xref: /openbmc/linux/net/ipv4/tcp.c (revision 6774def6)
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 <linux/kernel.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/init.h>
256 #include <linux/fs.h>
257 #include <linux/skbuff.h>
258 #include <linux/scatterlist.h>
259 #include <linux/splice.h>
260 #include <linux/net.h>
261 #include <linux/socket.h>
262 #include <linux/random.h>
263 #include <linux/bootmem.h>
264 #include <linux/highmem.h>
265 #include <linux/swap.h>
266 #include <linux/cache.h>
267 #include <linux/err.h>
268 #include <linux/crypto.h>
269 #include <linux/time.h>
270 #include <linux/slab.h>
271 
272 #include <net/icmp.h>
273 #include <net/inet_common.h>
274 #include <net/tcp.h>
275 #include <net/xfrm.h>
276 #include <net/ip.h>
277 #include <net/sock.h>
278 
279 #include <asm/uaccess.h>
280 #include <asm/ioctls.h>
281 #include <net/busy_poll.h>
282 
283 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
284 
285 int sysctl_tcp_min_tso_segs __read_mostly = 2;
286 
287 int sysctl_tcp_autocorking __read_mostly = 1;
288 
289 struct percpu_counter tcp_orphan_count;
290 EXPORT_SYMBOL_GPL(tcp_orphan_count);
291 
292 long sysctl_tcp_mem[3] __read_mostly;
293 int sysctl_tcp_wmem[3] __read_mostly;
294 int sysctl_tcp_rmem[3] __read_mostly;
295 
296 EXPORT_SYMBOL(sysctl_tcp_mem);
297 EXPORT_SYMBOL(sysctl_tcp_rmem);
298 EXPORT_SYMBOL(sysctl_tcp_wmem);
299 
300 atomic_long_t tcp_memory_allocated;	/* Current allocated memory. */
301 EXPORT_SYMBOL(tcp_memory_allocated);
302 
303 /*
304  * Current number of TCP sockets.
305  */
306 struct percpu_counter tcp_sockets_allocated;
307 EXPORT_SYMBOL(tcp_sockets_allocated);
308 
309 /*
310  * TCP splice context
311  */
312 struct tcp_splice_state {
313 	struct pipe_inode_info *pipe;
314 	size_t len;
315 	unsigned int flags;
316 };
317 
318 /*
319  * Pressure flag: try to collapse.
320  * Technical note: it is used by multiple contexts non atomically.
321  * All the __sk_mem_schedule() is of this nature: accounting
322  * is strict, actions are advisory and have some latency.
323  */
324 int tcp_memory_pressure __read_mostly;
325 EXPORT_SYMBOL(tcp_memory_pressure);
326 
327 void tcp_enter_memory_pressure(struct sock *sk)
328 {
329 	if (!tcp_memory_pressure) {
330 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
331 		tcp_memory_pressure = 1;
332 	}
333 }
334 EXPORT_SYMBOL(tcp_enter_memory_pressure);
335 
336 /* Convert seconds to retransmits based on initial and max timeout */
337 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
338 {
339 	u8 res = 0;
340 
341 	if (seconds > 0) {
342 		int period = timeout;
343 
344 		res = 1;
345 		while (seconds > period && res < 255) {
346 			res++;
347 			timeout <<= 1;
348 			if (timeout > rto_max)
349 				timeout = rto_max;
350 			period += timeout;
351 		}
352 	}
353 	return res;
354 }
355 
356 /* Convert retransmits to seconds based on initial and max timeout */
357 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
358 {
359 	int period = 0;
360 
361 	if (retrans > 0) {
362 		period = timeout;
363 		while (--retrans) {
364 			timeout <<= 1;
365 			if (timeout > rto_max)
366 				timeout = rto_max;
367 			period += timeout;
368 		}
369 	}
370 	return period;
371 }
372 
373 /* Address-family independent initialization for a tcp_sock.
374  *
375  * NOTE: A lot of things set to zero explicitly by call to
376  *       sk_alloc() so need not be done here.
377  */
378 void tcp_init_sock(struct sock *sk)
379 {
380 	struct inet_connection_sock *icsk = inet_csk(sk);
381 	struct tcp_sock *tp = tcp_sk(sk);
382 
383 	__skb_queue_head_init(&tp->out_of_order_queue);
384 	tcp_init_xmit_timers(sk);
385 	tcp_prequeue_init(tp);
386 	INIT_LIST_HEAD(&tp->tsq_node);
387 
388 	icsk->icsk_rto = TCP_TIMEOUT_INIT;
389 	tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
390 
391 	/* So many TCP implementations out there (incorrectly) count the
392 	 * initial SYN frame in their delayed-ACK and congestion control
393 	 * algorithms that we must have the following bandaid to talk
394 	 * efficiently to them.  -DaveM
395 	 */
396 	tp->snd_cwnd = TCP_INIT_CWND;
397 
398 	/* See draft-stevens-tcpca-spec-01 for discussion of the
399 	 * initialization of these values.
400 	 */
401 	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
402 	tp->snd_cwnd_clamp = ~0;
403 	tp->mss_cache = TCP_MSS_DEFAULT;
404 
405 	tp->reordering = sysctl_tcp_reordering;
406 	tcp_enable_early_retrans(tp);
407 	tcp_assign_congestion_control(sk);
408 
409 	tp->tsoffset = 0;
410 
411 	sk->sk_state = TCP_CLOSE;
412 
413 	sk->sk_write_space = sk_stream_write_space;
414 	sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
415 
416 	icsk->icsk_sync_mss = tcp_sync_mss;
417 
418 	sk->sk_sndbuf = sysctl_tcp_wmem[1];
419 	sk->sk_rcvbuf = sysctl_tcp_rmem[1];
420 
421 	local_bh_disable();
422 	sock_update_memcg(sk);
423 	sk_sockets_allocated_inc(sk);
424 	local_bh_enable();
425 }
426 EXPORT_SYMBOL(tcp_init_sock);
427 
428 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb)
429 {
430 	if (sk->sk_tsflags) {
431 		struct skb_shared_info *shinfo = skb_shinfo(skb);
432 
433 		sock_tx_timestamp(sk, &shinfo->tx_flags);
434 		if (shinfo->tx_flags & SKBTX_ANY_TSTAMP)
435 			shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
436 	}
437 }
438 
439 /*
440  *	Wait for a TCP event.
441  *
442  *	Note that we don't need to lock the socket, as the upper poll layers
443  *	take care of normal races (between the test and the event) and we don't
444  *	go look at any of the socket buffers directly.
445  */
446 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
447 {
448 	unsigned int mask;
449 	struct sock *sk = sock->sk;
450 	const struct tcp_sock *tp = tcp_sk(sk);
451 
452 	sock_rps_record_flow(sk);
453 
454 	sock_poll_wait(file, sk_sleep(sk), wait);
455 	if (sk->sk_state == TCP_LISTEN)
456 		return inet_csk_listen_poll(sk);
457 
458 	/* Socket is not locked. We are protected from async events
459 	 * by poll logic and correct handling of state changes
460 	 * made by other threads is impossible in any case.
461 	 */
462 
463 	mask = 0;
464 
465 	/*
466 	 * POLLHUP is certainly not done right. But poll() doesn't
467 	 * have a notion of HUP in just one direction, and for a
468 	 * socket the read side is more interesting.
469 	 *
470 	 * Some poll() documentation says that POLLHUP is incompatible
471 	 * with the POLLOUT/POLLWR flags, so somebody should check this
472 	 * all. But careful, it tends to be safer to return too many
473 	 * bits than too few, and you can easily break real applications
474 	 * if you don't tell them that something has hung up!
475 	 *
476 	 * Check-me.
477 	 *
478 	 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
479 	 * our fs/select.c). It means that after we received EOF,
480 	 * poll always returns immediately, making impossible poll() on write()
481 	 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
482 	 * if and only if shutdown has been made in both directions.
483 	 * Actually, it is interesting to look how Solaris and DUX
484 	 * solve this dilemma. I would prefer, if POLLHUP were maskable,
485 	 * then we could set it on SND_SHUTDOWN. BTW examples given
486 	 * in Stevens' books assume exactly this behaviour, it explains
487 	 * why POLLHUP is incompatible with POLLOUT.	--ANK
488 	 *
489 	 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
490 	 * blocking on fresh not-connected or disconnected socket. --ANK
491 	 */
492 	if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
493 		mask |= POLLHUP;
494 	if (sk->sk_shutdown & RCV_SHUTDOWN)
495 		mask |= POLLIN | POLLRDNORM | POLLRDHUP;
496 
497 	/* Connected or passive Fast Open socket? */
498 	if (sk->sk_state != TCP_SYN_SENT &&
499 	    (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
500 		int target = sock_rcvlowat(sk, 0, INT_MAX);
501 
502 		if (tp->urg_seq == tp->copied_seq &&
503 		    !sock_flag(sk, SOCK_URGINLINE) &&
504 		    tp->urg_data)
505 			target++;
506 
507 		/* Potential race condition. If read of tp below will
508 		 * escape above sk->sk_state, we can be illegally awaken
509 		 * in SYN_* states. */
510 		if (tp->rcv_nxt - tp->copied_seq >= target)
511 			mask |= POLLIN | POLLRDNORM;
512 
513 		if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
514 			if (sk_stream_is_writeable(sk)) {
515 				mask |= POLLOUT | POLLWRNORM;
516 			} else {  /* send SIGIO later */
517 				set_bit(SOCK_ASYNC_NOSPACE,
518 					&sk->sk_socket->flags);
519 				set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
520 
521 				/* Race breaker. If space is freed after
522 				 * wspace test but before the flags are set,
523 				 * IO signal will be lost.
524 				 */
525 				if (sk_stream_is_writeable(sk))
526 					mask |= POLLOUT | POLLWRNORM;
527 			}
528 		} else
529 			mask |= POLLOUT | POLLWRNORM;
530 
531 		if (tp->urg_data & TCP_URG_VALID)
532 			mask |= POLLPRI;
533 	}
534 	/* This barrier is coupled with smp_wmb() in tcp_reset() */
535 	smp_rmb();
536 	if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
537 		mask |= POLLERR;
538 
539 	return mask;
540 }
541 EXPORT_SYMBOL(tcp_poll);
542 
543 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
544 {
545 	struct tcp_sock *tp = tcp_sk(sk);
546 	int answ;
547 	bool slow;
548 
549 	switch (cmd) {
550 	case SIOCINQ:
551 		if (sk->sk_state == TCP_LISTEN)
552 			return -EINVAL;
553 
554 		slow = lock_sock_fast(sk);
555 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
556 			answ = 0;
557 		else if (sock_flag(sk, SOCK_URGINLINE) ||
558 			 !tp->urg_data ||
559 			 before(tp->urg_seq, tp->copied_seq) ||
560 			 !before(tp->urg_seq, tp->rcv_nxt)) {
561 
562 			answ = tp->rcv_nxt - tp->copied_seq;
563 
564 			/* Subtract 1, if FIN was received */
565 			if (answ && sock_flag(sk, SOCK_DONE))
566 				answ--;
567 		} else
568 			answ = tp->urg_seq - tp->copied_seq;
569 		unlock_sock_fast(sk, slow);
570 		break;
571 	case SIOCATMARK:
572 		answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
573 		break;
574 	case SIOCOUTQ:
575 		if (sk->sk_state == TCP_LISTEN)
576 			return -EINVAL;
577 
578 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
579 			answ = 0;
580 		else
581 			answ = tp->write_seq - tp->snd_una;
582 		break;
583 	case SIOCOUTQNSD:
584 		if (sk->sk_state == TCP_LISTEN)
585 			return -EINVAL;
586 
587 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
588 			answ = 0;
589 		else
590 			answ = tp->write_seq - tp->snd_nxt;
591 		break;
592 	default:
593 		return -ENOIOCTLCMD;
594 	}
595 
596 	return put_user(answ, (int __user *)arg);
597 }
598 EXPORT_SYMBOL(tcp_ioctl);
599 
600 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
601 {
602 	TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
603 	tp->pushed_seq = tp->write_seq;
604 }
605 
606 static inline bool forced_push(const struct tcp_sock *tp)
607 {
608 	return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
609 }
610 
611 static void skb_entail(struct sock *sk, struct sk_buff *skb)
612 {
613 	struct tcp_sock *tp = tcp_sk(sk);
614 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
615 
616 	skb->csum    = 0;
617 	tcb->seq     = tcb->end_seq = tp->write_seq;
618 	tcb->tcp_flags = TCPHDR_ACK;
619 	tcb->sacked  = 0;
620 	__skb_header_release(skb);
621 	tcp_add_write_queue_tail(sk, skb);
622 	sk->sk_wmem_queued += skb->truesize;
623 	sk_mem_charge(sk, skb->truesize);
624 	if (tp->nonagle & TCP_NAGLE_PUSH)
625 		tp->nonagle &= ~TCP_NAGLE_PUSH;
626 }
627 
628 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
629 {
630 	if (flags & MSG_OOB)
631 		tp->snd_up = tp->write_seq;
632 }
633 
634 /* If a not yet filled skb is pushed, do not send it if
635  * we have data packets in Qdisc or NIC queues :
636  * Because TX completion will happen shortly, it gives a chance
637  * to coalesce future sendmsg() payload into this skb, without
638  * need for a timer, and with no latency trade off.
639  * As packets containing data payload have a bigger truesize
640  * than pure acks (dataless) packets, the last checks prevent
641  * autocorking if we only have an ACK in Qdisc/NIC queues,
642  * or if TX completion was delayed after we processed ACK packet.
643  */
644 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
645 				int size_goal)
646 {
647 	return skb->len < size_goal &&
648 	       sysctl_tcp_autocorking &&
649 	       skb != tcp_write_queue_head(sk) &&
650 	       atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
651 }
652 
653 static void tcp_push(struct sock *sk, int flags, int mss_now,
654 		     int nonagle, int size_goal)
655 {
656 	struct tcp_sock *tp = tcp_sk(sk);
657 	struct sk_buff *skb;
658 
659 	if (!tcp_send_head(sk))
660 		return;
661 
662 	skb = tcp_write_queue_tail(sk);
663 	if (!(flags & MSG_MORE) || forced_push(tp))
664 		tcp_mark_push(tp, skb);
665 
666 	tcp_mark_urg(tp, flags);
667 
668 	if (tcp_should_autocork(sk, skb, size_goal)) {
669 
670 		/* avoid atomic op if TSQ_THROTTLED bit is already set */
671 		if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
672 			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
673 			set_bit(TSQ_THROTTLED, &tp->tsq_flags);
674 		}
675 		/* It is possible TX completion already happened
676 		 * before we set TSQ_THROTTLED.
677 		 */
678 		if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
679 			return;
680 	}
681 
682 	if (flags & MSG_MORE)
683 		nonagle = TCP_NAGLE_CORK;
684 
685 	__tcp_push_pending_frames(sk, mss_now, nonagle);
686 }
687 
688 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
689 				unsigned int offset, size_t len)
690 {
691 	struct tcp_splice_state *tss = rd_desc->arg.data;
692 	int ret;
693 
694 	ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
695 			      tss->flags);
696 	if (ret > 0)
697 		rd_desc->count -= ret;
698 	return ret;
699 }
700 
701 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
702 {
703 	/* Store TCP splice context information in read_descriptor_t. */
704 	read_descriptor_t rd_desc = {
705 		.arg.data = tss,
706 		.count	  = tss->len,
707 	};
708 
709 	return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
710 }
711 
712 /**
713  *  tcp_splice_read - splice data from TCP socket to a pipe
714  * @sock:	socket to splice from
715  * @ppos:	position (not valid)
716  * @pipe:	pipe to splice to
717  * @len:	number of bytes to splice
718  * @flags:	splice modifier flags
719  *
720  * Description:
721  *    Will read pages from given socket and fill them into a pipe.
722  *
723  **/
724 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
725 			struct pipe_inode_info *pipe, size_t len,
726 			unsigned int flags)
727 {
728 	struct sock *sk = sock->sk;
729 	struct tcp_splice_state tss = {
730 		.pipe = pipe,
731 		.len = len,
732 		.flags = flags,
733 	};
734 	long timeo;
735 	ssize_t spliced;
736 	int ret;
737 
738 	sock_rps_record_flow(sk);
739 	/*
740 	 * We can't seek on a socket input
741 	 */
742 	if (unlikely(*ppos))
743 		return -ESPIPE;
744 
745 	ret = spliced = 0;
746 
747 	lock_sock(sk);
748 
749 	timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
750 	while (tss.len) {
751 		ret = __tcp_splice_read(sk, &tss);
752 		if (ret < 0)
753 			break;
754 		else if (!ret) {
755 			if (spliced)
756 				break;
757 			if (sock_flag(sk, SOCK_DONE))
758 				break;
759 			if (sk->sk_err) {
760 				ret = sock_error(sk);
761 				break;
762 			}
763 			if (sk->sk_shutdown & RCV_SHUTDOWN)
764 				break;
765 			if (sk->sk_state == TCP_CLOSE) {
766 				/*
767 				 * This occurs when user tries to read
768 				 * from never connected socket.
769 				 */
770 				if (!sock_flag(sk, SOCK_DONE))
771 					ret = -ENOTCONN;
772 				break;
773 			}
774 			if (!timeo) {
775 				ret = -EAGAIN;
776 				break;
777 			}
778 			sk_wait_data(sk, &timeo);
779 			if (signal_pending(current)) {
780 				ret = sock_intr_errno(timeo);
781 				break;
782 			}
783 			continue;
784 		}
785 		tss.len -= ret;
786 		spliced += ret;
787 
788 		if (!timeo)
789 			break;
790 		release_sock(sk);
791 		lock_sock(sk);
792 
793 		if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
794 		    (sk->sk_shutdown & RCV_SHUTDOWN) ||
795 		    signal_pending(current))
796 			break;
797 	}
798 
799 	release_sock(sk);
800 
801 	if (spliced)
802 		return spliced;
803 
804 	return ret;
805 }
806 EXPORT_SYMBOL(tcp_splice_read);
807 
808 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
809 {
810 	struct sk_buff *skb;
811 
812 	/* The TCP header must be at least 32-bit aligned.  */
813 	size = ALIGN(size, 4);
814 
815 	skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
816 	if (skb) {
817 		if (sk_wmem_schedule(sk, skb->truesize)) {
818 			skb_reserve(skb, sk->sk_prot->max_header);
819 			/*
820 			 * Make sure that we have exactly size bytes
821 			 * available to the caller, no more, no less.
822 			 */
823 			skb->reserved_tailroom = skb->end - skb->tail - size;
824 			return skb;
825 		}
826 		__kfree_skb(skb);
827 	} else {
828 		sk->sk_prot->enter_memory_pressure(sk);
829 		sk_stream_moderate_sndbuf(sk);
830 	}
831 	return NULL;
832 }
833 
834 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
835 				       int large_allowed)
836 {
837 	struct tcp_sock *tp = tcp_sk(sk);
838 	u32 xmit_size_goal, old_size_goal;
839 
840 	xmit_size_goal = mss_now;
841 
842 	if (large_allowed && sk_can_gso(sk)) {
843 		u32 gso_size, hlen;
844 
845 		/* Maybe we should/could use sk->sk_prot->max_header here ? */
846 		hlen = inet_csk(sk)->icsk_af_ops->net_header_len +
847 		       inet_csk(sk)->icsk_ext_hdr_len +
848 		       tp->tcp_header_len;
849 
850 		/* Goal is to send at least one packet per ms,
851 		 * not one big TSO packet every 100 ms.
852 		 * This preserves ACK clocking and is consistent
853 		 * with tcp_tso_should_defer() heuristic.
854 		 */
855 		gso_size = sk->sk_pacing_rate / (2 * MSEC_PER_SEC);
856 		gso_size = max_t(u32, gso_size,
857 				 sysctl_tcp_min_tso_segs * mss_now);
858 
859 		xmit_size_goal = min_t(u32, gso_size,
860 				       sk->sk_gso_max_size - 1 - hlen);
861 
862 		xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
863 
864 		/* We try hard to avoid divides here */
865 		old_size_goal = tp->xmit_size_goal_segs * mss_now;
866 
867 		if (likely(old_size_goal <= xmit_size_goal &&
868 			   old_size_goal + mss_now > xmit_size_goal)) {
869 			xmit_size_goal = old_size_goal;
870 		} else {
871 			tp->xmit_size_goal_segs =
872 				min_t(u16, xmit_size_goal / mss_now,
873 				      sk->sk_gso_max_segs);
874 			xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
875 		}
876 	}
877 
878 	return max(xmit_size_goal, mss_now);
879 }
880 
881 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
882 {
883 	int mss_now;
884 
885 	mss_now = tcp_current_mss(sk);
886 	*size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
887 
888 	return mss_now;
889 }
890 
891 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
892 				size_t size, int flags)
893 {
894 	struct tcp_sock *tp = tcp_sk(sk);
895 	int mss_now, size_goal;
896 	int err;
897 	ssize_t copied;
898 	long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
899 
900 	/* Wait for a connection to finish. One exception is TCP Fast Open
901 	 * (passive side) where data is allowed to be sent before a connection
902 	 * is fully established.
903 	 */
904 	if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
905 	    !tcp_passive_fastopen(sk)) {
906 		if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
907 			goto out_err;
908 	}
909 
910 	clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
911 
912 	mss_now = tcp_send_mss(sk, &size_goal, flags);
913 	copied = 0;
914 
915 	err = -EPIPE;
916 	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
917 		goto out_err;
918 
919 	while (size > 0) {
920 		struct sk_buff *skb = tcp_write_queue_tail(sk);
921 		int copy, i;
922 		bool can_coalesce;
923 
924 		if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
925 new_segment:
926 			if (!sk_stream_memory_free(sk))
927 				goto wait_for_sndbuf;
928 
929 			skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
930 			if (!skb)
931 				goto wait_for_memory;
932 
933 			skb_entail(sk, skb);
934 			copy = size_goal;
935 		}
936 
937 		if (copy > size)
938 			copy = size;
939 
940 		i = skb_shinfo(skb)->nr_frags;
941 		can_coalesce = skb_can_coalesce(skb, i, page, offset);
942 		if (!can_coalesce && i >= MAX_SKB_FRAGS) {
943 			tcp_mark_push(tp, skb);
944 			goto new_segment;
945 		}
946 		if (!sk_wmem_schedule(sk, copy))
947 			goto wait_for_memory;
948 
949 		if (can_coalesce) {
950 			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
951 		} else {
952 			get_page(page);
953 			skb_fill_page_desc(skb, i, page, offset, copy);
954 		}
955 		skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
956 
957 		skb->len += copy;
958 		skb->data_len += copy;
959 		skb->truesize += copy;
960 		sk->sk_wmem_queued += copy;
961 		sk_mem_charge(sk, copy);
962 		skb->ip_summed = CHECKSUM_PARTIAL;
963 		tp->write_seq += copy;
964 		TCP_SKB_CB(skb)->end_seq += copy;
965 		tcp_skb_pcount_set(skb, 0);
966 
967 		if (!copied)
968 			TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
969 
970 		copied += copy;
971 		offset += copy;
972 		if (!(size -= copy)) {
973 			tcp_tx_timestamp(sk, skb);
974 			goto out;
975 		}
976 
977 		if (skb->len < size_goal || (flags & MSG_OOB))
978 			continue;
979 
980 		if (forced_push(tp)) {
981 			tcp_mark_push(tp, skb);
982 			__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
983 		} else if (skb == tcp_send_head(sk))
984 			tcp_push_one(sk, mss_now);
985 		continue;
986 
987 wait_for_sndbuf:
988 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
989 wait_for_memory:
990 		tcp_push(sk, flags & ~MSG_MORE, mss_now,
991 			 TCP_NAGLE_PUSH, size_goal);
992 
993 		if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
994 			goto do_error;
995 
996 		mss_now = tcp_send_mss(sk, &size_goal, flags);
997 	}
998 
999 out:
1000 	if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
1001 		tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1002 	return copied;
1003 
1004 do_error:
1005 	if (copied)
1006 		goto out;
1007 out_err:
1008 	return sk_stream_error(sk, flags, err);
1009 }
1010 
1011 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1012 		 size_t size, int flags)
1013 {
1014 	ssize_t res;
1015 
1016 	if (!(sk->sk_route_caps & NETIF_F_SG) ||
1017 	    !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1018 		return sock_no_sendpage(sk->sk_socket, page, offset, size,
1019 					flags);
1020 
1021 	lock_sock(sk);
1022 	res = do_tcp_sendpages(sk, page, offset, size, flags);
1023 	release_sock(sk);
1024 	return res;
1025 }
1026 EXPORT_SYMBOL(tcp_sendpage);
1027 
1028 static inline int select_size(const struct sock *sk, bool sg)
1029 {
1030 	const struct tcp_sock *tp = tcp_sk(sk);
1031 	int tmp = tp->mss_cache;
1032 
1033 	if (sg) {
1034 		if (sk_can_gso(sk)) {
1035 			/* Small frames wont use a full page:
1036 			 * Payload will immediately follow tcp header.
1037 			 */
1038 			tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1039 		} else {
1040 			int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1041 
1042 			if (tmp >= pgbreak &&
1043 			    tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1044 				tmp = pgbreak;
1045 		}
1046 	}
1047 
1048 	return tmp;
1049 }
1050 
1051 void tcp_free_fastopen_req(struct tcp_sock *tp)
1052 {
1053 	if (tp->fastopen_req != NULL) {
1054 		kfree(tp->fastopen_req);
1055 		tp->fastopen_req = NULL;
1056 	}
1057 }
1058 
1059 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1060 				int *copied, size_t size)
1061 {
1062 	struct tcp_sock *tp = tcp_sk(sk);
1063 	int err, flags;
1064 
1065 	if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1066 		return -EOPNOTSUPP;
1067 	if (tp->fastopen_req != NULL)
1068 		return -EALREADY; /* Another Fast Open is in progress */
1069 
1070 	tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1071 				   sk->sk_allocation);
1072 	if (unlikely(tp->fastopen_req == NULL))
1073 		return -ENOBUFS;
1074 	tp->fastopen_req->data = msg;
1075 	tp->fastopen_req->size = size;
1076 
1077 	flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1078 	err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1079 				    msg->msg_namelen, flags);
1080 	*copied = tp->fastopen_req->copied;
1081 	tcp_free_fastopen_req(tp);
1082 	return err;
1083 }
1084 
1085 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1086 		size_t size)
1087 {
1088 	struct iovec *iov;
1089 	struct tcp_sock *tp = tcp_sk(sk);
1090 	struct sk_buff *skb;
1091 	int iovlen, flags, err, copied = 0;
1092 	int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1093 	bool sg;
1094 	long timeo;
1095 
1096 	lock_sock(sk);
1097 
1098 	flags = msg->msg_flags;
1099 	if (flags & MSG_FASTOPEN) {
1100 		err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1101 		if (err == -EINPROGRESS && copied_syn > 0)
1102 			goto out;
1103 		else if (err)
1104 			goto out_err;
1105 		offset = copied_syn;
1106 	}
1107 
1108 	timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1109 
1110 	/* Wait for a connection to finish. One exception is TCP Fast Open
1111 	 * (passive side) where data is allowed to be sent before a connection
1112 	 * is fully established.
1113 	 */
1114 	if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1115 	    !tcp_passive_fastopen(sk)) {
1116 		if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1117 			goto do_error;
1118 	}
1119 
1120 	if (unlikely(tp->repair)) {
1121 		if (tp->repair_queue == TCP_RECV_QUEUE) {
1122 			copied = tcp_send_rcvq(sk, msg, size);
1123 			goto out_nopush;
1124 		}
1125 
1126 		err = -EINVAL;
1127 		if (tp->repair_queue == TCP_NO_QUEUE)
1128 			goto out_err;
1129 
1130 		/* 'common' sending to sendq */
1131 	}
1132 
1133 	/* This should be in poll */
1134 	clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1135 
1136 	mss_now = tcp_send_mss(sk, &size_goal, flags);
1137 
1138 	/* Ok commence sending. */
1139 	iovlen = msg->msg_iovlen;
1140 	iov = msg->msg_iov;
1141 	copied = 0;
1142 
1143 	err = -EPIPE;
1144 	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1145 		goto out_err;
1146 
1147 	sg = !!(sk->sk_route_caps & NETIF_F_SG);
1148 
1149 	while (--iovlen >= 0) {
1150 		size_t seglen = iov->iov_len;
1151 		unsigned char __user *from = iov->iov_base;
1152 
1153 		iov++;
1154 		if (unlikely(offset > 0)) {  /* Skip bytes copied in SYN */
1155 			if (offset >= seglen) {
1156 				offset -= seglen;
1157 				continue;
1158 			}
1159 			seglen -= offset;
1160 			from += offset;
1161 			offset = 0;
1162 		}
1163 
1164 		while (seglen > 0) {
1165 			int copy = 0;
1166 			int max = size_goal;
1167 
1168 			skb = tcp_write_queue_tail(sk);
1169 			if (tcp_send_head(sk)) {
1170 				if (skb->ip_summed == CHECKSUM_NONE)
1171 					max = mss_now;
1172 				copy = max - skb->len;
1173 			}
1174 
1175 			if (copy <= 0) {
1176 new_segment:
1177 				/* Allocate new segment. If the interface is SG,
1178 				 * allocate skb fitting to single page.
1179 				 */
1180 				if (!sk_stream_memory_free(sk))
1181 					goto wait_for_sndbuf;
1182 
1183 				skb = sk_stream_alloc_skb(sk,
1184 							  select_size(sk, sg),
1185 							  sk->sk_allocation);
1186 				if (!skb)
1187 					goto wait_for_memory;
1188 
1189 				/*
1190 				 * Check whether we can use HW checksum.
1191 				 */
1192 				if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1193 					skb->ip_summed = CHECKSUM_PARTIAL;
1194 
1195 				skb_entail(sk, skb);
1196 				copy = size_goal;
1197 				max = size_goal;
1198 
1199 				/* All packets are restored as if they have
1200 				 * already been sent. skb_mstamp isn't set to
1201 				 * avoid wrong rtt estimation.
1202 				 */
1203 				if (tp->repair)
1204 					TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1205 			}
1206 
1207 			/* Try to append data to the end of skb. */
1208 			if (copy > seglen)
1209 				copy = seglen;
1210 
1211 			/* Where to copy to? */
1212 			if (skb_availroom(skb) > 0) {
1213 				/* We have some space in skb head. Superb! */
1214 				copy = min_t(int, copy, skb_availroom(skb));
1215 				err = skb_add_data_nocache(sk, skb, from, copy);
1216 				if (err)
1217 					goto do_fault;
1218 			} else {
1219 				bool merge = true;
1220 				int i = skb_shinfo(skb)->nr_frags;
1221 				struct page_frag *pfrag = sk_page_frag(sk);
1222 
1223 				if (!sk_page_frag_refill(sk, pfrag))
1224 					goto wait_for_memory;
1225 
1226 				if (!skb_can_coalesce(skb, i, pfrag->page,
1227 						      pfrag->offset)) {
1228 					if (i == MAX_SKB_FRAGS || !sg) {
1229 						tcp_mark_push(tp, skb);
1230 						goto new_segment;
1231 					}
1232 					merge = false;
1233 				}
1234 
1235 				copy = min_t(int, copy, pfrag->size - pfrag->offset);
1236 
1237 				if (!sk_wmem_schedule(sk, copy))
1238 					goto wait_for_memory;
1239 
1240 				err = skb_copy_to_page_nocache(sk, from, skb,
1241 							       pfrag->page,
1242 							       pfrag->offset,
1243 							       copy);
1244 				if (err)
1245 					goto do_error;
1246 
1247 				/* Update the skb. */
1248 				if (merge) {
1249 					skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1250 				} else {
1251 					skb_fill_page_desc(skb, i, pfrag->page,
1252 							   pfrag->offset, copy);
1253 					get_page(pfrag->page);
1254 				}
1255 				pfrag->offset += copy;
1256 			}
1257 
1258 			if (!copied)
1259 				TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1260 
1261 			tp->write_seq += copy;
1262 			TCP_SKB_CB(skb)->end_seq += copy;
1263 			tcp_skb_pcount_set(skb, 0);
1264 
1265 			from += copy;
1266 			copied += copy;
1267 			if ((seglen -= copy) == 0 && iovlen == 0) {
1268 				tcp_tx_timestamp(sk, skb);
1269 				goto out;
1270 			}
1271 
1272 			if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1273 				continue;
1274 
1275 			if (forced_push(tp)) {
1276 				tcp_mark_push(tp, skb);
1277 				__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1278 			} else if (skb == tcp_send_head(sk))
1279 				tcp_push_one(sk, mss_now);
1280 			continue;
1281 
1282 wait_for_sndbuf:
1283 			set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1284 wait_for_memory:
1285 			if (copied)
1286 				tcp_push(sk, flags & ~MSG_MORE, mss_now,
1287 					 TCP_NAGLE_PUSH, size_goal);
1288 
1289 			if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1290 				goto do_error;
1291 
1292 			mss_now = tcp_send_mss(sk, &size_goal, flags);
1293 		}
1294 	}
1295 
1296 out:
1297 	if (copied)
1298 		tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1299 out_nopush:
1300 	release_sock(sk);
1301 	return copied + copied_syn;
1302 
1303 do_fault:
1304 	if (!skb->len) {
1305 		tcp_unlink_write_queue(skb, sk);
1306 		/* It is the one place in all of TCP, except connection
1307 		 * reset, where we can be unlinking the send_head.
1308 		 */
1309 		tcp_check_send_head(sk, skb);
1310 		sk_wmem_free_skb(sk, skb);
1311 	}
1312 
1313 do_error:
1314 	if (copied + copied_syn)
1315 		goto out;
1316 out_err:
1317 	err = sk_stream_error(sk, flags, err);
1318 	release_sock(sk);
1319 	return err;
1320 }
1321 EXPORT_SYMBOL(tcp_sendmsg);
1322 
1323 /*
1324  *	Handle reading urgent data. BSD has very simple semantics for
1325  *	this, no blocking and very strange errors 8)
1326  */
1327 
1328 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1329 {
1330 	struct tcp_sock *tp = tcp_sk(sk);
1331 
1332 	/* No URG data to read. */
1333 	if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1334 	    tp->urg_data == TCP_URG_READ)
1335 		return -EINVAL;	/* Yes this is right ! */
1336 
1337 	if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1338 		return -ENOTCONN;
1339 
1340 	if (tp->urg_data & TCP_URG_VALID) {
1341 		int err = 0;
1342 		char c = tp->urg_data;
1343 
1344 		if (!(flags & MSG_PEEK))
1345 			tp->urg_data = TCP_URG_READ;
1346 
1347 		/* Read urgent data. */
1348 		msg->msg_flags |= MSG_OOB;
1349 
1350 		if (len > 0) {
1351 			if (!(flags & MSG_TRUNC))
1352 				err = memcpy_toiovec(msg->msg_iov, &c, 1);
1353 			len = 1;
1354 		} else
1355 			msg->msg_flags |= MSG_TRUNC;
1356 
1357 		return err ? -EFAULT : len;
1358 	}
1359 
1360 	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1361 		return 0;
1362 
1363 	/* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1364 	 * the available implementations agree in this case:
1365 	 * this call should never block, independent of the
1366 	 * blocking state of the socket.
1367 	 * Mike <pall@rz.uni-karlsruhe.de>
1368 	 */
1369 	return -EAGAIN;
1370 }
1371 
1372 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1373 {
1374 	struct sk_buff *skb;
1375 	int copied = 0, err = 0;
1376 
1377 	/* XXX -- need to support SO_PEEK_OFF */
1378 
1379 	skb_queue_walk(&sk->sk_write_queue, skb) {
1380 		err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1381 		if (err)
1382 			break;
1383 
1384 		copied += skb->len;
1385 	}
1386 
1387 	return err ?: copied;
1388 }
1389 
1390 /* Clean up the receive buffer for full frames taken by the user,
1391  * then send an ACK if necessary.  COPIED is the number of bytes
1392  * tcp_recvmsg has given to the user so far, it speeds up the
1393  * calculation of whether or not we must ACK for the sake of
1394  * a window update.
1395  */
1396 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1397 {
1398 	struct tcp_sock *tp = tcp_sk(sk);
1399 	bool time_to_ack = false;
1400 
1401 	struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1402 
1403 	WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1404 	     "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1405 	     tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1406 
1407 	if (inet_csk_ack_scheduled(sk)) {
1408 		const struct inet_connection_sock *icsk = inet_csk(sk);
1409 		   /* Delayed ACKs frequently hit locked sockets during bulk
1410 		    * receive. */
1411 		if (icsk->icsk_ack.blocked ||
1412 		    /* Once-per-two-segments ACK was not sent by tcp_input.c */
1413 		    tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1414 		    /*
1415 		     * If this read emptied read buffer, we send ACK, if
1416 		     * connection is not bidirectional, user drained
1417 		     * receive buffer and there was a small segment
1418 		     * in queue.
1419 		     */
1420 		    (copied > 0 &&
1421 		     ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1422 		      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1423 		       !icsk->icsk_ack.pingpong)) &&
1424 		      !atomic_read(&sk->sk_rmem_alloc)))
1425 			time_to_ack = true;
1426 	}
1427 
1428 	/* We send an ACK if we can now advertise a non-zero window
1429 	 * which has been raised "significantly".
1430 	 *
1431 	 * Even if window raised up to infinity, do not send window open ACK
1432 	 * in states, where we will not receive more. It is useless.
1433 	 */
1434 	if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1435 		__u32 rcv_window_now = tcp_receive_window(tp);
1436 
1437 		/* Optimize, __tcp_select_window() is not cheap. */
1438 		if (2*rcv_window_now <= tp->window_clamp) {
1439 			__u32 new_window = __tcp_select_window(sk);
1440 
1441 			/* Send ACK now, if this read freed lots of space
1442 			 * in our buffer. Certainly, new_window is new window.
1443 			 * We can advertise it now, if it is not less than current one.
1444 			 * "Lots" means "at least twice" here.
1445 			 */
1446 			if (new_window && new_window >= 2 * rcv_window_now)
1447 				time_to_ack = true;
1448 		}
1449 	}
1450 	if (time_to_ack)
1451 		tcp_send_ack(sk);
1452 }
1453 
1454 static void tcp_prequeue_process(struct sock *sk)
1455 {
1456 	struct sk_buff *skb;
1457 	struct tcp_sock *tp = tcp_sk(sk);
1458 
1459 	NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1460 
1461 	/* RX process wants to run with disabled BHs, though it is not
1462 	 * necessary */
1463 	local_bh_disable();
1464 	while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1465 		sk_backlog_rcv(sk, skb);
1466 	local_bh_enable();
1467 
1468 	/* Clear memory counter. */
1469 	tp->ucopy.memory = 0;
1470 }
1471 
1472 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1473 {
1474 	struct sk_buff *skb;
1475 	u32 offset;
1476 
1477 	while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1478 		offset = seq - TCP_SKB_CB(skb)->seq;
1479 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1480 			offset--;
1481 		if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1482 			*off = offset;
1483 			return skb;
1484 		}
1485 		/* This looks weird, but this can happen if TCP collapsing
1486 		 * splitted a fat GRO packet, while we released socket lock
1487 		 * in skb_splice_bits()
1488 		 */
1489 		sk_eat_skb(sk, skb);
1490 	}
1491 	return NULL;
1492 }
1493 
1494 /*
1495  * This routine provides an alternative to tcp_recvmsg() for routines
1496  * that would like to handle copying from skbuffs directly in 'sendfile'
1497  * fashion.
1498  * Note:
1499  *	- It is assumed that the socket was locked by the caller.
1500  *	- The routine does not block.
1501  *	- At present, there is no support for reading OOB data
1502  *	  or for 'peeking' the socket using this routine
1503  *	  (although both would be easy to implement).
1504  */
1505 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1506 		  sk_read_actor_t recv_actor)
1507 {
1508 	struct sk_buff *skb;
1509 	struct tcp_sock *tp = tcp_sk(sk);
1510 	u32 seq = tp->copied_seq;
1511 	u32 offset;
1512 	int copied = 0;
1513 
1514 	if (sk->sk_state == TCP_LISTEN)
1515 		return -ENOTCONN;
1516 	while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1517 		if (offset < skb->len) {
1518 			int used;
1519 			size_t len;
1520 
1521 			len = skb->len - offset;
1522 			/* Stop reading if we hit a patch of urgent data */
1523 			if (tp->urg_data) {
1524 				u32 urg_offset = tp->urg_seq - seq;
1525 				if (urg_offset < len)
1526 					len = urg_offset;
1527 				if (!len)
1528 					break;
1529 			}
1530 			used = recv_actor(desc, skb, offset, len);
1531 			if (used <= 0) {
1532 				if (!copied)
1533 					copied = used;
1534 				break;
1535 			} else if (used <= len) {
1536 				seq += used;
1537 				copied += used;
1538 				offset += used;
1539 			}
1540 			/* If recv_actor drops the lock (e.g. TCP splice
1541 			 * receive) the skb pointer might be invalid when
1542 			 * getting here: tcp_collapse might have deleted it
1543 			 * while aggregating skbs from the socket queue.
1544 			 */
1545 			skb = tcp_recv_skb(sk, seq - 1, &offset);
1546 			if (!skb)
1547 				break;
1548 			/* TCP coalescing might have appended data to the skb.
1549 			 * Try to splice more frags
1550 			 */
1551 			if (offset + 1 != skb->len)
1552 				continue;
1553 		}
1554 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1555 			sk_eat_skb(sk, skb);
1556 			++seq;
1557 			break;
1558 		}
1559 		sk_eat_skb(sk, skb);
1560 		if (!desc->count)
1561 			break;
1562 		tp->copied_seq = seq;
1563 	}
1564 	tp->copied_seq = seq;
1565 
1566 	tcp_rcv_space_adjust(sk);
1567 
1568 	/* Clean up data we have read: This will do ACK frames. */
1569 	if (copied > 0) {
1570 		tcp_recv_skb(sk, seq, &offset);
1571 		tcp_cleanup_rbuf(sk, copied);
1572 	}
1573 	return copied;
1574 }
1575 EXPORT_SYMBOL(tcp_read_sock);
1576 
1577 /*
1578  *	This routine copies from a sock struct into the user buffer.
1579  *
1580  *	Technical note: in 2.3 we work on _locked_ socket, so that
1581  *	tricks with *seq access order and skb->users are not required.
1582  *	Probably, code can be easily improved even more.
1583  */
1584 
1585 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1586 		size_t len, int nonblock, int flags, int *addr_len)
1587 {
1588 	struct tcp_sock *tp = tcp_sk(sk);
1589 	int copied = 0;
1590 	u32 peek_seq;
1591 	u32 *seq;
1592 	unsigned long used;
1593 	int err;
1594 	int target;		/* Read at least this many bytes */
1595 	long timeo;
1596 	struct task_struct *user_recv = NULL;
1597 	struct sk_buff *skb;
1598 	u32 urg_hole = 0;
1599 
1600 	if (unlikely(flags & MSG_ERRQUEUE))
1601 		return ip_recv_error(sk, msg, len, addr_len);
1602 
1603 	if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1604 	    (sk->sk_state == TCP_ESTABLISHED))
1605 		sk_busy_loop(sk, nonblock);
1606 
1607 	lock_sock(sk);
1608 
1609 	err = -ENOTCONN;
1610 	if (sk->sk_state == TCP_LISTEN)
1611 		goto out;
1612 
1613 	timeo = sock_rcvtimeo(sk, nonblock);
1614 
1615 	/* Urgent data needs to be handled specially. */
1616 	if (flags & MSG_OOB)
1617 		goto recv_urg;
1618 
1619 	if (unlikely(tp->repair)) {
1620 		err = -EPERM;
1621 		if (!(flags & MSG_PEEK))
1622 			goto out;
1623 
1624 		if (tp->repair_queue == TCP_SEND_QUEUE)
1625 			goto recv_sndq;
1626 
1627 		err = -EINVAL;
1628 		if (tp->repair_queue == TCP_NO_QUEUE)
1629 			goto out;
1630 
1631 		/* 'common' recv queue MSG_PEEK-ing */
1632 	}
1633 
1634 	seq = &tp->copied_seq;
1635 	if (flags & MSG_PEEK) {
1636 		peek_seq = tp->copied_seq;
1637 		seq = &peek_seq;
1638 	}
1639 
1640 	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1641 
1642 	do {
1643 		u32 offset;
1644 
1645 		/* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1646 		if (tp->urg_data && tp->urg_seq == *seq) {
1647 			if (copied)
1648 				break;
1649 			if (signal_pending(current)) {
1650 				copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1651 				break;
1652 			}
1653 		}
1654 
1655 		/* Next get a buffer. */
1656 
1657 		skb_queue_walk(&sk->sk_receive_queue, skb) {
1658 			/* Now that we have two receive queues this
1659 			 * shouldn't happen.
1660 			 */
1661 			if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1662 				 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1663 				 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1664 				 flags))
1665 				break;
1666 
1667 			offset = *seq - TCP_SKB_CB(skb)->seq;
1668 			if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1669 				offset--;
1670 			if (offset < skb->len)
1671 				goto found_ok_skb;
1672 			if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1673 				goto found_fin_ok;
1674 			WARN(!(flags & MSG_PEEK),
1675 			     "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1676 			     *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1677 		}
1678 
1679 		/* Well, if we have backlog, try to process it now yet. */
1680 
1681 		if (copied >= target && !sk->sk_backlog.tail)
1682 			break;
1683 
1684 		if (copied) {
1685 			if (sk->sk_err ||
1686 			    sk->sk_state == TCP_CLOSE ||
1687 			    (sk->sk_shutdown & RCV_SHUTDOWN) ||
1688 			    !timeo ||
1689 			    signal_pending(current))
1690 				break;
1691 		} else {
1692 			if (sock_flag(sk, SOCK_DONE))
1693 				break;
1694 
1695 			if (sk->sk_err) {
1696 				copied = sock_error(sk);
1697 				break;
1698 			}
1699 
1700 			if (sk->sk_shutdown & RCV_SHUTDOWN)
1701 				break;
1702 
1703 			if (sk->sk_state == TCP_CLOSE) {
1704 				if (!sock_flag(sk, SOCK_DONE)) {
1705 					/* This occurs when user tries to read
1706 					 * from never connected socket.
1707 					 */
1708 					copied = -ENOTCONN;
1709 					break;
1710 				}
1711 				break;
1712 			}
1713 
1714 			if (!timeo) {
1715 				copied = -EAGAIN;
1716 				break;
1717 			}
1718 
1719 			if (signal_pending(current)) {
1720 				copied = sock_intr_errno(timeo);
1721 				break;
1722 			}
1723 		}
1724 
1725 		tcp_cleanup_rbuf(sk, copied);
1726 
1727 		if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1728 			/* Install new reader */
1729 			if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1730 				user_recv = current;
1731 				tp->ucopy.task = user_recv;
1732 				tp->ucopy.iov = msg->msg_iov;
1733 			}
1734 
1735 			tp->ucopy.len = len;
1736 
1737 			WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1738 				!(flags & (MSG_PEEK | MSG_TRUNC)));
1739 
1740 			/* Ugly... If prequeue is not empty, we have to
1741 			 * process it before releasing socket, otherwise
1742 			 * order will be broken at second iteration.
1743 			 * More elegant solution is required!!!
1744 			 *
1745 			 * Look: we have the following (pseudo)queues:
1746 			 *
1747 			 * 1. packets in flight
1748 			 * 2. backlog
1749 			 * 3. prequeue
1750 			 * 4. receive_queue
1751 			 *
1752 			 * Each queue can be processed only if the next ones
1753 			 * are empty. At this point we have empty receive_queue.
1754 			 * But prequeue _can_ be not empty after 2nd iteration,
1755 			 * when we jumped to start of loop because backlog
1756 			 * processing added something to receive_queue.
1757 			 * We cannot release_sock(), because backlog contains
1758 			 * packets arrived _after_ prequeued ones.
1759 			 *
1760 			 * Shortly, algorithm is clear --- to process all
1761 			 * the queues in order. We could make it more directly,
1762 			 * requeueing packets from backlog to prequeue, if
1763 			 * is not empty. It is more elegant, but eats cycles,
1764 			 * unfortunately.
1765 			 */
1766 			if (!skb_queue_empty(&tp->ucopy.prequeue))
1767 				goto do_prequeue;
1768 
1769 			/* __ Set realtime policy in scheduler __ */
1770 		}
1771 
1772 		if (copied >= target) {
1773 			/* Do not sleep, just process backlog. */
1774 			release_sock(sk);
1775 			lock_sock(sk);
1776 		} else
1777 			sk_wait_data(sk, &timeo);
1778 
1779 		if (user_recv) {
1780 			int chunk;
1781 
1782 			/* __ Restore normal policy in scheduler __ */
1783 
1784 			if ((chunk = len - tp->ucopy.len) != 0) {
1785 				NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1786 				len -= chunk;
1787 				copied += chunk;
1788 			}
1789 
1790 			if (tp->rcv_nxt == tp->copied_seq &&
1791 			    !skb_queue_empty(&tp->ucopy.prequeue)) {
1792 do_prequeue:
1793 				tcp_prequeue_process(sk);
1794 
1795 				if ((chunk = len - tp->ucopy.len) != 0) {
1796 					NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1797 					len -= chunk;
1798 					copied += chunk;
1799 				}
1800 			}
1801 		}
1802 		if ((flags & MSG_PEEK) &&
1803 		    (peek_seq - copied - urg_hole != tp->copied_seq)) {
1804 			net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1805 					    current->comm,
1806 					    task_pid_nr(current));
1807 			peek_seq = tp->copied_seq;
1808 		}
1809 		continue;
1810 
1811 	found_ok_skb:
1812 		/* Ok so how much can we use? */
1813 		used = skb->len - offset;
1814 		if (len < used)
1815 			used = len;
1816 
1817 		/* Do we have urgent data here? */
1818 		if (tp->urg_data) {
1819 			u32 urg_offset = tp->urg_seq - *seq;
1820 			if (urg_offset < used) {
1821 				if (!urg_offset) {
1822 					if (!sock_flag(sk, SOCK_URGINLINE)) {
1823 						++*seq;
1824 						urg_hole++;
1825 						offset++;
1826 						used--;
1827 						if (!used)
1828 							goto skip_copy;
1829 					}
1830 				} else
1831 					used = urg_offset;
1832 			}
1833 		}
1834 
1835 		if (!(flags & MSG_TRUNC)) {
1836 			err = skb_copy_datagram_iovec(skb, offset,
1837 						      msg->msg_iov, used);
1838 			if (err) {
1839 				/* Exception. Bailout! */
1840 				if (!copied)
1841 					copied = -EFAULT;
1842 				break;
1843 			}
1844 		}
1845 
1846 		*seq += used;
1847 		copied += used;
1848 		len -= used;
1849 
1850 		tcp_rcv_space_adjust(sk);
1851 
1852 skip_copy:
1853 		if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1854 			tp->urg_data = 0;
1855 			tcp_fast_path_check(sk);
1856 		}
1857 		if (used + offset < skb->len)
1858 			continue;
1859 
1860 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1861 			goto found_fin_ok;
1862 		if (!(flags & MSG_PEEK))
1863 			sk_eat_skb(sk, skb);
1864 		continue;
1865 
1866 	found_fin_ok:
1867 		/* Process the FIN. */
1868 		++*seq;
1869 		if (!(flags & MSG_PEEK))
1870 			sk_eat_skb(sk, skb);
1871 		break;
1872 	} while (len > 0);
1873 
1874 	if (user_recv) {
1875 		if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1876 			int chunk;
1877 
1878 			tp->ucopy.len = copied > 0 ? len : 0;
1879 
1880 			tcp_prequeue_process(sk);
1881 
1882 			if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1883 				NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1884 				len -= chunk;
1885 				copied += chunk;
1886 			}
1887 		}
1888 
1889 		tp->ucopy.task = NULL;
1890 		tp->ucopy.len = 0;
1891 	}
1892 
1893 	/* According to UNIX98, msg_name/msg_namelen are ignored
1894 	 * on connected socket. I was just happy when found this 8) --ANK
1895 	 */
1896 
1897 	/* Clean up data we have read: This will do ACK frames. */
1898 	tcp_cleanup_rbuf(sk, copied);
1899 
1900 	release_sock(sk);
1901 	return copied;
1902 
1903 out:
1904 	release_sock(sk);
1905 	return err;
1906 
1907 recv_urg:
1908 	err = tcp_recv_urg(sk, msg, len, flags);
1909 	goto out;
1910 
1911 recv_sndq:
1912 	err = tcp_peek_sndq(sk, msg, len);
1913 	goto out;
1914 }
1915 EXPORT_SYMBOL(tcp_recvmsg);
1916 
1917 void tcp_set_state(struct sock *sk, int state)
1918 {
1919 	int oldstate = sk->sk_state;
1920 
1921 	switch (state) {
1922 	case TCP_ESTABLISHED:
1923 		if (oldstate != TCP_ESTABLISHED)
1924 			TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1925 		break;
1926 
1927 	case TCP_CLOSE:
1928 		if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1929 			TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1930 
1931 		sk->sk_prot->unhash(sk);
1932 		if (inet_csk(sk)->icsk_bind_hash &&
1933 		    !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1934 			inet_put_port(sk);
1935 		/* fall through */
1936 	default:
1937 		if (oldstate == TCP_ESTABLISHED)
1938 			TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1939 	}
1940 
1941 	/* Change state AFTER socket is unhashed to avoid closed
1942 	 * socket sitting in hash tables.
1943 	 */
1944 	sk->sk_state = state;
1945 
1946 #ifdef STATE_TRACE
1947 	SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1948 #endif
1949 }
1950 EXPORT_SYMBOL_GPL(tcp_set_state);
1951 
1952 /*
1953  *	State processing on a close. This implements the state shift for
1954  *	sending our FIN frame. Note that we only send a FIN for some
1955  *	states. A shutdown() may have already sent the FIN, or we may be
1956  *	closed.
1957  */
1958 
1959 static const unsigned char new_state[16] = {
1960   /* current state:        new state:      action:	*/
1961   /* (Invalid)		*/ TCP_CLOSE,
1962   /* TCP_ESTABLISHED	*/ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1963   /* TCP_SYN_SENT	*/ TCP_CLOSE,
1964   /* TCP_SYN_RECV	*/ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1965   /* TCP_FIN_WAIT1	*/ TCP_FIN_WAIT1,
1966   /* TCP_FIN_WAIT2	*/ TCP_FIN_WAIT2,
1967   /* TCP_TIME_WAIT	*/ TCP_CLOSE,
1968   /* TCP_CLOSE		*/ TCP_CLOSE,
1969   /* TCP_CLOSE_WAIT	*/ TCP_LAST_ACK  | TCP_ACTION_FIN,
1970   /* TCP_LAST_ACK	*/ TCP_LAST_ACK,
1971   /* TCP_LISTEN		*/ TCP_CLOSE,
1972   /* TCP_CLOSING	*/ TCP_CLOSING,
1973 };
1974 
1975 static int tcp_close_state(struct sock *sk)
1976 {
1977 	int next = (int)new_state[sk->sk_state];
1978 	int ns = next & TCP_STATE_MASK;
1979 
1980 	tcp_set_state(sk, ns);
1981 
1982 	return next & TCP_ACTION_FIN;
1983 }
1984 
1985 /*
1986  *	Shutdown the sending side of a connection. Much like close except
1987  *	that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1988  */
1989 
1990 void tcp_shutdown(struct sock *sk, int how)
1991 {
1992 	/*	We need to grab some memory, and put together a FIN,
1993 	 *	and then put it into the queue to be sent.
1994 	 *		Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1995 	 */
1996 	if (!(how & SEND_SHUTDOWN))
1997 		return;
1998 
1999 	/* If we've already sent a FIN, or it's a closed state, skip this. */
2000 	if ((1 << sk->sk_state) &
2001 	    (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2002 	     TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2003 		/* Clear out any half completed packets.  FIN if needed. */
2004 		if (tcp_close_state(sk))
2005 			tcp_send_fin(sk);
2006 	}
2007 }
2008 EXPORT_SYMBOL(tcp_shutdown);
2009 
2010 bool tcp_check_oom(struct sock *sk, int shift)
2011 {
2012 	bool too_many_orphans, out_of_socket_memory;
2013 
2014 	too_many_orphans = tcp_too_many_orphans(sk, shift);
2015 	out_of_socket_memory = tcp_out_of_memory(sk);
2016 
2017 	if (too_many_orphans)
2018 		net_info_ratelimited("too many orphaned sockets\n");
2019 	if (out_of_socket_memory)
2020 		net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2021 	return too_many_orphans || out_of_socket_memory;
2022 }
2023 
2024 void tcp_close(struct sock *sk, long timeout)
2025 {
2026 	struct sk_buff *skb;
2027 	int data_was_unread = 0;
2028 	int state;
2029 
2030 	lock_sock(sk);
2031 	sk->sk_shutdown = SHUTDOWN_MASK;
2032 
2033 	if (sk->sk_state == TCP_LISTEN) {
2034 		tcp_set_state(sk, TCP_CLOSE);
2035 
2036 		/* Special case. */
2037 		inet_csk_listen_stop(sk);
2038 
2039 		goto adjudge_to_death;
2040 	}
2041 
2042 	/*  We need to flush the recv. buffs.  We do this only on the
2043 	 *  descriptor close, not protocol-sourced closes, because the
2044 	 *  reader process may not have drained the data yet!
2045 	 */
2046 	while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2047 		u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2048 
2049 		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2050 			len--;
2051 		data_was_unread += len;
2052 		__kfree_skb(skb);
2053 	}
2054 
2055 	sk_mem_reclaim(sk);
2056 
2057 	/* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2058 	if (sk->sk_state == TCP_CLOSE)
2059 		goto adjudge_to_death;
2060 
2061 	/* As outlined in RFC 2525, section 2.17, we send a RST here because
2062 	 * data was lost. To witness the awful effects of the old behavior of
2063 	 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2064 	 * GET in an FTP client, suspend the process, wait for the client to
2065 	 * advertise a zero window, then kill -9 the FTP client, wheee...
2066 	 * Note: timeout is always zero in such a case.
2067 	 */
2068 	if (unlikely(tcp_sk(sk)->repair)) {
2069 		sk->sk_prot->disconnect(sk, 0);
2070 	} else if (data_was_unread) {
2071 		/* Unread data was tossed, zap the connection. */
2072 		NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2073 		tcp_set_state(sk, TCP_CLOSE);
2074 		tcp_send_active_reset(sk, sk->sk_allocation);
2075 	} else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2076 		/* Check zero linger _after_ checking for unread data. */
2077 		sk->sk_prot->disconnect(sk, 0);
2078 		NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2079 	} else if (tcp_close_state(sk)) {
2080 		/* We FIN if the application ate all the data before
2081 		 * zapping the connection.
2082 		 */
2083 
2084 		/* RED-PEN. Formally speaking, we have broken TCP state
2085 		 * machine. State transitions:
2086 		 *
2087 		 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2088 		 * TCP_SYN_RECV	-> TCP_FIN_WAIT1 (forget it, it's impossible)
2089 		 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2090 		 *
2091 		 * are legal only when FIN has been sent (i.e. in window),
2092 		 * rather than queued out of window. Purists blame.
2093 		 *
2094 		 * F.e. "RFC state" is ESTABLISHED,
2095 		 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2096 		 *
2097 		 * The visible declinations are that sometimes
2098 		 * we enter time-wait state, when it is not required really
2099 		 * (harmless), do not send active resets, when they are
2100 		 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2101 		 * they look as CLOSING or LAST_ACK for Linux)
2102 		 * Probably, I missed some more holelets.
2103 		 * 						--ANK
2104 		 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2105 		 * in a single packet! (May consider it later but will
2106 		 * probably need API support or TCP_CORK SYN-ACK until
2107 		 * data is written and socket is closed.)
2108 		 */
2109 		tcp_send_fin(sk);
2110 	}
2111 
2112 	sk_stream_wait_close(sk, timeout);
2113 
2114 adjudge_to_death:
2115 	state = sk->sk_state;
2116 	sock_hold(sk);
2117 	sock_orphan(sk);
2118 
2119 	/* It is the last release_sock in its life. It will remove backlog. */
2120 	release_sock(sk);
2121 
2122 
2123 	/* Now socket is owned by kernel and we acquire BH lock
2124 	   to finish close. No need to check for user refs.
2125 	 */
2126 	local_bh_disable();
2127 	bh_lock_sock(sk);
2128 	WARN_ON(sock_owned_by_user(sk));
2129 
2130 	percpu_counter_inc(sk->sk_prot->orphan_count);
2131 
2132 	/* Have we already been destroyed by a softirq or backlog? */
2133 	if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2134 		goto out;
2135 
2136 	/*	This is a (useful) BSD violating of the RFC. There is a
2137 	 *	problem with TCP as specified in that the other end could
2138 	 *	keep a socket open forever with no application left this end.
2139 	 *	We use a 1 minute timeout (about the same as BSD) then kill
2140 	 *	our end. If they send after that then tough - BUT: long enough
2141 	 *	that we won't make the old 4*rto = almost no time - whoops
2142 	 *	reset mistake.
2143 	 *
2144 	 *	Nope, it was not mistake. It is really desired behaviour
2145 	 *	f.e. on http servers, when such sockets are useless, but
2146 	 *	consume significant resources. Let's do it with special
2147 	 *	linger2	option.					--ANK
2148 	 */
2149 
2150 	if (sk->sk_state == TCP_FIN_WAIT2) {
2151 		struct tcp_sock *tp = tcp_sk(sk);
2152 		if (tp->linger2 < 0) {
2153 			tcp_set_state(sk, TCP_CLOSE);
2154 			tcp_send_active_reset(sk, GFP_ATOMIC);
2155 			NET_INC_STATS_BH(sock_net(sk),
2156 					LINUX_MIB_TCPABORTONLINGER);
2157 		} else {
2158 			const int tmo = tcp_fin_time(sk);
2159 
2160 			if (tmo > TCP_TIMEWAIT_LEN) {
2161 				inet_csk_reset_keepalive_timer(sk,
2162 						tmo - TCP_TIMEWAIT_LEN);
2163 			} else {
2164 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2165 				goto out;
2166 			}
2167 		}
2168 	}
2169 	if (sk->sk_state != TCP_CLOSE) {
2170 		sk_mem_reclaim(sk);
2171 		if (tcp_check_oom(sk, 0)) {
2172 			tcp_set_state(sk, TCP_CLOSE);
2173 			tcp_send_active_reset(sk, GFP_ATOMIC);
2174 			NET_INC_STATS_BH(sock_net(sk),
2175 					LINUX_MIB_TCPABORTONMEMORY);
2176 		}
2177 	}
2178 
2179 	if (sk->sk_state == TCP_CLOSE) {
2180 		struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2181 		/* We could get here with a non-NULL req if the socket is
2182 		 * aborted (e.g., closed with unread data) before 3WHS
2183 		 * finishes.
2184 		 */
2185 		if (req != NULL)
2186 			reqsk_fastopen_remove(sk, req, false);
2187 		inet_csk_destroy_sock(sk);
2188 	}
2189 	/* Otherwise, socket is reprieved until protocol close. */
2190 
2191 out:
2192 	bh_unlock_sock(sk);
2193 	local_bh_enable();
2194 	sock_put(sk);
2195 }
2196 EXPORT_SYMBOL(tcp_close);
2197 
2198 /* These states need RST on ABORT according to RFC793 */
2199 
2200 static inline bool tcp_need_reset(int state)
2201 {
2202 	return (1 << state) &
2203 	       (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2204 		TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2205 }
2206 
2207 int tcp_disconnect(struct sock *sk, int flags)
2208 {
2209 	struct inet_sock *inet = inet_sk(sk);
2210 	struct inet_connection_sock *icsk = inet_csk(sk);
2211 	struct tcp_sock *tp = tcp_sk(sk);
2212 	int err = 0;
2213 	int old_state = sk->sk_state;
2214 
2215 	if (old_state != TCP_CLOSE)
2216 		tcp_set_state(sk, TCP_CLOSE);
2217 
2218 	/* ABORT function of RFC793 */
2219 	if (old_state == TCP_LISTEN) {
2220 		inet_csk_listen_stop(sk);
2221 	} else if (unlikely(tp->repair)) {
2222 		sk->sk_err = ECONNABORTED;
2223 	} else if (tcp_need_reset(old_state) ||
2224 		   (tp->snd_nxt != tp->write_seq &&
2225 		    (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2226 		/* The last check adjusts for discrepancy of Linux wrt. RFC
2227 		 * states
2228 		 */
2229 		tcp_send_active_reset(sk, gfp_any());
2230 		sk->sk_err = ECONNRESET;
2231 	} else if (old_state == TCP_SYN_SENT)
2232 		sk->sk_err = ECONNRESET;
2233 
2234 	tcp_clear_xmit_timers(sk);
2235 	__skb_queue_purge(&sk->sk_receive_queue);
2236 	tcp_write_queue_purge(sk);
2237 	__skb_queue_purge(&tp->out_of_order_queue);
2238 
2239 	inet->inet_dport = 0;
2240 
2241 	if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2242 		inet_reset_saddr(sk);
2243 
2244 	sk->sk_shutdown = 0;
2245 	sock_reset_flag(sk, SOCK_DONE);
2246 	tp->srtt_us = 0;
2247 	if ((tp->write_seq += tp->max_window + 2) == 0)
2248 		tp->write_seq = 1;
2249 	icsk->icsk_backoff = 0;
2250 	tp->snd_cwnd = 2;
2251 	icsk->icsk_probes_out = 0;
2252 	tp->packets_out = 0;
2253 	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2254 	tp->snd_cwnd_cnt = 0;
2255 	tp->window_clamp = 0;
2256 	tcp_set_ca_state(sk, TCP_CA_Open);
2257 	tcp_clear_retrans(tp);
2258 	inet_csk_delack_init(sk);
2259 	tcp_init_send_head(sk);
2260 	memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2261 	__sk_dst_reset(sk);
2262 
2263 	WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2264 
2265 	sk->sk_error_report(sk);
2266 	return err;
2267 }
2268 EXPORT_SYMBOL(tcp_disconnect);
2269 
2270 void tcp_sock_destruct(struct sock *sk)
2271 {
2272 	inet_sock_destruct(sk);
2273 
2274 	kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2275 }
2276 
2277 static inline bool tcp_can_repair_sock(const struct sock *sk)
2278 {
2279 	return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2280 		((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2281 }
2282 
2283 static int tcp_repair_options_est(struct tcp_sock *tp,
2284 		struct tcp_repair_opt __user *optbuf, unsigned int len)
2285 {
2286 	struct tcp_repair_opt opt;
2287 
2288 	while (len >= sizeof(opt)) {
2289 		if (copy_from_user(&opt, optbuf, sizeof(opt)))
2290 			return -EFAULT;
2291 
2292 		optbuf++;
2293 		len -= sizeof(opt);
2294 
2295 		switch (opt.opt_code) {
2296 		case TCPOPT_MSS:
2297 			tp->rx_opt.mss_clamp = opt.opt_val;
2298 			break;
2299 		case TCPOPT_WINDOW:
2300 			{
2301 				u16 snd_wscale = opt.opt_val & 0xFFFF;
2302 				u16 rcv_wscale = opt.opt_val >> 16;
2303 
2304 				if (snd_wscale > 14 || rcv_wscale > 14)
2305 					return -EFBIG;
2306 
2307 				tp->rx_opt.snd_wscale = snd_wscale;
2308 				tp->rx_opt.rcv_wscale = rcv_wscale;
2309 				tp->rx_opt.wscale_ok = 1;
2310 			}
2311 			break;
2312 		case TCPOPT_SACK_PERM:
2313 			if (opt.opt_val != 0)
2314 				return -EINVAL;
2315 
2316 			tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2317 			if (sysctl_tcp_fack)
2318 				tcp_enable_fack(tp);
2319 			break;
2320 		case TCPOPT_TIMESTAMP:
2321 			if (opt.opt_val != 0)
2322 				return -EINVAL;
2323 
2324 			tp->rx_opt.tstamp_ok = 1;
2325 			break;
2326 		}
2327 	}
2328 
2329 	return 0;
2330 }
2331 
2332 /*
2333  *	Socket option code for TCP.
2334  */
2335 static int do_tcp_setsockopt(struct sock *sk, int level,
2336 		int optname, char __user *optval, unsigned int optlen)
2337 {
2338 	struct tcp_sock *tp = tcp_sk(sk);
2339 	struct inet_connection_sock *icsk = inet_csk(sk);
2340 	int val;
2341 	int err = 0;
2342 
2343 	/* These are data/string values, all the others are ints */
2344 	switch (optname) {
2345 	case TCP_CONGESTION: {
2346 		char name[TCP_CA_NAME_MAX];
2347 
2348 		if (optlen < 1)
2349 			return -EINVAL;
2350 
2351 		val = strncpy_from_user(name, optval,
2352 					min_t(long, TCP_CA_NAME_MAX-1, optlen));
2353 		if (val < 0)
2354 			return -EFAULT;
2355 		name[val] = 0;
2356 
2357 		lock_sock(sk);
2358 		err = tcp_set_congestion_control(sk, name);
2359 		release_sock(sk);
2360 		return err;
2361 	}
2362 	default:
2363 		/* fallthru */
2364 		break;
2365 	}
2366 
2367 	if (optlen < sizeof(int))
2368 		return -EINVAL;
2369 
2370 	if (get_user(val, (int __user *)optval))
2371 		return -EFAULT;
2372 
2373 	lock_sock(sk);
2374 
2375 	switch (optname) {
2376 	case TCP_MAXSEG:
2377 		/* Values greater than interface MTU won't take effect. However
2378 		 * at the point when this call is done we typically don't yet
2379 		 * know which interface is going to be used */
2380 		if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2381 			err = -EINVAL;
2382 			break;
2383 		}
2384 		tp->rx_opt.user_mss = val;
2385 		break;
2386 
2387 	case TCP_NODELAY:
2388 		if (val) {
2389 			/* TCP_NODELAY is weaker than TCP_CORK, so that
2390 			 * this option on corked socket is remembered, but
2391 			 * it is not activated until cork is cleared.
2392 			 *
2393 			 * However, when TCP_NODELAY is set we make
2394 			 * an explicit push, which overrides even TCP_CORK
2395 			 * for currently queued segments.
2396 			 */
2397 			tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2398 			tcp_push_pending_frames(sk);
2399 		} else {
2400 			tp->nonagle &= ~TCP_NAGLE_OFF;
2401 		}
2402 		break;
2403 
2404 	case TCP_THIN_LINEAR_TIMEOUTS:
2405 		if (val < 0 || val > 1)
2406 			err = -EINVAL;
2407 		else
2408 			tp->thin_lto = val;
2409 		break;
2410 
2411 	case TCP_THIN_DUPACK:
2412 		if (val < 0 || val > 1)
2413 			err = -EINVAL;
2414 		else {
2415 			tp->thin_dupack = val;
2416 			if (tp->thin_dupack)
2417 				tcp_disable_early_retrans(tp);
2418 		}
2419 		break;
2420 
2421 	case TCP_REPAIR:
2422 		if (!tcp_can_repair_sock(sk))
2423 			err = -EPERM;
2424 		else if (val == 1) {
2425 			tp->repair = 1;
2426 			sk->sk_reuse = SK_FORCE_REUSE;
2427 			tp->repair_queue = TCP_NO_QUEUE;
2428 		} else if (val == 0) {
2429 			tp->repair = 0;
2430 			sk->sk_reuse = SK_NO_REUSE;
2431 			tcp_send_window_probe(sk);
2432 		} else
2433 			err = -EINVAL;
2434 
2435 		break;
2436 
2437 	case TCP_REPAIR_QUEUE:
2438 		if (!tp->repair)
2439 			err = -EPERM;
2440 		else if (val < TCP_QUEUES_NR)
2441 			tp->repair_queue = val;
2442 		else
2443 			err = -EINVAL;
2444 		break;
2445 
2446 	case TCP_QUEUE_SEQ:
2447 		if (sk->sk_state != TCP_CLOSE)
2448 			err = -EPERM;
2449 		else if (tp->repair_queue == TCP_SEND_QUEUE)
2450 			tp->write_seq = val;
2451 		else if (tp->repair_queue == TCP_RECV_QUEUE)
2452 			tp->rcv_nxt = val;
2453 		else
2454 			err = -EINVAL;
2455 		break;
2456 
2457 	case TCP_REPAIR_OPTIONS:
2458 		if (!tp->repair)
2459 			err = -EINVAL;
2460 		else if (sk->sk_state == TCP_ESTABLISHED)
2461 			err = tcp_repair_options_est(tp,
2462 					(struct tcp_repair_opt __user *)optval,
2463 					optlen);
2464 		else
2465 			err = -EPERM;
2466 		break;
2467 
2468 	case TCP_CORK:
2469 		/* When set indicates to always queue non-full frames.
2470 		 * Later the user clears this option and we transmit
2471 		 * any pending partial frames in the queue.  This is
2472 		 * meant to be used alongside sendfile() to get properly
2473 		 * filled frames when the user (for example) must write
2474 		 * out headers with a write() call first and then use
2475 		 * sendfile to send out the data parts.
2476 		 *
2477 		 * TCP_CORK can be set together with TCP_NODELAY and it is
2478 		 * stronger than TCP_NODELAY.
2479 		 */
2480 		if (val) {
2481 			tp->nonagle |= TCP_NAGLE_CORK;
2482 		} else {
2483 			tp->nonagle &= ~TCP_NAGLE_CORK;
2484 			if (tp->nonagle&TCP_NAGLE_OFF)
2485 				tp->nonagle |= TCP_NAGLE_PUSH;
2486 			tcp_push_pending_frames(sk);
2487 		}
2488 		break;
2489 
2490 	case TCP_KEEPIDLE:
2491 		if (val < 1 || val > MAX_TCP_KEEPIDLE)
2492 			err = -EINVAL;
2493 		else {
2494 			tp->keepalive_time = val * HZ;
2495 			if (sock_flag(sk, SOCK_KEEPOPEN) &&
2496 			    !((1 << sk->sk_state) &
2497 			      (TCPF_CLOSE | TCPF_LISTEN))) {
2498 				u32 elapsed = keepalive_time_elapsed(tp);
2499 				if (tp->keepalive_time > elapsed)
2500 					elapsed = tp->keepalive_time - elapsed;
2501 				else
2502 					elapsed = 0;
2503 				inet_csk_reset_keepalive_timer(sk, elapsed);
2504 			}
2505 		}
2506 		break;
2507 	case TCP_KEEPINTVL:
2508 		if (val < 1 || val > MAX_TCP_KEEPINTVL)
2509 			err = -EINVAL;
2510 		else
2511 			tp->keepalive_intvl = val * HZ;
2512 		break;
2513 	case TCP_KEEPCNT:
2514 		if (val < 1 || val > MAX_TCP_KEEPCNT)
2515 			err = -EINVAL;
2516 		else
2517 			tp->keepalive_probes = val;
2518 		break;
2519 	case TCP_SYNCNT:
2520 		if (val < 1 || val > MAX_TCP_SYNCNT)
2521 			err = -EINVAL;
2522 		else
2523 			icsk->icsk_syn_retries = val;
2524 		break;
2525 
2526 	case TCP_LINGER2:
2527 		if (val < 0)
2528 			tp->linger2 = -1;
2529 		else if (val > sysctl_tcp_fin_timeout / HZ)
2530 			tp->linger2 = 0;
2531 		else
2532 			tp->linger2 = val * HZ;
2533 		break;
2534 
2535 	case TCP_DEFER_ACCEPT:
2536 		/* Translate value in seconds to number of retransmits */
2537 		icsk->icsk_accept_queue.rskq_defer_accept =
2538 			secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2539 					TCP_RTO_MAX / HZ);
2540 		break;
2541 
2542 	case TCP_WINDOW_CLAMP:
2543 		if (!val) {
2544 			if (sk->sk_state != TCP_CLOSE) {
2545 				err = -EINVAL;
2546 				break;
2547 			}
2548 			tp->window_clamp = 0;
2549 		} else
2550 			tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2551 						SOCK_MIN_RCVBUF / 2 : val;
2552 		break;
2553 
2554 	case TCP_QUICKACK:
2555 		if (!val) {
2556 			icsk->icsk_ack.pingpong = 1;
2557 		} else {
2558 			icsk->icsk_ack.pingpong = 0;
2559 			if ((1 << sk->sk_state) &
2560 			    (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2561 			    inet_csk_ack_scheduled(sk)) {
2562 				icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2563 				tcp_cleanup_rbuf(sk, 1);
2564 				if (!(val & 1))
2565 					icsk->icsk_ack.pingpong = 1;
2566 			}
2567 		}
2568 		break;
2569 
2570 #ifdef CONFIG_TCP_MD5SIG
2571 	case TCP_MD5SIG:
2572 		/* Read the IP->Key mappings from userspace */
2573 		err = tp->af_specific->md5_parse(sk, optval, optlen);
2574 		break;
2575 #endif
2576 	case TCP_USER_TIMEOUT:
2577 		/* Cap the max time in ms TCP will retry or probe the window
2578 		 * before giving up and aborting (ETIMEDOUT) a connection.
2579 		 */
2580 		if (val < 0)
2581 			err = -EINVAL;
2582 		else
2583 			icsk->icsk_user_timeout = msecs_to_jiffies(val);
2584 		break;
2585 
2586 	case TCP_FASTOPEN:
2587 		if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2588 		    TCPF_LISTEN)))
2589 			err = fastopen_init_queue(sk, val);
2590 		else
2591 			err = -EINVAL;
2592 		break;
2593 	case TCP_TIMESTAMP:
2594 		if (!tp->repair)
2595 			err = -EPERM;
2596 		else
2597 			tp->tsoffset = val - tcp_time_stamp;
2598 		break;
2599 	case TCP_NOTSENT_LOWAT:
2600 		tp->notsent_lowat = val;
2601 		sk->sk_write_space(sk);
2602 		break;
2603 	default:
2604 		err = -ENOPROTOOPT;
2605 		break;
2606 	}
2607 
2608 	release_sock(sk);
2609 	return err;
2610 }
2611 
2612 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2613 		   unsigned int optlen)
2614 {
2615 	const struct inet_connection_sock *icsk = inet_csk(sk);
2616 
2617 	if (level != SOL_TCP)
2618 		return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2619 						     optval, optlen);
2620 	return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2621 }
2622 EXPORT_SYMBOL(tcp_setsockopt);
2623 
2624 #ifdef CONFIG_COMPAT
2625 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2626 			  char __user *optval, unsigned int optlen)
2627 {
2628 	if (level != SOL_TCP)
2629 		return inet_csk_compat_setsockopt(sk, level, optname,
2630 						  optval, optlen);
2631 	return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2632 }
2633 EXPORT_SYMBOL(compat_tcp_setsockopt);
2634 #endif
2635 
2636 /* Return information about state of tcp endpoint in API format. */
2637 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2638 {
2639 	const struct tcp_sock *tp = tcp_sk(sk);
2640 	const struct inet_connection_sock *icsk = inet_csk(sk);
2641 	u32 now = tcp_time_stamp;
2642 
2643 	memset(info, 0, sizeof(*info));
2644 
2645 	info->tcpi_state = sk->sk_state;
2646 	info->tcpi_ca_state = icsk->icsk_ca_state;
2647 	info->tcpi_retransmits = icsk->icsk_retransmits;
2648 	info->tcpi_probes = icsk->icsk_probes_out;
2649 	info->tcpi_backoff = icsk->icsk_backoff;
2650 
2651 	if (tp->rx_opt.tstamp_ok)
2652 		info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2653 	if (tcp_is_sack(tp))
2654 		info->tcpi_options |= TCPI_OPT_SACK;
2655 	if (tp->rx_opt.wscale_ok) {
2656 		info->tcpi_options |= TCPI_OPT_WSCALE;
2657 		info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2658 		info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2659 	}
2660 
2661 	if (tp->ecn_flags & TCP_ECN_OK)
2662 		info->tcpi_options |= TCPI_OPT_ECN;
2663 	if (tp->ecn_flags & TCP_ECN_SEEN)
2664 		info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2665 	if (tp->syn_data_acked)
2666 		info->tcpi_options |= TCPI_OPT_SYN_DATA;
2667 
2668 	info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2669 	info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2670 	info->tcpi_snd_mss = tp->mss_cache;
2671 	info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2672 
2673 	if (sk->sk_state == TCP_LISTEN) {
2674 		info->tcpi_unacked = sk->sk_ack_backlog;
2675 		info->tcpi_sacked = sk->sk_max_ack_backlog;
2676 	} else {
2677 		info->tcpi_unacked = tp->packets_out;
2678 		info->tcpi_sacked = tp->sacked_out;
2679 	}
2680 	info->tcpi_lost = tp->lost_out;
2681 	info->tcpi_retrans = tp->retrans_out;
2682 	info->tcpi_fackets = tp->fackets_out;
2683 
2684 	info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2685 	info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2686 	info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2687 
2688 	info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2689 	info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2690 	info->tcpi_rtt = tp->srtt_us >> 3;
2691 	info->tcpi_rttvar = tp->mdev_us >> 2;
2692 	info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2693 	info->tcpi_snd_cwnd = tp->snd_cwnd;
2694 	info->tcpi_advmss = tp->advmss;
2695 	info->tcpi_reordering = tp->reordering;
2696 
2697 	info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2698 	info->tcpi_rcv_space = tp->rcvq_space.space;
2699 
2700 	info->tcpi_total_retrans = tp->total_retrans;
2701 
2702 	info->tcpi_pacing_rate = sk->sk_pacing_rate != ~0U ?
2703 					sk->sk_pacing_rate : ~0ULL;
2704 	info->tcpi_max_pacing_rate = sk->sk_max_pacing_rate != ~0U ?
2705 					sk->sk_max_pacing_rate : ~0ULL;
2706 }
2707 EXPORT_SYMBOL_GPL(tcp_get_info);
2708 
2709 static int do_tcp_getsockopt(struct sock *sk, int level,
2710 		int optname, char __user *optval, int __user *optlen)
2711 {
2712 	struct inet_connection_sock *icsk = inet_csk(sk);
2713 	struct tcp_sock *tp = tcp_sk(sk);
2714 	int val, len;
2715 
2716 	if (get_user(len, optlen))
2717 		return -EFAULT;
2718 
2719 	len = min_t(unsigned int, len, sizeof(int));
2720 
2721 	if (len < 0)
2722 		return -EINVAL;
2723 
2724 	switch (optname) {
2725 	case TCP_MAXSEG:
2726 		val = tp->mss_cache;
2727 		if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2728 			val = tp->rx_opt.user_mss;
2729 		if (tp->repair)
2730 			val = tp->rx_opt.mss_clamp;
2731 		break;
2732 	case TCP_NODELAY:
2733 		val = !!(tp->nonagle&TCP_NAGLE_OFF);
2734 		break;
2735 	case TCP_CORK:
2736 		val = !!(tp->nonagle&TCP_NAGLE_CORK);
2737 		break;
2738 	case TCP_KEEPIDLE:
2739 		val = keepalive_time_when(tp) / HZ;
2740 		break;
2741 	case TCP_KEEPINTVL:
2742 		val = keepalive_intvl_when(tp) / HZ;
2743 		break;
2744 	case TCP_KEEPCNT:
2745 		val = keepalive_probes(tp);
2746 		break;
2747 	case TCP_SYNCNT:
2748 		val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2749 		break;
2750 	case TCP_LINGER2:
2751 		val = tp->linger2;
2752 		if (val >= 0)
2753 			val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2754 		break;
2755 	case TCP_DEFER_ACCEPT:
2756 		val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2757 				      TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2758 		break;
2759 	case TCP_WINDOW_CLAMP:
2760 		val = tp->window_clamp;
2761 		break;
2762 	case TCP_INFO: {
2763 		struct tcp_info info;
2764 
2765 		if (get_user(len, optlen))
2766 			return -EFAULT;
2767 
2768 		tcp_get_info(sk, &info);
2769 
2770 		len = min_t(unsigned int, len, sizeof(info));
2771 		if (put_user(len, optlen))
2772 			return -EFAULT;
2773 		if (copy_to_user(optval, &info, len))
2774 			return -EFAULT;
2775 		return 0;
2776 	}
2777 	case TCP_QUICKACK:
2778 		val = !icsk->icsk_ack.pingpong;
2779 		break;
2780 
2781 	case TCP_CONGESTION:
2782 		if (get_user(len, optlen))
2783 			return -EFAULT;
2784 		len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2785 		if (put_user(len, optlen))
2786 			return -EFAULT;
2787 		if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2788 			return -EFAULT;
2789 		return 0;
2790 
2791 	case TCP_THIN_LINEAR_TIMEOUTS:
2792 		val = tp->thin_lto;
2793 		break;
2794 	case TCP_THIN_DUPACK:
2795 		val = tp->thin_dupack;
2796 		break;
2797 
2798 	case TCP_REPAIR:
2799 		val = tp->repair;
2800 		break;
2801 
2802 	case TCP_REPAIR_QUEUE:
2803 		if (tp->repair)
2804 			val = tp->repair_queue;
2805 		else
2806 			return -EINVAL;
2807 		break;
2808 
2809 	case TCP_QUEUE_SEQ:
2810 		if (tp->repair_queue == TCP_SEND_QUEUE)
2811 			val = tp->write_seq;
2812 		else if (tp->repair_queue == TCP_RECV_QUEUE)
2813 			val = tp->rcv_nxt;
2814 		else
2815 			return -EINVAL;
2816 		break;
2817 
2818 	case TCP_USER_TIMEOUT:
2819 		val = jiffies_to_msecs(icsk->icsk_user_timeout);
2820 		break;
2821 
2822 	case TCP_FASTOPEN:
2823 		if (icsk->icsk_accept_queue.fastopenq != NULL)
2824 			val = icsk->icsk_accept_queue.fastopenq->max_qlen;
2825 		else
2826 			val = 0;
2827 		break;
2828 
2829 	case TCP_TIMESTAMP:
2830 		val = tcp_time_stamp + tp->tsoffset;
2831 		break;
2832 	case TCP_NOTSENT_LOWAT:
2833 		val = tp->notsent_lowat;
2834 		break;
2835 	default:
2836 		return -ENOPROTOOPT;
2837 	}
2838 
2839 	if (put_user(len, optlen))
2840 		return -EFAULT;
2841 	if (copy_to_user(optval, &val, len))
2842 		return -EFAULT;
2843 	return 0;
2844 }
2845 
2846 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2847 		   int __user *optlen)
2848 {
2849 	struct inet_connection_sock *icsk = inet_csk(sk);
2850 
2851 	if (level != SOL_TCP)
2852 		return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2853 						     optval, optlen);
2854 	return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2855 }
2856 EXPORT_SYMBOL(tcp_getsockopt);
2857 
2858 #ifdef CONFIG_COMPAT
2859 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2860 			  char __user *optval, int __user *optlen)
2861 {
2862 	if (level != SOL_TCP)
2863 		return inet_csk_compat_getsockopt(sk, level, optname,
2864 						  optval, optlen);
2865 	return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2866 }
2867 EXPORT_SYMBOL(compat_tcp_getsockopt);
2868 #endif
2869 
2870 #ifdef CONFIG_TCP_MD5SIG
2871 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2872 static DEFINE_MUTEX(tcp_md5sig_mutex);
2873 static bool tcp_md5sig_pool_populated = false;
2874 
2875 static void __tcp_alloc_md5sig_pool(void)
2876 {
2877 	int cpu;
2878 
2879 	for_each_possible_cpu(cpu) {
2880 		if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2881 			struct crypto_hash *hash;
2882 
2883 			hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2884 			if (IS_ERR_OR_NULL(hash))
2885 				return;
2886 			per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
2887 		}
2888 	}
2889 	/* before setting tcp_md5sig_pool_populated, we must commit all writes
2890 	 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
2891 	 */
2892 	smp_wmb();
2893 	tcp_md5sig_pool_populated = true;
2894 }
2895 
2896 bool tcp_alloc_md5sig_pool(void)
2897 {
2898 	if (unlikely(!tcp_md5sig_pool_populated)) {
2899 		mutex_lock(&tcp_md5sig_mutex);
2900 
2901 		if (!tcp_md5sig_pool_populated)
2902 			__tcp_alloc_md5sig_pool();
2903 
2904 		mutex_unlock(&tcp_md5sig_mutex);
2905 	}
2906 	return tcp_md5sig_pool_populated;
2907 }
2908 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2909 
2910 
2911 /**
2912  *	tcp_get_md5sig_pool - get md5sig_pool for this user
2913  *
2914  *	We use percpu structure, so if we succeed, we exit with preemption
2915  *	and BH disabled, to make sure another thread or softirq handling
2916  *	wont try to get same context.
2917  */
2918 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2919 {
2920 	local_bh_disable();
2921 
2922 	if (tcp_md5sig_pool_populated) {
2923 		/* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
2924 		smp_rmb();
2925 		return this_cpu_ptr(&tcp_md5sig_pool);
2926 	}
2927 	local_bh_enable();
2928 	return NULL;
2929 }
2930 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2931 
2932 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2933 			const struct tcphdr *th)
2934 {
2935 	struct scatterlist sg;
2936 	struct tcphdr hdr;
2937 	int err;
2938 
2939 	/* We are not allowed to change tcphdr, make a local copy */
2940 	memcpy(&hdr, th, sizeof(hdr));
2941 	hdr.check = 0;
2942 
2943 	/* options aren't included in the hash */
2944 	sg_init_one(&sg, &hdr, sizeof(hdr));
2945 	err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
2946 	return err;
2947 }
2948 EXPORT_SYMBOL(tcp_md5_hash_header);
2949 
2950 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2951 			  const struct sk_buff *skb, unsigned int header_len)
2952 {
2953 	struct scatterlist sg;
2954 	const struct tcphdr *tp = tcp_hdr(skb);
2955 	struct hash_desc *desc = &hp->md5_desc;
2956 	unsigned int i;
2957 	const unsigned int head_data_len = skb_headlen(skb) > header_len ?
2958 					   skb_headlen(skb) - header_len : 0;
2959 	const struct skb_shared_info *shi = skb_shinfo(skb);
2960 	struct sk_buff *frag_iter;
2961 
2962 	sg_init_table(&sg, 1);
2963 
2964 	sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2965 	if (crypto_hash_update(desc, &sg, head_data_len))
2966 		return 1;
2967 
2968 	for (i = 0; i < shi->nr_frags; ++i) {
2969 		const struct skb_frag_struct *f = &shi->frags[i];
2970 		unsigned int offset = f->page_offset;
2971 		struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
2972 
2973 		sg_set_page(&sg, page, skb_frag_size(f),
2974 			    offset_in_page(offset));
2975 		if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
2976 			return 1;
2977 	}
2978 
2979 	skb_walk_frags(skb, frag_iter)
2980 		if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
2981 			return 1;
2982 
2983 	return 0;
2984 }
2985 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2986 
2987 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
2988 {
2989 	struct scatterlist sg;
2990 
2991 	sg_init_one(&sg, key->key, key->keylen);
2992 	return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2993 }
2994 EXPORT_SYMBOL(tcp_md5_hash_key);
2995 
2996 #endif
2997 
2998 void tcp_done(struct sock *sk)
2999 {
3000 	struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3001 
3002 	if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3003 		TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3004 
3005 	tcp_set_state(sk, TCP_CLOSE);
3006 	tcp_clear_xmit_timers(sk);
3007 	if (req != NULL)
3008 		reqsk_fastopen_remove(sk, req, false);
3009 
3010 	sk->sk_shutdown = SHUTDOWN_MASK;
3011 
3012 	if (!sock_flag(sk, SOCK_DEAD))
3013 		sk->sk_state_change(sk);
3014 	else
3015 		inet_csk_destroy_sock(sk);
3016 }
3017 EXPORT_SYMBOL_GPL(tcp_done);
3018 
3019 extern struct tcp_congestion_ops tcp_reno;
3020 
3021 static __initdata unsigned long thash_entries;
3022 static int __init set_thash_entries(char *str)
3023 {
3024 	ssize_t ret;
3025 
3026 	if (!str)
3027 		return 0;
3028 
3029 	ret = kstrtoul(str, 0, &thash_entries);
3030 	if (ret)
3031 		return 0;
3032 
3033 	return 1;
3034 }
3035 __setup("thash_entries=", set_thash_entries);
3036 
3037 static void __init tcp_init_mem(void)
3038 {
3039 	unsigned long limit = nr_free_buffer_pages() / 8;
3040 	limit = max(limit, 128UL);
3041 	sysctl_tcp_mem[0] = limit / 4 * 3;
3042 	sysctl_tcp_mem[1] = limit;
3043 	sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3044 }
3045 
3046 void __init tcp_init(void)
3047 {
3048 	struct sk_buff *skb = NULL;
3049 	unsigned long limit;
3050 	int max_rshare, max_wshare, cnt;
3051 	unsigned int i;
3052 
3053 	BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3054 
3055 	percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3056 	percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3057 	tcp_hashinfo.bind_bucket_cachep =
3058 		kmem_cache_create("tcp_bind_bucket",
3059 				  sizeof(struct inet_bind_bucket), 0,
3060 				  SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3061 
3062 	/* Size and allocate the main established and bind bucket
3063 	 * hash tables.
3064 	 *
3065 	 * The methodology is similar to that of the buffer cache.
3066 	 */
3067 	tcp_hashinfo.ehash =
3068 		alloc_large_system_hash("TCP established",
3069 					sizeof(struct inet_ehash_bucket),
3070 					thash_entries,
3071 					17, /* one slot per 128 KB of memory */
3072 					0,
3073 					NULL,
3074 					&tcp_hashinfo.ehash_mask,
3075 					0,
3076 					thash_entries ? 0 : 512 * 1024);
3077 	for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3078 		INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3079 
3080 	if (inet_ehash_locks_alloc(&tcp_hashinfo))
3081 		panic("TCP: failed to alloc ehash_locks");
3082 	tcp_hashinfo.bhash =
3083 		alloc_large_system_hash("TCP bind",
3084 					sizeof(struct inet_bind_hashbucket),
3085 					tcp_hashinfo.ehash_mask + 1,
3086 					17, /* one slot per 128 KB of memory */
3087 					0,
3088 					&tcp_hashinfo.bhash_size,
3089 					NULL,
3090 					0,
3091 					64 * 1024);
3092 	tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3093 	for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3094 		spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3095 		INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3096 	}
3097 
3098 
3099 	cnt = tcp_hashinfo.ehash_mask + 1;
3100 
3101 	tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3102 	sysctl_tcp_max_orphans = cnt / 2;
3103 	sysctl_max_syn_backlog = max(128, cnt / 256);
3104 
3105 	tcp_init_mem();
3106 	/* Set per-socket limits to no more than 1/128 the pressure threshold */
3107 	limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3108 	max_wshare = min(4UL*1024*1024, limit);
3109 	max_rshare = min(6UL*1024*1024, limit);
3110 
3111 	sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3112 	sysctl_tcp_wmem[1] = 16*1024;
3113 	sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3114 
3115 	sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3116 	sysctl_tcp_rmem[1] = 87380;
3117 	sysctl_tcp_rmem[2] = max(87380, max_rshare);
3118 
3119 	pr_info("Hash tables configured (established %u bind %u)\n",
3120 		tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3121 
3122 	tcp_metrics_init();
3123 	BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3124 	tcp_tasklet_init();
3125 }
3126