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 int sysctl_tcp_min_tso_segs __read_mostly
= 2;
287 struct percpu_counter tcp_orphan_count
;
288 EXPORT_SYMBOL_GPL(tcp_orphan_count
);
290 int sysctl_tcp_wmem
[3] __read_mostly
;
291 int sysctl_tcp_rmem
[3] __read_mostly
;
293 EXPORT_SYMBOL(sysctl_tcp_rmem
);
294 EXPORT_SYMBOL(sysctl_tcp_wmem
);
296 atomic_long_t tcp_memory_allocated
; /* Current allocated memory. */
297 EXPORT_SYMBOL(tcp_memory_allocated
);
300 * Current number of TCP sockets.
302 struct percpu_counter tcp_sockets_allocated
;
303 EXPORT_SYMBOL(tcp_sockets_allocated
);
308 struct tcp_splice_state
{
309 struct pipe_inode_info
*pipe
;
315 * Pressure flag: try to collapse.
316 * Technical note: it is used by multiple contexts non atomically.
317 * All the __sk_mem_schedule() is of this nature: accounting
318 * is strict, actions are advisory and have some latency.
320 int tcp_memory_pressure __read_mostly
;
321 EXPORT_SYMBOL(tcp_memory_pressure
);
323 void tcp_enter_memory_pressure(struct sock
*sk
)
325 if (!tcp_memory_pressure
) {
326 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPMEMORYPRESSURES
);
327 tcp_memory_pressure
= 1;
330 EXPORT_SYMBOL(tcp_enter_memory_pressure
);
332 /* Convert seconds to retransmits based on initial and max timeout */
333 static u8
secs_to_retrans(int seconds
, int timeout
, int rto_max
)
338 int period
= timeout
;
341 while (seconds
> period
&& res
< 255) {
344 if (timeout
> rto_max
)
352 /* Convert retransmits to seconds based on initial and max timeout */
353 static int retrans_to_secs(u8 retrans
, int timeout
, int rto_max
)
361 if (timeout
> rto_max
)
369 /* Address-family independent initialization for a tcp_sock.
371 * NOTE: A lot of things set to zero explicitly by call to
372 * sk_alloc() so need not be done here.
374 void tcp_init_sock(struct sock
*sk
)
376 struct inet_connection_sock
*icsk
= inet_csk(sk
);
377 struct tcp_sock
*tp
= tcp_sk(sk
);
379 skb_queue_head_init(&tp
->out_of_order_queue
);
380 tcp_init_xmit_timers(sk
);
381 tcp_prequeue_init(tp
);
382 INIT_LIST_HEAD(&tp
->tsq_node
);
384 icsk
->icsk_rto
= TCP_TIMEOUT_INIT
;
385 tp
->mdev
= TCP_TIMEOUT_INIT
;
387 /* So many TCP implementations out there (incorrectly) count the
388 * initial SYN frame in their delayed-ACK and congestion control
389 * algorithms that we must have the following bandaid to talk
390 * efficiently to them. -DaveM
392 tp
->snd_cwnd
= TCP_INIT_CWND
;
394 /* See draft-stevens-tcpca-spec-01 for discussion of the
395 * initialization of these values.
397 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
398 tp
->snd_cwnd_clamp
= ~0;
399 tp
->mss_cache
= TCP_MSS_DEFAULT
;
401 tp
->reordering
= sysctl_tcp_reordering
;
402 tcp_enable_early_retrans(tp
);
403 icsk
->icsk_ca_ops
= &tcp_init_congestion_ops
;
407 sk
->sk_state
= TCP_CLOSE
;
409 sk
->sk_write_space
= sk_stream_write_space
;
410 sock_set_flag(sk
, SOCK_USE_WRITE_QUEUE
);
412 icsk
->icsk_sync_mss
= tcp_sync_mss
;
414 /* Presumed zeroed, in order of appearance:
415 * cookie_in_always, cookie_out_never,
416 * s_data_constant, s_data_in, s_data_out
418 sk
->sk_sndbuf
= sysctl_tcp_wmem
[1];
419 sk
->sk_rcvbuf
= sysctl_tcp_rmem
[1];
422 sock_update_memcg(sk
);
423 sk_sockets_allocated_inc(sk
);
426 EXPORT_SYMBOL(tcp_init_sock
);
429 * Wait for a TCP event.
431 * Note that we don't need to lock the socket, as the upper poll layers
432 * take care of normal races (between the test and the event) and we don't
433 * go look at any of the socket buffers directly.
435 unsigned int tcp_poll(struct file
*file
, struct socket
*sock
, poll_table
*wait
)
438 struct sock
*sk
= sock
->sk
;
439 const struct tcp_sock
*tp
= tcp_sk(sk
);
441 sock_poll_wait(file
, sk_sleep(sk
), wait
);
442 if (sk
->sk_state
== TCP_LISTEN
)
443 return inet_csk_listen_poll(sk
);
445 /* Socket is not locked. We are protected from async events
446 * by poll logic and correct handling of state changes
447 * made by other threads is impossible in any case.
453 * POLLHUP is certainly not done right. But poll() doesn't
454 * have a notion of HUP in just one direction, and for a
455 * socket the read side is more interesting.
457 * Some poll() documentation says that POLLHUP is incompatible
458 * with the POLLOUT/POLLWR flags, so somebody should check this
459 * all. But careful, it tends to be safer to return too many
460 * bits than too few, and you can easily break real applications
461 * if you don't tell them that something has hung up!
465 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
466 * our fs/select.c). It means that after we received EOF,
467 * poll always returns immediately, making impossible poll() on write()
468 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
469 * if and only if shutdown has been made in both directions.
470 * Actually, it is interesting to look how Solaris and DUX
471 * solve this dilemma. I would prefer, if POLLHUP were maskable,
472 * then we could set it on SND_SHUTDOWN. BTW examples given
473 * in Stevens' books assume exactly this behaviour, it explains
474 * why POLLHUP is incompatible with POLLOUT. --ANK
476 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
477 * blocking on fresh not-connected or disconnected socket. --ANK
479 if (sk
->sk_shutdown
== SHUTDOWN_MASK
|| sk
->sk_state
== TCP_CLOSE
)
481 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
482 mask
|= POLLIN
| POLLRDNORM
| POLLRDHUP
;
484 /* Connected or passive Fast Open socket? */
485 if (sk
->sk_state
!= TCP_SYN_SENT
&&
486 (sk
->sk_state
!= TCP_SYN_RECV
|| tp
->fastopen_rsk
!= NULL
)) {
487 int target
= sock_rcvlowat(sk
, 0, INT_MAX
);
489 if (tp
->urg_seq
== tp
->copied_seq
&&
490 !sock_flag(sk
, SOCK_URGINLINE
) &&
494 /* Potential race condition. If read of tp below will
495 * escape above sk->sk_state, we can be illegally awaken
496 * in SYN_* states. */
497 if (tp
->rcv_nxt
- tp
->copied_seq
>= target
)
498 mask
|= POLLIN
| POLLRDNORM
;
500 if (!(sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
501 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
)) {
502 mask
|= POLLOUT
| POLLWRNORM
;
503 } else { /* send SIGIO later */
504 set_bit(SOCK_ASYNC_NOSPACE
,
505 &sk
->sk_socket
->flags
);
506 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
508 /* Race breaker. If space is freed after
509 * wspace test but before the flags are set,
510 * IO signal will be lost.
512 if (sk_stream_wspace(sk
) >= sk_stream_min_wspace(sk
))
513 mask
|= POLLOUT
| POLLWRNORM
;
516 mask
|= POLLOUT
| POLLWRNORM
;
518 if (tp
->urg_data
& TCP_URG_VALID
)
521 /* This barrier is coupled with smp_wmb() in tcp_reset() */
528 EXPORT_SYMBOL(tcp_poll
);
530 int tcp_ioctl(struct sock
*sk
, int cmd
, unsigned long arg
)
532 struct tcp_sock
*tp
= tcp_sk(sk
);
538 if (sk
->sk_state
== TCP_LISTEN
)
541 slow
= lock_sock_fast(sk
);
542 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
544 else if (sock_flag(sk
, SOCK_URGINLINE
) ||
546 before(tp
->urg_seq
, tp
->copied_seq
) ||
547 !before(tp
->urg_seq
, tp
->rcv_nxt
)) {
549 answ
= tp
->rcv_nxt
- tp
->copied_seq
;
551 /* Subtract 1, if FIN was received */
552 if (answ
&& sock_flag(sk
, SOCK_DONE
))
555 answ
= tp
->urg_seq
- tp
->copied_seq
;
556 unlock_sock_fast(sk
, slow
);
559 answ
= tp
->urg_data
&& tp
->urg_seq
== tp
->copied_seq
;
562 if (sk
->sk_state
== TCP_LISTEN
)
565 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
568 answ
= tp
->write_seq
- tp
->snd_una
;
571 if (sk
->sk_state
== TCP_LISTEN
)
574 if ((1 << sk
->sk_state
) & (TCPF_SYN_SENT
| TCPF_SYN_RECV
))
577 answ
= tp
->write_seq
- tp
->snd_nxt
;
583 return put_user(answ
, (int __user
*)arg
);
585 EXPORT_SYMBOL(tcp_ioctl
);
587 static inline void tcp_mark_push(struct tcp_sock
*tp
, struct sk_buff
*skb
)
589 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
590 tp
->pushed_seq
= tp
->write_seq
;
593 static inline bool forced_push(const struct tcp_sock
*tp
)
595 return after(tp
->write_seq
, tp
->pushed_seq
+ (tp
->max_window
>> 1));
598 static inline void skb_entail(struct sock
*sk
, struct sk_buff
*skb
)
600 struct tcp_sock
*tp
= tcp_sk(sk
);
601 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
604 tcb
->seq
= tcb
->end_seq
= tp
->write_seq
;
605 tcb
->tcp_flags
= TCPHDR_ACK
;
607 skb_header_release(skb
);
608 tcp_add_write_queue_tail(sk
, skb
);
609 sk
->sk_wmem_queued
+= skb
->truesize
;
610 sk_mem_charge(sk
, skb
->truesize
);
611 if (tp
->nonagle
& TCP_NAGLE_PUSH
)
612 tp
->nonagle
&= ~TCP_NAGLE_PUSH
;
615 static inline void tcp_mark_urg(struct tcp_sock
*tp
, int flags
)
618 tp
->snd_up
= tp
->write_seq
;
621 static inline void tcp_push(struct sock
*sk
, int flags
, int mss_now
,
624 if (tcp_send_head(sk
)) {
625 struct tcp_sock
*tp
= tcp_sk(sk
);
627 if (!(flags
& MSG_MORE
) || forced_push(tp
))
628 tcp_mark_push(tp
, tcp_write_queue_tail(sk
));
630 tcp_mark_urg(tp
, flags
);
631 __tcp_push_pending_frames(sk
, mss_now
,
632 (flags
& MSG_MORE
) ? TCP_NAGLE_CORK
: nonagle
);
636 static int tcp_splice_data_recv(read_descriptor_t
*rd_desc
, struct sk_buff
*skb
,
637 unsigned int offset
, size_t len
)
639 struct tcp_splice_state
*tss
= rd_desc
->arg
.data
;
642 ret
= skb_splice_bits(skb
, offset
, tss
->pipe
, min(rd_desc
->count
, len
),
645 rd_desc
->count
-= ret
;
649 static int __tcp_splice_read(struct sock
*sk
, struct tcp_splice_state
*tss
)
651 /* Store TCP splice context information in read_descriptor_t. */
652 read_descriptor_t rd_desc
= {
657 return tcp_read_sock(sk
, &rd_desc
, tcp_splice_data_recv
);
661 * tcp_splice_read - splice data from TCP socket to a pipe
662 * @sock: socket to splice from
663 * @ppos: position (not valid)
664 * @pipe: pipe to splice to
665 * @len: number of bytes to splice
666 * @flags: splice modifier flags
669 * Will read pages from given socket and fill them into a pipe.
672 ssize_t
tcp_splice_read(struct socket
*sock
, loff_t
*ppos
,
673 struct pipe_inode_info
*pipe
, size_t len
,
676 struct sock
*sk
= sock
->sk
;
677 struct tcp_splice_state tss
= {
686 sock_rps_record_flow(sk
);
688 * We can't seek on a socket input
697 timeo
= sock_rcvtimeo(sk
, sock
->file
->f_flags
& O_NONBLOCK
);
699 ret
= __tcp_splice_read(sk
, &tss
);
705 if (sock_flag(sk
, SOCK_DONE
))
708 ret
= sock_error(sk
);
711 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
713 if (sk
->sk_state
== TCP_CLOSE
) {
715 * This occurs when user tries to read
716 * from never connected socket.
718 if (!sock_flag(sk
, SOCK_DONE
))
726 sk_wait_data(sk
, &timeo
);
727 if (signal_pending(current
)) {
728 ret
= sock_intr_errno(timeo
);
741 if (sk
->sk_err
|| sk
->sk_state
== TCP_CLOSE
||
742 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
743 signal_pending(current
))
754 EXPORT_SYMBOL(tcp_splice_read
);
756 struct sk_buff
*sk_stream_alloc_skb(struct sock
*sk
, int size
, gfp_t gfp
)
760 /* The TCP header must be at least 32-bit aligned. */
761 size
= ALIGN(size
, 4);
763 skb
= alloc_skb_fclone(size
+ sk
->sk_prot
->max_header
, gfp
);
765 if (sk_wmem_schedule(sk
, skb
->truesize
)) {
766 skb_reserve(skb
, sk
->sk_prot
->max_header
);
768 * Make sure that we have exactly size bytes
769 * available to the caller, no more, no less.
771 skb
->reserved_tailroom
= skb
->end
- skb
->tail
- size
;
776 sk
->sk_prot
->enter_memory_pressure(sk
);
777 sk_stream_moderate_sndbuf(sk
);
782 static unsigned int tcp_xmit_size_goal(struct sock
*sk
, u32 mss_now
,
785 struct tcp_sock
*tp
= tcp_sk(sk
);
786 u32 xmit_size_goal
, old_size_goal
;
788 xmit_size_goal
= mss_now
;
790 if (large_allowed
&& sk_can_gso(sk
)) {
793 /* Maybe we should/could use sk->sk_prot->max_header here ? */
794 hlen
= inet_csk(sk
)->icsk_af_ops
->net_header_len
+
795 inet_csk(sk
)->icsk_ext_hdr_len
+
798 /* Goal is to send at least one packet per ms,
799 * not one big TSO packet every 100 ms.
800 * This preserves ACK clocking and is consistent
801 * with tcp_tso_should_defer() heuristic.
803 gso_size
= sk
->sk_pacing_rate
/ (2 * MSEC_PER_SEC
);
804 gso_size
= max_t(u32
, gso_size
,
805 sysctl_tcp_min_tso_segs
* mss_now
);
807 xmit_size_goal
= min_t(u32
, gso_size
,
808 sk
->sk_gso_max_size
- 1 - hlen
);
810 xmit_size_goal
= tcp_bound_to_half_wnd(tp
, xmit_size_goal
);
812 /* We try hard to avoid divides here */
813 old_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
815 if (likely(old_size_goal
<= xmit_size_goal
&&
816 old_size_goal
+ mss_now
> xmit_size_goal
)) {
817 xmit_size_goal
= old_size_goal
;
819 tp
->xmit_size_goal_segs
=
820 min_t(u16
, xmit_size_goal
/ mss_now
,
821 sk
->sk_gso_max_segs
);
822 xmit_size_goal
= tp
->xmit_size_goal_segs
* mss_now
;
826 return max(xmit_size_goal
, mss_now
);
829 static int tcp_send_mss(struct sock
*sk
, int *size_goal
, int flags
)
833 mss_now
= tcp_current_mss(sk
);
834 *size_goal
= tcp_xmit_size_goal(sk
, mss_now
, !(flags
& MSG_OOB
));
839 static ssize_t
do_tcp_sendpages(struct sock
*sk
, struct page
*page
, int offset
,
840 size_t size
, int flags
)
842 struct tcp_sock
*tp
= tcp_sk(sk
);
843 int mss_now
, size_goal
;
846 long timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
848 /* Wait for a connection to finish. One exception is TCP Fast Open
849 * (passive side) where data is allowed to be sent before a connection
850 * is fully established.
852 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
853 !tcp_passive_fastopen(sk
)) {
854 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
858 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
860 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
864 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
868 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
872 if (!tcp_send_head(sk
) || (copy
= size_goal
- skb
->len
) <= 0) {
874 if (!sk_stream_memory_free(sk
))
875 goto wait_for_sndbuf
;
877 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
879 goto wait_for_memory
;
888 i
= skb_shinfo(skb
)->nr_frags
;
889 can_coalesce
= skb_can_coalesce(skb
, i
, page
, offset
);
890 if (!can_coalesce
&& i
>= MAX_SKB_FRAGS
) {
891 tcp_mark_push(tp
, skb
);
894 if (!sk_wmem_schedule(sk
, copy
))
895 goto wait_for_memory
;
898 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
901 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
903 skb_shinfo(skb
)->tx_flags
|= SKBTX_SHARED_FRAG
;
906 skb
->data_len
+= copy
;
907 skb
->truesize
+= copy
;
908 sk
->sk_wmem_queued
+= copy
;
909 sk_mem_charge(sk
, copy
);
910 skb
->ip_summed
= CHECKSUM_PARTIAL
;
911 tp
->write_seq
+= copy
;
912 TCP_SKB_CB(skb
)->end_seq
+= copy
;
913 skb_shinfo(skb
)->gso_segs
= 0;
916 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
923 if (skb
->len
< size_goal
|| (flags
& MSG_OOB
))
926 if (forced_push(tp
)) {
927 tcp_mark_push(tp
, skb
);
928 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
929 } else if (skb
== tcp_send_head(sk
))
930 tcp_push_one(sk
, mss_now
);
934 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
936 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
938 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
941 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
945 if (copied
&& !(flags
& MSG_SENDPAGE_NOTLAST
))
946 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
953 return sk_stream_error(sk
, flags
, err
);
956 int tcp_sendpage(struct sock
*sk
, struct page
*page
, int offset
,
957 size_t size
, int flags
)
961 if (!(sk
->sk_route_caps
& NETIF_F_SG
) ||
962 !(sk
->sk_route_caps
& NETIF_F_ALL_CSUM
))
963 return sock_no_sendpage(sk
->sk_socket
, page
, offset
, size
,
967 res
= do_tcp_sendpages(sk
, page
, offset
, size
, flags
);
971 EXPORT_SYMBOL(tcp_sendpage
);
973 static inline int select_size(const struct sock
*sk
, bool sg
)
975 const struct tcp_sock
*tp
= tcp_sk(sk
);
976 int tmp
= tp
->mss_cache
;
979 if (sk_can_gso(sk
)) {
980 /* Small frames wont use a full page:
981 * Payload will immediately follow tcp header.
983 tmp
= SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER
);
985 int pgbreak
= SKB_MAX_HEAD(MAX_TCP_HEADER
);
987 if (tmp
>= pgbreak
&&
988 tmp
<= pgbreak
+ (MAX_SKB_FRAGS
- 1) * PAGE_SIZE
)
996 void tcp_free_fastopen_req(struct tcp_sock
*tp
)
998 if (tp
->fastopen_req
!= NULL
) {
999 kfree(tp
->fastopen_req
);
1000 tp
->fastopen_req
= NULL
;
1004 static int tcp_sendmsg_fastopen(struct sock
*sk
, struct msghdr
*msg
,
1005 int *copied
, size_t size
)
1007 struct tcp_sock
*tp
= tcp_sk(sk
);
1010 if (!(sysctl_tcp_fastopen
& TFO_CLIENT_ENABLE
))
1012 if (tp
->fastopen_req
!= NULL
)
1013 return -EALREADY
; /* Another Fast Open is in progress */
1015 tp
->fastopen_req
= kzalloc(sizeof(struct tcp_fastopen_request
),
1017 if (unlikely(tp
->fastopen_req
== NULL
))
1019 tp
->fastopen_req
->data
= msg
;
1020 tp
->fastopen_req
->size
= size
;
1022 flags
= (msg
->msg_flags
& MSG_DONTWAIT
) ? O_NONBLOCK
: 0;
1023 err
= __inet_stream_connect(sk
->sk_socket
, msg
->msg_name
,
1024 msg
->msg_namelen
, flags
);
1025 *copied
= tp
->fastopen_req
->copied
;
1026 tcp_free_fastopen_req(tp
);
1030 int tcp_sendmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1034 struct tcp_sock
*tp
= tcp_sk(sk
);
1035 struct sk_buff
*skb
;
1036 int iovlen
, flags
, err
, copied
= 0;
1037 int mss_now
= 0, size_goal
, copied_syn
= 0, offset
= 0;
1043 flags
= msg
->msg_flags
;
1044 if (flags
& MSG_FASTOPEN
) {
1045 err
= tcp_sendmsg_fastopen(sk
, msg
, &copied_syn
, size
);
1046 if (err
== -EINPROGRESS
&& copied_syn
> 0)
1050 offset
= copied_syn
;
1053 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
1055 /* Wait for a connection to finish. One exception is TCP Fast Open
1056 * (passive side) where data is allowed to be sent before a connection
1057 * is fully established.
1059 if (((1 << sk
->sk_state
) & ~(TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
)) &&
1060 !tcp_passive_fastopen(sk
)) {
1061 if ((err
= sk_stream_wait_connect(sk
, &timeo
)) != 0)
1065 if (unlikely(tp
->repair
)) {
1066 if (tp
->repair_queue
== TCP_RECV_QUEUE
) {
1067 copied
= tcp_send_rcvq(sk
, msg
, size
);
1072 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1075 /* 'common' sending to sendq */
1078 /* This should be in poll */
1079 clear_bit(SOCK_ASYNC_NOSPACE
, &sk
->sk_socket
->flags
);
1081 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1083 /* Ok commence sending. */
1084 iovlen
= msg
->msg_iovlen
;
1089 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
))
1092 sg
= !!(sk
->sk_route_caps
& NETIF_F_SG
);
1094 while (--iovlen
>= 0) {
1095 size_t seglen
= iov
->iov_len
;
1096 unsigned char __user
*from
= iov
->iov_base
;
1099 if (unlikely(offset
> 0)) { /* Skip bytes copied in SYN */
1100 if (offset
>= seglen
) {
1109 while (seglen
> 0) {
1111 int max
= size_goal
;
1113 skb
= tcp_write_queue_tail(sk
);
1114 if (tcp_send_head(sk
)) {
1115 if (skb
->ip_summed
== CHECKSUM_NONE
)
1117 copy
= max
- skb
->len
;
1122 /* Allocate new segment. If the interface is SG,
1123 * allocate skb fitting to single page.
1125 if (!sk_stream_memory_free(sk
))
1126 goto wait_for_sndbuf
;
1128 skb
= sk_stream_alloc_skb(sk
,
1129 select_size(sk
, sg
),
1132 goto wait_for_memory
;
1135 * All packets are restored as if they have
1136 * already been sent.
1139 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1142 * Check whether we can use HW checksum.
1144 if (sk
->sk_route_caps
& NETIF_F_ALL_CSUM
)
1145 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1147 skb_entail(sk
, skb
);
1152 /* Try to append data to the end of skb. */
1156 /* Where to copy to? */
1157 if (skb_availroom(skb
) > 0) {
1158 /* We have some space in skb head. Superb! */
1159 copy
= min_t(int, copy
, skb_availroom(skb
));
1160 err
= skb_add_data_nocache(sk
, skb
, from
, copy
);
1165 int i
= skb_shinfo(skb
)->nr_frags
;
1166 struct page_frag
*pfrag
= sk_page_frag(sk
);
1168 if (!sk_page_frag_refill(sk
, pfrag
))
1169 goto wait_for_memory
;
1171 if (!skb_can_coalesce(skb
, i
, pfrag
->page
,
1173 if (i
== MAX_SKB_FRAGS
|| !sg
) {
1174 tcp_mark_push(tp
, skb
);
1180 copy
= min_t(int, copy
, pfrag
->size
- pfrag
->offset
);
1182 if (!sk_wmem_schedule(sk
, copy
))
1183 goto wait_for_memory
;
1185 err
= skb_copy_to_page_nocache(sk
, from
, skb
,
1192 /* Update the skb. */
1194 skb_frag_size_add(&skb_shinfo(skb
)->frags
[i
- 1], copy
);
1196 skb_fill_page_desc(skb
, i
, pfrag
->page
,
1197 pfrag
->offset
, copy
);
1198 get_page(pfrag
->page
);
1200 pfrag
->offset
+= copy
;
1204 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_PSH
;
1206 tp
->write_seq
+= copy
;
1207 TCP_SKB_CB(skb
)->end_seq
+= copy
;
1208 skb_shinfo(skb
)->gso_segs
= 0;
1212 if ((seglen
-= copy
) == 0 && iovlen
== 0)
1215 if (skb
->len
< max
|| (flags
& MSG_OOB
) || unlikely(tp
->repair
))
1218 if (forced_push(tp
)) {
1219 tcp_mark_push(tp
, skb
);
1220 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_PUSH
);
1221 } else if (skb
== tcp_send_head(sk
))
1222 tcp_push_one(sk
, mss_now
);
1226 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
1229 tcp_push(sk
, flags
& ~MSG_MORE
, mss_now
, TCP_NAGLE_PUSH
);
1231 if ((err
= sk_stream_wait_memory(sk
, &timeo
)) != 0)
1234 mss_now
= tcp_send_mss(sk
, &size_goal
, flags
);
1240 tcp_push(sk
, flags
, mss_now
, tp
->nonagle
);
1243 return copied
+ copied_syn
;
1247 tcp_unlink_write_queue(skb
, sk
);
1248 /* It is the one place in all of TCP, except connection
1249 * reset, where we can be unlinking the send_head.
1251 tcp_check_send_head(sk
, skb
);
1252 sk_wmem_free_skb(sk
, skb
);
1256 if (copied
+ copied_syn
)
1259 err
= sk_stream_error(sk
, flags
, err
);
1263 EXPORT_SYMBOL(tcp_sendmsg
);
1266 * Handle reading urgent data. BSD has very simple semantics for
1267 * this, no blocking and very strange errors 8)
1270 static int tcp_recv_urg(struct sock
*sk
, struct msghdr
*msg
, int len
, int flags
)
1272 struct tcp_sock
*tp
= tcp_sk(sk
);
1274 /* No URG data to read. */
1275 if (sock_flag(sk
, SOCK_URGINLINE
) || !tp
->urg_data
||
1276 tp
->urg_data
== TCP_URG_READ
)
1277 return -EINVAL
; /* Yes this is right ! */
1279 if (sk
->sk_state
== TCP_CLOSE
&& !sock_flag(sk
, SOCK_DONE
))
1282 if (tp
->urg_data
& TCP_URG_VALID
) {
1284 char c
= tp
->urg_data
;
1286 if (!(flags
& MSG_PEEK
))
1287 tp
->urg_data
= TCP_URG_READ
;
1289 /* Read urgent data. */
1290 msg
->msg_flags
|= MSG_OOB
;
1293 if (!(flags
& MSG_TRUNC
))
1294 err
= memcpy_toiovec(msg
->msg_iov
, &c
, 1);
1297 msg
->msg_flags
|= MSG_TRUNC
;
1299 return err
? -EFAULT
: len
;
1302 if (sk
->sk_state
== TCP_CLOSE
|| (sk
->sk_shutdown
& RCV_SHUTDOWN
))
1305 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1306 * the available implementations agree in this case:
1307 * this call should never block, independent of the
1308 * blocking state of the socket.
1309 * Mike <pall@rz.uni-karlsruhe.de>
1314 static int tcp_peek_sndq(struct sock
*sk
, struct msghdr
*msg
, int len
)
1316 struct sk_buff
*skb
;
1317 int copied
= 0, err
= 0;
1319 /* XXX -- need to support SO_PEEK_OFF */
1321 skb_queue_walk(&sk
->sk_write_queue
, skb
) {
1322 err
= skb_copy_datagram_iovec(skb
, 0, msg
->msg_iov
, skb
->len
);
1329 return err
?: copied
;
1332 /* Clean up the receive buffer for full frames taken by the user,
1333 * then send an ACK if necessary. COPIED is the number of bytes
1334 * tcp_recvmsg has given to the user so far, it speeds up the
1335 * calculation of whether or not we must ACK for the sake of
1338 void tcp_cleanup_rbuf(struct sock
*sk
, int copied
)
1340 struct tcp_sock
*tp
= tcp_sk(sk
);
1341 bool time_to_ack
= false;
1343 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
1345 WARN(skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
),
1346 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1347 tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
);
1349 if (inet_csk_ack_scheduled(sk
)) {
1350 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1351 /* Delayed ACKs frequently hit locked sockets during bulk
1353 if (icsk
->icsk_ack
.blocked
||
1354 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1355 tp
->rcv_nxt
- tp
->rcv_wup
> icsk
->icsk_ack
.rcv_mss
||
1357 * If this read emptied read buffer, we send ACK, if
1358 * connection is not bidirectional, user drained
1359 * receive buffer and there was a small segment
1363 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED2
) ||
1364 ((icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
) &&
1365 !icsk
->icsk_ack
.pingpong
)) &&
1366 !atomic_read(&sk
->sk_rmem_alloc
)))
1370 /* We send an ACK if we can now advertise a non-zero window
1371 * which has been raised "significantly".
1373 * Even if window raised up to infinity, do not send window open ACK
1374 * in states, where we will not receive more. It is useless.
1376 if (copied
> 0 && !time_to_ack
&& !(sk
->sk_shutdown
& RCV_SHUTDOWN
)) {
1377 __u32 rcv_window_now
= tcp_receive_window(tp
);
1379 /* Optimize, __tcp_select_window() is not cheap. */
1380 if (2*rcv_window_now
<= tp
->window_clamp
) {
1381 __u32 new_window
= __tcp_select_window(sk
);
1383 /* Send ACK now, if this read freed lots of space
1384 * in our buffer. Certainly, new_window is new window.
1385 * We can advertise it now, if it is not less than current one.
1386 * "Lots" means "at least twice" here.
1388 if (new_window
&& new_window
>= 2 * rcv_window_now
)
1396 static void tcp_prequeue_process(struct sock
*sk
)
1398 struct sk_buff
*skb
;
1399 struct tcp_sock
*tp
= tcp_sk(sk
);
1401 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPPREQUEUED
);
1403 /* RX process wants to run with disabled BHs, though it is not
1406 while ((skb
= __skb_dequeue(&tp
->ucopy
.prequeue
)) != NULL
)
1407 sk_backlog_rcv(sk
, skb
);
1410 /* Clear memory counter. */
1411 tp
->ucopy
.memory
= 0;
1414 #ifdef CONFIG_NET_DMA
1415 static void tcp_service_net_dma(struct sock
*sk
, bool wait
)
1417 dma_cookie_t done
, used
;
1418 dma_cookie_t last_issued
;
1419 struct tcp_sock
*tp
= tcp_sk(sk
);
1421 if (!tp
->ucopy
.dma_chan
)
1424 last_issued
= tp
->ucopy
.dma_cookie
;
1425 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1428 if (dma_async_is_tx_complete(tp
->ucopy
.dma_chan
,
1430 &used
) == DMA_SUCCESS
) {
1431 /* Safe to free early-copied skbs now */
1432 __skb_queue_purge(&sk
->sk_async_wait_queue
);
1435 struct sk_buff
*skb
;
1436 while ((skb
= skb_peek(&sk
->sk_async_wait_queue
)) &&
1437 (dma_async_is_complete(skb
->dma_cookie
, done
,
1438 used
) == DMA_SUCCESS
)) {
1439 __skb_dequeue(&sk
->sk_async_wait_queue
);
1447 static struct sk_buff
*tcp_recv_skb(struct sock
*sk
, u32 seq
, u32
*off
)
1449 struct sk_buff
*skb
;
1452 while ((skb
= skb_peek(&sk
->sk_receive_queue
)) != NULL
) {
1453 offset
= seq
- TCP_SKB_CB(skb
)->seq
;
1454 if (tcp_hdr(skb
)->syn
)
1456 if (offset
< skb
->len
|| tcp_hdr(skb
)->fin
) {
1460 /* This looks weird, but this can happen if TCP collapsing
1461 * splitted a fat GRO packet, while we released socket lock
1462 * in skb_splice_bits()
1464 sk_eat_skb(sk
, skb
, false);
1470 * This routine provides an alternative to tcp_recvmsg() for routines
1471 * that would like to handle copying from skbuffs directly in 'sendfile'
1474 * - It is assumed that the socket was locked by the caller.
1475 * - The routine does not block.
1476 * - At present, there is no support for reading OOB data
1477 * or for 'peeking' the socket using this routine
1478 * (although both would be easy to implement).
1480 int tcp_read_sock(struct sock
*sk
, read_descriptor_t
*desc
,
1481 sk_read_actor_t recv_actor
)
1483 struct sk_buff
*skb
;
1484 struct tcp_sock
*tp
= tcp_sk(sk
);
1485 u32 seq
= tp
->copied_seq
;
1489 if (sk
->sk_state
== TCP_LISTEN
)
1491 while ((skb
= tcp_recv_skb(sk
, seq
, &offset
)) != NULL
) {
1492 if (offset
< skb
->len
) {
1496 len
= skb
->len
- offset
;
1497 /* Stop reading if we hit a patch of urgent data */
1499 u32 urg_offset
= tp
->urg_seq
- seq
;
1500 if (urg_offset
< len
)
1505 used
= recv_actor(desc
, skb
, offset
, len
);
1510 } else if (used
<= len
) {
1515 /* If recv_actor drops the lock (e.g. TCP splice
1516 * receive) the skb pointer might be invalid when
1517 * getting here: tcp_collapse might have deleted it
1518 * while aggregating skbs from the socket queue.
1520 skb
= tcp_recv_skb(sk
, seq
- 1, &offset
);
1523 /* TCP coalescing might have appended data to the skb.
1524 * Try to splice more frags
1526 if (offset
+ 1 != skb
->len
)
1529 if (tcp_hdr(skb
)->fin
) {
1530 sk_eat_skb(sk
, skb
, false);
1534 sk_eat_skb(sk
, skb
, false);
1537 tp
->copied_seq
= seq
;
1539 tp
->copied_seq
= seq
;
1541 tcp_rcv_space_adjust(sk
);
1543 /* Clean up data we have read: This will do ACK frames. */
1545 tcp_recv_skb(sk
, seq
, &offset
);
1546 tcp_cleanup_rbuf(sk
, copied
);
1550 EXPORT_SYMBOL(tcp_read_sock
);
1553 * This routine copies from a sock struct into the user buffer.
1555 * Technical note: in 2.3 we work on _locked_ socket, so that
1556 * tricks with *seq access order and skb->users are not required.
1557 * Probably, code can be easily improved even more.
1560 int tcp_recvmsg(struct kiocb
*iocb
, struct sock
*sk
, struct msghdr
*msg
,
1561 size_t len
, int nonblock
, int flags
, int *addr_len
)
1563 struct tcp_sock
*tp
= tcp_sk(sk
);
1569 int target
; /* Read at least this many bytes */
1571 struct task_struct
*user_recv
= NULL
;
1572 bool copied_early
= false;
1573 struct sk_buff
*skb
;
1579 if (sk
->sk_state
== TCP_LISTEN
)
1582 timeo
= sock_rcvtimeo(sk
, nonblock
);
1584 /* Urgent data needs to be handled specially. */
1585 if (flags
& MSG_OOB
)
1588 if (unlikely(tp
->repair
)) {
1590 if (!(flags
& MSG_PEEK
))
1593 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
1597 if (tp
->repair_queue
== TCP_NO_QUEUE
)
1600 /* 'common' recv queue MSG_PEEK-ing */
1603 seq
= &tp
->copied_seq
;
1604 if (flags
& MSG_PEEK
) {
1605 peek_seq
= tp
->copied_seq
;
1609 target
= sock_rcvlowat(sk
, flags
& MSG_WAITALL
, len
);
1611 #ifdef CONFIG_NET_DMA
1612 tp
->ucopy
.dma_chan
= NULL
;
1614 skb
= skb_peek_tail(&sk
->sk_receive_queue
);
1619 available
= TCP_SKB_CB(skb
)->seq
+ skb
->len
- (*seq
);
1620 if ((available
< target
) &&
1621 (len
> sysctl_tcp_dma_copybreak
) && !(flags
& MSG_PEEK
) &&
1622 !sysctl_tcp_low_latency
&&
1623 net_dma_find_channel()) {
1624 preempt_enable_no_resched();
1625 tp
->ucopy
.pinned_list
=
1626 dma_pin_iovec_pages(msg
->msg_iov
, len
);
1628 preempt_enable_no_resched();
1636 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1637 if (tp
->urg_data
&& tp
->urg_seq
== *seq
) {
1640 if (signal_pending(current
)) {
1641 copied
= timeo
? sock_intr_errno(timeo
) : -EAGAIN
;
1646 /* Next get a buffer. */
1648 skb_queue_walk(&sk
->sk_receive_queue
, skb
) {
1649 /* Now that we have two receive queues this
1652 if (WARN(before(*seq
, TCP_SKB_CB(skb
)->seq
),
1653 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1654 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
,
1658 offset
= *seq
- TCP_SKB_CB(skb
)->seq
;
1659 if (tcp_hdr(skb
)->syn
)
1661 if (offset
< skb
->len
)
1663 if (tcp_hdr(skb
)->fin
)
1665 WARN(!(flags
& MSG_PEEK
),
1666 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1667 *seq
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
, flags
);
1670 /* Well, if we have backlog, try to process it now yet. */
1672 if (copied
>= target
&& !sk
->sk_backlog
.tail
)
1677 sk
->sk_state
== TCP_CLOSE
||
1678 (sk
->sk_shutdown
& RCV_SHUTDOWN
) ||
1680 signal_pending(current
))
1683 if (sock_flag(sk
, SOCK_DONE
))
1687 copied
= sock_error(sk
);
1691 if (sk
->sk_shutdown
& RCV_SHUTDOWN
)
1694 if (sk
->sk_state
== TCP_CLOSE
) {
1695 if (!sock_flag(sk
, SOCK_DONE
)) {
1696 /* This occurs when user tries to read
1697 * from never connected socket.
1710 if (signal_pending(current
)) {
1711 copied
= sock_intr_errno(timeo
);
1716 tcp_cleanup_rbuf(sk
, copied
);
1718 if (!sysctl_tcp_low_latency
&& tp
->ucopy
.task
== user_recv
) {
1719 /* Install new reader */
1720 if (!user_recv
&& !(flags
& (MSG_TRUNC
| MSG_PEEK
))) {
1721 user_recv
= current
;
1722 tp
->ucopy
.task
= user_recv
;
1723 tp
->ucopy
.iov
= msg
->msg_iov
;
1726 tp
->ucopy
.len
= len
;
1728 WARN_ON(tp
->copied_seq
!= tp
->rcv_nxt
&&
1729 !(flags
& (MSG_PEEK
| MSG_TRUNC
)));
1731 /* Ugly... If prequeue is not empty, we have to
1732 * process it before releasing socket, otherwise
1733 * order will be broken at second iteration.
1734 * More elegant solution is required!!!
1736 * Look: we have the following (pseudo)queues:
1738 * 1. packets in flight
1743 * Each queue can be processed only if the next ones
1744 * are empty. At this point we have empty receive_queue.
1745 * But prequeue _can_ be not empty after 2nd iteration,
1746 * when we jumped to start of loop because backlog
1747 * processing added something to receive_queue.
1748 * We cannot release_sock(), because backlog contains
1749 * packets arrived _after_ prequeued ones.
1751 * Shortly, algorithm is clear --- to process all
1752 * the queues in order. We could make it more directly,
1753 * requeueing packets from backlog to prequeue, if
1754 * is not empty. It is more elegant, but eats cycles,
1757 if (!skb_queue_empty(&tp
->ucopy
.prequeue
))
1760 /* __ Set realtime policy in scheduler __ */
1763 #ifdef CONFIG_NET_DMA
1764 if (tp
->ucopy
.dma_chan
) {
1765 if (tp
->rcv_wnd
== 0 &&
1766 !skb_queue_empty(&sk
->sk_async_wait_queue
)) {
1767 tcp_service_net_dma(sk
, true);
1768 tcp_cleanup_rbuf(sk
, copied
);
1770 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1773 if (copied
>= target
) {
1774 /* Do not sleep, just process backlog. */
1778 sk_wait_data(sk
, &timeo
);
1780 #ifdef CONFIG_NET_DMA
1781 tcp_service_net_dma(sk
, false); /* Don't block */
1782 tp
->ucopy
.wakeup
= 0;
1788 /* __ Restore normal policy in scheduler __ */
1790 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1791 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG
, chunk
);
1796 if (tp
->rcv_nxt
== tp
->copied_seq
&&
1797 !skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1799 tcp_prequeue_process(sk
);
1801 if ((chunk
= len
- tp
->ucopy
.len
) != 0) {
1802 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1808 if ((flags
& MSG_PEEK
) &&
1809 (peek_seq
- copied
- urg_hole
!= tp
->copied_seq
)) {
1810 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1812 task_pid_nr(current
));
1813 peek_seq
= tp
->copied_seq
;
1818 /* Ok so how much can we use? */
1819 used
= skb
->len
- offset
;
1823 /* Do we have urgent data here? */
1825 u32 urg_offset
= tp
->urg_seq
- *seq
;
1826 if (urg_offset
< used
) {
1828 if (!sock_flag(sk
, SOCK_URGINLINE
)) {
1841 if (!(flags
& MSG_TRUNC
)) {
1842 #ifdef CONFIG_NET_DMA
1843 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
1844 tp
->ucopy
.dma_chan
= net_dma_find_channel();
1846 if (tp
->ucopy
.dma_chan
) {
1847 tp
->ucopy
.dma_cookie
= dma_skb_copy_datagram_iovec(
1848 tp
->ucopy
.dma_chan
, skb
, offset
,
1850 tp
->ucopy
.pinned_list
);
1852 if (tp
->ucopy
.dma_cookie
< 0) {
1854 pr_alert("%s: dma_cookie < 0\n",
1857 /* Exception. Bailout! */
1863 dma_async_issue_pending(tp
->ucopy
.dma_chan
);
1865 if ((offset
+ used
) == skb
->len
)
1866 copied_early
= true;
1871 err
= skb_copy_datagram_iovec(skb
, offset
,
1872 msg
->msg_iov
, used
);
1874 /* Exception. Bailout! */
1886 tcp_rcv_space_adjust(sk
);
1889 if (tp
->urg_data
&& after(tp
->copied_seq
, tp
->urg_seq
)) {
1891 tcp_fast_path_check(sk
);
1893 if (used
+ offset
< skb
->len
)
1896 if (tcp_hdr(skb
)->fin
)
1898 if (!(flags
& MSG_PEEK
)) {
1899 sk_eat_skb(sk
, skb
, copied_early
);
1900 copied_early
= false;
1905 /* Process the FIN. */
1907 if (!(flags
& MSG_PEEK
)) {
1908 sk_eat_skb(sk
, skb
, copied_early
);
1909 copied_early
= false;
1915 if (!skb_queue_empty(&tp
->ucopy
.prequeue
)) {
1918 tp
->ucopy
.len
= copied
> 0 ? len
: 0;
1920 tcp_prequeue_process(sk
);
1922 if (copied
> 0 && (chunk
= len
- tp
->ucopy
.len
) != 0) {
1923 NET_ADD_STATS_USER(sock_net(sk
), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE
, chunk
);
1929 tp
->ucopy
.task
= NULL
;
1933 #ifdef CONFIG_NET_DMA
1934 tcp_service_net_dma(sk
, true); /* Wait for queue to drain */
1935 tp
->ucopy
.dma_chan
= NULL
;
1937 if (tp
->ucopy
.pinned_list
) {
1938 dma_unpin_iovec_pages(tp
->ucopy
.pinned_list
);
1939 tp
->ucopy
.pinned_list
= NULL
;
1943 /* According to UNIX98, msg_name/msg_namelen are ignored
1944 * on connected socket. I was just happy when found this 8) --ANK
1947 /* Clean up data we have read: This will do ACK frames. */
1948 tcp_cleanup_rbuf(sk
, copied
);
1958 err
= tcp_recv_urg(sk
, msg
, len
, flags
);
1962 err
= tcp_peek_sndq(sk
, msg
, len
);
1965 EXPORT_SYMBOL(tcp_recvmsg
);
1967 void tcp_set_state(struct sock
*sk
, int state
)
1969 int oldstate
= sk
->sk_state
;
1972 case TCP_ESTABLISHED
:
1973 if (oldstate
!= TCP_ESTABLISHED
)
1974 TCP_INC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1978 if (oldstate
== TCP_CLOSE_WAIT
|| oldstate
== TCP_ESTABLISHED
)
1979 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ESTABRESETS
);
1981 sk
->sk_prot
->unhash(sk
);
1982 if (inet_csk(sk
)->icsk_bind_hash
&&
1983 !(sk
->sk_userlocks
& SOCK_BINDPORT_LOCK
))
1987 if (oldstate
== TCP_ESTABLISHED
)
1988 TCP_DEC_STATS(sock_net(sk
), TCP_MIB_CURRESTAB
);
1991 /* Change state AFTER socket is unhashed to avoid closed
1992 * socket sitting in hash tables.
1994 sk
->sk_state
= state
;
1997 SOCK_DEBUG(sk
, "TCP sk=%p, State %s -> %s\n", sk
, statename
[oldstate
], statename
[state
]);
2000 EXPORT_SYMBOL_GPL(tcp_set_state
);
2003 * State processing on a close. This implements the state shift for
2004 * sending our FIN frame. Note that we only send a FIN for some
2005 * states. A shutdown() may have already sent the FIN, or we may be
2009 static const unsigned char new_state
[16] = {
2010 /* current state: new state: action: */
2011 /* (Invalid) */ TCP_CLOSE
,
2012 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2013 /* TCP_SYN_SENT */ TCP_CLOSE
,
2014 /* TCP_SYN_RECV */ TCP_FIN_WAIT1
| TCP_ACTION_FIN
,
2015 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1
,
2016 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2
,
2017 /* TCP_TIME_WAIT */ TCP_CLOSE
,
2018 /* TCP_CLOSE */ TCP_CLOSE
,
2019 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK
| TCP_ACTION_FIN
,
2020 /* TCP_LAST_ACK */ TCP_LAST_ACK
,
2021 /* TCP_LISTEN */ TCP_CLOSE
,
2022 /* TCP_CLOSING */ TCP_CLOSING
,
2025 static int tcp_close_state(struct sock
*sk
)
2027 int next
= (int)new_state
[sk
->sk_state
];
2028 int ns
= next
& TCP_STATE_MASK
;
2030 tcp_set_state(sk
, ns
);
2032 return next
& TCP_ACTION_FIN
;
2036 * Shutdown the sending side of a connection. Much like close except
2037 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2040 void tcp_shutdown(struct sock
*sk
, int how
)
2042 /* We need to grab some memory, and put together a FIN,
2043 * and then put it into the queue to be sent.
2044 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2046 if (!(how
& SEND_SHUTDOWN
))
2049 /* If we've already sent a FIN, or it's a closed state, skip this. */
2050 if ((1 << sk
->sk_state
) &
2051 (TCPF_ESTABLISHED
| TCPF_SYN_SENT
|
2052 TCPF_SYN_RECV
| TCPF_CLOSE_WAIT
)) {
2053 /* Clear out any half completed packets. FIN if needed. */
2054 if (tcp_close_state(sk
))
2058 EXPORT_SYMBOL(tcp_shutdown
);
2060 bool tcp_check_oom(struct sock
*sk
, int shift
)
2062 bool too_many_orphans
, out_of_socket_memory
;
2064 too_many_orphans
= tcp_too_many_orphans(sk
, shift
);
2065 out_of_socket_memory
= tcp_out_of_memory(sk
);
2067 if (too_many_orphans
)
2068 net_info_ratelimited("too many orphaned sockets\n");
2069 if (out_of_socket_memory
)
2070 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2071 return too_many_orphans
|| out_of_socket_memory
;
2074 void tcp_close(struct sock
*sk
, long timeout
)
2076 struct sk_buff
*skb
;
2077 int data_was_unread
= 0;
2081 sk
->sk_shutdown
= SHUTDOWN_MASK
;
2083 if (sk
->sk_state
== TCP_LISTEN
) {
2084 tcp_set_state(sk
, TCP_CLOSE
);
2087 inet_csk_listen_stop(sk
);
2089 goto adjudge_to_death
;
2092 /* We need to flush the recv. buffs. We do this only on the
2093 * descriptor close, not protocol-sourced closes, because the
2094 * reader process may not have drained the data yet!
2096 while ((skb
= __skb_dequeue(&sk
->sk_receive_queue
)) != NULL
) {
2097 u32 len
= TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
-
2099 data_was_unread
+= len
;
2105 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2106 if (sk
->sk_state
== TCP_CLOSE
)
2107 goto adjudge_to_death
;
2109 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2110 * data was lost. To witness the awful effects of the old behavior of
2111 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2112 * GET in an FTP client, suspend the process, wait for the client to
2113 * advertise a zero window, then kill -9 the FTP client, wheee...
2114 * Note: timeout is always zero in such a case.
2116 if (unlikely(tcp_sk(sk
)->repair
)) {
2117 sk
->sk_prot
->disconnect(sk
, 0);
2118 } else if (data_was_unread
) {
2119 /* Unread data was tossed, zap the connection. */
2120 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONCLOSE
);
2121 tcp_set_state(sk
, TCP_CLOSE
);
2122 tcp_send_active_reset(sk
, sk
->sk_allocation
);
2123 } else if (sock_flag(sk
, SOCK_LINGER
) && !sk
->sk_lingertime
) {
2124 /* Check zero linger _after_ checking for unread data. */
2125 sk
->sk_prot
->disconnect(sk
, 0);
2126 NET_INC_STATS_USER(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
2127 } else if (tcp_close_state(sk
)) {
2128 /* We FIN if the application ate all the data before
2129 * zapping the connection.
2132 /* RED-PEN. Formally speaking, we have broken TCP state
2133 * machine. State transitions:
2135 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2136 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2137 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2139 * are legal only when FIN has been sent (i.e. in window),
2140 * rather than queued out of window. Purists blame.
2142 * F.e. "RFC state" is ESTABLISHED,
2143 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2145 * The visible declinations are that sometimes
2146 * we enter time-wait state, when it is not required really
2147 * (harmless), do not send active resets, when they are
2148 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2149 * they look as CLOSING or LAST_ACK for Linux)
2150 * Probably, I missed some more holelets.
2152 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2153 * in a single packet! (May consider it later but will
2154 * probably need API support or TCP_CORK SYN-ACK until
2155 * data is written and socket is closed.)
2160 sk_stream_wait_close(sk
, timeout
);
2163 state
= sk
->sk_state
;
2167 /* It is the last release_sock in its life. It will remove backlog. */
2171 /* Now socket is owned by kernel and we acquire BH lock
2172 to finish close. No need to check for user refs.
2176 WARN_ON(sock_owned_by_user(sk
));
2178 percpu_counter_inc(sk
->sk_prot
->orphan_count
);
2180 /* Have we already been destroyed by a softirq or backlog? */
2181 if (state
!= TCP_CLOSE
&& sk
->sk_state
== TCP_CLOSE
)
2184 /* This is a (useful) BSD violating of the RFC. There is a
2185 * problem with TCP as specified in that the other end could
2186 * keep a socket open forever with no application left this end.
2187 * We use a 3 minute timeout (about the same as BSD) then kill
2188 * our end. If they send after that then tough - BUT: long enough
2189 * that we won't make the old 4*rto = almost no time - whoops
2192 * Nope, it was not mistake. It is really desired behaviour
2193 * f.e. on http servers, when such sockets are useless, but
2194 * consume significant resources. Let's do it with special
2195 * linger2 option. --ANK
2198 if (sk
->sk_state
== TCP_FIN_WAIT2
) {
2199 struct tcp_sock
*tp
= tcp_sk(sk
);
2200 if (tp
->linger2
< 0) {
2201 tcp_set_state(sk
, TCP_CLOSE
);
2202 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2203 NET_INC_STATS_BH(sock_net(sk
),
2204 LINUX_MIB_TCPABORTONLINGER
);
2206 const int tmo
= tcp_fin_time(sk
);
2208 if (tmo
> TCP_TIMEWAIT_LEN
) {
2209 inet_csk_reset_keepalive_timer(sk
,
2210 tmo
- TCP_TIMEWAIT_LEN
);
2212 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
2217 if (sk
->sk_state
!= TCP_CLOSE
) {
2219 if (tcp_check_oom(sk
, 0)) {
2220 tcp_set_state(sk
, TCP_CLOSE
);
2221 tcp_send_active_reset(sk
, GFP_ATOMIC
);
2222 NET_INC_STATS_BH(sock_net(sk
),
2223 LINUX_MIB_TCPABORTONMEMORY
);
2227 if (sk
->sk_state
== TCP_CLOSE
) {
2228 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
2229 /* We could get here with a non-NULL req if the socket is
2230 * aborted (e.g., closed with unread data) before 3WHS
2234 reqsk_fastopen_remove(sk
, req
, false);
2235 inet_csk_destroy_sock(sk
);
2237 /* Otherwise, socket is reprieved until protocol close. */
2244 EXPORT_SYMBOL(tcp_close
);
2246 /* These states need RST on ABORT according to RFC793 */
2248 static inline bool tcp_need_reset(int state
)
2250 return (1 << state
) &
2251 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
| TCPF_FIN_WAIT1
|
2252 TCPF_FIN_WAIT2
| TCPF_SYN_RECV
);
2255 int tcp_disconnect(struct sock
*sk
, int flags
)
2257 struct inet_sock
*inet
= inet_sk(sk
);
2258 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2259 struct tcp_sock
*tp
= tcp_sk(sk
);
2261 int old_state
= sk
->sk_state
;
2263 if (old_state
!= TCP_CLOSE
)
2264 tcp_set_state(sk
, TCP_CLOSE
);
2266 /* ABORT function of RFC793 */
2267 if (old_state
== TCP_LISTEN
) {
2268 inet_csk_listen_stop(sk
);
2269 } else if (unlikely(tp
->repair
)) {
2270 sk
->sk_err
= ECONNABORTED
;
2271 } else if (tcp_need_reset(old_state
) ||
2272 (tp
->snd_nxt
!= tp
->write_seq
&&
2273 (1 << old_state
) & (TCPF_CLOSING
| TCPF_LAST_ACK
))) {
2274 /* The last check adjusts for discrepancy of Linux wrt. RFC
2277 tcp_send_active_reset(sk
, gfp_any());
2278 sk
->sk_err
= ECONNRESET
;
2279 } else if (old_state
== TCP_SYN_SENT
)
2280 sk
->sk_err
= ECONNRESET
;
2282 tcp_clear_xmit_timers(sk
);
2283 __skb_queue_purge(&sk
->sk_receive_queue
);
2284 tcp_write_queue_purge(sk
);
2285 __skb_queue_purge(&tp
->out_of_order_queue
);
2286 #ifdef CONFIG_NET_DMA
2287 __skb_queue_purge(&sk
->sk_async_wait_queue
);
2290 inet
->inet_dport
= 0;
2292 if (!(sk
->sk_userlocks
& SOCK_BINDADDR_LOCK
))
2293 inet_reset_saddr(sk
);
2295 sk
->sk_shutdown
= 0;
2296 sock_reset_flag(sk
, SOCK_DONE
);
2298 if ((tp
->write_seq
+= tp
->max_window
+ 2) == 0)
2300 icsk
->icsk_backoff
= 0;
2302 icsk
->icsk_probes_out
= 0;
2303 tp
->packets_out
= 0;
2304 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
2305 tp
->snd_cwnd_cnt
= 0;
2306 tp
->window_clamp
= 0;
2307 tcp_set_ca_state(sk
, TCP_CA_Open
);
2308 tcp_clear_retrans(tp
);
2309 inet_csk_delack_init(sk
);
2310 tcp_init_send_head(sk
);
2311 memset(&tp
->rx_opt
, 0, sizeof(tp
->rx_opt
));
2314 WARN_ON(inet
->inet_num
&& !icsk
->icsk_bind_hash
);
2316 sk
->sk_error_report(sk
);
2319 EXPORT_SYMBOL(tcp_disconnect
);
2321 void tcp_sock_destruct(struct sock
*sk
)
2323 inet_sock_destruct(sk
);
2325 kfree(inet_csk(sk
)->icsk_accept_queue
.fastopenq
);
2328 static inline bool tcp_can_repair_sock(const struct sock
*sk
)
2330 return ns_capable(sock_net(sk
)->user_ns
, CAP_NET_ADMIN
) &&
2331 ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_ESTABLISHED
));
2334 static int tcp_repair_options_est(struct tcp_sock
*tp
,
2335 struct tcp_repair_opt __user
*optbuf
, unsigned int len
)
2337 struct tcp_repair_opt opt
;
2339 while (len
>= sizeof(opt
)) {
2340 if (copy_from_user(&opt
, optbuf
, sizeof(opt
)))
2346 switch (opt
.opt_code
) {
2348 tp
->rx_opt
.mss_clamp
= opt
.opt_val
;
2352 u16 snd_wscale
= opt
.opt_val
& 0xFFFF;
2353 u16 rcv_wscale
= opt
.opt_val
>> 16;
2355 if (snd_wscale
> 14 || rcv_wscale
> 14)
2358 tp
->rx_opt
.snd_wscale
= snd_wscale
;
2359 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2360 tp
->rx_opt
.wscale_ok
= 1;
2363 case TCPOPT_SACK_PERM
:
2364 if (opt
.opt_val
!= 0)
2367 tp
->rx_opt
.sack_ok
|= TCP_SACK_SEEN
;
2368 if (sysctl_tcp_fack
)
2369 tcp_enable_fack(tp
);
2371 case TCPOPT_TIMESTAMP
:
2372 if (opt
.opt_val
!= 0)
2375 tp
->rx_opt
.tstamp_ok
= 1;
2384 * Socket option code for TCP.
2386 static int do_tcp_setsockopt(struct sock
*sk
, int level
,
2387 int optname
, char __user
*optval
, unsigned int optlen
)
2389 struct tcp_sock
*tp
= tcp_sk(sk
);
2390 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2394 /* These are data/string values, all the others are ints */
2396 case TCP_CONGESTION
: {
2397 char name
[TCP_CA_NAME_MAX
];
2402 val
= strncpy_from_user(name
, optval
,
2403 min_t(long, TCP_CA_NAME_MAX
-1, optlen
));
2409 err
= tcp_set_congestion_control(sk
, name
);
2418 if (optlen
< sizeof(int))
2421 if (get_user(val
, (int __user
*)optval
))
2428 /* Values greater than interface MTU won't take effect. However
2429 * at the point when this call is done we typically don't yet
2430 * know which interface is going to be used */
2431 if (val
< TCP_MIN_MSS
|| val
> MAX_TCP_WINDOW
) {
2435 tp
->rx_opt
.user_mss
= val
;
2440 /* TCP_NODELAY is weaker than TCP_CORK, so that
2441 * this option on corked socket is remembered, but
2442 * it is not activated until cork is cleared.
2444 * However, when TCP_NODELAY is set we make
2445 * an explicit push, which overrides even TCP_CORK
2446 * for currently queued segments.
2448 tp
->nonagle
|= TCP_NAGLE_OFF
|TCP_NAGLE_PUSH
;
2449 tcp_push_pending_frames(sk
);
2451 tp
->nonagle
&= ~TCP_NAGLE_OFF
;
2455 case TCP_THIN_LINEAR_TIMEOUTS
:
2456 if (val
< 0 || val
> 1)
2462 case TCP_THIN_DUPACK
:
2463 if (val
< 0 || val
> 1)
2466 tp
->thin_dupack
= val
;
2467 if (tp
->thin_dupack
)
2468 tcp_disable_early_retrans(tp
);
2473 if (!tcp_can_repair_sock(sk
))
2475 else if (val
== 1) {
2477 sk
->sk_reuse
= SK_FORCE_REUSE
;
2478 tp
->repair_queue
= TCP_NO_QUEUE
;
2479 } else if (val
== 0) {
2481 sk
->sk_reuse
= SK_NO_REUSE
;
2482 tcp_send_window_probe(sk
);
2488 case TCP_REPAIR_QUEUE
:
2491 else if (val
< TCP_QUEUES_NR
)
2492 tp
->repair_queue
= val
;
2498 if (sk
->sk_state
!= TCP_CLOSE
)
2500 else if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2501 tp
->write_seq
= val
;
2502 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2508 case TCP_REPAIR_OPTIONS
:
2511 else if (sk
->sk_state
== TCP_ESTABLISHED
)
2512 err
= tcp_repair_options_est(tp
,
2513 (struct tcp_repair_opt __user
*)optval
,
2520 /* When set indicates to always queue non-full frames.
2521 * Later the user clears this option and we transmit
2522 * any pending partial frames in the queue. This is
2523 * meant to be used alongside sendfile() to get properly
2524 * filled frames when the user (for example) must write
2525 * out headers with a write() call first and then use
2526 * sendfile to send out the data parts.
2528 * TCP_CORK can be set together with TCP_NODELAY and it is
2529 * stronger than TCP_NODELAY.
2532 tp
->nonagle
|= TCP_NAGLE_CORK
;
2534 tp
->nonagle
&= ~TCP_NAGLE_CORK
;
2535 if (tp
->nonagle
&TCP_NAGLE_OFF
)
2536 tp
->nonagle
|= TCP_NAGLE_PUSH
;
2537 tcp_push_pending_frames(sk
);
2542 if (val
< 1 || val
> MAX_TCP_KEEPIDLE
)
2545 tp
->keepalive_time
= val
* HZ
;
2546 if (sock_flag(sk
, SOCK_KEEPOPEN
) &&
2547 !((1 << sk
->sk_state
) &
2548 (TCPF_CLOSE
| TCPF_LISTEN
))) {
2549 u32 elapsed
= keepalive_time_elapsed(tp
);
2550 if (tp
->keepalive_time
> elapsed
)
2551 elapsed
= tp
->keepalive_time
- elapsed
;
2554 inet_csk_reset_keepalive_timer(sk
, elapsed
);
2559 if (val
< 1 || val
> MAX_TCP_KEEPINTVL
)
2562 tp
->keepalive_intvl
= val
* HZ
;
2565 if (val
< 1 || val
> MAX_TCP_KEEPCNT
)
2568 tp
->keepalive_probes
= val
;
2571 if (val
< 1 || val
> MAX_TCP_SYNCNT
)
2574 icsk
->icsk_syn_retries
= val
;
2580 else if (val
> sysctl_tcp_fin_timeout
/ HZ
)
2583 tp
->linger2
= val
* HZ
;
2586 case TCP_DEFER_ACCEPT
:
2587 /* Translate value in seconds to number of retransmits */
2588 icsk
->icsk_accept_queue
.rskq_defer_accept
=
2589 secs_to_retrans(val
, TCP_TIMEOUT_INIT
/ HZ
,
2593 case TCP_WINDOW_CLAMP
:
2595 if (sk
->sk_state
!= TCP_CLOSE
) {
2599 tp
->window_clamp
= 0;
2601 tp
->window_clamp
= val
< SOCK_MIN_RCVBUF
/ 2 ?
2602 SOCK_MIN_RCVBUF
/ 2 : val
;
2607 icsk
->icsk_ack
.pingpong
= 1;
2609 icsk
->icsk_ack
.pingpong
= 0;
2610 if ((1 << sk
->sk_state
) &
2611 (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
) &&
2612 inet_csk_ack_scheduled(sk
)) {
2613 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
2614 tcp_cleanup_rbuf(sk
, 1);
2616 icsk
->icsk_ack
.pingpong
= 1;
2621 #ifdef CONFIG_TCP_MD5SIG
2623 /* Read the IP->Key mappings from userspace */
2624 err
= tp
->af_specific
->md5_parse(sk
, optval
, optlen
);
2627 case TCP_USER_TIMEOUT
:
2628 /* Cap the max timeout in ms TCP will retry/retrans
2629 * before giving up and aborting (ETIMEDOUT) a connection.
2634 icsk
->icsk_user_timeout
= msecs_to_jiffies(val
);
2638 if (val
>= 0 && ((1 << sk
->sk_state
) & (TCPF_CLOSE
|
2640 err
= fastopen_init_queue(sk
, val
);
2648 tp
->tsoffset
= val
- tcp_time_stamp
;
2659 int tcp_setsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2660 unsigned int optlen
)
2662 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2664 if (level
!= SOL_TCP
)
2665 return icsk
->icsk_af_ops
->setsockopt(sk
, level
, optname
,
2667 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2669 EXPORT_SYMBOL(tcp_setsockopt
);
2671 #ifdef CONFIG_COMPAT
2672 int compat_tcp_setsockopt(struct sock
*sk
, int level
, int optname
,
2673 char __user
*optval
, unsigned int optlen
)
2675 if (level
!= SOL_TCP
)
2676 return inet_csk_compat_setsockopt(sk
, level
, optname
,
2678 return do_tcp_setsockopt(sk
, level
, optname
, optval
, optlen
);
2680 EXPORT_SYMBOL(compat_tcp_setsockopt
);
2683 /* Return information about state of tcp endpoint in API format. */
2684 void tcp_get_info(const struct sock
*sk
, struct tcp_info
*info
)
2686 const struct tcp_sock
*tp
= tcp_sk(sk
);
2687 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2688 u32 now
= tcp_time_stamp
;
2690 memset(info
, 0, sizeof(*info
));
2692 info
->tcpi_state
= sk
->sk_state
;
2693 info
->tcpi_ca_state
= icsk
->icsk_ca_state
;
2694 info
->tcpi_retransmits
= icsk
->icsk_retransmits
;
2695 info
->tcpi_probes
= icsk
->icsk_probes_out
;
2696 info
->tcpi_backoff
= icsk
->icsk_backoff
;
2698 if (tp
->rx_opt
.tstamp_ok
)
2699 info
->tcpi_options
|= TCPI_OPT_TIMESTAMPS
;
2700 if (tcp_is_sack(tp
))
2701 info
->tcpi_options
|= TCPI_OPT_SACK
;
2702 if (tp
->rx_opt
.wscale_ok
) {
2703 info
->tcpi_options
|= TCPI_OPT_WSCALE
;
2704 info
->tcpi_snd_wscale
= tp
->rx_opt
.snd_wscale
;
2705 info
->tcpi_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
2708 if (tp
->ecn_flags
& TCP_ECN_OK
)
2709 info
->tcpi_options
|= TCPI_OPT_ECN
;
2710 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
2711 info
->tcpi_options
|= TCPI_OPT_ECN_SEEN
;
2712 if (tp
->syn_data_acked
)
2713 info
->tcpi_options
|= TCPI_OPT_SYN_DATA
;
2715 info
->tcpi_rto
= jiffies_to_usecs(icsk
->icsk_rto
);
2716 info
->tcpi_ato
= jiffies_to_usecs(icsk
->icsk_ack
.ato
);
2717 info
->tcpi_snd_mss
= tp
->mss_cache
;
2718 info
->tcpi_rcv_mss
= icsk
->icsk_ack
.rcv_mss
;
2720 if (sk
->sk_state
== TCP_LISTEN
) {
2721 info
->tcpi_unacked
= sk
->sk_ack_backlog
;
2722 info
->tcpi_sacked
= sk
->sk_max_ack_backlog
;
2724 info
->tcpi_unacked
= tp
->packets_out
;
2725 info
->tcpi_sacked
= tp
->sacked_out
;
2727 info
->tcpi_lost
= tp
->lost_out
;
2728 info
->tcpi_retrans
= tp
->retrans_out
;
2729 info
->tcpi_fackets
= tp
->fackets_out
;
2731 info
->tcpi_last_data_sent
= jiffies_to_msecs(now
- tp
->lsndtime
);
2732 info
->tcpi_last_data_recv
= jiffies_to_msecs(now
- icsk
->icsk_ack
.lrcvtime
);
2733 info
->tcpi_last_ack_recv
= jiffies_to_msecs(now
- tp
->rcv_tstamp
);
2735 info
->tcpi_pmtu
= icsk
->icsk_pmtu_cookie
;
2736 info
->tcpi_rcv_ssthresh
= tp
->rcv_ssthresh
;
2737 info
->tcpi_rtt
= jiffies_to_usecs(tp
->srtt
)>>3;
2738 info
->tcpi_rttvar
= jiffies_to_usecs(tp
->mdev
)>>2;
2739 info
->tcpi_snd_ssthresh
= tp
->snd_ssthresh
;
2740 info
->tcpi_snd_cwnd
= tp
->snd_cwnd
;
2741 info
->tcpi_advmss
= tp
->advmss
;
2742 info
->tcpi_reordering
= tp
->reordering
;
2744 info
->tcpi_rcv_rtt
= jiffies_to_usecs(tp
->rcv_rtt_est
.rtt
)>>3;
2745 info
->tcpi_rcv_space
= tp
->rcvq_space
.space
;
2747 info
->tcpi_total_retrans
= tp
->total_retrans
;
2749 EXPORT_SYMBOL_GPL(tcp_get_info
);
2751 static int do_tcp_getsockopt(struct sock
*sk
, int level
,
2752 int optname
, char __user
*optval
, int __user
*optlen
)
2754 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2755 struct tcp_sock
*tp
= tcp_sk(sk
);
2758 if (get_user(len
, optlen
))
2761 len
= min_t(unsigned int, len
, sizeof(int));
2768 val
= tp
->mss_cache
;
2769 if (!val
&& ((1 << sk
->sk_state
) & (TCPF_CLOSE
| TCPF_LISTEN
)))
2770 val
= tp
->rx_opt
.user_mss
;
2772 val
= tp
->rx_opt
.mss_clamp
;
2775 val
= !!(tp
->nonagle
&TCP_NAGLE_OFF
);
2778 val
= !!(tp
->nonagle
&TCP_NAGLE_CORK
);
2781 val
= keepalive_time_when(tp
) / HZ
;
2784 val
= keepalive_intvl_when(tp
) / HZ
;
2787 val
= keepalive_probes(tp
);
2790 val
= icsk
->icsk_syn_retries
? : sysctl_tcp_syn_retries
;
2795 val
= (val
? : sysctl_tcp_fin_timeout
) / HZ
;
2797 case TCP_DEFER_ACCEPT
:
2798 val
= retrans_to_secs(icsk
->icsk_accept_queue
.rskq_defer_accept
,
2799 TCP_TIMEOUT_INIT
/ HZ
, TCP_RTO_MAX
/ HZ
);
2801 case TCP_WINDOW_CLAMP
:
2802 val
= tp
->window_clamp
;
2805 struct tcp_info info
;
2807 if (get_user(len
, optlen
))
2810 tcp_get_info(sk
, &info
);
2812 len
= min_t(unsigned int, len
, sizeof(info
));
2813 if (put_user(len
, optlen
))
2815 if (copy_to_user(optval
, &info
, len
))
2820 val
= !icsk
->icsk_ack
.pingpong
;
2823 case TCP_CONGESTION
:
2824 if (get_user(len
, optlen
))
2826 len
= min_t(unsigned int, len
, TCP_CA_NAME_MAX
);
2827 if (put_user(len
, optlen
))
2829 if (copy_to_user(optval
, icsk
->icsk_ca_ops
->name
, len
))
2833 case TCP_THIN_LINEAR_TIMEOUTS
:
2836 case TCP_THIN_DUPACK
:
2837 val
= tp
->thin_dupack
;
2844 case TCP_REPAIR_QUEUE
:
2846 val
= tp
->repair_queue
;
2852 if (tp
->repair_queue
== TCP_SEND_QUEUE
)
2853 val
= tp
->write_seq
;
2854 else if (tp
->repair_queue
== TCP_RECV_QUEUE
)
2860 case TCP_USER_TIMEOUT
:
2861 val
= jiffies_to_msecs(icsk
->icsk_user_timeout
);
2864 val
= tcp_time_stamp
+ tp
->tsoffset
;
2867 return -ENOPROTOOPT
;
2870 if (put_user(len
, optlen
))
2872 if (copy_to_user(optval
, &val
, len
))
2877 int tcp_getsockopt(struct sock
*sk
, int level
, int optname
, char __user
*optval
,
2880 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2882 if (level
!= SOL_TCP
)
2883 return icsk
->icsk_af_ops
->getsockopt(sk
, level
, optname
,
2885 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2887 EXPORT_SYMBOL(tcp_getsockopt
);
2889 #ifdef CONFIG_COMPAT
2890 int compat_tcp_getsockopt(struct sock
*sk
, int level
, int optname
,
2891 char __user
*optval
, int __user
*optlen
)
2893 if (level
!= SOL_TCP
)
2894 return inet_csk_compat_getsockopt(sk
, level
, optname
,
2896 return do_tcp_getsockopt(sk
, level
, optname
, optval
, optlen
);
2898 EXPORT_SYMBOL(compat_tcp_getsockopt
);
2901 struct sk_buff
*tcp_tso_segment(struct sk_buff
*skb
,
2902 netdev_features_t features
)
2904 struct sk_buff
*segs
= ERR_PTR(-EINVAL
);
2905 unsigned int sum_truesize
= 0;
2910 unsigned int oldlen
;
2912 struct sk_buff
*gso_skb
= skb
;
2914 bool ooo_okay
, copy_destructor
;
2916 if (!pskb_may_pull(skb
, sizeof(*th
)))
2920 thlen
= th
->doff
* 4;
2921 if (thlen
< sizeof(*th
))
2924 if (!pskb_may_pull(skb
, thlen
))
2927 oldlen
= (u16
)~skb
->len
;
2928 __skb_pull(skb
, thlen
);
2930 mss
= skb_shinfo(skb
)->gso_size
;
2931 if (unlikely(skb
->len
<= mss
))
2934 if (skb_gso_ok(skb
, features
| NETIF_F_GSO_ROBUST
)) {
2935 /* Packet is from an untrusted source, reset gso_segs. */
2936 int type
= skb_shinfo(skb
)->gso_type
;
2944 SKB_GSO_UDP_TUNNEL
|
2946 !(type
& (SKB_GSO_TCPV4
| SKB_GSO_TCPV6
))))
2949 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss
);
2955 copy_destructor
= gso_skb
->destructor
== tcp_wfree
;
2956 ooo_okay
= gso_skb
->ooo_okay
;
2957 /* All segments but the first should have ooo_okay cleared */
2960 segs
= skb_segment(skb
, features
);
2964 /* Only first segment might have ooo_okay set */
2965 segs
->ooo_okay
= ooo_okay
;
2967 delta
= htonl(oldlen
+ (thlen
+ mss
));
2971 seq
= ntohl(th
->seq
);
2973 newcheck
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
2974 (__force u32
)delta
));
2977 th
->fin
= th
->psh
= 0;
2978 th
->check
= newcheck
;
2980 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2982 csum_fold(csum_partial(skb_transport_header(skb
),
2986 if (copy_destructor
) {
2987 skb
->destructor
= gso_skb
->destructor
;
2988 skb
->sk
= gso_skb
->sk
;
2989 sum_truesize
+= skb
->truesize
;
2994 th
->seq
= htonl(seq
);
2996 } while (skb
->next
);
2998 /* Following permits TCP Small Queues to work well with GSO :
2999 * The callback to TCP stack will be called at the time last frag
3000 * is freed at TX completion, and not right now when gso_skb
3001 * is freed by GSO engine
3003 if (copy_destructor
) {
3004 swap(gso_skb
->sk
, skb
->sk
);
3005 swap(gso_skb
->destructor
, skb
->destructor
);
3006 sum_truesize
+= skb
->truesize
;
3007 atomic_add(sum_truesize
- gso_skb
->truesize
,
3008 &skb
->sk
->sk_wmem_alloc
);
3011 delta
= htonl(oldlen
+ (skb
->tail
- skb
->transport_header
) +
3013 th
->check
= ~csum_fold((__force __wsum
)((__force u32
)th
->check
+
3014 (__force u32
)delta
));
3015 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
3016 th
->check
= csum_fold(csum_partial(skb_transport_header(skb
),
3022 EXPORT_SYMBOL(tcp_tso_segment
);
3024 struct sk_buff
**tcp_gro_receive(struct sk_buff
**head
, struct sk_buff
*skb
)
3026 struct sk_buff
**pp
= NULL
;
3033 unsigned int mss
= 1;
3039 off
= skb_gro_offset(skb
);
3040 hlen
= off
+ sizeof(*th
);
3041 th
= skb_gro_header_fast(skb
, off
);
3042 if (skb_gro_header_hard(skb
, hlen
)) {
3043 th
= skb_gro_header_slow(skb
, hlen
, off
);
3048 thlen
= th
->doff
* 4;
3049 if (thlen
< sizeof(*th
))
3053 if (skb_gro_header_hard(skb
, hlen
)) {
3054 th
= skb_gro_header_slow(skb
, hlen
, off
);
3059 skb_gro_pull(skb
, thlen
);
3061 len
= skb_gro_len(skb
);
3062 flags
= tcp_flag_word(th
);
3064 for (; (p
= *head
); head
= &p
->next
) {
3065 if (!NAPI_GRO_CB(p
)->same_flow
)
3070 if (*(u32
*)&th
->source
^ *(u32
*)&th2
->source
) {
3071 NAPI_GRO_CB(p
)->same_flow
= 0;
3078 goto out_check_final
;
3081 flush
= NAPI_GRO_CB(p
)->flush
;
3082 flush
|= (__force
int)(flags
& TCP_FLAG_CWR
);
3083 flush
|= (__force
int)((flags
^ tcp_flag_word(th2
)) &
3084 ~(TCP_FLAG_CWR
| TCP_FLAG_FIN
| TCP_FLAG_PSH
));
3085 flush
|= (__force
int)(th
->ack_seq
^ th2
->ack_seq
);
3086 for (i
= sizeof(*th
); i
< thlen
; i
+= 4)
3087 flush
|= *(u32
*)((u8
*)th
+ i
) ^
3088 *(u32
*)((u8
*)th2
+ i
);
3090 mss
= skb_shinfo(p
)->gso_size
;
3092 flush
|= (len
- 1) >= mss
;
3093 flush
|= (ntohl(th2
->seq
) + skb_gro_len(p
)) ^ ntohl(th
->seq
);
3095 if (flush
|| skb_gro_receive(head
, skb
)) {
3097 goto out_check_final
;
3102 tcp_flag_word(th2
) |= flags
& (TCP_FLAG_FIN
| TCP_FLAG_PSH
);
3106 flush
|= (__force
int)(flags
& (TCP_FLAG_URG
| TCP_FLAG_PSH
|
3107 TCP_FLAG_RST
| TCP_FLAG_SYN
|
3110 if (p
&& (!NAPI_GRO_CB(skb
)->same_flow
|| flush
))
3114 NAPI_GRO_CB(skb
)->flush
|= flush
;
3118 EXPORT_SYMBOL(tcp_gro_receive
);
3120 int tcp_gro_complete(struct sk_buff
*skb
)
3122 struct tcphdr
*th
= tcp_hdr(skb
);
3124 skb
->csum_start
= skb_transport_header(skb
) - skb
->head
;
3125 skb
->csum_offset
= offsetof(struct tcphdr
, check
);
3126 skb
->ip_summed
= CHECKSUM_PARTIAL
;
3128 skb_shinfo(skb
)->gso_segs
= NAPI_GRO_CB(skb
)->count
;
3131 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
3135 EXPORT_SYMBOL(tcp_gro_complete
);
3137 #ifdef CONFIG_TCP_MD5SIG
3138 static unsigned long tcp_md5sig_users
;
3139 static struct tcp_md5sig_pool __percpu
*tcp_md5sig_pool
;
3140 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock
);
3142 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu
*pool
)
3146 for_each_possible_cpu(cpu
) {
3147 struct tcp_md5sig_pool
*p
= per_cpu_ptr(pool
, cpu
);
3149 if (p
->md5_desc
.tfm
)
3150 crypto_free_hash(p
->md5_desc
.tfm
);
3155 void tcp_free_md5sig_pool(void)
3157 struct tcp_md5sig_pool __percpu
*pool
= NULL
;
3159 spin_lock_bh(&tcp_md5sig_pool_lock
);
3160 if (--tcp_md5sig_users
== 0) {
3161 pool
= tcp_md5sig_pool
;
3162 tcp_md5sig_pool
= NULL
;
3164 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3166 __tcp_free_md5sig_pool(pool
);
3168 EXPORT_SYMBOL(tcp_free_md5sig_pool
);
3170 static struct tcp_md5sig_pool __percpu
*
3171 __tcp_alloc_md5sig_pool(struct sock
*sk
)
3174 struct tcp_md5sig_pool __percpu
*pool
;
3176 pool
= alloc_percpu(struct tcp_md5sig_pool
);
3180 for_each_possible_cpu(cpu
) {
3181 struct crypto_hash
*hash
;
3183 hash
= crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC
);
3184 if (IS_ERR_OR_NULL(hash
))
3187 per_cpu_ptr(pool
, cpu
)->md5_desc
.tfm
= hash
;
3191 __tcp_free_md5sig_pool(pool
);
3195 struct tcp_md5sig_pool __percpu
*tcp_alloc_md5sig_pool(struct sock
*sk
)
3197 struct tcp_md5sig_pool __percpu
*pool
;
3201 spin_lock_bh(&tcp_md5sig_pool_lock
);
3202 pool
= tcp_md5sig_pool
;
3203 if (tcp_md5sig_users
++ == 0) {
3205 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3208 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3212 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3215 /* we cannot hold spinlock here because this may sleep. */
3216 struct tcp_md5sig_pool __percpu
*p
;
3218 p
= __tcp_alloc_md5sig_pool(sk
);
3219 spin_lock_bh(&tcp_md5sig_pool_lock
);
3222 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3225 pool
= tcp_md5sig_pool
;
3227 /* oops, it has already been assigned. */
3228 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3229 __tcp_free_md5sig_pool(p
);
3231 tcp_md5sig_pool
= pool
= p
;
3232 spin_unlock_bh(&tcp_md5sig_pool_lock
);
3237 EXPORT_SYMBOL(tcp_alloc_md5sig_pool
);
3241 * tcp_get_md5sig_pool - get md5sig_pool for this user
3243 * We use percpu structure, so if we succeed, we exit with preemption
3244 * and BH disabled, to make sure another thread or softirq handling
3245 * wont try to get same context.
3247 struct tcp_md5sig_pool
*tcp_get_md5sig_pool(void)
3249 struct tcp_md5sig_pool __percpu
*p
;
3253 spin_lock(&tcp_md5sig_pool_lock
);
3254 p
= tcp_md5sig_pool
;
3257 spin_unlock(&tcp_md5sig_pool_lock
);
3260 return this_cpu_ptr(p
);
3265 EXPORT_SYMBOL(tcp_get_md5sig_pool
);
3267 void tcp_put_md5sig_pool(void)
3270 tcp_free_md5sig_pool();
3272 EXPORT_SYMBOL(tcp_put_md5sig_pool
);
3274 int tcp_md5_hash_header(struct tcp_md5sig_pool
*hp
,
3275 const struct tcphdr
*th
)
3277 struct scatterlist sg
;
3281 /* We are not allowed to change tcphdr, make a local copy */
3282 memcpy(&hdr
, th
, sizeof(hdr
));
3285 /* options aren't included in the hash */
3286 sg_init_one(&sg
, &hdr
, sizeof(hdr
));
3287 err
= crypto_hash_update(&hp
->md5_desc
, &sg
, sizeof(hdr
));
3290 EXPORT_SYMBOL(tcp_md5_hash_header
);
3292 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool
*hp
,
3293 const struct sk_buff
*skb
, unsigned int header_len
)
3295 struct scatterlist sg
;
3296 const struct tcphdr
*tp
= tcp_hdr(skb
);
3297 struct hash_desc
*desc
= &hp
->md5_desc
;
3299 const unsigned int head_data_len
= skb_headlen(skb
) > header_len
?
3300 skb_headlen(skb
) - header_len
: 0;
3301 const struct skb_shared_info
*shi
= skb_shinfo(skb
);
3302 struct sk_buff
*frag_iter
;
3304 sg_init_table(&sg
, 1);
3306 sg_set_buf(&sg
, ((u8
*) tp
) + header_len
, head_data_len
);
3307 if (crypto_hash_update(desc
, &sg
, head_data_len
))
3310 for (i
= 0; i
< shi
->nr_frags
; ++i
) {
3311 const struct skb_frag_struct
*f
= &shi
->frags
[i
];
3312 unsigned int offset
= f
->page_offset
;
3313 struct page
*page
= skb_frag_page(f
) + (offset
>> PAGE_SHIFT
);
3315 sg_set_page(&sg
, page
, skb_frag_size(f
),
3316 offset_in_page(offset
));
3317 if (crypto_hash_update(desc
, &sg
, skb_frag_size(f
)))
3321 skb_walk_frags(skb
, frag_iter
)
3322 if (tcp_md5_hash_skb_data(hp
, frag_iter
, 0))
3327 EXPORT_SYMBOL(tcp_md5_hash_skb_data
);
3329 int tcp_md5_hash_key(struct tcp_md5sig_pool
*hp
, const struct tcp_md5sig_key
*key
)
3331 struct scatterlist sg
;
3333 sg_init_one(&sg
, key
->key
, key
->keylen
);
3334 return crypto_hash_update(&hp
->md5_desc
, &sg
, key
->keylen
);
3336 EXPORT_SYMBOL(tcp_md5_hash_key
);
3340 void tcp_done(struct sock
*sk
)
3342 struct request_sock
*req
= tcp_sk(sk
)->fastopen_rsk
;
3344 if (sk
->sk_state
== TCP_SYN_SENT
|| sk
->sk_state
== TCP_SYN_RECV
)
3345 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_ATTEMPTFAILS
);
3347 tcp_set_state(sk
, TCP_CLOSE
);
3348 tcp_clear_xmit_timers(sk
);
3350 reqsk_fastopen_remove(sk
, req
, false);
3352 sk
->sk_shutdown
= SHUTDOWN_MASK
;
3354 if (!sock_flag(sk
, SOCK_DEAD
))
3355 sk
->sk_state_change(sk
);
3357 inet_csk_destroy_sock(sk
);
3359 EXPORT_SYMBOL_GPL(tcp_done
);
3361 extern struct tcp_congestion_ops tcp_reno
;
3363 static __initdata
unsigned long thash_entries
;
3364 static int __init
set_thash_entries(char *str
)
3371 ret
= kstrtoul(str
, 0, &thash_entries
);
3377 __setup("thash_entries=", set_thash_entries
);
3379 void tcp_init_mem(struct net
*net
)
3381 unsigned long limit
= nr_free_buffer_pages() / 8;
3382 limit
= max(limit
, 128UL);
3383 net
->ipv4
.sysctl_tcp_mem
[0] = limit
/ 4 * 3;
3384 net
->ipv4
.sysctl_tcp_mem
[1] = limit
;
3385 net
->ipv4
.sysctl_tcp_mem
[2] = net
->ipv4
.sysctl_tcp_mem
[0] * 2;
3388 void __init
tcp_init(void)
3390 struct sk_buff
*skb
= NULL
;
3391 unsigned long limit
;
3392 int max_rshare
, max_wshare
, cnt
;
3395 BUILD_BUG_ON(sizeof(struct tcp_skb_cb
) > sizeof(skb
->cb
));
3397 percpu_counter_init(&tcp_sockets_allocated
, 0);
3398 percpu_counter_init(&tcp_orphan_count
, 0);
3399 tcp_hashinfo
.bind_bucket_cachep
=
3400 kmem_cache_create("tcp_bind_bucket",
3401 sizeof(struct inet_bind_bucket
), 0,
3402 SLAB_HWCACHE_ALIGN
|SLAB_PANIC
, NULL
);
3404 /* Size and allocate the main established and bind bucket
3407 * The methodology is similar to that of the buffer cache.
3409 tcp_hashinfo
.ehash
=
3410 alloc_large_system_hash("TCP established",
3411 sizeof(struct inet_ehash_bucket
),
3413 17, /* one slot per 128 KB of memory */
3416 &tcp_hashinfo
.ehash_mask
,
3418 thash_entries
? 0 : 512 * 1024);
3419 for (i
= 0; i
<= tcp_hashinfo
.ehash_mask
; i
++) {
3420 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].chain
, i
);
3421 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo
.ehash
[i
].twchain
, i
);
3423 if (inet_ehash_locks_alloc(&tcp_hashinfo
))
3424 panic("TCP: failed to alloc ehash_locks");
3425 tcp_hashinfo
.bhash
=
3426 alloc_large_system_hash("TCP bind",
3427 sizeof(struct inet_bind_hashbucket
),
3428 tcp_hashinfo
.ehash_mask
+ 1,
3429 17, /* one slot per 128 KB of memory */
3431 &tcp_hashinfo
.bhash_size
,
3435 tcp_hashinfo
.bhash_size
= 1U << tcp_hashinfo
.bhash_size
;
3436 for (i
= 0; i
< tcp_hashinfo
.bhash_size
; i
++) {
3437 spin_lock_init(&tcp_hashinfo
.bhash
[i
].lock
);
3438 INIT_HLIST_HEAD(&tcp_hashinfo
.bhash
[i
].chain
);
3442 cnt
= tcp_hashinfo
.ehash_mask
+ 1;
3444 tcp_death_row
.sysctl_max_tw_buckets
= cnt
/ 2;
3445 sysctl_tcp_max_orphans
= cnt
/ 2;
3446 sysctl_max_syn_backlog
= max(128, cnt
/ 256);
3448 tcp_init_mem(&init_net
);
3449 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3450 limit
= nr_free_buffer_pages() << (PAGE_SHIFT
- 7);
3451 max_wshare
= min(4UL*1024*1024, limit
);
3452 max_rshare
= min(6UL*1024*1024, limit
);
3454 sysctl_tcp_wmem
[0] = SK_MEM_QUANTUM
;
3455 sysctl_tcp_wmem
[1] = 16*1024;
3456 sysctl_tcp_wmem
[2] = max(64*1024, max_wshare
);
3458 sysctl_tcp_rmem
[0] = SK_MEM_QUANTUM
;
3459 sysctl_tcp_rmem
[1] = 87380;
3460 sysctl_tcp_rmem
[2] = max(87380, max_rshare
);
3462 pr_info("Hash tables configured (established %u bind %u)\n",
3463 tcp_hashinfo
.ehash_mask
+ 1, tcp_hashinfo
.bhash_size
);
3467 tcp_register_congestion_control(&tcp_reno
);