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
;
405 sk
->sk_state
= TCP_CLOSE
;
407 sk
->sk_write_space
= sk_stream_write_space
;
408 sock_set_flag(sk
, SOCK_USE_WRITE_QUEUE
);
410 icsk
->icsk_sync_mss
= tcp_sync_mss
;
412 /* TCP Cookie Transactions */
413 if (sysctl_tcp_cookie_size
> 0) {
414 /* Default, cookies without s_data_payload. */
416 kzalloc(sizeof(*tp
->cookie_values
),
418 if (tp
->cookie_values
!= NULL
)
419 kref_init(&tp
->cookie_values
->kref
);
421 /* Presumed zeroed, in order of appearance:
422 * cookie_in_always, cookie_out_never,
423 * s_data_constant, s_data_in, s_data_out
425 sk
->sk_sndbuf
= sysctl_tcp_wmem
[1];
426 sk
->sk_rcvbuf
= sysctl_tcp_rmem
[1];
429 sock_update_memcg(sk
);
430 sk_sockets_allocated_inc(sk
);
433 EXPORT_SYMBOL(tcp_init_sock
);
436 * Wait for a TCP event.
438 * Note that we don't need to lock the socket, as the upper poll layers
439 * take care of normal races (between the test and the event) and we don't
440 * go look at any of the socket buffers directly.
442 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
445 struct sock
*sk
= sock
->sk
;
446 const struct tcp_sock
*tp
= tcp_sk(sk
);
448 sock_poll_wait(file
, sk_sleep(sk
), wait
);
449 if (sk
->sk_state
== TCP_LISTEN
)
450 return inet_csk_listen_poll(sk
);
452 /* Socket is not locked. We are protected from async events
453 * by poll logic and correct handling of state changes
454 * made by other threads is impossible in any case.
460 * POLLHUP is certainly not done right. But poll() doesn't
461 * have a notion of HUP in just one direction, and for a
462 * socket the read side is more interesting.
464 * Some poll() documentation says that POLLHUP is incompatible
465 * with the POLLOUT/POLLWR flags, so somebody should check this
466 * all. But careful, it tends to be safer to return too many
467 * bits than too few, and you can easily break real applications
468 * if you don't tell them that something has hung up!
472 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
473 * our fs/select.c). It means that after we received EOF,
474 * poll always returns immediately, making impossible poll() on write()
475 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
476 * if and only if shutdown has been made in both directions.
477 * Actually, it is interesting to look how Solaris and DUX
478 * solve this dilemma. I would prefer, if POLLHUP were maskable,
479 * then we could set it on SND_SHUTDOWN. BTW examples given
480 * in Stevens' books assume exactly this behaviour, it explains
481 * why POLLHUP is incompatible with POLLOUT. --ANK
483 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
484 * blocking on fresh not-connected or disconnected socket. --ANK
486 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| sk
->sk_state
== TCP_CLOSE
)
488 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
489 mask
|= POLLIN
| POLLRDNORM
| POLLRDHUP
;
491 /* Connected or passive Fast Open socket? */
492 if (sk
->sk_state
!= TCP_SYN_SENT
&&
493 (sk
->sk_state
!= TCP_SYN_RECV
|| tp
->fastopen_rsk
!= NULL
)) {
494 int target
= sock_rcvlowat(sk
, 0, INT_MAX
);
496 if (tp
->urg_seq
== tp
->copied_seq
&&
497 !sock_flag(sk
, SOCK_URGINLINE
) &&
501 /* Potential race condition. If read of tp below will
502 * escape above sk->sk_state, we can be illegally awaken
503 * in SYN_* states. */
504 if (tp
->rcv_nxt
- tp
->copied_seq
>= target
)
505 mask
|= POLLIN
| POLLRDNORM
;
507 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
508 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
)) {
509 mask
|= POLLOUT
| POLLWRNORM
;
510 } else { /* send SIGIO later */
511 set_bit(SOCK_ASYNC_NOSPACE
,
512 &sk
->sk_socket
->flags
);
513 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
515 /* Race breaker. If space is freed after
516 * wspace test but before the flags are set,
517 * IO signal will be lost.
519 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
))
520 mask
|= POLLOUT
| POLLWRNORM
;
523 mask
|= POLLOUT
| POLLWRNORM
;
525 if (tp
->urg_data
& TCP_URG_VALID
)
528 /* This barrier is coupled with smp_wmb() in tcp_reset() */
535 EXPORT_SYMBOL(tcp_poll
);
537 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
539 struct tcp_sock
*tp
= tcp_sk(sk
);
545 if (sk
->sk_state
== TCP_LISTEN
)
548 slow
= lock_sock_fast(sk
);
549 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
551 else if (sock_flag(sk
, SOCK_URGINLINE
) ||
553 before(tp
->urg_seq
, tp
->copied_seq
) ||
554 !before(tp
->urg_seq
, tp
->rcv_nxt
)) {
556 answ
= tp
->rcv_nxt
- tp
->copied_seq
;
558 /* Subtract 1, if FIN was received */
559 if (answ
&& sock_flag(sk
, SOCK_DONE
))
562 answ
= tp
->urg_seq
- tp
->copied_seq
;
563 unlock_sock_fast(sk
, slow
);
566 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
569 if (sk
->sk_state
== TCP_LISTEN
)
572 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
575 answ
= tp
->write_seq
- tp
->snd_una
;
578 if (sk
->sk_state
== TCP_LISTEN
)
581 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
584 answ
= tp
->write_seq
- tp
->snd_nxt
;
590 return put_user(answ
, (int __user
*)arg
);
592 EXPORT_SYMBOL(tcp_ioctl
);
594 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
596 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
597 tp
->pushed_seq
= tp
->write_seq
;
600 static inline bool forced_push(const struct tcp_sock
*tp
)
602 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
605 static inline void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
607 struct tcp_sock
*tp
= tcp_sk(sk
);
608 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
611 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
612 tcb
->tcp_flags
= TCPHDR_ACK
;
614 skb_header_release(skb
);
615 tcp_add_write_queue_tail(sk
, skb
);
616 sk
->sk_wmem_queued
+= skb
->truesize
;
617 sk_mem_charge(sk
, skb
->truesize
);
618 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
619 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
622 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
)
625 tp
->snd_up
= tp
->write_seq
;
628 static inline void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
631 if (tcp_send_head(sk
)) {
632 struct tcp_sock
*tp
= tcp_sk(sk
);
634 if (!(flags
& MSG_MORE
) || forced_push(tp
))
635 tcp_mark_push(tp
, tcp_write_queue_tail(sk
));
637 tcp_mark_urg(tp
, flags
);
638 __tcp_push_pending_frames(sk
, mss_now
,
639 (flags
& MSG_MORE
) ? TCP_NAGLE_CORK
: nonagle
);
643 static int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
644 unsigned int offset
, size_t len
)
646 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
649 ret
= skb_splice_bits(skb
, offset
, tss
->pipe
, min(rd_desc
->count
, len
),
652 rd_desc
->count
-= ret
;
656 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
658 /* Store TCP splice context information in read_descriptor_t. */
659 read_descriptor_t rd_desc
= {
664 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
668 * tcp_splice_read - splice data from TCP socket to a pipe
669 * @sock: socket to splice from
670 * @ppos: position (not valid)
671 * @pipe: pipe to splice to
672 * @len: number of bytes to splice
673 * @flags: splice modifier flags
676 * Will read pages from given socket and fill them into a pipe.
679 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
680 struct pipe_inode_info
*pipe
, size_t len
,
683 struct sock
*sk
= sock
->sk
;
684 struct tcp_splice_state tss
= {
693 sock_rps_record_flow(sk
);
695 * We can't seek on a socket input
704 timeo
= sock_rcvtimeo(sk
, sock
->file
->f_flags
& O_NONBLOCK
);
706 ret
= __tcp_splice_read(sk
, &tss
);
712 if (sock_flag(sk
, SOCK_DONE
))
715 ret
= sock_error(sk
);
718 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
720 if (sk
->sk_state
== TCP_CLOSE
) {
722 * This occurs when user tries to read
723 * from never connected socket.
725 if (!sock_flag(sk
, SOCK_DONE
))
733 sk_wait_data(sk
, &timeo
);
734 if (signal_pending(current
)) {
735 ret
= sock_intr_errno(timeo
);
748 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
749 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
750 signal_pending(current
))
761 EXPORT_SYMBOL(tcp_splice_read
);
763 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
)
767 /* The TCP header must be at least 32-bit aligned. */
768 size
= ALIGN(size
, 4);
770 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
772 if (sk_wmem_schedule(sk
, skb
->truesize
)) {
773 skb_reserve(skb
, sk
->sk_prot
->max_header
);
775 * Make sure that we have exactly size bytes
776 * available to the caller, no more, no less.
778 skb
->avail_size
= size
;
783 sk
->sk_prot
->enter_memory_pressure(sk
);
784 sk_stream_moderate_sndbuf(sk
);
789 static unsigned int tcp_xmit_size_goal(struct sock
*sk
, u32 mss_now
,
792 struct tcp_sock
*tp
= tcp_sk(sk
);
793 u32 xmit_size_goal
, old_size_goal
;
795 xmit_size_goal
= mss_now
;
797 if (large_allowed
&& sk_can_gso(sk
)) {
798 xmit_size_goal
= ((sk
->sk_gso_max_size
- 1) -
799 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
800 inet_csk(sk
)->icsk_ext_hdr_len
-
803 /* TSQ : try to have two TSO segments in flight */
804 xmit_size_goal
= min_t(u32
, xmit_size_goal
,
805 sysctl_tcp_limit_output_bytes
>> 1);
807 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
809 /* We try hard to avoid divides here */
810 old_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
812 if (likely(old_size_goal
<= xmit_size_goal
&&
813 old_size_goal
+ mss_now
> xmit_size_goal
)) {
814 xmit_size_goal
= old_size_goal
;
816 tp
->xmit_size_goal_segs
=
817 min_t(u16
, xmit_size_goal
/ mss_now
,
818 sk
->sk_gso_max_segs
);
819 xmit_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
823 return max(xmit_size_goal
, mss_now
);
826 static int tcp_send_mss(struct sock
*sk
, int *size_goal
, int flags
)
830 mss_now
= tcp_current_mss(sk
);
831 *size_goal
= tcp_xmit_size_goal(sk
, mss_now
, !(flags
& MSG_OOB
));
836 static ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
*page
, int offset
,
837 size_t size
, int flags
)
839 struct tcp_sock
*tp
= tcp_sk(sk
);
840 int mss_now
, size_goal
;
843 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
845 /* Wait for a connection to finish. One exception is TCP Fast Open
846 * (passive side) where data is allowed to be sent before a connection
847 * is fully established.
849 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
850 !tcp_passive_fastopen(sk
)) {
851 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
855 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
857 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
861 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
865 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
869 if (!tcp_send_head(sk
) || (copy
= size_goal
- skb
->len
) <= 0) {
871 if (!sk_stream_memory_free(sk
))
872 goto wait_for_sndbuf
;
874 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
876 goto wait_for_memory
;
885 i
= skb_shinfo(skb
)->nr_frags
;
886 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
887 if (!can_coalesce
&& i
>= MAX_SKB_FRAGS
) {
888 tcp_mark_push(tp
, skb
);
891 if (!sk_wmem_schedule(sk
, copy
))
892 goto wait_for_memory
;
895 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
898 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
900 skb_shinfo(skb
)->tx_flags
|= SKBTX_SHARED_FRAG
;
903 skb
->data_len
+= copy
;
904 skb
->truesize
+= copy
;
905 sk
->sk_wmem_queued
+= copy
;
906 sk_mem_charge(sk
, copy
);
907 skb
->ip_summed
= CHECKSUM_PARTIAL
;
908 tp
->write_seq
+= copy
;
909 TCP_SKB_CB(skb
)->end_seq
+= copy
;
910 skb_shinfo(skb
)->gso_segs
= 0;
913 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
920 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
923 if (forced_push(tp
)) {
924 tcp_mark_push(tp
, skb
);
925 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
926 } else if (skb
== tcp_send_head(sk
))
927 tcp_push_one(sk
, mss_now
);
931 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
933 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
935 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
938 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
942 if (copied
&& !(flags
& MSG_SENDPAGE_NOTLAST
))
943 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
950 return sk_stream_error(sk
, flags
, err
);
953 int tcp_sendpage(struct sock
*sk
, struct page
*page
, int offset
,
954 size_t size
, int flags
)
958 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
959 !(sk
->sk_route_caps
& NETIF_F_ALL_CSUM
))
960 return sock_no_sendpage(sk
->sk_socket
, page
, offset
, size
,
964 res
= do_tcp_sendpages(sk
, page
, offset
, size
, flags
);
968 EXPORT_SYMBOL(tcp_sendpage
);
970 static inline int select_size(const struct sock
*sk
, bool sg
)
972 const struct tcp_sock
*tp
= tcp_sk(sk
);
973 int tmp
= tp
->mss_cache
;
976 if (sk_can_gso(sk
)) {
977 /* Small frames wont use a full page:
978 * Payload will immediately follow tcp header.
980 tmp
= SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER
);
982 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
984 if (tmp
>= pgbreak
&&
985 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
993 void tcp_free_fastopen_req(struct tcp_sock
*tp
)
995 if (tp
->fastopen_req
!= NULL
) {
996 kfree(tp
->fastopen_req
);
997 tp
->fastopen_req
= NULL
;
1001 static int tcp_sendmsg_fastopen(struct sock
*sk
, struct msghdr
*msg
, int *size
)
1003 struct tcp_sock
*tp
= tcp_sk(sk
);
1006 if (!(sysctl_tcp_fastopen
& TFO_CLIENT_ENABLE
))
1008 if (tp
->fastopen_req
!= NULL
)
1009 return -EALREADY
; /* Another Fast Open is in progress */
1011 tp
->fastopen_req
= kzalloc(sizeof(struct tcp_fastopen_request
),
1013 if (unlikely(tp
->fastopen_req
== NULL
))
1015 tp
->fastopen_req
->data
= msg
;
1017 flags
= (msg
->msg_flags
& MSG_DONTWAIT
) ? O_NONBLOCK
: 0;
1018 err
= __inet_stream_connect(sk
->sk_socket
, msg
->msg_name
,
1019 msg
->msg_namelen
, flags
);
1020 *size
= tp
->fastopen_req
->copied
;
1021 tcp_free_fastopen_req(tp
);
1025 int tcp_sendmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1029 struct tcp_sock
*tp
= tcp_sk(sk
);
1030 struct sk_buff
*skb
;
1031 int iovlen
, flags
, err
, copied
= 0;
1032 int mss_now
= 0, size_goal
, copied_syn
= 0, offset
= 0;
1038 flags
= msg
->msg_flags
;
1039 if (flags
& MSG_FASTOPEN
) {
1040 err
= tcp_sendmsg_fastopen(sk
, msg
, &copied_syn
);
1041 if (err
== -EINPROGRESS
&& copied_syn
> 0)
1045 offset
= copied_syn
;
1048 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
1050 /* Wait for a connection to finish. One exception is TCP Fast Open
1051 * (passive side) where data is allowed to be sent before a connection
1052 * is fully established.
1054 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
1055 !tcp_passive_fastopen(sk
)) {
1056 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
1060 if (unlikely(tp
->repair
)) {
1061 if (tp
->repair_queue
== TCP_RECV_QUEUE
) {
1062 copied
= tcp_send_rcvq(sk
, msg
, size
);
1067 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1070 /* 'common' sending to sendq */
1073 /* This should be in poll */
1074 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
1076 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1078 /* Ok commence sending. */
1079 iovlen
= msg
->msg_iovlen
;
1084 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
1087 sg
= !!(sk
->sk_route_caps
& NETIF_F_SG
);
1089 while (--iovlen
>= 0) {
1090 size_t seglen
= iov
->iov_len
;
1091 unsigned char __user
*from
= iov
->iov_base
;
1094 if (unlikely(offset
> 0)) { /* Skip bytes copied in SYN */
1095 if (offset
>= seglen
) {
1104 while (seglen
> 0) {
1106 int max
= size_goal
;
1108 skb
= tcp_write_queue_tail(sk
);
1109 if (tcp_send_head(sk
)) {
1110 if (skb
->ip_summed
== CHECKSUM_NONE
)
1112 copy
= max
- skb
->len
;
1117 /* Allocate new segment. If the interface is SG,
1118 * allocate skb fitting to single page.
1120 if (!sk_stream_memory_free(sk
))
1121 goto wait_for_sndbuf
;
1123 skb
= sk_stream_alloc_skb(sk
,
1124 select_size(sk
, sg
),
1127 goto wait_for_memory
;
1130 * Check whether we can use HW checksum.
1132 if (sk
->sk_route_caps
& NETIF_F_ALL_CSUM
)
1133 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1135 skb_entail(sk
, skb
);
1140 /* Try to append data to the end of skb. */
1144 /* Where to copy to? */
1145 if (skb_availroom(skb
) > 0) {
1146 /* We have some space in skb head. Superb! */
1147 copy
= min_t(int, copy
, skb_availroom(skb
));
1148 err
= skb_add_data_nocache(sk
, skb
, from
, copy
);
1153 int i
= skb_shinfo(skb
)->nr_frags
;
1154 struct page_frag
*pfrag
= sk_page_frag(sk
);
1156 if (!sk_page_frag_refill(sk
, pfrag
))
1157 goto wait_for_memory
;
1159 if (!skb_can_coalesce(skb
, i
, pfrag
->page
,
1161 if (i
== MAX_SKB_FRAGS
|| !sg
) {
1162 tcp_mark_push(tp
, skb
);
1168 copy
= min_t(int, copy
, pfrag
->size
- pfrag
->offset
);
1170 if (!sk_wmem_schedule(sk
, copy
))
1171 goto wait_for_memory
;
1173 err
= skb_copy_to_page_nocache(sk
, from
, skb
,
1180 /* Update the skb. */
1182 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
1184 skb_fill_page_desc(skb
, i
, pfrag
->page
,
1185 pfrag
->offset
, copy
);
1186 get_page(pfrag
->page
);
1188 pfrag
->offset
+= copy
;
1192 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
1194 tp
->write_seq
+= copy
;
1195 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1196 skb_shinfo(skb
)->gso_segs
= 0;
1200 if ((seglen
-= copy
) == 0 && iovlen
== 0)
1203 if (skb
->len
< max
|| (flags
& MSG_OOB
) || unlikely(tp
->repair
))
1206 if (forced_push(tp
)) {
1207 tcp_mark_push(tp
, skb
);
1208 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1209 } else if (skb
== tcp_send_head(sk
))
1210 tcp_push_one(sk
, mss_now
);
1214 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1217 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
1219 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
1222 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1228 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
1230 return copied
+ copied_syn
;
1234 tcp_unlink_write_queue(skb
, sk
);
1235 /* It is the one place in all of TCP, except connection
1236 * reset, where we can be unlinking the send_head.
1238 tcp_check_send_head(sk
, skb
);
1239 sk_wmem_free_skb(sk
, skb
);
1243 if (copied
+ copied_syn
)
1246 err
= sk_stream_error(sk
, flags
, err
);
1250 EXPORT_SYMBOL(tcp_sendmsg
);
1253 * Handle reading urgent data. BSD has very simple semantics for
1254 * this, no blocking and very strange errors 8)
1257 static int tcp_recv_urg(struct sock
*sk
, struct msghdr
*msg
, int len
, int flags
)
1259 struct tcp_sock
*tp
= tcp_sk(sk
);
1261 /* No URG data to read. */
1262 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1263 tp
->urg_data
== TCP_URG_READ
)
1264 return -EINVAL
; /* Yes this is right ! */
1266 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1269 if (tp
->urg_data
& TCP_URG_VALID
) {
1271 char c
= tp
->urg_data
;
1273 if (!(flags
& MSG_PEEK
))
1274 tp
->urg_data
= TCP_URG_READ
;
1276 /* Read urgent data. */
1277 msg
->msg_flags
|= MSG_OOB
;
1280 if (!(flags
& MSG_TRUNC
))
1281 err
= memcpy_toiovec(msg
->msg_iov
, &c
, 1);
1284 msg
->msg_flags
|= MSG_TRUNC
;
1286 return err
? -EFAULT
: len
;
1289 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1292 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1293 * the available implementations agree in this case:
1294 * this call should never block, independent of the
1295 * blocking state of the socket.
1296 * Mike <pall@rz.uni-karlsruhe.de>
1301 static int tcp_peek_sndq(struct sock
*sk
, struct msghdr
*msg
, int len
)
1303 struct sk_buff
*skb
;
1304 int copied
= 0, err
= 0;
1306 /* XXX -- need to support SO_PEEK_OFF */
1308 skb_queue_walk(&sk
->sk_write_queue
, skb
) {
1309 err
= skb_copy_datagram_iovec(skb
, 0, msg
->msg_iov
, skb
->len
);
1316 return err
?: copied
;
1319 /* Clean up the receive buffer for full frames taken by the user,
1320 * then send an ACK if necessary. COPIED is the number of bytes
1321 * tcp_recvmsg has given to the user so far, it speeds up the
1322 * calculation of whether or not we must ACK for the sake of
1325 void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1327 struct tcp_sock
*tp
= tcp_sk(sk
);
1328 bool time_to_ack
= false;
1330 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1332 WARN(skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
),
1333 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1334 tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
);
1336 if (inet_csk_ack_scheduled(sk
)) {
1337 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1338 /* Delayed ACKs frequently hit locked sockets during bulk
1340 if (icsk
->icsk_ack
.blocked
||
1341 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1342 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1344 * If this read emptied read buffer, we send ACK, if
1345 * connection is not bidirectional, user drained
1346 * receive buffer and there was a small segment
1350 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1351 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1352 !icsk
->icsk_ack
.pingpong
)) &&
1353 !atomic_read(&sk
->sk_rmem_alloc
)))
1357 /* We send an ACK if we can now advertise a non-zero window
1358 * which has been raised "significantly".
1360 * Even if window raised up to infinity, do not send window open ACK
1361 * in states, where we will not receive more. It is useless.
1363 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1364 __u32 rcv_window_now
= tcp_receive_window(tp
);
1366 /* Optimize, __tcp_select_window() is not cheap. */
1367 if (2*rcv_window_now
<= tp
->window_clamp
) {
1368 __u32 new_window
= __tcp_select_window(sk
);
1370 /* Send ACK now, if this read freed lots of space
1371 * in our buffer. Certainly, new_window is new window.
1372 * We can advertise it now, if it is not less than current one.
1373 * "Lots" means "at least twice" here.
1375 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1383 static void tcp_prequeue_process(struct sock
*sk
)
1385 struct sk_buff
*skb
;
1386 struct tcp_sock
*tp
= tcp_sk(sk
);
1388 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPPREQUEUED
);
1390 /* RX process wants to run with disabled BHs, though it is not
1393 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
1394 sk_backlog_rcv(sk
, skb
);
1397 /* Clear memory counter. */
1398 tp
->ucopy
.memory
= 0;
1401 #ifdef CONFIG_NET_DMA
1402 static void tcp_service_net_dma(struct sock
*sk
, bool wait
)
1404 dma_cookie_t done
, used
;
1405 dma_cookie_t last_issued
;
1406 struct tcp_sock
*tp
= tcp_sk(sk
);
1408 if (!tp
->ucopy
.dma_chan
)
1411 last_issued
= tp
->ucopy
.dma_cookie
;
1412 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1415 if (dma_async_is_tx_complete(tp
->ucopy
.dma_chan
,
1417 &used
) == DMA_SUCCESS
) {
1418 /* Safe to free early-copied skbs now */
1419 __skb_queue_purge(&sk
->sk_async_wait_queue
);
1422 struct sk_buff
*skb
;
1423 while ((skb
= skb_peek(&sk
->sk_async_wait_queue
)) &&
1424 (dma_async_is_complete(skb
->dma_cookie
, done
,
1425 used
) == DMA_SUCCESS
)) {
1426 __skb_dequeue(&sk
->sk_async_wait_queue
);
1434 static struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1436 struct sk_buff
*skb
;
1439 while ((skb
= skb_peek(&sk
->sk_receive_queue
)) != NULL
) {
1440 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1441 if (tcp_hdr(skb
)->syn
)
1443 if (offset
< skb
->len
|| tcp_hdr(skb
)->fin
) {
1447 /* This looks weird, but this can happen if TCP collapsing
1448 * splitted a fat GRO packet, while we released socket lock
1449 * in skb_splice_bits()
1451 sk_eat_skb(sk
, skb
, false);
1457 * This routine provides an alternative to tcp_recvmsg() for routines
1458 * that would like to handle copying from skbuffs directly in 'sendfile'
1461 * - It is assumed that the socket was locked by the caller.
1462 * - The routine does not block.
1463 * - At present, there is no support for reading OOB data
1464 * or for 'peeking' the socket using this routine
1465 * (although both would be easy to implement).
1467 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1468 sk_read_actor_t recv_actor
)
1470 struct sk_buff
*skb
;
1471 struct tcp_sock
*tp
= tcp_sk(sk
);
1472 u32 seq
= tp
->copied_seq
;
1476 if (sk
->sk_state
== TCP_LISTEN
)
1478 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1479 if (offset
< skb
->len
) {
1483 len
= skb
->len
- offset
;
1484 /* Stop reading if we hit a patch of urgent data */
1486 u32 urg_offset
= tp
->urg_seq
- seq
;
1487 if (urg_offset
< len
)
1492 used
= recv_actor(desc
, skb
, offset
, len
);
1497 } else if (used
<= len
) {
1502 /* If recv_actor drops the lock (e.g. TCP splice
1503 * receive) the skb pointer might be invalid when
1504 * getting here: tcp_collapse might have deleted it
1505 * while aggregating skbs from the socket queue.
1507 skb
= tcp_recv_skb(sk
, seq
- 1, &offset
);
1510 /* TCP coalescing might have appended data to the skb.
1511 * Try to splice more frags
1513 if (offset
+ 1 != skb
->len
)
1516 if (tcp_hdr(skb
)->fin
) {
1517 sk_eat_skb(sk
, skb
, false);
1521 sk_eat_skb(sk
, skb
, false);
1524 tp
->copied_seq
= seq
;
1526 tp
->copied_seq
= seq
;
1528 tcp_rcv_space_adjust(sk
);
1530 /* Clean up data we have read: This will do ACK frames. */
1532 tcp_recv_skb(sk
, seq
, &offset
);
1533 tcp_cleanup_rbuf(sk
, copied
);
1537 EXPORT_SYMBOL(tcp_read_sock
);
1540 * This routine copies from a sock struct into the user buffer.
1542 * Technical note: in 2.3 we work on _locked_ socket, so that
1543 * tricks with *seq access order and skb->users are not required.
1544 * Probably, code can be easily improved even more.
1547 int tcp_recvmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1548 size_t len
, int nonblock
, int flags
, int *addr_len
)
1550 struct tcp_sock
*tp
= tcp_sk(sk
);
1556 int target
; /* Read at least this many bytes */
1558 struct task_struct
*user_recv
= NULL
;
1559 bool copied_early
= false;
1560 struct sk_buff
*skb
;
1566 if (sk
->sk_state
== TCP_LISTEN
)
1569 timeo
= sock_rcvtimeo(sk
, nonblock
);
1571 /* Urgent data needs to be handled specially. */
1572 if (flags
& MSG_OOB
)
1575 if (unlikely(tp
->repair
)) {
1577 if (!(flags
& MSG_PEEK
))
1580 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
1584 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1587 /* 'common' recv queue MSG_PEEK-ing */
1590 seq
= &tp
->copied_seq
;
1591 if (flags
& MSG_PEEK
) {
1592 peek_seq
= tp
->copied_seq
;
1596 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1598 #ifdef CONFIG_NET_DMA
1599 tp
->ucopy
.dma_chan
= NULL
;
1601 skb
= skb_peek_tail(&sk
->sk_receive_queue
);
1606 available
= TCP_SKB_CB(skb
)->seq
+ skb
->len
- (*seq
);
1607 if ((available
< target
) &&
1608 (len
> sysctl_tcp_dma_copybreak
) && !(flags
& MSG_PEEK
) &&
1609 !sysctl_tcp_low_latency
&&
1610 net_dma_find_channel()) {
1611 preempt_enable_no_resched();
1612 tp
->ucopy
.pinned_list
=
1613 dma_pin_iovec_pages(msg
->msg_iov
, len
);
1615 preempt_enable_no_resched();
1623 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1624 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1627 if (signal_pending(current
)) {
1628 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1633 /* Next get a buffer. */
1635 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1636 /* Now that we have two receive queues this
1639 if (WARN(before(*seq
, TCP_SKB_CB(skb
)->seq
),
1640 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1641 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
,
1645 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1646 if (tcp_hdr(skb
)->syn
)
1648 if (offset
< skb
->len
)
1650 if (tcp_hdr(skb
)->fin
)
1652 WARN(!(flags
& MSG_PEEK
),
1653 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1654 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
, flags
);
1657 /* Well, if we have backlog, try to process it now yet. */
1659 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1664 sk
->sk_state
== TCP_CLOSE
||
1665 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1667 signal_pending(current
))
1670 if (sock_flag(sk
, SOCK_DONE
))
1674 copied
= sock_error(sk
);
1678 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1681 if (sk
->sk_state
== TCP_CLOSE
) {
1682 if (!sock_flag(sk
, SOCK_DONE
)) {
1683 /* This occurs when user tries to read
1684 * from never connected socket.
1697 if (signal_pending(current
)) {
1698 copied
= sock_intr_errno(timeo
);
1703 tcp_cleanup_rbuf(sk
, copied
);
1705 if (!sysctl_tcp_low_latency
&& tp
->ucopy
.task
== user_recv
) {
1706 /* Install new reader */
1707 if (!user_recv
&& !(flags
& (MSG_TRUNC
| MSG_PEEK
))) {
1708 user_recv
= current
;
1709 tp
->ucopy
.task
= user_recv
;
1710 tp
->ucopy
.iov
= msg
->msg_iov
;
1713 tp
->ucopy
.len
= len
;
1715 WARN_ON(tp
->copied_seq
!= tp
->rcv_nxt
&&
1716 !(flags
& (MSG_PEEK
| MSG_TRUNC
)));
1718 /* Ugly... If prequeue is not empty, we have to
1719 * process it before releasing socket, otherwise
1720 * order will be broken at second iteration.
1721 * More elegant solution is required!!!
1723 * Look: we have the following (pseudo)queues:
1725 * 1. packets in flight
1730 * Each queue can be processed only if the next ones
1731 * are empty. At this point we have empty receive_queue.
1732 * But prequeue _can_ be not empty after 2nd iteration,
1733 * when we jumped to start of loop because backlog
1734 * processing added something to receive_queue.
1735 * We cannot release_sock(), because backlog contains
1736 * packets arrived _after_ prequeued ones.
1738 * Shortly, algorithm is clear --- to process all
1739 * the queues in order. We could make it more directly,
1740 * requeueing packets from backlog to prequeue, if
1741 * is not empty. It is more elegant, but eats cycles,
1744 if (!skb_queue_empty(&tp
->ucopy
.prequeue
))
1747 /* __ Set realtime policy in scheduler __ */
1750 #ifdef CONFIG_NET_DMA
1751 if (tp
->ucopy
.dma_chan
) {
1752 if (tp
->rcv_wnd
== 0 &&
1753 !skb_queue_empty(&sk
->sk_async_wait_queue
)) {
1754 tcp_service_net_dma(sk
, true);
1755 tcp_cleanup_rbuf(sk
, copied
);
1757 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1760 if (copied
>= target
) {
1761 /* Do not sleep, just process backlog. */
1765 sk_wait_data(sk
, &timeo
);
1767 #ifdef CONFIG_NET_DMA
1768 tcp_service_net_dma(sk
, false); /* Don't block */
1769 tp
->ucopy
.wakeup
= 0;
1775 /* __ Restore normal policy in scheduler __ */
1777 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1778 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG
, chunk
);
1783 if (tp
->rcv_nxt
== tp
->copied_seq
&&
1784 !skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1786 tcp_prequeue_process(sk
);
1788 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1789 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1795 if ((flags
& MSG_PEEK
) &&
1796 (peek_seq
- copied
- urg_hole
!= tp
->copied_seq
)) {
1797 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1799 task_pid_nr(current
));
1800 peek_seq
= tp
->copied_seq
;
1805 /* Ok so how much can we use? */
1806 used
= skb
->len
- offset
;
1810 /* Do we have urgent data here? */
1812 u32 urg_offset
= tp
->urg_seq
- *seq
;
1813 if (urg_offset
< used
) {
1815 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1828 if (!(flags
& MSG_TRUNC
)) {
1829 #ifdef CONFIG_NET_DMA
1830 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
1831 tp
->ucopy
.dma_chan
= net_dma_find_channel();
1833 if (tp
->ucopy
.dma_chan
) {
1834 tp
->ucopy
.dma_cookie
= dma_skb_copy_datagram_iovec(
1835 tp
->ucopy
.dma_chan
, skb
, offset
,
1837 tp
->ucopy
.pinned_list
);
1839 if (tp
->ucopy
.dma_cookie
< 0) {
1841 pr_alert("%s: dma_cookie < 0\n",
1844 /* Exception. Bailout! */
1850 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1852 if ((offset
+ used
) == skb
->len
)
1853 copied_early
= true;
1858 err
= skb_copy_datagram_iovec(skb
, offset
,
1859 msg
->msg_iov
, used
);
1861 /* Exception. Bailout! */
1873 tcp_rcv_space_adjust(sk
);
1876 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1878 tcp_fast_path_check(sk
);
1880 if (used
+ offset
< skb
->len
)
1883 if (tcp_hdr(skb
)->fin
)
1885 if (!(flags
& MSG_PEEK
)) {
1886 sk_eat_skb(sk
, skb
, copied_early
);
1887 copied_early
= false;
1892 /* Process the FIN. */
1894 if (!(flags
& MSG_PEEK
)) {
1895 sk_eat_skb(sk
, skb
, copied_early
);
1896 copied_early
= false;
1902 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1905 tp
->ucopy
.len
= copied
> 0 ? len
: 0;
1907 tcp_prequeue_process(sk
);
1909 if (copied
> 0 && (chunk
= len
- tp
->ucopy
.len
) != 0) {
1910 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1916 tp
->ucopy
.task
= NULL
;
1920 #ifdef CONFIG_NET_DMA
1921 tcp_service_net_dma(sk
, true); /* Wait for queue to drain */
1922 tp
->ucopy
.dma_chan
= NULL
;
1924 if (tp
->ucopy
.pinned_list
) {
1925 dma_unpin_iovec_pages(tp
->ucopy
.pinned_list
);
1926 tp
->ucopy
.pinned_list
= NULL
;
1930 /* According to UNIX98, msg_name/msg_namelen are ignored
1931 * on connected socket. I was just happy when found this 8) --ANK
1934 /* Clean up data we have read: This will do ACK frames. */
1935 tcp_cleanup_rbuf(sk
, copied
);
1945 err
= tcp_recv_urg(sk
, msg
, len
, flags
);
1949 err
= tcp_peek_sndq(sk
, msg
, len
);
1952 EXPORT_SYMBOL(tcp_recvmsg
);
1954 void tcp_set_state(struct sock
*sk
, int state
)
1956 int oldstate
= sk
->sk_state
;
1959 case TCP_ESTABLISHED
:
1960 if (oldstate
!= TCP_ESTABLISHED
)
1961 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1965 if (oldstate
== TCP_CLOSE_WAIT
|| oldstate
== TCP_ESTABLISHED
)
1966 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ESTABRESETS
);
1968 sk
->sk_prot
->unhash(sk
);
1969 if (inet_csk(sk
)->icsk_bind_hash
&&
1970 !(sk
->sk_userlocks
& SOCK_BINDPORT_LOCK
))
1974 if (oldstate
== TCP_ESTABLISHED
)
1975 TCP_DEC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1978 /* Change state AFTER socket is unhashed to avoid closed
1979 * socket sitting in hash tables.
1981 sk
->sk_state
= state
;
1984 SOCK_DEBUG(sk
, "TCP sk=%p, State %s -> %s\n", sk
, statename
[oldstate
], statename
[state
]);
1987 EXPORT_SYMBOL_GPL(tcp_set_state
);
1990 * State processing on a close. This implements the state shift for
1991 * sending our FIN frame. Note that we only send a FIN for some
1992 * states. A shutdown() may have already sent the FIN, or we may be
1996 static const unsigned char new_state
[16] = {
1997 /* current state: new state: action: */
1998 /* (Invalid) */ TCP_CLOSE
,
1999 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2000 /* TCP_SYN_SENT */ TCP_CLOSE
,
2001 /* TCP_SYN_RECV */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2002 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1
,
2003 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2
,
2004 /* TCP_TIME_WAIT */ TCP_CLOSE
,
2005 /* TCP_CLOSE */ TCP_CLOSE
,
2006 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK
| TCP_ACTION_FIN
,
2007 /* TCP_LAST_ACK */ TCP_LAST_ACK
,
2008 /* TCP_LISTEN */ TCP_CLOSE
,
2009 /* TCP_CLOSING */ TCP_CLOSING
,
2012 static int tcp_close_state(struct sock
*sk
)
2014 int next
= (int)new_state
[sk
->sk_state
];
2015 int ns
= next
& TCP_STATE_MASK
;
2017 tcp_set_state(sk
, ns
);
2019 return next
& TCP_ACTION_FIN
;
2023 * Shutdown the sending side of a connection. Much like close except
2024 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2027 void tcp_shutdown(struct sock
*sk
, int how
)
2029 /* We need to grab some memory, and put together a FIN,
2030 * and then put it into the queue to be sent.
2031 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2033 if (!(how
& SEND_SHUTDOWN
))
2036 /* If we've already sent a FIN, or it's a closed state, skip this. */
2037 if ((1 << sk
->sk_state
) &
2038 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
2039 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
2040 /* Clear out any half completed packets. FIN if needed. */
2041 if (tcp_close_state(sk
))
2045 EXPORT_SYMBOL(tcp_shutdown
);
2047 bool tcp_check_oom(struct sock
*sk
, int shift
)
2049 bool too_many_orphans
, out_of_socket_memory
;
2051 too_many_orphans
= tcp_too_many_orphans(sk
, shift
);
2052 out_of_socket_memory
= tcp_out_of_memory(sk
);
2054 if (too_many_orphans
)
2055 net_info_ratelimited("too many orphaned sockets\n");
2056 if (out_of_socket_memory
)
2057 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2058 return too_many_orphans
|| out_of_socket_memory
;
2061 void tcp_close(struct sock
*sk
, long timeout
)
2063 struct sk_buff
*skb
;
2064 int data_was_unread
= 0;
2068 sk
->sk_shutdown
= SHUTDOWN_MASK
;
2070 if (sk
->sk_state
== TCP_LISTEN
) {
2071 tcp_set_state(sk
, TCP_CLOSE
);
2074 inet_csk_listen_stop(sk
);
2076 goto adjudge_to_death
;
2079 /* We need to flush the recv. buffs. We do this only on the
2080 * descriptor close, not protocol-sourced closes, because the
2081 * reader process may not have drained the data yet!
2083 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
2084 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
-
2086 data_was_unread
+= len
;
2092 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2093 if (sk
->sk_state
== TCP_CLOSE
)
2094 goto adjudge_to_death
;
2096 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2097 * data was lost. To witness the awful effects of the old behavior of
2098 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2099 * GET in an FTP client, suspend the process, wait for the client to
2100 * advertise a zero window, then kill -9 the FTP client, wheee...
2101 * Note: timeout is always zero in such a case.
2103 if (unlikely(tcp_sk(sk
)->repair
)) {
2104 sk
->sk_prot
->disconnect(sk
, 0);
2105 } else if (data_was_unread
) {
2106 /* Unread data was tossed, zap the connection. */
2107 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONCLOSE
);
2108 tcp_set_state(sk
, TCP_CLOSE
);
2109 tcp_send_active_reset(sk
, sk
->sk_allocation
);
2110 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
2111 /* Check zero linger _after_ checking for unread data. */
2112 sk
->sk_prot
->disconnect(sk
, 0);
2113 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
2114 } else if (tcp_close_state(sk
)) {
2115 /* We FIN if the application ate all the data before
2116 * zapping the connection.
2119 /* RED-PEN. Formally speaking, we have broken TCP state
2120 * machine. State transitions:
2122 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2123 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2124 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2126 * are legal only when FIN has been sent (i.e. in window),
2127 * rather than queued out of window. Purists blame.
2129 * F.e. "RFC state" is ESTABLISHED,
2130 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2132 * The visible declinations are that sometimes
2133 * we enter time-wait state, when it is not required really
2134 * (harmless), do not send active resets, when they are
2135 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2136 * they look as CLOSING or LAST_ACK for Linux)
2137 * Probably, I missed some more holelets.
2139 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2140 * in a single packet! (May consider it later but will
2141 * probably need API support or TCP_CORK SYN-ACK until
2142 * data is written and socket is closed.)
2147 sk_stream_wait_close(sk
, timeout
);
2150 state
= sk
->sk_state
;
2154 /* It is the last release_sock in its life. It will remove backlog. */
2158 /* Now socket is owned by kernel and we acquire BH lock
2159 to finish close. No need to check for user refs.
2163 WARN_ON(sock_owned_by_user(sk
));
2165 percpu_counter_inc(sk
->sk_prot
->orphan_count
);
2167 /* Have we already been destroyed by a softirq or backlog? */
2168 if (state
!= TCP_CLOSE
&& sk
->sk_state
== TCP_CLOSE
)
2171 /* This is a (useful) BSD violating of the RFC. There is a
2172 * problem with TCP as specified in that the other end could
2173 * keep a socket open forever with no application left this end.
2174 * We use a 3 minute timeout (about the same as BSD) then kill
2175 * our end. If they send after that then tough - BUT: long enough
2176 * that we won't make the old 4*rto = almost no time - whoops
2179 * Nope, it was not mistake. It is really desired behaviour
2180 * f.e. on http servers, when such sockets are useless, but
2181 * consume significant resources. Let's do it with special
2182 * linger2 option. --ANK
2185 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
2186 struct tcp_sock
*tp
= tcp_sk(sk
);
2187 if (tp
->linger2
< 0) {
2188 tcp_set_state(sk
, TCP_CLOSE
);
2189 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2190 NET_INC_STATS_BH(sock_net(sk
),
2191 LINUX_MIB_TCPABORTONLINGER
);
2193 const int tmo
= tcp_fin_time(sk
);
2195 if (tmo
> TCP_TIMEWAIT_LEN
) {
2196 inet_csk_reset_keepalive_timer(sk
,
2197 tmo
- TCP_TIMEWAIT_LEN
);
2199 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
2204 if (sk
->sk_state
!= TCP_CLOSE
) {
2206 if (tcp_check_oom(sk
, 0)) {
2207 tcp_set_state(sk
, TCP_CLOSE
);
2208 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2209 NET_INC_STATS_BH(sock_net(sk
),
2210 LINUX_MIB_TCPABORTONMEMORY
);
2214 if (sk
->sk_state
== TCP_CLOSE
) {
2215 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
2216 /* We could get here with a non-NULL req if the socket is
2217 * aborted (e.g., closed with unread data) before 3WHS
2221 reqsk_fastopen_remove(sk
, req
, false);
2222 inet_csk_destroy_sock(sk
);
2224 /* Otherwise, socket is reprieved until protocol close. */
2231 EXPORT_SYMBOL(tcp_close
);
2233 /* These states need RST on ABORT according to RFC793 */
2235 static inline bool tcp_need_reset(int state
)
2237 return (1 << state
) &
2238 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
2239 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
2242 int tcp_disconnect(struct sock
*sk
, int flags
)
2244 struct inet_sock
*inet
= inet_sk(sk
);
2245 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2246 struct tcp_sock
*tp
= tcp_sk(sk
);
2248 int old_state
= sk
->sk_state
;
2250 if (old_state
!= TCP_CLOSE
)
2251 tcp_set_state(sk
, TCP_CLOSE
);
2253 /* ABORT function of RFC793 */
2254 if (old_state
== TCP_LISTEN
) {
2255 inet_csk_listen_stop(sk
);
2256 } else if (unlikely(tp
->repair
)) {
2257 sk
->sk_err
= ECONNABORTED
;
2258 } else if (tcp_need_reset(old_state
) ||
2259 (tp
->snd_nxt
!= tp
->write_seq
&&
2260 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
2261 /* The last check adjusts for discrepancy of Linux wrt. RFC
2264 tcp_send_active_reset(sk
, gfp_any());
2265 sk
->sk_err
= ECONNRESET
;
2266 } else if (old_state
== TCP_SYN_SENT
)
2267 sk
->sk_err
= ECONNRESET
;
2269 tcp_clear_xmit_timers(sk
);
2270 __skb_queue_purge(&sk
->sk_receive_queue
);
2271 tcp_write_queue_purge(sk
);
2272 __skb_queue_purge(&tp
->out_of_order_queue
);
2273 #ifdef CONFIG_NET_DMA
2274 __skb_queue_purge(&sk
->sk_async_wait_queue
);
2277 inet
->inet_dport
= 0;
2279 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
2280 inet_reset_saddr(sk
);
2282 sk
->sk_shutdown
= 0;
2283 sock_reset_flag(sk
, SOCK_DONE
);
2285 if ((tp
->write_seq
+= tp
->max_window
+ 2) == 0)
2287 icsk
->icsk_backoff
= 0;
2289 icsk
->icsk_probes_out
= 0;
2290 tp
->packets_out
= 0;
2291 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
2292 tp
->snd_cwnd_cnt
= 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
);
2720 tp
->tsoffset
= val
- tcp_time_stamp
;
2731 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2732 unsigned int optlen
)
2734 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2736 if (level
!= SOL_TCP
)
2737 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
2739 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2741 EXPORT_SYMBOL(tcp_setsockopt
);
2743 #ifdef CONFIG_COMPAT
2744 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
2745 char __user
*optval
, unsigned int optlen
)
2747 if (level
!= SOL_TCP
)
2748 return inet_csk_compat_setsockopt(sk
, level
, optname
,
2750 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2752 EXPORT_SYMBOL(compat_tcp_setsockopt
);
2755 /* Return information about state of tcp endpoint in API format. */
2756 void tcp_get_info(const struct sock
*sk
, struct tcp_info
*info
)
2758 const struct tcp_sock
*tp
= tcp_sk(sk
);
2759 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2760 u32 now
= tcp_time_stamp
;
2762 memset(info
, 0, sizeof(*info
));
2764 info
->tcpi_state
= sk
->sk_state
;
2765 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
2766 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
2767 info
->tcpi_probes
= icsk
->icsk_probes_out
;
2768 info
->tcpi_backoff
= icsk
->icsk_backoff
;
2770 if (tp
->rx_opt
.tstamp_ok
)
2771 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
2772 if (tcp_is_sack(tp
))
2773 info
->tcpi_options
|= TCPI_OPT_SACK
;
2774 if (tp
->rx_opt
.wscale_ok
) {
2775 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
2776 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
2777 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
2780 if (tp
->ecn_flags
& TCP_ECN_OK
)
2781 info
->tcpi_options
|= TCPI_OPT_ECN
;
2782 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
2783 info
->tcpi_options
|= TCPI_OPT_ECN_SEEN
;
2784 if (tp
->syn_data_acked
)
2785 info
->tcpi_options
|= TCPI_OPT_SYN_DATA
;
2787 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
2788 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
2789 info
->tcpi_snd_mss
= tp
->mss_cache
;
2790 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
2792 if (sk
->sk_state
== TCP_LISTEN
) {
2793 info
->tcpi_unacked
= sk
->sk_ack_backlog
;
2794 info
->tcpi_sacked
= sk
->sk_max_ack_backlog
;
2796 info
->tcpi_unacked
= tp
->packets_out
;
2797 info
->tcpi_sacked
= tp
->sacked_out
;
2799 info
->tcpi_lost
= tp
->lost_out
;
2800 info
->tcpi_retrans
= tp
->retrans_out
;
2801 info
->tcpi_fackets
= tp
->fackets_out
;
2803 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
2804 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
2805 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
2807 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
2808 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
2809 info
->tcpi_rtt
= jiffies_to_usecs(tp
->srtt
)>>3;
2810 info
->tcpi_rttvar
= jiffies_to_usecs(tp
->mdev
)>>2;
2811 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
2812 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
2813 info
->tcpi_advmss
= tp
->advmss
;
2814 info
->tcpi_reordering
= tp
->reordering
;
2816 info
->tcpi_rcv_rtt
= jiffies_to_usecs(tp
->rcv_rtt_est
.rtt
)>>3;
2817 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
2819 info
->tcpi_total_retrans
= tp
->total_retrans
;
2821 EXPORT_SYMBOL_GPL(tcp_get_info
);
2823 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
2824 int optname
, char __user
*optval
, int __user
*optlen
)
2826 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2827 struct tcp_sock
*tp
= tcp_sk(sk
);
2830 if (get_user(len
, optlen
))
2833 len
= min_t(unsigned int, len
, sizeof(int));
2840 val
= tp
->mss_cache
;
2841 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
2842 val
= tp
->rx_opt
.user_mss
;
2844 val
= tp
->rx_opt
.mss_clamp
;
2847 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
2850 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
2853 val
= keepalive_time_when(tp
) / HZ
;
2856 val
= keepalive_intvl_when(tp
) / HZ
;
2859 val
= keepalive_probes(tp
);
2862 val
= icsk
->icsk_syn_retries
? : sysctl_tcp_syn_retries
;
2867 val
= (val
? : sysctl_tcp_fin_timeout
) / HZ
;
2869 case TCP_DEFER_ACCEPT
:
2870 val
= retrans_to_secs(icsk
->icsk_accept_queue
.rskq_defer_accept
,
2871 TCP_TIMEOUT_INIT
/ HZ
, TCP_RTO_MAX
/ HZ
);
2873 case TCP_WINDOW_CLAMP
:
2874 val
= tp
->window_clamp
;
2877 struct tcp_info info
;
2879 if (get_user(len
, optlen
))
2882 tcp_get_info(sk
, &info
);
2884 len
= min_t(unsigned int, len
, sizeof(info
));
2885 if (put_user(len
, optlen
))
2887 if (copy_to_user(optval
, &info
, len
))
2892 val
= !icsk
->icsk_ack
.pingpong
;
2895 case TCP_CONGESTION
:
2896 if (get_user(len
, optlen
))
2898 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
2899 if (put_user(len
, optlen
))
2901 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
2905 case TCP_COOKIE_TRANSACTIONS
: {
2906 struct tcp_cookie_transactions ctd
;
2907 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
2909 if (get_user(len
, optlen
))
2911 if (len
< sizeof(ctd
))
2914 memset(&ctd
, 0, sizeof(ctd
));
2915 ctd
.tcpct_flags
= (tp
->rx_opt
.cookie_in_always
?
2916 TCP_COOKIE_IN_ALWAYS
: 0)
2917 | (tp
->rx_opt
.cookie_out_never
?
2918 TCP_COOKIE_OUT_NEVER
: 0);
2921 ctd
.tcpct_flags
|= (cvp
->s_data_in
?
2923 | (cvp
->s_data_out
?
2924 TCP_S_DATA_OUT
: 0);
2926 ctd
.tcpct_cookie_desired
= cvp
->cookie_desired
;
2927 ctd
.tcpct_s_data_desired
= cvp
->s_data_desired
;
2929 memcpy(&ctd
.tcpct_value
[0], &cvp
->cookie_pair
[0],
2930 cvp
->cookie_pair_size
);
2931 ctd
.tcpct_used
= cvp
->cookie_pair_size
;
2934 if (put_user(sizeof(ctd
), optlen
))
2936 if (copy_to_user(optval
, &ctd
, sizeof(ctd
)))
2940 case TCP_THIN_LINEAR_TIMEOUTS
:
2943 case TCP_THIN_DUPACK
:
2944 val
= tp
->thin_dupack
;
2951 case TCP_REPAIR_QUEUE
:
2953 val
= tp
->repair_queue
;
2959 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2960 val
= tp
->write_seq
;
2961 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2967 case TCP_USER_TIMEOUT
:
2968 val
= jiffies_to_msecs(icsk
->icsk_user_timeout
);
2971 val
= tcp_time_stamp
+ tp
->tsoffset
;
2974 return -ENOPROTOOPT
;
2977 if (put_user(len
, optlen
))
2979 if (copy_to_user(optval
, &val
, len
))
2984 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2987 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2989 if (level
!= SOL_TCP
)
2990 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
2992 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2994 EXPORT_SYMBOL(tcp_getsockopt
);
2996 #ifdef CONFIG_COMPAT
2997 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
2998 char __user
*optval
, int __user
*optlen
)
3000 if (level
!= SOL_TCP
)
3001 return inet_csk_compat_getsockopt(sk
, level
, optname
,
3003 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
3005 EXPORT_SYMBOL(compat_tcp_getsockopt
);
3008 struct sk_buff
*tcp_tso_segment(struct sk_buff
*skb
,
3009 netdev_features_t features
)
3011 struct sk_buff
*segs
= ERR_PTR(-EINVAL
);
3016 unsigned int oldlen
;
3019 if (!pskb_may_pull(skb
, sizeof(*th
)))
3023 thlen
= th
->doff
* 4;
3024 if (thlen
< sizeof(*th
))
3027 if (!pskb_may_pull(skb
, thlen
))
3030 oldlen
= (u16
)~skb
->len
;
3031 __skb_pull(skb
, thlen
);
3033 mss
= skb_shinfo(skb
)->gso_size
;
3034 if (unlikely(skb
->len
<= mss
))
3037 if (skb_gso_ok(skb
, features
| NETIF_F_GSO_ROBUST
)) {
3038 /* Packet is from an untrusted source, reset gso_segs. */
3039 int type
= skb_shinfo(skb
)->gso_type
;
3048 !(type
& (SKB_GSO_TCPV4
| SKB_GSO_TCPV6
))))
3051 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss
);
3057 segs
= skb_segment(skb
, features
);
3061 delta
= htonl(oldlen
+ (thlen
+ mss
));
3065 seq
= ntohl(th
->seq
);
3068 th
->fin
= th
->psh
= 0;
3070 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
3071 (__force u32
)delta
));
3072 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
3074 csum_fold(csum_partial(skb_transport_header(skb
),
3081 th
->seq
= htonl(seq
);
3083 } while (skb
->next
);
3085 delta
= htonl(oldlen
+ (skb
->tail
- skb
->transport_header
) +
3087 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
3088 (__force u32
)delta
));
3089 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
3090 th
->check
= csum_fold(csum_partial(skb_transport_header(skb
),
3096 EXPORT_SYMBOL(tcp_tso_segment
);
3098 struct sk_buff
**tcp_gro_receive(struct sk_buff
**head
, struct sk_buff
*skb
)
3100 struct sk_buff
**pp
= NULL
;
3107 unsigned int mss
= 1;
3113 off
= skb_gro_offset(skb
);
3114 hlen
= off
+ sizeof(*th
);
3115 th
= skb_gro_header_fast(skb
, off
);
3116 if (skb_gro_header_hard(skb
, hlen
)) {
3117 th
= skb_gro_header_slow(skb
, hlen
, off
);
3122 thlen
= th
->doff
* 4;
3123 if (thlen
< sizeof(*th
))
3127 if (skb_gro_header_hard(skb
, hlen
)) {
3128 th
= skb_gro_header_slow(skb
, hlen
, off
);
3133 skb_gro_pull(skb
, thlen
);
3135 len
= skb_gro_len(skb
);
3136 flags
= tcp_flag_word(th
);
3138 for (; (p
= *head
); head
= &p
->next
) {
3139 if (!NAPI_GRO_CB(p
)->same_flow
)
3144 if (*(u32
*)&th
->source
^ *(u32
*)&th2
->source
) {
3145 NAPI_GRO_CB(p
)->same_flow
= 0;
3152 goto out_check_final
;
3155 flush
= NAPI_GRO_CB(p
)->flush
;
3156 flush
|= (__force
int)(flags
& TCP_FLAG_CWR
);
3157 flush
|= (__force
int)((flags
^ tcp_flag_word(th2
)) &
3158 ~(TCP_FLAG_CWR
| TCP_FLAG_FIN
| TCP_FLAG_PSH
));
3159 flush
|= (__force
int)(th
->ack_seq
^ th2
->ack_seq
);
3160 for (i
= sizeof(*th
); i
< thlen
; i
+= 4)
3161 flush
|= *(u32
*)((u8
*)th
+ i
) ^
3162 *(u32
*)((u8
*)th2
+ i
);
3164 mss
= skb_shinfo(p
)->gso_size
;
3166 flush
|= (len
- 1) >= mss
;
3167 flush
|= (ntohl(th2
->seq
) + skb_gro_len(p
)) ^ ntohl(th
->seq
);
3169 if (flush
|| skb_gro_receive(head
, skb
)) {
3171 goto out_check_final
;
3176 tcp_flag_word(th2
) |= flags
& (TCP_FLAG_FIN
| TCP_FLAG_PSH
);
3180 flush
|= (__force
int)(flags
& (TCP_FLAG_URG
| TCP_FLAG_PSH
|
3181 TCP_FLAG_RST
| TCP_FLAG_SYN
|
3184 if (p
&& (!NAPI_GRO_CB(skb
)->same_flow
|| flush
))
3188 NAPI_GRO_CB(skb
)->flush
|= flush
;
3192 EXPORT_SYMBOL(tcp_gro_receive
);
3194 int tcp_gro_complete(struct sk_buff
*skb
)
3196 struct tcphdr
*th
= tcp_hdr(skb
);
3198 skb
->csum_start
= skb_transport_header(skb
) - skb
->head
;
3199 skb
->csum_offset
= offsetof(struct tcphdr
, check
);
3200 skb
->ip_summed
= CHECKSUM_PARTIAL
;
3202 skb_shinfo(skb
)->gso_segs
= NAPI_GRO_CB(skb
)->count
;
3205 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
3209 EXPORT_SYMBOL(tcp_gro_complete
);
3211 #ifdef CONFIG_TCP_MD5SIG
3212 static unsigned long tcp_md5sig_users
;
3213 static struct tcp_md5sig_pool __percpu
*tcp_md5sig_pool
;
3214 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock
);
3216 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu
*pool
)
3220 for_each_possible_cpu(cpu
) {
3221 struct tcp_md5sig_pool
*p
= per_cpu_ptr(pool
, cpu
);
3223 if (p
->md5_desc
.tfm
)
3224 crypto_free_hash(p
->md5_desc
.tfm
);
3229 void tcp_free_md5sig_pool(void)
3231 struct tcp_md5sig_pool __percpu
*pool
= NULL
;
3233 spin_lock_bh(&tcp_md5sig_pool_lock
);
3234 if (--tcp_md5sig_users
== 0) {
3235 pool
= tcp_md5sig_pool
;
3236 tcp_md5sig_pool
= NULL
;
3238 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3240 __tcp_free_md5sig_pool(pool
);
3242 EXPORT_SYMBOL(tcp_free_md5sig_pool
);
3244 static struct tcp_md5sig_pool __percpu
*
3245 __tcp_alloc_md5sig_pool(struct sock
*sk
)
3248 struct tcp_md5sig_pool __percpu
*pool
;
3250 pool
= alloc_percpu(struct tcp_md5sig_pool
);
3254 for_each_possible_cpu(cpu
) {
3255 struct crypto_hash
*hash
;
3257 hash
= crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC
);
3258 if (IS_ERR_OR_NULL(hash
))
3261 per_cpu_ptr(pool
, cpu
)->md5_desc
.tfm
= hash
;
3265 __tcp_free_md5sig_pool(pool
);
3269 struct tcp_md5sig_pool __percpu
*tcp_alloc_md5sig_pool(struct sock
*sk
)
3271 struct tcp_md5sig_pool __percpu
*pool
;
3275 spin_lock_bh(&tcp_md5sig_pool_lock
);
3276 pool
= tcp_md5sig_pool
;
3277 if (tcp_md5sig_users
++ == 0) {
3279 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3282 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3286 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3289 /* we cannot hold spinlock here because this may sleep. */
3290 struct tcp_md5sig_pool __percpu
*p
;
3292 p
= __tcp_alloc_md5sig_pool(sk
);
3293 spin_lock_bh(&tcp_md5sig_pool_lock
);
3296 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3299 pool
= tcp_md5sig_pool
;
3301 /* oops, it has already been assigned. */
3302 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3303 __tcp_free_md5sig_pool(p
);
3305 tcp_md5sig_pool
= pool
= p
;
3306 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3311 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
3315 * tcp_get_md5sig_pool - get md5sig_pool for this user
3317 * We use percpu structure, so if we succeed, we exit with preemption
3318 * and BH disabled, to make sure another thread or softirq handling
3319 * wont try to get same context.
3321 struct tcp_md5sig_pool
*tcp_get_md5sig_pool(void)
3323 struct tcp_md5sig_pool __percpu
*p
;
3327 spin_lock(&tcp_md5sig_pool_lock
);
3328 p
= tcp_md5sig_pool
;
3331 spin_unlock(&tcp_md5sig_pool_lock
);
3334 return this_cpu_ptr(p
);
3339 EXPORT_SYMBOL(tcp_get_md5sig_pool
);
3341 void tcp_put_md5sig_pool(void)
3344 tcp_free_md5sig_pool();
3346 EXPORT_SYMBOL(tcp_put_md5sig_pool
);
3348 int tcp_md5_hash_header(struct tcp_md5sig_pool
*hp
,
3349 const struct tcphdr
*th
)
3351 struct scatterlist sg
;
3355 /* We are not allowed to change tcphdr, make a local copy */
3356 memcpy(&hdr
, th
, sizeof(hdr
));
3359 /* options aren't included in the hash */
3360 sg_init_one(&sg
, &hdr
, sizeof(hdr
));
3361 err
= crypto_hash_update(&hp
->md5_desc
, &sg
, sizeof(hdr
));
3364 EXPORT_SYMBOL(tcp_md5_hash_header
);
3366 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool
*hp
,
3367 const struct sk_buff
*skb
, unsigned int header_len
)
3369 struct scatterlist sg
;
3370 const struct tcphdr
*tp
= tcp_hdr(skb
);
3371 struct hash_desc
*desc
= &hp
->md5_desc
;
3373 const unsigned int head_data_len
= skb_headlen(skb
) > header_len
?
3374 skb_headlen(skb
) - header_len
: 0;
3375 const struct skb_shared_info
*shi
= skb_shinfo(skb
);
3376 struct sk_buff
*frag_iter
;
3378 sg_init_table(&sg
, 1);
3380 sg_set_buf(&sg
, ((u8
*) tp
) + header_len
, head_data_len
);
3381 if (crypto_hash_update(desc
, &sg
, head_data_len
))
3384 for (i
= 0; i
< shi
->nr_frags
; ++i
) {
3385 const struct skb_frag_struct
*f
= &shi
->frags
[i
];
3386 struct page
*page
= skb_frag_page(f
);
3387 sg_set_page(&sg
, page
, skb_frag_size(f
), f
->page_offset
);
3388 if (crypto_hash_update(desc
, &sg
, skb_frag_size(f
)))
3392 skb_walk_frags(skb
, frag_iter
)
3393 if (tcp_md5_hash_skb_data(hp
, frag_iter
, 0))
3398 EXPORT_SYMBOL(tcp_md5_hash_skb_data
);
3400 int tcp_md5_hash_key(struct tcp_md5sig_pool
*hp
, const struct tcp_md5sig_key
*key
)
3402 struct scatterlist sg
;
3404 sg_init_one(&sg
, key
->key
, key
->keylen
);
3405 return crypto_hash_update(&hp
->md5_desc
, &sg
, key
->keylen
);
3407 EXPORT_SYMBOL(tcp_md5_hash_key
);
3411 /* Each Responder maintains up to two secret values concurrently for
3412 * efficient secret rollover. Each secret value has 4 states:
3414 * Generating. (tcp_secret_generating != tcp_secret_primary)
3415 * Generates new Responder-Cookies, but not yet used for primary
3416 * verification. This is a short-term state, typically lasting only
3417 * one round trip time (RTT).
3419 * Primary. (tcp_secret_generating == tcp_secret_primary)
3420 * Used both for generation and primary verification.
3422 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3423 * Used for verification, until the first failure that can be
3424 * verified by the newer Generating secret. At that time, this
3425 * cookie's state is changed to Secondary, and the Generating
3426 * cookie's state is changed to Primary. This is a short-term state,
3427 * typically lasting only one round trip time (RTT).
3429 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3430 * Used for secondary verification, after primary verification
3431 * failures. This state lasts no more than twice the Maximum Segment
3432 * Lifetime (2MSL). Then, the secret is discarded.
3434 struct tcp_cookie_secret
{
3435 /* The secret is divided into two parts. The digest part is the
3436 * equivalent of previously hashing a secret and saving the state,
3437 * and serves as an initialization vector (IV). The message part
3438 * serves as the trailing secret.
3440 u32 secrets
[COOKIE_WORKSPACE_WORDS
];
3441 unsigned long expires
;
3444 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3445 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3446 #define TCP_SECRET_LIFE (HZ * 600)
3448 static struct tcp_cookie_secret tcp_secret_one
;
3449 static struct tcp_cookie_secret tcp_secret_two
;
3451 /* Essentially a circular list, without dynamic allocation. */
3452 static struct tcp_cookie_secret
*tcp_secret_generating
;
3453 static struct tcp_cookie_secret
*tcp_secret_primary
;
3454 static struct tcp_cookie_secret
*tcp_secret_retiring
;
3455 static struct tcp_cookie_secret
*tcp_secret_secondary
;
3457 static DEFINE_SPINLOCK(tcp_secret_locker
);
3459 /* Select a pseudo-random word in the cookie workspace.
3461 static inline u32
tcp_cookie_work(const u32
*ws
, const int n
)
3463 return ws
[COOKIE_DIGEST_WORDS
+ ((COOKIE_MESSAGE_WORDS
-1) & ws
[n
])];
3466 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3467 * Called in softirq context.
3468 * Returns: 0 for success.
3470 int tcp_cookie_generator(u32
*bakery
)
3472 unsigned long jiffy
= jiffies
;
3474 if (unlikely(time_after_eq(jiffy
, tcp_secret_generating
->expires
))) {
3475 spin_lock_bh(&tcp_secret_locker
);
3476 if (!time_after_eq(jiffy
, tcp_secret_generating
->expires
)) {
3477 /* refreshed by another */
3479 &tcp_secret_generating
->secrets
[0],
3480 COOKIE_WORKSPACE_WORDS
);
3482 /* still needs refreshing */
3483 get_random_bytes(bakery
, COOKIE_WORKSPACE_WORDS
);
3485 /* The first time, paranoia assumes that the
3486 * randomization function isn't as strong. But,
3487 * this secret initialization is delayed until
3488 * the last possible moment (packet arrival).
3489 * Although that time is observable, it is
3490 * unpredictably variable. Mash in the most
3491 * volatile clock bits available, and expire the
3492 * secret extra quickly.
3494 if (unlikely(tcp_secret_primary
->expires
==
3495 tcp_secret_secondary
->expires
)) {
3498 getnstimeofday(&tv
);
3499 bakery
[COOKIE_DIGEST_WORDS
+0] ^=
3502 tcp_secret_secondary
->expires
= jiffy
3504 + (0x0f & tcp_cookie_work(bakery
, 0));
3506 tcp_secret_secondary
->expires
= jiffy
3508 + (0xff & tcp_cookie_work(bakery
, 1));
3509 tcp_secret_primary
->expires
= jiffy
3511 + (0x1f & tcp_cookie_work(bakery
, 2));
3513 memcpy(&tcp_secret_secondary
->secrets
[0],
3514 bakery
, COOKIE_WORKSPACE_WORDS
);
3516 rcu_assign_pointer(tcp_secret_generating
,
3517 tcp_secret_secondary
);
3518 rcu_assign_pointer(tcp_secret_retiring
,
3519 tcp_secret_primary
);
3521 * Neither call_rcu() nor synchronize_rcu() needed.
3522 * Retiring data is not freed. It is replaced after
3523 * further (locked) pointer updates, and a quiet time
3524 * (minimum 1MSL, maximum LIFE - 2MSL).
3527 spin_unlock_bh(&tcp_secret_locker
);
3531 &rcu_dereference(tcp_secret_generating
)->secrets
[0],
3532 COOKIE_WORKSPACE_WORDS
);
3533 rcu_read_unlock_bh();
3537 EXPORT_SYMBOL(tcp_cookie_generator
);
3539 void tcp_done(struct sock
*sk
)
3541 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
3543 if (sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
3544 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_ATTEMPTFAILS
);
3546 tcp_set_state(sk
, TCP_CLOSE
);
3547 tcp_clear_xmit_timers(sk
);
3549 reqsk_fastopen_remove(sk
, req
, false);
3551 sk
->sk_shutdown
= SHUTDOWN_MASK
;
3553 if (!sock_flag(sk
, SOCK_DEAD
))
3554 sk
->sk_state_change(sk
);
3556 inet_csk_destroy_sock(sk
);
3558 EXPORT_SYMBOL_GPL(tcp_done
);
3560 extern struct tcp_congestion_ops tcp_reno
;
3562 static __initdata
unsigned long thash_entries
;
3563 static int __init
set_thash_entries(char *str
)
3570 ret
= kstrtoul(str
, 0, &thash_entries
);
3576 __setup("thash_entries=", set_thash_entries
);
3578 void tcp_init_mem(struct net
*net
)
3580 unsigned long limit
= nr_free_buffer_pages() / 8;
3581 limit
= max(limit
, 128UL);
3582 net
->ipv4
.sysctl_tcp_mem
[0] = limit
/ 4 * 3;
3583 net
->ipv4
.sysctl_tcp_mem
[1] = limit
;
3584 net
->ipv4
.sysctl_tcp_mem
[2] = net
->ipv4
.sysctl_tcp_mem
[0] * 2;
3587 void __init
tcp_init(void)
3589 struct sk_buff
*skb
= NULL
;
3590 unsigned long limit
;
3591 int max_rshare
, max_wshare
, cnt
;
3593 unsigned long jiffy
= jiffies
;
3595 BUILD_BUG_ON(sizeof(struct tcp_skb_cb
) > sizeof(skb
->cb
));
3597 percpu_counter_init(&tcp_sockets_allocated
, 0);
3598 percpu_counter_init(&tcp_orphan_count
, 0);
3599 tcp_hashinfo
.bind_bucket_cachep
=
3600 kmem_cache_create("tcp_bind_bucket",
3601 sizeof(struct inet_bind_bucket
), 0,
3602 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
);
3604 /* Size and allocate the main established and bind bucket
3607 * The methodology is similar to that of the buffer cache.
3609 tcp_hashinfo
.ehash
=
3610 alloc_large_system_hash("TCP established",
3611 sizeof(struct inet_ehash_bucket
),
3613 17, /* one slot per 128 KB of memory */
3616 &tcp_hashinfo
.ehash_mask
,
3618 thash_entries
? 0 : 512 * 1024);
3619 for (i
= 0; i
<= tcp_hashinfo
.ehash_mask
; i
++) {
3620 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].chain
, i
);
3621 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].twchain
, i
);
3623 if (inet_ehash_locks_alloc(&tcp_hashinfo
))
3624 panic("TCP: failed to alloc ehash_locks");
3625 tcp_hashinfo
.bhash
=
3626 alloc_large_system_hash("TCP bind",
3627 sizeof(struct inet_bind_hashbucket
),
3628 tcp_hashinfo
.ehash_mask
+ 1,
3629 17, /* one slot per 128 KB of memory */
3631 &tcp_hashinfo
.bhash_size
,
3635 tcp_hashinfo
.bhash_size
= 1U << tcp_hashinfo
.bhash_size
;
3636 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
3637 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
3638 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
3642 cnt
= tcp_hashinfo
.ehash_mask
+ 1;
3644 tcp_death_row
.sysctl_max_tw_buckets
= cnt
/ 2;
3645 sysctl_tcp_max_orphans
= cnt
/ 2;
3646 sysctl_max_syn_backlog
= max(128, cnt
/ 256);
3648 tcp_init_mem(&init_net
);
3649 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3650 limit
= nr_free_buffer_pages() << (PAGE_SHIFT
- 7);
3651 max_wshare
= min(4UL*1024*1024, limit
);
3652 max_rshare
= min(6UL*1024*1024, limit
);
3654 sysctl_tcp_wmem
[0] = SK_MEM_QUANTUM
;
3655 sysctl_tcp_wmem
[1] = 16*1024;
3656 sysctl_tcp_wmem
[2] = max(64*1024, max_wshare
);
3658 sysctl_tcp_rmem
[0] = SK_MEM_QUANTUM
;
3659 sysctl_tcp_rmem
[1] = 87380;
3660 sysctl_tcp_rmem
[2] = max(87380, max_rshare
);
3662 pr_info("Hash tables configured (established %u bind %u)\n",
3663 tcp_hashinfo
.ehash_mask
+ 1, tcp_hashinfo
.bhash_size
);
3667 tcp_register_congestion_control(&tcp_reno
);
3669 memset(&tcp_secret_one
.secrets
[0], 0, sizeof(tcp_secret_one
.secrets
));
3670 memset(&tcp_secret_two
.secrets
[0], 0, sizeof(tcp_secret_two
.secrets
));
3671 tcp_secret_one
.expires
= jiffy
; /* past due */
3672 tcp_secret_two
.expires
= jiffy
; /* past due */
3673 tcp_secret_generating
= &tcp_secret_one
;
3674 tcp_secret_primary
= &tcp_secret_one
;
3675 tcp_secret_retiring
= &tcp_secret_two
;
3676 tcp_secret_secondary
= &tcp_secret_two
;