1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Alan Cox : Numerous verify_area() calls
23 * Alan Cox : Set the ACK bit on a reset
24 * Alan Cox : Stopped it crashing if it closed while
25 * sk->inuse=1 and was trying to connect
27 * Alan Cox : All icmp error handling was broken
28 * pointers passed where wrong and the
29 * socket was looked up backwards. Nobody
30 * tested any icmp error code obviously.
31 * Alan Cox : tcp_err() now handled properly. It
32 * wakes people on errors. poll
33 * behaves and the icmp error race
34 * has gone by moving it into sock.c
35 * Alan Cox : tcp_send_reset() fixed to work for
36 * everything not just packets for
38 * Alan Cox : tcp option processing.
39 * Alan Cox : Reset tweaked (still not 100%) [Had
41 * Herp Rosmanith : More reset fixes
42 * Alan Cox : No longer acks invalid rst frames.
43 * Acking any kind of RST is right out.
44 * Alan Cox : Sets an ignore me flag on an rst
45 * receive otherwise odd bits of prattle
47 * Alan Cox : Fixed another acking RST frame bug.
48 * Should stop LAN workplace lockups.
49 * Alan Cox : Some tidyups using the new skb list
51 * Alan Cox : sk->keepopen now seems to work
52 * Alan Cox : Pulls options out correctly on accepts
53 * Alan Cox : Fixed assorted sk->rqueue->next errors
54 * Alan Cox : PSH doesn't end a TCP read. Switched a
56 * Alan Cox : Tidied tcp_data to avoid a potential
58 * Alan Cox : Added some better commenting, as the
59 * tcp is hard to follow
60 * Alan Cox : Removed incorrect check for 20 * psh
61 * Michael O'Reilly : ack < copied bug fix.
62 * Johannes Stille : Misc tcp fixes (not all in yet).
63 * Alan Cox : FIN with no memory -> CRASH
64 * Alan Cox : Added socket option proto entries.
65 * Also added awareness of them to accept.
66 * Alan Cox : Added TCP options (SOL_TCP)
67 * Alan Cox : Switched wakeup calls to callbacks,
68 * so the kernel can layer network
70 * Alan Cox : Use ip_tos/ip_ttl settings.
71 * Alan Cox : Handle FIN (more) properly (we hope).
72 * Alan Cox : RST frames sent on unsynchronised
74 * Alan Cox : Put in missing check for SYN bit.
75 * Alan Cox : Added tcp_select_window() aka NET2E
76 * window non shrink trick.
77 * Alan Cox : Added a couple of small NET2E timer
79 * Charles Hedrick : TCP fixes
80 * Toomas Tamm : TCP window fixes
81 * Alan Cox : Small URG fix to rlogin ^C ack fight
82 * Charles Hedrick : Rewrote most of it to actually work
83 * Linus : Rewrote tcp_read() and URG handling
85 * Gerhard Koerting: Fixed some missing timer handling
86 * Matthew Dillon : Reworked TCP machine states as per RFC
87 * Gerhard Koerting: PC/TCP workarounds
88 * Adam Caldwell : Assorted timer/timing errors
89 * Matthew Dillon : Fixed another RST bug
90 * Alan Cox : Move to kernel side addressing changes.
91 * Alan Cox : Beginning work on TCP fastpathing
93 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
94 * Alan Cox : TCP fast path debugging
95 * Alan Cox : Window clamping
96 * Michael Riepe : Bug in tcp_check()
97 * Matt Dillon : More TCP improvements and RST bug fixes
98 * Matt Dillon : Yet more small nasties remove from the
99 * TCP code (Be very nice to this man if
100 * tcp finally works 100%) 8)
101 * Alan Cox : BSD accept semantics.
102 * Alan Cox : Reset on closedown bug.
103 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
104 * Michael Pall : Handle poll() after URG properly in
106 * Michael Pall : Undo the last fix in tcp_read_urg()
107 * (multi URG PUSH broke rlogin).
108 * Michael Pall : Fix the multi URG PUSH problem in
109 * tcp_readable(), poll() after URG
111 * Michael Pall : recv(...,MSG_OOB) never blocks in the
113 * Alan Cox : Changed the semantics of sk->socket to
114 * fix a race and a signal problem with
115 * accept() and async I/O.
116 * Alan Cox : Relaxed the rules on tcp_sendto().
117 * Yury Shevchuk : Really fixed accept() blocking problem.
118 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
119 * clients/servers which listen in on
121 * Alan Cox : Cleaned the above up and shrank it to
122 * a sensible code size.
123 * Alan Cox : Self connect lockup fix.
124 * Alan Cox : No connect to multicast.
125 * Ross Biro : Close unaccepted children on master
127 * Alan Cox : Reset tracing code.
128 * Alan Cox : Spurious resets on shutdown.
129 * Alan Cox : Giant 15 minute/60 second timer error
130 * Alan Cox : Small whoops in polling before an
132 * Alan Cox : Kept the state trace facility since
133 * it's handy for debugging.
134 * Alan Cox : More reset handler fixes.
135 * Alan Cox : Started rewriting the code based on
136 * the RFC's for other useful protocol
137 * references see: Comer, KA9Q NOS, and
138 * for a reference on the difference
139 * between specifications and how BSD
140 * works see the 4.4lite source.
141 * A.N.Kuznetsov : Don't time wait on completion of tidy
143 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
144 * Linus Torvalds : Fixed BSD port reuse to work first syn
145 * Alan Cox : Reimplemented timers as per the RFC
146 * and using multiple timers for sanity.
147 * Alan Cox : Small bug fixes, and a lot of new
149 * Alan Cox : Fixed dual reader crash by locking
150 * the buffers (much like datagram.c)
151 * Alan Cox : Fixed stuck sockets in probe. A probe
152 * now gets fed up of retrying without
153 * (even a no space) answer.
154 * Alan Cox : Extracted closing code better
155 * Alan Cox : Fixed the closing state machine to
157 * Alan Cox : More 'per spec' fixes.
158 * Jorge Cwik : Even faster checksumming.
159 * Alan Cox : tcp_data() doesn't ack illegal PSH
160 * only frames. At least one pc tcp stack
162 * Alan Cox : Cache last socket.
163 * Alan Cox : Per route irtt.
164 * Matt Day : poll()->select() match BSD precisely on error
165 * Alan Cox : New buffers
166 * Marc Tamsky : Various sk->prot->retransmits and
167 * sk->retransmits misupdating fixed.
168 * Fixed tcp_write_timeout: stuck close,
169 * and TCP syn retries gets used now.
170 * Mark Yarvis : In tcp_read_wakeup(), don't send an
171 * ack if state is TCP_CLOSED.
172 * Alan Cox : Look up device on a retransmit - routes may
173 * change. Doesn't yet cope with MSS shrink right
175 * Marc Tamsky : Closing in closing fixes.
176 * Mike Shaver : RFC1122 verifications.
177 * Alan Cox : rcv_saddr errors.
178 * Alan Cox : Block double connect().
179 * Alan Cox : Small hooks for enSKIP.
180 * Alexey Kuznetsov: Path MTU discovery.
181 * Alan Cox : Support soft errors.
182 * Alan Cox : Fix MTU discovery pathological case
183 * when the remote claims no mtu!
184 * Marc Tamsky : TCP_CLOSE fix.
185 * Colin (G3TNE) : Send a reset on syn ack replies in
186 * window but wrong (fixes NT lpd problems)
187 * Pedro Roque : Better TCP window handling, delayed ack.
188 * Joerg Reuter : No modification of locked buffers in
189 * tcp_do_retransmit()
190 * Eric Schenk : Changed receiver side silly window
191 * avoidance algorithm to BSD style
192 * algorithm. This doubles throughput
193 * against machines running Solaris,
194 * and seems to result in general
196 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
197 * Willy Konynenberg : Transparent proxying support.
198 * Mike McLagan : Routing by source
199 * Keith Owens : Do proper merging with partial SKB's in
200 * tcp_do_sendmsg to avoid burstiness.
201 * Eric Schenk : Fix fast close down bug with
202 * shutdown() followed by close().
203 * Andi Kleen : Make poll agree with SIGIO
204 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
205 * lingertime == 0 (RFC 793 ABORT Call)
206 * Hirokazu Takahashi : Use copy_from_user() instead of
207 * csum_and_copy_from_user() if possible.
209 * Description of States:
211 * TCP_SYN_SENT sent a connection request, waiting for ack
213 * TCP_SYN_RECV received a connection request, sent ack,
214 * waiting for final ack in three-way handshake.
216 * TCP_ESTABLISHED connection established
218 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
219 * transmission of remaining buffered data
221 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
224 * TCP_CLOSING both sides have shutdown but we still have
225 * data we have to finish sending
227 * TCP_TIME_WAIT timeout to catch resent junk before entering
228 * closed, can only be entered from FIN_WAIT2
229 * or CLOSING. Required because the other end
230 * may not have gotten our last ACK causing it
231 * to retransmit the data packet (which we ignore)
233 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
234 * us to finish writing our data and to shutdown
235 * (we have to close() to move on to LAST_ACK)
237 * TCP_LAST_ACK out side has shutdown after remote has
238 * shutdown. There may still be data in our
239 * buffer that we have to finish sending
241 * TCP_CLOSE socket is finished
244 #define pr_fmt(fmt) "TCP: " fmt
246 #include <crypto/hash.h>
247 #include <linux/kernel.h>
248 #include <linux/module.h>
249 #include <linux/types.h>
250 #include <linux/fcntl.h>
251 #include <linux/poll.h>
252 #include <linux/inet_diag.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/memblock.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/time.h>
267 #include <linux/slab.h>
268 #include <linux/errqueue.h>
269 #include <linux/static_key.h>
271 #include <net/icmp.h>
272 #include <net/inet_common.h>
274 #include <net/xfrm.h>
276 #include <net/sock.h>
278 #include <linux/uaccess.h>
279 #include <asm/ioctls.h>
280 #include <net/busy_poll.h>
282 struct percpu_counter tcp_orphan_count
;
283 EXPORT_SYMBOL_GPL(tcp_orphan_count
);
285 long sysctl_tcp_mem
[3] __read_mostly
;
286 EXPORT_SYMBOL(sysctl_tcp_mem
);
288 atomic_long_t tcp_memory_allocated
; /* Current allocated memory. */
289 EXPORT_SYMBOL(tcp_memory_allocated
);
291 #if IS_ENABLED(CONFIG_SMC)
292 DEFINE_STATIC_KEY_FALSE(tcp_have_smc
);
293 EXPORT_SYMBOL(tcp_have_smc
);
297 * Current number of TCP sockets.
299 struct percpu_counter tcp_sockets_allocated
;
300 EXPORT_SYMBOL(tcp_sockets_allocated
);
305 struct tcp_splice_state
{
306 struct pipe_inode_info
*pipe
;
312 * Pressure flag: try to collapse.
313 * Technical note: it is used by multiple contexts non atomically.
314 * All the __sk_mem_schedule() is of this nature: accounting
315 * is strict, actions are advisory and have some latency.
317 unsigned long tcp_memory_pressure __read_mostly
;
318 EXPORT_SYMBOL_GPL(tcp_memory_pressure
);
320 DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key
);
321 EXPORT_SYMBOL(tcp_rx_skb_cache_key
);
323 DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key
);
325 void tcp_enter_memory_pressure(struct sock
*sk
)
329 if (tcp_memory_pressure
)
335 if (!cmpxchg(&tcp_memory_pressure
, 0, val
))
336 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMEMORYPRESSURES
);
338 EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure
);
340 void tcp_leave_memory_pressure(struct sock
*sk
)
344 if (!tcp_memory_pressure
)
346 val
= xchg(&tcp_memory_pressure
, 0);
348 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPMEMORYPRESSURESCHRONO
,
349 jiffies_to_msecs(jiffies
- val
));
351 EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure
);
353 /* Convert seconds to retransmits based on initial and max timeout */
354 static u8
secs_to_retrans(int seconds
, int timeout
, int rto_max
)
359 int period
= timeout
;
362 while (seconds
> period
&& res
< 255) {
365 if (timeout
> rto_max
)
373 /* Convert retransmits to seconds based on initial and max timeout */
374 static int retrans_to_secs(u8 retrans
, int timeout
, int rto_max
)
382 if (timeout
> rto_max
)
390 static u64
tcp_compute_delivery_rate(const struct tcp_sock
*tp
)
392 u32 rate
= READ_ONCE(tp
->rate_delivered
);
393 u32 intv
= READ_ONCE(tp
->rate_interval_us
);
397 rate64
= (u64
)rate
* tp
->mss_cache
* USEC_PER_SEC
;
398 do_div(rate64
, intv
);
403 /* Address-family independent initialization for a tcp_sock.
405 * NOTE: A lot of things set to zero explicitly by call to
406 * sk_alloc() so need not be done here.
408 void tcp_init_sock(struct sock
*sk
)
410 struct inet_connection_sock
*icsk
= inet_csk(sk
);
411 struct tcp_sock
*tp
= tcp_sk(sk
);
413 tp
->out_of_order_queue
= RB_ROOT
;
414 sk
->tcp_rtx_queue
= RB_ROOT
;
415 tcp_init_xmit_timers(sk
);
416 INIT_LIST_HEAD(&tp
->tsq_node
);
417 INIT_LIST_HEAD(&tp
->tsorted_sent_queue
);
419 icsk
->icsk_rto
= TCP_TIMEOUT_INIT
;
420 tp
->mdev_us
= jiffies_to_usecs(TCP_TIMEOUT_INIT
);
421 minmax_reset(&tp
->rtt_min
, tcp_jiffies32
, ~0U);
423 /* So many TCP implementations out there (incorrectly) count the
424 * initial SYN frame in their delayed-ACK and congestion control
425 * algorithms that we must have the following bandaid to talk
426 * efficiently to them. -DaveM
428 tp
->snd_cwnd
= TCP_INIT_CWND
;
430 /* There's a bubble in the pipe until at least the first ACK. */
431 tp
->app_limited
= ~0U;
433 /* See draft-stevens-tcpca-spec-01 for discussion of the
434 * initialization of these values.
436 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
437 tp
->snd_cwnd_clamp
= ~0;
438 tp
->mss_cache
= TCP_MSS_DEFAULT
;
440 tp
->reordering
= sock_net(sk
)->ipv4
.sysctl_tcp_reordering
;
441 tcp_assign_congestion_control(sk
);
444 tp
->rack
.reo_wnd_steps
= 1;
446 sk
->sk_state
= TCP_CLOSE
;
448 sk
->sk_write_space
= sk_stream_write_space
;
449 sock_set_flag(sk
, SOCK_USE_WRITE_QUEUE
);
451 icsk
->icsk_sync_mss
= tcp_sync_mss
;
453 sk
->sk_sndbuf
= sock_net(sk
)->ipv4
.sysctl_tcp_wmem
[1];
454 sk
->sk_rcvbuf
= sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[1];
456 sk_sockets_allocated_inc(sk
);
457 sk
->sk_route_forced_caps
= NETIF_F_GSO
;
459 EXPORT_SYMBOL(tcp_init_sock
);
461 static void tcp_tx_timestamp(struct sock
*sk
, u16 tsflags
)
463 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
465 if (tsflags
&& skb
) {
466 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
467 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
469 sock_tx_timestamp(sk
, tsflags
, &shinfo
->tx_flags
);
470 if (tsflags
& SOF_TIMESTAMPING_TX_ACK
)
471 tcb
->txstamp_ack
= 1;
472 if (tsflags
& SOF_TIMESTAMPING_TX_RECORD_MASK
)
473 shinfo
->tskey
= TCP_SKB_CB(skb
)->seq
+ skb
->len
- 1;
477 static inline bool tcp_stream_is_readable(const struct tcp_sock
*tp
,
478 int target
, struct sock
*sk
)
480 return (tp
->rcv_nxt
- tp
->copied_seq
>= target
) ||
481 (sk
->sk_prot
->stream_memory_read
?
482 sk
->sk_prot
->stream_memory_read(sk
) : false);
486 * Wait for a TCP event.
488 * Note that we don't need to lock the socket, as the upper poll layers
489 * take care of normal races (between the test and the event) and we don't
490 * go look at any of the socket buffers directly.
492 __poll_t
tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
495 struct sock
*sk
= sock
->sk
;
496 const struct tcp_sock
*tp
= tcp_sk(sk
);
499 sock_poll_wait(file
, sock
, wait
);
501 state
= inet_sk_state_load(sk
);
502 if (state
== TCP_LISTEN
)
503 return inet_csk_listen_poll(sk
);
505 /* Socket is not locked. We are protected from async events
506 * by poll logic and correct handling of state changes
507 * made by other threads is impossible in any case.
513 * EPOLLHUP is certainly not done right. But poll() doesn't
514 * have a notion of HUP in just one direction, and for a
515 * socket the read side is more interesting.
517 * Some poll() documentation says that EPOLLHUP is incompatible
518 * with the EPOLLOUT/POLLWR flags, so somebody should check this
519 * all. But careful, it tends to be safer to return too many
520 * bits than too few, and you can easily break real applications
521 * if you don't tell them that something has hung up!
525 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
526 * our fs/select.c). It means that after we received EOF,
527 * poll always returns immediately, making impossible poll() on write()
528 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
529 * if and only if shutdown has been made in both directions.
530 * Actually, it is interesting to look how Solaris and DUX
531 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
532 * then we could set it on SND_SHUTDOWN. BTW examples given
533 * in Stevens' books assume exactly this behaviour, it explains
534 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
536 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
537 * blocking on fresh not-connected or disconnected socket. --ANK
539 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| state
== TCP_CLOSE
)
541 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
542 mask
|= EPOLLIN
| EPOLLRDNORM
| EPOLLRDHUP
;
544 /* Connected or passive Fast Open socket? */
545 if (state
!= TCP_SYN_SENT
&&
546 (state
!= TCP_SYN_RECV
|| tp
->fastopen_rsk
)) {
547 int target
= sock_rcvlowat(sk
, 0, INT_MAX
);
549 if (tp
->urg_seq
== tp
->copied_seq
&&
550 !sock_flag(sk
, SOCK_URGINLINE
) &&
554 if (tcp_stream_is_readable(tp
, target
, sk
))
555 mask
|= EPOLLIN
| EPOLLRDNORM
;
557 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
558 if (sk_stream_is_writeable(sk
)) {
559 mask
|= EPOLLOUT
| EPOLLWRNORM
;
560 } else { /* send SIGIO later */
561 sk_set_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
562 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
564 /* Race breaker. If space is freed after
565 * wspace test but before the flags are set,
566 * IO signal will be lost. Memory barrier
567 * pairs with the input side.
569 smp_mb__after_atomic();
570 if (sk_stream_is_writeable(sk
))
571 mask
|= EPOLLOUT
| EPOLLWRNORM
;
574 mask
|= EPOLLOUT
| EPOLLWRNORM
;
576 if (tp
->urg_data
& TCP_URG_VALID
)
578 } else if (state
== TCP_SYN_SENT
&& inet_sk(sk
)->defer_connect
) {
579 /* Active TCP fastopen socket with defer_connect
580 * Return EPOLLOUT so application can call write()
581 * in order for kernel to generate SYN+data
583 mask
|= EPOLLOUT
| EPOLLWRNORM
;
585 /* This barrier is coupled with smp_wmb() in tcp_reset() */
587 if (sk
->sk_err
|| !skb_queue_empty_lockless(&sk
->sk_error_queue
))
592 EXPORT_SYMBOL(tcp_poll
);
594 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
596 struct tcp_sock
*tp
= tcp_sk(sk
);
602 if (sk
->sk_state
== TCP_LISTEN
)
605 slow
= lock_sock_fast(sk
);
607 unlock_sock_fast(sk
, slow
);
610 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
613 if (sk
->sk_state
== TCP_LISTEN
)
616 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
619 answ
= tp
->write_seq
- tp
->snd_una
;
622 if (sk
->sk_state
== TCP_LISTEN
)
625 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
628 answ
= tp
->write_seq
- tp
->snd_nxt
;
634 return put_user(answ
, (int __user
*)arg
);
636 EXPORT_SYMBOL(tcp_ioctl
);
638 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
640 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
641 tp
->pushed_seq
= tp
->write_seq
;
644 static inline bool forced_push(const struct tcp_sock
*tp
)
646 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
649 static void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
651 struct tcp_sock
*tp
= tcp_sk(sk
);
652 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
655 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
656 tcb
->tcp_flags
= TCPHDR_ACK
;
658 __skb_header_release(skb
);
659 tcp_add_write_queue_tail(sk
, skb
);
660 sk
->sk_wmem_queued
+= skb
->truesize
;
661 sk_mem_charge(sk
, skb
->truesize
);
662 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
663 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
665 tcp_slow_start_after_idle_check(sk
);
668 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
)
671 tp
->snd_up
= tp
->write_seq
;
674 /* If a not yet filled skb is pushed, do not send it if
675 * we have data packets in Qdisc or NIC queues :
676 * Because TX completion will happen shortly, it gives a chance
677 * to coalesce future sendmsg() payload into this skb, without
678 * need for a timer, and with no latency trade off.
679 * As packets containing data payload have a bigger truesize
680 * than pure acks (dataless) packets, the last checks prevent
681 * autocorking if we only have an ACK in Qdisc/NIC queues,
682 * or if TX completion was delayed after we processed ACK packet.
684 static bool tcp_should_autocork(struct sock
*sk
, struct sk_buff
*skb
,
687 return skb
->len
< size_goal
&&
688 sock_net(sk
)->ipv4
.sysctl_tcp_autocorking
&&
689 !tcp_rtx_queue_empty(sk
) &&
690 refcount_read(&sk
->sk_wmem_alloc
) > skb
->truesize
;
693 static void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
694 int nonagle
, int size_goal
)
696 struct tcp_sock
*tp
= tcp_sk(sk
);
699 skb
= tcp_write_queue_tail(sk
);
702 if (!(flags
& MSG_MORE
) || forced_push(tp
))
703 tcp_mark_push(tp
, skb
);
705 tcp_mark_urg(tp
, flags
);
707 if (tcp_should_autocork(sk
, skb
, size_goal
)) {
709 /* avoid atomic op if TSQ_THROTTLED bit is already set */
710 if (!test_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
)) {
711 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPAUTOCORKING
);
712 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
714 /* It is possible TX completion already happened
715 * before we set TSQ_THROTTLED.
717 if (refcount_read(&sk
->sk_wmem_alloc
) > skb
->truesize
)
721 if (flags
& MSG_MORE
)
722 nonagle
= TCP_NAGLE_CORK
;
724 __tcp_push_pending_frames(sk
, mss_now
, nonagle
);
727 static int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
728 unsigned int offset
, size_t len
)
730 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
733 ret
= skb_splice_bits(skb
, skb
->sk
, offset
, tss
->pipe
,
734 min(rd_desc
->count
, len
), tss
->flags
);
736 rd_desc
->count
-= ret
;
740 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
742 /* Store TCP splice context information in read_descriptor_t. */
743 read_descriptor_t rd_desc
= {
748 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
752 * tcp_splice_read - splice data from TCP socket to a pipe
753 * @sock: socket to splice from
754 * @ppos: position (not valid)
755 * @pipe: pipe to splice to
756 * @len: number of bytes to splice
757 * @flags: splice modifier flags
760 * Will read pages from given socket and fill them into a pipe.
763 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
764 struct pipe_inode_info
*pipe
, size_t len
,
767 struct sock
*sk
= sock
->sk
;
768 struct tcp_splice_state tss
= {
777 sock_rps_record_flow(sk
);
779 * We can't seek on a socket input
788 timeo
= sock_rcvtimeo(sk
, sock
->file
->f_flags
& O_NONBLOCK
);
790 ret
= __tcp_splice_read(sk
, &tss
);
796 if (sock_flag(sk
, SOCK_DONE
))
799 ret
= sock_error(sk
);
802 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
804 if (sk
->sk_state
== TCP_CLOSE
) {
806 * This occurs when user tries to read
807 * from never connected socket.
816 /* if __tcp_splice_read() got nothing while we have
817 * an skb in receive queue, we do not want to loop.
818 * This might happen with URG data.
820 if (!skb_queue_empty(&sk
->sk_receive_queue
))
822 sk_wait_data(sk
, &timeo
, NULL
);
823 if (signal_pending(current
)) {
824 ret
= sock_intr_errno(timeo
);
837 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
838 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
839 signal_pending(current
))
850 EXPORT_SYMBOL(tcp_splice_read
);
852 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
,
858 skb
= sk
->sk_tx_skb_cache
;
860 skb
->truesize
= SKB_TRUESIZE(skb_end_offset(skb
));
861 sk
->sk_tx_skb_cache
= NULL
;
863 INIT_LIST_HEAD(&skb
->tcp_tsorted_anchor
);
864 skb_shinfo(skb
)->tx_flags
= 0;
865 memset(TCP_SKB_CB(skb
), 0, sizeof(struct tcp_skb_cb
));
869 /* The TCP header must be at least 32-bit aligned. */
870 size
= ALIGN(size
, 4);
872 if (unlikely(tcp_under_memory_pressure(sk
)))
873 sk_mem_reclaim_partial(sk
);
875 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
879 if (force_schedule
) {
880 mem_scheduled
= true;
881 sk_forced_mem_schedule(sk
, skb
->truesize
);
883 mem_scheduled
= sk_wmem_schedule(sk
, skb
->truesize
);
885 if (likely(mem_scheduled
)) {
886 skb_reserve(skb
, sk
->sk_prot
->max_header
);
888 * Make sure that we have exactly size bytes
889 * available to the caller, no more, no less.
891 skb
->reserved_tailroom
= skb
->end
- skb
->tail
- size
;
892 INIT_LIST_HEAD(&skb
->tcp_tsorted_anchor
);
897 sk
->sk_prot
->enter_memory_pressure(sk
);
898 sk_stream_moderate_sndbuf(sk
);
903 static unsigned int tcp_xmit_size_goal(struct sock
*sk
, u32 mss_now
,
906 struct tcp_sock
*tp
= tcp_sk(sk
);
907 u32 new_size_goal
, size_goal
;
912 /* Note : tcp_tso_autosize() will eventually split this later */
913 new_size_goal
= sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
;
914 new_size_goal
= tcp_bound_to_half_wnd(tp
, new_size_goal
);
916 /* We try hard to avoid divides here */
917 size_goal
= tp
->gso_segs
* mss_now
;
918 if (unlikely(new_size_goal
< size_goal
||
919 new_size_goal
>= size_goal
+ mss_now
)) {
920 tp
->gso_segs
= min_t(u16
, new_size_goal
/ mss_now
,
921 sk
->sk_gso_max_segs
);
922 size_goal
= tp
->gso_segs
* mss_now
;
925 return max(size_goal
, mss_now
);
928 static int tcp_send_mss(struct sock
*sk
, int *size_goal
, int flags
)
932 mss_now
= tcp_current_mss(sk
);
933 *size_goal
= tcp_xmit_size_goal(sk
, mss_now
, !(flags
& MSG_OOB
));
938 /* In some cases, both sendpage() and sendmsg() could have added
939 * an skb to the write queue, but failed adding payload on it.
940 * We need to remove it to consume less memory, but more
941 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
944 static void tcp_remove_empty_skb(struct sock
*sk
, struct sk_buff
*skb
)
946 if (skb
&& !skb
->len
) {
947 tcp_unlink_write_queue(skb
, sk
);
948 if (tcp_write_queue_empty(sk
))
949 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
950 sk_wmem_free_skb(sk
, skb
);
954 ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
*page
, int offset
,
955 size_t size
, int flags
)
957 struct tcp_sock
*tp
= tcp_sk(sk
);
958 int mss_now
, size_goal
;
961 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
963 if (IS_ENABLED(CONFIG_DEBUG_VM
) &&
964 WARN_ONCE(PageSlab(page
), "page must not be a Slab one"))
967 /* Wait for a connection to finish. One exception is TCP Fast Open
968 * (passive side) where data is allowed to be sent before a connection
969 * is fully established.
971 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
972 !tcp_passive_fastopen(sk
)) {
973 err
= sk_stream_wait_connect(sk
, &timeo
);
978 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
980 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
984 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
988 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
992 if (!skb
|| (copy
= size_goal
- skb
->len
) <= 0 ||
993 !tcp_skb_can_collapse_to(skb
)) {
995 if (!sk_stream_memory_free(sk
))
996 goto wait_for_sndbuf
;
998 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
,
999 tcp_rtx_and_write_queues_empty(sk
));
1001 goto wait_for_memory
;
1003 #ifdef CONFIG_TLS_DEVICE
1004 skb
->decrypted
= !!(flags
& MSG_SENDPAGE_DECRYPTED
);
1006 skb_entail(sk
, skb
);
1013 i
= skb_shinfo(skb
)->nr_frags
;
1014 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
1015 if (!can_coalesce
&& i
>= sysctl_max_skb_frags
) {
1016 tcp_mark_push(tp
, skb
);
1019 if (!sk_wmem_schedule(sk
, copy
))
1020 goto wait_for_memory
;
1023 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
1026 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
1029 if (!(flags
& MSG_NO_SHARED_FRAGS
))
1030 skb_shinfo(skb
)->tx_flags
|= SKBTX_SHARED_FRAG
;
1033 skb
->data_len
+= copy
;
1034 skb
->truesize
+= copy
;
1035 sk
->sk_wmem_queued
+= copy
;
1036 sk_mem_charge(sk
, copy
);
1037 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1038 tp
->write_seq
+= copy
;
1039 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1040 tcp_skb_pcount_set(skb
, 0);
1043 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
1051 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
1054 if (forced_push(tp
)) {
1055 tcp_mark_push(tp
, skb
);
1056 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1057 } else if (skb
== tcp_send_head(sk
))
1058 tcp_push_one(sk
, mss_now
);
1062 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1064 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
,
1065 TCP_NAGLE_PUSH
, size_goal
);
1067 err
= sk_stream_wait_memory(sk
, &timeo
);
1071 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1076 tcp_tx_timestamp(sk
, sk
->sk_tsflags
);
1077 if (!(flags
& MSG_SENDPAGE_NOTLAST
))
1078 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
, size_goal
);
1083 tcp_remove_empty_skb(sk
, tcp_write_queue_tail(sk
));
1087 /* make sure we wake any epoll edge trigger waiter */
1088 if (unlikely(skb_queue_len(&sk
->sk_write_queue
) == 0 &&
1090 sk
->sk_write_space(sk
);
1091 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1093 return sk_stream_error(sk
, flags
, err
);
1095 EXPORT_SYMBOL_GPL(do_tcp_sendpages
);
1097 int tcp_sendpage_locked(struct sock
*sk
, struct page
*page
, int offset
,
1098 size_t size
, int flags
)
1100 if (!(sk
->sk_route_caps
& NETIF_F_SG
))
1101 return sock_no_sendpage_locked(sk
, page
, offset
, size
, flags
);
1103 tcp_rate_check_app_limited(sk
); /* is sending application-limited? */
1105 return do_tcp_sendpages(sk
, page
, offset
, size
, flags
);
1107 EXPORT_SYMBOL_GPL(tcp_sendpage_locked
);
1109 int tcp_sendpage(struct sock
*sk
, struct page
*page
, int offset
,
1110 size_t size
, int flags
)
1115 ret
= tcp_sendpage_locked(sk
, page
, offset
, size
, flags
);
1120 EXPORT_SYMBOL(tcp_sendpage
);
1122 void tcp_free_fastopen_req(struct tcp_sock
*tp
)
1124 if (tp
->fastopen_req
) {
1125 kfree(tp
->fastopen_req
);
1126 tp
->fastopen_req
= NULL
;
1130 static int tcp_sendmsg_fastopen(struct sock
*sk
, struct msghdr
*msg
,
1131 int *copied
, size_t size
,
1132 struct ubuf_info
*uarg
)
1134 struct tcp_sock
*tp
= tcp_sk(sk
);
1135 struct inet_sock
*inet
= inet_sk(sk
);
1136 struct sockaddr
*uaddr
= msg
->msg_name
;
1139 if (!(sock_net(sk
)->ipv4
.sysctl_tcp_fastopen
& TFO_CLIENT_ENABLE
) ||
1140 (uaddr
&& msg
->msg_namelen
>= sizeof(uaddr
->sa_family
) &&
1141 uaddr
->sa_family
== AF_UNSPEC
))
1143 if (tp
->fastopen_req
)
1144 return -EALREADY
; /* Another Fast Open is in progress */
1146 tp
->fastopen_req
= kzalloc(sizeof(struct tcp_fastopen_request
),
1148 if (unlikely(!tp
->fastopen_req
))
1150 tp
->fastopen_req
->data
= msg
;
1151 tp
->fastopen_req
->size
= size
;
1152 tp
->fastopen_req
->uarg
= uarg
;
1154 if (inet
->defer_connect
) {
1155 err
= tcp_connect(sk
);
1156 /* Same failure procedure as in tcp_v4/6_connect */
1158 tcp_set_state(sk
, TCP_CLOSE
);
1159 inet
->inet_dport
= 0;
1160 sk
->sk_route_caps
= 0;
1163 flags
= (msg
->msg_flags
& MSG_DONTWAIT
) ? O_NONBLOCK
: 0;
1164 err
= __inet_stream_connect(sk
->sk_socket
, uaddr
,
1165 msg
->msg_namelen
, flags
, 1);
1166 /* fastopen_req could already be freed in __inet_stream_connect
1167 * if the connection times out or gets rst
1169 if (tp
->fastopen_req
) {
1170 *copied
= tp
->fastopen_req
->copied
;
1171 tcp_free_fastopen_req(tp
);
1172 inet
->defer_connect
= 0;
1177 int tcp_sendmsg_locked(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
1179 struct tcp_sock
*tp
= tcp_sk(sk
);
1180 struct ubuf_info
*uarg
= NULL
;
1181 struct sk_buff
*skb
;
1182 struct sockcm_cookie sockc
;
1183 int flags
, err
, copied
= 0;
1184 int mss_now
= 0, size_goal
, copied_syn
= 0;
1185 bool process_backlog
= false;
1189 flags
= msg
->msg_flags
;
1191 if (flags
& MSG_ZEROCOPY
&& size
&& sock_flag(sk
, SOCK_ZEROCOPY
)) {
1192 skb
= tcp_write_queue_tail(sk
);
1193 uarg
= sock_zerocopy_realloc(sk
, size
, skb_zcopy(skb
));
1199 zc
= sk
->sk_route_caps
& NETIF_F_SG
;
1204 if (unlikely(flags
& MSG_FASTOPEN
|| inet_sk(sk
)->defer_connect
) &&
1206 err
= tcp_sendmsg_fastopen(sk
, msg
, &copied_syn
, size
, uarg
);
1207 if (err
== -EINPROGRESS
&& copied_syn
> 0)
1213 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
1215 tcp_rate_check_app_limited(sk
); /* is sending application-limited? */
1217 /* Wait for a connection to finish. One exception is TCP Fast Open
1218 * (passive side) where data is allowed to be sent before a connection
1219 * is fully established.
1221 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
1222 !tcp_passive_fastopen(sk
)) {
1223 err
= sk_stream_wait_connect(sk
, &timeo
);
1228 if (unlikely(tp
->repair
)) {
1229 if (tp
->repair_queue
== TCP_RECV_QUEUE
) {
1230 copied
= tcp_send_rcvq(sk
, msg
, size
);
1235 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1238 /* 'common' sending to sendq */
1241 sockcm_init(&sockc
, sk
);
1242 if (msg
->msg_controllen
) {
1243 err
= sock_cmsg_send(sk
, msg
, &sockc
);
1244 if (unlikely(err
)) {
1250 /* This should be in poll */
1251 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
1253 /* Ok commence sending. */
1257 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1260 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
1263 while (msg_data_left(msg
)) {
1266 skb
= tcp_write_queue_tail(sk
);
1268 copy
= size_goal
- skb
->len
;
1270 if (copy
<= 0 || !tcp_skb_can_collapse_to(skb
)) {
1274 if (!sk_stream_memory_free(sk
))
1275 goto wait_for_sndbuf
;
1277 if (process_backlog
&& sk_flush_backlog(sk
)) {
1278 process_backlog
= false;
1281 first_skb
= tcp_rtx_and_write_queues_empty(sk
);
1282 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
,
1285 goto wait_for_memory
;
1287 process_backlog
= true;
1288 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1290 skb_entail(sk
, skb
);
1293 /* All packets are restored as if they have
1294 * already been sent. skb_mstamp_ns isn't set to
1295 * avoid wrong rtt estimation.
1298 TCP_SKB_CB(skb
)->sacked
|= TCPCB_REPAIRED
;
1301 /* Try to append data to the end of skb. */
1302 if (copy
> msg_data_left(msg
))
1303 copy
= msg_data_left(msg
);
1305 /* Where to copy to? */
1306 if (skb_availroom(skb
) > 0 && !zc
) {
1307 /* We have some space in skb head. Superb! */
1308 copy
= min_t(int, copy
, skb_availroom(skb
));
1309 err
= skb_add_data_nocache(sk
, skb
, &msg
->msg_iter
, copy
);
1314 int i
= skb_shinfo(skb
)->nr_frags
;
1315 struct page_frag
*pfrag
= sk_page_frag(sk
);
1317 if (!sk_page_frag_refill(sk
, pfrag
))
1318 goto wait_for_memory
;
1320 if (!skb_can_coalesce(skb
, i
, pfrag
->page
,
1322 if (i
>= sysctl_max_skb_frags
) {
1323 tcp_mark_push(tp
, skb
);
1329 copy
= min_t(int, copy
, pfrag
->size
- pfrag
->offset
);
1331 if (!sk_wmem_schedule(sk
, copy
))
1332 goto wait_for_memory
;
1334 err
= skb_copy_to_page_nocache(sk
, &msg
->msg_iter
, skb
,
1341 /* Update the skb. */
1343 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
1345 skb_fill_page_desc(skb
, i
, pfrag
->page
,
1346 pfrag
->offset
, copy
);
1347 page_ref_inc(pfrag
->page
);
1349 pfrag
->offset
+= copy
;
1351 err
= skb_zerocopy_iter_stream(sk
, skb
, msg
, copy
, uarg
);
1352 if (err
== -EMSGSIZE
|| err
== -EEXIST
) {
1353 tcp_mark_push(tp
, skb
);
1362 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
1364 tp
->write_seq
+= copy
;
1365 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1366 tcp_skb_pcount_set(skb
, 0);
1369 if (!msg_data_left(msg
)) {
1370 if (unlikely(flags
& MSG_EOR
))
1371 TCP_SKB_CB(skb
)->eor
= 1;
1375 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
) || unlikely(tp
->repair
))
1378 if (forced_push(tp
)) {
1379 tcp_mark_push(tp
, skb
);
1380 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1381 } else if (skb
== tcp_send_head(sk
))
1382 tcp_push_one(sk
, mss_now
);
1386 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1389 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
,
1390 TCP_NAGLE_PUSH
, size_goal
);
1392 err
= sk_stream_wait_memory(sk
, &timeo
);
1396 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1401 tcp_tx_timestamp(sk
, sockc
.tsflags
);
1402 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
, size_goal
);
1405 sock_zerocopy_put(uarg
);
1406 return copied
+ copied_syn
;
1409 skb
= tcp_write_queue_tail(sk
);
1411 tcp_remove_empty_skb(sk
, skb
);
1413 if (copied
+ copied_syn
)
1416 sock_zerocopy_put_abort(uarg
, true);
1417 err
= sk_stream_error(sk
, flags
, err
);
1418 /* make sure we wake any epoll edge trigger waiter */
1419 if (unlikely(skb_queue_len(&sk
->sk_write_queue
) == 0 &&
1421 sk
->sk_write_space(sk
);
1422 tcp_chrono_stop(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1426 EXPORT_SYMBOL_GPL(tcp_sendmsg_locked
);
1428 int tcp_sendmsg(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
1433 ret
= tcp_sendmsg_locked(sk
, msg
, size
);
1438 EXPORT_SYMBOL(tcp_sendmsg
);
1441 * Handle reading urgent data. BSD has very simple semantics for
1442 * this, no blocking and very strange errors 8)
1445 static int tcp_recv_urg(struct sock
*sk
, struct msghdr
*msg
, int len
, int flags
)
1447 struct tcp_sock
*tp
= tcp_sk(sk
);
1449 /* No URG data to read. */
1450 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1451 tp
->urg_data
== TCP_URG_READ
)
1452 return -EINVAL
; /* Yes this is right ! */
1454 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1457 if (tp
->urg_data
& TCP_URG_VALID
) {
1459 char c
= tp
->urg_data
;
1461 if (!(flags
& MSG_PEEK
))
1462 tp
->urg_data
= TCP_URG_READ
;
1464 /* Read urgent data. */
1465 msg
->msg_flags
|= MSG_OOB
;
1468 if (!(flags
& MSG_TRUNC
))
1469 err
= memcpy_to_msg(msg
, &c
, 1);
1472 msg
->msg_flags
|= MSG_TRUNC
;
1474 return err
? -EFAULT
: len
;
1477 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1480 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1481 * the available implementations agree in this case:
1482 * this call should never block, independent of the
1483 * blocking state of the socket.
1484 * Mike <pall@rz.uni-karlsruhe.de>
1489 static int tcp_peek_sndq(struct sock
*sk
, struct msghdr
*msg
, int len
)
1491 struct sk_buff
*skb
;
1492 int copied
= 0, err
= 0;
1494 /* XXX -- need to support SO_PEEK_OFF */
1496 skb_rbtree_walk(skb
, &sk
->tcp_rtx_queue
) {
1497 err
= skb_copy_datagram_msg(skb
, 0, msg
, skb
->len
);
1503 skb_queue_walk(&sk
->sk_write_queue
, skb
) {
1504 err
= skb_copy_datagram_msg(skb
, 0, msg
, skb
->len
);
1511 return err
?: copied
;
1514 /* Clean up the receive buffer for full frames taken by the user,
1515 * then send an ACK if necessary. COPIED is the number of bytes
1516 * tcp_recvmsg has given to the user so far, it speeds up the
1517 * calculation of whether or not we must ACK for the sake of
1520 static void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1522 struct tcp_sock
*tp
= tcp_sk(sk
);
1523 bool time_to_ack
= false;
1525 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1527 WARN(skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
),
1528 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1529 tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
);
1531 if (inet_csk_ack_scheduled(sk
)) {
1532 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1533 /* Delayed ACKs frequently hit locked sockets during bulk
1535 if (icsk
->icsk_ack
.blocked
||
1536 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1537 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1539 * If this read emptied read buffer, we send ACK, if
1540 * connection is not bidirectional, user drained
1541 * receive buffer and there was a small segment
1545 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1546 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1547 !inet_csk_in_pingpong_mode(sk
))) &&
1548 !atomic_read(&sk
->sk_rmem_alloc
)))
1552 /* We send an ACK if we can now advertise a non-zero window
1553 * which has been raised "significantly".
1555 * Even if window raised up to infinity, do not send window open ACK
1556 * in states, where we will not receive more. It is useless.
1558 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1559 __u32 rcv_window_now
= tcp_receive_window(tp
);
1561 /* Optimize, __tcp_select_window() is not cheap. */
1562 if (2*rcv_window_now
<= tp
->window_clamp
) {
1563 __u32 new_window
= __tcp_select_window(sk
);
1565 /* Send ACK now, if this read freed lots of space
1566 * in our buffer. Certainly, new_window is new window.
1567 * We can advertise it now, if it is not less than current one.
1568 * "Lots" means "at least twice" here.
1570 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1578 static struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1580 struct sk_buff
*skb
;
1583 while ((skb
= skb_peek(&sk
->sk_receive_queue
)) != NULL
) {
1584 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1585 if (unlikely(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
1586 pr_err_once("%s: found a SYN, please report !\n", __func__
);
1589 if (offset
< skb
->len
|| (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)) {
1593 /* This looks weird, but this can happen if TCP collapsing
1594 * splitted a fat GRO packet, while we released socket lock
1595 * in skb_splice_bits()
1597 sk_eat_skb(sk
, skb
);
1603 * This routine provides an alternative to tcp_recvmsg() for routines
1604 * that would like to handle copying from skbuffs directly in 'sendfile'
1607 * - It is assumed that the socket was locked by the caller.
1608 * - The routine does not block.
1609 * - At present, there is no support for reading OOB data
1610 * or for 'peeking' the socket using this routine
1611 * (although both would be easy to implement).
1613 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1614 sk_read_actor_t recv_actor
)
1616 struct sk_buff
*skb
;
1617 struct tcp_sock
*tp
= tcp_sk(sk
);
1618 u32 seq
= tp
->copied_seq
;
1622 if (sk
->sk_state
== TCP_LISTEN
)
1624 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1625 if (offset
< skb
->len
) {
1629 len
= skb
->len
- offset
;
1630 /* Stop reading if we hit a patch of urgent data */
1632 u32 urg_offset
= tp
->urg_seq
- seq
;
1633 if (urg_offset
< len
)
1638 used
= recv_actor(desc
, skb
, offset
, len
);
1643 } else if (used
<= len
) {
1648 /* If recv_actor drops the lock (e.g. TCP splice
1649 * receive) the skb pointer might be invalid when
1650 * getting here: tcp_collapse might have deleted it
1651 * while aggregating skbs from the socket queue.
1653 skb
= tcp_recv_skb(sk
, seq
- 1, &offset
);
1656 /* TCP coalescing might have appended data to the skb.
1657 * Try to splice more frags
1659 if (offset
+ 1 != skb
->len
)
1662 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) {
1663 sk_eat_skb(sk
, skb
);
1667 sk_eat_skb(sk
, skb
);
1670 tp
->copied_seq
= seq
;
1672 tp
->copied_seq
= seq
;
1674 tcp_rcv_space_adjust(sk
);
1676 /* Clean up data we have read: This will do ACK frames. */
1678 tcp_recv_skb(sk
, seq
, &offset
);
1679 tcp_cleanup_rbuf(sk
, copied
);
1683 EXPORT_SYMBOL(tcp_read_sock
);
1685 int tcp_peek_len(struct socket
*sock
)
1687 return tcp_inq(sock
->sk
);
1689 EXPORT_SYMBOL(tcp_peek_len
);
1691 /* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1692 int tcp_set_rcvlowat(struct sock
*sk
, int val
)
1696 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)
1697 cap
= sk
->sk_rcvbuf
>> 1;
1699 cap
= sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2] >> 1;
1700 val
= min(val
, cap
);
1701 sk
->sk_rcvlowat
= val
? : 1;
1703 /* Check if we need to signal EPOLLIN right now */
1706 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)
1710 if (val
> sk
->sk_rcvbuf
) {
1711 sk
->sk_rcvbuf
= val
;
1712 tcp_sk(sk
)->window_clamp
= tcp_win_from_space(sk
, val
);
1716 EXPORT_SYMBOL(tcp_set_rcvlowat
);
1719 static const struct vm_operations_struct tcp_vm_ops
= {
1722 int tcp_mmap(struct file
*file
, struct socket
*sock
,
1723 struct vm_area_struct
*vma
)
1725 if (vma
->vm_flags
& (VM_WRITE
| VM_EXEC
))
1727 vma
->vm_flags
&= ~(VM_MAYWRITE
| VM_MAYEXEC
);
1729 /* Instruct vm_insert_page() to not down_read(mmap_sem) */
1730 vma
->vm_flags
|= VM_MIXEDMAP
;
1732 vma
->vm_ops
= &tcp_vm_ops
;
1735 EXPORT_SYMBOL(tcp_mmap
);
1737 static int tcp_zerocopy_receive(struct sock
*sk
,
1738 struct tcp_zerocopy_receive
*zc
)
1740 unsigned long address
= (unsigned long)zc
->address
;
1741 const skb_frag_t
*frags
= NULL
;
1742 u32 length
= 0, seq
, offset
;
1743 struct vm_area_struct
*vma
;
1744 struct sk_buff
*skb
= NULL
;
1745 struct tcp_sock
*tp
;
1749 if (address
& (PAGE_SIZE
- 1) || address
!= zc
->address
)
1752 if (sk
->sk_state
== TCP_LISTEN
)
1755 sock_rps_record_flow(sk
);
1757 down_read(¤t
->mm
->mmap_sem
);
1760 vma
= find_vma(current
->mm
, address
);
1761 if (!vma
|| vma
->vm_start
> address
|| vma
->vm_ops
!= &tcp_vm_ops
)
1763 zc
->length
= min_t(unsigned long, zc
->length
, vma
->vm_end
- address
);
1766 seq
= tp
->copied_seq
;
1768 zc
->length
= min_t(u32
, zc
->length
, inq
);
1769 zc
->length
&= ~(PAGE_SIZE
- 1);
1771 zap_page_range(vma
, address
, zc
->length
);
1772 zc
->recv_skip_hint
= 0;
1774 zc
->recv_skip_hint
= inq
;
1777 while (length
+ PAGE_SIZE
<= zc
->length
) {
1778 if (zc
->recv_skip_hint
< PAGE_SIZE
) {
1781 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1783 skb
= tcp_recv_skb(sk
, seq
, &offset
);
1786 zc
->recv_skip_hint
= skb
->len
- offset
;
1787 offset
-= skb_headlen(skb
);
1788 if ((int)offset
< 0 || skb_has_frag_list(skb
))
1790 frags
= skb_shinfo(skb
)->frags
;
1792 if (frags
->size
> offset
)
1794 offset
-= frags
->size
;
1798 if (frags
->size
!= PAGE_SIZE
|| frags
->page_offset
) {
1799 int remaining
= zc
->recv_skip_hint
;
1801 while (remaining
&& (frags
->size
!= PAGE_SIZE
||
1802 frags
->page_offset
)) {
1803 remaining
-= frags
->size
;
1806 zc
->recv_skip_hint
-= remaining
;
1809 ret
= vm_insert_page(vma
, address
+ length
,
1810 skb_frag_page(frags
));
1813 length
+= PAGE_SIZE
;
1815 zc
->recv_skip_hint
-= PAGE_SIZE
;
1819 up_read(¤t
->mm
->mmap_sem
);
1821 tp
->copied_seq
= seq
;
1822 tcp_rcv_space_adjust(sk
);
1824 /* Clean up data we have read: This will do ACK frames. */
1825 tcp_recv_skb(sk
, seq
, &offset
);
1826 tcp_cleanup_rbuf(sk
, length
);
1828 if (length
== zc
->length
)
1829 zc
->recv_skip_hint
= 0;
1831 if (!zc
->recv_skip_hint
&& sock_flag(sk
, SOCK_DONE
))
1834 zc
->length
= length
;
1839 static void tcp_update_recv_tstamps(struct sk_buff
*skb
,
1840 struct scm_timestamping_internal
*tss
)
1843 tss
->ts
[0] = ktime_to_timespec64(skb
->tstamp
);
1845 tss
->ts
[0] = (struct timespec64
) {0};
1847 if (skb_hwtstamps(skb
)->hwtstamp
)
1848 tss
->ts
[2] = ktime_to_timespec64(skb_hwtstamps(skb
)->hwtstamp
);
1850 tss
->ts
[2] = (struct timespec64
) {0};
1853 /* Similar to __sock_recv_timestamp, but does not require an skb */
1854 static void tcp_recv_timestamp(struct msghdr
*msg
, const struct sock
*sk
,
1855 struct scm_timestamping_internal
*tss
)
1857 int new_tstamp
= sock_flag(sk
, SOCK_TSTAMP_NEW
);
1858 bool has_timestamping
= false;
1860 if (tss
->ts
[0].tv_sec
|| tss
->ts
[0].tv_nsec
) {
1861 if (sock_flag(sk
, SOCK_RCVTSTAMP
)) {
1862 if (sock_flag(sk
, SOCK_RCVTSTAMPNS
)) {
1864 struct __kernel_timespec kts
= {tss
->ts
[0].tv_sec
, tss
->ts
[0].tv_nsec
};
1866 put_cmsg(msg
, SOL_SOCKET
, SO_TIMESTAMPNS_NEW
,
1869 struct timespec ts_old
= timespec64_to_timespec(tss
->ts
[0]);
1871 put_cmsg(msg
, SOL_SOCKET
, SO_TIMESTAMPNS_OLD
,
1872 sizeof(ts_old
), &ts_old
);
1876 struct __kernel_sock_timeval stv
;
1878 stv
.tv_sec
= tss
->ts
[0].tv_sec
;
1879 stv
.tv_usec
= tss
->ts
[0].tv_nsec
/ 1000;
1880 put_cmsg(msg
, SOL_SOCKET
, SO_TIMESTAMP_NEW
,
1883 struct __kernel_old_timeval tv
;
1885 tv
.tv_sec
= tss
->ts
[0].tv_sec
;
1886 tv
.tv_usec
= tss
->ts
[0].tv_nsec
/ 1000;
1887 put_cmsg(msg
, SOL_SOCKET
, SO_TIMESTAMP_OLD
,
1893 if (sk
->sk_tsflags
& SOF_TIMESTAMPING_SOFTWARE
)
1894 has_timestamping
= true;
1896 tss
->ts
[0] = (struct timespec64
) {0};
1899 if (tss
->ts
[2].tv_sec
|| tss
->ts
[2].tv_nsec
) {
1900 if (sk
->sk_tsflags
& SOF_TIMESTAMPING_RAW_HARDWARE
)
1901 has_timestamping
= true;
1903 tss
->ts
[2] = (struct timespec64
) {0};
1906 if (has_timestamping
) {
1907 tss
->ts
[1] = (struct timespec64
) {0};
1908 if (sock_flag(sk
, SOCK_TSTAMP_NEW
))
1909 put_cmsg_scm_timestamping64(msg
, tss
);
1911 put_cmsg_scm_timestamping(msg
, tss
);
1915 static int tcp_inq_hint(struct sock
*sk
)
1917 const struct tcp_sock
*tp
= tcp_sk(sk
);
1918 u32 copied_seq
= READ_ONCE(tp
->copied_seq
);
1919 u32 rcv_nxt
= READ_ONCE(tp
->rcv_nxt
);
1922 inq
= rcv_nxt
- copied_seq
;
1923 if (unlikely(inq
< 0 || copied_seq
!= READ_ONCE(tp
->copied_seq
))) {
1925 inq
= tp
->rcv_nxt
- tp
->copied_seq
;
1928 /* After receiving a FIN, tell the user-space to continue reading
1929 * by returning a non-zero inq.
1931 if (inq
== 0 && sock_flag(sk
, SOCK_DONE
))
1937 * This routine copies from a sock struct into the user buffer.
1939 * Technical note: in 2.3 we work on _locked_ socket, so that
1940 * tricks with *seq access order and skb->users are not required.
1941 * Probably, code can be easily improved even more.
1944 int tcp_recvmsg(struct sock
*sk
, struct msghdr
*msg
, size_t len
, int nonblock
,
1945 int flags
, int *addr_len
)
1947 struct tcp_sock
*tp
= tcp_sk(sk
);
1953 int target
; /* Read at least this many bytes */
1955 struct sk_buff
*skb
, *last
;
1957 struct scm_timestamping_internal tss
;
1958 bool has_tss
= false;
1961 if (unlikely(flags
& MSG_ERRQUEUE
))
1962 return inet_recv_error(sk
, msg
, len
, addr_len
);
1964 if (sk_can_busy_loop(sk
) && skb_queue_empty_lockless(&sk
->sk_receive_queue
) &&
1965 (sk
->sk_state
== TCP_ESTABLISHED
))
1966 sk_busy_loop(sk
, nonblock
);
1971 if (sk
->sk_state
== TCP_LISTEN
)
1974 has_cmsg
= tp
->recvmsg_inq
;
1975 timeo
= sock_rcvtimeo(sk
, nonblock
);
1977 /* Urgent data needs to be handled specially. */
1978 if (flags
& MSG_OOB
)
1981 if (unlikely(tp
->repair
)) {
1983 if (!(flags
& MSG_PEEK
))
1986 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
1990 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1993 /* 'common' recv queue MSG_PEEK-ing */
1996 seq
= &tp
->copied_seq
;
1997 if (flags
& MSG_PEEK
) {
1998 peek_seq
= tp
->copied_seq
;
2002 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
2007 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2008 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
2011 if (signal_pending(current
)) {
2012 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
2017 /* Next get a buffer. */
2019 last
= skb_peek_tail(&sk
->sk_receive_queue
);
2020 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
2022 /* Now that we have two receive queues this
2025 if (WARN(before(*seq
, TCP_SKB_CB(skb
)->seq
),
2026 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2027 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
,
2031 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
2032 if (unlikely(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2033 pr_err_once("%s: found a SYN, please report !\n", __func__
);
2036 if (offset
< skb
->len
)
2038 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
2040 WARN(!(flags
& MSG_PEEK
),
2041 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2042 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
, flags
);
2045 /* Well, if we have backlog, try to process it now yet. */
2047 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
2052 sk
->sk_state
== TCP_CLOSE
||
2053 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
2055 signal_pending(current
))
2058 if (sock_flag(sk
, SOCK_DONE
))
2062 copied
= sock_error(sk
);
2066 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
2069 if (sk
->sk_state
== TCP_CLOSE
) {
2070 /* This occurs when user tries to read
2071 * from never connected socket.
2082 if (signal_pending(current
)) {
2083 copied
= sock_intr_errno(timeo
);
2088 tcp_cleanup_rbuf(sk
, copied
);
2090 if (copied
>= target
) {
2091 /* Do not sleep, just process backlog. */
2095 sk_wait_data(sk
, &timeo
, last
);
2098 if ((flags
& MSG_PEEK
) &&
2099 (peek_seq
- copied
- urg_hole
!= tp
->copied_seq
)) {
2100 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2102 task_pid_nr(current
));
2103 peek_seq
= tp
->copied_seq
;
2108 /* Ok so how much can we use? */
2109 used
= skb
->len
- offset
;
2113 /* Do we have urgent data here? */
2115 u32 urg_offset
= tp
->urg_seq
- *seq
;
2116 if (urg_offset
< used
) {
2118 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
2131 if (!(flags
& MSG_TRUNC
)) {
2132 err
= skb_copy_datagram_msg(skb
, offset
, msg
, used
);
2134 /* Exception. Bailout! */
2145 tcp_rcv_space_adjust(sk
);
2148 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
2150 tcp_fast_path_check(sk
);
2152 if (used
+ offset
< skb
->len
)
2155 if (TCP_SKB_CB(skb
)->has_rxtstamp
) {
2156 tcp_update_recv_tstamps(skb
, &tss
);
2160 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
2162 if (!(flags
& MSG_PEEK
))
2163 sk_eat_skb(sk
, skb
);
2167 /* Process the FIN. */
2169 if (!(flags
& MSG_PEEK
))
2170 sk_eat_skb(sk
, skb
);
2174 /* According to UNIX98, msg_name/msg_namelen are ignored
2175 * on connected socket. I was just happy when found this 8) --ANK
2178 /* Clean up data we have read: This will do ACK frames. */
2179 tcp_cleanup_rbuf(sk
, copied
);
2185 tcp_recv_timestamp(msg
, sk
, &tss
);
2186 if (tp
->recvmsg_inq
) {
2187 inq
= tcp_inq_hint(sk
);
2188 put_cmsg(msg
, SOL_TCP
, TCP_CM_INQ
, sizeof(inq
), &inq
);
2199 err
= tcp_recv_urg(sk
, msg
, len
, flags
);
2203 err
= tcp_peek_sndq(sk
, msg
, len
);
2206 EXPORT_SYMBOL(tcp_recvmsg
);
2208 void tcp_set_state(struct sock
*sk
, int state
)
2210 int oldstate
= sk
->sk_state
;
2212 /* We defined a new enum for TCP states that are exported in BPF
2213 * so as not force the internal TCP states to be frozen. The
2214 * following checks will detect if an internal state value ever
2215 * differs from the BPF value. If this ever happens, then we will
2216 * need to remap the internal value to the BPF value before calling
2217 * tcp_call_bpf_2arg.
2219 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED
!= (int)TCP_ESTABLISHED
);
2220 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT
!= (int)TCP_SYN_SENT
);
2221 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV
!= (int)TCP_SYN_RECV
);
2222 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1
!= (int)TCP_FIN_WAIT1
);
2223 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2
!= (int)TCP_FIN_WAIT2
);
2224 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT
!= (int)TCP_TIME_WAIT
);
2225 BUILD_BUG_ON((int)BPF_TCP_CLOSE
!= (int)TCP_CLOSE
);
2226 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT
!= (int)TCP_CLOSE_WAIT
);
2227 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK
!= (int)TCP_LAST_ACK
);
2228 BUILD_BUG_ON((int)BPF_TCP_LISTEN
!= (int)TCP_LISTEN
);
2229 BUILD_BUG_ON((int)BPF_TCP_CLOSING
!= (int)TCP_CLOSING
);
2230 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV
!= (int)TCP_NEW_SYN_RECV
);
2231 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES
!= (int)TCP_MAX_STATES
);
2233 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk
), BPF_SOCK_OPS_STATE_CB_FLAG
))
2234 tcp_call_bpf_2arg(sk
, BPF_SOCK_OPS_STATE_CB
, oldstate
, state
);
2237 case TCP_ESTABLISHED
:
2238 if (oldstate
!= TCP_ESTABLISHED
)
2239 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
2243 if (oldstate
== TCP_CLOSE_WAIT
|| oldstate
== TCP_ESTABLISHED
)
2244 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ESTABRESETS
);
2246 sk
->sk_prot
->unhash(sk
);
2247 if (inet_csk(sk
)->icsk_bind_hash
&&
2248 !(sk
->sk_userlocks
& SOCK_BINDPORT_LOCK
))
2252 if (oldstate
== TCP_ESTABLISHED
)
2253 TCP_DEC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
2256 /* Change state AFTER socket is unhashed to avoid closed
2257 * socket sitting in hash tables.
2259 inet_sk_state_store(sk
, state
);
2261 EXPORT_SYMBOL_GPL(tcp_set_state
);
2264 * State processing on a close. This implements the state shift for
2265 * sending our FIN frame. Note that we only send a FIN for some
2266 * states. A shutdown() may have already sent the FIN, or we may be
2270 static const unsigned char new_state
[16] = {
2271 /* current state: new state: action: */
2272 [0 /* (Invalid) */] = TCP_CLOSE
,
2273 [TCP_ESTABLISHED
] = TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2274 [TCP_SYN_SENT
] = TCP_CLOSE
,
2275 [TCP_SYN_RECV
] = TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2276 [TCP_FIN_WAIT1
] = TCP_FIN_WAIT1
,
2277 [TCP_FIN_WAIT2
] = TCP_FIN_WAIT2
,
2278 [TCP_TIME_WAIT
] = TCP_CLOSE
,
2279 [TCP_CLOSE
] = TCP_CLOSE
,
2280 [TCP_CLOSE_WAIT
] = TCP_LAST_ACK
| TCP_ACTION_FIN
,
2281 [TCP_LAST_ACK
] = TCP_LAST_ACK
,
2282 [TCP_LISTEN
] = TCP_CLOSE
,
2283 [TCP_CLOSING
] = TCP_CLOSING
,
2284 [TCP_NEW_SYN_RECV
] = TCP_CLOSE
, /* should not happen ! */
2287 static int tcp_close_state(struct sock
*sk
)
2289 int next
= (int)new_state
[sk
->sk_state
];
2290 int ns
= next
& TCP_STATE_MASK
;
2292 tcp_set_state(sk
, ns
);
2294 return next
& TCP_ACTION_FIN
;
2298 * Shutdown the sending side of a connection. Much like close except
2299 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2302 void tcp_shutdown(struct sock
*sk
, int how
)
2304 /* We need to grab some memory, and put together a FIN,
2305 * and then put it into the queue to be sent.
2306 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2308 if (!(how
& SEND_SHUTDOWN
))
2311 /* If we've already sent a FIN, or it's a closed state, skip this. */
2312 if ((1 << sk
->sk_state
) &
2313 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
2314 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
2315 /* Clear out any half completed packets. FIN if needed. */
2316 if (tcp_close_state(sk
))
2320 EXPORT_SYMBOL(tcp_shutdown
);
2322 bool tcp_check_oom(struct sock
*sk
, int shift
)
2324 bool too_many_orphans
, out_of_socket_memory
;
2326 too_many_orphans
= tcp_too_many_orphans(sk
, shift
);
2327 out_of_socket_memory
= tcp_out_of_memory(sk
);
2329 if (too_many_orphans
)
2330 net_info_ratelimited("too many orphaned sockets\n");
2331 if (out_of_socket_memory
)
2332 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2333 return too_many_orphans
|| out_of_socket_memory
;
2336 void tcp_close(struct sock
*sk
, long timeout
)
2338 struct sk_buff
*skb
;
2339 int data_was_unread
= 0;
2343 sk
->sk_shutdown
= SHUTDOWN_MASK
;
2345 if (sk
->sk_state
== TCP_LISTEN
) {
2346 tcp_set_state(sk
, TCP_CLOSE
);
2349 inet_csk_listen_stop(sk
);
2351 goto adjudge_to_death
;
2354 /* We need to flush the recv. buffs. We do this only on the
2355 * descriptor close, not protocol-sourced closes, because the
2356 * reader process may not have drained the data yet!
2358 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
2359 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
;
2361 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
2363 data_was_unread
+= len
;
2369 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2370 if (sk
->sk_state
== TCP_CLOSE
)
2371 goto adjudge_to_death
;
2373 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2374 * data was lost. To witness the awful effects of the old behavior of
2375 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2376 * GET in an FTP client, suspend the process, wait for the client to
2377 * advertise a zero window, then kill -9 the FTP client, wheee...
2378 * Note: timeout is always zero in such a case.
2380 if (unlikely(tcp_sk(sk
)->repair
)) {
2381 sk
->sk_prot
->disconnect(sk
, 0);
2382 } else if (data_was_unread
) {
2383 /* Unread data was tossed, zap the connection. */
2384 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONCLOSE
);
2385 tcp_set_state(sk
, TCP_CLOSE
);
2386 tcp_send_active_reset(sk
, sk
->sk_allocation
);
2387 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
2388 /* Check zero linger _after_ checking for unread data. */
2389 sk
->sk_prot
->disconnect(sk
, 0);
2390 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
2391 } else if (tcp_close_state(sk
)) {
2392 /* We FIN if the application ate all the data before
2393 * zapping the connection.
2396 /* RED-PEN. Formally speaking, we have broken TCP state
2397 * machine. State transitions:
2399 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2400 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2401 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2403 * are legal only when FIN has been sent (i.e. in window),
2404 * rather than queued out of window. Purists blame.
2406 * F.e. "RFC state" is ESTABLISHED,
2407 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2409 * The visible declinations are that sometimes
2410 * we enter time-wait state, when it is not required really
2411 * (harmless), do not send active resets, when they are
2412 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2413 * they look as CLOSING or LAST_ACK for Linux)
2414 * Probably, I missed some more holelets.
2416 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2417 * in a single packet! (May consider it later but will
2418 * probably need API support or TCP_CORK SYN-ACK until
2419 * data is written and socket is closed.)
2424 sk_stream_wait_close(sk
, timeout
);
2427 state
= sk
->sk_state
;
2433 /* remove backlog if any, without releasing ownership. */
2436 percpu_counter_inc(sk
->sk_prot
->orphan_count
);
2438 /* Have we already been destroyed by a softirq or backlog? */
2439 if (state
!= TCP_CLOSE
&& sk
->sk_state
== TCP_CLOSE
)
2442 /* This is a (useful) BSD violating of the RFC. There is a
2443 * problem with TCP as specified in that the other end could
2444 * keep a socket open forever with no application left this end.
2445 * We use a 1 minute timeout (about the same as BSD) then kill
2446 * our end. If they send after that then tough - BUT: long enough
2447 * that we won't make the old 4*rto = almost no time - whoops
2450 * Nope, it was not mistake. It is really desired behaviour
2451 * f.e. on http servers, when such sockets are useless, but
2452 * consume significant resources. Let's do it with special
2453 * linger2 option. --ANK
2456 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
2457 struct tcp_sock
*tp
= tcp_sk(sk
);
2458 if (tp
->linger2
< 0) {
2459 tcp_set_state(sk
, TCP_CLOSE
);
2460 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2461 __NET_INC_STATS(sock_net(sk
),
2462 LINUX_MIB_TCPABORTONLINGER
);
2464 const int tmo
= tcp_fin_time(sk
);
2466 if (tmo
> TCP_TIMEWAIT_LEN
) {
2467 inet_csk_reset_keepalive_timer(sk
,
2468 tmo
- TCP_TIMEWAIT_LEN
);
2470 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
2475 if (sk
->sk_state
!= TCP_CLOSE
) {
2477 if (tcp_check_oom(sk
, 0)) {
2478 tcp_set_state(sk
, TCP_CLOSE
);
2479 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2480 __NET_INC_STATS(sock_net(sk
),
2481 LINUX_MIB_TCPABORTONMEMORY
);
2482 } else if (!check_net(sock_net(sk
))) {
2483 /* Not possible to send reset; just close */
2484 tcp_set_state(sk
, TCP_CLOSE
);
2488 if (sk
->sk_state
== TCP_CLOSE
) {
2489 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
2490 /* We could get here with a non-NULL req if the socket is
2491 * aborted (e.g., closed with unread data) before 3WHS
2495 reqsk_fastopen_remove(sk
, req
, false);
2496 inet_csk_destroy_sock(sk
);
2498 /* Otherwise, socket is reprieved until protocol close. */
2506 EXPORT_SYMBOL(tcp_close
);
2508 /* These states need RST on ABORT according to RFC793 */
2510 static inline bool tcp_need_reset(int state
)
2512 return (1 << state
) &
2513 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
2514 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
2517 static void tcp_rtx_queue_purge(struct sock
*sk
)
2519 struct rb_node
*p
= rb_first(&sk
->tcp_rtx_queue
);
2522 struct sk_buff
*skb
= rb_to_skb(p
);
2525 /* Since we are deleting whole queue, no need to
2526 * list_del(&skb->tcp_tsorted_anchor)
2528 tcp_rtx_queue_unlink(skb
, sk
);
2529 sk_wmem_free_skb(sk
, skb
);
2533 void tcp_write_queue_purge(struct sock
*sk
)
2535 struct sk_buff
*skb
;
2537 tcp_chrono_stop(sk
, TCP_CHRONO_BUSY
);
2538 while ((skb
= __skb_dequeue(&sk
->sk_write_queue
)) != NULL
) {
2539 tcp_skb_tsorted_anchor_cleanup(skb
);
2540 sk_wmem_free_skb(sk
, skb
);
2542 tcp_rtx_queue_purge(sk
);
2543 skb
= sk
->sk_tx_skb_cache
;
2546 sk
->sk_tx_skb_cache
= NULL
;
2548 INIT_LIST_HEAD(&tcp_sk(sk
)->tsorted_sent_queue
);
2550 tcp_clear_all_retrans_hints(tcp_sk(sk
));
2551 tcp_sk(sk
)->packets_out
= 0;
2552 inet_csk(sk
)->icsk_backoff
= 0;
2555 int tcp_disconnect(struct sock
*sk
, int flags
)
2557 struct inet_sock
*inet
= inet_sk(sk
);
2558 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2559 struct tcp_sock
*tp
= tcp_sk(sk
);
2560 int old_state
= sk
->sk_state
;
2562 if (old_state
!= TCP_CLOSE
)
2563 tcp_set_state(sk
, TCP_CLOSE
);
2565 /* ABORT function of RFC793 */
2566 if (old_state
== TCP_LISTEN
) {
2567 inet_csk_listen_stop(sk
);
2568 } else if (unlikely(tp
->repair
)) {
2569 sk
->sk_err
= ECONNABORTED
;
2570 } else if (tcp_need_reset(old_state
) ||
2571 (tp
->snd_nxt
!= tp
->write_seq
&&
2572 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
2573 /* The last check adjusts for discrepancy of Linux wrt. RFC
2576 tcp_send_active_reset(sk
, gfp_any());
2577 sk
->sk_err
= ECONNRESET
;
2578 } else if (old_state
== TCP_SYN_SENT
)
2579 sk
->sk_err
= ECONNRESET
;
2581 tcp_clear_xmit_timers(sk
);
2582 __skb_queue_purge(&sk
->sk_receive_queue
);
2583 if (sk
->sk_rx_skb_cache
) {
2584 __kfree_skb(sk
->sk_rx_skb_cache
);
2585 sk
->sk_rx_skb_cache
= NULL
;
2587 tp
->copied_seq
= tp
->rcv_nxt
;
2589 tcp_write_queue_purge(sk
);
2590 tcp_fastopen_active_disable_ofo_check(sk
);
2591 skb_rbtree_purge(&tp
->out_of_order_queue
);
2593 inet
->inet_dport
= 0;
2595 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
2596 inet_reset_saddr(sk
);
2598 sk
->sk_shutdown
= 0;
2599 sock_reset_flag(sk
, SOCK_DONE
);
2601 tp
->mdev_us
= jiffies_to_usecs(TCP_TIMEOUT_INIT
);
2602 tp
->rcv_rtt_last_tsecr
= 0;
2603 tp
->write_seq
+= tp
->max_window
+ 2;
2604 if (tp
->write_seq
== 0)
2606 icsk
->icsk_backoff
= 0;
2608 icsk
->icsk_probes_out
= 0;
2609 icsk
->icsk_rto
= TCP_TIMEOUT_INIT
;
2610 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
2611 tp
->snd_cwnd
= TCP_INIT_CWND
;
2612 tp
->snd_cwnd_cnt
= 0;
2613 tp
->window_clamp
= 0;
2614 tp
->delivered_ce
= 0;
2615 tcp_set_ca_state(sk
, TCP_CA_Open
);
2616 tp
->is_sack_reneg
= 0;
2617 tcp_clear_retrans(tp
);
2618 inet_csk_delack_init(sk
);
2619 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2620 * issue in __tcp_select_window()
2622 icsk
->icsk_ack
.rcv_mss
= TCP_MIN_MSS
;
2623 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
2625 dst_release(sk
->sk_rx_dst
);
2626 sk
->sk_rx_dst
= NULL
;
2627 tcp_saved_syn_free(tp
);
2628 tp
->compressed_ack
= 0;
2630 tp
->bytes_acked
= 0;
2631 tp
->bytes_received
= 0;
2632 tp
->bytes_retrans
= 0;
2633 tp
->duplicate_sack
[0].start_seq
= 0;
2634 tp
->duplicate_sack
[0].end_seq
= 0;
2637 tp
->retrans_out
= 0;
2639 tp
->tlp_high_seq
= 0;
2640 tp
->last_oow_ack_time
= 0;
2641 /* There's a bubble in the pipe until at least the first ACK. */
2642 tp
->app_limited
= ~0U;
2643 tp
->rack
.mstamp
= 0;
2644 tp
->rack
.advanced
= 0;
2645 tp
->rack
.reo_wnd_steps
= 1;
2646 tp
->rack
.last_delivered
= 0;
2647 tp
->rack
.reo_wnd_persist
= 0;
2648 tp
->rack
.dsack_seen
= 0;
2649 tp
->syn_data_acked
= 0;
2650 tp
->rx_opt
.saw_tstamp
= 0;
2651 tp
->rx_opt
.dsack
= 0;
2652 tp
->rx_opt
.num_sacks
= 0;
2655 /* Clean up fastopen related fields */
2656 tcp_free_fastopen_req(tp
);
2657 inet
->defer_connect
= 0;
2659 WARN_ON(inet
->inet_num
&& !icsk
->icsk_bind_hash
);
2661 if (sk
->sk_frag
.page
) {
2662 put_page(sk
->sk_frag
.page
);
2663 sk
->sk_frag
.page
= NULL
;
2664 sk
->sk_frag
.offset
= 0;
2667 sk
->sk_error_report(sk
);
2670 EXPORT_SYMBOL(tcp_disconnect
);
2672 static inline bool tcp_can_repair_sock(const struct sock
*sk
)
2674 return ns_capable(sock_net(sk
)->user_ns
, CAP_NET_ADMIN
) &&
2675 (sk
->sk_state
!= TCP_LISTEN
);
2678 static int tcp_repair_set_window(struct tcp_sock
*tp
, char __user
*optbuf
, int len
)
2680 struct tcp_repair_window opt
;
2685 if (len
!= sizeof(opt
))
2688 if (copy_from_user(&opt
, optbuf
, sizeof(opt
)))
2691 if (opt
.max_window
< opt
.snd_wnd
)
2694 if (after(opt
.snd_wl1
, tp
->rcv_nxt
+ opt
.rcv_wnd
))
2697 if (after(opt
.rcv_wup
, tp
->rcv_nxt
))
2700 tp
->snd_wl1
= opt
.snd_wl1
;
2701 tp
->snd_wnd
= opt
.snd_wnd
;
2702 tp
->max_window
= opt
.max_window
;
2704 tp
->rcv_wnd
= opt
.rcv_wnd
;
2705 tp
->rcv_wup
= opt
.rcv_wup
;
2710 static int tcp_repair_options_est(struct sock
*sk
,
2711 struct tcp_repair_opt __user
*optbuf
, unsigned int len
)
2713 struct tcp_sock
*tp
= tcp_sk(sk
);
2714 struct tcp_repair_opt opt
;
2716 while (len
>= sizeof(opt
)) {
2717 if (copy_from_user(&opt
, optbuf
, sizeof(opt
)))
2723 switch (opt
.opt_code
) {
2725 tp
->rx_opt
.mss_clamp
= opt
.opt_val
;
2730 u16 snd_wscale
= opt
.opt_val
& 0xFFFF;
2731 u16 rcv_wscale
= opt
.opt_val
>> 16;
2733 if (snd_wscale
> TCP_MAX_WSCALE
|| rcv_wscale
> TCP_MAX_WSCALE
)
2736 tp
->rx_opt
.snd_wscale
= snd_wscale
;
2737 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2738 tp
->rx_opt
.wscale_ok
= 1;
2741 case TCPOPT_SACK_PERM
:
2742 if (opt
.opt_val
!= 0)
2745 tp
->rx_opt
.sack_ok
|= TCP_SACK_SEEN
;
2747 case TCPOPT_TIMESTAMP
:
2748 if (opt
.opt_val
!= 0)
2751 tp
->rx_opt
.tstamp_ok
= 1;
2759 DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled
);
2760 EXPORT_SYMBOL(tcp_tx_delay_enabled
);
2762 static void tcp_enable_tx_delay(void)
2764 if (!static_branch_unlikely(&tcp_tx_delay_enabled
)) {
2765 static int __tcp_tx_delay_enabled
= 0;
2767 if (cmpxchg(&__tcp_tx_delay_enabled
, 0, 1) == 0) {
2768 static_branch_enable(&tcp_tx_delay_enabled
);
2769 pr_info("TCP_TX_DELAY enabled\n");
2775 * Socket option code for TCP.
2777 static int do_tcp_setsockopt(struct sock
*sk
, int level
,
2778 int optname
, char __user
*optval
, unsigned int optlen
)
2780 struct tcp_sock
*tp
= tcp_sk(sk
);
2781 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2782 struct net
*net
= sock_net(sk
);
2786 /* These are data/string values, all the others are ints */
2788 case TCP_CONGESTION
: {
2789 char name
[TCP_CA_NAME_MAX
];
2794 val
= strncpy_from_user(name
, optval
,
2795 min_t(long, TCP_CA_NAME_MAX
-1, optlen
));
2801 err
= tcp_set_congestion_control(sk
, name
, true, true,
2802 ns_capable(sock_net(sk
)->user_ns
,
2808 char name
[TCP_ULP_NAME_MAX
];
2813 val
= strncpy_from_user(name
, optval
,
2814 min_t(long, TCP_ULP_NAME_MAX
- 1,
2821 err
= tcp_set_ulp(sk
, name
);
2825 case TCP_FASTOPEN_KEY
: {
2826 __u8 key
[TCP_FASTOPEN_KEY_BUF_LENGTH
];
2827 __u8
*backup_key
= NULL
;
2829 /* Allow a backup key as well to facilitate key rotation
2830 * First key is the active one.
2832 if (optlen
!= TCP_FASTOPEN_KEY_LENGTH
&&
2833 optlen
!= TCP_FASTOPEN_KEY_BUF_LENGTH
)
2836 if (copy_from_user(key
, optval
, optlen
))
2839 if (optlen
== TCP_FASTOPEN_KEY_BUF_LENGTH
)
2840 backup_key
= key
+ TCP_FASTOPEN_KEY_LENGTH
;
2842 return tcp_fastopen_reset_cipher(net
, sk
, key
, backup_key
);
2849 if (optlen
< sizeof(int))
2852 if (get_user(val
, (int __user
*)optval
))
2859 /* Values greater than interface MTU won't take effect. However
2860 * at the point when this call is done we typically don't yet
2861 * know which interface is going to be used
2863 if (val
&& (val
< TCP_MIN_MSS
|| val
> MAX_TCP_WINDOW
)) {
2867 tp
->rx_opt
.user_mss
= val
;
2872 /* TCP_NODELAY is weaker than TCP_CORK, so that
2873 * this option on corked socket is remembered, but
2874 * it is not activated until cork is cleared.
2876 * However, when TCP_NODELAY is set we make
2877 * an explicit push, which overrides even TCP_CORK
2878 * for currently queued segments.
2880 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
2881 tcp_push_pending_frames(sk
);
2883 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
2887 case TCP_THIN_LINEAR_TIMEOUTS
:
2888 if (val
< 0 || val
> 1)
2894 case TCP_THIN_DUPACK
:
2895 if (val
< 0 || val
> 1)
2900 if (!tcp_can_repair_sock(sk
))
2902 else if (val
== TCP_REPAIR_ON
) {
2904 sk
->sk_reuse
= SK_FORCE_REUSE
;
2905 tp
->repair_queue
= TCP_NO_QUEUE
;
2906 } else if (val
== TCP_REPAIR_OFF
) {
2908 sk
->sk_reuse
= SK_NO_REUSE
;
2909 tcp_send_window_probe(sk
);
2910 } else if (val
== TCP_REPAIR_OFF_NO_WP
) {
2912 sk
->sk_reuse
= SK_NO_REUSE
;
2918 case TCP_REPAIR_QUEUE
:
2921 else if ((unsigned int)val
< TCP_QUEUES_NR
)
2922 tp
->repair_queue
= val
;
2928 if (sk
->sk_state
!= TCP_CLOSE
)
2930 else if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2931 tp
->write_seq
= val
;
2932 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2938 case TCP_REPAIR_OPTIONS
:
2941 else if (sk
->sk_state
== TCP_ESTABLISHED
)
2942 err
= tcp_repair_options_est(sk
,
2943 (struct tcp_repair_opt __user
*)optval
,
2950 /* When set indicates to always queue non-full frames.
2951 * Later the user clears this option and we transmit
2952 * any pending partial frames in the queue. This is
2953 * meant to be used alongside sendfile() to get properly
2954 * filled frames when the user (for example) must write
2955 * out headers with a write() call first and then use
2956 * sendfile to send out the data parts.
2958 * TCP_CORK can be set together with TCP_NODELAY and it is
2959 * stronger than TCP_NODELAY.
2962 tp
->nonagle
|= TCP_NAGLE_CORK
;
2964 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
2965 if (tp
->nonagle
&TCP_NAGLE_OFF
)
2966 tp
->nonagle
|= TCP_NAGLE_PUSH
;
2967 tcp_push_pending_frames(sk
);
2972 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
2975 tp
->keepalive_time
= val
* HZ
;
2976 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
2977 !((1 << sk
->sk_state
) &
2978 (TCPF_CLOSE
| TCPF_LISTEN
))) {
2979 u32 elapsed
= keepalive_time_elapsed(tp
);
2980 if (tp
->keepalive_time
> elapsed
)
2981 elapsed
= tp
->keepalive_time
- elapsed
;
2984 inet_csk_reset_keepalive_timer(sk
, elapsed
);
2989 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
2992 tp
->keepalive_intvl
= val
* HZ
;
2995 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
2998 tp
->keepalive_probes
= val
;
3001 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
3004 icsk
->icsk_syn_retries
= val
;
3008 if (val
< 0 || val
> 1)
3017 else if (val
> net
->ipv4
.sysctl_tcp_fin_timeout
/ HZ
)
3020 tp
->linger2
= val
* HZ
;
3023 case TCP_DEFER_ACCEPT
:
3024 /* Translate value in seconds to number of retransmits */
3025 icsk
->icsk_accept_queue
.rskq_defer_accept
=
3026 secs_to_retrans(val
, TCP_TIMEOUT_INIT
/ HZ
,
3030 case TCP_WINDOW_CLAMP
:
3032 if (sk
->sk_state
!= TCP_CLOSE
) {
3036 tp
->window_clamp
= 0;
3038 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
3039 SOCK_MIN_RCVBUF
/ 2 : val
;
3044 inet_csk_enter_pingpong_mode(sk
);
3046 inet_csk_exit_pingpong_mode(sk
);
3047 if ((1 << sk
->sk_state
) &
3048 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
3049 inet_csk_ack_scheduled(sk
)) {
3050 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
3051 tcp_cleanup_rbuf(sk
, 1);
3053 inet_csk_enter_pingpong_mode(sk
);
3058 #ifdef CONFIG_TCP_MD5SIG
3060 case TCP_MD5SIG_EXT
:
3061 if ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
))
3062 err
= tp
->af_specific
->md5_parse(sk
, optname
, optval
, optlen
);
3067 case TCP_USER_TIMEOUT
:
3068 /* Cap the max time in ms TCP will retry or probe the window
3069 * before giving up and aborting (ETIMEDOUT) a connection.
3074 icsk
->icsk_user_timeout
= val
;
3078 if (val
>= 0 && ((1 << sk
->sk_state
) & (TCPF_CLOSE
|
3080 tcp_fastopen_init_key_once(net
);
3082 fastopen_queue_tune(sk
, val
);
3087 case TCP_FASTOPEN_CONNECT
:
3088 if (val
> 1 || val
< 0) {
3090 } else if (net
->ipv4
.sysctl_tcp_fastopen
& TFO_CLIENT_ENABLE
) {
3091 if (sk
->sk_state
== TCP_CLOSE
)
3092 tp
->fastopen_connect
= val
;
3099 case TCP_FASTOPEN_NO_COOKIE
:
3100 if (val
> 1 || val
< 0)
3102 else if (!((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
3105 tp
->fastopen_no_cookie
= val
;
3111 tp
->tsoffset
= val
- tcp_time_stamp_raw();
3113 case TCP_REPAIR_WINDOW
:
3114 err
= tcp_repair_set_window(tp
, optval
, optlen
);
3116 case TCP_NOTSENT_LOWAT
:
3117 tp
->notsent_lowat
= val
;
3118 sk
->sk_write_space(sk
);
3121 if (val
> 1 || val
< 0)
3124 tp
->recvmsg_inq
= val
;
3128 tcp_enable_tx_delay();
3129 tp
->tcp_tx_delay
= val
;
3140 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
3141 unsigned int optlen
)
3143 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3145 if (level
!= SOL_TCP
)
3146 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
3148 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
3150 EXPORT_SYMBOL(tcp_setsockopt
);
3152 #ifdef CONFIG_COMPAT
3153 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
3154 char __user
*optval
, unsigned int optlen
)
3156 if (level
!= SOL_TCP
)
3157 return inet_csk_compat_setsockopt(sk
, level
, optname
,
3159 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
3161 EXPORT_SYMBOL(compat_tcp_setsockopt
);
3164 static void tcp_get_info_chrono_stats(const struct tcp_sock
*tp
,
3165 struct tcp_info
*info
)
3167 u64 stats
[__TCP_CHRONO_MAX
], total
= 0;
3170 for (i
= TCP_CHRONO_BUSY
; i
< __TCP_CHRONO_MAX
; ++i
) {
3171 stats
[i
] = tp
->chrono_stat
[i
- 1];
3172 if (i
== tp
->chrono_type
)
3173 stats
[i
] += tcp_jiffies32
- tp
->chrono_start
;
3174 stats
[i
] *= USEC_PER_SEC
/ HZ
;
3178 info
->tcpi_busy_time
= total
;
3179 info
->tcpi_rwnd_limited
= stats
[TCP_CHRONO_RWND_LIMITED
];
3180 info
->tcpi_sndbuf_limited
= stats
[TCP_CHRONO_SNDBUF_LIMITED
];
3183 /* Return information about state of tcp endpoint in API format. */
3184 void tcp_get_info(struct sock
*sk
, struct tcp_info
*info
)
3186 const struct tcp_sock
*tp
= tcp_sk(sk
); /* iff sk_type == SOCK_STREAM */
3187 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3193 memset(info
, 0, sizeof(*info
));
3194 if (sk
->sk_type
!= SOCK_STREAM
)
3197 info
->tcpi_state
= inet_sk_state_load(sk
);
3199 /* Report meaningful fields for all TCP states, including listeners */
3200 rate
= READ_ONCE(sk
->sk_pacing_rate
);
3201 rate64
= (rate
!= ~0UL) ? rate
: ~0ULL;
3202 info
->tcpi_pacing_rate
= rate64
;
3204 rate
= READ_ONCE(sk
->sk_max_pacing_rate
);
3205 rate64
= (rate
!= ~0UL) ? rate
: ~0ULL;
3206 info
->tcpi_max_pacing_rate
= rate64
;
3208 info
->tcpi_reordering
= tp
->reordering
;
3209 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
3211 if (info
->tcpi_state
== TCP_LISTEN
) {
3212 /* listeners aliased fields :
3213 * tcpi_unacked -> Number of children ready for accept()
3214 * tcpi_sacked -> max backlog
3216 info
->tcpi_unacked
= sk
->sk_ack_backlog
;
3217 info
->tcpi_sacked
= sk
->sk_max_ack_backlog
;
3221 slow
= lock_sock_fast(sk
);
3223 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
3224 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
3225 info
->tcpi_probes
= icsk
->icsk_probes_out
;
3226 info
->tcpi_backoff
= icsk
->icsk_backoff
;
3228 if (tp
->rx_opt
.tstamp_ok
)
3229 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
3230 if (tcp_is_sack(tp
))
3231 info
->tcpi_options
|= TCPI_OPT_SACK
;
3232 if (tp
->rx_opt
.wscale_ok
) {
3233 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
3234 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
3235 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
3238 if (tp
->ecn_flags
& TCP_ECN_OK
)
3239 info
->tcpi_options
|= TCPI_OPT_ECN
;
3240 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
3241 info
->tcpi_options
|= TCPI_OPT_ECN_SEEN
;
3242 if (tp
->syn_data_acked
)
3243 info
->tcpi_options
|= TCPI_OPT_SYN_DATA
;
3245 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
3246 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
3247 info
->tcpi_snd_mss
= tp
->mss_cache
;
3248 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
3250 info
->tcpi_unacked
= tp
->packets_out
;
3251 info
->tcpi_sacked
= tp
->sacked_out
;
3253 info
->tcpi_lost
= tp
->lost_out
;
3254 info
->tcpi_retrans
= tp
->retrans_out
;
3256 now
= tcp_jiffies32
;
3257 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
3258 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
3259 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
3261 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
3262 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
3263 info
->tcpi_rtt
= tp
->srtt_us
>> 3;
3264 info
->tcpi_rttvar
= tp
->mdev_us
>> 2;
3265 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
3266 info
->tcpi_advmss
= tp
->advmss
;
3268 info
->tcpi_rcv_rtt
= tp
->rcv_rtt_est
.rtt_us
>> 3;
3269 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
3271 info
->tcpi_total_retrans
= tp
->total_retrans
;
3273 info
->tcpi_bytes_acked
= tp
->bytes_acked
;
3274 info
->tcpi_bytes_received
= tp
->bytes_received
;
3275 info
->tcpi_notsent_bytes
= max_t(int, 0, tp
->write_seq
- tp
->snd_nxt
);
3276 tcp_get_info_chrono_stats(tp
, info
);
3278 info
->tcpi_segs_out
= tp
->segs_out
;
3279 info
->tcpi_segs_in
= tp
->segs_in
;
3281 info
->tcpi_min_rtt
= tcp_min_rtt(tp
);
3282 info
->tcpi_data_segs_in
= tp
->data_segs_in
;
3283 info
->tcpi_data_segs_out
= tp
->data_segs_out
;
3285 info
->tcpi_delivery_rate_app_limited
= tp
->rate_app_limited
? 1 : 0;
3286 rate64
= tcp_compute_delivery_rate(tp
);
3288 info
->tcpi_delivery_rate
= rate64
;
3289 info
->tcpi_delivered
= tp
->delivered
;
3290 info
->tcpi_delivered_ce
= tp
->delivered_ce
;
3291 info
->tcpi_bytes_sent
= tp
->bytes_sent
;
3292 info
->tcpi_bytes_retrans
= tp
->bytes_retrans
;
3293 info
->tcpi_dsack_dups
= tp
->dsack_dups
;
3294 info
->tcpi_reord_seen
= tp
->reord_seen
;
3295 unlock_sock_fast(sk
, slow
);
3297 EXPORT_SYMBOL_GPL(tcp_get_info
);
3299 static size_t tcp_opt_stats_get_size(void)
3302 nla_total_size_64bit(sizeof(u64
)) + /* TCP_NLA_BUSY */
3303 nla_total_size_64bit(sizeof(u64
)) + /* TCP_NLA_RWND_LIMITED */
3304 nla_total_size_64bit(sizeof(u64
)) + /* TCP_NLA_SNDBUF_LIMITED */
3305 nla_total_size_64bit(sizeof(u64
)) + /* TCP_NLA_DATA_SEGS_OUT */
3306 nla_total_size_64bit(sizeof(u64
)) + /* TCP_NLA_TOTAL_RETRANS */
3307 nla_total_size_64bit(sizeof(u64
)) + /* TCP_NLA_PACING_RATE */
3308 nla_total_size_64bit(sizeof(u64
)) + /* TCP_NLA_DELIVERY_RATE */
3309 nla_total_size(sizeof(u32
)) + /* TCP_NLA_SND_CWND */
3310 nla_total_size(sizeof(u32
)) + /* TCP_NLA_REORDERING */
3311 nla_total_size(sizeof(u32
)) + /* TCP_NLA_MIN_RTT */
3312 nla_total_size(sizeof(u8
)) + /* TCP_NLA_RECUR_RETRANS */
3313 nla_total_size(sizeof(u8
)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3314 nla_total_size(sizeof(u32
)) + /* TCP_NLA_SNDQ_SIZE */
3315 nla_total_size(sizeof(u8
)) + /* TCP_NLA_CA_STATE */
3316 nla_total_size(sizeof(u32
)) + /* TCP_NLA_SND_SSTHRESH */
3317 nla_total_size(sizeof(u32
)) + /* TCP_NLA_DELIVERED */
3318 nla_total_size(sizeof(u32
)) + /* TCP_NLA_DELIVERED_CE */
3319 nla_total_size_64bit(sizeof(u64
)) + /* TCP_NLA_BYTES_SENT */
3320 nla_total_size_64bit(sizeof(u64
)) + /* TCP_NLA_BYTES_RETRANS */
3321 nla_total_size(sizeof(u32
)) + /* TCP_NLA_DSACK_DUPS */
3322 nla_total_size(sizeof(u32
)) + /* TCP_NLA_REORD_SEEN */
3323 nla_total_size(sizeof(u32
)) + /* TCP_NLA_SRTT */
3327 struct sk_buff
*tcp_get_timestamping_opt_stats(const struct sock
*sk
)
3329 const struct tcp_sock
*tp
= tcp_sk(sk
);
3330 struct sk_buff
*stats
;
3331 struct tcp_info info
;
3335 stats
= alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC
);
3339 tcp_get_info_chrono_stats(tp
, &info
);
3340 nla_put_u64_64bit(stats
, TCP_NLA_BUSY
,
3341 info
.tcpi_busy_time
, TCP_NLA_PAD
);
3342 nla_put_u64_64bit(stats
, TCP_NLA_RWND_LIMITED
,
3343 info
.tcpi_rwnd_limited
, TCP_NLA_PAD
);
3344 nla_put_u64_64bit(stats
, TCP_NLA_SNDBUF_LIMITED
,
3345 info
.tcpi_sndbuf_limited
, TCP_NLA_PAD
);
3346 nla_put_u64_64bit(stats
, TCP_NLA_DATA_SEGS_OUT
,
3347 tp
->data_segs_out
, TCP_NLA_PAD
);
3348 nla_put_u64_64bit(stats
, TCP_NLA_TOTAL_RETRANS
,
3349 tp
->total_retrans
, TCP_NLA_PAD
);
3351 rate
= READ_ONCE(sk
->sk_pacing_rate
);
3352 rate64
= (rate
!= ~0UL) ? rate
: ~0ULL;
3353 nla_put_u64_64bit(stats
, TCP_NLA_PACING_RATE
, rate64
, TCP_NLA_PAD
);
3355 rate64
= tcp_compute_delivery_rate(tp
);
3356 nla_put_u64_64bit(stats
, TCP_NLA_DELIVERY_RATE
, rate64
, TCP_NLA_PAD
);
3358 nla_put_u32(stats
, TCP_NLA_SND_CWND
, tp
->snd_cwnd
);
3359 nla_put_u32(stats
, TCP_NLA_REORDERING
, tp
->reordering
);
3360 nla_put_u32(stats
, TCP_NLA_MIN_RTT
, tcp_min_rtt(tp
));
3362 nla_put_u8(stats
, TCP_NLA_RECUR_RETRANS
, inet_csk(sk
)->icsk_retransmits
);
3363 nla_put_u8(stats
, TCP_NLA_DELIVERY_RATE_APP_LMT
, !!tp
->rate_app_limited
);
3364 nla_put_u32(stats
, TCP_NLA_SND_SSTHRESH
, tp
->snd_ssthresh
);
3365 nla_put_u32(stats
, TCP_NLA_DELIVERED
, tp
->delivered
);
3366 nla_put_u32(stats
, TCP_NLA_DELIVERED_CE
, tp
->delivered_ce
);
3368 nla_put_u32(stats
, TCP_NLA_SNDQ_SIZE
, tp
->write_seq
- tp
->snd_una
);
3369 nla_put_u8(stats
, TCP_NLA_CA_STATE
, inet_csk(sk
)->icsk_ca_state
);
3371 nla_put_u64_64bit(stats
, TCP_NLA_BYTES_SENT
, tp
->bytes_sent
,
3373 nla_put_u64_64bit(stats
, TCP_NLA_BYTES_RETRANS
, tp
->bytes_retrans
,
3375 nla_put_u32(stats
, TCP_NLA_DSACK_DUPS
, tp
->dsack_dups
);
3376 nla_put_u32(stats
, TCP_NLA_REORD_SEEN
, tp
->reord_seen
);
3377 nla_put_u32(stats
, TCP_NLA_SRTT
, tp
->srtt_us
>> 3);
3382 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
3383 int optname
, char __user
*optval
, int __user
*optlen
)
3385 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3386 struct tcp_sock
*tp
= tcp_sk(sk
);
3387 struct net
*net
= sock_net(sk
);
3390 if (get_user(len
, optlen
))
3393 len
= min_t(unsigned int, len
, sizeof(int));
3400 val
= tp
->mss_cache
;
3401 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
3402 val
= tp
->rx_opt
.user_mss
;
3404 val
= tp
->rx_opt
.mss_clamp
;
3407 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
3410 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
3413 val
= keepalive_time_when(tp
) / HZ
;
3416 val
= keepalive_intvl_when(tp
) / HZ
;
3419 val
= keepalive_probes(tp
);
3422 val
= icsk
->icsk_syn_retries
? : net
->ipv4
.sysctl_tcp_syn_retries
;
3427 val
= (val
? : net
->ipv4
.sysctl_tcp_fin_timeout
) / HZ
;
3429 case TCP_DEFER_ACCEPT
:
3430 val
= retrans_to_secs(icsk
->icsk_accept_queue
.rskq_defer_accept
,
3431 TCP_TIMEOUT_INIT
/ HZ
, TCP_RTO_MAX
/ HZ
);
3433 case TCP_WINDOW_CLAMP
:
3434 val
= tp
->window_clamp
;
3437 struct tcp_info info
;
3439 if (get_user(len
, optlen
))
3442 tcp_get_info(sk
, &info
);
3444 len
= min_t(unsigned int, len
, sizeof(info
));
3445 if (put_user(len
, optlen
))
3447 if (copy_to_user(optval
, &info
, len
))
3452 const struct tcp_congestion_ops
*ca_ops
;
3453 union tcp_cc_info info
;
3457 if (get_user(len
, optlen
))
3460 ca_ops
= icsk
->icsk_ca_ops
;
3461 if (ca_ops
&& ca_ops
->get_info
)
3462 sz
= ca_ops
->get_info(sk
, ~0U, &attr
, &info
);
3464 len
= min_t(unsigned int, len
, sz
);
3465 if (put_user(len
, optlen
))
3467 if (copy_to_user(optval
, &info
, len
))
3472 val
= !inet_csk_in_pingpong_mode(sk
);
3475 case TCP_CONGESTION
:
3476 if (get_user(len
, optlen
))
3478 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
3479 if (put_user(len
, optlen
))
3481 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
3486 if (get_user(len
, optlen
))
3488 len
= min_t(unsigned int, len
, TCP_ULP_NAME_MAX
);
3489 if (!icsk
->icsk_ulp_ops
) {
3490 if (put_user(0, optlen
))
3494 if (put_user(len
, optlen
))
3496 if (copy_to_user(optval
, icsk
->icsk_ulp_ops
->name
, len
))
3500 case TCP_FASTOPEN_KEY
: {
3501 __u8 key
[TCP_FASTOPEN_KEY_BUF_LENGTH
];
3502 struct tcp_fastopen_context
*ctx
;
3503 unsigned int key_len
= 0;
3505 if (get_user(len
, optlen
))
3509 ctx
= rcu_dereference(icsk
->icsk_accept_queue
.fastopenq
.ctx
);
3511 key_len
= tcp_fastopen_context_len(ctx
) *
3512 TCP_FASTOPEN_KEY_LENGTH
;
3513 memcpy(&key
[0], &ctx
->key
[0], key_len
);
3517 len
= min_t(unsigned int, len
, key_len
);
3518 if (put_user(len
, optlen
))
3520 if (copy_to_user(optval
, key
, len
))
3524 case TCP_THIN_LINEAR_TIMEOUTS
:
3528 case TCP_THIN_DUPACK
:
3536 case TCP_REPAIR_QUEUE
:
3538 val
= tp
->repair_queue
;
3543 case TCP_REPAIR_WINDOW
: {
3544 struct tcp_repair_window opt
;
3546 if (get_user(len
, optlen
))
3549 if (len
!= sizeof(opt
))
3555 opt
.snd_wl1
= tp
->snd_wl1
;
3556 opt
.snd_wnd
= tp
->snd_wnd
;
3557 opt
.max_window
= tp
->max_window
;
3558 opt
.rcv_wnd
= tp
->rcv_wnd
;
3559 opt
.rcv_wup
= tp
->rcv_wup
;
3561 if (copy_to_user(optval
, &opt
, len
))
3566 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
3567 val
= tp
->write_seq
;
3568 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
3574 case TCP_USER_TIMEOUT
:
3575 val
= icsk
->icsk_user_timeout
;
3579 val
= icsk
->icsk_accept_queue
.fastopenq
.max_qlen
;
3582 case TCP_FASTOPEN_CONNECT
:
3583 val
= tp
->fastopen_connect
;
3586 case TCP_FASTOPEN_NO_COOKIE
:
3587 val
= tp
->fastopen_no_cookie
;
3591 val
= tp
->tcp_tx_delay
;
3595 val
= tcp_time_stamp_raw() + tp
->tsoffset
;
3597 case TCP_NOTSENT_LOWAT
:
3598 val
= tp
->notsent_lowat
;
3601 val
= tp
->recvmsg_inq
;
3606 case TCP_SAVED_SYN
: {
3607 if (get_user(len
, optlen
))
3611 if (tp
->saved_syn
) {
3612 if (len
< tp
->saved_syn
[0]) {
3613 if (put_user(tp
->saved_syn
[0], optlen
)) {
3620 len
= tp
->saved_syn
[0];
3621 if (put_user(len
, optlen
)) {
3625 if (copy_to_user(optval
, tp
->saved_syn
+ 1, len
)) {
3629 tcp_saved_syn_free(tp
);
3634 if (put_user(len
, optlen
))
3640 case TCP_ZEROCOPY_RECEIVE
: {
3641 struct tcp_zerocopy_receive zc
;
3644 if (get_user(len
, optlen
))
3646 if (len
!= sizeof(zc
))
3648 if (copy_from_user(&zc
, optval
, len
))
3651 err
= tcp_zerocopy_receive(sk
, &zc
);
3653 if (!err
&& copy_to_user(optval
, &zc
, len
))
3659 return -ENOPROTOOPT
;
3662 if (put_user(len
, optlen
))
3664 if (copy_to_user(optval
, &val
, len
))
3669 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
3672 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3674 if (level
!= SOL_TCP
)
3675 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
3677 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
3679 EXPORT_SYMBOL(tcp_getsockopt
);
3681 #ifdef CONFIG_COMPAT
3682 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
3683 char __user
*optval
, int __user
*optlen
)
3685 if (level
!= SOL_TCP
)
3686 return inet_csk_compat_getsockopt(sk
, level
, optname
,
3688 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
3690 EXPORT_SYMBOL(compat_tcp_getsockopt
);
3693 #ifdef CONFIG_TCP_MD5SIG
3694 static DEFINE_PER_CPU(struct tcp_md5sig_pool
, tcp_md5sig_pool
);
3695 static DEFINE_MUTEX(tcp_md5sig_mutex
);
3696 static bool tcp_md5sig_pool_populated
= false;
3698 static void __tcp_alloc_md5sig_pool(void)
3700 struct crypto_ahash
*hash
;
3703 hash
= crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC
);
3707 for_each_possible_cpu(cpu
) {
3708 void *scratch
= per_cpu(tcp_md5sig_pool
, cpu
).scratch
;
3709 struct ahash_request
*req
;
3712 scratch
= kmalloc_node(sizeof(union tcp_md5sum_block
) +
3713 sizeof(struct tcphdr
),
3718 per_cpu(tcp_md5sig_pool
, cpu
).scratch
= scratch
;
3720 if (per_cpu(tcp_md5sig_pool
, cpu
).md5_req
)
3723 req
= ahash_request_alloc(hash
, GFP_KERNEL
);
3727 ahash_request_set_callback(req
, 0, NULL
, NULL
);
3729 per_cpu(tcp_md5sig_pool
, cpu
).md5_req
= req
;
3731 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3732 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3735 tcp_md5sig_pool_populated
= true;
3738 bool tcp_alloc_md5sig_pool(void)
3740 if (unlikely(!tcp_md5sig_pool_populated
)) {
3741 mutex_lock(&tcp_md5sig_mutex
);
3743 if (!tcp_md5sig_pool_populated
) {
3744 __tcp_alloc_md5sig_pool();
3745 if (tcp_md5sig_pool_populated
)
3746 static_branch_inc(&tcp_md5_needed
);
3749 mutex_unlock(&tcp_md5sig_mutex
);
3751 return tcp_md5sig_pool_populated
;
3753 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
3757 * tcp_get_md5sig_pool - get md5sig_pool for this user
3759 * We use percpu structure, so if we succeed, we exit with preemption
3760 * and BH disabled, to make sure another thread or softirq handling
3761 * wont try to get same context.
3763 struct tcp_md5sig_pool
*tcp_get_md5sig_pool(void)
3767 if (tcp_md5sig_pool_populated
) {
3768 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3770 return this_cpu_ptr(&tcp_md5sig_pool
);
3775 EXPORT_SYMBOL(tcp_get_md5sig_pool
);
3777 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool
*hp
,
3778 const struct sk_buff
*skb
, unsigned int header_len
)
3780 struct scatterlist sg
;
3781 const struct tcphdr
*tp
= tcp_hdr(skb
);
3782 struct ahash_request
*req
= hp
->md5_req
;
3784 const unsigned int head_data_len
= skb_headlen(skb
) > header_len
?
3785 skb_headlen(skb
) - header_len
: 0;
3786 const struct skb_shared_info
*shi
= skb_shinfo(skb
);
3787 struct sk_buff
*frag_iter
;
3789 sg_init_table(&sg
, 1);
3791 sg_set_buf(&sg
, ((u8
*) tp
) + header_len
, head_data_len
);
3792 ahash_request_set_crypt(req
, &sg
, NULL
, head_data_len
);
3793 if (crypto_ahash_update(req
))
3796 for (i
= 0; i
< shi
->nr_frags
; ++i
) {
3797 const struct skb_frag_struct
*f
= &shi
->frags
[i
];
3798 unsigned int offset
= f
->page_offset
;
3799 struct page
*page
= skb_frag_page(f
) + (offset
>> PAGE_SHIFT
);
3801 sg_set_page(&sg
, page
, skb_frag_size(f
),
3802 offset_in_page(offset
));
3803 ahash_request_set_crypt(req
, &sg
, NULL
, skb_frag_size(f
));
3804 if (crypto_ahash_update(req
))
3808 skb_walk_frags(skb
, frag_iter
)
3809 if (tcp_md5_hash_skb_data(hp
, frag_iter
, 0))
3814 EXPORT_SYMBOL(tcp_md5_hash_skb_data
);
3816 int tcp_md5_hash_key(struct tcp_md5sig_pool
*hp
, const struct tcp_md5sig_key
*key
)
3818 struct scatterlist sg
;
3820 sg_init_one(&sg
, key
->key
, key
->keylen
);
3821 ahash_request_set_crypt(hp
->md5_req
, &sg
, NULL
, key
->keylen
);
3822 return crypto_ahash_update(hp
->md5_req
);
3824 EXPORT_SYMBOL(tcp_md5_hash_key
);
3828 void tcp_done(struct sock
*sk
)
3830 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
3832 if (sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
3833 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ATTEMPTFAILS
);
3835 tcp_set_state(sk
, TCP_CLOSE
);
3836 tcp_clear_xmit_timers(sk
);
3838 reqsk_fastopen_remove(sk
, req
, false);
3840 sk
->sk_shutdown
= SHUTDOWN_MASK
;
3842 if (!sock_flag(sk
, SOCK_DEAD
))
3843 sk
->sk_state_change(sk
);
3845 inet_csk_destroy_sock(sk
);
3847 EXPORT_SYMBOL_GPL(tcp_done
);
3849 int tcp_abort(struct sock
*sk
, int err
)
3851 if (!sk_fullsock(sk
)) {
3852 if (sk
->sk_state
== TCP_NEW_SYN_RECV
) {
3853 struct request_sock
*req
= inet_reqsk(sk
);
3856 inet_csk_reqsk_queue_drop(req
->rsk_listener
, req
);
3863 /* Don't race with userspace socket closes such as tcp_close. */
3866 if (sk
->sk_state
== TCP_LISTEN
) {
3867 tcp_set_state(sk
, TCP_CLOSE
);
3868 inet_csk_listen_stop(sk
);
3871 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3875 if (!sock_flag(sk
, SOCK_DEAD
)) {
3877 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3879 sk
->sk_error_report(sk
);
3880 if (tcp_need_reset(sk
->sk_state
))
3881 tcp_send_active_reset(sk
, GFP_ATOMIC
);
3887 tcp_write_queue_purge(sk
);
3891 EXPORT_SYMBOL_GPL(tcp_abort
);
3893 extern struct tcp_congestion_ops tcp_reno
;
3895 static __initdata
unsigned long thash_entries
;
3896 static int __init
set_thash_entries(char *str
)
3903 ret
= kstrtoul(str
, 0, &thash_entries
);
3909 __setup("thash_entries=", set_thash_entries
);
3911 static void __init
tcp_init_mem(void)
3913 unsigned long limit
= nr_free_buffer_pages() / 16;
3915 limit
= max(limit
, 128UL);
3916 sysctl_tcp_mem
[0] = limit
/ 4 * 3; /* 4.68 % */
3917 sysctl_tcp_mem
[1] = limit
; /* 6.25 % */
3918 sysctl_tcp_mem
[2] = sysctl_tcp_mem
[0] * 2; /* 9.37 % */
3921 void __init
tcp_init(void)
3923 int max_rshare
, max_wshare
, cnt
;
3924 unsigned long limit
;
3927 BUILD_BUG_ON(TCP_MIN_SND_MSS
<= MAX_TCP_OPTION_SPACE
);
3928 BUILD_BUG_ON(sizeof(struct tcp_skb_cb
) >
3929 FIELD_SIZEOF(struct sk_buff
, cb
));
3931 percpu_counter_init(&tcp_sockets_allocated
, 0, GFP_KERNEL
);
3932 percpu_counter_init(&tcp_orphan_count
, 0, GFP_KERNEL
);
3933 inet_hashinfo_init(&tcp_hashinfo
);
3934 inet_hashinfo2_init(&tcp_hashinfo
, "tcp_listen_portaddr_hash",
3935 thash_entries
, 21, /* one slot per 2 MB*/
3937 tcp_hashinfo
.bind_bucket_cachep
=
3938 kmem_cache_create("tcp_bind_bucket",
3939 sizeof(struct inet_bind_bucket
), 0,
3940 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
);
3942 /* Size and allocate the main established and bind bucket
3945 * The methodology is similar to that of the buffer cache.
3947 tcp_hashinfo
.ehash
=
3948 alloc_large_system_hash("TCP established",
3949 sizeof(struct inet_ehash_bucket
),
3951 17, /* one slot per 128 KB of memory */
3954 &tcp_hashinfo
.ehash_mask
,
3956 thash_entries
? 0 : 512 * 1024);
3957 for (i
= 0; i
<= tcp_hashinfo
.ehash_mask
; i
++)
3958 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].chain
, i
);
3960 if (inet_ehash_locks_alloc(&tcp_hashinfo
))
3961 panic("TCP: failed to alloc ehash_locks");
3962 tcp_hashinfo
.bhash
=
3963 alloc_large_system_hash("TCP bind",
3964 sizeof(struct inet_bind_hashbucket
),
3965 tcp_hashinfo
.ehash_mask
+ 1,
3966 17, /* one slot per 128 KB of memory */
3968 &tcp_hashinfo
.bhash_size
,
3972 tcp_hashinfo
.bhash_size
= 1U << tcp_hashinfo
.bhash_size
;
3973 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
3974 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
3975 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
3979 cnt
= tcp_hashinfo
.ehash_mask
+ 1;
3980 sysctl_tcp_max_orphans
= cnt
/ 2;
3983 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3984 limit
= nr_free_buffer_pages() << (PAGE_SHIFT
- 7);
3985 max_wshare
= min(4UL*1024*1024, limit
);
3986 max_rshare
= min(6UL*1024*1024, limit
);
3988 init_net
.ipv4
.sysctl_tcp_wmem
[0] = SK_MEM_QUANTUM
;
3989 init_net
.ipv4
.sysctl_tcp_wmem
[1] = 16*1024;
3990 init_net
.ipv4
.sysctl_tcp_wmem
[2] = max(64*1024, max_wshare
);
3992 init_net
.ipv4
.sysctl_tcp_rmem
[0] = SK_MEM_QUANTUM
;
3993 init_net
.ipv4
.sysctl_tcp_rmem
[1] = 131072;
3994 init_net
.ipv4
.sysctl_tcp_rmem
[2] = max(131072, max_rshare
);
3996 pr_info("Hash tables configured (established %u bind %u)\n",
3997 tcp_hashinfo
.ehash_mask
+ 1, tcp_hashinfo
.bhash_size
);
4001 BUG_ON(tcp_register_congestion_control(&tcp_reno
) != 0);