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>
272 #include <net/icmp.h>
273 #include <net/inet_common.h>
275 #include <net/xfrm.h>
277 #include <net/netdma.h>
278 #include <net/sock.h>
280 #include <asm/uaccess.h>
281 #include <asm/ioctls.h>
283 int sysctl_tcp_fin_timeout __read_mostly
= TCP_FIN_TIMEOUT
;
285 struct percpu_counter tcp_orphan_count
;
286 EXPORT_SYMBOL_GPL(tcp_orphan_count
);
288 int sysctl_tcp_wmem
[3] __read_mostly
;
289 int sysctl_tcp_rmem
[3] __read_mostly
;
291 EXPORT_SYMBOL(sysctl_tcp_rmem
);
292 EXPORT_SYMBOL(sysctl_tcp_wmem
);
294 atomic_long_t tcp_memory_allocated
; /* Current allocated memory. */
295 EXPORT_SYMBOL(tcp_memory_allocated
);
298 * Current number of TCP sockets.
300 struct percpu_counter tcp_sockets_allocated
;
301 EXPORT_SYMBOL(tcp_sockets_allocated
);
306 struct tcp_splice_state
{
307 struct pipe_inode_info
*pipe
;
313 * Pressure flag: try to collapse.
314 * Technical note: it is used by multiple contexts non atomically.
315 * All the __sk_mem_schedule() is of this nature: accounting
316 * is strict, actions are advisory and have some latency.
318 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;
328 EXPORT_SYMBOL(tcp_enter_memory_pressure
);
330 /* Convert seconds to retransmits based on initial and max timeout */
331 static u8
secs_to_retrans(int seconds
, int timeout
, int rto_max
)
336 int period
= timeout
;
339 while (seconds
> period
&& res
< 255) {
342 if (timeout
> rto_max
)
350 /* Convert retransmits to seconds based on initial and max timeout */
351 static int retrans_to_secs(u8 retrans
, int timeout
, int rto_max
)
359 if (timeout
> rto_max
)
367 /* Address-family independent initialization for a tcp_sock.
369 * NOTE: A lot of things set to zero explicitly by call to
370 * sk_alloc() so need not be done here.
372 void tcp_init_sock(struct sock
*sk
)
374 struct inet_connection_sock
*icsk
= inet_csk(sk
);
375 struct tcp_sock
*tp
= tcp_sk(sk
);
377 skb_queue_head_init(&tp
->out_of_order_queue
);
378 tcp_init_xmit_timers(sk
);
379 tcp_prequeue_init(tp
);
380 INIT_LIST_HEAD(&tp
->tsq_node
);
382 icsk
->icsk_rto
= TCP_TIMEOUT_INIT
;
383 tp
->mdev
= TCP_TIMEOUT_INIT
;
385 /* So many TCP implementations out there (incorrectly) count the
386 * initial SYN frame in their delayed-ACK and congestion control
387 * algorithms that we must have the following bandaid to talk
388 * efficiently to them. -DaveM
390 tp
->snd_cwnd
= TCP_INIT_CWND
;
392 /* See draft-stevens-tcpca-spec-01 for discussion of the
393 * initialization of these values.
395 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
396 tp
->snd_cwnd_clamp
= ~0;
397 tp
->mss_cache
= TCP_MSS_DEFAULT
;
399 tp
->reordering
= sysctl_tcp_reordering
;
400 tcp_enable_early_retrans(tp
);
401 icsk
->icsk_ca_ops
= &tcp_init_congestion_ops
;
403 sk
->sk_state
= TCP_CLOSE
;
405 sk
->sk_write_space
= sk_stream_write_space
;
406 sock_set_flag(sk
, SOCK_USE_WRITE_QUEUE
);
408 icsk
->icsk_sync_mss
= tcp_sync_mss
;
410 /* TCP Cookie Transactions */
411 if (sysctl_tcp_cookie_size
> 0) {
412 /* Default, cookies without s_data_payload. */
414 kzalloc(sizeof(*tp
->cookie_values
),
416 if (tp
->cookie_values
!= NULL
)
417 kref_init(&tp
->cookie_values
->kref
);
419 /* Presumed zeroed, in order of appearance:
420 * cookie_in_always, cookie_out_never,
421 * s_data_constant, s_data_in, s_data_out
423 sk
->sk_sndbuf
= sysctl_tcp_wmem
[1];
424 sk
->sk_rcvbuf
= sysctl_tcp_rmem
[1];
427 sock_update_memcg(sk
);
428 sk_sockets_allocated_inc(sk
);
431 EXPORT_SYMBOL(tcp_init_sock
);
434 * Wait for a TCP event.
436 * Note that we don't need to lock the socket, as the upper poll layers
437 * take care of normal races (between the test and the event) and we don't
438 * go look at any of the socket buffers directly.
440 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
443 struct sock
*sk
= sock
->sk
;
444 const struct tcp_sock
*tp
= tcp_sk(sk
);
446 sock_poll_wait(file
, sk_sleep(sk
), wait
);
447 if (sk
->sk_state
== TCP_LISTEN
)
448 return inet_csk_listen_poll(sk
);
450 /* Socket is not locked. We are protected from async events
451 * by poll logic and correct handling of state changes
452 * made by other threads is impossible in any case.
458 * POLLHUP is certainly not done right. But poll() doesn't
459 * have a notion of HUP in just one direction, and for a
460 * socket the read side is more interesting.
462 * Some poll() documentation says that POLLHUP is incompatible
463 * with the POLLOUT/POLLWR flags, so somebody should check this
464 * all. But careful, it tends to be safer to return too many
465 * bits than too few, and you can easily break real applications
466 * if you don't tell them that something has hung up!
470 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
471 * our fs/select.c). It means that after we received EOF,
472 * poll always returns immediately, making impossible poll() on write()
473 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
474 * if and only if shutdown has been made in both directions.
475 * Actually, it is interesting to look how Solaris and DUX
476 * solve this dilemma. I would prefer, if POLLHUP were maskable,
477 * then we could set it on SND_SHUTDOWN. BTW examples given
478 * in Stevens' books assume exactly this behaviour, it explains
479 * why POLLHUP is incompatible with POLLOUT. --ANK
481 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
482 * blocking on fresh not-connected or disconnected socket. --ANK
484 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| sk
->sk_state
== TCP_CLOSE
)
486 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
487 mask
|= POLLIN
| POLLRDNORM
| POLLRDHUP
;
489 /* Connected or passive Fast Open socket? */
490 if (sk
->sk_state
!= TCP_SYN_SENT
&&
491 (sk
->sk_state
!= TCP_SYN_RECV
|| tp
->fastopen_rsk
!= NULL
)) {
492 int target
= sock_rcvlowat(sk
, 0, INT_MAX
);
494 if (tp
->urg_seq
== tp
->copied_seq
&&
495 !sock_flag(sk
, SOCK_URGINLINE
) &&
499 /* Potential race condition. If read of tp below will
500 * escape above sk->sk_state, we can be illegally awaken
501 * in SYN_* states. */
502 if (tp
->rcv_nxt
- tp
->copied_seq
>= target
)
503 mask
|= POLLIN
| POLLRDNORM
;
505 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
506 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
)) {
507 mask
|= POLLOUT
| POLLWRNORM
;
508 } else { /* send SIGIO later */
509 set_bit(SOCK_ASYNC_NOSPACE
,
510 &sk
->sk_socket
->flags
);
511 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
513 /* Race breaker. If space is freed after
514 * wspace test but before the flags are set,
515 * IO signal will be lost.
517 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
))
518 mask
|= POLLOUT
| POLLWRNORM
;
521 mask
|= POLLOUT
| POLLWRNORM
;
523 if (tp
->urg_data
& TCP_URG_VALID
)
526 /* This barrier is coupled with smp_wmb() in tcp_reset() */
533 EXPORT_SYMBOL(tcp_poll
);
535 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
537 struct tcp_sock
*tp
= tcp_sk(sk
);
543 if (sk
->sk_state
== TCP_LISTEN
)
546 slow
= lock_sock_fast(sk
);
547 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
549 else if (sock_flag(sk
, SOCK_URGINLINE
) ||
551 before(tp
->urg_seq
, tp
->copied_seq
) ||
552 !before(tp
->urg_seq
, tp
->rcv_nxt
)) {
554 answ
= tp
->rcv_nxt
- tp
->copied_seq
;
556 /* Subtract 1, if FIN was received */
557 if (answ
&& sock_flag(sk
, SOCK_DONE
))
560 answ
= tp
->urg_seq
- tp
->copied_seq
;
561 unlock_sock_fast(sk
, slow
);
564 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
567 if (sk
->sk_state
== TCP_LISTEN
)
570 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
573 answ
= tp
->write_seq
- tp
->snd_una
;
576 if (sk
->sk_state
== TCP_LISTEN
)
579 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
582 answ
= tp
->write_seq
- tp
->snd_nxt
;
588 return put_user(answ
, (int __user
*)arg
);
590 EXPORT_SYMBOL(tcp_ioctl
);
592 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
594 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
595 tp
->pushed_seq
= tp
->write_seq
;
598 static inline bool forced_push(const struct tcp_sock
*tp
)
600 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
603 static inline void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
605 struct tcp_sock
*tp
= tcp_sk(sk
);
606 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
609 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
610 tcb
->tcp_flags
= TCPHDR_ACK
;
612 skb_header_release(skb
);
613 tcp_add_write_queue_tail(sk
, skb
);
614 sk
->sk_wmem_queued
+= skb
->truesize
;
615 sk_mem_charge(sk
, skb
->truesize
);
616 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
617 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
620 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
)
623 tp
->snd_up
= tp
->write_seq
;
626 static inline void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
629 if (tcp_send_head(sk
)) {
630 struct tcp_sock
*tp
= tcp_sk(sk
);
632 if (!(flags
& MSG_MORE
) || forced_push(tp
))
633 tcp_mark_push(tp
, tcp_write_queue_tail(sk
));
635 tcp_mark_urg(tp
, flags
);
636 __tcp_push_pending_frames(sk
, mss_now
,
637 (flags
& MSG_MORE
) ? TCP_NAGLE_CORK
: nonagle
);
641 static int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
642 unsigned int offset
, size_t len
)
644 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
647 ret
= skb_splice_bits(skb
, offset
, tss
->pipe
, min(rd_desc
->count
, len
),
650 rd_desc
->count
-= ret
;
654 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
656 /* Store TCP splice context information in read_descriptor_t. */
657 read_descriptor_t rd_desc
= {
662 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
666 * tcp_splice_read - splice data from TCP socket to a pipe
667 * @sock: socket to splice from
668 * @ppos: position (not valid)
669 * @pipe: pipe to splice to
670 * @len: number of bytes to splice
671 * @flags: splice modifier flags
674 * Will read pages from given socket and fill them into a pipe.
677 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
678 struct pipe_inode_info
*pipe
, size_t len
,
681 struct sock
*sk
= sock
->sk
;
682 struct tcp_splice_state tss
= {
691 sock_rps_record_flow(sk
);
693 * We can't seek on a socket input
702 timeo
= sock_rcvtimeo(sk
, sock
->file
->f_flags
& O_NONBLOCK
);
704 ret
= __tcp_splice_read(sk
, &tss
);
710 if (sock_flag(sk
, SOCK_DONE
))
713 ret
= sock_error(sk
);
716 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
718 if (sk
->sk_state
== TCP_CLOSE
) {
720 * This occurs when user tries to read
721 * from never connected socket.
723 if (!sock_flag(sk
, SOCK_DONE
))
731 sk_wait_data(sk
, &timeo
);
732 if (signal_pending(current
)) {
733 ret
= sock_intr_errno(timeo
);
746 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
747 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
748 signal_pending(current
))
759 EXPORT_SYMBOL(tcp_splice_read
);
761 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
)
765 /* The TCP header must be at least 32-bit aligned. */
766 size
= ALIGN(size
, 4);
768 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
770 if (sk_wmem_schedule(sk
, skb
->truesize
)) {
771 skb_reserve(skb
, sk
->sk_prot
->max_header
);
773 * Make sure that we have exactly size bytes
774 * available to the caller, no more, no less.
776 skb
->avail_size
= size
;
781 sk
->sk_prot
->enter_memory_pressure(sk
);
782 sk_stream_moderate_sndbuf(sk
);
787 static unsigned int tcp_xmit_size_goal(struct sock
*sk
, u32 mss_now
,
790 struct tcp_sock
*tp
= tcp_sk(sk
);
791 u32 xmit_size_goal
, old_size_goal
;
793 xmit_size_goal
= mss_now
;
795 if (large_allowed
&& sk_can_gso(sk
)) {
796 xmit_size_goal
= ((sk
->sk_gso_max_size
- 1) -
797 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
798 inet_csk(sk
)->icsk_ext_hdr_len
-
801 /* TSQ : try to have two TSO segments in flight */
802 xmit_size_goal
= min_t(u32
, xmit_size_goal
,
803 sysctl_tcp_limit_output_bytes
>> 1);
805 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
807 /* We try hard to avoid divides here */
808 old_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
810 if (likely(old_size_goal
<= xmit_size_goal
&&
811 old_size_goal
+ mss_now
> xmit_size_goal
)) {
812 xmit_size_goal
= old_size_goal
;
814 tp
->xmit_size_goal_segs
=
815 min_t(u16
, xmit_size_goal
/ mss_now
,
816 sk
->sk_gso_max_segs
);
817 xmit_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
821 return max(xmit_size_goal
, mss_now
);
824 static int tcp_send_mss(struct sock
*sk
, int *size_goal
, int flags
)
828 mss_now
= tcp_current_mss(sk
);
829 *size_goal
= tcp_xmit_size_goal(sk
, mss_now
, !(flags
& MSG_OOB
));
834 static ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
*page
, int offset
,
835 size_t size
, int flags
)
837 struct tcp_sock
*tp
= tcp_sk(sk
);
838 int mss_now
, size_goal
;
841 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
843 /* Wait for a connection to finish. One exception is TCP Fast Open
844 * (passive side) where data is allowed to be sent before a connection
845 * is fully established.
847 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
848 !tcp_passive_fastopen(sk
)) {
849 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
853 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
855 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
859 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
863 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
867 if (!tcp_send_head(sk
) || (copy
= size_goal
- skb
->len
) <= 0) {
869 if (!sk_stream_memory_free(sk
))
870 goto wait_for_sndbuf
;
872 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
874 goto wait_for_memory
;
883 i
= skb_shinfo(skb
)->nr_frags
;
884 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
885 if (!can_coalesce
&& i
>= MAX_SKB_FRAGS
) {
886 tcp_mark_push(tp
, skb
);
889 if (!sk_wmem_schedule(sk
, copy
))
890 goto wait_for_memory
;
893 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
896 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
899 skb_shinfo(skb
)->gso_type
|= SKB_GSO_SHARED_FRAG
;
902 skb
->data_len
+= copy
;
903 skb
->truesize
+= copy
;
904 sk
->sk_wmem_queued
+= copy
;
905 sk_mem_charge(sk
, copy
);
906 skb
->ip_summed
= CHECKSUM_PARTIAL
;
907 tp
->write_seq
+= copy
;
908 TCP_SKB_CB(skb
)->end_seq
+= copy
;
909 skb_shinfo(skb
)->gso_segs
= 0;
912 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
919 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
922 if (forced_push(tp
)) {
923 tcp_mark_push(tp
, skb
);
924 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
925 } else if (skb
== tcp_send_head(sk
))
926 tcp_push_one(sk
, mss_now
);
930 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
932 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
934 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
937 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
941 if (copied
&& !(flags
& MSG_SENDPAGE_NOTLAST
))
942 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
949 return sk_stream_error(sk
, flags
, err
);
952 int tcp_sendpage(struct sock
*sk
, struct page
*page
, int offset
,
953 size_t size
, int flags
)
957 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
958 !(sk
->sk_route_caps
& NETIF_F_ALL_CSUM
))
959 return sock_no_sendpage(sk
->sk_socket
, page
, offset
, size
,
963 res
= do_tcp_sendpages(sk
, page
, offset
, size
, flags
);
967 EXPORT_SYMBOL(tcp_sendpage
);
969 static inline int select_size(const struct sock
*sk
, bool sg
)
971 const struct tcp_sock
*tp
= tcp_sk(sk
);
972 int tmp
= tp
->mss_cache
;
975 if (sk_can_gso(sk
)) {
976 /* Small frames wont use a full page:
977 * Payload will immediately follow tcp header.
979 tmp
= SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER
);
981 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
983 if (tmp
>= pgbreak
&&
984 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
992 void tcp_free_fastopen_req(struct tcp_sock
*tp
)
994 if (tp
->fastopen_req
!= NULL
) {
995 kfree(tp
->fastopen_req
);
996 tp
->fastopen_req
= NULL
;
1000 static int tcp_sendmsg_fastopen(struct sock
*sk
, struct msghdr
*msg
, int *size
)
1002 struct tcp_sock
*tp
= tcp_sk(sk
);
1005 if (!(sysctl_tcp_fastopen
& TFO_CLIENT_ENABLE
))
1007 if (tp
->fastopen_req
!= NULL
)
1008 return -EALREADY
; /* Another Fast Open is in progress */
1010 tp
->fastopen_req
= kzalloc(sizeof(struct tcp_fastopen_request
),
1012 if (unlikely(tp
->fastopen_req
== NULL
))
1014 tp
->fastopen_req
->data
= msg
;
1016 flags
= (msg
->msg_flags
& MSG_DONTWAIT
) ? O_NONBLOCK
: 0;
1017 err
= __inet_stream_connect(sk
->sk_socket
, msg
->msg_name
,
1018 msg
->msg_namelen
, flags
);
1019 *size
= tp
->fastopen_req
->copied
;
1020 tcp_free_fastopen_req(tp
);
1024 int tcp_sendmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1028 struct tcp_sock
*tp
= tcp_sk(sk
);
1029 struct sk_buff
*skb
;
1030 int iovlen
, flags
, err
, copied
= 0;
1031 int mss_now
= 0, size_goal
, copied_syn
= 0, offset
= 0;
1037 flags
= msg
->msg_flags
;
1038 if (flags
& MSG_FASTOPEN
) {
1039 err
= tcp_sendmsg_fastopen(sk
, msg
, &copied_syn
);
1040 if (err
== -EINPROGRESS
&& copied_syn
> 0)
1044 offset
= copied_syn
;
1047 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
1049 /* Wait for a connection to finish. One exception is TCP Fast Open
1050 * (passive side) where data is allowed to be sent before a connection
1051 * is fully established.
1053 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
1054 !tcp_passive_fastopen(sk
)) {
1055 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
1059 if (unlikely(tp
->repair
)) {
1060 if (tp
->repair_queue
== TCP_RECV_QUEUE
) {
1061 copied
= tcp_send_rcvq(sk
, msg
, size
);
1066 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1069 /* 'common' sending to sendq */
1072 /* This should be in poll */
1073 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
1075 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1077 /* Ok commence sending. */
1078 iovlen
= msg
->msg_iovlen
;
1083 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
1086 sg
= !!(sk
->sk_route_caps
& NETIF_F_SG
);
1088 while (--iovlen
>= 0) {
1089 size_t seglen
= iov
->iov_len
;
1090 unsigned char __user
*from
= iov
->iov_base
;
1093 if (unlikely(offset
> 0)) { /* Skip bytes copied in SYN */
1094 if (offset
>= seglen
) {
1103 while (seglen
> 0) {
1105 int max
= size_goal
;
1107 skb
= tcp_write_queue_tail(sk
);
1108 if (tcp_send_head(sk
)) {
1109 if (skb
->ip_summed
== CHECKSUM_NONE
)
1111 copy
= max
- skb
->len
;
1116 /* Allocate new segment. If the interface is SG,
1117 * allocate skb fitting to single page.
1119 if (!sk_stream_memory_free(sk
))
1120 goto wait_for_sndbuf
;
1122 skb
= sk_stream_alloc_skb(sk
,
1123 select_size(sk
, sg
),
1126 goto wait_for_memory
;
1129 * Check whether we can use HW checksum.
1131 if (sk
->sk_route_caps
& NETIF_F_ALL_CSUM
)
1132 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1134 skb_entail(sk
, skb
);
1139 /* Try to append data to the end of skb. */
1143 /* Where to copy to? */
1144 if (skb_availroom(skb
) > 0) {
1145 /* We have some space in skb head. Superb! */
1146 copy
= min_t(int, copy
, skb_availroom(skb
));
1147 err
= skb_add_data_nocache(sk
, skb
, from
, copy
);
1152 int i
= skb_shinfo(skb
)->nr_frags
;
1153 struct page_frag
*pfrag
= sk_page_frag(sk
);
1155 if (!sk_page_frag_refill(sk
, pfrag
))
1156 goto wait_for_memory
;
1158 if (!skb_can_coalesce(skb
, i
, pfrag
->page
,
1160 if (i
== MAX_SKB_FRAGS
|| !sg
) {
1161 tcp_mark_push(tp
, skb
);
1167 copy
= min_t(int, copy
, pfrag
->size
- pfrag
->offset
);
1169 if (!sk_wmem_schedule(sk
, copy
))
1170 goto wait_for_memory
;
1172 err
= skb_copy_to_page_nocache(sk
, from
, skb
,
1179 /* Update the skb. */
1181 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
1183 skb_fill_page_desc(skb
, i
, pfrag
->page
,
1184 pfrag
->offset
, copy
);
1185 get_page(pfrag
->page
);
1187 pfrag
->offset
+= copy
;
1191 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
1193 tp
->write_seq
+= copy
;
1194 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1195 skb_shinfo(skb
)->gso_segs
= 0;
1199 if ((seglen
-= copy
) == 0 && iovlen
== 0)
1202 if (skb
->len
< max
|| (flags
& MSG_OOB
) || unlikely(tp
->repair
))
1205 if (forced_push(tp
)) {
1206 tcp_mark_push(tp
, skb
);
1207 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1208 } else if (skb
== tcp_send_head(sk
))
1209 tcp_push_one(sk
, mss_now
);
1213 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1216 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
1218 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
1221 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1227 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
1229 return copied
+ copied_syn
;
1233 tcp_unlink_write_queue(skb
, sk
);
1234 /* It is the one place in all of TCP, except connection
1235 * reset, where we can be unlinking the send_head.
1237 tcp_check_send_head(sk
, skb
);
1238 sk_wmem_free_skb(sk
, skb
);
1242 if (copied
+ copied_syn
)
1245 err
= sk_stream_error(sk
, flags
, err
);
1249 EXPORT_SYMBOL(tcp_sendmsg
);
1252 * Handle reading urgent data. BSD has very simple semantics for
1253 * this, no blocking and very strange errors 8)
1256 static int tcp_recv_urg(struct sock
*sk
, struct msghdr
*msg
, int len
, int flags
)
1258 struct tcp_sock
*tp
= tcp_sk(sk
);
1260 /* No URG data to read. */
1261 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1262 tp
->urg_data
== TCP_URG_READ
)
1263 return -EINVAL
; /* Yes this is right ! */
1265 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1268 if (tp
->urg_data
& TCP_URG_VALID
) {
1270 char c
= tp
->urg_data
;
1272 if (!(flags
& MSG_PEEK
))
1273 tp
->urg_data
= TCP_URG_READ
;
1275 /* Read urgent data. */
1276 msg
->msg_flags
|= MSG_OOB
;
1279 if (!(flags
& MSG_TRUNC
))
1280 err
= memcpy_toiovec(msg
->msg_iov
, &c
, 1);
1283 msg
->msg_flags
|= MSG_TRUNC
;
1285 return err
? -EFAULT
: len
;
1288 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1291 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1292 * the available implementations agree in this case:
1293 * this call should never block, independent of the
1294 * blocking state of the socket.
1295 * Mike <pall@rz.uni-karlsruhe.de>
1300 static int tcp_peek_sndq(struct sock
*sk
, struct msghdr
*msg
, int len
)
1302 struct sk_buff
*skb
;
1303 int copied
= 0, err
= 0;
1305 /* XXX -- need to support SO_PEEK_OFF */
1307 skb_queue_walk(&sk
->sk_write_queue
, skb
) {
1308 err
= skb_copy_datagram_iovec(skb
, 0, msg
->msg_iov
, skb
->len
);
1315 return err
?: copied
;
1318 /* Clean up the receive buffer for full frames taken by the user,
1319 * then send an ACK if necessary. COPIED is the number of bytes
1320 * tcp_recvmsg has given to the user so far, it speeds up the
1321 * calculation of whether or not we must ACK for the sake of
1324 void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1326 struct tcp_sock
*tp
= tcp_sk(sk
);
1327 bool time_to_ack
= false;
1329 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1331 WARN(skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
),
1332 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1333 tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
);
1335 if (inet_csk_ack_scheduled(sk
)) {
1336 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1337 /* Delayed ACKs frequently hit locked sockets during bulk
1339 if (icsk
->icsk_ack
.blocked
||
1340 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1341 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1343 * If this read emptied read buffer, we send ACK, if
1344 * connection is not bidirectional, user drained
1345 * receive buffer and there was a small segment
1349 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1350 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1351 !icsk
->icsk_ack
.pingpong
)) &&
1352 !atomic_read(&sk
->sk_rmem_alloc
)))
1356 /* We send an ACK if we can now advertise a non-zero window
1357 * which has been raised "significantly".
1359 * Even if window raised up to infinity, do not send window open ACK
1360 * in states, where we will not receive more. It is useless.
1362 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1363 __u32 rcv_window_now
= tcp_receive_window(tp
);
1365 /* Optimize, __tcp_select_window() is not cheap. */
1366 if (2*rcv_window_now
<= tp
->window_clamp
) {
1367 __u32 new_window
= __tcp_select_window(sk
);
1369 /* Send ACK now, if this read freed lots of space
1370 * in our buffer. Certainly, new_window is new window.
1371 * We can advertise it now, if it is not less than current one.
1372 * "Lots" means "at least twice" here.
1374 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1382 static void tcp_prequeue_process(struct sock
*sk
)
1384 struct sk_buff
*skb
;
1385 struct tcp_sock
*tp
= tcp_sk(sk
);
1387 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPPREQUEUED
);
1389 /* RX process wants to run with disabled BHs, though it is not
1392 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
1393 sk_backlog_rcv(sk
, skb
);
1396 /* Clear memory counter. */
1397 tp
->ucopy
.memory
= 0;
1400 #ifdef CONFIG_NET_DMA
1401 static void tcp_service_net_dma(struct sock
*sk
, bool wait
)
1403 dma_cookie_t done
, used
;
1404 dma_cookie_t last_issued
;
1405 struct tcp_sock
*tp
= tcp_sk(sk
);
1407 if (!tp
->ucopy
.dma_chan
)
1410 last_issued
= tp
->ucopy
.dma_cookie
;
1411 dma_async_memcpy_issue_pending(tp
->ucopy
.dma_chan
);
1414 if (dma_async_memcpy_complete(tp
->ucopy
.dma_chan
,
1416 &used
) == DMA_SUCCESS
) {
1417 /* Safe to free early-copied skbs now */
1418 __skb_queue_purge(&sk
->sk_async_wait_queue
);
1421 struct sk_buff
*skb
;
1422 while ((skb
= skb_peek(&sk
->sk_async_wait_queue
)) &&
1423 (dma_async_is_complete(skb
->dma_cookie
, done
,
1424 used
) == DMA_SUCCESS
)) {
1425 __skb_dequeue(&sk
->sk_async_wait_queue
);
1433 static struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1435 struct sk_buff
*skb
;
1438 while ((skb
= skb_peek(&sk
->sk_receive_queue
)) != NULL
) {
1439 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1440 if (tcp_hdr(skb
)->syn
)
1442 if (offset
< skb
->len
|| tcp_hdr(skb
)->fin
) {
1446 /* This looks weird, but this can happen if TCP collapsing
1447 * splitted a fat GRO packet, while we released socket lock
1448 * in skb_splice_bits()
1450 sk_eat_skb(sk
, skb
, false);
1456 * This routine provides an alternative to tcp_recvmsg() for routines
1457 * that would like to handle copying from skbuffs directly in 'sendfile'
1460 * - It is assumed that the socket was locked by the caller.
1461 * - The routine does not block.
1462 * - At present, there is no support for reading OOB data
1463 * or for 'peeking' the socket using this routine
1464 * (although both would be easy to implement).
1466 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1467 sk_read_actor_t recv_actor
)
1469 struct sk_buff
*skb
;
1470 struct tcp_sock
*tp
= tcp_sk(sk
);
1471 u32 seq
= tp
->copied_seq
;
1475 if (sk
->sk_state
== TCP_LISTEN
)
1477 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1478 if (offset
< skb
->len
) {
1482 len
= skb
->len
- offset
;
1483 /* Stop reading if we hit a patch of urgent data */
1485 u32 urg_offset
= tp
->urg_seq
- seq
;
1486 if (urg_offset
< len
)
1491 used
= recv_actor(desc
, skb
, offset
, len
);
1496 } else if (used
<= len
) {
1501 /* If recv_actor drops the lock (e.g. TCP splice
1502 * receive) the skb pointer might be invalid when
1503 * getting here: tcp_collapse might have deleted it
1504 * while aggregating skbs from the socket queue.
1506 skb
= tcp_recv_skb(sk
, seq
- 1, &offset
);
1509 /* TCP coalescing might have appended data to the skb.
1510 * Try to splice more frags
1512 if (offset
+ 1 != skb
->len
)
1515 if (tcp_hdr(skb
)->fin
) {
1516 sk_eat_skb(sk
, skb
, false);
1520 sk_eat_skb(sk
, skb
, false);
1523 tp
->copied_seq
= seq
;
1525 tp
->copied_seq
= seq
;
1527 tcp_rcv_space_adjust(sk
);
1529 /* Clean up data we have read: This will do ACK frames. */
1531 tcp_recv_skb(sk
, seq
, &offset
);
1532 tcp_cleanup_rbuf(sk
, copied
);
1536 EXPORT_SYMBOL(tcp_read_sock
);
1539 * This routine copies from a sock struct into the user buffer.
1541 * Technical note: in 2.3 we work on _locked_ socket, so that
1542 * tricks with *seq access order and skb->users are not required.
1543 * Probably, code can be easily improved even more.
1546 int tcp_recvmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1547 size_t len
, int nonblock
, int flags
, int *addr_len
)
1549 struct tcp_sock
*tp
= tcp_sk(sk
);
1555 int target
; /* Read at least this many bytes */
1557 struct task_struct
*user_recv
= NULL
;
1558 bool copied_early
= false;
1559 struct sk_buff
*skb
;
1565 if (sk
->sk_state
== TCP_LISTEN
)
1568 timeo
= sock_rcvtimeo(sk
, nonblock
);
1570 /* Urgent data needs to be handled specially. */
1571 if (flags
& MSG_OOB
)
1574 if (unlikely(tp
->repair
)) {
1576 if (!(flags
& MSG_PEEK
))
1579 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
1583 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1586 /* 'common' recv queue MSG_PEEK-ing */
1589 seq
= &tp
->copied_seq
;
1590 if (flags
& MSG_PEEK
) {
1591 peek_seq
= tp
->copied_seq
;
1595 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1597 #ifdef CONFIG_NET_DMA
1598 tp
->ucopy
.dma_chan
= NULL
;
1600 skb
= skb_peek_tail(&sk
->sk_receive_queue
);
1605 available
= TCP_SKB_CB(skb
)->seq
+ skb
->len
- (*seq
);
1606 if ((available
< target
) &&
1607 (len
> sysctl_tcp_dma_copybreak
) && !(flags
& MSG_PEEK
) &&
1608 !sysctl_tcp_low_latency
&&
1609 net_dma_find_channel()) {
1610 preempt_enable_no_resched();
1611 tp
->ucopy
.pinned_list
=
1612 dma_pin_iovec_pages(msg
->msg_iov
, len
);
1614 preempt_enable_no_resched();
1622 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1623 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1626 if (signal_pending(current
)) {
1627 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1632 /* Next get a buffer. */
1634 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1635 /* Now that we have two receive queues this
1638 if (WARN(before(*seq
, TCP_SKB_CB(skb
)->seq
),
1639 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1640 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
,
1644 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1645 if (tcp_hdr(skb
)->syn
)
1647 if (offset
< skb
->len
)
1649 if (tcp_hdr(skb
)->fin
)
1651 WARN(!(flags
& MSG_PEEK
),
1652 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1653 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
, flags
);
1656 /* Well, if we have backlog, try to process it now yet. */
1658 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1663 sk
->sk_state
== TCP_CLOSE
||
1664 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1666 signal_pending(current
))
1669 if (sock_flag(sk
, SOCK_DONE
))
1673 copied
= sock_error(sk
);
1677 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1680 if (sk
->sk_state
== TCP_CLOSE
) {
1681 if (!sock_flag(sk
, SOCK_DONE
)) {
1682 /* This occurs when user tries to read
1683 * from never connected socket.
1696 if (signal_pending(current
)) {
1697 copied
= sock_intr_errno(timeo
);
1702 tcp_cleanup_rbuf(sk
, copied
);
1704 if (!sysctl_tcp_low_latency
&& tp
->ucopy
.task
== user_recv
) {
1705 /* Install new reader */
1706 if (!user_recv
&& !(flags
& (MSG_TRUNC
| MSG_PEEK
))) {
1707 user_recv
= current
;
1708 tp
->ucopy
.task
= user_recv
;
1709 tp
->ucopy
.iov
= msg
->msg_iov
;
1712 tp
->ucopy
.len
= len
;
1714 WARN_ON(tp
->copied_seq
!= tp
->rcv_nxt
&&
1715 !(flags
& (MSG_PEEK
| MSG_TRUNC
)));
1717 /* Ugly... If prequeue is not empty, we have to
1718 * process it before releasing socket, otherwise
1719 * order will be broken at second iteration.
1720 * More elegant solution is required!!!
1722 * Look: we have the following (pseudo)queues:
1724 * 1. packets in flight
1729 * Each queue can be processed only if the next ones
1730 * are empty. At this point we have empty receive_queue.
1731 * But prequeue _can_ be not empty after 2nd iteration,
1732 * when we jumped to start of loop because backlog
1733 * processing added something to receive_queue.
1734 * We cannot release_sock(), because backlog contains
1735 * packets arrived _after_ prequeued ones.
1737 * Shortly, algorithm is clear --- to process all
1738 * the queues in order. We could make it more directly,
1739 * requeueing packets from backlog to prequeue, if
1740 * is not empty. It is more elegant, but eats cycles,
1743 if (!skb_queue_empty(&tp
->ucopy
.prequeue
))
1746 /* __ Set realtime policy in scheduler __ */
1749 #ifdef CONFIG_NET_DMA
1750 if (tp
->ucopy
.dma_chan
) {
1751 if (tp
->rcv_wnd
== 0 &&
1752 !skb_queue_empty(&sk
->sk_async_wait_queue
)) {
1753 tcp_service_net_dma(sk
, true);
1754 tcp_cleanup_rbuf(sk
, copied
);
1756 dma_async_memcpy_issue_pending(tp
->ucopy
.dma_chan
);
1759 if (copied
>= target
) {
1760 /* Do not sleep, just process backlog. */
1764 sk_wait_data(sk
, &timeo
);
1766 #ifdef CONFIG_NET_DMA
1767 tcp_service_net_dma(sk
, false); /* Don't block */
1768 tp
->ucopy
.wakeup
= 0;
1774 /* __ Restore normal policy in scheduler __ */
1776 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1777 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG
, chunk
);
1782 if (tp
->rcv_nxt
== tp
->copied_seq
&&
1783 !skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1785 tcp_prequeue_process(sk
);
1787 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1788 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1794 if ((flags
& MSG_PEEK
) &&
1795 (peek_seq
- copied
- urg_hole
!= tp
->copied_seq
)) {
1796 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1798 task_pid_nr(current
));
1799 peek_seq
= tp
->copied_seq
;
1804 /* Ok so how much can we use? */
1805 used
= skb
->len
- offset
;
1809 /* Do we have urgent data here? */
1811 u32 urg_offset
= tp
->urg_seq
- *seq
;
1812 if (urg_offset
< used
) {
1814 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1827 if (!(flags
& MSG_TRUNC
)) {
1828 #ifdef CONFIG_NET_DMA
1829 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
1830 tp
->ucopy
.dma_chan
= net_dma_find_channel();
1832 if (tp
->ucopy
.dma_chan
) {
1833 tp
->ucopy
.dma_cookie
= dma_skb_copy_datagram_iovec(
1834 tp
->ucopy
.dma_chan
, skb
, offset
,
1836 tp
->ucopy
.pinned_list
);
1838 if (tp
->ucopy
.dma_cookie
< 0) {
1840 pr_alert("%s: dma_cookie < 0\n",
1843 /* Exception. Bailout! */
1849 dma_async_memcpy_issue_pending(tp
->ucopy
.dma_chan
);
1851 if ((offset
+ used
) == skb
->len
)
1852 copied_early
= true;
1857 err
= skb_copy_datagram_iovec(skb
, offset
,
1858 msg
->msg_iov
, used
);
1860 /* Exception. Bailout! */
1872 tcp_rcv_space_adjust(sk
);
1875 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1877 tcp_fast_path_check(sk
);
1879 if (used
+ offset
< skb
->len
)
1882 if (tcp_hdr(skb
)->fin
)
1884 if (!(flags
& MSG_PEEK
)) {
1885 sk_eat_skb(sk
, skb
, copied_early
);
1886 copied_early
= false;
1891 /* Process the FIN. */
1893 if (!(flags
& MSG_PEEK
)) {
1894 sk_eat_skb(sk
, skb
, copied_early
);
1895 copied_early
= false;
1901 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1904 tp
->ucopy
.len
= copied
> 0 ? len
: 0;
1906 tcp_prequeue_process(sk
);
1908 if (copied
> 0 && (chunk
= len
- tp
->ucopy
.len
) != 0) {
1909 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1915 tp
->ucopy
.task
= NULL
;
1919 #ifdef CONFIG_NET_DMA
1920 tcp_service_net_dma(sk
, true); /* Wait for queue to drain */
1921 tp
->ucopy
.dma_chan
= NULL
;
1923 if (tp
->ucopy
.pinned_list
) {
1924 dma_unpin_iovec_pages(tp
->ucopy
.pinned_list
);
1925 tp
->ucopy
.pinned_list
= NULL
;
1929 /* According to UNIX98, msg_name/msg_namelen are ignored
1930 * on connected socket. I was just happy when found this 8) --ANK
1933 /* Clean up data we have read: This will do ACK frames. */
1934 tcp_cleanup_rbuf(sk
, copied
);
1944 err
= tcp_recv_urg(sk
, msg
, len
, flags
);
1948 err
= tcp_peek_sndq(sk
, msg
, len
);
1951 EXPORT_SYMBOL(tcp_recvmsg
);
1953 void tcp_set_state(struct sock
*sk
, int state
)
1955 int oldstate
= sk
->sk_state
;
1958 case TCP_ESTABLISHED
:
1959 if (oldstate
!= TCP_ESTABLISHED
)
1960 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1964 if (oldstate
== TCP_CLOSE_WAIT
|| oldstate
== TCP_ESTABLISHED
)
1965 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ESTABRESETS
);
1967 sk
->sk_prot
->unhash(sk
);
1968 if (inet_csk(sk
)->icsk_bind_hash
&&
1969 !(sk
->sk_userlocks
& SOCK_BINDPORT_LOCK
))
1973 if (oldstate
== TCP_ESTABLISHED
)
1974 TCP_DEC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1977 /* Change state AFTER socket is unhashed to avoid closed
1978 * socket sitting in hash tables.
1980 sk
->sk_state
= state
;
1983 SOCK_DEBUG(sk
, "TCP sk=%p, State %s -> %s\n", sk
, statename
[oldstate
], statename
[state
]);
1986 EXPORT_SYMBOL_GPL(tcp_set_state
);
1989 * State processing on a close. This implements the state shift for
1990 * sending our FIN frame. Note that we only send a FIN for some
1991 * states. A shutdown() may have already sent the FIN, or we may be
1995 static const unsigned char new_state
[16] = {
1996 /* current state: new state: action: */
1997 /* (Invalid) */ TCP_CLOSE
,
1998 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1999 /* TCP_SYN_SENT */ TCP_CLOSE
,
2000 /* TCP_SYN_RECV */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2001 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1
,
2002 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2
,
2003 /* TCP_TIME_WAIT */ TCP_CLOSE
,
2004 /* TCP_CLOSE */ TCP_CLOSE
,
2005 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK
| TCP_ACTION_FIN
,
2006 /* TCP_LAST_ACK */ TCP_LAST_ACK
,
2007 /* TCP_LISTEN */ TCP_CLOSE
,
2008 /* TCP_CLOSING */ TCP_CLOSING
,
2011 static int tcp_close_state(struct sock
*sk
)
2013 int next
= (int)new_state
[sk
->sk_state
];
2014 int ns
= next
& TCP_STATE_MASK
;
2016 tcp_set_state(sk
, ns
);
2018 return next
& TCP_ACTION_FIN
;
2022 * Shutdown the sending side of a connection. Much like close except
2023 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2026 void tcp_shutdown(struct sock
*sk
, int how
)
2028 /* We need to grab some memory, and put together a FIN,
2029 * and then put it into the queue to be sent.
2030 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2032 if (!(how
& SEND_SHUTDOWN
))
2035 /* If we've already sent a FIN, or it's a closed state, skip this. */
2036 if ((1 << sk
->sk_state
) &
2037 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
2038 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
2039 /* Clear out any half completed packets. FIN if needed. */
2040 if (tcp_close_state(sk
))
2044 EXPORT_SYMBOL(tcp_shutdown
);
2046 bool tcp_check_oom(struct sock
*sk
, int shift
)
2048 bool too_many_orphans
, out_of_socket_memory
;
2050 too_many_orphans
= tcp_too_many_orphans(sk
, shift
);
2051 out_of_socket_memory
= tcp_out_of_memory(sk
);
2053 if (too_many_orphans
)
2054 net_info_ratelimited("too many orphaned sockets\n");
2055 if (out_of_socket_memory
)
2056 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2057 return too_many_orphans
|| out_of_socket_memory
;
2060 void tcp_close(struct sock
*sk
, long timeout
)
2062 struct sk_buff
*skb
;
2063 int data_was_unread
= 0;
2067 sk
->sk_shutdown
= SHUTDOWN_MASK
;
2069 if (sk
->sk_state
== TCP_LISTEN
) {
2070 tcp_set_state(sk
, TCP_CLOSE
);
2073 inet_csk_listen_stop(sk
);
2075 goto adjudge_to_death
;
2078 /* We need to flush the recv. buffs. We do this only on the
2079 * descriptor close, not protocol-sourced closes, because the
2080 * reader process may not have drained the data yet!
2082 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
2083 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
-
2085 data_was_unread
+= len
;
2091 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2092 if (sk
->sk_state
== TCP_CLOSE
)
2093 goto adjudge_to_death
;
2095 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2096 * data was lost. To witness the awful effects of the old behavior of
2097 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2098 * GET in an FTP client, suspend the process, wait for the client to
2099 * advertise a zero window, then kill -9 the FTP client, wheee...
2100 * Note: timeout is always zero in such a case.
2102 if (unlikely(tcp_sk(sk
)->repair
)) {
2103 sk
->sk_prot
->disconnect(sk
, 0);
2104 } else if (data_was_unread
) {
2105 /* Unread data was tossed, zap the connection. */
2106 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONCLOSE
);
2107 tcp_set_state(sk
, TCP_CLOSE
);
2108 tcp_send_active_reset(sk
, sk
->sk_allocation
);
2109 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
2110 /* Check zero linger _after_ checking for unread data. */
2111 sk
->sk_prot
->disconnect(sk
, 0);
2112 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
2113 } else if (tcp_close_state(sk
)) {
2114 /* We FIN if the application ate all the data before
2115 * zapping the connection.
2118 /* RED-PEN. Formally speaking, we have broken TCP state
2119 * machine. State transitions:
2121 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2122 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2123 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2125 * are legal only when FIN has been sent (i.e. in window),
2126 * rather than queued out of window. Purists blame.
2128 * F.e. "RFC state" is ESTABLISHED,
2129 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2131 * The visible declinations are that sometimes
2132 * we enter time-wait state, when it is not required really
2133 * (harmless), do not send active resets, when they are
2134 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2135 * they look as CLOSING or LAST_ACK for Linux)
2136 * Probably, I missed some more holelets.
2138 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2139 * in a single packet! (May consider it later but will
2140 * probably need API support or TCP_CORK SYN-ACK until
2141 * data is written and socket is closed.)
2146 sk_stream_wait_close(sk
, timeout
);
2149 state
= sk
->sk_state
;
2153 /* It is the last release_sock in its life. It will remove backlog. */
2157 /* Now socket is owned by kernel and we acquire BH lock
2158 to finish close. No need to check for user refs.
2162 WARN_ON(sock_owned_by_user(sk
));
2164 percpu_counter_inc(sk
->sk_prot
->orphan_count
);
2166 /* Have we already been destroyed by a softirq or backlog? */
2167 if (state
!= TCP_CLOSE
&& sk
->sk_state
== TCP_CLOSE
)
2170 /* This is a (useful) BSD violating of the RFC. There is a
2171 * problem with TCP as specified in that the other end could
2172 * keep a socket open forever with no application left this end.
2173 * We use a 3 minute timeout (about the same as BSD) then kill
2174 * our end. If they send after that then tough - BUT: long enough
2175 * that we won't make the old 4*rto = almost no time - whoops
2178 * Nope, it was not mistake. It is really desired behaviour
2179 * f.e. on http servers, when such sockets are useless, but
2180 * consume significant resources. Let's do it with special
2181 * linger2 option. --ANK
2184 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
2185 struct tcp_sock
*tp
= tcp_sk(sk
);
2186 if (tp
->linger2
< 0) {
2187 tcp_set_state(sk
, TCP_CLOSE
);
2188 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2189 NET_INC_STATS_BH(sock_net(sk
),
2190 LINUX_MIB_TCPABORTONLINGER
);
2192 const int tmo
= tcp_fin_time(sk
);
2194 if (tmo
> TCP_TIMEWAIT_LEN
) {
2195 inet_csk_reset_keepalive_timer(sk
,
2196 tmo
- TCP_TIMEWAIT_LEN
);
2198 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
2203 if (sk
->sk_state
!= TCP_CLOSE
) {
2205 if (tcp_check_oom(sk
, 0)) {
2206 tcp_set_state(sk
, TCP_CLOSE
);
2207 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2208 NET_INC_STATS_BH(sock_net(sk
),
2209 LINUX_MIB_TCPABORTONMEMORY
);
2213 if (sk
->sk_state
== TCP_CLOSE
) {
2214 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
2215 /* We could get here with a non-NULL req if the socket is
2216 * aborted (e.g., closed with unread data) before 3WHS
2220 reqsk_fastopen_remove(sk
, req
, false);
2221 inet_csk_destroy_sock(sk
);
2223 /* Otherwise, socket is reprieved until protocol close. */
2230 EXPORT_SYMBOL(tcp_close
);
2232 /* These states need RST on ABORT according to RFC793 */
2234 static inline bool tcp_need_reset(int state
)
2236 return (1 << state
) &
2237 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
2238 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
2241 int tcp_disconnect(struct sock
*sk
, int flags
)
2243 struct inet_sock
*inet
= inet_sk(sk
);
2244 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2245 struct tcp_sock
*tp
= tcp_sk(sk
);
2247 int old_state
= sk
->sk_state
;
2249 if (old_state
!= TCP_CLOSE
)
2250 tcp_set_state(sk
, TCP_CLOSE
);
2252 /* ABORT function of RFC793 */
2253 if (old_state
== TCP_LISTEN
) {
2254 inet_csk_listen_stop(sk
);
2255 } else if (unlikely(tp
->repair
)) {
2256 sk
->sk_err
= ECONNABORTED
;
2257 } else if (tcp_need_reset(old_state
) ||
2258 (tp
->snd_nxt
!= tp
->write_seq
&&
2259 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
2260 /* The last check adjusts for discrepancy of Linux wrt. RFC
2263 tcp_send_active_reset(sk
, gfp_any());
2264 sk
->sk_err
= ECONNRESET
;
2265 } else if (old_state
== TCP_SYN_SENT
)
2266 sk
->sk_err
= ECONNRESET
;
2268 tcp_clear_xmit_timers(sk
);
2269 __skb_queue_purge(&sk
->sk_receive_queue
);
2270 tcp_write_queue_purge(sk
);
2271 __skb_queue_purge(&tp
->out_of_order_queue
);
2272 #ifdef CONFIG_NET_DMA
2273 __skb_queue_purge(&sk
->sk_async_wait_queue
);
2276 inet
->inet_dport
= 0;
2278 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
2279 inet_reset_saddr(sk
);
2281 sk
->sk_shutdown
= 0;
2282 sock_reset_flag(sk
, SOCK_DONE
);
2284 if ((tp
->write_seq
+= tp
->max_window
+ 2) == 0)
2286 icsk
->icsk_backoff
= 0;
2288 icsk
->icsk_probes_out
= 0;
2289 tp
->packets_out
= 0;
2290 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
2291 tp
->snd_cwnd_cnt
= 0;
2292 tp
->bytes_acked
= 0;
2293 tp
->window_clamp
= 0;
2294 tcp_set_ca_state(sk
, TCP_CA_Open
);
2295 tcp_clear_retrans(tp
);
2296 inet_csk_delack_init(sk
);
2297 tcp_init_send_head(sk
);
2298 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
2301 WARN_ON(inet
->inet_num
&& !icsk
->icsk_bind_hash
);
2303 sk
->sk_error_report(sk
);
2306 EXPORT_SYMBOL(tcp_disconnect
);
2308 void tcp_sock_destruct(struct sock
*sk
)
2310 inet_sock_destruct(sk
);
2312 kfree(inet_csk(sk
)->icsk_accept_queue
.fastopenq
);
2315 static inline bool tcp_can_repair_sock(const struct sock
*sk
)
2317 return ns_capable(sock_net(sk
)->user_ns
, CAP_NET_ADMIN
) &&
2318 ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_ESTABLISHED
));
2321 static int tcp_repair_options_est(struct tcp_sock
*tp
,
2322 struct tcp_repair_opt __user
*optbuf
, unsigned int len
)
2324 struct tcp_repair_opt opt
;
2326 while (len
>= sizeof(opt
)) {
2327 if (copy_from_user(&opt
, optbuf
, sizeof(opt
)))
2333 switch (opt
.opt_code
) {
2335 tp
->rx_opt
.mss_clamp
= opt
.opt_val
;
2339 u16 snd_wscale
= opt
.opt_val
& 0xFFFF;
2340 u16 rcv_wscale
= opt
.opt_val
>> 16;
2342 if (snd_wscale
> 14 || rcv_wscale
> 14)
2345 tp
->rx_opt
.snd_wscale
= snd_wscale
;
2346 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2347 tp
->rx_opt
.wscale_ok
= 1;
2350 case TCPOPT_SACK_PERM
:
2351 if (opt
.opt_val
!= 0)
2354 tp
->rx_opt
.sack_ok
|= TCP_SACK_SEEN
;
2355 if (sysctl_tcp_fack
)
2356 tcp_enable_fack(tp
);
2358 case TCPOPT_TIMESTAMP
:
2359 if (opt
.opt_val
!= 0)
2362 tp
->rx_opt
.tstamp_ok
= 1;
2371 * Socket option code for TCP.
2373 static int do_tcp_setsockopt(struct sock
*sk
, int level
,
2374 int optname
, char __user
*optval
, unsigned int optlen
)
2376 struct tcp_sock
*tp
= tcp_sk(sk
);
2377 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2381 /* These are data/string values, all the others are ints */
2383 case TCP_CONGESTION
: {
2384 char name
[TCP_CA_NAME_MAX
];
2389 val
= strncpy_from_user(name
, optval
,
2390 min_t(long, TCP_CA_NAME_MAX
-1, optlen
));
2396 err
= tcp_set_congestion_control(sk
, name
);
2400 case TCP_COOKIE_TRANSACTIONS
: {
2401 struct tcp_cookie_transactions ctd
;
2402 struct tcp_cookie_values
*cvp
= NULL
;
2404 if (sizeof(ctd
) > optlen
)
2406 if (copy_from_user(&ctd
, optval
, sizeof(ctd
)))
2409 if (ctd
.tcpct_used
> sizeof(ctd
.tcpct_value
) ||
2410 ctd
.tcpct_s_data_desired
> TCP_MSS_DESIRED
)
2413 if (ctd
.tcpct_cookie_desired
== 0) {
2414 /* default to global value */
2415 } else if ((0x1 & ctd
.tcpct_cookie_desired
) ||
2416 ctd
.tcpct_cookie_desired
> TCP_COOKIE_MAX
||
2417 ctd
.tcpct_cookie_desired
< TCP_COOKIE_MIN
) {
2421 if (TCP_COOKIE_OUT_NEVER
& ctd
.tcpct_flags
) {
2422 /* Supercedes all other values */
2424 if (tp
->cookie_values
!= NULL
) {
2425 kref_put(&tp
->cookie_values
->kref
,
2426 tcp_cookie_values_release
);
2427 tp
->cookie_values
= NULL
;
2429 tp
->rx_opt
.cookie_in_always
= 0; /* false */
2430 tp
->rx_opt
.cookie_out_never
= 1; /* true */
2435 /* Allocate ancillary memory before locking.
2437 if (ctd
.tcpct_used
> 0 ||
2438 (tp
->cookie_values
== NULL
&&
2439 (sysctl_tcp_cookie_size
> 0 ||
2440 ctd
.tcpct_cookie_desired
> 0 ||
2441 ctd
.tcpct_s_data_desired
> 0))) {
2442 cvp
= kzalloc(sizeof(*cvp
) + ctd
.tcpct_used
,
2447 kref_init(&cvp
->kref
);
2450 tp
->rx_opt
.cookie_in_always
=
2451 (TCP_COOKIE_IN_ALWAYS
& ctd
.tcpct_flags
);
2452 tp
->rx_opt
.cookie_out_never
= 0; /* false */
2454 if (tp
->cookie_values
!= NULL
) {
2456 /* Changed values are recorded by a changed
2457 * pointer, ensuring the cookie will differ,
2458 * without separately hashing each value later.
2460 kref_put(&tp
->cookie_values
->kref
,
2461 tcp_cookie_values_release
);
2463 cvp
= tp
->cookie_values
;
2468 cvp
->cookie_desired
= ctd
.tcpct_cookie_desired
;
2470 if (ctd
.tcpct_used
> 0) {
2471 memcpy(cvp
->s_data_payload
, ctd
.tcpct_value
,
2473 cvp
->s_data_desired
= ctd
.tcpct_used
;
2474 cvp
->s_data_constant
= 1; /* true */
2476 /* No constant payload data. */
2477 cvp
->s_data_desired
= ctd
.tcpct_s_data_desired
;
2478 cvp
->s_data_constant
= 0; /* false */
2481 tp
->cookie_values
= cvp
;
2491 if (optlen
< sizeof(int))
2494 if (get_user(val
, (int __user
*)optval
))
2501 /* Values greater than interface MTU won't take effect. However
2502 * at the point when this call is done we typically don't yet
2503 * know which interface is going to be used */
2504 if (val
< TCP_MIN_MSS
|| val
> MAX_TCP_WINDOW
) {
2508 tp
->rx_opt
.user_mss
= val
;
2513 /* TCP_NODELAY is weaker than TCP_CORK, so that
2514 * this option on corked socket is remembered, but
2515 * it is not activated until cork is cleared.
2517 * However, when TCP_NODELAY is set we make
2518 * an explicit push, which overrides even TCP_CORK
2519 * for currently queued segments.
2521 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
2522 tcp_push_pending_frames(sk
);
2524 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
2528 case TCP_THIN_LINEAR_TIMEOUTS
:
2529 if (val
< 0 || val
> 1)
2535 case TCP_THIN_DUPACK
:
2536 if (val
< 0 || val
> 1)
2539 tp
->thin_dupack
= val
;
2540 if (tp
->thin_dupack
)
2541 tcp_disable_early_retrans(tp
);
2545 if (!tcp_can_repair_sock(sk
))
2547 else if (val
== 1) {
2549 sk
->sk_reuse
= SK_FORCE_REUSE
;
2550 tp
->repair_queue
= TCP_NO_QUEUE
;
2551 } else if (val
== 0) {
2553 sk
->sk_reuse
= SK_NO_REUSE
;
2554 tcp_send_window_probe(sk
);
2560 case TCP_REPAIR_QUEUE
:
2563 else if (val
< TCP_QUEUES_NR
)
2564 tp
->repair_queue
= val
;
2570 if (sk
->sk_state
!= TCP_CLOSE
)
2572 else if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2573 tp
->write_seq
= val
;
2574 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2580 case TCP_REPAIR_OPTIONS
:
2583 else if (sk
->sk_state
== TCP_ESTABLISHED
)
2584 err
= tcp_repair_options_est(tp
,
2585 (struct tcp_repair_opt __user
*)optval
,
2592 /* When set indicates to always queue non-full frames.
2593 * Later the user clears this option and we transmit
2594 * any pending partial frames in the queue. This is
2595 * meant to be used alongside sendfile() to get properly
2596 * filled frames when the user (for example) must write
2597 * out headers with a write() call first and then use
2598 * sendfile to send out the data parts.
2600 * TCP_CORK can be set together with TCP_NODELAY and it is
2601 * stronger than TCP_NODELAY.
2604 tp
->nonagle
|= TCP_NAGLE_CORK
;
2606 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
2607 if (tp
->nonagle
&TCP_NAGLE_OFF
)
2608 tp
->nonagle
|= TCP_NAGLE_PUSH
;
2609 tcp_push_pending_frames(sk
);
2614 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
2617 tp
->keepalive_time
= val
* HZ
;
2618 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
2619 !((1 << sk
->sk_state
) &
2620 (TCPF_CLOSE
| TCPF_LISTEN
))) {
2621 u32 elapsed
= keepalive_time_elapsed(tp
);
2622 if (tp
->keepalive_time
> elapsed
)
2623 elapsed
= tp
->keepalive_time
- elapsed
;
2626 inet_csk_reset_keepalive_timer(sk
, elapsed
);
2631 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
2634 tp
->keepalive_intvl
= val
* HZ
;
2637 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
2640 tp
->keepalive_probes
= val
;
2643 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
2646 icsk
->icsk_syn_retries
= val
;
2652 else if (val
> sysctl_tcp_fin_timeout
/ HZ
)
2655 tp
->linger2
= val
* HZ
;
2658 case TCP_DEFER_ACCEPT
:
2659 /* Translate value in seconds to number of retransmits */
2660 icsk
->icsk_accept_queue
.rskq_defer_accept
=
2661 secs_to_retrans(val
, TCP_TIMEOUT_INIT
/ HZ
,
2665 case TCP_WINDOW_CLAMP
:
2667 if (sk
->sk_state
!= TCP_CLOSE
) {
2671 tp
->window_clamp
= 0;
2673 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
2674 SOCK_MIN_RCVBUF
/ 2 : val
;
2679 icsk
->icsk_ack
.pingpong
= 1;
2681 icsk
->icsk_ack
.pingpong
= 0;
2682 if ((1 << sk
->sk_state
) &
2683 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
2684 inet_csk_ack_scheduled(sk
)) {
2685 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
2686 tcp_cleanup_rbuf(sk
, 1);
2688 icsk
->icsk_ack
.pingpong
= 1;
2693 #ifdef CONFIG_TCP_MD5SIG
2695 /* Read the IP->Key mappings from userspace */
2696 err
= tp
->af_specific
->md5_parse(sk
, optval
, optlen
);
2699 case TCP_USER_TIMEOUT
:
2700 /* Cap the max timeout in ms TCP will retry/retrans
2701 * before giving up and aborting (ETIMEDOUT) a connection.
2706 icsk
->icsk_user_timeout
= msecs_to_jiffies(val
);
2710 if (val
>= 0 && ((1 << sk
->sk_state
) & (TCPF_CLOSE
|
2712 err
= fastopen_init_queue(sk
, val
);
2725 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2726 unsigned int optlen
)
2728 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2730 if (level
!= SOL_TCP
)
2731 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
2733 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2735 EXPORT_SYMBOL(tcp_setsockopt
);
2737 #ifdef CONFIG_COMPAT
2738 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
2739 char __user
*optval
, unsigned int optlen
)
2741 if (level
!= SOL_TCP
)
2742 return inet_csk_compat_setsockopt(sk
, level
, optname
,
2744 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2746 EXPORT_SYMBOL(compat_tcp_setsockopt
);
2749 /* Return information about state of tcp endpoint in API format. */
2750 void tcp_get_info(const struct sock
*sk
, struct tcp_info
*info
)
2752 const struct tcp_sock
*tp
= tcp_sk(sk
);
2753 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2754 u32 now
= tcp_time_stamp
;
2756 memset(info
, 0, sizeof(*info
));
2758 info
->tcpi_state
= sk
->sk_state
;
2759 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
2760 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
2761 info
->tcpi_probes
= icsk
->icsk_probes_out
;
2762 info
->tcpi_backoff
= icsk
->icsk_backoff
;
2764 if (tp
->rx_opt
.tstamp_ok
)
2765 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
2766 if (tcp_is_sack(tp
))
2767 info
->tcpi_options
|= TCPI_OPT_SACK
;
2768 if (tp
->rx_opt
.wscale_ok
) {
2769 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
2770 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
2771 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
2774 if (tp
->ecn_flags
& TCP_ECN_OK
)
2775 info
->tcpi_options
|= TCPI_OPT_ECN
;
2776 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
2777 info
->tcpi_options
|= TCPI_OPT_ECN_SEEN
;
2778 if (tp
->syn_data_acked
)
2779 info
->tcpi_options
|= TCPI_OPT_SYN_DATA
;
2781 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
2782 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
2783 info
->tcpi_snd_mss
= tp
->mss_cache
;
2784 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
2786 if (sk
->sk_state
== TCP_LISTEN
) {
2787 info
->tcpi_unacked
= sk
->sk_ack_backlog
;
2788 info
->tcpi_sacked
= sk
->sk_max_ack_backlog
;
2790 info
->tcpi_unacked
= tp
->packets_out
;
2791 info
->tcpi_sacked
= tp
->sacked_out
;
2793 info
->tcpi_lost
= tp
->lost_out
;
2794 info
->tcpi_retrans
= tp
->retrans_out
;
2795 info
->tcpi_fackets
= tp
->fackets_out
;
2797 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
2798 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
2799 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
2801 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
2802 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
2803 info
->tcpi_rtt
= jiffies_to_usecs(tp
->srtt
)>>3;
2804 info
->tcpi_rttvar
= jiffies_to_usecs(tp
->mdev
)>>2;
2805 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
2806 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
2807 info
->tcpi_advmss
= tp
->advmss
;
2808 info
->tcpi_reordering
= tp
->reordering
;
2810 info
->tcpi_rcv_rtt
= jiffies_to_usecs(tp
->rcv_rtt_est
.rtt
)>>3;
2811 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
2813 info
->tcpi_total_retrans
= tp
->total_retrans
;
2815 EXPORT_SYMBOL_GPL(tcp_get_info
);
2817 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
2818 int optname
, char __user
*optval
, int __user
*optlen
)
2820 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2821 struct tcp_sock
*tp
= tcp_sk(sk
);
2824 if (get_user(len
, optlen
))
2827 len
= min_t(unsigned int, len
, sizeof(int));
2834 val
= tp
->mss_cache
;
2835 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
2836 val
= tp
->rx_opt
.user_mss
;
2838 val
= tp
->rx_opt
.mss_clamp
;
2841 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
2844 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
2847 val
= keepalive_time_when(tp
) / HZ
;
2850 val
= keepalive_intvl_when(tp
) / HZ
;
2853 val
= keepalive_probes(tp
);
2856 val
= icsk
->icsk_syn_retries
? : sysctl_tcp_syn_retries
;
2861 val
= (val
? : sysctl_tcp_fin_timeout
) / HZ
;
2863 case TCP_DEFER_ACCEPT
:
2864 val
= retrans_to_secs(icsk
->icsk_accept_queue
.rskq_defer_accept
,
2865 TCP_TIMEOUT_INIT
/ HZ
, TCP_RTO_MAX
/ HZ
);
2867 case TCP_WINDOW_CLAMP
:
2868 val
= tp
->window_clamp
;
2871 struct tcp_info info
;
2873 if (get_user(len
, optlen
))
2876 tcp_get_info(sk
, &info
);
2878 len
= min_t(unsigned int, len
, sizeof(info
));
2879 if (put_user(len
, optlen
))
2881 if (copy_to_user(optval
, &info
, len
))
2886 val
= !icsk
->icsk_ack
.pingpong
;
2889 case TCP_CONGESTION
:
2890 if (get_user(len
, optlen
))
2892 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
2893 if (put_user(len
, optlen
))
2895 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
2899 case TCP_COOKIE_TRANSACTIONS
: {
2900 struct tcp_cookie_transactions ctd
;
2901 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
2903 if (get_user(len
, optlen
))
2905 if (len
< sizeof(ctd
))
2908 memset(&ctd
, 0, sizeof(ctd
));
2909 ctd
.tcpct_flags
= (tp
->rx_opt
.cookie_in_always
?
2910 TCP_COOKIE_IN_ALWAYS
: 0)
2911 | (tp
->rx_opt
.cookie_out_never
?
2912 TCP_COOKIE_OUT_NEVER
: 0);
2915 ctd
.tcpct_flags
|= (cvp
->s_data_in
?
2917 | (cvp
->s_data_out
?
2918 TCP_S_DATA_OUT
: 0);
2920 ctd
.tcpct_cookie_desired
= cvp
->cookie_desired
;
2921 ctd
.tcpct_s_data_desired
= cvp
->s_data_desired
;
2923 memcpy(&ctd
.tcpct_value
[0], &cvp
->cookie_pair
[0],
2924 cvp
->cookie_pair_size
);
2925 ctd
.tcpct_used
= cvp
->cookie_pair_size
;
2928 if (put_user(sizeof(ctd
), optlen
))
2930 if (copy_to_user(optval
, &ctd
, sizeof(ctd
)))
2934 case TCP_THIN_LINEAR_TIMEOUTS
:
2937 case TCP_THIN_DUPACK
:
2938 val
= tp
->thin_dupack
;
2945 case TCP_REPAIR_QUEUE
:
2947 val
= tp
->repair_queue
;
2953 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2954 val
= tp
->write_seq
;
2955 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2961 case TCP_USER_TIMEOUT
:
2962 val
= jiffies_to_msecs(icsk
->icsk_user_timeout
);
2965 return -ENOPROTOOPT
;
2968 if (put_user(len
, optlen
))
2970 if (copy_to_user(optval
, &val
, len
))
2975 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2978 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2980 if (level
!= SOL_TCP
)
2981 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
2983 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2985 EXPORT_SYMBOL(tcp_getsockopt
);
2987 #ifdef CONFIG_COMPAT
2988 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
2989 char __user
*optval
, int __user
*optlen
)
2991 if (level
!= SOL_TCP
)
2992 return inet_csk_compat_getsockopt(sk
, level
, optname
,
2994 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2996 EXPORT_SYMBOL(compat_tcp_getsockopt
);
2999 struct sk_buff
*tcp_tso_segment(struct sk_buff
*skb
,
3000 netdev_features_t features
)
3002 struct sk_buff
*segs
= ERR_PTR(-EINVAL
);
3007 unsigned int oldlen
;
3010 if (!pskb_may_pull(skb
, sizeof(*th
)))
3014 thlen
= th
->doff
* 4;
3015 if (thlen
< sizeof(*th
))
3018 if (!pskb_may_pull(skb
, thlen
))
3021 oldlen
= (u16
)~skb
->len
;
3022 __skb_pull(skb
, thlen
);
3024 mss
= skb_shinfo(skb
)->gso_size
;
3025 if (unlikely(skb
->len
<= mss
))
3028 if (skb_gso_ok(skb
, features
| NETIF_F_GSO_ROBUST
)) {
3029 /* Packet is from an untrusted source, reset gso_segs. */
3030 int type
= skb_shinfo(skb
)->gso_type
;
3037 SKB_GSO_SHARED_FRAG
|
3039 !(type
& (SKB_GSO_TCPV4
| SKB_GSO_TCPV6
))))
3042 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss
);
3048 segs
= skb_segment(skb
, features
);
3052 delta
= htonl(oldlen
+ (thlen
+ mss
));
3056 seq
= ntohl(th
->seq
);
3059 th
->fin
= th
->psh
= 0;
3061 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
3062 (__force u32
)delta
));
3063 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
3065 csum_fold(csum_partial(skb_transport_header(skb
),
3072 th
->seq
= htonl(seq
);
3074 } while (skb
->next
);
3076 delta
= htonl(oldlen
+ (skb
->tail
- skb
->transport_header
) +
3078 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
3079 (__force u32
)delta
));
3080 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
3081 th
->check
= csum_fold(csum_partial(skb_transport_header(skb
),
3087 EXPORT_SYMBOL(tcp_tso_segment
);
3089 struct sk_buff
**tcp_gro_receive(struct sk_buff
**head
, struct sk_buff
*skb
)
3091 struct sk_buff
**pp
= NULL
;
3098 unsigned int mss
= 1;
3104 off
= skb_gro_offset(skb
);
3105 hlen
= off
+ sizeof(*th
);
3106 th
= skb_gro_header_fast(skb
, off
);
3107 if (skb_gro_header_hard(skb
, hlen
)) {
3108 th
= skb_gro_header_slow(skb
, hlen
, off
);
3113 thlen
= th
->doff
* 4;
3114 if (thlen
< sizeof(*th
))
3118 if (skb_gro_header_hard(skb
, hlen
)) {
3119 th
= skb_gro_header_slow(skb
, hlen
, off
);
3124 skb_gro_pull(skb
, thlen
);
3126 len
= skb_gro_len(skb
);
3127 flags
= tcp_flag_word(th
);
3129 for (; (p
= *head
); head
= &p
->next
) {
3130 if (!NAPI_GRO_CB(p
)->same_flow
)
3135 if (*(u32
*)&th
->source
^ *(u32
*)&th2
->source
) {
3136 NAPI_GRO_CB(p
)->same_flow
= 0;
3143 goto out_check_final
;
3146 flush
= NAPI_GRO_CB(p
)->flush
;
3147 flush
|= (__force
int)(flags
& TCP_FLAG_CWR
);
3148 flush
|= (__force
int)((flags
^ tcp_flag_word(th2
)) &
3149 ~(TCP_FLAG_CWR
| TCP_FLAG_FIN
| TCP_FLAG_PSH
));
3150 flush
|= (__force
int)(th
->ack_seq
^ th2
->ack_seq
);
3151 for (i
= sizeof(*th
); i
< thlen
; i
+= 4)
3152 flush
|= *(u32
*)((u8
*)th
+ i
) ^
3153 *(u32
*)((u8
*)th2
+ i
);
3155 mss
= skb_shinfo(p
)->gso_size
;
3157 flush
|= (len
- 1) >= mss
;
3158 flush
|= (ntohl(th2
->seq
) + skb_gro_len(p
)) ^ ntohl(th
->seq
);
3160 if (flush
|| skb_gro_receive(head
, skb
)) {
3162 goto out_check_final
;
3167 tcp_flag_word(th2
) |= flags
& (TCP_FLAG_FIN
| TCP_FLAG_PSH
);
3171 flush
|= (__force
int)(flags
& (TCP_FLAG_URG
| TCP_FLAG_PSH
|
3172 TCP_FLAG_RST
| TCP_FLAG_SYN
|
3175 if (p
&& (!NAPI_GRO_CB(skb
)->same_flow
|| flush
))
3179 NAPI_GRO_CB(skb
)->flush
|= flush
;
3183 EXPORT_SYMBOL(tcp_gro_receive
);
3185 int tcp_gro_complete(struct sk_buff
*skb
)
3187 struct tcphdr
*th
= tcp_hdr(skb
);
3189 skb
->csum_start
= skb_transport_header(skb
) - skb
->head
;
3190 skb
->csum_offset
= offsetof(struct tcphdr
, check
);
3191 skb
->ip_summed
= CHECKSUM_PARTIAL
;
3193 skb_shinfo(skb
)->gso_segs
= NAPI_GRO_CB(skb
)->count
;
3196 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
3200 EXPORT_SYMBOL(tcp_gro_complete
);
3202 #ifdef CONFIG_TCP_MD5SIG
3203 static unsigned long tcp_md5sig_users
;
3204 static struct tcp_md5sig_pool __percpu
*tcp_md5sig_pool
;
3205 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock
);
3207 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu
*pool
)
3211 for_each_possible_cpu(cpu
) {
3212 struct tcp_md5sig_pool
*p
= per_cpu_ptr(pool
, cpu
);
3214 if (p
->md5_desc
.tfm
)
3215 crypto_free_hash(p
->md5_desc
.tfm
);
3220 void tcp_free_md5sig_pool(void)
3222 struct tcp_md5sig_pool __percpu
*pool
= NULL
;
3224 spin_lock_bh(&tcp_md5sig_pool_lock
);
3225 if (--tcp_md5sig_users
== 0) {
3226 pool
= tcp_md5sig_pool
;
3227 tcp_md5sig_pool
= NULL
;
3229 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3231 __tcp_free_md5sig_pool(pool
);
3233 EXPORT_SYMBOL(tcp_free_md5sig_pool
);
3235 static struct tcp_md5sig_pool __percpu
*
3236 __tcp_alloc_md5sig_pool(struct sock
*sk
)
3239 struct tcp_md5sig_pool __percpu
*pool
;
3241 pool
= alloc_percpu(struct tcp_md5sig_pool
);
3245 for_each_possible_cpu(cpu
) {
3246 struct crypto_hash
*hash
;
3248 hash
= crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC
);
3249 if (IS_ERR_OR_NULL(hash
))
3252 per_cpu_ptr(pool
, cpu
)->md5_desc
.tfm
= hash
;
3256 __tcp_free_md5sig_pool(pool
);
3260 struct tcp_md5sig_pool __percpu
*tcp_alloc_md5sig_pool(struct sock
*sk
)
3262 struct tcp_md5sig_pool __percpu
*pool
;
3266 spin_lock_bh(&tcp_md5sig_pool_lock
);
3267 pool
= tcp_md5sig_pool
;
3268 if (tcp_md5sig_users
++ == 0) {
3270 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3273 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3277 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3280 /* we cannot hold spinlock here because this may sleep. */
3281 struct tcp_md5sig_pool __percpu
*p
;
3283 p
= __tcp_alloc_md5sig_pool(sk
);
3284 spin_lock_bh(&tcp_md5sig_pool_lock
);
3287 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3290 pool
= tcp_md5sig_pool
;
3292 /* oops, it has already been assigned. */
3293 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3294 __tcp_free_md5sig_pool(p
);
3296 tcp_md5sig_pool
= pool
= p
;
3297 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3302 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
3306 * tcp_get_md5sig_pool - get md5sig_pool for this user
3308 * We use percpu structure, so if we succeed, we exit with preemption
3309 * and BH disabled, to make sure another thread or softirq handling
3310 * wont try to get same context.
3312 struct tcp_md5sig_pool
*tcp_get_md5sig_pool(void)
3314 struct tcp_md5sig_pool __percpu
*p
;
3318 spin_lock(&tcp_md5sig_pool_lock
);
3319 p
= tcp_md5sig_pool
;
3322 spin_unlock(&tcp_md5sig_pool_lock
);
3325 return this_cpu_ptr(p
);
3330 EXPORT_SYMBOL(tcp_get_md5sig_pool
);
3332 void tcp_put_md5sig_pool(void)
3335 tcp_free_md5sig_pool();
3337 EXPORT_SYMBOL(tcp_put_md5sig_pool
);
3339 int tcp_md5_hash_header(struct tcp_md5sig_pool
*hp
,
3340 const struct tcphdr
*th
)
3342 struct scatterlist sg
;
3346 /* We are not allowed to change tcphdr, make a local copy */
3347 memcpy(&hdr
, th
, sizeof(hdr
));
3350 /* options aren't included in the hash */
3351 sg_init_one(&sg
, &hdr
, sizeof(hdr
));
3352 err
= crypto_hash_update(&hp
->md5_desc
, &sg
, sizeof(hdr
));
3355 EXPORT_SYMBOL(tcp_md5_hash_header
);
3357 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool
*hp
,
3358 const struct sk_buff
*skb
, unsigned int header_len
)
3360 struct scatterlist sg
;
3361 const struct tcphdr
*tp
= tcp_hdr(skb
);
3362 struct hash_desc
*desc
= &hp
->md5_desc
;
3364 const unsigned int head_data_len
= skb_headlen(skb
) > header_len
?
3365 skb_headlen(skb
) - header_len
: 0;
3366 const struct skb_shared_info
*shi
= skb_shinfo(skb
);
3367 struct sk_buff
*frag_iter
;
3369 sg_init_table(&sg
, 1);
3371 sg_set_buf(&sg
, ((u8
*) tp
) + header_len
, head_data_len
);
3372 if (crypto_hash_update(desc
, &sg
, head_data_len
))
3375 for (i
= 0; i
< shi
->nr_frags
; ++i
) {
3376 const struct skb_frag_struct
*f
= &shi
->frags
[i
];
3377 struct page
*page
= skb_frag_page(f
);
3378 sg_set_page(&sg
, page
, skb_frag_size(f
), f
->page_offset
);
3379 if (crypto_hash_update(desc
, &sg
, skb_frag_size(f
)))
3383 skb_walk_frags(skb
, frag_iter
)
3384 if (tcp_md5_hash_skb_data(hp
, frag_iter
, 0))
3389 EXPORT_SYMBOL(tcp_md5_hash_skb_data
);
3391 int tcp_md5_hash_key(struct tcp_md5sig_pool
*hp
, const struct tcp_md5sig_key
*key
)
3393 struct scatterlist sg
;
3395 sg_init_one(&sg
, key
->key
, key
->keylen
);
3396 return crypto_hash_update(&hp
->md5_desc
, &sg
, key
->keylen
);
3398 EXPORT_SYMBOL(tcp_md5_hash_key
);
3402 /* Each Responder maintains up to two secret values concurrently for
3403 * efficient secret rollover. Each secret value has 4 states:
3405 * Generating. (tcp_secret_generating != tcp_secret_primary)
3406 * Generates new Responder-Cookies, but not yet used for primary
3407 * verification. This is a short-term state, typically lasting only
3408 * one round trip time (RTT).
3410 * Primary. (tcp_secret_generating == tcp_secret_primary)
3411 * Used both for generation and primary verification.
3413 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3414 * Used for verification, until the first failure that can be
3415 * verified by the newer Generating secret. At that time, this
3416 * cookie's state is changed to Secondary, and the Generating
3417 * cookie's state is changed to Primary. This is a short-term state,
3418 * typically lasting only one round trip time (RTT).
3420 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3421 * Used for secondary verification, after primary verification
3422 * failures. This state lasts no more than twice the Maximum Segment
3423 * Lifetime (2MSL). Then, the secret is discarded.
3425 struct tcp_cookie_secret
{
3426 /* The secret is divided into two parts. The digest part is the
3427 * equivalent of previously hashing a secret and saving the state,
3428 * and serves as an initialization vector (IV). The message part
3429 * serves as the trailing secret.
3431 u32 secrets
[COOKIE_WORKSPACE_WORDS
];
3432 unsigned long expires
;
3435 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3436 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3437 #define TCP_SECRET_LIFE (HZ * 600)
3439 static struct tcp_cookie_secret tcp_secret_one
;
3440 static struct tcp_cookie_secret tcp_secret_two
;
3442 /* Essentially a circular list, without dynamic allocation. */
3443 static struct tcp_cookie_secret
*tcp_secret_generating
;
3444 static struct tcp_cookie_secret
*tcp_secret_primary
;
3445 static struct tcp_cookie_secret
*tcp_secret_retiring
;
3446 static struct tcp_cookie_secret
*tcp_secret_secondary
;
3448 static DEFINE_SPINLOCK(tcp_secret_locker
);
3450 /* Select a pseudo-random word in the cookie workspace.
3452 static inline u32
tcp_cookie_work(const u32
*ws
, const int n
)
3454 return ws
[COOKIE_DIGEST_WORDS
+ ((COOKIE_MESSAGE_WORDS
-1) & ws
[n
])];
3457 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3458 * Called in softirq context.
3459 * Returns: 0 for success.
3461 int tcp_cookie_generator(u32
*bakery
)
3463 unsigned long jiffy
= jiffies
;
3465 if (unlikely(time_after_eq(jiffy
, tcp_secret_generating
->expires
))) {
3466 spin_lock_bh(&tcp_secret_locker
);
3467 if (!time_after_eq(jiffy
, tcp_secret_generating
->expires
)) {
3468 /* refreshed by another */
3470 &tcp_secret_generating
->secrets
[0],
3471 COOKIE_WORKSPACE_WORDS
);
3473 /* still needs refreshing */
3474 get_random_bytes(bakery
, COOKIE_WORKSPACE_WORDS
);
3476 /* The first time, paranoia assumes that the
3477 * randomization function isn't as strong. But,
3478 * this secret initialization is delayed until
3479 * the last possible moment (packet arrival).
3480 * Although that time is observable, it is
3481 * unpredictably variable. Mash in the most
3482 * volatile clock bits available, and expire the
3483 * secret extra quickly.
3485 if (unlikely(tcp_secret_primary
->expires
==
3486 tcp_secret_secondary
->expires
)) {
3489 getnstimeofday(&tv
);
3490 bakery
[COOKIE_DIGEST_WORDS
+0] ^=
3493 tcp_secret_secondary
->expires
= jiffy
3495 + (0x0f & tcp_cookie_work(bakery
, 0));
3497 tcp_secret_secondary
->expires
= jiffy
3499 + (0xff & tcp_cookie_work(bakery
, 1));
3500 tcp_secret_primary
->expires
= jiffy
3502 + (0x1f & tcp_cookie_work(bakery
, 2));
3504 memcpy(&tcp_secret_secondary
->secrets
[0],
3505 bakery
, COOKIE_WORKSPACE_WORDS
);
3507 rcu_assign_pointer(tcp_secret_generating
,
3508 tcp_secret_secondary
);
3509 rcu_assign_pointer(tcp_secret_retiring
,
3510 tcp_secret_primary
);
3512 * Neither call_rcu() nor synchronize_rcu() needed.
3513 * Retiring data is not freed. It is replaced after
3514 * further (locked) pointer updates, and a quiet time
3515 * (minimum 1MSL, maximum LIFE - 2MSL).
3518 spin_unlock_bh(&tcp_secret_locker
);
3522 &rcu_dereference(tcp_secret_generating
)->secrets
[0],
3523 COOKIE_WORKSPACE_WORDS
);
3524 rcu_read_unlock_bh();
3528 EXPORT_SYMBOL(tcp_cookie_generator
);
3530 void tcp_done(struct sock
*sk
)
3532 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
3534 if (sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
3535 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_ATTEMPTFAILS
);
3537 tcp_set_state(sk
, TCP_CLOSE
);
3538 tcp_clear_xmit_timers(sk
);
3540 reqsk_fastopen_remove(sk
, req
, false);
3542 sk
->sk_shutdown
= SHUTDOWN_MASK
;
3544 if (!sock_flag(sk
, SOCK_DEAD
))
3545 sk
->sk_state_change(sk
);
3547 inet_csk_destroy_sock(sk
);
3549 EXPORT_SYMBOL_GPL(tcp_done
);
3551 extern struct tcp_congestion_ops tcp_reno
;
3553 static __initdata
unsigned long thash_entries
;
3554 static int __init
set_thash_entries(char *str
)
3561 ret
= kstrtoul(str
, 0, &thash_entries
);
3567 __setup("thash_entries=", set_thash_entries
);
3569 void tcp_init_mem(struct net
*net
)
3571 unsigned long limit
= nr_free_buffer_pages() / 8;
3572 limit
= max(limit
, 128UL);
3573 net
->ipv4
.sysctl_tcp_mem
[0] = limit
/ 4 * 3;
3574 net
->ipv4
.sysctl_tcp_mem
[1] = limit
;
3575 net
->ipv4
.sysctl_tcp_mem
[2] = net
->ipv4
.sysctl_tcp_mem
[0] * 2;
3578 void __init
tcp_init(void)
3580 struct sk_buff
*skb
= NULL
;
3581 unsigned long limit
;
3582 int max_rshare
, max_wshare
, cnt
;
3584 unsigned long jiffy
= jiffies
;
3586 BUILD_BUG_ON(sizeof(struct tcp_skb_cb
) > sizeof(skb
->cb
));
3588 percpu_counter_init(&tcp_sockets_allocated
, 0);
3589 percpu_counter_init(&tcp_orphan_count
, 0);
3590 tcp_hashinfo
.bind_bucket_cachep
=
3591 kmem_cache_create("tcp_bind_bucket",
3592 sizeof(struct inet_bind_bucket
), 0,
3593 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
);
3595 /* Size and allocate the main established and bind bucket
3598 * The methodology is similar to that of the buffer cache.
3600 tcp_hashinfo
.ehash
=
3601 alloc_large_system_hash("TCP established",
3602 sizeof(struct inet_ehash_bucket
),
3604 17, /* one slot per 128 KB of memory */
3607 &tcp_hashinfo
.ehash_mask
,
3609 thash_entries
? 0 : 512 * 1024);
3610 for (i
= 0; i
<= tcp_hashinfo
.ehash_mask
; i
++) {
3611 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].chain
, i
);
3612 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].twchain
, i
);
3614 if (inet_ehash_locks_alloc(&tcp_hashinfo
))
3615 panic("TCP: failed to alloc ehash_locks");
3616 tcp_hashinfo
.bhash
=
3617 alloc_large_system_hash("TCP bind",
3618 sizeof(struct inet_bind_hashbucket
),
3619 tcp_hashinfo
.ehash_mask
+ 1,
3620 17, /* one slot per 128 KB of memory */
3622 &tcp_hashinfo
.bhash_size
,
3626 tcp_hashinfo
.bhash_size
= 1U << tcp_hashinfo
.bhash_size
;
3627 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
3628 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
3629 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
3633 cnt
= tcp_hashinfo
.ehash_mask
+ 1;
3635 tcp_death_row
.sysctl_max_tw_buckets
= cnt
/ 2;
3636 sysctl_tcp_max_orphans
= cnt
/ 2;
3637 sysctl_max_syn_backlog
= max(128, cnt
/ 256);
3639 tcp_init_mem(&init_net
);
3640 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3641 limit
= nr_free_buffer_pages() << (PAGE_SHIFT
- 7);
3642 max_wshare
= min(4UL*1024*1024, limit
);
3643 max_rshare
= min(6UL*1024*1024, limit
);
3645 sysctl_tcp_wmem
[0] = SK_MEM_QUANTUM
;
3646 sysctl_tcp_wmem
[1] = 16*1024;
3647 sysctl_tcp_wmem
[2] = max(64*1024, max_wshare
);
3649 sysctl_tcp_rmem
[0] = SK_MEM_QUANTUM
;
3650 sysctl_tcp_rmem
[1] = 87380;
3651 sysctl_tcp_rmem
[2] = max(87380, max_rshare
);
3653 pr_info("Hash tables configured (established %u bind %u)\n",
3654 tcp_hashinfo
.ehash_mask
+ 1, tcp_hashinfo
.bhash_size
);
3658 tcp_register_congestion_control(&tcp_reno
);
3660 memset(&tcp_secret_one
.secrets
[0], 0, sizeof(tcp_secret_one
.secrets
));
3661 memset(&tcp_secret_two
.secrets
[0], 0, sizeof(tcp_secret_two
.secrets
));
3662 tcp_secret_one
.expires
= jiffy
; /* past due */
3663 tcp_secret_two
.expires
= jiffy
; /* past due */
3664 tcp_secret_generating
= &tcp_secret_one
;
3665 tcp_secret_primary
= &tcp_secret_one
;
3666 tcp_secret_retiring
= &tcp_secret_two
;
3667 tcp_secret_secondary
= &tcp_secret_two
;