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/config.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/smp_lock.h>
257 #include <linux/fs.h>
258 #include <linux/random.h>
259 #include <linux/bootmem.h>
261 #include <net/icmp.h>
263 #include <net/xfrm.h>
267 #include <asm/uaccess.h>
268 #include <asm/ioctls.h>
270 int sysctl_tcp_fin_timeout
= TCP_FIN_TIMEOUT
;
272 DEFINE_SNMP_STAT(struct tcp_mib
, tcp_statistics
) __read_mostly
;
274 atomic_t tcp_orphan_count
= ATOMIC_INIT(0);
276 EXPORT_SYMBOL_GPL(tcp_orphan_count
);
278 int sysctl_tcp_mem
[3];
279 int sysctl_tcp_wmem
[3] = { 4 * 1024, 16 * 1024, 128 * 1024 };
280 int sysctl_tcp_rmem
[3] = { 4 * 1024, 87380, 87380 * 2 };
282 EXPORT_SYMBOL(sysctl_tcp_mem
);
283 EXPORT_SYMBOL(sysctl_tcp_rmem
);
284 EXPORT_SYMBOL(sysctl_tcp_wmem
);
286 atomic_t tcp_memory_allocated
; /* Current allocated memory. */
287 atomic_t tcp_sockets_allocated
; /* Current number of TCP sockets. */
289 EXPORT_SYMBOL(tcp_memory_allocated
);
290 EXPORT_SYMBOL(tcp_sockets_allocated
);
293 * Pressure flag: try to collapse.
294 * Technical note: it is used by multiple contexts non atomically.
295 * All the sk_stream_mem_schedule() is of this nature: accounting
296 * is strict, actions are advisory and have some latency.
298 int tcp_memory_pressure
;
300 EXPORT_SYMBOL(tcp_memory_pressure
);
302 void tcp_enter_memory_pressure(void)
304 if (!tcp_memory_pressure
) {
305 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES
);
306 tcp_memory_pressure
= 1;
310 EXPORT_SYMBOL(tcp_enter_memory_pressure
);
313 * Wait for a TCP event.
315 * Note that we don't need to lock the socket, as the upper poll layers
316 * take care of normal races (between the test and the event) and we don't
317 * go look at any of the socket buffers directly.
319 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
322 struct sock
*sk
= sock
->sk
;
323 struct tcp_sock
*tp
= tcp_sk(sk
);
325 poll_wait(file
, sk
->sk_sleep
, wait
);
326 if (sk
->sk_state
== TCP_LISTEN
)
327 return inet_csk_listen_poll(sk
);
329 /* Socket is not locked. We are protected from async events
330 by poll logic and correct handling of state changes
331 made by another threads is impossible in any case.
339 * POLLHUP is certainly not done right. But poll() doesn't
340 * have a notion of HUP in just one direction, and for a
341 * socket the read side is more interesting.
343 * Some poll() documentation says that POLLHUP is incompatible
344 * with the POLLOUT/POLLWR flags, so somebody should check this
345 * all. But careful, it tends to be safer to return too many
346 * bits than too few, and you can easily break real applications
347 * if you don't tell them that something has hung up!
351 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
352 * our fs/select.c). It means that after we received EOF,
353 * poll always returns immediately, making impossible poll() on write()
354 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
355 * if and only if shutdown has been made in both directions.
356 * Actually, it is interesting to look how Solaris and DUX
357 * solve this dilemma. I would prefer, if PULLHUP were maskable,
358 * then we could set it on SND_SHUTDOWN. BTW examples given
359 * in Stevens' books assume exactly this behaviour, it explains
360 * why PULLHUP is incompatible with POLLOUT. --ANK
362 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
363 * blocking on fresh not-connected or disconnected socket. --ANK
365 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| sk
->sk_state
== TCP_CLOSE
)
367 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
368 mask
|= POLLIN
| POLLRDNORM
;
371 if ((1 << sk
->sk_state
) & ~(TCPF_SYN_SENT
| TCPF_SYN_RECV
)) {
372 /* Potential race condition. If read of tp below will
373 * escape above sk->sk_state, we can be illegally awaken
374 * in SYN_* states. */
375 if ((tp
->rcv_nxt
!= tp
->copied_seq
) &&
376 (tp
->urg_seq
!= tp
->copied_seq
||
377 tp
->rcv_nxt
!= tp
->copied_seq
+ 1 ||
378 sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
))
379 mask
|= POLLIN
| POLLRDNORM
;
381 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
382 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
)) {
383 mask
|= POLLOUT
| POLLWRNORM
;
384 } else { /* send SIGIO later */
385 set_bit(SOCK_ASYNC_NOSPACE
,
386 &sk
->sk_socket
->flags
);
387 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
389 /* Race breaker. If space is freed after
390 * wspace test but before the flags are set,
391 * IO signal will be lost.
393 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
))
394 mask
|= POLLOUT
| POLLWRNORM
;
398 if (tp
->urg_data
& TCP_URG_VALID
)
404 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
406 struct tcp_sock
*tp
= tcp_sk(sk
);
411 if (sk
->sk_state
== TCP_LISTEN
)
415 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
417 else if (sock_flag(sk
, SOCK_URGINLINE
) ||
419 before(tp
->urg_seq
, tp
->copied_seq
) ||
420 !before(tp
->urg_seq
, tp
->rcv_nxt
)) {
421 answ
= tp
->rcv_nxt
- tp
->copied_seq
;
423 /* Subtract 1, if FIN is in queue. */
424 if (answ
&& !skb_queue_empty(&sk
->sk_receive_queue
))
426 ((struct sk_buff
*)sk
->sk_receive_queue
.prev
)->h
.th
->fin
;
428 answ
= tp
->urg_seq
- tp
->copied_seq
;
432 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
435 if (sk
->sk_state
== TCP_LISTEN
)
438 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
441 answ
= tp
->write_seq
- tp
->snd_una
;
447 return put_user(answ
, (int __user
*)arg
);
450 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
452 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
453 tp
->pushed_seq
= tp
->write_seq
;
456 static inline int forced_push(struct tcp_sock
*tp
)
458 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
461 static inline void skb_entail(struct sock
*sk
, struct tcp_sock
*tp
,
465 TCP_SKB_CB(skb
)->seq
= tp
->write_seq
;
466 TCP_SKB_CB(skb
)->end_seq
= tp
->write_seq
;
467 TCP_SKB_CB(skb
)->flags
= TCPCB_FLAG_ACK
;
468 TCP_SKB_CB(skb
)->sacked
= 0;
469 skb_header_release(skb
);
470 __skb_queue_tail(&sk
->sk_write_queue
, skb
);
471 sk_charge_skb(sk
, skb
);
472 if (!sk
->sk_send_head
)
473 sk
->sk_send_head
= skb
;
474 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
475 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
478 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
,
481 if (flags
& MSG_OOB
) {
483 tp
->snd_up
= tp
->write_seq
;
484 TCP_SKB_CB(skb
)->sacked
|= TCPCB_URG
;
488 static inline void tcp_push(struct sock
*sk
, struct tcp_sock
*tp
, int flags
,
489 int mss_now
, int nonagle
)
491 if (sk
->sk_send_head
) {
492 struct sk_buff
*skb
= sk
->sk_write_queue
.prev
;
493 if (!(flags
& MSG_MORE
) || forced_push(tp
))
494 tcp_mark_push(tp
, skb
);
495 tcp_mark_urg(tp
, flags
, skb
);
496 __tcp_push_pending_frames(sk
, tp
, mss_now
,
497 (flags
& MSG_MORE
) ? TCP_NAGLE_CORK
: nonagle
);
501 static ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
**pages
, int poffset
,
502 size_t psize
, int flags
)
504 struct tcp_sock
*tp
= tcp_sk(sk
);
505 int mss_now
, size_goal
;
508 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
510 /* Wait for a connection to finish. */
511 if ((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
512 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
515 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
517 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
518 size_goal
= tp
->xmit_size_goal
;
522 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
526 struct sk_buff
*skb
= sk
->sk_write_queue
.prev
;
527 struct page
*page
= pages
[poffset
/ PAGE_SIZE
];
528 int copy
, i
, can_coalesce
;
529 int offset
= poffset
% PAGE_SIZE
;
530 int size
= min_t(size_t, psize
, PAGE_SIZE
- offset
);
532 if (!sk
->sk_send_head
|| (copy
= size_goal
- skb
->len
) <= 0) {
534 if (!sk_stream_memory_free(sk
))
535 goto wait_for_sndbuf
;
537 skb
= sk_stream_alloc_pskb(sk
, 0, 0,
540 goto wait_for_memory
;
542 skb_entail(sk
, tp
, skb
);
549 i
= skb_shinfo(skb
)->nr_frags
;
550 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
551 if (!can_coalesce
&& i
>= MAX_SKB_FRAGS
) {
552 tcp_mark_push(tp
, skb
);
555 if (!sk_stream_wmem_schedule(sk
, copy
))
556 goto wait_for_memory
;
559 skb_shinfo(skb
)->frags
[i
- 1].size
+= copy
;
562 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
566 skb
->data_len
+= copy
;
567 skb
->truesize
+= copy
;
568 sk
->sk_wmem_queued
+= copy
;
569 sk
->sk_forward_alloc
-= copy
;
570 skb
->ip_summed
= CHECKSUM_HW
;
571 tp
->write_seq
+= copy
;
572 TCP_SKB_CB(skb
)->end_seq
+= copy
;
573 skb_shinfo(skb
)->tso_segs
= 0;
576 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_PSH
;
580 if (!(psize
-= copy
))
583 if (skb
->len
< mss_now
|| (flags
& MSG_OOB
))
586 if (forced_push(tp
)) {
587 tcp_mark_push(tp
, skb
);
588 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_PUSH
);
589 } else if (skb
== sk
->sk_send_head
)
590 tcp_push_one(sk
, mss_now
);
594 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
597 tcp_push(sk
, tp
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
599 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
602 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
603 size_goal
= tp
->xmit_size_goal
;
608 tcp_push(sk
, tp
, flags
, mss_now
, tp
->nonagle
);
615 return sk_stream_error(sk
, flags
, err
);
618 ssize_t
tcp_sendpage(struct socket
*sock
, struct page
*page
, int offset
,
619 size_t size
, int flags
)
622 struct sock
*sk
= sock
->sk
;
624 #define TCP_ZC_CSUM_FLAGS (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM)
626 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
627 !(sk
->sk_route_caps
& TCP_ZC_CSUM_FLAGS
))
628 return sock_no_sendpage(sock
, page
, offset
, size
, flags
);
630 #undef TCP_ZC_CSUM_FLAGS
634 res
= do_tcp_sendpages(sk
, &page
, offset
, size
, flags
);
640 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
641 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
643 static inline int select_size(struct sock
*sk
, struct tcp_sock
*tp
)
645 int tmp
= tp
->mss_cache
;
647 if (sk
->sk_route_caps
& NETIF_F_SG
) {
648 if (sk
->sk_route_caps
& NETIF_F_TSO
)
651 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
653 if (tmp
>= pgbreak
&&
654 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
662 int tcp_sendmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
666 struct tcp_sock
*tp
= tcp_sk(sk
);
669 int mss_now
, size_goal
;
676 flags
= msg
->msg_flags
;
677 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
679 /* Wait for a connection to finish. */
680 if ((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
681 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
684 /* This should be in poll */
685 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
687 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
688 size_goal
= tp
->xmit_size_goal
;
690 /* Ok commence sending. */
691 iovlen
= msg
->msg_iovlen
;
696 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
699 while (--iovlen
>= 0) {
700 int seglen
= iov
->iov_len
;
701 unsigned char __user
*from
= iov
->iov_base
;
708 skb
= sk
->sk_write_queue
.prev
;
710 if (!sk
->sk_send_head
||
711 (copy
= size_goal
- skb
->len
) <= 0) {
714 /* Allocate new segment. If the interface is SG,
715 * allocate skb fitting to single page.
717 if (!sk_stream_memory_free(sk
))
718 goto wait_for_sndbuf
;
720 skb
= sk_stream_alloc_pskb(sk
, select_size(sk
, tp
),
721 0, sk
->sk_allocation
);
723 goto wait_for_memory
;
726 * Check whether we can use HW checksum.
728 if (sk
->sk_route_caps
&
729 (NETIF_F_IP_CSUM
| NETIF_F_NO_CSUM
|
731 skb
->ip_summed
= CHECKSUM_HW
;
733 skb_entail(sk
, tp
, skb
);
737 /* Try to append data to the end of skb. */
741 /* Where to copy to? */
742 if (skb_tailroom(skb
) > 0) {
743 /* We have some space in skb head. Superb! */
744 if (copy
> skb_tailroom(skb
))
745 copy
= skb_tailroom(skb
);
746 if ((err
= skb_add_data(skb
, from
, copy
)) != 0)
750 int i
= skb_shinfo(skb
)->nr_frags
;
751 struct page
*page
= TCP_PAGE(sk
);
752 int off
= TCP_OFF(sk
);
754 if (skb_can_coalesce(skb
, i
, page
, off
) &&
756 /* We can extend the last page
759 } else if (i
== MAX_SKB_FRAGS
||
761 !(sk
->sk_route_caps
& NETIF_F_SG
))) {
762 /* Need to add new fragment and cannot
763 * do this because interface is non-SG,
764 * or because all the page slots are
766 tcp_mark_push(tp
, skb
);
769 if (off
== PAGE_SIZE
) {
771 TCP_PAGE(sk
) = page
= NULL
;
777 if (copy
> PAGE_SIZE
- off
)
778 copy
= PAGE_SIZE
- off
;
780 if (!sk_stream_wmem_schedule(sk
, copy
))
781 goto wait_for_memory
;
784 /* Allocate new cache page. */
785 if (!(page
= sk_stream_alloc_page(sk
)))
786 goto wait_for_memory
;
789 /* Time to copy data. We are close to
791 err
= skb_copy_to_page(sk
, from
, skb
, page
,
794 /* If this page was new, give it to the
795 * socket so it does not get leaked.
804 /* Update the skb. */
806 skb_shinfo(skb
)->frags
[i
- 1].size
+=
809 skb_fill_page_desc(skb
, i
, page
, off
, copy
);
812 } else if (off
+ copy
< PAGE_SIZE
) {
818 TCP_OFF(sk
) = off
+ copy
;
822 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_PSH
;
824 tp
->write_seq
+= copy
;
825 TCP_SKB_CB(skb
)->end_seq
+= copy
;
826 skb_shinfo(skb
)->tso_segs
= 0;
830 if ((seglen
-= copy
) == 0 && iovlen
== 0)
833 if (skb
->len
< mss_now
|| (flags
& MSG_OOB
))
836 if (forced_push(tp
)) {
837 tcp_mark_push(tp
, skb
);
838 __tcp_push_pending_frames(sk
, tp
, mss_now
, TCP_NAGLE_PUSH
);
839 } else if (skb
== sk
->sk_send_head
)
840 tcp_push_one(sk
, mss_now
);
844 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
847 tcp_push(sk
, tp
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
849 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
852 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
853 size_goal
= tp
->xmit_size_goal
;
859 tcp_push(sk
, tp
, flags
, mss_now
, tp
->nonagle
);
866 if (sk
->sk_send_head
== skb
)
867 sk
->sk_send_head
= NULL
;
868 __skb_unlink(skb
, &sk
->sk_write_queue
);
869 sk_stream_free_skb(sk
, skb
);
876 err
= sk_stream_error(sk
, flags
, err
);
883 * Handle reading urgent data. BSD has very simple semantics for
884 * this, no blocking and very strange errors 8)
887 static int tcp_recv_urg(struct sock
*sk
, long timeo
,
888 struct msghdr
*msg
, int len
, int flags
,
891 struct tcp_sock
*tp
= tcp_sk(sk
);
893 /* No URG data to read. */
894 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
895 tp
->urg_data
== TCP_URG_READ
)
896 return -EINVAL
; /* Yes this is right ! */
898 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
901 if (tp
->urg_data
& TCP_URG_VALID
) {
903 char c
= tp
->urg_data
;
905 if (!(flags
& MSG_PEEK
))
906 tp
->urg_data
= TCP_URG_READ
;
908 /* Read urgent data. */
909 msg
->msg_flags
|= MSG_OOB
;
912 if (!(flags
& MSG_TRUNC
))
913 err
= memcpy_toiovec(msg
->msg_iov
, &c
, 1);
916 msg
->msg_flags
|= MSG_TRUNC
;
918 return err
? -EFAULT
: len
;
921 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
924 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
925 * the available implementations agree in this case:
926 * this call should never block, independent of the
927 * blocking state of the socket.
928 * Mike <pall@rz.uni-karlsruhe.de>
933 /* Clean up the receive buffer for full frames taken by the user,
934 * then send an ACK if necessary. COPIED is the number of bytes
935 * tcp_recvmsg has given to the user so far, it speeds up the
936 * calculation of whether or not we must ACK for the sake of
939 static void cleanup_rbuf(struct sock
*sk
, int copied
)
941 struct tcp_sock
*tp
= tcp_sk(sk
);
945 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
947 BUG_TRAP(!skb
|| before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
));
950 if (inet_csk_ack_scheduled(sk
)) {
951 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
952 /* Delayed ACKs frequently hit locked sockets during bulk
954 if (icsk
->icsk_ack
.blocked
||
955 /* Once-per-two-segments ACK was not sent by tcp_input.c */
956 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
958 * If this read emptied read buffer, we send ACK, if
959 * connection is not bidirectional, user drained
960 * receive buffer and there was a small segment
963 (copied
> 0 && (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
964 !icsk
->icsk_ack
.pingpong
&& !atomic_read(&sk
->sk_rmem_alloc
)))
968 /* We send an ACK if we can now advertise a non-zero window
969 * which has been raised "significantly".
971 * Even if window raised up to infinity, do not send window open ACK
972 * in states, where we will not receive more. It is useless.
974 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
975 __u32 rcv_window_now
= tcp_receive_window(tp
);
977 /* Optimize, __tcp_select_window() is not cheap. */
978 if (2*rcv_window_now
<= tp
->window_clamp
) {
979 __u32 new_window
= __tcp_select_window(sk
);
981 /* Send ACK now, if this read freed lots of space
982 * in our buffer. Certainly, new_window is new window.
983 * We can advertise it now, if it is not less than current one.
984 * "Lots" means "at least twice" here.
986 if (new_window
&& new_window
>= 2 * rcv_window_now
)
994 static void tcp_prequeue_process(struct sock
*sk
)
997 struct tcp_sock
*tp
= tcp_sk(sk
);
999 NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED
);
1001 /* RX process wants to run with disabled BHs, though it is not
1004 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
1005 sk
->sk_backlog_rcv(sk
, skb
);
1008 /* Clear memory counter. */
1009 tp
->ucopy
.memory
= 0;
1012 static inline struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1014 struct sk_buff
*skb
;
1017 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1018 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1021 if (offset
< skb
->len
|| skb
->h
.th
->fin
) {
1030 * This routine provides an alternative to tcp_recvmsg() for routines
1031 * that would like to handle copying from skbuffs directly in 'sendfile'
1034 * - It is assumed that the socket was locked by the caller.
1035 * - The routine does not block.
1036 * - At present, there is no support for reading OOB data
1037 * or for 'peeking' the socket using this routine
1038 * (although both would be easy to implement).
1040 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1041 sk_read_actor_t recv_actor
)
1043 struct sk_buff
*skb
;
1044 struct tcp_sock
*tp
= tcp_sk(sk
);
1045 u32 seq
= tp
->copied_seq
;
1049 if (sk
->sk_state
== TCP_LISTEN
)
1051 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1052 if (offset
< skb
->len
) {
1055 len
= skb
->len
- offset
;
1056 /* Stop reading if we hit a patch of urgent data */
1058 u32 urg_offset
= tp
->urg_seq
- seq
;
1059 if (urg_offset
< len
)
1064 used
= recv_actor(desc
, skb
, offset
, len
);
1070 if (offset
!= skb
->len
)
1073 if (skb
->h
.th
->fin
) {
1074 sk_eat_skb(sk
, skb
);
1078 sk_eat_skb(sk
, skb
);
1082 tp
->copied_seq
= seq
;
1084 tcp_rcv_space_adjust(sk
);
1086 /* Clean up data we have read: This will do ACK frames. */
1088 cleanup_rbuf(sk
, copied
);
1093 * This routine copies from a sock struct into the user buffer.
1095 * Technical note: in 2.3 we work on _locked_ socket, so that
1096 * tricks with *seq access order and skb->users are not required.
1097 * Probably, code can be easily improved even more.
1100 int tcp_recvmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1101 size_t len
, int nonblock
, int flags
, int *addr_len
)
1103 struct tcp_sock
*tp
= tcp_sk(sk
);
1109 int target
; /* Read at least this many bytes */
1111 struct task_struct
*user_recv
= NULL
;
1115 TCP_CHECK_TIMER(sk
);
1118 if (sk
->sk_state
== TCP_LISTEN
)
1121 timeo
= sock_rcvtimeo(sk
, nonblock
);
1123 /* Urgent data needs to be handled specially. */
1124 if (flags
& MSG_OOB
)
1127 seq
= &tp
->copied_seq
;
1128 if (flags
& MSG_PEEK
) {
1129 peek_seq
= tp
->copied_seq
;
1133 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1136 struct sk_buff
*skb
;
1139 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1140 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1143 if (signal_pending(current
)) {
1144 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1149 /* Next get a buffer. */
1151 skb
= skb_peek(&sk
->sk_receive_queue
);
1156 /* Now that we have two receive queues this
1159 if (before(*seq
, TCP_SKB_CB(skb
)->seq
)) {
1160 printk(KERN_INFO
"recvmsg bug: copied %X "
1161 "seq %X\n", *seq
, TCP_SKB_CB(skb
)->seq
);
1164 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1167 if (offset
< skb
->len
)
1171 BUG_TRAP(flags
& MSG_PEEK
);
1173 } while (skb
!= (struct sk_buff
*)&sk
->sk_receive_queue
);
1175 /* Well, if we have backlog, try to process it now yet. */
1177 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1182 sk
->sk_state
== TCP_CLOSE
||
1183 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1185 signal_pending(current
) ||
1189 if (sock_flag(sk
, SOCK_DONE
))
1193 copied
= sock_error(sk
);
1197 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1200 if (sk
->sk_state
== TCP_CLOSE
) {
1201 if (!sock_flag(sk
, SOCK_DONE
)) {
1202 /* This occurs when user tries to read
1203 * from never connected socket.
1216 if (signal_pending(current
)) {
1217 copied
= sock_intr_errno(timeo
);
1222 cleanup_rbuf(sk
, copied
);
1224 if (!sysctl_tcp_low_latency
&& tp
->ucopy
.task
== user_recv
) {
1225 /* Install new reader */
1226 if (!user_recv
&& !(flags
& (MSG_TRUNC
| MSG_PEEK
))) {
1227 user_recv
= current
;
1228 tp
->ucopy
.task
= user_recv
;
1229 tp
->ucopy
.iov
= msg
->msg_iov
;
1232 tp
->ucopy
.len
= len
;
1234 BUG_TRAP(tp
->copied_seq
== tp
->rcv_nxt
||
1235 (flags
& (MSG_PEEK
| MSG_TRUNC
)));
1237 /* Ugly... If prequeue is not empty, we have to
1238 * process it before releasing socket, otherwise
1239 * order will be broken at second iteration.
1240 * More elegant solution is required!!!
1242 * Look: we have the following (pseudo)queues:
1244 * 1. packets in flight
1249 * Each queue can be processed only if the next ones
1250 * are empty. At this point we have empty receive_queue.
1251 * But prequeue _can_ be not empty after 2nd iteration,
1252 * when we jumped to start of loop because backlog
1253 * processing added something to receive_queue.
1254 * We cannot release_sock(), because backlog contains
1255 * packets arrived _after_ prequeued ones.
1257 * Shortly, algorithm is clear --- to process all
1258 * the queues in order. We could make it more directly,
1259 * requeueing packets from backlog to prequeue, if
1260 * is not empty. It is more elegant, but eats cycles,
1263 if (!skb_queue_empty(&tp
->ucopy
.prequeue
))
1266 /* __ Set realtime policy in scheduler __ */
1269 if (copied
>= target
) {
1270 /* Do not sleep, just process backlog. */
1274 sk_wait_data(sk
, &timeo
);
1279 /* __ Restore normal policy in scheduler __ */
1281 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1282 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG
, chunk
);
1287 if (tp
->rcv_nxt
== tp
->copied_seq
&&
1288 !skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1290 tcp_prequeue_process(sk
);
1292 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1293 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1299 if ((flags
& MSG_PEEK
) && peek_seq
!= tp
->copied_seq
) {
1300 if (net_ratelimit())
1301 printk(KERN_DEBUG
"TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1302 current
->comm
, current
->pid
);
1303 peek_seq
= tp
->copied_seq
;
1308 /* Ok so how much can we use? */
1309 used
= skb
->len
- offset
;
1313 /* Do we have urgent data here? */
1315 u32 urg_offset
= tp
->urg_seq
- *seq
;
1316 if (urg_offset
< used
) {
1318 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1330 if (!(flags
& MSG_TRUNC
)) {
1331 err
= skb_copy_datagram_iovec(skb
, offset
,
1332 msg
->msg_iov
, used
);
1334 /* Exception. Bailout! */
1345 tcp_rcv_space_adjust(sk
);
1348 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1350 tcp_fast_path_check(sk
, tp
);
1352 if (used
+ offset
< skb
->len
)
1357 if (!(flags
& MSG_PEEK
))
1358 sk_eat_skb(sk
, skb
);
1362 /* Process the FIN. */
1364 if (!(flags
& MSG_PEEK
))
1365 sk_eat_skb(sk
, skb
);
1370 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1373 tp
->ucopy
.len
= copied
> 0 ? len
: 0;
1375 tcp_prequeue_process(sk
);
1377 if (copied
> 0 && (chunk
= len
- tp
->ucopy
.len
) != 0) {
1378 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1384 tp
->ucopy
.task
= NULL
;
1388 /* According to UNIX98, msg_name/msg_namelen are ignored
1389 * on connected socket. I was just happy when found this 8) --ANK
1392 /* Clean up data we have read: This will do ACK frames. */
1393 cleanup_rbuf(sk
, copied
);
1395 TCP_CHECK_TIMER(sk
);
1400 TCP_CHECK_TIMER(sk
);
1405 err
= tcp_recv_urg(sk
, timeo
, msg
, len
, flags
, addr_len
);
1410 * State processing on a close. This implements the state shift for
1411 * sending our FIN frame. Note that we only send a FIN for some
1412 * states. A shutdown() may have already sent the FIN, or we may be
1416 static const unsigned char new_state
[16] = {
1417 /* current state: new state: action: */
1418 /* (Invalid) */ TCP_CLOSE
,
1419 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1420 /* TCP_SYN_SENT */ TCP_CLOSE
,
1421 /* TCP_SYN_RECV */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1422 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1
,
1423 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2
,
1424 /* TCP_TIME_WAIT */ TCP_CLOSE
,
1425 /* TCP_CLOSE */ TCP_CLOSE
,
1426 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK
| TCP_ACTION_FIN
,
1427 /* TCP_LAST_ACK */ TCP_LAST_ACK
,
1428 /* TCP_LISTEN */ TCP_CLOSE
,
1429 /* TCP_CLOSING */ TCP_CLOSING
,
1432 static int tcp_close_state(struct sock
*sk
)
1434 int next
= (int)new_state
[sk
->sk_state
];
1435 int ns
= next
& TCP_STATE_MASK
;
1437 tcp_set_state(sk
, ns
);
1439 return next
& TCP_ACTION_FIN
;
1443 * Shutdown the sending side of a connection. Much like close except
1444 * that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1447 void tcp_shutdown(struct sock
*sk
, int how
)
1449 /* We need to grab some memory, and put together a FIN,
1450 * and then put it into the queue to be sent.
1451 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1453 if (!(how
& SEND_SHUTDOWN
))
1456 /* If we've already sent a FIN, or it's a closed state, skip this. */
1457 if ((1 << sk
->sk_state
) &
1458 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
1459 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
1460 /* Clear out any half completed packets. FIN if needed. */
1461 if (tcp_close_state(sk
))
1466 void tcp_close(struct sock
*sk
, long timeout
)
1468 struct sk_buff
*skb
;
1469 int data_was_unread
= 0;
1472 sk
->sk_shutdown
= SHUTDOWN_MASK
;
1474 if (sk
->sk_state
== TCP_LISTEN
) {
1475 tcp_set_state(sk
, TCP_CLOSE
);
1478 inet_csk_listen_stop(sk
);
1480 goto adjudge_to_death
;
1483 /* We need to flush the recv. buffs. We do this only on the
1484 * descriptor close, not protocol-sourced closes, because the
1485 * reader process may not have drained the data yet!
1487 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
1488 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
-
1490 data_was_unread
+= len
;
1494 sk_stream_mem_reclaim(sk
);
1496 /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1497 * 3.10, we send a RST here because data was lost. To
1498 * witness the awful effects of the old behavior of always
1499 * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1500 * a bulk GET in an FTP client, suspend the process, wait
1501 * for the client to advertise a zero window, then kill -9
1502 * the FTP client, wheee... Note: timeout is always zero
1505 if (data_was_unread
) {
1506 /* Unread data was tossed, zap the connection. */
1507 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE
);
1508 tcp_set_state(sk
, TCP_CLOSE
);
1509 tcp_send_active_reset(sk
, GFP_KERNEL
);
1510 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
1511 /* Check zero linger _after_ checking for unread data. */
1512 sk
->sk_prot
->disconnect(sk
, 0);
1513 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA
);
1514 } else if (tcp_close_state(sk
)) {
1515 /* We FIN if the application ate all the data before
1516 * zapping the connection.
1519 /* RED-PEN. Formally speaking, we have broken TCP state
1520 * machine. State transitions:
1522 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1523 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1524 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1526 * are legal only when FIN has been sent (i.e. in window),
1527 * rather than queued out of window. Purists blame.
1529 * F.e. "RFC state" is ESTABLISHED,
1530 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1532 * The visible declinations are that sometimes
1533 * we enter time-wait state, when it is not required really
1534 * (harmless), do not send active resets, when they are
1535 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1536 * they look as CLOSING or LAST_ACK for Linux)
1537 * Probably, I missed some more holelets.
1543 sk_stream_wait_close(sk
, timeout
);
1546 /* It is the last release_sock in its life. It will remove backlog. */
1550 /* Now socket is owned by kernel and we acquire BH lock
1551 to finish close. No need to check for user refs.
1555 BUG_TRAP(!sock_owned_by_user(sk
));
1560 /* This is a (useful) BSD violating of the RFC. There is a
1561 * problem with TCP as specified in that the other end could
1562 * keep a socket open forever with no application left this end.
1563 * We use a 3 minute timeout (about the same as BSD) then kill
1564 * our end. If they send after that then tough - BUT: long enough
1565 * that we won't make the old 4*rto = almost no time - whoops
1568 * Nope, it was not mistake. It is really desired behaviour
1569 * f.e. on http servers, when such sockets are useless, but
1570 * consume significant resources. Let's do it with special
1571 * linger2 option. --ANK
1574 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
1575 struct tcp_sock
*tp
= tcp_sk(sk
);
1576 if (tp
->linger2
< 0) {
1577 tcp_set_state(sk
, TCP_CLOSE
);
1578 tcp_send_active_reset(sk
, GFP_ATOMIC
);
1579 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER
);
1581 const int tmo
= tcp_fin_time(sk
);
1583 if (tmo
> TCP_TIMEWAIT_LEN
) {
1584 inet_csk_reset_keepalive_timer(sk
, tcp_fin_time(sk
));
1586 atomic_inc(sk
->sk_prot
->orphan_count
);
1587 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
1592 if (sk
->sk_state
!= TCP_CLOSE
) {
1593 sk_stream_mem_reclaim(sk
);
1594 if (atomic_read(sk
->sk_prot
->orphan_count
) > sysctl_tcp_max_orphans
||
1595 (sk
->sk_wmem_queued
> SOCK_MIN_SNDBUF
&&
1596 atomic_read(&tcp_memory_allocated
) > sysctl_tcp_mem
[2])) {
1597 if (net_ratelimit())
1598 printk(KERN_INFO
"TCP: too many of orphaned "
1600 tcp_set_state(sk
, TCP_CLOSE
);
1601 tcp_send_active_reset(sk
, GFP_ATOMIC
);
1602 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY
);
1605 atomic_inc(sk
->sk_prot
->orphan_count
);
1607 if (sk
->sk_state
== TCP_CLOSE
)
1608 inet_csk_destroy_sock(sk
);
1609 /* Otherwise, socket is reprieved until protocol close. */
1617 /* These states need RST on ABORT according to RFC793 */
1619 static inline int tcp_need_reset(int state
)
1621 return (1 << state
) &
1622 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
1623 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
1626 int tcp_disconnect(struct sock
*sk
, int flags
)
1628 struct inet_sock
*inet
= inet_sk(sk
);
1629 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1630 struct tcp_sock
*tp
= tcp_sk(sk
);
1632 int old_state
= sk
->sk_state
;
1634 if (old_state
!= TCP_CLOSE
)
1635 tcp_set_state(sk
, TCP_CLOSE
);
1637 /* ABORT function of RFC793 */
1638 if (old_state
== TCP_LISTEN
) {
1639 inet_csk_listen_stop(sk
);
1640 } else if (tcp_need_reset(old_state
) ||
1641 (tp
->snd_nxt
!= tp
->write_seq
&&
1642 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
1643 /* The last check adjusts for discrepancy of Linux wrt. RFC
1646 tcp_send_active_reset(sk
, gfp_any());
1647 sk
->sk_err
= ECONNRESET
;
1648 } else if (old_state
== TCP_SYN_SENT
)
1649 sk
->sk_err
= ECONNRESET
;
1651 tcp_clear_xmit_timers(sk
);
1652 __skb_queue_purge(&sk
->sk_receive_queue
);
1653 sk_stream_writequeue_purge(sk
);
1654 __skb_queue_purge(&tp
->out_of_order_queue
);
1658 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
1659 inet_reset_saddr(sk
);
1661 sk
->sk_shutdown
= 0;
1662 sock_reset_flag(sk
, SOCK_DONE
);
1664 if ((tp
->write_seq
+= tp
->max_window
+ 2) == 0)
1666 icsk
->icsk_backoff
= 0;
1668 icsk
->icsk_probes_out
= 0;
1669 tp
->packets_out
= 0;
1670 tp
->snd_ssthresh
= 0x7fffffff;
1671 tp
->snd_cwnd_cnt
= 0;
1672 tp
->bytes_acked
= 0;
1673 tcp_set_ca_state(sk
, TCP_CA_Open
);
1674 tcp_clear_retrans(tp
);
1675 inet_csk_delack_init(sk
);
1676 sk
->sk_send_head
= NULL
;
1677 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
1680 BUG_TRAP(!inet
->num
|| icsk
->icsk_bind_hash
);
1682 sk
->sk_error_report(sk
);
1687 * Socket option code for TCP.
1689 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
1692 struct tcp_sock
*tp
= tcp_sk(sk
);
1693 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1697 if (level
!= SOL_TCP
)
1698 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
1701 /* This is a string value all the others are int's */
1702 if (optname
== TCP_CONGESTION
) {
1703 char name
[TCP_CA_NAME_MAX
];
1708 val
= strncpy_from_user(name
, optval
,
1709 min(TCP_CA_NAME_MAX
-1, optlen
));
1715 err
= tcp_set_congestion_control(sk
, name
);
1720 if (optlen
< sizeof(int))
1723 if (get_user(val
, (int __user
*)optval
))
1730 /* Values greater than interface MTU won't take effect. However
1731 * at the point when this call is done we typically don't yet
1732 * know which interface is going to be used */
1733 if (val
< 8 || val
> MAX_TCP_WINDOW
) {
1737 tp
->rx_opt
.user_mss
= val
;
1742 /* TCP_NODELAY is weaker than TCP_CORK, so that
1743 * this option on corked socket is remembered, but
1744 * it is not activated until cork is cleared.
1746 * However, when TCP_NODELAY is set we make
1747 * an explicit push, which overrides even TCP_CORK
1748 * for currently queued segments.
1750 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
1751 tcp_push_pending_frames(sk
, tp
);
1753 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
1758 /* When set indicates to always queue non-full frames.
1759 * Later the user clears this option and we transmit
1760 * any pending partial frames in the queue. This is
1761 * meant to be used alongside sendfile() to get properly
1762 * filled frames when the user (for example) must write
1763 * out headers with a write() call first and then use
1764 * sendfile to send out the data parts.
1766 * TCP_CORK can be set together with TCP_NODELAY and it is
1767 * stronger than TCP_NODELAY.
1770 tp
->nonagle
|= TCP_NAGLE_CORK
;
1772 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
1773 if (tp
->nonagle
&TCP_NAGLE_OFF
)
1774 tp
->nonagle
|= TCP_NAGLE_PUSH
;
1775 tcp_push_pending_frames(sk
, tp
);
1780 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
1783 tp
->keepalive_time
= val
* HZ
;
1784 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
1785 !((1 << sk
->sk_state
) &
1786 (TCPF_CLOSE
| TCPF_LISTEN
))) {
1787 __u32 elapsed
= tcp_time_stamp
- tp
->rcv_tstamp
;
1788 if (tp
->keepalive_time
> elapsed
)
1789 elapsed
= tp
->keepalive_time
- elapsed
;
1792 inet_csk_reset_keepalive_timer(sk
, elapsed
);
1797 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
1800 tp
->keepalive_intvl
= val
* HZ
;
1803 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
1806 tp
->keepalive_probes
= val
;
1809 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
1812 icsk
->icsk_syn_retries
= val
;
1818 else if (val
> sysctl_tcp_fin_timeout
/ HZ
)
1821 tp
->linger2
= val
* HZ
;
1824 case TCP_DEFER_ACCEPT
:
1825 icsk
->icsk_accept_queue
.rskq_defer_accept
= 0;
1827 /* Translate value in seconds to number of
1829 while (icsk
->icsk_accept_queue
.rskq_defer_accept
< 32 &&
1830 val
> ((TCP_TIMEOUT_INIT
/ HZ
) <<
1831 icsk
->icsk_accept_queue
.rskq_defer_accept
))
1832 icsk
->icsk_accept_queue
.rskq_defer_accept
++;
1833 icsk
->icsk_accept_queue
.rskq_defer_accept
++;
1837 case TCP_WINDOW_CLAMP
:
1839 if (sk
->sk_state
!= TCP_CLOSE
) {
1843 tp
->window_clamp
= 0;
1845 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
1846 SOCK_MIN_RCVBUF
/ 2 : val
;
1851 icsk
->icsk_ack
.pingpong
= 1;
1853 icsk
->icsk_ack
.pingpong
= 0;
1854 if ((1 << sk
->sk_state
) &
1855 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
1856 inet_csk_ack_scheduled(sk
)) {
1857 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
1858 cleanup_rbuf(sk
, 1);
1860 icsk
->icsk_ack
.pingpong
= 1;
1873 /* Return information about state of tcp endpoint in API format. */
1874 void tcp_get_info(struct sock
*sk
, struct tcp_info
*info
)
1876 struct tcp_sock
*tp
= tcp_sk(sk
);
1877 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1878 u32 now
= tcp_time_stamp
;
1880 memset(info
, 0, sizeof(*info
));
1882 info
->tcpi_state
= sk
->sk_state
;
1883 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
1884 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
1885 info
->tcpi_probes
= icsk
->icsk_probes_out
;
1886 info
->tcpi_backoff
= icsk
->icsk_backoff
;
1888 if (tp
->rx_opt
.tstamp_ok
)
1889 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
1890 if (tp
->rx_opt
.sack_ok
)
1891 info
->tcpi_options
|= TCPI_OPT_SACK
;
1892 if (tp
->rx_opt
.wscale_ok
) {
1893 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
1894 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
1895 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
1898 if (tp
->ecn_flags
&TCP_ECN_OK
)
1899 info
->tcpi_options
|= TCPI_OPT_ECN
;
1901 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
1902 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
1903 info
->tcpi_snd_mss
= tp
->mss_cache
;
1904 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
1906 info
->tcpi_unacked
= tp
->packets_out
;
1907 info
->tcpi_sacked
= tp
->sacked_out
;
1908 info
->tcpi_lost
= tp
->lost_out
;
1909 info
->tcpi_retrans
= tp
->retrans_out
;
1910 info
->tcpi_fackets
= tp
->fackets_out
;
1912 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
1913 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
1914 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
1916 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
1917 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
1918 info
->tcpi_rtt
= jiffies_to_usecs(tp
->srtt
)>>3;
1919 info
->tcpi_rttvar
= jiffies_to_usecs(tp
->mdev
)>>2;
1920 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
1921 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
1922 info
->tcpi_advmss
= tp
->advmss
;
1923 info
->tcpi_reordering
= tp
->reordering
;
1925 info
->tcpi_rcv_rtt
= jiffies_to_usecs(tp
->rcv_rtt_est
.rtt
)>>3;
1926 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
1928 info
->tcpi_total_retrans
= tp
->total_retrans
;
1931 EXPORT_SYMBOL_GPL(tcp_get_info
);
1933 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
1936 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1937 struct tcp_sock
*tp
= tcp_sk(sk
);
1940 if (level
!= SOL_TCP
)
1941 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
1944 if (get_user(len
, optlen
))
1947 len
= min_t(unsigned int, len
, sizeof(int));
1954 val
= tp
->mss_cache
;
1955 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
1956 val
= tp
->rx_opt
.user_mss
;
1959 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
1962 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
1965 val
= (tp
->keepalive_time
? : sysctl_tcp_keepalive_time
) / HZ
;
1968 val
= (tp
->keepalive_intvl
? : sysctl_tcp_keepalive_intvl
) / HZ
;
1971 val
= tp
->keepalive_probes
? : sysctl_tcp_keepalive_probes
;
1974 val
= icsk
->icsk_syn_retries
? : sysctl_tcp_syn_retries
;
1979 val
= (val
? : sysctl_tcp_fin_timeout
) / HZ
;
1981 case TCP_DEFER_ACCEPT
:
1982 val
= !icsk
->icsk_accept_queue
.rskq_defer_accept
? 0 :
1983 ((TCP_TIMEOUT_INIT
/ HZ
) << (icsk
->icsk_accept_queue
.rskq_defer_accept
- 1));
1985 case TCP_WINDOW_CLAMP
:
1986 val
= tp
->window_clamp
;
1989 struct tcp_info info
;
1991 if (get_user(len
, optlen
))
1994 tcp_get_info(sk
, &info
);
1996 len
= min_t(unsigned int, len
, sizeof(info
));
1997 if (put_user(len
, optlen
))
1999 if (copy_to_user(optval
, &info
, len
))
2004 val
= !icsk
->icsk_ack
.pingpong
;
2007 case TCP_CONGESTION
:
2008 if (get_user(len
, optlen
))
2010 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
2011 if (put_user(len
, optlen
))
2013 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
2017 return -ENOPROTOOPT
;
2020 if (put_user(len
, optlen
))
2022 if (copy_to_user(optval
, &val
, len
))
2028 extern void __skb_cb_too_small_for_tcp(int, int);
2029 extern struct tcp_congestion_ops tcp_reno
;
2031 static __initdata
unsigned long thash_entries
;
2032 static int __init
set_thash_entries(char *str
)
2036 thash_entries
= simple_strtoul(str
, &str
, 0);
2039 __setup("thash_entries=", set_thash_entries
);
2041 void __init
tcp_init(void)
2043 struct sk_buff
*skb
= NULL
;
2046 if (sizeof(struct tcp_skb_cb
) > sizeof(skb
->cb
))
2047 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb
),
2050 tcp_hashinfo
.bind_bucket_cachep
=
2051 kmem_cache_create("tcp_bind_bucket",
2052 sizeof(struct inet_bind_bucket
), 0,
2053 SLAB_HWCACHE_ALIGN
, NULL
, NULL
);
2054 if (!tcp_hashinfo
.bind_bucket_cachep
)
2055 panic("tcp_init: Cannot alloc tcp_bind_bucket cache.");
2057 /* Size and allocate the main established and bind bucket
2060 * The methodology is similar to that of the buffer cache.
2062 tcp_hashinfo
.ehash
=
2063 alloc_large_system_hash("TCP established",
2064 sizeof(struct inet_ehash_bucket
),
2066 (num_physpages
>= 128 * 1024) ?
2069 &tcp_hashinfo
.ehash_size
,
2072 tcp_hashinfo
.ehash_size
= (1 << tcp_hashinfo
.ehash_size
) >> 1;
2073 for (i
= 0; i
< (tcp_hashinfo
.ehash_size
<< 1); i
++) {
2074 rwlock_init(&tcp_hashinfo
.ehash
[i
].lock
);
2075 INIT_HLIST_HEAD(&tcp_hashinfo
.ehash
[i
].chain
);
2078 tcp_hashinfo
.bhash
=
2079 alloc_large_system_hash("TCP bind",
2080 sizeof(struct inet_bind_hashbucket
),
2081 tcp_hashinfo
.ehash_size
,
2082 (num_physpages
>= 128 * 1024) ?
2085 &tcp_hashinfo
.bhash_size
,
2088 tcp_hashinfo
.bhash_size
= 1 << tcp_hashinfo
.bhash_size
;
2089 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
2090 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
2091 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
2094 /* Try to be a bit smarter and adjust defaults depending
2095 * on available memory.
2097 for (order
= 0; ((1 << order
) << PAGE_SHIFT
) <
2098 (tcp_hashinfo
.bhash_size
* sizeof(struct inet_bind_hashbucket
));
2102 tcp_death_row
.sysctl_max_tw_buckets
= 180000;
2103 sysctl_tcp_max_orphans
= 4096 << (order
- 4);
2104 sysctl_max_syn_backlog
= 1024;
2105 } else if (order
< 3) {
2106 tcp_death_row
.sysctl_max_tw_buckets
>>= (3 - order
);
2107 sysctl_tcp_max_orphans
>>= (3 - order
);
2108 sysctl_max_syn_backlog
= 128;
2111 sysctl_tcp_mem
[0] = 768 << order
;
2112 sysctl_tcp_mem
[1] = 1024 << order
;
2113 sysctl_tcp_mem
[2] = 1536 << order
;
2116 sysctl_tcp_wmem
[2] = 64 * 1024;
2117 sysctl_tcp_rmem
[0] = PAGE_SIZE
;
2118 sysctl_tcp_rmem
[1] = 43689;
2119 sysctl_tcp_rmem
[2] = 2 * 43689;
2122 printk(KERN_INFO
"TCP: Hash tables configured "
2123 "(established %d bind %d)\n",
2124 tcp_hashinfo
.ehash_size
<< 1, tcp_hashinfo
.bhash_size
);
2126 tcp_register_congestion_control(&tcp_reno
);
2129 EXPORT_SYMBOL(tcp_close
);
2130 EXPORT_SYMBOL(tcp_disconnect
);
2131 EXPORT_SYMBOL(tcp_getsockopt
);
2132 EXPORT_SYMBOL(tcp_ioctl
);
2133 EXPORT_SYMBOL(tcp_poll
);
2134 EXPORT_SYMBOL(tcp_read_sock
);
2135 EXPORT_SYMBOL(tcp_recvmsg
);
2136 EXPORT_SYMBOL(tcp_sendmsg
);
2137 EXPORT_SYMBOL(tcp_sendpage
);
2138 EXPORT_SYMBOL(tcp_setsockopt
);
2139 EXPORT_SYMBOL(tcp_shutdown
);
2140 EXPORT_SYMBOL(tcp_statistics
);