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>
282 #include <net/busy_poll.h>
284 int sysctl_tcp_fin_timeout __read_mostly
= TCP_FIN_TIMEOUT
;
286 int sysctl_tcp_min_tso_segs __read_mostly
= 2;
288 struct percpu_counter tcp_orphan_count
;
289 EXPORT_SYMBOL_GPL(tcp_orphan_count
);
291 long sysctl_tcp_mem
[3] __read_mostly
;
292 int sysctl_tcp_wmem
[3] __read_mostly
;
293 int sysctl_tcp_rmem
[3] __read_mostly
;
295 EXPORT_SYMBOL(sysctl_tcp_mem
);
296 EXPORT_SYMBOL(sysctl_tcp_rmem
);
297 EXPORT_SYMBOL(sysctl_tcp_wmem
);
299 atomic_long_t tcp_memory_allocated
; /* Current allocated memory. */
300 EXPORT_SYMBOL(tcp_memory_allocated
);
303 * Current number of TCP sockets.
305 struct percpu_counter tcp_sockets_allocated
;
306 EXPORT_SYMBOL(tcp_sockets_allocated
);
311 struct tcp_splice_state
{
312 struct pipe_inode_info
*pipe
;
318 * Pressure flag: try to collapse.
319 * Technical note: it is used by multiple contexts non atomically.
320 * All the __sk_mem_schedule() is of this nature: accounting
321 * is strict, actions are advisory and have some latency.
323 int tcp_memory_pressure __read_mostly
;
324 EXPORT_SYMBOL(tcp_memory_pressure
);
326 void tcp_enter_memory_pressure(struct sock
*sk
)
328 if (!tcp_memory_pressure
) {
329 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMEMORYPRESSURES
);
330 tcp_memory_pressure
= 1;
333 EXPORT_SYMBOL(tcp_enter_memory_pressure
);
335 /* Convert seconds to retransmits based on initial and max timeout */
336 static u8
secs_to_retrans(int seconds
, int timeout
, int rto_max
)
341 int period
= timeout
;
344 while (seconds
> period
&& res
< 255) {
347 if (timeout
> rto_max
)
355 /* Convert retransmits to seconds based on initial and max timeout */
356 static int retrans_to_secs(u8 retrans
, int timeout
, int rto_max
)
364 if (timeout
> rto_max
)
372 /* Address-family independent initialization for a tcp_sock.
374 * NOTE: A lot of things set to zero explicitly by call to
375 * sk_alloc() so need not be done here.
377 void tcp_init_sock(struct sock
*sk
)
379 struct inet_connection_sock
*icsk
= inet_csk(sk
);
380 struct tcp_sock
*tp
= tcp_sk(sk
);
382 skb_queue_head_init(&tp
->out_of_order_queue
);
383 tcp_init_xmit_timers(sk
);
384 tcp_prequeue_init(tp
);
385 INIT_LIST_HEAD(&tp
->tsq_node
);
387 icsk
->icsk_rto
= TCP_TIMEOUT_INIT
;
388 tp
->mdev
= TCP_TIMEOUT_INIT
;
390 /* So many TCP implementations out there (incorrectly) count the
391 * initial SYN frame in their delayed-ACK and congestion control
392 * algorithms that we must have the following bandaid to talk
393 * efficiently to them. -DaveM
395 tp
->snd_cwnd
= TCP_INIT_CWND
;
397 /* See draft-stevens-tcpca-spec-01 for discussion of the
398 * initialization of these values.
400 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
401 tp
->snd_cwnd_clamp
= ~0;
402 tp
->mss_cache
= TCP_MSS_DEFAULT
;
404 tp
->reordering
= sysctl_tcp_reordering
;
405 tcp_enable_early_retrans(tp
);
406 icsk
->icsk_ca_ops
= &tcp_init_congestion_ops
;
410 sk
->sk_state
= TCP_CLOSE
;
412 sk
->sk_write_space
= sk_stream_write_space
;
413 sock_set_flag(sk
, SOCK_USE_WRITE_QUEUE
);
415 icsk
->icsk_sync_mss
= tcp_sync_mss
;
417 sk
->sk_sndbuf
= sysctl_tcp_wmem
[1];
418 sk
->sk_rcvbuf
= sysctl_tcp_rmem
[1];
421 sock_update_memcg(sk
);
422 sk_sockets_allocated_inc(sk
);
425 EXPORT_SYMBOL(tcp_init_sock
);
428 * Wait for a TCP event.
430 * Note that we don't need to lock the socket, as the upper poll layers
431 * take care of normal races (between the test and the event) and we don't
432 * go look at any of the socket buffers directly.
434 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
437 struct sock
*sk
= sock
->sk
;
438 const struct tcp_sock
*tp
= tcp_sk(sk
);
440 sock_rps_record_flow(sk
);
442 sock_poll_wait(file
, sk_sleep(sk
), wait
);
443 if (sk
->sk_state
== TCP_LISTEN
)
444 return inet_csk_listen_poll(sk
);
446 /* Socket is not locked. We are protected from async events
447 * by poll logic and correct handling of state changes
448 * made by other threads is impossible in any case.
454 * POLLHUP is certainly not done right. But poll() doesn't
455 * have a notion of HUP in just one direction, and for a
456 * socket the read side is more interesting.
458 * Some poll() documentation says that POLLHUP is incompatible
459 * with the POLLOUT/POLLWR flags, so somebody should check this
460 * all. But careful, it tends to be safer to return too many
461 * bits than too few, and you can easily break real applications
462 * if you don't tell them that something has hung up!
466 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
467 * our fs/select.c). It means that after we received EOF,
468 * poll always returns immediately, making impossible poll() on write()
469 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
470 * if and only if shutdown has been made in both directions.
471 * Actually, it is interesting to look how Solaris and DUX
472 * solve this dilemma. I would prefer, if POLLHUP were maskable,
473 * then we could set it on SND_SHUTDOWN. BTW examples given
474 * in Stevens' books assume exactly this behaviour, it explains
475 * why POLLHUP is incompatible with POLLOUT. --ANK
477 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
478 * blocking on fresh not-connected or disconnected socket. --ANK
480 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| sk
->sk_state
== TCP_CLOSE
)
482 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
483 mask
|= POLLIN
| POLLRDNORM
| POLLRDHUP
;
485 /* Connected or passive Fast Open socket? */
486 if (sk
->sk_state
!= TCP_SYN_SENT
&&
487 (sk
->sk_state
!= TCP_SYN_RECV
|| tp
->fastopen_rsk
!= NULL
)) {
488 int target
= sock_rcvlowat(sk
, 0, INT_MAX
);
490 if (tp
->urg_seq
== tp
->copied_seq
&&
491 !sock_flag(sk
, SOCK_URGINLINE
) &&
495 /* Potential race condition. If read of tp below will
496 * escape above sk->sk_state, we can be illegally awaken
497 * in SYN_* states. */
498 if (tp
->rcv_nxt
- tp
->copied_seq
>= target
)
499 mask
|= POLLIN
| POLLRDNORM
;
501 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
502 if (sk_stream_is_writeable(sk
)) {
503 mask
|= POLLOUT
| POLLWRNORM
;
504 } else { /* send SIGIO later */
505 set_bit(SOCK_ASYNC_NOSPACE
,
506 &sk
->sk_socket
->flags
);
507 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
509 /* Race breaker. If space is freed after
510 * wspace test but before the flags are set,
511 * IO signal will be lost.
513 if (sk_stream_is_writeable(sk
))
514 mask
|= POLLOUT
| POLLWRNORM
;
517 mask
|= POLLOUT
| POLLWRNORM
;
519 if (tp
->urg_data
& TCP_URG_VALID
)
522 /* This barrier is coupled with smp_wmb() in tcp_reset() */
529 EXPORT_SYMBOL(tcp_poll
);
531 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
533 struct tcp_sock
*tp
= tcp_sk(sk
);
539 if (sk
->sk_state
== TCP_LISTEN
)
542 slow
= lock_sock_fast(sk
);
543 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
545 else if (sock_flag(sk
, SOCK_URGINLINE
) ||
547 before(tp
->urg_seq
, tp
->copied_seq
) ||
548 !before(tp
->urg_seq
, tp
->rcv_nxt
)) {
550 answ
= tp
->rcv_nxt
- tp
->copied_seq
;
552 /* Subtract 1, if FIN was received */
553 if (answ
&& sock_flag(sk
, SOCK_DONE
))
556 answ
= tp
->urg_seq
- tp
->copied_seq
;
557 unlock_sock_fast(sk
, slow
);
560 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
563 if (sk
->sk_state
== TCP_LISTEN
)
566 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
569 answ
= tp
->write_seq
- tp
->snd_una
;
572 if (sk
->sk_state
== TCP_LISTEN
)
575 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
578 answ
= tp
->write_seq
- tp
->snd_nxt
;
584 return put_user(answ
, (int __user
*)arg
);
586 EXPORT_SYMBOL(tcp_ioctl
);
588 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
590 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
591 tp
->pushed_seq
= tp
->write_seq
;
594 static inline bool forced_push(const struct tcp_sock
*tp
)
596 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
599 static inline void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
601 struct tcp_sock
*tp
= tcp_sk(sk
);
602 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
605 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
606 tcb
->tcp_flags
= TCPHDR_ACK
;
608 skb_header_release(skb
);
609 tcp_add_write_queue_tail(sk
, skb
);
610 sk
->sk_wmem_queued
+= skb
->truesize
;
611 sk_mem_charge(sk
, skb
->truesize
);
612 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
613 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
616 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
)
619 tp
->snd_up
= tp
->write_seq
;
622 static inline void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
625 if (tcp_send_head(sk
)) {
626 struct tcp_sock
*tp
= tcp_sk(sk
);
628 if (!(flags
& MSG_MORE
) || forced_push(tp
))
629 tcp_mark_push(tp
, tcp_write_queue_tail(sk
));
631 tcp_mark_urg(tp
, flags
);
632 __tcp_push_pending_frames(sk
, mss_now
,
633 (flags
& MSG_MORE
) ? TCP_NAGLE_CORK
: nonagle
);
637 static int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
638 unsigned int offset
, size_t len
)
640 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
643 ret
= skb_splice_bits(skb
, offset
, tss
->pipe
, min(rd_desc
->count
, len
),
646 rd_desc
->count
-= ret
;
650 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
652 /* Store TCP splice context information in read_descriptor_t. */
653 read_descriptor_t rd_desc
= {
658 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
662 * tcp_splice_read - splice data from TCP socket to a pipe
663 * @sock: socket to splice from
664 * @ppos: position (not valid)
665 * @pipe: pipe to splice to
666 * @len: number of bytes to splice
667 * @flags: splice modifier flags
670 * Will read pages from given socket and fill them into a pipe.
673 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
674 struct pipe_inode_info
*pipe
, size_t len
,
677 struct sock
*sk
= sock
->sk
;
678 struct tcp_splice_state tss
= {
687 sock_rps_record_flow(sk
);
689 * We can't seek on a socket input
698 timeo
= sock_rcvtimeo(sk
, sock
->file
->f_flags
& O_NONBLOCK
);
700 ret
= __tcp_splice_read(sk
, &tss
);
706 if (sock_flag(sk
, SOCK_DONE
))
709 ret
= sock_error(sk
);
712 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
714 if (sk
->sk_state
== TCP_CLOSE
) {
716 * This occurs when user tries to read
717 * from never connected socket.
719 if (!sock_flag(sk
, SOCK_DONE
))
727 sk_wait_data(sk
, &timeo
);
728 if (signal_pending(current
)) {
729 ret
= sock_intr_errno(timeo
);
742 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
743 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
744 signal_pending(current
))
755 EXPORT_SYMBOL(tcp_splice_read
);
757 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
)
761 /* The TCP header must be at least 32-bit aligned. */
762 size
= ALIGN(size
, 4);
764 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
766 if (sk_wmem_schedule(sk
, skb
->truesize
)) {
767 skb_reserve(skb
, sk
->sk_prot
->max_header
);
769 * Make sure that we have exactly size bytes
770 * available to the caller, no more, no less.
772 skb
->reserved_tailroom
= skb
->end
- skb
->tail
- size
;
777 sk
->sk_prot
->enter_memory_pressure(sk
);
778 sk_stream_moderate_sndbuf(sk
);
783 static unsigned int tcp_xmit_size_goal(struct sock
*sk
, u32 mss_now
,
786 struct tcp_sock
*tp
= tcp_sk(sk
);
787 u32 xmit_size_goal
, old_size_goal
;
789 xmit_size_goal
= mss_now
;
791 if (large_allowed
&& sk_can_gso(sk
)) {
794 /* Maybe we should/could use sk->sk_prot->max_header here ? */
795 hlen
= inet_csk(sk
)->icsk_af_ops
->net_header_len
+
796 inet_csk(sk
)->icsk_ext_hdr_len
+
799 /* Goal is to send at least one packet per ms,
800 * not one big TSO packet every 100 ms.
801 * This preserves ACK clocking and is consistent
802 * with tcp_tso_should_defer() heuristic.
804 gso_size
= sk
->sk_pacing_rate
/ (2 * MSEC_PER_SEC
);
805 gso_size
= max_t(u32
, gso_size
,
806 sysctl_tcp_min_tso_segs
* mss_now
);
808 xmit_size_goal
= min_t(u32
, gso_size
,
809 sk
->sk_gso_max_size
- 1 - hlen
);
811 /* TSQ : try to have at least two segments in flight
812 * (one in NIC TX ring, another in Qdisc)
814 xmit_size_goal
= min_t(u32
, xmit_size_goal
,
815 sysctl_tcp_limit_output_bytes
>> 1);
817 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
819 /* We try hard to avoid divides here */
820 old_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
822 if (likely(old_size_goal
<= xmit_size_goal
&&
823 old_size_goal
+ mss_now
> xmit_size_goal
)) {
824 xmit_size_goal
= old_size_goal
;
826 tp
->xmit_size_goal_segs
=
827 min_t(u16
, xmit_size_goal
/ mss_now
,
828 sk
->sk_gso_max_segs
);
829 xmit_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
833 return max(xmit_size_goal
, mss_now
);
836 static int tcp_send_mss(struct sock
*sk
, int *size_goal
, int flags
)
840 mss_now
= tcp_current_mss(sk
);
841 *size_goal
= tcp_xmit_size_goal(sk
, mss_now
, !(flags
& MSG_OOB
));
846 static ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
*page
, int offset
,
847 size_t size
, int flags
)
849 struct tcp_sock
*tp
= tcp_sk(sk
);
850 int mss_now
, size_goal
;
853 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
855 /* Wait for a connection to finish. One exception is TCP Fast Open
856 * (passive side) where data is allowed to be sent before a connection
857 * is fully established.
859 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
860 !tcp_passive_fastopen(sk
)) {
861 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
865 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
867 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
871 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
875 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
879 if (!tcp_send_head(sk
) || (copy
= size_goal
- skb
->len
) <= 0) {
881 if (!sk_stream_memory_free(sk
))
882 goto wait_for_sndbuf
;
884 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
886 goto wait_for_memory
;
895 i
= skb_shinfo(skb
)->nr_frags
;
896 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
897 if (!can_coalesce
&& i
>= MAX_SKB_FRAGS
) {
898 tcp_mark_push(tp
, skb
);
901 if (!sk_wmem_schedule(sk
, copy
))
902 goto wait_for_memory
;
905 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
908 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
910 skb_shinfo(skb
)->tx_flags
|= SKBTX_SHARED_FRAG
;
913 skb
->data_len
+= copy
;
914 skb
->truesize
+= copy
;
915 sk
->sk_wmem_queued
+= copy
;
916 sk_mem_charge(sk
, copy
);
917 skb
->ip_summed
= CHECKSUM_PARTIAL
;
918 tp
->write_seq
+= copy
;
919 TCP_SKB_CB(skb
)->end_seq
+= copy
;
920 skb_shinfo(skb
)->gso_segs
= 0;
923 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
930 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
933 if (forced_push(tp
)) {
934 tcp_mark_push(tp
, skb
);
935 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
936 } else if (skb
== tcp_send_head(sk
))
937 tcp_push_one(sk
, mss_now
);
941 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
943 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
945 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
948 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
952 if (copied
&& !(flags
& MSG_SENDPAGE_NOTLAST
))
953 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
960 return sk_stream_error(sk
, flags
, err
);
963 int tcp_sendpage(struct sock
*sk
, struct page
*page
, int offset
,
964 size_t size
, int flags
)
968 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
969 !(sk
->sk_route_caps
& NETIF_F_ALL_CSUM
))
970 return sock_no_sendpage(sk
->sk_socket
, page
, offset
, size
,
974 res
= do_tcp_sendpages(sk
, page
, offset
, size
, flags
);
978 EXPORT_SYMBOL(tcp_sendpage
);
980 static inline int select_size(const struct sock
*sk
, bool sg
)
982 const struct tcp_sock
*tp
= tcp_sk(sk
);
983 int tmp
= tp
->mss_cache
;
986 if (sk_can_gso(sk
)) {
987 /* Small frames wont use a full page:
988 * Payload will immediately follow tcp header.
990 tmp
= SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER
);
992 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
994 if (tmp
>= pgbreak
&&
995 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
1003 void tcp_free_fastopen_req(struct tcp_sock
*tp
)
1005 if (tp
->fastopen_req
!= NULL
) {
1006 kfree(tp
->fastopen_req
);
1007 tp
->fastopen_req
= NULL
;
1011 static int tcp_sendmsg_fastopen(struct sock
*sk
, struct msghdr
*msg
, int *size
)
1013 struct tcp_sock
*tp
= tcp_sk(sk
);
1016 if (!(sysctl_tcp_fastopen
& TFO_CLIENT_ENABLE
))
1018 if (tp
->fastopen_req
!= NULL
)
1019 return -EALREADY
; /* Another Fast Open is in progress */
1021 tp
->fastopen_req
= kzalloc(sizeof(struct tcp_fastopen_request
),
1023 if (unlikely(tp
->fastopen_req
== NULL
))
1025 tp
->fastopen_req
->data
= msg
;
1027 flags
= (msg
->msg_flags
& MSG_DONTWAIT
) ? O_NONBLOCK
: 0;
1028 err
= __inet_stream_connect(sk
->sk_socket
, msg
->msg_name
,
1029 msg
->msg_namelen
, flags
);
1030 *size
= tp
->fastopen_req
->copied
;
1031 tcp_free_fastopen_req(tp
);
1035 int tcp_sendmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1039 struct tcp_sock
*tp
= tcp_sk(sk
);
1040 struct sk_buff
*skb
;
1041 int iovlen
, flags
, err
, copied
= 0;
1042 int mss_now
= 0, size_goal
, copied_syn
= 0, offset
= 0;
1048 flags
= msg
->msg_flags
;
1049 if (flags
& MSG_FASTOPEN
) {
1050 err
= tcp_sendmsg_fastopen(sk
, msg
, &copied_syn
);
1051 if (err
== -EINPROGRESS
&& copied_syn
> 0)
1055 offset
= copied_syn
;
1058 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
1060 /* Wait for a connection to finish. One exception is TCP Fast Open
1061 * (passive side) where data is allowed to be sent before a connection
1062 * is fully established.
1064 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
1065 !tcp_passive_fastopen(sk
)) {
1066 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
1070 if (unlikely(tp
->repair
)) {
1071 if (tp
->repair_queue
== TCP_RECV_QUEUE
) {
1072 copied
= tcp_send_rcvq(sk
, msg
, size
);
1077 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1080 /* 'common' sending to sendq */
1083 /* This should be in poll */
1084 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
1086 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1088 /* Ok commence sending. */
1089 iovlen
= msg
->msg_iovlen
;
1094 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
1097 sg
= !!(sk
->sk_route_caps
& NETIF_F_SG
);
1099 while (--iovlen
>= 0) {
1100 size_t seglen
= iov
->iov_len
;
1101 unsigned char __user
*from
= iov
->iov_base
;
1104 if (unlikely(offset
> 0)) { /* Skip bytes copied in SYN */
1105 if (offset
>= seglen
) {
1114 while (seglen
> 0) {
1116 int max
= size_goal
;
1118 skb
= tcp_write_queue_tail(sk
);
1119 if (tcp_send_head(sk
)) {
1120 if (skb
->ip_summed
== CHECKSUM_NONE
)
1122 copy
= max
- skb
->len
;
1127 /* Allocate new segment. If the interface is SG,
1128 * allocate skb fitting to single page.
1130 if (!sk_stream_memory_free(sk
))
1131 goto wait_for_sndbuf
;
1133 skb
= sk_stream_alloc_skb(sk
,
1134 select_size(sk
, sg
),
1137 goto wait_for_memory
;
1140 * All packets are restored as if they have
1141 * already been sent.
1144 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1147 * Check whether we can use HW checksum.
1149 if (sk
->sk_route_caps
& NETIF_F_ALL_CSUM
)
1150 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1152 skb_entail(sk
, skb
);
1157 /* Try to append data to the end of skb. */
1161 /* Where to copy to? */
1162 if (skb_availroom(skb
) > 0) {
1163 /* We have some space in skb head. Superb! */
1164 copy
= min_t(int, copy
, skb_availroom(skb
));
1165 err
= skb_add_data_nocache(sk
, skb
, from
, copy
);
1170 int i
= skb_shinfo(skb
)->nr_frags
;
1171 struct page_frag
*pfrag
= sk_page_frag(sk
);
1173 if (!sk_page_frag_refill(sk
, pfrag
))
1174 goto wait_for_memory
;
1176 if (!skb_can_coalesce(skb
, i
, pfrag
->page
,
1178 if (i
== MAX_SKB_FRAGS
|| !sg
) {
1179 tcp_mark_push(tp
, skb
);
1185 copy
= min_t(int, copy
, pfrag
->size
- pfrag
->offset
);
1187 if (!sk_wmem_schedule(sk
, copy
))
1188 goto wait_for_memory
;
1190 err
= skb_copy_to_page_nocache(sk
, from
, skb
,
1197 /* Update the skb. */
1199 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
1201 skb_fill_page_desc(skb
, i
, pfrag
->page
,
1202 pfrag
->offset
, copy
);
1203 get_page(pfrag
->page
);
1205 pfrag
->offset
+= copy
;
1209 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
1211 tp
->write_seq
+= copy
;
1212 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1213 skb_shinfo(skb
)->gso_segs
= 0;
1217 if ((seglen
-= copy
) == 0 && iovlen
== 0)
1220 if (skb
->len
< max
|| (flags
& MSG_OOB
) || unlikely(tp
->repair
))
1223 if (forced_push(tp
)) {
1224 tcp_mark_push(tp
, skb
);
1225 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1226 } else if (skb
== tcp_send_head(sk
))
1227 tcp_push_one(sk
, mss_now
);
1231 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1234 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
1236 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
1239 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1245 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
1247 return copied
+ copied_syn
;
1251 tcp_unlink_write_queue(skb
, sk
);
1252 /* It is the one place in all of TCP, except connection
1253 * reset, where we can be unlinking the send_head.
1255 tcp_check_send_head(sk
, skb
);
1256 sk_wmem_free_skb(sk
, skb
);
1260 if (copied
+ copied_syn
)
1263 err
= sk_stream_error(sk
, flags
, err
);
1267 EXPORT_SYMBOL(tcp_sendmsg
);
1270 * Handle reading urgent data. BSD has very simple semantics for
1271 * this, no blocking and very strange errors 8)
1274 static int tcp_recv_urg(struct sock
*sk
, struct msghdr
*msg
, int len
, int flags
)
1276 struct tcp_sock
*tp
= tcp_sk(sk
);
1278 /* No URG data to read. */
1279 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1280 tp
->urg_data
== TCP_URG_READ
)
1281 return -EINVAL
; /* Yes this is right ! */
1283 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1286 if (tp
->urg_data
& TCP_URG_VALID
) {
1288 char c
= tp
->urg_data
;
1290 if (!(flags
& MSG_PEEK
))
1291 tp
->urg_data
= TCP_URG_READ
;
1293 /* Read urgent data. */
1294 msg
->msg_flags
|= MSG_OOB
;
1297 if (!(flags
& MSG_TRUNC
))
1298 err
= memcpy_toiovec(msg
->msg_iov
, &c
, 1);
1301 msg
->msg_flags
|= MSG_TRUNC
;
1303 return err
? -EFAULT
: len
;
1306 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1309 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1310 * the available implementations agree in this case:
1311 * this call should never block, independent of the
1312 * blocking state of the socket.
1313 * Mike <pall@rz.uni-karlsruhe.de>
1318 static int tcp_peek_sndq(struct sock
*sk
, struct msghdr
*msg
, int len
)
1320 struct sk_buff
*skb
;
1321 int copied
= 0, err
= 0;
1323 /* XXX -- need to support SO_PEEK_OFF */
1325 skb_queue_walk(&sk
->sk_write_queue
, skb
) {
1326 err
= skb_copy_datagram_iovec(skb
, 0, msg
->msg_iov
, skb
->len
);
1333 return err
?: copied
;
1336 /* Clean up the receive buffer for full frames taken by the user,
1337 * then send an ACK if necessary. COPIED is the number of bytes
1338 * tcp_recvmsg has given to the user so far, it speeds up the
1339 * calculation of whether or not we must ACK for the sake of
1342 void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1344 struct tcp_sock
*tp
= tcp_sk(sk
);
1345 bool time_to_ack
= false;
1347 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1349 WARN(skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
),
1350 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1351 tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
);
1353 if (inet_csk_ack_scheduled(sk
)) {
1354 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1355 /* Delayed ACKs frequently hit locked sockets during bulk
1357 if (icsk
->icsk_ack
.blocked
||
1358 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1359 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1361 * If this read emptied read buffer, we send ACK, if
1362 * connection is not bidirectional, user drained
1363 * receive buffer and there was a small segment
1367 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1368 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1369 !icsk
->icsk_ack
.pingpong
)) &&
1370 !atomic_read(&sk
->sk_rmem_alloc
)))
1374 /* We send an ACK if we can now advertise a non-zero window
1375 * which has been raised "significantly".
1377 * Even if window raised up to infinity, do not send window open ACK
1378 * in states, where we will not receive more. It is useless.
1380 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1381 __u32 rcv_window_now
= tcp_receive_window(tp
);
1383 /* Optimize, __tcp_select_window() is not cheap. */
1384 if (2*rcv_window_now
<= tp
->window_clamp
) {
1385 __u32 new_window
= __tcp_select_window(sk
);
1387 /* Send ACK now, if this read freed lots of space
1388 * in our buffer. Certainly, new_window is new window.
1389 * We can advertise it now, if it is not less than current one.
1390 * "Lots" means "at least twice" here.
1392 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1400 static void tcp_prequeue_process(struct sock
*sk
)
1402 struct sk_buff
*skb
;
1403 struct tcp_sock
*tp
= tcp_sk(sk
);
1405 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPPREQUEUED
);
1407 /* RX process wants to run with disabled BHs, though it is not
1410 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
1411 sk_backlog_rcv(sk
, skb
);
1414 /* Clear memory counter. */
1415 tp
->ucopy
.memory
= 0;
1418 #ifdef CONFIG_NET_DMA
1419 static void tcp_service_net_dma(struct sock
*sk
, bool wait
)
1421 dma_cookie_t done
, used
;
1422 dma_cookie_t last_issued
;
1423 struct tcp_sock
*tp
= tcp_sk(sk
);
1425 if (!tp
->ucopy
.dma_chan
)
1428 last_issued
= tp
->ucopy
.dma_cookie
;
1429 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1432 if (dma_async_is_tx_complete(tp
->ucopy
.dma_chan
,
1434 &used
) == DMA_SUCCESS
) {
1435 /* Safe to free early-copied skbs now */
1436 __skb_queue_purge(&sk
->sk_async_wait_queue
);
1439 struct sk_buff
*skb
;
1440 while ((skb
= skb_peek(&sk
->sk_async_wait_queue
)) &&
1441 (dma_async_is_complete(skb
->dma_cookie
, done
,
1442 used
) == DMA_SUCCESS
)) {
1443 __skb_dequeue(&sk
->sk_async_wait_queue
);
1451 static struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1453 struct sk_buff
*skb
;
1456 while ((skb
= skb_peek(&sk
->sk_receive_queue
)) != NULL
) {
1457 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1458 if (tcp_hdr(skb
)->syn
)
1460 if (offset
< skb
->len
|| tcp_hdr(skb
)->fin
) {
1464 /* This looks weird, but this can happen if TCP collapsing
1465 * splitted a fat GRO packet, while we released socket lock
1466 * in skb_splice_bits()
1468 sk_eat_skb(sk
, skb
, false);
1474 * This routine provides an alternative to tcp_recvmsg() for routines
1475 * that would like to handle copying from skbuffs directly in 'sendfile'
1478 * - It is assumed that the socket was locked by the caller.
1479 * - The routine does not block.
1480 * - At present, there is no support for reading OOB data
1481 * or for 'peeking' the socket using this routine
1482 * (although both would be easy to implement).
1484 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1485 sk_read_actor_t recv_actor
)
1487 struct sk_buff
*skb
;
1488 struct tcp_sock
*tp
= tcp_sk(sk
);
1489 u32 seq
= tp
->copied_seq
;
1493 if (sk
->sk_state
== TCP_LISTEN
)
1495 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1496 if (offset
< skb
->len
) {
1500 len
= skb
->len
- offset
;
1501 /* Stop reading if we hit a patch of urgent data */
1503 u32 urg_offset
= tp
->urg_seq
- seq
;
1504 if (urg_offset
< len
)
1509 used
= recv_actor(desc
, skb
, offset
, len
);
1514 } else if (used
<= len
) {
1519 /* If recv_actor drops the lock (e.g. TCP splice
1520 * receive) the skb pointer might be invalid when
1521 * getting here: tcp_collapse might have deleted it
1522 * while aggregating skbs from the socket queue.
1524 skb
= tcp_recv_skb(sk
, seq
- 1, &offset
);
1527 /* TCP coalescing might have appended data to the skb.
1528 * Try to splice more frags
1530 if (offset
+ 1 != skb
->len
)
1533 if (tcp_hdr(skb
)->fin
) {
1534 sk_eat_skb(sk
, skb
, false);
1538 sk_eat_skb(sk
, skb
, false);
1541 tp
->copied_seq
= seq
;
1543 tp
->copied_seq
= seq
;
1545 tcp_rcv_space_adjust(sk
);
1547 /* Clean up data we have read: This will do ACK frames. */
1549 tcp_recv_skb(sk
, seq
, &offset
);
1550 tcp_cleanup_rbuf(sk
, copied
);
1554 EXPORT_SYMBOL(tcp_read_sock
);
1557 * This routine copies from a sock struct into the user buffer.
1559 * Technical note: in 2.3 we work on _locked_ socket, so that
1560 * tricks with *seq access order and skb->users are not required.
1561 * Probably, code can be easily improved even more.
1564 int tcp_recvmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1565 size_t len
, int nonblock
, int flags
, int *addr_len
)
1567 struct tcp_sock
*tp
= tcp_sk(sk
);
1573 int target
; /* Read at least this many bytes */
1575 struct task_struct
*user_recv
= NULL
;
1576 bool copied_early
= false;
1577 struct sk_buff
*skb
;
1580 if (sk_can_busy_loop(sk
) && skb_queue_empty(&sk
->sk_receive_queue
) &&
1581 (sk
->sk_state
== TCP_ESTABLISHED
))
1582 sk_busy_loop(sk
, nonblock
);
1587 if (sk
->sk_state
== TCP_LISTEN
)
1590 timeo
= sock_rcvtimeo(sk
, nonblock
);
1592 /* Urgent data needs to be handled specially. */
1593 if (flags
& MSG_OOB
)
1596 if (unlikely(tp
->repair
)) {
1598 if (!(flags
& MSG_PEEK
))
1601 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
1605 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1608 /* 'common' recv queue MSG_PEEK-ing */
1611 seq
= &tp
->copied_seq
;
1612 if (flags
& MSG_PEEK
) {
1613 peek_seq
= tp
->copied_seq
;
1617 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1619 #ifdef CONFIG_NET_DMA
1620 tp
->ucopy
.dma_chan
= NULL
;
1622 skb
= skb_peek_tail(&sk
->sk_receive_queue
);
1627 available
= TCP_SKB_CB(skb
)->seq
+ skb
->len
- (*seq
);
1628 if ((available
< target
) &&
1629 (len
> sysctl_tcp_dma_copybreak
) && !(flags
& MSG_PEEK
) &&
1630 !sysctl_tcp_low_latency
&&
1631 net_dma_find_channel()) {
1632 preempt_enable_no_resched();
1633 tp
->ucopy
.pinned_list
=
1634 dma_pin_iovec_pages(msg
->msg_iov
, len
);
1636 preempt_enable_no_resched();
1644 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1645 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1648 if (signal_pending(current
)) {
1649 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1654 /* Next get a buffer. */
1656 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1657 /* Now that we have two receive queues this
1660 if (WARN(before(*seq
, TCP_SKB_CB(skb
)->seq
),
1661 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1662 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
,
1666 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1667 if (tcp_hdr(skb
)->syn
)
1669 if (offset
< skb
->len
)
1671 if (tcp_hdr(skb
)->fin
)
1673 WARN(!(flags
& MSG_PEEK
),
1674 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1675 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
, flags
);
1678 /* Well, if we have backlog, try to process it now yet. */
1680 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1685 sk
->sk_state
== TCP_CLOSE
||
1686 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1688 signal_pending(current
))
1691 if (sock_flag(sk
, SOCK_DONE
))
1695 copied
= sock_error(sk
);
1699 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1702 if (sk
->sk_state
== TCP_CLOSE
) {
1703 if (!sock_flag(sk
, SOCK_DONE
)) {
1704 /* This occurs when user tries to read
1705 * from never connected socket.
1718 if (signal_pending(current
)) {
1719 copied
= sock_intr_errno(timeo
);
1724 tcp_cleanup_rbuf(sk
, copied
);
1726 if (!sysctl_tcp_low_latency
&& tp
->ucopy
.task
== user_recv
) {
1727 /* Install new reader */
1728 if (!user_recv
&& !(flags
& (MSG_TRUNC
| MSG_PEEK
))) {
1729 user_recv
= current
;
1730 tp
->ucopy
.task
= user_recv
;
1731 tp
->ucopy
.iov
= msg
->msg_iov
;
1734 tp
->ucopy
.len
= len
;
1736 WARN_ON(tp
->copied_seq
!= tp
->rcv_nxt
&&
1737 !(flags
& (MSG_PEEK
| MSG_TRUNC
)));
1739 /* Ugly... If prequeue is not empty, we have to
1740 * process it before releasing socket, otherwise
1741 * order will be broken at second iteration.
1742 * More elegant solution is required!!!
1744 * Look: we have the following (pseudo)queues:
1746 * 1. packets in flight
1751 * Each queue can be processed only if the next ones
1752 * are empty. At this point we have empty receive_queue.
1753 * But prequeue _can_ be not empty after 2nd iteration,
1754 * when we jumped to start of loop because backlog
1755 * processing added something to receive_queue.
1756 * We cannot release_sock(), because backlog contains
1757 * packets arrived _after_ prequeued ones.
1759 * Shortly, algorithm is clear --- to process all
1760 * the queues in order. We could make it more directly,
1761 * requeueing packets from backlog to prequeue, if
1762 * is not empty. It is more elegant, but eats cycles,
1765 if (!skb_queue_empty(&tp
->ucopy
.prequeue
))
1768 /* __ Set realtime policy in scheduler __ */
1771 #ifdef CONFIG_NET_DMA
1772 if (tp
->ucopy
.dma_chan
) {
1773 if (tp
->rcv_wnd
== 0 &&
1774 !skb_queue_empty(&sk
->sk_async_wait_queue
)) {
1775 tcp_service_net_dma(sk
, true);
1776 tcp_cleanup_rbuf(sk
, copied
);
1778 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1781 if (copied
>= target
) {
1782 /* Do not sleep, just process backlog. */
1786 sk_wait_data(sk
, &timeo
);
1788 #ifdef CONFIG_NET_DMA
1789 tcp_service_net_dma(sk
, false); /* Don't block */
1790 tp
->ucopy
.wakeup
= 0;
1796 /* __ Restore normal policy in scheduler __ */
1798 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1799 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG
, chunk
);
1804 if (tp
->rcv_nxt
== tp
->copied_seq
&&
1805 !skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1807 tcp_prequeue_process(sk
);
1809 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1810 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1816 if ((flags
& MSG_PEEK
) &&
1817 (peek_seq
- copied
- urg_hole
!= tp
->copied_seq
)) {
1818 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1820 task_pid_nr(current
));
1821 peek_seq
= tp
->copied_seq
;
1826 /* Ok so how much can we use? */
1827 used
= skb
->len
- offset
;
1831 /* Do we have urgent data here? */
1833 u32 urg_offset
= tp
->urg_seq
- *seq
;
1834 if (urg_offset
< used
) {
1836 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1849 if (!(flags
& MSG_TRUNC
)) {
1850 #ifdef CONFIG_NET_DMA
1851 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
1852 tp
->ucopy
.dma_chan
= net_dma_find_channel();
1854 if (tp
->ucopy
.dma_chan
) {
1855 tp
->ucopy
.dma_cookie
= dma_skb_copy_datagram_iovec(
1856 tp
->ucopy
.dma_chan
, skb
, offset
,
1858 tp
->ucopy
.pinned_list
);
1860 if (tp
->ucopy
.dma_cookie
< 0) {
1862 pr_alert("%s: dma_cookie < 0\n",
1865 /* Exception. Bailout! */
1871 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1873 if ((offset
+ used
) == skb
->len
)
1874 copied_early
= true;
1879 err
= skb_copy_datagram_iovec(skb
, offset
,
1880 msg
->msg_iov
, used
);
1882 /* Exception. Bailout! */
1894 tcp_rcv_space_adjust(sk
);
1897 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1899 tcp_fast_path_check(sk
);
1901 if (used
+ offset
< skb
->len
)
1904 if (tcp_hdr(skb
)->fin
)
1906 if (!(flags
& MSG_PEEK
)) {
1907 sk_eat_skb(sk
, skb
, copied_early
);
1908 copied_early
= false;
1913 /* Process the FIN. */
1915 if (!(flags
& MSG_PEEK
)) {
1916 sk_eat_skb(sk
, skb
, copied_early
);
1917 copied_early
= false;
1923 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1926 tp
->ucopy
.len
= copied
> 0 ? len
: 0;
1928 tcp_prequeue_process(sk
);
1930 if (copied
> 0 && (chunk
= len
- tp
->ucopy
.len
) != 0) {
1931 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1937 tp
->ucopy
.task
= NULL
;
1941 #ifdef CONFIG_NET_DMA
1942 tcp_service_net_dma(sk
, true); /* Wait for queue to drain */
1943 tp
->ucopy
.dma_chan
= NULL
;
1945 if (tp
->ucopy
.pinned_list
) {
1946 dma_unpin_iovec_pages(tp
->ucopy
.pinned_list
);
1947 tp
->ucopy
.pinned_list
= NULL
;
1951 /* According to UNIX98, msg_name/msg_namelen are ignored
1952 * on connected socket. I was just happy when found this 8) --ANK
1955 /* Clean up data we have read: This will do ACK frames. */
1956 tcp_cleanup_rbuf(sk
, copied
);
1966 err
= tcp_recv_urg(sk
, msg
, len
, flags
);
1970 err
= tcp_peek_sndq(sk
, msg
, len
);
1973 EXPORT_SYMBOL(tcp_recvmsg
);
1975 void tcp_set_state(struct sock
*sk
, int state
)
1977 int oldstate
= sk
->sk_state
;
1980 case TCP_ESTABLISHED
:
1981 if (oldstate
!= TCP_ESTABLISHED
)
1982 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1986 if (oldstate
== TCP_CLOSE_WAIT
|| oldstate
== TCP_ESTABLISHED
)
1987 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ESTABRESETS
);
1989 sk
->sk_prot
->unhash(sk
);
1990 if (inet_csk(sk
)->icsk_bind_hash
&&
1991 !(sk
->sk_userlocks
& SOCK_BINDPORT_LOCK
))
1995 if (oldstate
== TCP_ESTABLISHED
)
1996 TCP_DEC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1999 /* Change state AFTER socket is unhashed to avoid closed
2000 * socket sitting in hash tables.
2002 sk
->sk_state
= state
;
2005 SOCK_DEBUG(sk
, "TCP sk=%p, State %s -> %s\n", sk
, statename
[oldstate
], statename
[state
]);
2008 EXPORT_SYMBOL_GPL(tcp_set_state
);
2011 * State processing on a close. This implements the state shift for
2012 * sending our FIN frame. Note that we only send a FIN for some
2013 * states. A shutdown() may have already sent the FIN, or we may be
2017 static const unsigned char new_state
[16] = {
2018 /* current state: new state: action: */
2019 /* (Invalid) */ TCP_CLOSE
,
2020 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2021 /* TCP_SYN_SENT */ TCP_CLOSE
,
2022 /* TCP_SYN_RECV */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2023 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1
,
2024 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2
,
2025 /* TCP_TIME_WAIT */ TCP_CLOSE
,
2026 /* TCP_CLOSE */ TCP_CLOSE
,
2027 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK
| TCP_ACTION_FIN
,
2028 /* TCP_LAST_ACK */ TCP_LAST_ACK
,
2029 /* TCP_LISTEN */ TCP_CLOSE
,
2030 /* TCP_CLOSING */ TCP_CLOSING
,
2033 static int tcp_close_state(struct sock
*sk
)
2035 int next
= (int)new_state
[sk
->sk_state
];
2036 int ns
= next
& TCP_STATE_MASK
;
2038 tcp_set_state(sk
, ns
);
2040 return next
& TCP_ACTION_FIN
;
2044 * Shutdown the sending side of a connection. Much like close except
2045 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2048 void tcp_shutdown(struct sock
*sk
, int how
)
2050 /* We need to grab some memory, and put together a FIN,
2051 * and then put it into the queue to be sent.
2052 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2054 if (!(how
& SEND_SHUTDOWN
))
2057 /* If we've already sent a FIN, or it's a closed state, skip this. */
2058 if ((1 << sk
->sk_state
) &
2059 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
2060 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
2061 /* Clear out any half completed packets. FIN if needed. */
2062 if (tcp_close_state(sk
))
2066 EXPORT_SYMBOL(tcp_shutdown
);
2068 bool tcp_check_oom(struct sock
*sk
, int shift
)
2070 bool too_many_orphans
, out_of_socket_memory
;
2072 too_many_orphans
= tcp_too_many_orphans(sk
, shift
);
2073 out_of_socket_memory
= tcp_out_of_memory(sk
);
2075 if (too_many_orphans
)
2076 net_info_ratelimited("too many orphaned sockets\n");
2077 if (out_of_socket_memory
)
2078 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2079 return too_many_orphans
|| out_of_socket_memory
;
2082 void tcp_close(struct sock
*sk
, long timeout
)
2084 struct sk_buff
*skb
;
2085 int data_was_unread
= 0;
2089 sk
->sk_shutdown
= SHUTDOWN_MASK
;
2091 if (sk
->sk_state
== TCP_LISTEN
) {
2092 tcp_set_state(sk
, TCP_CLOSE
);
2095 inet_csk_listen_stop(sk
);
2097 goto adjudge_to_death
;
2100 /* We need to flush the recv. buffs. We do this only on the
2101 * descriptor close, not protocol-sourced closes, because the
2102 * reader process may not have drained the data yet!
2104 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
2105 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
-
2107 data_was_unread
+= len
;
2113 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2114 if (sk
->sk_state
== TCP_CLOSE
)
2115 goto adjudge_to_death
;
2117 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2118 * data was lost. To witness the awful effects of the old behavior of
2119 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2120 * GET in an FTP client, suspend the process, wait for the client to
2121 * advertise a zero window, then kill -9 the FTP client, wheee...
2122 * Note: timeout is always zero in such a case.
2124 if (unlikely(tcp_sk(sk
)->repair
)) {
2125 sk
->sk_prot
->disconnect(sk
, 0);
2126 } else if (data_was_unread
) {
2127 /* Unread data was tossed, zap the connection. */
2128 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONCLOSE
);
2129 tcp_set_state(sk
, TCP_CLOSE
);
2130 tcp_send_active_reset(sk
, sk
->sk_allocation
);
2131 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
2132 /* Check zero linger _after_ checking for unread data. */
2133 sk
->sk_prot
->disconnect(sk
, 0);
2134 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
2135 } else if (tcp_close_state(sk
)) {
2136 /* We FIN if the application ate all the data before
2137 * zapping the connection.
2140 /* RED-PEN. Formally speaking, we have broken TCP state
2141 * machine. State transitions:
2143 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2144 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2145 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2147 * are legal only when FIN has been sent (i.e. in window),
2148 * rather than queued out of window. Purists blame.
2150 * F.e. "RFC state" is ESTABLISHED,
2151 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2153 * The visible declinations are that sometimes
2154 * we enter time-wait state, when it is not required really
2155 * (harmless), do not send active resets, when they are
2156 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2157 * they look as CLOSING or LAST_ACK for Linux)
2158 * Probably, I missed some more holelets.
2160 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2161 * in a single packet! (May consider it later but will
2162 * probably need API support or TCP_CORK SYN-ACK until
2163 * data is written and socket is closed.)
2168 sk_stream_wait_close(sk
, timeout
);
2171 state
= sk
->sk_state
;
2175 /* It is the last release_sock in its life. It will remove backlog. */
2179 /* Now socket is owned by kernel and we acquire BH lock
2180 to finish close. No need to check for user refs.
2184 WARN_ON(sock_owned_by_user(sk
));
2186 percpu_counter_inc(sk
->sk_prot
->orphan_count
);
2188 /* Have we already been destroyed by a softirq or backlog? */
2189 if (state
!= TCP_CLOSE
&& sk
->sk_state
== TCP_CLOSE
)
2192 /* This is a (useful) BSD violating of the RFC. There is a
2193 * problem with TCP as specified in that the other end could
2194 * keep a socket open forever with no application left this end.
2195 * We use a 3 minute timeout (about the same as BSD) then kill
2196 * our end. If they send after that then tough - BUT: long enough
2197 * that we won't make the old 4*rto = almost no time - whoops
2200 * Nope, it was not mistake. It is really desired behaviour
2201 * f.e. on http servers, when such sockets are useless, but
2202 * consume significant resources. Let's do it with special
2203 * linger2 option. --ANK
2206 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
2207 struct tcp_sock
*tp
= tcp_sk(sk
);
2208 if (tp
->linger2
< 0) {
2209 tcp_set_state(sk
, TCP_CLOSE
);
2210 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2211 NET_INC_STATS_BH(sock_net(sk
),
2212 LINUX_MIB_TCPABORTONLINGER
);
2214 const int tmo
= tcp_fin_time(sk
);
2216 if (tmo
> TCP_TIMEWAIT_LEN
) {
2217 inet_csk_reset_keepalive_timer(sk
,
2218 tmo
- TCP_TIMEWAIT_LEN
);
2220 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
2225 if (sk
->sk_state
!= TCP_CLOSE
) {
2227 if (tcp_check_oom(sk
, 0)) {
2228 tcp_set_state(sk
, TCP_CLOSE
);
2229 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2230 NET_INC_STATS_BH(sock_net(sk
),
2231 LINUX_MIB_TCPABORTONMEMORY
);
2235 if (sk
->sk_state
== TCP_CLOSE
) {
2236 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
2237 /* We could get here with a non-NULL req if the socket is
2238 * aborted (e.g., closed with unread data) before 3WHS
2242 reqsk_fastopen_remove(sk
, req
, false);
2243 inet_csk_destroy_sock(sk
);
2245 /* Otherwise, socket is reprieved until protocol close. */
2252 EXPORT_SYMBOL(tcp_close
);
2254 /* These states need RST on ABORT according to RFC793 */
2256 static inline bool tcp_need_reset(int state
)
2258 return (1 << state
) &
2259 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
2260 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
2263 int tcp_disconnect(struct sock
*sk
, int flags
)
2265 struct inet_sock
*inet
= inet_sk(sk
);
2266 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2267 struct tcp_sock
*tp
= tcp_sk(sk
);
2269 int old_state
= sk
->sk_state
;
2271 if (old_state
!= TCP_CLOSE
)
2272 tcp_set_state(sk
, TCP_CLOSE
);
2274 /* ABORT function of RFC793 */
2275 if (old_state
== TCP_LISTEN
) {
2276 inet_csk_listen_stop(sk
);
2277 } else if (unlikely(tp
->repair
)) {
2278 sk
->sk_err
= ECONNABORTED
;
2279 } else if (tcp_need_reset(old_state
) ||
2280 (tp
->snd_nxt
!= tp
->write_seq
&&
2281 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
2282 /* The last check adjusts for discrepancy of Linux wrt. RFC
2285 tcp_send_active_reset(sk
, gfp_any());
2286 sk
->sk_err
= ECONNRESET
;
2287 } else if (old_state
== TCP_SYN_SENT
)
2288 sk
->sk_err
= ECONNRESET
;
2290 tcp_clear_xmit_timers(sk
);
2291 __skb_queue_purge(&sk
->sk_receive_queue
);
2292 tcp_write_queue_purge(sk
);
2293 __skb_queue_purge(&tp
->out_of_order_queue
);
2294 #ifdef CONFIG_NET_DMA
2295 __skb_queue_purge(&sk
->sk_async_wait_queue
);
2298 inet
->inet_dport
= 0;
2300 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
2301 inet_reset_saddr(sk
);
2303 sk
->sk_shutdown
= 0;
2304 sock_reset_flag(sk
, SOCK_DONE
);
2306 if ((tp
->write_seq
+= tp
->max_window
+ 2) == 0)
2308 icsk
->icsk_backoff
= 0;
2310 icsk
->icsk_probes_out
= 0;
2311 tp
->packets_out
= 0;
2312 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
2313 tp
->snd_cwnd_cnt
= 0;
2314 tp
->window_clamp
= 0;
2315 tcp_set_ca_state(sk
, TCP_CA_Open
);
2316 tcp_clear_retrans(tp
);
2317 inet_csk_delack_init(sk
);
2318 tcp_init_send_head(sk
);
2319 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
2322 WARN_ON(inet
->inet_num
&& !icsk
->icsk_bind_hash
);
2324 sk
->sk_error_report(sk
);
2327 EXPORT_SYMBOL(tcp_disconnect
);
2329 void tcp_sock_destruct(struct sock
*sk
)
2331 inet_sock_destruct(sk
);
2333 kfree(inet_csk(sk
)->icsk_accept_queue
.fastopenq
);
2336 static inline bool tcp_can_repair_sock(const struct sock
*sk
)
2338 return ns_capable(sock_net(sk
)->user_ns
, CAP_NET_ADMIN
) &&
2339 ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_ESTABLISHED
));
2342 static int tcp_repair_options_est(struct tcp_sock
*tp
,
2343 struct tcp_repair_opt __user
*optbuf
, unsigned int len
)
2345 struct tcp_repair_opt opt
;
2347 while (len
>= sizeof(opt
)) {
2348 if (copy_from_user(&opt
, optbuf
, sizeof(opt
)))
2354 switch (opt
.opt_code
) {
2356 tp
->rx_opt
.mss_clamp
= opt
.opt_val
;
2360 u16 snd_wscale
= opt
.opt_val
& 0xFFFF;
2361 u16 rcv_wscale
= opt
.opt_val
>> 16;
2363 if (snd_wscale
> 14 || rcv_wscale
> 14)
2366 tp
->rx_opt
.snd_wscale
= snd_wscale
;
2367 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2368 tp
->rx_opt
.wscale_ok
= 1;
2371 case TCPOPT_SACK_PERM
:
2372 if (opt
.opt_val
!= 0)
2375 tp
->rx_opt
.sack_ok
|= TCP_SACK_SEEN
;
2376 if (sysctl_tcp_fack
)
2377 tcp_enable_fack(tp
);
2379 case TCPOPT_TIMESTAMP
:
2380 if (opt
.opt_val
!= 0)
2383 tp
->rx_opt
.tstamp_ok
= 1;
2392 * Socket option code for TCP.
2394 static int do_tcp_setsockopt(struct sock
*sk
, int level
,
2395 int optname
, char __user
*optval
, unsigned int optlen
)
2397 struct tcp_sock
*tp
= tcp_sk(sk
);
2398 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2402 /* These are data/string values, all the others are ints */
2404 case TCP_CONGESTION
: {
2405 char name
[TCP_CA_NAME_MAX
];
2410 val
= strncpy_from_user(name
, optval
,
2411 min_t(long, TCP_CA_NAME_MAX
-1, optlen
));
2417 err
= tcp_set_congestion_control(sk
, name
);
2426 if (optlen
< sizeof(int))
2429 if (get_user(val
, (int __user
*)optval
))
2436 /* Values greater than interface MTU won't take effect. However
2437 * at the point when this call is done we typically don't yet
2438 * know which interface is going to be used */
2439 if (val
< TCP_MIN_MSS
|| val
> MAX_TCP_WINDOW
) {
2443 tp
->rx_opt
.user_mss
= val
;
2448 /* TCP_NODELAY is weaker than TCP_CORK, so that
2449 * this option on corked socket is remembered, but
2450 * it is not activated until cork is cleared.
2452 * However, when TCP_NODELAY is set we make
2453 * an explicit push, which overrides even TCP_CORK
2454 * for currently queued segments.
2456 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
2457 tcp_push_pending_frames(sk
);
2459 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
2463 case TCP_THIN_LINEAR_TIMEOUTS
:
2464 if (val
< 0 || val
> 1)
2470 case TCP_THIN_DUPACK
:
2471 if (val
< 0 || val
> 1)
2474 tp
->thin_dupack
= val
;
2475 if (tp
->thin_dupack
)
2476 tcp_disable_early_retrans(tp
);
2481 if (!tcp_can_repair_sock(sk
))
2483 else if (val
== 1) {
2485 sk
->sk_reuse
= SK_FORCE_REUSE
;
2486 tp
->repair_queue
= TCP_NO_QUEUE
;
2487 } else if (val
== 0) {
2489 sk
->sk_reuse
= SK_NO_REUSE
;
2490 tcp_send_window_probe(sk
);
2496 case TCP_REPAIR_QUEUE
:
2499 else if (val
< TCP_QUEUES_NR
)
2500 tp
->repair_queue
= val
;
2506 if (sk
->sk_state
!= TCP_CLOSE
)
2508 else if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2509 tp
->write_seq
= val
;
2510 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2516 case TCP_REPAIR_OPTIONS
:
2519 else if (sk
->sk_state
== TCP_ESTABLISHED
)
2520 err
= tcp_repair_options_est(tp
,
2521 (struct tcp_repair_opt __user
*)optval
,
2528 /* When set indicates to always queue non-full frames.
2529 * Later the user clears this option and we transmit
2530 * any pending partial frames in the queue. This is
2531 * meant to be used alongside sendfile() to get properly
2532 * filled frames when the user (for example) must write
2533 * out headers with a write() call first and then use
2534 * sendfile to send out the data parts.
2536 * TCP_CORK can be set together with TCP_NODELAY and it is
2537 * stronger than TCP_NODELAY.
2540 tp
->nonagle
|= TCP_NAGLE_CORK
;
2542 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
2543 if (tp
->nonagle
&TCP_NAGLE_OFF
)
2544 tp
->nonagle
|= TCP_NAGLE_PUSH
;
2545 tcp_push_pending_frames(sk
);
2550 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
2553 tp
->keepalive_time
= val
* HZ
;
2554 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
2555 !((1 << sk
->sk_state
) &
2556 (TCPF_CLOSE
| TCPF_LISTEN
))) {
2557 u32 elapsed
= keepalive_time_elapsed(tp
);
2558 if (tp
->keepalive_time
> elapsed
)
2559 elapsed
= tp
->keepalive_time
- elapsed
;
2562 inet_csk_reset_keepalive_timer(sk
, elapsed
);
2567 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
2570 tp
->keepalive_intvl
= val
* HZ
;
2573 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
2576 tp
->keepalive_probes
= val
;
2579 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
2582 icsk
->icsk_syn_retries
= val
;
2588 else if (val
> sysctl_tcp_fin_timeout
/ HZ
)
2591 tp
->linger2
= val
* HZ
;
2594 case TCP_DEFER_ACCEPT
:
2595 /* Translate value in seconds to number of retransmits */
2596 icsk
->icsk_accept_queue
.rskq_defer_accept
=
2597 secs_to_retrans(val
, TCP_TIMEOUT_INIT
/ HZ
,
2601 case TCP_WINDOW_CLAMP
:
2603 if (sk
->sk_state
!= TCP_CLOSE
) {
2607 tp
->window_clamp
= 0;
2609 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
2610 SOCK_MIN_RCVBUF
/ 2 : val
;
2615 icsk
->icsk_ack
.pingpong
= 1;
2617 icsk
->icsk_ack
.pingpong
= 0;
2618 if ((1 << sk
->sk_state
) &
2619 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
2620 inet_csk_ack_scheduled(sk
)) {
2621 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
2622 tcp_cleanup_rbuf(sk
, 1);
2624 icsk
->icsk_ack
.pingpong
= 1;
2629 #ifdef CONFIG_TCP_MD5SIG
2631 /* Read the IP->Key mappings from userspace */
2632 err
= tp
->af_specific
->md5_parse(sk
, optval
, optlen
);
2635 case TCP_USER_TIMEOUT
:
2636 /* Cap the max timeout in ms TCP will retry/retrans
2637 * before giving up and aborting (ETIMEDOUT) a connection.
2642 icsk
->icsk_user_timeout
= msecs_to_jiffies(val
);
2646 if (val
>= 0 && ((1 << sk
->sk_state
) & (TCPF_CLOSE
|
2648 err
= fastopen_init_queue(sk
, val
);
2656 tp
->tsoffset
= val
- tcp_time_stamp
;
2658 case TCP_NOTSENT_LOWAT
:
2659 tp
->notsent_lowat
= val
;
2660 sk
->sk_write_space(sk
);
2671 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2672 unsigned int optlen
)
2674 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2676 if (level
!= SOL_TCP
)
2677 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
2679 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2681 EXPORT_SYMBOL(tcp_setsockopt
);
2683 #ifdef CONFIG_COMPAT
2684 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
2685 char __user
*optval
, unsigned int optlen
)
2687 if (level
!= SOL_TCP
)
2688 return inet_csk_compat_setsockopt(sk
, level
, optname
,
2690 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2692 EXPORT_SYMBOL(compat_tcp_setsockopt
);
2695 /* Return information about state of tcp endpoint in API format. */
2696 void tcp_get_info(const struct sock
*sk
, struct tcp_info
*info
)
2698 const struct tcp_sock
*tp
= tcp_sk(sk
);
2699 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2700 u32 now
= tcp_time_stamp
;
2702 memset(info
, 0, sizeof(*info
));
2704 info
->tcpi_state
= sk
->sk_state
;
2705 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
2706 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
2707 info
->tcpi_probes
= icsk
->icsk_probes_out
;
2708 info
->tcpi_backoff
= icsk
->icsk_backoff
;
2710 if (tp
->rx_opt
.tstamp_ok
)
2711 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
2712 if (tcp_is_sack(tp
))
2713 info
->tcpi_options
|= TCPI_OPT_SACK
;
2714 if (tp
->rx_opt
.wscale_ok
) {
2715 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
2716 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
2717 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
2720 if (tp
->ecn_flags
& TCP_ECN_OK
)
2721 info
->tcpi_options
|= TCPI_OPT_ECN
;
2722 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
2723 info
->tcpi_options
|= TCPI_OPT_ECN_SEEN
;
2724 if (tp
->syn_data_acked
)
2725 info
->tcpi_options
|= TCPI_OPT_SYN_DATA
;
2727 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
2728 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
2729 info
->tcpi_snd_mss
= tp
->mss_cache
;
2730 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
2732 if (sk
->sk_state
== TCP_LISTEN
) {
2733 info
->tcpi_unacked
= sk
->sk_ack_backlog
;
2734 info
->tcpi_sacked
= sk
->sk_max_ack_backlog
;
2736 info
->tcpi_unacked
= tp
->packets_out
;
2737 info
->tcpi_sacked
= tp
->sacked_out
;
2739 info
->tcpi_lost
= tp
->lost_out
;
2740 info
->tcpi_retrans
= tp
->retrans_out
;
2741 info
->tcpi_fackets
= tp
->fackets_out
;
2743 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
2744 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
2745 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
2747 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
2748 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
2749 info
->tcpi_rtt
= jiffies_to_usecs(tp
->srtt
)>>3;
2750 info
->tcpi_rttvar
= jiffies_to_usecs(tp
->mdev
)>>2;
2751 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
2752 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
2753 info
->tcpi_advmss
= tp
->advmss
;
2754 info
->tcpi_reordering
= tp
->reordering
;
2756 info
->tcpi_rcv_rtt
= jiffies_to_usecs(tp
->rcv_rtt_est
.rtt
)>>3;
2757 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
2759 info
->tcpi_total_retrans
= tp
->total_retrans
;
2761 EXPORT_SYMBOL_GPL(tcp_get_info
);
2763 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
2764 int optname
, char __user
*optval
, int __user
*optlen
)
2766 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2767 struct tcp_sock
*tp
= tcp_sk(sk
);
2770 if (get_user(len
, optlen
))
2773 len
= min_t(unsigned int, len
, sizeof(int));
2780 val
= tp
->mss_cache
;
2781 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
2782 val
= tp
->rx_opt
.user_mss
;
2784 val
= tp
->rx_opt
.mss_clamp
;
2787 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
2790 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
2793 val
= keepalive_time_when(tp
) / HZ
;
2796 val
= keepalive_intvl_when(tp
) / HZ
;
2799 val
= keepalive_probes(tp
);
2802 val
= icsk
->icsk_syn_retries
? : sysctl_tcp_syn_retries
;
2807 val
= (val
? : sysctl_tcp_fin_timeout
) / HZ
;
2809 case TCP_DEFER_ACCEPT
:
2810 val
= retrans_to_secs(icsk
->icsk_accept_queue
.rskq_defer_accept
,
2811 TCP_TIMEOUT_INIT
/ HZ
, TCP_RTO_MAX
/ HZ
);
2813 case TCP_WINDOW_CLAMP
:
2814 val
= tp
->window_clamp
;
2817 struct tcp_info info
;
2819 if (get_user(len
, optlen
))
2822 tcp_get_info(sk
, &info
);
2824 len
= min_t(unsigned int, len
, sizeof(info
));
2825 if (put_user(len
, optlen
))
2827 if (copy_to_user(optval
, &info
, len
))
2832 val
= !icsk
->icsk_ack
.pingpong
;
2835 case TCP_CONGESTION
:
2836 if (get_user(len
, optlen
))
2838 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
2839 if (put_user(len
, optlen
))
2841 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
2845 case TCP_THIN_LINEAR_TIMEOUTS
:
2848 case TCP_THIN_DUPACK
:
2849 val
= tp
->thin_dupack
;
2856 case TCP_REPAIR_QUEUE
:
2858 val
= tp
->repair_queue
;
2864 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2865 val
= tp
->write_seq
;
2866 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2872 case TCP_USER_TIMEOUT
:
2873 val
= jiffies_to_msecs(icsk
->icsk_user_timeout
);
2876 val
= tcp_time_stamp
+ tp
->tsoffset
;
2878 case TCP_NOTSENT_LOWAT
:
2879 val
= tp
->notsent_lowat
;
2882 return -ENOPROTOOPT
;
2885 if (put_user(len
, optlen
))
2887 if (copy_to_user(optval
, &val
, len
))
2892 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2895 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2897 if (level
!= SOL_TCP
)
2898 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
2900 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2902 EXPORT_SYMBOL(tcp_getsockopt
);
2904 #ifdef CONFIG_COMPAT
2905 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
2906 char __user
*optval
, int __user
*optlen
)
2908 if (level
!= SOL_TCP
)
2909 return inet_csk_compat_getsockopt(sk
, level
, optname
,
2911 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2913 EXPORT_SYMBOL(compat_tcp_getsockopt
);
2916 #ifdef CONFIG_TCP_MD5SIG
2917 static struct tcp_md5sig_pool __percpu
*tcp_md5sig_pool __read_mostly
;
2918 static DEFINE_MUTEX(tcp_md5sig_mutex
);
2920 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu
*pool
)
2924 for_each_possible_cpu(cpu
) {
2925 struct tcp_md5sig_pool
*p
= per_cpu_ptr(pool
, cpu
);
2927 if (p
->md5_desc
.tfm
)
2928 crypto_free_hash(p
->md5_desc
.tfm
);
2933 static void __tcp_alloc_md5sig_pool(void)
2936 struct tcp_md5sig_pool __percpu
*pool
;
2938 pool
= alloc_percpu(struct tcp_md5sig_pool
);
2942 for_each_possible_cpu(cpu
) {
2943 struct crypto_hash
*hash
;
2945 hash
= crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC
);
2946 if (IS_ERR_OR_NULL(hash
))
2949 per_cpu_ptr(pool
, cpu
)->md5_desc
.tfm
= hash
;
2951 /* before setting tcp_md5sig_pool, we must commit all writes
2952 * to memory. See ACCESS_ONCE() in tcp_get_md5sig_pool()
2955 tcp_md5sig_pool
= pool
;
2958 __tcp_free_md5sig_pool(pool
);
2961 bool tcp_alloc_md5sig_pool(void)
2963 if (unlikely(!tcp_md5sig_pool
)) {
2964 mutex_lock(&tcp_md5sig_mutex
);
2966 if (!tcp_md5sig_pool
)
2967 __tcp_alloc_md5sig_pool();
2969 mutex_unlock(&tcp_md5sig_mutex
);
2971 return tcp_md5sig_pool
!= NULL
;
2973 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
2977 * tcp_get_md5sig_pool - get md5sig_pool for this user
2979 * We use percpu structure, so if we succeed, we exit with preemption
2980 * and BH disabled, to make sure another thread or softirq handling
2981 * wont try to get same context.
2983 struct tcp_md5sig_pool
*tcp_get_md5sig_pool(void)
2985 struct tcp_md5sig_pool __percpu
*p
;
2988 p
= ACCESS_ONCE(tcp_md5sig_pool
);
2990 return __this_cpu_ptr(p
);
2995 EXPORT_SYMBOL(tcp_get_md5sig_pool
);
2997 int tcp_md5_hash_header(struct tcp_md5sig_pool
*hp
,
2998 const struct tcphdr
*th
)
3000 struct scatterlist sg
;
3004 /* We are not allowed to change tcphdr, make a local copy */
3005 memcpy(&hdr
, th
, sizeof(hdr
));
3008 /* options aren't included in the hash */
3009 sg_init_one(&sg
, &hdr
, sizeof(hdr
));
3010 err
= crypto_hash_update(&hp
->md5_desc
, &sg
, sizeof(hdr
));
3013 EXPORT_SYMBOL(tcp_md5_hash_header
);
3015 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool
*hp
,
3016 const struct sk_buff
*skb
, unsigned int header_len
)
3018 struct scatterlist sg
;
3019 const struct tcphdr
*tp
= tcp_hdr(skb
);
3020 struct hash_desc
*desc
= &hp
->md5_desc
;
3022 const unsigned int head_data_len
= skb_headlen(skb
) > header_len
?
3023 skb_headlen(skb
) - header_len
: 0;
3024 const struct skb_shared_info
*shi
= skb_shinfo(skb
);
3025 struct sk_buff
*frag_iter
;
3027 sg_init_table(&sg
, 1);
3029 sg_set_buf(&sg
, ((u8
*) tp
) + header_len
, head_data_len
);
3030 if (crypto_hash_update(desc
, &sg
, head_data_len
))
3033 for (i
= 0; i
< shi
->nr_frags
; ++i
) {
3034 const struct skb_frag_struct
*f
= &shi
->frags
[i
];
3035 unsigned int offset
= f
->page_offset
;
3036 struct page
*page
= skb_frag_page(f
) + (offset
>> PAGE_SHIFT
);
3038 sg_set_page(&sg
, page
, skb_frag_size(f
),
3039 offset_in_page(offset
));
3040 if (crypto_hash_update(desc
, &sg
, skb_frag_size(f
)))
3044 skb_walk_frags(skb
, frag_iter
)
3045 if (tcp_md5_hash_skb_data(hp
, frag_iter
, 0))
3050 EXPORT_SYMBOL(tcp_md5_hash_skb_data
);
3052 int tcp_md5_hash_key(struct tcp_md5sig_pool
*hp
, const struct tcp_md5sig_key
*key
)
3054 struct scatterlist sg
;
3056 sg_init_one(&sg
, key
->key
, key
->keylen
);
3057 return crypto_hash_update(&hp
->md5_desc
, &sg
, key
->keylen
);
3059 EXPORT_SYMBOL(tcp_md5_hash_key
);
3063 void tcp_done(struct sock
*sk
)
3065 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
3067 if (sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
3068 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_ATTEMPTFAILS
);
3070 tcp_set_state(sk
, TCP_CLOSE
);
3071 tcp_clear_xmit_timers(sk
);
3073 reqsk_fastopen_remove(sk
, req
, false);
3075 sk
->sk_shutdown
= SHUTDOWN_MASK
;
3077 if (!sock_flag(sk
, SOCK_DEAD
))
3078 sk
->sk_state_change(sk
);
3080 inet_csk_destroy_sock(sk
);
3082 EXPORT_SYMBOL_GPL(tcp_done
);
3084 extern struct tcp_congestion_ops tcp_reno
;
3086 static __initdata
unsigned long thash_entries
;
3087 static int __init
set_thash_entries(char *str
)
3094 ret
= kstrtoul(str
, 0, &thash_entries
);
3100 __setup("thash_entries=", set_thash_entries
);
3102 static void tcp_init_mem(void)
3104 unsigned long limit
= nr_free_buffer_pages() / 8;
3105 limit
= max(limit
, 128UL);
3106 sysctl_tcp_mem
[0] = limit
/ 4 * 3;
3107 sysctl_tcp_mem
[1] = limit
;
3108 sysctl_tcp_mem
[2] = sysctl_tcp_mem
[0] * 2;
3111 void __init
tcp_init(void)
3113 struct sk_buff
*skb
= NULL
;
3114 unsigned long limit
;
3115 int max_rshare
, max_wshare
, cnt
;
3118 BUILD_BUG_ON(sizeof(struct tcp_skb_cb
) > sizeof(skb
->cb
));
3120 percpu_counter_init(&tcp_sockets_allocated
, 0);
3121 percpu_counter_init(&tcp_orphan_count
, 0);
3122 tcp_hashinfo
.bind_bucket_cachep
=
3123 kmem_cache_create("tcp_bind_bucket",
3124 sizeof(struct inet_bind_bucket
), 0,
3125 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
);
3127 /* Size and allocate the main established and bind bucket
3130 * The methodology is similar to that of the buffer cache.
3132 tcp_hashinfo
.ehash
=
3133 alloc_large_system_hash("TCP established",
3134 sizeof(struct inet_ehash_bucket
),
3136 17, /* one slot per 128 KB of memory */
3139 &tcp_hashinfo
.ehash_mask
,
3141 thash_entries
? 0 : 512 * 1024);
3142 for (i
= 0; i
<= tcp_hashinfo
.ehash_mask
; i
++)
3143 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].chain
, i
);
3145 if (inet_ehash_locks_alloc(&tcp_hashinfo
))
3146 panic("TCP: failed to alloc ehash_locks");
3147 tcp_hashinfo
.bhash
=
3148 alloc_large_system_hash("TCP bind",
3149 sizeof(struct inet_bind_hashbucket
),
3150 tcp_hashinfo
.ehash_mask
+ 1,
3151 17, /* one slot per 128 KB of memory */
3153 &tcp_hashinfo
.bhash_size
,
3157 tcp_hashinfo
.bhash_size
= 1U << tcp_hashinfo
.bhash_size
;
3158 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
3159 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
3160 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
3164 cnt
= tcp_hashinfo
.ehash_mask
+ 1;
3166 tcp_death_row
.sysctl_max_tw_buckets
= cnt
/ 2;
3167 sysctl_tcp_max_orphans
= cnt
/ 2;
3168 sysctl_max_syn_backlog
= max(128, cnt
/ 256);
3171 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3172 limit
= nr_free_buffer_pages() << (PAGE_SHIFT
- 7);
3173 max_wshare
= min(4UL*1024*1024, limit
);
3174 max_rshare
= min(6UL*1024*1024, limit
);
3176 sysctl_tcp_wmem
[0] = SK_MEM_QUANTUM
;
3177 sysctl_tcp_wmem
[1] = 16*1024;
3178 sysctl_tcp_wmem
[2] = max(64*1024, max_wshare
);
3180 sysctl_tcp_rmem
[0] = SK_MEM_QUANTUM
;
3181 sysctl_tcp_rmem
[1] = 87380;
3182 sysctl_tcp_rmem
[2] = max(87380, max_rshare
);
3184 pr_info("Hash tables configured (established %u bind %u)\n",
3185 tcp_hashinfo
.ehash_mask
+ 1, tcp_hashinfo
.bhash_size
);
3189 tcp_register_congestion_control(&tcp_reno
);