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/module.h>
251 #include <linux/types.h>
252 #include <linux/fcntl.h>
253 #include <linux/poll.h>
254 #include <linux/init.h>
255 #include <linux/fs.h>
256 #include <linux/skbuff.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/cache.h>
263 #include <linux/err.h>
264 #include <linux/crypto.h>
266 #include <net/icmp.h>
268 #include <net/xfrm.h>
270 #include <net/netdma.h>
271 #include <net/sock.h>
273 #include <asm/uaccess.h>
274 #include <asm/ioctls.h>
276 int sysctl_tcp_fin_timeout __read_mostly
= TCP_FIN_TIMEOUT
;
278 DEFINE_SNMP_STAT(struct tcp_mib
, tcp_statistics
) __read_mostly
;
280 atomic_t tcp_orphan_count
= ATOMIC_INIT(0);
282 EXPORT_SYMBOL_GPL(tcp_orphan_count
);
284 int sysctl_tcp_mem
[3] __read_mostly
;
285 int sysctl_tcp_wmem
[3] __read_mostly
;
286 int sysctl_tcp_rmem
[3] __read_mostly
;
288 EXPORT_SYMBOL(sysctl_tcp_mem
);
289 EXPORT_SYMBOL(sysctl_tcp_rmem
);
290 EXPORT_SYMBOL(sysctl_tcp_wmem
);
292 atomic_t tcp_memory_allocated
; /* Current allocated memory. */
293 atomic_t tcp_sockets_allocated
; /* Current number of TCP sockets. */
295 EXPORT_SYMBOL(tcp_memory_allocated
);
296 EXPORT_SYMBOL(tcp_sockets_allocated
);
301 struct tcp_splice_state
{
302 struct pipe_inode_info
*pipe
;
308 * Pressure flag: try to collapse.
309 * Technical note: it is used by multiple contexts non atomically.
310 * All the sk_stream_mem_schedule() is of this nature: accounting
311 * is strict, actions are advisory and have some latency.
313 int tcp_memory_pressure __read_mostly
;
315 EXPORT_SYMBOL(tcp_memory_pressure
);
317 void tcp_enter_memory_pressure(void)
319 if (!tcp_memory_pressure
) {
320 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES
);
321 tcp_memory_pressure
= 1;
325 EXPORT_SYMBOL(tcp_enter_memory_pressure
);
327 /* Convert seconds to retransmits based on initial and max timeout */
328 static u8
secs_to_retrans(int seconds
, int timeout
, int rto_max
)
333 int period
= timeout
;
336 while (seconds
> period
&& res
< 255) {
339 if (timeout
> rto_max
)
347 /* Convert retransmits to seconds based on initial and max timeout */
348 static int retrans_to_secs(u8 retrans
, int timeout
, int rto_max
)
356 if (timeout
> rto_max
)
365 * Wait for a TCP event.
367 * Note that we don't need to lock the socket, as the upper poll layers
368 * take care of normal races (between the test and the event) and we don't
369 * go look at any of the socket buffers directly.
371 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
374 struct sock
*sk
= sock
->sk
;
375 struct tcp_sock
*tp
= tcp_sk(sk
);
377 poll_wait(file
, sk
->sk_sleep
, wait
);
378 if (sk
->sk_state
== TCP_LISTEN
)
379 return inet_csk_listen_poll(sk
);
381 /* Socket is not locked. We are protected from async events
382 by poll logic and correct handling of state changes
383 made by another threads is impossible in any case.
391 * POLLHUP is certainly not done right. But poll() doesn't
392 * have a notion of HUP in just one direction, and for a
393 * socket the read side is more interesting.
395 * Some poll() documentation says that POLLHUP is incompatible
396 * with the POLLOUT/POLLWR flags, so somebody should check this
397 * all. But careful, it tends to be safer to return too many
398 * bits than too few, and you can easily break real applications
399 * if you don't tell them that something has hung up!
403 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
404 * our fs/select.c). It means that after we received EOF,
405 * poll always returns immediately, making impossible poll() on write()
406 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
407 * if and only if shutdown has been made in both directions.
408 * Actually, it is interesting to look how Solaris and DUX
409 * solve this dilemma. I would prefer, if PULLHUP were maskable,
410 * then we could set it on SND_SHUTDOWN. BTW examples given
411 * in Stevens' books assume exactly this behaviour, it explains
412 * why PULLHUP is incompatible with POLLOUT. --ANK
414 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
415 * blocking on fresh not-connected or disconnected socket. --ANK
417 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| sk
->sk_state
== TCP_CLOSE
)
419 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
420 mask
|= POLLIN
| POLLRDNORM
| POLLRDHUP
;
423 if ((1 << sk
->sk_state
) & ~(TCPF_SYN_SENT
| TCPF_SYN_RECV
)) {
424 /* Potential race condition. If read of tp below will
425 * escape above sk->sk_state, we can be illegally awaken
426 * in SYN_* states. */
427 if ((tp
->rcv_nxt
!= tp
->copied_seq
) &&
428 (tp
->urg_seq
!= tp
->copied_seq
||
429 tp
->rcv_nxt
!= tp
->copied_seq
+ 1 ||
430 sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
))
431 mask
|= POLLIN
| POLLRDNORM
;
433 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
434 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
)) {
435 mask
|= POLLOUT
| POLLWRNORM
;
436 } else { /* send SIGIO later */
437 set_bit(SOCK_ASYNC_NOSPACE
,
438 &sk
->sk_socket
->flags
);
439 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
441 /* Race breaker. If space is freed after
442 * wspace test but before the flags are set,
443 * IO signal will be lost.
445 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
))
446 mask
|= POLLOUT
| POLLWRNORM
;
449 mask
|= POLLOUT
| POLLWRNORM
;
451 if (tp
->urg_data
& TCP_URG_VALID
)
457 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
459 struct tcp_sock
*tp
= tcp_sk(sk
);
464 if (sk
->sk_state
== TCP_LISTEN
)
468 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
470 else if (sock_flag(sk
, SOCK_URGINLINE
) ||
472 before(tp
->urg_seq
, tp
->copied_seq
) ||
473 !before(tp
->urg_seq
, tp
->rcv_nxt
)) {
474 answ
= tp
->rcv_nxt
- tp
->copied_seq
;
476 /* Subtract 1, if FIN is in queue. */
477 if (answ
&& !skb_queue_empty(&sk
->sk_receive_queue
))
479 tcp_hdr((struct sk_buff
*)sk
->sk_receive_queue
.prev
)->fin
;
481 answ
= tp
->urg_seq
- tp
->copied_seq
;
485 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
488 if (sk
->sk_state
== TCP_LISTEN
)
491 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
494 answ
= tp
->write_seq
- tp
->snd_una
;
500 return put_user(answ
, (int __user
*)arg
);
503 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
505 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
506 tp
->pushed_seq
= tp
->write_seq
;
509 static inline int forced_push(struct tcp_sock
*tp
)
511 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
514 static inline void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
516 struct tcp_sock
*tp
= tcp_sk(sk
);
517 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
520 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
521 tcb
->flags
= TCPCB_FLAG_ACK
;
523 skb_header_release(skb
);
524 tcp_add_write_queue_tail(sk
, skb
);
525 sk_charge_skb(sk
, skb
);
526 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
527 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
530 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
,
533 if (flags
& MSG_OOB
) {
535 tp
->snd_up
= tp
->write_seq
;
536 TCP_SKB_CB(skb
)->sacked
|= TCPCB_URG
;
540 static inline void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
543 struct tcp_sock
*tp
= tcp_sk(sk
);
545 if (tcp_send_head(sk
)) {
546 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
547 if (!(flags
& MSG_MORE
) || forced_push(tp
))
548 tcp_mark_push(tp
, skb
);
549 tcp_mark_urg(tp
, flags
, skb
);
550 __tcp_push_pending_frames(sk
, mss_now
,
551 (flags
& MSG_MORE
) ? TCP_NAGLE_CORK
: nonagle
);
555 int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
556 unsigned int offset
, size_t len
)
558 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
561 ret
= skb_splice_bits(skb
, offset
, tss
->pipe
, min(rd_desc
->count
, len
),
564 rd_desc
->count
-= ret
;
568 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
570 /* Store TCP splice context information in read_descriptor_t. */
571 read_descriptor_t rd_desc
= {
576 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
580 * tcp_splice_read - splice data from TCP socket to a pipe
581 * @sock: socket to splice from
582 * @ppos: position (not valid)
583 * @pipe: pipe to splice to
584 * @len: number of bytes to splice
585 * @flags: splice modifier flags
588 * Will read pages from given socket and fill them into a pipe.
591 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
592 struct pipe_inode_info
*pipe
, size_t len
,
595 struct sock
*sk
= sock
->sk
;
596 struct tcp_splice_state tss
= {
606 * We can't seek on a socket input
615 timeo
= sock_rcvtimeo(sk
, flags
& SPLICE_F_NONBLOCK
);
617 ret
= __tcp_splice_read(sk
, &tss
);
623 if (flags
& SPLICE_F_NONBLOCK
) {
627 if (sock_flag(sk
, SOCK_DONE
))
630 ret
= sock_error(sk
);
633 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
635 if (sk
->sk_state
== TCP_CLOSE
) {
637 * This occurs when user tries to read
638 * from never connected socket.
640 if (!sock_flag(sk
, SOCK_DONE
))
648 sk_wait_data(sk
, &timeo
);
649 if (signal_pending(current
)) {
650 ret
= sock_intr_errno(timeo
);
663 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
664 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
665 signal_pending(current
))
677 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
)
681 /* The TCP header must be at least 32-bit aligned. */
682 size
= ALIGN(size
, 4);
684 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
686 if (sk_stream_wmem_schedule(sk
, skb
->truesize
)) {
688 * Make sure that we have exactly size bytes
689 * available to the caller, no more, no less.
691 skb_reserve(skb
, skb_tailroom(skb
) - size
);
696 sk
->sk_prot
->enter_memory_pressure();
697 sk_stream_moderate_sndbuf(sk
);
702 static ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
**pages
, int poffset
,
703 size_t psize
, int flags
)
705 struct tcp_sock
*tp
= tcp_sk(sk
);
706 int mss_now
, size_goal
;
709 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
711 /* Wait for a connection to finish. */
712 if ((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
713 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
716 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
718 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
719 size_goal
= tp
->xmit_size_goal
;
723 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
727 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
728 struct page
*page
= pages
[poffset
/ PAGE_SIZE
];
729 int copy
, i
, can_coalesce
;
730 int offset
= poffset
% PAGE_SIZE
;
731 int size
= min_t(size_t, psize
, PAGE_SIZE
- offset
);
733 if (!tcp_send_head(sk
) || (copy
= size_goal
- skb
->len
) <= 0) {
735 if (!sk_stream_memory_free(sk
))
736 goto wait_for_sndbuf
;
738 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
740 goto wait_for_memory
;
749 i
= skb_shinfo(skb
)->nr_frags
;
750 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
751 if (!can_coalesce
&& i
>= MAX_SKB_FRAGS
) {
752 tcp_mark_push(tp
, skb
);
755 if (!sk_stream_wmem_schedule(sk
, copy
))
756 goto wait_for_memory
;
759 skb_shinfo(skb
)->frags
[i
- 1].size
+= copy
;
762 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
766 skb
->data_len
+= copy
;
767 skb
->truesize
+= copy
;
768 sk
->sk_wmem_queued
+= copy
;
769 sk
->sk_forward_alloc
-= copy
;
770 skb
->ip_summed
= CHECKSUM_PARTIAL
;
771 tp
->write_seq
+= copy
;
772 TCP_SKB_CB(skb
)->end_seq
+= copy
;
773 skb_shinfo(skb
)->gso_segs
= 0;
776 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_PSH
;
780 if (!(psize
-= copy
))
783 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
786 if (forced_push(tp
)) {
787 tcp_mark_push(tp
, skb
);
788 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
789 } else if (skb
== tcp_send_head(sk
))
790 tcp_push_one(sk
, mss_now
);
794 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
797 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
799 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
802 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
803 size_goal
= tp
->xmit_size_goal
;
808 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
815 return sk_stream_error(sk
, flags
, err
);
818 ssize_t
tcp_sendpage(struct socket
*sock
, struct page
*page
, int offset
,
819 size_t size
, int flags
)
822 struct sock
*sk
= sock
->sk
;
824 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
825 !(sk
->sk_route_caps
& NETIF_F_ALL_CSUM
))
826 return sock_no_sendpage(sock
, page
, offset
, size
, flags
);
830 res
= do_tcp_sendpages(sk
, &page
, offset
, size
, flags
);
836 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
837 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
839 static inline int select_size(struct sock
*sk
)
841 struct tcp_sock
*tp
= tcp_sk(sk
);
842 int tmp
= tp
->mss_cache
;
844 if (sk
->sk_route_caps
& NETIF_F_SG
) {
848 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
850 if (tmp
>= pgbreak
&&
851 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
859 int tcp_sendmsg(struct kiocb
*iocb
, struct socket
*sock
, struct msghdr
*msg
,
862 struct sock
*sk
= sock
->sk
;
864 struct tcp_sock
*tp
= tcp_sk(sk
);
867 int mss_now
, size_goal
;
874 flags
= msg
->msg_flags
;
875 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
877 /* Wait for a connection to finish. */
878 if ((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
879 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
882 /* This should be in poll */
883 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
885 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
886 size_goal
= tp
->xmit_size_goal
;
888 /* Ok commence sending. */
889 iovlen
= msg
->msg_iovlen
;
894 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
897 while (--iovlen
>= 0) {
898 int seglen
= iov
->iov_len
;
899 unsigned char __user
*from
= iov
->iov_base
;
907 skb
= tcp_write_queue_tail(sk
);
908 if (tcp_send_head(sk
)) {
909 if (skb
->ip_summed
== CHECKSUM_NONE
)
911 copy
= max
- skb
->len
;
916 /* Allocate new segment. If the interface is SG,
917 * allocate skb fitting to single page.
919 if (!sk_stream_memory_free(sk
))
920 goto wait_for_sndbuf
;
922 skb
= sk_stream_alloc_skb(sk
, select_size(sk
),
925 goto wait_for_memory
;
928 * Check whether we can use HW checksum.
930 if (sk
->sk_route_caps
& NETIF_F_ALL_CSUM
)
931 skb
->ip_summed
= CHECKSUM_PARTIAL
;
938 /* Try to append data to the end of skb. */
942 /* Where to copy to? */
943 if (skb_tailroom(skb
) > 0) {
944 /* We have some space in skb head. Superb! */
945 if (copy
> skb_tailroom(skb
))
946 copy
= skb_tailroom(skb
);
947 if ((err
= skb_add_data(skb
, from
, copy
)) != 0)
951 int i
= skb_shinfo(skb
)->nr_frags
;
952 struct page
*page
= TCP_PAGE(sk
);
953 int off
= TCP_OFF(sk
);
955 if (skb_can_coalesce(skb
, i
, page
, off
) &&
957 /* We can extend the last page
960 } else if (i
== MAX_SKB_FRAGS
||
962 !(sk
->sk_route_caps
& NETIF_F_SG
))) {
963 /* Need to add new fragment and cannot
964 * do this because interface is non-SG,
965 * or because all the page slots are
967 tcp_mark_push(tp
, skb
);
970 if (off
== PAGE_SIZE
) {
972 TCP_PAGE(sk
) = page
= NULL
;
978 if (copy
> PAGE_SIZE
- off
)
979 copy
= PAGE_SIZE
- off
;
981 if (!sk_stream_wmem_schedule(sk
, copy
))
982 goto wait_for_memory
;
985 /* Allocate new cache page. */
986 if (!(page
= sk_stream_alloc_page(sk
)))
987 goto wait_for_memory
;
990 /* Time to copy data. We are close to
992 err
= skb_copy_to_page(sk
, from
, skb
, page
,
995 /* If this page was new, give it to the
996 * socket so it does not get leaked.
1005 /* Update the skb. */
1007 skb_shinfo(skb
)->frags
[i
- 1].size
+=
1010 skb_fill_page_desc(skb
, i
, page
, off
, copy
);
1013 } else if (off
+ copy
< PAGE_SIZE
) {
1015 TCP_PAGE(sk
) = page
;
1019 TCP_OFF(sk
) = off
+ copy
;
1023 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_PSH
;
1025 tp
->write_seq
+= copy
;
1026 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1027 skb_shinfo(skb
)->gso_segs
= 0;
1031 if ((seglen
-= copy
) == 0 && iovlen
== 0)
1034 if (skb
->len
< max
|| (flags
& MSG_OOB
))
1037 #ifdef CONFIG_INET_GSO
1038 if (iov
->iov_len
> PAGE_SIZE
)
1040 #endif /* CONFIG_INET_GSO */
1042 if (forced_push(tp
)) {
1043 tcp_mark_push(tp
, skb
);
1044 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1045 } else if (skb
== tcp_send_head(sk
))
1046 tcp_push_one(sk
, mss_now
);
1050 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1053 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
1055 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
1058 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
1059 size_goal
= tp
->xmit_size_goal
;
1065 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
1066 TCP_CHECK_TIMER(sk
);
1072 tcp_unlink_write_queue(skb
, sk
);
1073 /* It is the one place in all of TCP, except connection
1074 * reset, where we can be unlinking the send_head.
1076 tcp_check_send_head(sk
, skb
);
1077 sk_stream_free_skb(sk
, skb
);
1084 err
= sk_stream_error(sk
, flags
, err
);
1085 TCP_CHECK_TIMER(sk
);
1091 * Handle reading urgent data. BSD has very simple semantics for
1092 * this, no blocking and very strange errors 8)
1095 static int tcp_recv_urg(struct sock
*sk
, long timeo
,
1096 struct msghdr
*msg
, int len
, int flags
,
1099 struct tcp_sock
*tp
= tcp_sk(sk
);
1101 /* No URG data to read. */
1102 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1103 tp
->urg_data
== TCP_URG_READ
)
1104 return -EINVAL
; /* Yes this is right ! */
1106 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1109 if (tp
->urg_data
& TCP_URG_VALID
) {
1111 char c
= tp
->urg_data
;
1113 if (!(flags
& MSG_PEEK
))
1114 tp
->urg_data
= TCP_URG_READ
;
1116 /* Read urgent data. */
1117 msg
->msg_flags
|= MSG_OOB
;
1120 if (!(flags
& MSG_TRUNC
))
1121 err
= memcpy_toiovec(msg
->msg_iov
, &c
, 1);
1124 msg
->msg_flags
|= MSG_TRUNC
;
1126 return err
? -EFAULT
: len
;
1129 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1132 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1133 * the available implementations agree in this case:
1134 * this call should never block, independent of the
1135 * blocking state of the socket.
1136 * Mike <pall@rz.uni-karlsruhe.de>
1141 /* Clean up the receive buffer for full frames taken by the user,
1142 * then send an ACK if necessary. COPIED is the number of bytes
1143 * tcp_recvmsg has given to the user so far, it speeds up the
1144 * calculation of whether or not we must ACK for the sake of
1147 void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1149 struct tcp_sock
*tp
= tcp_sk(sk
);
1150 int time_to_ack
= 0;
1153 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1155 BUG_TRAP(!skb
|| before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
));
1158 if (inet_csk_ack_scheduled(sk
)) {
1159 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1160 /* Delayed ACKs frequently hit locked sockets during bulk
1162 if (icsk
->icsk_ack
.blocked
||
1163 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1164 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1166 * If this read emptied read buffer, we send ACK, if
1167 * connection is not bidirectional, user drained
1168 * receive buffer and there was a small segment
1172 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1173 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1174 !icsk
->icsk_ack
.pingpong
)) &&
1175 !atomic_read(&sk
->sk_rmem_alloc
)))
1179 /* We send an ACK if we can now advertise a non-zero window
1180 * which has been raised "significantly".
1182 * Even if window raised up to infinity, do not send window open ACK
1183 * in states, where we will not receive more. It is useless.
1185 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1186 __u32 rcv_window_now
= tcp_receive_window(tp
);
1188 /* Optimize, __tcp_select_window() is not cheap. */
1189 if (2*rcv_window_now
<= tp
->window_clamp
) {
1190 __u32 new_window
= __tcp_select_window(sk
);
1192 /* Send ACK now, if this read freed lots of space
1193 * in our buffer. Certainly, new_window is new window.
1194 * We can advertise it now, if it is not less than current one.
1195 * "Lots" means "at least twice" here.
1197 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1205 static void tcp_prequeue_process(struct sock
*sk
)
1207 struct sk_buff
*skb
;
1208 struct tcp_sock
*tp
= tcp_sk(sk
);
1210 NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED
);
1212 /* RX process wants to run with disabled BHs, though it is not
1215 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
1216 sk
->sk_backlog_rcv(sk
, skb
);
1219 /* Clear memory counter. */
1220 tp
->ucopy
.memory
= 0;
1223 static inline struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1225 struct sk_buff
*skb
;
1228 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1229 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1230 if (tcp_hdr(skb
)->syn
)
1232 if (offset
< skb
->len
|| tcp_hdr(skb
)->fin
) {
1241 * This routine provides an alternative to tcp_recvmsg() for routines
1242 * that would like to handle copying from skbuffs directly in 'sendfile'
1245 * - It is assumed that the socket was locked by the caller.
1246 * - The routine does not block.
1247 * - At present, there is no support for reading OOB data
1248 * or for 'peeking' the socket using this routine
1249 * (although both would be easy to implement).
1251 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1252 sk_read_actor_t recv_actor
)
1254 struct sk_buff
*skb
;
1255 struct tcp_sock
*tp
= tcp_sk(sk
);
1256 u32 seq
= tp
->copied_seq
;
1260 if (sk
->sk_state
== TCP_LISTEN
)
1262 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1263 if (offset
< skb
->len
) {
1266 len
= skb
->len
- offset
;
1267 /* Stop reading if we hit a patch of urgent data */
1269 u32 urg_offset
= tp
->urg_seq
- seq
;
1270 if (urg_offset
< len
)
1275 used
= recv_actor(desc
, skb
, offset
, len
);
1280 } else if (used
<= len
) {
1286 * If recv_actor drops the lock (e.g. TCP splice
1287 * receive) the skb pointer might be invalid when
1288 * getting here: tcp_collapse might have deleted it
1289 * while aggregating skbs from the socket queue.
1291 skb
= tcp_recv_skb(sk
, seq
-1, &offset
);
1292 if (!skb
|| (offset
+1 != skb
->len
))
1295 if (tcp_hdr(skb
)->fin
) {
1296 sk_eat_skb(sk
, skb
, 0);
1300 sk_eat_skb(sk
, skb
, 0);
1303 tp
->copied_seq
= seq
;
1305 tp
->copied_seq
= seq
;
1307 tcp_rcv_space_adjust(sk
);
1309 /* Clean up data we have read: This will do ACK frames. */
1311 tcp_cleanup_rbuf(sk
, copied
);
1316 * This routine copies from a sock struct into the user buffer.
1318 * Technical note: in 2.3 we work on _locked_ socket, so that
1319 * tricks with *seq access order and skb->users are not required.
1320 * Probably, code can be easily improved even more.
1323 int tcp_recvmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1324 size_t len
, int nonblock
, int flags
, int *addr_len
)
1326 struct tcp_sock
*tp
= tcp_sk(sk
);
1332 int target
; /* Read at least this many bytes */
1334 struct task_struct
*user_recv
= NULL
;
1335 int copied_early
= 0;
1340 TCP_CHECK_TIMER(sk
);
1343 if (sk
->sk_state
== TCP_LISTEN
)
1346 timeo
= sock_rcvtimeo(sk
, nonblock
);
1348 /* Urgent data needs to be handled specially. */
1349 if (flags
& MSG_OOB
)
1352 seq
= &tp
->copied_seq
;
1353 if (flags
& MSG_PEEK
) {
1354 peek_seq
= tp
->copied_seq
;
1358 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1360 #ifdef CONFIG_NET_DMA
1361 tp
->ucopy
.dma_chan
= NULL
;
1363 if ((len
> sysctl_tcp_dma_copybreak
) && !(flags
& MSG_PEEK
) &&
1364 !sysctl_tcp_low_latency
&& __get_cpu_var(softnet_data
).net_dma
) {
1365 preempt_enable_no_resched();
1366 tp
->ucopy
.pinned_list
= dma_pin_iovec_pages(msg
->msg_iov
, len
);
1368 preempt_enable_no_resched();
1372 struct sk_buff
*skb
;
1375 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1376 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1379 if (signal_pending(current
)) {
1380 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1385 /* Next get a buffer. */
1387 skb
= skb_peek(&sk
->sk_receive_queue
);
1392 /* Now that we have two receive queues this
1395 if (before(*seq
, TCP_SKB_CB(skb
)->seq
)) {
1396 printk(KERN_INFO
"recvmsg bug: copied %X "
1397 "seq %X\n", *seq
, TCP_SKB_CB(skb
)->seq
);
1400 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1401 if (tcp_hdr(skb
)->syn
)
1403 if (offset
< skb
->len
)
1405 if (tcp_hdr(skb
)->fin
)
1407 BUG_TRAP(flags
& MSG_PEEK
);
1409 } while (skb
!= (struct sk_buff
*)&sk
->sk_receive_queue
);
1411 /* Well, if we have backlog, try to process it now yet. */
1413 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1418 sk
->sk_state
== TCP_CLOSE
||
1419 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1421 signal_pending(current
))
1424 if (sock_flag(sk
, SOCK_DONE
))
1428 copied
= sock_error(sk
);
1432 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1435 if (sk
->sk_state
== TCP_CLOSE
) {
1436 if (!sock_flag(sk
, SOCK_DONE
)) {
1437 /* This occurs when user tries to read
1438 * from never connected socket.
1451 if (signal_pending(current
)) {
1452 copied
= sock_intr_errno(timeo
);
1457 tcp_cleanup_rbuf(sk
, copied
);
1459 if (!sysctl_tcp_low_latency
&& tp
->ucopy
.task
== user_recv
) {
1460 /* Install new reader */
1461 if (!user_recv
&& !(flags
& (MSG_TRUNC
| MSG_PEEK
))) {
1462 user_recv
= current
;
1463 tp
->ucopy
.task
= user_recv
;
1464 tp
->ucopy
.iov
= msg
->msg_iov
;
1467 tp
->ucopy
.len
= len
;
1469 BUG_TRAP(tp
->copied_seq
== tp
->rcv_nxt
||
1470 (flags
& (MSG_PEEK
| MSG_TRUNC
)));
1472 /* Ugly... If prequeue is not empty, we have to
1473 * process it before releasing socket, otherwise
1474 * order will be broken at second iteration.
1475 * More elegant solution is required!!!
1477 * Look: we have the following (pseudo)queues:
1479 * 1. packets in flight
1484 * Each queue can be processed only if the next ones
1485 * are empty. At this point we have empty receive_queue.
1486 * But prequeue _can_ be not empty after 2nd iteration,
1487 * when we jumped to start of loop because backlog
1488 * processing added something to receive_queue.
1489 * We cannot release_sock(), because backlog contains
1490 * packets arrived _after_ prequeued ones.
1492 * Shortly, algorithm is clear --- to process all
1493 * the queues in order. We could make it more directly,
1494 * requeueing packets from backlog to prequeue, if
1495 * is not empty. It is more elegant, but eats cycles,
1498 if (!skb_queue_empty(&tp
->ucopy
.prequeue
))
1501 /* __ Set realtime policy in scheduler __ */
1504 if (copied
>= target
) {
1505 /* Do not sleep, just process backlog. */
1509 sk_wait_data(sk
, &timeo
);
1511 #ifdef CONFIG_NET_DMA
1512 tp
->ucopy
.wakeup
= 0;
1518 /* __ Restore normal policy in scheduler __ */
1520 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1521 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG
, chunk
);
1526 if (tp
->rcv_nxt
== tp
->copied_seq
&&
1527 !skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1529 tcp_prequeue_process(sk
);
1531 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1532 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1538 if ((flags
& MSG_PEEK
) &&
1539 (peek_seq
- copied
- urg_hole
!= tp
->copied_seq
)) {
1540 if (net_ratelimit())
1541 printk(KERN_DEBUG
"TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1542 current
->comm
, current
->pid
);
1543 peek_seq
= tp
->copied_seq
;
1548 /* Ok so how much can we use? */
1549 used
= skb
->len
- offset
;
1553 /* Do we have urgent data here? */
1555 u32 urg_offset
= tp
->urg_seq
- *seq
;
1556 if (urg_offset
< used
) {
1558 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1571 if (!(flags
& MSG_TRUNC
)) {
1572 #ifdef CONFIG_NET_DMA
1573 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
1574 tp
->ucopy
.dma_chan
= get_softnet_dma();
1576 if (tp
->ucopy
.dma_chan
) {
1577 tp
->ucopy
.dma_cookie
= dma_skb_copy_datagram_iovec(
1578 tp
->ucopy
.dma_chan
, skb
, offset
,
1580 tp
->ucopy
.pinned_list
);
1582 if (tp
->ucopy
.dma_cookie
< 0) {
1584 printk(KERN_ALERT
"dma_cookie < 0\n");
1586 /* Exception. Bailout! */
1591 if ((offset
+ used
) == skb
->len
)
1597 err
= skb_copy_datagram_iovec(skb
, offset
,
1598 msg
->msg_iov
, used
);
1600 /* Exception. Bailout! */
1612 tcp_rcv_space_adjust(sk
);
1615 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1617 tcp_fast_path_check(sk
);
1619 if (used
+ offset
< skb
->len
)
1622 if (tcp_hdr(skb
)->fin
)
1624 if (!(flags
& MSG_PEEK
)) {
1625 sk_eat_skb(sk
, skb
, copied_early
);
1631 /* Process the FIN. */
1633 if (!(flags
& MSG_PEEK
)) {
1634 sk_eat_skb(sk
, skb
, copied_early
);
1641 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1644 tp
->ucopy
.len
= copied
> 0 ? len
: 0;
1646 tcp_prequeue_process(sk
);
1648 if (copied
> 0 && (chunk
= len
- tp
->ucopy
.len
) != 0) {
1649 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1655 tp
->ucopy
.task
= NULL
;
1659 #ifdef CONFIG_NET_DMA
1660 if (tp
->ucopy
.dma_chan
) {
1661 struct sk_buff
*skb
;
1662 dma_cookie_t done
, used
;
1664 dma_async_memcpy_issue_pending(tp
->ucopy
.dma_chan
);
1666 while (dma_async_memcpy_complete(tp
->ucopy
.dma_chan
,
1667 tp
->ucopy
.dma_cookie
, &done
,
1668 &used
) == DMA_IN_PROGRESS
) {
1669 /* do partial cleanup of sk_async_wait_queue */
1670 while ((skb
= skb_peek(&sk
->sk_async_wait_queue
)) &&
1671 (dma_async_is_complete(skb
->dma_cookie
, done
,
1672 used
) == DMA_SUCCESS
)) {
1673 __skb_dequeue(&sk
->sk_async_wait_queue
);
1678 /* Safe to free early-copied skbs now */
1679 __skb_queue_purge(&sk
->sk_async_wait_queue
);
1680 dma_chan_put(tp
->ucopy
.dma_chan
);
1681 tp
->ucopy
.dma_chan
= NULL
;
1683 if (tp
->ucopy
.pinned_list
) {
1684 dma_unpin_iovec_pages(tp
->ucopy
.pinned_list
);
1685 tp
->ucopy
.pinned_list
= NULL
;
1689 /* According to UNIX98, msg_name/msg_namelen are ignored
1690 * on connected socket. I was just happy when found this 8) --ANK
1693 /* Clean up data we have read: This will do ACK frames. */
1694 tcp_cleanup_rbuf(sk
, copied
);
1696 TCP_CHECK_TIMER(sk
);
1701 TCP_CHECK_TIMER(sk
);
1706 err
= tcp_recv_urg(sk
, timeo
, msg
, len
, flags
, addr_len
);
1710 void tcp_set_state(struct sock
*sk
, int state
)
1712 int oldstate
= sk
->sk_state
;
1715 case TCP_ESTABLISHED
:
1716 if (oldstate
!= TCP_ESTABLISHED
)
1717 TCP_INC_STATS(TCP_MIB_CURRESTAB
);
1721 if (oldstate
== TCP_CLOSE_WAIT
|| oldstate
== TCP_ESTABLISHED
)
1722 TCP_INC_STATS(TCP_MIB_ESTABRESETS
);
1724 sk
->sk_prot
->unhash(sk
);
1725 if (inet_csk(sk
)->icsk_bind_hash
&&
1726 !(sk
->sk_userlocks
& SOCK_BINDPORT_LOCK
))
1727 inet_put_port(&tcp_hashinfo
, sk
);
1730 if (oldstate
==TCP_ESTABLISHED
)
1731 TCP_DEC_STATS(TCP_MIB_CURRESTAB
);
1734 /* Change state AFTER socket is unhashed to avoid closed
1735 * socket sitting in hash tables.
1737 sk
->sk_state
= state
;
1740 SOCK_DEBUG(sk
, "TCP sk=%p, State %s -> %s\n",sk
, statename
[oldstate
],statename
[state
]);
1743 EXPORT_SYMBOL_GPL(tcp_set_state
);
1746 * State processing on a close. This implements the state shift for
1747 * sending our FIN frame. Note that we only send a FIN for some
1748 * states. A shutdown() may have already sent the FIN, or we may be
1752 static const unsigned char new_state
[16] = {
1753 /* current state: new state: action: */
1754 /* (Invalid) */ TCP_CLOSE
,
1755 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1756 /* TCP_SYN_SENT */ TCP_CLOSE
,
1757 /* TCP_SYN_RECV */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1758 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1
,
1759 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2
,
1760 /* TCP_TIME_WAIT */ TCP_CLOSE
,
1761 /* TCP_CLOSE */ TCP_CLOSE
,
1762 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK
| TCP_ACTION_FIN
,
1763 /* TCP_LAST_ACK */ TCP_LAST_ACK
,
1764 /* TCP_LISTEN */ TCP_CLOSE
,
1765 /* TCP_CLOSING */ TCP_CLOSING
,
1768 static int tcp_close_state(struct sock
*sk
)
1770 int next
= (int)new_state
[sk
->sk_state
];
1771 int ns
= next
& TCP_STATE_MASK
;
1773 tcp_set_state(sk
, ns
);
1775 return next
& TCP_ACTION_FIN
;
1779 * Shutdown the sending side of a connection. Much like close except
1780 * that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1783 void tcp_shutdown(struct sock
*sk
, int how
)
1785 /* We need to grab some memory, and put together a FIN,
1786 * and then put it into the queue to be sent.
1787 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1789 if (!(how
& SEND_SHUTDOWN
))
1792 /* If we've already sent a FIN, or it's a closed state, skip this. */
1793 if ((1 << sk
->sk_state
) &
1794 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
1795 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
1796 /* Clear out any half completed packets. FIN if needed. */
1797 if (tcp_close_state(sk
))
1802 void tcp_close(struct sock
*sk
, long timeout
)
1804 struct sk_buff
*skb
;
1805 int data_was_unread
= 0;
1809 sk
->sk_shutdown
= SHUTDOWN_MASK
;
1811 if (sk
->sk_state
== TCP_LISTEN
) {
1812 tcp_set_state(sk
, TCP_CLOSE
);
1815 inet_csk_listen_stop(sk
);
1817 goto adjudge_to_death
;
1820 /* We need to flush the recv. buffs. We do this only on the
1821 * descriptor close, not protocol-sourced closes, because the
1822 * reader process may not have drained the data yet!
1824 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
1825 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
-
1827 data_was_unread
+= len
;
1831 sk_stream_mem_reclaim(sk
);
1833 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1834 if (sk
->sk_state
== TCP_CLOSE
)
1835 goto adjudge_to_death
;
1837 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1838 * data was lost. To witness the awful effects of the old behavior of
1839 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1840 * GET in an FTP client, suspend the process, wait for the client to
1841 * advertise a zero window, then kill -9 the FTP client, wheee...
1842 * Note: timeout is always zero in such a case.
1844 if (data_was_unread
) {
1845 /* Unread data was tossed, zap the connection. */
1846 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE
);
1847 tcp_set_state(sk
, TCP_CLOSE
);
1848 tcp_send_active_reset(sk
, GFP_KERNEL
);
1849 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
1850 /* Check zero linger _after_ checking for unread data. */
1851 sk
->sk_prot
->disconnect(sk
, 0);
1852 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA
);
1853 } else if (tcp_close_state(sk
)) {
1854 /* We FIN if the application ate all the data before
1855 * zapping the connection.
1858 /* RED-PEN. Formally speaking, we have broken TCP state
1859 * machine. State transitions:
1861 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1862 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1863 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1865 * are legal only when FIN has been sent (i.e. in window),
1866 * rather than queued out of window. Purists blame.
1868 * F.e. "RFC state" is ESTABLISHED,
1869 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1871 * The visible declinations are that sometimes
1872 * we enter time-wait state, when it is not required really
1873 * (harmless), do not send active resets, when they are
1874 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1875 * they look as CLOSING or LAST_ACK for Linux)
1876 * Probably, I missed some more holelets.
1882 sk_stream_wait_close(sk
, timeout
);
1885 state
= sk
->sk_state
;
1888 atomic_inc(sk
->sk_prot
->orphan_count
);
1890 /* It is the last release_sock in its life. It will remove backlog. */
1894 /* Now socket is owned by kernel and we acquire BH lock
1895 to finish close. No need to check for user refs.
1899 BUG_TRAP(!sock_owned_by_user(sk
));
1901 /* Have we already been destroyed by a softirq or backlog? */
1902 if (state
!= TCP_CLOSE
&& sk
->sk_state
== TCP_CLOSE
)
1905 /* This is a (useful) BSD violating of the RFC. There is a
1906 * problem with TCP as specified in that the other end could
1907 * keep a socket open forever with no application left this end.
1908 * We use a 3 minute timeout (about the same as BSD) then kill
1909 * our end. If they send after that then tough - BUT: long enough
1910 * that we won't make the old 4*rto = almost no time - whoops
1913 * Nope, it was not mistake. It is really desired behaviour
1914 * f.e. on http servers, when such sockets are useless, but
1915 * consume significant resources. Let's do it with special
1916 * linger2 option. --ANK
1919 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
1920 struct tcp_sock
*tp
= tcp_sk(sk
);
1921 if (tp
->linger2
< 0) {
1922 tcp_set_state(sk
, TCP_CLOSE
);
1923 tcp_send_active_reset(sk
, GFP_ATOMIC
);
1924 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER
);
1926 const int tmo
= tcp_fin_time(sk
);
1928 if (tmo
> TCP_TIMEWAIT_LEN
) {
1929 inet_csk_reset_keepalive_timer(sk
,
1930 tmo
- TCP_TIMEWAIT_LEN
);
1932 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
1937 if (sk
->sk_state
!= TCP_CLOSE
) {
1938 sk_stream_mem_reclaim(sk
);
1939 if (tcp_too_many_orphans(sk
,
1940 atomic_read(sk
->sk_prot
->orphan_count
))) {
1941 if (net_ratelimit())
1942 printk(KERN_INFO
"TCP: too many of orphaned "
1944 tcp_set_state(sk
, TCP_CLOSE
);
1945 tcp_send_active_reset(sk
, GFP_ATOMIC
);
1946 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY
);
1950 if (sk
->sk_state
== TCP_CLOSE
)
1951 inet_csk_destroy_sock(sk
);
1952 /* Otherwise, socket is reprieved until protocol close. */
1960 /* These states need RST on ABORT according to RFC793 */
1962 static inline int tcp_need_reset(int state
)
1964 return (1 << state
) &
1965 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
1966 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
1969 int tcp_disconnect(struct sock
*sk
, int flags
)
1971 struct inet_sock
*inet
= inet_sk(sk
);
1972 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1973 struct tcp_sock
*tp
= tcp_sk(sk
);
1975 int old_state
= sk
->sk_state
;
1977 if (old_state
!= TCP_CLOSE
)
1978 tcp_set_state(sk
, TCP_CLOSE
);
1980 /* ABORT function of RFC793 */
1981 if (old_state
== TCP_LISTEN
) {
1982 inet_csk_listen_stop(sk
);
1983 } else if (tcp_need_reset(old_state
) ||
1984 (tp
->snd_nxt
!= tp
->write_seq
&&
1985 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
1986 /* The last check adjusts for discrepancy of Linux wrt. RFC
1989 tcp_send_active_reset(sk
, gfp_any());
1990 sk
->sk_err
= ECONNRESET
;
1991 } else if (old_state
== TCP_SYN_SENT
)
1992 sk
->sk_err
= ECONNRESET
;
1994 tcp_clear_xmit_timers(sk
);
1995 __skb_queue_purge(&sk
->sk_receive_queue
);
1996 tcp_write_queue_purge(sk
);
1997 __skb_queue_purge(&tp
->out_of_order_queue
);
1998 #ifdef CONFIG_NET_DMA
1999 __skb_queue_purge(&sk
->sk_async_wait_queue
);
2004 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
2005 inet_reset_saddr(sk
);
2007 sk
->sk_shutdown
= 0;
2008 sock_reset_flag(sk
, SOCK_DONE
);
2010 if ((tp
->write_seq
+= tp
->max_window
+ 2) == 0)
2012 icsk
->icsk_backoff
= 0;
2014 icsk
->icsk_probes_out
= 0;
2015 tp
->packets_out
= 0;
2016 tp
->snd_ssthresh
= 0x7fffffff;
2017 tp
->snd_cwnd_cnt
= 0;
2018 tp
->bytes_acked
= 0;
2019 tcp_set_ca_state(sk
, TCP_CA_Open
);
2020 tcp_clear_retrans(tp
);
2021 inet_csk_delack_init(sk
);
2022 tcp_init_send_head(sk
);
2023 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
2026 BUG_TRAP(!inet
->num
|| icsk
->icsk_bind_hash
);
2028 sk
->sk_error_report(sk
);
2033 * Socket option code for TCP.
2035 static int do_tcp_setsockopt(struct sock
*sk
, int level
,
2036 int optname
, char __user
*optval
, int optlen
)
2038 struct tcp_sock
*tp
= tcp_sk(sk
);
2039 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2043 /* This is a string value all the others are int's */
2044 if (optname
== TCP_CONGESTION
) {
2045 char name
[TCP_CA_NAME_MAX
];
2050 val
= strncpy_from_user(name
, optval
,
2051 min(TCP_CA_NAME_MAX
-1, optlen
));
2057 err
= tcp_set_congestion_control(sk
, name
);
2062 if (optlen
< sizeof(int))
2065 if (get_user(val
, (int __user
*)optval
))
2072 /* Values greater than interface MTU won't take effect. However
2073 * at the point when this call is done we typically don't yet
2074 * know which interface is going to be used */
2075 if (val
< TCP_MIN_MSS
|| val
> MAX_TCP_WINDOW
) {
2079 tp
->rx_opt
.user_mss
= val
;
2084 /* TCP_NODELAY is weaker than TCP_CORK, so that
2085 * this option on corked socket is remembered, but
2086 * it is not activated until cork is cleared.
2088 * However, when TCP_NODELAY is set we make
2089 * an explicit push, which overrides even TCP_CORK
2090 * for currently queued segments.
2092 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
2093 tcp_push_pending_frames(sk
);
2095 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
2100 /* When set indicates to always queue non-full frames.
2101 * Later the user clears this option and we transmit
2102 * any pending partial frames in the queue. This is
2103 * meant to be used alongside sendfile() to get properly
2104 * filled frames when the user (for example) must write
2105 * out headers with a write() call first and then use
2106 * sendfile to send out the data parts.
2108 * TCP_CORK can be set together with TCP_NODELAY and it is
2109 * stronger than TCP_NODELAY.
2112 tp
->nonagle
|= TCP_NAGLE_CORK
;
2114 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
2115 if (tp
->nonagle
&TCP_NAGLE_OFF
)
2116 tp
->nonagle
|= TCP_NAGLE_PUSH
;
2117 tcp_push_pending_frames(sk
);
2122 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
2125 tp
->keepalive_time
= val
* HZ
;
2126 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
2127 !((1 << sk
->sk_state
) &
2128 (TCPF_CLOSE
| TCPF_LISTEN
))) {
2129 u32 elapsed
= keepalive_time_elapsed(tp
);
2130 if (tp
->keepalive_time
> elapsed
)
2131 elapsed
= tp
->keepalive_time
- elapsed
;
2134 inet_csk_reset_keepalive_timer(sk
, elapsed
);
2139 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
2142 tp
->keepalive_intvl
= val
* HZ
;
2145 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
2148 tp
->keepalive_probes
= val
;
2151 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
2154 icsk
->icsk_syn_retries
= val
;
2160 else if (val
> sysctl_tcp_fin_timeout
/ HZ
)
2163 tp
->linger2
= val
* HZ
;
2166 case TCP_DEFER_ACCEPT
:
2167 /* Translate value in seconds to number of retransmits */
2168 icsk
->icsk_accept_queue
.rskq_defer_accept
=
2169 secs_to_retrans(val
, TCP_TIMEOUT_INIT
/ HZ
,
2173 case TCP_WINDOW_CLAMP
:
2175 if (sk
->sk_state
!= TCP_CLOSE
) {
2179 tp
->window_clamp
= 0;
2181 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
2182 SOCK_MIN_RCVBUF
/ 2 : val
;
2187 icsk
->icsk_ack
.pingpong
= 1;
2189 icsk
->icsk_ack
.pingpong
= 0;
2190 if ((1 << sk
->sk_state
) &
2191 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
2192 inet_csk_ack_scheduled(sk
)) {
2193 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
2194 tcp_cleanup_rbuf(sk
, 1);
2196 icsk
->icsk_ack
.pingpong
= 1;
2201 #ifdef CONFIG_TCP_MD5SIG
2203 /* Read the IP->Key mappings from userspace */
2204 err
= tp
->af_specific
->md5_parse(sk
, optval
, optlen
);
2217 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2220 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2222 if (level
!= SOL_TCP
)
2223 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
2225 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2228 #ifdef CONFIG_COMPAT
2229 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
2230 char __user
*optval
, int optlen
)
2232 if (level
!= SOL_TCP
)
2233 return inet_csk_compat_setsockopt(sk
, level
, optname
,
2235 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2238 EXPORT_SYMBOL(compat_tcp_setsockopt
);
2241 /* Return information about state of tcp endpoint in API format. */
2242 void tcp_get_info(struct sock
*sk
, struct tcp_info
*info
)
2244 struct tcp_sock
*tp
= tcp_sk(sk
);
2245 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2246 u32 now
= tcp_time_stamp
;
2248 memset(info
, 0, sizeof(*info
));
2250 info
->tcpi_state
= sk
->sk_state
;
2251 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
2252 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
2253 info
->tcpi_probes
= icsk
->icsk_probes_out
;
2254 info
->tcpi_backoff
= icsk
->icsk_backoff
;
2256 if (tp
->rx_opt
.tstamp_ok
)
2257 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
2258 if (tp
->rx_opt
.sack_ok
)
2259 info
->tcpi_options
|= TCPI_OPT_SACK
;
2260 if (tp
->rx_opt
.wscale_ok
) {
2261 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
2262 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
2263 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
2266 if (tp
->ecn_flags
&TCP_ECN_OK
)
2267 info
->tcpi_options
|= TCPI_OPT_ECN
;
2269 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
2270 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
2271 info
->tcpi_snd_mss
= tp
->mss_cache
;
2272 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
2274 info
->tcpi_unacked
= tp
->packets_out
;
2275 info
->tcpi_sacked
= tp
->sacked_out
;
2276 info
->tcpi_lost
= tp
->lost_out
;
2277 info
->tcpi_retrans
= tp
->retrans_out
;
2278 info
->tcpi_fackets
= tp
->fackets_out
;
2280 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
2281 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
2282 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
2284 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
2285 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
2286 info
->tcpi_rtt
= jiffies_to_usecs(tp
->srtt
)>>3;
2287 info
->tcpi_rttvar
= jiffies_to_usecs(tp
->mdev
)>>2;
2288 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
2289 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
2290 info
->tcpi_advmss
= tp
->advmss
;
2291 info
->tcpi_reordering
= tp
->reordering
;
2293 info
->tcpi_rcv_rtt
= jiffies_to_usecs(tp
->rcv_rtt_est
.rtt
)>>3;
2294 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
2296 info
->tcpi_total_retrans
= tp
->total_retrans
;
2299 EXPORT_SYMBOL_GPL(tcp_get_info
);
2301 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
2302 int optname
, char __user
*optval
, int __user
*optlen
)
2304 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2305 struct tcp_sock
*tp
= tcp_sk(sk
);
2308 if (get_user(len
, optlen
))
2311 len
= min_t(unsigned int, len
, sizeof(int));
2318 val
= tp
->mss_cache
;
2319 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
2320 val
= tp
->rx_opt
.user_mss
;
2323 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
2326 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
2329 val
= (tp
->keepalive_time
? : sysctl_tcp_keepalive_time
) / HZ
;
2332 val
= (tp
->keepalive_intvl
? : sysctl_tcp_keepalive_intvl
) / HZ
;
2335 val
= tp
->keepalive_probes
? : sysctl_tcp_keepalive_probes
;
2338 val
= icsk
->icsk_syn_retries
? : sysctl_tcp_syn_retries
;
2343 val
= (val
? : sysctl_tcp_fin_timeout
) / HZ
;
2345 case TCP_DEFER_ACCEPT
:
2346 val
= retrans_to_secs(icsk
->icsk_accept_queue
.rskq_defer_accept
,
2347 TCP_TIMEOUT_INIT
/ HZ
, TCP_RTO_MAX
/ HZ
);
2349 case TCP_WINDOW_CLAMP
:
2350 val
= tp
->window_clamp
;
2353 struct tcp_info info
;
2355 if (get_user(len
, optlen
))
2358 tcp_get_info(sk
, &info
);
2360 len
= min_t(unsigned int, len
, sizeof(info
));
2361 if (put_user(len
, optlen
))
2363 if (copy_to_user(optval
, &info
, len
))
2368 val
= !icsk
->icsk_ack
.pingpong
;
2371 case TCP_CONGESTION
:
2372 if (get_user(len
, optlen
))
2374 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
2375 if (put_user(len
, optlen
))
2377 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
2381 return -ENOPROTOOPT
;
2384 if (put_user(len
, optlen
))
2386 if (copy_to_user(optval
, &val
, len
))
2391 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2394 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2396 if (level
!= SOL_TCP
)
2397 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
2399 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2402 #ifdef CONFIG_COMPAT
2403 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
2404 char __user
*optval
, int __user
*optlen
)
2406 if (level
!= SOL_TCP
)
2407 return inet_csk_compat_getsockopt(sk
, level
, optname
,
2409 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2412 EXPORT_SYMBOL(compat_tcp_getsockopt
);
2415 struct sk_buff
*tcp_tso_segment(struct sk_buff
*skb
, int features
)
2417 struct sk_buff
*segs
= ERR_PTR(-EINVAL
);
2422 unsigned int oldlen
;
2425 if (!pskb_may_pull(skb
, sizeof(*th
)))
2429 thlen
= th
->doff
* 4;
2430 if (thlen
< sizeof(*th
))
2433 if (!pskb_may_pull(skb
, thlen
))
2436 oldlen
= (u16
)~skb
->len
;
2437 __skb_pull(skb
, thlen
);
2439 if (skb_gso_ok(skb
, features
| NETIF_F_GSO_ROBUST
)) {
2440 /* Packet is from an untrusted source, reset gso_segs. */
2441 int type
= skb_shinfo(skb
)->gso_type
;
2450 !(type
& (SKB_GSO_TCPV4
| SKB_GSO_TCPV6
))))
2453 mss
= skb_shinfo(skb
)->gso_size
;
2454 skb_shinfo(skb
)->gso_segs
= (skb
->len
+ mss
- 1) / mss
;
2460 segs
= skb_segment(skb
, features
);
2464 len
= skb_shinfo(skb
)->gso_size
;
2465 delta
= htonl(oldlen
+ (thlen
+ len
));
2469 seq
= ntohl(th
->seq
);
2472 th
->fin
= th
->psh
= 0;
2474 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
2475 (__force u32
)delta
));
2476 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2478 csum_fold(csum_partial(skb_transport_header(skb
),
2485 th
->seq
= htonl(seq
);
2487 } while (skb
->next
);
2489 delta
= htonl(oldlen
+ (skb
->tail
- skb
->transport_header
) +
2491 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
2492 (__force u32
)delta
));
2493 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2494 th
->check
= csum_fold(csum_partial(skb_transport_header(skb
),
2500 EXPORT_SYMBOL(tcp_tso_segment
);
2502 #ifdef CONFIG_TCP_MD5SIG
2503 static unsigned long tcp_md5sig_users
;
2504 static struct tcp_md5sig_pool
**tcp_md5sig_pool
;
2505 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock
);
2507 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool
**pool
)
2510 for_each_possible_cpu(cpu
) {
2511 struct tcp_md5sig_pool
*p
= *per_cpu_ptr(pool
, cpu
);
2513 if (p
->md5_desc
.tfm
)
2514 crypto_free_hash(p
->md5_desc
.tfm
);
2522 void tcp_free_md5sig_pool(void)
2524 struct tcp_md5sig_pool
**pool
= NULL
;
2526 spin_lock_bh(&tcp_md5sig_pool_lock
);
2527 if (--tcp_md5sig_users
== 0) {
2528 pool
= tcp_md5sig_pool
;
2529 tcp_md5sig_pool
= NULL
;
2531 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2533 __tcp_free_md5sig_pool(pool
);
2536 EXPORT_SYMBOL(tcp_free_md5sig_pool
);
2538 static struct tcp_md5sig_pool
**__tcp_alloc_md5sig_pool(void)
2541 struct tcp_md5sig_pool
**pool
;
2543 pool
= alloc_percpu(struct tcp_md5sig_pool
*);
2547 for_each_possible_cpu(cpu
) {
2548 struct tcp_md5sig_pool
*p
;
2549 struct crypto_hash
*hash
;
2551 p
= kzalloc(sizeof(*p
), GFP_KERNEL
);
2554 *per_cpu_ptr(pool
, cpu
) = p
;
2556 hash
= crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC
);
2557 if (!hash
|| IS_ERR(hash
))
2560 p
->md5_desc
.tfm
= hash
;
2564 __tcp_free_md5sig_pool(pool
);
2568 struct tcp_md5sig_pool
**tcp_alloc_md5sig_pool(void)
2570 struct tcp_md5sig_pool
**pool
;
2574 spin_lock_bh(&tcp_md5sig_pool_lock
);
2575 pool
= tcp_md5sig_pool
;
2576 if (tcp_md5sig_users
++ == 0) {
2578 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2581 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2585 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2588 /* we cannot hold spinlock here because this may sleep. */
2589 struct tcp_md5sig_pool
**p
= __tcp_alloc_md5sig_pool();
2590 spin_lock_bh(&tcp_md5sig_pool_lock
);
2593 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2596 pool
= tcp_md5sig_pool
;
2598 /* oops, it has already been assigned. */
2599 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2600 __tcp_free_md5sig_pool(p
);
2602 tcp_md5sig_pool
= pool
= p
;
2603 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2609 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
2611 struct tcp_md5sig_pool
*__tcp_get_md5sig_pool(int cpu
)
2613 struct tcp_md5sig_pool
**p
;
2614 spin_lock_bh(&tcp_md5sig_pool_lock
);
2615 p
= tcp_md5sig_pool
;
2618 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2619 return (p
? *per_cpu_ptr(p
, cpu
) : NULL
);
2622 EXPORT_SYMBOL(__tcp_get_md5sig_pool
);
2624 void __tcp_put_md5sig_pool(void)
2626 tcp_free_md5sig_pool();
2629 EXPORT_SYMBOL(__tcp_put_md5sig_pool
);
2632 void tcp_done(struct sock
*sk
)
2634 if(sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
2635 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS
);
2637 tcp_set_state(sk
, TCP_CLOSE
);
2638 tcp_clear_xmit_timers(sk
);
2640 sk
->sk_shutdown
= SHUTDOWN_MASK
;
2642 if (!sock_flag(sk
, SOCK_DEAD
))
2643 sk
->sk_state_change(sk
);
2645 inet_csk_destroy_sock(sk
);
2647 EXPORT_SYMBOL_GPL(tcp_done
);
2649 extern void __skb_cb_too_small_for_tcp(int, int);
2650 extern struct tcp_congestion_ops tcp_reno
;
2652 static __initdata
unsigned long thash_entries
;
2653 static int __init
set_thash_entries(char *str
)
2657 thash_entries
= simple_strtoul(str
, &str
, 0);
2660 __setup("thash_entries=", set_thash_entries
);
2662 void __init
tcp_init(void)
2664 struct sk_buff
*skb
= NULL
;
2665 unsigned long limit
;
2666 int order
, i
, max_share
;
2668 if (sizeof(struct tcp_skb_cb
) > sizeof(skb
->cb
))
2669 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb
),
2672 tcp_hashinfo
.bind_bucket_cachep
=
2673 kmem_cache_create("tcp_bind_bucket",
2674 sizeof(struct inet_bind_bucket
), 0,
2675 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
, NULL
);
2677 /* Size and allocate the main established and bind bucket
2680 * The methodology is similar to that of the buffer cache.
2682 tcp_hashinfo
.ehash
=
2683 alloc_large_system_hash("TCP established",
2684 sizeof(struct inet_ehash_bucket
),
2686 (num_physpages
>= 128 * 1024) ?
2689 &tcp_hashinfo
.ehash_size
,
2692 tcp_hashinfo
.ehash_size
= 1 << tcp_hashinfo
.ehash_size
;
2693 for (i
= 0; i
< tcp_hashinfo
.ehash_size
; i
++) {
2694 rwlock_init(&tcp_hashinfo
.ehash
[i
].lock
);
2695 INIT_HLIST_HEAD(&tcp_hashinfo
.ehash
[i
].chain
);
2696 INIT_HLIST_HEAD(&tcp_hashinfo
.ehash
[i
].twchain
);
2699 tcp_hashinfo
.bhash
=
2700 alloc_large_system_hash("TCP bind",
2701 sizeof(struct inet_bind_hashbucket
),
2702 tcp_hashinfo
.ehash_size
,
2703 (num_physpages
>= 128 * 1024) ?
2706 &tcp_hashinfo
.bhash_size
,
2709 tcp_hashinfo
.bhash_size
= 1 << tcp_hashinfo
.bhash_size
;
2710 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
2711 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
2712 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
2715 /* Try to be a bit smarter and adjust defaults depending
2716 * on available memory.
2718 for (order
= 0; ((1 << order
) << PAGE_SHIFT
) <
2719 (tcp_hashinfo
.bhash_size
* sizeof(struct inet_bind_hashbucket
));
2723 tcp_death_row
.sysctl_max_tw_buckets
= 180000;
2724 sysctl_tcp_max_orphans
= 4096 << (order
- 4);
2725 sysctl_max_syn_backlog
= 1024;
2726 } else if (order
< 3) {
2727 tcp_death_row
.sysctl_max_tw_buckets
>>= (3 - order
);
2728 sysctl_tcp_max_orphans
>>= (3 - order
);
2729 sysctl_max_syn_backlog
= 128;
2732 /* Set the pressure threshold to be a fraction of global memory that
2733 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2734 * memory, with a floor of 128 pages.
2736 limit
= min(nr_all_pages
, 1UL<<(28-PAGE_SHIFT
)) >> (20-PAGE_SHIFT
);
2737 limit
= (limit
* (nr_all_pages
>> (20-PAGE_SHIFT
))) >> (PAGE_SHIFT
-11);
2738 limit
= max(limit
, 128UL);
2739 sysctl_tcp_mem
[0] = limit
/ 4 * 3;
2740 sysctl_tcp_mem
[1] = limit
;
2741 sysctl_tcp_mem
[2] = sysctl_tcp_mem
[0] * 2;
2743 /* Set per-socket limits to no more than 1/128 the pressure threshold */
2744 limit
= ((unsigned long)sysctl_tcp_mem
[1]) << (PAGE_SHIFT
- 7);
2745 max_share
= min(4UL*1024*1024, limit
);
2747 sysctl_tcp_wmem
[0] = SK_STREAM_MEM_QUANTUM
;
2748 sysctl_tcp_wmem
[1] = 16*1024;
2749 sysctl_tcp_wmem
[2] = max(64*1024, max_share
);
2751 sysctl_tcp_rmem
[0] = SK_STREAM_MEM_QUANTUM
;
2752 sysctl_tcp_rmem
[1] = 87380;
2753 sysctl_tcp_rmem
[2] = max(87380, max_share
);
2755 printk(KERN_INFO
"TCP: Hash tables configured "
2756 "(established %d bind %d)\n",
2757 tcp_hashinfo
.ehash_size
, tcp_hashinfo
.bhash_size
);
2759 tcp_register_congestion_control(&tcp_reno
);
2762 EXPORT_SYMBOL(tcp_close
);
2763 EXPORT_SYMBOL(tcp_disconnect
);
2764 EXPORT_SYMBOL(tcp_getsockopt
);
2765 EXPORT_SYMBOL(tcp_ioctl
);
2766 EXPORT_SYMBOL(tcp_poll
);
2767 EXPORT_SYMBOL(tcp_read_sock
);
2768 EXPORT_SYMBOL(tcp_recvmsg
);
2769 EXPORT_SYMBOL(tcp_sendmsg
);
2770 EXPORT_SYMBOL(tcp_splice_read
);
2771 EXPORT_SYMBOL(tcp_sendpage
);
2772 EXPORT_SYMBOL(tcp_setsockopt
);
2773 EXPORT_SYMBOL(tcp_shutdown
);
2774 EXPORT_SYMBOL(tcp_statistics
);