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.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp.c,v 1.216 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Alan Cox : Numerous verify_area() calls
24 * Alan Cox : Set the ACK bit on a reset
25 * Alan Cox : Stopped it crashing if it closed while
26 * sk->inuse=1 and was trying to connect
28 * Alan Cox : All icmp error handling was broken
29 * pointers passed where wrong and the
30 * socket was looked up backwards. Nobody
31 * tested any icmp error code obviously.
32 * Alan Cox : tcp_err() now handled properly. It
33 * wakes people on errors. poll
34 * behaves and the icmp error race
35 * has gone by moving it into sock.c
36 * Alan Cox : tcp_send_reset() fixed to work for
37 * everything not just packets for
39 * Alan Cox : tcp option processing.
40 * Alan Cox : Reset tweaked (still not 100%) [Had
42 * Herp Rosmanith : More reset fixes
43 * Alan Cox : No longer acks invalid rst frames.
44 * Acking any kind of RST is right out.
45 * Alan Cox : Sets an ignore me flag on an rst
46 * receive otherwise odd bits of prattle
48 * Alan Cox : Fixed another acking RST frame bug.
49 * Should stop LAN workplace lockups.
50 * Alan Cox : Some tidyups using the new skb list
52 * Alan Cox : sk->keepopen now seems to work
53 * Alan Cox : Pulls options out correctly on accepts
54 * Alan Cox : Fixed assorted sk->rqueue->next errors
55 * Alan Cox : PSH doesn't end a TCP read. Switched a
57 * Alan Cox : Tidied tcp_data to avoid a potential
59 * Alan Cox : Added some better commenting, as the
60 * tcp is hard to follow
61 * Alan Cox : Removed incorrect check for 20 * psh
62 * Michael O'Reilly : ack < copied bug fix.
63 * Johannes Stille : Misc tcp fixes (not all in yet).
64 * Alan Cox : FIN with no memory -> CRASH
65 * Alan Cox : Added socket option proto entries.
66 * Also added awareness of them to accept.
67 * Alan Cox : Added TCP options (SOL_TCP)
68 * Alan Cox : Switched wakeup calls to callbacks,
69 * so the kernel can layer network
71 * Alan Cox : Use ip_tos/ip_ttl settings.
72 * Alan Cox : Handle FIN (more) properly (we hope).
73 * Alan Cox : RST frames sent on unsynchronised
75 * Alan Cox : Put in missing check for SYN bit.
76 * Alan Cox : Added tcp_select_window() aka NET2E
77 * window non shrink trick.
78 * Alan Cox : Added a couple of small NET2E timer
80 * Charles Hedrick : TCP fixes
81 * Toomas Tamm : TCP window fixes
82 * Alan Cox : Small URG fix to rlogin ^C ack fight
83 * Charles Hedrick : Rewrote most of it to actually work
84 * Linus : Rewrote tcp_read() and URG handling
86 * Gerhard Koerting: Fixed some missing timer handling
87 * Matthew Dillon : Reworked TCP machine states as per RFC
88 * Gerhard Koerting: PC/TCP workarounds
89 * Adam Caldwell : Assorted timer/timing errors
90 * Matthew Dillon : Fixed another RST bug
91 * Alan Cox : Move to kernel side addressing changes.
92 * Alan Cox : Beginning work on TCP fastpathing
94 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
95 * Alan Cox : TCP fast path debugging
96 * Alan Cox : Window clamping
97 * Michael Riepe : Bug in tcp_check()
98 * Matt Dillon : More TCP improvements and RST bug fixes
99 * Matt Dillon : Yet more small nasties remove from the
100 * TCP code (Be very nice to this man if
101 * tcp finally works 100%) 8)
102 * Alan Cox : BSD accept semantics.
103 * Alan Cox : Reset on closedown bug.
104 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
105 * Michael Pall : Handle poll() after URG properly in
107 * Michael Pall : Undo the last fix in tcp_read_urg()
108 * (multi URG PUSH broke rlogin).
109 * Michael Pall : Fix the multi URG PUSH problem in
110 * tcp_readable(), poll() after URG
112 * Michael Pall : recv(...,MSG_OOB) never blocks in the
114 * Alan Cox : Changed the semantics of sk->socket to
115 * fix a race and a signal problem with
116 * accept() and async I/O.
117 * Alan Cox : Relaxed the rules on tcp_sendto().
118 * Yury Shevchuk : Really fixed accept() blocking problem.
119 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
120 * clients/servers which listen in on
122 * Alan Cox : Cleaned the above up and shrank it to
123 * a sensible code size.
124 * Alan Cox : Self connect lockup fix.
125 * Alan Cox : No connect to multicast.
126 * Ross Biro : Close unaccepted children on master
128 * Alan Cox : Reset tracing code.
129 * Alan Cox : Spurious resets on shutdown.
130 * Alan Cox : Giant 15 minute/60 second timer error
131 * Alan Cox : Small whoops in polling before an
133 * Alan Cox : Kept the state trace facility since
134 * it's handy for debugging.
135 * Alan Cox : More reset handler fixes.
136 * Alan Cox : Started rewriting the code based on
137 * the RFC's for other useful protocol
138 * references see: Comer, KA9Q NOS, and
139 * for a reference on the difference
140 * between specifications and how BSD
141 * works see the 4.4lite source.
142 * A.N.Kuznetsov : Don't time wait on completion of tidy
144 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
145 * Linus Torvalds : Fixed BSD port reuse to work first syn
146 * Alan Cox : Reimplemented timers as per the RFC
147 * and using multiple timers for sanity.
148 * Alan Cox : Small bug fixes, and a lot of new
150 * Alan Cox : Fixed dual reader crash by locking
151 * the buffers (much like datagram.c)
152 * Alan Cox : Fixed stuck sockets in probe. A probe
153 * now gets fed up of retrying without
154 * (even a no space) answer.
155 * Alan Cox : Extracted closing code better
156 * Alan Cox : Fixed the closing state machine to
158 * Alan Cox : More 'per spec' fixes.
159 * Jorge Cwik : Even faster checksumming.
160 * Alan Cox : tcp_data() doesn't ack illegal PSH
161 * only frames. At least one pc tcp stack
163 * Alan Cox : Cache last socket.
164 * Alan Cox : Per route irtt.
165 * Matt Day : poll()->select() match BSD precisely on error
166 * Alan Cox : New buffers
167 * Marc Tamsky : Various sk->prot->retransmits and
168 * sk->retransmits misupdating fixed.
169 * Fixed tcp_write_timeout: stuck close,
170 * and TCP syn retries gets used now.
171 * Mark Yarvis : In tcp_read_wakeup(), don't send an
172 * ack if state is TCP_CLOSED.
173 * Alan Cox : Look up device on a retransmit - routes may
174 * change. Doesn't yet cope with MSS shrink right
176 * Marc Tamsky : Closing in closing fixes.
177 * Mike Shaver : RFC1122 verifications.
178 * Alan Cox : rcv_saddr errors.
179 * Alan Cox : Block double connect().
180 * Alan Cox : Small hooks for enSKIP.
181 * Alexey Kuznetsov: Path MTU discovery.
182 * Alan Cox : Support soft errors.
183 * Alan Cox : Fix MTU discovery pathological case
184 * when the remote claims no mtu!
185 * Marc Tamsky : TCP_CLOSE fix.
186 * Colin (G3TNE) : Send a reset on syn ack replies in
187 * window but wrong (fixes NT lpd problems)
188 * Pedro Roque : Better TCP window handling, delayed ack.
189 * Joerg Reuter : No modification of locked buffers in
190 * tcp_do_retransmit()
191 * Eric Schenk : Changed receiver side silly window
192 * avoidance algorithm to BSD style
193 * algorithm. This doubles throughput
194 * against machines running Solaris,
195 * and seems to result in general
197 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
198 * Willy Konynenberg : Transparent proxying support.
199 * Mike McLagan : Routing by source
200 * Keith Owens : Do proper merging with partial SKB's in
201 * tcp_do_sendmsg to avoid burstiness.
202 * Eric Schenk : Fix fast close down bug with
203 * shutdown() followed by close().
204 * Andi Kleen : Make poll agree with SIGIO
205 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
206 * lingertime == 0 (RFC 793 ABORT Call)
207 * Hirokazu Takahashi : Use copy_from_user() instead of
208 * csum_and_copy_from_user() if possible.
210 * This program is free software; you can redistribute it and/or
211 * modify it under the terms of the GNU General Public License
212 * as published by the Free Software Foundation; either version
213 * 2 of the License, or(at your option) any later version.
215 * Description of States:
217 * TCP_SYN_SENT sent a connection request, waiting for ack
219 * TCP_SYN_RECV received a connection request, sent ack,
220 * waiting for final ack in three-way handshake.
222 * TCP_ESTABLISHED connection established
224 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
225 * transmission of remaining buffered data
227 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
230 * TCP_CLOSING both sides have shutdown but we still have
231 * data we have to finish sending
233 * TCP_TIME_WAIT timeout to catch resent junk before entering
234 * closed, can only be entered from FIN_WAIT2
235 * or CLOSING. Required because the other end
236 * may not have gotten our last ACK causing it
237 * to retransmit the data packet (which we ignore)
239 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
240 * us to finish writing our data and to shutdown
241 * (we have to close() to move on to LAST_ACK)
243 * TCP_LAST_ACK out side has shutdown after remote has
244 * shutdown. There may still be data in our
245 * buffer that we have to finish sending
247 * TCP_CLOSE socket is finished
250 #include <linux/kernel.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/init.h>
256 #include <linux/fs.h>
257 #include <linux/skbuff.h>
258 #include <linux/scatterlist.h>
259 #include <linux/splice.h>
260 #include <linux/net.h>
261 #include <linux/socket.h>
262 #include <linux/random.h>
263 #include <linux/bootmem.h>
264 #include <linux/highmem.h>
265 #include <linux/swap.h>
266 #include <linux/cache.h>
267 #include <linux/err.h>
268 #include <linux/crypto.h>
270 #include <net/icmp.h>
272 #include <net/xfrm.h>
274 #include <net/netdma.h>
275 #include <net/sock.h>
277 #include <asm/uaccess.h>
278 #include <asm/ioctls.h>
280 int sysctl_tcp_fin_timeout __read_mostly
= TCP_FIN_TIMEOUT
;
282 DEFINE_SNMP_STAT(struct tcp_mib
, tcp_statistics
) __read_mostly
;
284 atomic_t tcp_orphan_count
= ATOMIC_INIT(0);
286 EXPORT_SYMBOL_GPL(tcp_orphan_count
);
288 int sysctl_tcp_mem
[3] __read_mostly
;
289 int sysctl_tcp_wmem
[3] __read_mostly
;
290 int sysctl_tcp_rmem
[3] __read_mostly
;
292 EXPORT_SYMBOL(sysctl_tcp_mem
);
293 EXPORT_SYMBOL(sysctl_tcp_rmem
);
294 EXPORT_SYMBOL(sysctl_tcp_wmem
);
296 atomic_t tcp_memory_allocated
; /* Current allocated memory. */
297 atomic_t tcp_sockets_allocated
; /* Current number of TCP sockets. */
299 EXPORT_SYMBOL(tcp_memory_allocated
);
300 EXPORT_SYMBOL(tcp_sockets_allocated
);
305 struct tcp_splice_state
{
306 struct pipe_inode_info
*pipe
;
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.
317 int tcp_memory_pressure __read_mostly
;
319 EXPORT_SYMBOL(tcp_memory_pressure
);
321 void tcp_enter_memory_pressure(void)
323 if (!tcp_memory_pressure
) {
324 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES
);
325 tcp_memory_pressure
= 1;
329 EXPORT_SYMBOL(tcp_enter_memory_pressure
);
332 * Wait for a TCP event.
334 * Note that we don't need to lock the socket, as the upper poll layers
335 * take care of normal races (between the test and the event) and we don't
336 * go look at any of the socket buffers directly.
338 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
341 struct sock
*sk
= sock
->sk
;
342 struct tcp_sock
*tp
= tcp_sk(sk
);
344 poll_wait(file
, sk
->sk_sleep
, wait
);
345 if (sk
->sk_state
== TCP_LISTEN
)
346 return inet_csk_listen_poll(sk
);
348 /* Socket is not locked. We are protected from async events
349 by poll logic and correct handling of state changes
350 made by another threads is impossible in any case.
358 * POLLHUP is certainly not done right. But poll() doesn't
359 * have a notion of HUP in just one direction, and for a
360 * socket the read side is more interesting.
362 * Some poll() documentation says that POLLHUP is incompatible
363 * with the POLLOUT/POLLWR flags, so somebody should check this
364 * all. But careful, it tends to be safer to return too many
365 * bits than too few, and you can easily break real applications
366 * if you don't tell them that something has hung up!
370 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
371 * our fs/select.c). It means that after we received EOF,
372 * poll always returns immediately, making impossible poll() on write()
373 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
374 * if and only if shutdown has been made in both directions.
375 * Actually, it is interesting to look how Solaris and DUX
376 * solve this dilemma. I would prefer, if PULLHUP were maskable,
377 * then we could set it on SND_SHUTDOWN. BTW examples given
378 * in Stevens' books assume exactly this behaviour, it explains
379 * why PULLHUP is incompatible with POLLOUT. --ANK
381 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
382 * blocking on fresh not-connected or disconnected socket. --ANK
384 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| sk
->sk_state
== TCP_CLOSE
)
386 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
387 mask
|= POLLIN
| POLLRDNORM
| POLLRDHUP
;
390 if ((1 << sk
->sk_state
) & ~(TCPF_SYN_SENT
| TCPF_SYN_RECV
)) {
391 /* Potential race condition. If read of tp below will
392 * escape above sk->sk_state, we can be illegally awaken
393 * in SYN_* states. */
394 if ((tp
->rcv_nxt
!= tp
->copied_seq
) &&
395 (tp
->urg_seq
!= tp
->copied_seq
||
396 tp
->rcv_nxt
!= tp
->copied_seq
+ 1 ||
397 sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
))
398 mask
|= POLLIN
| POLLRDNORM
;
400 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
401 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
)) {
402 mask
|= POLLOUT
| POLLWRNORM
;
403 } else { /* send SIGIO later */
404 set_bit(SOCK_ASYNC_NOSPACE
,
405 &sk
->sk_socket
->flags
);
406 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
408 /* Race breaker. If space is freed after
409 * wspace test but before the flags are set,
410 * IO signal will be lost.
412 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
))
413 mask
|= POLLOUT
| POLLWRNORM
;
417 if (tp
->urg_data
& TCP_URG_VALID
)
423 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
425 struct tcp_sock
*tp
= tcp_sk(sk
);
430 if (sk
->sk_state
== TCP_LISTEN
)
434 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
436 else if (sock_flag(sk
, SOCK_URGINLINE
) ||
438 before(tp
->urg_seq
, tp
->copied_seq
) ||
439 !before(tp
->urg_seq
, tp
->rcv_nxt
)) {
440 answ
= tp
->rcv_nxt
- tp
->copied_seq
;
442 /* Subtract 1, if FIN is in queue. */
443 if (answ
&& !skb_queue_empty(&sk
->sk_receive_queue
))
445 tcp_hdr((struct sk_buff
*)sk
->sk_receive_queue
.prev
)->fin
;
447 answ
= tp
->urg_seq
- tp
->copied_seq
;
451 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
454 if (sk
->sk_state
== TCP_LISTEN
)
457 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
460 answ
= tp
->write_seq
- tp
->snd_una
;
466 return put_user(answ
, (int __user
*)arg
);
469 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
471 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
472 tp
->pushed_seq
= tp
->write_seq
;
475 static inline int forced_push(struct tcp_sock
*tp
)
477 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
480 static inline void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
482 struct tcp_sock
*tp
= tcp_sk(sk
);
483 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
486 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
487 tcb
->flags
= TCPCB_FLAG_ACK
;
489 skb_header_release(skb
);
490 tcp_add_write_queue_tail(sk
, skb
);
491 sk
->sk_wmem_queued
+= skb
->truesize
;
492 sk_mem_charge(sk
, skb
->truesize
);
493 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
494 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
497 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
,
500 if (flags
& MSG_OOB
) {
502 tp
->snd_up
= tp
->write_seq
;
506 static inline void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
509 struct tcp_sock
*tp
= tcp_sk(sk
);
511 if (tcp_send_head(sk
)) {
512 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
513 if (!(flags
& MSG_MORE
) || forced_push(tp
))
514 tcp_mark_push(tp
, skb
);
515 tcp_mark_urg(tp
, flags
, skb
);
516 __tcp_push_pending_frames(sk
, mss_now
,
517 (flags
& MSG_MORE
) ? TCP_NAGLE_CORK
: nonagle
);
521 static int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
522 unsigned int offset
, size_t len
)
524 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
526 return skb_splice_bits(skb
, offset
, tss
->pipe
, tss
->len
, tss
->flags
);
529 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
531 /* Store TCP splice context information in read_descriptor_t. */
532 read_descriptor_t rd_desc
= {
536 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
540 * tcp_splice_read - splice data from TCP socket to a pipe
541 * @sock: socket to splice from
542 * @ppos: position (not valid)
543 * @pipe: pipe to splice to
544 * @len: number of bytes to splice
545 * @flags: splice modifier flags
548 * Will read pages from given socket and fill them into a pipe.
551 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
552 struct pipe_inode_info
*pipe
, size_t len
,
555 struct sock
*sk
= sock
->sk
;
556 struct tcp_splice_state tss
= {
566 * We can't seek on a socket input
575 timeo
= sock_rcvtimeo(sk
, flags
& SPLICE_F_NONBLOCK
);
577 ret
= __tcp_splice_read(sk
, &tss
);
583 if (flags
& SPLICE_F_NONBLOCK
) {
587 if (sock_flag(sk
, SOCK_DONE
))
590 ret
= sock_error(sk
);
593 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
595 if (sk
->sk_state
== TCP_CLOSE
) {
597 * This occurs when user tries to read
598 * from never connected socket.
600 if (!sock_flag(sk
, SOCK_DONE
))
608 sk_wait_data(sk
, &timeo
);
609 if (signal_pending(current
)) {
610 ret
= sock_intr_errno(timeo
);
621 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
622 (sk
->sk_shutdown
& RCV_SHUTDOWN
) || !timeo
||
623 signal_pending(current
))
635 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
)
639 /* The TCP header must be at least 32-bit aligned. */
640 size
= ALIGN(size
, 4);
642 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
644 if (sk_wmem_schedule(sk
, skb
->truesize
)) {
646 * Make sure that we have exactly size bytes
647 * available to the caller, no more, no less.
649 skb_reserve(skb
, skb_tailroom(skb
) - size
);
654 sk
->sk_prot
->enter_memory_pressure();
655 sk_stream_moderate_sndbuf(sk
);
660 static ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
**pages
, int poffset
,
661 size_t psize
, int flags
)
663 struct tcp_sock
*tp
= tcp_sk(sk
);
664 int mss_now
, size_goal
;
667 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
669 /* Wait for a connection to finish. */
670 if ((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
671 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
674 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
676 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
677 size_goal
= tp
->xmit_size_goal
;
681 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
685 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
686 struct page
*page
= pages
[poffset
/ PAGE_SIZE
];
687 int copy
, i
, can_coalesce
;
688 int offset
= poffset
% PAGE_SIZE
;
689 int size
= min_t(size_t, psize
, PAGE_SIZE
- offset
);
691 if (!tcp_send_head(sk
) || (copy
= size_goal
- skb
->len
) <= 0) {
693 if (!sk_stream_memory_free(sk
))
694 goto wait_for_sndbuf
;
696 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
698 goto wait_for_memory
;
707 i
= skb_shinfo(skb
)->nr_frags
;
708 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
709 if (!can_coalesce
&& i
>= MAX_SKB_FRAGS
) {
710 tcp_mark_push(tp
, skb
);
713 if (!sk_wmem_schedule(sk
, copy
))
714 goto wait_for_memory
;
717 skb_shinfo(skb
)->frags
[i
- 1].size
+= copy
;
720 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
724 skb
->data_len
+= copy
;
725 skb
->truesize
+= copy
;
726 sk
->sk_wmem_queued
+= copy
;
727 sk_mem_charge(sk
, copy
);
728 skb
->ip_summed
= CHECKSUM_PARTIAL
;
729 tp
->write_seq
+= copy
;
730 TCP_SKB_CB(skb
)->end_seq
+= copy
;
731 skb_shinfo(skb
)->gso_segs
= 0;
734 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_PSH
;
738 if (!(psize
-= copy
))
741 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
744 if (forced_push(tp
)) {
745 tcp_mark_push(tp
, skb
);
746 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
747 } else if (skb
== tcp_send_head(sk
))
748 tcp_push_one(sk
, mss_now
);
752 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
755 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
757 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
760 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
761 size_goal
= tp
->xmit_size_goal
;
766 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
773 return sk_stream_error(sk
, flags
, err
);
776 ssize_t
tcp_sendpage(struct socket
*sock
, struct page
*page
, int offset
,
777 size_t size
, int flags
)
780 struct sock
*sk
= sock
->sk
;
782 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
783 !(sk
->sk_route_caps
& NETIF_F_ALL_CSUM
))
784 return sock_no_sendpage(sock
, page
, offset
, size
, flags
);
788 res
= do_tcp_sendpages(sk
, &page
, offset
, size
, flags
);
794 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
795 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
797 static inline int select_size(struct sock
*sk
)
799 struct tcp_sock
*tp
= tcp_sk(sk
);
800 int tmp
= tp
->mss_cache
;
802 if (sk
->sk_route_caps
& NETIF_F_SG
) {
806 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
808 if (tmp
>= pgbreak
&&
809 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
817 int tcp_sendmsg(struct kiocb
*iocb
, struct socket
*sock
, struct msghdr
*msg
,
820 struct sock
*sk
= sock
->sk
;
822 struct tcp_sock
*tp
= tcp_sk(sk
);
825 int mss_now
, size_goal
;
832 flags
= msg
->msg_flags
;
833 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
835 /* Wait for a connection to finish. */
836 if ((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
837 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
840 /* This should be in poll */
841 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
843 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
844 size_goal
= tp
->xmit_size_goal
;
846 /* Ok commence sending. */
847 iovlen
= msg
->msg_iovlen
;
852 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
855 while (--iovlen
>= 0) {
856 int seglen
= iov
->iov_len
;
857 unsigned char __user
*from
= iov
->iov_base
;
864 skb
= tcp_write_queue_tail(sk
);
866 if (!tcp_send_head(sk
) ||
867 (copy
= size_goal
- skb
->len
) <= 0) {
870 /* Allocate new segment. If the interface is SG,
871 * allocate skb fitting to single page.
873 if (!sk_stream_memory_free(sk
))
874 goto wait_for_sndbuf
;
876 skb
= sk_stream_alloc_skb(sk
, select_size(sk
),
879 goto wait_for_memory
;
882 * Check whether we can use HW checksum.
884 if (sk
->sk_route_caps
& NETIF_F_ALL_CSUM
)
885 skb
->ip_summed
= CHECKSUM_PARTIAL
;
891 /* Try to append data to the end of skb. */
895 /* Where to copy to? */
896 if (skb_tailroom(skb
) > 0) {
897 /* We have some space in skb head. Superb! */
898 if (copy
> skb_tailroom(skb
))
899 copy
= skb_tailroom(skb
);
900 if ((err
= skb_add_data(skb
, from
, copy
)) != 0)
904 int i
= skb_shinfo(skb
)->nr_frags
;
905 struct page
*page
= TCP_PAGE(sk
);
906 int off
= TCP_OFF(sk
);
908 if (skb_can_coalesce(skb
, i
, page
, off
) &&
910 /* We can extend the last page
913 } else if (i
== MAX_SKB_FRAGS
||
915 !(sk
->sk_route_caps
& NETIF_F_SG
))) {
916 /* Need to add new fragment and cannot
917 * do this because interface is non-SG,
918 * or because all the page slots are
920 tcp_mark_push(tp
, skb
);
923 if (off
== PAGE_SIZE
) {
925 TCP_PAGE(sk
) = page
= NULL
;
931 if (copy
> PAGE_SIZE
- off
)
932 copy
= PAGE_SIZE
- off
;
934 if (!sk_wmem_schedule(sk
, copy
))
935 goto wait_for_memory
;
938 /* Allocate new cache page. */
939 if (!(page
= sk_stream_alloc_page(sk
)))
940 goto wait_for_memory
;
943 /* Time to copy data. We are close to
945 err
= skb_copy_to_page(sk
, from
, skb
, page
,
948 /* If this page was new, give it to the
949 * socket so it does not get leaked.
958 /* Update the skb. */
960 skb_shinfo(skb
)->frags
[i
- 1].size
+=
963 skb_fill_page_desc(skb
, i
, page
, off
, copy
);
966 } else if (off
+ copy
< PAGE_SIZE
) {
972 TCP_OFF(sk
) = off
+ copy
;
976 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_PSH
;
978 tp
->write_seq
+= copy
;
979 TCP_SKB_CB(skb
)->end_seq
+= copy
;
980 skb_shinfo(skb
)->gso_segs
= 0;
984 if ((seglen
-= copy
) == 0 && iovlen
== 0)
987 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
990 if (forced_push(tp
)) {
991 tcp_mark_push(tp
, skb
);
992 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
993 } else if (skb
== tcp_send_head(sk
))
994 tcp_push_one(sk
, mss_now
);
998 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1001 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
1003 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
1006 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
1007 size_goal
= tp
->xmit_size_goal
;
1013 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
1014 TCP_CHECK_TIMER(sk
);
1020 tcp_unlink_write_queue(skb
, sk
);
1021 /* It is the one place in all of TCP, except connection
1022 * reset, where we can be unlinking the send_head.
1024 tcp_check_send_head(sk
, skb
);
1025 sk_wmem_free_skb(sk
, skb
);
1032 err
= sk_stream_error(sk
, flags
, err
);
1033 TCP_CHECK_TIMER(sk
);
1039 * Handle reading urgent data. BSD has very simple semantics for
1040 * this, no blocking and very strange errors 8)
1043 static int tcp_recv_urg(struct sock
*sk
, long timeo
,
1044 struct msghdr
*msg
, int len
, int flags
,
1047 struct tcp_sock
*tp
= tcp_sk(sk
);
1049 /* No URG data to read. */
1050 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1051 tp
->urg_data
== TCP_URG_READ
)
1052 return -EINVAL
; /* Yes this is right ! */
1054 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1057 if (tp
->urg_data
& TCP_URG_VALID
) {
1059 char c
= tp
->urg_data
;
1061 if (!(flags
& MSG_PEEK
))
1062 tp
->urg_data
= TCP_URG_READ
;
1064 /* Read urgent data. */
1065 msg
->msg_flags
|= MSG_OOB
;
1068 if (!(flags
& MSG_TRUNC
))
1069 err
= memcpy_toiovec(msg
->msg_iov
, &c
, 1);
1072 msg
->msg_flags
|= MSG_TRUNC
;
1074 return err
? -EFAULT
: len
;
1077 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1080 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1081 * the available implementations agree in this case:
1082 * this call should never block, independent of the
1083 * blocking state of the socket.
1084 * Mike <pall@rz.uni-karlsruhe.de>
1089 /* Clean up the receive buffer for full frames taken by the user,
1090 * then send an ACK if necessary. COPIED is the number of bytes
1091 * tcp_recvmsg has given to the user so far, it speeds up the
1092 * calculation of whether or not we must ACK for the sake of
1095 void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1097 struct tcp_sock
*tp
= tcp_sk(sk
);
1098 int time_to_ack
= 0;
1101 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1103 BUG_TRAP(!skb
|| before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
));
1106 if (inet_csk_ack_scheduled(sk
)) {
1107 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1108 /* Delayed ACKs frequently hit locked sockets during bulk
1110 if (icsk
->icsk_ack
.blocked
||
1111 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1112 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1114 * If this read emptied read buffer, we send ACK, if
1115 * connection is not bidirectional, user drained
1116 * receive buffer and there was a small segment
1120 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1121 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1122 !icsk
->icsk_ack
.pingpong
)) &&
1123 !atomic_read(&sk
->sk_rmem_alloc
)))
1127 /* We send an ACK if we can now advertise a non-zero window
1128 * which has been raised "significantly".
1130 * Even if window raised up to infinity, do not send window open ACK
1131 * in states, where we will not receive more. It is useless.
1133 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1134 __u32 rcv_window_now
= tcp_receive_window(tp
);
1136 /* Optimize, __tcp_select_window() is not cheap. */
1137 if (2*rcv_window_now
<= tp
->window_clamp
) {
1138 __u32 new_window
= __tcp_select_window(sk
);
1140 /* Send ACK now, if this read freed lots of space
1141 * in our buffer. Certainly, new_window is new window.
1142 * We can advertise it now, if it is not less than current one.
1143 * "Lots" means "at least twice" here.
1145 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1153 static void tcp_prequeue_process(struct sock
*sk
)
1155 struct sk_buff
*skb
;
1156 struct tcp_sock
*tp
= tcp_sk(sk
);
1158 NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED
);
1160 /* RX process wants to run with disabled BHs, though it is not
1163 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
1164 sk
->sk_backlog_rcv(sk
, skb
);
1167 /* Clear memory counter. */
1168 tp
->ucopy
.memory
= 0;
1171 static inline struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1173 struct sk_buff
*skb
;
1176 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1177 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1178 if (tcp_hdr(skb
)->syn
)
1180 if (offset
< skb
->len
|| tcp_hdr(skb
)->fin
) {
1189 * This routine provides an alternative to tcp_recvmsg() for routines
1190 * that would like to handle copying from skbuffs directly in 'sendfile'
1193 * - It is assumed that the socket was locked by the caller.
1194 * - The routine does not block.
1195 * - At present, there is no support for reading OOB data
1196 * or for 'peeking' the socket using this routine
1197 * (although both would be easy to implement).
1199 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1200 sk_read_actor_t recv_actor
)
1202 struct sk_buff
*skb
;
1203 struct tcp_sock
*tp
= tcp_sk(sk
);
1204 u32 seq
= tp
->copied_seq
;
1208 if (sk
->sk_state
== TCP_LISTEN
)
1210 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1211 if (offset
< skb
->len
) {
1214 len
= skb
->len
- offset
;
1215 /* Stop reading if we hit a patch of urgent data */
1217 u32 urg_offset
= tp
->urg_seq
- seq
;
1218 if (urg_offset
< len
)
1223 used
= recv_actor(desc
, skb
, offset
, len
);
1228 } else if (used
<= len
) {
1234 * If recv_actor drops the lock (e.g. TCP splice
1235 * receive) the skb pointer might be invalid when
1236 * getting here: tcp_collapse might have deleted it
1237 * while aggregating skbs from the socket queue.
1239 skb
= tcp_recv_skb(sk
, seq
-1, &offset
);
1240 if (!skb
|| (offset
+1 != skb
->len
))
1243 if (tcp_hdr(skb
)->fin
) {
1244 sk_eat_skb(sk
, skb
, 0);
1248 sk_eat_skb(sk
, skb
, 0);
1252 tp
->copied_seq
= seq
;
1254 tcp_rcv_space_adjust(sk
);
1256 /* Clean up data we have read: This will do ACK frames. */
1258 tcp_cleanup_rbuf(sk
, copied
);
1263 * This routine copies from a sock struct into the user buffer.
1265 * Technical note: in 2.3 we work on _locked_ socket, so that
1266 * tricks with *seq access order and skb->users are not required.
1267 * Probably, code can be easily improved even more.
1270 int tcp_recvmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1271 size_t len
, int nonblock
, int flags
, int *addr_len
)
1273 struct tcp_sock
*tp
= tcp_sk(sk
);
1279 int target
; /* Read at least this many bytes */
1281 struct task_struct
*user_recv
= NULL
;
1282 int copied_early
= 0;
1283 struct sk_buff
*skb
;
1287 TCP_CHECK_TIMER(sk
);
1290 if (sk
->sk_state
== TCP_LISTEN
)
1293 timeo
= sock_rcvtimeo(sk
, nonblock
);
1295 /* Urgent data needs to be handled specially. */
1296 if (flags
& MSG_OOB
)
1299 seq
= &tp
->copied_seq
;
1300 if (flags
& MSG_PEEK
) {
1301 peek_seq
= tp
->copied_seq
;
1305 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1307 #ifdef CONFIG_NET_DMA
1308 tp
->ucopy
.dma_chan
= NULL
;
1310 skb
= skb_peek_tail(&sk
->sk_receive_queue
);
1315 available
= TCP_SKB_CB(skb
)->seq
+ skb
->len
- (*seq
);
1316 if ((available
< target
) &&
1317 (len
> sysctl_tcp_dma_copybreak
) && !(flags
& MSG_PEEK
) &&
1318 !sysctl_tcp_low_latency
&&
1319 __get_cpu_var(softnet_data
).net_dma
) {
1320 preempt_enable_no_resched();
1321 tp
->ucopy
.pinned_list
=
1322 dma_pin_iovec_pages(msg
->msg_iov
, len
);
1324 preempt_enable_no_resched();
1332 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1333 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1336 if (signal_pending(current
)) {
1337 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1342 /* Next get a buffer. */
1344 skb
= skb_peek(&sk
->sk_receive_queue
);
1349 /* Now that we have two receive queues this
1352 if (before(*seq
, TCP_SKB_CB(skb
)->seq
)) {
1353 printk(KERN_INFO
"recvmsg bug: copied %X "
1354 "seq %X\n", *seq
, TCP_SKB_CB(skb
)->seq
);
1357 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1358 if (tcp_hdr(skb
)->syn
)
1360 if (offset
< skb
->len
)
1362 if (tcp_hdr(skb
)->fin
)
1364 BUG_TRAP(flags
& MSG_PEEK
);
1366 } while (skb
!= (struct sk_buff
*)&sk
->sk_receive_queue
);
1368 /* Well, if we have backlog, try to process it now yet. */
1370 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1375 sk
->sk_state
== TCP_CLOSE
||
1376 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1378 signal_pending(current
) ||
1382 if (sock_flag(sk
, SOCK_DONE
))
1386 copied
= sock_error(sk
);
1390 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1393 if (sk
->sk_state
== TCP_CLOSE
) {
1394 if (!sock_flag(sk
, SOCK_DONE
)) {
1395 /* This occurs when user tries to read
1396 * from never connected socket.
1409 if (signal_pending(current
)) {
1410 copied
= sock_intr_errno(timeo
);
1415 tcp_cleanup_rbuf(sk
, copied
);
1417 if (!sysctl_tcp_low_latency
&& tp
->ucopy
.task
== user_recv
) {
1418 /* Install new reader */
1419 if (!user_recv
&& !(flags
& (MSG_TRUNC
| MSG_PEEK
))) {
1420 user_recv
= current
;
1421 tp
->ucopy
.task
= user_recv
;
1422 tp
->ucopy
.iov
= msg
->msg_iov
;
1425 tp
->ucopy
.len
= len
;
1427 BUG_TRAP(tp
->copied_seq
== tp
->rcv_nxt
||
1428 (flags
& (MSG_PEEK
| MSG_TRUNC
)));
1430 /* Ugly... If prequeue is not empty, we have to
1431 * process it before releasing socket, otherwise
1432 * order will be broken at second iteration.
1433 * More elegant solution is required!!!
1435 * Look: we have the following (pseudo)queues:
1437 * 1. packets in flight
1442 * Each queue can be processed only if the next ones
1443 * are empty. At this point we have empty receive_queue.
1444 * But prequeue _can_ be not empty after 2nd iteration,
1445 * when we jumped to start of loop because backlog
1446 * processing added something to receive_queue.
1447 * We cannot release_sock(), because backlog contains
1448 * packets arrived _after_ prequeued ones.
1450 * Shortly, algorithm is clear --- to process all
1451 * the queues in order. We could make it more directly,
1452 * requeueing packets from backlog to prequeue, if
1453 * is not empty. It is more elegant, but eats cycles,
1456 if (!skb_queue_empty(&tp
->ucopy
.prequeue
))
1459 /* __ Set realtime policy in scheduler __ */
1462 if (copied
>= target
) {
1463 /* Do not sleep, just process backlog. */
1467 sk_wait_data(sk
, &timeo
);
1469 #ifdef CONFIG_NET_DMA
1470 tp
->ucopy
.wakeup
= 0;
1476 /* __ Restore normal policy in scheduler __ */
1478 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1479 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG
, chunk
);
1484 if (tp
->rcv_nxt
== tp
->copied_seq
&&
1485 !skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1487 tcp_prequeue_process(sk
);
1489 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1490 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1496 if ((flags
& MSG_PEEK
) && peek_seq
!= tp
->copied_seq
) {
1497 if (net_ratelimit())
1498 printk(KERN_DEBUG
"TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1499 current
->comm
, task_pid_nr(current
));
1500 peek_seq
= tp
->copied_seq
;
1505 /* Ok so how much can we use? */
1506 used
= skb
->len
- offset
;
1510 /* Do we have urgent data here? */
1512 u32 urg_offset
= tp
->urg_seq
- *seq
;
1513 if (urg_offset
< used
) {
1515 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1527 if (!(flags
& MSG_TRUNC
)) {
1528 #ifdef CONFIG_NET_DMA
1529 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
1530 tp
->ucopy
.dma_chan
= get_softnet_dma();
1532 if (tp
->ucopy
.dma_chan
) {
1533 tp
->ucopy
.dma_cookie
= dma_skb_copy_datagram_iovec(
1534 tp
->ucopy
.dma_chan
, skb
, offset
,
1536 tp
->ucopy
.pinned_list
);
1538 if (tp
->ucopy
.dma_cookie
< 0) {
1540 printk(KERN_ALERT
"dma_cookie < 0\n");
1542 /* Exception. Bailout! */
1547 if ((offset
+ used
) == skb
->len
)
1553 err
= skb_copy_datagram_iovec(skb
, offset
,
1554 msg
->msg_iov
, used
);
1556 /* Exception. Bailout! */
1568 tcp_rcv_space_adjust(sk
);
1571 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1573 tcp_fast_path_check(sk
);
1575 if (used
+ offset
< skb
->len
)
1578 if (tcp_hdr(skb
)->fin
)
1580 if (!(flags
& MSG_PEEK
)) {
1581 sk_eat_skb(sk
, skb
, copied_early
);
1587 /* Process the FIN. */
1589 if (!(flags
& MSG_PEEK
)) {
1590 sk_eat_skb(sk
, skb
, copied_early
);
1597 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1600 tp
->ucopy
.len
= copied
> 0 ? len
: 0;
1602 tcp_prequeue_process(sk
);
1604 if (copied
> 0 && (chunk
= len
- tp
->ucopy
.len
) != 0) {
1605 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1611 tp
->ucopy
.task
= NULL
;
1615 #ifdef CONFIG_NET_DMA
1616 if (tp
->ucopy
.dma_chan
) {
1617 dma_cookie_t done
, used
;
1619 dma_async_memcpy_issue_pending(tp
->ucopy
.dma_chan
);
1621 while (dma_async_memcpy_complete(tp
->ucopy
.dma_chan
,
1622 tp
->ucopy
.dma_cookie
, &done
,
1623 &used
) == DMA_IN_PROGRESS
) {
1624 /* do partial cleanup of sk_async_wait_queue */
1625 while ((skb
= skb_peek(&sk
->sk_async_wait_queue
)) &&
1626 (dma_async_is_complete(skb
->dma_cookie
, done
,
1627 used
) == DMA_SUCCESS
)) {
1628 __skb_dequeue(&sk
->sk_async_wait_queue
);
1633 /* Safe to free early-copied skbs now */
1634 __skb_queue_purge(&sk
->sk_async_wait_queue
);
1635 dma_chan_put(tp
->ucopy
.dma_chan
);
1636 tp
->ucopy
.dma_chan
= NULL
;
1638 if (tp
->ucopy
.pinned_list
) {
1639 dma_unpin_iovec_pages(tp
->ucopy
.pinned_list
);
1640 tp
->ucopy
.pinned_list
= NULL
;
1644 /* According to UNIX98, msg_name/msg_namelen are ignored
1645 * on connected socket. I was just happy when found this 8) --ANK
1648 /* Clean up data we have read: This will do ACK frames. */
1649 tcp_cleanup_rbuf(sk
, copied
);
1651 TCP_CHECK_TIMER(sk
);
1656 TCP_CHECK_TIMER(sk
);
1661 err
= tcp_recv_urg(sk
, timeo
, msg
, len
, flags
, addr_len
);
1665 void tcp_set_state(struct sock
*sk
, int state
)
1667 int oldstate
= sk
->sk_state
;
1670 case TCP_ESTABLISHED
:
1671 if (oldstate
!= TCP_ESTABLISHED
)
1672 TCP_INC_STATS(TCP_MIB_CURRESTAB
);
1676 if (oldstate
== TCP_CLOSE_WAIT
|| oldstate
== TCP_ESTABLISHED
)
1677 TCP_INC_STATS(TCP_MIB_ESTABRESETS
);
1679 sk
->sk_prot
->unhash(sk
);
1680 if (inet_csk(sk
)->icsk_bind_hash
&&
1681 !(sk
->sk_userlocks
& SOCK_BINDPORT_LOCK
))
1685 if (oldstate
==TCP_ESTABLISHED
)
1686 TCP_DEC_STATS(TCP_MIB_CURRESTAB
);
1689 /* Change state AFTER socket is unhashed to avoid closed
1690 * socket sitting in hash tables.
1692 sk
->sk_state
= state
;
1695 SOCK_DEBUG(sk
, "TCP sk=%p, State %s -> %s\n",sk
, statename
[oldstate
],statename
[state
]);
1698 EXPORT_SYMBOL_GPL(tcp_set_state
);
1701 * State processing on a close. This implements the state shift for
1702 * sending our FIN frame. Note that we only send a FIN for some
1703 * states. A shutdown() may have already sent the FIN, or we may be
1707 static const unsigned char new_state
[16] = {
1708 /* current state: new state: action: */
1709 /* (Invalid) */ TCP_CLOSE
,
1710 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1711 /* TCP_SYN_SENT */ TCP_CLOSE
,
1712 /* TCP_SYN_RECV */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1713 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1
,
1714 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2
,
1715 /* TCP_TIME_WAIT */ TCP_CLOSE
,
1716 /* TCP_CLOSE */ TCP_CLOSE
,
1717 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK
| TCP_ACTION_FIN
,
1718 /* TCP_LAST_ACK */ TCP_LAST_ACK
,
1719 /* TCP_LISTEN */ TCP_CLOSE
,
1720 /* TCP_CLOSING */ TCP_CLOSING
,
1723 static int tcp_close_state(struct sock
*sk
)
1725 int next
= (int)new_state
[sk
->sk_state
];
1726 int ns
= next
& TCP_STATE_MASK
;
1728 tcp_set_state(sk
, ns
);
1730 return next
& TCP_ACTION_FIN
;
1734 * Shutdown the sending side of a connection. Much like close except
1735 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1738 void tcp_shutdown(struct sock
*sk
, int how
)
1740 /* We need to grab some memory, and put together a FIN,
1741 * and then put it into the queue to be sent.
1742 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1744 if (!(how
& SEND_SHUTDOWN
))
1747 /* If we've already sent a FIN, or it's a closed state, skip this. */
1748 if ((1 << sk
->sk_state
) &
1749 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
1750 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
1751 /* Clear out any half completed packets. FIN if needed. */
1752 if (tcp_close_state(sk
))
1757 void tcp_close(struct sock
*sk
, long timeout
)
1759 struct sk_buff
*skb
;
1760 int data_was_unread
= 0;
1764 sk
->sk_shutdown
= SHUTDOWN_MASK
;
1766 if (sk
->sk_state
== TCP_LISTEN
) {
1767 tcp_set_state(sk
, TCP_CLOSE
);
1770 inet_csk_listen_stop(sk
);
1772 goto adjudge_to_death
;
1775 /* We need to flush the recv. buffs. We do this only on the
1776 * descriptor close, not protocol-sourced closes, because the
1777 * reader process may not have drained the data yet!
1779 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
1780 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
-
1782 data_was_unread
+= len
;
1788 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1789 * data was lost. To witness the awful effects of the old behavior of
1790 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1791 * GET in an FTP client, suspend the process, wait for the client to
1792 * advertise a zero window, then kill -9 the FTP client, wheee...
1793 * Note: timeout is always zero in such a case.
1795 if (data_was_unread
) {
1796 /* Unread data was tossed, zap the connection. */
1797 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE
);
1798 tcp_set_state(sk
, TCP_CLOSE
);
1799 tcp_send_active_reset(sk
, GFP_KERNEL
);
1800 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
1801 /* Check zero linger _after_ checking for unread data. */
1802 sk
->sk_prot
->disconnect(sk
, 0);
1803 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA
);
1804 } else if (tcp_close_state(sk
)) {
1805 /* We FIN if the application ate all the data before
1806 * zapping the connection.
1809 /* RED-PEN. Formally speaking, we have broken TCP state
1810 * machine. State transitions:
1812 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1813 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1814 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1816 * are legal only when FIN has been sent (i.e. in window),
1817 * rather than queued out of window. Purists blame.
1819 * F.e. "RFC state" is ESTABLISHED,
1820 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1822 * The visible declinations are that sometimes
1823 * we enter time-wait state, when it is not required really
1824 * (harmless), do not send active resets, when they are
1825 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1826 * they look as CLOSING or LAST_ACK for Linux)
1827 * Probably, I missed some more holelets.
1833 sk_stream_wait_close(sk
, timeout
);
1836 state
= sk
->sk_state
;
1839 atomic_inc(sk
->sk_prot
->orphan_count
);
1841 /* It is the last release_sock in its life. It will remove backlog. */
1845 /* Now socket is owned by kernel and we acquire BH lock
1846 to finish close. No need to check for user refs.
1850 BUG_TRAP(!sock_owned_by_user(sk
));
1852 /* Have we already been destroyed by a softirq or backlog? */
1853 if (state
!= TCP_CLOSE
&& sk
->sk_state
== TCP_CLOSE
)
1856 /* This is a (useful) BSD violating of the RFC. There is a
1857 * problem with TCP as specified in that the other end could
1858 * keep a socket open forever with no application left this end.
1859 * We use a 3 minute timeout (about the same as BSD) then kill
1860 * our end. If they send after that then tough - BUT: long enough
1861 * that we won't make the old 4*rto = almost no time - whoops
1864 * Nope, it was not mistake. It is really desired behaviour
1865 * f.e. on http servers, when such sockets are useless, but
1866 * consume significant resources. Let's do it with special
1867 * linger2 option. --ANK
1870 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
1871 struct tcp_sock
*tp
= tcp_sk(sk
);
1872 if (tp
->linger2
< 0) {
1873 tcp_set_state(sk
, TCP_CLOSE
);
1874 tcp_send_active_reset(sk
, GFP_ATOMIC
);
1875 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER
);
1877 const int tmo
= tcp_fin_time(sk
);
1879 if (tmo
> TCP_TIMEWAIT_LEN
) {
1880 inet_csk_reset_keepalive_timer(sk
,
1881 tmo
- TCP_TIMEWAIT_LEN
);
1883 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
1888 if (sk
->sk_state
!= TCP_CLOSE
) {
1890 if (tcp_too_many_orphans(sk
,
1891 atomic_read(sk
->sk_prot
->orphan_count
))) {
1892 if (net_ratelimit())
1893 printk(KERN_INFO
"TCP: too many of orphaned "
1895 tcp_set_state(sk
, TCP_CLOSE
);
1896 tcp_send_active_reset(sk
, GFP_ATOMIC
);
1897 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY
);
1901 if (sk
->sk_state
== TCP_CLOSE
)
1902 inet_csk_destroy_sock(sk
);
1903 /* Otherwise, socket is reprieved until protocol close. */
1911 /* These states need RST on ABORT according to RFC793 */
1913 static inline int tcp_need_reset(int state
)
1915 return (1 << state
) &
1916 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
1917 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
1920 int tcp_disconnect(struct sock
*sk
, int flags
)
1922 struct inet_sock
*inet
= inet_sk(sk
);
1923 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1924 struct tcp_sock
*tp
= tcp_sk(sk
);
1926 int old_state
= sk
->sk_state
;
1928 if (old_state
!= TCP_CLOSE
)
1929 tcp_set_state(sk
, TCP_CLOSE
);
1931 /* ABORT function of RFC793 */
1932 if (old_state
== TCP_LISTEN
) {
1933 inet_csk_listen_stop(sk
);
1934 } else if (tcp_need_reset(old_state
) ||
1935 (tp
->snd_nxt
!= tp
->write_seq
&&
1936 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
1937 /* The last check adjusts for discrepancy of Linux wrt. RFC
1940 tcp_send_active_reset(sk
, gfp_any());
1941 sk
->sk_err
= ECONNRESET
;
1942 } else if (old_state
== TCP_SYN_SENT
)
1943 sk
->sk_err
= ECONNRESET
;
1945 tcp_clear_xmit_timers(sk
);
1946 __skb_queue_purge(&sk
->sk_receive_queue
);
1947 tcp_write_queue_purge(sk
);
1948 __skb_queue_purge(&tp
->out_of_order_queue
);
1949 #ifdef CONFIG_NET_DMA
1950 __skb_queue_purge(&sk
->sk_async_wait_queue
);
1955 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
1956 inet_reset_saddr(sk
);
1958 sk
->sk_shutdown
= 0;
1959 sock_reset_flag(sk
, SOCK_DONE
);
1961 if ((tp
->write_seq
+= tp
->max_window
+ 2) == 0)
1963 icsk
->icsk_backoff
= 0;
1965 icsk
->icsk_probes_out
= 0;
1966 tp
->packets_out
= 0;
1967 tp
->snd_ssthresh
= 0x7fffffff;
1968 tp
->snd_cwnd_cnt
= 0;
1969 tp
->bytes_acked
= 0;
1970 tcp_set_ca_state(sk
, TCP_CA_Open
);
1971 tcp_clear_retrans(tp
);
1972 inet_csk_delack_init(sk
);
1973 tcp_init_send_head(sk
);
1974 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
1977 BUG_TRAP(!inet
->num
|| icsk
->icsk_bind_hash
);
1979 sk
->sk_error_report(sk
);
1984 * Socket option code for TCP.
1986 static int do_tcp_setsockopt(struct sock
*sk
, int level
,
1987 int optname
, char __user
*optval
, int optlen
)
1989 struct tcp_sock
*tp
= tcp_sk(sk
);
1990 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1994 /* This is a string value all the others are int's */
1995 if (optname
== TCP_CONGESTION
) {
1996 char name
[TCP_CA_NAME_MAX
];
2001 val
= strncpy_from_user(name
, optval
,
2002 min(TCP_CA_NAME_MAX
-1, optlen
));
2008 err
= tcp_set_congestion_control(sk
, name
);
2013 if (optlen
< sizeof(int))
2016 if (get_user(val
, (int __user
*)optval
))
2023 /* Values greater than interface MTU won't take effect. However
2024 * at the point when this call is done we typically don't yet
2025 * know which interface is going to be used */
2026 if (val
< 8 || val
> MAX_TCP_WINDOW
) {
2030 tp
->rx_opt
.user_mss
= val
;
2035 /* TCP_NODELAY is weaker than TCP_CORK, so that
2036 * this option on corked socket is remembered, but
2037 * it is not activated until cork is cleared.
2039 * However, when TCP_NODELAY is set we make
2040 * an explicit push, which overrides even TCP_CORK
2041 * for currently queued segments.
2043 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
2044 tcp_push_pending_frames(sk
);
2046 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
2051 /* When set indicates to always queue non-full frames.
2052 * Later the user clears this option and we transmit
2053 * any pending partial frames in the queue. This is
2054 * meant to be used alongside sendfile() to get properly
2055 * filled frames when the user (for example) must write
2056 * out headers with a write() call first and then use
2057 * sendfile to send out the data parts.
2059 * TCP_CORK can be set together with TCP_NODELAY and it is
2060 * stronger than TCP_NODELAY.
2063 tp
->nonagle
|= TCP_NAGLE_CORK
;
2065 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
2066 if (tp
->nonagle
&TCP_NAGLE_OFF
)
2067 tp
->nonagle
|= TCP_NAGLE_PUSH
;
2068 tcp_push_pending_frames(sk
);
2073 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
2076 tp
->keepalive_time
= val
* HZ
;
2077 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
2078 !((1 << sk
->sk_state
) &
2079 (TCPF_CLOSE
| TCPF_LISTEN
))) {
2080 __u32 elapsed
= tcp_time_stamp
- tp
->rcv_tstamp
;
2081 if (tp
->keepalive_time
> elapsed
)
2082 elapsed
= tp
->keepalive_time
- elapsed
;
2085 inet_csk_reset_keepalive_timer(sk
, elapsed
);
2090 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
2093 tp
->keepalive_intvl
= val
* HZ
;
2096 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
2099 tp
->keepalive_probes
= val
;
2102 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
2105 icsk
->icsk_syn_retries
= val
;
2111 else if (val
> sysctl_tcp_fin_timeout
/ HZ
)
2114 tp
->linger2
= val
* HZ
;
2117 case TCP_DEFER_ACCEPT
:
2118 icsk
->icsk_accept_queue
.rskq_defer_accept
= 0;
2120 /* Translate value in seconds to number of
2122 while (icsk
->icsk_accept_queue
.rskq_defer_accept
< 32 &&
2123 val
> ((TCP_TIMEOUT_INIT
/ HZ
) <<
2124 icsk
->icsk_accept_queue
.rskq_defer_accept
))
2125 icsk
->icsk_accept_queue
.rskq_defer_accept
++;
2126 icsk
->icsk_accept_queue
.rskq_defer_accept
++;
2130 case TCP_WINDOW_CLAMP
:
2132 if (sk
->sk_state
!= TCP_CLOSE
) {
2136 tp
->window_clamp
= 0;
2138 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
2139 SOCK_MIN_RCVBUF
/ 2 : val
;
2144 icsk
->icsk_ack
.pingpong
= 1;
2146 icsk
->icsk_ack
.pingpong
= 0;
2147 if ((1 << sk
->sk_state
) &
2148 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
2149 inet_csk_ack_scheduled(sk
)) {
2150 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
2151 tcp_cleanup_rbuf(sk
, 1);
2153 icsk
->icsk_ack
.pingpong
= 1;
2158 #ifdef CONFIG_TCP_MD5SIG
2160 /* Read the IP->Key mappings from userspace */
2161 err
= tp
->af_specific
->md5_parse(sk
, optval
, optlen
);
2174 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2177 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2179 if (level
!= SOL_TCP
)
2180 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
2182 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2185 #ifdef CONFIG_COMPAT
2186 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
2187 char __user
*optval
, int optlen
)
2189 if (level
!= SOL_TCP
)
2190 return inet_csk_compat_setsockopt(sk
, level
, optname
,
2192 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2195 EXPORT_SYMBOL(compat_tcp_setsockopt
);
2198 /* Return information about state of tcp endpoint in API format. */
2199 void tcp_get_info(struct sock
*sk
, struct tcp_info
*info
)
2201 struct tcp_sock
*tp
= tcp_sk(sk
);
2202 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2203 u32 now
= tcp_time_stamp
;
2205 memset(info
, 0, sizeof(*info
));
2207 info
->tcpi_state
= sk
->sk_state
;
2208 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
2209 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
2210 info
->tcpi_probes
= icsk
->icsk_probes_out
;
2211 info
->tcpi_backoff
= icsk
->icsk_backoff
;
2213 if (tp
->rx_opt
.tstamp_ok
)
2214 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
2215 if (tcp_is_sack(tp
))
2216 info
->tcpi_options
|= TCPI_OPT_SACK
;
2217 if (tp
->rx_opt
.wscale_ok
) {
2218 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
2219 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
2220 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
2223 if (tp
->ecn_flags
&TCP_ECN_OK
)
2224 info
->tcpi_options
|= TCPI_OPT_ECN
;
2226 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
2227 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
2228 info
->tcpi_snd_mss
= tp
->mss_cache
;
2229 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
2231 if (sk
->sk_state
== TCP_LISTEN
) {
2232 info
->tcpi_unacked
= sk
->sk_ack_backlog
;
2233 info
->tcpi_sacked
= sk
->sk_max_ack_backlog
;
2235 info
->tcpi_unacked
= tp
->packets_out
;
2236 info
->tcpi_sacked
= tp
->sacked_out
;
2238 info
->tcpi_lost
= tp
->lost_out
;
2239 info
->tcpi_retrans
= tp
->retrans_out
;
2240 info
->tcpi_fackets
= tp
->fackets_out
;
2242 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
2243 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
2244 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
2246 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
2247 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
2248 info
->tcpi_rtt
= jiffies_to_usecs(tp
->srtt
)>>3;
2249 info
->tcpi_rttvar
= jiffies_to_usecs(tp
->mdev
)>>2;
2250 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
2251 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
2252 info
->tcpi_advmss
= tp
->advmss
;
2253 info
->tcpi_reordering
= tp
->reordering
;
2255 info
->tcpi_rcv_rtt
= jiffies_to_usecs(tp
->rcv_rtt_est
.rtt
)>>3;
2256 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
2258 info
->tcpi_total_retrans
= tp
->total_retrans
;
2261 EXPORT_SYMBOL_GPL(tcp_get_info
);
2263 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
2264 int optname
, char __user
*optval
, int __user
*optlen
)
2266 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2267 struct tcp_sock
*tp
= tcp_sk(sk
);
2270 if (get_user(len
, optlen
))
2273 len
= min_t(unsigned int, len
, sizeof(int));
2280 val
= tp
->mss_cache
;
2281 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
2282 val
= tp
->rx_opt
.user_mss
;
2285 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
2288 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
2291 val
= (tp
->keepalive_time
? : sysctl_tcp_keepalive_time
) / HZ
;
2294 val
= (tp
->keepalive_intvl
? : sysctl_tcp_keepalive_intvl
) / HZ
;
2297 val
= tp
->keepalive_probes
? : sysctl_tcp_keepalive_probes
;
2300 val
= icsk
->icsk_syn_retries
? : sysctl_tcp_syn_retries
;
2305 val
= (val
? : sysctl_tcp_fin_timeout
) / HZ
;
2307 case TCP_DEFER_ACCEPT
:
2308 val
= !icsk
->icsk_accept_queue
.rskq_defer_accept
? 0 :
2309 ((TCP_TIMEOUT_INIT
/ HZ
) << (icsk
->icsk_accept_queue
.rskq_defer_accept
- 1));
2311 case TCP_WINDOW_CLAMP
:
2312 val
= tp
->window_clamp
;
2315 struct tcp_info info
;
2317 if (get_user(len
, optlen
))
2320 tcp_get_info(sk
, &info
);
2322 len
= min_t(unsigned int, len
, sizeof(info
));
2323 if (put_user(len
, optlen
))
2325 if (copy_to_user(optval
, &info
, len
))
2330 val
= !icsk
->icsk_ack
.pingpong
;
2333 case TCP_CONGESTION
:
2334 if (get_user(len
, optlen
))
2336 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
2337 if (put_user(len
, optlen
))
2339 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
2343 return -ENOPROTOOPT
;
2346 if (put_user(len
, optlen
))
2348 if (copy_to_user(optval
, &val
, len
))
2353 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2356 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2358 if (level
!= SOL_TCP
)
2359 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
2361 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2364 #ifdef CONFIG_COMPAT
2365 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
2366 char __user
*optval
, int __user
*optlen
)
2368 if (level
!= SOL_TCP
)
2369 return inet_csk_compat_getsockopt(sk
, level
, optname
,
2371 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2374 EXPORT_SYMBOL(compat_tcp_getsockopt
);
2377 struct sk_buff
*tcp_tso_segment(struct sk_buff
*skb
, int features
)
2379 struct sk_buff
*segs
= ERR_PTR(-EINVAL
);
2384 unsigned int oldlen
;
2387 if (!pskb_may_pull(skb
, sizeof(*th
)))
2391 thlen
= th
->doff
* 4;
2392 if (thlen
< sizeof(*th
))
2395 if (!pskb_may_pull(skb
, thlen
))
2398 oldlen
= (u16
)~skb
->len
;
2399 __skb_pull(skb
, thlen
);
2401 if (skb_gso_ok(skb
, features
| NETIF_F_GSO_ROBUST
)) {
2402 /* Packet is from an untrusted source, reset gso_segs. */
2403 int type
= skb_shinfo(skb
)->gso_type
;
2412 !(type
& (SKB_GSO_TCPV4
| SKB_GSO_TCPV6
))))
2415 mss
= skb_shinfo(skb
)->gso_size
;
2416 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss
);
2422 segs
= skb_segment(skb
, features
);
2426 len
= skb_shinfo(skb
)->gso_size
;
2427 delta
= htonl(oldlen
+ (thlen
+ len
));
2431 seq
= ntohl(th
->seq
);
2434 th
->fin
= th
->psh
= 0;
2436 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
2437 (__force u32
)delta
));
2438 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2440 csum_fold(csum_partial(skb_transport_header(skb
),
2447 th
->seq
= htonl(seq
);
2449 } while (skb
->next
);
2451 delta
= htonl(oldlen
+ (skb
->tail
- skb
->transport_header
) +
2453 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
2454 (__force u32
)delta
));
2455 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2456 th
->check
= csum_fold(csum_partial(skb_transport_header(skb
),
2462 EXPORT_SYMBOL(tcp_tso_segment
);
2464 #ifdef CONFIG_TCP_MD5SIG
2465 static unsigned long tcp_md5sig_users
;
2466 static struct tcp_md5sig_pool
**tcp_md5sig_pool
;
2467 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock
);
2469 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool
**pool
)
2472 for_each_possible_cpu(cpu
) {
2473 struct tcp_md5sig_pool
*p
= *per_cpu_ptr(pool
, cpu
);
2475 if (p
->md5_desc
.tfm
)
2476 crypto_free_hash(p
->md5_desc
.tfm
);
2484 void tcp_free_md5sig_pool(void)
2486 struct tcp_md5sig_pool
**pool
= NULL
;
2488 spin_lock_bh(&tcp_md5sig_pool_lock
);
2489 if (--tcp_md5sig_users
== 0) {
2490 pool
= tcp_md5sig_pool
;
2491 tcp_md5sig_pool
= NULL
;
2493 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2495 __tcp_free_md5sig_pool(pool
);
2498 EXPORT_SYMBOL(tcp_free_md5sig_pool
);
2500 static struct tcp_md5sig_pool
**__tcp_alloc_md5sig_pool(void)
2503 struct tcp_md5sig_pool
**pool
;
2505 pool
= alloc_percpu(struct tcp_md5sig_pool
*);
2509 for_each_possible_cpu(cpu
) {
2510 struct tcp_md5sig_pool
*p
;
2511 struct crypto_hash
*hash
;
2513 p
= kzalloc(sizeof(*p
), GFP_KERNEL
);
2516 *per_cpu_ptr(pool
, cpu
) = p
;
2518 hash
= crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC
);
2519 if (!hash
|| IS_ERR(hash
))
2522 p
->md5_desc
.tfm
= hash
;
2526 __tcp_free_md5sig_pool(pool
);
2530 struct tcp_md5sig_pool
**tcp_alloc_md5sig_pool(void)
2532 struct tcp_md5sig_pool
**pool
;
2536 spin_lock_bh(&tcp_md5sig_pool_lock
);
2537 pool
= tcp_md5sig_pool
;
2538 if (tcp_md5sig_users
++ == 0) {
2540 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2543 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2547 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2550 /* we cannot hold spinlock here because this may sleep. */
2551 struct tcp_md5sig_pool
**p
= __tcp_alloc_md5sig_pool();
2552 spin_lock_bh(&tcp_md5sig_pool_lock
);
2555 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2558 pool
= tcp_md5sig_pool
;
2560 /* oops, it has already been assigned. */
2561 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2562 __tcp_free_md5sig_pool(p
);
2564 tcp_md5sig_pool
= pool
= p
;
2565 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2571 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
2573 struct tcp_md5sig_pool
*__tcp_get_md5sig_pool(int cpu
)
2575 struct tcp_md5sig_pool
**p
;
2576 spin_lock_bh(&tcp_md5sig_pool_lock
);
2577 p
= tcp_md5sig_pool
;
2580 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2581 return (p
? *per_cpu_ptr(p
, cpu
) : NULL
);
2584 EXPORT_SYMBOL(__tcp_get_md5sig_pool
);
2586 void __tcp_put_md5sig_pool(void)
2588 tcp_free_md5sig_pool();
2591 EXPORT_SYMBOL(__tcp_put_md5sig_pool
);
2594 void tcp_done(struct sock
*sk
)
2596 if(sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
2597 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS
);
2599 tcp_set_state(sk
, TCP_CLOSE
);
2600 tcp_clear_xmit_timers(sk
);
2602 sk
->sk_shutdown
= SHUTDOWN_MASK
;
2604 if (!sock_flag(sk
, SOCK_DEAD
))
2605 sk
->sk_state_change(sk
);
2607 inet_csk_destroy_sock(sk
);
2609 EXPORT_SYMBOL_GPL(tcp_done
);
2611 extern struct tcp_congestion_ops tcp_reno
;
2613 static __initdata
unsigned long thash_entries
;
2614 static int __init
set_thash_entries(char *str
)
2618 thash_entries
= simple_strtoul(str
, &str
, 0);
2621 __setup("thash_entries=", set_thash_entries
);
2623 void __init
tcp_init(void)
2625 struct sk_buff
*skb
= NULL
;
2626 unsigned long nr_pages
, limit
;
2627 int order
, i
, max_share
;
2629 BUILD_BUG_ON(sizeof(struct tcp_skb_cb
) > sizeof(skb
->cb
));
2631 tcp_hashinfo
.bind_bucket_cachep
=
2632 kmem_cache_create("tcp_bind_bucket",
2633 sizeof(struct inet_bind_bucket
), 0,
2634 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
);
2636 /* Size and allocate the main established and bind bucket
2639 * The methodology is similar to that of the buffer cache.
2641 tcp_hashinfo
.ehash
=
2642 alloc_large_system_hash("TCP established",
2643 sizeof(struct inet_ehash_bucket
),
2645 (num_physpages
>= 128 * 1024) ?
2648 &tcp_hashinfo
.ehash_size
,
2650 thash_entries
? 0 : 512 * 1024);
2651 tcp_hashinfo
.ehash_size
= 1 << tcp_hashinfo
.ehash_size
;
2652 for (i
= 0; i
< tcp_hashinfo
.ehash_size
; i
++) {
2653 INIT_HLIST_HEAD(&tcp_hashinfo
.ehash
[i
].chain
);
2654 INIT_HLIST_HEAD(&tcp_hashinfo
.ehash
[i
].twchain
);
2656 if (inet_ehash_locks_alloc(&tcp_hashinfo
))
2657 panic("TCP: failed to alloc ehash_locks");
2658 tcp_hashinfo
.bhash
=
2659 alloc_large_system_hash("TCP bind",
2660 sizeof(struct inet_bind_hashbucket
),
2661 tcp_hashinfo
.ehash_size
,
2662 (num_physpages
>= 128 * 1024) ?
2665 &tcp_hashinfo
.bhash_size
,
2668 tcp_hashinfo
.bhash_size
= 1 << tcp_hashinfo
.bhash_size
;
2669 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
2670 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
2671 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
2674 /* Try to be a bit smarter and adjust defaults depending
2675 * on available memory.
2677 for (order
= 0; ((1 << order
) << PAGE_SHIFT
) <
2678 (tcp_hashinfo
.bhash_size
* sizeof(struct inet_bind_hashbucket
));
2682 tcp_death_row
.sysctl_max_tw_buckets
= 180000;
2683 sysctl_tcp_max_orphans
= 4096 << (order
- 4);
2684 sysctl_max_syn_backlog
= 1024;
2685 } else if (order
< 3) {
2686 tcp_death_row
.sysctl_max_tw_buckets
>>= (3 - order
);
2687 sysctl_tcp_max_orphans
>>= (3 - order
);
2688 sysctl_max_syn_backlog
= 128;
2691 /* Set the pressure threshold to be a fraction of global memory that
2692 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2693 * memory, with a floor of 128 pages.
2695 nr_pages
= totalram_pages
- totalhigh_pages
;
2696 limit
= min(nr_pages
, 1UL<<(28-PAGE_SHIFT
)) >> (20-PAGE_SHIFT
);
2697 limit
= (limit
* (nr_pages
>> (20-PAGE_SHIFT
))) >> (PAGE_SHIFT
-11);
2698 limit
= max(limit
, 128UL);
2699 sysctl_tcp_mem
[0] = limit
/ 4 * 3;
2700 sysctl_tcp_mem
[1] = limit
;
2701 sysctl_tcp_mem
[2] = sysctl_tcp_mem
[0] * 2;
2703 /* Set per-socket limits to no more than 1/128 the pressure threshold */
2704 limit
= ((unsigned long)sysctl_tcp_mem
[1]) << (PAGE_SHIFT
- 7);
2705 max_share
= min(4UL*1024*1024, limit
);
2707 sysctl_tcp_wmem
[0] = SK_MEM_QUANTUM
;
2708 sysctl_tcp_wmem
[1] = 16*1024;
2709 sysctl_tcp_wmem
[2] = max(64*1024, max_share
);
2711 sysctl_tcp_rmem
[0] = SK_MEM_QUANTUM
;
2712 sysctl_tcp_rmem
[1] = 87380;
2713 sysctl_tcp_rmem
[2] = max(87380, max_share
);
2715 printk(KERN_INFO
"TCP: Hash tables configured "
2716 "(established %d bind %d)\n",
2717 tcp_hashinfo
.ehash_size
, tcp_hashinfo
.bhash_size
);
2719 tcp_register_congestion_control(&tcp_reno
);
2722 EXPORT_SYMBOL(tcp_close
);
2723 EXPORT_SYMBOL(tcp_disconnect
);
2724 EXPORT_SYMBOL(tcp_getsockopt
);
2725 EXPORT_SYMBOL(tcp_ioctl
);
2726 EXPORT_SYMBOL(tcp_poll
);
2727 EXPORT_SYMBOL(tcp_read_sock
);
2728 EXPORT_SYMBOL(tcp_recvmsg
);
2729 EXPORT_SYMBOL(tcp_sendmsg
);
2730 EXPORT_SYMBOL(tcp_splice_read
);
2731 EXPORT_SYMBOL(tcp_sendpage
);
2732 EXPORT_SYMBOL(tcp_setsockopt
);
2733 EXPORT_SYMBOL(tcp_shutdown
);
2734 EXPORT_SYMBOL(tcp_statistics
);