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).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
21 * Alan Cox : Numerous verify_area() calls
22 * Alan Cox : Set the ACK bit on a reset
23 * Alan Cox : Stopped it crashing if it closed while
24 * sk->inuse=1 and was trying to connect
26 * Alan Cox : All icmp error handling was broken
27 * pointers passed where wrong and the
28 * socket was looked up backwards. Nobody
29 * tested any icmp error code obviously.
30 * Alan Cox : tcp_err() now handled properly. It
31 * wakes people on errors. poll
32 * behaves and the icmp error race
33 * has gone by moving it into sock.c
34 * Alan Cox : tcp_send_reset() fixed to work for
35 * everything not just packets for
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
40 * Herp Rosmanith : More reset fixes
41 * Alan Cox : No longer acks invalid rst frames.
42 * Acking any kind of RST is right out.
43 * Alan Cox : Sets an ignore me flag on an rst
44 * receive otherwise odd bits of prattle
46 * Alan Cox : Fixed another acking RST frame bug.
47 * Should stop LAN workplace lockups.
48 * Alan Cox : Some tidyups using the new skb list
50 * Alan Cox : sk->keepopen now seems to work
51 * Alan Cox : Pulls options out correctly on accepts
52 * Alan Cox : Fixed assorted sk->rqueue->next errors
53 * Alan Cox : PSH doesn't end a TCP read. Switched a
55 * Alan Cox : Tidied tcp_data to avoid a potential
57 * Alan Cox : Added some better commenting, as the
58 * tcp is hard to follow
59 * Alan Cox : Removed incorrect check for 20 * psh
60 * Michael O'Reilly : ack < copied bug fix.
61 * Johannes Stille : Misc tcp fixes (not all in yet).
62 * Alan Cox : FIN with no memory -> CRASH
63 * Alan Cox : Added socket option proto entries.
64 * Also added awareness of them to accept.
65 * Alan Cox : Added TCP options (SOL_TCP)
66 * Alan Cox : Switched wakeup calls to callbacks,
67 * so the kernel can layer network
69 * Alan Cox : Use ip_tos/ip_ttl settings.
70 * Alan Cox : Handle FIN (more) properly (we hope).
71 * Alan Cox : RST frames sent on unsynchronised
73 * Alan Cox : Put in missing check for SYN bit.
74 * Alan Cox : Added tcp_select_window() aka NET2E
75 * window non shrink trick.
76 * Alan Cox : Added a couple of small NET2E timer
78 * Charles Hedrick : TCP fixes
79 * Toomas Tamm : TCP window fixes
80 * Alan Cox : Small URG fix to rlogin ^C ack fight
81 * Charles Hedrick : Rewrote most of it to actually work
82 * Linus : Rewrote tcp_read() and URG handling
84 * Gerhard Koerting: Fixed some missing timer handling
85 * Matthew Dillon : Reworked TCP machine states as per RFC
86 * Gerhard Koerting: PC/TCP workarounds
87 * Adam Caldwell : Assorted timer/timing errors
88 * Matthew Dillon : Fixed another RST bug
89 * Alan Cox : Move to kernel side addressing changes.
90 * Alan Cox : Beginning work on TCP fastpathing
92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
93 * Alan Cox : TCP fast path debugging
94 * Alan Cox : Window clamping
95 * Michael Riepe : Bug in tcp_check()
96 * Matt Dillon : More TCP improvements and RST bug fixes
97 * Matt Dillon : Yet more small nasties remove from the
98 * TCP code (Be very nice to this man if
99 * tcp finally works 100%) 8)
100 * Alan Cox : BSD accept semantics.
101 * Alan Cox : Reset on closedown bug.
102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
103 * Michael Pall : Handle poll() after URG properly in
105 * Michael Pall : Undo the last fix in tcp_read_urg()
106 * (multi URG PUSH broke rlogin).
107 * Michael Pall : Fix the multi URG PUSH problem in
108 * tcp_readable(), poll() after URG
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
112 * Alan Cox : Changed the semantics of sk->socket to
113 * fix a race and a signal problem with
114 * accept() and async I/O.
115 * Alan Cox : Relaxed the rules on tcp_sendto().
116 * Yury Shevchuk : Really fixed accept() blocking problem.
117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
118 * clients/servers which listen in on
120 * Alan Cox : Cleaned the above up and shrank it to
121 * a sensible code size.
122 * Alan Cox : Self connect lockup fix.
123 * Alan Cox : No connect to multicast.
124 * Ross Biro : Close unaccepted children on master
126 * Alan Cox : Reset tracing code.
127 * Alan Cox : Spurious resets on shutdown.
128 * Alan Cox : Giant 15 minute/60 second timer error
129 * Alan Cox : Small whoops in polling before an
131 * Alan Cox : Kept the state trace facility since
132 * it's handy for debugging.
133 * Alan Cox : More reset handler fixes.
134 * Alan Cox : Started rewriting the code based on
135 * the RFC's for other useful protocol
136 * references see: Comer, KA9Q NOS, and
137 * for a reference on the difference
138 * between specifications and how BSD
139 * works see the 4.4lite source.
140 * A.N.Kuznetsov : Don't time wait on completion of tidy
142 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
143 * Linus Torvalds : Fixed BSD port reuse to work first syn
144 * Alan Cox : Reimplemented timers as per the RFC
145 * and using multiple timers for sanity.
146 * Alan Cox : Small bug fixes, and a lot of new
148 * Alan Cox : Fixed dual reader crash by locking
149 * the buffers (much like datagram.c)
150 * Alan Cox : Fixed stuck sockets in probe. A probe
151 * now gets fed up of retrying without
152 * (even a no space) answer.
153 * Alan Cox : Extracted closing code better
154 * Alan Cox : Fixed the closing state machine to
156 * Alan Cox : More 'per spec' fixes.
157 * Jorge Cwik : Even faster checksumming.
158 * Alan Cox : tcp_data() doesn't ack illegal PSH
159 * only frames. At least one pc tcp stack
161 * Alan Cox : Cache last socket.
162 * Alan Cox : Per route irtt.
163 * Matt Day : poll()->select() match BSD precisely on error
164 * Alan Cox : New buffers
165 * Marc Tamsky : Various sk->prot->retransmits and
166 * sk->retransmits misupdating fixed.
167 * Fixed tcp_write_timeout: stuck close,
168 * and TCP syn retries gets used now.
169 * Mark Yarvis : In tcp_read_wakeup(), don't send an
170 * ack if state is TCP_CLOSED.
171 * Alan Cox : Look up device on a retransmit - routes may
172 * change. Doesn't yet cope with MSS shrink right
174 * Marc Tamsky : Closing in closing fixes.
175 * Mike Shaver : RFC1122 verifications.
176 * Alan Cox : rcv_saddr errors.
177 * Alan Cox : Block double connect().
178 * Alan Cox : Small hooks for enSKIP.
179 * Alexey Kuznetsov: Path MTU discovery.
180 * Alan Cox : Support soft errors.
181 * Alan Cox : Fix MTU discovery pathological case
182 * when the remote claims no mtu!
183 * Marc Tamsky : TCP_CLOSE fix.
184 * Colin (G3TNE) : Send a reset on syn ack replies in
185 * window but wrong (fixes NT lpd problems)
186 * Pedro Roque : Better TCP window handling, delayed ack.
187 * Joerg Reuter : No modification of locked buffers in
188 * tcp_do_retransmit()
189 * Eric Schenk : Changed receiver side silly window
190 * avoidance algorithm to BSD style
191 * algorithm. This doubles throughput
192 * against machines running Solaris,
193 * and seems to result in general
195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
196 * Willy Konynenberg : Transparent proxying support.
197 * Mike McLagan : Routing by source
198 * Keith Owens : Do proper merging with partial SKB's in
199 * tcp_do_sendmsg to avoid burstiness.
200 * Eric Schenk : Fix fast close down bug with
201 * shutdown() followed by close().
202 * Andi Kleen : Make poll agree with SIGIO
203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
204 * lingertime == 0 (RFC 793 ABORT Call)
205 * Hirokazu Takahashi : Use copy_from_user() instead of
206 * csum_and_copy_from_user() if possible.
208 * This program is free software; you can redistribute it and/or
209 * modify it under the terms of the GNU General Public License
210 * as published by the Free Software Foundation; either version
211 * 2 of the License, or(at your option) any later version.
213 * Description of States:
215 * TCP_SYN_SENT sent a connection request, waiting for ack
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
220 * TCP_ESTABLISHED connection established
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
231 * TCP_TIME_WAIT timeout to catch resent junk before entering
232 * closed, can only be entered from FIN_WAIT2
233 * or CLOSING. Required because the other end
234 * may not have gotten our last ACK causing it
235 * to retransmit the data packet (which we ignore)
237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
238 * us to finish writing our data and to shutdown
239 * (we have to close() to move on to LAST_ACK)
241 * TCP_LAST_ACK out side has shutdown after remote has
242 * shutdown. There may still be data in our
243 * buffer that we have to finish sending
245 * TCP_CLOSE socket is finished
248 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
267 #include <linux/time.h>
268 #include <linux/slab.h>
270 #include <net/icmp.h>
272 #include <net/xfrm.h>
274 #include <net/netdma.h>
275 #include <net/sock.h>
277 #include <asm/uaccess.h>
278 #include <asm/ioctls.h>
280 int sysctl_tcp_fin_timeout __read_mostly
= TCP_FIN_TIMEOUT
;
282 struct percpu_counter tcp_orphan_count
;
283 EXPORT_SYMBOL_GPL(tcp_orphan_count
);
285 long 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_long_t tcp_memory_allocated
; /* Current allocated memory. */
294 EXPORT_SYMBOL(tcp_memory_allocated
);
297 * Current number of TCP sockets.
299 struct percpu_counter tcp_sockets_allocated
;
300 EXPORT_SYMBOL(tcp_sockets_allocated
);
305 struct tcp_splice_state
{
306 struct pipe_inode_info
*pipe
;
312 * Pressure flag: try to collapse.
313 * Technical note: it is used by multiple contexts non atomically.
314 * All the __sk_mem_schedule() is of this nature: accounting
315 * is strict, actions are advisory and have some latency.
317 int tcp_memory_pressure __read_mostly
;
319 EXPORT_SYMBOL(tcp_memory_pressure
);
321 void tcp_enter_memory_pressure(struct sock
*sk
)
323 if (!tcp_memory_pressure
) {
324 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMEMORYPRESSURES
);
325 tcp_memory_pressure
= 1;
329 EXPORT_SYMBOL(tcp_enter_memory_pressure
);
331 /* Convert seconds to retransmits based on initial and max timeout */
332 static u8
secs_to_retrans(int seconds
, int timeout
, int rto_max
)
337 int period
= timeout
;
340 while (seconds
> period
&& res
< 255) {
343 if (timeout
> rto_max
)
351 /* Convert retransmits to seconds based on initial and max timeout */
352 static int retrans_to_secs(u8 retrans
, int timeout
, int rto_max
)
360 if (timeout
> rto_max
)
369 * Wait for a TCP event.
371 * Note that we don't need to lock the socket, as the upper poll layers
372 * take care of normal races (between the test and the event) and we don't
373 * go look at any of the socket buffers directly.
375 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
378 struct sock
*sk
= sock
->sk
;
379 struct tcp_sock
*tp
= tcp_sk(sk
);
381 sock_poll_wait(file
, sk
->sk_sleep
, wait
);
382 if (sk
->sk_state
== TCP_LISTEN
)
383 return inet_csk_listen_poll(sk
);
385 /* Socket is not locked. We are protected from async events
386 * by poll logic and correct handling of state changes
387 * made by other threads is impossible in any case.
393 * POLLHUP is certainly not done right. But poll() doesn't
394 * have a notion of HUP in just one direction, and for a
395 * socket the read side is more interesting.
397 * Some poll() documentation says that POLLHUP is incompatible
398 * with the POLLOUT/POLLWR flags, so somebody should check this
399 * all. But careful, it tends to be safer to return too many
400 * bits than too few, and you can easily break real applications
401 * if you don't tell them that something has hung up!
405 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
406 * our fs/select.c). It means that after we received EOF,
407 * poll always returns immediately, making impossible poll() on write()
408 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
409 * if and only if shutdown has been made in both directions.
410 * Actually, it is interesting to look how Solaris and DUX
411 * solve this dilemma. I would prefer, if POLLHUP were maskable,
412 * then we could set it on SND_SHUTDOWN. BTW examples given
413 * in Stevens' books assume exactly this behaviour, it explains
414 * why POLLHUP is incompatible with POLLOUT. --ANK
416 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
417 * blocking on fresh not-connected or disconnected socket. --ANK
419 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| sk
->sk_state
== TCP_CLOSE
)
421 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
422 mask
|= POLLIN
| POLLRDNORM
| POLLRDHUP
;
425 if ((1 << sk
->sk_state
) & ~(TCPF_SYN_SENT
| TCPF_SYN_RECV
)) {
426 int target
= sock_rcvlowat(sk
, 0, INT_MAX
);
428 if (tp
->urg_seq
== tp
->copied_seq
&&
429 !sock_flag(sk
, SOCK_URGINLINE
) &&
433 /* Potential race condition. If read of tp below will
434 * escape above sk->sk_state, we can be illegally awaken
435 * in SYN_* states. */
436 if (tp
->rcv_nxt
- tp
->copied_seq
>= target
)
437 mask
|= POLLIN
| POLLRDNORM
;
439 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
440 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
)) {
441 mask
|= POLLOUT
| POLLWRNORM
;
442 } else { /* send SIGIO later */
443 set_bit(SOCK_ASYNC_NOSPACE
,
444 &sk
->sk_socket
->flags
);
445 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
447 /* Race breaker. If space is freed after
448 * wspace test but before the flags are set,
449 * IO signal will be lost.
451 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
))
452 mask
|= POLLOUT
| POLLWRNORM
;
455 mask
|= POLLOUT
| POLLWRNORM
;
457 if (tp
->urg_data
& TCP_URG_VALID
)
460 /* This barrier is coupled with smp_wmb() in tcp_reset() */
468 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
470 struct tcp_sock
*tp
= tcp_sk(sk
);
475 if (sk
->sk_state
== TCP_LISTEN
)
479 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
481 else if (sock_flag(sk
, SOCK_URGINLINE
) ||
483 before(tp
->urg_seq
, tp
->copied_seq
) ||
484 !before(tp
->urg_seq
, tp
->rcv_nxt
)) {
487 answ
= tp
->rcv_nxt
- tp
->copied_seq
;
489 /* Subtract 1, if FIN is in queue. */
490 skb
= skb_peek_tail(&sk
->sk_receive_queue
);
492 answ
-= tcp_hdr(skb
)->fin
;
494 answ
= tp
->urg_seq
- tp
->copied_seq
;
498 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
501 if (sk
->sk_state
== TCP_LISTEN
)
504 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
507 answ
= tp
->write_seq
- tp
->snd_una
;
513 return put_user(answ
, (int __user
*)arg
);
516 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
518 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
519 tp
->pushed_seq
= tp
->write_seq
;
522 static inline int forced_push(struct tcp_sock
*tp
)
524 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
527 static inline void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
529 struct tcp_sock
*tp
= tcp_sk(sk
);
530 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
533 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
534 tcb
->flags
= TCPCB_FLAG_ACK
;
536 skb_header_release(skb
);
537 tcp_add_write_queue_tail(sk
, skb
);
538 sk
->sk_wmem_queued
+= skb
->truesize
;
539 sk_mem_charge(sk
, skb
->truesize
);
540 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
541 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
544 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
)
547 tp
->snd_up
= tp
->write_seq
;
550 static inline void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
553 if (tcp_send_head(sk
)) {
554 struct tcp_sock
*tp
= tcp_sk(sk
);
556 if (!(flags
& MSG_MORE
) || forced_push(tp
))
557 tcp_mark_push(tp
, tcp_write_queue_tail(sk
));
559 tcp_mark_urg(tp
, flags
);
560 __tcp_push_pending_frames(sk
, mss_now
,
561 (flags
& MSG_MORE
) ? TCP_NAGLE_CORK
: nonagle
);
565 static int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
566 unsigned int offset
, size_t len
)
568 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
571 ret
= skb_splice_bits(skb
, offset
, tss
->pipe
, min(rd_desc
->count
, len
),
574 rd_desc
->count
-= ret
;
578 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
580 /* Store TCP splice context information in read_descriptor_t. */
581 read_descriptor_t rd_desc
= {
586 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
590 * tcp_splice_read - splice data from TCP socket to a pipe
591 * @sock: socket to splice from
592 * @ppos: position (not valid)
593 * @pipe: pipe to splice to
594 * @len: number of bytes to splice
595 * @flags: splice modifier flags
598 * Will read pages from given socket and fill them into a pipe.
601 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
602 struct pipe_inode_info
*pipe
, size_t len
,
605 struct sock
*sk
= sock
->sk
;
606 struct tcp_splice_state tss
= {
616 * We can't seek on a socket input
625 timeo
= sock_rcvtimeo(sk
, sock
->file
->f_flags
& O_NONBLOCK
);
627 ret
= __tcp_splice_read(sk
, &tss
);
633 if (sock_flag(sk
, SOCK_DONE
))
636 ret
= sock_error(sk
);
639 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
641 if (sk
->sk_state
== TCP_CLOSE
) {
643 * This occurs when user tries to read
644 * from never connected socket.
646 if (!sock_flag(sk
, SOCK_DONE
))
654 sk_wait_data(sk
, &timeo
);
655 if (signal_pending(current
)) {
656 ret
= sock_intr_errno(timeo
);
669 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
670 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
671 signal_pending(current
))
683 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
)
687 /* The TCP header must be at least 32-bit aligned. */
688 size
= ALIGN(size
, 4);
690 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
692 if (sk_wmem_schedule(sk
, skb
->truesize
)) {
694 * Make sure that we have exactly size bytes
695 * available to the caller, no more, no less.
697 skb_reserve(skb
, skb_tailroom(skb
) - size
);
702 sk
->sk_prot
->enter_memory_pressure(sk
);
703 sk_stream_moderate_sndbuf(sk
);
708 static unsigned int tcp_xmit_size_goal(struct sock
*sk
, u32 mss_now
,
711 struct tcp_sock
*tp
= tcp_sk(sk
);
712 u32 xmit_size_goal
, old_size_goal
;
714 xmit_size_goal
= mss_now
;
716 if (large_allowed
&& sk_can_gso(sk
)) {
717 xmit_size_goal
= ((sk
->sk_gso_max_size
- 1) -
718 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
719 inet_csk(sk
)->icsk_ext_hdr_len
-
722 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
724 /* We try hard to avoid divides here */
725 old_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
727 if (likely(old_size_goal
<= xmit_size_goal
&&
728 old_size_goal
+ mss_now
> xmit_size_goal
)) {
729 xmit_size_goal
= old_size_goal
;
731 tp
->xmit_size_goal_segs
= xmit_size_goal
/ mss_now
;
732 xmit_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
736 return max(xmit_size_goal
, mss_now
);
739 static int tcp_send_mss(struct sock
*sk
, int *size_goal
, int flags
)
743 mss_now
= tcp_current_mss(sk
);
744 *size_goal
= tcp_xmit_size_goal(sk
, mss_now
, !(flags
& MSG_OOB
));
749 static ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
**pages
, int poffset
,
750 size_t psize
, int flags
)
752 struct tcp_sock
*tp
= tcp_sk(sk
);
753 int mss_now
, size_goal
;
756 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
758 /* Wait for a connection to finish. */
759 if ((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
760 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
763 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
765 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
769 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
773 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
774 struct page
*page
= pages
[poffset
/ PAGE_SIZE
];
775 int copy
, i
, can_coalesce
;
776 int offset
= poffset
% PAGE_SIZE
;
777 int size
= min_t(size_t, psize
, PAGE_SIZE
- offset
);
779 if (!tcp_send_head(sk
) || (copy
= size_goal
- skb
->len
) <= 0) {
781 if (!sk_stream_memory_free(sk
))
782 goto wait_for_sndbuf
;
784 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
786 goto wait_for_memory
;
795 i
= skb_shinfo(skb
)->nr_frags
;
796 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
797 if (!can_coalesce
&& i
>= MAX_SKB_FRAGS
) {
798 tcp_mark_push(tp
, skb
);
801 if (!sk_wmem_schedule(sk
, copy
))
802 goto wait_for_memory
;
805 skb_shinfo(skb
)->frags
[i
- 1].size
+= copy
;
808 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
812 skb
->data_len
+= copy
;
813 skb
->truesize
+= copy
;
814 sk
->sk_wmem_queued
+= copy
;
815 sk_mem_charge(sk
, copy
);
816 skb
->ip_summed
= CHECKSUM_PARTIAL
;
817 tp
->write_seq
+= copy
;
818 TCP_SKB_CB(skb
)->end_seq
+= copy
;
819 skb_shinfo(skb
)->gso_segs
= 0;
822 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_PSH
;
826 if (!(psize
-= copy
))
829 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
832 if (forced_push(tp
)) {
833 tcp_mark_push(tp
, skb
);
834 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
835 } else if (skb
== tcp_send_head(sk
))
836 tcp_push_one(sk
, mss_now
);
840 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
843 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
845 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
848 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
853 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
860 return sk_stream_error(sk
, flags
, err
);
863 ssize_t
tcp_sendpage(struct socket
*sock
, struct page
*page
, int offset
,
864 size_t size
, int flags
)
867 struct sock
*sk
= sock
->sk
;
869 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
870 !(sk
->sk_route_caps
& NETIF_F_ALL_CSUM
))
871 return sock_no_sendpage(sock
, page
, offset
, size
, flags
);
875 res
= do_tcp_sendpages(sk
, &page
, offset
, size
, flags
);
881 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
882 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
884 static inline int select_size(struct sock
*sk
, int sg
)
886 struct tcp_sock
*tp
= tcp_sk(sk
);
887 int tmp
= tp
->mss_cache
;
893 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
895 if (tmp
>= pgbreak
&&
896 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
904 int tcp_sendmsg(struct kiocb
*iocb
, struct socket
*sock
, struct msghdr
*msg
,
907 struct sock
*sk
= sock
->sk
;
909 struct tcp_sock
*tp
= tcp_sk(sk
);
912 int mss_now
, size_goal
;
919 flags
= msg
->msg_flags
;
920 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
922 /* Wait for a connection to finish. */
923 if ((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
924 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
927 /* This should be in poll */
928 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
930 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
932 /* Ok commence sending. */
933 iovlen
= msg
->msg_iovlen
;
938 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
941 sg
= sk
->sk_route_caps
& NETIF_F_SG
;
943 while (--iovlen
>= 0) {
944 size_t seglen
= iov
->iov_len
;
945 unsigned char __user
*from
= iov
->iov_base
;
953 skb
= tcp_write_queue_tail(sk
);
954 if (tcp_send_head(sk
)) {
955 if (skb
->ip_summed
== CHECKSUM_NONE
)
957 copy
= max
- skb
->len
;
962 /* Allocate new segment. If the interface is SG,
963 * allocate skb fitting to single page.
965 if (!sk_stream_memory_free(sk
))
966 goto wait_for_sndbuf
;
968 skb
= sk_stream_alloc_skb(sk
,
972 goto wait_for_memory
;
975 * Check whether we can use HW checksum.
977 if (sk
->sk_route_caps
& NETIF_F_ALL_CSUM
)
978 skb
->ip_summed
= CHECKSUM_PARTIAL
;
985 /* Try to append data to the end of skb. */
989 /* Where to copy to? */
990 if (skb_tailroom(skb
) > 0) {
991 /* We have some space in skb head. Superb! */
992 if (copy
> skb_tailroom(skb
))
993 copy
= skb_tailroom(skb
);
994 if ((err
= skb_add_data(skb
, from
, copy
)) != 0)
998 int i
= skb_shinfo(skb
)->nr_frags
;
999 struct page
*page
= TCP_PAGE(sk
);
1000 int off
= TCP_OFF(sk
);
1002 if (skb_can_coalesce(skb
, i
, page
, off
) &&
1004 /* We can extend the last page
1007 } else if (i
== MAX_SKB_FRAGS
|| !sg
) {
1008 /* Need to add new fragment and cannot
1009 * do this because interface is non-SG,
1010 * or because all the page slots are
1012 tcp_mark_push(tp
, skb
);
1015 if (off
== PAGE_SIZE
) {
1017 TCP_PAGE(sk
) = page
= NULL
;
1023 if (copy
> PAGE_SIZE
- off
)
1024 copy
= PAGE_SIZE
- off
;
1026 if (!sk_wmem_schedule(sk
, copy
))
1027 goto wait_for_memory
;
1030 /* Allocate new cache page. */
1031 if (!(page
= sk_stream_alloc_page(sk
)))
1032 goto wait_for_memory
;
1035 /* Time to copy data. We are close to
1037 err
= skb_copy_to_page(sk
, from
, skb
, page
,
1040 /* If this page was new, give it to the
1041 * socket so it does not get leaked.
1043 if (!TCP_PAGE(sk
)) {
1044 TCP_PAGE(sk
) = page
;
1050 /* Update the skb. */
1052 skb_shinfo(skb
)->frags
[i
- 1].size
+=
1055 skb_fill_page_desc(skb
, i
, page
, off
, copy
);
1058 } else if (off
+ copy
< PAGE_SIZE
) {
1060 TCP_PAGE(sk
) = page
;
1064 TCP_OFF(sk
) = off
+ copy
;
1068 TCP_SKB_CB(skb
)->flags
&= ~TCPCB_FLAG_PSH
;
1070 tp
->write_seq
+= copy
;
1071 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1072 skb_shinfo(skb
)->gso_segs
= 0;
1076 if ((seglen
-= copy
) == 0 && iovlen
== 0)
1079 if (skb
->len
< max
|| (flags
& MSG_OOB
))
1082 if (forced_push(tp
)) {
1083 tcp_mark_push(tp
, skb
);
1084 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1085 } else if (skb
== tcp_send_head(sk
))
1086 tcp_push_one(sk
, mss_now
);
1090 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1093 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
1095 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
1098 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1104 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
1105 TCP_CHECK_TIMER(sk
);
1111 tcp_unlink_write_queue(skb
, sk
);
1112 /* It is the one place in all of TCP, except connection
1113 * reset, where we can be unlinking the send_head.
1115 tcp_check_send_head(sk
, skb
);
1116 sk_wmem_free_skb(sk
, skb
);
1123 err
= sk_stream_error(sk
, flags
, err
);
1124 TCP_CHECK_TIMER(sk
);
1130 * Handle reading urgent data. BSD has very simple semantics for
1131 * this, no blocking and very strange errors 8)
1134 static int tcp_recv_urg(struct sock
*sk
, struct msghdr
*msg
, int len
, int flags
)
1136 struct tcp_sock
*tp
= tcp_sk(sk
);
1138 /* No URG data to read. */
1139 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1140 tp
->urg_data
== TCP_URG_READ
)
1141 return -EINVAL
; /* Yes this is right ! */
1143 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1146 if (tp
->urg_data
& TCP_URG_VALID
) {
1148 char c
= tp
->urg_data
;
1150 if (!(flags
& MSG_PEEK
))
1151 tp
->urg_data
= TCP_URG_READ
;
1153 /* Read urgent data. */
1154 msg
->msg_flags
|= MSG_OOB
;
1157 if (!(flags
& MSG_TRUNC
))
1158 err
= memcpy_toiovec(msg
->msg_iov
, &c
, 1);
1161 msg
->msg_flags
|= MSG_TRUNC
;
1163 return err
? -EFAULT
: len
;
1166 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1169 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1170 * the available implementations agree in this case:
1171 * this call should never block, independent of the
1172 * blocking state of the socket.
1173 * Mike <pall@rz.uni-karlsruhe.de>
1178 /* Clean up the receive buffer for full frames taken by the user,
1179 * then send an ACK if necessary. COPIED is the number of bytes
1180 * tcp_recvmsg has given to the user so far, it speeds up the
1181 * calculation of whether or not we must ACK for the sake of
1184 void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1186 struct tcp_sock
*tp
= tcp_sk(sk
);
1187 int time_to_ack
= 0;
1190 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1192 WARN(skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
),
1193 KERN_INFO
"cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1194 tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
);
1197 if (inet_csk_ack_scheduled(sk
)) {
1198 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1199 /* Delayed ACKs frequently hit locked sockets during bulk
1201 if (icsk
->icsk_ack
.blocked
||
1202 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1203 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1205 * If this read emptied read buffer, we send ACK, if
1206 * connection is not bidirectional, user drained
1207 * receive buffer and there was a small segment
1211 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1212 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1213 !icsk
->icsk_ack
.pingpong
)) &&
1214 !atomic_read(&sk
->sk_rmem_alloc
)))
1218 /* We send an ACK if we can now advertise a non-zero window
1219 * which has been raised "significantly".
1221 * Even if window raised up to infinity, do not send window open ACK
1222 * in states, where we will not receive more. It is useless.
1224 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1225 __u32 rcv_window_now
= tcp_receive_window(tp
);
1227 /* Optimize, __tcp_select_window() is not cheap. */
1228 if (2*rcv_window_now
<= tp
->window_clamp
) {
1229 __u32 new_window
= __tcp_select_window(sk
);
1231 /* Send ACK now, if this read freed lots of space
1232 * in our buffer. Certainly, new_window is new window.
1233 * We can advertise it now, if it is not less than current one.
1234 * "Lots" means "at least twice" here.
1236 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1244 static void tcp_prequeue_process(struct sock
*sk
)
1246 struct sk_buff
*skb
;
1247 struct tcp_sock
*tp
= tcp_sk(sk
);
1249 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPPREQUEUED
);
1251 /* RX process wants to run with disabled BHs, though it is not
1254 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
1255 sk_backlog_rcv(sk
, skb
);
1258 /* Clear memory counter. */
1259 tp
->ucopy
.memory
= 0;
1262 #ifdef CONFIG_NET_DMA
1263 static void tcp_service_net_dma(struct sock
*sk
, bool wait
)
1265 dma_cookie_t done
, used
;
1266 dma_cookie_t last_issued
;
1267 struct tcp_sock
*tp
= tcp_sk(sk
);
1269 if (!tp
->ucopy
.dma_chan
)
1272 last_issued
= tp
->ucopy
.dma_cookie
;
1273 dma_async_memcpy_issue_pending(tp
->ucopy
.dma_chan
);
1276 if (dma_async_memcpy_complete(tp
->ucopy
.dma_chan
,
1278 &used
) == DMA_SUCCESS
) {
1279 /* Safe to free early-copied skbs now */
1280 __skb_queue_purge(&sk
->sk_async_wait_queue
);
1283 struct sk_buff
*skb
;
1284 while ((skb
= skb_peek(&sk
->sk_async_wait_queue
)) &&
1285 (dma_async_is_complete(skb
->dma_cookie
, done
,
1286 used
) == DMA_SUCCESS
)) {
1287 __skb_dequeue(&sk
->sk_async_wait_queue
);
1295 static inline struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1297 struct sk_buff
*skb
;
1300 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1301 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1302 if (tcp_hdr(skb
)->syn
)
1304 if (offset
< skb
->len
|| tcp_hdr(skb
)->fin
) {
1313 * This routine provides an alternative to tcp_recvmsg() for routines
1314 * that would like to handle copying from skbuffs directly in 'sendfile'
1317 * - It is assumed that the socket was locked by the caller.
1318 * - The routine does not block.
1319 * - At present, there is no support for reading OOB data
1320 * or for 'peeking' the socket using this routine
1321 * (although both would be easy to implement).
1323 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1324 sk_read_actor_t recv_actor
)
1326 struct sk_buff
*skb
;
1327 struct tcp_sock
*tp
= tcp_sk(sk
);
1328 u32 seq
= tp
->copied_seq
;
1332 if (sk
->sk_state
== TCP_LISTEN
)
1334 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1335 if (offset
< skb
->len
) {
1339 len
= skb
->len
- offset
;
1340 /* Stop reading if we hit a patch of urgent data */
1342 u32 urg_offset
= tp
->urg_seq
- seq
;
1343 if (urg_offset
< len
)
1348 used
= recv_actor(desc
, skb
, offset
, len
);
1353 } else if (used
<= len
) {
1359 * If recv_actor drops the lock (e.g. TCP splice
1360 * receive) the skb pointer might be invalid when
1361 * getting here: tcp_collapse might have deleted it
1362 * while aggregating skbs from the socket queue.
1364 skb
= tcp_recv_skb(sk
, seq
-1, &offset
);
1365 if (!skb
|| (offset
+1 != skb
->len
))
1368 if (tcp_hdr(skb
)->fin
) {
1369 sk_eat_skb(sk
, skb
, 0);
1373 sk_eat_skb(sk
, skb
, 0);
1376 tp
->copied_seq
= seq
;
1378 tp
->copied_seq
= seq
;
1380 tcp_rcv_space_adjust(sk
);
1382 /* Clean up data we have read: This will do ACK frames. */
1384 tcp_cleanup_rbuf(sk
, copied
);
1389 * This routine copies from a sock struct into the user buffer.
1391 * Technical note: in 2.3 we work on _locked_ socket, so that
1392 * tricks with *seq access order and skb->users are not required.
1393 * Probably, code can be easily improved even more.
1396 int tcp_recvmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1397 size_t len
, int nonblock
, int flags
, int *addr_len
)
1399 struct tcp_sock
*tp
= tcp_sk(sk
);
1405 int target
; /* Read at least this many bytes */
1407 struct task_struct
*user_recv
= NULL
;
1408 int copied_early
= 0;
1409 struct sk_buff
*skb
;
1414 TCP_CHECK_TIMER(sk
);
1417 if (sk
->sk_state
== TCP_LISTEN
)
1420 timeo
= sock_rcvtimeo(sk
, nonblock
);
1422 /* Urgent data needs to be handled specially. */
1423 if (flags
& MSG_OOB
)
1426 seq
= &tp
->copied_seq
;
1427 if (flags
& MSG_PEEK
) {
1428 peek_seq
= tp
->copied_seq
;
1432 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1434 #ifdef CONFIG_NET_DMA
1435 tp
->ucopy
.dma_chan
= NULL
;
1437 skb
= skb_peek_tail(&sk
->sk_receive_queue
);
1442 available
= TCP_SKB_CB(skb
)->seq
+ skb
->len
- (*seq
);
1443 if ((available
< target
) &&
1444 (len
> sysctl_tcp_dma_copybreak
) && !(flags
& MSG_PEEK
) &&
1445 !sysctl_tcp_low_latency
&&
1446 dma_find_channel(DMA_MEMCPY
)) {
1447 preempt_enable_no_resched();
1448 tp
->ucopy
.pinned_list
=
1449 dma_pin_iovec_pages(msg
->msg_iov
, len
);
1451 preempt_enable_no_resched();
1459 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1460 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1463 if (signal_pending(current
)) {
1464 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1469 /* Next get a buffer. */
1471 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1472 /* Now that we have two receive queues this
1475 if (WARN(before(*seq
, TCP_SKB_CB(skb
)->seq
),
1476 KERN_INFO
"recvmsg bug: copied %X "
1477 "seq %X rcvnxt %X fl %X\n", *seq
,
1478 TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
,
1482 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1483 if (tcp_hdr(skb
)->syn
)
1485 if (offset
< skb
->len
)
1487 if (tcp_hdr(skb
)->fin
)
1489 WARN(!(flags
& MSG_PEEK
), KERN_INFO
"recvmsg bug 2: "
1490 "copied %X seq %X rcvnxt %X fl %X\n",
1491 *seq
, TCP_SKB_CB(skb
)->seq
,
1492 tp
->rcv_nxt
, flags
);
1495 /* Well, if we have backlog, try to process it now yet. */
1497 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1502 sk
->sk_state
== TCP_CLOSE
||
1503 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1505 signal_pending(current
))
1508 if (sock_flag(sk
, SOCK_DONE
))
1512 copied
= sock_error(sk
);
1516 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1519 if (sk
->sk_state
== TCP_CLOSE
) {
1520 if (!sock_flag(sk
, SOCK_DONE
)) {
1521 /* This occurs when user tries to read
1522 * from never connected socket.
1535 if (signal_pending(current
)) {
1536 copied
= sock_intr_errno(timeo
);
1541 tcp_cleanup_rbuf(sk
, copied
);
1543 if (!sysctl_tcp_low_latency
&& tp
->ucopy
.task
== user_recv
) {
1544 /* Install new reader */
1545 if (!user_recv
&& !(flags
& (MSG_TRUNC
| MSG_PEEK
))) {
1546 user_recv
= current
;
1547 tp
->ucopy
.task
= user_recv
;
1548 tp
->ucopy
.iov
= msg
->msg_iov
;
1551 tp
->ucopy
.len
= len
;
1553 WARN_ON(tp
->copied_seq
!= tp
->rcv_nxt
&&
1554 !(flags
& (MSG_PEEK
| MSG_TRUNC
)));
1556 /* Ugly... If prequeue is not empty, we have to
1557 * process it before releasing socket, otherwise
1558 * order will be broken at second iteration.
1559 * More elegant solution is required!!!
1561 * Look: we have the following (pseudo)queues:
1563 * 1. packets in flight
1568 * Each queue can be processed only if the next ones
1569 * are empty. At this point we have empty receive_queue.
1570 * But prequeue _can_ be not empty after 2nd iteration,
1571 * when we jumped to start of loop because backlog
1572 * processing added something to receive_queue.
1573 * We cannot release_sock(), because backlog contains
1574 * packets arrived _after_ prequeued ones.
1576 * Shortly, algorithm is clear --- to process all
1577 * the queues in order. We could make it more directly,
1578 * requeueing packets from backlog to prequeue, if
1579 * is not empty. It is more elegant, but eats cycles,
1582 if (!skb_queue_empty(&tp
->ucopy
.prequeue
))
1585 /* __ Set realtime policy in scheduler __ */
1588 #ifdef CONFIG_NET_DMA
1589 if (tp
->ucopy
.dma_chan
)
1590 dma_async_memcpy_issue_pending(tp
->ucopy
.dma_chan
);
1592 if (copied
>= target
) {
1593 /* Do not sleep, just process backlog. */
1597 sk_wait_data(sk
, &timeo
);
1599 #ifdef CONFIG_NET_DMA
1600 tcp_service_net_dma(sk
, false); /* Don't block */
1601 tp
->ucopy
.wakeup
= 0;
1607 /* __ Restore normal policy in scheduler __ */
1609 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1610 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG
, chunk
);
1615 if (tp
->rcv_nxt
== tp
->copied_seq
&&
1616 !skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1618 tcp_prequeue_process(sk
);
1620 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1621 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1627 if ((flags
& MSG_PEEK
) &&
1628 (peek_seq
- copied
- urg_hole
!= tp
->copied_seq
)) {
1629 if (net_ratelimit())
1630 printk(KERN_DEBUG
"TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1631 current
->comm
, task_pid_nr(current
));
1632 peek_seq
= tp
->copied_seq
;
1637 /* Ok so how much can we use? */
1638 used
= skb
->len
- offset
;
1642 /* Do we have urgent data here? */
1644 u32 urg_offset
= tp
->urg_seq
- *seq
;
1645 if (urg_offset
< used
) {
1647 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1660 if (!(flags
& MSG_TRUNC
)) {
1661 #ifdef CONFIG_NET_DMA
1662 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
1663 tp
->ucopy
.dma_chan
= dma_find_channel(DMA_MEMCPY
);
1665 if (tp
->ucopy
.dma_chan
) {
1666 tp
->ucopy
.dma_cookie
= dma_skb_copy_datagram_iovec(
1667 tp
->ucopy
.dma_chan
, skb
, offset
,
1669 tp
->ucopy
.pinned_list
);
1671 if (tp
->ucopy
.dma_cookie
< 0) {
1673 printk(KERN_ALERT
"dma_cookie < 0\n");
1675 /* Exception. Bailout! */
1681 dma_async_memcpy_issue_pending(tp
->ucopy
.dma_chan
);
1683 if ((offset
+ used
) == skb
->len
)
1689 err
= skb_copy_datagram_iovec(skb
, offset
,
1690 msg
->msg_iov
, used
);
1692 /* Exception. Bailout! */
1704 tcp_rcv_space_adjust(sk
);
1707 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1709 tcp_fast_path_check(sk
);
1711 if (used
+ offset
< skb
->len
)
1714 if (tcp_hdr(skb
)->fin
)
1716 if (!(flags
& MSG_PEEK
)) {
1717 sk_eat_skb(sk
, skb
, copied_early
);
1723 /* Process the FIN. */
1725 if (!(flags
& MSG_PEEK
)) {
1726 sk_eat_skb(sk
, skb
, copied_early
);
1733 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1736 tp
->ucopy
.len
= copied
> 0 ? len
: 0;
1738 tcp_prequeue_process(sk
);
1740 if (copied
> 0 && (chunk
= len
- tp
->ucopy
.len
) != 0) {
1741 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1747 tp
->ucopy
.task
= NULL
;
1751 #ifdef CONFIG_NET_DMA
1752 tcp_service_net_dma(sk
, true); /* Wait for queue to drain */
1753 tp
->ucopy
.dma_chan
= NULL
;
1755 if (tp
->ucopy
.pinned_list
) {
1756 dma_unpin_iovec_pages(tp
->ucopy
.pinned_list
);
1757 tp
->ucopy
.pinned_list
= NULL
;
1761 /* According to UNIX98, msg_name/msg_namelen are ignored
1762 * on connected socket. I was just happy when found this 8) --ANK
1765 /* Clean up data we have read: This will do ACK frames. */
1766 tcp_cleanup_rbuf(sk
, copied
);
1768 TCP_CHECK_TIMER(sk
);
1773 TCP_CHECK_TIMER(sk
);
1778 err
= tcp_recv_urg(sk
, msg
, len
, flags
);
1782 void tcp_set_state(struct sock
*sk
, int state
)
1784 int oldstate
= sk
->sk_state
;
1787 case TCP_ESTABLISHED
:
1788 if (oldstate
!= TCP_ESTABLISHED
)
1789 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1793 if (oldstate
== TCP_CLOSE_WAIT
|| oldstate
== TCP_ESTABLISHED
)
1794 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ESTABRESETS
);
1796 sk
->sk_prot
->unhash(sk
);
1797 if (inet_csk(sk
)->icsk_bind_hash
&&
1798 !(sk
->sk_userlocks
& SOCK_BINDPORT_LOCK
))
1802 if (oldstate
== TCP_ESTABLISHED
)
1803 TCP_DEC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1806 /* Change state AFTER socket is unhashed to avoid closed
1807 * socket sitting in hash tables.
1809 sk
->sk_state
= state
;
1812 SOCK_DEBUG(sk
, "TCP sk=%p, State %s -> %s\n", sk
, statename
[oldstate
], statename
[state
]);
1815 EXPORT_SYMBOL_GPL(tcp_set_state
);
1818 * State processing on a close. This implements the state shift for
1819 * sending our FIN frame. Note that we only send a FIN for some
1820 * states. A shutdown() may have already sent the FIN, or we may be
1824 static const unsigned char new_state
[16] = {
1825 /* current state: new state: action: */
1826 /* (Invalid) */ TCP_CLOSE
,
1827 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1828 /* TCP_SYN_SENT */ TCP_CLOSE
,
1829 /* TCP_SYN_RECV */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1830 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1
,
1831 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2
,
1832 /* TCP_TIME_WAIT */ TCP_CLOSE
,
1833 /* TCP_CLOSE */ TCP_CLOSE
,
1834 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK
| TCP_ACTION_FIN
,
1835 /* TCP_LAST_ACK */ TCP_LAST_ACK
,
1836 /* TCP_LISTEN */ TCP_CLOSE
,
1837 /* TCP_CLOSING */ TCP_CLOSING
,
1840 static int tcp_close_state(struct sock
*sk
)
1842 int next
= (int)new_state
[sk
->sk_state
];
1843 int ns
= next
& TCP_STATE_MASK
;
1845 tcp_set_state(sk
, ns
);
1847 return next
& TCP_ACTION_FIN
;
1851 * Shutdown the sending side of a connection. Much like close except
1852 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1855 void tcp_shutdown(struct sock
*sk
, int how
)
1857 /* We need to grab some memory, and put together a FIN,
1858 * and then put it into the queue to be sent.
1859 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1861 if (!(how
& SEND_SHUTDOWN
))
1864 /* If we've already sent a FIN, or it's a closed state, skip this. */
1865 if ((1 << sk
->sk_state
) &
1866 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
1867 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
1868 /* Clear out any half completed packets. FIN if needed. */
1869 if (tcp_close_state(sk
))
1874 void tcp_close(struct sock
*sk
, long timeout
)
1876 struct sk_buff
*skb
;
1877 int data_was_unread
= 0;
1881 sk
->sk_shutdown
= SHUTDOWN_MASK
;
1883 if (sk
->sk_state
== TCP_LISTEN
) {
1884 tcp_set_state(sk
, TCP_CLOSE
);
1887 inet_csk_listen_stop(sk
);
1889 goto adjudge_to_death
;
1892 /* We need to flush the recv. buffs. We do this only on the
1893 * descriptor close, not protocol-sourced closes, because the
1894 * reader process may not have drained the data yet!
1896 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
1897 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
-
1899 data_was_unread
+= len
;
1905 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1906 * data was lost. To witness the awful effects of the old behavior of
1907 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1908 * GET in an FTP client, suspend the process, wait for the client to
1909 * advertise a zero window, then kill -9 the FTP client, wheee...
1910 * Note: timeout is always zero in such a case.
1912 if (data_was_unread
) {
1913 /* Unread data was tossed, zap the connection. */
1914 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONCLOSE
);
1915 tcp_set_state(sk
, TCP_CLOSE
);
1916 tcp_send_active_reset(sk
, sk
->sk_allocation
);
1917 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
1918 /* Check zero linger _after_ checking for unread data. */
1919 sk
->sk_prot
->disconnect(sk
, 0);
1920 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
1921 } else if (tcp_close_state(sk
)) {
1922 /* We FIN if the application ate all the data before
1923 * zapping the connection.
1926 /* RED-PEN. Formally speaking, we have broken TCP state
1927 * machine. State transitions:
1929 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1930 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1931 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1933 * are legal only when FIN has been sent (i.e. in window),
1934 * rather than queued out of window. Purists blame.
1936 * F.e. "RFC state" is ESTABLISHED,
1937 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1939 * The visible declinations are that sometimes
1940 * we enter time-wait state, when it is not required really
1941 * (harmless), do not send active resets, when they are
1942 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1943 * they look as CLOSING or LAST_ACK for Linux)
1944 * Probably, I missed some more holelets.
1950 sk_stream_wait_close(sk
, timeout
);
1953 state
= sk
->sk_state
;
1957 /* It is the last release_sock in its life. It will remove backlog. */
1961 /* Now socket is owned by kernel and we acquire BH lock
1962 to finish close. No need to check for user refs.
1966 WARN_ON(sock_owned_by_user(sk
));
1968 percpu_counter_inc(sk
->sk_prot
->orphan_count
);
1970 /* Have we already been destroyed by a softirq or backlog? */
1971 if (state
!= TCP_CLOSE
&& sk
->sk_state
== TCP_CLOSE
)
1974 /* This is a (useful) BSD violating of the RFC. There is a
1975 * problem with TCP as specified in that the other end could
1976 * keep a socket open forever with no application left this end.
1977 * We use a 3 minute timeout (about the same as BSD) then kill
1978 * our end. If they send after that then tough - BUT: long enough
1979 * that we won't make the old 4*rto = almost no time - whoops
1982 * Nope, it was not mistake. It is really desired behaviour
1983 * f.e. on http servers, when such sockets are useless, but
1984 * consume significant resources. Let's do it with special
1985 * linger2 option. --ANK
1988 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
1989 struct tcp_sock
*tp
= tcp_sk(sk
);
1990 if (tp
->linger2
< 0) {
1991 tcp_set_state(sk
, TCP_CLOSE
);
1992 tcp_send_active_reset(sk
, GFP_ATOMIC
);
1993 NET_INC_STATS_BH(sock_net(sk
),
1994 LINUX_MIB_TCPABORTONLINGER
);
1996 const int tmo
= tcp_fin_time(sk
);
1998 if (tmo
> TCP_TIMEWAIT_LEN
) {
1999 inet_csk_reset_keepalive_timer(sk
,
2000 tmo
- TCP_TIMEWAIT_LEN
);
2002 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
2007 if (sk
->sk_state
!= TCP_CLOSE
) {
2009 if (tcp_too_many_orphans(sk
, 0)) {
2010 if (net_ratelimit())
2011 printk(KERN_INFO
"TCP: too many of orphaned "
2013 tcp_set_state(sk
, TCP_CLOSE
);
2014 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2015 NET_INC_STATS_BH(sock_net(sk
),
2016 LINUX_MIB_TCPABORTONMEMORY
);
2020 if (sk
->sk_state
== TCP_CLOSE
)
2021 inet_csk_destroy_sock(sk
);
2022 /* Otherwise, socket is reprieved until protocol close. */
2030 /* These states need RST on ABORT according to RFC793 */
2032 static inline int tcp_need_reset(int state
)
2034 return (1 << state
) &
2035 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
2036 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
2039 int tcp_disconnect(struct sock
*sk
, int flags
)
2041 struct inet_sock
*inet
= inet_sk(sk
);
2042 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2043 struct tcp_sock
*tp
= tcp_sk(sk
);
2045 int old_state
= sk
->sk_state
;
2047 if (old_state
!= TCP_CLOSE
)
2048 tcp_set_state(sk
, TCP_CLOSE
);
2050 /* ABORT function of RFC793 */
2051 if (old_state
== TCP_LISTEN
) {
2052 inet_csk_listen_stop(sk
);
2053 } else if (tcp_need_reset(old_state
) ||
2054 (tp
->snd_nxt
!= tp
->write_seq
&&
2055 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
2056 /* The last check adjusts for discrepancy of Linux wrt. RFC
2059 tcp_send_active_reset(sk
, gfp_any());
2060 sk
->sk_err
= ECONNRESET
;
2061 } else if (old_state
== TCP_SYN_SENT
)
2062 sk
->sk_err
= ECONNRESET
;
2064 tcp_clear_xmit_timers(sk
);
2065 __skb_queue_purge(&sk
->sk_receive_queue
);
2066 tcp_write_queue_purge(sk
);
2067 __skb_queue_purge(&tp
->out_of_order_queue
);
2068 #ifdef CONFIG_NET_DMA
2069 __skb_queue_purge(&sk
->sk_async_wait_queue
);
2072 inet
->inet_dport
= 0;
2074 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
2075 inet_reset_saddr(sk
);
2077 sk
->sk_shutdown
= 0;
2078 sock_reset_flag(sk
, SOCK_DONE
);
2080 if ((tp
->write_seq
+= tp
->max_window
+ 2) == 0)
2082 icsk
->icsk_backoff
= 0;
2084 icsk
->icsk_probes_out
= 0;
2085 tp
->packets_out
= 0;
2086 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
2087 tp
->snd_cwnd_cnt
= 0;
2088 tp
->bytes_acked
= 0;
2089 tp
->window_clamp
= 0;
2090 tcp_set_ca_state(sk
, TCP_CA_Open
);
2091 tcp_clear_retrans(tp
);
2092 inet_csk_delack_init(sk
);
2093 tcp_init_send_head(sk
);
2094 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
2097 WARN_ON(inet
->inet_num
&& !icsk
->icsk_bind_hash
);
2099 sk
->sk_error_report(sk
);
2104 * Socket option code for TCP.
2106 static int do_tcp_setsockopt(struct sock
*sk
, int level
,
2107 int optname
, char __user
*optval
, unsigned int optlen
)
2109 struct tcp_sock
*tp
= tcp_sk(sk
);
2110 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2114 /* These are data/string values, all the others are ints */
2116 case TCP_CONGESTION
: {
2117 char name
[TCP_CA_NAME_MAX
];
2122 val
= strncpy_from_user(name
, optval
,
2123 min_t(long, TCP_CA_NAME_MAX
-1, optlen
));
2129 err
= tcp_set_congestion_control(sk
, name
);
2133 case TCP_COOKIE_TRANSACTIONS
: {
2134 struct tcp_cookie_transactions ctd
;
2135 struct tcp_cookie_values
*cvp
= NULL
;
2137 if (sizeof(ctd
) > optlen
)
2139 if (copy_from_user(&ctd
, optval
, sizeof(ctd
)))
2142 if (ctd
.tcpct_used
> sizeof(ctd
.tcpct_value
) ||
2143 ctd
.tcpct_s_data_desired
> TCP_MSS_DESIRED
)
2146 if (ctd
.tcpct_cookie_desired
== 0) {
2147 /* default to global value */
2148 } else if ((0x1 & ctd
.tcpct_cookie_desired
) ||
2149 ctd
.tcpct_cookie_desired
> TCP_COOKIE_MAX
||
2150 ctd
.tcpct_cookie_desired
< TCP_COOKIE_MIN
) {
2154 if (TCP_COOKIE_OUT_NEVER
& ctd
.tcpct_flags
) {
2155 /* Supercedes all other values */
2157 if (tp
->cookie_values
!= NULL
) {
2158 kref_put(&tp
->cookie_values
->kref
,
2159 tcp_cookie_values_release
);
2160 tp
->cookie_values
= NULL
;
2162 tp
->rx_opt
.cookie_in_always
= 0; /* false */
2163 tp
->rx_opt
.cookie_out_never
= 1; /* true */
2168 /* Allocate ancillary memory before locking.
2170 if (ctd
.tcpct_used
> 0 ||
2171 (tp
->cookie_values
== NULL
&&
2172 (sysctl_tcp_cookie_size
> 0 ||
2173 ctd
.tcpct_cookie_desired
> 0 ||
2174 ctd
.tcpct_s_data_desired
> 0))) {
2175 cvp
= kzalloc(sizeof(*cvp
) + ctd
.tcpct_used
,
2180 kref_init(&cvp
->kref
);
2183 tp
->rx_opt
.cookie_in_always
=
2184 (TCP_COOKIE_IN_ALWAYS
& ctd
.tcpct_flags
);
2185 tp
->rx_opt
.cookie_out_never
= 0; /* false */
2187 if (tp
->cookie_values
!= NULL
) {
2189 /* Changed values are recorded by a changed
2190 * pointer, ensuring the cookie will differ,
2191 * without separately hashing each value later.
2193 kref_put(&tp
->cookie_values
->kref
,
2194 tcp_cookie_values_release
);
2196 cvp
= tp
->cookie_values
;
2201 cvp
->cookie_desired
= ctd
.tcpct_cookie_desired
;
2203 if (ctd
.tcpct_used
> 0) {
2204 memcpy(cvp
->s_data_payload
, ctd
.tcpct_value
,
2206 cvp
->s_data_desired
= ctd
.tcpct_used
;
2207 cvp
->s_data_constant
= 1; /* true */
2209 /* No constant payload data. */
2210 cvp
->s_data_desired
= ctd
.tcpct_s_data_desired
;
2211 cvp
->s_data_constant
= 0; /* false */
2214 tp
->cookie_values
= cvp
;
2224 if (optlen
< sizeof(int))
2227 if (get_user(val
, (int __user
*)optval
))
2234 /* Values greater than interface MTU won't take effect. However
2235 * at the point when this call is done we typically don't yet
2236 * know which interface is going to be used */
2237 if (val
< TCP_MIN_MSS
|| val
> MAX_TCP_WINDOW
) {
2241 tp
->rx_opt
.user_mss
= val
;
2246 /* TCP_NODELAY is weaker than TCP_CORK, so that
2247 * this option on corked socket is remembered, but
2248 * it is not activated until cork is cleared.
2250 * However, when TCP_NODELAY is set we make
2251 * an explicit push, which overrides even TCP_CORK
2252 * for currently queued segments.
2254 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
2255 tcp_push_pending_frames(sk
);
2257 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
2261 case TCP_THIN_LINEAR_TIMEOUTS
:
2262 if (val
< 0 || val
> 1)
2268 case TCP_THIN_DUPACK
:
2269 if (val
< 0 || val
> 1)
2272 tp
->thin_dupack
= val
;
2276 /* When set indicates to always queue non-full frames.
2277 * Later the user clears this option and we transmit
2278 * any pending partial frames in the queue. This is
2279 * meant to be used alongside sendfile() to get properly
2280 * filled frames when the user (for example) must write
2281 * out headers with a write() call first and then use
2282 * sendfile to send out the data parts.
2284 * TCP_CORK can be set together with TCP_NODELAY and it is
2285 * stronger than TCP_NODELAY.
2288 tp
->nonagle
|= TCP_NAGLE_CORK
;
2290 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
2291 if (tp
->nonagle
&TCP_NAGLE_OFF
)
2292 tp
->nonagle
|= TCP_NAGLE_PUSH
;
2293 tcp_push_pending_frames(sk
);
2298 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
2301 tp
->keepalive_time
= val
* HZ
;
2302 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
2303 !((1 << sk
->sk_state
) &
2304 (TCPF_CLOSE
| TCPF_LISTEN
))) {
2305 __u32 elapsed
= tcp_time_stamp
- tp
->rcv_tstamp
;
2306 if (tp
->keepalive_time
> elapsed
)
2307 elapsed
= tp
->keepalive_time
- elapsed
;
2310 inet_csk_reset_keepalive_timer(sk
, elapsed
);
2315 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
2318 tp
->keepalive_intvl
= val
* HZ
;
2321 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
2324 tp
->keepalive_probes
= val
;
2327 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
2330 icsk
->icsk_syn_retries
= val
;
2336 else if (val
> sysctl_tcp_fin_timeout
/ HZ
)
2339 tp
->linger2
= val
* HZ
;
2342 case TCP_DEFER_ACCEPT
:
2343 /* Translate value in seconds to number of retransmits */
2344 icsk
->icsk_accept_queue
.rskq_defer_accept
=
2345 secs_to_retrans(val
, TCP_TIMEOUT_INIT
/ HZ
,
2349 case TCP_WINDOW_CLAMP
:
2351 if (sk
->sk_state
!= TCP_CLOSE
) {
2355 tp
->window_clamp
= 0;
2357 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
2358 SOCK_MIN_RCVBUF
/ 2 : val
;
2363 icsk
->icsk_ack
.pingpong
= 1;
2365 icsk
->icsk_ack
.pingpong
= 0;
2366 if ((1 << sk
->sk_state
) &
2367 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
2368 inet_csk_ack_scheduled(sk
)) {
2369 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
2370 tcp_cleanup_rbuf(sk
, 1);
2372 icsk
->icsk_ack
.pingpong
= 1;
2377 #ifdef CONFIG_TCP_MD5SIG
2379 /* Read the IP->Key mappings from userspace */
2380 err
= tp
->af_specific
->md5_parse(sk
, optval
, optlen
);
2393 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2394 unsigned int optlen
)
2396 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2398 if (level
!= SOL_TCP
)
2399 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
2401 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2404 #ifdef CONFIG_COMPAT
2405 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
2406 char __user
*optval
, unsigned int optlen
)
2408 if (level
!= SOL_TCP
)
2409 return inet_csk_compat_setsockopt(sk
, level
, optname
,
2411 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2414 EXPORT_SYMBOL(compat_tcp_setsockopt
);
2417 /* Return information about state of tcp endpoint in API format. */
2418 void tcp_get_info(struct sock
*sk
, struct tcp_info
*info
)
2420 struct tcp_sock
*tp
= tcp_sk(sk
);
2421 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2422 u32 now
= tcp_time_stamp
;
2424 memset(info
, 0, sizeof(*info
));
2426 info
->tcpi_state
= sk
->sk_state
;
2427 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
2428 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
2429 info
->tcpi_probes
= icsk
->icsk_probes_out
;
2430 info
->tcpi_backoff
= icsk
->icsk_backoff
;
2432 if (tp
->rx_opt
.tstamp_ok
)
2433 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
2434 if (tcp_is_sack(tp
))
2435 info
->tcpi_options
|= TCPI_OPT_SACK
;
2436 if (tp
->rx_opt
.wscale_ok
) {
2437 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
2438 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
2439 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
2442 if (tp
->ecn_flags
&TCP_ECN_OK
)
2443 info
->tcpi_options
|= TCPI_OPT_ECN
;
2445 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
2446 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
2447 info
->tcpi_snd_mss
= tp
->mss_cache
;
2448 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
2450 if (sk
->sk_state
== TCP_LISTEN
) {
2451 info
->tcpi_unacked
= sk
->sk_ack_backlog
;
2452 info
->tcpi_sacked
= sk
->sk_max_ack_backlog
;
2454 info
->tcpi_unacked
= tp
->packets_out
;
2455 info
->tcpi_sacked
= tp
->sacked_out
;
2457 info
->tcpi_lost
= tp
->lost_out
;
2458 info
->tcpi_retrans
= tp
->retrans_out
;
2459 info
->tcpi_fackets
= tp
->fackets_out
;
2461 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
2462 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
2463 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
2465 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
2466 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
2467 info
->tcpi_rtt
= jiffies_to_usecs(tp
->srtt
)>>3;
2468 info
->tcpi_rttvar
= jiffies_to_usecs(tp
->mdev
)>>2;
2469 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
2470 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
2471 info
->tcpi_advmss
= tp
->advmss
;
2472 info
->tcpi_reordering
= tp
->reordering
;
2474 info
->tcpi_rcv_rtt
= jiffies_to_usecs(tp
->rcv_rtt_est
.rtt
)>>3;
2475 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
2477 info
->tcpi_total_retrans
= tp
->total_retrans
;
2480 EXPORT_SYMBOL_GPL(tcp_get_info
);
2482 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
2483 int optname
, char __user
*optval
, int __user
*optlen
)
2485 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2486 struct tcp_sock
*tp
= tcp_sk(sk
);
2489 if (get_user(len
, optlen
))
2492 len
= min_t(unsigned int, len
, sizeof(int));
2499 val
= tp
->mss_cache
;
2500 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
2501 val
= tp
->rx_opt
.user_mss
;
2504 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
2507 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
2510 val
= keepalive_time_when(tp
) / HZ
;
2513 val
= keepalive_intvl_when(tp
) / HZ
;
2516 val
= keepalive_probes(tp
);
2519 val
= icsk
->icsk_syn_retries
? : sysctl_tcp_syn_retries
;
2524 val
= (val
? : sysctl_tcp_fin_timeout
) / HZ
;
2526 case TCP_DEFER_ACCEPT
:
2527 val
= retrans_to_secs(icsk
->icsk_accept_queue
.rskq_defer_accept
,
2528 TCP_TIMEOUT_INIT
/ HZ
, TCP_RTO_MAX
/ HZ
);
2530 case TCP_WINDOW_CLAMP
:
2531 val
= tp
->window_clamp
;
2534 struct tcp_info info
;
2536 if (get_user(len
, optlen
))
2539 tcp_get_info(sk
, &info
);
2541 len
= min_t(unsigned int, len
, sizeof(info
));
2542 if (put_user(len
, optlen
))
2544 if (copy_to_user(optval
, &info
, len
))
2549 val
= !icsk
->icsk_ack
.pingpong
;
2552 case TCP_CONGESTION
:
2553 if (get_user(len
, optlen
))
2555 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
2556 if (put_user(len
, optlen
))
2558 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
2562 case TCP_COOKIE_TRANSACTIONS
: {
2563 struct tcp_cookie_transactions ctd
;
2564 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
2566 if (get_user(len
, optlen
))
2568 if (len
< sizeof(ctd
))
2571 memset(&ctd
, 0, sizeof(ctd
));
2572 ctd
.tcpct_flags
= (tp
->rx_opt
.cookie_in_always
?
2573 TCP_COOKIE_IN_ALWAYS
: 0)
2574 | (tp
->rx_opt
.cookie_out_never
?
2575 TCP_COOKIE_OUT_NEVER
: 0);
2578 ctd
.tcpct_flags
|= (cvp
->s_data_in
?
2580 | (cvp
->s_data_out
?
2581 TCP_S_DATA_OUT
: 0);
2583 ctd
.tcpct_cookie_desired
= cvp
->cookie_desired
;
2584 ctd
.tcpct_s_data_desired
= cvp
->s_data_desired
;
2586 memcpy(&ctd
.tcpct_value
[0], &cvp
->cookie_pair
[0],
2587 cvp
->cookie_pair_size
);
2588 ctd
.tcpct_used
= cvp
->cookie_pair_size
;
2591 if (put_user(sizeof(ctd
), optlen
))
2593 if (copy_to_user(optval
, &ctd
, sizeof(ctd
)))
2598 return -ENOPROTOOPT
;
2601 if (put_user(len
, optlen
))
2603 if (copy_to_user(optval
, &val
, len
))
2608 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2611 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2613 if (level
!= SOL_TCP
)
2614 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
2616 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2619 #ifdef CONFIG_COMPAT
2620 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
2621 char __user
*optval
, int __user
*optlen
)
2623 if (level
!= SOL_TCP
)
2624 return inet_csk_compat_getsockopt(sk
, level
, optname
,
2626 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2629 EXPORT_SYMBOL(compat_tcp_getsockopt
);
2632 struct sk_buff
*tcp_tso_segment(struct sk_buff
*skb
, int features
)
2634 struct sk_buff
*segs
= ERR_PTR(-EINVAL
);
2639 unsigned int oldlen
;
2642 if (!pskb_may_pull(skb
, sizeof(*th
)))
2646 thlen
= th
->doff
* 4;
2647 if (thlen
< sizeof(*th
))
2650 if (!pskb_may_pull(skb
, thlen
))
2653 oldlen
= (u16
)~skb
->len
;
2654 __skb_pull(skb
, thlen
);
2656 mss
= skb_shinfo(skb
)->gso_size
;
2657 if (unlikely(skb
->len
<= mss
))
2660 if (skb_gso_ok(skb
, features
| NETIF_F_GSO_ROBUST
)) {
2661 /* Packet is from an untrusted source, reset gso_segs. */
2662 int type
= skb_shinfo(skb
)->gso_type
;
2670 !(type
& (SKB_GSO_TCPV4
| SKB_GSO_TCPV6
))))
2673 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss
);
2679 segs
= skb_segment(skb
, features
);
2683 delta
= htonl(oldlen
+ (thlen
+ mss
));
2687 seq
= ntohl(th
->seq
);
2690 th
->fin
= th
->psh
= 0;
2692 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
2693 (__force u32
)delta
));
2694 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2696 csum_fold(csum_partial(skb_transport_header(skb
),
2703 th
->seq
= htonl(seq
);
2705 } while (skb
->next
);
2707 delta
= htonl(oldlen
+ (skb
->tail
- skb
->transport_header
) +
2709 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
2710 (__force u32
)delta
));
2711 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2712 th
->check
= csum_fold(csum_partial(skb_transport_header(skb
),
2718 EXPORT_SYMBOL(tcp_tso_segment
);
2720 struct sk_buff
**tcp_gro_receive(struct sk_buff
**head
, struct sk_buff
*skb
)
2722 struct sk_buff
**pp
= NULL
;
2729 unsigned int mss
= 1;
2735 off
= skb_gro_offset(skb
);
2736 hlen
= off
+ sizeof(*th
);
2737 th
= skb_gro_header_fast(skb
, off
);
2738 if (skb_gro_header_hard(skb
, hlen
)) {
2739 th
= skb_gro_header_slow(skb
, hlen
, off
);
2744 thlen
= th
->doff
* 4;
2745 if (thlen
< sizeof(*th
))
2749 if (skb_gro_header_hard(skb
, hlen
)) {
2750 th
= skb_gro_header_slow(skb
, hlen
, off
);
2755 skb_gro_pull(skb
, thlen
);
2757 len
= skb_gro_len(skb
);
2758 flags
= tcp_flag_word(th
);
2760 for (; (p
= *head
); head
= &p
->next
) {
2761 if (!NAPI_GRO_CB(p
)->same_flow
)
2766 if (*(u32
*)&th
->source
^ *(u32
*)&th2
->source
) {
2767 NAPI_GRO_CB(p
)->same_flow
= 0;
2774 goto out_check_final
;
2777 flush
= NAPI_GRO_CB(p
)->flush
;
2778 flush
|= flags
& TCP_FLAG_CWR
;
2779 flush
|= (flags
^ tcp_flag_word(th2
)) &
2780 ~(TCP_FLAG_CWR
| TCP_FLAG_FIN
| TCP_FLAG_PSH
);
2781 flush
|= th
->ack_seq
^ th2
->ack_seq
;
2782 for (i
= sizeof(*th
); i
< thlen
; i
+= 4)
2783 flush
|= *(u32
*)((u8
*)th
+ i
) ^
2784 *(u32
*)((u8
*)th2
+ i
);
2786 mss
= skb_shinfo(p
)->gso_size
;
2788 flush
|= (len
- 1) >= mss
;
2789 flush
|= (ntohl(th2
->seq
) + skb_gro_len(p
)) ^ ntohl(th
->seq
);
2791 if (flush
|| skb_gro_receive(head
, skb
)) {
2793 goto out_check_final
;
2798 tcp_flag_word(th2
) |= flags
& (TCP_FLAG_FIN
| TCP_FLAG_PSH
);
2802 flush
|= flags
& (TCP_FLAG_URG
| TCP_FLAG_PSH
| TCP_FLAG_RST
|
2803 TCP_FLAG_SYN
| TCP_FLAG_FIN
);
2805 if (p
&& (!NAPI_GRO_CB(skb
)->same_flow
|| flush
))
2809 NAPI_GRO_CB(skb
)->flush
|= flush
;
2813 EXPORT_SYMBOL(tcp_gro_receive
);
2815 int tcp_gro_complete(struct sk_buff
*skb
)
2817 struct tcphdr
*th
= tcp_hdr(skb
);
2819 skb
->csum_start
= skb_transport_header(skb
) - skb
->head
;
2820 skb
->csum_offset
= offsetof(struct tcphdr
, check
);
2821 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2823 skb_shinfo(skb
)->gso_segs
= NAPI_GRO_CB(skb
)->count
;
2826 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
2830 EXPORT_SYMBOL(tcp_gro_complete
);
2832 #ifdef CONFIG_TCP_MD5SIG
2833 static unsigned long tcp_md5sig_users
;
2834 static struct tcp_md5sig_pool
* __percpu
*tcp_md5sig_pool
;
2835 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock
);
2837 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool
* __percpu
*pool
)
2840 for_each_possible_cpu(cpu
) {
2841 struct tcp_md5sig_pool
*p
= *per_cpu_ptr(pool
, cpu
);
2843 if (p
->md5_desc
.tfm
)
2844 crypto_free_hash(p
->md5_desc
.tfm
);
2851 void tcp_free_md5sig_pool(void)
2853 struct tcp_md5sig_pool
* __percpu
*pool
= NULL
;
2855 spin_lock_bh(&tcp_md5sig_pool_lock
);
2856 if (--tcp_md5sig_users
== 0) {
2857 pool
= tcp_md5sig_pool
;
2858 tcp_md5sig_pool
= NULL
;
2860 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2862 __tcp_free_md5sig_pool(pool
);
2865 EXPORT_SYMBOL(tcp_free_md5sig_pool
);
2867 static struct tcp_md5sig_pool
* __percpu
*
2868 __tcp_alloc_md5sig_pool(struct sock
*sk
)
2871 struct tcp_md5sig_pool
* __percpu
*pool
;
2873 pool
= alloc_percpu(struct tcp_md5sig_pool
*);
2877 for_each_possible_cpu(cpu
) {
2878 struct tcp_md5sig_pool
*p
;
2879 struct crypto_hash
*hash
;
2881 p
= kzalloc(sizeof(*p
), sk
->sk_allocation
);
2884 *per_cpu_ptr(pool
, cpu
) = p
;
2886 hash
= crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC
);
2887 if (!hash
|| IS_ERR(hash
))
2890 p
->md5_desc
.tfm
= hash
;
2894 __tcp_free_md5sig_pool(pool
);
2898 struct tcp_md5sig_pool
* __percpu
*tcp_alloc_md5sig_pool(struct sock
*sk
)
2900 struct tcp_md5sig_pool
* __percpu
*pool
;
2904 spin_lock_bh(&tcp_md5sig_pool_lock
);
2905 pool
= tcp_md5sig_pool
;
2906 if (tcp_md5sig_users
++ == 0) {
2908 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2911 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2915 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2918 /* we cannot hold spinlock here because this may sleep. */
2919 struct tcp_md5sig_pool
* __percpu
*p
;
2921 p
= __tcp_alloc_md5sig_pool(sk
);
2922 spin_lock_bh(&tcp_md5sig_pool_lock
);
2925 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2928 pool
= tcp_md5sig_pool
;
2930 /* oops, it has already been assigned. */
2931 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2932 __tcp_free_md5sig_pool(p
);
2934 tcp_md5sig_pool
= pool
= p
;
2935 spin_unlock_bh(&tcp_md5sig_pool_lock
);
2941 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
2945 * tcp_get_md5sig_pool - get md5sig_pool for this user
2947 * We use percpu structure, so if we succeed, we exit with preemption
2948 * and BH disabled, to make sure another thread or softirq handling
2949 * wont try to get same context.
2951 struct tcp_md5sig_pool
*tcp_get_md5sig_pool(void)
2953 struct tcp_md5sig_pool
* __percpu
*p
;
2957 spin_lock(&tcp_md5sig_pool_lock
);
2958 p
= tcp_md5sig_pool
;
2961 spin_unlock(&tcp_md5sig_pool_lock
);
2964 return *per_cpu_ptr(p
, smp_processor_id());
2969 EXPORT_SYMBOL(tcp_get_md5sig_pool
);
2971 void tcp_put_md5sig_pool(void)
2974 tcp_free_md5sig_pool();
2976 EXPORT_SYMBOL(tcp_put_md5sig_pool
);
2978 int tcp_md5_hash_header(struct tcp_md5sig_pool
*hp
,
2981 struct scatterlist sg
;
2984 __sum16 old_checksum
= th
->check
;
2986 /* options aren't included in the hash */
2987 sg_init_one(&sg
, th
, sizeof(struct tcphdr
));
2988 err
= crypto_hash_update(&hp
->md5_desc
, &sg
, sizeof(struct tcphdr
));
2989 th
->check
= old_checksum
;
2993 EXPORT_SYMBOL(tcp_md5_hash_header
);
2995 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool
*hp
,
2996 struct sk_buff
*skb
, unsigned header_len
)
2998 struct scatterlist sg
;
2999 const struct tcphdr
*tp
= tcp_hdr(skb
);
3000 struct hash_desc
*desc
= &hp
->md5_desc
;
3002 const unsigned head_data_len
= skb_headlen(skb
) > header_len
?
3003 skb_headlen(skb
) - header_len
: 0;
3004 const struct skb_shared_info
*shi
= skb_shinfo(skb
);
3006 sg_init_table(&sg
, 1);
3008 sg_set_buf(&sg
, ((u8
*) tp
) + header_len
, head_data_len
);
3009 if (crypto_hash_update(desc
, &sg
, head_data_len
))
3012 for (i
= 0; i
< shi
->nr_frags
; ++i
) {
3013 const struct skb_frag_struct
*f
= &shi
->frags
[i
];
3014 sg_set_page(&sg
, f
->page
, f
->size
, f
->page_offset
);
3015 if (crypto_hash_update(desc
, &sg
, f
->size
))
3022 EXPORT_SYMBOL(tcp_md5_hash_skb_data
);
3024 int tcp_md5_hash_key(struct tcp_md5sig_pool
*hp
, struct tcp_md5sig_key
*key
)
3026 struct scatterlist sg
;
3028 sg_init_one(&sg
, key
->key
, key
->keylen
);
3029 return crypto_hash_update(&hp
->md5_desc
, &sg
, key
->keylen
);
3032 EXPORT_SYMBOL(tcp_md5_hash_key
);
3037 * Each Responder maintains up to two secret values concurrently for
3038 * efficient secret rollover. Each secret value has 4 states:
3040 * Generating. (tcp_secret_generating != tcp_secret_primary)
3041 * Generates new Responder-Cookies, but not yet used for primary
3042 * verification. This is a short-term state, typically lasting only
3043 * one round trip time (RTT).
3045 * Primary. (tcp_secret_generating == tcp_secret_primary)
3046 * Used both for generation and primary verification.
3048 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3049 * Used for verification, until the first failure that can be
3050 * verified by the newer Generating secret. At that time, this
3051 * cookie's state is changed to Secondary, and the Generating
3052 * cookie's state is changed to Primary. This is a short-term state,
3053 * typically lasting only one round trip time (RTT).
3055 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3056 * Used for secondary verification, after primary verification
3057 * failures. This state lasts no more than twice the Maximum Segment
3058 * Lifetime (2MSL). Then, the secret is discarded.
3060 struct tcp_cookie_secret
{
3061 /* The secret is divided into two parts. The digest part is the
3062 * equivalent of previously hashing a secret and saving the state,
3063 * and serves as an initialization vector (IV). The message part
3064 * serves as the trailing secret.
3066 u32 secrets
[COOKIE_WORKSPACE_WORDS
];
3067 unsigned long expires
;
3070 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3071 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3072 #define TCP_SECRET_LIFE (HZ * 600)
3074 static struct tcp_cookie_secret tcp_secret_one
;
3075 static struct tcp_cookie_secret tcp_secret_two
;
3077 /* Essentially a circular list, without dynamic allocation. */
3078 static struct tcp_cookie_secret
*tcp_secret_generating
;
3079 static struct tcp_cookie_secret
*tcp_secret_primary
;
3080 static struct tcp_cookie_secret
*tcp_secret_retiring
;
3081 static struct tcp_cookie_secret
*tcp_secret_secondary
;
3083 static DEFINE_SPINLOCK(tcp_secret_locker
);
3085 /* Select a pseudo-random word in the cookie workspace.
3087 static inline u32
tcp_cookie_work(const u32
*ws
, const int n
)
3089 return ws
[COOKIE_DIGEST_WORDS
+ ((COOKIE_MESSAGE_WORDS
-1) & ws
[n
])];
3092 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3093 * Called in softirq context.
3094 * Returns: 0 for success.
3096 int tcp_cookie_generator(u32
*bakery
)
3098 unsigned long jiffy
= jiffies
;
3100 if (unlikely(time_after_eq(jiffy
, tcp_secret_generating
->expires
))) {
3101 spin_lock_bh(&tcp_secret_locker
);
3102 if (!time_after_eq(jiffy
, tcp_secret_generating
->expires
)) {
3103 /* refreshed by another */
3105 &tcp_secret_generating
->secrets
[0],
3106 COOKIE_WORKSPACE_WORDS
);
3108 /* still needs refreshing */
3109 get_random_bytes(bakery
, COOKIE_WORKSPACE_WORDS
);
3111 /* The first time, paranoia assumes that the
3112 * randomization function isn't as strong. But,
3113 * this secret initialization is delayed until
3114 * the last possible moment (packet arrival).
3115 * Although that time is observable, it is
3116 * unpredictably variable. Mash in the most
3117 * volatile clock bits available, and expire the
3118 * secret extra quickly.
3120 if (unlikely(tcp_secret_primary
->expires
==
3121 tcp_secret_secondary
->expires
)) {
3124 getnstimeofday(&tv
);
3125 bakery
[COOKIE_DIGEST_WORDS
+0] ^=
3128 tcp_secret_secondary
->expires
= jiffy
3130 + (0x0f & tcp_cookie_work(bakery
, 0));
3132 tcp_secret_secondary
->expires
= jiffy
3134 + (0xff & tcp_cookie_work(bakery
, 1));
3135 tcp_secret_primary
->expires
= jiffy
3137 + (0x1f & tcp_cookie_work(bakery
, 2));
3139 memcpy(&tcp_secret_secondary
->secrets
[0],
3140 bakery
, COOKIE_WORKSPACE_WORDS
);
3142 rcu_assign_pointer(tcp_secret_generating
,
3143 tcp_secret_secondary
);
3144 rcu_assign_pointer(tcp_secret_retiring
,
3145 tcp_secret_primary
);
3147 * Neither call_rcu() nor synchronize_rcu() needed.
3148 * Retiring data is not freed. It is replaced after
3149 * further (locked) pointer updates, and a quiet time
3150 * (minimum 1MSL, maximum LIFE - 2MSL).
3153 spin_unlock_bh(&tcp_secret_locker
);
3157 &rcu_dereference(tcp_secret_generating
)->secrets
[0],
3158 COOKIE_WORKSPACE_WORDS
);
3159 rcu_read_unlock_bh();
3163 EXPORT_SYMBOL(tcp_cookie_generator
);
3165 void tcp_done(struct sock
*sk
)
3167 if (sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
3168 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_ATTEMPTFAILS
);
3170 tcp_set_state(sk
, TCP_CLOSE
);
3171 tcp_clear_xmit_timers(sk
);
3173 sk
->sk_shutdown
= SHUTDOWN_MASK
;
3175 if (!sock_flag(sk
, SOCK_DEAD
))
3176 sk
->sk_state_change(sk
);
3178 inet_csk_destroy_sock(sk
);
3180 EXPORT_SYMBOL_GPL(tcp_done
);
3182 extern struct tcp_congestion_ops tcp_reno
;
3184 static __initdata
unsigned long thash_entries
;
3185 static int __init
set_thash_entries(char *str
)
3189 thash_entries
= simple_strtoul(str
, &str
, 0);
3192 __setup("thash_entries=", set_thash_entries
);
3194 void __init
tcp_init(void)
3196 struct sk_buff
*skb
= NULL
;
3197 unsigned long nr_pages
, limit
;
3198 int i
, max_share
, cnt
;
3199 unsigned long jiffy
= jiffies
;
3201 BUILD_BUG_ON(sizeof(struct tcp_skb_cb
) > sizeof(skb
->cb
));
3203 percpu_counter_init(&tcp_sockets_allocated
, 0);
3204 percpu_counter_init(&tcp_orphan_count
, 0);
3205 tcp_hashinfo
.bind_bucket_cachep
=
3206 kmem_cache_create("tcp_bind_bucket",
3207 sizeof(struct inet_bind_bucket
), 0,
3208 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
);
3210 /* Size and allocate the main established and bind bucket
3213 * The methodology is similar to that of the buffer cache.
3215 tcp_hashinfo
.ehash
=
3216 alloc_large_system_hash("TCP established",
3217 sizeof(struct inet_ehash_bucket
),
3219 (totalram_pages
>= 128 * 1024) ?
3223 &tcp_hashinfo
.ehash_mask
,
3224 thash_entries
? 0 : 512 * 1024);
3225 for (i
= 0; i
<= tcp_hashinfo
.ehash_mask
; i
++) {
3226 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].chain
, i
);
3227 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].twchain
, i
);
3229 if (inet_ehash_locks_alloc(&tcp_hashinfo
))
3230 panic("TCP: failed to alloc ehash_locks");
3231 tcp_hashinfo
.bhash
=
3232 alloc_large_system_hash("TCP bind",
3233 sizeof(struct inet_bind_hashbucket
),
3234 tcp_hashinfo
.ehash_mask
+ 1,
3235 (totalram_pages
>= 128 * 1024) ?
3238 &tcp_hashinfo
.bhash_size
,
3241 tcp_hashinfo
.bhash_size
= 1 << tcp_hashinfo
.bhash_size
;
3242 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
3243 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
3244 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
3248 cnt
= tcp_hashinfo
.ehash_mask
+ 1;
3250 tcp_death_row
.sysctl_max_tw_buckets
= cnt
/ 2;
3251 sysctl_tcp_max_orphans
= cnt
/ 2;
3252 sysctl_max_syn_backlog
= max(128, cnt
/ 256);
3254 /* Set the pressure threshold to be a fraction of global memory that
3255 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
3256 * memory, with a floor of 128 pages.
3258 nr_pages
= totalram_pages
- totalhigh_pages
;
3259 limit
= min(nr_pages
, 1UL<<(28-PAGE_SHIFT
)) >> (20-PAGE_SHIFT
);
3260 limit
= (limit
* (nr_pages
>> (20-PAGE_SHIFT
))) >> (PAGE_SHIFT
-11);
3261 limit
= max(limit
, 128UL);
3262 sysctl_tcp_mem
[0] = limit
/ 4 * 3;
3263 sysctl_tcp_mem
[1] = limit
;
3264 sysctl_tcp_mem
[2] = sysctl_tcp_mem
[0] * 2;
3266 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3267 limit
= ((unsigned long)sysctl_tcp_mem
[1]) << (PAGE_SHIFT
- 7);
3268 max_share
= min(4UL*1024*1024, limit
);
3270 sysctl_tcp_wmem
[0] = SK_MEM_QUANTUM
;
3271 sysctl_tcp_wmem
[1] = 16*1024;
3272 sysctl_tcp_wmem
[2] = max(64*1024, max_share
);
3274 sysctl_tcp_rmem
[0] = SK_MEM_QUANTUM
;
3275 sysctl_tcp_rmem
[1] = 87380;
3276 sysctl_tcp_rmem
[2] = max(87380, max_share
);
3278 printk(KERN_INFO
"TCP: Hash tables configured "
3279 "(established %u bind %u)\n",
3280 tcp_hashinfo
.ehash_mask
+ 1, tcp_hashinfo
.bhash_size
);
3282 tcp_register_congestion_control(&tcp_reno
);
3284 memset(&tcp_secret_one
.secrets
[0], 0, sizeof(tcp_secret_one
.secrets
));
3285 memset(&tcp_secret_two
.secrets
[0], 0, sizeof(tcp_secret_two
.secrets
));
3286 tcp_secret_one
.expires
= jiffy
; /* past due */
3287 tcp_secret_two
.expires
= jiffy
; /* past due */
3288 tcp_secret_generating
= &tcp_secret_one
;
3289 tcp_secret_primary
= &tcp_secret_one
;
3290 tcp_secret_retiring
= &tcp_secret_two
;
3291 tcp_secret_secondary
= &tcp_secret_two
;
3294 EXPORT_SYMBOL(tcp_close
);
3295 EXPORT_SYMBOL(tcp_disconnect
);
3296 EXPORT_SYMBOL(tcp_getsockopt
);
3297 EXPORT_SYMBOL(tcp_ioctl
);
3298 EXPORT_SYMBOL(tcp_poll
);
3299 EXPORT_SYMBOL(tcp_read_sock
);
3300 EXPORT_SYMBOL(tcp_recvmsg
);
3301 EXPORT_SYMBOL(tcp_sendmsg
);
3302 EXPORT_SYMBOL(tcp_splice_read
);
3303 EXPORT_SYMBOL(tcp_sendpage
);
3304 EXPORT_SYMBOL(tcp_setsockopt
);
3305 EXPORT_SYMBOL(tcp_shutdown
);