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 #define pr_fmt(fmt) "TCP: " fmt
250 #include <linux/kernel.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/init.h>
256 #include <linux/fs.h>
257 #include <linux/skbuff.h>
258 #include <linux/scatterlist.h>
259 #include <linux/splice.h>
260 #include <linux/net.h>
261 #include <linux/socket.h>
262 #include <linux/random.h>
263 #include <linux/bootmem.h>
264 #include <linux/highmem.h>
265 #include <linux/swap.h>
266 #include <linux/cache.h>
267 #include <linux/err.h>
268 #include <linux/crypto.h>
269 #include <linux/time.h>
270 #include <linux/slab.h>
271 #include <linux/uid_stat.h>
273 #include <net/icmp.h>
274 #include <net/inet_common.h>
276 #include <net/xfrm.h>
278 #include <net/ip6_route.h>
279 #include <net/ipv6.h>
280 #include <net/transp_v6.h>
281 #include <net/netdma.h>
282 #include <net/sock.h>
284 #include <asm/uaccess.h>
285 #include <asm/ioctls.h>
287 int sysctl_tcp_fin_timeout __read_mostly
= TCP_FIN_TIMEOUT
;
289 int sysctl_tcp_min_tso_segs __read_mostly
= 2;
291 struct percpu_counter tcp_orphan_count
;
292 EXPORT_SYMBOL_GPL(tcp_orphan_count
);
294 int sysctl_tcp_wmem
[3] __read_mostly
;
295 int sysctl_tcp_rmem
[3] __read_mostly
;
297 EXPORT_SYMBOL(sysctl_tcp_rmem
);
298 EXPORT_SYMBOL(sysctl_tcp_wmem
);
300 atomic_long_t tcp_memory_allocated
; /* Current allocated memory. */
301 EXPORT_SYMBOL(tcp_memory_allocated
);
304 * Current number of TCP sockets.
306 struct percpu_counter tcp_sockets_allocated
;
307 EXPORT_SYMBOL(tcp_sockets_allocated
);
312 struct tcp_splice_state
{
313 struct pipe_inode_info
*pipe
;
319 * Pressure flag: try to collapse.
320 * Technical note: it is used by multiple contexts non atomically.
321 * All the __sk_mem_schedule() is of this nature: accounting
322 * is strict, actions are advisory and have some latency.
324 int tcp_memory_pressure __read_mostly
;
325 EXPORT_SYMBOL(tcp_memory_pressure
);
327 void tcp_enter_memory_pressure(struct sock
*sk
)
329 if (!tcp_memory_pressure
) {
330 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMEMORYPRESSURES
);
331 tcp_memory_pressure
= 1;
334 EXPORT_SYMBOL(tcp_enter_memory_pressure
);
336 /* Convert seconds to retransmits based on initial and max timeout */
337 static u8
secs_to_retrans(int seconds
, int timeout
, int rto_max
)
342 int period
= timeout
;
345 while (seconds
> period
&& res
< 255) {
348 if (timeout
> rto_max
)
356 /* Convert retransmits to seconds based on initial and max timeout */
357 static int retrans_to_secs(u8 retrans
, int timeout
, int rto_max
)
365 if (timeout
> rto_max
)
373 /* Address-family independent initialization for a tcp_sock.
375 * NOTE: A lot of things set to zero explicitly by call to
376 * sk_alloc() so need not be done here.
378 void tcp_init_sock(struct sock
*sk
)
380 struct inet_connection_sock
*icsk
= inet_csk(sk
);
381 struct tcp_sock
*tp
= tcp_sk(sk
);
383 skb_queue_head_init(&tp
->out_of_order_queue
);
384 tcp_init_xmit_timers(sk
);
385 tcp_prequeue_init(tp
);
386 INIT_LIST_HEAD(&tp
->tsq_node
);
388 icsk
->icsk_rto
= TCP_TIMEOUT_INIT
;
389 tp
->mdev
= TCP_TIMEOUT_INIT
;
391 /* So many TCP implementations out there (incorrectly) count the
392 * initial SYN frame in their delayed-ACK and congestion control
393 * algorithms that we must have the following bandaid to talk
394 * efficiently to them. -DaveM
396 tp
->snd_cwnd
= TCP_INIT_CWND
;
398 /* See draft-stevens-tcpca-spec-01 for discussion of the
399 * initialization of these values.
401 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
402 tp
->snd_cwnd_clamp
= ~0;
403 tp
->mss_cache
= TCP_MSS_DEFAULT
;
405 tp
->reordering
= sysctl_tcp_reordering
;
406 tcp_enable_early_retrans(tp
);
407 icsk
->icsk_ca_ops
= &tcp_init_congestion_ops
;
411 sk
->sk_state
= TCP_CLOSE
;
413 sk
->sk_write_space
= sk_stream_write_space
;
414 sock_set_flag(sk
, SOCK_USE_WRITE_QUEUE
);
416 icsk
->icsk_sync_mss
= tcp_sync_mss
;
418 /* Presumed zeroed, in order of appearance:
419 * cookie_in_always, cookie_out_never,
420 * s_data_constant, s_data_in, s_data_out
422 sk
->sk_sndbuf
= sysctl_tcp_wmem
[1];
423 sk
->sk_rcvbuf
= sysctl_tcp_rmem
[1];
426 sock_update_memcg(sk
);
427 sk_sockets_allocated_inc(sk
);
430 EXPORT_SYMBOL(tcp_init_sock
);
433 * Wait for a TCP event.
435 * Note that we don't need to lock the socket, as the upper poll layers
436 * take care of normal races (between the test and the event) and we don't
437 * go look at any of the socket buffers directly.
439 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
442 struct sock
*sk
= sock
->sk
;
443 const struct tcp_sock
*tp
= tcp_sk(sk
);
445 sock_poll_wait(file
, sk_sleep(sk
), wait
);
446 if (sk
->sk_state
== TCP_LISTEN
)
447 return inet_csk_listen_poll(sk
);
449 /* Socket is not locked. We are protected from async events
450 * by poll logic and correct handling of state changes
451 * made by other threads is impossible in any case.
457 * POLLHUP is certainly not done right. But poll() doesn't
458 * have a notion of HUP in just one direction, and for a
459 * socket the read side is more interesting.
461 * Some poll() documentation says that POLLHUP is incompatible
462 * with the POLLOUT/POLLWR flags, so somebody should check this
463 * all. But careful, it tends to be safer to return too many
464 * bits than too few, and you can easily break real applications
465 * if you don't tell them that something has hung up!
469 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
470 * our fs/select.c). It means that after we received EOF,
471 * poll always returns immediately, making impossible poll() on write()
472 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
473 * if and only if shutdown has been made in both directions.
474 * Actually, it is interesting to look how Solaris and DUX
475 * solve this dilemma. I would prefer, if POLLHUP were maskable,
476 * then we could set it on SND_SHUTDOWN. BTW examples given
477 * in Stevens' books assume exactly this behaviour, it explains
478 * why POLLHUP is incompatible with POLLOUT. --ANK
480 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
481 * blocking on fresh not-connected or disconnected socket. --ANK
483 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| sk
->sk_state
== TCP_CLOSE
)
485 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
486 mask
|= POLLIN
| POLLRDNORM
| POLLRDHUP
;
488 /* Connected or passive Fast Open socket? */
489 if (sk
->sk_state
!= TCP_SYN_SENT
&&
490 (sk
->sk_state
!= TCP_SYN_RECV
|| tp
->fastopen_rsk
!= NULL
)) {
491 int target
= sock_rcvlowat(sk
, 0, INT_MAX
);
493 if (tp
->urg_seq
== tp
->copied_seq
&&
494 !sock_flag(sk
, SOCK_URGINLINE
) &&
498 /* Potential race condition. If read of tp below will
499 * escape above sk->sk_state, we can be illegally awaken
500 * in SYN_* states. */
501 if (tp
->rcv_nxt
- tp
->copied_seq
>= target
)
502 mask
|= POLLIN
| POLLRDNORM
;
504 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
505 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
)) {
506 mask
|= POLLOUT
| POLLWRNORM
;
507 } else { /* send SIGIO later */
508 set_bit(SOCK_ASYNC_NOSPACE
,
509 &sk
->sk_socket
->flags
);
510 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
512 /* Race breaker. If space is freed after
513 * wspace test but before the flags are set,
514 * IO signal will be lost.
516 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
))
517 mask
|= POLLOUT
| POLLWRNORM
;
520 mask
|= POLLOUT
| POLLWRNORM
;
522 if (tp
->urg_data
& TCP_URG_VALID
)
525 /* This barrier is coupled with smp_wmb() in tcp_reset() */
532 EXPORT_SYMBOL(tcp_poll
);
534 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
536 struct tcp_sock
*tp
= tcp_sk(sk
);
542 if (sk
->sk_state
== TCP_LISTEN
)
545 slow
= lock_sock_fast(sk
);
546 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
548 else if (sock_flag(sk
, SOCK_URGINLINE
) ||
550 before(tp
->urg_seq
, tp
->copied_seq
) ||
551 !before(tp
->urg_seq
, tp
->rcv_nxt
)) {
553 answ
= tp
->rcv_nxt
- tp
->copied_seq
;
555 /* Subtract 1, if FIN was received */
556 if (answ
&& sock_flag(sk
, SOCK_DONE
))
559 answ
= tp
->urg_seq
- tp
->copied_seq
;
560 unlock_sock_fast(sk
, slow
);
563 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
566 if (sk
->sk_state
== TCP_LISTEN
)
569 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
572 answ
= tp
->write_seq
- tp
->snd_una
;
575 if (sk
->sk_state
== TCP_LISTEN
)
578 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
581 answ
= tp
->write_seq
- tp
->snd_nxt
;
583 /* MTK_NET_CHANGES */
586 struct uid_err uid_e
;
587 if (copy_from_user(&uid_e
, (char __user
*)arg
, sizeof(uid_e
)))
589 printk(KERN_WARNING
"SIOCKILLSOCK uid = %d , err = %d",
590 uid_e
.appuid
, uid_e
.errNum
);
591 if (uid_e
.errNum
== 0)
593 // handle BR release problem
594 tcp_v4_handle_retrans_time_by_uid(uid_e
);
598 tcp_v4_reset_connections_by_uid(uid_e
);
607 return put_user(answ
, (int __user
*)arg
);
609 EXPORT_SYMBOL(tcp_ioctl
);
611 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
613 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
614 tp
->pushed_seq
= tp
->write_seq
;
617 static inline bool forced_push(const struct tcp_sock
*tp
)
619 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
622 static inline void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
624 struct tcp_sock
*tp
= tcp_sk(sk
);
625 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
628 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
629 tcb
->tcp_flags
= TCPHDR_ACK
;
631 skb_header_release(skb
);
632 tcp_add_write_queue_tail(sk
, skb
);
633 sk
->sk_wmem_queued
+= skb
->truesize
;
634 sk_mem_charge(sk
, skb
->truesize
);
635 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
636 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
639 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
)
642 tp
->snd_up
= tp
->write_seq
;
645 static inline void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
648 if (tcp_send_head(sk
)) {
649 struct tcp_sock
*tp
= tcp_sk(sk
);
651 if (!(flags
& MSG_MORE
) || forced_push(tp
))
652 tcp_mark_push(tp
, tcp_write_queue_tail(sk
));
654 tcp_mark_urg(tp
, flags
);
655 __tcp_push_pending_frames(sk
, mss_now
,
656 (flags
& MSG_MORE
) ? TCP_NAGLE_CORK
: nonagle
);
660 static int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
661 unsigned int offset
, size_t len
)
663 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
666 ret
= skb_splice_bits(skb
, offset
, tss
->pipe
, min(rd_desc
->count
, len
),
669 rd_desc
->count
-= ret
;
673 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
675 /* Store TCP splice context information in read_descriptor_t. */
676 read_descriptor_t rd_desc
= {
681 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
685 * tcp_splice_read - splice data from TCP socket to a pipe
686 * @sock: socket to splice from
687 * @ppos: position (not valid)
688 * @pipe: pipe to splice to
689 * @len: number of bytes to splice
690 * @flags: splice modifier flags
693 * Will read pages from given socket and fill them into a pipe.
696 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
697 struct pipe_inode_info
*pipe
, size_t len
,
700 struct sock
*sk
= sock
->sk
;
701 struct tcp_splice_state tss
= {
710 sock_rps_record_flow(sk
);
712 * We can't seek on a socket input
721 timeo
= sock_rcvtimeo(sk
, sock
->file
->f_flags
& O_NONBLOCK
);
723 ret
= __tcp_splice_read(sk
, &tss
);
729 if (sock_flag(sk
, SOCK_DONE
))
732 ret
= sock_error(sk
);
735 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
737 if (sk
->sk_state
== TCP_CLOSE
) {
739 * This occurs when user tries to read
740 * from never connected socket.
742 if (!sock_flag(sk
, SOCK_DONE
))
750 /* if __tcp_splice_read() got nothing while we have
751 * an skb in receive queue, we do not want to loop.
752 * This might happen with URG data.
754 if (!skb_queue_empty(&sk
->sk_receive_queue
))
756 sk_wait_data(sk
, &timeo
);
757 if (signal_pending(current
)) {
758 ret
= sock_intr_errno(timeo
);
771 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
772 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
773 signal_pending(current
))
784 EXPORT_SYMBOL(tcp_splice_read
);
786 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
)
790 /* The TCP header must be at least 32-bit aligned. */
791 size
= ALIGN(size
, 4);
793 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
795 if (sk_wmem_schedule(sk
, skb
->truesize
)) {
796 skb_reserve(skb
, sk
->sk_prot
->max_header
);
798 * Make sure that we have exactly size bytes
799 * available to the caller, no more, no less.
801 skb
->reserved_tailroom
= skb
->end
- skb
->tail
- size
;
806 sk
->sk_prot
->enter_memory_pressure(sk
);
807 sk_stream_moderate_sndbuf(sk
);
812 static unsigned int tcp_xmit_size_goal(struct sock
*sk
, u32 mss_now
,
815 struct tcp_sock
*tp
= tcp_sk(sk
);
816 u32 xmit_size_goal
, old_size_goal
;
818 xmit_size_goal
= mss_now
;
820 if (large_allowed
&& sk_can_gso(sk
)) {
823 /* Maybe we should/could use sk->sk_prot->max_header here ? */
824 hlen
= inet_csk(sk
)->icsk_af_ops
->net_header_len
+
825 inet_csk(sk
)->icsk_ext_hdr_len
+
828 /* Goal is to send at least one packet per ms,
829 * not one big TSO packet every 100 ms.
830 * This preserves ACK clocking and is consistent
831 * with tcp_tso_should_defer() heuristic.
833 gso_size
= sk
->sk_pacing_rate
/ (2 * MSEC_PER_SEC
);
834 gso_size
= max_t(u32
, gso_size
,
835 sysctl_tcp_min_tso_segs
* mss_now
);
837 xmit_size_goal
= min_t(u32
, gso_size
,
838 sk
->sk_gso_max_size
- 1 - hlen
);
840 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
842 /* We try hard to avoid divides here */
843 old_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
845 if (likely(old_size_goal
<= xmit_size_goal
&&
846 old_size_goal
+ mss_now
> xmit_size_goal
)) {
847 xmit_size_goal
= old_size_goal
;
849 tp
->xmit_size_goal_segs
=
850 min_t(u16
, xmit_size_goal
/ mss_now
,
851 sk
->sk_gso_max_segs
);
852 xmit_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
856 return max(xmit_size_goal
, mss_now
);
859 static int tcp_send_mss(struct sock
*sk
, int *size_goal
, int flags
)
863 mss_now
= tcp_current_mss(sk
);
864 *size_goal
= tcp_xmit_size_goal(sk
, mss_now
, !(flags
& MSG_OOB
));
869 static ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
*page
, int offset
,
870 size_t size
, int flags
)
872 struct tcp_sock
*tp
= tcp_sk(sk
);
873 int mss_now
, size_goal
;
876 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
878 /* Wait for a connection to finish. One exception is TCP Fast Open
879 * (passive side) where data is allowed to be sent before a connection
880 * is fully established.
882 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
883 !tcp_passive_fastopen(sk
)) {
884 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
888 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
890 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
894 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
898 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
902 if (!tcp_send_head(sk
) || (copy
= size_goal
- skb
->len
) <= 0) {
904 if (!sk_stream_memory_free(sk
))
905 goto wait_for_sndbuf
;
907 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
909 goto wait_for_memory
;
918 i
= skb_shinfo(skb
)->nr_frags
;
919 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
920 if (!can_coalesce
&& i
>= MAX_SKB_FRAGS
) {
921 tcp_mark_push(tp
, skb
);
924 if (!sk_wmem_schedule(sk
, copy
))
925 goto wait_for_memory
;
928 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
931 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
933 skb_shinfo(skb
)->tx_flags
|= SKBTX_SHARED_FRAG
;
936 skb
->data_len
+= copy
;
937 skb
->truesize
+= copy
;
938 sk
->sk_wmem_queued
+= copy
;
939 sk_mem_charge(sk
, copy
);
940 skb
->ip_summed
= CHECKSUM_PARTIAL
;
941 tp
->write_seq
+= copy
;
942 TCP_SKB_CB(skb
)->end_seq
+= copy
;
943 skb_shinfo(skb
)->gso_segs
= 0;
946 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
953 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
956 if (forced_push(tp
)) {
957 tcp_mark_push(tp
, skb
);
958 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
959 } else if (skb
== tcp_send_head(sk
))
960 tcp_push_one(sk
, mss_now
);
964 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
966 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
968 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
971 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
975 if (copied
&& !(flags
& MSG_SENDPAGE_NOTLAST
))
976 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
983 return sk_stream_error(sk
, flags
, err
);
986 int tcp_sendpage(struct sock
*sk
, struct page
*page
, int offset
,
987 size_t size
, int flags
)
991 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
992 !(sk
->sk_route_caps
& NETIF_F_ALL_CSUM
))
993 return sock_no_sendpage(sk
->sk_socket
, page
, offset
, size
,
997 res
= do_tcp_sendpages(sk
, page
, offset
, size
, flags
);
1001 EXPORT_SYMBOL(tcp_sendpage
);
1003 static inline int select_size(const struct sock
*sk
, bool sg
)
1005 const struct tcp_sock
*tp
= tcp_sk(sk
);
1006 int tmp
= tp
->mss_cache
;
1009 if (sk_can_gso(sk
)) {
1010 /* Small frames wont use a full page:
1011 * Payload will immediately follow tcp header.
1013 tmp
= SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER
);
1015 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
1017 if (tmp
>= pgbreak
&&
1018 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
1026 void tcp_free_fastopen_req(struct tcp_sock
*tp
)
1028 if (tp
->fastopen_req
!= NULL
) {
1029 kfree(tp
->fastopen_req
);
1030 tp
->fastopen_req
= NULL
;
1034 static int tcp_sendmsg_fastopen(struct sock
*sk
, struct msghdr
*msg
,
1035 int *copied
, size_t size
)
1037 struct tcp_sock
*tp
= tcp_sk(sk
);
1040 if (!(sysctl_tcp_fastopen
& TFO_CLIENT_ENABLE
))
1042 if (tp
->fastopen_req
!= NULL
)
1043 return -EALREADY
; /* Another Fast Open is in progress */
1045 tp
->fastopen_req
= kzalloc(sizeof(struct tcp_fastopen_request
),
1047 if (unlikely(tp
->fastopen_req
== NULL
))
1049 tp
->fastopen_req
->data
= msg
;
1050 tp
->fastopen_req
->size
= size
;
1052 flags
= (msg
->msg_flags
& MSG_DONTWAIT
) ? O_NONBLOCK
: 0;
1053 err
= __inet_stream_connect(sk
->sk_socket
, msg
->msg_name
,
1054 msg
->msg_namelen
, flags
);
1055 *copied
= tp
->fastopen_req
->copied
;
1056 tcp_free_fastopen_req(tp
);
1060 int tcp_sendmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1064 struct tcp_sock
*tp
= tcp_sk(sk
);
1065 struct sk_buff
*skb
;
1066 int iovlen
, flags
, err
, copied
= 0;
1067 int mss_now
= 0, size_goal
, copied_syn
= 0, offset
= 0;
1073 flags
= msg
->msg_flags
;
1074 if (flags
& MSG_FASTOPEN
) {
1075 err
= tcp_sendmsg_fastopen(sk
, msg
, &copied_syn
, size
);
1076 if (err
== -EINPROGRESS
&& copied_syn
> 0)
1080 offset
= copied_syn
;
1083 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
1085 /* Wait for a connection to finish. One exception is TCP Fast Open
1086 * (passive side) where data is allowed to be sent before a connection
1087 * is fully established.
1089 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
1090 !tcp_passive_fastopen(sk
)) {
1091 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
1095 if (unlikely(tp
->repair
)) {
1096 if (tp
->repair_queue
== TCP_RECV_QUEUE
) {
1097 copied
= tcp_send_rcvq(sk
, msg
, size
);
1102 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1105 /* 'common' sending to sendq */
1108 /* This should be in poll */
1109 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
1111 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1113 /* Ok commence sending. */
1114 iovlen
= msg
->msg_iovlen
;
1119 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
1122 sg
= !!(sk
->sk_route_caps
& NETIF_F_SG
);
1124 while (--iovlen
>= 0) {
1125 size_t seglen
= iov
->iov_len
;
1126 unsigned char __user
*from
= iov
->iov_base
;
1129 if (unlikely(offset
> 0)) { /* Skip bytes copied in SYN */
1130 if (offset
>= seglen
) {
1139 while (seglen
> 0) {
1141 int max
= size_goal
;
1143 skb
= tcp_write_queue_tail(sk
);
1144 if (tcp_send_head(sk
)) {
1145 if (skb
->ip_summed
== CHECKSUM_NONE
)
1147 copy
= max
- skb
->len
;
1152 /* Allocate new segment. If the interface is SG,
1153 * allocate skb fitting to single page.
1155 if (!sk_stream_memory_free(sk
))
1156 goto wait_for_sndbuf
;
1158 skb
= sk_stream_alloc_skb(sk
,
1159 select_size(sk
, sg
),
1162 goto wait_for_memory
;
1165 * All packets are restored as if they have
1166 * already been sent.
1169 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1172 * Check whether we can use HW checksum.
1174 if (sk
->sk_route_caps
& NETIF_F_ALL_CSUM
)
1175 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1177 skb_entail(sk
, skb
);
1182 /* Try to append data to the end of skb. */
1186 /* Where to copy to? */
1187 if (skb_availroom(skb
) > 0) {
1188 /* We have some space in skb head. Superb! */
1189 copy
= min_t(int, copy
, skb_availroom(skb
));
1190 err
= skb_add_data_nocache(sk
, skb
, from
, copy
);
1195 int i
= skb_shinfo(skb
)->nr_frags
;
1196 struct page_frag
*pfrag
= sk_page_frag(sk
);
1198 if (!sk_page_frag_refill(sk
, pfrag
))
1199 goto wait_for_memory
;
1201 if (!skb_can_coalesce(skb
, i
, pfrag
->page
,
1203 if (i
== MAX_SKB_FRAGS
|| !sg
) {
1204 tcp_mark_push(tp
, skb
);
1210 copy
= min_t(int, copy
, pfrag
->size
- pfrag
->offset
);
1212 if (!sk_wmem_schedule(sk
, copy
))
1213 goto wait_for_memory
;
1215 err
= skb_copy_to_page_nocache(sk
, from
, skb
,
1222 /* Update the skb. */
1224 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
1226 skb_fill_page_desc(skb
, i
, pfrag
->page
,
1227 pfrag
->offset
, copy
);
1228 get_page(pfrag
->page
);
1230 pfrag
->offset
+= copy
;
1234 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
1236 tp
->write_seq
+= copy
;
1237 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1238 skb_shinfo(skb
)->gso_segs
= 0;
1242 if ((seglen
-= copy
) == 0 && iovlen
== 0)
1245 if (skb
->len
< max
|| (flags
& MSG_OOB
) || unlikely(tp
->repair
))
1248 if (forced_push(tp
)) {
1249 tcp_mark_push(tp
, skb
);
1250 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1251 } else if (skb
== tcp_send_head(sk
))
1252 tcp_push_one(sk
, mss_now
);
1256 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1259 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
1261 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
1264 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1270 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
1274 if (copied
+ copied_syn
)
1275 uid_stat_tcp_snd(current_uid(), copied
+ copied_syn
);
1276 return copied
+ copied_syn
;
1280 tcp_unlink_write_queue(skb
, sk
);
1281 /* It is the one place in all of TCP, except connection
1282 * reset, where we can be unlinking the send_head.
1284 tcp_check_send_head(sk
, skb
);
1285 sk_wmem_free_skb(sk
, skb
);
1289 if (copied
+ copied_syn
)
1292 err
= sk_stream_error(sk
, flags
, err
);
1296 EXPORT_SYMBOL(tcp_sendmsg
);
1299 * Handle reading urgent data. BSD has very simple semantics for
1300 * this, no blocking and very strange errors 8)
1303 static int tcp_recv_urg(struct sock
*sk
, struct msghdr
*msg
, int len
, int flags
)
1305 struct tcp_sock
*tp
= tcp_sk(sk
);
1307 /* No URG data to read. */
1308 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1309 tp
->urg_data
== TCP_URG_READ
)
1310 return -EINVAL
; /* Yes this is right ! */
1312 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1315 if (tp
->urg_data
& TCP_URG_VALID
) {
1317 char c
= tp
->urg_data
;
1319 if (!(flags
& MSG_PEEK
))
1320 tp
->urg_data
= TCP_URG_READ
;
1322 /* Read urgent data. */
1323 msg
->msg_flags
|= MSG_OOB
;
1326 if (!(flags
& MSG_TRUNC
))
1327 err
= memcpy_toiovec(msg
->msg_iov
, &c
, 1);
1330 msg
->msg_flags
|= MSG_TRUNC
;
1332 return err
? -EFAULT
: len
;
1335 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1338 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1339 * the available implementations agree in this case:
1340 * this call should never block, independent of the
1341 * blocking state of the socket.
1342 * Mike <pall@rz.uni-karlsruhe.de>
1347 static int tcp_peek_sndq(struct sock
*sk
, struct msghdr
*msg
, int len
)
1349 struct sk_buff
*skb
;
1350 int copied
= 0, err
= 0;
1352 /* XXX -- need to support SO_PEEK_OFF */
1354 skb_queue_walk(&sk
->sk_write_queue
, skb
) {
1355 err
= skb_copy_datagram_iovec(skb
, 0, msg
->msg_iov
, skb
->len
);
1362 return err
?: copied
;
1365 /* Clean up the receive buffer for full frames taken by the user,
1366 * then send an ACK if necessary. COPIED is the number of bytes
1367 * tcp_recvmsg has given to the user so far, it speeds up the
1368 * calculation of whether or not we must ACK for the sake of
1371 void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1373 struct tcp_sock
*tp
= tcp_sk(sk
);
1374 bool time_to_ack
= false;
1376 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1378 WARN(skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
),
1379 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1380 tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
);
1382 if (inet_csk_ack_scheduled(sk
)) {
1383 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1384 /* Delayed ACKs frequently hit locked sockets during bulk
1386 if (icsk
->icsk_ack
.blocked
||
1387 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1388 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1390 * If this read emptied read buffer, we send ACK, if
1391 * connection is not bidirectional, user drained
1392 * receive buffer and there was a small segment
1396 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1397 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1398 !icsk
->icsk_ack
.pingpong
)) &&
1399 !atomic_read(&sk
->sk_rmem_alloc
)))
1403 /* We send an ACK if we can now advertise a non-zero window
1404 * which has been raised "significantly".
1406 * Even if window raised up to infinity, do not send window open ACK
1407 * in states, where we will not receive more. It is useless.
1409 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1410 __u32 rcv_window_now
= tcp_receive_window(tp
);
1412 /* Optimize, __tcp_select_window() is not cheap. */
1413 if (2*rcv_window_now
<= tp
->window_clamp
) {
1414 __u32 new_window
= __tcp_select_window(sk
);
1416 /* Send ACK now, if this read freed lots of space
1417 * in our buffer. Certainly, new_window is new window.
1418 * We can advertise it now, if it is not less than current one.
1419 * "Lots" means "at least twice" here.
1421 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1429 static void tcp_prequeue_process(struct sock
*sk
)
1431 struct sk_buff
*skb
;
1432 struct tcp_sock
*tp
= tcp_sk(sk
);
1434 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPPREQUEUED
);
1436 /* RX process wants to run with disabled BHs, though it is not
1439 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
1440 sk_backlog_rcv(sk
, skb
);
1443 /* Clear memory counter. */
1444 tp
->ucopy
.memory
= 0;
1447 #ifdef CONFIG_NET_DMA
1448 static void tcp_service_net_dma(struct sock
*sk
, bool wait
)
1450 dma_cookie_t done
, used
;
1451 dma_cookie_t last_issued
;
1452 struct tcp_sock
*tp
= tcp_sk(sk
);
1454 if (!tp
->ucopy
.dma_chan
)
1457 last_issued
= tp
->ucopy
.dma_cookie
;
1458 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1461 if (dma_async_is_tx_complete(tp
->ucopy
.dma_chan
,
1463 &used
) == DMA_SUCCESS
) {
1464 /* Safe to free early-copied skbs now */
1465 __skb_queue_purge(&sk
->sk_async_wait_queue
);
1468 struct sk_buff
*skb
;
1469 while ((skb
= skb_peek(&sk
->sk_async_wait_queue
)) &&
1470 (dma_async_is_complete(skb
->dma_cookie
, done
,
1471 used
) == DMA_SUCCESS
)) {
1472 __skb_dequeue(&sk
->sk_async_wait_queue
);
1480 static struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1482 struct sk_buff
*skb
;
1485 while ((skb
= skb_peek(&sk
->sk_receive_queue
)) != NULL
) {
1486 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1487 if (tcp_hdr(skb
)->syn
)
1489 if (offset
< skb
->len
|| tcp_hdr(skb
)->fin
) {
1493 /* This looks weird, but this can happen if TCP collapsing
1494 * splitted a fat GRO packet, while we released socket lock
1495 * in skb_splice_bits()
1497 sk_eat_skb(sk
, skb
, false);
1503 * This routine provides an alternative to tcp_recvmsg() for routines
1504 * that would like to handle copying from skbuffs directly in 'sendfile'
1507 * - It is assumed that the socket was locked by the caller.
1508 * - The routine does not block.
1509 * - At present, there is no support for reading OOB data
1510 * or for 'peeking' the socket using this routine
1511 * (although both would be easy to implement).
1513 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1514 sk_read_actor_t recv_actor
)
1516 struct sk_buff
*skb
;
1517 struct tcp_sock
*tp
= tcp_sk(sk
);
1518 u32 seq
= tp
->copied_seq
;
1522 if (sk
->sk_state
== TCP_LISTEN
)
1524 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1525 if (offset
< skb
->len
) {
1529 len
= skb
->len
- offset
;
1530 /* Stop reading if we hit a patch of urgent data */
1532 u32 urg_offset
= tp
->urg_seq
- seq
;
1533 if (urg_offset
< len
)
1538 used
= recv_actor(desc
, skb
, offset
, len
);
1543 } else if (used
<= len
) {
1548 /* If recv_actor drops the lock (e.g. TCP splice
1549 * receive) the skb pointer might be invalid when
1550 * getting here: tcp_collapse might have deleted it
1551 * while aggregating skbs from the socket queue.
1553 skb
= tcp_recv_skb(sk
, seq
- 1, &offset
);
1556 /* TCP coalescing might have appended data to the skb.
1557 * Try to splice more frags
1559 if (offset
+ 1 != skb
->len
)
1562 if (tcp_hdr(skb
)->fin
) {
1563 sk_eat_skb(sk
, skb
, false);
1567 sk_eat_skb(sk
, skb
, false);
1570 tp
->copied_seq
= seq
;
1572 tp
->copied_seq
= seq
;
1574 tcp_rcv_space_adjust(sk
);
1576 /* Clean up data we have read: This will do ACK frames. */
1578 tcp_recv_skb(sk
, seq
, &offset
);
1579 tcp_cleanup_rbuf(sk
, copied
);
1580 uid_stat_tcp_rcv(current_uid(), copied
);
1584 EXPORT_SYMBOL(tcp_read_sock
);
1587 * This routine copies from a sock struct into the user buffer.
1589 * Technical note: in 2.3 we work on _locked_ socket, so that
1590 * tricks with *seq access order and skb->users are not required.
1591 * Probably, code can be easily improved even more.
1594 int tcp_recvmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1595 size_t len
, int nonblock
, int flags
, int *addr_len
)
1597 struct tcp_sock
*tp
= tcp_sk(sk
);
1603 int target
; /* Read at least this many bytes */
1605 struct task_struct
*user_recv
= NULL
;
1606 bool copied_early
= false;
1607 struct sk_buff
*skb
;
1613 if (sk
->sk_state
== TCP_LISTEN
)
1616 timeo
= sock_rcvtimeo(sk
, nonblock
);
1618 /* Urgent data needs to be handled specially. */
1619 if (flags
& MSG_OOB
)
1622 if (unlikely(tp
->repair
)) {
1624 if (!(flags
& MSG_PEEK
))
1627 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
1631 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1634 /* 'common' recv queue MSG_PEEK-ing */
1637 seq
= &tp
->copied_seq
;
1638 if (flags
& MSG_PEEK
) {
1639 peek_seq
= tp
->copied_seq
;
1643 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1645 #ifdef CONFIG_NET_DMA
1646 tp
->ucopy
.dma_chan
= NULL
;
1648 skb
= skb_peek_tail(&sk
->sk_receive_queue
);
1653 available
= TCP_SKB_CB(skb
)->seq
+ skb
->len
- (*seq
);
1654 if ((available
< target
) &&
1655 (len
> sysctl_tcp_dma_copybreak
) && !(flags
& MSG_PEEK
) &&
1656 !sysctl_tcp_low_latency
&&
1657 net_dma_find_channel()) {
1658 preempt_enable_no_resched();
1659 tp
->ucopy
.pinned_list
=
1660 dma_pin_iovec_pages(msg
->msg_iov
, len
);
1662 preempt_enable_no_resched();
1670 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1671 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1674 if (signal_pending(current
)) {
1675 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1680 /* Next get a buffer. */
1682 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1683 /* Now that we have two receive queues this
1686 if (WARN(before(*seq
, TCP_SKB_CB(skb
)->seq
),
1687 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1688 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
,
1692 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1693 if (tcp_hdr(skb
)->syn
)
1695 if (offset
< skb
->len
)
1697 if (tcp_hdr(skb
)->fin
)
1699 WARN(!(flags
& MSG_PEEK
),
1700 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1701 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
, flags
);
1704 /* Well, if we have backlog, try to process it now yet. */
1706 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1711 sk
->sk_state
== TCP_CLOSE
||
1712 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1714 signal_pending(current
))
1717 if (sock_flag(sk
, SOCK_DONE
))
1721 copied
= sock_error(sk
);
1725 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1728 if (sk
->sk_state
== TCP_CLOSE
) {
1729 if (!sock_flag(sk
, SOCK_DONE
)) {
1730 /* This occurs when user tries to read
1731 * from never connected socket.
1744 if (signal_pending(current
)) {
1745 copied
= sock_intr_errno(timeo
);
1750 tcp_cleanup_rbuf(sk
, copied
);
1752 if (!sysctl_tcp_low_latency
&& tp
->ucopy
.task
== user_recv
) {
1753 /* Install new reader */
1754 if (!user_recv
&& !(flags
& (MSG_TRUNC
| MSG_PEEK
))) {
1755 user_recv
= current
;
1756 tp
->ucopy
.task
= user_recv
;
1757 tp
->ucopy
.iov
= msg
->msg_iov
;
1760 tp
->ucopy
.len
= len
;
1762 WARN_ON(tp
->copied_seq
!= tp
->rcv_nxt
&&
1763 !(flags
& (MSG_PEEK
| MSG_TRUNC
)));
1765 /* Ugly... If prequeue is not empty, we have to
1766 * process it before releasing socket, otherwise
1767 * order will be broken at second iteration.
1768 * More elegant solution is required!!!
1770 * Look: we have the following (pseudo)queues:
1772 * 1. packets in flight
1777 * Each queue can be processed only if the next ones
1778 * are empty. At this point we have empty receive_queue.
1779 * But prequeue _can_ be not empty after 2nd iteration,
1780 * when we jumped to start of loop because backlog
1781 * processing added something to receive_queue.
1782 * We cannot release_sock(), because backlog contains
1783 * packets arrived _after_ prequeued ones.
1785 * Shortly, algorithm is clear --- to process all
1786 * the queues in order. We could make it more directly,
1787 * requeueing packets from backlog to prequeue, if
1788 * is not empty. It is more elegant, but eats cycles,
1791 if (!skb_queue_empty(&tp
->ucopy
.prequeue
))
1794 /* __ Set realtime policy in scheduler __ */
1797 #ifdef CONFIG_NET_DMA
1798 if (tp
->ucopy
.dma_chan
) {
1799 if (tp
->rcv_wnd
== 0 &&
1800 !skb_queue_empty(&sk
->sk_async_wait_queue
)) {
1801 tcp_service_net_dma(sk
, true);
1802 tcp_cleanup_rbuf(sk
, copied
);
1804 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1807 if (copied
>= target
) {
1808 /* Do not sleep, just process backlog. */
1812 sk_wait_data(sk
, &timeo
);
1814 #ifdef CONFIG_NET_DMA
1815 tcp_service_net_dma(sk
, false); /* Don't block */
1816 tp
->ucopy
.wakeup
= 0;
1822 /* __ Restore normal policy in scheduler __ */
1824 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1825 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG
, chunk
);
1830 if (tp
->rcv_nxt
== tp
->copied_seq
&&
1831 !skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1833 tcp_prequeue_process(sk
);
1835 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1836 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1842 if ((flags
& MSG_PEEK
) &&
1843 (peek_seq
- copied
- urg_hole
!= tp
->copied_seq
)) {
1844 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1846 task_pid_nr(current
));
1847 peek_seq
= tp
->copied_seq
;
1852 /* Ok so how much can we use? */
1853 used
= skb
->len
- offset
;
1857 /* Do we have urgent data here? */
1859 u32 urg_offset
= tp
->urg_seq
- *seq
;
1860 if (urg_offset
< used
) {
1862 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1875 if (!(flags
& MSG_TRUNC
)) {
1876 #ifdef CONFIG_NET_DMA
1877 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
1878 tp
->ucopy
.dma_chan
= net_dma_find_channel();
1880 if (tp
->ucopy
.dma_chan
) {
1881 tp
->ucopy
.dma_cookie
= dma_skb_copy_datagram_iovec(
1882 tp
->ucopy
.dma_chan
, skb
, offset
,
1884 tp
->ucopy
.pinned_list
);
1886 if (tp
->ucopy
.dma_cookie
< 0) {
1888 pr_alert("%s: dma_cookie < 0\n",
1891 /* Exception. Bailout! */
1897 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1899 if ((offset
+ used
) == skb
->len
)
1900 copied_early
= true;
1905 err
= skb_copy_datagram_iovec(skb
, offset
,
1906 msg
->msg_iov
, used
);
1908 /* Exception. Bailout! */
1920 tcp_rcv_space_adjust(sk
);
1923 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1925 tcp_fast_path_check(sk
);
1927 if (used
+ offset
< skb
->len
)
1930 if (tcp_hdr(skb
)->fin
)
1932 if (!(flags
& MSG_PEEK
)) {
1933 sk_eat_skb(sk
, skb
, copied_early
);
1934 copied_early
= false;
1939 /* Process the FIN. */
1941 if (!(flags
& MSG_PEEK
)) {
1942 sk_eat_skb(sk
, skb
, copied_early
);
1943 copied_early
= false;
1949 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1952 tp
->ucopy
.len
= copied
> 0 ? len
: 0;
1954 tcp_prequeue_process(sk
);
1956 if (copied
> 0 && (chunk
= len
- tp
->ucopy
.len
) != 0) {
1957 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1963 tp
->ucopy
.task
= NULL
;
1967 #ifdef CONFIG_NET_DMA
1968 tcp_service_net_dma(sk
, true); /* Wait for queue to drain */
1969 tp
->ucopy
.dma_chan
= NULL
;
1971 if (tp
->ucopy
.pinned_list
) {
1972 dma_unpin_iovec_pages(tp
->ucopy
.pinned_list
);
1973 tp
->ucopy
.pinned_list
= NULL
;
1977 /* According to UNIX98, msg_name/msg_namelen are ignored
1978 * on connected socket. I was just happy when found this 8) --ANK
1981 /* Clean up data we have read: This will do ACK frames. */
1982 tcp_cleanup_rbuf(sk
, copied
);
1987 uid_stat_tcp_rcv(current_uid(), copied
);
1995 err
= tcp_recv_urg(sk
, msg
, len
, flags
);
1997 uid_stat_tcp_rcv(current_uid(), err
);
2001 err
= tcp_peek_sndq(sk
, msg
, len
);
2004 EXPORT_SYMBOL(tcp_recvmsg
);
2006 void tcp_set_state(struct sock
*sk
, int state
)
2008 int oldstate
= sk
->sk_state
;
2011 case TCP_ESTABLISHED
:
2012 if (oldstate
!= TCP_ESTABLISHED
)
2013 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
2017 if (oldstate
== TCP_CLOSE_WAIT
|| oldstate
== TCP_ESTABLISHED
)
2018 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ESTABRESETS
);
2020 sk
->sk_prot
->unhash(sk
);
2021 if (inet_csk(sk
)->icsk_bind_hash
&&
2022 !(sk
->sk_userlocks
& SOCK_BINDPORT_LOCK
))
2026 if (oldstate
== TCP_ESTABLISHED
)
2027 TCP_DEC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
2030 /* Change state AFTER socket is unhashed to avoid closed
2031 * socket sitting in hash tables.
2033 sk
->sk_state
= state
;
2036 SOCK_DEBUG(sk
, "TCP sk=%p, State %s -> %s\n", sk
, statename
[oldstate
], statename
[state
]);
2039 EXPORT_SYMBOL_GPL(tcp_set_state
);
2042 * State processing on a close. This implements the state shift for
2043 * sending our FIN frame. Note that we only send a FIN for some
2044 * states. A shutdown() may have already sent the FIN, or we may be
2048 static const unsigned char new_state
[16] = {
2049 /* current state: new state: action: */
2050 /* (Invalid) */ TCP_CLOSE
,
2051 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2052 /* TCP_SYN_SENT */ TCP_CLOSE
,
2053 /* TCP_SYN_RECV */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2054 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1
,
2055 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2
,
2056 /* TCP_TIME_WAIT */ TCP_CLOSE
,
2057 /* TCP_CLOSE */ TCP_CLOSE
,
2058 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK
| TCP_ACTION_FIN
,
2059 /* TCP_LAST_ACK */ TCP_LAST_ACK
,
2060 /* TCP_LISTEN */ TCP_CLOSE
,
2061 /* TCP_CLOSING */ TCP_CLOSING
,
2064 static int tcp_close_state(struct sock
*sk
)
2066 int next
= (int)new_state
[sk
->sk_state
];
2067 int ns
= next
& TCP_STATE_MASK
;
2069 tcp_set_state(sk
, ns
);
2071 return next
& TCP_ACTION_FIN
;
2075 * Shutdown the sending side of a connection. Much like close except
2076 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2079 void tcp_shutdown(struct sock
*sk
, int how
)
2081 /* We need to grab some memory, and put together a FIN,
2082 * and then put it into the queue to be sent.
2083 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2085 if (!(how
& SEND_SHUTDOWN
))
2088 /* If we've already sent a FIN, or it's a closed state, skip this. */
2089 if ((1 << sk
->sk_state
) &
2090 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
2091 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
2092 /* Clear out any half completed packets. FIN if needed. */
2093 if (tcp_close_state(sk
))
2097 EXPORT_SYMBOL(tcp_shutdown
);
2099 bool tcp_check_oom(struct sock
*sk
, int shift
)
2101 bool too_many_orphans
, out_of_socket_memory
;
2103 too_many_orphans
= tcp_too_many_orphans(sk
, shift
);
2104 out_of_socket_memory
= tcp_out_of_memory(sk
);
2106 if (too_many_orphans
)
2107 net_info_ratelimited("too many orphaned sockets\n");
2108 if (out_of_socket_memory
)
2109 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2110 return too_many_orphans
|| out_of_socket_memory
;
2113 void tcp_close(struct sock
*sk
, long timeout
)
2115 struct sk_buff
*skb
;
2116 int data_was_unread
= 0;
2120 sk
->sk_shutdown
= SHUTDOWN_MASK
;
2122 if (sk
->sk_state
== TCP_LISTEN
) {
2123 tcp_set_state(sk
, TCP_CLOSE
);
2126 inet_csk_listen_stop(sk
);
2128 goto adjudge_to_death
;
2131 /* We need to flush the recv. buffs. We do this only on the
2132 * descriptor close, not protocol-sourced closes, because the
2133 * reader process may not have drained the data yet!
2135 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
2136 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
-
2138 data_was_unread
+= len
;
2144 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2145 if (sk
->sk_state
== TCP_CLOSE
)
2146 goto adjudge_to_death
;
2148 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2149 * data was lost. To witness the awful effects of the old behavior of
2150 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2151 * GET in an FTP client, suspend the process, wait for the client to
2152 * advertise a zero window, then kill -9 the FTP client, wheee...
2153 * Note: timeout is always zero in such a case.
2155 if (unlikely(tcp_sk(sk
)->repair
)) {
2156 sk
->sk_prot
->disconnect(sk
, 0);
2157 } else if (data_was_unread
) {
2158 /* Unread data was tossed, zap the connection. */
2159 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONCLOSE
);
2160 tcp_set_state(sk
, TCP_CLOSE
);
2161 tcp_send_active_reset(sk
, sk
->sk_allocation
);
2162 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
2163 /* Check zero linger _after_ checking for unread data. */
2164 sk
->sk_prot
->disconnect(sk
, 0);
2165 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
2166 } else if (tcp_close_state(sk
)) {
2167 /* We FIN if the application ate all the data before
2168 * zapping the connection.
2171 /* RED-PEN. Formally speaking, we have broken TCP state
2172 * machine. State transitions:
2174 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2175 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2176 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2178 * are legal only when FIN has been sent (i.e. in window),
2179 * rather than queued out of window. Purists blame.
2181 * F.e. "RFC state" is ESTABLISHED,
2182 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2184 * The visible declinations are that sometimes
2185 * we enter time-wait state, when it is not required really
2186 * (harmless), do not send active resets, when they are
2187 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2188 * they look as CLOSING or LAST_ACK for Linux)
2189 * Probably, I missed some more holelets.
2191 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2192 * in a single packet! (May consider it later but will
2193 * probably need API support or TCP_CORK SYN-ACK until
2194 * data is written and socket is closed.)
2199 sk_stream_wait_close(sk
, timeout
);
2202 state
= sk
->sk_state
;
2206 /* It is the last release_sock in its life. It will remove backlog. */
2210 /* Now socket is owned by kernel and we acquire BH lock
2211 to finish close. No need to check for user refs.
2215 WARN_ON(sock_owned_by_user(sk
));
2217 percpu_counter_inc(sk
->sk_prot
->orphan_count
);
2219 /* Have we already been destroyed by a softirq or backlog? */
2220 if (state
!= TCP_CLOSE
&& sk
->sk_state
== TCP_CLOSE
)
2223 /* This is a (useful) BSD violating of the RFC. There is a
2224 * problem with TCP as specified in that the other end could
2225 * keep a socket open forever with no application left this end.
2226 * We use a 3 minute timeout (about the same as BSD) then kill
2227 * our end. If they send after that then tough - BUT: long enough
2228 * that we won't make the old 4*rto = almost no time - whoops
2231 * Nope, it was not mistake. It is really desired behaviour
2232 * f.e. on http servers, when such sockets are useless, but
2233 * consume significant resources. Let's do it with special
2234 * linger2 option. --ANK
2237 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
2238 struct tcp_sock
*tp
= tcp_sk(sk
);
2239 if (tp
->linger2
< 0) {
2240 tcp_set_state(sk
, TCP_CLOSE
);
2241 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2242 NET_INC_STATS_BH(sock_net(sk
),
2243 LINUX_MIB_TCPABORTONLINGER
);
2245 const int tmo
= tcp_fin_time(sk
);
2247 if (tmo
> TCP_TIMEWAIT_LEN
) {
2248 inet_csk_reset_keepalive_timer(sk
,
2249 tmo
- TCP_TIMEWAIT_LEN
);
2251 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
2256 if (sk
->sk_state
!= TCP_CLOSE
) {
2258 if (tcp_check_oom(sk
, 0)) {
2259 tcp_set_state(sk
, TCP_CLOSE
);
2260 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2261 NET_INC_STATS_BH(sock_net(sk
),
2262 LINUX_MIB_TCPABORTONMEMORY
);
2266 if (sk
->sk_state
== TCP_CLOSE
) {
2267 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
2268 /* We could get here with a non-NULL req if the socket is
2269 * aborted (e.g., closed with unread data) before 3WHS
2273 reqsk_fastopen_remove(sk
, req
, false);
2274 inet_csk_destroy_sock(sk
);
2276 /* Otherwise, socket is reprieved until protocol close. */
2283 EXPORT_SYMBOL(tcp_close
);
2285 /* These states need RST on ABORT according to RFC793 */
2287 static inline bool tcp_need_reset(int state
)
2289 return (1 << state
) &
2290 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
2291 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
2294 int tcp_disconnect(struct sock
*sk
, int flags
)
2296 struct inet_sock
*inet
= inet_sk(sk
);
2297 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2298 struct tcp_sock
*tp
= tcp_sk(sk
);
2300 int old_state
= sk
->sk_state
;
2302 if (old_state
!= TCP_CLOSE
)
2303 tcp_set_state(sk
, TCP_CLOSE
);
2305 /* ABORT function of RFC793 */
2306 if (old_state
== TCP_LISTEN
) {
2307 inet_csk_listen_stop(sk
);
2308 } else if (unlikely(tp
->repair
)) {
2309 sk
->sk_err
= ECONNABORTED
;
2310 } else if (tcp_need_reset(old_state
) ||
2311 (tp
->snd_nxt
!= tp
->write_seq
&&
2312 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
2313 /* The last check adjusts for discrepancy of Linux wrt. RFC
2316 tcp_send_active_reset(sk
, gfp_any());
2317 sk
->sk_err
= ECONNRESET
;
2318 } else if (old_state
== TCP_SYN_SENT
)
2319 sk
->sk_err
= ECONNRESET
;
2321 tcp_clear_xmit_timers(sk
);
2322 __skb_queue_purge(&sk
->sk_receive_queue
);
2323 tcp_write_queue_purge(sk
);
2324 __skb_queue_purge(&tp
->out_of_order_queue
);
2325 #ifdef CONFIG_NET_DMA
2326 __skb_queue_purge(&sk
->sk_async_wait_queue
);
2329 inet
->inet_dport
= 0;
2331 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
2332 inet_reset_saddr(sk
);
2334 sk
->sk_shutdown
= 0;
2335 sock_reset_flag(sk
, SOCK_DONE
);
2337 if ((tp
->write_seq
+= tp
->max_window
+ 2) == 0)
2339 icsk
->icsk_backoff
= 0;
2341 icsk
->icsk_probes_out
= 0;
2342 tp
->packets_out
= 0;
2343 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
2344 tp
->snd_cwnd_cnt
= 0;
2345 tp
->window_clamp
= 0;
2346 tcp_set_ca_state(sk
, TCP_CA_Open
);
2347 tcp_clear_retrans(tp
);
2348 inet_csk_delack_init(sk
);
2349 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2350 * issue in __tcp_select_window()
2352 icsk
->icsk_ack
.rcv_mss
= TCP_MIN_MSS
;
2353 tcp_init_send_head(sk
);
2354 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
2356 dst_release(sk
->sk_rx_dst
);
2357 sk
->sk_rx_dst
= NULL
;
2359 WARN_ON(inet
->inet_num
&& !icsk
->icsk_bind_hash
);
2361 sk
->sk_error_report(sk
);
2364 EXPORT_SYMBOL(tcp_disconnect
);
2366 void tcp_sock_destruct(struct sock
*sk
)
2368 inet_sock_destruct(sk
);
2370 kfree(inet_csk(sk
)->icsk_accept_queue
.fastopenq
);
2373 static inline bool tcp_can_repair_sock(const struct sock
*sk
)
2375 return ns_capable(sock_net(sk
)->user_ns
, CAP_NET_ADMIN
) &&
2376 ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_ESTABLISHED
));
2379 static int tcp_repair_options_est(struct tcp_sock
*tp
,
2380 struct tcp_repair_opt __user
*optbuf
, unsigned int len
)
2382 struct tcp_repair_opt opt
;
2384 while (len
>= sizeof(opt
)) {
2385 if (copy_from_user(&opt
, optbuf
, sizeof(opt
)))
2391 switch (opt
.opt_code
) {
2393 tp
->rx_opt
.mss_clamp
= opt
.opt_val
;
2397 u16 snd_wscale
= opt
.opt_val
& 0xFFFF;
2398 u16 rcv_wscale
= opt
.opt_val
>> 16;
2400 if (snd_wscale
> 14 || rcv_wscale
> 14)
2403 tp
->rx_opt
.snd_wscale
= snd_wscale
;
2404 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2405 tp
->rx_opt
.wscale_ok
= 1;
2408 case TCPOPT_SACK_PERM
:
2409 if (opt
.opt_val
!= 0)
2412 tp
->rx_opt
.sack_ok
|= TCP_SACK_SEEN
;
2413 if (sysctl_tcp_fack
)
2414 tcp_enable_fack(tp
);
2416 case TCPOPT_TIMESTAMP
:
2417 if (opt
.opt_val
!= 0)
2420 tp
->rx_opt
.tstamp_ok
= 1;
2429 * Socket option code for TCP.
2431 static int do_tcp_setsockopt(struct sock
*sk
, int level
,
2432 int optname
, char __user
*optval
, unsigned int optlen
)
2434 struct tcp_sock
*tp
= tcp_sk(sk
);
2435 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2439 /* These are data/string values, all the others are ints */
2441 case TCP_CONGESTION
: {
2442 char name
[TCP_CA_NAME_MAX
];
2447 val
= strncpy_from_user(name
, optval
,
2448 min_t(long, TCP_CA_NAME_MAX
-1, optlen
));
2454 err
= tcp_set_congestion_control(sk
, name
);
2463 if (optlen
< sizeof(int))
2466 if (get_user(val
, (int __user
*)optval
))
2473 /* Values greater than interface MTU won't take effect. However
2474 * at the point when this call is done we typically don't yet
2475 * know which interface is going to be used */
2476 if (val
< TCP_MIN_MSS
|| val
> MAX_TCP_WINDOW
) {
2480 tp
->rx_opt
.user_mss
= val
;
2485 /* TCP_NODELAY is weaker than TCP_CORK, so that
2486 * this option on corked socket is remembered, but
2487 * it is not activated until cork is cleared.
2489 * However, when TCP_NODELAY is set we make
2490 * an explicit push, which overrides even TCP_CORK
2491 * for currently queued segments.
2493 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
2494 tcp_push_pending_frames(sk
);
2496 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
2500 case TCP_THIN_LINEAR_TIMEOUTS
:
2501 if (val
< 0 || val
> 1)
2507 case TCP_THIN_DUPACK
:
2508 if (val
< 0 || val
> 1)
2511 tp
->thin_dupack
= val
;
2512 if (tp
->thin_dupack
)
2513 tcp_disable_early_retrans(tp
);
2518 if (!tcp_can_repair_sock(sk
))
2520 else if (val
== 1) {
2522 sk
->sk_reuse
= SK_FORCE_REUSE
;
2523 tp
->repair_queue
= TCP_NO_QUEUE
;
2524 } else if (val
== 0) {
2526 sk
->sk_reuse
= SK_NO_REUSE
;
2527 tcp_send_window_probe(sk
);
2533 case TCP_REPAIR_QUEUE
:
2536 else if (val
< TCP_QUEUES_NR
)
2537 tp
->repair_queue
= val
;
2543 if (sk
->sk_state
!= TCP_CLOSE
)
2545 else if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2546 tp
->write_seq
= val
;
2547 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2553 case TCP_REPAIR_OPTIONS
:
2556 else if (sk
->sk_state
== TCP_ESTABLISHED
)
2557 err
= tcp_repair_options_est(tp
,
2558 (struct tcp_repair_opt __user
*)optval
,
2565 /* When set indicates to always queue non-full frames.
2566 * Later the user clears this option and we transmit
2567 * any pending partial frames in the queue. This is
2568 * meant to be used alongside sendfile() to get properly
2569 * filled frames when the user (for example) must write
2570 * out headers with a write() call first and then use
2571 * sendfile to send out the data parts.
2573 * TCP_CORK can be set together with TCP_NODELAY and it is
2574 * stronger than TCP_NODELAY.
2577 tp
->nonagle
|= TCP_NAGLE_CORK
;
2579 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
2580 if (tp
->nonagle
&TCP_NAGLE_OFF
)
2581 tp
->nonagle
|= TCP_NAGLE_PUSH
;
2582 tcp_push_pending_frames(sk
);
2587 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
2590 tp
->keepalive_time
= val
* HZ
;
2591 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
2592 !((1 << sk
->sk_state
) &
2593 (TCPF_CLOSE
| TCPF_LISTEN
))) {
2594 u32 elapsed
= keepalive_time_elapsed(tp
);
2595 if (tp
->keepalive_time
> elapsed
)
2596 elapsed
= tp
->keepalive_time
- elapsed
;
2599 inet_csk_reset_keepalive_timer(sk
, elapsed
);
2604 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
2607 tp
->keepalive_intvl
= val
* HZ
;
2610 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
2613 tp
->keepalive_probes
= val
;
2616 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
2619 icsk
->icsk_syn_retries
= val
;
2625 else if (val
> sysctl_tcp_fin_timeout
/ HZ
)
2628 tp
->linger2
= val
* HZ
;
2631 case TCP_DEFER_ACCEPT
:
2632 /* Translate value in seconds to number of retransmits */
2633 icsk
->icsk_accept_queue
.rskq_defer_accept
=
2634 secs_to_retrans(val
, TCP_TIMEOUT_INIT
/ HZ
,
2635 sysctl_tcp_rto_max
/ HZ
);
2638 case TCP_WINDOW_CLAMP
:
2640 if (sk
->sk_state
!= TCP_CLOSE
) {
2644 tp
->window_clamp
= 0;
2646 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
2647 SOCK_MIN_RCVBUF
/ 2 : val
;
2652 icsk
->icsk_ack
.pingpong
= 1;
2654 icsk
->icsk_ack
.pingpong
= 0;
2655 if ((1 << sk
->sk_state
) &
2656 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
2657 inet_csk_ack_scheduled(sk
)) {
2658 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
2659 tcp_cleanup_rbuf(sk
, 1);
2661 icsk
->icsk_ack
.pingpong
= 1;
2666 #ifdef CONFIG_TCP_MD5SIG
2668 /* Read the IP->Key mappings from userspace */
2669 err
= tp
->af_specific
->md5_parse(sk
, optval
, optlen
);
2672 case TCP_USER_TIMEOUT
:
2673 /* Cap the max timeout in ms TCP will retry/retrans
2674 * before giving up and aborting (ETIMEDOUT) a connection.
2679 icsk
->icsk_user_timeout
= msecs_to_jiffies(val
);
2683 if (val
>= 0 && ((1 << sk
->sk_state
) & (TCPF_CLOSE
|
2685 err
= fastopen_init_queue(sk
, val
);
2693 tp
->tsoffset
= val
- tcp_time_stamp
;
2704 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2705 unsigned int optlen
)
2707 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2709 if (level
!= SOL_TCP
)
2710 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
2712 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2714 EXPORT_SYMBOL(tcp_setsockopt
);
2716 #ifdef CONFIG_COMPAT
2717 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
2718 char __user
*optval
, unsigned int optlen
)
2720 if (level
!= SOL_TCP
)
2721 return inet_csk_compat_setsockopt(sk
, level
, optname
,
2723 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2725 EXPORT_SYMBOL(compat_tcp_setsockopt
);
2728 /* Return information about state of tcp endpoint in API format. */
2729 void tcp_get_info(const struct sock
*sk
, struct tcp_info
*info
)
2731 const struct tcp_sock
*tp
= tcp_sk(sk
);
2732 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2733 u32 now
= tcp_time_stamp
;
2735 memset(info
, 0, sizeof(*info
));
2737 info
->tcpi_state
= sk
->sk_state
;
2738 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
2739 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
2740 info
->tcpi_probes
= icsk
->icsk_probes_out
;
2741 info
->tcpi_backoff
= icsk
->icsk_backoff
;
2743 if (tp
->rx_opt
.tstamp_ok
)
2744 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
2745 if (tcp_is_sack(tp
))
2746 info
->tcpi_options
|= TCPI_OPT_SACK
;
2747 if (tp
->rx_opt
.wscale_ok
) {
2748 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
2749 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
2750 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
2753 if (tp
->ecn_flags
& TCP_ECN_OK
)
2754 info
->tcpi_options
|= TCPI_OPT_ECN
;
2755 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
2756 info
->tcpi_options
|= TCPI_OPT_ECN_SEEN
;
2757 if (tp
->syn_data_acked
)
2758 info
->tcpi_options
|= TCPI_OPT_SYN_DATA
;
2760 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
2761 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
2762 info
->tcpi_snd_mss
= tp
->mss_cache
;
2763 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
2765 if (sk
->sk_state
== TCP_LISTEN
) {
2766 info
->tcpi_unacked
= sk
->sk_ack_backlog
;
2767 info
->tcpi_sacked
= sk
->sk_max_ack_backlog
;
2769 info
->tcpi_unacked
= tp
->packets_out
;
2770 info
->tcpi_sacked
= tp
->sacked_out
;
2772 info
->tcpi_lost
= tp
->lost_out
;
2773 info
->tcpi_retrans
= tp
->retrans_out
;
2774 info
->tcpi_fackets
= tp
->fackets_out
;
2776 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
2777 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
2778 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
2780 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
2781 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
2782 info
->tcpi_rtt
= jiffies_to_usecs(tp
->srtt
)>>3;
2783 info
->tcpi_rttvar
= jiffies_to_usecs(tp
->mdev
)>>2;
2784 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
2785 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
2786 info
->tcpi_advmss
= tp
->advmss
;
2787 info
->tcpi_reordering
= tp
->reordering
;
2789 info
->tcpi_rcv_rtt
= jiffies_to_usecs(tp
->rcv_rtt_est
.rtt
)>>3;
2790 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
2792 info
->tcpi_total_retrans
= tp
->total_retrans
;
2794 EXPORT_SYMBOL_GPL(tcp_get_info
);
2796 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
2797 int optname
, char __user
*optval
, int __user
*optlen
)
2799 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2800 struct tcp_sock
*tp
= tcp_sk(sk
);
2803 if (get_user(len
, optlen
))
2806 len
= min_t(unsigned int, len
, sizeof(int));
2813 val
= tp
->mss_cache
;
2814 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
2815 val
= tp
->rx_opt
.user_mss
;
2817 val
= tp
->rx_opt
.mss_clamp
;
2820 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
2823 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
2826 val
= keepalive_time_when(tp
) / HZ
;
2829 val
= keepalive_intvl_when(tp
) / HZ
;
2832 val
= keepalive_probes(tp
);
2835 val
= icsk
->icsk_syn_retries
? : sysctl_tcp_syn_retries
;
2840 val
= (val
? : sysctl_tcp_fin_timeout
) / HZ
;
2842 case TCP_DEFER_ACCEPT
:
2843 val
= retrans_to_secs(icsk
->icsk_accept_queue
.rskq_defer_accept
,
2844 TCP_TIMEOUT_INIT
/ HZ
, sysctl_tcp_rto_max
/ HZ
);
2846 case TCP_WINDOW_CLAMP
:
2847 val
= tp
->window_clamp
;
2850 struct tcp_info info
;
2852 if (get_user(len
, optlen
))
2855 tcp_get_info(sk
, &info
);
2857 len
= min_t(unsigned int, len
, sizeof(info
));
2858 if (put_user(len
, optlen
))
2860 if (copy_to_user(optval
, &info
, len
))
2865 val
= !icsk
->icsk_ack
.pingpong
;
2868 case TCP_CONGESTION
:
2869 if (get_user(len
, optlen
))
2871 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
2872 if (put_user(len
, optlen
))
2874 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
2878 case TCP_THIN_LINEAR_TIMEOUTS
:
2881 case TCP_THIN_DUPACK
:
2882 val
= tp
->thin_dupack
;
2889 case TCP_REPAIR_QUEUE
:
2891 val
= tp
->repair_queue
;
2897 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2898 val
= tp
->write_seq
;
2899 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2905 case TCP_USER_TIMEOUT
:
2906 val
= jiffies_to_msecs(icsk
->icsk_user_timeout
);
2909 val
= tcp_time_stamp
+ tp
->tsoffset
;
2912 return -ENOPROTOOPT
;
2915 if (put_user(len
, optlen
))
2917 if (copy_to_user(optval
, &val
, len
))
2922 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2925 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2927 if (level
!= SOL_TCP
)
2928 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
2930 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2932 EXPORT_SYMBOL(tcp_getsockopt
);
2934 #ifdef CONFIG_COMPAT
2935 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
2936 char __user
*optval
, int __user
*optlen
)
2938 if (level
!= SOL_TCP
)
2939 return inet_csk_compat_getsockopt(sk
, level
, optname
,
2941 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2943 EXPORT_SYMBOL(compat_tcp_getsockopt
);
2946 struct sk_buff
*tcp_tso_segment(struct sk_buff
*skb
,
2947 netdev_features_t features
)
2949 struct sk_buff
*segs
= ERR_PTR(-EINVAL
);
2950 unsigned int sum_truesize
= 0;
2955 unsigned int oldlen
;
2957 struct sk_buff
*gso_skb
= skb
;
2959 bool ooo_okay
, copy_destructor
;
2961 if (!pskb_may_pull(skb
, sizeof(*th
)))
2965 thlen
= th
->doff
* 4;
2966 if (thlen
< sizeof(*th
))
2969 if (!pskb_may_pull(skb
, thlen
))
2972 oldlen
= (u16
)~skb
->len
;
2973 __skb_pull(skb
, thlen
);
2975 mss
= skb_shinfo(skb
)->gso_size
;
2976 if (unlikely(skb
->len
<= mss
))
2979 if (skb_gso_ok(skb
, features
| NETIF_F_GSO_ROBUST
)) {
2980 /* Packet is from an untrusted source, reset gso_segs. */
2981 int type
= skb_shinfo(skb
)->gso_type
;
2989 SKB_GSO_UDP_TUNNEL
|
2991 !(type
& (SKB_GSO_TCPV4
| SKB_GSO_TCPV6
))))
2994 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss
);
3000 copy_destructor
= gso_skb
->destructor
== tcp_wfree
;
3001 ooo_okay
= gso_skb
->ooo_okay
;
3002 /* All segments but the first should have ooo_okay cleared */
3005 segs
= skb_segment(skb
, features
);
3009 /* Only first segment might have ooo_okay set */
3010 segs
->ooo_okay
= ooo_okay
;
3012 delta
= htonl(oldlen
+ (thlen
+ mss
));
3016 seq
= ntohl(th
->seq
);
3018 newcheck
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
3019 (__force u32
)delta
));
3022 th
->fin
= th
->psh
= 0;
3023 th
->check
= newcheck
;
3025 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
3027 csum_fold(csum_partial(skb_transport_header(skb
),
3031 if (copy_destructor
) {
3032 skb
->destructor
= gso_skb
->destructor
;
3033 skb
->sk
= gso_skb
->sk
;
3034 sum_truesize
+= skb
->truesize
;
3039 th
->seq
= htonl(seq
);
3041 } while (skb
->next
);
3043 /* Following permits TCP Small Queues to work well with GSO :
3044 * The callback to TCP stack will be called at the time last frag
3045 * is freed at TX completion, and not right now when gso_skb
3046 * is freed by GSO engine
3048 if (copy_destructor
) {
3049 swap(gso_skb
->sk
, skb
->sk
);
3050 swap(gso_skb
->destructor
, skb
->destructor
);
3051 sum_truesize
+= skb
->truesize
;
3052 atomic_add(sum_truesize
- gso_skb
->truesize
,
3053 &skb
->sk
->sk_wmem_alloc
);
3056 delta
= htonl(oldlen
+ (skb
->tail
- skb
->transport_header
) +
3058 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
3059 (__force u32
)delta
));
3060 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
3061 th
->check
= csum_fold(csum_partial(skb_transport_header(skb
),
3067 EXPORT_SYMBOL(tcp_tso_segment
);
3069 struct sk_buff
**tcp_gro_receive(struct sk_buff
**head
, struct sk_buff
*skb
)
3071 struct sk_buff
**pp
= NULL
;
3078 unsigned int mss
= 1;
3084 off
= skb_gro_offset(skb
);
3085 hlen
= off
+ sizeof(*th
);
3086 th
= skb_gro_header_fast(skb
, off
);
3087 if (skb_gro_header_hard(skb
, hlen
)) {
3088 th
= skb_gro_header_slow(skb
, hlen
, off
);
3093 thlen
= th
->doff
* 4;
3094 if (thlen
< sizeof(*th
))
3098 if (skb_gro_header_hard(skb
, hlen
)) {
3099 th
= skb_gro_header_slow(skb
, hlen
, off
);
3104 skb_gro_pull(skb
, thlen
);
3106 len
= skb_gro_len(skb
);
3107 flags
= tcp_flag_word(th
);
3109 for (; (p
= *head
); head
= &p
->next
) {
3110 if (!NAPI_GRO_CB(p
)->same_flow
)
3115 if (*(u32
*)&th
->source
^ *(u32
*)&th2
->source
) {
3116 NAPI_GRO_CB(p
)->same_flow
= 0;
3123 goto out_check_final
;
3126 flush
= NAPI_GRO_CB(p
)->flush
;
3127 flush
|= (__force
int)(flags
& TCP_FLAG_CWR
);
3128 flush
|= (__force
int)((flags
^ tcp_flag_word(th2
)) &
3129 ~(TCP_FLAG_CWR
| TCP_FLAG_FIN
| TCP_FLAG_PSH
));
3130 flush
|= (__force
int)(th
->ack_seq
^ th2
->ack_seq
);
3131 for (i
= sizeof(*th
); i
< thlen
; i
+= 4)
3132 flush
|= *(u32
*)((u8
*)th
+ i
) ^
3133 *(u32
*)((u8
*)th2
+ i
);
3135 mss
= skb_shinfo(p
)->gso_size
;
3137 flush
|= (len
- 1) >= mss
;
3138 flush
|= (ntohl(th2
->seq
) + skb_gro_len(p
)) ^ ntohl(th
->seq
);
3140 if (flush
|| skb_gro_receive(head
, skb
)) {
3142 goto out_check_final
;
3147 tcp_flag_word(th2
) |= flags
& (TCP_FLAG_FIN
| TCP_FLAG_PSH
);
3151 flush
|= (__force
int)(flags
& (TCP_FLAG_URG
| TCP_FLAG_PSH
|
3152 TCP_FLAG_RST
| TCP_FLAG_SYN
|
3155 if (p
&& (!NAPI_GRO_CB(skb
)->same_flow
|| flush
))
3159 NAPI_GRO_CB(skb
)->flush
|= flush
;
3163 EXPORT_SYMBOL(tcp_gro_receive
);
3165 int tcp_gro_complete(struct sk_buff
*skb
)
3167 struct tcphdr
*th
= tcp_hdr(skb
);
3169 skb
->csum_start
= skb_transport_header(skb
) - skb
->head
;
3170 skb
->csum_offset
= offsetof(struct tcphdr
, check
);
3171 skb
->ip_summed
= CHECKSUM_PARTIAL
;
3173 skb_shinfo(skb
)->gso_segs
= NAPI_GRO_CB(skb
)->count
;
3176 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
3180 EXPORT_SYMBOL(tcp_gro_complete
);
3182 #ifdef CONFIG_TCP_MD5SIG
3183 static unsigned long tcp_md5sig_users
;
3184 static struct tcp_md5sig_pool __percpu
*tcp_md5sig_pool
;
3185 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock
);
3187 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu
*pool
)
3191 for_each_possible_cpu(cpu
) {
3192 struct tcp_md5sig_pool
*p
= per_cpu_ptr(pool
, cpu
);
3194 if (p
->md5_desc
.tfm
)
3195 crypto_free_hash(p
->md5_desc
.tfm
);
3200 void tcp_free_md5sig_pool(void)
3202 struct tcp_md5sig_pool __percpu
*pool
= NULL
;
3204 spin_lock_bh(&tcp_md5sig_pool_lock
);
3205 if (--tcp_md5sig_users
== 0) {
3206 pool
= tcp_md5sig_pool
;
3207 tcp_md5sig_pool
= NULL
;
3209 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3211 __tcp_free_md5sig_pool(pool
);
3213 EXPORT_SYMBOL(tcp_free_md5sig_pool
);
3215 static struct tcp_md5sig_pool __percpu
*
3216 __tcp_alloc_md5sig_pool(struct sock
*sk
)
3219 struct tcp_md5sig_pool __percpu
*pool
;
3221 pool
= alloc_percpu(struct tcp_md5sig_pool
);
3225 for_each_possible_cpu(cpu
) {
3226 struct crypto_hash
*hash
;
3228 hash
= crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC
);
3229 if (IS_ERR_OR_NULL(hash
))
3232 per_cpu_ptr(pool
, cpu
)->md5_desc
.tfm
= hash
;
3236 __tcp_free_md5sig_pool(pool
);
3240 struct tcp_md5sig_pool __percpu
*tcp_alloc_md5sig_pool(struct sock
*sk
)
3242 struct tcp_md5sig_pool __percpu
*pool
;
3246 spin_lock_bh(&tcp_md5sig_pool_lock
);
3247 pool
= tcp_md5sig_pool
;
3248 if (tcp_md5sig_users
++ == 0) {
3250 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3253 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3257 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3260 /* we cannot hold spinlock here because this may sleep. */
3261 struct tcp_md5sig_pool __percpu
*p
;
3263 p
= __tcp_alloc_md5sig_pool(sk
);
3264 spin_lock_bh(&tcp_md5sig_pool_lock
);
3267 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3270 pool
= tcp_md5sig_pool
;
3272 /* oops, it has already been assigned. */
3273 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3274 __tcp_free_md5sig_pool(p
);
3276 tcp_md5sig_pool
= pool
= p
;
3277 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3282 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
3286 * tcp_get_md5sig_pool - get md5sig_pool for this user
3288 * We use percpu structure, so if we succeed, we exit with preemption
3289 * and BH disabled, to make sure another thread or softirq handling
3290 * wont try to get same context.
3292 struct tcp_md5sig_pool
*tcp_get_md5sig_pool(void)
3294 struct tcp_md5sig_pool __percpu
*p
;
3298 spin_lock(&tcp_md5sig_pool_lock
);
3299 p
= tcp_md5sig_pool
;
3302 spin_unlock(&tcp_md5sig_pool_lock
);
3305 return this_cpu_ptr(p
);
3310 EXPORT_SYMBOL(tcp_get_md5sig_pool
);
3312 void tcp_put_md5sig_pool(void)
3315 tcp_free_md5sig_pool();
3317 EXPORT_SYMBOL(tcp_put_md5sig_pool
);
3319 int tcp_md5_hash_header(struct tcp_md5sig_pool
*hp
,
3320 const struct tcphdr
*th
)
3322 struct scatterlist sg
;
3326 /* We are not allowed to change tcphdr, make a local copy */
3327 memcpy(&hdr
, th
, sizeof(hdr
));
3330 /* options aren't included in the hash */
3331 sg_init_one(&sg
, &hdr
, sizeof(hdr
));
3332 err
= crypto_hash_update(&hp
->md5_desc
, &sg
, sizeof(hdr
));
3335 EXPORT_SYMBOL(tcp_md5_hash_header
);
3337 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool
*hp
,
3338 const struct sk_buff
*skb
, unsigned int header_len
)
3340 struct scatterlist sg
;
3341 const struct tcphdr
*tp
= tcp_hdr(skb
);
3342 struct hash_desc
*desc
= &hp
->md5_desc
;
3344 const unsigned int head_data_len
= skb_headlen(skb
) > header_len
?
3345 skb_headlen(skb
) - header_len
: 0;
3346 const struct skb_shared_info
*shi
= skb_shinfo(skb
);
3347 struct sk_buff
*frag_iter
;
3349 sg_init_table(&sg
, 1);
3351 sg_set_buf(&sg
, ((u8
*) tp
) + header_len
, head_data_len
);
3352 if (crypto_hash_update(desc
, &sg
, head_data_len
))
3355 for (i
= 0; i
< shi
->nr_frags
; ++i
) {
3356 const struct skb_frag_struct
*f
= &shi
->frags
[i
];
3357 unsigned int offset
= f
->page_offset
;
3358 struct page
*page
= skb_frag_page(f
) + (offset
>> PAGE_SHIFT
);
3360 sg_set_page(&sg
, page
, skb_frag_size(f
),
3361 offset_in_page(offset
));
3362 if (crypto_hash_update(desc
, &sg
, skb_frag_size(f
)))
3366 skb_walk_frags(skb
, frag_iter
)
3367 if (tcp_md5_hash_skb_data(hp
, frag_iter
, 0))
3372 EXPORT_SYMBOL(tcp_md5_hash_skb_data
);
3374 int tcp_md5_hash_key(struct tcp_md5sig_pool
*hp
, const struct tcp_md5sig_key
*key
)
3376 struct scatterlist sg
;
3378 sg_init_one(&sg
, key
->key
, key
->keylen
);
3379 return crypto_hash_update(&hp
->md5_desc
, &sg
, key
->keylen
);
3381 EXPORT_SYMBOL(tcp_md5_hash_key
);
3385 void tcp_done(struct sock
*sk
)
3387 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
3389 if (sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
3390 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_ATTEMPTFAILS
);
3392 tcp_set_state(sk
, TCP_CLOSE
);
3393 tcp_clear_xmit_timers(sk
);
3395 reqsk_fastopen_remove(sk
, req
, false);
3397 sk
->sk_shutdown
= SHUTDOWN_MASK
;
3399 if (!sock_flag(sk
, SOCK_DEAD
))
3400 sk
->sk_state_change(sk
);
3402 inet_csk_destroy_sock(sk
);
3404 EXPORT_SYMBOL_GPL(tcp_done
);
3406 extern struct tcp_congestion_ops tcp_reno
;
3408 static __initdata
unsigned long thash_entries
;
3409 static int __init
set_thash_entries(char *str
)
3416 ret
= kstrtoul(str
, 0, &thash_entries
);
3422 __setup("thash_entries=", set_thash_entries
);
3424 void tcp_init_mem(struct net
*net
)
3426 unsigned long limit
= nr_free_buffer_pages() / 8;
3427 limit
= max(limit
, 128UL);
3428 net
->ipv4
.sysctl_tcp_mem
[0] = limit
/ 4 * 3;
3429 net
->ipv4
.sysctl_tcp_mem
[1] = limit
;
3430 net
->ipv4
.sysctl_tcp_mem
[2] = net
->ipv4
.sysctl_tcp_mem
[0] * 2;
3433 void __init
tcp_init(void)
3435 struct sk_buff
*skb
= NULL
;
3436 unsigned long limit
;
3437 int max_rshare
, max_wshare
, cnt
;
3440 BUILD_BUG_ON(sizeof(struct tcp_skb_cb
) > sizeof(skb
->cb
));
3442 percpu_counter_init(&tcp_sockets_allocated
, 0);
3443 percpu_counter_init(&tcp_orphan_count
, 0);
3444 tcp_hashinfo
.bind_bucket_cachep
=
3445 kmem_cache_create("tcp_bind_bucket",
3446 sizeof(struct inet_bind_bucket
), 0,
3447 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
);
3449 /* Size and allocate the main established and bind bucket
3452 * The methodology is similar to that of the buffer cache.
3454 tcp_hashinfo
.ehash
=
3455 alloc_large_system_hash("TCP established",
3456 sizeof(struct inet_ehash_bucket
),
3458 17, /* one slot per 128 KB of memory */
3461 &tcp_hashinfo
.ehash_mask
,
3463 thash_entries
? 0 : 512 * 1024);
3464 for (i
= 0; i
<= tcp_hashinfo
.ehash_mask
; i
++) {
3465 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].chain
, i
);
3466 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].twchain
, i
);
3468 if (inet_ehash_locks_alloc(&tcp_hashinfo
))
3469 panic("TCP: failed to alloc ehash_locks");
3470 tcp_hashinfo
.bhash
=
3471 alloc_large_system_hash("TCP bind",
3472 sizeof(struct inet_bind_hashbucket
),
3473 tcp_hashinfo
.ehash_mask
+ 1,
3474 17, /* one slot per 128 KB of memory */
3476 &tcp_hashinfo
.bhash_size
,
3480 tcp_hashinfo
.bhash_size
= 1U << tcp_hashinfo
.bhash_size
;
3481 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
3482 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
3483 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
3487 cnt
= tcp_hashinfo
.ehash_mask
+ 1;
3489 tcp_death_row
.sysctl_max_tw_buckets
= cnt
/ 2;
3490 sysctl_tcp_max_orphans
= cnt
/ 2;
3491 sysctl_max_syn_backlog
= max(128, cnt
/ 256);
3493 tcp_init_mem(&init_net
);
3494 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3495 limit
= nr_free_buffer_pages() << (PAGE_SHIFT
- 7);
3496 max_wshare
= min(4UL*1024*1024, limit
);
3497 max_rshare
= min(6UL*1024*1024, limit
);
3499 sysctl_tcp_wmem
[0] = SK_MEM_QUANTUM
;
3500 sysctl_tcp_wmem
[1] = 16*1024;
3501 sysctl_tcp_wmem
[2] = max(64*1024, max_wshare
);
3503 sysctl_tcp_rmem
[0] = SK_MEM_QUANTUM
;
3504 sysctl_tcp_rmem
[1] = 87380;
3505 sysctl_tcp_rmem
[2] = max(87380, max_rshare
);
3507 pr_info("Hash tables configured (established %u bind %u)\n",
3508 tcp_hashinfo
.ehash_mask
+ 1, tcp_hashinfo
.bhash_size
);
3512 tcp_register_congestion_control(&tcp_reno
);
3517 static int tcp_is_local(struct net
*net
, __be32 addr
) {
3519 struct flowi4 fl4
= { .daddr
= addr
};
3520 rt
= ip_route_output_key(net
, &fl4
);
3521 if (IS_ERR_OR_NULL(rt
))
3523 return rt
->dst
.dev
&& (rt
->dst
.dev
->flags
& IFF_LOOPBACK
);
3526 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3527 static int tcp_is_local6(struct net
*net
, struct in6_addr
*addr
) {
3528 struct rt6_info
*rt6
= rt6_lookup(net
, addr
, addr
, 0, 0);
3529 return rt6
&& rt6
->dst
.dev
&& (rt6
->dst
.dev
->flags
& IFF_LOOPBACK
);
3534 * tcp_nuke_addr - destroy all sockets on the given local address
3535 * if local address is the unspecified address (0.0.0.0 or ::), destroy all
3536 * sockets with local addresses that are not configured.
3538 int tcp_nuke_addr(struct net
*net
, struct sockaddr
*addr
)
3540 int family
= addr
->sa_family
;
3541 unsigned int bucket
;
3544 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3545 struct in6_addr
*in6
= NULL
;
3547 if (family
== AF_INET
) {
3548 in
= &((struct sockaddr_in
*)addr
)->sin_addr
;
3549 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3550 } else if (family
== AF_INET6
) {
3551 in6
= &((struct sockaddr_in6
*)addr
)->sin6_addr
;
3554 return -EAFNOSUPPORT
;
3557 for (bucket
= 0; bucket
< tcp_hashinfo
.ehash_mask
; bucket
++) {
3558 struct hlist_nulls_node
*node
;
3560 spinlock_t
*lock
= inet_ehash_lockp(&tcp_hashinfo
, bucket
);
3564 sk_nulls_for_each(sk
, node
, &tcp_hashinfo
.ehash
[bucket
].chain
) {
3565 struct inet_sock
*inet
= inet_sk(sk
);
3567 if (sysctl_ip_dynaddr
&& sk
->sk_state
== TCP_SYN_SENT
)
3569 if (sock_flag(sk
, SOCK_DEAD
))
3572 if (family
== AF_INET
) {
3573 __be32 s4
= inet
->inet_rcv_saddr
;
3574 if (s4
== LOOPBACK4_IPV6
)
3577 if (in
->s_addr
!= s4
&&
3578 !(in
->s_addr
== INADDR_ANY
&&
3579 !tcp_is_local(net
, s4
)))
3583 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3584 if (family
== AF_INET6
) {
3585 struct in6_addr
*s6
;
3589 s6
= &inet
->pinet6
->rcv_saddr
;
3590 if (ipv6_addr_type(s6
) == IPV6_ADDR_MAPPED
)
3593 if (!ipv6_addr_equal(in6
, s6
) &&
3594 !(ipv6_addr_equal(in6
, &in6addr_any
) &&
3595 !tcp_is_local6(net
, s6
)))
3601 spin_unlock_bh(lock
);
3605 sk
->sk_err
= ETIMEDOUT
;
3606 sk
->sk_error_report(sk
);
3615 spin_unlock_bh(lock
);