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>
272 #include <linux/errqueue.h>
273 #include <linux/static_key.h>
275 #include <net/icmp.h>
276 #include <net/inet_common.h>
278 #include <net/xfrm.h>
280 #include <net/sock.h>
282 #include <linux/uaccess.h>
283 #include <asm/ioctls.h>
284 #include <net/busy_poll.h>
286 #include <trace/events/tcp.h>
288 struct percpu_counter tcp_orphan_count
;
289 EXPORT_SYMBOL_GPL(tcp_orphan_count
);
291 long sysctl_tcp_mem
[3] __read_mostly
;
292 EXPORT_SYMBOL(sysctl_tcp_mem
);
294 atomic_long_t tcp_memory_allocated
; /* Current allocated memory. */
295 EXPORT_SYMBOL(tcp_memory_allocated
);
297 #if IS_ENABLED(CONFIG_SMC)
298 DEFINE_STATIC_KEY_FALSE(tcp_have_smc
);
299 EXPORT_SYMBOL(tcp_have_smc
);
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 unsigned long tcp_memory_pressure __read_mostly
;
324 EXPORT_SYMBOL_GPL(tcp_memory_pressure
);
326 void tcp_enter_memory_pressure(struct sock
*sk
)
330 if (tcp_memory_pressure
)
336 if (!cmpxchg(&tcp_memory_pressure
, 0, val
))
337 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMEMORYPRESSURES
);
339 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure
);
341 void tcp_leave_memory_pressure(struct sock
*sk
)
345 if (!tcp_memory_pressure
)
347 val
= xchg(&tcp_memory_pressure
, 0);
349 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPMEMORYPRESSURESCHRONO
,
350 jiffies_to_msecs(jiffies
- val
));
352 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure
);
354 /* Convert seconds to retransmits based on initial and max timeout */
355 static u8
secs_to_retrans(int seconds
, int timeout
, int rto_max
)
360 int period
= timeout
;
363 while (seconds
> period
&& res
< 255) {
366 if (timeout
> rto_max
)
374 /* Convert retransmits to seconds based on initial and max timeout */
375 static int retrans_to_secs(u8 retrans
, int timeout
, int rto_max
)
383 if (timeout
> rto_max
)
391 static u64
tcp_compute_delivery_rate(const struct tcp_sock
*tp
)
393 u32 rate
= READ_ONCE(tp
->rate_delivered
);
394 u32 intv
= READ_ONCE(tp
->rate_interval_us
);
398 rate64
= (u64
)rate
* tp
->mss_cache
* USEC_PER_SEC
;
399 do_div(rate64
, intv
);
404 /* Address-family independent initialization for a tcp_sock.
406 * NOTE: A lot of things set to zero explicitly by call to
407 * sk_alloc() so need not be done here.
409 void tcp_init_sock(struct sock
*sk
)
411 struct inet_connection_sock
*icsk
= inet_csk(sk
);
412 struct tcp_sock
*tp
= tcp_sk(sk
);
414 tp
->out_of_order_queue
= RB_ROOT
;
415 sk
->tcp_rtx_queue
= RB_ROOT
;
416 tcp_init_xmit_timers(sk
);
417 INIT_LIST_HEAD(&tp
->tsq_node
);
418 INIT_LIST_HEAD(&tp
->tsorted_sent_queue
);
420 icsk
->icsk_rto
= TCP_TIMEOUT_INIT
;
421 tp
->mdev_us
= jiffies_to_usecs(TCP_TIMEOUT_INIT
);
422 minmax_reset(&tp
->rtt_min
, tcp_jiffies32
, ~0U);
424 /* So many TCP implementations out there (incorrectly) count the
425 * initial SYN frame in their delayed-ACK and congestion control
426 * algorithms that we must have the following bandaid to talk
427 * efficiently to them. -DaveM
429 tp
->snd_cwnd
= TCP_INIT_CWND
;
431 /* There's a bubble in the pipe until at least the first ACK. */
432 tp
->app_limited
= ~0U;
434 /* See draft-stevens-tcpca-spec-01 for discussion of the
435 * initialization of these values.
437 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
438 tp
->snd_cwnd_clamp
= ~0;
439 tp
->mss_cache
= TCP_MSS_DEFAULT
;
441 tp
->reordering
= sock_net(sk
)->ipv4
.sysctl_tcp_reordering
;
442 tcp_assign_congestion_control(sk
);
445 tp
->rack
.reo_wnd_steps
= 1;
447 sk
->sk_state
= TCP_CLOSE
;
449 sk
->sk_write_space
= sk_stream_write_space
;
450 sock_set_flag(sk
, SOCK_USE_WRITE_QUEUE
);
452 icsk
->icsk_sync_mss
= tcp_sync_mss
;
454 sk
->sk_sndbuf
= sock_net(sk
)->ipv4
.sysctl_tcp_wmem
[1];
455 sk
->sk_rcvbuf
= sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[1];
457 sk_sockets_allocated_inc(sk
);
459 EXPORT_SYMBOL(tcp_init_sock
);
461 void tcp_init_transfer(struct sock
*sk
, int bpf_op
)
463 struct inet_connection_sock
*icsk
= inet_csk(sk
);
466 icsk
->icsk_af_ops
->rebuild_header(sk
);
467 tcp_init_metrics(sk
);
468 tcp_call_bpf(sk
, bpf_op
);
469 tcp_init_congestion_control(sk
);
470 tcp_init_buffer_space(sk
);
473 static void tcp_tx_timestamp(struct sock
*sk
, u16 tsflags
)
475 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
477 if (tsflags
&& skb
) {
478 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
479 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
481 sock_tx_timestamp(sk
, tsflags
, &shinfo
->tx_flags
);
482 if (tsflags
& SOF_TIMESTAMPING_TX_ACK
)
483 tcb
->txstamp_ack
= 1;
484 if (tsflags
& SOF_TIMESTAMPING_TX_RECORD_MASK
)
485 shinfo
->tskey
= TCP_SKB_CB(skb
)->seq
+ skb
->len
- 1;
490 * Wait for a TCP event.
492 * Note that we don't need to lock the socket, as the upper poll layers
493 * take care of normal races (between the test and the event) and we don't
494 * go look at any of the socket buffers directly.
496 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
499 struct sock
*sk
= sock
->sk
;
500 const struct tcp_sock
*tp
= tcp_sk(sk
);
503 sock_rps_record_flow(sk
);
505 sock_poll_wait(file
, sk_sleep(sk
), wait
);
507 state
= sk_state_load(sk
);
508 if (state
== TCP_LISTEN
)
509 return inet_csk_listen_poll(sk
);
511 /* Socket is not locked. We are protected from async events
512 * by poll logic and correct handling of state changes
513 * made by other threads is impossible in any case.
519 * POLLHUP is certainly not done right. But poll() doesn't
520 * have a notion of HUP in just one direction, and for a
521 * socket the read side is more interesting.
523 * Some poll() documentation says that POLLHUP is incompatible
524 * with the POLLOUT/POLLWR flags, so somebody should check this
525 * all. But careful, it tends to be safer to return too many
526 * bits than too few, and you can easily break real applications
527 * if you don't tell them that something has hung up!
531 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
532 * our fs/select.c). It means that after we received EOF,
533 * poll always returns immediately, making impossible poll() on write()
534 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
535 * if and only if shutdown has been made in both directions.
536 * Actually, it is interesting to look how Solaris and DUX
537 * solve this dilemma. I would prefer, if POLLHUP were maskable,
538 * then we could set it on SND_SHUTDOWN. BTW examples given
539 * in Stevens' books assume exactly this behaviour, it explains
540 * why POLLHUP is incompatible with POLLOUT. --ANK
542 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
543 * blocking on fresh not-connected or disconnected socket. --ANK
545 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| state
== TCP_CLOSE
)
547 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
548 mask
|= POLLIN
| POLLRDNORM
| POLLRDHUP
;
550 /* Connected or passive Fast Open socket? */
551 if (state
!= TCP_SYN_SENT
&&
552 (state
!= TCP_SYN_RECV
|| tp
->fastopen_rsk
)) {
553 int target
= sock_rcvlowat(sk
, 0, INT_MAX
);
555 if (tp
->urg_seq
== tp
->copied_seq
&&
556 !sock_flag(sk
, SOCK_URGINLINE
) &&
560 if (tp
->rcv_nxt
- tp
->copied_seq
>= target
)
561 mask
|= POLLIN
| POLLRDNORM
;
563 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
564 if (sk_stream_is_writeable(sk
)) {
565 mask
|= POLLOUT
| POLLWRNORM
;
566 } else { /* send SIGIO later */
567 sk_set_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
568 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
570 /* Race breaker. If space is freed after
571 * wspace test but before the flags are set,
572 * IO signal will be lost. Memory barrier
573 * pairs with the input side.
575 smp_mb__after_atomic();
576 if (sk_stream_is_writeable(sk
))
577 mask
|= POLLOUT
| POLLWRNORM
;
580 mask
|= POLLOUT
| POLLWRNORM
;
582 if (tp
->urg_data
& TCP_URG_VALID
)
584 } else if (state
== TCP_SYN_SENT
&& inet_sk(sk
)->defer_connect
) {
585 /* Active TCP fastopen socket with defer_connect
586 * Return POLLOUT so application can call write()
587 * in order for kernel to generate SYN+data
589 mask
|= POLLOUT
| POLLWRNORM
;
591 /* This barrier is coupled with smp_wmb() in tcp_reset() */
593 if (sk
->sk_err
|| !skb_queue_empty(&sk
->sk_error_queue
))
598 EXPORT_SYMBOL(tcp_poll
);
600 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
602 struct tcp_sock
*tp
= tcp_sk(sk
);
608 if (sk
->sk_state
== TCP_LISTEN
)
611 slow
= lock_sock_fast(sk
);
613 unlock_sock_fast(sk
, slow
);
616 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
619 if (sk
->sk_state
== TCP_LISTEN
)
622 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
625 answ
= tp
->write_seq
- tp
->snd_una
;
628 if (sk
->sk_state
== TCP_LISTEN
)
631 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
634 answ
= tp
->write_seq
- tp
->snd_nxt
;
640 return put_user(answ
, (int __user
*)arg
);
642 EXPORT_SYMBOL(tcp_ioctl
);
644 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
646 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
647 tp
->pushed_seq
= tp
->write_seq
;
650 static inline bool forced_push(const struct tcp_sock
*tp
)
652 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
655 static void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
657 struct tcp_sock
*tp
= tcp_sk(sk
);
658 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
661 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
662 tcb
->tcp_flags
= TCPHDR_ACK
;
664 __skb_header_release(skb
);
665 tcp_add_write_queue_tail(sk
, skb
);
666 sk
->sk_wmem_queued
+= skb
->truesize
;
667 sk_mem_charge(sk
, skb
->truesize
);
668 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
669 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
671 tcp_slow_start_after_idle_check(sk
);
674 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
)
677 tp
->snd_up
= tp
->write_seq
;
680 /* If a not yet filled skb is pushed, do not send it if
681 * we have data packets in Qdisc or NIC queues :
682 * Because TX completion will happen shortly, it gives a chance
683 * to coalesce future sendmsg() payload into this skb, without
684 * need for a timer, and with no latency trade off.
685 * As packets containing data payload have a bigger truesize
686 * than pure acks (dataless) packets, the last checks prevent
687 * autocorking if we only have an ACK in Qdisc/NIC queues,
688 * or if TX completion was delayed after we processed ACK packet.
690 static bool tcp_should_autocork(struct sock
*sk
, struct sk_buff
*skb
,
693 return skb
->len
< size_goal
&&
694 sock_net(sk
)->ipv4
.sysctl_tcp_autocorking
&&
695 !tcp_rtx_queue_empty(sk
) &&
696 refcount_read(&sk
->sk_wmem_alloc
) > skb
->truesize
;
699 static void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
700 int nonagle
, int size_goal
)
702 struct tcp_sock
*tp
= tcp_sk(sk
);
705 skb
= tcp_write_queue_tail(sk
);
708 if (!(flags
& MSG_MORE
) || forced_push(tp
))
709 tcp_mark_push(tp
, skb
);
711 tcp_mark_urg(tp
, flags
);
713 if (tcp_should_autocork(sk
, skb
, size_goal
)) {
715 /* avoid atomic op if TSQ_THROTTLED bit is already set */
716 if (!test_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
)) {
717 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPAUTOCORKING
);
718 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
720 /* It is possible TX completion already happened
721 * before we set TSQ_THROTTLED.
723 if (refcount_read(&sk
->sk_wmem_alloc
) > skb
->truesize
)
727 if (flags
& MSG_MORE
)
728 nonagle
= TCP_NAGLE_CORK
;
730 __tcp_push_pending_frames(sk
, mss_now
, nonagle
);
733 static int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
734 unsigned int offset
, size_t len
)
736 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
739 ret
= skb_splice_bits(skb
, skb
->sk
, offset
, tss
->pipe
,
740 min(rd_desc
->count
, len
), tss
->flags
);
742 rd_desc
->count
-= ret
;
746 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
748 /* Store TCP splice context information in read_descriptor_t. */
749 read_descriptor_t rd_desc
= {
754 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
758 * tcp_splice_read - splice data from TCP socket to a pipe
759 * @sock: socket to splice from
760 * @ppos: position (not valid)
761 * @pipe: pipe to splice to
762 * @len: number of bytes to splice
763 * @flags: splice modifier flags
766 * Will read pages from given socket and fill them into a pipe.
769 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
770 struct pipe_inode_info
*pipe
, size_t len
,
773 struct sock
*sk
= sock
->sk
;
774 struct tcp_splice_state tss
= {
783 sock_rps_record_flow(sk
);
785 * We can't seek on a socket input
794 timeo
= sock_rcvtimeo(sk
, sock
->file
->f_flags
& O_NONBLOCK
);
796 ret
= __tcp_splice_read(sk
, &tss
);
802 if (sock_flag(sk
, SOCK_DONE
))
805 ret
= sock_error(sk
);
808 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
810 if (sk
->sk_state
== TCP_CLOSE
) {
812 * This occurs when user tries to read
813 * from never connected socket.
815 if (!sock_flag(sk
, SOCK_DONE
))
823 /* if __tcp_splice_read() got nothing while we have
824 * an skb in receive queue, we do not want to loop.
825 * This might happen with URG data.
827 if (!skb_queue_empty(&sk
->sk_receive_queue
))
829 sk_wait_data(sk
, &timeo
, NULL
);
830 if (signal_pending(current
)) {
831 ret
= sock_intr_errno(timeo
);
844 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
845 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
846 signal_pending(current
))
857 EXPORT_SYMBOL(tcp_splice_read
);
859 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
,
864 /* The TCP header must be at least 32-bit aligned. */
865 size
= ALIGN(size
, 4);
867 if (unlikely(tcp_under_memory_pressure(sk
)))
868 sk_mem_reclaim_partial(sk
);
870 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
874 if (force_schedule
) {
875 mem_scheduled
= true;
876 sk_forced_mem_schedule(sk
, skb
->truesize
);
878 mem_scheduled
= sk_wmem_schedule(sk
, skb
->truesize
);
880 if (likely(mem_scheduled
)) {
881 skb_reserve(skb
, sk
->sk_prot
->max_header
);
883 * Make sure that we have exactly size bytes
884 * available to the caller, no more, no less.
886 skb
->reserved_tailroom
= skb
->end
- skb
->tail
- size
;
887 INIT_LIST_HEAD(&skb
->tcp_tsorted_anchor
);
892 sk
->sk_prot
->enter_memory_pressure(sk
);
893 sk_stream_moderate_sndbuf(sk
);
898 static unsigned int tcp_xmit_size_goal(struct sock
*sk
, u32 mss_now
,
901 struct tcp_sock
*tp
= tcp_sk(sk
);
902 u32 new_size_goal
, size_goal
;
904 if (!large_allowed
|| !sk_can_gso(sk
))
907 /* Note : tcp_tso_autosize() will eventually split this later */
908 new_size_goal
= sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
;
909 new_size_goal
= tcp_bound_to_half_wnd(tp
, new_size_goal
);
911 /* We try hard to avoid divides here */
912 size_goal
= tp
->gso_segs
* mss_now
;
913 if (unlikely(new_size_goal
< size_goal
||
914 new_size_goal
>= size_goal
+ mss_now
)) {
915 tp
->gso_segs
= min_t(u16
, new_size_goal
/ mss_now
,
916 sk
->sk_gso_max_segs
);
917 size_goal
= tp
->gso_segs
* mss_now
;
920 return max(size_goal
, mss_now
);
923 static int tcp_send_mss(struct sock
*sk
, int *size_goal
, int flags
)
927 mss_now
= tcp_current_mss(sk
);
928 *size_goal
= tcp_xmit_size_goal(sk
, mss_now
, !(flags
& MSG_OOB
));
933 ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
*page
, int offset
,
934 size_t size
, int flags
)
936 struct tcp_sock
*tp
= tcp_sk(sk
);
937 int mss_now
, size_goal
;
940 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
942 /* Wait for a connection to finish. One exception is TCP Fast Open
943 * (passive side) where data is allowed to be sent before a connection
944 * is fully established.
946 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
947 !tcp_passive_fastopen(sk
)) {
948 err
= sk_stream_wait_connect(sk
, &timeo
);
953 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
955 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
959 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
963 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
967 if (!skb
|| (copy
= size_goal
- skb
->len
) <= 0 ||
968 !tcp_skb_can_collapse_to(skb
)) {
970 if (!sk_stream_memory_free(sk
))
971 goto wait_for_sndbuf
;
973 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
,
974 tcp_rtx_and_write_queues_empty(sk
));
976 goto wait_for_memory
;
985 i
= skb_shinfo(skb
)->nr_frags
;
986 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
987 if (!can_coalesce
&& i
>= sysctl_max_skb_frags
) {
988 tcp_mark_push(tp
, skb
);
991 if (!sk_wmem_schedule(sk
, copy
))
992 goto wait_for_memory
;
995 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
998 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
1000 skb_shinfo(skb
)->tx_flags
|= SKBTX_SHARED_FRAG
;
1003 skb
->data_len
+= copy
;
1004 skb
->truesize
+= copy
;
1005 sk
->sk_wmem_queued
+= copy
;
1006 sk_mem_charge(sk
, copy
);
1007 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1008 tp
->write_seq
+= copy
;
1009 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1010 tcp_skb_pcount_set(skb
, 0);
1013 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
1021 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
1024 if (forced_push(tp
)) {
1025 tcp_mark_push(tp
, skb
);
1026 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1027 } else if (skb
== tcp_send_head(sk
))
1028 tcp_push_one(sk
, mss_now
);
1032 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1034 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
,
1035 TCP_NAGLE_PUSH
, size_goal
);
1037 err
= sk_stream_wait_memory(sk
, &timeo
);
1041 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1046 tcp_tx_timestamp(sk
, sk
->sk_tsflags
);
1047 if (!(flags
& MSG_SENDPAGE_NOTLAST
))
1048 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
, size_goal
);
1056 /* make sure we wake any epoll edge trigger waiter */
1057 if (unlikely(skb_queue_len(&sk
->sk_write_queue
) == 0 &&
1059 sk
->sk_write_space(sk
);
1060 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1062 return sk_stream_error(sk
, flags
, err
);
1064 EXPORT_SYMBOL_GPL(do_tcp_sendpages
);
1066 int tcp_sendpage_locked(struct sock
*sk
, struct page
*page
, int offset
,
1067 size_t size
, int flags
)
1069 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
1070 !sk_check_csum_caps(sk
))
1071 return sock_no_sendpage_locked(sk
, page
, offset
, size
, flags
);
1073 tcp_rate_check_app_limited(sk
); /* is sending application-limited? */
1075 return do_tcp_sendpages(sk
, page
, offset
, size
, flags
);
1077 EXPORT_SYMBOL_GPL(tcp_sendpage_locked
);
1079 int tcp_sendpage(struct sock
*sk
, struct page
*page
, int offset
,
1080 size_t size
, int flags
)
1085 ret
= tcp_sendpage_locked(sk
, page
, offset
, size
, flags
);
1090 EXPORT_SYMBOL(tcp_sendpage
);
1092 /* Do not bother using a page frag for very small frames.
1093 * But use this heuristic only for the first skb in write queue.
1095 * Having no payload in skb->head allows better SACK shifting
1096 * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1097 * write queue has less skbs.
1098 * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1099 * This also speeds up tso_fragment(), since it wont fallback
1100 * to tcp_fragment().
1102 static int linear_payload_sz(bool first_skb
)
1105 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER
);
1109 static int select_size(const struct sock
*sk
, bool sg
, bool first_skb
)
1111 const struct tcp_sock
*tp
= tcp_sk(sk
);
1112 int tmp
= tp
->mss_cache
;
1115 if (sk_can_gso(sk
)) {
1116 tmp
= linear_payload_sz(first_skb
);
1118 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
1120 if (tmp
>= pgbreak
&&
1121 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
1129 void tcp_free_fastopen_req(struct tcp_sock
*tp
)
1131 if (tp
->fastopen_req
) {
1132 kfree(tp
->fastopen_req
);
1133 tp
->fastopen_req
= NULL
;
1137 static int tcp_sendmsg_fastopen(struct sock
*sk
, struct msghdr
*msg
,
1138 int *copied
, size_t size
)
1140 struct tcp_sock
*tp
= tcp_sk(sk
);
1141 struct inet_sock
*inet
= inet_sk(sk
);
1142 struct sockaddr
*uaddr
= msg
->msg_name
;
1145 if (!(sock_net(sk
)->ipv4
.sysctl_tcp_fastopen
& TFO_CLIENT_ENABLE
) ||
1146 (uaddr
&& msg
->msg_namelen
>= sizeof(uaddr
->sa_family
) &&
1147 uaddr
->sa_family
== AF_UNSPEC
))
1149 if (tp
->fastopen_req
)
1150 return -EALREADY
; /* Another Fast Open is in progress */
1152 tp
->fastopen_req
= kzalloc(sizeof(struct tcp_fastopen_request
),
1154 if (unlikely(!tp
->fastopen_req
))
1156 tp
->fastopen_req
->data
= msg
;
1157 tp
->fastopen_req
->size
= size
;
1159 if (inet
->defer_connect
) {
1160 err
= tcp_connect(sk
);
1161 /* Same failure procedure as in tcp_v4/6_connect */
1163 tcp_set_state(sk
, TCP_CLOSE
);
1164 inet
->inet_dport
= 0;
1165 sk
->sk_route_caps
= 0;
1168 flags
= (msg
->msg_flags
& MSG_DONTWAIT
) ? O_NONBLOCK
: 0;
1169 err
= __inet_stream_connect(sk
->sk_socket
, uaddr
,
1170 msg
->msg_namelen
, flags
, 1);
1171 /* fastopen_req could already be freed in __inet_stream_connect
1172 * if the connection times out or gets rst
1174 if (tp
->fastopen_req
) {
1175 *copied
= tp
->fastopen_req
->copied
;
1176 tcp_free_fastopen_req(tp
);
1177 inet
->defer_connect
= 0;
1182 int tcp_sendmsg_locked(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
1184 struct tcp_sock
*tp
= tcp_sk(sk
);
1185 struct ubuf_info
*uarg
= NULL
;
1186 struct sk_buff
*skb
;
1187 struct sockcm_cookie sockc
;
1188 int flags
, err
, copied
= 0;
1189 int mss_now
= 0, size_goal
, copied_syn
= 0;
1190 bool process_backlog
= false;
1194 flags
= msg
->msg_flags
;
1196 if (flags
& MSG_ZEROCOPY
&& size
&& sock_flag(sk
, SOCK_ZEROCOPY
)) {
1197 if ((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) {
1202 skb
= tcp_write_queue_tail(sk
);
1203 uarg
= sock_zerocopy_realloc(sk
, size
, skb_zcopy(skb
));
1209 if (!(sk_check_csum_caps(sk
) && sk
->sk_route_caps
& NETIF_F_SG
))
1213 if (unlikely(flags
& MSG_FASTOPEN
|| inet_sk(sk
)->defer_connect
) &&
1215 err
= tcp_sendmsg_fastopen(sk
, msg
, &copied_syn
, size
);
1216 if (err
== -EINPROGRESS
&& copied_syn
> 0)
1222 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
1224 tcp_rate_check_app_limited(sk
); /* is sending application-limited? */
1226 /* Wait for a connection to finish. One exception is TCP Fast Open
1227 * (passive side) where data is allowed to be sent before a connection
1228 * is fully established.
1230 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
1231 !tcp_passive_fastopen(sk
)) {
1232 err
= sk_stream_wait_connect(sk
, &timeo
);
1237 if (unlikely(tp
->repair
)) {
1238 if (tp
->repair_queue
== TCP_RECV_QUEUE
) {
1239 copied
= tcp_send_rcvq(sk
, msg
, size
);
1244 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1247 /* 'common' sending to sendq */
1250 sockc
.tsflags
= sk
->sk_tsflags
;
1251 if (msg
->msg_controllen
) {
1252 err
= sock_cmsg_send(sk
, msg
, &sockc
);
1253 if (unlikely(err
)) {
1259 /* This should be in poll */
1260 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
1262 /* Ok commence sending. */
1266 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1269 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
1272 sg
= !!(sk
->sk_route_caps
& NETIF_F_SG
);
1274 while (msg_data_left(msg
)) {
1276 int max
= size_goal
;
1278 skb
= tcp_write_queue_tail(sk
);
1280 if (skb
->ip_summed
== CHECKSUM_NONE
)
1282 copy
= max
- skb
->len
;
1285 if (copy
<= 0 || !tcp_skb_can_collapse_to(skb
)) {
1289 /* Allocate new segment. If the interface is SG,
1290 * allocate skb fitting to single page.
1292 if (!sk_stream_memory_free(sk
))
1293 goto wait_for_sndbuf
;
1295 if (process_backlog
&& sk_flush_backlog(sk
)) {
1296 process_backlog
= false;
1299 first_skb
= tcp_rtx_and_write_queues_empty(sk
);
1300 skb
= sk_stream_alloc_skb(sk
,
1301 select_size(sk
, sg
, first_skb
),
1305 goto wait_for_memory
;
1307 process_backlog
= true;
1309 * Check whether we can use HW checksum.
1311 if (sk_check_csum_caps(sk
))
1312 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1314 skb_entail(sk
, skb
);
1318 /* All packets are restored as if they have
1319 * already been sent. skb_mstamp isn't set to
1320 * avoid wrong rtt estimation.
1323 TCP_SKB_CB(skb
)->sacked
|= TCPCB_REPAIRED
;
1326 /* Try to append data to the end of skb. */
1327 if (copy
> msg_data_left(msg
))
1328 copy
= msg_data_left(msg
);
1330 /* Where to copy to? */
1331 if (skb_availroom(skb
) > 0) {
1332 /* We have some space in skb head. Superb! */
1333 copy
= min_t(int, copy
, skb_availroom(skb
));
1334 err
= skb_add_data_nocache(sk
, skb
, &msg
->msg_iter
, copy
);
1337 } else if (!uarg
|| !uarg
->zerocopy
) {
1339 int i
= skb_shinfo(skb
)->nr_frags
;
1340 struct page_frag
*pfrag
= sk_page_frag(sk
);
1342 if (!sk_page_frag_refill(sk
, pfrag
))
1343 goto wait_for_memory
;
1345 if (!skb_can_coalesce(skb
, i
, pfrag
->page
,
1347 if (i
>= sysctl_max_skb_frags
|| !sg
) {
1348 tcp_mark_push(tp
, skb
);
1354 copy
= min_t(int, copy
, pfrag
->size
- pfrag
->offset
);
1356 if (!sk_wmem_schedule(sk
, copy
))
1357 goto wait_for_memory
;
1359 err
= skb_copy_to_page_nocache(sk
, &msg
->msg_iter
, skb
,
1366 /* Update the skb. */
1368 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
1370 skb_fill_page_desc(skb
, i
, pfrag
->page
,
1371 pfrag
->offset
, copy
);
1372 page_ref_inc(pfrag
->page
);
1374 pfrag
->offset
+= copy
;
1376 err
= skb_zerocopy_iter_stream(sk
, skb
, msg
, copy
, uarg
);
1377 if (err
== -EMSGSIZE
|| err
== -EEXIST
)
1385 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
1387 tp
->write_seq
+= copy
;
1388 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1389 tcp_skb_pcount_set(skb
, 0);
1392 if (!msg_data_left(msg
)) {
1393 if (unlikely(flags
& MSG_EOR
))
1394 TCP_SKB_CB(skb
)->eor
= 1;
1398 if (skb
->len
< max
|| (flags
& MSG_OOB
) || unlikely(tp
->repair
))
1401 if (forced_push(tp
)) {
1402 tcp_mark_push(tp
, skb
);
1403 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1404 } else if (skb
== tcp_send_head(sk
))
1405 tcp_push_one(sk
, mss_now
);
1409 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1412 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
,
1413 TCP_NAGLE_PUSH
, size_goal
);
1415 err
= sk_stream_wait_memory(sk
, &timeo
);
1419 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1424 tcp_tx_timestamp(sk
, sockc
.tsflags
);
1425 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
, size_goal
);
1428 sock_zerocopy_put(uarg
);
1429 return copied
+ copied_syn
;
1433 tcp_unlink_write_queue(skb
, sk
);
1434 /* It is the one place in all of TCP, except connection
1435 * reset, where we can be unlinking the send_head.
1437 tcp_check_send_head(sk
, skb
);
1438 sk_wmem_free_skb(sk
, skb
);
1442 if (copied
+ copied_syn
)
1445 sock_zerocopy_put_abort(uarg
);
1446 err
= sk_stream_error(sk
, flags
, err
);
1447 /* make sure we wake any epoll edge trigger waiter */
1448 if (unlikely(skb_queue_len(&sk
->sk_write_queue
) == 0 &&
1450 sk
->sk_write_space(sk
);
1451 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1455 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked
);
1457 int tcp_sendmsg(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
1462 ret
= tcp_sendmsg_locked(sk
, msg
, size
);
1467 EXPORT_SYMBOL(tcp_sendmsg
);
1470 * Handle reading urgent data. BSD has very simple semantics for
1471 * this, no blocking and very strange errors 8)
1474 static int tcp_recv_urg(struct sock
*sk
, struct msghdr
*msg
, int len
, int flags
)
1476 struct tcp_sock
*tp
= tcp_sk(sk
);
1478 /* No URG data to read. */
1479 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1480 tp
->urg_data
== TCP_URG_READ
)
1481 return -EINVAL
; /* Yes this is right ! */
1483 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1486 if (tp
->urg_data
& TCP_URG_VALID
) {
1488 char c
= tp
->urg_data
;
1490 if (!(flags
& MSG_PEEK
))
1491 tp
->urg_data
= TCP_URG_READ
;
1493 /* Read urgent data. */
1494 msg
->msg_flags
|= MSG_OOB
;
1497 if (!(flags
& MSG_TRUNC
))
1498 err
= memcpy_to_msg(msg
, &c
, 1);
1501 msg
->msg_flags
|= MSG_TRUNC
;
1503 return err
? -EFAULT
: len
;
1506 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1509 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1510 * the available implementations agree in this case:
1511 * this call should never block, independent of the
1512 * blocking state of the socket.
1513 * Mike <pall@rz.uni-karlsruhe.de>
1518 static int tcp_peek_sndq(struct sock
*sk
, struct msghdr
*msg
, int len
)
1520 struct sk_buff
*skb
;
1521 int copied
= 0, err
= 0;
1523 /* XXX -- need to support SO_PEEK_OFF */
1525 skb_rbtree_walk(skb
, &sk
->tcp_rtx_queue
) {
1526 err
= skb_copy_datagram_msg(skb
, 0, msg
, skb
->len
);
1532 skb_queue_walk(&sk
->sk_write_queue
, skb
) {
1533 err
= skb_copy_datagram_msg(skb
, 0, msg
, skb
->len
);
1540 return err
?: copied
;
1543 /* Clean up the receive buffer for full frames taken by the user,
1544 * then send an ACK if necessary. COPIED is the number of bytes
1545 * tcp_recvmsg has given to the user so far, it speeds up the
1546 * calculation of whether or not we must ACK for the sake of
1549 static void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1551 struct tcp_sock
*tp
= tcp_sk(sk
);
1552 bool time_to_ack
= false;
1554 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1556 WARN(skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
),
1557 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1558 tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
);
1560 if (inet_csk_ack_scheduled(sk
)) {
1561 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1562 /* Delayed ACKs frequently hit locked sockets during bulk
1564 if (icsk
->icsk_ack
.blocked
||
1565 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1566 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1568 * If this read emptied read buffer, we send ACK, if
1569 * connection is not bidirectional, user drained
1570 * receive buffer and there was a small segment
1574 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1575 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1576 !icsk
->icsk_ack
.pingpong
)) &&
1577 !atomic_read(&sk
->sk_rmem_alloc
)))
1581 /* We send an ACK if we can now advertise a non-zero window
1582 * which has been raised "significantly".
1584 * Even if window raised up to infinity, do not send window open ACK
1585 * in states, where we will not receive more. It is useless.
1587 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1588 __u32 rcv_window_now
= tcp_receive_window(tp
);
1590 /* Optimize, __tcp_select_window() is not cheap. */
1591 if (2*rcv_window_now
<= tp
->window_clamp
) {
1592 __u32 new_window
= __tcp_select_window(sk
);
1594 /* Send ACK now, if this read freed lots of space
1595 * in our buffer. Certainly, new_window is new window.
1596 * We can advertise it now, if it is not less than current one.
1597 * "Lots" means "at least twice" here.
1599 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1607 static struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1609 struct sk_buff
*skb
;
1612 while ((skb
= skb_peek(&sk
->sk_receive_queue
)) != NULL
) {
1613 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1614 if (unlikely(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
1615 pr_err_once("%s: found a SYN, please report !\n", __func__
);
1618 if (offset
< skb
->len
|| (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)) {
1622 /* This looks weird, but this can happen if TCP collapsing
1623 * splitted a fat GRO packet, while we released socket lock
1624 * in skb_splice_bits()
1626 sk_eat_skb(sk
, skb
);
1632 * This routine provides an alternative to tcp_recvmsg() for routines
1633 * that would like to handle copying from skbuffs directly in 'sendfile'
1636 * - It is assumed that the socket was locked by the caller.
1637 * - The routine does not block.
1638 * - At present, there is no support for reading OOB data
1639 * or for 'peeking' the socket using this routine
1640 * (although both would be easy to implement).
1642 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1643 sk_read_actor_t recv_actor
)
1645 struct sk_buff
*skb
;
1646 struct tcp_sock
*tp
= tcp_sk(sk
);
1647 u32 seq
= tp
->copied_seq
;
1651 if (sk
->sk_state
== TCP_LISTEN
)
1653 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1654 if (offset
< skb
->len
) {
1658 len
= skb
->len
- offset
;
1659 /* Stop reading if we hit a patch of urgent data */
1661 u32 urg_offset
= tp
->urg_seq
- seq
;
1662 if (urg_offset
< len
)
1667 used
= recv_actor(desc
, skb
, offset
, len
);
1672 } else if (used
<= len
) {
1677 /* If recv_actor drops the lock (e.g. TCP splice
1678 * receive) the skb pointer might be invalid when
1679 * getting here: tcp_collapse might have deleted it
1680 * while aggregating skbs from the socket queue.
1682 skb
= tcp_recv_skb(sk
, seq
- 1, &offset
);
1685 /* TCP coalescing might have appended data to the skb.
1686 * Try to splice more frags
1688 if (offset
+ 1 != skb
->len
)
1691 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) {
1692 sk_eat_skb(sk
, skb
);
1696 sk_eat_skb(sk
, skb
);
1699 tp
->copied_seq
= seq
;
1701 tp
->copied_seq
= seq
;
1703 tcp_rcv_space_adjust(sk
);
1705 /* Clean up data we have read: This will do ACK frames. */
1707 tcp_recv_skb(sk
, seq
, &offset
);
1708 tcp_cleanup_rbuf(sk
, copied
);
1712 EXPORT_SYMBOL(tcp_read_sock
);
1714 int tcp_peek_len(struct socket
*sock
)
1716 return tcp_inq(sock
->sk
);
1718 EXPORT_SYMBOL(tcp_peek_len
);
1720 static void tcp_update_recv_tstamps(struct sk_buff
*skb
,
1721 struct scm_timestamping
*tss
)
1724 tss
->ts
[0] = ktime_to_timespec(skb
->tstamp
);
1726 tss
->ts
[0] = (struct timespec
) {0};
1728 if (skb_hwtstamps(skb
)->hwtstamp
)
1729 tss
->ts
[2] = ktime_to_timespec(skb_hwtstamps(skb
)->hwtstamp
);
1731 tss
->ts
[2] = (struct timespec
) {0};
1734 /* Similar to __sock_recv_timestamp, but does not require an skb */
1735 void tcp_recv_timestamp(struct msghdr
*msg
, const struct sock
*sk
,
1736 struct scm_timestamping
*tss
)
1739 bool has_timestamping
= false;
1741 if (tss
->ts
[0].tv_sec
|| tss
->ts
[0].tv_nsec
) {
1742 if (sock_flag(sk
, SOCK_RCVTSTAMP
)) {
1743 if (sock_flag(sk
, SOCK_RCVTSTAMPNS
)) {
1744 put_cmsg(msg
, SOL_SOCKET
, SCM_TIMESTAMPNS
,
1745 sizeof(tss
->ts
[0]), &tss
->ts
[0]);
1747 tv
.tv_sec
= tss
->ts
[0].tv_sec
;
1748 tv
.tv_usec
= tss
->ts
[0].tv_nsec
/ 1000;
1750 put_cmsg(msg
, SOL_SOCKET
, SCM_TIMESTAMP
,
1755 if (sk
->sk_tsflags
& SOF_TIMESTAMPING_SOFTWARE
)
1756 has_timestamping
= true;
1758 tss
->ts
[0] = (struct timespec
) {0};
1761 if (tss
->ts
[2].tv_sec
|| tss
->ts
[2].tv_nsec
) {
1762 if (sk
->sk_tsflags
& SOF_TIMESTAMPING_RAW_HARDWARE
)
1763 has_timestamping
= true;
1765 tss
->ts
[2] = (struct timespec
) {0};
1768 if (has_timestamping
) {
1769 tss
->ts
[1] = (struct timespec
) {0};
1770 put_cmsg(msg
, SOL_SOCKET
, SCM_TIMESTAMPING
,
1776 * This routine copies from a sock struct into the user buffer.
1778 * Technical note: in 2.3 we work on _locked_ socket, so that
1779 * tricks with *seq access order and skb->users are not required.
1780 * Probably, code can be easily improved even more.
1783 int tcp_recvmsg(struct sock
*sk
, struct msghdr
*msg
, size_t len
, int nonblock
,
1784 int flags
, int *addr_len
)
1786 struct tcp_sock
*tp
= tcp_sk(sk
);
1792 int target
; /* Read at least this many bytes */
1794 struct sk_buff
*skb
, *last
;
1796 struct scm_timestamping tss
;
1797 bool has_tss
= false;
1799 if (unlikely(flags
& MSG_ERRQUEUE
))
1800 return inet_recv_error(sk
, msg
, len
, addr_len
);
1802 if (sk_can_busy_loop(sk
) && skb_queue_empty(&sk
->sk_receive_queue
) &&
1803 (sk
->sk_state
== TCP_ESTABLISHED
))
1804 sk_busy_loop(sk
, nonblock
);
1809 if (sk
->sk_state
== TCP_LISTEN
)
1812 timeo
= sock_rcvtimeo(sk
, nonblock
);
1814 /* Urgent data needs to be handled specially. */
1815 if (flags
& MSG_OOB
)
1818 if (unlikely(tp
->repair
)) {
1820 if (!(flags
& MSG_PEEK
))
1823 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
1827 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1830 /* 'common' recv queue MSG_PEEK-ing */
1833 seq
= &tp
->copied_seq
;
1834 if (flags
& MSG_PEEK
) {
1835 peek_seq
= tp
->copied_seq
;
1839 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1844 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1845 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1848 if (signal_pending(current
)) {
1849 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1854 /* Next get a buffer. */
1856 last
= skb_peek_tail(&sk
->sk_receive_queue
);
1857 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1859 /* Now that we have two receive queues this
1862 if (WARN(before(*seq
, TCP_SKB_CB(skb
)->seq
),
1863 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
1864 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
,
1868 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1869 if (unlikely(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
1870 pr_err_once("%s: found a SYN, please report !\n", __func__
);
1873 if (offset
< skb
->len
)
1875 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1877 WARN(!(flags
& MSG_PEEK
),
1878 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
1879 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
, flags
);
1882 /* Well, if we have backlog, try to process it now yet. */
1884 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1889 sk
->sk_state
== TCP_CLOSE
||
1890 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1892 signal_pending(current
))
1895 if (sock_flag(sk
, SOCK_DONE
))
1899 copied
= sock_error(sk
);
1903 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1906 if (sk
->sk_state
== TCP_CLOSE
) {
1907 if (!sock_flag(sk
, SOCK_DONE
)) {
1908 /* This occurs when user tries to read
1909 * from never connected socket.
1922 if (signal_pending(current
)) {
1923 copied
= sock_intr_errno(timeo
);
1928 tcp_cleanup_rbuf(sk
, copied
);
1930 if (copied
>= target
) {
1931 /* Do not sleep, just process backlog. */
1935 sk_wait_data(sk
, &timeo
, last
);
1938 if ((flags
& MSG_PEEK
) &&
1939 (peek_seq
- copied
- urg_hole
!= tp
->copied_seq
)) {
1940 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1942 task_pid_nr(current
));
1943 peek_seq
= tp
->copied_seq
;
1948 /* Ok so how much can we use? */
1949 used
= skb
->len
- offset
;
1953 /* Do we have urgent data here? */
1955 u32 urg_offset
= tp
->urg_seq
- *seq
;
1956 if (urg_offset
< used
) {
1958 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1971 if (!(flags
& MSG_TRUNC
)) {
1972 err
= skb_copy_datagram_msg(skb
, offset
, msg
, used
);
1974 /* Exception. Bailout! */
1985 tcp_rcv_space_adjust(sk
);
1988 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1990 tcp_fast_path_check(sk
);
1992 if (used
+ offset
< skb
->len
)
1995 if (TCP_SKB_CB(skb
)->has_rxtstamp
) {
1996 tcp_update_recv_tstamps(skb
, &tss
);
1999 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
2001 if (!(flags
& MSG_PEEK
))
2002 sk_eat_skb(sk
, skb
);
2006 /* Process the FIN. */
2008 if (!(flags
& MSG_PEEK
))
2009 sk_eat_skb(sk
, skb
);
2013 /* According to UNIX98, msg_name/msg_namelen are ignored
2014 * on connected socket. I was just happy when found this 8) --ANK
2018 tcp_recv_timestamp(msg
, sk
, &tss
);
2020 /* Clean up data we have read: This will do ACK frames. */
2021 tcp_cleanup_rbuf(sk
, copied
);
2031 err
= tcp_recv_urg(sk
, msg
, len
, flags
);
2035 err
= tcp_peek_sndq(sk
, msg
, len
);
2038 EXPORT_SYMBOL(tcp_recvmsg
);
2040 void tcp_set_state(struct sock
*sk
, int state
)
2042 int oldstate
= sk
->sk_state
;
2044 trace_tcp_set_state(sk
, oldstate
, state
);
2047 case TCP_ESTABLISHED
:
2048 if (oldstate
!= TCP_ESTABLISHED
)
2049 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
2053 if (oldstate
== TCP_CLOSE_WAIT
|| oldstate
== TCP_ESTABLISHED
)
2054 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ESTABRESETS
);
2056 sk
->sk_prot
->unhash(sk
);
2057 if (inet_csk(sk
)->icsk_bind_hash
&&
2058 !(sk
->sk_userlocks
& SOCK_BINDPORT_LOCK
))
2062 if (oldstate
== TCP_ESTABLISHED
)
2063 TCP_DEC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
2066 /* Change state AFTER socket is unhashed to avoid closed
2067 * socket sitting in hash tables.
2069 sk_state_store(sk
, state
);
2072 SOCK_DEBUG(sk
, "TCP sk=%p, State %s -> %s\n", sk
, statename
[oldstate
], statename
[state
]);
2075 EXPORT_SYMBOL_GPL(tcp_set_state
);
2078 * State processing on a close. This implements the state shift for
2079 * sending our FIN frame. Note that we only send a FIN for some
2080 * states. A shutdown() may have already sent the FIN, or we may be
2084 static const unsigned char new_state
[16] = {
2085 /* current state: new state: action: */
2086 [0 /* (Invalid) */] = TCP_CLOSE
,
2087 [TCP_ESTABLISHED
] = TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2088 [TCP_SYN_SENT
] = TCP_CLOSE
,
2089 [TCP_SYN_RECV
] = TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2090 [TCP_FIN_WAIT1
] = TCP_FIN_WAIT1
,
2091 [TCP_FIN_WAIT2
] = TCP_FIN_WAIT2
,
2092 [TCP_TIME_WAIT
] = TCP_CLOSE
,
2093 [TCP_CLOSE
] = TCP_CLOSE
,
2094 [TCP_CLOSE_WAIT
] = TCP_LAST_ACK
| TCP_ACTION_FIN
,
2095 [TCP_LAST_ACK
] = TCP_LAST_ACK
,
2096 [TCP_LISTEN
] = TCP_CLOSE
,
2097 [TCP_CLOSING
] = TCP_CLOSING
,
2098 [TCP_NEW_SYN_RECV
] = TCP_CLOSE
, /* should not happen ! */
2101 static int tcp_close_state(struct sock
*sk
)
2103 int next
= (int)new_state
[sk
->sk_state
];
2104 int ns
= next
& TCP_STATE_MASK
;
2106 tcp_set_state(sk
, ns
);
2108 return next
& TCP_ACTION_FIN
;
2112 * Shutdown the sending side of a connection. Much like close except
2113 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2116 void tcp_shutdown(struct sock
*sk
, int how
)
2118 /* We need to grab some memory, and put together a FIN,
2119 * and then put it into the queue to be sent.
2120 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2122 if (!(how
& SEND_SHUTDOWN
))
2125 /* If we've already sent a FIN, or it's a closed state, skip this. */
2126 if ((1 << sk
->sk_state
) &
2127 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
2128 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
2129 /* Clear out any half completed packets. FIN if needed. */
2130 if (tcp_close_state(sk
))
2134 EXPORT_SYMBOL(tcp_shutdown
);
2136 bool tcp_check_oom(struct sock
*sk
, int shift
)
2138 bool too_many_orphans
, out_of_socket_memory
;
2140 too_many_orphans
= tcp_too_many_orphans(sk
, shift
);
2141 out_of_socket_memory
= tcp_out_of_memory(sk
);
2143 if (too_many_orphans
)
2144 net_info_ratelimited("too many orphaned sockets\n");
2145 if (out_of_socket_memory
)
2146 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2147 return too_many_orphans
|| out_of_socket_memory
;
2150 void tcp_close(struct sock
*sk
, long timeout
)
2152 struct sk_buff
*skb
;
2153 int data_was_unread
= 0;
2157 sk
->sk_shutdown
= SHUTDOWN_MASK
;
2159 if (sk
->sk_state
== TCP_LISTEN
) {
2160 tcp_set_state(sk
, TCP_CLOSE
);
2163 inet_csk_listen_stop(sk
);
2165 goto adjudge_to_death
;
2168 /* We need to flush the recv. buffs. We do this only on the
2169 * descriptor close, not protocol-sourced closes, because the
2170 * reader process may not have drained the data yet!
2172 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
2173 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
;
2175 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
2177 data_was_unread
+= len
;
2183 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2184 if (sk
->sk_state
== TCP_CLOSE
)
2185 goto adjudge_to_death
;
2187 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2188 * data was lost. To witness the awful effects of the old behavior of
2189 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2190 * GET in an FTP client, suspend the process, wait for the client to
2191 * advertise a zero window, then kill -9 the FTP client, wheee...
2192 * Note: timeout is always zero in such a case.
2194 if (unlikely(tcp_sk(sk
)->repair
)) {
2195 sk
->sk_prot
->disconnect(sk
, 0);
2196 } else if (data_was_unread
) {
2197 /* Unread data was tossed, zap the connection. */
2198 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONCLOSE
);
2199 tcp_set_state(sk
, TCP_CLOSE
);
2200 tcp_send_active_reset(sk
, sk
->sk_allocation
);
2201 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
2202 /* Check zero linger _after_ checking for unread data. */
2203 sk
->sk_prot
->disconnect(sk
, 0);
2204 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
2205 } else if (tcp_close_state(sk
)) {
2206 /* We FIN if the application ate all the data before
2207 * zapping the connection.
2210 /* RED-PEN. Formally speaking, we have broken TCP state
2211 * machine. State transitions:
2213 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2214 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2215 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2217 * are legal only when FIN has been sent (i.e. in window),
2218 * rather than queued out of window. Purists blame.
2220 * F.e. "RFC state" is ESTABLISHED,
2221 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2223 * The visible declinations are that sometimes
2224 * we enter time-wait state, when it is not required really
2225 * (harmless), do not send active resets, when they are
2226 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2227 * they look as CLOSING or LAST_ACK for Linux)
2228 * Probably, I missed some more holelets.
2230 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2231 * in a single packet! (May consider it later but will
2232 * probably need API support or TCP_CORK SYN-ACK until
2233 * data is written and socket is closed.)
2238 sk_stream_wait_close(sk
, timeout
);
2241 state
= sk
->sk_state
;
2247 /* remove backlog if any, without releasing ownership. */
2250 percpu_counter_inc(sk
->sk_prot
->orphan_count
);
2252 /* Have we already been destroyed by a softirq or backlog? */
2253 if (state
!= TCP_CLOSE
&& sk
->sk_state
== TCP_CLOSE
)
2256 /* This is a (useful) BSD violating of the RFC. There is a
2257 * problem with TCP as specified in that the other end could
2258 * keep a socket open forever with no application left this end.
2259 * We use a 1 minute timeout (about the same as BSD) then kill
2260 * our end. If they send after that then tough - BUT: long enough
2261 * that we won't make the old 4*rto = almost no time - whoops
2264 * Nope, it was not mistake. It is really desired behaviour
2265 * f.e. on http servers, when such sockets are useless, but
2266 * consume significant resources. Let's do it with special
2267 * linger2 option. --ANK
2270 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
2271 struct tcp_sock
*tp
= tcp_sk(sk
);
2272 if (tp
->linger2
< 0) {
2273 tcp_set_state(sk
, TCP_CLOSE
);
2274 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2275 __NET_INC_STATS(sock_net(sk
),
2276 LINUX_MIB_TCPABORTONLINGER
);
2278 const int tmo
= tcp_fin_time(sk
);
2280 if (tmo
> TCP_TIMEWAIT_LEN
) {
2281 inet_csk_reset_keepalive_timer(sk
,
2282 tmo
- TCP_TIMEWAIT_LEN
);
2284 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
2289 if (sk
->sk_state
!= TCP_CLOSE
) {
2291 if (tcp_check_oom(sk
, 0)) {
2292 tcp_set_state(sk
, TCP_CLOSE
);
2293 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2294 __NET_INC_STATS(sock_net(sk
),
2295 LINUX_MIB_TCPABORTONMEMORY
);
2296 } else if (!check_net(sock_net(sk
))) {
2297 /* Not possible to send reset; just close */
2298 tcp_set_state(sk
, TCP_CLOSE
);
2302 if (sk
->sk_state
== TCP_CLOSE
) {
2303 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
2304 /* We could get here with a non-NULL req if the socket is
2305 * aborted (e.g., closed with unread data) before 3WHS
2309 reqsk_fastopen_remove(sk
, req
, false);
2310 inet_csk_destroy_sock(sk
);
2312 /* Otherwise, socket is reprieved until protocol close. */
2320 EXPORT_SYMBOL(tcp_close
);
2322 /* These states need RST on ABORT according to RFC793 */
2324 static inline bool tcp_need_reset(int state
)
2326 return (1 << state
) &
2327 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
2328 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
2331 static void tcp_rtx_queue_purge(struct sock
*sk
)
2333 struct rb_node
*p
= rb_first(&sk
->tcp_rtx_queue
);
2336 struct sk_buff
*skb
= rb_to_skb(p
);
2339 /* Since we are deleting whole queue, no need to
2340 * list_del(&skb->tcp_tsorted_anchor)
2342 tcp_rtx_queue_unlink(skb
, sk
);
2343 sk_wmem_free_skb(sk
, skb
);
2347 void tcp_write_queue_purge(struct sock
*sk
)
2349 struct sk_buff
*skb
;
2351 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
2352 while ((skb
= __skb_dequeue(&sk
->sk_write_queue
)) != NULL
) {
2353 tcp_skb_tsorted_anchor_cleanup(skb
);
2354 sk_wmem_free_skb(sk
, skb
);
2356 tcp_rtx_queue_purge(sk
);
2357 INIT_LIST_HEAD(&tcp_sk(sk
)->tsorted_sent_queue
);
2359 tcp_clear_all_retrans_hints(tcp_sk(sk
));
2360 tcp_sk(sk
)->packets_out
= 0;
2363 int tcp_disconnect(struct sock
*sk
, int flags
)
2365 struct inet_sock
*inet
= inet_sk(sk
);
2366 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2367 struct tcp_sock
*tp
= tcp_sk(sk
);
2369 int old_state
= sk
->sk_state
;
2371 if (old_state
!= TCP_CLOSE
)
2372 tcp_set_state(sk
, TCP_CLOSE
);
2374 /* ABORT function of RFC793 */
2375 if (old_state
== TCP_LISTEN
) {
2376 inet_csk_listen_stop(sk
);
2377 } else if (unlikely(tp
->repair
)) {
2378 sk
->sk_err
= ECONNABORTED
;
2379 } else if (tcp_need_reset(old_state
) ||
2380 (tp
->snd_nxt
!= tp
->write_seq
&&
2381 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
2382 /* The last check adjusts for discrepancy of Linux wrt. RFC
2385 tcp_send_active_reset(sk
, gfp_any());
2386 sk
->sk_err
= ECONNRESET
;
2387 } else if (old_state
== TCP_SYN_SENT
)
2388 sk
->sk_err
= ECONNRESET
;
2390 tcp_clear_xmit_timers(sk
);
2391 __skb_queue_purge(&sk
->sk_receive_queue
);
2392 tcp_write_queue_purge(sk
);
2393 tcp_fastopen_active_disable_ofo_check(sk
);
2394 skb_rbtree_purge(&tp
->out_of_order_queue
);
2396 inet
->inet_dport
= 0;
2398 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
2399 inet_reset_saddr(sk
);
2401 sk
->sk_shutdown
= 0;
2402 sock_reset_flag(sk
, SOCK_DONE
);
2404 tp
->write_seq
+= tp
->max_window
+ 2;
2405 if (tp
->write_seq
== 0)
2407 icsk
->icsk_backoff
= 0;
2409 icsk
->icsk_probes_out
= 0;
2410 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
2411 tp
->snd_cwnd_cnt
= 0;
2412 tp
->window_clamp
= 0;
2413 tcp_set_ca_state(sk
, TCP_CA_Open
);
2414 tp
->is_sack_reneg
= 0;
2415 tcp_clear_retrans(tp
);
2416 inet_csk_delack_init(sk
);
2417 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2418 * issue in __tcp_select_window()
2420 icsk
->icsk_ack
.rcv_mss
= TCP_MIN_MSS
;
2421 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
2423 dst_release(sk
->sk_rx_dst
);
2424 sk
->sk_rx_dst
= NULL
;
2425 tcp_saved_syn_free(tp
);
2427 /* Clean up fastopen related fields */
2428 tcp_free_fastopen_req(tp
);
2429 inet
->defer_connect
= 0;
2431 WARN_ON(inet
->inet_num
&& !icsk
->icsk_bind_hash
);
2433 if (sk
->sk_frag
.page
) {
2434 put_page(sk
->sk_frag
.page
);
2435 sk
->sk_frag
.page
= NULL
;
2436 sk
->sk_frag
.offset
= 0;
2439 sk
->sk_error_report(sk
);
2442 EXPORT_SYMBOL(tcp_disconnect
);
2444 static inline bool tcp_can_repair_sock(const struct sock
*sk
)
2446 return ns_capable(sock_net(sk
)->user_ns
, CAP_NET_ADMIN
) &&
2447 (sk
->sk_state
!= TCP_LISTEN
);
2450 static int tcp_repair_set_window(struct tcp_sock
*tp
, char __user
*optbuf
, int len
)
2452 struct tcp_repair_window opt
;
2457 if (len
!= sizeof(opt
))
2460 if (copy_from_user(&opt
, optbuf
, sizeof(opt
)))
2463 if (opt
.max_window
< opt
.snd_wnd
)
2466 if (after(opt
.snd_wl1
, tp
->rcv_nxt
+ opt
.rcv_wnd
))
2469 if (after(opt
.rcv_wup
, tp
->rcv_nxt
))
2472 tp
->snd_wl1
= opt
.snd_wl1
;
2473 tp
->snd_wnd
= opt
.snd_wnd
;
2474 tp
->max_window
= opt
.max_window
;
2476 tp
->rcv_wnd
= opt
.rcv_wnd
;
2477 tp
->rcv_wup
= opt
.rcv_wup
;
2482 static int tcp_repair_options_est(struct sock
*sk
,
2483 struct tcp_repair_opt __user
*optbuf
, unsigned int len
)
2485 struct tcp_sock
*tp
= tcp_sk(sk
);
2486 struct tcp_repair_opt opt
;
2488 while (len
>= sizeof(opt
)) {
2489 if (copy_from_user(&opt
, optbuf
, sizeof(opt
)))
2495 switch (opt
.opt_code
) {
2497 tp
->rx_opt
.mss_clamp
= opt
.opt_val
;
2502 u16 snd_wscale
= opt
.opt_val
& 0xFFFF;
2503 u16 rcv_wscale
= opt
.opt_val
>> 16;
2505 if (snd_wscale
> TCP_MAX_WSCALE
|| rcv_wscale
> TCP_MAX_WSCALE
)
2508 tp
->rx_opt
.snd_wscale
= snd_wscale
;
2509 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2510 tp
->rx_opt
.wscale_ok
= 1;
2513 case TCPOPT_SACK_PERM
:
2514 if (opt
.opt_val
!= 0)
2517 tp
->rx_opt
.sack_ok
|= TCP_SACK_SEEN
;
2519 case TCPOPT_TIMESTAMP
:
2520 if (opt
.opt_val
!= 0)
2523 tp
->rx_opt
.tstamp_ok
= 1;
2532 * Socket option code for TCP.
2534 static int do_tcp_setsockopt(struct sock
*sk
, int level
,
2535 int optname
, char __user
*optval
, unsigned int optlen
)
2537 struct tcp_sock
*tp
= tcp_sk(sk
);
2538 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2539 struct net
*net
= sock_net(sk
);
2543 /* These are data/string values, all the others are ints */
2545 case TCP_CONGESTION
: {
2546 char name
[TCP_CA_NAME_MAX
];
2551 val
= strncpy_from_user(name
, optval
,
2552 min_t(long, TCP_CA_NAME_MAX
-1, optlen
));
2558 err
= tcp_set_congestion_control(sk
, name
, true, true);
2563 char name
[TCP_ULP_NAME_MAX
];
2568 val
= strncpy_from_user(name
, optval
,
2569 min_t(long, TCP_ULP_NAME_MAX
- 1,
2576 err
= tcp_set_ulp(sk
, name
);
2580 case TCP_FASTOPEN_KEY
: {
2581 __u8 key
[TCP_FASTOPEN_KEY_LENGTH
];
2583 if (optlen
!= sizeof(key
))
2586 if (copy_from_user(key
, optval
, optlen
))
2589 return tcp_fastopen_reset_cipher(net
, sk
, key
, sizeof(key
));
2596 if (optlen
< sizeof(int))
2599 if (get_user(val
, (int __user
*)optval
))
2606 /* Values greater than interface MTU won't take effect. However
2607 * at the point when this call is done we typically don't yet
2608 * know which interface is going to be used
2610 if (val
&& (val
< TCP_MIN_MSS
|| val
> MAX_TCP_WINDOW
)) {
2614 tp
->rx_opt
.user_mss
= val
;
2619 /* TCP_NODELAY is weaker than TCP_CORK, so that
2620 * this option on corked socket is remembered, but
2621 * it is not activated until cork is cleared.
2623 * However, when TCP_NODELAY is set we make
2624 * an explicit push, which overrides even TCP_CORK
2625 * for currently queued segments.
2627 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
2628 tcp_push_pending_frames(sk
);
2630 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
2634 case TCP_THIN_LINEAR_TIMEOUTS
:
2635 if (val
< 0 || val
> 1)
2641 case TCP_THIN_DUPACK
:
2642 if (val
< 0 || val
> 1)
2647 if (!tcp_can_repair_sock(sk
))
2649 else if (val
== 1) {
2651 sk
->sk_reuse
= SK_FORCE_REUSE
;
2652 tp
->repair_queue
= TCP_NO_QUEUE
;
2653 } else if (val
== 0) {
2655 sk
->sk_reuse
= SK_NO_REUSE
;
2656 tcp_send_window_probe(sk
);
2662 case TCP_REPAIR_QUEUE
:
2665 else if ((unsigned int)val
< TCP_QUEUES_NR
)
2666 tp
->repair_queue
= val
;
2672 if (sk
->sk_state
!= TCP_CLOSE
)
2674 else if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2675 tp
->write_seq
= val
;
2676 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2682 case TCP_REPAIR_OPTIONS
:
2685 else if (sk
->sk_state
== TCP_ESTABLISHED
)
2686 err
= tcp_repair_options_est(sk
,
2687 (struct tcp_repair_opt __user
*)optval
,
2694 /* When set indicates to always queue non-full frames.
2695 * Later the user clears this option and we transmit
2696 * any pending partial frames in the queue. This is
2697 * meant to be used alongside sendfile() to get properly
2698 * filled frames when the user (for example) must write
2699 * out headers with a write() call first and then use
2700 * sendfile to send out the data parts.
2702 * TCP_CORK can be set together with TCP_NODELAY and it is
2703 * stronger than TCP_NODELAY.
2706 tp
->nonagle
|= TCP_NAGLE_CORK
;
2708 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
2709 if (tp
->nonagle
&TCP_NAGLE_OFF
)
2710 tp
->nonagle
|= TCP_NAGLE_PUSH
;
2711 tcp_push_pending_frames(sk
);
2716 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
2719 tp
->keepalive_time
= val
* HZ
;
2720 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
2721 !((1 << sk
->sk_state
) &
2722 (TCPF_CLOSE
| TCPF_LISTEN
))) {
2723 u32 elapsed
= keepalive_time_elapsed(tp
);
2724 if (tp
->keepalive_time
> elapsed
)
2725 elapsed
= tp
->keepalive_time
- elapsed
;
2728 inet_csk_reset_keepalive_timer(sk
, elapsed
);
2733 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
2736 tp
->keepalive_intvl
= val
* HZ
;
2739 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
2742 tp
->keepalive_probes
= val
;
2745 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
2748 icsk
->icsk_syn_retries
= val
;
2752 if (val
< 0 || val
> 1)
2761 else if (val
> net
->ipv4
.sysctl_tcp_fin_timeout
/ HZ
)
2764 tp
->linger2
= val
* HZ
;
2767 case TCP_DEFER_ACCEPT
:
2768 /* Translate value in seconds to number of retransmits */
2769 icsk
->icsk_accept_queue
.rskq_defer_accept
=
2770 secs_to_retrans(val
, TCP_TIMEOUT_INIT
/ HZ
,
2774 case TCP_WINDOW_CLAMP
:
2776 if (sk
->sk_state
!= TCP_CLOSE
) {
2780 tp
->window_clamp
= 0;
2782 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
2783 SOCK_MIN_RCVBUF
/ 2 : val
;
2788 icsk
->icsk_ack
.pingpong
= 1;
2790 icsk
->icsk_ack
.pingpong
= 0;
2791 if ((1 << sk
->sk_state
) &
2792 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
2793 inet_csk_ack_scheduled(sk
)) {
2794 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
2795 tcp_cleanup_rbuf(sk
, 1);
2797 icsk
->icsk_ack
.pingpong
= 1;
2802 #ifdef CONFIG_TCP_MD5SIG
2804 case TCP_MD5SIG_EXT
:
2805 if ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
))
2806 err
= tp
->af_specific
->md5_parse(sk
, optname
, optval
, optlen
);
2811 case TCP_USER_TIMEOUT
:
2812 /* Cap the max time in ms TCP will retry or probe the window
2813 * before giving up and aborting (ETIMEDOUT) a connection.
2818 icsk
->icsk_user_timeout
= msecs_to_jiffies(val
);
2822 if (val
>= 0 && ((1 << sk
->sk_state
) & (TCPF_CLOSE
|
2824 tcp_fastopen_init_key_once(net
);
2826 fastopen_queue_tune(sk
, val
);
2831 case TCP_FASTOPEN_CONNECT
:
2832 if (val
> 1 || val
< 0) {
2834 } else if (net
->ipv4
.sysctl_tcp_fastopen
& TFO_CLIENT_ENABLE
) {
2835 if (sk
->sk_state
== TCP_CLOSE
)
2836 tp
->fastopen_connect
= val
;
2843 case TCP_FASTOPEN_NO_COOKIE
:
2844 if (val
> 1 || val
< 0)
2846 else if (!((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
2849 tp
->fastopen_no_cookie
= val
;
2855 tp
->tsoffset
= val
- tcp_time_stamp_raw();
2857 case TCP_REPAIR_WINDOW
:
2858 err
= tcp_repair_set_window(tp
, optval
, optlen
);
2860 case TCP_NOTSENT_LOWAT
:
2861 tp
->notsent_lowat
= val
;
2862 sk
->sk_write_space(sk
);
2873 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2874 unsigned int optlen
)
2876 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2878 if (level
!= SOL_TCP
)
2879 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
2881 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2883 EXPORT_SYMBOL(tcp_setsockopt
);
2885 #ifdef CONFIG_COMPAT
2886 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
2887 char __user
*optval
, unsigned int optlen
)
2889 if (level
!= SOL_TCP
)
2890 return inet_csk_compat_setsockopt(sk
, level
, optname
,
2892 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2894 EXPORT_SYMBOL(compat_tcp_setsockopt
);
2897 static void tcp_get_info_chrono_stats(const struct tcp_sock
*tp
,
2898 struct tcp_info
*info
)
2900 u64 stats
[__TCP_CHRONO_MAX
], total
= 0;
2903 for (i
= TCP_CHRONO_BUSY
; i
< __TCP_CHRONO_MAX
; ++i
) {
2904 stats
[i
] = tp
->chrono_stat
[i
- 1];
2905 if (i
== tp
->chrono_type
)
2906 stats
[i
] += tcp_jiffies32
- tp
->chrono_start
;
2907 stats
[i
] *= USEC_PER_SEC
/ HZ
;
2911 info
->tcpi_busy_time
= total
;
2912 info
->tcpi_rwnd_limited
= stats
[TCP_CHRONO_RWND_LIMITED
];
2913 info
->tcpi_sndbuf_limited
= stats
[TCP_CHRONO_SNDBUF_LIMITED
];
2916 /* Return information about state of tcp endpoint in API format. */
2917 void tcp_get_info(struct sock
*sk
, struct tcp_info
*info
)
2919 const struct tcp_sock
*tp
= tcp_sk(sk
); /* iff sk_type == SOCK_STREAM */
2920 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2926 memset(info
, 0, sizeof(*info
));
2927 if (sk
->sk_type
!= SOCK_STREAM
)
2930 info
->tcpi_state
= sk_state_load(sk
);
2932 /* Report meaningful fields for all TCP states, including listeners */
2933 rate
= READ_ONCE(sk
->sk_pacing_rate
);
2934 rate64
= rate
!= ~0U ? rate
: ~0ULL;
2935 info
->tcpi_pacing_rate
= rate64
;
2937 rate
= READ_ONCE(sk
->sk_max_pacing_rate
);
2938 rate64
= rate
!= ~0U ? rate
: ~0ULL;
2939 info
->tcpi_max_pacing_rate
= rate64
;
2941 info
->tcpi_reordering
= tp
->reordering
;
2942 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
2944 if (info
->tcpi_state
== TCP_LISTEN
) {
2945 /* listeners aliased fields :
2946 * tcpi_unacked -> Number of children ready for accept()
2947 * tcpi_sacked -> max backlog
2949 info
->tcpi_unacked
= sk
->sk_ack_backlog
;
2950 info
->tcpi_sacked
= sk
->sk_max_ack_backlog
;
2954 slow
= lock_sock_fast(sk
);
2956 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
2957 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
2958 info
->tcpi_probes
= icsk
->icsk_probes_out
;
2959 info
->tcpi_backoff
= icsk
->icsk_backoff
;
2961 if (tp
->rx_opt
.tstamp_ok
)
2962 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
2963 if (tcp_is_sack(tp
))
2964 info
->tcpi_options
|= TCPI_OPT_SACK
;
2965 if (tp
->rx_opt
.wscale_ok
) {
2966 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
2967 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
2968 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
2971 if (tp
->ecn_flags
& TCP_ECN_OK
)
2972 info
->tcpi_options
|= TCPI_OPT_ECN
;
2973 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
2974 info
->tcpi_options
|= TCPI_OPT_ECN_SEEN
;
2975 if (tp
->syn_data_acked
)
2976 info
->tcpi_options
|= TCPI_OPT_SYN_DATA
;
2978 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
2979 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
2980 info
->tcpi_snd_mss
= tp
->mss_cache
;
2981 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
2983 info
->tcpi_unacked
= tp
->packets_out
;
2984 info
->tcpi_sacked
= tp
->sacked_out
;
2986 info
->tcpi_lost
= tp
->lost_out
;
2987 info
->tcpi_retrans
= tp
->retrans_out
;
2989 now
= tcp_jiffies32
;
2990 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
2991 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
2992 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
2994 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
2995 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
2996 info
->tcpi_rtt
= tp
->srtt_us
>> 3;
2997 info
->tcpi_rttvar
= tp
->mdev_us
>> 2;
2998 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
2999 info
->tcpi_advmss
= tp
->advmss
;
3001 info
->tcpi_rcv_rtt
= tp
->rcv_rtt_est
.rtt_us
>> 3;
3002 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
3004 info
->tcpi_total_retrans
= tp
->total_retrans
;
3006 info
->tcpi_bytes_acked
= tp
->bytes_acked
;
3007 info
->tcpi_bytes_received
= tp
->bytes_received
;
3008 info
->tcpi_notsent_bytes
= max_t(int, 0, tp
->write_seq
- tp
->snd_nxt
);
3009 tcp_get_info_chrono_stats(tp
, info
);
3011 info
->tcpi_segs_out
= tp
->segs_out
;
3012 info
->tcpi_segs_in
= tp
->segs_in
;
3014 info
->tcpi_min_rtt
= tcp_min_rtt(tp
);
3015 info
->tcpi_data_segs_in
= tp
->data_segs_in
;
3016 info
->tcpi_data_segs_out
= tp
->data_segs_out
;
3018 info
->tcpi_delivery_rate_app_limited
= tp
->rate_app_limited
? 1 : 0;
3019 rate64
= tcp_compute_delivery_rate(tp
);
3021 info
->tcpi_delivery_rate
= rate64
;
3022 unlock_sock_fast(sk
, slow
);
3024 EXPORT_SYMBOL_GPL(tcp_get_info
);
3026 struct sk_buff
*tcp_get_timestamping_opt_stats(const struct sock
*sk
)
3028 const struct tcp_sock
*tp
= tcp_sk(sk
);
3029 struct sk_buff
*stats
;
3030 struct tcp_info info
;
3034 stats
= alloc_skb(7 * nla_total_size_64bit(sizeof(u64
)) +
3035 3 * nla_total_size(sizeof(u32
)) +
3036 2 * nla_total_size(sizeof(u8
)), GFP_ATOMIC
);
3040 tcp_get_info_chrono_stats(tp
, &info
);
3041 nla_put_u64_64bit(stats
, TCP_NLA_BUSY
,
3042 info
.tcpi_busy_time
, TCP_NLA_PAD
);
3043 nla_put_u64_64bit(stats
, TCP_NLA_RWND_LIMITED
,
3044 info
.tcpi_rwnd_limited
, TCP_NLA_PAD
);
3045 nla_put_u64_64bit(stats
, TCP_NLA_SNDBUF_LIMITED
,
3046 info
.tcpi_sndbuf_limited
, TCP_NLA_PAD
);
3047 nla_put_u64_64bit(stats
, TCP_NLA_DATA_SEGS_OUT
,
3048 tp
->data_segs_out
, TCP_NLA_PAD
);
3049 nla_put_u64_64bit(stats
, TCP_NLA_TOTAL_RETRANS
,
3050 tp
->total_retrans
, TCP_NLA_PAD
);
3052 rate
= READ_ONCE(sk
->sk_pacing_rate
);
3053 rate64
= rate
!= ~0U ? rate
: ~0ULL;
3054 nla_put_u64_64bit(stats
, TCP_NLA_PACING_RATE
, rate64
, TCP_NLA_PAD
);
3056 rate64
= tcp_compute_delivery_rate(tp
);
3057 nla_put_u64_64bit(stats
, TCP_NLA_DELIVERY_RATE
, rate64
, TCP_NLA_PAD
);
3059 nla_put_u32(stats
, TCP_NLA_SND_CWND
, tp
->snd_cwnd
);
3060 nla_put_u32(stats
, TCP_NLA_REORDERING
, tp
->reordering
);
3061 nla_put_u32(stats
, TCP_NLA_MIN_RTT
, tcp_min_rtt(tp
));
3063 nla_put_u8(stats
, TCP_NLA_RECUR_RETRANS
, inet_csk(sk
)->icsk_retransmits
);
3064 nla_put_u8(stats
, TCP_NLA_DELIVERY_RATE_APP_LMT
, !!tp
->rate_app_limited
);
3068 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
3069 int optname
, char __user
*optval
, int __user
*optlen
)
3071 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3072 struct tcp_sock
*tp
= tcp_sk(sk
);
3073 struct net
*net
= sock_net(sk
);
3076 if (get_user(len
, optlen
))
3079 len
= min_t(unsigned int, len
, sizeof(int));
3086 val
= tp
->mss_cache
;
3087 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
3088 val
= tp
->rx_opt
.user_mss
;
3090 val
= tp
->rx_opt
.mss_clamp
;
3093 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
3096 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
3099 val
= keepalive_time_when(tp
) / HZ
;
3102 val
= keepalive_intvl_when(tp
) / HZ
;
3105 val
= keepalive_probes(tp
);
3108 val
= icsk
->icsk_syn_retries
? : net
->ipv4
.sysctl_tcp_syn_retries
;
3113 val
= (val
? : net
->ipv4
.sysctl_tcp_fin_timeout
) / HZ
;
3115 case TCP_DEFER_ACCEPT
:
3116 val
= retrans_to_secs(icsk
->icsk_accept_queue
.rskq_defer_accept
,
3117 TCP_TIMEOUT_INIT
/ HZ
, TCP_RTO_MAX
/ HZ
);
3119 case TCP_WINDOW_CLAMP
:
3120 val
= tp
->window_clamp
;
3123 struct tcp_info info
;
3125 if (get_user(len
, optlen
))
3128 tcp_get_info(sk
, &info
);
3130 len
= min_t(unsigned int, len
, sizeof(info
));
3131 if (put_user(len
, optlen
))
3133 if (copy_to_user(optval
, &info
, len
))
3138 const struct tcp_congestion_ops
*ca_ops
;
3139 union tcp_cc_info info
;
3143 if (get_user(len
, optlen
))
3146 ca_ops
= icsk
->icsk_ca_ops
;
3147 if (ca_ops
&& ca_ops
->get_info
)
3148 sz
= ca_ops
->get_info(sk
, ~0U, &attr
, &info
);
3150 len
= min_t(unsigned int, len
, sz
);
3151 if (put_user(len
, optlen
))
3153 if (copy_to_user(optval
, &info
, len
))
3158 val
= !icsk
->icsk_ack
.pingpong
;
3161 case TCP_CONGESTION
:
3162 if (get_user(len
, optlen
))
3164 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
3165 if (put_user(len
, optlen
))
3167 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
3172 if (get_user(len
, optlen
))
3174 len
= min_t(unsigned int, len
, TCP_ULP_NAME_MAX
);
3175 if (!icsk
->icsk_ulp_ops
) {
3176 if (put_user(0, optlen
))
3180 if (put_user(len
, optlen
))
3182 if (copy_to_user(optval
, icsk
->icsk_ulp_ops
->name
, len
))
3186 case TCP_FASTOPEN_KEY
: {
3187 __u8 key
[TCP_FASTOPEN_KEY_LENGTH
];
3188 struct tcp_fastopen_context
*ctx
;
3190 if (get_user(len
, optlen
))
3194 ctx
= rcu_dereference(icsk
->icsk_accept_queue
.fastopenq
.ctx
);
3196 memcpy(key
, ctx
->key
, sizeof(key
));
3201 len
= min_t(unsigned int, len
, sizeof(key
));
3202 if (put_user(len
, optlen
))
3204 if (copy_to_user(optval
, key
, len
))
3208 case TCP_THIN_LINEAR_TIMEOUTS
:
3212 case TCP_THIN_DUPACK
:
3220 case TCP_REPAIR_QUEUE
:
3222 val
= tp
->repair_queue
;
3227 case TCP_REPAIR_WINDOW
: {
3228 struct tcp_repair_window opt
;
3230 if (get_user(len
, optlen
))
3233 if (len
!= sizeof(opt
))
3239 opt
.snd_wl1
= tp
->snd_wl1
;
3240 opt
.snd_wnd
= tp
->snd_wnd
;
3241 opt
.max_window
= tp
->max_window
;
3242 opt
.rcv_wnd
= tp
->rcv_wnd
;
3243 opt
.rcv_wup
= tp
->rcv_wup
;
3245 if (copy_to_user(optval
, &opt
, len
))
3250 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
3251 val
= tp
->write_seq
;
3252 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
3258 case TCP_USER_TIMEOUT
:
3259 val
= jiffies_to_msecs(icsk
->icsk_user_timeout
);
3263 val
= icsk
->icsk_accept_queue
.fastopenq
.max_qlen
;
3266 case TCP_FASTOPEN_CONNECT
:
3267 val
= tp
->fastopen_connect
;
3270 case TCP_FASTOPEN_NO_COOKIE
:
3271 val
= tp
->fastopen_no_cookie
;
3275 val
= tcp_time_stamp_raw() + tp
->tsoffset
;
3277 case TCP_NOTSENT_LOWAT
:
3278 val
= tp
->notsent_lowat
;
3283 case TCP_SAVED_SYN
: {
3284 if (get_user(len
, optlen
))
3288 if (tp
->saved_syn
) {
3289 if (len
< tp
->saved_syn
[0]) {
3290 if (put_user(tp
->saved_syn
[0], optlen
)) {
3297 len
= tp
->saved_syn
[0];
3298 if (put_user(len
, optlen
)) {
3302 if (copy_to_user(optval
, tp
->saved_syn
+ 1, len
)) {
3306 tcp_saved_syn_free(tp
);
3311 if (put_user(len
, optlen
))
3317 return -ENOPROTOOPT
;
3320 if (put_user(len
, optlen
))
3322 if (copy_to_user(optval
, &val
, len
))
3327 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
3330 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3332 if (level
!= SOL_TCP
)
3333 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
3335 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
3337 EXPORT_SYMBOL(tcp_getsockopt
);
3339 #ifdef CONFIG_COMPAT
3340 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
3341 char __user
*optval
, int __user
*optlen
)
3343 if (level
!= SOL_TCP
)
3344 return inet_csk_compat_getsockopt(sk
, level
, optname
,
3346 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
3348 EXPORT_SYMBOL(compat_tcp_getsockopt
);
3351 #ifdef CONFIG_TCP_MD5SIG
3352 static DEFINE_PER_CPU(struct tcp_md5sig_pool
, tcp_md5sig_pool
);
3353 static DEFINE_MUTEX(tcp_md5sig_mutex
);
3354 static bool tcp_md5sig_pool_populated
= false;
3356 static void __tcp_alloc_md5sig_pool(void)
3358 struct crypto_ahash
*hash
;
3361 hash
= crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC
);
3365 for_each_possible_cpu(cpu
) {
3366 void *scratch
= per_cpu(tcp_md5sig_pool
, cpu
).scratch
;
3367 struct ahash_request
*req
;
3370 scratch
= kmalloc_node(sizeof(union tcp_md5sum_block
) +
3371 sizeof(struct tcphdr
),
3376 per_cpu(tcp_md5sig_pool
, cpu
).scratch
= scratch
;
3378 if (per_cpu(tcp_md5sig_pool
, cpu
).md5_req
)
3381 req
= ahash_request_alloc(hash
, GFP_KERNEL
);
3385 ahash_request_set_callback(req
, 0, NULL
, NULL
);
3387 per_cpu(tcp_md5sig_pool
, cpu
).md5_req
= req
;
3389 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3390 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3393 tcp_md5sig_pool_populated
= true;
3396 bool tcp_alloc_md5sig_pool(void)
3398 if (unlikely(!tcp_md5sig_pool_populated
)) {
3399 mutex_lock(&tcp_md5sig_mutex
);
3401 if (!tcp_md5sig_pool_populated
)
3402 __tcp_alloc_md5sig_pool();
3404 mutex_unlock(&tcp_md5sig_mutex
);
3406 return tcp_md5sig_pool_populated
;
3408 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
3412 * tcp_get_md5sig_pool - get md5sig_pool for this user
3414 * We use percpu structure, so if we succeed, we exit with preemption
3415 * and BH disabled, to make sure another thread or softirq handling
3416 * wont try to get same context.
3418 struct tcp_md5sig_pool
*tcp_get_md5sig_pool(void)
3422 if (tcp_md5sig_pool_populated
) {
3423 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3425 return this_cpu_ptr(&tcp_md5sig_pool
);
3430 EXPORT_SYMBOL(tcp_get_md5sig_pool
);
3432 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool
*hp
,
3433 const struct sk_buff
*skb
, unsigned int header_len
)
3435 struct scatterlist sg
;
3436 const struct tcphdr
*tp
= tcp_hdr(skb
);
3437 struct ahash_request
*req
= hp
->md5_req
;
3439 const unsigned int head_data_len
= skb_headlen(skb
) > header_len
?
3440 skb_headlen(skb
) - header_len
: 0;
3441 const struct skb_shared_info
*shi
= skb_shinfo(skb
);
3442 struct sk_buff
*frag_iter
;
3444 sg_init_table(&sg
, 1);
3446 sg_set_buf(&sg
, ((u8
*) tp
) + header_len
, head_data_len
);
3447 ahash_request_set_crypt(req
, &sg
, NULL
, head_data_len
);
3448 if (crypto_ahash_update(req
))
3451 for (i
= 0; i
< shi
->nr_frags
; ++i
) {
3452 const struct skb_frag_struct
*f
= &shi
->frags
[i
];
3453 unsigned int offset
= f
->page_offset
;
3454 struct page
*page
= skb_frag_page(f
) + (offset
>> PAGE_SHIFT
);
3456 sg_set_page(&sg
, page
, skb_frag_size(f
),
3457 offset_in_page(offset
));
3458 ahash_request_set_crypt(req
, &sg
, NULL
, skb_frag_size(f
));
3459 if (crypto_ahash_update(req
))
3463 skb_walk_frags(skb
, frag_iter
)
3464 if (tcp_md5_hash_skb_data(hp
, frag_iter
, 0))
3469 EXPORT_SYMBOL(tcp_md5_hash_skb_data
);
3471 int tcp_md5_hash_key(struct tcp_md5sig_pool
*hp
, const struct tcp_md5sig_key
*key
)
3473 struct scatterlist sg
;
3475 sg_init_one(&sg
, key
->key
, key
->keylen
);
3476 ahash_request_set_crypt(hp
->md5_req
, &sg
, NULL
, key
->keylen
);
3477 return crypto_ahash_update(hp
->md5_req
);
3479 EXPORT_SYMBOL(tcp_md5_hash_key
);
3483 void tcp_done(struct sock
*sk
)
3485 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
3487 if (sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
3488 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ATTEMPTFAILS
);
3490 tcp_set_state(sk
, TCP_CLOSE
);
3491 tcp_clear_xmit_timers(sk
);
3493 reqsk_fastopen_remove(sk
, req
, false);
3495 sk
->sk_shutdown
= SHUTDOWN_MASK
;
3497 if (!sock_flag(sk
, SOCK_DEAD
))
3498 sk
->sk_state_change(sk
);
3500 inet_csk_destroy_sock(sk
);
3502 EXPORT_SYMBOL_GPL(tcp_done
);
3504 int tcp_abort(struct sock
*sk
, int err
)
3506 if (!sk_fullsock(sk
)) {
3507 if (sk
->sk_state
== TCP_NEW_SYN_RECV
) {
3508 struct request_sock
*req
= inet_reqsk(sk
);
3511 inet_csk_reqsk_queue_drop(req
->rsk_listener
, req
);
3518 /* Don't race with userspace socket closes such as tcp_close. */
3521 if (sk
->sk_state
== TCP_LISTEN
) {
3522 tcp_set_state(sk
, TCP_CLOSE
);
3523 inet_csk_listen_stop(sk
);
3526 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3530 if (!sock_flag(sk
, SOCK_DEAD
)) {
3532 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3534 sk
->sk_error_report(sk
);
3535 if (tcp_need_reset(sk
->sk_state
))
3536 tcp_send_active_reset(sk
, GFP_ATOMIC
);
3542 tcp_write_queue_purge(sk
);
3546 EXPORT_SYMBOL_GPL(tcp_abort
);
3548 extern struct tcp_congestion_ops tcp_reno
;
3550 static __initdata
unsigned long thash_entries
;
3551 static int __init
set_thash_entries(char *str
)
3558 ret
= kstrtoul(str
, 0, &thash_entries
);
3564 __setup("thash_entries=", set_thash_entries
);
3566 static void __init
tcp_init_mem(void)
3568 unsigned long limit
= nr_free_buffer_pages() / 16;
3570 limit
= max(limit
, 128UL);
3571 sysctl_tcp_mem
[0] = limit
/ 4 * 3; /* 4.68 % */
3572 sysctl_tcp_mem
[1] = limit
; /* 6.25 % */
3573 sysctl_tcp_mem
[2] = sysctl_tcp_mem
[0] * 2; /* 9.37 % */
3576 void __init
tcp_init(void)
3578 int max_rshare
, max_wshare
, cnt
;
3579 unsigned long limit
;
3582 BUILD_BUG_ON(TCP_MIN_SND_MSS
<= MAX_TCP_OPTION_SPACE
);
3583 BUILD_BUG_ON(sizeof(struct tcp_skb_cb
) >
3584 FIELD_SIZEOF(struct sk_buff
, cb
));
3586 percpu_counter_init(&tcp_sockets_allocated
, 0, GFP_KERNEL
);
3587 percpu_counter_init(&tcp_orphan_count
, 0, GFP_KERNEL
);
3588 inet_hashinfo_init(&tcp_hashinfo
);
3589 tcp_hashinfo
.bind_bucket_cachep
=
3590 kmem_cache_create("tcp_bind_bucket",
3591 sizeof(struct inet_bind_bucket
), 0,
3592 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
);
3594 /* Size and allocate the main established and bind bucket
3597 * The methodology is similar to that of the buffer cache.
3599 tcp_hashinfo
.ehash
=
3600 alloc_large_system_hash("TCP established",
3601 sizeof(struct inet_ehash_bucket
),
3603 17, /* one slot per 128 KB of memory */
3606 &tcp_hashinfo
.ehash_mask
,
3608 thash_entries
? 0 : 512 * 1024);
3609 for (i
= 0; i
<= tcp_hashinfo
.ehash_mask
; i
++)
3610 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].chain
, i
);
3612 if (inet_ehash_locks_alloc(&tcp_hashinfo
))
3613 panic("TCP: failed to alloc ehash_locks");
3614 tcp_hashinfo
.bhash
=
3615 alloc_large_system_hash("TCP bind",
3616 sizeof(struct inet_bind_hashbucket
),
3617 tcp_hashinfo
.ehash_mask
+ 1,
3618 17, /* one slot per 128 KB of memory */
3620 &tcp_hashinfo
.bhash_size
,
3624 tcp_hashinfo
.bhash_size
= 1U << tcp_hashinfo
.bhash_size
;
3625 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
3626 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
3627 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
3631 cnt
= tcp_hashinfo
.ehash_mask
+ 1;
3632 sysctl_tcp_max_orphans
= cnt
/ 2;
3635 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3636 limit
= nr_free_buffer_pages() << (PAGE_SHIFT
- 7);
3637 max_wshare
= min(4UL*1024*1024, limit
);
3638 max_rshare
= min(6UL*1024*1024, limit
);
3640 init_net
.ipv4
.sysctl_tcp_wmem
[0] = SK_MEM_QUANTUM
;
3641 init_net
.ipv4
.sysctl_tcp_wmem
[1] = 16*1024;
3642 init_net
.ipv4
.sysctl_tcp_wmem
[2] = max(64*1024, max_wshare
);
3644 init_net
.ipv4
.sysctl_tcp_rmem
[0] = SK_MEM_QUANTUM
;
3645 init_net
.ipv4
.sysctl_tcp_rmem
[1] = 87380;
3646 init_net
.ipv4
.sysctl_tcp_rmem
[2] = max(87380, max_rshare
);
3648 pr_info("Hash tables configured (established %u bind %u)\n",
3649 tcp_hashinfo
.ehash_mask
+ 1, tcp_hashinfo
.bhash_size
);
3653 BUG_ON(tcp_register_congestion_control(&tcp_reno
) != 0);