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 <crypto/hash.h>
251 #include <linux/kernel.h>
252 #include <linux/module.h>
253 #include <linux/types.h>
254 #include <linux/fcntl.h>
255 #include <linux/poll.h>
256 #include <linux/inet_diag.h>
257 #include <linux/init.h>
258 #include <linux/fs.h>
259 #include <linux/skbuff.h>
260 #include <linux/scatterlist.h>
261 #include <linux/splice.h>
262 #include <linux/net.h>
263 #include <linux/socket.h>
264 #include <linux/random.h>
265 #include <linux/bootmem.h>
266 #include <linux/highmem.h>
267 #include <linux/swap.h>
268 #include <linux/cache.h>
269 #include <linux/err.h>
270 #include <linux/time.h>
271 #include <linux/slab.h>
273 #include <net/icmp.h>
274 #include <net/inet_common.h>
276 #include <net/xfrm.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_min_tso_segs __read_mostly
= 2;
286 int sysctl_tcp_autocorking __read_mostly
= 1;
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 tp
->out_of_order_queue
= RB_ROOT
;
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_us
= jiffies_to_usecs(TCP_TIMEOUT_INIT
);
389 minmax_reset(&tp
->rtt_min
, tcp_time_stamp
, ~0U);
391 /* So many TCP implementations out there (incorrectly) count the
392 * initial SYN frame in their delayed-ACK and congestion control
393 * algorithms that we must have the following bandaid to talk
394 * efficiently to them. -DaveM
396 tp
->snd_cwnd
= TCP_INIT_CWND
;
398 /* There's a bubble in the pipe until at least the first ACK. */
399 tp
->app_limited
= ~0U;
401 /* See draft-stevens-tcpca-spec-01 for discussion of the
402 * initialization of these values.
404 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
405 tp
->snd_cwnd_clamp
= ~0;
406 tp
->mss_cache
= TCP_MSS_DEFAULT
;
408 tp
->reordering
= sock_net(sk
)->ipv4
.sysctl_tcp_reordering
;
409 tcp_enable_early_retrans(tp
);
410 tcp_assign_congestion_control(sk
);
414 sk
->sk_state
= TCP_CLOSE
;
416 sk
->sk_write_space
= sk_stream_write_space
;
417 sock_set_flag(sk
, SOCK_USE_WRITE_QUEUE
);
419 icsk
->icsk_sync_mss
= tcp_sync_mss
;
421 sk
->sk_sndbuf
= sysctl_tcp_wmem
[1];
422 sk
->sk_rcvbuf
= sysctl_tcp_rmem
[1];
425 sk_sockets_allocated_inc(sk
);
428 EXPORT_SYMBOL(tcp_init_sock
);
430 static void tcp_tx_timestamp(struct sock
*sk
, u16 tsflags
, struct sk_buff
*skb
)
433 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
434 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
436 sock_tx_timestamp(sk
, tsflags
, &shinfo
->tx_flags
);
437 if (tsflags
& SOF_TIMESTAMPING_TX_ACK
)
438 tcb
->txstamp_ack
= 1;
439 if (tsflags
& SOF_TIMESTAMPING_TX_RECORD_MASK
)
440 shinfo
->tskey
= TCP_SKB_CB(skb
)->seq
+ skb
->len
- 1;
445 * Wait for a TCP event.
447 * Note that we don't need to lock the socket, as the upper poll layers
448 * take care of normal races (between the test and the event) and we don't
449 * go look at any of the socket buffers directly.
451 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
454 struct sock
*sk
= sock
->sk
;
455 const struct tcp_sock
*tp
= tcp_sk(sk
);
458 sock_rps_record_flow(sk
);
460 sock_poll_wait(file
, sk_sleep(sk
), wait
);
462 state
= sk_state_load(sk
);
463 if (state
== TCP_LISTEN
)
464 return inet_csk_listen_poll(sk
);
466 /* Socket is not locked. We are protected from async events
467 * by poll logic and correct handling of state changes
468 * made by other threads is impossible in any case.
474 * POLLHUP is certainly not done right. But poll() doesn't
475 * have a notion of HUP in just one direction, and for a
476 * socket the read side is more interesting.
478 * Some poll() documentation says that POLLHUP is incompatible
479 * with the POLLOUT/POLLWR flags, so somebody should check this
480 * all. But careful, it tends to be safer to return too many
481 * bits than too few, and you can easily break real applications
482 * if you don't tell them that something has hung up!
486 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
487 * our fs/select.c). It means that after we received EOF,
488 * poll always returns immediately, making impossible poll() on write()
489 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
490 * if and only if shutdown has been made in both directions.
491 * Actually, it is interesting to look how Solaris and DUX
492 * solve this dilemma. I would prefer, if POLLHUP were maskable,
493 * then we could set it on SND_SHUTDOWN. BTW examples given
494 * in Stevens' books assume exactly this behaviour, it explains
495 * why POLLHUP is incompatible with POLLOUT. --ANK
497 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
498 * blocking on fresh not-connected or disconnected socket. --ANK
500 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| state
== TCP_CLOSE
)
502 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
503 mask
|= POLLIN
| POLLRDNORM
| POLLRDHUP
;
505 /* Connected or passive Fast Open socket? */
506 if (state
!= TCP_SYN_SENT
&&
507 (state
!= TCP_SYN_RECV
|| tp
->fastopen_rsk
)) {
508 int target
= sock_rcvlowat(sk
, 0, INT_MAX
);
510 if (tp
->urg_seq
== tp
->copied_seq
&&
511 !sock_flag(sk
, SOCK_URGINLINE
) &&
515 if (tp
->rcv_nxt
- tp
->copied_seq
>= target
)
516 mask
|= POLLIN
| POLLRDNORM
;
518 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
519 if (sk_stream_is_writeable(sk
)) {
520 mask
|= POLLOUT
| POLLWRNORM
;
521 } else { /* send SIGIO later */
522 sk_set_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
523 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
525 /* Race breaker. If space is freed after
526 * wspace test but before the flags are set,
527 * IO signal will be lost. Memory barrier
528 * pairs with the input side.
530 smp_mb__after_atomic();
531 if (sk_stream_is_writeable(sk
))
532 mask
|= POLLOUT
| POLLWRNORM
;
535 mask
|= POLLOUT
| POLLWRNORM
;
537 if (tp
->urg_data
& TCP_URG_VALID
)
540 /* This barrier is coupled with smp_wmb() in tcp_reset() */
542 if (sk
->sk_err
|| !skb_queue_empty(&sk
->sk_error_queue
))
547 EXPORT_SYMBOL(tcp_poll
);
549 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
551 struct tcp_sock
*tp
= tcp_sk(sk
);
557 if (sk
->sk_state
== TCP_LISTEN
)
560 slow
= lock_sock_fast(sk
);
562 unlock_sock_fast(sk
, slow
);
565 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
568 if (sk
->sk_state
== TCP_LISTEN
)
571 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
574 answ
= tp
->write_seq
- tp
->snd_una
;
577 if (sk
->sk_state
== TCP_LISTEN
)
580 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
583 answ
= tp
->write_seq
- tp
->snd_nxt
;
589 return put_user(answ
, (int __user
*)arg
);
591 EXPORT_SYMBOL(tcp_ioctl
);
593 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
595 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
596 tp
->pushed_seq
= tp
->write_seq
;
599 static inline bool forced_push(const struct tcp_sock
*tp
)
601 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
604 static void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
606 struct tcp_sock
*tp
= tcp_sk(sk
);
607 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
610 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
611 tcb
->tcp_flags
= TCPHDR_ACK
;
613 __skb_header_release(skb
);
614 tcp_add_write_queue_tail(sk
, skb
);
615 sk
->sk_wmem_queued
+= skb
->truesize
;
616 sk_mem_charge(sk
, skb
->truesize
);
617 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
618 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
620 tcp_slow_start_after_idle_check(sk
);
623 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
)
626 tp
->snd_up
= tp
->write_seq
;
629 /* If a not yet filled skb is pushed, do not send it if
630 * we have data packets in Qdisc or NIC queues :
631 * Because TX completion will happen shortly, it gives a chance
632 * to coalesce future sendmsg() payload into this skb, without
633 * need for a timer, and with no latency trade off.
634 * As packets containing data payload have a bigger truesize
635 * than pure acks (dataless) packets, the last checks prevent
636 * autocorking if we only have an ACK in Qdisc/NIC queues,
637 * or if TX completion was delayed after we processed ACK packet.
639 static bool tcp_should_autocork(struct sock
*sk
, struct sk_buff
*skb
,
642 return skb
->len
< size_goal
&&
643 sysctl_tcp_autocorking
&&
644 skb
!= tcp_write_queue_head(sk
) &&
645 atomic_read(&sk
->sk_wmem_alloc
) > skb
->truesize
;
648 static void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
649 int nonagle
, int size_goal
)
651 struct tcp_sock
*tp
= tcp_sk(sk
);
654 if (!tcp_send_head(sk
))
657 skb
= tcp_write_queue_tail(sk
);
658 if (!(flags
& MSG_MORE
) || forced_push(tp
))
659 tcp_mark_push(tp
, skb
);
661 tcp_mark_urg(tp
, flags
);
663 if (tcp_should_autocork(sk
, skb
, size_goal
)) {
665 /* avoid atomic op if TSQ_THROTTLED bit is already set */
666 if (!test_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
)) {
667 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPAUTOCORKING
);
668 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
670 /* It is possible TX completion already happened
671 * before we set TSQ_THROTTLED.
673 if (atomic_read(&sk
->sk_wmem_alloc
) > skb
->truesize
)
677 if (flags
& MSG_MORE
)
678 nonagle
= TCP_NAGLE_CORK
;
680 __tcp_push_pending_frames(sk
, mss_now
, nonagle
);
683 static int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
684 unsigned int offset
, size_t len
)
686 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
689 ret
= skb_splice_bits(skb
, skb
->sk
, offset
, tss
->pipe
,
690 min(rd_desc
->count
, len
), tss
->flags
);
692 rd_desc
->count
-= ret
;
696 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
698 /* Store TCP splice context information in read_descriptor_t. */
699 read_descriptor_t rd_desc
= {
704 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
708 * tcp_splice_read - splice data from TCP socket to a pipe
709 * @sock: socket to splice from
710 * @ppos: position (not valid)
711 * @pipe: pipe to splice to
712 * @len: number of bytes to splice
713 * @flags: splice modifier flags
716 * Will read pages from given socket and fill them into a pipe.
719 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
720 struct pipe_inode_info
*pipe
, size_t len
,
723 struct sock
*sk
= sock
->sk
;
724 struct tcp_splice_state tss
= {
733 sock_rps_record_flow(sk
);
735 * We can't seek on a socket input
744 timeo
= sock_rcvtimeo(sk
, sock
->file
->f_flags
& O_NONBLOCK
);
746 ret
= __tcp_splice_read(sk
, &tss
);
752 if (sock_flag(sk
, SOCK_DONE
))
755 ret
= sock_error(sk
);
758 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
760 if (sk
->sk_state
== TCP_CLOSE
) {
762 * This occurs when user tries to read
763 * from never connected socket.
765 if (!sock_flag(sk
, SOCK_DONE
))
773 sk_wait_data(sk
, &timeo
, NULL
);
774 if (signal_pending(current
)) {
775 ret
= sock_intr_errno(timeo
);
788 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
789 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
790 signal_pending(current
))
801 EXPORT_SYMBOL(tcp_splice_read
);
803 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
,
808 /* The TCP header must be at least 32-bit aligned. */
809 size
= ALIGN(size
, 4);
811 if (unlikely(tcp_under_memory_pressure(sk
)))
812 sk_mem_reclaim_partial(sk
);
814 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
818 if (force_schedule
) {
819 mem_scheduled
= true;
820 sk_forced_mem_schedule(sk
, skb
->truesize
);
822 mem_scheduled
= sk_wmem_schedule(sk
, skb
->truesize
);
824 if (likely(mem_scheduled
)) {
825 skb_reserve(skb
, sk
->sk_prot
->max_header
);
827 * Make sure that we have exactly size bytes
828 * available to the caller, no more, no less.
830 skb
->reserved_tailroom
= skb
->end
- skb
->tail
- size
;
835 sk
->sk_prot
->enter_memory_pressure(sk
);
836 sk_stream_moderate_sndbuf(sk
);
841 static unsigned int tcp_xmit_size_goal(struct sock
*sk
, u32 mss_now
,
844 struct tcp_sock
*tp
= tcp_sk(sk
);
845 u32 new_size_goal
, size_goal
;
847 if (!large_allowed
|| !sk_can_gso(sk
))
850 /* Note : tcp_tso_autosize() will eventually split this later */
851 new_size_goal
= sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
;
852 new_size_goal
= tcp_bound_to_half_wnd(tp
, new_size_goal
);
854 /* We try hard to avoid divides here */
855 size_goal
= tp
->gso_segs
* mss_now
;
856 if (unlikely(new_size_goal
< size_goal
||
857 new_size_goal
>= size_goal
+ mss_now
)) {
858 tp
->gso_segs
= min_t(u16
, new_size_goal
/ mss_now
,
859 sk
->sk_gso_max_segs
);
860 size_goal
= tp
->gso_segs
* mss_now
;
863 return max(size_goal
, mss_now
);
866 static int tcp_send_mss(struct sock
*sk
, int *size_goal
, int flags
)
870 mss_now
= tcp_current_mss(sk
);
871 *size_goal
= tcp_xmit_size_goal(sk
, mss_now
, !(flags
& MSG_OOB
));
876 static ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
*page
, int offset
,
877 size_t size
, int flags
)
879 struct tcp_sock
*tp
= tcp_sk(sk
);
880 int mss_now
, size_goal
;
883 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
885 /* Wait for a connection to finish. One exception is TCP Fast Open
886 * (passive side) where data is allowed to be sent before a connection
887 * is fully established.
889 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
890 !tcp_passive_fastopen(sk
)) {
891 err
= sk_stream_wait_connect(sk
, &timeo
);
896 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
898 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
902 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
906 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
910 if (!tcp_send_head(sk
) || (copy
= size_goal
- skb
->len
) <= 0 ||
911 !tcp_skb_can_collapse_to(skb
)) {
913 if (!sk_stream_memory_free(sk
))
914 goto wait_for_sndbuf
;
916 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
,
917 skb_queue_empty(&sk
->sk_write_queue
));
919 goto wait_for_memory
;
928 i
= skb_shinfo(skb
)->nr_frags
;
929 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
930 if (!can_coalesce
&& i
>= sysctl_max_skb_frags
) {
931 tcp_mark_push(tp
, skb
);
934 if (!sk_wmem_schedule(sk
, copy
))
935 goto wait_for_memory
;
938 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
941 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
943 skb_shinfo(skb
)->tx_flags
|= SKBTX_SHARED_FRAG
;
946 skb
->data_len
+= copy
;
947 skb
->truesize
+= copy
;
948 sk
->sk_wmem_queued
+= copy
;
949 sk_mem_charge(sk
, copy
);
950 skb
->ip_summed
= CHECKSUM_PARTIAL
;
951 tp
->write_seq
+= copy
;
952 TCP_SKB_CB(skb
)->end_seq
+= copy
;
953 tcp_skb_pcount_set(skb
, 0);
956 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
962 tcp_tx_timestamp(sk
, sk
->sk_tsflags
, skb
);
966 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
969 if (forced_push(tp
)) {
970 tcp_mark_push(tp
, skb
);
971 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
972 } else if (skb
== tcp_send_head(sk
))
973 tcp_push_one(sk
, mss_now
);
977 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
979 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
,
980 TCP_NAGLE_PUSH
, size_goal
);
982 err
= sk_stream_wait_memory(sk
, &timeo
);
986 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
990 if (copied
&& !(flags
& MSG_SENDPAGE_NOTLAST
))
991 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
, size_goal
);
998 /* make sure we wake any epoll edge trigger waiter */
999 if (unlikely(skb_queue_len(&sk
->sk_write_queue
) == 0 &&
1001 sk
->sk_write_space(sk
);
1002 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1004 return sk_stream_error(sk
, flags
, err
);
1007 int tcp_sendpage(struct sock
*sk
, struct page
*page
, int offset
,
1008 size_t size
, int flags
)
1012 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
1013 !sk_check_csum_caps(sk
))
1014 return sock_no_sendpage(sk
->sk_socket
, page
, offset
, size
,
1019 tcp_rate_check_app_limited(sk
); /* is sending application-limited? */
1021 res
= do_tcp_sendpages(sk
, page
, offset
, size
, flags
);
1025 EXPORT_SYMBOL(tcp_sendpage
);
1027 /* Do not bother using a page frag for very small frames.
1028 * But use this heuristic only for the first skb in write queue.
1030 * Having no payload in skb->head allows better SACK shifting
1031 * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1032 * write queue has less skbs.
1033 * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1034 * This also speeds up tso_fragment(), since it wont fallback
1035 * to tcp_fragment().
1037 static int linear_payload_sz(bool first_skb
)
1040 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER
);
1044 static int select_size(const struct sock
*sk
, bool sg
, bool first_skb
)
1046 const struct tcp_sock
*tp
= tcp_sk(sk
);
1047 int tmp
= tp
->mss_cache
;
1050 if (sk_can_gso(sk
)) {
1051 tmp
= linear_payload_sz(first_skb
);
1053 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
1055 if (tmp
>= pgbreak
&&
1056 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
1064 void tcp_free_fastopen_req(struct tcp_sock
*tp
)
1066 if (tp
->fastopen_req
) {
1067 kfree(tp
->fastopen_req
);
1068 tp
->fastopen_req
= NULL
;
1072 static int tcp_sendmsg_fastopen(struct sock
*sk
, struct msghdr
*msg
,
1073 int *copied
, size_t size
)
1075 struct tcp_sock
*tp
= tcp_sk(sk
);
1078 if (!(sysctl_tcp_fastopen
& TFO_CLIENT_ENABLE
))
1080 if (tp
->fastopen_req
)
1081 return -EALREADY
; /* Another Fast Open is in progress */
1083 tp
->fastopen_req
= kzalloc(sizeof(struct tcp_fastopen_request
),
1085 if (unlikely(!tp
->fastopen_req
))
1087 tp
->fastopen_req
->data
= msg
;
1088 tp
->fastopen_req
->size
= size
;
1090 flags
= (msg
->msg_flags
& MSG_DONTWAIT
) ? O_NONBLOCK
: 0;
1091 err
= __inet_stream_connect(sk
->sk_socket
, msg
->msg_name
,
1092 msg
->msg_namelen
, flags
);
1093 *copied
= tp
->fastopen_req
->copied
;
1094 tcp_free_fastopen_req(tp
);
1098 int tcp_sendmsg(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
1100 struct tcp_sock
*tp
= tcp_sk(sk
);
1101 struct sk_buff
*skb
;
1102 struct sockcm_cookie sockc
;
1103 int flags
, err
, copied
= 0;
1104 int mss_now
= 0, size_goal
, copied_syn
= 0;
1105 bool process_backlog
= false;
1111 flags
= msg
->msg_flags
;
1112 if (flags
& MSG_FASTOPEN
) {
1113 err
= tcp_sendmsg_fastopen(sk
, msg
, &copied_syn
, size
);
1114 if (err
== -EINPROGRESS
&& copied_syn
> 0)
1120 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
1122 tcp_rate_check_app_limited(sk
); /* is sending application-limited? */
1124 /* Wait for a connection to finish. One exception is TCP Fast Open
1125 * (passive side) where data is allowed to be sent before a connection
1126 * is fully established.
1128 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
1129 !tcp_passive_fastopen(sk
)) {
1130 err
= sk_stream_wait_connect(sk
, &timeo
);
1135 if (unlikely(tp
->repair
)) {
1136 if (tp
->repair_queue
== TCP_RECV_QUEUE
) {
1137 copied
= tcp_send_rcvq(sk
, msg
, size
);
1142 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1145 /* 'common' sending to sendq */
1148 sockc
.tsflags
= sk
->sk_tsflags
;
1149 if (msg
->msg_controllen
) {
1150 err
= sock_cmsg_send(sk
, msg
, &sockc
);
1151 if (unlikely(err
)) {
1157 /* This should be in poll */
1158 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
1160 /* Ok commence sending. */
1164 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1167 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
1170 sg
= !!(sk
->sk_route_caps
& NETIF_F_SG
);
1172 while (msg_data_left(msg
)) {
1174 int max
= size_goal
;
1176 skb
= tcp_write_queue_tail(sk
);
1177 if (tcp_send_head(sk
)) {
1178 if (skb
->ip_summed
== CHECKSUM_NONE
)
1180 copy
= max
- skb
->len
;
1183 if (copy
<= 0 || !tcp_skb_can_collapse_to(skb
)) {
1187 /* Allocate new segment. If the interface is SG,
1188 * allocate skb fitting to single page.
1190 if (!sk_stream_memory_free(sk
))
1191 goto wait_for_sndbuf
;
1193 if (process_backlog
&& sk_flush_backlog(sk
)) {
1194 process_backlog
= false;
1197 first_skb
= skb_queue_empty(&sk
->sk_write_queue
);
1198 skb
= sk_stream_alloc_skb(sk
,
1199 select_size(sk
, sg
, first_skb
),
1203 goto wait_for_memory
;
1205 process_backlog
= true;
1207 * Check whether we can use HW checksum.
1209 if (sk_check_csum_caps(sk
))
1210 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1212 skb_entail(sk
, skb
);
1216 /* All packets are restored as if they have
1217 * already been sent. skb_mstamp isn't set to
1218 * avoid wrong rtt estimation.
1221 TCP_SKB_CB(skb
)->sacked
|= TCPCB_REPAIRED
;
1224 /* Try to append data to the end of skb. */
1225 if (copy
> msg_data_left(msg
))
1226 copy
= msg_data_left(msg
);
1228 /* Where to copy to? */
1229 if (skb_availroom(skb
) > 0) {
1230 /* We have some space in skb head. Superb! */
1231 copy
= min_t(int, copy
, skb_availroom(skb
));
1232 err
= skb_add_data_nocache(sk
, skb
, &msg
->msg_iter
, copy
);
1237 int i
= skb_shinfo(skb
)->nr_frags
;
1238 struct page_frag
*pfrag
= sk_page_frag(sk
);
1240 if (!sk_page_frag_refill(sk
, pfrag
))
1241 goto wait_for_memory
;
1243 if (!skb_can_coalesce(skb
, i
, pfrag
->page
,
1245 if (i
>= sysctl_max_skb_frags
|| !sg
) {
1246 tcp_mark_push(tp
, skb
);
1252 copy
= min_t(int, copy
, pfrag
->size
- pfrag
->offset
);
1254 if (!sk_wmem_schedule(sk
, copy
))
1255 goto wait_for_memory
;
1257 err
= skb_copy_to_page_nocache(sk
, &msg
->msg_iter
, skb
,
1264 /* Update the skb. */
1266 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
1268 skb_fill_page_desc(skb
, i
, pfrag
->page
,
1269 pfrag
->offset
, copy
);
1270 get_page(pfrag
->page
);
1272 pfrag
->offset
+= copy
;
1276 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
1278 tp
->write_seq
+= copy
;
1279 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1280 tcp_skb_pcount_set(skb
, 0);
1283 if (!msg_data_left(msg
)) {
1284 tcp_tx_timestamp(sk
, sockc
.tsflags
, skb
);
1285 if (unlikely(flags
& MSG_EOR
))
1286 TCP_SKB_CB(skb
)->eor
= 1;
1290 if (skb
->len
< max
|| (flags
& MSG_OOB
) || unlikely(tp
->repair
))
1293 if (forced_push(tp
)) {
1294 tcp_mark_push(tp
, skb
);
1295 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1296 } else if (skb
== tcp_send_head(sk
))
1297 tcp_push_one(sk
, mss_now
);
1301 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1304 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
,
1305 TCP_NAGLE_PUSH
, size_goal
);
1307 err
= sk_stream_wait_memory(sk
, &timeo
);
1311 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1316 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
, size_goal
);
1319 return copied
+ copied_syn
;
1323 tcp_unlink_write_queue(skb
, sk
);
1324 /* It is the one place in all of TCP, except connection
1325 * reset, where we can be unlinking the send_head.
1327 tcp_check_send_head(sk
, skb
);
1328 sk_wmem_free_skb(sk
, skb
);
1332 if (copied
+ copied_syn
)
1335 err
= sk_stream_error(sk
, flags
, err
);
1336 /* make sure we wake any epoll edge trigger waiter */
1337 if (unlikely(skb_queue_len(&sk
->sk_write_queue
) == 0 &&
1339 sk
->sk_write_space(sk
);
1340 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1345 EXPORT_SYMBOL(tcp_sendmsg
);
1348 * Handle reading urgent data. BSD has very simple semantics for
1349 * this, no blocking and very strange errors 8)
1352 static int tcp_recv_urg(struct sock
*sk
, struct msghdr
*msg
, int len
, int flags
)
1354 struct tcp_sock
*tp
= tcp_sk(sk
);
1356 /* No URG data to read. */
1357 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1358 tp
->urg_data
== TCP_URG_READ
)
1359 return -EINVAL
; /* Yes this is right ! */
1361 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1364 if (tp
->urg_data
& TCP_URG_VALID
) {
1366 char c
= tp
->urg_data
;
1368 if (!(flags
& MSG_PEEK
))
1369 tp
->urg_data
= TCP_URG_READ
;
1371 /* Read urgent data. */
1372 msg
->msg_flags
|= MSG_OOB
;
1375 if (!(flags
& MSG_TRUNC
))
1376 err
= memcpy_to_msg(msg
, &c
, 1);
1379 msg
->msg_flags
|= MSG_TRUNC
;
1381 return err
? -EFAULT
: len
;
1384 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1387 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1388 * the available implementations agree in this case:
1389 * this call should never block, independent of the
1390 * blocking state of the socket.
1391 * Mike <pall@rz.uni-karlsruhe.de>
1396 static int tcp_peek_sndq(struct sock
*sk
, struct msghdr
*msg
, int len
)
1398 struct sk_buff
*skb
;
1399 int copied
= 0, err
= 0;
1401 /* XXX -- need to support SO_PEEK_OFF */
1403 skb_queue_walk(&sk
->sk_write_queue
, skb
) {
1404 err
= skb_copy_datagram_msg(skb
, 0, msg
, skb
->len
);
1411 return err
?: copied
;
1414 /* Clean up the receive buffer for full frames taken by the user,
1415 * then send an ACK if necessary. COPIED is the number of bytes
1416 * tcp_recvmsg has given to the user so far, it speeds up the
1417 * calculation of whether or not we must ACK for the sake of
1420 static void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1422 struct tcp_sock
*tp
= tcp_sk(sk
);
1423 bool time_to_ack
= false;
1425 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1427 WARN(skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
),
1428 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1429 tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
);
1431 if (inet_csk_ack_scheduled(sk
)) {
1432 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1433 /* Delayed ACKs frequently hit locked sockets during bulk
1435 if (icsk
->icsk_ack
.blocked
||
1436 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1437 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1439 * If this read emptied read buffer, we send ACK, if
1440 * connection is not bidirectional, user drained
1441 * receive buffer and there was a small segment
1445 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1446 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1447 !icsk
->icsk_ack
.pingpong
)) &&
1448 !atomic_read(&sk
->sk_rmem_alloc
)))
1452 /* We send an ACK if we can now advertise a non-zero window
1453 * which has been raised "significantly".
1455 * Even if window raised up to infinity, do not send window open ACK
1456 * in states, where we will not receive more. It is useless.
1458 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1459 __u32 rcv_window_now
= tcp_receive_window(tp
);
1461 /* Optimize, __tcp_select_window() is not cheap. */
1462 if (2*rcv_window_now
<= tp
->window_clamp
) {
1463 __u32 new_window
= __tcp_select_window(sk
);
1465 /* Send ACK now, if this read freed lots of space
1466 * in our buffer. Certainly, new_window is new window.
1467 * We can advertise it now, if it is not less than current one.
1468 * "Lots" means "at least twice" here.
1470 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1478 static void tcp_prequeue_process(struct sock
*sk
)
1480 struct sk_buff
*skb
;
1481 struct tcp_sock
*tp
= tcp_sk(sk
);
1483 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPPREQUEUED
);
1485 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
1486 sk_backlog_rcv(sk
, skb
);
1488 /* Clear memory counter. */
1489 tp
->ucopy
.memory
= 0;
1492 static struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1494 struct sk_buff
*skb
;
1497 while ((skb
= skb_peek(&sk
->sk_receive_queue
)) != NULL
) {
1498 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1499 if (unlikely(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
1500 pr_err_once("%s: found a SYN, please report !\n", __func__
);
1503 if (offset
< skb
->len
|| (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)) {
1507 /* This looks weird, but this can happen if TCP collapsing
1508 * splitted a fat GRO packet, while we released socket lock
1509 * in skb_splice_bits()
1511 sk_eat_skb(sk
, skb
);
1517 * This routine provides an alternative to tcp_recvmsg() for routines
1518 * that would like to handle copying from skbuffs directly in 'sendfile'
1521 * - It is assumed that the socket was locked by the caller.
1522 * - The routine does not block.
1523 * - At present, there is no support for reading OOB data
1524 * or for 'peeking' the socket using this routine
1525 * (although both would be easy to implement).
1527 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1528 sk_read_actor_t recv_actor
)
1530 struct sk_buff
*skb
;
1531 struct tcp_sock
*tp
= tcp_sk(sk
);
1532 u32 seq
= tp
->copied_seq
;
1536 if (sk
->sk_state
== TCP_LISTEN
)
1538 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1539 if (offset
< skb
->len
) {
1543 len
= skb
->len
- offset
;
1544 /* Stop reading if we hit a patch of urgent data */
1546 u32 urg_offset
= tp
->urg_seq
- seq
;
1547 if (urg_offset
< len
)
1552 used
= recv_actor(desc
, skb
, offset
, len
);
1557 } else if (used
<= len
) {
1562 /* If recv_actor drops the lock (e.g. TCP splice
1563 * receive) the skb pointer might be invalid when
1564 * getting here: tcp_collapse might have deleted it
1565 * while aggregating skbs from the socket queue.
1567 skb
= tcp_recv_skb(sk
, seq
- 1, &offset
);
1570 /* TCP coalescing might have appended data to the skb.
1571 * Try to splice more frags
1573 if (offset
+ 1 != skb
->len
)
1576 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) {
1577 sk_eat_skb(sk
, skb
);
1581 sk_eat_skb(sk
, skb
);
1584 tp
->copied_seq
= seq
;
1586 tp
->copied_seq
= seq
;
1588 tcp_rcv_space_adjust(sk
);
1590 /* Clean up data we have read: This will do ACK frames. */
1592 tcp_recv_skb(sk
, seq
, &offset
);
1593 tcp_cleanup_rbuf(sk
, copied
);
1597 EXPORT_SYMBOL(tcp_read_sock
);
1599 int tcp_peek_len(struct socket
*sock
)
1601 return tcp_inq(sock
->sk
);
1603 EXPORT_SYMBOL(tcp_peek_len
);
1606 * This routine copies from a sock struct into the user buffer.
1608 * Technical note: in 2.3 we work on _locked_ socket, so that
1609 * tricks with *seq access order and skb->users are not required.
1610 * Probably, code can be easily improved even more.
1613 int tcp_recvmsg(struct sock
*sk
, struct msghdr
*msg
, size_t len
, int nonblock
,
1614 int flags
, int *addr_len
)
1616 struct tcp_sock
*tp
= tcp_sk(sk
);
1622 int target
; /* Read at least this many bytes */
1624 struct task_struct
*user_recv
= NULL
;
1625 struct sk_buff
*skb
, *last
;
1628 if (unlikely(flags
& MSG_ERRQUEUE
))
1629 return inet_recv_error(sk
, msg
, len
, addr_len
);
1631 if (sk_can_busy_loop(sk
) && skb_queue_empty(&sk
->sk_receive_queue
) &&
1632 (sk
->sk_state
== TCP_ESTABLISHED
))
1633 sk_busy_loop(sk
, nonblock
);
1638 if (sk
->sk_state
== TCP_LISTEN
)
1641 timeo
= sock_rcvtimeo(sk
, nonblock
);
1643 /* Urgent data needs to be handled specially. */
1644 if (flags
& MSG_OOB
)
1647 if (unlikely(tp
->repair
)) {
1649 if (!(flags
& MSG_PEEK
))
1652 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
1656 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1659 /* 'common' recv queue MSG_PEEK-ing */
1662 seq
= &tp
->copied_seq
;
1663 if (flags
& MSG_PEEK
) {
1664 peek_seq
= tp
->copied_seq
;
1668 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1673 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1674 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1677 if (signal_pending(current
)) {
1678 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1683 /* Next get a buffer. */
1685 last
= skb_peek_tail(&sk
->sk_receive_queue
);
1686 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1688 /* Now that we have two receive queues this
1691 if (WARN(before(*seq
, TCP_SKB_CB(skb
)->seq
),
1692 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1693 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
,
1697 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1698 if (unlikely(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
1699 pr_err_once("%s: found a SYN, please report !\n", __func__
);
1702 if (offset
< skb
->len
)
1704 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1706 WARN(!(flags
& MSG_PEEK
),
1707 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1708 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
, flags
);
1711 /* Well, if we have backlog, try to process it now yet. */
1713 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1718 sk
->sk_state
== TCP_CLOSE
||
1719 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1721 signal_pending(current
))
1724 if (sock_flag(sk
, SOCK_DONE
))
1728 copied
= sock_error(sk
);
1732 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1735 if (sk
->sk_state
== TCP_CLOSE
) {
1736 if (!sock_flag(sk
, SOCK_DONE
)) {
1737 /* This occurs when user tries to read
1738 * from never connected socket.
1751 if (signal_pending(current
)) {
1752 copied
= sock_intr_errno(timeo
);
1757 tcp_cleanup_rbuf(sk
, copied
);
1759 if (!sysctl_tcp_low_latency
&& tp
->ucopy
.task
== user_recv
) {
1760 /* Install new reader */
1761 if (!user_recv
&& !(flags
& (MSG_TRUNC
| MSG_PEEK
))) {
1762 user_recv
= current
;
1763 tp
->ucopy
.task
= user_recv
;
1764 tp
->ucopy
.msg
= msg
;
1767 tp
->ucopy
.len
= len
;
1769 WARN_ON(tp
->copied_seq
!= tp
->rcv_nxt
&&
1770 !(flags
& (MSG_PEEK
| MSG_TRUNC
)));
1772 /* Ugly... If prequeue is not empty, we have to
1773 * process it before releasing socket, otherwise
1774 * order will be broken at second iteration.
1775 * More elegant solution is required!!!
1777 * Look: we have the following (pseudo)queues:
1779 * 1. packets in flight
1784 * Each queue can be processed only if the next ones
1785 * are empty. At this point we have empty receive_queue.
1786 * But prequeue _can_ be not empty after 2nd iteration,
1787 * when we jumped to start of loop because backlog
1788 * processing added something to receive_queue.
1789 * We cannot release_sock(), because backlog contains
1790 * packets arrived _after_ prequeued ones.
1792 * Shortly, algorithm is clear --- to process all
1793 * the queues in order. We could make it more directly,
1794 * requeueing packets from backlog to prequeue, if
1795 * is not empty. It is more elegant, but eats cycles,
1798 if (!skb_queue_empty(&tp
->ucopy
.prequeue
))
1801 /* __ Set realtime policy in scheduler __ */
1804 if (copied
>= target
) {
1805 /* Do not sleep, just process backlog. */
1809 sk_wait_data(sk
, &timeo
, last
);
1815 /* __ Restore normal policy in scheduler __ */
1817 chunk
= len
- tp
->ucopy
.len
;
1819 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG
, chunk
);
1824 if (tp
->rcv_nxt
== tp
->copied_seq
&&
1825 !skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1827 tcp_prequeue_process(sk
);
1829 chunk
= len
- tp
->ucopy
.len
;
1831 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1837 if ((flags
& MSG_PEEK
) &&
1838 (peek_seq
- copied
- urg_hole
!= tp
->copied_seq
)) {
1839 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1841 task_pid_nr(current
));
1842 peek_seq
= tp
->copied_seq
;
1847 /* Ok so how much can we use? */
1848 used
= skb
->len
- offset
;
1852 /* Do we have urgent data here? */
1854 u32 urg_offset
= tp
->urg_seq
- *seq
;
1855 if (urg_offset
< used
) {
1857 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1870 if (!(flags
& MSG_TRUNC
)) {
1871 err
= skb_copy_datagram_msg(skb
, offset
, msg
, used
);
1873 /* Exception. Bailout! */
1884 tcp_rcv_space_adjust(sk
);
1887 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1889 tcp_fast_path_check(sk
);
1891 if (used
+ offset
< skb
->len
)
1894 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1896 if (!(flags
& MSG_PEEK
))
1897 sk_eat_skb(sk
, skb
);
1901 /* Process the FIN. */
1903 if (!(flags
& MSG_PEEK
))
1904 sk_eat_skb(sk
, skb
);
1909 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1912 tp
->ucopy
.len
= copied
> 0 ? len
: 0;
1914 tcp_prequeue_process(sk
);
1916 if (copied
> 0 && (chunk
= len
- tp
->ucopy
.len
) != 0) {
1917 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1923 tp
->ucopy
.task
= NULL
;
1927 /* According to UNIX98, msg_name/msg_namelen are ignored
1928 * on connected socket. I was just happy when found this 8) --ANK
1931 /* Clean up data we have read: This will do ACK frames. */
1932 tcp_cleanup_rbuf(sk
, copied
);
1942 err
= tcp_recv_urg(sk
, msg
, len
, flags
);
1946 err
= tcp_peek_sndq(sk
, msg
, len
);
1949 EXPORT_SYMBOL(tcp_recvmsg
);
1951 void tcp_set_state(struct sock
*sk
, int state
)
1953 int oldstate
= sk
->sk_state
;
1956 case TCP_ESTABLISHED
:
1957 if (oldstate
!= TCP_ESTABLISHED
)
1958 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1962 if (oldstate
== TCP_CLOSE_WAIT
|| oldstate
== TCP_ESTABLISHED
)
1963 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ESTABRESETS
);
1965 sk
->sk_prot
->unhash(sk
);
1966 if (inet_csk(sk
)->icsk_bind_hash
&&
1967 !(sk
->sk_userlocks
& SOCK_BINDPORT_LOCK
))
1971 if (oldstate
== TCP_ESTABLISHED
)
1972 TCP_DEC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1975 /* Change state AFTER socket is unhashed to avoid closed
1976 * socket sitting in hash tables.
1978 sk_state_store(sk
, state
);
1981 SOCK_DEBUG(sk
, "TCP sk=%p, State %s -> %s\n", sk
, statename
[oldstate
], statename
[state
]);
1984 EXPORT_SYMBOL_GPL(tcp_set_state
);
1987 * State processing on a close. This implements the state shift for
1988 * sending our FIN frame. Note that we only send a FIN for some
1989 * states. A shutdown() may have already sent the FIN, or we may be
1993 static const unsigned char new_state
[16] = {
1994 /* current state: new state: action: */
1995 [0 /* (Invalid) */] = TCP_CLOSE
,
1996 [TCP_ESTABLISHED
] = TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1997 [TCP_SYN_SENT
] = TCP_CLOSE
,
1998 [TCP_SYN_RECV
] = TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1999 [TCP_FIN_WAIT1
] = TCP_FIN_WAIT1
,
2000 [TCP_FIN_WAIT2
] = TCP_FIN_WAIT2
,
2001 [TCP_TIME_WAIT
] = TCP_CLOSE
,
2002 [TCP_CLOSE
] = TCP_CLOSE
,
2003 [TCP_CLOSE_WAIT
] = TCP_LAST_ACK
| TCP_ACTION_FIN
,
2004 [TCP_LAST_ACK
] = TCP_LAST_ACK
,
2005 [TCP_LISTEN
] = TCP_CLOSE
,
2006 [TCP_CLOSING
] = TCP_CLOSING
,
2007 [TCP_NEW_SYN_RECV
] = TCP_CLOSE
, /* should not happen ! */
2010 static int tcp_close_state(struct sock
*sk
)
2012 int next
= (int)new_state
[sk
->sk_state
];
2013 int ns
= next
& TCP_STATE_MASK
;
2015 tcp_set_state(sk
, ns
);
2017 return next
& TCP_ACTION_FIN
;
2021 * Shutdown the sending side of a connection. Much like close except
2022 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2025 void tcp_shutdown(struct sock
*sk
, int how
)
2027 /* We need to grab some memory, and put together a FIN,
2028 * and then put it into the queue to be sent.
2029 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2031 if (!(how
& SEND_SHUTDOWN
))
2034 /* If we've already sent a FIN, or it's a closed state, skip this. */
2035 if ((1 << sk
->sk_state
) &
2036 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
2037 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
2038 /* Clear out any half completed packets. FIN if needed. */
2039 if (tcp_close_state(sk
))
2043 EXPORT_SYMBOL(tcp_shutdown
);
2045 bool tcp_check_oom(struct sock
*sk
, int shift
)
2047 bool too_many_orphans
, out_of_socket_memory
;
2049 too_many_orphans
= tcp_too_many_orphans(sk
, shift
);
2050 out_of_socket_memory
= tcp_out_of_memory(sk
);
2052 if (too_many_orphans
)
2053 net_info_ratelimited("too many orphaned sockets\n");
2054 if (out_of_socket_memory
)
2055 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2056 return too_many_orphans
|| out_of_socket_memory
;
2059 void tcp_close(struct sock
*sk
, long timeout
)
2061 struct sk_buff
*skb
;
2062 int data_was_unread
= 0;
2066 sk
->sk_shutdown
= SHUTDOWN_MASK
;
2068 if (sk
->sk_state
== TCP_LISTEN
) {
2069 tcp_set_state(sk
, TCP_CLOSE
);
2072 inet_csk_listen_stop(sk
);
2074 goto adjudge_to_death
;
2077 /* We need to flush the recv. buffs. We do this only on the
2078 * descriptor close, not protocol-sourced closes, because the
2079 * reader process may not have drained the data yet!
2081 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
2082 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
;
2084 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
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(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(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 1 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(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(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_rbtree_purge(&tp
->out_of_order_queue
);
2274 inet
->inet_dport
= 0;
2276 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
2277 inet_reset_saddr(sk
);
2279 sk
->sk_shutdown
= 0;
2280 sock_reset_flag(sk
, SOCK_DONE
);
2282 tp
->write_seq
+= tp
->max_window
+ 2;
2283 if (tp
->write_seq
== 0)
2285 icsk
->icsk_backoff
= 0;
2287 icsk
->icsk_probes_out
= 0;
2288 tp
->packets_out
= 0;
2289 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
2290 tp
->snd_cwnd_cnt
= 0;
2291 tp
->window_clamp
= 0;
2292 tcp_set_ca_state(sk
, TCP_CA_Open
);
2293 tcp_clear_retrans(tp
);
2294 inet_csk_delack_init(sk
);
2295 tcp_init_send_head(sk
);
2296 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
2299 WARN_ON(inet
->inet_num
&& !icsk
->icsk_bind_hash
);
2301 sk
->sk_error_report(sk
);
2304 EXPORT_SYMBOL(tcp_disconnect
);
2306 static inline bool tcp_can_repair_sock(const struct sock
*sk
)
2308 return ns_capable(sock_net(sk
)->user_ns
, CAP_NET_ADMIN
) &&
2309 (sk
->sk_state
!= TCP_LISTEN
);
2312 static int tcp_repair_set_window(struct tcp_sock
*tp
, char __user
*optbuf
, int len
)
2314 struct tcp_repair_window opt
;
2319 if (len
!= sizeof(opt
))
2322 if (copy_from_user(&opt
, optbuf
, sizeof(opt
)))
2325 if (opt
.max_window
< opt
.snd_wnd
)
2328 if (after(opt
.snd_wl1
, tp
->rcv_nxt
+ opt
.rcv_wnd
))
2331 if (after(opt
.rcv_wup
, tp
->rcv_nxt
))
2334 tp
->snd_wl1
= opt
.snd_wl1
;
2335 tp
->snd_wnd
= opt
.snd_wnd
;
2336 tp
->max_window
= opt
.max_window
;
2338 tp
->rcv_wnd
= opt
.rcv_wnd
;
2339 tp
->rcv_wup
= opt
.rcv_wup
;
2344 static int tcp_repair_options_est(struct tcp_sock
*tp
,
2345 struct tcp_repair_opt __user
*optbuf
, unsigned int len
)
2347 struct tcp_repair_opt opt
;
2349 while (len
>= sizeof(opt
)) {
2350 if (copy_from_user(&opt
, optbuf
, sizeof(opt
)))
2356 switch (opt
.opt_code
) {
2358 tp
->rx_opt
.mss_clamp
= opt
.opt_val
;
2362 u16 snd_wscale
= opt
.opt_val
& 0xFFFF;
2363 u16 rcv_wscale
= opt
.opt_val
>> 16;
2365 if (snd_wscale
> 14 || rcv_wscale
> 14)
2368 tp
->rx_opt
.snd_wscale
= snd_wscale
;
2369 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2370 tp
->rx_opt
.wscale_ok
= 1;
2373 case TCPOPT_SACK_PERM
:
2374 if (opt
.opt_val
!= 0)
2377 tp
->rx_opt
.sack_ok
|= TCP_SACK_SEEN
;
2378 if (sysctl_tcp_fack
)
2379 tcp_enable_fack(tp
);
2381 case TCPOPT_TIMESTAMP
:
2382 if (opt
.opt_val
!= 0)
2385 tp
->rx_opt
.tstamp_ok
= 1;
2394 * Socket option code for TCP.
2396 static int do_tcp_setsockopt(struct sock
*sk
, int level
,
2397 int optname
, char __user
*optval
, unsigned int optlen
)
2399 struct tcp_sock
*tp
= tcp_sk(sk
);
2400 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2401 struct net
*net
= sock_net(sk
);
2405 /* These are data/string values, all the others are ints */
2407 case TCP_CONGESTION
: {
2408 char name
[TCP_CA_NAME_MAX
];
2413 val
= strncpy_from_user(name
, optval
,
2414 min_t(long, TCP_CA_NAME_MAX
-1, optlen
));
2420 err
= tcp_set_congestion_control(sk
, name
);
2429 if (optlen
< sizeof(int))
2432 if (get_user(val
, (int __user
*)optval
))
2439 /* Values greater than interface MTU won't take effect. However
2440 * at the point when this call is done we typically don't yet
2441 * know which interface is going to be used */
2442 if (val
< TCP_MIN_MSS
|| val
> MAX_TCP_WINDOW
) {
2446 tp
->rx_opt
.user_mss
= val
;
2451 /* TCP_NODELAY is weaker than TCP_CORK, so that
2452 * this option on corked socket is remembered, but
2453 * it is not activated until cork is cleared.
2455 * However, when TCP_NODELAY is set we make
2456 * an explicit push, which overrides even TCP_CORK
2457 * for currently queued segments.
2459 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
2460 tcp_push_pending_frames(sk
);
2462 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
2466 case TCP_THIN_LINEAR_TIMEOUTS
:
2467 if (val
< 0 || val
> 1)
2473 case TCP_THIN_DUPACK
:
2474 if (val
< 0 || val
> 1)
2477 tp
->thin_dupack
= val
;
2478 if (tp
->thin_dupack
)
2479 tcp_disable_early_retrans(tp
);
2484 if (!tcp_can_repair_sock(sk
))
2486 else if (val
== 1) {
2488 sk
->sk_reuse
= SK_FORCE_REUSE
;
2489 tp
->repair_queue
= TCP_NO_QUEUE
;
2490 } else if (val
== 0) {
2492 sk
->sk_reuse
= SK_NO_REUSE
;
2493 tcp_send_window_probe(sk
);
2499 case TCP_REPAIR_QUEUE
:
2502 else if (val
< TCP_QUEUES_NR
)
2503 tp
->repair_queue
= val
;
2509 if (sk
->sk_state
!= TCP_CLOSE
)
2511 else if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2512 tp
->write_seq
= val
;
2513 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2519 case TCP_REPAIR_OPTIONS
:
2522 else if (sk
->sk_state
== TCP_ESTABLISHED
)
2523 err
= tcp_repair_options_est(tp
,
2524 (struct tcp_repair_opt __user
*)optval
,
2531 /* When set indicates to always queue non-full frames.
2532 * Later the user clears this option and we transmit
2533 * any pending partial frames in the queue. This is
2534 * meant to be used alongside sendfile() to get properly
2535 * filled frames when the user (for example) must write
2536 * out headers with a write() call first and then use
2537 * sendfile to send out the data parts.
2539 * TCP_CORK can be set together with TCP_NODELAY and it is
2540 * stronger than TCP_NODELAY.
2543 tp
->nonagle
|= TCP_NAGLE_CORK
;
2545 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
2546 if (tp
->nonagle
&TCP_NAGLE_OFF
)
2547 tp
->nonagle
|= TCP_NAGLE_PUSH
;
2548 tcp_push_pending_frames(sk
);
2553 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
2556 tp
->keepalive_time
= val
* HZ
;
2557 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
2558 !((1 << sk
->sk_state
) &
2559 (TCPF_CLOSE
| TCPF_LISTEN
))) {
2560 u32 elapsed
= keepalive_time_elapsed(tp
);
2561 if (tp
->keepalive_time
> elapsed
)
2562 elapsed
= tp
->keepalive_time
- elapsed
;
2565 inet_csk_reset_keepalive_timer(sk
, elapsed
);
2570 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
2573 tp
->keepalive_intvl
= val
* HZ
;
2576 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
2579 tp
->keepalive_probes
= val
;
2582 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
2585 icsk
->icsk_syn_retries
= val
;
2589 if (val
< 0 || val
> 1)
2598 else if (val
> net
->ipv4
.sysctl_tcp_fin_timeout
/ HZ
)
2601 tp
->linger2
= val
* HZ
;
2604 case TCP_DEFER_ACCEPT
:
2605 /* Translate value in seconds to number of retransmits */
2606 icsk
->icsk_accept_queue
.rskq_defer_accept
=
2607 secs_to_retrans(val
, TCP_TIMEOUT_INIT
/ HZ
,
2611 case TCP_WINDOW_CLAMP
:
2613 if (sk
->sk_state
!= TCP_CLOSE
) {
2617 tp
->window_clamp
= 0;
2619 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
2620 SOCK_MIN_RCVBUF
/ 2 : val
;
2625 icsk
->icsk_ack
.pingpong
= 1;
2627 icsk
->icsk_ack
.pingpong
= 0;
2628 if ((1 << sk
->sk_state
) &
2629 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
2630 inet_csk_ack_scheduled(sk
)) {
2631 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
2632 tcp_cleanup_rbuf(sk
, 1);
2634 icsk
->icsk_ack
.pingpong
= 1;
2639 #ifdef CONFIG_TCP_MD5SIG
2641 /* Read the IP->Key mappings from userspace */
2642 err
= tp
->af_specific
->md5_parse(sk
, optval
, optlen
);
2645 case TCP_USER_TIMEOUT
:
2646 /* Cap the max time in ms TCP will retry or probe the window
2647 * before giving up and aborting (ETIMEDOUT) a connection.
2652 icsk
->icsk_user_timeout
= msecs_to_jiffies(val
);
2656 if (val
>= 0 && ((1 << sk
->sk_state
) & (TCPF_CLOSE
|
2658 tcp_fastopen_init_key_once(true);
2660 fastopen_queue_tune(sk
, val
);
2669 tp
->tsoffset
= val
- tcp_time_stamp
;
2671 case TCP_REPAIR_WINDOW
:
2672 err
= tcp_repair_set_window(tp
, optval
, optlen
);
2674 case TCP_NOTSENT_LOWAT
:
2675 tp
->notsent_lowat
= val
;
2676 sk
->sk_write_space(sk
);
2687 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2688 unsigned int optlen
)
2690 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2692 if (level
!= SOL_TCP
)
2693 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
2695 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2697 EXPORT_SYMBOL(tcp_setsockopt
);
2699 #ifdef CONFIG_COMPAT
2700 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
2701 char __user
*optval
, unsigned int optlen
)
2703 if (level
!= SOL_TCP
)
2704 return inet_csk_compat_setsockopt(sk
, level
, optname
,
2706 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2708 EXPORT_SYMBOL(compat_tcp_setsockopt
);
2711 static void tcp_get_info_chrono_stats(const struct tcp_sock
*tp
,
2712 struct tcp_info
*info
)
2714 u64 stats
[__TCP_CHRONO_MAX
], total
= 0;
2717 for (i
= TCP_CHRONO_BUSY
; i
< __TCP_CHRONO_MAX
; ++i
) {
2718 stats
[i
] = tp
->chrono_stat
[i
- 1];
2719 if (i
== tp
->chrono_type
)
2720 stats
[i
] += tcp_time_stamp
- tp
->chrono_start
;
2721 stats
[i
] *= USEC_PER_SEC
/ HZ
;
2725 info
->tcpi_busy_time
= total
;
2726 info
->tcpi_rwnd_limited
= stats
[TCP_CHRONO_RWND_LIMITED
];
2727 info
->tcpi_sndbuf_limited
= stats
[TCP_CHRONO_SNDBUF_LIMITED
];
2730 /* Return information about state of tcp endpoint in API format. */
2731 void tcp_get_info(struct sock
*sk
, struct tcp_info
*info
)
2733 const struct tcp_sock
*tp
= tcp_sk(sk
); /* iff sk_type == SOCK_STREAM */
2734 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2735 u32 now
= tcp_time_stamp
, intv
;
2740 memset(info
, 0, sizeof(*info
));
2741 if (sk
->sk_type
!= SOCK_STREAM
)
2744 info
->tcpi_state
= sk_state_load(sk
);
2746 /* Report meaningful fields for all TCP states, including listeners */
2747 rate
= READ_ONCE(sk
->sk_pacing_rate
);
2748 rate64
= rate
!= ~0U ? rate
: ~0ULL;
2749 info
->tcpi_pacing_rate
= rate64
;
2751 rate
= READ_ONCE(sk
->sk_max_pacing_rate
);
2752 rate64
= rate
!= ~0U ? rate
: ~0ULL;
2753 info
->tcpi_max_pacing_rate
= rate64
;
2755 info
->tcpi_reordering
= tp
->reordering
;
2756 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
2758 if (info
->tcpi_state
== TCP_LISTEN
) {
2759 /* listeners aliased fields :
2760 * tcpi_unacked -> Number of children ready for accept()
2761 * tcpi_sacked -> max backlog
2763 info
->tcpi_unacked
= sk
->sk_ack_backlog
;
2764 info
->tcpi_sacked
= sk
->sk_max_ack_backlog
;
2767 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
2768 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
2769 info
->tcpi_probes
= icsk
->icsk_probes_out
;
2770 info
->tcpi_backoff
= icsk
->icsk_backoff
;
2772 if (tp
->rx_opt
.tstamp_ok
)
2773 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
2774 if (tcp_is_sack(tp
))
2775 info
->tcpi_options
|= TCPI_OPT_SACK
;
2776 if (tp
->rx_opt
.wscale_ok
) {
2777 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
2778 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
2779 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
2782 if (tp
->ecn_flags
& TCP_ECN_OK
)
2783 info
->tcpi_options
|= TCPI_OPT_ECN
;
2784 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
2785 info
->tcpi_options
|= TCPI_OPT_ECN_SEEN
;
2786 if (tp
->syn_data_acked
)
2787 info
->tcpi_options
|= TCPI_OPT_SYN_DATA
;
2789 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
2790 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
2791 info
->tcpi_snd_mss
= tp
->mss_cache
;
2792 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
2794 info
->tcpi_unacked
= tp
->packets_out
;
2795 info
->tcpi_sacked
= tp
->sacked_out
;
2797 info
->tcpi_lost
= tp
->lost_out
;
2798 info
->tcpi_retrans
= tp
->retrans_out
;
2799 info
->tcpi_fackets
= tp
->fackets_out
;
2801 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
2802 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
2803 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
2805 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
2806 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
2807 info
->tcpi_rtt
= tp
->srtt_us
>> 3;
2808 info
->tcpi_rttvar
= tp
->mdev_us
>> 2;
2809 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
2810 info
->tcpi_advmss
= tp
->advmss
;
2812 info
->tcpi_rcv_rtt
= jiffies_to_usecs(tp
->rcv_rtt_est
.rtt
)>>3;
2813 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
2815 info
->tcpi_total_retrans
= tp
->total_retrans
;
2817 slow
= lock_sock_fast(sk
);
2819 info
->tcpi_bytes_acked
= tp
->bytes_acked
;
2820 info
->tcpi_bytes_received
= tp
->bytes_received
;
2821 info
->tcpi_notsent_bytes
= max_t(int, 0, tp
->write_seq
- tp
->snd_nxt
);
2822 tcp_get_info_chrono_stats(tp
, info
);
2824 unlock_sock_fast(sk
, slow
);
2826 info
->tcpi_segs_out
= tp
->segs_out
;
2827 info
->tcpi_segs_in
= tp
->segs_in
;
2829 info
->tcpi_min_rtt
= tcp_min_rtt(tp
);
2830 info
->tcpi_data_segs_in
= tp
->data_segs_in
;
2831 info
->tcpi_data_segs_out
= tp
->data_segs_out
;
2833 info
->tcpi_delivery_rate_app_limited
= tp
->rate_app_limited
? 1 : 0;
2834 rate
= READ_ONCE(tp
->rate_delivered
);
2835 intv
= READ_ONCE(tp
->rate_interval_us
);
2837 rate64
= (u64
)rate
* tp
->mss_cache
* USEC_PER_SEC
;
2838 do_div(rate64
, intv
);
2839 info
->tcpi_delivery_rate
= rate64
;
2842 EXPORT_SYMBOL_GPL(tcp_get_info
);
2844 struct sk_buff
*tcp_get_timestamping_opt_stats(const struct sock
*sk
)
2846 const struct tcp_sock
*tp
= tcp_sk(sk
);
2847 struct sk_buff
*stats
;
2848 struct tcp_info info
;
2850 stats
= alloc_skb(3 * nla_total_size_64bit(sizeof(u64
)), GFP_ATOMIC
);
2854 tcp_get_info_chrono_stats(tp
, &info
);
2855 nla_put_u64_64bit(stats
, TCP_NLA_BUSY
,
2856 info
.tcpi_busy_time
, TCP_NLA_PAD
);
2857 nla_put_u64_64bit(stats
, TCP_NLA_RWND_LIMITED
,
2858 info
.tcpi_rwnd_limited
, TCP_NLA_PAD
);
2859 nla_put_u64_64bit(stats
, TCP_NLA_SNDBUF_LIMITED
,
2860 info
.tcpi_sndbuf_limited
, TCP_NLA_PAD
);
2864 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
2865 int optname
, char __user
*optval
, int __user
*optlen
)
2867 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2868 struct tcp_sock
*tp
= tcp_sk(sk
);
2869 struct net
*net
= sock_net(sk
);
2872 if (get_user(len
, optlen
))
2875 len
= min_t(unsigned int, len
, sizeof(int));
2882 val
= tp
->mss_cache
;
2883 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
2884 val
= tp
->rx_opt
.user_mss
;
2886 val
= tp
->rx_opt
.mss_clamp
;
2889 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
2892 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
2895 val
= keepalive_time_when(tp
) / HZ
;
2898 val
= keepalive_intvl_when(tp
) / HZ
;
2901 val
= keepalive_probes(tp
);
2904 val
= icsk
->icsk_syn_retries
? : net
->ipv4
.sysctl_tcp_syn_retries
;
2909 val
= (val
? : net
->ipv4
.sysctl_tcp_fin_timeout
) / HZ
;
2911 case TCP_DEFER_ACCEPT
:
2912 val
= retrans_to_secs(icsk
->icsk_accept_queue
.rskq_defer_accept
,
2913 TCP_TIMEOUT_INIT
/ HZ
, TCP_RTO_MAX
/ HZ
);
2915 case TCP_WINDOW_CLAMP
:
2916 val
= tp
->window_clamp
;
2919 struct tcp_info info
;
2921 if (get_user(len
, optlen
))
2924 tcp_get_info(sk
, &info
);
2926 len
= min_t(unsigned int, len
, sizeof(info
));
2927 if (put_user(len
, optlen
))
2929 if (copy_to_user(optval
, &info
, len
))
2934 const struct tcp_congestion_ops
*ca_ops
;
2935 union tcp_cc_info info
;
2939 if (get_user(len
, optlen
))
2942 ca_ops
= icsk
->icsk_ca_ops
;
2943 if (ca_ops
&& ca_ops
->get_info
)
2944 sz
= ca_ops
->get_info(sk
, ~0U, &attr
, &info
);
2946 len
= min_t(unsigned int, len
, sz
);
2947 if (put_user(len
, optlen
))
2949 if (copy_to_user(optval
, &info
, len
))
2954 val
= !icsk
->icsk_ack
.pingpong
;
2957 case TCP_CONGESTION
:
2958 if (get_user(len
, optlen
))
2960 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
2961 if (put_user(len
, optlen
))
2963 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
2967 case TCP_THIN_LINEAR_TIMEOUTS
:
2970 case TCP_THIN_DUPACK
:
2971 val
= tp
->thin_dupack
;
2978 case TCP_REPAIR_QUEUE
:
2980 val
= tp
->repair_queue
;
2985 case TCP_REPAIR_WINDOW
: {
2986 struct tcp_repair_window opt
;
2988 if (get_user(len
, optlen
))
2991 if (len
!= sizeof(opt
))
2997 opt
.snd_wl1
= tp
->snd_wl1
;
2998 opt
.snd_wnd
= tp
->snd_wnd
;
2999 opt
.max_window
= tp
->max_window
;
3000 opt
.rcv_wnd
= tp
->rcv_wnd
;
3001 opt
.rcv_wup
= tp
->rcv_wup
;
3003 if (copy_to_user(optval
, &opt
, len
))
3008 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
3009 val
= tp
->write_seq
;
3010 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
3016 case TCP_USER_TIMEOUT
:
3017 val
= jiffies_to_msecs(icsk
->icsk_user_timeout
);
3021 val
= icsk
->icsk_accept_queue
.fastopenq
.max_qlen
;
3025 val
= tcp_time_stamp
+ tp
->tsoffset
;
3027 case TCP_NOTSENT_LOWAT
:
3028 val
= tp
->notsent_lowat
;
3033 case TCP_SAVED_SYN
: {
3034 if (get_user(len
, optlen
))
3038 if (tp
->saved_syn
) {
3039 if (len
< tp
->saved_syn
[0]) {
3040 if (put_user(tp
->saved_syn
[0], optlen
)) {
3047 len
= tp
->saved_syn
[0];
3048 if (put_user(len
, optlen
)) {
3052 if (copy_to_user(optval
, tp
->saved_syn
+ 1, len
)) {
3056 tcp_saved_syn_free(tp
);
3061 if (put_user(len
, optlen
))
3067 return -ENOPROTOOPT
;
3070 if (put_user(len
, optlen
))
3072 if (copy_to_user(optval
, &val
, len
))
3077 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
3080 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3082 if (level
!= SOL_TCP
)
3083 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
3085 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
3087 EXPORT_SYMBOL(tcp_getsockopt
);
3089 #ifdef CONFIG_COMPAT
3090 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
3091 char __user
*optval
, int __user
*optlen
)
3093 if (level
!= SOL_TCP
)
3094 return inet_csk_compat_getsockopt(sk
, level
, optname
,
3096 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
3098 EXPORT_SYMBOL(compat_tcp_getsockopt
);
3101 #ifdef CONFIG_TCP_MD5SIG
3102 static DEFINE_PER_CPU(struct tcp_md5sig_pool
, tcp_md5sig_pool
);
3103 static DEFINE_MUTEX(tcp_md5sig_mutex
);
3104 static bool tcp_md5sig_pool_populated
= false;
3106 static void __tcp_alloc_md5sig_pool(void)
3108 struct crypto_ahash
*hash
;
3111 hash
= crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC
);
3115 for_each_possible_cpu(cpu
) {
3116 void *scratch
= per_cpu(tcp_md5sig_pool
, cpu
).scratch
;
3117 struct ahash_request
*req
;
3120 scratch
= kmalloc_node(sizeof(union tcp_md5sum_block
) +
3121 sizeof(struct tcphdr
),
3126 per_cpu(tcp_md5sig_pool
, cpu
).scratch
= scratch
;
3128 if (per_cpu(tcp_md5sig_pool
, cpu
).md5_req
)
3131 req
= ahash_request_alloc(hash
, GFP_KERNEL
);
3135 ahash_request_set_callback(req
, 0, NULL
, NULL
);
3137 per_cpu(tcp_md5sig_pool
, cpu
).md5_req
= req
;
3139 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3140 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3143 tcp_md5sig_pool_populated
= true;
3146 bool tcp_alloc_md5sig_pool(void)
3148 if (unlikely(!tcp_md5sig_pool_populated
)) {
3149 mutex_lock(&tcp_md5sig_mutex
);
3151 if (!tcp_md5sig_pool_populated
)
3152 __tcp_alloc_md5sig_pool();
3154 mutex_unlock(&tcp_md5sig_mutex
);
3156 return tcp_md5sig_pool_populated
;
3158 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
3162 * tcp_get_md5sig_pool - get md5sig_pool for this user
3164 * We use percpu structure, so if we succeed, we exit with preemption
3165 * and BH disabled, to make sure another thread or softirq handling
3166 * wont try to get same context.
3168 struct tcp_md5sig_pool
*tcp_get_md5sig_pool(void)
3172 if (tcp_md5sig_pool_populated
) {
3173 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3175 return this_cpu_ptr(&tcp_md5sig_pool
);
3180 EXPORT_SYMBOL(tcp_get_md5sig_pool
);
3182 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool
*hp
,
3183 const struct sk_buff
*skb
, unsigned int header_len
)
3185 struct scatterlist sg
;
3186 const struct tcphdr
*tp
= tcp_hdr(skb
);
3187 struct ahash_request
*req
= hp
->md5_req
;
3189 const unsigned int head_data_len
= skb_headlen(skb
) > header_len
?
3190 skb_headlen(skb
) - header_len
: 0;
3191 const struct skb_shared_info
*shi
= skb_shinfo(skb
);
3192 struct sk_buff
*frag_iter
;
3194 sg_init_table(&sg
, 1);
3196 sg_set_buf(&sg
, ((u8
*) tp
) + header_len
, head_data_len
);
3197 ahash_request_set_crypt(req
, &sg
, NULL
, head_data_len
);
3198 if (crypto_ahash_update(req
))
3201 for (i
= 0; i
< shi
->nr_frags
; ++i
) {
3202 const struct skb_frag_struct
*f
= &shi
->frags
[i
];
3203 unsigned int offset
= f
->page_offset
;
3204 struct page
*page
= skb_frag_page(f
) + (offset
>> PAGE_SHIFT
);
3206 sg_set_page(&sg
, page
, skb_frag_size(f
),
3207 offset_in_page(offset
));
3208 ahash_request_set_crypt(req
, &sg
, NULL
, skb_frag_size(f
));
3209 if (crypto_ahash_update(req
))
3213 skb_walk_frags(skb
, frag_iter
)
3214 if (tcp_md5_hash_skb_data(hp
, frag_iter
, 0))
3219 EXPORT_SYMBOL(tcp_md5_hash_skb_data
);
3221 int tcp_md5_hash_key(struct tcp_md5sig_pool
*hp
, const struct tcp_md5sig_key
*key
)
3223 struct scatterlist sg
;
3225 sg_init_one(&sg
, key
->key
, key
->keylen
);
3226 ahash_request_set_crypt(hp
->md5_req
, &sg
, NULL
, key
->keylen
);
3227 return crypto_ahash_update(hp
->md5_req
);
3229 EXPORT_SYMBOL(tcp_md5_hash_key
);
3233 void tcp_done(struct sock
*sk
)
3235 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
3237 if (sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
3238 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ATTEMPTFAILS
);
3240 tcp_set_state(sk
, TCP_CLOSE
);
3241 tcp_clear_xmit_timers(sk
);
3243 reqsk_fastopen_remove(sk
, req
, false);
3245 sk
->sk_shutdown
= SHUTDOWN_MASK
;
3247 if (!sock_flag(sk
, SOCK_DEAD
))
3248 sk
->sk_state_change(sk
);
3250 inet_csk_destroy_sock(sk
);
3252 EXPORT_SYMBOL_GPL(tcp_done
);
3254 int tcp_abort(struct sock
*sk
, int err
)
3256 if (!sk_fullsock(sk
)) {
3257 if (sk
->sk_state
== TCP_NEW_SYN_RECV
) {
3258 struct request_sock
*req
= inet_reqsk(sk
);
3261 inet_csk_reqsk_queue_drop_and_put(req
->rsk_listener
,
3269 /* Don't race with userspace socket closes such as tcp_close. */
3272 if (sk
->sk_state
== TCP_LISTEN
) {
3273 tcp_set_state(sk
, TCP_CLOSE
);
3274 inet_csk_listen_stop(sk
);
3277 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3281 if (!sock_flag(sk
, SOCK_DEAD
)) {
3283 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3285 sk
->sk_error_report(sk
);
3286 if (tcp_need_reset(sk
->sk_state
))
3287 tcp_send_active_reset(sk
, GFP_ATOMIC
);
3296 EXPORT_SYMBOL_GPL(tcp_abort
);
3298 extern struct tcp_congestion_ops tcp_reno
;
3300 static __initdata
unsigned long thash_entries
;
3301 static int __init
set_thash_entries(char *str
)
3308 ret
= kstrtoul(str
, 0, &thash_entries
);
3314 __setup("thash_entries=", set_thash_entries
);
3316 static void __init
tcp_init_mem(void)
3318 unsigned long limit
= nr_free_buffer_pages() / 16;
3320 limit
= max(limit
, 128UL);
3321 sysctl_tcp_mem
[0] = limit
/ 4 * 3; /* 4.68 % */
3322 sysctl_tcp_mem
[1] = limit
; /* 6.25 % */
3323 sysctl_tcp_mem
[2] = sysctl_tcp_mem
[0] * 2; /* 9.37 % */
3326 void __init
tcp_init(void)
3328 int max_rshare
, max_wshare
, cnt
;
3329 unsigned long limit
;
3332 BUILD_BUG_ON(sizeof(struct tcp_skb_cb
) >
3333 FIELD_SIZEOF(struct sk_buff
, cb
));
3335 percpu_counter_init(&tcp_sockets_allocated
, 0, GFP_KERNEL
);
3336 percpu_counter_init(&tcp_orphan_count
, 0, GFP_KERNEL
);
3337 tcp_hashinfo
.bind_bucket_cachep
=
3338 kmem_cache_create("tcp_bind_bucket",
3339 sizeof(struct inet_bind_bucket
), 0,
3340 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
);
3342 /* Size and allocate the main established and bind bucket
3345 * The methodology is similar to that of the buffer cache.
3347 tcp_hashinfo
.ehash
=
3348 alloc_large_system_hash("TCP established",
3349 sizeof(struct inet_ehash_bucket
),
3351 17, /* one slot per 128 KB of memory */
3354 &tcp_hashinfo
.ehash_mask
,
3356 thash_entries
? 0 : 512 * 1024);
3357 for (i
= 0; i
<= tcp_hashinfo
.ehash_mask
; i
++)
3358 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].chain
, i
);
3360 if (inet_ehash_locks_alloc(&tcp_hashinfo
))
3361 panic("TCP: failed to alloc ehash_locks");
3362 tcp_hashinfo
.bhash
=
3363 alloc_large_system_hash("TCP bind",
3364 sizeof(struct inet_bind_hashbucket
),
3365 tcp_hashinfo
.ehash_mask
+ 1,
3366 17, /* one slot per 128 KB of memory */
3368 &tcp_hashinfo
.bhash_size
,
3372 tcp_hashinfo
.bhash_size
= 1U << tcp_hashinfo
.bhash_size
;
3373 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
3374 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
3375 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
3379 cnt
= tcp_hashinfo
.ehash_mask
+ 1;
3381 tcp_death_row
.sysctl_max_tw_buckets
= cnt
/ 2;
3382 sysctl_tcp_max_orphans
= cnt
/ 2;
3383 sysctl_max_syn_backlog
= max(128, cnt
/ 256);
3386 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3387 limit
= nr_free_buffer_pages() << (PAGE_SHIFT
- 7);
3388 max_wshare
= min(4UL*1024*1024, limit
);
3389 max_rshare
= min(6UL*1024*1024, limit
);
3391 sysctl_tcp_wmem
[0] = SK_MEM_QUANTUM
;
3392 sysctl_tcp_wmem
[1] = 16*1024;
3393 sysctl_tcp_wmem
[2] = max(64*1024, max_wshare
);
3395 sysctl_tcp_rmem
[0] = SK_MEM_QUANTUM
;
3396 sysctl_tcp_rmem
[1] = 87380;
3397 sysctl_tcp_rmem
[2] = max(87380, max_rshare
);
3399 pr_info("Hash tables configured (established %u bind %u)\n",
3400 tcp_hashinfo
.ehash_mask
+ 1, tcp_hashinfo
.bhash_size
);
3403 BUG_ON(tcp_register_congestion_control(&tcp_reno
) != 0);