xref: /linux/net/ipv4/tcp.c (revision ac8b422838046ffc26be4874a3cbae0d313f4209)
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 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
267 #include <linux/time.h>
268 #include <linux/slab.h>
269 
270 #include <net/icmp.h>
271 #include <net/tcp.h>
272 #include <net/xfrm.h>
273 #include <net/ip.h>
274 #include <net/netdma.h>
275 #include <net/sock.h>
276 
277 #include <asm/uaccess.h>
278 #include <asm/ioctls.h>
279 
280 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
281 
282 struct percpu_counter tcp_orphan_count;
283 EXPORT_SYMBOL_GPL(tcp_orphan_count);
284 
285 int sysctl_tcp_mem[3] __read_mostly;
286 int sysctl_tcp_wmem[3] __read_mostly;
287 int sysctl_tcp_rmem[3] __read_mostly;
288 
289 EXPORT_SYMBOL(sysctl_tcp_mem);
290 EXPORT_SYMBOL(sysctl_tcp_rmem);
291 EXPORT_SYMBOL(sysctl_tcp_wmem);
292 
293 atomic_t tcp_memory_allocated;	/* Current allocated memory. */
294 EXPORT_SYMBOL(tcp_memory_allocated);
295 
296 /*
297  * Current number of TCP sockets.
298  */
299 struct percpu_counter tcp_sockets_allocated;
300 EXPORT_SYMBOL(tcp_sockets_allocated);
301 
302 /*
303  * TCP splice context
304  */
305 struct tcp_splice_state {
306 	struct pipe_inode_info *pipe;
307 	size_t len;
308 	unsigned int flags;
309 };
310 
311 /*
312  * Pressure flag: try to collapse.
313  * Technical note: it is used by multiple contexts non atomically.
314  * All the __sk_mem_schedule() is of this nature: accounting
315  * is strict, actions are advisory and have some latency.
316  */
317 int tcp_memory_pressure __read_mostly;
318 
319 EXPORT_SYMBOL(tcp_memory_pressure);
320 
321 void tcp_enter_memory_pressure(struct sock *sk)
322 {
323 	if (!tcp_memory_pressure) {
324 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
325 		tcp_memory_pressure = 1;
326 	}
327 }
328 
329 EXPORT_SYMBOL(tcp_enter_memory_pressure);
330 
331 /* Convert seconds to retransmits based on initial and max timeout */
332 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
333 {
334 	u8 res = 0;
335 
336 	if (seconds > 0) {
337 		int period = timeout;
338 
339 		res = 1;
340 		while (seconds > period && res < 255) {
341 			res++;
342 			timeout <<= 1;
343 			if (timeout > rto_max)
344 				timeout = rto_max;
345 			period += timeout;
346 		}
347 	}
348 	return res;
349 }
350 
351 /* Convert retransmits to seconds based on initial and max timeout */
352 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
353 {
354 	int period = 0;
355 
356 	if (retrans > 0) {
357 		period = timeout;
358 		while (--retrans) {
359 			timeout <<= 1;
360 			if (timeout > rto_max)
361 				timeout = rto_max;
362 			period += timeout;
363 		}
364 	}
365 	return period;
366 }
367 
368 /*
369  *	Wait for a TCP event.
370  *
371  *	Note that we don't need to lock the socket, as the upper poll layers
372  *	take care of normal races (between the test and the event) and we don't
373  *	go look at any of the socket buffers directly.
374  */
375 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
376 {
377 	unsigned int mask;
378 	struct sock *sk = sock->sk;
379 	struct tcp_sock *tp = tcp_sk(sk);
380 
381 	sock_poll_wait(file, sk_sleep(sk), wait);
382 	if (sk->sk_state == TCP_LISTEN)
383 		return inet_csk_listen_poll(sk);
384 
385 	/* Socket is not locked. We are protected from async events
386 	 * by poll logic and correct handling of state changes
387 	 * made by other threads is impossible in any case.
388 	 */
389 
390 	mask = 0;
391 	if (sk->sk_err)
392 		mask = POLLERR;
393 
394 	/*
395 	 * POLLHUP is certainly not done right. But poll() doesn't
396 	 * have a notion of HUP in just one direction, and for a
397 	 * socket the read side is more interesting.
398 	 *
399 	 * Some poll() documentation says that POLLHUP is incompatible
400 	 * with the POLLOUT/POLLWR flags, so somebody should check this
401 	 * all. But careful, it tends to be safer to return too many
402 	 * bits than too few, and you can easily break real applications
403 	 * if you don't tell them that something has hung up!
404 	 *
405 	 * Check-me.
406 	 *
407 	 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
408 	 * our fs/select.c). It means that after we received EOF,
409 	 * poll always returns immediately, making impossible poll() on write()
410 	 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
411 	 * if and only if shutdown has been made in both directions.
412 	 * Actually, it is interesting to look how Solaris and DUX
413 	 * solve this dilemma. I would prefer, if POLLHUP were maskable,
414 	 * then we could set it on SND_SHUTDOWN. BTW examples given
415 	 * in Stevens' books assume exactly this behaviour, it explains
416 	 * why POLLHUP is incompatible with POLLOUT.	--ANK
417 	 *
418 	 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
419 	 * blocking on fresh not-connected or disconnected socket. --ANK
420 	 */
421 	if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
422 		mask |= POLLHUP;
423 	if (sk->sk_shutdown & RCV_SHUTDOWN)
424 		mask |= POLLIN | POLLRDNORM | POLLRDHUP;
425 
426 	/* Connected? */
427 	if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
428 		int target = sock_rcvlowat(sk, 0, INT_MAX);
429 
430 		if (tp->urg_seq == tp->copied_seq &&
431 		    !sock_flag(sk, SOCK_URGINLINE) &&
432 		    tp->urg_data)
433 			target++;
434 
435 		/* Potential race condition. If read of tp below will
436 		 * escape above sk->sk_state, we can be illegally awaken
437 		 * in SYN_* states. */
438 		if (tp->rcv_nxt - tp->copied_seq >= target)
439 			mask |= POLLIN | POLLRDNORM;
440 
441 		if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
442 			if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
443 				mask |= POLLOUT | POLLWRNORM;
444 			} else {  /* send SIGIO later */
445 				set_bit(SOCK_ASYNC_NOSPACE,
446 					&sk->sk_socket->flags);
447 				set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
448 
449 				/* Race breaker. If space is freed after
450 				 * wspace test but before the flags are set,
451 				 * IO signal will be lost.
452 				 */
453 				if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
454 					mask |= POLLOUT | POLLWRNORM;
455 			}
456 		}
457 
458 		if (tp->urg_data & TCP_URG_VALID)
459 			mask |= POLLPRI;
460 	}
461 	return mask;
462 }
463 
464 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
465 {
466 	struct tcp_sock *tp = tcp_sk(sk);
467 	int answ;
468 
469 	switch (cmd) {
470 	case SIOCINQ:
471 		if (sk->sk_state == TCP_LISTEN)
472 			return -EINVAL;
473 
474 		lock_sock(sk);
475 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
476 			answ = 0;
477 		else if (sock_flag(sk, SOCK_URGINLINE) ||
478 			 !tp->urg_data ||
479 			 before(tp->urg_seq, tp->copied_seq) ||
480 			 !before(tp->urg_seq, tp->rcv_nxt)) {
481 			struct sk_buff *skb;
482 
483 			answ = tp->rcv_nxt - tp->copied_seq;
484 
485 			/* Subtract 1, if FIN is in queue. */
486 			skb = skb_peek_tail(&sk->sk_receive_queue);
487 			if (answ && skb)
488 				answ -= tcp_hdr(skb)->fin;
489 		} else
490 			answ = tp->urg_seq - tp->copied_seq;
491 		release_sock(sk);
492 		break;
493 	case SIOCATMARK:
494 		answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
495 		break;
496 	case SIOCOUTQ:
497 		if (sk->sk_state == TCP_LISTEN)
498 			return -EINVAL;
499 
500 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
501 			answ = 0;
502 		else
503 			answ = tp->write_seq - tp->snd_una;
504 		break;
505 	default:
506 		return -ENOIOCTLCMD;
507 	}
508 
509 	return put_user(answ, (int __user *)arg);
510 }
511 
512 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
513 {
514 	TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
515 	tp->pushed_seq = tp->write_seq;
516 }
517 
518 static inline int forced_push(struct tcp_sock *tp)
519 {
520 	return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
521 }
522 
523 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
524 {
525 	struct tcp_sock *tp = tcp_sk(sk);
526 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
527 
528 	skb->csum    = 0;
529 	tcb->seq     = tcb->end_seq = tp->write_seq;
530 	tcb->flags   = TCPCB_FLAG_ACK;
531 	tcb->sacked  = 0;
532 	skb_header_release(skb);
533 	tcp_add_write_queue_tail(sk, skb);
534 	sk->sk_wmem_queued += skb->truesize;
535 	sk_mem_charge(sk, skb->truesize);
536 	if (tp->nonagle & TCP_NAGLE_PUSH)
537 		tp->nonagle &= ~TCP_NAGLE_PUSH;
538 }
539 
540 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
541 {
542 	if (flags & MSG_OOB)
543 		tp->snd_up = tp->write_seq;
544 }
545 
546 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
547 			    int nonagle)
548 {
549 	if (tcp_send_head(sk)) {
550 		struct tcp_sock *tp = tcp_sk(sk);
551 
552 		if (!(flags & MSG_MORE) || forced_push(tp))
553 			tcp_mark_push(tp, tcp_write_queue_tail(sk));
554 
555 		tcp_mark_urg(tp, flags);
556 		__tcp_push_pending_frames(sk, mss_now,
557 					  (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
558 	}
559 }
560 
561 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
562 				unsigned int offset, size_t len)
563 {
564 	struct tcp_splice_state *tss = rd_desc->arg.data;
565 	int ret;
566 
567 	ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
568 			      tss->flags);
569 	if (ret > 0)
570 		rd_desc->count -= ret;
571 	return ret;
572 }
573 
574 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
575 {
576 	/* Store TCP splice context information in read_descriptor_t. */
577 	read_descriptor_t rd_desc = {
578 		.arg.data = tss,
579 		.count	  = tss->len,
580 	};
581 
582 	return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
583 }
584 
585 /**
586  *  tcp_splice_read - splice data from TCP socket to a pipe
587  * @sock:	socket to splice from
588  * @ppos:	position (not valid)
589  * @pipe:	pipe to splice to
590  * @len:	number of bytes to splice
591  * @flags:	splice modifier flags
592  *
593  * Description:
594  *    Will read pages from given socket and fill them into a pipe.
595  *
596  **/
597 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
598 			struct pipe_inode_info *pipe, size_t len,
599 			unsigned int flags)
600 {
601 	struct sock *sk = sock->sk;
602 	struct tcp_splice_state tss = {
603 		.pipe = pipe,
604 		.len = len,
605 		.flags = flags,
606 	};
607 	long timeo;
608 	ssize_t spliced;
609 	int ret;
610 
611 	sock_rps_record_flow(sk);
612 	/*
613 	 * We can't seek on a socket input
614 	 */
615 	if (unlikely(*ppos))
616 		return -ESPIPE;
617 
618 	ret = spliced = 0;
619 
620 	lock_sock(sk);
621 
622 	timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
623 	while (tss.len) {
624 		ret = __tcp_splice_read(sk, &tss);
625 		if (ret < 0)
626 			break;
627 		else if (!ret) {
628 			if (spliced)
629 				break;
630 			if (sock_flag(sk, SOCK_DONE))
631 				break;
632 			if (sk->sk_err) {
633 				ret = sock_error(sk);
634 				break;
635 			}
636 			if (sk->sk_shutdown & RCV_SHUTDOWN)
637 				break;
638 			if (sk->sk_state == TCP_CLOSE) {
639 				/*
640 				 * This occurs when user tries to read
641 				 * from never connected socket.
642 				 */
643 				if (!sock_flag(sk, SOCK_DONE))
644 					ret = -ENOTCONN;
645 				break;
646 			}
647 			if (!timeo) {
648 				ret = -EAGAIN;
649 				break;
650 			}
651 			sk_wait_data(sk, &timeo);
652 			if (signal_pending(current)) {
653 				ret = sock_intr_errno(timeo);
654 				break;
655 			}
656 			continue;
657 		}
658 		tss.len -= ret;
659 		spliced += ret;
660 
661 		if (!timeo)
662 			break;
663 		release_sock(sk);
664 		lock_sock(sk);
665 
666 		if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
667 		    (sk->sk_shutdown & RCV_SHUTDOWN) ||
668 		    signal_pending(current))
669 			break;
670 	}
671 
672 	release_sock(sk);
673 
674 	if (spliced)
675 		return spliced;
676 
677 	return ret;
678 }
679 
680 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
681 {
682 	struct sk_buff *skb;
683 
684 	/* The TCP header must be at least 32-bit aligned.  */
685 	size = ALIGN(size, 4);
686 
687 	skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
688 	if (skb) {
689 		if (sk_wmem_schedule(sk, skb->truesize)) {
690 			/*
691 			 * Make sure that we have exactly size bytes
692 			 * available to the caller, no more, no less.
693 			 */
694 			skb_reserve(skb, skb_tailroom(skb) - size);
695 			return skb;
696 		}
697 		__kfree_skb(skb);
698 	} else {
699 		sk->sk_prot->enter_memory_pressure(sk);
700 		sk_stream_moderate_sndbuf(sk);
701 	}
702 	return NULL;
703 }
704 
705 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
706 				       int large_allowed)
707 {
708 	struct tcp_sock *tp = tcp_sk(sk);
709 	u32 xmit_size_goal, old_size_goal;
710 
711 	xmit_size_goal = mss_now;
712 
713 	if (large_allowed && sk_can_gso(sk)) {
714 		xmit_size_goal = ((sk->sk_gso_max_size - 1) -
715 				  inet_csk(sk)->icsk_af_ops->net_header_len -
716 				  inet_csk(sk)->icsk_ext_hdr_len -
717 				  tp->tcp_header_len);
718 
719 		xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
720 
721 		/* We try hard to avoid divides here */
722 		old_size_goal = tp->xmit_size_goal_segs * mss_now;
723 
724 		if (likely(old_size_goal <= xmit_size_goal &&
725 			   old_size_goal + mss_now > xmit_size_goal)) {
726 			xmit_size_goal = old_size_goal;
727 		} else {
728 			tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
729 			xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
730 		}
731 	}
732 
733 	return max(xmit_size_goal, mss_now);
734 }
735 
736 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
737 {
738 	int mss_now;
739 
740 	mss_now = tcp_current_mss(sk);
741 	*size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
742 
743 	return mss_now;
744 }
745 
746 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
747 			 size_t psize, int flags)
748 {
749 	struct tcp_sock *tp = tcp_sk(sk);
750 	int mss_now, size_goal;
751 	int err;
752 	ssize_t copied;
753 	long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
754 
755 	/* Wait for a connection to finish. */
756 	if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
757 		if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
758 			goto out_err;
759 
760 	clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
761 
762 	mss_now = tcp_send_mss(sk, &size_goal, flags);
763 	copied = 0;
764 
765 	err = -EPIPE;
766 	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
767 		goto out_err;
768 
769 	while (psize > 0) {
770 		struct sk_buff *skb = tcp_write_queue_tail(sk);
771 		struct page *page = pages[poffset / PAGE_SIZE];
772 		int copy, i, can_coalesce;
773 		int offset = poffset % PAGE_SIZE;
774 		int size = min_t(size_t, psize, PAGE_SIZE - offset);
775 
776 		if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
777 new_segment:
778 			if (!sk_stream_memory_free(sk))
779 				goto wait_for_sndbuf;
780 
781 			skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
782 			if (!skb)
783 				goto wait_for_memory;
784 
785 			skb_entail(sk, skb);
786 			copy = size_goal;
787 		}
788 
789 		if (copy > size)
790 			copy = size;
791 
792 		i = skb_shinfo(skb)->nr_frags;
793 		can_coalesce = skb_can_coalesce(skb, i, page, offset);
794 		if (!can_coalesce && i >= MAX_SKB_FRAGS) {
795 			tcp_mark_push(tp, skb);
796 			goto new_segment;
797 		}
798 		if (!sk_wmem_schedule(sk, copy))
799 			goto wait_for_memory;
800 
801 		if (can_coalesce) {
802 			skb_shinfo(skb)->frags[i - 1].size += copy;
803 		} else {
804 			get_page(page);
805 			skb_fill_page_desc(skb, i, page, offset, copy);
806 		}
807 
808 		skb->len += copy;
809 		skb->data_len += copy;
810 		skb->truesize += copy;
811 		sk->sk_wmem_queued += copy;
812 		sk_mem_charge(sk, copy);
813 		skb->ip_summed = CHECKSUM_PARTIAL;
814 		tp->write_seq += copy;
815 		TCP_SKB_CB(skb)->end_seq += copy;
816 		skb_shinfo(skb)->gso_segs = 0;
817 
818 		if (!copied)
819 			TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
820 
821 		copied += copy;
822 		poffset += copy;
823 		if (!(psize -= copy))
824 			goto out;
825 
826 		if (skb->len < size_goal || (flags & MSG_OOB))
827 			continue;
828 
829 		if (forced_push(tp)) {
830 			tcp_mark_push(tp, skb);
831 			__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
832 		} else if (skb == tcp_send_head(sk))
833 			tcp_push_one(sk, mss_now);
834 		continue;
835 
836 wait_for_sndbuf:
837 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
838 wait_for_memory:
839 		if (copied)
840 			tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
841 
842 		if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
843 			goto do_error;
844 
845 		mss_now = tcp_send_mss(sk, &size_goal, flags);
846 	}
847 
848 out:
849 	if (copied)
850 		tcp_push(sk, flags, mss_now, tp->nonagle);
851 	return copied;
852 
853 do_error:
854 	if (copied)
855 		goto out;
856 out_err:
857 	return sk_stream_error(sk, flags, err);
858 }
859 
860 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
861 		     size_t size, int flags)
862 {
863 	ssize_t res;
864 	struct sock *sk = sock->sk;
865 
866 	if (!(sk->sk_route_caps & NETIF_F_SG) ||
867 	    !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
868 		return sock_no_sendpage(sock, page, offset, size, flags);
869 
870 	lock_sock(sk);
871 	TCP_CHECK_TIMER(sk);
872 	res = do_tcp_sendpages(sk, &page, offset, size, flags);
873 	TCP_CHECK_TIMER(sk);
874 	release_sock(sk);
875 	return res;
876 }
877 
878 #define TCP_PAGE(sk)	(sk->sk_sndmsg_page)
879 #define TCP_OFF(sk)	(sk->sk_sndmsg_off)
880 
881 static inline int select_size(struct sock *sk, int sg)
882 {
883 	struct tcp_sock *tp = tcp_sk(sk);
884 	int tmp = tp->mss_cache;
885 
886 	if (sg) {
887 		if (sk_can_gso(sk))
888 			tmp = 0;
889 		else {
890 			int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
891 
892 			if (tmp >= pgbreak &&
893 			    tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
894 				tmp = pgbreak;
895 		}
896 	}
897 
898 	return tmp;
899 }
900 
901 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
902 		size_t size)
903 {
904 	struct sock *sk = sock->sk;
905 	struct iovec *iov;
906 	struct tcp_sock *tp = tcp_sk(sk);
907 	struct sk_buff *skb;
908 	int iovlen, flags;
909 	int mss_now, size_goal;
910 	int sg, err, copied;
911 	long timeo;
912 
913 	lock_sock(sk);
914 	TCP_CHECK_TIMER(sk);
915 
916 	flags = msg->msg_flags;
917 	timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
918 
919 	/* Wait for a connection to finish. */
920 	if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
921 		if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
922 			goto out_err;
923 
924 	/* This should be in poll */
925 	clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
926 
927 	mss_now = tcp_send_mss(sk, &size_goal, flags);
928 
929 	/* Ok commence sending. */
930 	iovlen = msg->msg_iovlen;
931 	iov = msg->msg_iov;
932 	copied = 0;
933 
934 	err = -EPIPE;
935 	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
936 		goto out_err;
937 
938 	sg = sk->sk_route_caps & NETIF_F_SG;
939 
940 	while (--iovlen >= 0) {
941 		int seglen = iov->iov_len;
942 		unsigned char __user *from = iov->iov_base;
943 
944 		iov++;
945 
946 		while (seglen > 0) {
947 			int copy = 0;
948 			int max = size_goal;
949 
950 			skb = tcp_write_queue_tail(sk);
951 			if (tcp_send_head(sk)) {
952 				if (skb->ip_summed == CHECKSUM_NONE)
953 					max = mss_now;
954 				copy = max - skb->len;
955 			}
956 
957 			if (copy <= 0) {
958 new_segment:
959 				/* Allocate new segment. If the interface is SG,
960 				 * allocate skb fitting to single page.
961 				 */
962 				if (!sk_stream_memory_free(sk))
963 					goto wait_for_sndbuf;
964 
965 				skb = sk_stream_alloc_skb(sk,
966 							  select_size(sk, sg),
967 							  sk->sk_allocation);
968 				if (!skb)
969 					goto wait_for_memory;
970 
971 				/*
972 				 * Check whether we can use HW checksum.
973 				 */
974 				if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
975 					skb->ip_summed = CHECKSUM_PARTIAL;
976 
977 				skb_entail(sk, skb);
978 				copy = size_goal;
979 				max = size_goal;
980 			}
981 
982 			/* Try to append data to the end of skb. */
983 			if (copy > seglen)
984 				copy = seglen;
985 
986 			/* Where to copy to? */
987 			if (skb_tailroom(skb) > 0) {
988 				/* We have some space in skb head. Superb! */
989 				if (copy > skb_tailroom(skb))
990 					copy = skb_tailroom(skb);
991 				if ((err = skb_add_data(skb, from, copy)) != 0)
992 					goto do_fault;
993 			} else {
994 				int merge = 0;
995 				int i = skb_shinfo(skb)->nr_frags;
996 				struct page *page = TCP_PAGE(sk);
997 				int off = TCP_OFF(sk);
998 
999 				if (skb_can_coalesce(skb, i, page, off) &&
1000 				    off != PAGE_SIZE) {
1001 					/* We can extend the last page
1002 					 * fragment. */
1003 					merge = 1;
1004 				} else if (i == MAX_SKB_FRAGS || !sg) {
1005 					/* Need to add new fragment and cannot
1006 					 * do this because interface is non-SG,
1007 					 * or because all the page slots are
1008 					 * busy. */
1009 					tcp_mark_push(tp, skb);
1010 					goto new_segment;
1011 				} else if (page) {
1012 					if (off == PAGE_SIZE) {
1013 						put_page(page);
1014 						TCP_PAGE(sk) = page = NULL;
1015 						off = 0;
1016 					}
1017 				} else
1018 					off = 0;
1019 
1020 				if (copy > PAGE_SIZE - off)
1021 					copy = PAGE_SIZE - off;
1022 
1023 				if (!sk_wmem_schedule(sk, copy))
1024 					goto wait_for_memory;
1025 
1026 				if (!page) {
1027 					/* Allocate new cache page. */
1028 					if (!(page = sk_stream_alloc_page(sk)))
1029 						goto wait_for_memory;
1030 				}
1031 
1032 				/* Time to copy data. We are close to
1033 				 * the end! */
1034 				err = skb_copy_to_page(sk, from, skb, page,
1035 						       off, copy);
1036 				if (err) {
1037 					/* If this page was new, give it to the
1038 					 * socket so it does not get leaked.
1039 					 */
1040 					if (!TCP_PAGE(sk)) {
1041 						TCP_PAGE(sk) = page;
1042 						TCP_OFF(sk) = 0;
1043 					}
1044 					goto do_error;
1045 				}
1046 
1047 				/* Update the skb. */
1048 				if (merge) {
1049 					skb_shinfo(skb)->frags[i - 1].size +=
1050 									copy;
1051 				} else {
1052 					skb_fill_page_desc(skb, i, page, off, copy);
1053 					if (TCP_PAGE(sk)) {
1054 						get_page(page);
1055 					} else if (off + copy < PAGE_SIZE) {
1056 						get_page(page);
1057 						TCP_PAGE(sk) = page;
1058 					}
1059 				}
1060 
1061 				TCP_OFF(sk) = off + copy;
1062 			}
1063 
1064 			if (!copied)
1065 				TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
1066 
1067 			tp->write_seq += copy;
1068 			TCP_SKB_CB(skb)->end_seq += copy;
1069 			skb_shinfo(skb)->gso_segs = 0;
1070 
1071 			from += copy;
1072 			copied += copy;
1073 			if ((seglen -= copy) == 0 && iovlen == 0)
1074 				goto out;
1075 
1076 			if (skb->len < max || (flags & MSG_OOB))
1077 				continue;
1078 
1079 			if (forced_push(tp)) {
1080 				tcp_mark_push(tp, skb);
1081 				__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1082 			} else if (skb == tcp_send_head(sk))
1083 				tcp_push_one(sk, mss_now);
1084 			continue;
1085 
1086 wait_for_sndbuf:
1087 			set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1088 wait_for_memory:
1089 			if (copied)
1090 				tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1091 
1092 			if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1093 				goto do_error;
1094 
1095 			mss_now = tcp_send_mss(sk, &size_goal, flags);
1096 		}
1097 	}
1098 
1099 out:
1100 	if (copied)
1101 		tcp_push(sk, flags, mss_now, tp->nonagle);
1102 	TCP_CHECK_TIMER(sk);
1103 	release_sock(sk);
1104 	return copied;
1105 
1106 do_fault:
1107 	if (!skb->len) {
1108 		tcp_unlink_write_queue(skb, sk);
1109 		/* It is the one place in all of TCP, except connection
1110 		 * reset, where we can be unlinking the send_head.
1111 		 */
1112 		tcp_check_send_head(sk, skb);
1113 		sk_wmem_free_skb(sk, skb);
1114 	}
1115 
1116 do_error:
1117 	if (copied)
1118 		goto out;
1119 out_err:
1120 	err = sk_stream_error(sk, flags, err);
1121 	TCP_CHECK_TIMER(sk);
1122 	release_sock(sk);
1123 	return err;
1124 }
1125 
1126 /*
1127  *	Handle reading urgent data. BSD has very simple semantics for
1128  *	this, no blocking and very strange errors 8)
1129  */
1130 
1131 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1132 {
1133 	struct tcp_sock *tp = tcp_sk(sk);
1134 
1135 	/* No URG data to read. */
1136 	if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1137 	    tp->urg_data == TCP_URG_READ)
1138 		return -EINVAL;	/* Yes this is right ! */
1139 
1140 	if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1141 		return -ENOTCONN;
1142 
1143 	if (tp->urg_data & TCP_URG_VALID) {
1144 		int err = 0;
1145 		char c = tp->urg_data;
1146 
1147 		if (!(flags & MSG_PEEK))
1148 			tp->urg_data = TCP_URG_READ;
1149 
1150 		/* Read urgent data. */
1151 		msg->msg_flags |= MSG_OOB;
1152 
1153 		if (len > 0) {
1154 			if (!(flags & MSG_TRUNC))
1155 				err = memcpy_toiovec(msg->msg_iov, &c, 1);
1156 			len = 1;
1157 		} else
1158 			msg->msg_flags |= MSG_TRUNC;
1159 
1160 		return err ? -EFAULT : len;
1161 	}
1162 
1163 	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1164 		return 0;
1165 
1166 	/* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1167 	 * the available implementations agree in this case:
1168 	 * this call should never block, independent of the
1169 	 * blocking state of the socket.
1170 	 * Mike <pall@rz.uni-karlsruhe.de>
1171 	 */
1172 	return -EAGAIN;
1173 }
1174 
1175 /* Clean up the receive buffer for full frames taken by the user,
1176  * then send an ACK if necessary.  COPIED is the number of bytes
1177  * tcp_recvmsg has given to the user so far, it speeds up the
1178  * calculation of whether or not we must ACK for the sake of
1179  * a window update.
1180  */
1181 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1182 {
1183 	struct tcp_sock *tp = tcp_sk(sk);
1184 	int time_to_ack = 0;
1185 
1186 #if TCP_DEBUG
1187 	struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1188 
1189 	WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1190 	     KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1191 	     tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1192 #endif
1193 
1194 	if (inet_csk_ack_scheduled(sk)) {
1195 		const struct inet_connection_sock *icsk = inet_csk(sk);
1196 		   /* Delayed ACKs frequently hit locked sockets during bulk
1197 		    * receive. */
1198 		if (icsk->icsk_ack.blocked ||
1199 		    /* Once-per-two-segments ACK was not sent by tcp_input.c */
1200 		    tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1201 		    /*
1202 		     * If this read emptied read buffer, we send ACK, if
1203 		     * connection is not bidirectional, user drained
1204 		     * receive buffer and there was a small segment
1205 		     * in queue.
1206 		     */
1207 		    (copied > 0 &&
1208 		     ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1209 		      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1210 		       !icsk->icsk_ack.pingpong)) &&
1211 		      !atomic_read(&sk->sk_rmem_alloc)))
1212 			time_to_ack = 1;
1213 	}
1214 
1215 	/* We send an ACK if we can now advertise a non-zero window
1216 	 * which has been raised "significantly".
1217 	 *
1218 	 * Even if window raised up to infinity, do not send window open ACK
1219 	 * in states, where we will not receive more. It is useless.
1220 	 */
1221 	if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1222 		__u32 rcv_window_now = tcp_receive_window(tp);
1223 
1224 		/* Optimize, __tcp_select_window() is not cheap. */
1225 		if (2*rcv_window_now <= tp->window_clamp) {
1226 			__u32 new_window = __tcp_select_window(sk);
1227 
1228 			/* Send ACK now, if this read freed lots of space
1229 			 * in our buffer. Certainly, new_window is new window.
1230 			 * We can advertise it now, if it is not less than current one.
1231 			 * "Lots" means "at least twice" here.
1232 			 */
1233 			if (new_window && new_window >= 2 * rcv_window_now)
1234 				time_to_ack = 1;
1235 		}
1236 	}
1237 	if (time_to_ack)
1238 		tcp_send_ack(sk);
1239 }
1240 
1241 static void tcp_prequeue_process(struct sock *sk)
1242 {
1243 	struct sk_buff *skb;
1244 	struct tcp_sock *tp = tcp_sk(sk);
1245 
1246 	NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1247 
1248 	/* RX process wants to run with disabled BHs, though it is not
1249 	 * necessary */
1250 	local_bh_disable();
1251 	while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1252 		sk_backlog_rcv(sk, skb);
1253 	local_bh_enable();
1254 
1255 	/* Clear memory counter. */
1256 	tp->ucopy.memory = 0;
1257 }
1258 
1259 #ifdef CONFIG_NET_DMA
1260 static void tcp_service_net_dma(struct sock *sk, bool wait)
1261 {
1262 	dma_cookie_t done, used;
1263 	dma_cookie_t last_issued;
1264 	struct tcp_sock *tp = tcp_sk(sk);
1265 
1266 	if (!tp->ucopy.dma_chan)
1267 		return;
1268 
1269 	last_issued = tp->ucopy.dma_cookie;
1270 	dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1271 
1272 	do {
1273 		if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1274 					      last_issued, &done,
1275 					      &used) == DMA_SUCCESS) {
1276 			/* Safe to free early-copied skbs now */
1277 			__skb_queue_purge(&sk->sk_async_wait_queue);
1278 			break;
1279 		} else {
1280 			struct sk_buff *skb;
1281 			while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1282 			       (dma_async_is_complete(skb->dma_cookie, done,
1283 						      used) == DMA_SUCCESS)) {
1284 				__skb_dequeue(&sk->sk_async_wait_queue);
1285 				kfree_skb(skb);
1286 			}
1287 		}
1288 	} while (wait);
1289 }
1290 #endif
1291 
1292 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1293 {
1294 	struct sk_buff *skb;
1295 	u32 offset;
1296 
1297 	skb_queue_walk(&sk->sk_receive_queue, skb) {
1298 		offset = seq - TCP_SKB_CB(skb)->seq;
1299 		if (tcp_hdr(skb)->syn)
1300 			offset--;
1301 		if (offset < skb->len || tcp_hdr(skb)->fin) {
1302 			*off = offset;
1303 			return skb;
1304 		}
1305 	}
1306 	return NULL;
1307 }
1308 
1309 /*
1310  * This routine provides an alternative to tcp_recvmsg() for routines
1311  * that would like to handle copying from skbuffs directly in 'sendfile'
1312  * fashion.
1313  * Note:
1314  *	- It is assumed that the socket was locked by the caller.
1315  *	- The routine does not block.
1316  *	- At present, there is no support for reading OOB data
1317  *	  or for 'peeking' the socket using this routine
1318  *	  (although both would be easy to implement).
1319  */
1320 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1321 		  sk_read_actor_t recv_actor)
1322 {
1323 	struct sk_buff *skb;
1324 	struct tcp_sock *tp = tcp_sk(sk);
1325 	u32 seq = tp->copied_seq;
1326 	u32 offset;
1327 	int copied = 0;
1328 
1329 	if (sk->sk_state == TCP_LISTEN)
1330 		return -ENOTCONN;
1331 	while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1332 		if (offset < skb->len) {
1333 			int used;
1334 			size_t len;
1335 
1336 			len = skb->len - offset;
1337 			/* Stop reading if we hit a patch of urgent data */
1338 			if (tp->urg_data) {
1339 				u32 urg_offset = tp->urg_seq - seq;
1340 				if (urg_offset < len)
1341 					len = urg_offset;
1342 				if (!len)
1343 					break;
1344 			}
1345 			used = recv_actor(desc, skb, offset, len);
1346 			if (used < 0) {
1347 				if (!copied)
1348 					copied = used;
1349 				break;
1350 			} else if (used <= len) {
1351 				seq += used;
1352 				copied += used;
1353 				offset += used;
1354 			}
1355 			/*
1356 			 * If recv_actor drops the lock (e.g. TCP splice
1357 			 * receive) the skb pointer might be invalid when
1358 			 * getting here: tcp_collapse might have deleted it
1359 			 * while aggregating skbs from the socket queue.
1360 			 */
1361 			skb = tcp_recv_skb(sk, seq-1, &offset);
1362 			if (!skb || (offset+1 != skb->len))
1363 				break;
1364 		}
1365 		if (tcp_hdr(skb)->fin) {
1366 			sk_eat_skb(sk, skb, 0);
1367 			++seq;
1368 			break;
1369 		}
1370 		sk_eat_skb(sk, skb, 0);
1371 		if (!desc->count)
1372 			break;
1373 		tp->copied_seq = seq;
1374 	}
1375 	tp->copied_seq = seq;
1376 
1377 	tcp_rcv_space_adjust(sk);
1378 
1379 	/* Clean up data we have read: This will do ACK frames. */
1380 	if (copied > 0)
1381 		tcp_cleanup_rbuf(sk, copied);
1382 	return copied;
1383 }
1384 
1385 /*
1386  *	This routine copies from a sock struct into the user buffer.
1387  *
1388  *	Technical note: in 2.3 we work on _locked_ socket, so that
1389  *	tricks with *seq access order and skb->users are not required.
1390  *	Probably, code can be easily improved even more.
1391  */
1392 
1393 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1394 		size_t len, int nonblock, int flags, int *addr_len)
1395 {
1396 	struct tcp_sock *tp = tcp_sk(sk);
1397 	int copied = 0;
1398 	u32 peek_seq;
1399 	u32 *seq;
1400 	unsigned long used;
1401 	int err;
1402 	int target;		/* Read at least this many bytes */
1403 	long timeo;
1404 	struct task_struct *user_recv = NULL;
1405 	int copied_early = 0;
1406 	struct sk_buff *skb;
1407 	u32 urg_hole = 0;
1408 
1409 	lock_sock(sk);
1410 
1411 	TCP_CHECK_TIMER(sk);
1412 
1413 	err = -ENOTCONN;
1414 	if (sk->sk_state == TCP_LISTEN)
1415 		goto out;
1416 
1417 	timeo = sock_rcvtimeo(sk, nonblock);
1418 
1419 	/* Urgent data needs to be handled specially. */
1420 	if (flags & MSG_OOB)
1421 		goto recv_urg;
1422 
1423 	seq = &tp->copied_seq;
1424 	if (flags & MSG_PEEK) {
1425 		peek_seq = tp->copied_seq;
1426 		seq = &peek_seq;
1427 	}
1428 
1429 	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1430 
1431 #ifdef CONFIG_NET_DMA
1432 	tp->ucopy.dma_chan = NULL;
1433 	preempt_disable();
1434 	skb = skb_peek_tail(&sk->sk_receive_queue);
1435 	{
1436 		int available = 0;
1437 
1438 		if (skb)
1439 			available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1440 		if ((available < target) &&
1441 		    (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1442 		    !sysctl_tcp_low_latency &&
1443 		    dma_find_channel(DMA_MEMCPY)) {
1444 			preempt_enable_no_resched();
1445 			tp->ucopy.pinned_list =
1446 					dma_pin_iovec_pages(msg->msg_iov, len);
1447 		} else {
1448 			preempt_enable_no_resched();
1449 		}
1450 	}
1451 #endif
1452 
1453 	do {
1454 		u32 offset;
1455 
1456 		/* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1457 		if (tp->urg_data && tp->urg_seq == *seq) {
1458 			if (copied)
1459 				break;
1460 			if (signal_pending(current)) {
1461 				copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1462 				break;
1463 			}
1464 		}
1465 
1466 		/* Next get a buffer. */
1467 
1468 		skb_queue_walk(&sk->sk_receive_queue, skb) {
1469 			/* Now that we have two receive queues this
1470 			 * shouldn't happen.
1471 			 */
1472 			if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1473 			     KERN_INFO "recvmsg bug: copied %X "
1474 				       "seq %X rcvnxt %X fl %X\n", *seq,
1475 				       TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1476 				       flags))
1477 				break;
1478 
1479 			offset = *seq - TCP_SKB_CB(skb)->seq;
1480 			if (tcp_hdr(skb)->syn)
1481 				offset--;
1482 			if (offset < skb->len)
1483 				goto found_ok_skb;
1484 			if (tcp_hdr(skb)->fin)
1485 				goto found_fin_ok;
1486 			WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: "
1487 					"copied %X seq %X rcvnxt %X fl %X\n",
1488 					*seq, TCP_SKB_CB(skb)->seq,
1489 					tp->rcv_nxt, flags);
1490 		}
1491 
1492 		/* Well, if we have backlog, try to process it now yet. */
1493 
1494 		if (copied >= target && !sk->sk_backlog.tail)
1495 			break;
1496 
1497 		if (copied) {
1498 			if (sk->sk_err ||
1499 			    sk->sk_state == TCP_CLOSE ||
1500 			    (sk->sk_shutdown & RCV_SHUTDOWN) ||
1501 			    !timeo ||
1502 			    signal_pending(current))
1503 				break;
1504 		} else {
1505 			if (sock_flag(sk, SOCK_DONE))
1506 				break;
1507 
1508 			if (sk->sk_err) {
1509 				copied = sock_error(sk);
1510 				break;
1511 			}
1512 
1513 			if (sk->sk_shutdown & RCV_SHUTDOWN)
1514 				break;
1515 
1516 			if (sk->sk_state == TCP_CLOSE) {
1517 				if (!sock_flag(sk, SOCK_DONE)) {
1518 					/* This occurs when user tries to read
1519 					 * from never connected socket.
1520 					 */
1521 					copied = -ENOTCONN;
1522 					break;
1523 				}
1524 				break;
1525 			}
1526 
1527 			if (!timeo) {
1528 				copied = -EAGAIN;
1529 				break;
1530 			}
1531 
1532 			if (signal_pending(current)) {
1533 				copied = sock_intr_errno(timeo);
1534 				break;
1535 			}
1536 		}
1537 
1538 		tcp_cleanup_rbuf(sk, copied);
1539 
1540 		if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1541 			/* Install new reader */
1542 			if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1543 				user_recv = current;
1544 				tp->ucopy.task = user_recv;
1545 				tp->ucopy.iov = msg->msg_iov;
1546 			}
1547 
1548 			tp->ucopy.len = len;
1549 
1550 			WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1551 				!(flags & (MSG_PEEK | MSG_TRUNC)));
1552 
1553 			/* Ugly... If prequeue is not empty, we have to
1554 			 * process it before releasing socket, otherwise
1555 			 * order will be broken at second iteration.
1556 			 * More elegant solution is required!!!
1557 			 *
1558 			 * Look: we have the following (pseudo)queues:
1559 			 *
1560 			 * 1. packets in flight
1561 			 * 2. backlog
1562 			 * 3. prequeue
1563 			 * 4. receive_queue
1564 			 *
1565 			 * Each queue can be processed only if the next ones
1566 			 * are empty. At this point we have empty receive_queue.
1567 			 * But prequeue _can_ be not empty after 2nd iteration,
1568 			 * when we jumped to start of loop because backlog
1569 			 * processing added something to receive_queue.
1570 			 * We cannot release_sock(), because backlog contains
1571 			 * packets arrived _after_ prequeued ones.
1572 			 *
1573 			 * Shortly, algorithm is clear --- to process all
1574 			 * the queues in order. We could make it more directly,
1575 			 * requeueing packets from backlog to prequeue, if
1576 			 * is not empty. It is more elegant, but eats cycles,
1577 			 * unfortunately.
1578 			 */
1579 			if (!skb_queue_empty(&tp->ucopy.prequeue))
1580 				goto do_prequeue;
1581 
1582 			/* __ Set realtime policy in scheduler __ */
1583 		}
1584 
1585 #ifdef CONFIG_NET_DMA
1586 		if (tp->ucopy.dma_chan)
1587 			dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1588 #endif
1589 		if (copied >= target) {
1590 			/* Do not sleep, just process backlog. */
1591 			release_sock(sk);
1592 			lock_sock(sk);
1593 		} else
1594 			sk_wait_data(sk, &timeo);
1595 
1596 #ifdef CONFIG_NET_DMA
1597 		tcp_service_net_dma(sk, false);  /* Don't block */
1598 		tp->ucopy.wakeup = 0;
1599 #endif
1600 
1601 		if (user_recv) {
1602 			int chunk;
1603 
1604 			/* __ Restore normal policy in scheduler __ */
1605 
1606 			if ((chunk = len - tp->ucopy.len) != 0) {
1607 				NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1608 				len -= chunk;
1609 				copied += chunk;
1610 			}
1611 
1612 			if (tp->rcv_nxt == tp->copied_seq &&
1613 			    !skb_queue_empty(&tp->ucopy.prequeue)) {
1614 do_prequeue:
1615 				tcp_prequeue_process(sk);
1616 
1617 				if ((chunk = len - tp->ucopy.len) != 0) {
1618 					NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1619 					len -= chunk;
1620 					copied += chunk;
1621 				}
1622 			}
1623 		}
1624 		if ((flags & MSG_PEEK) &&
1625 		    (peek_seq - copied - urg_hole != tp->copied_seq)) {
1626 			if (net_ratelimit())
1627 				printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1628 				       current->comm, task_pid_nr(current));
1629 			peek_seq = tp->copied_seq;
1630 		}
1631 		continue;
1632 
1633 	found_ok_skb:
1634 		/* Ok so how much can we use? */
1635 		used = skb->len - offset;
1636 		if (len < used)
1637 			used = len;
1638 
1639 		/* Do we have urgent data here? */
1640 		if (tp->urg_data) {
1641 			u32 urg_offset = tp->urg_seq - *seq;
1642 			if (urg_offset < used) {
1643 				if (!urg_offset) {
1644 					if (!sock_flag(sk, SOCK_URGINLINE)) {
1645 						++*seq;
1646 						urg_hole++;
1647 						offset++;
1648 						used--;
1649 						if (!used)
1650 							goto skip_copy;
1651 					}
1652 				} else
1653 					used = urg_offset;
1654 			}
1655 		}
1656 
1657 		if (!(flags & MSG_TRUNC)) {
1658 #ifdef CONFIG_NET_DMA
1659 			if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1660 				tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1661 
1662 			if (tp->ucopy.dma_chan) {
1663 				tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1664 					tp->ucopy.dma_chan, skb, offset,
1665 					msg->msg_iov, used,
1666 					tp->ucopy.pinned_list);
1667 
1668 				if (tp->ucopy.dma_cookie < 0) {
1669 
1670 					printk(KERN_ALERT "dma_cookie < 0\n");
1671 
1672 					/* Exception. Bailout! */
1673 					if (!copied)
1674 						copied = -EFAULT;
1675 					break;
1676 				}
1677 
1678 				dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1679 
1680 				if ((offset + used) == skb->len)
1681 					copied_early = 1;
1682 
1683 			} else
1684 #endif
1685 			{
1686 				err = skb_copy_datagram_iovec(skb, offset,
1687 						msg->msg_iov, used);
1688 				if (err) {
1689 					/* Exception. Bailout! */
1690 					if (!copied)
1691 						copied = -EFAULT;
1692 					break;
1693 				}
1694 			}
1695 		}
1696 
1697 		*seq += used;
1698 		copied += used;
1699 		len -= used;
1700 
1701 		tcp_rcv_space_adjust(sk);
1702 
1703 skip_copy:
1704 		if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1705 			tp->urg_data = 0;
1706 			tcp_fast_path_check(sk);
1707 		}
1708 		if (used + offset < skb->len)
1709 			continue;
1710 
1711 		if (tcp_hdr(skb)->fin)
1712 			goto found_fin_ok;
1713 		if (!(flags & MSG_PEEK)) {
1714 			sk_eat_skb(sk, skb, copied_early);
1715 			copied_early = 0;
1716 		}
1717 		continue;
1718 
1719 	found_fin_ok:
1720 		/* Process the FIN. */
1721 		++*seq;
1722 		if (!(flags & MSG_PEEK)) {
1723 			sk_eat_skb(sk, skb, copied_early);
1724 			copied_early = 0;
1725 		}
1726 		break;
1727 	} while (len > 0);
1728 
1729 	if (user_recv) {
1730 		if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1731 			int chunk;
1732 
1733 			tp->ucopy.len = copied > 0 ? len : 0;
1734 
1735 			tcp_prequeue_process(sk);
1736 
1737 			if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1738 				NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1739 				len -= chunk;
1740 				copied += chunk;
1741 			}
1742 		}
1743 
1744 		tp->ucopy.task = NULL;
1745 		tp->ucopy.len = 0;
1746 	}
1747 
1748 #ifdef CONFIG_NET_DMA
1749 	tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1750 	tp->ucopy.dma_chan = NULL;
1751 
1752 	if (tp->ucopy.pinned_list) {
1753 		dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1754 		tp->ucopy.pinned_list = NULL;
1755 	}
1756 #endif
1757 
1758 	/* According to UNIX98, msg_name/msg_namelen are ignored
1759 	 * on connected socket. I was just happy when found this 8) --ANK
1760 	 */
1761 
1762 	/* Clean up data we have read: This will do ACK frames. */
1763 	tcp_cleanup_rbuf(sk, copied);
1764 
1765 	TCP_CHECK_TIMER(sk);
1766 	release_sock(sk);
1767 	return copied;
1768 
1769 out:
1770 	TCP_CHECK_TIMER(sk);
1771 	release_sock(sk);
1772 	return err;
1773 
1774 recv_urg:
1775 	err = tcp_recv_urg(sk, msg, len, flags);
1776 	goto out;
1777 }
1778 
1779 void tcp_set_state(struct sock *sk, int state)
1780 {
1781 	int oldstate = sk->sk_state;
1782 
1783 	switch (state) {
1784 	case TCP_ESTABLISHED:
1785 		if (oldstate != TCP_ESTABLISHED)
1786 			TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1787 		break;
1788 
1789 	case TCP_CLOSE:
1790 		if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1791 			TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1792 
1793 		sk->sk_prot->unhash(sk);
1794 		if (inet_csk(sk)->icsk_bind_hash &&
1795 		    !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1796 			inet_put_port(sk);
1797 		/* fall through */
1798 	default:
1799 		if (oldstate == TCP_ESTABLISHED)
1800 			TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1801 	}
1802 
1803 	/* Change state AFTER socket is unhashed to avoid closed
1804 	 * socket sitting in hash tables.
1805 	 */
1806 	sk->sk_state = state;
1807 
1808 #ifdef STATE_TRACE
1809 	SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1810 #endif
1811 }
1812 EXPORT_SYMBOL_GPL(tcp_set_state);
1813 
1814 /*
1815  *	State processing on a close. This implements the state shift for
1816  *	sending our FIN frame. Note that we only send a FIN for some
1817  *	states. A shutdown() may have already sent the FIN, or we may be
1818  *	closed.
1819  */
1820 
1821 static const unsigned char new_state[16] = {
1822   /* current state:        new state:      action:	*/
1823   /* (Invalid)		*/ TCP_CLOSE,
1824   /* TCP_ESTABLISHED	*/ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1825   /* TCP_SYN_SENT	*/ TCP_CLOSE,
1826   /* TCP_SYN_RECV	*/ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1827   /* TCP_FIN_WAIT1	*/ TCP_FIN_WAIT1,
1828   /* TCP_FIN_WAIT2	*/ TCP_FIN_WAIT2,
1829   /* TCP_TIME_WAIT	*/ TCP_CLOSE,
1830   /* TCP_CLOSE		*/ TCP_CLOSE,
1831   /* TCP_CLOSE_WAIT	*/ TCP_LAST_ACK  | TCP_ACTION_FIN,
1832   /* TCP_LAST_ACK	*/ TCP_LAST_ACK,
1833   /* TCP_LISTEN		*/ TCP_CLOSE,
1834   /* TCP_CLOSING	*/ TCP_CLOSING,
1835 };
1836 
1837 static int tcp_close_state(struct sock *sk)
1838 {
1839 	int next = (int)new_state[sk->sk_state];
1840 	int ns = next & TCP_STATE_MASK;
1841 
1842 	tcp_set_state(sk, ns);
1843 
1844 	return next & TCP_ACTION_FIN;
1845 }
1846 
1847 /*
1848  *	Shutdown the sending side of a connection. Much like close except
1849  *	that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1850  */
1851 
1852 void tcp_shutdown(struct sock *sk, int how)
1853 {
1854 	/*	We need to grab some memory, and put together a FIN,
1855 	 *	and then put it into the queue to be sent.
1856 	 *		Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1857 	 */
1858 	if (!(how & SEND_SHUTDOWN))
1859 		return;
1860 
1861 	/* If we've already sent a FIN, or it's a closed state, skip this. */
1862 	if ((1 << sk->sk_state) &
1863 	    (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1864 	     TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1865 		/* Clear out any half completed packets.  FIN if needed. */
1866 		if (tcp_close_state(sk))
1867 			tcp_send_fin(sk);
1868 	}
1869 }
1870 
1871 void tcp_close(struct sock *sk, long timeout)
1872 {
1873 	struct sk_buff *skb;
1874 	int data_was_unread = 0;
1875 	int state;
1876 
1877 	lock_sock(sk);
1878 	sk->sk_shutdown = SHUTDOWN_MASK;
1879 
1880 	if (sk->sk_state == TCP_LISTEN) {
1881 		tcp_set_state(sk, TCP_CLOSE);
1882 
1883 		/* Special case. */
1884 		inet_csk_listen_stop(sk);
1885 
1886 		goto adjudge_to_death;
1887 	}
1888 
1889 	/*  We need to flush the recv. buffs.  We do this only on the
1890 	 *  descriptor close, not protocol-sourced closes, because the
1891 	 *  reader process may not have drained the data yet!
1892 	 */
1893 	while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1894 		u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1895 			  tcp_hdr(skb)->fin;
1896 		data_was_unread += len;
1897 		__kfree_skb(skb);
1898 	}
1899 
1900 	sk_mem_reclaim(sk);
1901 
1902 	/* As outlined in RFC 2525, section 2.17, we send a RST here because
1903 	 * data was lost. To witness the awful effects of the old behavior of
1904 	 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1905 	 * GET in an FTP client, suspend the process, wait for the client to
1906 	 * advertise a zero window, then kill -9 the FTP client, wheee...
1907 	 * Note: timeout is always zero in such a case.
1908 	 */
1909 	if (data_was_unread) {
1910 		/* Unread data was tossed, zap the connection. */
1911 		NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1912 		tcp_set_state(sk, TCP_CLOSE);
1913 		tcp_send_active_reset(sk, sk->sk_allocation);
1914 	} else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1915 		/* Check zero linger _after_ checking for unread data. */
1916 		sk->sk_prot->disconnect(sk, 0);
1917 		NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1918 	} else if (tcp_close_state(sk)) {
1919 		/* We FIN if the application ate all the data before
1920 		 * zapping the connection.
1921 		 */
1922 
1923 		/* RED-PEN. Formally speaking, we have broken TCP state
1924 		 * machine. State transitions:
1925 		 *
1926 		 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1927 		 * TCP_SYN_RECV	-> TCP_FIN_WAIT1 (forget it, it's impossible)
1928 		 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1929 		 *
1930 		 * are legal only when FIN has been sent (i.e. in window),
1931 		 * rather than queued out of window. Purists blame.
1932 		 *
1933 		 * F.e. "RFC state" is ESTABLISHED,
1934 		 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1935 		 *
1936 		 * The visible declinations are that sometimes
1937 		 * we enter time-wait state, when it is not required really
1938 		 * (harmless), do not send active resets, when they are
1939 		 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1940 		 * they look as CLOSING or LAST_ACK for Linux)
1941 		 * Probably, I missed some more holelets.
1942 		 * 						--ANK
1943 		 */
1944 		tcp_send_fin(sk);
1945 	}
1946 
1947 	sk_stream_wait_close(sk, timeout);
1948 
1949 adjudge_to_death:
1950 	state = sk->sk_state;
1951 	sock_hold(sk);
1952 	sock_orphan(sk);
1953 
1954 	/* It is the last release_sock in its life. It will remove backlog. */
1955 	release_sock(sk);
1956 
1957 
1958 	/* Now socket is owned by kernel and we acquire BH lock
1959 	   to finish close. No need to check for user refs.
1960 	 */
1961 	local_bh_disable();
1962 	bh_lock_sock(sk);
1963 	WARN_ON(sock_owned_by_user(sk));
1964 
1965 	percpu_counter_inc(sk->sk_prot->orphan_count);
1966 
1967 	/* Have we already been destroyed by a softirq or backlog? */
1968 	if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1969 		goto out;
1970 
1971 	/*	This is a (useful) BSD violating of the RFC. There is a
1972 	 *	problem with TCP as specified in that the other end could
1973 	 *	keep a socket open forever with no application left this end.
1974 	 *	We use a 3 minute timeout (about the same as BSD) then kill
1975 	 *	our end. If they send after that then tough - BUT: long enough
1976 	 *	that we won't make the old 4*rto = almost no time - whoops
1977 	 *	reset mistake.
1978 	 *
1979 	 *	Nope, it was not mistake. It is really desired behaviour
1980 	 *	f.e. on http servers, when such sockets are useless, but
1981 	 *	consume significant resources. Let's do it with special
1982 	 *	linger2	option.					--ANK
1983 	 */
1984 
1985 	if (sk->sk_state == TCP_FIN_WAIT2) {
1986 		struct tcp_sock *tp = tcp_sk(sk);
1987 		if (tp->linger2 < 0) {
1988 			tcp_set_state(sk, TCP_CLOSE);
1989 			tcp_send_active_reset(sk, GFP_ATOMIC);
1990 			NET_INC_STATS_BH(sock_net(sk),
1991 					LINUX_MIB_TCPABORTONLINGER);
1992 		} else {
1993 			const int tmo = tcp_fin_time(sk);
1994 
1995 			if (tmo > TCP_TIMEWAIT_LEN) {
1996 				inet_csk_reset_keepalive_timer(sk,
1997 						tmo - TCP_TIMEWAIT_LEN);
1998 			} else {
1999 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2000 				goto out;
2001 			}
2002 		}
2003 	}
2004 	if (sk->sk_state != TCP_CLOSE) {
2005 		int orphan_count = percpu_counter_read_positive(
2006 						sk->sk_prot->orphan_count);
2007 
2008 		sk_mem_reclaim(sk);
2009 		if (tcp_too_many_orphans(sk, orphan_count)) {
2010 			if (net_ratelimit())
2011 				printk(KERN_INFO "TCP: too many of orphaned "
2012 				       "sockets\n");
2013 			tcp_set_state(sk, TCP_CLOSE);
2014 			tcp_send_active_reset(sk, GFP_ATOMIC);
2015 			NET_INC_STATS_BH(sock_net(sk),
2016 					LINUX_MIB_TCPABORTONMEMORY);
2017 		}
2018 	}
2019 
2020 	if (sk->sk_state == TCP_CLOSE)
2021 		inet_csk_destroy_sock(sk);
2022 	/* Otherwise, socket is reprieved until protocol close. */
2023 
2024 out:
2025 	bh_unlock_sock(sk);
2026 	local_bh_enable();
2027 	sock_put(sk);
2028 }
2029 
2030 /* These states need RST on ABORT according to RFC793 */
2031 
2032 static inline int tcp_need_reset(int state)
2033 {
2034 	return (1 << state) &
2035 	       (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2036 		TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2037 }
2038 
2039 int tcp_disconnect(struct sock *sk, int flags)
2040 {
2041 	struct inet_sock *inet = inet_sk(sk);
2042 	struct inet_connection_sock *icsk = inet_csk(sk);
2043 	struct tcp_sock *tp = tcp_sk(sk);
2044 	int err = 0;
2045 	int old_state = sk->sk_state;
2046 
2047 	if (old_state != TCP_CLOSE)
2048 		tcp_set_state(sk, TCP_CLOSE);
2049 
2050 	/* ABORT function of RFC793 */
2051 	if (old_state == TCP_LISTEN) {
2052 		inet_csk_listen_stop(sk);
2053 	} else if (tcp_need_reset(old_state) ||
2054 		   (tp->snd_nxt != tp->write_seq &&
2055 		    (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2056 		/* The last check adjusts for discrepancy of Linux wrt. RFC
2057 		 * states
2058 		 */
2059 		tcp_send_active_reset(sk, gfp_any());
2060 		sk->sk_err = ECONNRESET;
2061 	} else if (old_state == TCP_SYN_SENT)
2062 		sk->sk_err = ECONNRESET;
2063 
2064 	tcp_clear_xmit_timers(sk);
2065 	__skb_queue_purge(&sk->sk_receive_queue);
2066 	tcp_write_queue_purge(sk);
2067 	__skb_queue_purge(&tp->out_of_order_queue);
2068 #ifdef CONFIG_NET_DMA
2069 	__skb_queue_purge(&sk->sk_async_wait_queue);
2070 #endif
2071 
2072 	inet->inet_dport = 0;
2073 
2074 	if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2075 		inet_reset_saddr(sk);
2076 
2077 	sk->sk_shutdown = 0;
2078 	sock_reset_flag(sk, SOCK_DONE);
2079 	tp->srtt = 0;
2080 	if ((tp->write_seq += tp->max_window + 2) == 0)
2081 		tp->write_seq = 1;
2082 	icsk->icsk_backoff = 0;
2083 	tp->snd_cwnd = 2;
2084 	icsk->icsk_probes_out = 0;
2085 	tp->packets_out = 0;
2086 	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2087 	tp->snd_cwnd_cnt = 0;
2088 	tp->bytes_acked = 0;
2089 	tp->window_clamp = 0;
2090 	tcp_set_ca_state(sk, TCP_CA_Open);
2091 	tcp_clear_retrans(tp);
2092 	inet_csk_delack_init(sk);
2093 	tcp_init_send_head(sk);
2094 	memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2095 	__sk_dst_reset(sk);
2096 
2097 	WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2098 
2099 	sk->sk_error_report(sk);
2100 	return err;
2101 }
2102 
2103 /*
2104  *	Socket option code for TCP.
2105  */
2106 static int do_tcp_setsockopt(struct sock *sk, int level,
2107 		int optname, char __user *optval, unsigned int optlen)
2108 {
2109 	struct tcp_sock *tp = tcp_sk(sk);
2110 	struct inet_connection_sock *icsk = inet_csk(sk);
2111 	int val;
2112 	int err = 0;
2113 
2114 	/* These are data/string values, all the others are ints */
2115 	switch (optname) {
2116 	case TCP_CONGESTION: {
2117 		char name[TCP_CA_NAME_MAX];
2118 
2119 		if (optlen < 1)
2120 			return -EINVAL;
2121 
2122 		val = strncpy_from_user(name, optval,
2123 					min_t(long, TCP_CA_NAME_MAX-1, optlen));
2124 		if (val < 0)
2125 			return -EFAULT;
2126 		name[val] = 0;
2127 
2128 		lock_sock(sk);
2129 		err = tcp_set_congestion_control(sk, name);
2130 		release_sock(sk);
2131 		return err;
2132 	}
2133 	case TCP_COOKIE_TRANSACTIONS: {
2134 		struct tcp_cookie_transactions ctd;
2135 		struct tcp_cookie_values *cvp = NULL;
2136 
2137 		if (sizeof(ctd) > optlen)
2138 			return -EINVAL;
2139 		if (copy_from_user(&ctd, optval, sizeof(ctd)))
2140 			return -EFAULT;
2141 
2142 		if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2143 		    ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2144 			return -EINVAL;
2145 
2146 		if (ctd.tcpct_cookie_desired == 0) {
2147 			/* default to global value */
2148 		} else if ((0x1 & ctd.tcpct_cookie_desired) ||
2149 			   ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2150 			   ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2151 			return -EINVAL;
2152 		}
2153 
2154 		if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2155 			/* Supercedes all other values */
2156 			lock_sock(sk);
2157 			if (tp->cookie_values != NULL) {
2158 				kref_put(&tp->cookie_values->kref,
2159 					 tcp_cookie_values_release);
2160 				tp->cookie_values = NULL;
2161 			}
2162 			tp->rx_opt.cookie_in_always = 0; /* false */
2163 			tp->rx_opt.cookie_out_never = 1; /* true */
2164 			release_sock(sk);
2165 			return err;
2166 		}
2167 
2168 		/* Allocate ancillary memory before locking.
2169 		 */
2170 		if (ctd.tcpct_used > 0 ||
2171 		    (tp->cookie_values == NULL &&
2172 		     (sysctl_tcp_cookie_size > 0 ||
2173 		      ctd.tcpct_cookie_desired > 0 ||
2174 		      ctd.tcpct_s_data_desired > 0))) {
2175 			cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2176 				      GFP_KERNEL);
2177 			if (cvp == NULL)
2178 				return -ENOMEM;
2179 		}
2180 		lock_sock(sk);
2181 		tp->rx_opt.cookie_in_always =
2182 			(TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2183 		tp->rx_opt.cookie_out_never = 0; /* false */
2184 
2185 		if (tp->cookie_values != NULL) {
2186 			if (cvp != NULL) {
2187 				/* Changed values are recorded by a changed
2188 				 * pointer, ensuring the cookie will differ,
2189 				 * without separately hashing each value later.
2190 				 */
2191 				kref_put(&tp->cookie_values->kref,
2192 					 tcp_cookie_values_release);
2193 				kref_init(&cvp->kref);
2194 				tp->cookie_values = cvp;
2195 			} else {
2196 				cvp = tp->cookie_values;
2197 			}
2198 		}
2199 		if (cvp != NULL) {
2200 			cvp->cookie_desired = ctd.tcpct_cookie_desired;
2201 
2202 			if (ctd.tcpct_used > 0) {
2203 				memcpy(cvp->s_data_payload, ctd.tcpct_value,
2204 				       ctd.tcpct_used);
2205 				cvp->s_data_desired = ctd.tcpct_used;
2206 				cvp->s_data_constant = 1; /* true */
2207 			} else {
2208 				/* No constant payload data. */
2209 				cvp->s_data_desired = ctd.tcpct_s_data_desired;
2210 				cvp->s_data_constant = 0; /* false */
2211 			}
2212 		}
2213 		release_sock(sk);
2214 		return err;
2215 	}
2216 	default:
2217 		/* fallthru */
2218 		break;
2219 	}
2220 
2221 	if (optlen < sizeof(int))
2222 		return -EINVAL;
2223 
2224 	if (get_user(val, (int __user *)optval))
2225 		return -EFAULT;
2226 
2227 	lock_sock(sk);
2228 
2229 	switch (optname) {
2230 	case TCP_MAXSEG:
2231 		/* Values greater than interface MTU won't take effect. However
2232 		 * at the point when this call is done we typically don't yet
2233 		 * know which interface is going to be used */
2234 		if (val < 8 || val > MAX_TCP_WINDOW) {
2235 			err = -EINVAL;
2236 			break;
2237 		}
2238 		tp->rx_opt.user_mss = val;
2239 		break;
2240 
2241 	case TCP_NODELAY:
2242 		if (val) {
2243 			/* TCP_NODELAY is weaker than TCP_CORK, so that
2244 			 * this option on corked socket is remembered, but
2245 			 * it is not activated until cork is cleared.
2246 			 *
2247 			 * However, when TCP_NODELAY is set we make
2248 			 * an explicit push, which overrides even TCP_CORK
2249 			 * for currently queued segments.
2250 			 */
2251 			tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2252 			tcp_push_pending_frames(sk);
2253 		} else {
2254 			tp->nonagle &= ~TCP_NAGLE_OFF;
2255 		}
2256 		break;
2257 
2258 	case TCP_THIN_LINEAR_TIMEOUTS:
2259 		if (val < 0 || val > 1)
2260 			err = -EINVAL;
2261 		else
2262 			tp->thin_lto = val;
2263 		break;
2264 
2265 	case TCP_THIN_DUPACK:
2266 		if (val < 0 || val > 1)
2267 			err = -EINVAL;
2268 		else
2269 			tp->thin_dupack = val;
2270 		break;
2271 
2272 	case TCP_CORK:
2273 		/* When set indicates to always queue non-full frames.
2274 		 * Later the user clears this option and we transmit
2275 		 * any pending partial frames in the queue.  This is
2276 		 * meant to be used alongside sendfile() to get properly
2277 		 * filled frames when the user (for example) must write
2278 		 * out headers with a write() call first and then use
2279 		 * sendfile to send out the data parts.
2280 		 *
2281 		 * TCP_CORK can be set together with TCP_NODELAY and it is
2282 		 * stronger than TCP_NODELAY.
2283 		 */
2284 		if (val) {
2285 			tp->nonagle |= TCP_NAGLE_CORK;
2286 		} else {
2287 			tp->nonagle &= ~TCP_NAGLE_CORK;
2288 			if (tp->nonagle&TCP_NAGLE_OFF)
2289 				tp->nonagle |= TCP_NAGLE_PUSH;
2290 			tcp_push_pending_frames(sk);
2291 		}
2292 		break;
2293 
2294 	case TCP_KEEPIDLE:
2295 		if (val < 1 || val > MAX_TCP_KEEPIDLE)
2296 			err = -EINVAL;
2297 		else {
2298 			tp->keepalive_time = val * HZ;
2299 			if (sock_flag(sk, SOCK_KEEPOPEN) &&
2300 			    !((1 << sk->sk_state) &
2301 			      (TCPF_CLOSE | TCPF_LISTEN))) {
2302 				u32 elapsed = keepalive_time_elapsed(tp);
2303 				if (tp->keepalive_time > elapsed)
2304 					elapsed = tp->keepalive_time - elapsed;
2305 				else
2306 					elapsed = 0;
2307 				inet_csk_reset_keepalive_timer(sk, elapsed);
2308 			}
2309 		}
2310 		break;
2311 	case TCP_KEEPINTVL:
2312 		if (val < 1 || val > MAX_TCP_KEEPINTVL)
2313 			err = -EINVAL;
2314 		else
2315 			tp->keepalive_intvl = val * HZ;
2316 		break;
2317 	case TCP_KEEPCNT:
2318 		if (val < 1 || val > MAX_TCP_KEEPCNT)
2319 			err = -EINVAL;
2320 		else
2321 			tp->keepalive_probes = val;
2322 		break;
2323 	case TCP_SYNCNT:
2324 		if (val < 1 || val > MAX_TCP_SYNCNT)
2325 			err = -EINVAL;
2326 		else
2327 			icsk->icsk_syn_retries = val;
2328 		break;
2329 
2330 	case TCP_LINGER2:
2331 		if (val < 0)
2332 			tp->linger2 = -1;
2333 		else if (val > sysctl_tcp_fin_timeout / HZ)
2334 			tp->linger2 = 0;
2335 		else
2336 			tp->linger2 = val * HZ;
2337 		break;
2338 
2339 	case TCP_DEFER_ACCEPT:
2340 		/* Translate value in seconds to number of retransmits */
2341 		icsk->icsk_accept_queue.rskq_defer_accept =
2342 			secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2343 					TCP_RTO_MAX / HZ);
2344 		break;
2345 
2346 	case TCP_WINDOW_CLAMP:
2347 		if (!val) {
2348 			if (sk->sk_state != TCP_CLOSE) {
2349 				err = -EINVAL;
2350 				break;
2351 			}
2352 			tp->window_clamp = 0;
2353 		} else
2354 			tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2355 						SOCK_MIN_RCVBUF / 2 : val;
2356 		break;
2357 
2358 	case TCP_QUICKACK:
2359 		if (!val) {
2360 			icsk->icsk_ack.pingpong = 1;
2361 		} else {
2362 			icsk->icsk_ack.pingpong = 0;
2363 			if ((1 << sk->sk_state) &
2364 			    (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2365 			    inet_csk_ack_scheduled(sk)) {
2366 				icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2367 				tcp_cleanup_rbuf(sk, 1);
2368 				if (!(val & 1))
2369 					icsk->icsk_ack.pingpong = 1;
2370 			}
2371 		}
2372 		break;
2373 
2374 #ifdef CONFIG_TCP_MD5SIG
2375 	case TCP_MD5SIG:
2376 		/* Read the IP->Key mappings from userspace */
2377 		err = tp->af_specific->md5_parse(sk, optval, optlen);
2378 		break;
2379 #endif
2380 
2381 	default:
2382 		err = -ENOPROTOOPT;
2383 		break;
2384 	}
2385 
2386 	release_sock(sk);
2387 	return err;
2388 }
2389 
2390 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2391 		   unsigned int optlen)
2392 {
2393 	struct inet_connection_sock *icsk = inet_csk(sk);
2394 
2395 	if (level != SOL_TCP)
2396 		return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2397 						     optval, optlen);
2398 	return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2399 }
2400 
2401 #ifdef CONFIG_COMPAT
2402 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2403 			  char __user *optval, unsigned int optlen)
2404 {
2405 	if (level != SOL_TCP)
2406 		return inet_csk_compat_setsockopt(sk, level, optname,
2407 						  optval, optlen);
2408 	return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2409 }
2410 
2411 EXPORT_SYMBOL(compat_tcp_setsockopt);
2412 #endif
2413 
2414 /* Return information about state of tcp endpoint in API format. */
2415 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2416 {
2417 	struct tcp_sock *tp = tcp_sk(sk);
2418 	const struct inet_connection_sock *icsk = inet_csk(sk);
2419 	u32 now = tcp_time_stamp;
2420 
2421 	memset(info, 0, sizeof(*info));
2422 
2423 	info->tcpi_state = sk->sk_state;
2424 	info->tcpi_ca_state = icsk->icsk_ca_state;
2425 	info->tcpi_retransmits = icsk->icsk_retransmits;
2426 	info->tcpi_probes = icsk->icsk_probes_out;
2427 	info->tcpi_backoff = icsk->icsk_backoff;
2428 
2429 	if (tp->rx_opt.tstamp_ok)
2430 		info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2431 	if (tcp_is_sack(tp))
2432 		info->tcpi_options |= TCPI_OPT_SACK;
2433 	if (tp->rx_opt.wscale_ok) {
2434 		info->tcpi_options |= TCPI_OPT_WSCALE;
2435 		info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2436 		info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2437 	}
2438 
2439 	if (tp->ecn_flags&TCP_ECN_OK)
2440 		info->tcpi_options |= TCPI_OPT_ECN;
2441 
2442 	info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2443 	info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2444 	info->tcpi_snd_mss = tp->mss_cache;
2445 	info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2446 
2447 	if (sk->sk_state == TCP_LISTEN) {
2448 		info->tcpi_unacked = sk->sk_ack_backlog;
2449 		info->tcpi_sacked = sk->sk_max_ack_backlog;
2450 	} else {
2451 		info->tcpi_unacked = tp->packets_out;
2452 		info->tcpi_sacked = tp->sacked_out;
2453 	}
2454 	info->tcpi_lost = tp->lost_out;
2455 	info->tcpi_retrans = tp->retrans_out;
2456 	info->tcpi_fackets = tp->fackets_out;
2457 
2458 	info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2459 	info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2460 	info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2461 
2462 	info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2463 	info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2464 	info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2465 	info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2466 	info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2467 	info->tcpi_snd_cwnd = tp->snd_cwnd;
2468 	info->tcpi_advmss = tp->advmss;
2469 	info->tcpi_reordering = tp->reordering;
2470 
2471 	info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2472 	info->tcpi_rcv_space = tp->rcvq_space.space;
2473 
2474 	info->tcpi_total_retrans = tp->total_retrans;
2475 }
2476 
2477 EXPORT_SYMBOL_GPL(tcp_get_info);
2478 
2479 static int do_tcp_getsockopt(struct sock *sk, int level,
2480 		int optname, char __user *optval, int __user *optlen)
2481 {
2482 	struct inet_connection_sock *icsk = inet_csk(sk);
2483 	struct tcp_sock *tp = tcp_sk(sk);
2484 	int val, len;
2485 
2486 	if (get_user(len, optlen))
2487 		return -EFAULT;
2488 
2489 	len = min_t(unsigned int, len, sizeof(int));
2490 
2491 	if (len < 0)
2492 		return -EINVAL;
2493 
2494 	switch (optname) {
2495 	case TCP_MAXSEG:
2496 		val = tp->mss_cache;
2497 		if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2498 			val = tp->rx_opt.user_mss;
2499 		break;
2500 	case TCP_NODELAY:
2501 		val = !!(tp->nonagle&TCP_NAGLE_OFF);
2502 		break;
2503 	case TCP_CORK:
2504 		val = !!(tp->nonagle&TCP_NAGLE_CORK);
2505 		break;
2506 	case TCP_KEEPIDLE:
2507 		val = keepalive_time_when(tp) / HZ;
2508 		break;
2509 	case TCP_KEEPINTVL:
2510 		val = keepalive_intvl_when(tp) / HZ;
2511 		break;
2512 	case TCP_KEEPCNT:
2513 		val = keepalive_probes(tp);
2514 		break;
2515 	case TCP_SYNCNT:
2516 		val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2517 		break;
2518 	case TCP_LINGER2:
2519 		val = tp->linger2;
2520 		if (val >= 0)
2521 			val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2522 		break;
2523 	case TCP_DEFER_ACCEPT:
2524 		val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2525 				      TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2526 		break;
2527 	case TCP_WINDOW_CLAMP:
2528 		val = tp->window_clamp;
2529 		break;
2530 	case TCP_INFO: {
2531 		struct tcp_info info;
2532 
2533 		if (get_user(len, optlen))
2534 			return -EFAULT;
2535 
2536 		tcp_get_info(sk, &info);
2537 
2538 		len = min_t(unsigned int, len, sizeof(info));
2539 		if (put_user(len, optlen))
2540 			return -EFAULT;
2541 		if (copy_to_user(optval, &info, len))
2542 			return -EFAULT;
2543 		return 0;
2544 	}
2545 	case TCP_QUICKACK:
2546 		val = !icsk->icsk_ack.pingpong;
2547 		break;
2548 
2549 	case TCP_CONGESTION:
2550 		if (get_user(len, optlen))
2551 			return -EFAULT;
2552 		len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2553 		if (put_user(len, optlen))
2554 			return -EFAULT;
2555 		if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2556 			return -EFAULT;
2557 		return 0;
2558 
2559 	case TCP_COOKIE_TRANSACTIONS: {
2560 		struct tcp_cookie_transactions ctd;
2561 		struct tcp_cookie_values *cvp = tp->cookie_values;
2562 
2563 		if (get_user(len, optlen))
2564 			return -EFAULT;
2565 		if (len < sizeof(ctd))
2566 			return -EINVAL;
2567 
2568 		memset(&ctd, 0, sizeof(ctd));
2569 		ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2570 				   TCP_COOKIE_IN_ALWAYS : 0)
2571 				| (tp->rx_opt.cookie_out_never ?
2572 				   TCP_COOKIE_OUT_NEVER : 0);
2573 
2574 		if (cvp != NULL) {
2575 			ctd.tcpct_flags |= (cvp->s_data_in ?
2576 					    TCP_S_DATA_IN : 0)
2577 					 | (cvp->s_data_out ?
2578 					    TCP_S_DATA_OUT : 0);
2579 
2580 			ctd.tcpct_cookie_desired = cvp->cookie_desired;
2581 			ctd.tcpct_s_data_desired = cvp->s_data_desired;
2582 
2583 			memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2584 			       cvp->cookie_pair_size);
2585 			ctd.tcpct_used = cvp->cookie_pair_size;
2586 		}
2587 
2588 		if (put_user(sizeof(ctd), optlen))
2589 			return -EFAULT;
2590 		if (copy_to_user(optval, &ctd, sizeof(ctd)))
2591 			return -EFAULT;
2592 		return 0;
2593 	}
2594 	default:
2595 		return -ENOPROTOOPT;
2596 	}
2597 
2598 	if (put_user(len, optlen))
2599 		return -EFAULT;
2600 	if (copy_to_user(optval, &val, len))
2601 		return -EFAULT;
2602 	return 0;
2603 }
2604 
2605 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2606 		   int __user *optlen)
2607 {
2608 	struct inet_connection_sock *icsk = inet_csk(sk);
2609 
2610 	if (level != SOL_TCP)
2611 		return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2612 						     optval, optlen);
2613 	return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2614 }
2615 
2616 #ifdef CONFIG_COMPAT
2617 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2618 			  char __user *optval, int __user *optlen)
2619 {
2620 	if (level != SOL_TCP)
2621 		return inet_csk_compat_getsockopt(sk, level, optname,
2622 						  optval, optlen);
2623 	return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2624 }
2625 
2626 EXPORT_SYMBOL(compat_tcp_getsockopt);
2627 #endif
2628 
2629 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2630 {
2631 	struct sk_buff *segs = ERR_PTR(-EINVAL);
2632 	struct tcphdr *th;
2633 	unsigned thlen;
2634 	unsigned int seq;
2635 	__be32 delta;
2636 	unsigned int oldlen;
2637 	unsigned int mss;
2638 
2639 	if (!pskb_may_pull(skb, sizeof(*th)))
2640 		goto out;
2641 
2642 	th = tcp_hdr(skb);
2643 	thlen = th->doff * 4;
2644 	if (thlen < sizeof(*th))
2645 		goto out;
2646 
2647 	if (!pskb_may_pull(skb, thlen))
2648 		goto out;
2649 
2650 	oldlen = (u16)~skb->len;
2651 	__skb_pull(skb, thlen);
2652 
2653 	mss = skb_shinfo(skb)->gso_size;
2654 	if (unlikely(skb->len <= mss))
2655 		goto out;
2656 
2657 	if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2658 		/* Packet is from an untrusted source, reset gso_segs. */
2659 		int type = skb_shinfo(skb)->gso_type;
2660 
2661 		if (unlikely(type &
2662 			     ~(SKB_GSO_TCPV4 |
2663 			       SKB_GSO_DODGY |
2664 			       SKB_GSO_TCP_ECN |
2665 			       SKB_GSO_TCPV6 |
2666 			       0) ||
2667 			     !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2668 			goto out;
2669 
2670 		skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2671 
2672 		segs = NULL;
2673 		goto out;
2674 	}
2675 
2676 	segs = skb_segment(skb, features);
2677 	if (IS_ERR(segs))
2678 		goto out;
2679 
2680 	delta = htonl(oldlen + (thlen + mss));
2681 
2682 	skb = segs;
2683 	th = tcp_hdr(skb);
2684 	seq = ntohl(th->seq);
2685 
2686 	do {
2687 		th->fin = th->psh = 0;
2688 
2689 		th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2690 				       (__force u32)delta));
2691 		if (skb->ip_summed != CHECKSUM_PARTIAL)
2692 			th->check =
2693 			     csum_fold(csum_partial(skb_transport_header(skb),
2694 						    thlen, skb->csum));
2695 
2696 		seq += mss;
2697 		skb = skb->next;
2698 		th = tcp_hdr(skb);
2699 
2700 		th->seq = htonl(seq);
2701 		th->cwr = 0;
2702 	} while (skb->next);
2703 
2704 	delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2705 		      skb->data_len);
2706 	th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2707 				(__force u32)delta));
2708 	if (skb->ip_summed != CHECKSUM_PARTIAL)
2709 		th->check = csum_fold(csum_partial(skb_transport_header(skb),
2710 						   thlen, skb->csum));
2711 
2712 out:
2713 	return segs;
2714 }
2715 EXPORT_SYMBOL(tcp_tso_segment);
2716 
2717 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2718 {
2719 	struct sk_buff **pp = NULL;
2720 	struct sk_buff *p;
2721 	struct tcphdr *th;
2722 	struct tcphdr *th2;
2723 	unsigned int len;
2724 	unsigned int thlen;
2725 	__be32 flags;
2726 	unsigned int mss = 1;
2727 	unsigned int hlen;
2728 	unsigned int off;
2729 	int flush = 1;
2730 	int i;
2731 
2732 	off = skb_gro_offset(skb);
2733 	hlen = off + sizeof(*th);
2734 	th = skb_gro_header_fast(skb, off);
2735 	if (skb_gro_header_hard(skb, hlen)) {
2736 		th = skb_gro_header_slow(skb, hlen, off);
2737 		if (unlikely(!th))
2738 			goto out;
2739 	}
2740 
2741 	thlen = th->doff * 4;
2742 	if (thlen < sizeof(*th))
2743 		goto out;
2744 
2745 	hlen = off + thlen;
2746 	if (skb_gro_header_hard(skb, hlen)) {
2747 		th = skb_gro_header_slow(skb, hlen, off);
2748 		if (unlikely(!th))
2749 			goto out;
2750 	}
2751 
2752 	skb_gro_pull(skb, thlen);
2753 
2754 	len = skb_gro_len(skb);
2755 	flags = tcp_flag_word(th);
2756 
2757 	for (; (p = *head); head = &p->next) {
2758 		if (!NAPI_GRO_CB(p)->same_flow)
2759 			continue;
2760 
2761 		th2 = tcp_hdr(p);
2762 
2763 		if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2764 			NAPI_GRO_CB(p)->same_flow = 0;
2765 			continue;
2766 		}
2767 
2768 		goto found;
2769 	}
2770 
2771 	goto out_check_final;
2772 
2773 found:
2774 	flush = NAPI_GRO_CB(p)->flush;
2775 	flush |= (__force int)(flags & TCP_FLAG_CWR);
2776 	flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2777 		  ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2778 	flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2779 	for (i = sizeof(*th); i < thlen; i += 4)
2780 		flush |= *(u32 *)((u8 *)th + i) ^
2781 			 *(u32 *)((u8 *)th2 + i);
2782 
2783 	mss = skb_shinfo(p)->gso_size;
2784 
2785 	flush |= (len - 1) >= mss;
2786 	flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2787 
2788 	if (flush || skb_gro_receive(head, skb)) {
2789 		mss = 1;
2790 		goto out_check_final;
2791 	}
2792 
2793 	p = *head;
2794 	th2 = tcp_hdr(p);
2795 	tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2796 
2797 out_check_final:
2798 	flush = len < mss;
2799 	flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2800 					TCP_FLAG_RST | TCP_FLAG_SYN |
2801 					TCP_FLAG_FIN));
2802 
2803 	if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2804 		pp = head;
2805 
2806 out:
2807 	NAPI_GRO_CB(skb)->flush |= flush;
2808 
2809 	return pp;
2810 }
2811 EXPORT_SYMBOL(tcp_gro_receive);
2812 
2813 int tcp_gro_complete(struct sk_buff *skb)
2814 {
2815 	struct tcphdr *th = tcp_hdr(skb);
2816 
2817 	skb->csum_start = skb_transport_header(skb) - skb->head;
2818 	skb->csum_offset = offsetof(struct tcphdr, check);
2819 	skb->ip_summed = CHECKSUM_PARTIAL;
2820 
2821 	skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2822 
2823 	if (th->cwr)
2824 		skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2825 
2826 	return 0;
2827 }
2828 EXPORT_SYMBOL(tcp_gro_complete);
2829 
2830 #ifdef CONFIG_TCP_MD5SIG
2831 static unsigned long tcp_md5sig_users;
2832 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool;
2833 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2834 
2835 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool)
2836 {
2837 	int cpu;
2838 	for_each_possible_cpu(cpu) {
2839 		struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2840 		if (p) {
2841 			if (p->md5_desc.tfm)
2842 				crypto_free_hash(p->md5_desc.tfm);
2843 			kfree(p);
2844 		}
2845 	}
2846 	free_percpu(pool);
2847 }
2848 
2849 void tcp_free_md5sig_pool(void)
2850 {
2851 	struct tcp_md5sig_pool * __percpu *pool = NULL;
2852 
2853 	spin_lock_bh(&tcp_md5sig_pool_lock);
2854 	if (--tcp_md5sig_users == 0) {
2855 		pool = tcp_md5sig_pool;
2856 		tcp_md5sig_pool = NULL;
2857 	}
2858 	spin_unlock_bh(&tcp_md5sig_pool_lock);
2859 	if (pool)
2860 		__tcp_free_md5sig_pool(pool);
2861 }
2862 
2863 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2864 
2865 static struct tcp_md5sig_pool * __percpu *
2866 __tcp_alloc_md5sig_pool(struct sock *sk)
2867 {
2868 	int cpu;
2869 	struct tcp_md5sig_pool * __percpu *pool;
2870 
2871 	pool = alloc_percpu(struct tcp_md5sig_pool *);
2872 	if (!pool)
2873 		return NULL;
2874 
2875 	for_each_possible_cpu(cpu) {
2876 		struct tcp_md5sig_pool *p;
2877 		struct crypto_hash *hash;
2878 
2879 		p = kzalloc(sizeof(*p), sk->sk_allocation);
2880 		if (!p)
2881 			goto out_free;
2882 		*per_cpu_ptr(pool, cpu) = p;
2883 
2884 		hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2885 		if (!hash || IS_ERR(hash))
2886 			goto out_free;
2887 
2888 		p->md5_desc.tfm = hash;
2889 	}
2890 	return pool;
2891 out_free:
2892 	__tcp_free_md5sig_pool(pool);
2893 	return NULL;
2894 }
2895 
2896 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2897 {
2898 	struct tcp_md5sig_pool * __percpu *pool;
2899 	int alloc = 0;
2900 
2901 retry:
2902 	spin_lock_bh(&tcp_md5sig_pool_lock);
2903 	pool = tcp_md5sig_pool;
2904 	if (tcp_md5sig_users++ == 0) {
2905 		alloc = 1;
2906 		spin_unlock_bh(&tcp_md5sig_pool_lock);
2907 	} else if (!pool) {
2908 		tcp_md5sig_users--;
2909 		spin_unlock_bh(&tcp_md5sig_pool_lock);
2910 		cpu_relax();
2911 		goto retry;
2912 	} else
2913 		spin_unlock_bh(&tcp_md5sig_pool_lock);
2914 
2915 	if (alloc) {
2916 		/* we cannot hold spinlock here because this may sleep. */
2917 		struct tcp_md5sig_pool * __percpu *p;
2918 
2919 		p = __tcp_alloc_md5sig_pool(sk);
2920 		spin_lock_bh(&tcp_md5sig_pool_lock);
2921 		if (!p) {
2922 			tcp_md5sig_users--;
2923 			spin_unlock_bh(&tcp_md5sig_pool_lock);
2924 			return NULL;
2925 		}
2926 		pool = tcp_md5sig_pool;
2927 		if (pool) {
2928 			/* oops, it has already been assigned. */
2929 			spin_unlock_bh(&tcp_md5sig_pool_lock);
2930 			__tcp_free_md5sig_pool(p);
2931 		} else {
2932 			tcp_md5sig_pool = pool = p;
2933 			spin_unlock_bh(&tcp_md5sig_pool_lock);
2934 		}
2935 	}
2936 	return pool;
2937 }
2938 
2939 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2940 
2941 
2942 /**
2943  *	tcp_get_md5sig_pool - get md5sig_pool for this user
2944  *
2945  *	We use percpu structure, so if we succeed, we exit with preemption
2946  *	and BH disabled, to make sure another thread or softirq handling
2947  *	wont try to get same context.
2948  */
2949 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2950 {
2951 	struct tcp_md5sig_pool * __percpu *p;
2952 
2953 	local_bh_disable();
2954 
2955 	spin_lock(&tcp_md5sig_pool_lock);
2956 	p = tcp_md5sig_pool;
2957 	if (p)
2958 		tcp_md5sig_users++;
2959 	spin_unlock(&tcp_md5sig_pool_lock);
2960 
2961 	if (p)
2962 		return *per_cpu_ptr(p, smp_processor_id());
2963 
2964 	local_bh_enable();
2965 	return NULL;
2966 }
2967 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2968 
2969 void tcp_put_md5sig_pool(void)
2970 {
2971 	local_bh_enable();
2972 	tcp_free_md5sig_pool();
2973 }
2974 EXPORT_SYMBOL(tcp_put_md5sig_pool);
2975 
2976 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2977 			struct tcphdr *th)
2978 {
2979 	struct scatterlist sg;
2980 	int err;
2981 
2982 	__sum16 old_checksum = th->check;
2983 	th->check = 0;
2984 	/* options aren't included in the hash */
2985 	sg_init_one(&sg, th, sizeof(struct tcphdr));
2986 	err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2987 	th->check = old_checksum;
2988 	return err;
2989 }
2990 
2991 EXPORT_SYMBOL(tcp_md5_hash_header);
2992 
2993 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2994 			  struct sk_buff *skb, unsigned header_len)
2995 {
2996 	struct scatterlist sg;
2997 	const struct tcphdr *tp = tcp_hdr(skb);
2998 	struct hash_desc *desc = &hp->md5_desc;
2999 	unsigned i;
3000 	const unsigned head_data_len = skb_headlen(skb) > header_len ?
3001 				       skb_headlen(skb) - header_len : 0;
3002 	const struct skb_shared_info *shi = skb_shinfo(skb);
3003 
3004 	sg_init_table(&sg, 1);
3005 
3006 	sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3007 	if (crypto_hash_update(desc, &sg, head_data_len))
3008 		return 1;
3009 
3010 	for (i = 0; i < shi->nr_frags; ++i) {
3011 		const struct skb_frag_struct *f = &shi->frags[i];
3012 		sg_set_page(&sg, f->page, f->size, f->page_offset);
3013 		if (crypto_hash_update(desc, &sg, f->size))
3014 			return 1;
3015 	}
3016 
3017 	return 0;
3018 }
3019 
3020 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3021 
3022 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
3023 {
3024 	struct scatterlist sg;
3025 
3026 	sg_init_one(&sg, key->key, key->keylen);
3027 	return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3028 }
3029 
3030 EXPORT_SYMBOL(tcp_md5_hash_key);
3031 
3032 #endif
3033 
3034 /**
3035  * Each Responder maintains up to two secret values concurrently for
3036  * efficient secret rollover.  Each secret value has 4 states:
3037  *
3038  * Generating.  (tcp_secret_generating != tcp_secret_primary)
3039  *    Generates new Responder-Cookies, but not yet used for primary
3040  *    verification.  This is a short-term state, typically lasting only
3041  *    one round trip time (RTT).
3042  *
3043  * Primary.  (tcp_secret_generating == tcp_secret_primary)
3044  *    Used both for generation and primary verification.
3045  *
3046  * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3047  *    Used for verification, until the first failure that can be
3048  *    verified by the newer Generating secret.  At that time, this
3049  *    cookie's state is changed to Secondary, and the Generating
3050  *    cookie's state is changed to Primary.  This is a short-term state,
3051  *    typically lasting only one round trip time (RTT).
3052  *
3053  * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3054  *    Used for secondary verification, after primary verification
3055  *    failures.  This state lasts no more than twice the Maximum Segment
3056  *    Lifetime (2MSL).  Then, the secret is discarded.
3057  */
3058 struct tcp_cookie_secret {
3059 	/* The secret is divided into two parts.  The digest part is the
3060 	 * equivalent of previously hashing a secret and saving the state,
3061 	 * and serves as an initialization vector (IV).  The message part
3062 	 * serves as the trailing secret.
3063 	 */
3064 	u32				secrets[COOKIE_WORKSPACE_WORDS];
3065 	unsigned long			expires;
3066 };
3067 
3068 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3069 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3070 #define TCP_SECRET_LIFE (HZ * 600)
3071 
3072 static struct tcp_cookie_secret tcp_secret_one;
3073 static struct tcp_cookie_secret tcp_secret_two;
3074 
3075 /* Essentially a circular list, without dynamic allocation. */
3076 static struct tcp_cookie_secret *tcp_secret_generating;
3077 static struct tcp_cookie_secret *tcp_secret_primary;
3078 static struct tcp_cookie_secret *tcp_secret_retiring;
3079 static struct tcp_cookie_secret *tcp_secret_secondary;
3080 
3081 static DEFINE_SPINLOCK(tcp_secret_locker);
3082 
3083 /* Select a pseudo-random word in the cookie workspace.
3084  */
3085 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3086 {
3087 	return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3088 }
3089 
3090 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3091  * Called in softirq context.
3092  * Returns: 0 for success.
3093  */
3094 int tcp_cookie_generator(u32 *bakery)
3095 {
3096 	unsigned long jiffy = jiffies;
3097 
3098 	if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3099 		spin_lock_bh(&tcp_secret_locker);
3100 		if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3101 			/* refreshed by another */
3102 			memcpy(bakery,
3103 			       &tcp_secret_generating->secrets[0],
3104 			       COOKIE_WORKSPACE_WORDS);
3105 		} else {
3106 			/* still needs refreshing */
3107 			get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3108 
3109 			/* The first time, paranoia assumes that the
3110 			 * randomization function isn't as strong.  But,
3111 			 * this secret initialization is delayed until
3112 			 * the last possible moment (packet arrival).
3113 			 * Although that time is observable, it is
3114 			 * unpredictably variable.  Mash in the most
3115 			 * volatile clock bits available, and expire the
3116 			 * secret extra quickly.
3117 			 */
3118 			if (unlikely(tcp_secret_primary->expires ==
3119 				     tcp_secret_secondary->expires)) {
3120 				struct timespec tv;
3121 
3122 				getnstimeofday(&tv);
3123 				bakery[COOKIE_DIGEST_WORDS+0] ^=
3124 					(u32)tv.tv_nsec;
3125 
3126 				tcp_secret_secondary->expires = jiffy
3127 					+ TCP_SECRET_1MSL
3128 					+ (0x0f & tcp_cookie_work(bakery, 0));
3129 			} else {
3130 				tcp_secret_secondary->expires = jiffy
3131 					+ TCP_SECRET_LIFE
3132 					+ (0xff & tcp_cookie_work(bakery, 1));
3133 				tcp_secret_primary->expires = jiffy
3134 					+ TCP_SECRET_2MSL
3135 					+ (0x1f & tcp_cookie_work(bakery, 2));
3136 			}
3137 			memcpy(&tcp_secret_secondary->secrets[0],
3138 			       bakery, COOKIE_WORKSPACE_WORDS);
3139 
3140 			rcu_assign_pointer(tcp_secret_generating,
3141 					   tcp_secret_secondary);
3142 			rcu_assign_pointer(tcp_secret_retiring,
3143 					   tcp_secret_primary);
3144 			/*
3145 			 * Neither call_rcu() nor synchronize_rcu() needed.
3146 			 * Retiring data is not freed.  It is replaced after
3147 			 * further (locked) pointer updates, and a quiet time
3148 			 * (minimum 1MSL, maximum LIFE - 2MSL).
3149 			 */
3150 		}
3151 		spin_unlock_bh(&tcp_secret_locker);
3152 	} else {
3153 		rcu_read_lock_bh();
3154 		memcpy(bakery,
3155 		       &rcu_dereference(tcp_secret_generating)->secrets[0],
3156 		       COOKIE_WORKSPACE_WORDS);
3157 		rcu_read_unlock_bh();
3158 	}
3159 	return 0;
3160 }
3161 EXPORT_SYMBOL(tcp_cookie_generator);
3162 
3163 void tcp_done(struct sock *sk)
3164 {
3165 	if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3166 		TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3167 
3168 	tcp_set_state(sk, TCP_CLOSE);
3169 	tcp_clear_xmit_timers(sk);
3170 
3171 	sk->sk_shutdown = SHUTDOWN_MASK;
3172 
3173 	if (!sock_flag(sk, SOCK_DEAD))
3174 		sk->sk_state_change(sk);
3175 	else
3176 		inet_csk_destroy_sock(sk);
3177 }
3178 EXPORT_SYMBOL_GPL(tcp_done);
3179 
3180 extern struct tcp_congestion_ops tcp_reno;
3181 
3182 static __initdata unsigned long thash_entries;
3183 static int __init set_thash_entries(char *str)
3184 {
3185 	if (!str)
3186 		return 0;
3187 	thash_entries = simple_strtoul(str, &str, 0);
3188 	return 1;
3189 }
3190 __setup("thash_entries=", set_thash_entries);
3191 
3192 void __init tcp_init(void)
3193 {
3194 	struct sk_buff *skb = NULL;
3195 	unsigned long nr_pages, limit;
3196 	int order, i, max_share;
3197 	unsigned long jiffy = jiffies;
3198 
3199 	BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3200 
3201 	percpu_counter_init(&tcp_sockets_allocated, 0);
3202 	percpu_counter_init(&tcp_orphan_count, 0);
3203 	tcp_hashinfo.bind_bucket_cachep =
3204 		kmem_cache_create("tcp_bind_bucket",
3205 				  sizeof(struct inet_bind_bucket), 0,
3206 				  SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3207 
3208 	/* Size and allocate the main established and bind bucket
3209 	 * hash tables.
3210 	 *
3211 	 * The methodology is similar to that of the buffer cache.
3212 	 */
3213 	tcp_hashinfo.ehash =
3214 		alloc_large_system_hash("TCP established",
3215 					sizeof(struct inet_ehash_bucket),
3216 					thash_entries,
3217 					(totalram_pages >= 128 * 1024) ?
3218 					13 : 15,
3219 					0,
3220 					NULL,
3221 					&tcp_hashinfo.ehash_mask,
3222 					thash_entries ? 0 : 512 * 1024);
3223 	for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3224 		INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3225 		INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3226 	}
3227 	if (inet_ehash_locks_alloc(&tcp_hashinfo))
3228 		panic("TCP: failed to alloc ehash_locks");
3229 	tcp_hashinfo.bhash =
3230 		alloc_large_system_hash("TCP bind",
3231 					sizeof(struct inet_bind_hashbucket),
3232 					tcp_hashinfo.ehash_mask + 1,
3233 					(totalram_pages >= 128 * 1024) ?
3234 					13 : 15,
3235 					0,
3236 					&tcp_hashinfo.bhash_size,
3237 					NULL,
3238 					64 * 1024);
3239 	tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3240 	for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3241 		spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3242 		INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3243 	}
3244 
3245 	/* Try to be a bit smarter and adjust defaults depending
3246 	 * on available memory.
3247 	 */
3248 	for (order = 0; ((1 << order) << PAGE_SHIFT) <
3249 			(tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
3250 			order++)
3251 		;
3252 	if (order >= 4) {
3253 		tcp_death_row.sysctl_max_tw_buckets = 180000;
3254 		sysctl_tcp_max_orphans = 4096 << (order - 4);
3255 		sysctl_max_syn_backlog = 1024;
3256 	} else if (order < 3) {
3257 		tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
3258 		sysctl_tcp_max_orphans >>= (3 - order);
3259 		sysctl_max_syn_backlog = 128;
3260 	}
3261 
3262 	/* Set the pressure threshold to be a fraction of global memory that
3263 	 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
3264 	 * memory, with a floor of 128 pages.
3265 	 */
3266 	nr_pages = totalram_pages - totalhigh_pages;
3267 	limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
3268 	limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
3269 	limit = max(limit, 128UL);
3270 	sysctl_tcp_mem[0] = limit / 4 * 3;
3271 	sysctl_tcp_mem[1] = limit;
3272 	sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3273 
3274 	/* Set per-socket limits to no more than 1/128 the pressure threshold */
3275 	limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3276 	max_share = min(4UL*1024*1024, limit);
3277 
3278 	sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3279 	sysctl_tcp_wmem[1] = 16*1024;
3280 	sysctl_tcp_wmem[2] = max(64*1024, max_share);
3281 
3282 	sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3283 	sysctl_tcp_rmem[1] = 87380;
3284 	sysctl_tcp_rmem[2] = max(87380, max_share);
3285 
3286 	printk(KERN_INFO "TCP: Hash tables configured "
3287 	       "(established %u bind %u)\n",
3288 	       tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3289 
3290 	tcp_register_congestion_control(&tcp_reno);
3291 
3292 	memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3293 	memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3294 	tcp_secret_one.expires = jiffy; /* past due */
3295 	tcp_secret_two.expires = jiffy; /* past due */
3296 	tcp_secret_generating = &tcp_secret_one;
3297 	tcp_secret_primary = &tcp_secret_one;
3298 	tcp_secret_retiring = &tcp_secret_two;
3299 	tcp_secret_secondary = &tcp_secret_two;
3300 }
3301 
3302 EXPORT_SYMBOL(tcp_close);
3303 EXPORT_SYMBOL(tcp_disconnect);
3304 EXPORT_SYMBOL(tcp_getsockopt);
3305 EXPORT_SYMBOL(tcp_ioctl);
3306 EXPORT_SYMBOL(tcp_poll);
3307 EXPORT_SYMBOL(tcp_read_sock);
3308 EXPORT_SYMBOL(tcp_recvmsg);
3309 EXPORT_SYMBOL(tcp_sendmsg);
3310 EXPORT_SYMBOL(tcp_splice_read);
3311 EXPORT_SYMBOL(tcp_sendpage);
3312 EXPORT_SYMBOL(tcp_setsockopt);
3313 EXPORT_SYMBOL(tcp_shutdown);
3314