2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
21 * Alan Cox : Numerous verify_area() calls
22 * Alan Cox : Set the ACK bit on a reset
23 * Alan Cox : Stopped it crashing if it closed while
24 * sk->inuse=1 and was trying to connect
26 * Alan Cox : All icmp error handling was broken
27 * pointers passed where wrong and the
28 * socket was looked up backwards. Nobody
29 * tested any icmp error code obviously.
30 * Alan Cox : tcp_err() now handled properly. It
31 * wakes people on errors. poll
32 * behaves and the icmp error race
33 * has gone by moving it into sock.c
34 * Alan Cox : tcp_send_reset() fixed to work for
35 * everything not just packets for
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
40 * Herp Rosmanith : More reset fixes
41 * Alan Cox : No longer acks invalid rst frames.
42 * Acking any kind of RST is right out.
43 * Alan Cox : Sets an ignore me flag on an rst
44 * receive otherwise odd bits of prattle
46 * Alan Cox : Fixed another acking RST frame bug.
47 * Should stop LAN workplace lockups.
48 * Alan Cox : Some tidyups using the new skb list
50 * Alan Cox : sk->keepopen now seems to work
51 * Alan Cox : Pulls options out correctly on accepts
52 * Alan Cox : Fixed assorted sk->rqueue->next errors
53 * Alan Cox : PSH doesn't end a TCP read. Switched a
55 * Alan Cox : Tidied tcp_data to avoid a potential
57 * Alan Cox : Added some better commenting, as the
58 * tcp is hard to follow
59 * Alan Cox : Removed incorrect check for 20 * psh
60 * Michael O'Reilly : ack < copied bug fix.
61 * Johannes Stille : Misc tcp fixes (not all in yet).
62 * Alan Cox : FIN with no memory -> CRASH
63 * Alan Cox : Added socket option proto entries.
64 * Also added awareness of them to accept.
65 * Alan Cox : Added TCP options (SOL_TCP)
66 * Alan Cox : Switched wakeup calls to callbacks,
67 * so the kernel can layer network
69 * Alan Cox : Use ip_tos/ip_ttl settings.
70 * Alan Cox : Handle FIN (more) properly (we hope).
71 * Alan Cox : RST frames sent on unsynchronised
73 * Alan Cox : Put in missing check for SYN bit.
74 * Alan Cox : Added tcp_select_window() aka NET2E
75 * window non shrink trick.
76 * Alan Cox : Added a couple of small NET2E timer
78 * Charles Hedrick : TCP fixes
79 * Toomas Tamm : TCP window fixes
80 * Alan Cox : Small URG fix to rlogin ^C ack fight
81 * Charles Hedrick : Rewrote most of it to actually work
82 * Linus : Rewrote tcp_read() and URG handling
84 * Gerhard Koerting: Fixed some missing timer handling
85 * Matthew Dillon : Reworked TCP machine states as per RFC
86 * Gerhard Koerting: PC/TCP workarounds
87 * Adam Caldwell : Assorted timer/timing errors
88 * Matthew Dillon : Fixed another RST bug
89 * Alan Cox : Move to kernel side addressing changes.
90 * Alan Cox : Beginning work on TCP fastpathing
92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
93 * Alan Cox : TCP fast path debugging
94 * Alan Cox : Window clamping
95 * Michael Riepe : Bug in tcp_check()
96 * Matt Dillon : More TCP improvements and RST bug fixes
97 * Matt Dillon : Yet more small nasties remove from the
98 * TCP code (Be very nice to this man if
99 * tcp finally works 100%) 8)
100 * Alan Cox : BSD accept semantics.
101 * Alan Cox : Reset on closedown bug.
102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
103 * Michael Pall : Handle poll() after URG properly in
105 * Michael Pall : Undo the last fix in tcp_read_urg()
106 * (multi URG PUSH broke rlogin).
107 * Michael Pall : Fix the multi URG PUSH problem in
108 * tcp_readable(), poll() after URG
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
112 * Alan Cox : Changed the semantics of sk->socket to
113 * fix a race and a signal problem with
114 * accept() and async I/O.
115 * Alan Cox : Relaxed the rules on tcp_sendto().
116 * Yury Shevchuk : Really fixed accept() blocking problem.
117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
118 * clients/servers which listen in on
120 * Alan Cox : Cleaned the above up and shrank it to
121 * a sensible code size.
122 * Alan Cox : Self connect lockup fix.
123 * Alan Cox : No connect to multicast.
124 * Ross Biro : Close unaccepted children on master
126 * Alan Cox : Reset tracing code.
127 * Alan Cox : Spurious resets on shutdown.
128 * Alan Cox : Giant 15 minute/60 second timer error
129 * Alan Cox : Small whoops in polling before an
131 * Alan Cox : Kept the state trace facility since
132 * it's handy for debugging.
133 * Alan Cox : More reset handler fixes.
134 * Alan Cox : Started rewriting the code based on
135 * the RFC's for other useful protocol
136 * references see: Comer, KA9Q NOS, and
137 * for a reference on the difference
138 * between specifications and how BSD
139 * works see the 4.4lite source.
140 * A.N.Kuznetsov : Don't time wait on completion of tidy
142 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
143 * Linus Torvalds : Fixed BSD port reuse to work first syn
144 * Alan Cox : Reimplemented timers as per the RFC
145 * and using multiple timers for sanity.
146 * Alan Cox : Small bug fixes, and a lot of new
148 * Alan Cox : Fixed dual reader crash by locking
149 * the buffers (much like datagram.c)
150 * Alan Cox : Fixed stuck sockets in probe. A probe
151 * now gets fed up of retrying without
152 * (even a no space) answer.
153 * Alan Cox : Extracted closing code better
154 * Alan Cox : Fixed the closing state machine to
156 * Alan Cox : More 'per spec' fixes.
157 * Jorge Cwik : Even faster checksumming.
158 * Alan Cox : tcp_data() doesn't ack illegal PSH
159 * only frames. At least one pc tcp stack
161 * Alan Cox : Cache last socket.
162 * Alan Cox : Per route irtt.
163 * Matt Day : poll()->select() match BSD precisely on error
164 * Alan Cox : New buffers
165 * Marc Tamsky : Various sk->prot->retransmits and
166 * sk->retransmits misupdating fixed.
167 * Fixed tcp_write_timeout: stuck close,
168 * and TCP syn retries gets used now.
169 * Mark Yarvis : In tcp_read_wakeup(), don't send an
170 * ack if state is TCP_CLOSED.
171 * Alan Cox : Look up device on a retransmit - routes may
172 * change. Doesn't yet cope with MSS shrink right
174 * Marc Tamsky : Closing in closing fixes.
175 * Mike Shaver : RFC1122 verifications.
176 * Alan Cox : rcv_saddr errors.
177 * Alan Cox : Block double connect().
178 * Alan Cox : Small hooks for enSKIP.
179 * Alexey Kuznetsov: Path MTU discovery.
180 * Alan Cox : Support soft errors.
181 * Alan Cox : Fix MTU discovery pathological case
182 * when the remote claims no mtu!
183 * Marc Tamsky : TCP_CLOSE fix.
184 * Colin (G3TNE) : Send a reset on syn ack replies in
185 * window but wrong (fixes NT lpd problems)
186 * Pedro Roque : Better TCP window handling, delayed ack.
187 * Joerg Reuter : No modification of locked buffers in
188 * tcp_do_retransmit()
189 * Eric Schenk : Changed receiver side silly window
190 * avoidance algorithm to BSD style
191 * algorithm. This doubles throughput
192 * against machines running Solaris,
193 * and seems to result in general
195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
196 * Willy Konynenberg : Transparent proxying support.
197 * Mike McLagan : Routing by source
198 * Keith Owens : Do proper merging with partial SKB's in
199 * tcp_do_sendmsg to avoid burstiness.
200 * Eric Schenk : Fix fast close down bug with
201 * shutdown() followed by close().
202 * Andi Kleen : Make poll agree with SIGIO
203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
204 * lingertime == 0 (RFC 793 ABORT Call)
205 * Hirokazu Takahashi : Use copy_from_user() instead of
206 * csum_and_copy_from_user() if possible.
208 * This program is free software; you can redistribute it and/or
209 * modify it under the terms of the GNU General Public License
210 * as published by the Free Software Foundation; either version
211 * 2 of the License, or(at your option) any later version.
213 * Description of States:
215 * TCP_SYN_SENT sent a connection request, waiting for ack
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
220 * TCP_ESTABLISHED connection established
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
231 * TCP_TIME_WAIT timeout to catch resent junk before entering
232 * closed, can only be entered from FIN_WAIT2
233 * or CLOSING. Required because the other end
234 * may not have gotten our last ACK causing it
235 * to retransmit the data packet (which we ignore)
237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
238 * us to finish writing our data and to shutdown
239 * (we have to close() to move on to LAST_ACK)
241 * TCP_LAST_ACK out side has shutdown after remote has
242 * shutdown. There may still be data in our
243 * buffer that we have to finish sending
245 * TCP_CLOSE socket is finished
248 #define pr_fmt(fmt) "TCP: " fmt
250 #include <linux/kernel.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/init.h>
256 #include <linux/fs.h>
257 #include <linux/skbuff.h>
258 #include <linux/scatterlist.h>
259 #include <linux/splice.h>
260 #include <linux/net.h>
261 #include <linux/socket.h>
262 #include <linux/random.h>
263 #include <linux/bootmem.h>
264 #include <linux/highmem.h>
265 #include <linux/swap.h>
266 #include <linux/cache.h>
267 #include <linux/err.h>
268 #include <linux/crypto.h>
269 #include <linux/time.h>
270 #include <linux/slab.h>
272 #include <net/icmp.h>
273 #include <net/inet_common.h>
275 #include <net/xfrm.h>
277 #include <net/netdma.h>
278 #include <net/sock.h>
280 #include <asm/uaccess.h>
281 #include <asm/ioctls.h>
283 int sysctl_tcp_fin_timeout __read_mostly
= TCP_FIN_TIMEOUT
;
285 struct percpu_counter tcp_orphan_count
;
286 EXPORT_SYMBOL_GPL(tcp_orphan_count
);
288 int sysctl_tcp_wmem
[3] __read_mostly
;
289 int sysctl_tcp_rmem
[3] __read_mostly
;
291 EXPORT_SYMBOL(sysctl_tcp_rmem
);
292 EXPORT_SYMBOL(sysctl_tcp_wmem
);
294 atomic_long_t tcp_memory_allocated
; /* Current allocated memory. */
295 EXPORT_SYMBOL(tcp_memory_allocated
);
298 * Current number of TCP sockets.
300 struct percpu_counter tcp_sockets_allocated
;
301 EXPORT_SYMBOL(tcp_sockets_allocated
);
306 struct tcp_splice_state
{
307 struct pipe_inode_info
*pipe
;
313 * Pressure flag: try to collapse.
314 * Technical note: it is used by multiple contexts non atomically.
315 * All the __sk_mem_schedule() is of this nature: accounting
316 * is strict, actions are advisory and have some latency.
318 int tcp_memory_pressure __read_mostly
;
319 EXPORT_SYMBOL(tcp_memory_pressure
);
321 void tcp_enter_memory_pressure(struct sock
*sk
)
323 if (!tcp_memory_pressure
) {
324 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMEMORYPRESSURES
);
325 tcp_memory_pressure
= 1;
328 EXPORT_SYMBOL(tcp_enter_memory_pressure
);
330 /* Convert seconds to retransmits based on initial and max timeout */
331 static u8
secs_to_retrans(int seconds
, int timeout
, int rto_max
)
336 int period
= timeout
;
339 while (seconds
> period
&& res
< 255) {
342 if (timeout
> rto_max
)
350 /* Convert retransmits to seconds based on initial and max timeout */
351 static int retrans_to_secs(u8 retrans
, int timeout
, int rto_max
)
359 if (timeout
> rto_max
)
367 /* Address-family independent initialization for a tcp_sock.
369 * NOTE: A lot of things set to zero explicitly by call to
370 * sk_alloc() so need not be done here.
372 void tcp_init_sock(struct sock
*sk
)
374 struct inet_connection_sock
*icsk
= inet_csk(sk
);
375 struct tcp_sock
*tp
= tcp_sk(sk
);
377 skb_queue_head_init(&tp
->out_of_order_queue
);
378 tcp_init_xmit_timers(sk
);
379 tcp_prequeue_init(tp
);
380 INIT_LIST_HEAD(&tp
->tsq_node
);
382 icsk
->icsk_rto
= TCP_TIMEOUT_INIT
;
383 tp
->mdev
= TCP_TIMEOUT_INIT
;
385 /* So many TCP implementations out there (incorrectly) count the
386 * initial SYN frame in their delayed-ACK and congestion control
387 * algorithms that we must have the following bandaid to talk
388 * efficiently to them. -DaveM
390 tp
->snd_cwnd
= TCP_INIT_CWND
;
392 /* See draft-stevens-tcpca-spec-01 for discussion of the
393 * initialization of these values.
395 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
396 tp
->snd_cwnd_clamp
= ~0;
397 tp
->mss_cache
= TCP_MSS_DEFAULT
;
399 tp
->reordering
= sysctl_tcp_reordering
;
400 tcp_enable_early_retrans(tp
);
401 icsk
->icsk_ca_ops
= &tcp_init_congestion_ops
;
405 sk
->sk_state
= TCP_CLOSE
;
407 sk
->sk_write_space
= sk_stream_write_space
;
408 sock_set_flag(sk
, SOCK_USE_WRITE_QUEUE
);
410 icsk
->icsk_sync_mss
= tcp_sync_mss
;
412 /* Presumed zeroed, in order of appearance:
413 * cookie_in_always, cookie_out_never,
414 * s_data_constant, s_data_in, s_data_out
416 sk
->sk_sndbuf
= sysctl_tcp_wmem
[1];
417 sk
->sk_rcvbuf
= sysctl_tcp_rmem
[1];
420 sock_update_memcg(sk
);
421 sk_sockets_allocated_inc(sk
);
424 EXPORT_SYMBOL(tcp_init_sock
);
427 * Wait for a TCP event.
429 * Note that we don't need to lock the socket, as the upper poll layers
430 * take care of normal races (between the test and the event) and we don't
431 * go look at any of the socket buffers directly.
433 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
436 struct sock
*sk
= sock
->sk
;
437 const struct tcp_sock
*tp
= tcp_sk(sk
);
439 sock_poll_wait(file
, sk_sleep(sk
), wait
);
440 if (sk
->sk_state
== TCP_LISTEN
)
441 return inet_csk_listen_poll(sk
);
443 /* Socket is not locked. We are protected from async events
444 * by poll logic and correct handling of state changes
445 * made by other threads is impossible in any case.
451 * POLLHUP is certainly not done right. But poll() doesn't
452 * have a notion of HUP in just one direction, and for a
453 * socket the read side is more interesting.
455 * Some poll() documentation says that POLLHUP is incompatible
456 * with the POLLOUT/POLLWR flags, so somebody should check this
457 * all. But careful, it tends to be safer to return too many
458 * bits than too few, and you can easily break real applications
459 * if you don't tell them that something has hung up!
463 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
464 * our fs/select.c). It means that after we received EOF,
465 * poll always returns immediately, making impossible poll() on write()
466 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
467 * if and only if shutdown has been made in both directions.
468 * Actually, it is interesting to look how Solaris and DUX
469 * solve this dilemma. I would prefer, if POLLHUP were maskable,
470 * then we could set it on SND_SHUTDOWN. BTW examples given
471 * in Stevens' books assume exactly this behaviour, it explains
472 * why POLLHUP is incompatible with POLLOUT. --ANK
474 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
475 * blocking on fresh not-connected or disconnected socket. --ANK
477 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| sk
->sk_state
== TCP_CLOSE
)
479 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
480 mask
|= POLLIN
| POLLRDNORM
| POLLRDHUP
;
482 /* Connected or passive Fast Open socket? */
483 if (sk
->sk_state
!= TCP_SYN_SENT
&&
484 (sk
->sk_state
!= TCP_SYN_RECV
|| tp
->fastopen_rsk
!= NULL
)) {
485 int target
= sock_rcvlowat(sk
, 0, INT_MAX
);
487 if (tp
->urg_seq
== tp
->copied_seq
&&
488 !sock_flag(sk
, SOCK_URGINLINE
) &&
492 /* Potential race condition. If read of tp below will
493 * escape above sk->sk_state, we can be illegally awaken
494 * in SYN_* states. */
495 if (tp
->rcv_nxt
- tp
->copied_seq
>= target
)
496 mask
|= POLLIN
| POLLRDNORM
;
498 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
499 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
)) {
500 mask
|= POLLOUT
| POLLWRNORM
;
501 } else { /* send SIGIO later */
502 set_bit(SOCK_ASYNC_NOSPACE
,
503 &sk
->sk_socket
->flags
);
504 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
506 /* Race breaker. If space is freed after
507 * wspace test but before the flags are set,
508 * IO signal will be lost.
510 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
))
511 mask
|= POLLOUT
| POLLWRNORM
;
514 mask
|= POLLOUT
| POLLWRNORM
;
516 if (tp
->urg_data
& TCP_URG_VALID
)
519 /* This barrier is coupled with smp_wmb() in tcp_reset() */
526 EXPORT_SYMBOL(tcp_poll
);
528 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
530 struct tcp_sock
*tp
= tcp_sk(sk
);
536 if (sk
->sk_state
== TCP_LISTEN
)
539 slow
= lock_sock_fast(sk
);
540 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
542 else if (sock_flag(sk
, SOCK_URGINLINE
) ||
544 before(tp
->urg_seq
, tp
->copied_seq
) ||
545 !before(tp
->urg_seq
, tp
->rcv_nxt
)) {
547 answ
= tp
->rcv_nxt
- tp
->copied_seq
;
549 /* Subtract 1, if FIN was received */
550 if (answ
&& sock_flag(sk
, SOCK_DONE
))
553 answ
= tp
->urg_seq
- tp
->copied_seq
;
554 unlock_sock_fast(sk
, slow
);
557 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
560 if (sk
->sk_state
== TCP_LISTEN
)
563 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
566 answ
= tp
->write_seq
- tp
->snd_una
;
569 if (sk
->sk_state
== TCP_LISTEN
)
572 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
575 answ
= tp
->write_seq
- tp
->snd_nxt
;
581 return put_user(answ
, (int __user
*)arg
);
583 EXPORT_SYMBOL(tcp_ioctl
);
585 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
587 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
588 tp
->pushed_seq
= tp
->write_seq
;
591 static inline bool forced_push(const struct tcp_sock
*tp
)
593 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
596 static inline void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
598 struct tcp_sock
*tp
= tcp_sk(sk
);
599 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
602 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
603 tcb
->tcp_flags
= TCPHDR_ACK
;
605 skb_header_release(skb
);
606 tcp_add_write_queue_tail(sk
, skb
);
607 sk
->sk_wmem_queued
+= skb
->truesize
;
608 sk_mem_charge(sk
, skb
->truesize
);
609 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
610 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
613 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
)
616 tp
->snd_up
= tp
->write_seq
;
619 static inline void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
622 if (tcp_send_head(sk
)) {
623 struct tcp_sock
*tp
= tcp_sk(sk
);
625 if (!(flags
& MSG_MORE
) || forced_push(tp
))
626 tcp_mark_push(tp
, tcp_write_queue_tail(sk
));
628 tcp_mark_urg(tp
, flags
);
629 __tcp_push_pending_frames(sk
, mss_now
,
630 (flags
& MSG_MORE
) ? TCP_NAGLE_CORK
: nonagle
);
634 static int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
635 unsigned int offset
, size_t len
)
637 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
640 ret
= skb_splice_bits(skb
, offset
, tss
->pipe
, min(rd_desc
->count
, len
),
643 rd_desc
->count
-= ret
;
647 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
649 /* Store TCP splice context information in read_descriptor_t. */
650 read_descriptor_t rd_desc
= {
655 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
659 * tcp_splice_read - splice data from TCP socket to a pipe
660 * @sock: socket to splice from
661 * @ppos: position (not valid)
662 * @pipe: pipe to splice to
663 * @len: number of bytes to splice
664 * @flags: splice modifier flags
667 * Will read pages from given socket and fill them into a pipe.
670 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
671 struct pipe_inode_info
*pipe
, size_t len
,
674 struct sock
*sk
= sock
->sk
;
675 struct tcp_splice_state tss
= {
684 sock_rps_record_flow(sk
);
686 * We can't seek on a socket input
695 timeo
= sock_rcvtimeo(sk
, sock
->file
->f_flags
& O_NONBLOCK
);
697 ret
= __tcp_splice_read(sk
, &tss
);
703 if (sock_flag(sk
, SOCK_DONE
))
706 ret
= sock_error(sk
);
709 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
711 if (sk
->sk_state
== TCP_CLOSE
) {
713 * This occurs when user tries to read
714 * from never connected socket.
716 if (!sock_flag(sk
, SOCK_DONE
))
724 sk_wait_data(sk
, &timeo
);
725 if (signal_pending(current
)) {
726 ret
= sock_intr_errno(timeo
);
739 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
740 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
741 signal_pending(current
))
752 EXPORT_SYMBOL(tcp_splice_read
);
754 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
)
758 /* The TCP header must be at least 32-bit aligned. */
759 size
= ALIGN(size
, 4);
761 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
763 if (sk_wmem_schedule(sk
, skb
->truesize
)) {
764 skb_reserve(skb
, sk
->sk_prot
->max_header
);
766 * Make sure that we have exactly size bytes
767 * available to the caller, no more, no less.
769 skb
->reserved_tailroom
= skb
->end
- skb
->tail
- size
;
774 sk
->sk_prot
->enter_memory_pressure(sk
);
775 sk_stream_moderate_sndbuf(sk
);
780 static unsigned int tcp_xmit_size_goal(struct sock
*sk
, u32 mss_now
,
783 struct tcp_sock
*tp
= tcp_sk(sk
);
784 u32 xmit_size_goal
, old_size_goal
;
786 xmit_size_goal
= mss_now
;
788 if (large_allowed
&& sk_can_gso(sk
)) {
789 xmit_size_goal
= ((sk
->sk_gso_max_size
- 1) -
790 inet_csk(sk
)->icsk_af_ops
->net_header_len
-
791 inet_csk(sk
)->icsk_ext_hdr_len
-
794 /* TSQ : try to have two TSO segments in flight */
795 xmit_size_goal
= min_t(u32
, xmit_size_goal
,
796 sysctl_tcp_limit_output_bytes
>> 1);
798 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
800 /* We try hard to avoid divides here */
801 old_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
803 if (likely(old_size_goal
<= xmit_size_goal
&&
804 old_size_goal
+ mss_now
> xmit_size_goal
)) {
805 xmit_size_goal
= old_size_goal
;
807 tp
->xmit_size_goal_segs
=
808 min_t(u16
, xmit_size_goal
/ mss_now
,
809 sk
->sk_gso_max_segs
);
810 xmit_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
814 return max(xmit_size_goal
, mss_now
);
817 static int tcp_send_mss(struct sock
*sk
, int *size_goal
, int flags
)
821 mss_now
= tcp_current_mss(sk
);
822 *size_goal
= tcp_xmit_size_goal(sk
, mss_now
, !(flags
& MSG_OOB
));
827 static ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
*page
, int offset
,
828 size_t size
, int flags
)
830 struct tcp_sock
*tp
= tcp_sk(sk
);
831 int mss_now
, size_goal
;
834 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
836 /* Wait for a connection to finish. One exception is TCP Fast Open
837 * (passive side) where data is allowed to be sent before a connection
838 * is fully established.
840 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
841 !tcp_passive_fastopen(sk
)) {
842 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
846 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
848 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
852 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
856 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
860 if (!tcp_send_head(sk
) || (copy
= size_goal
- skb
->len
) <= 0) {
862 if (!sk_stream_memory_free(sk
))
863 goto wait_for_sndbuf
;
865 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
867 goto wait_for_memory
;
876 i
= skb_shinfo(skb
)->nr_frags
;
877 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
878 if (!can_coalesce
&& i
>= MAX_SKB_FRAGS
) {
879 tcp_mark_push(tp
, skb
);
882 if (!sk_wmem_schedule(sk
, copy
))
883 goto wait_for_memory
;
886 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
889 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
891 skb_shinfo(skb
)->tx_flags
|= SKBTX_SHARED_FRAG
;
894 skb
->data_len
+= copy
;
895 skb
->truesize
+= copy
;
896 sk
->sk_wmem_queued
+= copy
;
897 sk_mem_charge(sk
, copy
);
898 skb
->ip_summed
= CHECKSUM_PARTIAL
;
899 tp
->write_seq
+= copy
;
900 TCP_SKB_CB(skb
)->end_seq
+= copy
;
901 skb_shinfo(skb
)->gso_segs
= 0;
904 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
911 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
914 if (forced_push(tp
)) {
915 tcp_mark_push(tp
, skb
);
916 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
917 } else if (skb
== tcp_send_head(sk
))
918 tcp_push_one(sk
, mss_now
);
922 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
924 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
926 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
929 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
933 if (copied
&& !(flags
& MSG_SENDPAGE_NOTLAST
))
934 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
941 return sk_stream_error(sk
, flags
, err
);
944 int tcp_sendpage(struct sock
*sk
, struct page
*page
, int offset
,
945 size_t size
, int flags
)
949 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
950 !(sk
->sk_route_caps
& NETIF_F_ALL_CSUM
))
951 return sock_no_sendpage(sk
->sk_socket
, page
, offset
, size
,
955 res
= do_tcp_sendpages(sk
, page
, offset
, size
, flags
);
959 EXPORT_SYMBOL(tcp_sendpage
);
961 static inline int select_size(const struct sock
*sk
, bool sg
)
963 const struct tcp_sock
*tp
= tcp_sk(sk
);
964 int tmp
= tp
->mss_cache
;
967 if (sk_can_gso(sk
)) {
968 /* Small frames wont use a full page:
969 * Payload will immediately follow tcp header.
971 tmp
= SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER
);
973 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
975 if (tmp
>= pgbreak
&&
976 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
984 void tcp_free_fastopen_req(struct tcp_sock
*tp
)
986 if (tp
->fastopen_req
!= NULL
) {
987 kfree(tp
->fastopen_req
);
988 tp
->fastopen_req
= NULL
;
992 static int tcp_sendmsg_fastopen(struct sock
*sk
, struct msghdr
*msg
, int *size
)
994 struct tcp_sock
*tp
= tcp_sk(sk
);
997 if (!(sysctl_tcp_fastopen
& TFO_CLIENT_ENABLE
))
999 if (tp
->fastopen_req
!= NULL
)
1000 return -EALREADY
; /* Another Fast Open is in progress */
1002 tp
->fastopen_req
= kzalloc(sizeof(struct tcp_fastopen_request
),
1004 if (unlikely(tp
->fastopen_req
== NULL
))
1006 tp
->fastopen_req
->data
= msg
;
1008 flags
= (msg
->msg_flags
& MSG_DONTWAIT
) ? O_NONBLOCK
: 0;
1009 err
= __inet_stream_connect(sk
->sk_socket
, msg
->msg_name
,
1010 msg
->msg_namelen
, flags
);
1011 *size
= tp
->fastopen_req
->copied
;
1012 tcp_free_fastopen_req(tp
);
1016 int tcp_sendmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1020 struct tcp_sock
*tp
= tcp_sk(sk
);
1021 struct sk_buff
*skb
;
1022 int iovlen
, flags
, err
, copied
= 0;
1023 int mss_now
= 0, size_goal
, copied_syn
= 0, offset
= 0;
1029 flags
= msg
->msg_flags
;
1030 if (flags
& MSG_FASTOPEN
) {
1031 err
= tcp_sendmsg_fastopen(sk
, msg
, &copied_syn
);
1032 if (err
== -EINPROGRESS
&& copied_syn
> 0)
1036 offset
= copied_syn
;
1039 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
1041 /* Wait for a connection to finish. One exception is TCP Fast Open
1042 * (passive side) where data is allowed to be sent before a connection
1043 * is fully established.
1045 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
1046 !tcp_passive_fastopen(sk
)) {
1047 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
1051 if (unlikely(tp
->repair
)) {
1052 if (tp
->repair_queue
== TCP_RECV_QUEUE
) {
1053 copied
= tcp_send_rcvq(sk
, msg
, size
);
1058 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1061 /* 'common' sending to sendq */
1064 /* This should be in poll */
1065 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
1067 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1069 /* Ok commence sending. */
1070 iovlen
= msg
->msg_iovlen
;
1075 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
1078 sg
= !!(sk
->sk_route_caps
& NETIF_F_SG
);
1080 while (--iovlen
>= 0) {
1081 size_t seglen
= iov
->iov_len
;
1082 unsigned char __user
*from
= iov
->iov_base
;
1085 if (unlikely(offset
> 0)) { /* Skip bytes copied in SYN */
1086 if (offset
>= seglen
) {
1095 while (seglen
> 0) {
1097 int max
= size_goal
;
1099 skb
= tcp_write_queue_tail(sk
);
1100 if (tcp_send_head(sk
)) {
1101 if (skb
->ip_summed
== CHECKSUM_NONE
)
1103 copy
= max
- skb
->len
;
1108 /* Allocate new segment. If the interface is SG,
1109 * allocate skb fitting to single page.
1111 if (!sk_stream_memory_free(sk
))
1112 goto wait_for_sndbuf
;
1114 skb
= sk_stream_alloc_skb(sk
,
1115 select_size(sk
, sg
),
1118 goto wait_for_memory
;
1121 * Check whether we can use HW checksum.
1123 if (sk
->sk_route_caps
& NETIF_F_ALL_CSUM
)
1124 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1126 skb_entail(sk
, skb
);
1131 /* Try to append data to the end of skb. */
1135 /* Where to copy to? */
1136 if (skb_availroom(skb
) > 0) {
1137 /* We have some space in skb head. Superb! */
1138 copy
= min_t(int, copy
, skb_availroom(skb
));
1139 err
= skb_add_data_nocache(sk
, skb
, from
, copy
);
1144 int i
= skb_shinfo(skb
)->nr_frags
;
1145 struct page_frag
*pfrag
= sk_page_frag(sk
);
1147 if (!sk_page_frag_refill(sk
, pfrag
))
1148 goto wait_for_memory
;
1150 if (!skb_can_coalesce(skb
, i
, pfrag
->page
,
1152 if (i
== MAX_SKB_FRAGS
|| !sg
) {
1153 tcp_mark_push(tp
, skb
);
1159 copy
= min_t(int, copy
, pfrag
->size
- pfrag
->offset
);
1161 if (!sk_wmem_schedule(sk
, copy
))
1162 goto wait_for_memory
;
1164 err
= skb_copy_to_page_nocache(sk
, from
, skb
,
1171 /* Update the skb. */
1173 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
1175 skb_fill_page_desc(skb
, i
, pfrag
->page
,
1176 pfrag
->offset
, copy
);
1177 get_page(pfrag
->page
);
1179 pfrag
->offset
+= copy
;
1183 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
1185 tp
->write_seq
+= copy
;
1186 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1187 skb_shinfo(skb
)->gso_segs
= 0;
1191 if ((seglen
-= copy
) == 0 && iovlen
== 0)
1194 if (skb
->len
< max
|| (flags
& MSG_OOB
) || unlikely(tp
->repair
))
1197 if (forced_push(tp
)) {
1198 tcp_mark_push(tp
, skb
);
1199 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1200 } else if (skb
== tcp_send_head(sk
))
1201 tcp_push_one(sk
, mss_now
);
1205 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1208 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
1210 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
1213 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1219 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
1221 return copied
+ copied_syn
;
1225 tcp_unlink_write_queue(skb
, sk
);
1226 /* It is the one place in all of TCP, except connection
1227 * reset, where we can be unlinking the send_head.
1229 tcp_check_send_head(sk
, skb
);
1230 sk_wmem_free_skb(sk
, skb
);
1234 if (copied
+ copied_syn
)
1237 err
= sk_stream_error(sk
, flags
, err
);
1241 EXPORT_SYMBOL(tcp_sendmsg
);
1244 * Handle reading urgent data. BSD has very simple semantics for
1245 * this, no blocking and very strange errors 8)
1248 static int tcp_recv_urg(struct sock
*sk
, struct msghdr
*msg
, int len
, int flags
)
1250 struct tcp_sock
*tp
= tcp_sk(sk
);
1252 /* No URG data to read. */
1253 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1254 tp
->urg_data
== TCP_URG_READ
)
1255 return -EINVAL
; /* Yes this is right ! */
1257 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1260 if (tp
->urg_data
& TCP_URG_VALID
) {
1262 char c
= tp
->urg_data
;
1264 if (!(flags
& MSG_PEEK
))
1265 tp
->urg_data
= TCP_URG_READ
;
1267 /* Read urgent data. */
1268 msg
->msg_flags
|= MSG_OOB
;
1271 if (!(flags
& MSG_TRUNC
))
1272 err
= memcpy_toiovec(msg
->msg_iov
, &c
, 1);
1275 msg
->msg_flags
|= MSG_TRUNC
;
1277 return err
? -EFAULT
: len
;
1280 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1283 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1284 * the available implementations agree in this case:
1285 * this call should never block, independent of the
1286 * blocking state of the socket.
1287 * Mike <pall@rz.uni-karlsruhe.de>
1292 static int tcp_peek_sndq(struct sock
*sk
, struct msghdr
*msg
, int len
)
1294 struct sk_buff
*skb
;
1295 int copied
= 0, err
= 0;
1297 /* XXX -- need to support SO_PEEK_OFF */
1299 skb_queue_walk(&sk
->sk_write_queue
, skb
) {
1300 err
= skb_copy_datagram_iovec(skb
, 0, msg
->msg_iov
, skb
->len
);
1307 return err
?: copied
;
1310 /* Clean up the receive buffer for full frames taken by the user,
1311 * then send an ACK if necessary. COPIED is the number of bytes
1312 * tcp_recvmsg has given to the user so far, it speeds up the
1313 * calculation of whether or not we must ACK for the sake of
1316 void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1318 struct tcp_sock
*tp
= tcp_sk(sk
);
1319 bool time_to_ack
= false;
1321 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1323 WARN(skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
),
1324 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1325 tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
);
1327 if (inet_csk_ack_scheduled(sk
)) {
1328 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1329 /* Delayed ACKs frequently hit locked sockets during bulk
1331 if (icsk
->icsk_ack
.blocked
||
1332 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1333 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1335 * If this read emptied read buffer, we send ACK, if
1336 * connection is not bidirectional, user drained
1337 * receive buffer and there was a small segment
1341 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1342 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1343 !icsk
->icsk_ack
.pingpong
)) &&
1344 !atomic_read(&sk
->sk_rmem_alloc
)))
1348 /* We send an ACK if we can now advertise a non-zero window
1349 * which has been raised "significantly".
1351 * Even if window raised up to infinity, do not send window open ACK
1352 * in states, where we will not receive more. It is useless.
1354 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1355 __u32 rcv_window_now
= tcp_receive_window(tp
);
1357 /* Optimize, __tcp_select_window() is not cheap. */
1358 if (2*rcv_window_now
<= tp
->window_clamp
) {
1359 __u32 new_window
= __tcp_select_window(sk
);
1361 /* Send ACK now, if this read freed lots of space
1362 * in our buffer. Certainly, new_window is new window.
1363 * We can advertise it now, if it is not less than current one.
1364 * "Lots" means "at least twice" here.
1366 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1374 static void tcp_prequeue_process(struct sock
*sk
)
1376 struct sk_buff
*skb
;
1377 struct tcp_sock
*tp
= tcp_sk(sk
);
1379 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPPREQUEUED
);
1381 /* RX process wants to run with disabled BHs, though it is not
1384 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
1385 sk_backlog_rcv(sk
, skb
);
1388 /* Clear memory counter. */
1389 tp
->ucopy
.memory
= 0;
1392 #ifdef CONFIG_NET_DMA
1393 static void tcp_service_net_dma(struct sock
*sk
, bool wait
)
1395 dma_cookie_t done
, used
;
1396 dma_cookie_t last_issued
;
1397 struct tcp_sock
*tp
= tcp_sk(sk
);
1399 if (!tp
->ucopy
.dma_chan
)
1402 last_issued
= tp
->ucopy
.dma_cookie
;
1403 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1406 if (dma_async_is_tx_complete(tp
->ucopy
.dma_chan
,
1408 &used
) == DMA_SUCCESS
) {
1409 /* Safe to free early-copied skbs now */
1410 __skb_queue_purge(&sk
->sk_async_wait_queue
);
1413 struct sk_buff
*skb
;
1414 while ((skb
= skb_peek(&sk
->sk_async_wait_queue
)) &&
1415 (dma_async_is_complete(skb
->dma_cookie
, done
,
1416 used
) == DMA_SUCCESS
)) {
1417 __skb_dequeue(&sk
->sk_async_wait_queue
);
1425 static struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1427 struct sk_buff
*skb
;
1430 while ((skb
= skb_peek(&sk
->sk_receive_queue
)) != NULL
) {
1431 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1432 if (tcp_hdr(skb
)->syn
)
1434 if (offset
< skb
->len
|| tcp_hdr(skb
)->fin
) {
1438 /* This looks weird, but this can happen if TCP collapsing
1439 * splitted a fat GRO packet, while we released socket lock
1440 * in skb_splice_bits()
1442 sk_eat_skb(sk
, skb
, false);
1448 * This routine provides an alternative to tcp_recvmsg() for routines
1449 * that would like to handle copying from skbuffs directly in 'sendfile'
1452 * - It is assumed that the socket was locked by the caller.
1453 * - The routine does not block.
1454 * - At present, there is no support for reading OOB data
1455 * or for 'peeking' the socket using this routine
1456 * (although both would be easy to implement).
1458 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1459 sk_read_actor_t recv_actor
)
1461 struct sk_buff
*skb
;
1462 struct tcp_sock
*tp
= tcp_sk(sk
);
1463 u32 seq
= tp
->copied_seq
;
1467 if (sk
->sk_state
== TCP_LISTEN
)
1469 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1470 if (offset
< skb
->len
) {
1474 len
= skb
->len
- offset
;
1475 /* Stop reading if we hit a patch of urgent data */
1477 u32 urg_offset
= tp
->urg_seq
- seq
;
1478 if (urg_offset
< len
)
1483 used
= recv_actor(desc
, skb
, offset
, len
);
1488 } else if (used
<= len
) {
1493 /* If recv_actor drops the lock (e.g. TCP splice
1494 * receive) the skb pointer might be invalid when
1495 * getting here: tcp_collapse might have deleted it
1496 * while aggregating skbs from the socket queue.
1498 skb
= tcp_recv_skb(sk
, seq
- 1, &offset
);
1501 /* TCP coalescing might have appended data to the skb.
1502 * Try to splice more frags
1504 if (offset
+ 1 != skb
->len
)
1507 if (tcp_hdr(skb
)->fin
) {
1508 sk_eat_skb(sk
, skb
, false);
1512 sk_eat_skb(sk
, skb
, false);
1515 tp
->copied_seq
= seq
;
1517 tp
->copied_seq
= seq
;
1519 tcp_rcv_space_adjust(sk
);
1521 /* Clean up data we have read: This will do ACK frames. */
1523 tcp_recv_skb(sk
, seq
, &offset
);
1524 tcp_cleanup_rbuf(sk
, copied
);
1528 EXPORT_SYMBOL(tcp_read_sock
);
1531 * This routine copies from a sock struct into the user buffer.
1533 * Technical note: in 2.3 we work on _locked_ socket, so that
1534 * tricks with *seq access order and skb->users are not required.
1535 * Probably, code can be easily improved even more.
1538 int tcp_recvmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1539 size_t len
, int nonblock
, int flags
, int *addr_len
)
1541 struct tcp_sock
*tp
= tcp_sk(sk
);
1547 int target
; /* Read at least this many bytes */
1549 struct task_struct
*user_recv
= NULL
;
1550 bool copied_early
= false;
1551 struct sk_buff
*skb
;
1557 if (sk
->sk_state
== TCP_LISTEN
)
1560 timeo
= sock_rcvtimeo(sk
, nonblock
);
1562 /* Urgent data needs to be handled specially. */
1563 if (flags
& MSG_OOB
)
1566 if (unlikely(tp
->repair
)) {
1568 if (!(flags
& MSG_PEEK
))
1571 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
1575 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1578 /* 'common' recv queue MSG_PEEK-ing */
1581 seq
= &tp
->copied_seq
;
1582 if (flags
& MSG_PEEK
) {
1583 peek_seq
= tp
->copied_seq
;
1587 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1589 #ifdef CONFIG_NET_DMA
1590 tp
->ucopy
.dma_chan
= NULL
;
1592 skb
= skb_peek_tail(&sk
->sk_receive_queue
);
1597 available
= TCP_SKB_CB(skb
)->seq
+ skb
->len
- (*seq
);
1598 if ((available
< target
) &&
1599 (len
> sysctl_tcp_dma_copybreak
) && !(flags
& MSG_PEEK
) &&
1600 !sysctl_tcp_low_latency
&&
1601 net_dma_find_channel()) {
1602 preempt_enable_no_resched();
1603 tp
->ucopy
.pinned_list
=
1604 dma_pin_iovec_pages(msg
->msg_iov
, len
);
1606 preempt_enable_no_resched();
1614 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1615 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1618 if (signal_pending(current
)) {
1619 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1624 /* Next get a buffer. */
1626 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1627 /* Now that we have two receive queues this
1630 if (WARN(before(*seq
, TCP_SKB_CB(skb
)->seq
),
1631 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1632 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
,
1636 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1637 if (tcp_hdr(skb
)->syn
)
1639 if (offset
< skb
->len
)
1641 if (tcp_hdr(skb
)->fin
)
1643 WARN(!(flags
& MSG_PEEK
),
1644 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1645 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
, flags
);
1648 /* Well, if we have backlog, try to process it now yet. */
1650 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1655 sk
->sk_state
== TCP_CLOSE
||
1656 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1658 signal_pending(current
))
1661 if (sock_flag(sk
, SOCK_DONE
))
1665 copied
= sock_error(sk
);
1669 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1672 if (sk
->sk_state
== TCP_CLOSE
) {
1673 if (!sock_flag(sk
, SOCK_DONE
)) {
1674 /* This occurs when user tries to read
1675 * from never connected socket.
1688 if (signal_pending(current
)) {
1689 copied
= sock_intr_errno(timeo
);
1694 tcp_cleanup_rbuf(sk
, copied
);
1696 if (!sysctl_tcp_low_latency
&& tp
->ucopy
.task
== user_recv
) {
1697 /* Install new reader */
1698 if (!user_recv
&& !(flags
& (MSG_TRUNC
| MSG_PEEK
))) {
1699 user_recv
= current
;
1700 tp
->ucopy
.task
= user_recv
;
1701 tp
->ucopy
.iov
= msg
->msg_iov
;
1704 tp
->ucopy
.len
= len
;
1706 WARN_ON(tp
->copied_seq
!= tp
->rcv_nxt
&&
1707 !(flags
& (MSG_PEEK
| MSG_TRUNC
)));
1709 /* Ugly... If prequeue is not empty, we have to
1710 * process it before releasing socket, otherwise
1711 * order will be broken at second iteration.
1712 * More elegant solution is required!!!
1714 * Look: we have the following (pseudo)queues:
1716 * 1. packets in flight
1721 * Each queue can be processed only if the next ones
1722 * are empty. At this point we have empty receive_queue.
1723 * But prequeue _can_ be not empty after 2nd iteration,
1724 * when we jumped to start of loop because backlog
1725 * processing added something to receive_queue.
1726 * We cannot release_sock(), because backlog contains
1727 * packets arrived _after_ prequeued ones.
1729 * Shortly, algorithm is clear --- to process all
1730 * the queues in order. We could make it more directly,
1731 * requeueing packets from backlog to prequeue, if
1732 * is not empty. It is more elegant, but eats cycles,
1735 if (!skb_queue_empty(&tp
->ucopy
.prequeue
))
1738 /* __ Set realtime policy in scheduler __ */
1741 #ifdef CONFIG_NET_DMA
1742 if (tp
->ucopy
.dma_chan
) {
1743 if (tp
->rcv_wnd
== 0 &&
1744 !skb_queue_empty(&sk
->sk_async_wait_queue
)) {
1745 tcp_service_net_dma(sk
, true);
1746 tcp_cleanup_rbuf(sk
, copied
);
1748 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1751 if (copied
>= target
) {
1752 /* Do not sleep, just process backlog. */
1756 sk_wait_data(sk
, &timeo
);
1758 #ifdef CONFIG_NET_DMA
1759 tcp_service_net_dma(sk
, false); /* Don't block */
1760 tp
->ucopy
.wakeup
= 0;
1766 /* __ Restore normal policy in scheduler __ */
1768 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1769 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG
, chunk
);
1774 if (tp
->rcv_nxt
== tp
->copied_seq
&&
1775 !skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1777 tcp_prequeue_process(sk
);
1779 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1780 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1786 if ((flags
& MSG_PEEK
) &&
1787 (peek_seq
- copied
- urg_hole
!= tp
->copied_seq
)) {
1788 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1790 task_pid_nr(current
));
1791 peek_seq
= tp
->copied_seq
;
1796 /* Ok so how much can we use? */
1797 used
= skb
->len
- offset
;
1801 /* Do we have urgent data here? */
1803 u32 urg_offset
= tp
->urg_seq
- *seq
;
1804 if (urg_offset
< used
) {
1806 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1819 if (!(flags
& MSG_TRUNC
)) {
1820 #ifdef CONFIG_NET_DMA
1821 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
1822 tp
->ucopy
.dma_chan
= net_dma_find_channel();
1824 if (tp
->ucopy
.dma_chan
) {
1825 tp
->ucopy
.dma_cookie
= dma_skb_copy_datagram_iovec(
1826 tp
->ucopy
.dma_chan
, skb
, offset
,
1828 tp
->ucopy
.pinned_list
);
1830 if (tp
->ucopy
.dma_cookie
< 0) {
1832 pr_alert("%s: dma_cookie < 0\n",
1835 /* Exception. Bailout! */
1841 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1843 if ((offset
+ used
) == skb
->len
)
1844 copied_early
= true;
1849 err
= skb_copy_datagram_iovec(skb
, offset
,
1850 msg
->msg_iov
, used
);
1852 /* Exception. Bailout! */
1864 tcp_rcv_space_adjust(sk
);
1867 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1869 tcp_fast_path_check(sk
);
1871 if (used
+ offset
< skb
->len
)
1874 if (tcp_hdr(skb
)->fin
)
1876 if (!(flags
& MSG_PEEK
)) {
1877 sk_eat_skb(sk
, skb
, copied_early
);
1878 copied_early
= false;
1883 /* Process the FIN. */
1885 if (!(flags
& MSG_PEEK
)) {
1886 sk_eat_skb(sk
, skb
, copied_early
);
1887 copied_early
= false;
1893 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1896 tp
->ucopy
.len
= copied
> 0 ? len
: 0;
1898 tcp_prequeue_process(sk
);
1900 if (copied
> 0 && (chunk
= len
- tp
->ucopy
.len
) != 0) {
1901 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1907 tp
->ucopy
.task
= NULL
;
1911 #ifdef CONFIG_NET_DMA
1912 tcp_service_net_dma(sk
, true); /* Wait for queue to drain */
1913 tp
->ucopy
.dma_chan
= NULL
;
1915 if (tp
->ucopy
.pinned_list
) {
1916 dma_unpin_iovec_pages(tp
->ucopy
.pinned_list
);
1917 tp
->ucopy
.pinned_list
= NULL
;
1921 /* According to UNIX98, msg_name/msg_namelen are ignored
1922 * on connected socket. I was just happy when found this 8) --ANK
1925 /* Clean up data we have read: This will do ACK frames. */
1926 tcp_cleanup_rbuf(sk
, copied
);
1936 err
= tcp_recv_urg(sk
, msg
, len
, flags
);
1940 err
= tcp_peek_sndq(sk
, msg
, len
);
1943 EXPORT_SYMBOL(tcp_recvmsg
);
1945 void tcp_set_state(struct sock
*sk
, int state
)
1947 int oldstate
= sk
->sk_state
;
1950 case TCP_ESTABLISHED
:
1951 if (oldstate
!= TCP_ESTABLISHED
)
1952 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1956 if (oldstate
== TCP_CLOSE_WAIT
|| oldstate
== TCP_ESTABLISHED
)
1957 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ESTABRESETS
);
1959 sk
->sk_prot
->unhash(sk
);
1960 if (inet_csk(sk
)->icsk_bind_hash
&&
1961 !(sk
->sk_userlocks
& SOCK_BINDPORT_LOCK
))
1965 if (oldstate
== TCP_ESTABLISHED
)
1966 TCP_DEC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1969 /* Change state AFTER socket is unhashed to avoid closed
1970 * socket sitting in hash tables.
1972 sk
->sk_state
= state
;
1975 SOCK_DEBUG(sk
, "TCP sk=%p, State %s -> %s\n", sk
, statename
[oldstate
], statename
[state
]);
1978 EXPORT_SYMBOL_GPL(tcp_set_state
);
1981 * State processing on a close. This implements the state shift for
1982 * sending our FIN frame. Note that we only send a FIN for some
1983 * states. A shutdown() may have already sent the FIN, or we may be
1987 static const unsigned char new_state
[16] = {
1988 /* current state: new state: action: */
1989 /* (Invalid) */ TCP_CLOSE
,
1990 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1991 /* TCP_SYN_SENT */ TCP_CLOSE
,
1992 /* TCP_SYN_RECV */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
1993 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1
,
1994 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2
,
1995 /* TCP_TIME_WAIT */ TCP_CLOSE
,
1996 /* TCP_CLOSE */ TCP_CLOSE
,
1997 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK
| TCP_ACTION_FIN
,
1998 /* TCP_LAST_ACK */ TCP_LAST_ACK
,
1999 /* TCP_LISTEN */ TCP_CLOSE
,
2000 /* TCP_CLOSING */ TCP_CLOSING
,
2003 static int tcp_close_state(struct sock
*sk
)
2005 int next
= (int)new_state
[sk
->sk_state
];
2006 int ns
= next
& TCP_STATE_MASK
;
2008 tcp_set_state(sk
, ns
);
2010 return next
& TCP_ACTION_FIN
;
2014 * Shutdown the sending side of a connection. Much like close except
2015 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2018 void tcp_shutdown(struct sock
*sk
, int how
)
2020 /* We need to grab some memory, and put together a FIN,
2021 * and then put it into the queue to be sent.
2022 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2024 if (!(how
& SEND_SHUTDOWN
))
2027 /* If we've already sent a FIN, or it's a closed state, skip this. */
2028 if ((1 << sk
->sk_state
) &
2029 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
2030 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
2031 /* Clear out any half completed packets. FIN if needed. */
2032 if (tcp_close_state(sk
))
2036 EXPORT_SYMBOL(tcp_shutdown
);
2038 bool tcp_check_oom(struct sock
*sk
, int shift
)
2040 bool too_many_orphans
, out_of_socket_memory
;
2042 too_many_orphans
= tcp_too_many_orphans(sk
, shift
);
2043 out_of_socket_memory
= tcp_out_of_memory(sk
);
2045 if (too_many_orphans
)
2046 net_info_ratelimited("too many orphaned sockets\n");
2047 if (out_of_socket_memory
)
2048 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2049 return too_many_orphans
|| out_of_socket_memory
;
2052 void tcp_close(struct sock
*sk
, long timeout
)
2054 struct sk_buff
*skb
;
2055 int data_was_unread
= 0;
2059 sk
->sk_shutdown
= SHUTDOWN_MASK
;
2061 if (sk
->sk_state
== TCP_LISTEN
) {
2062 tcp_set_state(sk
, TCP_CLOSE
);
2065 inet_csk_listen_stop(sk
);
2067 goto adjudge_to_death
;
2070 /* We need to flush the recv. buffs. We do this only on the
2071 * descriptor close, not protocol-sourced closes, because the
2072 * reader process may not have drained the data yet!
2074 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
2075 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
-
2077 data_was_unread
+= len
;
2083 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2084 if (sk
->sk_state
== TCP_CLOSE
)
2085 goto adjudge_to_death
;
2087 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2088 * data was lost. To witness the awful effects of the old behavior of
2089 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2090 * GET in an FTP client, suspend the process, wait for the client to
2091 * advertise a zero window, then kill -9 the FTP client, wheee...
2092 * Note: timeout is always zero in such a case.
2094 if (unlikely(tcp_sk(sk
)->repair
)) {
2095 sk
->sk_prot
->disconnect(sk
, 0);
2096 } else if (data_was_unread
) {
2097 /* Unread data was tossed, zap the connection. */
2098 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONCLOSE
);
2099 tcp_set_state(sk
, TCP_CLOSE
);
2100 tcp_send_active_reset(sk
, sk
->sk_allocation
);
2101 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
2102 /* Check zero linger _after_ checking for unread data. */
2103 sk
->sk_prot
->disconnect(sk
, 0);
2104 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
2105 } else if (tcp_close_state(sk
)) {
2106 /* We FIN if the application ate all the data before
2107 * zapping the connection.
2110 /* RED-PEN. Formally speaking, we have broken TCP state
2111 * machine. State transitions:
2113 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2114 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2115 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2117 * are legal only when FIN has been sent (i.e. in window),
2118 * rather than queued out of window. Purists blame.
2120 * F.e. "RFC state" is ESTABLISHED,
2121 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2123 * The visible declinations are that sometimes
2124 * we enter time-wait state, when it is not required really
2125 * (harmless), do not send active resets, when they are
2126 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2127 * they look as CLOSING or LAST_ACK for Linux)
2128 * Probably, I missed some more holelets.
2130 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2131 * in a single packet! (May consider it later but will
2132 * probably need API support or TCP_CORK SYN-ACK until
2133 * data is written and socket is closed.)
2138 sk_stream_wait_close(sk
, timeout
);
2141 state
= sk
->sk_state
;
2145 /* It is the last release_sock in its life. It will remove backlog. */
2149 /* Now socket is owned by kernel and we acquire BH lock
2150 to finish close. No need to check for user refs.
2154 WARN_ON(sock_owned_by_user(sk
));
2156 percpu_counter_inc(sk
->sk_prot
->orphan_count
);
2158 /* Have we already been destroyed by a softirq or backlog? */
2159 if (state
!= TCP_CLOSE
&& sk
->sk_state
== TCP_CLOSE
)
2162 /* This is a (useful) BSD violating of the RFC. There is a
2163 * problem with TCP as specified in that the other end could
2164 * keep a socket open forever with no application left this end.
2165 * We use a 3 minute timeout (about the same as BSD) then kill
2166 * our end. If they send after that then tough - BUT: long enough
2167 * that we won't make the old 4*rto = almost no time - whoops
2170 * Nope, it was not mistake. It is really desired behaviour
2171 * f.e. on http servers, when such sockets are useless, but
2172 * consume significant resources. Let's do it with special
2173 * linger2 option. --ANK
2176 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
2177 struct tcp_sock
*tp
= tcp_sk(sk
);
2178 if (tp
->linger2
< 0) {
2179 tcp_set_state(sk
, TCP_CLOSE
);
2180 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2181 NET_INC_STATS_BH(sock_net(sk
),
2182 LINUX_MIB_TCPABORTONLINGER
);
2184 const int tmo
= tcp_fin_time(sk
);
2186 if (tmo
> TCP_TIMEWAIT_LEN
) {
2187 inet_csk_reset_keepalive_timer(sk
,
2188 tmo
- TCP_TIMEWAIT_LEN
);
2190 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
2195 if (sk
->sk_state
!= TCP_CLOSE
) {
2197 if (tcp_check_oom(sk
, 0)) {
2198 tcp_set_state(sk
, TCP_CLOSE
);
2199 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2200 NET_INC_STATS_BH(sock_net(sk
),
2201 LINUX_MIB_TCPABORTONMEMORY
);
2205 if (sk
->sk_state
== TCP_CLOSE
) {
2206 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
2207 /* We could get here with a non-NULL req if the socket is
2208 * aborted (e.g., closed with unread data) before 3WHS
2212 reqsk_fastopen_remove(sk
, req
, false);
2213 inet_csk_destroy_sock(sk
);
2215 /* Otherwise, socket is reprieved until protocol close. */
2222 EXPORT_SYMBOL(tcp_close
);
2224 /* These states need RST on ABORT according to RFC793 */
2226 static inline bool tcp_need_reset(int state
)
2228 return (1 << state
) &
2229 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
2230 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
2233 int tcp_disconnect(struct sock
*sk
, int flags
)
2235 struct inet_sock
*inet
= inet_sk(sk
);
2236 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2237 struct tcp_sock
*tp
= tcp_sk(sk
);
2239 int old_state
= sk
->sk_state
;
2241 if (old_state
!= TCP_CLOSE
)
2242 tcp_set_state(sk
, TCP_CLOSE
);
2244 /* ABORT function of RFC793 */
2245 if (old_state
== TCP_LISTEN
) {
2246 inet_csk_listen_stop(sk
);
2247 } else if (unlikely(tp
->repair
)) {
2248 sk
->sk_err
= ECONNABORTED
;
2249 } else if (tcp_need_reset(old_state
) ||
2250 (tp
->snd_nxt
!= tp
->write_seq
&&
2251 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
2252 /* The last check adjusts for discrepancy of Linux wrt. RFC
2255 tcp_send_active_reset(sk
, gfp_any());
2256 sk
->sk_err
= ECONNRESET
;
2257 } else if (old_state
== TCP_SYN_SENT
)
2258 sk
->sk_err
= ECONNRESET
;
2260 tcp_clear_xmit_timers(sk
);
2261 __skb_queue_purge(&sk
->sk_receive_queue
);
2262 tcp_write_queue_purge(sk
);
2263 __skb_queue_purge(&tp
->out_of_order_queue
);
2264 #ifdef CONFIG_NET_DMA
2265 __skb_queue_purge(&sk
->sk_async_wait_queue
);
2268 inet
->inet_dport
= 0;
2270 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
2271 inet_reset_saddr(sk
);
2273 sk
->sk_shutdown
= 0;
2274 sock_reset_flag(sk
, SOCK_DONE
);
2276 if ((tp
->write_seq
+= tp
->max_window
+ 2) == 0)
2278 icsk
->icsk_backoff
= 0;
2280 icsk
->icsk_probes_out
= 0;
2281 tp
->packets_out
= 0;
2282 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
2283 tp
->snd_cwnd_cnt
= 0;
2284 tp
->window_clamp
= 0;
2285 tcp_set_ca_state(sk
, TCP_CA_Open
);
2286 tcp_clear_retrans(tp
);
2287 inet_csk_delack_init(sk
);
2288 tcp_init_send_head(sk
);
2289 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
2292 WARN_ON(inet
->inet_num
&& !icsk
->icsk_bind_hash
);
2294 sk
->sk_error_report(sk
);
2297 EXPORT_SYMBOL(tcp_disconnect
);
2299 void tcp_sock_destruct(struct sock
*sk
)
2301 inet_sock_destruct(sk
);
2303 kfree(inet_csk(sk
)->icsk_accept_queue
.fastopenq
);
2306 static inline bool tcp_can_repair_sock(const struct sock
*sk
)
2308 return ns_capable(sock_net(sk
)->user_ns
, CAP_NET_ADMIN
) &&
2309 ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_ESTABLISHED
));
2312 static int tcp_repair_options_est(struct tcp_sock
*tp
,
2313 struct tcp_repair_opt __user
*optbuf
, unsigned int len
)
2315 struct tcp_repair_opt opt
;
2317 while (len
>= sizeof(opt
)) {
2318 if (copy_from_user(&opt
, optbuf
, sizeof(opt
)))
2324 switch (opt
.opt_code
) {
2326 tp
->rx_opt
.mss_clamp
= opt
.opt_val
;
2330 u16 snd_wscale
= opt
.opt_val
& 0xFFFF;
2331 u16 rcv_wscale
= opt
.opt_val
>> 16;
2333 if (snd_wscale
> 14 || rcv_wscale
> 14)
2336 tp
->rx_opt
.snd_wscale
= snd_wscale
;
2337 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2338 tp
->rx_opt
.wscale_ok
= 1;
2341 case TCPOPT_SACK_PERM
:
2342 if (opt
.opt_val
!= 0)
2345 tp
->rx_opt
.sack_ok
|= TCP_SACK_SEEN
;
2346 if (sysctl_tcp_fack
)
2347 tcp_enable_fack(tp
);
2349 case TCPOPT_TIMESTAMP
:
2350 if (opt
.opt_val
!= 0)
2353 tp
->rx_opt
.tstamp_ok
= 1;
2362 * Socket option code for TCP.
2364 static int do_tcp_setsockopt(struct sock
*sk
, int level
,
2365 int optname
, char __user
*optval
, unsigned int optlen
)
2367 struct tcp_sock
*tp
= tcp_sk(sk
);
2368 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2372 /* These are data/string values, all the others are ints */
2374 case TCP_CONGESTION
: {
2375 char name
[TCP_CA_NAME_MAX
];
2380 val
= strncpy_from_user(name
, optval
,
2381 min_t(long, TCP_CA_NAME_MAX
-1, optlen
));
2387 err
= tcp_set_congestion_control(sk
, name
);
2396 if (optlen
< sizeof(int))
2399 if (get_user(val
, (int __user
*)optval
))
2406 /* Values greater than interface MTU won't take effect. However
2407 * at the point when this call is done we typically don't yet
2408 * know which interface is going to be used */
2409 if (val
< TCP_MIN_MSS
|| val
> MAX_TCP_WINDOW
) {
2413 tp
->rx_opt
.user_mss
= val
;
2418 /* TCP_NODELAY is weaker than TCP_CORK, so that
2419 * this option on corked socket is remembered, but
2420 * it is not activated until cork is cleared.
2422 * However, when TCP_NODELAY is set we make
2423 * an explicit push, which overrides even TCP_CORK
2424 * for currently queued segments.
2426 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
2427 tcp_push_pending_frames(sk
);
2429 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
2433 case TCP_THIN_LINEAR_TIMEOUTS
:
2434 if (val
< 0 || val
> 1)
2440 case TCP_THIN_DUPACK
:
2441 if (val
< 0 || val
> 1)
2444 tp
->thin_dupack
= val
;
2445 if (tp
->thin_dupack
)
2446 tcp_disable_early_retrans(tp
);
2450 if (!tcp_can_repair_sock(sk
))
2452 else if (val
== 1) {
2454 sk
->sk_reuse
= SK_FORCE_REUSE
;
2455 tp
->repair_queue
= TCP_NO_QUEUE
;
2456 } else if (val
== 0) {
2458 sk
->sk_reuse
= SK_NO_REUSE
;
2459 tcp_send_window_probe(sk
);
2465 case TCP_REPAIR_QUEUE
:
2468 else if (val
< TCP_QUEUES_NR
)
2469 tp
->repair_queue
= val
;
2475 if (sk
->sk_state
!= TCP_CLOSE
)
2477 else if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2478 tp
->write_seq
= val
;
2479 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2485 case TCP_REPAIR_OPTIONS
:
2488 else if (sk
->sk_state
== TCP_ESTABLISHED
)
2489 err
= tcp_repair_options_est(tp
,
2490 (struct tcp_repair_opt __user
*)optval
,
2497 /* When set indicates to always queue non-full frames.
2498 * Later the user clears this option and we transmit
2499 * any pending partial frames in the queue. This is
2500 * meant to be used alongside sendfile() to get properly
2501 * filled frames when the user (for example) must write
2502 * out headers with a write() call first and then use
2503 * sendfile to send out the data parts.
2505 * TCP_CORK can be set together with TCP_NODELAY and it is
2506 * stronger than TCP_NODELAY.
2509 tp
->nonagle
|= TCP_NAGLE_CORK
;
2511 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
2512 if (tp
->nonagle
&TCP_NAGLE_OFF
)
2513 tp
->nonagle
|= TCP_NAGLE_PUSH
;
2514 tcp_push_pending_frames(sk
);
2519 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
2522 tp
->keepalive_time
= val
* HZ
;
2523 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
2524 !((1 << sk
->sk_state
) &
2525 (TCPF_CLOSE
| TCPF_LISTEN
))) {
2526 u32 elapsed
= keepalive_time_elapsed(tp
);
2527 if (tp
->keepalive_time
> elapsed
)
2528 elapsed
= tp
->keepalive_time
- elapsed
;
2531 inet_csk_reset_keepalive_timer(sk
, elapsed
);
2536 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
2539 tp
->keepalive_intvl
= val
* HZ
;
2542 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
2545 tp
->keepalive_probes
= val
;
2548 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
2551 icsk
->icsk_syn_retries
= val
;
2557 else if (val
> sysctl_tcp_fin_timeout
/ HZ
)
2560 tp
->linger2
= val
* HZ
;
2563 case TCP_DEFER_ACCEPT
:
2564 /* Translate value in seconds to number of retransmits */
2565 icsk
->icsk_accept_queue
.rskq_defer_accept
=
2566 secs_to_retrans(val
, TCP_TIMEOUT_INIT
/ HZ
,
2570 case TCP_WINDOW_CLAMP
:
2572 if (sk
->sk_state
!= TCP_CLOSE
) {
2576 tp
->window_clamp
= 0;
2578 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
2579 SOCK_MIN_RCVBUF
/ 2 : val
;
2584 icsk
->icsk_ack
.pingpong
= 1;
2586 icsk
->icsk_ack
.pingpong
= 0;
2587 if ((1 << sk
->sk_state
) &
2588 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
2589 inet_csk_ack_scheduled(sk
)) {
2590 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
2591 tcp_cleanup_rbuf(sk
, 1);
2593 icsk
->icsk_ack
.pingpong
= 1;
2598 #ifdef CONFIG_TCP_MD5SIG
2600 /* Read the IP->Key mappings from userspace */
2601 err
= tp
->af_specific
->md5_parse(sk
, optval
, optlen
);
2604 case TCP_USER_TIMEOUT
:
2605 /* Cap the max timeout in ms TCP will retry/retrans
2606 * before giving up and aborting (ETIMEDOUT) a connection.
2611 icsk
->icsk_user_timeout
= msecs_to_jiffies(val
);
2615 if (val
>= 0 && ((1 << sk
->sk_state
) & (TCPF_CLOSE
|
2617 err
= fastopen_init_queue(sk
, val
);
2625 tp
->tsoffset
= val
- tcp_time_stamp
;
2636 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2637 unsigned int optlen
)
2639 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2641 if (level
!= SOL_TCP
)
2642 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
2644 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2646 EXPORT_SYMBOL(tcp_setsockopt
);
2648 #ifdef CONFIG_COMPAT
2649 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
2650 char __user
*optval
, unsigned int optlen
)
2652 if (level
!= SOL_TCP
)
2653 return inet_csk_compat_setsockopt(sk
, level
, optname
,
2655 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2657 EXPORT_SYMBOL(compat_tcp_setsockopt
);
2660 /* Return information about state of tcp endpoint in API format. */
2661 void tcp_get_info(const struct sock
*sk
, struct tcp_info
*info
)
2663 const struct tcp_sock
*tp
= tcp_sk(sk
);
2664 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2665 u32 now
= tcp_time_stamp
;
2667 memset(info
, 0, sizeof(*info
));
2669 info
->tcpi_state
= sk
->sk_state
;
2670 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
2671 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
2672 info
->tcpi_probes
= icsk
->icsk_probes_out
;
2673 info
->tcpi_backoff
= icsk
->icsk_backoff
;
2675 if (tp
->rx_opt
.tstamp_ok
)
2676 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
2677 if (tcp_is_sack(tp
))
2678 info
->tcpi_options
|= TCPI_OPT_SACK
;
2679 if (tp
->rx_opt
.wscale_ok
) {
2680 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
2681 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
2682 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
2685 if (tp
->ecn_flags
& TCP_ECN_OK
)
2686 info
->tcpi_options
|= TCPI_OPT_ECN
;
2687 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
2688 info
->tcpi_options
|= TCPI_OPT_ECN_SEEN
;
2689 if (tp
->syn_data_acked
)
2690 info
->tcpi_options
|= TCPI_OPT_SYN_DATA
;
2692 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
2693 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
2694 info
->tcpi_snd_mss
= tp
->mss_cache
;
2695 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
2697 if (sk
->sk_state
== TCP_LISTEN
) {
2698 info
->tcpi_unacked
= sk
->sk_ack_backlog
;
2699 info
->tcpi_sacked
= sk
->sk_max_ack_backlog
;
2701 info
->tcpi_unacked
= tp
->packets_out
;
2702 info
->tcpi_sacked
= tp
->sacked_out
;
2704 info
->tcpi_lost
= tp
->lost_out
;
2705 info
->tcpi_retrans
= tp
->retrans_out
;
2706 info
->tcpi_fackets
= tp
->fackets_out
;
2708 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
2709 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
2710 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
2712 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
2713 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
2714 info
->tcpi_rtt
= jiffies_to_usecs(tp
->srtt
)>>3;
2715 info
->tcpi_rttvar
= jiffies_to_usecs(tp
->mdev
)>>2;
2716 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
2717 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
2718 info
->tcpi_advmss
= tp
->advmss
;
2719 info
->tcpi_reordering
= tp
->reordering
;
2721 info
->tcpi_rcv_rtt
= jiffies_to_usecs(tp
->rcv_rtt_est
.rtt
)>>3;
2722 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
2724 info
->tcpi_total_retrans
= tp
->total_retrans
;
2726 EXPORT_SYMBOL_GPL(tcp_get_info
);
2728 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
2729 int optname
, char __user
*optval
, int __user
*optlen
)
2731 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2732 struct tcp_sock
*tp
= tcp_sk(sk
);
2735 if (get_user(len
, optlen
))
2738 len
= min_t(unsigned int, len
, sizeof(int));
2745 val
= tp
->mss_cache
;
2746 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
2747 val
= tp
->rx_opt
.user_mss
;
2749 val
= tp
->rx_opt
.mss_clamp
;
2752 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
2755 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
2758 val
= keepalive_time_when(tp
) / HZ
;
2761 val
= keepalive_intvl_when(tp
) / HZ
;
2764 val
= keepalive_probes(tp
);
2767 val
= icsk
->icsk_syn_retries
? : sysctl_tcp_syn_retries
;
2772 val
= (val
? : sysctl_tcp_fin_timeout
) / HZ
;
2774 case TCP_DEFER_ACCEPT
:
2775 val
= retrans_to_secs(icsk
->icsk_accept_queue
.rskq_defer_accept
,
2776 TCP_TIMEOUT_INIT
/ HZ
, TCP_RTO_MAX
/ HZ
);
2778 case TCP_WINDOW_CLAMP
:
2779 val
= tp
->window_clamp
;
2782 struct tcp_info info
;
2784 if (get_user(len
, optlen
))
2787 tcp_get_info(sk
, &info
);
2789 len
= min_t(unsigned int, len
, sizeof(info
));
2790 if (put_user(len
, optlen
))
2792 if (copy_to_user(optval
, &info
, len
))
2797 val
= !icsk
->icsk_ack
.pingpong
;
2800 case TCP_CONGESTION
:
2801 if (get_user(len
, optlen
))
2803 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
2804 if (put_user(len
, optlen
))
2806 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
2810 case TCP_THIN_LINEAR_TIMEOUTS
:
2813 case TCP_THIN_DUPACK
:
2814 val
= tp
->thin_dupack
;
2821 case TCP_REPAIR_QUEUE
:
2823 val
= tp
->repair_queue
;
2829 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2830 val
= tp
->write_seq
;
2831 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2837 case TCP_USER_TIMEOUT
:
2838 val
= jiffies_to_msecs(icsk
->icsk_user_timeout
);
2841 val
= tcp_time_stamp
+ tp
->tsoffset
;
2844 return -ENOPROTOOPT
;
2847 if (put_user(len
, optlen
))
2849 if (copy_to_user(optval
, &val
, len
))
2854 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2857 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2859 if (level
!= SOL_TCP
)
2860 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
2862 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2864 EXPORT_SYMBOL(tcp_getsockopt
);
2866 #ifdef CONFIG_COMPAT
2867 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
2868 char __user
*optval
, int __user
*optlen
)
2870 if (level
!= SOL_TCP
)
2871 return inet_csk_compat_getsockopt(sk
, level
, optname
,
2873 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2875 EXPORT_SYMBOL(compat_tcp_getsockopt
);
2878 struct sk_buff
*tcp_tso_segment(struct sk_buff
*skb
,
2879 netdev_features_t features
)
2881 struct sk_buff
*segs
= ERR_PTR(-EINVAL
);
2886 unsigned int oldlen
;
2888 struct sk_buff
*gso_skb
= skb
;
2890 bool ooo_okay
, copy_destructor
;
2892 if (!pskb_may_pull(skb
, sizeof(*th
)))
2896 thlen
= th
->doff
* 4;
2897 if (thlen
< sizeof(*th
))
2900 if (!pskb_may_pull(skb
, thlen
))
2903 oldlen
= (u16
)~skb
->len
;
2904 __skb_pull(skb
, thlen
);
2906 mss
= skb_shinfo(skb
)->gso_size
;
2907 if (unlikely(skb
->len
<= mss
))
2910 if (skb_gso_ok(skb
, features
| NETIF_F_GSO_ROBUST
)) {
2911 /* Packet is from an untrusted source, reset gso_segs. */
2912 int type
= skb_shinfo(skb
)->gso_type
;
2920 SKB_GSO_UDP_TUNNEL
|
2922 !(type
& (SKB_GSO_TCPV4
| SKB_GSO_TCPV6
))))
2925 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss
);
2931 copy_destructor
= gso_skb
->destructor
== tcp_wfree
;
2932 ooo_okay
= gso_skb
->ooo_okay
;
2933 /* All segments but the first should have ooo_okay cleared */
2936 segs
= skb_segment(skb
, features
);
2940 /* Only first segment might have ooo_okay set */
2941 segs
->ooo_okay
= ooo_okay
;
2943 delta
= htonl(oldlen
+ (thlen
+ mss
));
2947 seq
= ntohl(th
->seq
);
2949 newcheck
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
2950 (__force u32
)delta
));
2953 th
->fin
= th
->psh
= 0;
2954 th
->check
= newcheck
;
2956 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2958 csum_fold(csum_partial(skb_transport_header(skb
),
2962 if (copy_destructor
) {
2963 skb
->destructor
= gso_skb
->destructor
;
2964 skb
->sk
= gso_skb
->sk
;
2965 /* {tcp|sock}_wfree() use exact truesize accounting :
2966 * sum(skb->truesize) MUST be exactly be gso_skb->truesize
2967 * So we account mss bytes of 'true size' for each segment.
2968 * The last segment will contain the remaining.
2970 skb
->truesize
= mss
;
2971 gso_skb
->truesize
-= mss
;
2976 th
->seq
= htonl(seq
);
2978 } while (skb
->next
);
2980 /* Following permits TCP Small Queues to work well with GSO :
2981 * The callback to TCP stack will be called at the time last frag
2982 * is freed at TX completion, and not right now when gso_skb
2983 * is freed by GSO engine
2985 if (copy_destructor
) {
2986 swap(gso_skb
->sk
, skb
->sk
);
2987 swap(gso_skb
->destructor
, skb
->destructor
);
2988 swap(gso_skb
->truesize
, skb
->truesize
);
2991 delta
= htonl(oldlen
+ (skb
->tail
- skb
->transport_header
) +
2993 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
2994 (__force u32
)delta
));
2995 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2996 th
->check
= csum_fold(csum_partial(skb_transport_header(skb
),
3002 EXPORT_SYMBOL(tcp_tso_segment
);
3004 struct sk_buff
**tcp_gro_receive(struct sk_buff
**head
, struct sk_buff
*skb
)
3006 struct sk_buff
**pp
= NULL
;
3013 unsigned int mss
= 1;
3019 off
= skb_gro_offset(skb
);
3020 hlen
= off
+ sizeof(*th
);
3021 th
= skb_gro_header_fast(skb
, off
);
3022 if (skb_gro_header_hard(skb
, hlen
)) {
3023 th
= skb_gro_header_slow(skb
, hlen
, off
);
3028 thlen
= th
->doff
* 4;
3029 if (thlen
< sizeof(*th
))
3033 if (skb_gro_header_hard(skb
, hlen
)) {
3034 th
= skb_gro_header_slow(skb
, hlen
, off
);
3039 skb_gro_pull(skb
, thlen
);
3041 len
= skb_gro_len(skb
);
3042 flags
= tcp_flag_word(th
);
3044 for (; (p
= *head
); head
= &p
->next
) {
3045 if (!NAPI_GRO_CB(p
)->same_flow
)
3050 if (*(u32
*)&th
->source
^ *(u32
*)&th2
->source
) {
3051 NAPI_GRO_CB(p
)->same_flow
= 0;
3058 goto out_check_final
;
3061 flush
= NAPI_GRO_CB(p
)->flush
;
3062 flush
|= (__force
int)(flags
& TCP_FLAG_CWR
);
3063 flush
|= (__force
int)((flags
^ tcp_flag_word(th2
)) &
3064 ~(TCP_FLAG_CWR
| TCP_FLAG_FIN
| TCP_FLAG_PSH
));
3065 flush
|= (__force
int)(th
->ack_seq
^ th2
->ack_seq
);
3066 for (i
= sizeof(*th
); i
< thlen
; i
+= 4)
3067 flush
|= *(u32
*)((u8
*)th
+ i
) ^
3068 *(u32
*)((u8
*)th2
+ i
);
3070 mss
= skb_shinfo(p
)->gso_size
;
3072 flush
|= (len
- 1) >= mss
;
3073 flush
|= (ntohl(th2
->seq
) + skb_gro_len(p
)) ^ ntohl(th
->seq
);
3075 if (flush
|| skb_gro_receive(head
, skb
)) {
3077 goto out_check_final
;
3082 tcp_flag_word(th2
) |= flags
& (TCP_FLAG_FIN
| TCP_FLAG_PSH
);
3086 flush
|= (__force
int)(flags
& (TCP_FLAG_URG
| TCP_FLAG_PSH
|
3087 TCP_FLAG_RST
| TCP_FLAG_SYN
|
3090 if (p
&& (!NAPI_GRO_CB(skb
)->same_flow
|| flush
))
3094 NAPI_GRO_CB(skb
)->flush
|= flush
;
3098 EXPORT_SYMBOL(tcp_gro_receive
);
3100 int tcp_gro_complete(struct sk_buff
*skb
)
3102 struct tcphdr
*th
= tcp_hdr(skb
);
3104 skb
->csum_start
= skb_transport_header(skb
) - skb
->head
;
3105 skb
->csum_offset
= offsetof(struct tcphdr
, check
);
3106 skb
->ip_summed
= CHECKSUM_PARTIAL
;
3108 skb_shinfo(skb
)->gso_segs
= NAPI_GRO_CB(skb
)->count
;
3111 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
3115 EXPORT_SYMBOL(tcp_gro_complete
);
3117 #ifdef CONFIG_TCP_MD5SIG
3118 static unsigned long tcp_md5sig_users
;
3119 static struct tcp_md5sig_pool __percpu
*tcp_md5sig_pool
;
3120 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock
);
3122 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu
*pool
)
3126 for_each_possible_cpu(cpu
) {
3127 struct tcp_md5sig_pool
*p
= per_cpu_ptr(pool
, cpu
);
3129 if (p
->md5_desc
.tfm
)
3130 crypto_free_hash(p
->md5_desc
.tfm
);
3135 void tcp_free_md5sig_pool(void)
3137 struct tcp_md5sig_pool __percpu
*pool
= NULL
;
3139 spin_lock_bh(&tcp_md5sig_pool_lock
);
3140 if (--tcp_md5sig_users
== 0) {
3141 pool
= tcp_md5sig_pool
;
3142 tcp_md5sig_pool
= NULL
;
3144 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3146 __tcp_free_md5sig_pool(pool
);
3148 EXPORT_SYMBOL(tcp_free_md5sig_pool
);
3150 static struct tcp_md5sig_pool __percpu
*
3151 __tcp_alloc_md5sig_pool(struct sock
*sk
)
3154 struct tcp_md5sig_pool __percpu
*pool
;
3156 pool
= alloc_percpu(struct tcp_md5sig_pool
);
3160 for_each_possible_cpu(cpu
) {
3161 struct crypto_hash
*hash
;
3163 hash
= crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC
);
3164 if (IS_ERR_OR_NULL(hash
))
3167 per_cpu_ptr(pool
, cpu
)->md5_desc
.tfm
= hash
;
3171 __tcp_free_md5sig_pool(pool
);
3175 struct tcp_md5sig_pool __percpu
*tcp_alloc_md5sig_pool(struct sock
*sk
)
3177 struct tcp_md5sig_pool __percpu
*pool
;
3181 spin_lock_bh(&tcp_md5sig_pool_lock
);
3182 pool
= tcp_md5sig_pool
;
3183 if (tcp_md5sig_users
++ == 0) {
3185 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3188 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3192 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3195 /* we cannot hold spinlock here because this may sleep. */
3196 struct tcp_md5sig_pool __percpu
*p
;
3198 p
= __tcp_alloc_md5sig_pool(sk
);
3199 spin_lock_bh(&tcp_md5sig_pool_lock
);
3202 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3205 pool
= tcp_md5sig_pool
;
3207 /* oops, it has already been assigned. */
3208 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3209 __tcp_free_md5sig_pool(p
);
3211 tcp_md5sig_pool
= pool
= p
;
3212 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3217 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
3221 * tcp_get_md5sig_pool - get md5sig_pool for this user
3223 * We use percpu structure, so if we succeed, we exit with preemption
3224 * and BH disabled, to make sure another thread or softirq handling
3225 * wont try to get same context.
3227 struct tcp_md5sig_pool
*tcp_get_md5sig_pool(void)
3229 struct tcp_md5sig_pool __percpu
*p
;
3233 spin_lock(&tcp_md5sig_pool_lock
);
3234 p
= tcp_md5sig_pool
;
3237 spin_unlock(&tcp_md5sig_pool_lock
);
3240 return this_cpu_ptr(p
);
3245 EXPORT_SYMBOL(tcp_get_md5sig_pool
);
3247 void tcp_put_md5sig_pool(void)
3250 tcp_free_md5sig_pool();
3252 EXPORT_SYMBOL(tcp_put_md5sig_pool
);
3254 int tcp_md5_hash_header(struct tcp_md5sig_pool
*hp
,
3255 const struct tcphdr
*th
)
3257 struct scatterlist sg
;
3261 /* We are not allowed to change tcphdr, make a local copy */
3262 memcpy(&hdr
, th
, sizeof(hdr
));
3265 /* options aren't included in the hash */
3266 sg_init_one(&sg
, &hdr
, sizeof(hdr
));
3267 err
= crypto_hash_update(&hp
->md5_desc
, &sg
, sizeof(hdr
));
3270 EXPORT_SYMBOL(tcp_md5_hash_header
);
3272 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool
*hp
,
3273 const struct sk_buff
*skb
, unsigned int header_len
)
3275 struct scatterlist sg
;
3276 const struct tcphdr
*tp
= tcp_hdr(skb
);
3277 struct hash_desc
*desc
= &hp
->md5_desc
;
3279 const unsigned int head_data_len
= skb_headlen(skb
) > header_len
?
3280 skb_headlen(skb
) - header_len
: 0;
3281 const struct skb_shared_info
*shi
= skb_shinfo(skb
);
3282 struct sk_buff
*frag_iter
;
3284 sg_init_table(&sg
, 1);
3286 sg_set_buf(&sg
, ((u8
*) tp
) + header_len
, head_data_len
);
3287 if (crypto_hash_update(desc
, &sg
, head_data_len
))
3290 for (i
= 0; i
< shi
->nr_frags
; ++i
) {
3291 const struct skb_frag_struct
*f
= &shi
->frags
[i
];
3292 unsigned int offset
= f
->page_offset
;
3293 struct page
*page
= skb_frag_page(f
) + (offset
>> PAGE_SHIFT
);
3295 sg_set_page(&sg
, page
, skb_frag_size(f
),
3296 offset_in_page(offset
));
3297 if (crypto_hash_update(desc
, &sg
, skb_frag_size(f
)))
3301 skb_walk_frags(skb
, frag_iter
)
3302 if (tcp_md5_hash_skb_data(hp
, frag_iter
, 0))
3307 EXPORT_SYMBOL(tcp_md5_hash_skb_data
);
3309 int tcp_md5_hash_key(struct tcp_md5sig_pool
*hp
, const struct tcp_md5sig_key
*key
)
3311 struct scatterlist sg
;
3313 sg_init_one(&sg
, key
->key
, key
->keylen
);
3314 return crypto_hash_update(&hp
->md5_desc
, &sg
, key
->keylen
);
3316 EXPORT_SYMBOL(tcp_md5_hash_key
);
3320 void tcp_done(struct sock
*sk
)
3322 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
3324 if (sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
3325 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_ATTEMPTFAILS
);
3327 tcp_set_state(sk
, TCP_CLOSE
);
3328 tcp_clear_xmit_timers(sk
);
3330 reqsk_fastopen_remove(sk
, req
, false);
3332 sk
->sk_shutdown
= SHUTDOWN_MASK
;
3334 if (!sock_flag(sk
, SOCK_DEAD
))
3335 sk
->sk_state_change(sk
);
3337 inet_csk_destroy_sock(sk
);
3339 EXPORT_SYMBOL_GPL(tcp_done
);
3341 extern struct tcp_congestion_ops tcp_reno
;
3343 static __initdata
unsigned long thash_entries
;
3344 static int __init
set_thash_entries(char *str
)
3351 ret
= kstrtoul(str
, 0, &thash_entries
);
3357 __setup("thash_entries=", set_thash_entries
);
3359 void tcp_init_mem(struct net
*net
)
3361 unsigned long limit
= nr_free_buffer_pages() / 8;
3362 limit
= max(limit
, 128UL);
3363 net
->ipv4
.sysctl_tcp_mem
[0] = limit
/ 4 * 3;
3364 net
->ipv4
.sysctl_tcp_mem
[1] = limit
;
3365 net
->ipv4
.sysctl_tcp_mem
[2] = net
->ipv4
.sysctl_tcp_mem
[0] * 2;
3368 void __init
tcp_init(void)
3370 struct sk_buff
*skb
= NULL
;
3371 unsigned long limit
;
3372 int max_rshare
, max_wshare
, cnt
;
3375 BUILD_BUG_ON(sizeof(struct tcp_skb_cb
) > sizeof(skb
->cb
));
3377 percpu_counter_init(&tcp_sockets_allocated
, 0);
3378 percpu_counter_init(&tcp_orphan_count
, 0);
3379 tcp_hashinfo
.bind_bucket_cachep
=
3380 kmem_cache_create("tcp_bind_bucket",
3381 sizeof(struct inet_bind_bucket
), 0,
3382 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
);
3384 /* Size and allocate the main established and bind bucket
3387 * The methodology is similar to that of the buffer cache.
3389 tcp_hashinfo
.ehash
=
3390 alloc_large_system_hash("TCP established",
3391 sizeof(struct inet_ehash_bucket
),
3393 17, /* one slot per 128 KB of memory */
3396 &tcp_hashinfo
.ehash_mask
,
3398 thash_entries
? 0 : 512 * 1024);
3399 for (i
= 0; i
<= tcp_hashinfo
.ehash_mask
; i
++) {
3400 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].chain
, i
);
3401 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].twchain
, i
);
3403 if (inet_ehash_locks_alloc(&tcp_hashinfo
))
3404 panic("TCP: failed to alloc ehash_locks");
3405 tcp_hashinfo
.bhash
=
3406 alloc_large_system_hash("TCP bind",
3407 sizeof(struct inet_bind_hashbucket
),
3408 tcp_hashinfo
.ehash_mask
+ 1,
3409 17, /* one slot per 128 KB of memory */
3411 &tcp_hashinfo
.bhash_size
,
3415 tcp_hashinfo
.bhash_size
= 1U << tcp_hashinfo
.bhash_size
;
3416 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
3417 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
3418 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
3422 cnt
= tcp_hashinfo
.ehash_mask
+ 1;
3424 tcp_death_row
.sysctl_max_tw_buckets
= cnt
/ 2;
3425 sysctl_tcp_max_orphans
= cnt
/ 2;
3426 sysctl_max_syn_backlog
= max(128, cnt
/ 256);
3428 tcp_init_mem(&init_net
);
3429 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3430 limit
= nr_free_buffer_pages() << (PAGE_SHIFT
- 7);
3431 max_wshare
= min(4UL*1024*1024, limit
);
3432 max_rshare
= min(6UL*1024*1024, limit
);
3434 sysctl_tcp_wmem
[0] = SK_MEM_QUANTUM
;
3435 sysctl_tcp_wmem
[1] = 16*1024;
3436 sysctl_tcp_wmem
[2] = max(64*1024, max_wshare
);
3438 sysctl_tcp_rmem
[0] = SK_MEM_QUANTUM
;
3439 sysctl_tcp_rmem
[1] = 87380;
3440 sysctl_tcp_rmem
[2] = max(87380, max_rshare
);
3442 pr_info("Hash tables configured (established %u bind %u)\n",
3443 tcp_hashinfo
.ehash_mask
+ 1, tcp_hashinfo
.bhash_size
);
3447 tcp_register_congestion_control(&tcp_reno
);