Description:
763  *    Will read pages from given socket and fill them into a pipe.
764  *
765  **/
766 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
767 			struct pipe_inode_info *pipe, size_t len,
768 			unsigned int flags)
769 {
770 	struct sock *sk = sock->sk;
771 	struct tcp_splice_state tss = {
772 		.pipe = pipe,
773 		.len = len,
774 		.flags = flags,
775 	};
776 	long timeo;
777 	ssize_t spliced;
778 	int ret;
779 
780 	sock_rps_record_flow(sk);
781 	/*
782 	 * We can't seek on a socket input
783 	 */
784 	if (unlikely(*ppos))
785 		return -ESPIPE;
786 
787 	ret = spliced = 0;
788 
789 	lock_sock(sk);
790 
791 	timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
792 	while (tss.len) {
793 		ret = __tcp_splice_read(sk, &tss);
794 		if (ret < 0)
795 			break;
796 		else if (!ret) {
797 			if (spliced)
798 				break;
799 			if (sock_flag(sk, SOCK_DONE))
800 				break;
801 			if (sk->sk_err) {
802 				ret = sock_error(sk);
803 				break;
804 			}
805 			if (sk->sk_shutdown & RCV_SHUTDOWN)
806 				break;
807 			if (sk->sk_state == TCP_CLOSE) {
808 				/*
809 				 * This occurs when user tries to read
810 				 * from never connected socket.
811 				 */
812 				if (!sock_flag(sk, SOCK_DONE))
813 					ret = -ENOTCONN;
814 				break;
815 			}
816 			if (!timeo) {
817 				ret = -EAGAIN;
818 				break;
819 			}
820 			/* if __tcp_splice_read() got nothing while we have
821 			 * an skb in receive queue, we do not want to loop.
822 			 * This might happen with URG data.
823 			 */
824 			if (!skb_queue_empty(&sk->sk_receive_queue))
825 				break;
826 			sk_wait_data(sk, &timeo, NULL);
827 			if (signal_pending(current)) {
828 				ret = sock_intr_errno(timeo);
829 				break;
830 			}
831 			continue;
832 		}
833 		tss.len -= ret;
834 		spliced += ret;
835 
836 		if (!timeo)
837 			break;
838 		release_sock(sk);
839 		lock_sock(sk);
840 
841 		if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
842 		    (sk->sk_shutdown & RCV_SHUTDOWN) ||
843 		    signal_pending(current))
844 			break;
845 	}
846 
847 	release_sock(sk);
848 
849 	if (spliced)
850 		return spliced;
851 
852 	return ret;
853 }
854 EXPORT_SYMBOL(tcp_splice_read);
855 
856 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
857 				    bool force_schedule)
858 {
859 	struct sk_buff *skb;
860 
861 	/* The TCP header must be at least 32-bit aligned.  */
862 	size = ALIGN(size, 4);
863 
864 	if (unlikely(tcp_under_memory_pressure(sk)))
865 		sk_mem_reclaim_partial(sk);
866 
867 	skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
868 	if (likely(skb)) {
869 		bool mem_scheduled;
870 
871 		if (force_schedule) {
872 			mem_scheduled = true;
873 			sk_forced_mem_schedule(sk, skb->truesize);
874 		} else {
875 			mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
876 		}
877 		if (likely(mem_scheduled)) {
878 			skb_reserve(skb, sk->sk_prot->max_header);
879 			/*
880 			 * Make sure that we have exactly size bytes
881 			 * available to the caller, no more, no less.
882 			 */
883 			skb->reserved_tailroom = skb->end - skb->tail - size;
884 			INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
885 			return skb;
886 		}
887 		__kfree_skb(skb);
888 	} else {
889 		sk->sk_prot->enter_memory_pressure(sk);
890 		sk_stream_moderate_sndbuf(sk);
891 	}
892 	return NULL;
893 }
894 
895 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
896 				       int large_allowed)
897 {
898 	struct tcp_sock *tp = tcp_sk(sk);
899 	u32 new_size_goal, size_goal;
900 
901 	if (!large_allowed)
902 		return mss_now;
903 
904 	/* Note : tcp_tso_autosize() will eventually split this later */
905 	new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
906 	new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
907 
908 	/* We try hard to avoid divides here */
909 	size_goal = tp->gso_segs * mss_now;
910 	if (unlikely(new_size_goal < size_goal ||
911 		     new_size_goal >= size_goal + mss_now)) {
912 		tp->gso_segs = min_t(u16, new_size_goal / mss_now,
913 				     sk->sk_gso_max_segs);
914 		size_goal = tp->gso_segs * mss_now;
915 	}
916 
917 	return max(size_goal, mss_now);
918 }
919 
920 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
921 {
922 	int mss_now;
923 
924 	mss_now = tcp_current_mss(sk);
925 	*size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
926 
927 	return mss_now;
928 }
929 
930 ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
931 			 size_t size, int flags)
932 {
933 	struct tcp_sock *tp = tcp_sk(sk);
934 	int mss_now, size_goal;
935 	int err;
936 	ssize_t copied;
937 	long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
938 
939 	/* Wait for a connection to finish. One exception is TCP Fast Open
940 	 * (passive side) where data is allowed to be sent before a connection
941 	 * is fully established.
942 	 */
943 	if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
944 	    !tcp_passive_fastopen(sk)) {
945 		err = sk_stream_wait_connect(sk, &timeo);
946 		if (err != 0)
947 			goto out_err;
948 	}
949 
950 	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
951 
952 	mss_now = tcp_send_mss(sk, &size_goal, flags);
953 	copied = 0;
954 
955 	err = -EPIPE;
956 	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
957 		goto out_err;
958 
959 	while (size > 0) {
960 		struct sk_buff *skb = tcp_write_queue_tail(sk);
961 		int copy, i;
962 		bool can_coalesce;
963 
964 		if (!skb || (copy = size_goal - skb->len) <= 0 ||
965 		    !tcp_skb_can_collapse_to(skb)) {
966 new_segment:
967 			if (!sk_stream_memory_free(sk))
968 				goto wait_for_sndbuf;
969 
970 			skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
971 					tcp_rtx_and_write_queues_empty(sk));
972 			if (!skb)
973 				goto wait_for_memory;
974 
975 			skb_entail(sk, skb);
976 			copy = size_goal;
977 		}
978 
979 		if (copy > size)
980 			copy = size;
981 
982 		i = skb_shinfo(skb)->nr_frags;
983 		can_coalesce = skb_can_coalesce(skb, i, page, offset);
984 		if (!can_coalesce && i >= sysctl_max_skb_frags) {
985 			tcp_mark_push(tp, skb);
986 			goto new_segment;
987 		}
988 		if (!sk_wmem_schedule(sk, copy))
989 			goto wait_for_memory;
990 
991 		if (can_coalesce) {
992 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
993 		} else {
994 			get_page(page);
995 			skb_fill_page_desc(skb, i, page, offset, copy);
996 		}
997 		skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
998 
999 		skb->len += copy;
1000 		skb->data_len += copy;
1001 		skb->truesize += copy;
1002 		sk->sk_wmem_queued += copy;
1003 		sk_mem_charge(sk, copy);
1004 		skb->ip_summed = CHECKSUM_PARTIAL;
1005 		tp->write_seq += copy;
1006 		TCP_SKB_CB(skb)->end_seq += copy;
1007 		tcp_skb_pcount_set(skb, 0);
1008 
1009 		if (!copied)
1010 			TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1011 
1012 		copied += copy;
1013 		offset += copy;
1014 		size -= copy;
1015 		if (!size)
1016 			goto out;
1017 
1018 		if (skb->len < size_goal || (flags & MSG_OOB))
1019 			continue;
1020 
1021 		if (forced_push(tp)) {
1022 			tcp_mark_push(tp, skb);
1023 			__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1024 		} else if (skb == tcp_send_head(sk))
1025 			tcp_push_one(sk, mss_now);
1026 		continue;
1027 
1028 wait_for_sndbuf:
1029 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1030 wait_for_memory:
1031 		tcp_push(sk, flags & ~MSG_MORE, mss_now,
1032 			 TCP_NAGLE_PUSH, size_goal);
1033 
1034 		err = sk_stream_wait_memory(sk, &timeo);
1035 		if (err != 0)
1036 			goto do_error;
1037 
1038 		mss_now = tcp_send_mss(sk, &size_goal, flags);
1039 	}
1040 
1041 out:
1042 	if (copied) {
1043 		tcp_tx_timestamp(sk, sk->sk_tsflags);
1044 		if (!(flags & MSG_SENDPAGE_NOTLAST))
1045 			tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1046 	}
1047 	return copied;
1048 
1049 do_error:
1050 	if (copied)
1051 		goto out;
1052 out_err:
1053 	/* make sure we wake any epoll edge trigger waiter */
1054 	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1055 		     err == -EAGAIN)) {
1056 		sk->sk_write_space(sk);
1057 		tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1058 	}
1059 	return sk_stream_error(sk, flags, err);
1060 }
1061 EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1062 
1063 int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1064 			size_t size, int flags)
1065 {
1066 	if (!(sk->sk_route_caps & NETIF_F_SG))
1067 		return sock_no_sendpage_locked(sk, page, offset, size, flags);
1068 
1069 	tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1070 
1071 	return do_tcp_sendpages(sk, page, offset, size, flags);
1072 }
1073 EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1074 
1075 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1076 		 size_t size, int flags)
1077 {
1078 	int ret;
1079 
1080 	lock_sock(sk);
1081 	ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1082 	release_sock(sk);
1083 
1084 	return ret;
1085 }
1086 EXPORT_SYMBOL(tcp_sendpage);
1087 
1088 /* Do not bother using a page frag for very small frames.
1089  * But use this heuristic only for the first skb in write queue.
1090  *
1091  * Having no payload in skb->head allows better SACK shifting
1092  * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1093  * write queue has less skbs.
1094  * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1095  * This also speeds up tso_fragment(), since it wont fallback
1096  * to tcp_fragment().
1097  */
1098 static int linear_payload_sz(bool first_skb)
1099 {
1100 	if (first_skb)
1101 		return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1102 	return 0;
1103 }
1104 
1105 static int select_size(bool first_skb, bool zc)
1106 {
1107 	if (zc)
1108 		return 0;
1109 	return linear_payload_sz(first_skb);
1110 }
1111 
1112 void tcp_free_fastopen_req(struct tcp_sock *tp)
1113 {
1114 	if (tp->fastopen_req) {
1115 		kfree(tp->fastopen_req);
1116 		tp->fastopen_req = NULL;
1117 	}
1118 }
1119 
1120 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1121 				int *copied, size_t size)
1122 {
1123 	struct tcp_sock *tp = tcp_sk(sk);
1124 	struct inet_sock *inet = inet_sk(sk);
1125 	struct sockaddr *uaddr = msg->msg_name;
1126 	int err, flags;
1127 
1128 	if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1129 	    (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1130 	     uaddr->sa_family == AF_UNSPEC))
1131 		return -EOPNOTSUPP;
1132 	if (tp->fastopen_req)
1133 		return -EALREADY; /* Another Fast Open is in progress */
1134 
1135 	tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1136 				   sk->sk_allocation);
1137 	if (unlikely(!tp->fastopen_req))
1138 		return -ENOBUFS;
1139 	tp->fastopen_req->data = msg;
1140 	tp->fastopen_req->size = size;
1141 
1142 	if (inet->defer_connect) {
1143 		err = tcp_connect(sk);
1144 		/* Same failure procedure as in tcp_v4/6_connect */
1145 		if (err) {
1146 			tcp_set_state(sk, TCP_CLOSE);
1147 			inet->inet_dport = 0;
1148 			sk->sk_route_caps = 0;
1149 		}
1150 	}
1151 	flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1152 	err = __inet_stream_connect(sk->sk_socket, uaddr,
1153 				    msg->msg_namelen, flags, 1);
1154 	/* fastopen_req could already be freed in __inet_stream_connect
1155 	 * if the connection times out or gets rst
1156 	 */
1157 	if (tp->fastopen_req) {
1158 		*copied = tp->fastopen_req->copied;
1159 		tcp_free_fastopen_req(tp);
1160 		inet->defer_connect = 0;
1161 	}
1162 	return err;
1163 }
1164 
1165 int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1166 {
1167 	struct tcp_sock *tp = tcp_sk(sk);
1168 	struct ubuf_info *uarg = NULL;
1169 	struct sk_buff *skb;
1170 	struct sockcm_cookie sockc;
1171 	int flags, err, copied = 0;
1172 	int mss_now = 0, size_goal, copied_syn = 0;
1173 	bool process_backlog = false;
1174 	bool zc = false;
1175 	long timeo;
1176 
1177 	flags = msg->msg_flags;
1178 
1179 	if (flags & MSG_ZEROCOPY && size) {
1180 		if (sk->sk_state != TCP_ESTABLISHED) {
1181 			err = -EINVAL;
1182 			goto out_err;
1183 		}
1184 
1185 		skb = tcp_write_queue_tail(sk);
1186 		uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1187 		if (!uarg) {
1188 			err = -ENOBUFS;
1189 			goto out_err;
1190 		}
1191 
1192 		zc = sk->sk_route_caps & NETIF_F_SG;
1193 		if (!zc)
1194 			uarg->zerocopy = 0;
1195 	}
1196 
1197 	if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect)) {
1198 		err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1199 		if (err == -EINPROGRESS && copied_syn > 0)
1200 			goto out;
1201 		else if (err)
1202 			goto out_err;
1203 	}
1204 
1205 	timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1206 
1207 	tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1208 
1209 	/* Wait for a connection to finish. One exception is TCP Fast Open
1210 	 * (passive side) where data is allowed to be sent before a connection
1211 	 * is fully established.
1212 	 */
1213 	if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1214 	    !tcp_passive_fastopen(sk)) {
1215 		err = sk_stream_wait_connect(sk, &timeo);
1216 		if (err != 0)
1217 			goto do_error;
1218 	}
1219 
1220 	if (unlikely(tp->repair)) {
1221 		if (tp->repair_queue == TCP_RECV_QUEUE) {
1222 			copied = tcp_send_rcvq(sk, msg, size);
1223 			goto out_nopush;
1224 		}
1225 
1226 		err = -EINVAL;
1227 		if (tp->repair_queue == TCP_NO_QUEUE)
1228 			goto out_err;
1229 
1230 		/* 'common' sending to sendq */
1231 	}
1232 
1233 	sockc.tsflags = sk->sk_tsflags;
1234 	if (msg->msg_controllen) {
1235 		err = sock_cmsg_send(sk, msg, &sockc);
1236 		if (unlikely(err)) {
1237 			err = -EINVAL;
1238 			goto out_err;
1239 		}
1240 	}
1241 
1242 	/* This should be in poll */
1243 	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1244 
1245 	/* Ok commence sending. */
1246 	copied = 0;
1247 
1248 restart:
1249 	mss_now = tcp_send_mss(sk, &size_goal, flags);
1250 
1251 	err = -EPIPE;
1252 	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1253 		goto do_error;
1254 
1255 	while (msg_data_left(msg)) {
1256 		int copy = 0;
1257 
1258 		skb = tcp_write_queue_tail(sk);
1259 		if (skb)
1260 			copy = size_goal - skb->len;
1261 
1262 		if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1263 			bool first_skb;
1264 			int linear;
1265 
1266 new_segment:
1267 			/* Allocate new segment. If the interface is SG,
1268 			 * allocate skb fitting to single page.
1269 			 */
1270 			if (!sk_stream_memory_free(sk))
1271 				goto wait_for_sndbuf;
1272 
1273 			if (process_backlog && sk_flush_backlog(sk)) {
1274 				process_backlog = false;
1275 				goto restart;
1276 			}
1277 			first_skb = tcp_rtx_and_write_queues_empty(sk);
1278 			linear = select_size(first_skb, zc);
1279 			skb = sk_stream_alloc_skb(sk, linear, sk->sk_allocation,
1280 						  first_skb);
1281 			if (!skb)
1282 				goto wait_for_memory;
1283 
1284 			process_backlog = true;
1285 			skb->ip_summed = CHECKSUM_PARTIAL;
1286 
1287 			skb_entail(sk, skb);
1288 			copy = size_goal;
1289 
1290 			/* All packets are restored as if they have
1291 			 * already been sent. skb_mstamp isn't set to
1292 			 * avoid wrong rtt estimation.
1293 			 */
1294 			if (tp->repair)
1295 				TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1296 		}
1297 
1298 		/* Try to append data to the end of skb. */
1299 		if (copy > msg_data_left(msg))
1300 			copy = msg_data_left(msg);
1301 
1302 		/* Where to copy to? */
1303 		if (skb_availroom(skb) > 0 && !zc) {
1304 			/* We have some space in skb head. Superb! */
1305 			copy = min_t(int, copy, skb_availroom(skb));
1306 			err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1307 			if (err)
1308 				goto do_fault;
1309 		} else if (!zc) {
1310 			bool merge = true;
1311 			int i = skb_shinfo(skb)->nr_frags;
1312 			struct page_frag *pfrag = sk_page_frag(sk);
1313 
1314 			if (!sk_page_frag_refill(sk, pfrag))
1315 				goto wait_for_memory;
1316 
1317 			if (!skb_can_coalesce(skb, i, pfrag->page,
1318 					      pfrag->offset)) {
1319 				if (i >= sysctl_max_skb_frags) {
1320 					tcp_mark_push(tp, skb);
1321 					goto new_segment;
1322 				}
1323 				merge = false;
1324 			}
1325 
1326 			copy = min_t(int, copy, pfrag->size - pfrag->offset);
1327 
1328 			if (!sk_wmem_schedule(sk, copy))
1329 				goto wait_for_memory;
1330 
1331 			err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1332 						       pfrag->page,
1333 						       pfrag->offset,
1334 						       copy);
1335 			if (err)
1336 				goto do_error;
1337 
1338 			/* Update the skb. */
1339 			if (merge) {
1340 				skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1341 			} else {
1342 				skb_fill_page_desc(skb, i, pfrag->page,
1343 						   pfrag->offset, copy);
1344 				page_ref_inc(pfrag->page);
1345 			}
1346 			pfrag->offset += copy;
1347 		} else {
1348 			err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1349 			if (err == -EMSGSIZE || err == -EEXIST) {
1350 				tcp_mark_push(tp, skb);
1351 				goto new_segment;
1352 			}
1353 			if (err < 0)
1354 				goto do_error;
1355 			copy = err;
1356 		}
1357 
1358 		if (!copied)
1359 			TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1360 
1361 		tp->write_seq += copy;
1362 		TCP_SKB_CB(skb)->end_seq += copy;
1363 		tcp_skb_pcount_set(skb, 0);
1364 
1365 		copied += copy;
1366 		if (!msg_data_left(msg)) {
1367 			if (unlikely(flags & MSG_EOR))
1368 				TCP_SKB_CB(skb)->eor = 1;
1369 			goto out;
1370 		}
1371 
1372 		if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1373 			continue;
1374 
1375 		if (forced_push(tp)) {
1376 			tcp_mark_push(tp, skb);
1377 			__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1378 		} else if (skb == tcp_send_head(sk))
1379 			tcp_push_one(sk, mss_now);
1380 		continue;
1381 
1382 wait_for_sndbuf:
1383 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1384 wait_for_memory:
1385 		if (copied)
1386 			tcp_push(sk, flags & ~MSG_MORE, mss_now,
1387 				 TCP_NAGLE_PUSH, size_goal);
1388 
1389 		err = sk_stream_wait_memory(sk, &timeo);
1390 		if (err != 0)
1391 			goto do_error;
1392 
1393 		mss_now = tcp_send_mss(sk, &size_goal, flags);
1394 	}
1395 
1396 out:
1397 	if (copied) {
1398 		tcp_tx_timestamp(sk, sockc.tsflags);
1399 		tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1400 	}
1401 out_nopush:
1402 	sock_zerocopy_put(uarg);
1403 	return copied + copied_syn;
1404 
1405 do_fault:
1406 	if (!skb->len) {
1407 		tcp_unlink_write_queue(skb, sk);
1408 		/* It is the one place in all of TCP, except connection
1409 		 * reset, where we can be unlinking the send_head.
1410 		 */
1411 		tcp_check_send_head(sk, skb);
1412 		sk_wmem_free_skb(sk, skb);
1413 	}
1414 
1415 do_error:
1416 	if (copied + copied_syn)
1417 		goto out;
1418 out_err:
1419 	sock_zerocopy_put_abort(uarg);
1420 	err = sk_stream_error(sk, flags, err);
1421 	/* make sure we wake any epoll edge trigger waiter */
1422 	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1423 		     err == -EAGAIN)) {
1424 		sk->sk_write_space(sk);
1425 		tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1426 	}
1427 	return err;
1428 }
1429 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1430 
1431 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1432 {
1433 	int ret;
1434 
1435 	lock_sock(sk);
1436 	ret = tcp_sendmsg_locked(sk, msg, size);
1437 	release_sock(sk);
1438 
1439 	return ret;
1440 }
1441 EXPORT_SYMBOL(tcp_sendmsg);
1442 
1443 /*
1444  *	Handle reading urgent data. BSD has very simple semantics for
1445  *	this, no blocking and very strange errors 8)
1446  */
1447 
1448 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1449 {
1450 	struct tcp_sock *tp = tcp_sk(sk);
1451 
1452 	/* No URG data to read. */
1453 	if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1454 	    tp->urg_data == TCP_URG_READ)
1455 		return -EINVAL;	/* Yes this is right ! */
1456 
1457 	if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1458 		return -ENOTCONN;
1459 
1460 	if (tp->urg_data & TCP_URG_VALID) {
1461 		int err = 0;
1462 		char c = tp->urg_data;
1463 
1464 		if (!(flags & MSG_PEEK))
1465 			tp->urg_data = TCP_URG_READ;
1466 
1467 		/* Read urgent data. */
1468 		msg->msg_flags |= MSG_OOB;
1469 
1470 		if (len > 0) {
1471 			if (!(flags & MSG_TRUNC))
1472 				err = memcpy_to_msg(msg, &c, 1);
1473 			len = 1;
1474 		} else
1475 			msg->msg_flags |= MSG_TRUNC;
1476 
1477 		return err ? -EFAULT : len;
1478 	}
1479 
1480 	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1481 		return 0;
1482 
1483 	/* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1484 	 * the available implementations agree in this case:
1485 	 * this call should never block, independent of the
1486 	 * blocking state of the socket.
1487 	 * Mike <pall@rz.uni-karlsruhe.de>
1488 	 */
1489 	return -EAGAIN;
1490 }
1491 
1492 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1493 {
1494 	struct sk_buff *skb;
1495 	int copied = 0, err = 0;
1496 
1497 	/* XXX -- need to support SO_PEEK_OFF */
1498 
1499 	skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1500 		err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1501 		if (err)
1502 			return err;
1503 		copied += skb->len;
1504 	}
1505 
1506 	skb_queue_walk(&sk->sk_write_queue, skb) {
1507 		err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1508 		if (err)
1509 			break;
1510 
1511 		copied += skb->len;
1512 	}
1513 
1514 	return err ?: copied;
1515 }
1516 
1517 /* Clean up the receive buffer for full frames taken by the user,
1518  * then send an ACK if necessary.  COPIED is the number of bytes
1519  * tcp_recvmsg has given to the user so far, it speeds up the
1520  * calculation of whether or not we must ACK for the sake of
1521  * a window update.
1522  */
1523 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1524 {
1525 	struct tcp_sock *tp = tcp_sk(sk);
1526 	bool time_to_ack = false;
1527 
1528 	struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1529 
1530 	WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1531 	     "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1532 	     tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1533 
1534 	if (inet_csk_ack_scheduled(sk)) {
1535 		const struct inet_connection_sock *icsk = inet_csk(sk);
1536 		   /* Delayed ACKs frequently hit locked sockets during bulk
1537 		    * receive. */
1538 		if (icsk->icsk_ack.blocked ||
1539 		    /* Once-per-two-segments ACK was not sent by tcp_input.c */
1540 		    tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1541 		    /*
1542 		     * If this read emptied read buffer, we send ACK, if
1543 		     * connection is not bidirectional, user drained
1544 		     * receive buffer and there was a small segment
1545 		     * in queue.
1546 		     */
1547 		    (copied > 0 &&
1548 		     ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1549 		      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1550 		       !icsk->icsk_ack.pingpong)) &&
1551 		      !atomic_read(&sk->sk_rmem_alloc)))
1552 			time_to_ack = true;
1553 	}
1554 
1555 	/* We send an ACK if we can now advertise a non-zero window
1556 	 * which has been raised "significantly".
1557 	 *
1558 	 * Even if window raised up to infinity, do not send window open ACK
1559 	 * in states, where we will not receive more. It is useless.
1560 	 */
1561 	if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1562 		__u32 rcv_window_now = tcp_receive_window(tp);
1563 
1564 		/* Optimize, __tcp_select_window() is not cheap. */
1565 		if (2*rcv_window_now <= tp->window_clamp) {
1566 			__u32 new_window = __tcp_select_window(sk);
1567 
1568 			/* Send ACK now, if this read freed lots of space
1569 			 * in our buffer. Certainly, new_window is new window.
1570 			 * We can advertise it now, if it is not less than current one.
1571 			 * "Lots" means "at least twice" here.
1572 			 */
1573 			if (new_window && new_window >= 2 * rcv_window_now)
1574 				time_to_ack = true;
1575 		}
1576 	}
1577 	if (time_to_ack)
1578 		tcp_send_ack(sk);
1579 }
1580 
1581 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1582 {
1583 	struct sk_buff *skb;
1584 	u32 offset;
1585 
1586 	while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1587 		offset = seq - TCP_SKB_CB(skb)->seq;
1588 		if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1589 			pr_err_once("%s: found a SYN, please report !\n", __func__);
1590 			offset--;
1591 		}
1592 		if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1593 			*off = offset;
1594 			return skb;
1595 		}
1596 		/* This looks weird, but this can happen if TCP collapsing
1597 		 * splitted a fat GRO packet, while we released socket lock
1598 		 * in skb_splice_bits()
1599 		 */
1600 		sk_eat_skb(sk, skb);
1601 	}
1602 	return NULL;
1603 }
1604 
1605 /*
1606  * This routine provides an alternative to tcp_recvmsg() for routines
1607  * that would like to handle copying from skbuffs directly in 'sendfile'
1608  * fashion.
1609  * Note:
1610  *	- It is assumed that the socket was locked by the caller.
1611  *	- The routine does not block.
1612  *	- At present, there is no support for reading OOB data
1613  *	  or for 'peeking' the socket using this routine
1614  *	  (although both would be easy to implement).
1615  */
1616 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1617 		  sk_read_actor_t recv_actor)
1618 {
1619 	struct sk_buff *skb;
1620 	struct tcp_sock *tp = tcp_sk(sk);
1621 	u32 seq = tp->copied_seq;
1622 	u32 offset;
1623 	int copied = 0;
1624 
1625 	if (sk->sk_state == TCP_LISTEN)
1626 		return -ENOTCONN;
1627 	while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1628 		if (offset < skb->len) {
1629 			int used;
1630 			size_t len;
1631 
1632 			len = skb->len - offset;
1633 			/* Stop reading if we hit a patch of urgent data */
1634 			if (tp->urg_data) {
1635 				u32 urg_offset = tp->urg_seq - seq;
1636 				if (urg_offset < len)
1637 					len = urg_offset;
1638 				if (!len)
1639 					break;
1640 			}
1641 			used = recv_actor(desc, skb, offset, len);
1642 			if (used <= 0) {
1643 				if (!copied)
1644 					copied = used;
1645 				break;
1646 			} else if (used <= len) {
1647 				seq += used;
1648 				copied += used;
1649 				offset += used;
1650 			}
1651 			/* If recv_actor drops the lock (e.g. TCP splice
1652 			 * receive) the skb pointer might be invalid when
1653 			 * getting here: tcp_collapse might have deleted it
1654 			 * while aggregating skbs from the socket queue.
1655 			 */
1656 			skb = tcp_recv_skb(sk, seq - 1, &offset);
1657 			if (!skb)
1658 				break;
1659 			/* TCP coalescing might have appended data to the skb.
1660 			 * Try to splice more frags
1661 			 */
1662 			if (offset + 1 != skb->len)
1663 				continue;
1664 		}
1665 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1666 			sk_eat_skb(sk, skb);
1667 			++seq;
1668 			break;
1669 		}
1670 		sk_eat_skb(sk, skb);
1671 		if (!desc->count)
1672 			break;
1673 		tp->copied_seq = seq;
1674 	}
1675 	tp->copied_seq = seq;
1676 
1677 	tcp_rcv_space_adjust(sk);
1678 
1679 	/* Clean up data we have read: This will do ACK frames. */
1680 	if (copied > 0) {
1681 		tcp_recv_skb(sk, seq, &offset);
1682 		tcp_cleanup_rbuf(sk, copied);
1683 	}
1684 	return copied;
1685 }
1686 EXPORT_SYMBOL(tcp_read_sock);
1687 
1688 int tcp_peek_len(struct socket *sock)
1689 {
1690 	return tcp_inq(sock->sk);
1691 }
1692 EXPORT_SYMBOL(tcp_peek_len);
1693 
1694 static void tcp_update_recv_tstamps(struct sk_buff *skb,
1695 				    struct scm_timestamping *tss)
1696 {
1697 	if (skb->tstamp)
1698 		tss->ts[0] = ktime_to_timespec(skb->tstamp);
1699 	else
1700 		tss->ts[0] = (struct timespec) {0};
1701 
1702 	if (skb_hwtstamps(skb)->hwtstamp)
1703 		tss->ts[2] = ktime_to_timespec(skb_hwtstamps(skb)->hwtstamp);
1704 	else
1705 		tss->ts[2] = (struct timespec) {0};
1706 }
1707 
1708 /* Similar to __sock_recv_timestamp, but does not require an skb */
1709 static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1710 			       struct scm_timestamping *tss)
1711 {
1712 	struct timeval tv;
1713 	bool has_timestamping = false;
1714 
1715 	if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1716 		if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1717 			if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1718 				put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPNS,
1719 					 sizeof(tss->ts[0]), &tss->ts[0]);
1720 			} else {
1721 				tv.tv_sec = tss->ts[0].tv_sec;
1722 				tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1723 
1724 				put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP,
1725 					 sizeof(tv), &tv);
1726 			}
1727 		}
1728 
1729 		if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
1730 			has_timestamping = true;
1731 		else
1732 			tss->ts[0] = (struct timespec) {0};
1733 	}
1734 
1735 	if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1736 		if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
1737 			has_timestamping = true;
1738 		else
1739 			tss->ts[2] = (struct timespec) {0};
1740 	}
1741 
1742 	if (has_timestamping) {
1743 		tss->ts[1] = (struct timespec) {0};
1744 		put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMPING,
1745 			 sizeof(*tss), tss);
1746 	}
1747 }
1748 
1749 /*
1750  *	This routine copies from a sock struct into the user buffer.
1751  *
1752  *	Technical note: in 2.3 we work on _locked_ socket, so that
1753  *	tricks with *seq access order and skb->users are not required.
1754  *	Probably, code can be easily improved even more.
1755  */
1756 
1757 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1758 		int flags, int *addr_len)
1759 {
1760 	struct tcp_sock *tp = tcp_sk(sk);
1761 	int copied = 0;
1762 	u32 peek_seq;
1763 	u32 *seq;
1764 	unsigned long used;
1765 	int err;
1766 	int target;		/* Read at least this many bytes */
1767 	long timeo;
1768 	struct sk_buff *skb, *last;
1769 	u32 urg_hole = 0;
1770 	struct scm_timestamping tss;
1771 	bool has_tss = false;
1772 
1773 	if (unlikely(flags & MSG_ERRQUEUE))
1774 		return inet_recv_error(sk, msg, len, addr_len);
1775 
1776 	if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1777 	    (sk->sk_state == TCP_ESTABLISHED))
1778 		sk_busy_loop(sk, nonblock);
1779 
1780 	lock_sock(sk);
1781 
1782 	err = -ENOTCONN;
1783 	if (sk->sk_state == TCP_LISTEN)
1784 		goto out;
1785 
1786 	timeo = sock_rcvtimeo(sk, nonblock);
1787 
1788 	/* Urgent data needs to be handled specially. */
1789 	if (flags & MSG_OOB)
1790 		goto recv_urg;
1791 
1792 	if (unlikely(tp->repair)) {
1793 		err = -EPERM;
1794 		if (!(flags & MSG_PEEK))
1795 			goto out;
1796 
1797 		if (tp->repair_queue == TCP_SEND_QUEUE)
1798 			goto recv_sndq;
1799 
1800 		err = -EINVAL;
1801 		if (tp->repair_queue == TCP_NO_QUEUE)
1802 			goto out;
1803 
1804 		/* 'common' recv queue MSG_PEEK-ing */
1805 	}
1806 
1807 	seq = &tp->copied_seq;
1808 	if (flags & MSG_PEEK) {
1809 		peek_seq = tp->copied_seq;
1810 		seq = &peek_seq;
1811 	}
1812 
1813 	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1814 
1815 	do {
1816 		u32 offset;
1817 
1818 		/* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1819 		if (tp->urg_data && tp->urg_seq == *seq) {
1820 			if (copied)
1821 				break;
1822 			if (signal_pending(current)) {
1823 				copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1824 				break;
1825 			}
1826 		}
1827 
1828 		/* Next get a buffer. */
1829 
1830 		last = skb_peek_tail(&sk->sk_receive_queue);
1831 		skb_queue_walk(&sk->sk_receive_queue, skb) {
1832 			last = skb;
1833 			/* Now that we have two receive queues this
1834 			 * shouldn't happen.
1835 			 */
1836 			if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1837 				 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1838 				 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1839 				 flags))
1840 				break;
1841 
1842 			offset = *seq - TCP_SKB_CB(skb)->seq;
1843 			if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1844 				pr_err_once("%s: found a SYN, please report !\n", __func__);
1845 				offset--;
1846 			}
1847 			if (offset < skb->len)
1848 				goto found_ok_skb;
1849 			if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1850 				goto found_fin_ok;
1851 			WARN(!(flags & MSG_PEEK),
1852 			     "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1853 			     *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1854 		}
1855 
1856 		/* Well, if we have backlog, try to process it now yet. */
1857 
1858 		if (copied >= target && !sk->sk_backlog.tail)
1859 			break;
1860 
1861 		if (copied) {
1862 			if (sk->sk_err ||
1863 			    sk->sk_state == TCP_CLOSE ||
1864 			    (sk->sk_shutdown & RCV_SHUTDOWN) ||
1865 			    !timeo ||
1866 			    signal_pending(current))
1867 				break;
1868 		} else {
1869 			if (sock_flag(sk, SOCK_DONE))
1870 				break;
1871 
1872 			if (sk->sk_err) {
1873 				copied = sock_error(sk);
1874 				break;
1875 			}
1876 
1877 			if (sk->sk_shutdown & RCV_SHUTDOWN)
1878 				break;
1879 
1880 			if (sk->sk_state == TCP_CLOSE) {
1881 				if (!sock_flag(sk, SOCK_DONE)) {
1882 					/* This occurs when user tries to read
1883 					 * from never connected socket.
1884 					 */
1885 					copied = -ENOTCONN;
1886 					break;
1887 				}
1888 				break;
1889 			}
1890 
1891 			if (!timeo) {
1892 				copied = -EAGAIN;
1893 				break;
1894 			}
1895 
1896 			if (signal_pending(current)) {
1897 				copied = sock_intr_errno(timeo);
1898 				break;
1899 			}
1900 		}
1901 
1902 		tcp_cleanup_rbuf(sk, copied);
1903 
1904 		if (copied >= target) {
1905 			/* Do not sleep, just process backlog. */
1906 			release_sock(sk);
1907 			lock_sock(sk);
1908 		} else {
1909 			sk_wait_data(sk, &timeo, last);
1910 		}
1911 
1912 		if ((flags & MSG_PEEK) &&
1913 		    (peek_seq - copied - urg_hole != tp->copied_seq)) {
1914 			net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1915 					    current->comm,
1916 					    task_pid_nr(current));
1917 			peek_seq = tp->copied_seq;
1918 		}
1919 		continue;
1920 
1921 	found_ok_skb:
1922 		/* Ok so how much can we use? */
1923 		used = skb->len - offset;
1924 		if (len < used)
1925 			used = len;
1926 
1927 		/* Do we have urgent data here? */
1928 		if (tp->urg_data) {
1929 			u32 urg_offset = tp->urg_seq - *seq;
1930 			if (urg_offset < used) {
1931 				if (!urg_offset) {
1932 					if (!sock_flag(sk, SOCK_URGINLINE)) {
1933 						++*seq;
1934 						urg_hole++;
1935 						offset++;
1936 						used--;
1937 						if (!used)
1938 							goto skip_copy;
1939 					}
1940 				} else
1941 					used = urg_offset;
1942 			}
1943 		}
1944 
1945 		if (!(flags & MSG_TRUNC)) {
1946 			err = skb_copy_datagram_msg(skb, offset, msg, used);
1947 			if (err) {
1948 				/* Exception. Bailout! */
1949 				if (!copied)
1950 					copied = -EFAULT;
1951 				break;
1952 			}
1953 		}
1954 
1955 		*seq += used;
1956 		copied += used;
1957 		len -= used;
1958 
1959 		tcp_rcv_space_adjust(sk);
1960 
1961 skip_copy:
1962 		if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1963 			tp->urg_data = 0;
1964 			tcp_fast_path_check(sk);
1965 		}
1966 		if (used + offset < skb->len)
1967 			continue;
1968 
1969 		if (TCP_SKB_CB(skb)->has_rxtstamp) {
1970 			tcp_update_recv_tstamps(skb, &tss);
1971 			has_tss = true;
1972 		}
1973 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1974 			goto found_fin_ok;
1975 		if (!(flags & MSG_PEEK))
1976 			sk_eat_skb(sk, skb);
1977 		continue;
1978 
1979 	found_fin_ok:
1980 		/* Process the FIN. */
1981 		++*seq;
1982 		if (!(flags & MSG_PEEK))
1983 			sk_eat_skb(sk, skb);
1984 		break;
1985 	} while (len > 0);
1986 
1987 	/* According to UNIX98, msg_name/msg_namelen are ignored
1988 	 * on connected socket. I was just happy when found this 8) --ANK
1989 	 */
1990 
1991 	if (has_tss)
1992 		tcp_recv_timestamp(msg, sk, &tss);
1993 
1994 	/* Clean up data we have read: This will do ACK frames. */
1995 	tcp_cleanup_rbuf(sk, copied);
1996 
1997 	release_sock(sk);
1998 	return copied;
1999 
2000 out:
2001 	release_sock(sk);
2002 	return err;
2003 
2004 recv_urg:
2005 	err = tcp_recv_urg(sk, msg, len, flags);
2006 	goto out;
2007 
2008 recv_sndq:
2009 	err = tcp_peek_sndq(sk, msg, len);
2010 	goto out;
2011 }
2012 EXPORT_SYMBOL(tcp_recvmsg);
2013 
2014 void tcp_set_state(struct sock *sk, int state)
2015 {
2016 	int oldstate = sk->sk_state;
2017 
2018 	/* We defined a new enum for TCP states that are exported in BPF
2019 	 * so as not force the internal TCP states to be frozen. The
2020 	 * following checks will detect if an internal state value ever
2021 	 * differs from the BPF value. If this ever happens, then we will
2022 	 * need to remap the internal value to the BPF value before calling
2023 	 * tcp_call_bpf_2arg.
2024 	 */
2025 	BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2026 	BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2027 	BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2028 	BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2029 	BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2030 	BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2031 	BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2032 	BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2033 	BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2034 	BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2035 	BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2036 	BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2037 	BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2038 
2039 	if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2040 		tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2041 
2042 	switch (state) {
2043 	case TCP_ESTABLISHED:
2044 		if (oldstate != TCP_ESTABLISHED)
2045 			TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2046 		break;
2047 
2048 	case TCP_CLOSE:
2049 		if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2050 			TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2051 
2052 		sk->sk_prot->unhash(sk);
2053 		if (inet_csk(sk)->icsk_bind_hash &&
2054 		    !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2055 			inet_put_port(sk);
2056 		/* fall through */
2057 	default:
2058 		if (oldstate == TCP_ESTABLISHED)
2059 			TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2060 	}
2061 
2062 	/* Change state AFTER socket is unhashed to avoid closed
2063 	 * socket sitting in hash tables.
2064 	 */
2065 	inet_sk_state_store(sk, state);
2066 
2067 #ifdef STATE_TRACE
2068 	SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2069 #endif
2070 }
2071 EXPORT_SYMBOL_GPL(tcp_set_state);
2072 
2073 /*
2074  *	State processing on a close. This implements the state shift for
2075  *	sending our FIN frame. Note that we only send a FIN for some
2076  *	states. A shutdown() may have already sent the FIN, or we may be
2077  *	closed.
2078  */
2079 
2080 static const unsigned char new_state[16] = {
2081   /* current state:        new state:      action:	*/
2082   [0 /* (Invalid) */]	= TCP_CLOSE,
2083   [TCP_ESTABLISHED]	= TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2084   [TCP_SYN_SENT]	= TCP_CLOSE,
2085   [TCP_SYN_RECV]	= TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2086   [TCP_FIN_WAIT1]	= TCP_FIN_WAIT1,
2087   [TCP_FIN_WAIT2]	= TCP_FIN_WAIT2,
2088   [TCP_TIME_WAIT]	= TCP_CLOSE,
2089   [TCP_CLOSE]		= TCP_CLOSE,
2090   [TCP_CLOSE_WAIT]	= TCP_LAST_ACK  | TCP_ACTION_FIN,
2091   [TCP_LAST_ACK]	= TCP_LAST_ACK,
2092   [TCP_LISTEN]		= TCP_CLOSE,
2093   [TCP_CLOSING]		= TCP_CLOSING,
2094   [TCP_NEW_SYN_RECV]	= TCP_CLOSE,	/* should not happen ! */
2095 };
2096 
2097 static int tcp_close_state(struct sock *sk)
2098 {
2099 	int next = (int)new_state[sk->sk_state];
2100 	int ns = next & TCP_STATE_MASK;
2101 
2102 	tcp_set_state(sk, ns);
2103 
2104 	return next & TCP_ACTION_FIN;
2105 }
2106 
2107 /*
2108  *	Shutdown the sending side of a connection. Much like close except
2109  *	that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2110  */
2111 
2112 void tcp_shutdown(struct sock *sk, int how)
2113 {
2114 	/*	We need to grab some memory, and put together a FIN,
2115 	 *	and then put it into the queue to be sent.
2116 	 *		Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2117 	 */
2118 	if (!(how & SEND_SHUTDOWN))
2119 		return;
2120 
2121 	/* If we've already sent a FIN, or it's a closed state, skip this. */
2122 	if ((1 << sk->sk_state) &
2123 	    (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2124 	     TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2125 		/* Clear out any half completed packets.  FIN if needed. */
2126 		if (tcp_close_state(sk))
2127 			tcp_send_fin(sk);
2128 	}
2129 }
2130 EXPORT_SYMBOL(tcp_shutdown);
2131 
2132 bool tcp_check_oom(struct sock *sk, int shift)
2133 {
2134 	bool too_many_orphans, out_of_socket_memory;
2135 
2136 	too_many_orphans = tcp_too_many_orphans(sk, shift);
2137 	out_of_socket_memory = tcp_out_of_memory(sk);
2138 
2139 	if (too_many_orphans)
2140 		net_info_ratelimited("too many orphaned sockets\n");
2141 	if (out_of_socket_memory)
2142 		net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2143 	return too_many_orphans || out_of_socket_memory;
2144 }
2145 
2146 void tcp_close(struct sock *sk, long timeout)
2147 {
2148 	struct sk_buff *skb;
2149 	int data_was_unread = 0;
2150 	int state;
2151 
2152 	lock_sock(sk);
2153 	sk->sk_shutdown = SHUTDOWN_MASK;
2154 
2155 	if (sk->sk_state == TCP_LISTEN) {
2156 		tcp_set_state(sk, TCP_CLOSE);
2157 
2158 		/* Special case. */
2159 		inet_csk_listen_stop(sk);
2160 
2161 		goto adjudge_to_death;
2162 	}
2163 
2164 	/*  We need to flush the recv. buffs.  We do this only on the
2165 	 *  descriptor close, not protocol-sourced closes, because the
2166 	 *  reader process may not have drained the data yet!
2167 	 */
2168 	while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2169 		u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2170 
2171 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2172 			len--;
2173 		data_was_unread += len;
2174 		__kfree_skb(skb);
2175 	}
2176 
2177 	sk_mem_reclaim(sk);
2178 
2179 	/* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2180 	if (sk->sk_state == TCP_CLOSE)
2181 		goto adjudge_to_death;
2182 
2183 	/* As outlined in RFC 2525, section 2.17, we send a RST here because
2184 	 * data was lost. To witness the awful effects of the old behavior of
2185 	 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2186 	 * GET in an FTP client, suspend the process, wait for the client to
2187 	 * advertise a zero window, then kill -9 the FTP client, wheee...
2188 	 * Note: timeout is always zero in such a case.
2189 	 */
2190 	if (unlikely(tcp_sk(sk)->repair)) {
2191 		sk->sk_prot->disconnect(sk, 0);
2192 	} else if (data_was_unread) {
2193 		/* Unread data was tossed, zap the connection. */
2194 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2195 		tcp_set_state(sk, TCP_CLOSE);
2196 		tcp_send_active_reset(sk, sk->sk_allocation);
2197 	} else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2198 		/* Check zero linger _after_ checking for unread data. */
2199 		sk->sk_prot->disconnect(sk, 0);
2200 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2201 	} else if (tcp_close_state(sk)) {
2202 		/* We FIN if the application ate all the data before
2203 		 * zapping the connection.
2204 		 */
2205 
2206 		/* RED-PEN. Formally speaking, we have broken TCP state
2207 		 * machine. State transitions:
2208 		 *
2209 		 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2210 		 * TCP_SYN_RECV	-> TCP_FIN_WAIT1 (forget it, it's impossible)
2211 		 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2212 		 *
2213 		 * are legal only when FIN has been sent (i.e. in window),
2214 		 * rather than queued out of window. Purists blame.
2215 		 *
2216 		 * F.e. "RFC state" is ESTABLISHED,
2217 		 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2218 		 *
2219 		 * The visible declinations are that sometimes
2220 		 * we enter time-wait state, when it is not required really
2221 		 * (harmless), do not send active resets, when they are
2222 		 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2223 		 * they look as CLOSING or LAST_ACK for Linux)
2224 		 * Probably, I missed some more holelets.
2225 		 * 						--ANK
2226 		 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2227 		 * in a single packet! (May consider it later but will
2228 		 * probably need API support or TCP_CORK SYN-ACK until
2229 		 * data is written and socket is closed.)
2230 		 */
2231 		tcp_send_fin(sk);
2232 	}
2233 
2234 	sk_stream_wait_close(sk, timeout);
2235 
2236 adjudge_to_death:
2237 	state = sk->sk_state;
2238 	sock_hold(sk);
2239 	sock_orphan(sk);
2240 
2241 	/* It is the last release_sock in its life. It will remove backlog. */
2242 	release_sock(sk);
2243 
2244 
2245 	/* Now socket is owned by kernel and we acquire BH lock
2246 	 *  to finish close. No need to check for user refs.
2247 	 */
2248 	local_bh_disable();
2249 	bh_lock_sock(sk);
2250 	WARN_ON(sock_owned_by_user(sk));
2251 
2252 	percpu_counter_inc(sk->sk_prot->orphan_count);
2253 
2254 	/* Have we already been destroyed by a softirq or backlog? */
2255 	if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2256 		goto out;
2257 
2258 	/*	This is a (useful) BSD violating of the RFC. There is a
2259 	 *	problem with TCP as specified in that the other end could
2260 	 *	keep a socket open forever with no application left this end.
2261 	 *	We use a 1 minute timeout (about the same as BSD) then kill
2262 	 *	our end. If they send after that then tough - BUT: long enough
2263 	 *	that we won't make the old 4*rto = almost no time - whoops
2264 	 *	reset mistake.
2265 	 *
2266 	 *	Nope, it was not mistake. It is really desired behaviour
2267 	 *	f.e. on http servers, when such sockets are useless, but
2268 	 *	consume significant resources. Let's do it with special
2269 	 *	linger2	option.					--ANK
2270 	 */
2271 
2272 	if (sk->sk_state == TCP_FIN_WAIT2) {
2273 		struct tcp_sock *tp = tcp_sk(sk);
2274 		if (tp->linger2 < 0) {
2275 			tcp_set_state(sk, TCP_CLOSE);
2276 			tcp_send_active_reset(sk, GFP_ATOMIC);
2277 			__NET_INC_STATS(sock_net(sk),
2278 					LINUX_MIB_TCPABORTONLINGER);
2279 		} else {
2280 			const int tmo = tcp_fin_time(sk);
2281 
2282 			if (tmo > TCP_TIMEWAIT_LEN) {
2283 				inet_csk_reset_keepalive_timer(sk,
2284 						tmo - TCP_TIMEWAIT_LEN);
2285 			} else {
2286 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2287 				goto out;
2288 			}
2289 		}
2290 	}
2291 	if (sk->sk_state != TCP_CLOSE) {
2292 		sk_mem_reclaim(sk);
2293 		if (tcp_check_oom(sk, 0)) {
2294 			tcp_set_state(sk, TCP_CLOSE);
2295 			tcp_send_active_reset(sk, GFP_ATOMIC);
2296 			__NET_INC_STATS(sock_net(sk),
2297 					LINUX_MIB_TCPABORTONMEMORY);
2298 		} else if (!check_net(sock_net(sk))) {
2299 			/* Not possible to send reset; just close */
2300 			tcp_set_state(sk, TCP_CLOSE);
2301 		}
2302 	}
2303 
2304 	if (sk->sk_state == TCP_CLOSE) {
2305 		struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2306 		/* We could get here with a non-NULL req if the socket is
2307 		 * aborted (e.g., closed with unread data) before 3WHS
2308 		 * finishes.
2309 		 */
2310 		if (req)
2311 			reqsk_fastopen_remove(sk, req, false);
2312 		inet_csk_destroy_sock(sk);
2313 	}
2314 	/* Otherwise, socket is reprieved until protocol close. */
2315 
2316 out:
2317 	bh_unlock_sock(sk);
2318 	local_bh_enable();
2319 	sock_put(sk);
2320 }
2321 EXPORT_SYMBOL(tcp_close);
2322 
2323 /* These states need RST on ABORT according to RFC793 */
2324 
2325 static inline bool tcp_need_reset(int state)
2326 {
2327 	return (1 << state) &
2328 	       (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2329 		TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2330 }
2331 
2332 static void tcp_rtx_queue_purge(struct sock *sk)
2333 {
2334 	struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2335 
2336 	while (p) {
2337 		struct sk_buff *skb = rb_to_skb(p);
2338 
2339 		p = rb_next(p);
2340 		/* Since we are deleting whole queue, no need to
2341 		 * list_del(&skb->tcp_tsorted_anchor)
2342 		 */
2343 		tcp_rtx_queue_unlink(skb, sk);
2344 		sk_wmem_free_skb(sk, skb);
2345 	}
2346 }
2347 
2348 void tcp_write_queue_purge(struct sock *sk)
2349 {
2350 	struct sk_buff *skb;
2351 
2352 	tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2353 	while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2354 		tcp_skb_tsorted_anchor_cleanup(skb);
2355 		sk_wmem_free_skb(sk, skb);
2356 	}
2357 	tcp_rtx_queue_purge(sk);
2358 	INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2359 	sk_mem_reclaim(sk);
2360 	tcp_clear_all_retrans_hints(tcp_sk(sk));
2361 }
2362 
2363 int tcp_disconnect(struct sock *sk, int flags)
2364 {
2365 	struct inet_sock *inet = inet_sk(sk);
2366 	struct inet_connection_sock *icsk = inet_csk(sk);
2367 	struct tcp_sock *tp = tcp_sk(sk);
2368 	int err = 0;
2369 	int old_state = sk->sk_state;
2370 
2371 	if (old_state != TCP_CLOSE)
2372 		tcp_set_state(sk, TCP_CLOSE);
2373 
2374 	/* ABORT function of RFC793 */
2375 	if (old_state == TCP_LISTEN) {
2376 		inet_csk_listen_stop(sk);
2377 	} else if (unlikely(tp->repair)) {
2378 		sk->sk_err = ECONNABORTED;
2379 	} else if (tcp_need_reset(old_state) ||
2380 		   (tp->snd_nxt != tp->write_seq &&
2381 		    (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2382 		/* The last check adjusts for discrepancy of Linux wrt. RFC
2383 		 * states
2384 		 */
2385 		tcp_send_active_reset(sk, gfp_any());
2386 		sk->sk_err = ECONNRESET;
2387 	} else if (old_state == TCP_SYN_SENT)
2388 		sk->sk_err = ECONNRESET;
2389 
2390 	tcp_clear_xmit_timers(sk);
2391 	__skb_queue_purge(&sk->sk_receive_queue);
2392 	tcp_write_queue_purge(sk);
2393 	tcp_fastopen_active_disable_ofo_check(sk);
2394 	skb_rbtree_purge(&tp->out_of_order_queue);
2395 
2396 	inet->inet_dport = 0;
2397 
2398 	if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2399 		inet_reset_saddr(sk);
2400 
2401 	sk->sk_shutdown = 0;
2402 	sock_reset_flag(sk, SOCK_DONE);
2403 	tp->srtt_us = 0;
2404 	tp->write_seq += tp->max_window + 2;
2405 	if (tp->write_seq == 0)
2406 		tp->write_seq = 1;
2407 	icsk->icsk_backoff = 0;
2408 	tp->snd_cwnd = 2;
2409 	icsk->icsk_probes_out = 0;
2410 	tp->packets_out = 0;
2411 	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2412 	tp->snd_cwnd_cnt = 0;
2413 	tp->window_clamp = 0;
2414 	tcp_set_ca_state(sk, TCP_CA_Open);
2415 	tp->is_sack_reneg = 0;
2416 	tcp_clear_retrans(tp);
2417 	inet_csk_delack_init(sk);
2418 	/* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2419 	 * issue in __tcp_select_window()
2420 	 */
2421 	icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2422 	memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2423 	__sk_dst_reset(sk);
2424 	dst_release(sk->sk_rx_dst);
2425 	sk->sk_rx_dst = NULL;
2426 	tcp_saved_syn_free(tp);
2427 
2428 	/* Clean up fastopen related fields */
2429 	tcp_free_fastopen_req(tp);
2430 	inet->defer_connect = 0;
2431 
2432 	WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2433 
2434 	if (sk->sk_frag.page) {
2435 		put_page(sk->sk_frag.page);
2436 		sk->sk_frag.page = NULL;
2437 		sk->sk_frag.offset = 0;
2438 	}
2439 
2440 	sk->sk_error_report(sk);
2441 	return err;
2442 }
2443 EXPORT_SYMBOL(tcp_disconnect);
2444 
2445 static inline bool tcp_can_repair_sock(const struct sock *sk)
2446 {
2447 	return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2448 		(sk->sk_state != TCP_LISTEN);
2449 }
2450 
2451 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2452 {
2453 	struct tcp_repair_window opt;
2454 
2455 	if (!tp->repair)
2456 		return -EPERM;
2457 
2458 	if (len != sizeof(opt))
2459 		return -EINVAL;
2460 
2461 	if (copy_from_user(&opt, optbuf, sizeof(opt)))
2462 		return -EFAULT;
2463 
2464 	if (opt.max_window < opt.snd_wnd)
2465 		return -EINVAL;
2466 
2467 	if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2468 		return -EINVAL;
2469 
2470 	if (after(opt.rcv_wup, tp->rcv_nxt))
2471 		return -EINVAL;
2472 
2473 	tp->snd_wl1	= opt.snd_wl1;
2474 	tp->snd_wnd	= opt.snd_wnd;
2475 	tp->max_window	= opt.max_window;
2476 
2477 	tp->rcv_wnd	= opt.rcv_wnd;
2478 	tp->rcv_wup	= opt.rcv_wup;
2479 
2480 	return 0;
2481 }
2482 
2483 static int tcp_repair_options_est(struct sock *sk,
2484 		struct tcp_repair_opt __user *optbuf, unsigned int len)
2485 {
2486 	struct tcp_sock *tp = tcp_sk(sk);
2487 	struct tcp_repair_opt opt;
2488 
2489 	while (len >= sizeof(opt)) {
2490 		if (copy_from_user(&opt, optbuf, sizeof(opt)))
2491 			return -EFAULT;
2492 
2493 		optbuf++;
2494 		len -= sizeof(opt);
2495 
2496 		switch (opt.opt_code) {
2497 		case TCPOPT_MSS:
2498 			tp->rx_opt.mss_clamp = opt.opt_val;
2499 			tcp_mtup_init(sk);
2500 			break;
2501 		case TCPOPT_WINDOW:
2502 			{
2503 				u16 snd_wscale = opt.opt_val & 0xFFFF;
2504 				u16 rcv_wscale = opt.opt_val >> 16;
2505 
2506 				if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2507 					return -EFBIG;
2508 
2509 				tp->rx_opt.snd_wscale = snd_wscale;
2510 				tp->rx_opt.rcv_wscale = rcv_wscale;
2511 				tp->rx_opt.wscale_ok = 1;
2512 			}
2513 			break;
2514 		case TCPOPT_SACK_PERM:
2515 			if (opt.opt_val != 0)
2516 				return -EINVAL;
2517 
2518 			tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2519 			break;
2520 		case TCPOPT_TIMESTAMP:
2521 			if (opt.opt_val != 0)
2522 				return -EINVAL;
2523 
2524 			tp->rx_opt.tstamp_ok = 1;
2525 			break;
2526 		}
2527 	}
2528 
2529 	return 0;
2530 }
2531 
2532 /*
2533  *	Socket option code for TCP.
2534  */
2535 static int do_tcp_setsockopt(struct sock *sk, int level,
2536 		int optname, char __user *optval, unsigned int optlen)
2537 {
2538 	struct tcp_sock *tp = tcp_sk(sk);
2539 	struct inet_connection_sock *icsk = inet_csk(sk);
2540 	struct net *net = sock_net(sk);
2541 	int val;
2542 	int err = 0;
2543 
2544 	/* These are data/string values, all the others are ints */
2545 	switch (optname) {
2546 	case TCP_CONGESTION: {
2547 		char name[TCP_CA_NAME_MAX];
2548 
2549 		if (optlen < 1)
2550 			return -EINVAL;
2551 
2552 		val = strncpy_from_user(name, optval,
2553 					min_t(long, TCP_CA_NAME_MAX-1, optlen));
2554 		if (val < 0)
2555 			return -EFAULT;
2556 		name[val] = 0;
2557 
2558 		lock_sock(sk);
2559 		err = tcp_set_congestion_control(sk, name, true, true);
2560 		release_sock(sk);
2561 		return err;
2562 	}
2563 	case TCP_ULP: {
2564 		char name[TCP_ULP_NAME_MAX];
2565 
2566 		if (optlen < 1)
2567 			return -EINVAL;
2568 
2569 		val = strncpy_from_user(name, optval,
2570 					min_t(long, TCP_ULP_NAME_MAX - 1,
2571 					      optlen));
2572 		if (val < 0)
2573 			return -EFAULT;
2574 		name[val] = 0;
2575 
2576 		lock_sock(sk);
2577 		err = tcp_set_ulp(sk, name);
2578 		release_sock(sk);
2579 		return err;
2580 	}
2581 	case TCP_FASTOPEN_KEY: {
2582 		__u8 key[TCP_FASTOPEN_KEY_LENGTH];
2583 
2584 		if (optlen != sizeof(key))
2585 			return -EINVAL;
2586 
2587 		if (copy_from_user(key, optval, optlen))
2588 			return -EFAULT;
2589 
2590 		return tcp_fastopen_reset_cipher(net, sk, key, sizeof(key));
2591 	}
2592 	default:
2593 		/* fallthru */
2594 		break;
2595 	}
2596 
2597 	if (optlen < sizeof(int))
2598 		return -EINVAL;
2599 
2600 	if (get_user(val, (int __user *)optval))
2601 		return -EFAULT;
2602 
2603 	lock_sock(sk);
2604 
2605 	switch (optname) {
2606 	case TCP_MAXSEG:
2607 		/* Values greater than interface MTU won't take effect. However
2608 		 * at the point when this call is done we typically don't yet
2609 		 * know which interface is going to be used
2610 		 */
2611 		if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2612 			err = -EINVAL;
2613 			break;
2614 		}
2615 		tp->rx_opt.user_mss = val;
2616 		break;
2617 
2618 	case TCP_NODELAY:
2619 		if (val) {
2620 			/* TCP_NODELAY is weaker than TCP_CORK, so that
2621 			 * this option on corked socket is remembered, but
2622 			 * it is not activated until cork is cleared.
2623 			 *
2624 			 * However, when TCP_NODELAY is set we make
2625 			 * an explicit push, which overrides even TCP_CORK
2626 			 * for currently queued segments.
2627 			 */
2628 			tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2629 			tcp_push_pending_frames(sk);
2630 		} else {
2631 			tp->nonagle &= ~TCP_NAGLE_OFF;
2632 		}
2633 		break;
2634 
2635 	case TCP_THIN_LINEAR_TIMEOUTS:
2636 		if (val < 0 || val > 1)
2637 			err = -EINVAL;
2638 		else
2639 			tp->thin_lto = val;
2640 		break;
2641 
2642 	case TCP_THIN_DUPACK:
2643 		if (val < 0 || val > 1)
2644 			err = -EINVAL;
2645 		break;
2646 
2647 	case TCP_REPAIR:
2648 		if (!tcp_can_repair_sock(sk))
2649 			err = -EPERM;
2650 		else if (val == 1) {
2651 			tp->repair = 1;
2652 			sk->sk_reuse = SK_FORCE_REUSE;
2653 			tp->repair_queue = TCP_NO_QUEUE;
2654 		} else if (val == 0) {
2655 			tp->repair = 0;
2656 			sk->sk_reuse = SK_NO_REUSE;
2657 			tcp_send_window_probe(sk);
2658 		} else
2659 			err = -EINVAL;
2660 
2661 		break;
2662 
2663 	case TCP_REPAIR_QUEUE:
2664 		if (!tp->repair)
2665 			err = -EPERM;
2666 		else if (val < TCP_QUEUES_NR)
2667 			tp->repair_queue = val;
2668 		else
2669 			err = -EINVAL;
2670 		break;
2671 
2672 	case TCP_QUEUE_SEQ:
2673 		if (sk->sk_state != TCP_CLOSE)
2674 			err = -EPERM;
2675 		else if (tp->repair_queue == TCP_SEND_QUEUE)
2676 			tp->write_seq = val;
2677 		else if (tp->repair_queue == TCP_RECV_QUEUE)
2678 			tp->rcv_nxt = val;
2679 		else
2680 			err = -EINVAL;
2681 		break;
2682 
2683 	case TCP_REPAIR_OPTIONS:
2684 		if (!tp->repair)
2685 			err = -EINVAL;
2686 		else if (sk->sk_state == TCP_ESTABLISHED)
2687 			err = tcp_repair_options_est(sk,
2688 					(struct tcp_repair_opt __user *)optval,
2689 					optlen);
2690 		else
2691 			err = -EPERM;
2692 		break;
2693 
2694 	case TCP_CORK:
2695 		/* When set indicates to always queue non-full frames.
2696 		 * Later the user clears this option and we transmit
2697 		 * any pending partial frames in the queue.  This is
2698 		 * meant to be used alongside sendfile() to get properly
2699 		 * filled frames when the user (for example) must write
2700 		 * out headers with a write() call first and then use
2701 		 * sendfile to send out the data parts.
2702 		 *
2703 		 * TCP_CORK can be set together with TCP_NODELAY and it is
2704 		 * stronger than TCP_NODELAY.
2705 		 */
2706 		if (val) {
2707 			tp->nonagle |= TCP_NAGLE_CORK;
2708 		} else {
2709 			tp->nonagle &= ~TCP_NAGLE_CORK;
2710 			if (tp->nonagle&TCP_NAGLE_OFF)
2711 				tp->nonagle |= TCP_NAGLE_PUSH;
2712 			tcp_push_pending_frames(sk);
2713 		}
2714 		break;
2715 
2716 	case TCP_KEEPIDLE:
2717 		if (val < 1 || val > MAX_TCP_KEEPIDLE)
2718 			err = -EINVAL;
2719 		else {
2720 			tp->keepalive_time = val * HZ;
2721 			if (sock_flag(sk, SOCK_KEEPOPEN) &&
2722 			    !((1 << sk->sk_state) &
2723 			      (TCPF_CLOSE | TCPF_LISTEN))) {
2724 				u32 elapsed = keepalive_time_elapsed(tp);
2725 				if (tp->keepalive_time > elapsed)
2726 					elapsed = tp->keepalive_time - elapsed;
2727 				else
2728 					elapsed = 0;
2729 				inet_csk_reset_keepalive_timer(sk, elapsed);
2730 			}
2731 		}
2732 		break;
2733 	case TCP_KEEPINTVL:
2734 		if (val < 1 || val > MAX_TCP_KEEPINTVL)
2735 			err = -EINVAL;
2736 		else
2737 			tp->keepalive_intvl = val * HZ;
2738 		break;
2739 	case TCP_KEEPCNT:
2740 		if (val < 1 || val > MAX_TCP_KEEPCNT)
2741 			err = -EINVAL;
2742 		else
2743 			tp->keepalive_probes = val;
2744 		break;
2745 	case TCP_SYNCNT:
2746 		if (val < 1 || val > MAX_TCP_SYNCNT)
2747 			err = -EINVAL;
2748 		else
2749 			icsk->icsk_syn_retries = val;
2750 		break;
2751 
2752 	case TCP_SAVE_SYN:
2753 		if (val < 0 || val > 1)
2754 			err = -EINVAL;
2755 		else
2756 			tp->save_syn = val;
2757 		break;
2758 
2759 	case TCP_LINGER2:
2760 		if (val < 0)
2761 			tp->linger2 = -1;
2762 		else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2763 			tp->linger2 = 0;
2764 		else
2765 			tp->linger2 = val * HZ;
2766 		break;
2767 
2768 	case TCP_DEFER_ACCEPT:
2769 		/* Translate value in seconds to number of retransmits */
2770 		icsk->icsk_accept_queue.rskq_defer_accept =
2771 			secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2772 					TCP_RTO_MAX / HZ);
2773 		break;
2774 
2775 	case TCP_WINDOW_CLAMP:
2776 		if (!val) {
2777 			if (sk->sk_state != TCP_CLOSE) {
2778 				err = -EINVAL;
2779 				break;
2780 			}
2781 			tp->window_clamp = 0;
2782 		} else
2783 			tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2784 						SOCK_MIN_RCVBUF / 2 : val;
2785 		break;
2786 
2787 	case TCP_QUICKACK:
2788 		if (!val) {
2789 			icsk->icsk_ack.pingpong = 1;
2790 		} else {
2791 			icsk->icsk_ack.pingpong = 0;
2792 			if ((1 << sk->sk_state) &
2793 			    (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2794 			    inet_csk_ack_scheduled(sk)) {
2795 				icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2796 				tcp_cleanup_rbuf(sk, 1);
2797 				if (!(val & 1))
2798 					icsk->icsk_ack.pingpong = 1;
2799 			}
2800 		}
2801 		break;
2802 
2803 #ifdef CONFIG_TCP_MD5SIG
2804 	case TCP_MD5SIG:
2805 	case TCP_MD5SIG_EXT:
2806 		/* Read the IP->Key mappings from userspace */
2807 		err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
2808 		break;
2809 #endif
2810 	case TCP_USER_TIMEOUT:
2811 		/* Cap the max time in ms TCP will retry or probe the window
2812 		 * before giving up and aborting (ETIMEDOUT) a connection.
2813 		 */
2814 		if (val < 0)
2815 			err = -EINVAL;
2816 		else
2817 			icsk->icsk_user_timeout = msecs_to_jiffies(val);
2818 		break;
2819 
2820 	case TCP_FASTOPEN:
2821 		if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2822 		    TCPF_LISTEN))) {
2823 			tcp_fastopen_init_key_once(net);
2824 
2825 			fastopen_queue_tune(sk, val);
2826 		} else {
2827 			err = -EINVAL;
2828 		}
2829 		break;
2830 	case TCP_FASTOPEN_CONNECT:
2831 		if (val > 1 || val < 0) {
2832 			err = -EINVAL;
2833 		} else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
2834 			if (sk->sk_state == TCP_CLOSE)
2835 				tp->fastopen_connect = val;
2836 			else
2837 				err = -EINVAL;
2838 		} else {
2839 			err = -EOPNOTSUPP;
2840 		}
2841 		break;
2842 	case TCP_FASTOPEN_NO_COOKIE:
2843 		if (val > 1 || val < 0)
2844 			err = -EINVAL;
2845 		else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2846 			err = -EINVAL;
2847 		else
2848 			tp->fastopen_no_cookie = val;
2849 		break;
2850 	case TCP_TIMESTAMP:
2851 		if (!tp->repair)
2852 			err = -EPERM;
2853 		else
2854 			tp->tsoffset = val - tcp_time_stamp_raw();
2855 		break;
2856 	case TCP_REPAIR_WINDOW:
2857 		err = tcp_repair_set_window(tp, optval, optlen);
2858 		break;
2859 	case TCP_NOTSENT_LOWAT:
2860 		tp->notsent_lowat = val;
2861 		sk->sk_write_space(sk);
2862 		break;
2863 	default:
2864 		err = -ENOPROTOOPT;
2865 		break;
2866 	}
2867 
2868 	release_sock(sk);
2869 	return err;
2870 }
2871 
2872 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2873 		   unsigned int optlen)
2874 {
2875 	const struct inet_connection_sock *icsk = inet_csk(sk);
2876 
2877 	if (level != SOL_TCP)
2878 		return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2879 						     optval, optlen);
2880 	return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2881 }
2882 EXPORT_SYMBOL(tcp_setsockopt);
2883 
2884 #ifdef CONFIG_COMPAT
2885 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2886 			  char __user *optval, unsigned int optlen)
2887 {
2888 	if (level != SOL_TCP)
2889 		return inet_csk_compat_setsockopt(sk, level, optname,
2890 						  optval, optlen);
2891 	return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2892 }
2893 EXPORT_SYMBOL(compat_tcp_setsockopt);
2894 #endif
2895 
2896 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
2897 				      struct tcp_info *info)
2898 {
2899 	u64 stats[__TCP_CHRONO_MAX], total = 0;
2900 	enum tcp_chrono i;
2901 
2902 	for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
2903 		stats[i] = tp->chrono_stat[i - 1];
2904 		if (i == tp->chrono_type)
2905 			stats[i] += tcp_jiffies32 - tp->chrono_start;
2906 		stats[i] *= USEC_PER_SEC / HZ;
2907 		total += stats[i];
2908 	}
2909 
2910 	info->tcpi_busy_time = total;
2911 	info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
2912 	info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
2913 }
2914 
2915 /* Return information about state of tcp endpoint in API format. */
2916 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2917 {
2918 	const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2919 	const struct inet_connection_sock *icsk = inet_csk(sk);
2920 	u32 now;
2921 	u64 rate64;
2922 	bool slow;
2923 	u32 rate;
2924 
2925 	memset(info, 0, sizeof(*info));
2926 	if (sk->sk_type != SOCK_STREAM)
2927 		return;
2928 
2929 	info->tcpi_state = inet_sk_state_load(sk);
2930 
2931 	/* Report meaningful fields for all TCP states, including listeners */
2932 	rate = READ_ONCE(sk->sk_pacing_rate);
2933 	rate64 = rate != ~0U ? rate : ~0ULL;
2934 	info->tcpi_pacing_rate = rate64;
2935 
2936 	rate = READ_ONCE(sk->sk_max_pacing_rate);
2937 	rate64 = rate != ~0U ? rate : ~0ULL;
2938 	info->tcpi_max_pacing_rate = rate64;
2939 
2940 	info->tcpi_reordering = tp->reordering;
2941 	info->tcpi_snd_cwnd = tp->snd_cwnd;
2942 
2943 	if (info->tcpi_state == TCP_LISTEN) {
2944 		/* listeners aliased fields :
2945 		 * tcpi_unacked -> Number of children ready for accept()
2946 		 * tcpi_sacked  -> max backlog
2947 		 */
2948 		info->tcpi_unacked = sk->sk_ack_backlog;
2949 		info->tcpi_sacked = sk->sk_max_ack_backlog;
2950 		return;
2951 	}
2952 
2953 	slow = lock_sock_fast(sk);
2954 
2955 	info->tcpi_ca_state = icsk->icsk_ca_state;
2956 	info->tcpi_retransmits = icsk->icsk_retransmits;
2957 	info->tcpi_probes = icsk->icsk_probes_out;
2958 	info->tcpi_backoff = icsk->icsk_backoff;
2959 
2960 	if (tp->rx_opt.tstamp_ok)
2961 		info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2962 	if (tcp_is_sack(tp))
2963 		info->tcpi_options |= TCPI_OPT_SACK;
2964 	if (tp->rx_opt.wscale_ok) {
2965 		info->tcpi_options |= TCPI_OPT_WSCALE;
2966 		info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2967 		info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2968 	}
2969 
2970 	if (tp->ecn_flags & TCP_ECN_OK)
2971 		info->tcpi_options |= TCPI_OPT_ECN;
2972 	if (tp->ecn_flags & TCP_ECN_SEEN)
2973 		info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2974 	if (tp->syn_data_acked)
2975 		info->tcpi_options |= TCPI_OPT_SYN_DATA;
2976 
2977 	info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2978 	info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2979 	info->tcpi_snd_mss = tp->mss_cache;
2980 	info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2981 
2982 	info->tcpi_unacked = tp->packets_out;
2983 	info->tcpi_sacked = tp->sacked_out;
2984 
2985 	info->tcpi_lost = tp->lost_out;
2986 	info->tcpi_retrans = tp->retrans_out;
2987 
2988 	now = tcp_jiffies32;
2989 	info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2990 	info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2991 	info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2992 
2993 	info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2994 	info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2995 	info->tcpi_rtt = tp->srtt_us >> 3;
2996 	info->tcpi_rttvar = tp->mdev_us >> 2;
2997 	info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2998 	info->tcpi_advmss = tp->advmss;
2999 
3000 	info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3001 	info->tcpi_rcv_space = tp->rcvq_space.space;
3002 
3003 	info->tcpi_total_retrans = tp->total_retrans;
3004 
3005 	info->tcpi_bytes_acked = tp->bytes_acked;
3006 	info->tcpi_bytes_received = tp->bytes_received;
3007 	info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3008 	tcp_get_info_chrono_stats(tp, info);
3009 
3010 	info->tcpi_segs_out = tp->segs_out;
3011 	info->tcpi_segs_in = tp->segs_in;
3012 
3013 	info->tcpi_min_rtt = tcp_min_rtt(tp);
3014 	info->tcpi_data_segs_in = tp->data_segs_in;
3015 	info->tcpi_data_segs_out = tp->data_segs_out;
3016 
3017 	info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3018 	rate64 = tcp_compute_delivery_rate(tp);
3019 	if (rate64)
3020 		info->tcpi_delivery_rate = rate64;
3021 	unlock_sock_fast(sk, slow);
3022 }
3023 EXPORT_SYMBOL_GPL(tcp_get_info);
3024 
3025 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
3026 {
3027 	const struct tcp_sock *tp = tcp_sk(sk);
3028 	struct sk_buff *stats;
3029 	struct tcp_info info;
3030 	u64 rate64;
3031 	u32 rate;
3032 
3033 	stats = alloc_skb(7 * nla_total_size_64bit(sizeof(u64)) +
3034 			  3 * nla_total_size(sizeof(u32)) +
3035 			  2 * nla_total_size(sizeof(u8)), GFP_ATOMIC);
3036 	if (!stats)
3037 		return NULL;
3038 
3039 	tcp_get_info_chrono_stats(tp, &info);
3040 	nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3041 			  info.tcpi_busy_time, TCP_NLA_PAD);
3042 	nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3043 			  info.tcpi_rwnd_limited, TCP_NLA_PAD);
3044 	nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3045 			  info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3046 	nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3047 			  tp->data_segs_out, TCP_NLA_PAD);
3048 	nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3049 			  tp->total_retrans, TCP_NLA_PAD);
3050 
3051 	rate = READ_ONCE(sk->sk_pacing_rate);
3052 	rate64 = rate != ~0U ? rate : ~0ULL;
3053 	nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3054 
3055 	rate64 = tcp_compute_delivery_rate(tp);
3056 	nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3057 
3058 	nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3059 	nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3060 	nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3061 
3062 	nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3063 	nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3064 	return stats;
3065 }
3066 
3067 static int do_tcp_getsockopt(struct sock *sk, int level,
3068 		int optname, char __user *optval, int __user *optlen)
3069 {
3070 	struct inet_connection_sock *icsk = inet_csk(sk);
3071 	struct tcp_sock *tp = tcp_sk(sk);
3072 	struct net *net = sock_net(sk);
3073 	int val, len;
3074 
3075 	if (get_user(len, optlen))
3076 		return -EFAULT;
3077 
3078 	len = min_t(unsigned int, len, sizeof(int));
3079 
3080 	if (len < 0)
3081 		return -EINVAL;
3082 
3083 	switch (optname) {
3084 	case TCP_MAXSEG:
3085 		val = tp->mss_cache;
3086 		if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3087 			val = tp->rx_opt.user_mss;
3088 		if (tp->repair)
3089 			val = tp->rx_opt.mss_clamp;
3090 		break;
3091 	case TCP_NODELAY:
3092 		val = !!(tp->nonagle&TCP_NAGLE_OFF);
3093 		break;
3094 	case TCP_CORK:
3095 		val = !!(tp->nonagle&TCP_NAGLE_CORK);
3096 		break;
3097 	case TCP_KEEPIDLE:
3098 		val = keepalive_time_when(tp) / HZ;
3099 		break;
3100 	case TCP_KEEPINTVL:
3101 		val = keepalive_intvl_when(tp) / HZ;
3102 		break;
3103 	case TCP_KEEPCNT:
3104 		val = keepalive_probes(tp);
3105 		break;
3106 	case TCP_SYNCNT:
3107 		val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3108 		break;
3109 	case TCP_LINGER2:
3110 		val = tp->linger2;
3111 		if (val >= 0)
3112 			val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3113 		break;
3114 	case TCP_DEFER_ACCEPT:
3115 		val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3116 				      TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3117 		break;
3118 	case TCP_WINDOW_CLAMP:
3119 		val = tp->window_clamp;
3120 		break;
3121 	case TCP_INFO: {
3122 		struct tcp_info info;
3123 
3124 		if (get_user(len, optlen))
3125 			return -EFAULT;
3126 
3127 		tcp_get_info(sk, &info);
3128 
3129 		len = min_t(unsigned int, len, sizeof(info));
3130 		if (put_user(len, optlen))
3131 			return -EFAULT;
3132 		if (copy_to_user(optval, &info, len))
3133 			return -EFAULT;
3134 		return 0;
3135 	}
3136 	case TCP_CC_INFO: {
3137 		const struct tcp_congestion_ops *ca_ops;
3138 		union tcp_cc_info info;
3139 		size_t sz = 0;
3140 		int attr;
3141 
3142 		if (get_user(len, optlen))
3143 			return -EFAULT;
3144 
3145 		ca_ops = icsk->icsk_ca_ops;
3146 		if (ca_ops && ca_ops->get_info)
3147 			sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3148 
3149 		len = min_t(unsigned int, len, sz);
3150 		if (put_user(len, optlen))
3151 			return -EFAULT;
3152 		if (copy_to_user(optval, &info, len))
3153 			return -EFAULT;
3154 		return 0;
3155 	}
3156 	case TCP_QUICKACK:
3157 		val = !icsk->icsk_ack.pingpong;
3158 		break;
3159 
3160 	case TCP_CONGESTION:
3161 		if (get_user(len, optlen))
3162 			return -EFAULT;
3163 		len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3164 		if (put_user(len, optlen))
3165 			return -EFAULT;
3166 		if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3167 			return -EFAULT;
3168 		return 0;
3169 
3170 	case TCP_ULP:
3171 		if (get_user(len, optlen))
3172 			return -EFAULT;
3173 		len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3174 		if (!icsk->icsk_ulp_ops) {
3175 			if (put_user(0, optlen))
3176 				return -EFAULT;
3177 			return 0;
3178 		}
3179 		if (put_user(len, optlen))
3180 			return -EFAULT;
3181 		if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3182 			return -EFAULT;
3183 		return 0;
3184 
3185 	case TCP_FASTOPEN_KEY: {
3186 		__u8 key[TCP_FASTOPEN_KEY_LENGTH];
3187 		struct tcp_fastopen_context *ctx;
3188 
3189 		if (get_user(len, optlen))
3190 			return -EFAULT;
3191 
3192 		rcu_read_lock();
3193 		ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3194 		if (ctx)
3195 			memcpy(key, ctx->key, sizeof(key));
3196 		else
3197 			len = 0;
3198 		rcu_read_unlock();
3199 
3200 		len = min_t(unsigned int, len, sizeof(key));
3201 		if (put_user(len, optlen))
3202 			return -EFAULT;
3203 		if (copy_to_user(optval, key, len))
3204 			return -EFAULT;
3205 		return 0;
3206 	}
3207 	case TCP_THIN_LINEAR_TIMEOUTS:
3208 		val = tp->thin_lto;
3209 		break;
3210 
3211 	case TCP_THIN_DUPACK:
3212 		val = 0;
3213 		break;
3214 
3215 	case TCP_REPAIR:
3216 		val = tp->repair;
3217 		break;
3218 
3219 	case TCP_REPAIR_QUEUE:
3220 		if (tp->repair)
3221 			val = tp->repair_queue;
3222 		else
3223 			return -EINVAL;
3224 		break;
3225 
3226 	case TCP_REPAIR_WINDOW: {
3227 		struct tcp_repair_window opt;
3228 
3229 		if (get_user(len, optlen))
3230 			return -EFAULT;
3231 
3232 		if (len != sizeof(opt))
3233 			return -EINVAL;
3234 
3235 		if (!tp->repair)
3236 			return -EPERM;
3237 
3238 		opt.snd_wl1	= tp->snd_wl1;
3239 		opt.snd_wnd	= tp->snd_wnd;
3240 		opt.max_window	= tp->max_window;
3241 		opt.rcv_wnd	= tp->rcv_wnd;
3242 		opt.rcv_wup	= tp->rcv_wup;
3243 
3244 		if (copy_to_user(optval, &opt, len))
3245 			return -EFAULT;
3246 		return 0;
3247 	}
3248 	case TCP_QUEUE_SEQ:
3249 		if (tp->repair_queue == TCP_SEND_QUEUE)
3250 			val = tp->write_seq;
3251 		else if (tp->repair_queue == TCP_RECV_QUEUE)
3252 			val = tp->rcv_nxt;
3253 		else
3254 			return -EINVAL;
3255 		break;
3256 
3257 	case TCP_USER_TIMEOUT:
3258 		val = jiffies_to_msecs(icsk->icsk_user_timeout);
3259 		break;
3260 
3261 	case TCP_FASTOPEN:
3262 		val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3263 		break;
3264 
3265 	case TCP_FASTOPEN_CONNECT:
3266 		val = tp->fastopen_connect;
3267 		break;
3268 
3269 	case TCP_FASTOPEN_NO_COOKIE:
3270 		val = tp->fastopen_no_cookie;
3271 		break;
3272 
3273 	case TCP_TIMESTAMP:
3274 		val = tcp_time_stamp_raw() + tp->tsoffset;
3275 		break;
3276 	case TCP_NOTSENT_LOWAT:
3277 		val = tp->notsent_lowat;
3278 		break;
3279 	case TCP_SAVE_SYN:
3280 		val = tp->save_syn;
3281 		break;
3282 	case TCP_SAVED_SYN: {
3283 		if (get_user(len, optlen))
3284 			return -EFAULT;
3285 
3286 		lock_sock(sk);
3287 		if (tp->saved_syn) {
3288 			if (len < tp->saved_syn[0]) {
3289 				if (put_user(tp->saved_syn[0], optlen)) {
3290 					release_sock(sk);
3291 					return -EFAULT;
3292 				}
3293 				release_sock(sk);
3294 				return -EINVAL;
3295 			}
3296 			len = tp->saved_syn[0];
3297 			if (put_user(len, optlen)) {
3298 				release_sock(sk);
3299 				return -EFAULT;
3300 			}
3301 			if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3302 				release_sock(sk);
3303 				return -EFAULT;
3304 			}
3305 			tcp_saved_syn_free(tp);
3306 			release_sock(sk);
3307 		} else {
3308 			release_sock(sk);
3309 			len = 0;
3310 			if (put_user(len, optlen))
3311 				return -EFAULT;
3312 		}
3313 		return 0;
3314 	}
3315 	default:
3316 		return -ENOPROTOOPT;
3317 	}
3318 
3319 	if (put_user(len, optlen))
3320 		return -EFAULT;
3321 	if (copy_to_user(optval, &val, len))
3322 		return -EFAULT;
3323 	return 0;
3324 }
3325 
3326 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3327 		   int __user *optlen)
3328 {
3329 	struct inet_connection_sock *icsk = inet_csk(sk);
3330 
3331 	if (level != SOL_TCP)
3332 		return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3333 						     optval, optlen);
3334 	return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3335 }
3336 EXPORT_SYMBOL(tcp_getsockopt);
3337 
3338 #ifdef CONFIG_COMPAT
3339 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3340 			  char __user *optval, int __user *optlen)
3341 {
3342 	if (level != SOL_TCP)
3343 		return inet_csk_compat_getsockopt(sk, level, optname,
3344 						  optval, optlen);
3345 	return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3346 }
3347 EXPORT_SYMBOL(compat_tcp_getsockopt);
3348 #endif
3349 
3350 #ifdef CONFIG_TCP_MD5SIG
3351 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3352 static DEFINE_MUTEX(tcp_md5sig_mutex);
3353 static bool tcp_md5sig_pool_populated = false;
3354 
3355 static void __tcp_alloc_md5sig_pool(void)
3356 {
3357 	struct crypto_ahash *hash;
3358 	int cpu;
3359 
3360 	hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3361 	if (IS_ERR(hash))
3362 		return;
3363 
3364 	for_each_possible_cpu(cpu) {
3365 		void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3366 		struct ahash_request *req;
3367 
3368 		if (!scratch) {
3369 			scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3370 					       sizeof(struct tcphdr),
3371 					       GFP_KERNEL,
3372 					       cpu_to_node(cpu));
3373 			if (!scratch)
3374 				return;
3375 			per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3376 		}
3377 		if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3378 			continue;
3379 
3380 		req = ahash_request_alloc(hash, GFP_KERNEL);
3381 		if (!req)
3382 			return;
3383 
3384 		ahash_request_set_callback(req, 0, NULL, NULL);
3385 
3386 		per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3387 	}
3388 	/* before setting tcp_md5sig_pool_populated, we must commit all writes
3389 	 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3390 	 */
3391 	smp_wmb();
3392 	tcp_md5sig_pool_populated = true;
3393 }
3394 
3395 bool tcp_alloc_md5sig_pool(void)
3396 {
3397 	if (unlikely(!tcp_md5sig_pool_populated)) {
3398 		mutex_lock(&tcp_md5sig_mutex);
3399 
3400 		if (!tcp_md5sig_pool_populated)
3401 			__tcp_alloc_md5sig_pool();
3402 
3403 		mutex_unlock(&tcp_md5sig_mutex);
3404 	}
3405 	return tcp_md5sig_pool_populated;
3406 }
3407 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3408 
3409 
3410 /**
3411  *	tcp_get_md5sig_pool - get md5sig_pool for this user
3412  *
3413  *	We use percpu structure, so if we succeed, we exit with preemption
3414  *	and BH disabled, to make sure another thread or softirq handling
3415  *	wont try to get same context.
3416  */
3417 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3418 {
3419 	local_bh_disable();
3420 
3421 	if (tcp_md5sig_pool_populated) {
3422 		/* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3423 		smp_rmb();
3424 		return this_cpu_ptr(&tcp_md5sig_pool);
3425 	}
3426 	local_bh_enable();
3427 	return NULL;
3428 }
3429 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3430 
3431 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3432 			  const struct sk_buff *skb, unsigned int header_len)
3433 {
3434 	struct scatterlist sg;
3435 	const struct tcphdr *tp = tcp_hdr(skb);
3436 	struct ahash_request *req = hp->md5_req;
3437 	unsigned int i;
3438 	const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3439 					   skb_headlen(skb) - header_len : 0;
3440 	const struct skb_shared_info *shi = skb_shinfo(skb);
3441 	struct sk_buff *frag_iter;
3442 
3443 	sg_init_table(&sg, 1);
3444 
3445 	sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3446 	ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3447 	if (crypto_ahash_update(req))
3448 		return 1;
3449 
3450 	for (i = 0; i < shi->nr_frags; ++i) {
3451 		const struct skb_frag_struct *f = &shi->frags[i];
3452 		unsigned int offset = f->page_offset;
3453 		struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3454 
3455 		sg_set_page(&sg, page, skb_frag_size(f),
3456 			    offset_in_page(offset));
3457 		ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3458 		if (crypto_ahash_update(req))
3459 			return 1;
3460 	}
3461 
3462 	skb_walk_frags(skb, frag_iter)
3463 		if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3464 			return 1;
3465 
3466 	return 0;
3467 }
3468 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3469 
3470 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3471 {
3472 	struct scatterlist sg;
3473 
3474 	sg_init_one(&sg, key->key, key->keylen);
3475 	ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3476 	return crypto_ahash_update(hp->md5_req);
3477 }
3478 EXPORT_SYMBOL(tcp_md5_hash_key);
3479 
3480 #endif
3481 
3482 void tcp_done(struct sock *sk)
3483 {
3484 	struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3485 
3486 	if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3487 		TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3488 
3489 	tcp_set_state(sk, TCP_CLOSE);
3490 	tcp_clear_xmit_timers(sk);
3491 	if (req)
3492 		reqsk_fastopen_remove(sk, req, false);
3493 
3494 	sk->sk_shutdown = SHUTDOWN_MASK;
3495 
3496 	if (!sock_flag(sk, SOCK_DEAD))
3497 		sk->sk_state_change(sk);
3498 	else
3499 		inet_csk_destroy_sock(sk);
3500 }
3501 EXPORT_SYMBOL_GPL(tcp_done);
3502 
3503 int tcp_abort(struct sock *sk, int err)
3504 {
3505 	if (!sk_fullsock(sk)) {
3506 		if (sk->sk_state == TCP_NEW_SYN_RECV) {
3507 			struct request_sock *req = inet_reqsk(sk);
3508 
3509 			local_bh_disable();
3510 			inet_csk_reqsk_queue_drop_and_put(req->rsk_listener,
3511 							  req);
3512 			local_bh_enable();
3513 			return 0;
3514 		}
3515 		return -EOPNOTSUPP;
3516 	}
3517 
3518 	/* Don't race with userspace socket closes such as tcp_close. */
3519 	lock_sock(sk);
3520 
3521 	if (sk->sk_state == TCP_LISTEN) {
3522 		tcp_set_state(sk, TCP_CLOSE);
3523 		inet_csk_listen_stop(sk);
3524 	}
3525 
3526 	/* Don't race with BH socket closes such as inet_csk_listen_stop. */
3527 	local_bh_disable();
3528 	bh_lock_sock(sk);
3529 
3530 	if (!sock_flag(sk, SOCK_DEAD)) {
3531 		sk->sk_err = err;
3532 		/* This barrier is coupled with smp_rmb() in tcp_poll() */
3533 		smp_wmb();
3534 		sk->sk_error_report(sk);
3535 		if (tcp_need_reset(sk->sk_state))
3536 			tcp_send_active_reset(sk, GFP_ATOMIC);
3537 		tcp_done(sk);
3538 	}
3539 
3540 	bh_unlock_sock(sk);
3541 	local_bh_enable();
3542 	release_sock(sk);
3543 	return 0;
3544 }
3545 EXPORT_SYMBOL_GPL(tcp_abort);
3546 
3547 extern struct tcp_congestion_ops tcp_reno;
3548 
3549 static __initdata unsigned long thash_entries;
3550 static int __init set_thash_entries(char *str)
3551 {
3552 	ssize_t ret;
3553 
3554 	if (!str)
3555 		return 0;
3556 
3557 	ret = kstrtoul(str, 0, &thash_entries);
3558 	if (ret)
3559 		return 0;
3560 
3561 	return 1;
3562 }
3563 __setup("thash_entries=", set_thash_entries);
3564 
3565 static void __init tcp_init_mem(void)
3566 {
3567 	unsigned long limit = nr_free_buffer_pages() / 16;
3568 
3569 	limit = max(limit, 128UL);
3570 	sysctl_tcp_mem[0] = limit / 4 * 3;		/* 4.68 % */
3571 	sysctl_tcp_mem[1] = limit;			/* 6.25 % */
3572 	sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;	/* 9.37 % */
3573 }
3574 
3575 void __init tcp_init(void)
3576 {
3577 	int max_rshare, max_wshare, cnt;
3578 	unsigned long limit;
3579 	unsigned int i;
3580 
3581 	BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3582 		     FIELD_SIZEOF(struct sk_buff, cb));
3583 
3584 	percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3585 	percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3586 	inet_hashinfo_init(&tcp_hashinfo);
3587 	inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3588 			    thash_entries, 21,  /* one slot per 2 MB*/
3589 			    0, 64 * 1024);
3590 	tcp_hashinfo.bind_bucket_cachep =
3591 		kmem_cache_create("tcp_bind_bucket",
3592 				  sizeof(struct inet_bind_bucket), 0,
3593 				  SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3594 
3595 	/* Size and allocate the main established and bind bucket
3596 	 * hash tables.
3597 	 *
3598 	 * The methodology is similar to that of the buffer cache.
3599 	 */
3600 	tcp_hashinfo.ehash =
3601 		alloc_large_system_hash("TCP established",
3602 					sizeof(struct inet_ehash_bucket),
3603 					thash_entries,
3604 					17, /* one slot per 128 KB of memory */
3605 					0,
3606 					NULL,
3607 					&tcp_hashinfo.ehash_mask,
3608 					0,
3609 					thash_entries ? 0 : 512 * 1024);
3610 	for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3611 		INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3612 
3613 	if (inet_ehash_locks_alloc(&tcp_hashinfo))
3614 		panic("TCP: failed to alloc ehash_locks");
3615 	tcp_hashinfo.bhash =
3616 		alloc_large_system_hash("TCP bind",
3617 					sizeof(struct inet_bind_hashbucket),
3618 					tcp_hashinfo.ehash_mask + 1,
3619 					17, /* one slot per 128 KB of memory */
3620 					0,
3621 					&tcp_hashinfo.bhash_size,
3622 					NULL,
3623 					0,
3624 					64 * 1024);
3625 	tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3626 	for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3627 		spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3628 		INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3629 	}
3630 
3631 
3632 	cnt = tcp_hashinfo.ehash_mask + 1;
3633 	sysctl_tcp_max_orphans = cnt / 2;
3634 
3635 	tcp_init_mem();
3636 	/* Set per-socket limits to no more than 1/128 the pressure threshold */
3637 	limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3638 	max_wshare = min(4UL*1024*1024, limit);
3639 	max_rshare = min(6UL*1024*1024, limit);
3640 
3641 	init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3642 	init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
3643 	init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3644 
3645 	init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3646 	init_net.ipv4.sysctl_tcp_rmem[1] = 87380;
3647 	init_net.ipv4.sysctl_tcp_rmem[2] = max(87380, max_rshare);
3648 
3649 	pr_info("Hash tables configured (established %u bind %u)\n",
3650 		tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3651 
3652 	tcp_v4_init();
3653 	tcp_metrics_init();
3654 	BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3655 	tcp_tasklet_init();
3656 }
3657