2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp.c,v 1.216 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Alan Cox : Numerous verify_area() calls
24 * Alan Cox : Set the ACK bit on a reset
25 * Alan Cox : Stopped it crashing if it closed while
26 * sk->inuse=1 and was trying to connect
28 * Alan Cox : All icmp error handling was broken
29 * pointers passed where wrong and the
30 * socket was looked up backwards. Nobody
31 * tested any icmp error code obviously.
32 * Alan Cox : tcp_err() now handled properly. It
33 * wakes people on errors. poll
34 * behaves and the icmp error race
35 * has gone by moving it into sock.c
36 * Alan Cox : tcp_send_reset() fixed to work for
37 * everything not just packets for
39 * Alan Cox : tcp option processing.
40 * Alan Cox : Reset tweaked (still not 100%) [Had
42 * Herp Rosmanith : More reset fixes
43 * Alan Cox : No longer acks invalid rst frames.
44 * Acking any kind of RST is right out.
45 * Alan Cox : Sets an ignore me flag on an rst
46 * receive otherwise odd bits of prattle
48 * Alan Cox : Fixed another acking RST frame bug.
49 * Should stop LAN workplace lockups.
50 * Alan Cox : Some tidyups using the new skb list
52 * Alan Cox : sk->keepopen now seems to work
53 * Alan Cox : Pulls options out correctly on accepts
54 * Alan Cox : Fixed assorted sk->rqueue->next errors
55 * Alan Cox : PSH doesn't end a TCP read. Switched a
57 * Alan Cox : Tidied tcp_data to avoid a potential
59 * Alan Cox : Added some better commenting, as the
60 * tcp is hard to follow
61 * Alan Cox : Removed incorrect check for 20 * psh
62 * Michael O'Reilly : ack < copied bug fix.
63 * Johannes Stille : Misc tcp fixes (not all in yet).
64 * Alan Cox : FIN with no memory -> CRASH
65 * Alan Cox : Added socket option proto entries.
66 * Also added awareness of them to accept.
67 * Alan Cox : Added TCP options (SOL_TCP)
68 * Alan Cox : Switched wakeup calls to callbacks,
69 * so the kernel can layer network
71 * Alan Cox : Use ip_tos/ip_ttl settings.
72 * Alan Cox : Handle FIN (more) properly (we hope).
73 * Alan Cox : RST frames sent on unsynchronised
75 * Alan Cox : Put in missing check for SYN bit.
76 * Alan Cox : Added tcp_select_window() aka NET2E
77 * window non shrink trick.
78 * Alan Cox : Added a couple of small NET2E timer
80 * Charles Hedrick : TCP fixes
81 * Toomas Tamm : TCP window fixes
82 * Alan Cox : Small URG fix to rlogin ^C ack fight
83 * Charles Hedrick : Rewrote most of it to actually work
84 * Linus : Rewrote tcp_read() and URG handling
86 * Gerhard Koerting: Fixed some missing timer handling
87 * Matthew Dillon : Reworked TCP machine states as per RFC
88 * Gerhard Koerting: PC/TCP workarounds
89 * Adam Caldwell : Assorted timer/timing errors
90 * Matthew Dillon : Fixed another RST bug
91 * Alan Cox : Move to kernel side addressing changes.
92 * Alan Cox : Beginning work on TCP fastpathing
94 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
95 * Alan Cox : TCP fast path debugging
96 * Alan Cox : Window clamping
97 * Michael Riepe : Bug in tcp_check()
98 * Matt Dillon : More TCP improvements and RST bug fixes
99 * Matt Dillon : Yet more small nasties remove from the
100 * TCP code (Be very nice to this man if
101 * tcp finally works 100%) 8)
102 * Alan Cox : BSD accept semantics.
103 * Alan Cox : Reset on closedown bug.
104 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
105 * Michael Pall : Handle poll() after URG properly in
107 * Michael Pall : Undo the last fix in tcp_read_urg()
108 * (multi URG PUSH broke rlogin).
109 * Michael Pall : Fix the multi URG PUSH problem in
110 * tcp_readable(), poll() after URG
112 * Michael Pall : recv(...,MSG_OOB) never blocks in the
114 * Alan Cox : Changed the semantics of sk->socket to
115 * fix a race and a signal problem with
116 * accept() and async I/O.
117 * Alan Cox : Relaxed the rules on tcp_sendto().
118 * Yury Shevchuk : Really fixed accept() blocking problem.
119 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
120 * clients/servers which listen in on
122 * Alan Cox : Cleaned the above up and shrank it to
123 * a sensible code size.
124 * Alan Cox : Self connect lockup fix.
125 * Alan Cox : No connect to multicast.
126 * Ross Biro : Close unaccepted children on master
128 * Alan Cox : Reset tracing code.
129 * Alan Cox : Spurious resets on shutdown.
130 * Alan Cox : Giant 15 minute/60 second timer error
131 * Alan Cox : Small whoops in polling before an
133 * Alan Cox : Kept the state trace facility since
134 * it's handy for debugging.
135 * Alan Cox : More reset handler fixes.
136 * Alan Cox : Started rewriting the code based on
137 * the RFC's for other useful protocol
138 * references see: Comer, KA9Q NOS, and
139 * for a reference on the difference
140 * between specifications and how BSD
141 * works see the 4.4lite source.
142 * A.N.Kuznetsov : Don't time wait on completion of tidy
144 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
145 * Linus Torvalds : Fixed BSD port reuse to work first syn
146 * Alan Cox : Reimplemented timers as per the RFC
147 * and using multiple timers for sanity.
148 * Alan Cox : Small bug fixes, and a lot of new
150 * Alan Cox : Fixed dual reader crash by locking
151 * the buffers (much like datagram.c)
152 * Alan Cox : Fixed stuck sockets in probe. A probe
153 * now gets fed up of retrying without
154 * (even a no space) answer.
155 * Alan Cox : Extracted closing code better
156 * Alan Cox : Fixed the closing state machine to
158 * Alan Cox : More 'per spec' fixes.
159 * Jorge Cwik : Even faster checksumming.
160 * Alan Cox : tcp_data() doesn't ack illegal PSH
161 * only frames. At least one pc tcp stack
163 * Alan Cox : Cache last socket.
164 * Alan Cox : Per route irtt.
165 * Matt Day : poll()->select() match BSD precisely on error
166 * Alan Cox : New buffers
167 * Marc Tamsky : Various sk->prot->retransmits and
168 * sk->retransmits misupdating fixed.
169 * Fixed tcp_write_timeout: stuck close,
170 * and TCP syn retries gets used now.
171 * Mark Yarvis : In tcp_read_wakeup(), don't send an
172 * ack if state is TCP_CLOSED.
173 * Alan Cox : Look up device on a retransmit - routes may
174 * change. Doesn't yet cope with MSS shrink right
176 * Marc Tamsky : Closing in closing fixes.
177 * Mike Shaver : RFC1122 verifications.
178 * Alan Cox : rcv_saddr errors.
179 * Alan Cox : Block double connect().
180 * Alan Cox : Small hooks for enSKIP.
181 * Alexey Kuznetsov: Path MTU discovery.
182 * Alan Cox : Support soft errors.
183 * Alan Cox : Fix MTU discovery pathological case
184 * when the remote claims no mtu!
185 * Marc Tamsky : TCP_CLOSE fix.
186 * Colin (G3TNE) : Send a reset on syn ack replies in
187 * window but wrong (fixes NT lpd problems)
188 * Pedro Roque : Better TCP window handling, delayed ack.
189 * Joerg Reuter : No modification of locked buffers in
190 * tcp_do_retransmit()
191 * Eric Schenk : Changed receiver side silly window
192 * avoidance algorithm to BSD style
193 * algorithm. This doubles throughput
194 * against machines running Solaris,
195 * and seems to result in general
197 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
198 * Willy Konynenberg : Transparent proxying support.
199 * Mike McLagan : Routing by source
200 * Keith Owens : Do proper merging with partial SKB's in
201 * tcp_do_sendmsg to avoid burstiness.
202 * Eric Schenk : Fix fast close down bug with
203 * shutdown() followed by close().
204 * Andi Kleen : Make poll agree with SIGIO
205 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
206 * lingertime == 0 (RFC 793 ABORT Call)
207 * Hirokazu Takahashi : Use copy_from_user() instead of
208 * csum_and_copy_from_user() if possible.
210 * This program is free software; you can redistribute it and/or
211 * modify it under the terms of the GNU General Public License
212 * as published by the Free Software Foundation; either version
213 * 2 of the License, or(at your option) any later version.
215 * Description of States:
217 * TCP_SYN_SENT sent a connection request, waiting for ack
219 * TCP_SYN_RECV received a connection request, sent ack,
220 * waiting for final ack in three-way handshake.
222 * TCP_ESTABLISHED connection established
224 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
225 * transmission of remaining buffered data
227 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
230 * TCP_CLOSING both sides have shutdown but we still have
231 * data we have to finish sending
233 * TCP_TIME_WAIT timeout to catch resent junk before entering
234 * closed, can only be entered from FIN_WAIT2
235 * or CLOSING. Required because the other end
236 * may not have gotten our last ACK causing it
237 * to retransmit the data packet (which we ignore)
239 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
240 * us to finish writing our data and to shutdown
241 * (we have to close() to move on to LAST_ACK)
243 * TCP_LAST_ACK out side has shutdown after remote has
244 * shutdown. There may still be data in our
245 * buffer that we have to finish sending
247 * TCP_CLOSE socket is finished
250 #include <linux/kernel.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/init.h>
256 #include <linux/fs.h>
257 #include <linux/skbuff.h>
258 #include <linux/splice.h>
259 #include <linux/net.h>
260 #include <linux/socket.h>
261 #include <linux/random.h>
262 #include <linux/bootmem.h>
263 #include <linux/cache.h>
264 #include <linux/err.h>
265 #include <linux/crypto.h>
267 #include <net/icmp.h>
269 #include <net/xfrm.h>
271 #include <net/netdma.h>
272 #include <net/sock.h>
274 #include <asm/uaccess.h>
275 #include <asm/ioctls.h>
277 int sysctl_tcp_fin_timeout __read_mostly
= TCP_FIN_TIMEOUT
;
279 DEFINE_SNMP_STAT(struct tcp_mib
, tcp_statistics
) __read_mostly
;
281 atomic_t tcp_orphan_count
= ATOMIC_INIT(0);
283 EXPORT_SYMBOL_GPL(tcp_orphan_count
);
285 int sysctl_tcp_mem
[3] __read_mostly
;
286 int sysctl_tcp_wmem
[3] __read_mostly
;
287 int sysctl_tcp_rmem
[3] __read_mostly
;
289 EXPORT_SYMBOL(sysctl_tcp_mem
);
290 EXPORT_SYMBOL(sysctl_tcp_rmem
);
291 EXPORT_SYMBOL(sysctl_tcp_wmem
);
293 atomic_t tcp_memory_allocated
; /* Current allocated memory. */
294 atomic_t tcp_sockets_allocated
; /* Current number of TCP sockets. */
296 EXPORT_SYMBOL(tcp_memory_allocated
);
297 EXPORT_SYMBOL(tcp_sockets_allocated
);
302 struct tcp_splice_state
{
303 struct pipe_inode_info
*pipe
;
309 * Pressure flag: try to collapse.
310 * Technical note: it is used by multiple contexts non atomically.
311 * All the sk_stream_mem_schedule() is of this nature: accounting
312 * is strict, actions are advisory and have some latency.
314 int tcp_memory_pressure __read_mostly
;
316 EXPORT_SYMBOL(tcp_memory_pressure
);
318 void tcp_enter_memory_pressure(void)
320 if (!tcp_memory_pressure
) {
321 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES
);
322 tcp_memory_pressure
= 1;
326 EXPORT_SYMBOL(tcp_enter_memory_pressure
);
329 * Wait for a TCP event.
331 * Note that we don't need to lock the socket, as the upper poll layers
332 * take care of normal races (between the test and the event) and we don't
333 * go look at any of the socket buffers directly.
335 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
338 struct sock
*sk
= sock
->sk
;
339 struct tcp_sock
*tp
= tcp_sk(sk
);
341 poll_wait(file
, sk
->sk_sleep
, wait
);
342 if (sk
->sk_state
== TCP_LISTEN
)
343 return inet_csk_listen_poll(sk
);
345 /* Socket is not locked. We are protected from async events
346 by poll logic and correct handling of state changes
347 made by another threads is impossible in any case.
355 * POLLHUP is certainly not done right. But poll() doesn't
356 * have a notion of HUP in just one direction, and for a
357 * socket the read side is more interesting.
359 * Some poll() documentation says that POLLHUP is incompatible
360 * with the POLLOUT/POLLWR flags, so somebody should check this
361 * all. But careful, it tends to be safer to return too many
362 * bits than too few, and you can easily break real applications
363 * if you don't tell them that something has hung up!
367 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
368 * our fs/select.c). It means that after we received EOF,
369 * poll always returns immediately, making impossible poll() on write()
370 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
371 * if and only if shutdown has been made in both directions.
372 * Actually, it is interesting to look how Solaris and DUX
373 * solve this dilemma. I would prefer, if PULLHUP were maskable,
374 * then we could set it on SND_SHUTDOWN. BTW examples given
375 * in Stevens' books assume exactly this behaviour, it explains
376 * why PULLHUP is incompatible with POLLOUT. --ANK
378 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
379 * blocking on fresh not-connected or disconnected socket. --ANK
381 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| sk
->sk_state
== TCP_CLOSE
)
383 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
384 mask
|= POLLIN
| POLLRDNORM
| POLLRDHUP
;
387 if ((1 << sk
->sk_state
) & ~(TCPF_SYN_SENT
| TCPF_SYN_RECV
)) {
388 /* Potential race condition. If read of tp below will
389 * escape above sk->sk_state, we can be illegally awaken
390 * in SYN_* states. */
391 if ((tp
->rcv_nxt
!= tp
->copied_seq
) &&
392 (tp
->urg_seq
!= tp
->copied_seq
||
393 tp
->rcv_nxt
!= tp
->copied_seq
+ 1 ||
394 sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
))
395 mask
|= POLLIN
| POLLRDNORM
;
397 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
398 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
)) {
399 mask
|= POLLOUT
| POLLWRNORM
;
400 } else { /* send SIGIO later */
401 set_bit(SOCK_ASYNC_NOSPACE
,
402 &sk
->sk_socket
->flags
);
403 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
405 /* Race breaker. If space is freed after
406 * wspace test but before the flags are set,
407 * IO signal will be lost.
409 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
))
410 mask
|= POLLOUT
| POLLWRNORM
;
414 if (tp
->urg_data
& TCP_URG_VALID
)
420 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
422 struct tcp_sock
*tp
= tcp_sk(sk
);
427 if (sk
->sk_state
== TCP_LISTEN
)
431 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
433 else if (sock_flag(sk
, SOCK_URGINLINE
) ||
435 before(tp
->urg_seq
, tp
->copied_seq
) ||
436 !before(tp
->urg_seq
, tp
->rcv_nxt
)) {
437 answ
= tp
->rcv_nxt
- tp
->copied_seq
;
439 /* Subtract 1, if FIN is in queue. */
440 if (answ
&& !skb_queue_empty(&sk
->sk_receive_queue
))
442 tcp_hdr((struct sk_buff
*)sk
->sk_receive_queue
.prev
)->fin
;
444 answ
= tp
->urg_seq
- tp
->copied_seq
;
448 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
451 if (sk
->sk_state
== TCP_LISTEN
)
454 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
457 answ
= tp
->write_seq
- tp
->snd_una
;
463 return put_user(answ
, (int __user
*)arg
);
466 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
468 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
469 tp
->pushed_seq
= tp
->write_seq
;
472 static inline int forced_push(struct tcp_sock
*tp
)
474 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
477 static inline void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
479 struct tcp_sock
*tp
= tcp_sk(sk
);
480 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
483 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
484 tcb
->flags
= TCPCB_FLAG_ACK
;
486 skb_header_release(skb
);
487 tcp_add_write_queue_tail(sk
, skb
);
488 sk_charge_skb(sk
, skb
);
489 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
490 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
493 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
,
496 if (flags
& MSG_OOB
) {
498 tp
->snd_up
= tp
->write_seq
;
499 TCP_SKB_CB(skb
)->sacked
|= TCPCB_URG
;
503 static inline void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
506 struct tcp_sock
*tp
= tcp_sk(sk
);
508 if (tcp_send_head(sk
)) {
509 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
510 if (!(flags
& MSG_MORE
) || forced_push(tp
))
511 tcp_mark_push(tp
, skb
);
512 tcp_mark_urg(tp
, flags
, skb
);
513 __tcp_push_pending_frames(sk
, mss_now
,
514 (flags
& MSG_MORE
) ? TCP_NAGLE_CORK
: nonagle
);
518 static int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
519 unsigned int offset
, size_t len
)
521 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
523 return skb_splice_bits(skb
, offset
, tss
->pipe
, tss
->len
, tss
->flags
);
526 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
528 /* Store TCP splice context information in read_descriptor_t. */
529 read_descriptor_t rd_desc
= {
533 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
537 * tcp_splice_read - splice data from TCP socket to a pipe
538 * @sock: socket to splice from
539 * @ppos: position (not valid)
540 * @pipe: pipe to splice to
541 * @len: number of bytes to splice
542 * @flags: splice modifier flags
545 * Will read pages from given socket and fill them into a pipe.
548 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
549 struct pipe_inode_info
*pipe
, size_t len
,
552 struct sock
*sk
= sock
->sk
;
553 struct tcp_splice_state tss
= {
563 * We can't seek on a socket input
572 timeo
= sock_rcvtimeo(sk
, flags
& SPLICE_F_NONBLOCK
);
574 ret
= __tcp_splice_read(sk
, &tss
);
580 if (flags
& SPLICE_F_NONBLOCK
) {
584 if (sock_flag(sk
, SOCK_DONE
))
587 ret
= sock_error(sk
);
590 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
592 if (sk
->sk_state
== TCP_CLOSE
) {
594 * This occurs when user tries to read
595 * from never connected socket.
597 if (!sock_flag(sk
, SOCK_DONE
))
605 sk_wait_data(sk
, &timeo
);
606 if (signal_pending(current
)) {
607 ret
= sock_intr_errno(timeo
);
618 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
619 (sk
->sk_shutdown
& RCV_SHUTDOWN
) || !timeo
||
620 signal_pending(current
))
632 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
)
636 /* The TCP header must be at least 32-bit aligned. */
637 size
= ALIGN(size
, 4);
639 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
641 if (sk_stream_wmem_schedule(sk
, skb
->truesize
)) {
643 * Make sure that we have exactly size bytes
644 * available to the caller, no more, no less.
646 skb_reserve(skb
, skb_tailroom(skb
) - size
);
651 sk
->sk_prot
->enter_memory_pressure();
652 sk_stream_moderate_sndbuf(sk
);
657 static ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
**pages
, int poffset
,
658 size_t psize
, int flags
)
660 struct tcp_sock
*tp
= tcp_sk(sk
);
661 int mss_now
, size_goal
;
664 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
666 /* Wait for a connection to finish. */
667 if ((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
668 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
671 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
673 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
674 size_goal
= tp
->xmit_size_goal
;
678 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
682 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
683 struct page
*page
= pages
[poffset
/ PAGE_SIZE
];
684 int copy
, i
, can_coalesce
;
685 int offset
= poffset
% PAGE_SIZE
;
686 int size
= min_t(size_t, psize
, PAGE_SIZE
- offset
);
688 if (!tcp_send_head(sk
) || (copy
= size_goal
- skb
->len
) <= 0) {
690 if (!sk_stream_memory_free(sk
))
691 goto wait_for_sndbuf
;
693 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
695 goto wait_for_memory
;
704 i
= skb_shinfo(skb
)->nr_frags
;
705 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
706 if (!can_coalesce
&& i
>= MAX_SKB_FRAGS
) {
707 tcp_mark_push(tp
, skb
);
710 if (!sk_stream_wmem_schedule(sk
, copy
))
711 goto wait_for_memory
;
714 skb_shinfo(skb
)->frags
[i
- 1].size
+= copy
;
717 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
721 skb
->data_len
+= copy
;
722 skb
->truesize
+= copy
;
723 sk
->sk_wmem_queued
+= copy
;
724 sk
->sk_forward_alloc
-= copy
;
725 skb
->ip_summed
= CHECKSUM_PARTIAL
;
726 tp
->write_seq
+= copy
;
727 TCP_SKB_CB(skb
)->end_seq
+= copy
;
728 skb_shinfo(skb
)->gso_segs
= 0;
731 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_PSH
;
735 if (!(psize
-= copy
))
738 if (skb
->len
< mss_now
|| (flags
& MSG_OOB
))
741 if (forced_push(tp
)) {
742 tcp_mark_push(tp
, skb
);
743 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
744 } else if (skb
== tcp_send_head(sk
))
745 tcp_push_one(sk
, mss_now
);
749 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
752 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
754 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
757 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
758 size_goal
= tp
->xmit_size_goal
;
763 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
770 return sk_stream_error(sk
, flags
, err
);
773 ssize_t
tcp_sendpage(struct socket
*sock
, struct page
*page
, int offset
,
774 size_t size
, int flags
)
777 struct sock
*sk
= sock
->sk
;
779 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
780 !(sk
->sk_route_caps
& NETIF_F_ALL_CSUM
))
781 return sock_no_sendpage(sock
, page
, offset
, size
, flags
);
785 res
= do_tcp_sendpages(sk
, &page
, offset
, size
, flags
);
791 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
792 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
794 static inline int select_size(struct sock
*sk
)
796 struct tcp_sock
*tp
= tcp_sk(sk
);
797 int tmp
= tp
->mss_cache
;
799 if (sk
->sk_route_caps
& NETIF_F_SG
) {
803 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
805 if (tmp
>= pgbreak
&&
806 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
814 int tcp_sendmsg(struct kiocb
*iocb
, struct socket
*sock
, struct msghdr
*msg
,
817 struct sock
*sk
= sock
->sk
;
819 struct tcp_sock
*tp
= tcp_sk(sk
);
822 int mss_now
, size_goal
;
829 flags
= msg
->msg_flags
;
830 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
832 /* Wait for a connection to finish. */
833 if ((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
834 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
837 /* This should be in poll */
838 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
840 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
841 size_goal
= tp
->xmit_size_goal
;
843 /* Ok commence sending. */
844 iovlen
= msg
->msg_iovlen
;
849 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
852 while (--iovlen
>= 0) {
853 int seglen
= iov
->iov_len
;
854 unsigned char __user
*from
= iov
->iov_base
;
861 skb
= tcp_write_queue_tail(sk
);
863 if (!tcp_send_head(sk
) ||
864 (copy
= size_goal
- skb
->len
) <= 0) {
867 /* Allocate new segment. If the interface is SG,
868 * allocate skb fitting to single page.
870 if (!sk_stream_memory_free(sk
))
871 goto wait_for_sndbuf
;
873 skb
= sk_stream_alloc_skb(sk
, select_size(sk
),
876 goto wait_for_memory
;
879 * Check whether we can use HW checksum.
881 if (sk
->sk_route_caps
& NETIF_F_ALL_CSUM
)
882 skb
->ip_summed
= CHECKSUM_PARTIAL
;
888 /* Try to append data to the end of skb. */
892 /* Where to copy to? */
893 if (skb_tailroom(skb
) > 0) {
894 /* We have some space in skb head. Superb! */
895 if (copy
> skb_tailroom(skb
))
896 copy
= skb_tailroom(skb
);
897 if ((err
= skb_add_data(skb
, from
, copy
)) != 0)
901 int i
= skb_shinfo(skb
)->nr_frags
;
902 struct page
*page
= TCP_PAGE(sk
);
903 int off
= TCP_OFF(sk
);
905 if (skb_can_coalesce(skb
, i
, page
, off
) &&
907 /* We can extend the last page
910 } else if (i
== MAX_SKB_FRAGS
||
912 !(sk
->sk_route_caps
& NETIF_F_SG
))) {
913 /* Need to add new fragment and cannot
914 * do this because interface is non-SG,
915 * or because all the page slots are
917 tcp_mark_push(tp
, skb
);
920 if (off
== PAGE_SIZE
) {
922 TCP_PAGE(sk
) = page
= NULL
;
928 if (copy
> PAGE_SIZE
- off
)
929 copy
= PAGE_SIZE
- off
;
931 if (!sk_stream_wmem_schedule(sk
, copy
))
932 goto wait_for_memory
;
935 /* Allocate new cache page. */
936 if (!(page
= sk_stream_alloc_page(sk
)))
937 goto wait_for_memory
;
940 /* Time to copy data. We are close to
942 err
= skb_copy_to_page(sk
, from
, skb
, page
,
945 /* If this page was new, give it to the
946 * socket so it does not get leaked.
955 /* Update the skb. */
957 skb_shinfo(skb
)->frags
[i
- 1].size
+=
960 skb_fill_page_desc(skb
, i
, page
, off
, copy
);
963 } else if (off
+ copy
< PAGE_SIZE
) {
969 TCP_OFF(sk
) = off
+ copy
;
973 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_PSH
;
975 tp
->write_seq
+= copy
;
976 TCP_SKB_CB(skb
)->end_seq
+= copy
;
977 skb_shinfo(skb
)->gso_segs
= 0;
981 if ((seglen
-= copy
) == 0 && iovlen
== 0)
984 if (skb
->len
< mss_now
|| (flags
& MSG_OOB
))
987 if (forced_push(tp
)) {
988 tcp_mark_push(tp
, skb
);
989 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
990 } else if (skb
== tcp_send_head(sk
))
991 tcp_push_one(sk
, mss_now
);
995 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
998 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
1000 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
1003 mss_now
= tcp_current_mss(sk
, !(flags
&MSG_OOB
));
1004 size_goal
= tp
->xmit_size_goal
;
1010 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
1011 TCP_CHECK_TIMER(sk
);
1017 tcp_unlink_write_queue(skb
, sk
);
1018 /* It is the one place in all of TCP, except connection
1019 * reset, where we can be unlinking the send_head.
1021 tcp_check_send_head(sk
, skb
);
1022 sk_stream_free_skb(sk
, skb
);
1029 err
= sk_stream_error(sk
, flags
, err
);
1030 TCP_CHECK_TIMER(sk
);
1036 * Handle reading urgent data. BSD has very simple semantics for
1037 * this, no blocking and very strange errors 8)
1040 static int tcp_recv_urg(struct sock
*sk
, long timeo
,
1041 struct msghdr
*msg
, int len
, int flags
,
1044 struct tcp_sock
*tp
= tcp_sk(sk
);
1046 /* No URG data to read. */
1047 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1048 tp
->urg_data
== TCP_URG_READ
)
1049 return -EINVAL
; /* Yes this is right ! */
1051 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1054 if (tp
->urg_data
& TCP_URG_VALID
) {
1056 char c
= tp
->urg_data
;
1058 if (!(flags
& MSG_PEEK
))
1059 tp
->urg_data
= TCP_URG_READ
;
1061 /* Read urgent data. */
1062 msg
->msg_flags
|= MSG_OOB
;
1065 if (!(flags
& MSG_TRUNC
))
1066 err
= memcpy_toiovec(msg
->msg_iov
, &c
, 1);
1069 msg
->msg_flags
|= MSG_TRUNC
;
1071 return err
? -EFAULT
: len
;
1074 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1077 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1078 * the available implementations agree in this case:
1079 * this call should never block, independent of the
1080 * blocking state of the socket.
1081 * Mike <pall@rz.uni-karlsruhe.de>
1086 /* Clean up the receive buffer for full frames taken by the user,
1087 * then send an ACK if necessary. COPIED is the number of bytes
1088 * tcp_recvmsg has given to the user so far, it speeds up the
1089 * calculation of whether or not we must ACK for the sake of
1092 void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1094 struct tcp_sock
*tp
= tcp_sk(sk
);
1095 int time_to_ack
= 0;
1098 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1100 BUG_TRAP(!skb
|| before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
));
1103 if (inet_csk_ack_scheduled(sk
)) {
1104 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1105 /* Delayed ACKs frequently hit locked sockets during bulk
1107 if (icsk
->icsk_ack
.blocked
||
1108 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1109 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1111 * If this read emptied read buffer, we send ACK, if
1112 * connection is not bidirectional, user drained
1113 * receive buffer and there was a small segment
1117 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1118 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1119 !icsk
->icsk_ack
.pingpong
)) &&
1120 !atomic_read(&sk
->sk_rmem_alloc
)))
1124 /* We send an ACK if we can now advertise a non-zero window
1125 * which has been raised "significantly".
1127 * Even if window raised up to infinity, do not send window open ACK
1128 * in states, where we will not receive more. It is useless.
1130 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1131 __u32 rcv_window_now
= tcp_receive_window(tp
);
1133 /* Optimize, __tcp_select_window() is not cheap. */
1134 if (2*rcv_window_now
<= tp
->window_clamp
) {
1135 __u32 new_window
= __tcp_select_window(sk
);
1137 /* Send ACK now, if this read freed lots of space
1138 * in our buffer. Certainly, new_window is new window.
1139 * We can advertise it now, if it is not less than current one.
1140 * "Lots" means "at least twice" here.
1142 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1150 static void tcp_prequeue_process(struct sock
*sk
)
1152 struct sk_buff
*skb
;
1153 struct tcp_sock
*tp
= tcp_sk(sk
);
1155 NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED
);
1157 /* RX process wants to run with disabled BHs, though it is not
1160 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
1161 sk
->sk_backlog_rcv(sk
, skb
);
1164 /* Clear memory counter. */
1165 tp
->ucopy
.memory
= 0;
1168 static inline struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1170 struct sk_buff
*skb
;
1173 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1174 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1175 if (tcp_hdr(skb
)->syn
)
1177 if (offset
< skb
->len
|| tcp_hdr(skb
)->fin
) {
1186 * This routine provides an alternative to tcp_recvmsg() for routines
1187 * that would like to handle copying from skbuffs directly in 'sendfile'
1190 * - It is assumed that the socket was locked by the caller.
1191 * - The routine does not block.
1192 * - At present, there is no support for reading OOB data
1193 * or for 'peeking' the socket using this routine
1194 * (although both would be easy to implement).
1196 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1197 sk_read_actor_t recv_actor
)
1199 struct sk_buff
*skb
;
1200 struct tcp_sock
*tp
= tcp_sk(sk
);
1201 u32 seq
= tp
->copied_seq
;
1205 if (sk
->sk_state
== TCP_LISTEN
)
1207 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1208 if (offset
< skb
->len
) {
1211 len
= skb
->len
- offset
;
1212 /* Stop reading if we hit a patch of urgent data */
1214 u32 urg_offset
= tp
->urg_seq
- seq
;
1215 if (urg_offset
< len
)
1220 used
= recv_actor(desc
, skb
, offset
, len
);
1225 } else if (used
<= len
) {
1230 if (offset
!= skb
->len
)
1233 if (tcp_hdr(skb
)->fin
) {
1234 sk_eat_skb(sk
, skb
, 0);
1238 sk_eat_skb(sk
, skb
, 0);
1242 tp
->copied_seq
= seq
;
1244 tcp_rcv_space_adjust(sk
);
1246 /* Clean up data we have read: This will do ACK frames. */
1248 tcp_cleanup_rbuf(sk
, copied
);
1253 * This routine copies from a sock struct into the user buffer.
1255 * Technical note: in 2.3 we work on _locked_ socket, so that
1256 * tricks with *seq access order and skb->users are not required.
1257 * Probably, code can be easily improved even more.
1260 int tcp_recvmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1261 size_t len
, int nonblock
, int flags
, int *addr_len
)
1263 struct tcp_sock
*tp
= tcp_sk(sk
);
1269 int target
; /* Read at least this many bytes */
1271 struct task_struct
*user_recv
= NULL
;
1272 int copied_early
= 0;
1273 struct sk_buff
*skb
;
1277 TCP_CHECK_TIMER(sk
);
1280 if (sk
->sk_state
== TCP_LISTEN
)
1283 timeo
= sock_rcvtimeo(sk
, nonblock
);
1285 /* Urgent data needs to be handled specially. */
1286 if (flags
& MSG_OOB
)
1289 seq
= &tp
->copied_seq
;
1290 if (flags
& MSG_PEEK
) {
1291 peek_seq
= tp
->copied_seq
;
1295 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1297 #ifdef CONFIG_NET_DMA
1298 tp
->ucopy
.dma_chan
= NULL
;
1300 skb
= skb_peek_tail(&sk
->sk_receive_queue
);
1305 available
= TCP_SKB_CB(skb
)->seq
+ skb
->len
- (*seq
);
1306 if ((available
< target
) &&
1307 (len
> sysctl_tcp_dma_copybreak
) && !(flags
& MSG_PEEK
) &&
1308 !sysctl_tcp_low_latency
&&
1309 __get_cpu_var(softnet_data
).net_dma
) {
1310 preempt_enable_no_resched();
1311 tp
->ucopy
.pinned_list
=
1312 dma_pin_iovec_pages(msg
->msg_iov
, len
);
1314 preempt_enable_no_resched();
1322 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1323 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1326 if (signal_pending(current
)) {
1327 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1332 /* Next get a buffer. */
1334 skb
= skb_peek(&sk
->sk_receive_queue
);
1339 /* Now that we have two receive queues this
1342 if (before(*seq
, TCP_SKB_CB(skb
)->seq
)) {
1343 printk(KERN_INFO
"recvmsg bug: copied %X "
1344 "seq %X\n", *seq
, TCP_SKB_CB(skb
)->seq
);
1347 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1348 if (tcp_hdr(skb
)->syn
)
1350 if (offset
< skb
->len
)
1352 if (tcp_hdr(skb
)->fin
)
1354 BUG_TRAP(flags
& MSG_PEEK
);
1356 } while (skb
!= (struct sk_buff
*)&sk
->sk_receive_queue
);
1358 /* Well, if we have backlog, try to process it now yet. */
1360 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1365 sk
->sk_state
== TCP_CLOSE
||
1366 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1368 signal_pending(current
) ||
1372 if (sock_flag(sk
, SOCK_DONE
))
1376 copied
= sock_error(sk
);
1380 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1383 if (sk
->sk_state
== TCP_CLOSE
) {
1384 if (!sock_flag(sk
, SOCK_DONE
)) {
1385 /* This occurs when user tries to read
1386 * from never connected socket.
1399 if (signal_pending(current
)) {
1400 copied
= sock_intr_errno(timeo
);
1405 tcp_cleanup_rbuf(sk
, copied
);
1407 if (!sysctl_tcp_low_latency
&& tp
->ucopy
.task
== user_recv
) {
1408 /* Install new reader */
1409 if (!user_recv
&& !(flags
& (MSG_TRUNC
| MSG_PEEK
))) {
1410 user_recv
= current
;
1411 tp
->ucopy
.task
= user_recv
;
1412 tp
->ucopy
.iov
= msg
->msg_iov
;
1415 tp
->ucopy
.len
= len
;
1417 BUG_TRAP(tp
->copied_seq
== tp
->rcv_nxt
||
1418 (flags
& (MSG_PEEK
| MSG_TRUNC
)));
1420 /* Ugly... If prequeue is not empty, we have to
1421 * process it before releasing socket, otherwise
1422 * order will be broken at second iteration.
1423 * More elegant solution is required!!!
1425 * Look: we have the following (pseudo)queues:
1427 * 1. packets in flight
1432 * Each queue can be processed only if the next ones
1433 * are empty. At this point we have empty receive_queue.
1434 * But prequeue _can_ be not empty after 2nd iteration,
1435 * when we jumped to start of loop because backlog
1436 * processing added something to receive_queue.
1437 * We cannot release_sock(), because backlog contains
1438 * packets arrived _after_ prequeued ones.
1440 * Shortly, algorithm is clear --- to process all
1441 * the queues in order. We could make it more directly,
1442 * requeueing packets from backlog to prequeue, if
1443 * is not empty. It is more elegant, but eats cycles,
1446 if (!skb_queue_empty(&tp
->ucopy
.prequeue
))
1449 /* __ Set realtime policy in scheduler __ */
1452 if (copied
>= target
) {
1453 /* Do not sleep, just process backlog. */
1457 sk_wait_data(sk
, &timeo
);
1459 #ifdef CONFIG_NET_DMA
1460 tp
->ucopy
.wakeup
= 0;
1466 /* __ Restore normal policy in scheduler __ */
1468 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1469 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG
, chunk
);
1474 if (tp
->rcv_nxt
== tp
->copied_seq
&&
1475 !skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1477 tcp_prequeue_process(sk
);
1479 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1480 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1486 if ((flags
& MSG_PEEK
) && peek_seq
!= tp
->copied_seq
) {
1487 if (net_ratelimit())
1488 printk(KERN_DEBUG
"TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1489 current
->comm
, task_pid_nr(current
));
1490 peek_seq
= tp
->copied_seq
;
1495 /* Ok so how much can we use? */
1496 used
= skb
->len
- offset
;
1500 /* Do we have urgent data here? */
1502 u32 urg_offset
= tp
->urg_seq
- *seq
;
1503 if (urg_offset
< used
) {
1505 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1517 if (!(flags
& MSG_TRUNC
)) {
1518 #ifdef CONFIG_NET_DMA
1519 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
1520 tp
->ucopy
.dma_chan
= get_softnet_dma();
1522 if (tp
->ucopy
.dma_chan
) {
1523 tp
->ucopy
.dma_cookie
= dma_skb_copy_datagram_iovec(
1524 tp
->ucopy
.dma_chan
, skb
, offset
,
1526 tp
->ucopy
.pinned_list
);
1528 if (tp
->ucopy
.dma_cookie
< 0) {
1530 printk(KERN_ALERT
"dma_cookie < 0\n");
1532 /* Exception. Bailout! */
1537 if ((offset
+ used
) == skb
->len
)
1543 err
= skb_copy_datagram_iovec(skb
, offset
,
1544 msg
->msg_iov
, used
);
1546 /* Exception. Bailout! */
1558 tcp_rcv_space_adjust(sk
);
1561 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1563 tcp_fast_path_check(sk
);
1565 if (used
+ offset
< skb
->len
)
1568 if (tcp_hdr(skb
)->fin
)
1570 if (!(flags
& MSG_PEEK
)) {
1571 sk_eat_skb(sk
, skb
, copied_early
);
1577 /* Process the FIN. */
1579 if (!(flags
& MSG_PEEK
)) {
1580 sk_eat_skb(sk
, skb
, copied_early
);
1587 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1590 tp
->ucopy
.len
= copied
> 0 ? len
: 0;
1592 tcp_prequeue_process(sk
);
1594 if (copied
> 0 && (chunk
= len
- tp
->ucopy
.len
) != 0) {
1595 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1601 tp
->ucopy
.task
= NULL
;
1605 #ifdef CONFIG_NET_DMA
1606 if (tp
->ucopy
.dma_chan
) {
1607 dma_cookie_t done
, used
;
1609 dma_async_memcpy_issue_pending(tp
->ucopy
.dma_chan
);
1611 while (dma_async_memcpy_complete(tp
->ucopy
.dma_chan
,
1612 tp
->ucopy
.dma_cookie
, &done
,
1613 &used
) == DMA_IN_PROGRESS
) {
1614 /* do partial cleanup of sk_async_wait_queue */
1615 while ((skb
= skb_peek(&sk
->sk_async_wait_queue
)) &&
1616 (dma_async_is_complete(skb
->dma_cookie
, done
,
1617 used
) == DMA_SUCCESS
)) {
1618 __skb_dequeue(&sk
->sk_async_wait_queue
);
1623 /* Safe to free early-copied skbs now */
1624 __skb_queue_purge(&sk
->sk_async_wait_queue
);
1625 dma_chan_put(tp
->ucopy
.dma_chan
);
1626 tp
->ucopy
.dma_chan
= NULL
;
1628 if (tp
->ucopy
.pinned_list
) {
1629 dma_unpin_iovec_pages(tp
->ucopy
.pinned_list
);
1630 tp
->ucopy
.pinned_list
= NULL
;
1634 /* According to UNIX98, msg_name/msg_namelen are ignored
1635 * on connected socket. I was just happy when found this 8) --ANK
1638 /* Clean up data we have read: This will do ACK frames. */
1639 tcp_cleanup_rbuf(sk
, copied
);
1641 TCP_CHECK_TIMER(sk
);
1646 TCP_CHECK_TIMER(sk
);
1651 err
= tcp_recv_urg(sk
, timeo
, msg
, len
, flags
, addr_len
);
1656 * State processing on a close. This implements the state shift for
1657 * sending our FIN frame. Note that we only send a FIN for some
1658 * states. A shutdown() may have already sent the FIN, or we may be
1662 static const unsigned char new_state
[16] = {
1663 /* current state: new state: action: */
1664 /* (Invalid) */ TCP_CLOSE
,
1665 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1666 /* TCP_SYN_SENT */ TCP_CLOSE
,
1667 /* TCP_SYN_RECV */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1668 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1
,
1669 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2
,
1670 /* TCP_TIME_WAIT */ TCP_CLOSE
,
1671 /* TCP_CLOSE */ TCP_CLOSE
,
1672 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK
| TCP_ACTION_FIN
,
1673 /* TCP_LAST_ACK */ TCP_LAST_ACK
,
1674 /* TCP_LISTEN */ TCP_CLOSE
,
1675 /* TCP_CLOSING */ TCP_CLOSING
,
1678 static int tcp_close_state(struct sock
*sk
)
1680 int next
= (int)new_state
[sk
->sk_state
];
1681 int ns
= next
& TCP_STATE_MASK
;
1683 tcp_set_state(sk
, ns
);
1685 return next
& TCP_ACTION_FIN
;
1689 * Shutdown the sending side of a connection. Much like close except
1690 * that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1693 void tcp_shutdown(struct sock
*sk
, int how
)
1695 /* We need to grab some memory, and put together a FIN,
1696 * and then put it into the queue to be sent.
1697 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1699 if (!(how
& SEND_SHUTDOWN
))
1702 /* If we've already sent a FIN, or it's a closed state, skip this. */
1703 if ((1 << sk
->sk_state
) &
1704 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
1705 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
1706 /* Clear out any half completed packets. FIN if needed. */
1707 if (tcp_close_state(sk
))
1712 void tcp_close(struct sock
*sk
, long timeout
)
1714 struct sk_buff
*skb
;
1715 int data_was_unread
= 0;
1719 sk
->sk_shutdown
= SHUTDOWN_MASK
;
1721 if (sk
->sk_state
== TCP_LISTEN
) {
1722 tcp_set_state(sk
, TCP_CLOSE
);
1725 inet_csk_listen_stop(sk
);
1727 goto adjudge_to_death
;
1730 /* We need to flush the recv. buffs. We do this only on the
1731 * descriptor close, not protocol-sourced closes, because the
1732 * reader process may not have drained the data yet!
1734 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
1735 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
-
1737 data_was_unread
+= len
;
1741 sk_stream_mem_reclaim(sk
);
1743 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1744 * data was lost. To witness the awful effects of the old behavior of
1745 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1746 * GET in an FTP client, suspend the process, wait for the client to
1747 * advertise a zero window, then kill -9 the FTP client, wheee...
1748 * Note: timeout is always zero in such a case.
1750 if (data_was_unread
) {
1751 /* Unread data was tossed, zap the connection. */
1752 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE
);
1753 tcp_set_state(sk
, TCP_CLOSE
);
1754 tcp_send_active_reset(sk
, GFP_KERNEL
);
1755 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
1756 /* Check zero linger _after_ checking for unread data. */
1757 sk
->sk_prot
->disconnect(sk
, 0);
1758 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA
);
1759 } else if (tcp_close_state(sk
)) {
1760 /* We FIN if the application ate all the data before
1761 * zapping the connection.
1764 /* RED-PEN. Formally speaking, we have broken TCP state
1765 * machine. State transitions:
1767 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1768 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1769 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1771 * are legal only when FIN has been sent (i.e. in window),
1772 * rather than queued out of window. Purists blame.
1774 * F.e. "RFC state" is ESTABLISHED,
1775 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1777 * The visible declinations are that sometimes
1778 * we enter time-wait state, when it is not required really
1779 * (harmless), do not send active resets, when they are
1780 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1781 * they look as CLOSING or LAST_ACK for Linux)
1782 * Probably, I missed some more holelets.
1788 sk_stream_wait_close(sk
, timeout
);
1791 state
= sk
->sk_state
;
1794 atomic_inc(sk
->sk_prot
->orphan_count
);
1796 /* It is the last release_sock in its life. It will remove backlog. */
1800 /* Now socket is owned by kernel and we acquire BH lock
1801 to finish close. No need to check for user refs.
1805 BUG_TRAP(!sock_owned_by_user(sk
));
1807 /* Have we already been destroyed by a softirq or backlog? */
1808 if (state
!= TCP_CLOSE
&& sk
->sk_state
== TCP_CLOSE
)
1811 /* This is a (useful) BSD violating of the RFC. There is a
1812 * problem with TCP as specified in that the other end could
1813 * keep a socket open forever with no application left this end.
1814 * We use a 3 minute timeout (about the same as BSD) then kill
1815 * our end. If they send after that then tough - BUT: long enough
1816 * that we won't make the old 4*rto = almost no time - whoops
1819 * Nope, it was not mistake. It is really desired behaviour
1820 * f.e. on http servers, when such sockets are useless, but
1821 * consume significant resources. Let's do it with special
1822 * linger2 option. --ANK
1825 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
1826 struct tcp_sock
*tp
= tcp_sk(sk
);
1827 if (tp
->linger2
< 0) {
1828 tcp_set_state(sk
, TCP_CLOSE
);
1829 tcp_send_active_reset(sk
, GFP_ATOMIC
);
1830 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER
);
1832 const int tmo
= tcp_fin_time(sk
);
1834 if (tmo
> TCP_TIMEWAIT_LEN
) {
1835 inet_csk_reset_keepalive_timer(sk
,
1836 tmo
- TCP_TIMEWAIT_LEN
);
1838 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
1843 if (sk
->sk_state
!= TCP_CLOSE
) {
1844 sk_stream_mem_reclaim(sk
);
1845 if (tcp_too_many_orphans(sk
,
1846 atomic_read(sk
->sk_prot
->orphan_count
))) {
1847 if (net_ratelimit())
1848 printk(KERN_INFO
"TCP: too many of orphaned "
1850 tcp_set_state(sk
, TCP_CLOSE
);
1851 tcp_send_active_reset(sk
, GFP_ATOMIC
);
1852 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY
);
1856 if (sk
->sk_state
== TCP_CLOSE
)
1857 inet_csk_destroy_sock(sk
);
1858 /* Otherwise, socket is reprieved until protocol close. */
1866 /* These states need RST on ABORT according to RFC793 */
1868 static inline int tcp_need_reset(int state
)
1870 return (1 << state
) &
1871 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
1872 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
1875 int tcp_disconnect(struct sock
*sk
, int flags
)
1877 struct inet_sock
*inet
= inet_sk(sk
);
1878 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1879 struct tcp_sock
*tp
= tcp_sk(sk
);
1881 int old_state
= sk
->sk_state
;
1883 if (old_state
!= TCP_CLOSE
)
1884 tcp_set_state(sk
, TCP_CLOSE
);
1886 /* ABORT function of RFC793 */
1887 if (old_state
== TCP_LISTEN
) {
1888 inet_csk_listen_stop(sk
);
1889 } else if (tcp_need_reset(old_state
) ||
1890 (tp
->snd_nxt
!= tp
->write_seq
&&
1891 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
1892 /* The last check adjusts for discrepancy of Linux wrt. RFC
1895 tcp_send_active_reset(sk
, gfp_any());
1896 sk
->sk_err
= ECONNRESET
;
1897 } else if (old_state
== TCP_SYN_SENT
)
1898 sk
->sk_err
= ECONNRESET
;
1900 tcp_clear_xmit_timers(sk
);
1901 __skb_queue_purge(&sk
->sk_receive_queue
);
1902 tcp_write_queue_purge(sk
);
1903 __skb_queue_purge(&tp
->out_of_order_queue
);
1904 #ifdef CONFIG_NET_DMA
1905 __skb_queue_purge(&sk
->sk_async_wait_queue
);
1910 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
1911 inet_reset_saddr(sk
);
1913 sk
->sk_shutdown
= 0;
1914 sock_reset_flag(sk
, SOCK_DONE
);
1916 if ((tp
->write_seq
+= tp
->max_window
+ 2) == 0)
1918 icsk
->icsk_backoff
= 0;
1920 icsk
->icsk_probes_out
= 0;
1921 tp
->packets_out
= 0;
1922 tp
->snd_ssthresh
= 0x7fffffff;
1923 tp
->snd_cwnd_cnt
= 0;
1924 tp
->bytes_acked
= 0;
1925 tcp_set_ca_state(sk
, TCP_CA_Open
);
1926 tcp_clear_retrans(tp
);
1927 inet_csk_delack_init(sk
);
1928 tcp_init_send_head(sk
);
1929 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
1932 BUG_TRAP(!inet
->num
|| icsk
->icsk_bind_hash
);
1934 sk
->sk_error_report(sk
);
1939 * Socket option code for TCP.
1941 static int do_tcp_setsockopt(struct sock
*sk
, int level
,
1942 int optname
, char __user
*optval
, int optlen
)
1944 struct tcp_sock
*tp
= tcp_sk(sk
);
1945 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1949 /* This is a string value all the others are int's */
1950 if (optname
== TCP_CONGESTION
) {
1951 char name
[TCP_CA_NAME_MAX
];
1956 val
= strncpy_from_user(name
, optval
,
1957 min(TCP_CA_NAME_MAX
-1, optlen
));
1963 err
= tcp_set_congestion_control(sk
, name
);
1968 if (optlen
< sizeof(int))
1971 if (get_user(val
, (int __user
*)optval
))
1978 /* Values greater than interface MTU won't take effect. However
1979 * at the point when this call is done we typically don't yet
1980 * know which interface is going to be used */
1981 if (val
< 8 || val
> MAX_TCP_WINDOW
) {
1985 tp
->rx_opt
.user_mss
= val
;
1990 /* TCP_NODELAY is weaker than TCP_CORK, so that
1991 * this option on corked socket is remembered, but
1992 * it is not activated until cork is cleared.
1994 * However, when TCP_NODELAY is set we make
1995 * an explicit push, which overrides even TCP_CORK
1996 * for currently queued segments.
1998 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
1999 tcp_push_pending_frames(sk
);
2001 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
2006 /* When set indicates to always queue non-full frames.
2007 * Later the user clears this option and we transmit
2008 * any pending partial frames in the queue. This is
2009 * meant to be used alongside sendfile() to get properly
2010 * filled frames when the user (for example) must write
2011 * out headers with a write() call first and then use
2012 * sendfile to send out the data parts.
2014 * TCP_CORK can be set together with TCP_NODELAY and it is
2015 * stronger than TCP_NODELAY.
2018 tp
->nonagle
|= TCP_NAGLE_CORK
;
2020 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
2021 if (tp
->nonagle
&TCP_NAGLE_OFF
)
2022 tp
->nonagle
|= TCP_NAGLE_PUSH
;
2023 tcp_push_pending_frames(sk
);
2028 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
2031 tp
->keepalive_time
= val
* HZ
;
2032 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
2033 !((1 << sk
->sk_state
) &
2034 (TCPF_CLOSE
| TCPF_LISTEN
))) {
2035 __u32 elapsed
= tcp_time_stamp
- tp
->rcv_tstamp
;
2036 if (tp
->keepalive_time
> elapsed
)
2037 elapsed
= tp
->keepalive_time
- elapsed
;
2040 inet_csk_reset_keepalive_timer(sk
, elapsed
);
2045 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
2048 tp
->keepalive_intvl
= val
* HZ
;
2051 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
2054 tp
->keepalive_probes
= val
;
2057 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
2060 icsk
->icsk_syn_retries
= val
;
2066 else if (val
> sysctl_tcp_fin_timeout
/ HZ
)
2069 tp
->linger2
= val
* HZ
;
2072 case TCP_DEFER_ACCEPT
:
2073 icsk
->icsk_accept_queue
.rskq_defer_accept
= 0;
2075 /* Translate value in seconds to number of
2077 while (icsk
->icsk_accept_queue
.rskq_defer_accept
< 32 &&
2078 val
> ((TCP_TIMEOUT_INIT
/ HZ
) <<
2079 icsk
->icsk_accept_queue
.rskq_defer_accept
))
2080 icsk
->icsk_accept_queue
.rskq_defer_accept
++;
2081 icsk
->icsk_accept_queue
.rskq_defer_accept
++;
2085 case TCP_WINDOW_CLAMP
:
2087 if (sk
->sk_state
!= TCP_CLOSE
) {
2091 tp
->window_clamp
= 0;
2093 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
2094 SOCK_MIN_RCVBUF
/ 2 : val
;
2099 icsk
->icsk_ack
.pingpong
= 1;
2101 icsk
->icsk_ack
.pingpong
= 0;
2102 if ((1 << sk
->sk_state
) &
2103 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
2104 inet_csk_ack_scheduled(sk
)) {
2105 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
2106 tcp_cleanup_rbuf(sk
, 1);
2108 icsk
->icsk_ack
.pingpong
= 1;
2113 #ifdef CONFIG_TCP_MD5SIG
2115 /* Read the IP->Key mappings from userspace */
2116 err
= tp
->af_specific
->md5_parse(sk
, optval
, optlen
);
2129 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2132 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2134 if (level
!= SOL_TCP
)
2135 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
2137 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2140 #ifdef CONFIG_COMPAT
2141 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
2142 char __user
*optval
, int optlen
)
2144 if (level
!= SOL_TCP
)
2145 return inet_csk_compat_setsockopt(sk
, level
, optname
,
2147 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2150 EXPORT_SYMBOL(compat_tcp_setsockopt
);
2153 /* Return information about state of tcp endpoint in API format. */
2154 void tcp_get_info(struct sock
*sk
, struct tcp_info
*info
)
2156 struct tcp_sock
*tp
= tcp_sk(sk
);
2157 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2158 u32 now
= tcp_time_stamp
;
2160 memset(info
, 0, sizeof(*info
));
2162 info
->tcpi_state
= sk
->sk_state
;
2163 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
2164 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
2165 info
->tcpi_probes
= icsk
->icsk_probes_out
;
2166 info
->tcpi_backoff
= icsk
->icsk_backoff
;
2168 if (tp
->rx_opt
.tstamp_ok
)
2169 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
2170 if (tcp_is_sack(tp
))
2171 info
->tcpi_options
|= TCPI_OPT_SACK
;
2172 if (tp
->rx_opt
.wscale_ok
) {
2173 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
2174 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
2175 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
2178 if (tp
->ecn_flags
&TCP_ECN_OK
)
2179 info
->tcpi_options
|= TCPI_OPT_ECN
;
2181 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
2182 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
2183 info
->tcpi_snd_mss
= tp
->mss_cache
;
2184 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
2186 if (sk
->sk_state
== TCP_LISTEN
) {
2187 info
->tcpi_unacked
= sk
->sk_ack_backlog
;
2188 info
->tcpi_sacked
= sk
->sk_max_ack_backlog
;
2190 info
->tcpi_unacked
= tp
->packets_out
;
2191 info
->tcpi_sacked
= tp
->sacked_out
;
2193 info
->tcpi_lost
= tp
->lost_out
;
2194 info
->tcpi_retrans
= tp
->retrans_out
;
2195 info
->tcpi_fackets
= tp
->fackets_out
;
2197 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
2198 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
2199 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
2201 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
2202 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
2203 info
->tcpi_rtt
= jiffies_to_usecs(tp
->srtt
)>>3;
2204 info
->tcpi_rttvar
= jiffies_to_usecs(tp
->mdev
)>>2;
2205 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
2206 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
2207 info
->tcpi_advmss
= tp
->advmss
;
2208 info
->tcpi_reordering
= tp
->reordering
;
2210 info
->tcpi_rcv_rtt
= jiffies_to_usecs(tp
->rcv_rtt_est
.rtt
)>>3;
2211 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
2213 info
->tcpi_total_retrans
= tp
->total_retrans
;
2216 EXPORT_SYMBOL_GPL(tcp_get_info
);
2218 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
2219 int optname
, char __user
*optval
, int __user
*optlen
)
2221 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2222 struct tcp_sock
*tp
= tcp_sk(sk
);
2225 if (get_user(len
, optlen
))
2228 len
= min_t(unsigned int, len
, sizeof(int));
2235 val
= tp
->mss_cache
;
2236 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
2237 val
= tp
->rx_opt
.user_mss
;
2240 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
2243 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
2246 val
= (tp
->keepalive_time
? : sysctl_tcp_keepalive_time
) / HZ
;
2249 val
= (tp
->keepalive_intvl
? : sysctl_tcp_keepalive_intvl
) / HZ
;
2252 val
= tp
->keepalive_probes
? : sysctl_tcp_keepalive_probes
;
2255 val
= icsk
->icsk_syn_retries
? : sysctl_tcp_syn_retries
;
2260 val
= (val
? : sysctl_tcp_fin_timeout
) / HZ
;
2262 case TCP_DEFER_ACCEPT
:
2263 val
= !icsk
->icsk_accept_queue
.rskq_defer_accept
? 0 :
2264 ((TCP_TIMEOUT_INIT
/ HZ
) << (icsk
->icsk_accept_queue
.rskq_defer_accept
- 1));
2266 case TCP_WINDOW_CLAMP
:
2267 val
= tp
->window_clamp
;
2270 struct tcp_info info
;
2272 if (get_user(len
, optlen
))
2275 tcp_get_info(sk
, &info
);
2277 len
= min_t(unsigned int, len
, sizeof(info
));
2278 if (put_user(len
, optlen
))
2280 if (copy_to_user(optval
, &info
, len
))
2285 val
= !icsk
->icsk_ack
.pingpong
;
2288 case TCP_CONGESTION
:
2289 if (get_user(len
, optlen
))
2291 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
2292 if (put_user(len
, optlen
))
2294 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
2298 return -ENOPROTOOPT
;
2301 if (put_user(len
, optlen
))
2303 if (copy_to_user(optval
, &val
, len
))
2308 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2311 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2313 if (level
!= SOL_TCP
)
2314 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
2316 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2319 #ifdef CONFIG_COMPAT
2320 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
2321 char __user
*optval
, int __user
*optlen
)
2323 if (level
!= SOL_TCP
)
2324 return inet_csk_compat_getsockopt(sk
, level
, optname
,
2326 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2329 EXPORT_SYMBOL(compat_tcp_getsockopt
);
2332 struct sk_buff
*tcp_tso_segment(struct sk_buff
*skb
, int features
)
2334 struct sk_buff
*segs
= ERR_PTR(-EINVAL
);
2339 unsigned int oldlen
;
2342 if (!pskb_may_pull(skb
, sizeof(*th
)))
2346 thlen
= th
->doff
* 4;
2347 if (thlen
< sizeof(*th
))
2350 if (!pskb_may_pull(skb
, thlen
))
2353 oldlen
= (u16
)~skb
->len
;
2354 __skb_pull(skb
, thlen
);
2356 if (skb_gso_ok(skb
, features
| NETIF_F_GSO_ROBUST
)) {
2357 /* Packet is from an untrusted source, reset gso_segs. */
2358 int type
= skb_shinfo(skb
)->gso_type
;
2367 !(type
& (SKB_GSO_TCPV4
| SKB_GSO_TCPV6
))))
2370 mss
= skb_shinfo(skb
)->gso_size
;
2371 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss
);
2377 segs
= skb_segment(skb
, features
);
2381 len
= skb_shinfo(skb
)->gso_size
;
2382 delta
= htonl(oldlen
+ (thlen
+ len
));
2386 seq
= ntohl(th
->seq
);
2389 th
->fin
= th
->psh
= 0;
2391 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
2392 (__force u32
)delta
));
2393 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2395 csum_fold(csum_partial(skb_transport_header(skb
),
2402 th
->seq
= htonl(seq
);
2404 } while (skb
->next
);
2406 delta
= htonl(oldlen
+ (skb
->tail
- skb
->transport_header
) +
2408 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
2409 (__force u32
)delta
));
2410 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2411 th
->check
= csum_fold(csum_partial(skb_transport_header(skb
),
2417 EXPORT_SYMBOL(tcp_tso_segment
);
2419 #ifdef CONFIG_TCP_MD5SIG
2420 static unsigned long tcp_md5sig_users
;
2421 static struct tcp_md5sig_pool
**tcp_md5sig_pool
;
2422 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock
);
2424 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool
**pool
)
2427 for_each_possible_cpu(cpu
) {
2428 struct tcp_md5sig_pool
*p
= *per_cpu_ptr(pool
, cpu
);
2430 if (p
->md5_desc
.tfm
)
2431 crypto_free_hash(p
->md5_desc
.tfm
);
2439 void tcp_free_md5sig_pool(void)
2441 struct tcp_md5sig_pool
**pool
= NULL
;
2443 spin_lock_bh(&tcp_md5sig_pool_lock
);
2444 if (--tcp_md5sig_users
== 0) {
2445 pool
= tcp_md5sig_pool
;
2446 tcp_md5sig_pool
= NULL
;
2448 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2450 __tcp_free_md5sig_pool(pool
);
2453 EXPORT_SYMBOL(tcp_free_md5sig_pool
);
2455 static struct tcp_md5sig_pool
**__tcp_alloc_md5sig_pool(void)
2458 struct tcp_md5sig_pool
**pool
;
2460 pool
= alloc_percpu(struct tcp_md5sig_pool
*);
2464 for_each_possible_cpu(cpu
) {
2465 struct tcp_md5sig_pool
*p
;
2466 struct crypto_hash
*hash
;
2468 p
= kzalloc(sizeof(*p
), GFP_KERNEL
);
2471 *per_cpu_ptr(pool
, cpu
) = p
;
2473 hash
= crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC
);
2474 if (!hash
|| IS_ERR(hash
))
2477 p
->md5_desc
.tfm
= hash
;
2481 __tcp_free_md5sig_pool(pool
);
2485 struct tcp_md5sig_pool
**tcp_alloc_md5sig_pool(void)
2487 struct tcp_md5sig_pool
**pool
;
2491 spin_lock_bh(&tcp_md5sig_pool_lock
);
2492 pool
= tcp_md5sig_pool
;
2493 if (tcp_md5sig_users
++ == 0) {
2495 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2498 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2502 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2505 /* we cannot hold spinlock here because this may sleep. */
2506 struct tcp_md5sig_pool
**p
= __tcp_alloc_md5sig_pool();
2507 spin_lock_bh(&tcp_md5sig_pool_lock
);
2510 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2513 pool
= tcp_md5sig_pool
;
2515 /* oops, it has already been assigned. */
2516 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2517 __tcp_free_md5sig_pool(p
);
2519 tcp_md5sig_pool
= pool
= p
;
2520 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2526 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
2528 struct tcp_md5sig_pool
*__tcp_get_md5sig_pool(int cpu
)
2530 struct tcp_md5sig_pool
**p
;
2531 spin_lock_bh(&tcp_md5sig_pool_lock
);
2532 p
= tcp_md5sig_pool
;
2535 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2536 return (p
? *per_cpu_ptr(p
, cpu
) : NULL
);
2539 EXPORT_SYMBOL(__tcp_get_md5sig_pool
);
2541 void __tcp_put_md5sig_pool(void)
2543 tcp_free_md5sig_pool();
2546 EXPORT_SYMBOL(__tcp_put_md5sig_pool
);
2549 void tcp_done(struct sock
*sk
)
2551 if(sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
2552 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS
);
2554 tcp_set_state(sk
, TCP_CLOSE
);
2555 tcp_clear_xmit_timers(sk
);
2557 sk
->sk_shutdown
= SHUTDOWN_MASK
;
2559 if (!sock_flag(sk
, SOCK_DEAD
))
2560 sk
->sk_state_change(sk
);
2562 inet_csk_destroy_sock(sk
);
2564 EXPORT_SYMBOL_GPL(tcp_done
);
2566 extern struct tcp_congestion_ops tcp_reno
;
2568 static __initdata
unsigned long thash_entries
;
2569 static int __init
set_thash_entries(char *str
)
2573 thash_entries
= simple_strtoul(str
, &str
, 0);
2576 __setup("thash_entries=", set_thash_entries
);
2578 void __init
tcp_init(void)
2580 struct sk_buff
*skb
= NULL
;
2581 unsigned long limit
;
2582 int order
, i
, max_share
;
2584 BUILD_BUG_ON(sizeof(struct tcp_skb_cb
) > sizeof(skb
->cb
));
2586 tcp_hashinfo
.bind_bucket_cachep
=
2587 kmem_cache_create("tcp_bind_bucket",
2588 sizeof(struct inet_bind_bucket
), 0,
2589 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
);
2591 /* Size and allocate the main established and bind bucket
2594 * The methodology is similar to that of the buffer cache.
2596 tcp_hashinfo
.ehash
=
2597 alloc_large_system_hash("TCP established",
2598 sizeof(struct inet_ehash_bucket
),
2600 (num_physpages
>= 128 * 1024) ?
2603 &tcp_hashinfo
.ehash_size
,
2605 thash_entries
? 0 : 512 * 1024);
2606 tcp_hashinfo
.ehash_size
= 1 << tcp_hashinfo
.ehash_size
;
2607 for (i
= 0; i
< tcp_hashinfo
.ehash_size
; i
++) {
2608 INIT_HLIST_HEAD(&tcp_hashinfo
.ehash
[i
].chain
);
2609 INIT_HLIST_HEAD(&tcp_hashinfo
.ehash
[i
].twchain
);
2611 if (inet_ehash_locks_alloc(&tcp_hashinfo
))
2612 panic("TCP: failed to alloc ehash_locks");
2613 tcp_hashinfo
.bhash
=
2614 alloc_large_system_hash("TCP bind",
2615 sizeof(struct inet_bind_hashbucket
),
2616 tcp_hashinfo
.ehash_size
,
2617 (num_physpages
>= 128 * 1024) ?
2620 &tcp_hashinfo
.bhash_size
,
2623 tcp_hashinfo
.bhash_size
= 1 << tcp_hashinfo
.bhash_size
;
2624 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
2625 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
2626 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
2629 /* Try to be a bit smarter and adjust defaults depending
2630 * on available memory.
2632 for (order
= 0; ((1 << order
) << PAGE_SHIFT
) <
2633 (tcp_hashinfo
.bhash_size
* sizeof(struct inet_bind_hashbucket
));
2637 tcp_death_row
.sysctl_max_tw_buckets
= 180000;
2638 sysctl_tcp_max_orphans
= 4096 << (order
- 4);
2639 sysctl_max_syn_backlog
= 1024;
2640 } else if (order
< 3) {
2641 tcp_death_row
.sysctl_max_tw_buckets
>>= (3 - order
);
2642 sysctl_tcp_max_orphans
>>= (3 - order
);
2643 sysctl_max_syn_backlog
= 128;
2646 /* Set the pressure threshold to be a fraction of global memory that
2647 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2648 * memory, with a floor of 128 pages.
2650 limit
= min(nr_all_pages
, 1UL<<(28-PAGE_SHIFT
)) >> (20-PAGE_SHIFT
);
2651 limit
= (limit
* (nr_all_pages
>> (20-PAGE_SHIFT
))) >> (PAGE_SHIFT
-11);
2652 limit
= max(limit
, 128UL);
2653 sysctl_tcp_mem
[0] = limit
/ 4 * 3;
2654 sysctl_tcp_mem
[1] = limit
;
2655 sysctl_tcp_mem
[2] = sysctl_tcp_mem
[0] * 2;
2657 /* Set per-socket limits to no more than 1/128 the pressure threshold */
2658 limit
= ((unsigned long)sysctl_tcp_mem
[1]) << (PAGE_SHIFT
- 7);
2659 max_share
= min(4UL*1024*1024, limit
);
2661 sysctl_tcp_wmem
[0] = SK_STREAM_MEM_QUANTUM
;
2662 sysctl_tcp_wmem
[1] = 16*1024;
2663 sysctl_tcp_wmem
[2] = max(64*1024, max_share
);
2665 sysctl_tcp_rmem
[0] = SK_STREAM_MEM_QUANTUM
;
2666 sysctl_tcp_rmem
[1] = 87380;
2667 sysctl_tcp_rmem
[2] = max(87380, max_share
);
2669 printk(KERN_INFO
"TCP: Hash tables configured "
2670 "(established %d bind %d)\n",
2671 tcp_hashinfo
.ehash_size
, tcp_hashinfo
.bhash_size
);
2673 tcp_register_congestion_control(&tcp_reno
);
2676 EXPORT_SYMBOL(tcp_close
);
2677 EXPORT_SYMBOL(tcp_disconnect
);
2678 EXPORT_SYMBOL(tcp_getsockopt
);
2679 EXPORT_SYMBOL(tcp_ioctl
);
2680 EXPORT_SYMBOL(tcp_poll
);
2681 EXPORT_SYMBOL(tcp_read_sock
);
2682 EXPORT_SYMBOL(tcp_recvmsg
);
2683 EXPORT_SYMBOL(tcp_sendmsg
);
2684 EXPORT_SYMBOL(tcp_splice_read
);
2685 EXPORT_SYMBOL(tcp_sendpage
);
2686 EXPORT_SYMBOL(tcp_setsockopt
);
2687 EXPORT_SYMBOL(tcp_shutdown
);
2688 EXPORT_SYMBOL(tcp_statistics
);