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/inet_diag.h>
256 #include <linux/init.h>
257 #include <linux/fs.h>
258 #include <linux/skbuff.h>
259 #include <linux/scatterlist.h>
260 #include <linux/splice.h>
261 #include <linux/net.h>
262 #include <linux/socket.h>
263 #include <linux/random.h>
264 #include <linux/bootmem.h>
265 #include <linux/highmem.h>
266 #include <linux/swap.h>
267 #include <linux/cache.h>
268 #include <linux/err.h>
269 #include <linux/crypto.h>
270 #include <linux/time.h>
271 #include <linux/slab.h>
272 #include <linux/uid_stat.h>
274 #include <net/icmp.h>
275 #include <net/inet_common.h>
277 #include <net/xfrm.h>
279 #include <net/ip6_route.h>
280 #include <net/ipv6.h>
281 #include <net/transp_v6.h>
282 #include <net/sock.h>
284 #include <asm/uaccess.h>
285 #include <asm/ioctls.h>
286 #include <net/busy_poll.h>
288 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
290 int sysctl_tcp_min_tso_segs __read_mostly = 2;
292 int sysctl_tcp_autocorking __read_mostly = 1;
294 struct percpu_counter tcp_orphan_count;
295 EXPORT_SYMBOL_GPL(tcp_orphan_count);
297 long sysctl_tcp_mem[3] __read_mostly;
298 int sysctl_tcp_wmem[3] __read_mostly;
299 int sysctl_tcp_rmem[3] __read_mostly;
301 EXPORT_SYMBOL(sysctl_tcp_mem);
302 EXPORT_SYMBOL(sysctl_tcp_rmem);
303 EXPORT_SYMBOL(sysctl_tcp_wmem);
305 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
306 EXPORT_SYMBOL(tcp_memory_allocated);
309 * Current number of TCP sockets.
311 struct percpu_counter tcp_sockets_allocated;
312 EXPORT_SYMBOL(tcp_sockets_allocated);
317 struct tcp_splice_state {
318 struct pipe_inode_info *pipe;
324 * Pressure flag: try to collapse.
325 * Technical note: it is used by multiple contexts non atomically.
326 * All the __sk_mem_schedule() is of this nature: accounting
327 * is strict, actions are advisory and have some latency.
329 int tcp_memory_pressure __read_mostly;
330 EXPORT_SYMBOL(tcp_memory_pressure);
332 void tcp_enter_memory_pressure(struct sock *sk)
334 if (!tcp_memory_pressure) {
335 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
336 tcp_memory_pressure = 1;
339 EXPORT_SYMBOL(tcp_enter_memory_pressure);
341 /* Convert seconds to retransmits based on initial and max timeout */
342 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
347 int period = timeout;
350 while (seconds > period && res < 255) {
353 if (timeout > rto_max)
361 /* Convert retransmits to seconds based on initial and max timeout */
362 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
370 if (timeout > rto_max)
378 /* Address-family independent initialization for a tcp_sock.
380 * NOTE: A lot of things set to zero explicitly by call to
381 * sk_alloc() so need not be done here.
383 void tcp_init_sock(struct sock *sk)
385 struct inet_connection_sock *icsk = inet_csk(sk);
386 struct tcp_sock *tp = tcp_sk(sk);
388 __skb_queue_head_init(&tp->out_of_order_queue);
389 tcp_init_xmit_timers(sk);
390 tcp_prequeue_init(tp);
391 INIT_LIST_HEAD(&tp->tsq_node);
393 icsk->icsk_rto = TCP_TIMEOUT_INIT;
394 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
395 tp->rtt_min[0].rtt = ~0U;
397 /* So many TCP implementations out there (incorrectly) count the
398 * initial SYN frame in their delayed-ACK and congestion control
399 * algorithms that we must have the following bandaid to talk
400 * efficiently to them. -DaveM
402 tp->snd_cwnd = TCP_INIT_CWND;
404 /* See draft-stevens-tcpca-spec-01 for discussion of the
405 * initialization of these values.
407 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
408 tp->snd_cwnd_clamp = ~0;
409 tp->mss_cache = TCP_MSS_DEFAULT;
410 u64_stats_init(&tp->syncp);
412 tp->reordering = sysctl_tcp_reordering;
413 tcp_enable_early_retrans(tp);
414 tcp_assign_congestion_control(sk);
418 sk->sk_state = TCP_CLOSE;
420 sk->sk_write_space = sk_stream_write_space;
421 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
423 icsk->icsk_sync_mss = tcp_sync_mss;
425 sk->sk_sndbuf = sysctl_tcp_wmem[1];
426 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
429 sock_update_memcg(sk);
430 sk_sockets_allocated_inc(sk);
433 EXPORT_SYMBOL(tcp_init_sock);
435 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb)
437 if (sk->sk_tsflags) {
438 struct skb_shared_info *shinfo = skb_shinfo(skb);
440 sock_tx_timestamp(sk, &shinfo->tx_flags);
441 if (shinfo->tx_flags & SKBTX_ANY_TSTAMP)
442 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
447 * Wait for a TCP event.
449 * Note that we don't need to lock the socket, as the upper poll layers
450 * take care of normal races (between the test and the event) and we don't
451 * go look at any of the socket buffers directly.
453 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
456 struct sock *sk = sock->sk;
457 const struct tcp_sock *tp = tcp_sk(sk);
460 sock_rps_record_flow(sk);
462 sock_poll_wait(file, sk_sleep(sk), wait);
464 state = sk_state_load(sk);
465 if (state == TCP_LISTEN)
466 return inet_csk_listen_poll(sk);
468 /* Socket is not locked. We are protected from async events
469 * by poll logic and correct handling of state changes
470 * made by other threads is impossible in any case.
476 * POLLHUP is certainly not done right. But poll() doesn't
477 * have a notion of HUP in just one direction, and for a
478 * socket the read side is more interesting.
480 * Some poll() documentation says that POLLHUP is incompatible
481 * with the POLLOUT/POLLWR flags, so somebody should check this
482 * all. But careful, it tends to be safer to return too many
483 * bits than too few, and you can easily break real applications
484 * if you don't tell them that something has hung up!
488 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
489 * our fs/select.c). It means that after we received EOF,
490 * poll always returns immediately, making impossible poll() on write()
491 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
492 * if and only if shutdown has been made in both directions.
493 * Actually, it is interesting to look how Solaris and DUX
494 * solve this dilemma. I would prefer, if POLLHUP were maskable,
495 * then we could set it on SND_SHUTDOWN. BTW examples given
496 * in Stevens' books assume exactly this behaviour, it explains
497 * why POLLHUP is incompatible with POLLOUT. --ANK
499 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
500 * blocking on fresh not-connected or disconnected socket. --ANK
502 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
504 if (sk->sk_shutdown & RCV_SHUTDOWN)
505 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
507 /* Connected or passive Fast Open socket? */
508 if (state != TCP_SYN_SENT &&
509 (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
510 int target = sock_rcvlowat(sk, 0, INT_MAX);
512 if (tp->urg_seq == tp->copied_seq &&
513 !sock_flag(sk, SOCK_URGINLINE) &&
517 if (tp->rcv_nxt - tp->copied_seq >= target)
518 mask |= POLLIN | POLLRDNORM;
520 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
521 if (sk_stream_is_writeable(sk)) {
522 mask |= POLLOUT | POLLWRNORM;
523 } else { /* send SIGIO later */
524 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
525 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
527 /* Race breaker. If space is freed after
528 * wspace test but before the flags are set,
529 * IO signal will be lost. Memory barrier
530 * pairs with the input side.
532 smp_mb__after_atomic();
533 if (sk_stream_is_writeable(sk))
534 mask |= POLLOUT | POLLWRNORM;
537 mask |= POLLOUT | POLLWRNORM;
539 if (tp->urg_data & TCP_URG_VALID)
542 /* This barrier is coupled with smp_wmb() in tcp_reset() */
544 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
549 EXPORT_SYMBOL(tcp_poll);
551 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
553 struct tcp_sock *tp = tcp_sk(sk);
559 if (sk->sk_state == TCP_LISTEN)
562 slow = lock_sock_fast(sk);
563 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
565 else if (sock_flag(sk, SOCK_URGINLINE) ||
567 before(tp->urg_seq, tp->copied_seq) ||
568 !before(tp->urg_seq, tp->rcv_nxt)) {
570 answ = tp->rcv_nxt - tp->copied_seq;
572 /* Subtract 1, if FIN was received */
573 if (answ && sock_flag(sk, SOCK_DONE))
576 answ = tp->urg_seq - tp->copied_seq;
577 unlock_sock_fast(sk, slow);
580 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
583 if (sk->sk_state == TCP_LISTEN)
586 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
589 answ = tp->write_seq - tp->snd_una;
592 if (sk->sk_state == TCP_LISTEN)
595 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
598 answ = tp->write_seq - tp->snd_nxt;
604 return put_user(answ, (int __user *)arg);
606 EXPORT_SYMBOL(tcp_ioctl);
608 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
610 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
611 tp->pushed_seq = tp->write_seq;
614 static inline bool forced_push(const struct tcp_sock *tp)
616 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
619 static void skb_entail(struct sock *sk, struct sk_buff *skb)
621 struct tcp_sock *tp = tcp_sk(sk);
622 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
625 tcb->seq = tcb->end_seq = tp->write_seq;
626 tcb->tcp_flags = TCPHDR_ACK;
628 __skb_header_release(skb);
629 tcp_add_write_queue_tail(sk, skb);
630 sk->sk_wmem_queued += skb->truesize;
631 sk_mem_charge(sk, skb->truesize);
632 if (tp->nonagle & TCP_NAGLE_PUSH)
633 tp->nonagle &= ~TCP_NAGLE_PUSH;
635 tcp_slow_start_after_idle_check(sk);
638 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
641 tp->snd_up = tp->write_seq;
644 /* If a not yet filled skb is pushed, do not send it if
645 * we have data packets in Qdisc or NIC queues :
646 * Because TX completion will happen shortly, it gives a chance
647 * to coalesce future sendmsg() payload into this skb, without
648 * need for a timer, and with no latency trade off.
649 * As packets containing data payload have a bigger truesize
650 * than pure acks (dataless) packets, the last checks prevent
651 * autocorking if we only have an ACK in Qdisc/NIC queues,
652 * or if TX completion was delayed after we processed ACK packet.
654 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
657 return skb->len < size_goal &&
658 sysctl_tcp_autocorking &&
659 skb != tcp_write_queue_head(sk) &&
660 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
663 static void tcp_push(struct sock *sk, int flags, int mss_now,
664 int nonagle, int size_goal)
666 struct tcp_sock *tp = tcp_sk(sk);
669 if (!tcp_send_head(sk))
672 skb = tcp_write_queue_tail(sk);
673 if (!(flags & MSG_MORE) || forced_push(tp))
674 tcp_mark_push(tp, skb);
676 tcp_mark_urg(tp, flags);
678 if (tcp_should_autocork(sk, skb, size_goal)) {
680 /* avoid atomic op if TSQ_THROTTLED bit is already set */
681 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
682 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
683 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
685 /* It is possible TX completion already happened
686 * before we set TSQ_THROTTLED.
688 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
692 if (flags & MSG_MORE)
693 nonagle = TCP_NAGLE_CORK;
695 __tcp_push_pending_frames(sk, mss_now, nonagle);
698 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
699 unsigned int offset, size_t len)
701 struct tcp_splice_state *tss = rd_desc->arg.data;
704 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
705 min(rd_desc->count, len), tss->flags,
708 rd_desc->count -= ret;
712 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
714 /* Store TCP splice context information in read_descriptor_t. */
715 read_descriptor_t rd_desc = {
720 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
724 * tcp_splice_read - splice data from TCP socket to a pipe
725 * @sock: socket to splice from
726 * @ppos: position (not valid)
727 * @pipe: pipe to splice to
728 * @len: number of bytes to splice
729 * @flags: splice modifier flags
732 * Will read pages from given socket and fill them into a pipe.
735 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
736 struct pipe_inode_info *pipe, size_t len,
739 struct sock *sk = sock->sk;
740 struct tcp_splice_state tss = {
749 sock_rps_record_flow(sk);
751 * We can't seek on a socket input
760 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
762 ret = __tcp_splice_read(sk, &tss);
768 if (sock_flag(sk, SOCK_DONE))
771 ret = sock_error(sk);
774 if (sk->sk_shutdown & RCV_SHUTDOWN)
776 if (sk->sk_state == TCP_CLOSE) {
778 * This occurs when user tries to read
779 * from never connected socket.
781 if (!sock_flag(sk, SOCK_DONE))
789 sk_wait_data(sk, &timeo, NULL);
790 if (signal_pending(current)) {
791 ret = sock_intr_errno(timeo);
804 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
805 (sk->sk_shutdown & RCV_SHUTDOWN) ||
806 signal_pending(current))
817 EXPORT_SYMBOL(tcp_splice_read);
819 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
824 /* The TCP header must be at least 32-bit aligned. */
825 size = ALIGN(size, 4);
827 if (unlikely(tcp_under_memory_pressure(sk)))
828 sk_mem_reclaim_partial(sk);
830 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
834 if (force_schedule) {
835 mem_scheduled = true;
836 sk_forced_mem_schedule(sk, skb->truesize);
838 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
840 if (likely(mem_scheduled)) {
841 skb_reserve(skb, sk->sk_prot->max_header);
843 * Make sure that we have exactly size bytes
844 * available to the caller, no more, no less.
846 skb->reserved_tailroom = skb->end - skb->tail - size;
851 sk->sk_prot->enter_memory_pressure(sk);
852 sk_stream_moderate_sndbuf(sk);
857 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
860 struct tcp_sock *tp = tcp_sk(sk);
861 u32 new_size_goal, size_goal;
863 if (!large_allowed || !sk_can_gso(sk))
866 /* Note : tcp_tso_autosize() will eventually split this later */
867 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
868 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
870 /* We try hard to avoid divides here */
871 size_goal = tp->gso_segs * mss_now;
872 if (unlikely(new_size_goal < size_goal ||
873 new_size_goal >= size_goal + mss_now)) {
874 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
875 sk->sk_gso_max_segs);
876 size_goal = tp->gso_segs * mss_now;
879 return max(size_goal, mss_now);
882 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
886 mss_now = tcp_current_mss(sk);
887 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
892 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
893 size_t size, int flags)
895 struct tcp_sock *tp = tcp_sk(sk);
896 int mss_now, size_goal;
899 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
901 /* Wait for a connection to finish. One exception is TCP Fast Open
902 * (passive side) where data is allowed to be sent before a connection
903 * is fully established.
905 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
906 !tcp_passive_fastopen(sk)) {
907 err = sk_stream_wait_connect(sk, &timeo);
912 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
914 mss_now = tcp_send_mss(sk, &size_goal, flags);
918 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
922 struct sk_buff *skb = tcp_write_queue_tail(sk);
926 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
928 if (!sk_stream_memory_free(sk))
929 goto wait_for_sndbuf;
931 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
932 skb_queue_empty(&sk->sk_write_queue));
934 goto wait_for_memory;
943 i = skb_shinfo(skb)->nr_frags;
944 can_coalesce = skb_can_coalesce(skb, i, page, offset);
945 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
946 tcp_mark_push(tp, skb);
949 if (!sk_wmem_schedule(sk, copy))
950 goto wait_for_memory;
953 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
956 skb_fill_page_desc(skb, i, page, offset, copy);
958 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
961 skb->data_len += copy;
962 skb->truesize += copy;
963 sk->sk_wmem_queued += copy;
964 sk_mem_charge(sk, copy);
965 skb->ip_summed = CHECKSUM_PARTIAL;
966 tp->write_seq += copy;
967 TCP_SKB_CB(skb)->end_seq += copy;
968 tcp_skb_pcount_set(skb, 0);
971 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
977 tcp_tx_timestamp(sk, skb);
981 if (skb->len < size_goal || (flags & MSG_OOB))
984 if (forced_push(tp)) {
985 tcp_mark_push(tp, skb);
986 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
987 } else if (skb == tcp_send_head(sk))
988 tcp_push_one(sk, mss_now);
992 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
994 tcp_push(sk, flags & ~MSG_MORE, mss_now,
995 TCP_NAGLE_PUSH, size_goal);
997 err = sk_stream_wait_memory(sk, &timeo);
1001 mss_now = tcp_send_mss(sk, &size_goal, flags);
1005 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
1006 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1013 /* make sure we wake any epoll edge trigger waiter */
1014 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1015 sk->sk_write_space(sk);
1016 return sk_stream_error(sk, flags, err);
1019 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1020 size_t size, int flags)
1024 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1025 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1026 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1030 res = do_tcp_sendpages(sk, page, offset, size, flags);
1034 EXPORT_SYMBOL(tcp_sendpage);
1036 static inline int select_size(const struct sock *sk, bool sg)
1038 const struct tcp_sock *tp = tcp_sk(sk);
1039 int tmp = tp->mss_cache;
1042 if (sk_can_gso(sk)) {
1043 /* Small frames wont use a full page:
1044 * Payload will immediately follow tcp header.
1046 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1048 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1050 if (tmp >= pgbreak &&
1051 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1059 void tcp_free_fastopen_req(struct tcp_sock *tp)
1061 if (tp->fastopen_req) {
1062 kfree(tp->fastopen_req);
1063 tp->fastopen_req = NULL;
1067 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1068 int *copied, size_t size)
1070 struct tcp_sock *tp = tcp_sk(sk);
1073 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1075 if (tp->fastopen_req)
1076 return -EALREADY; /* Another Fast Open is in progress */
1078 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1080 if (unlikely(!tp->fastopen_req))
1082 tp->fastopen_req->data = msg;
1083 tp->fastopen_req->size = size;
1085 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1086 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1087 msg->msg_namelen, flags);
1088 *copied = tp->fastopen_req->copied;
1089 tcp_free_fastopen_req(tp);
1093 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1095 struct tcp_sock *tp = tcp_sk(sk);
1096 struct sk_buff *skb;
1097 int flags, err, copied = 0;
1098 int mss_now = 0, size_goal, copied_syn = 0;
1104 flags = msg->msg_flags;
1105 if (flags & MSG_FASTOPEN) {
1106 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1107 if (err == -EINPROGRESS && copied_syn > 0)
1113 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1115 /* Wait for a connection to finish. One exception is TCP Fast Open
1116 * (passive side) where data is allowed to be sent before a connection
1117 * is fully established.
1119 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1120 !tcp_passive_fastopen(sk)) {
1121 err = sk_stream_wait_connect(sk, &timeo);
1126 if (unlikely(tp->repair)) {
1127 if (tp->repair_queue == TCP_RECV_QUEUE) {
1128 copied = tcp_send_rcvq(sk, msg, size);
1133 if (tp->repair_queue == TCP_NO_QUEUE)
1136 /* 'common' sending to sendq */
1139 /* This should be in poll */
1140 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1142 mss_now = tcp_send_mss(sk, &size_goal, flags);
1144 /* Ok commence sending. */
1148 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1151 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1153 while (msg_data_left(msg)) {
1155 int max = size_goal;
1157 skb = tcp_write_queue_tail(sk);
1158 if (tcp_send_head(sk)) {
1159 if (skb->ip_summed == CHECKSUM_NONE)
1161 copy = max - skb->len;
1166 /* Allocate new segment. If the interface is SG,
1167 * allocate skb fitting to single page.
1169 if (!sk_stream_memory_free(sk))
1170 goto wait_for_sndbuf;
1172 skb = sk_stream_alloc_skb(sk,
1173 select_size(sk, sg),
1175 skb_queue_empty(&sk->sk_write_queue));
1177 goto wait_for_memory;
1180 * Check whether we can use HW checksum.
1182 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1183 skb->ip_summed = CHECKSUM_PARTIAL;
1185 skb_entail(sk, skb);
1189 /* All packets are restored as if they have
1190 * already been sent. skb_mstamp isn't set to
1191 * avoid wrong rtt estimation.
1194 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1197 /* Try to append data to the end of skb. */
1198 if (copy > msg_data_left(msg))
1199 copy = msg_data_left(msg);
1201 /* Where to copy to? */
1202 if (skb_availroom(skb) > 0) {
1203 /* We have some space in skb head. Superb! */
1204 copy = min_t(int, copy, skb_availroom(skb));
1205 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1210 int i = skb_shinfo(skb)->nr_frags;
1211 struct page_frag *pfrag = sk_page_frag(sk);
1213 if (!sk_page_frag_refill(sk, pfrag))
1214 goto wait_for_memory;
1216 if (!skb_can_coalesce(skb, i, pfrag->page,
1218 if (i == MAX_SKB_FRAGS || !sg) {
1219 tcp_mark_push(tp, skb);
1225 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1227 if (!sk_wmem_schedule(sk, copy))
1228 goto wait_for_memory;
1230 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1237 /* Update the skb. */
1239 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1241 skb_fill_page_desc(skb, i, pfrag->page,
1242 pfrag->offset, copy);
1243 get_page(pfrag->page);
1245 pfrag->offset += copy;
1249 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1251 tp->write_seq += copy;
1252 TCP_SKB_CB(skb)->end_seq += copy;
1253 tcp_skb_pcount_set(skb, 0);
1256 if (!msg_data_left(msg)) {
1257 tcp_tx_timestamp(sk, skb);
1261 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1264 if (forced_push(tp)) {
1265 tcp_mark_push(tp, skb);
1266 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1267 } else if (skb == tcp_send_head(sk))
1268 tcp_push_one(sk, mss_now);
1272 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1275 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1276 TCP_NAGLE_PUSH, size_goal);
1278 err = sk_stream_wait_memory(sk, &timeo);
1282 mss_now = tcp_send_mss(sk, &size_goal, flags);
1287 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1291 if (copied + copied_syn)
1292 uid_stat_tcp_snd(current_uid(), copied + copied_syn);
1293 return copied + copied_syn;
1297 tcp_unlink_write_queue(skb, sk);
1298 /* It is the one place in all of TCP, except connection
1299 * reset, where we can be unlinking the send_head.
1301 tcp_check_send_head(sk, skb);
1302 sk_wmem_free_skb(sk, skb);
1306 if (copied + copied_syn)
1309 err = sk_stream_error(sk, flags, err);
1310 /* make sure we wake any epoll edge trigger waiter */
1311 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1312 sk->sk_write_space(sk);
1316 EXPORT_SYMBOL(tcp_sendmsg);
1319 * Handle reading urgent data. BSD has very simple semantics for
1320 * this, no blocking and very strange errors 8)
1323 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1325 struct tcp_sock *tp = tcp_sk(sk);
1327 /* No URG data to read. */
1328 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1329 tp->urg_data == TCP_URG_READ)
1330 return -EINVAL; /* Yes this is right ! */
1332 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1335 if (tp->urg_data & TCP_URG_VALID) {
1337 char c = tp->urg_data;
1339 if (!(flags & MSG_PEEK))
1340 tp->urg_data = TCP_URG_READ;
1342 /* Read urgent data. */
1343 msg->msg_flags |= MSG_OOB;
1346 if (!(flags & MSG_TRUNC))
1347 err = memcpy_to_msg(msg, &c, 1);
1350 msg->msg_flags |= MSG_TRUNC;
1352 return err ? -EFAULT : len;
1355 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1358 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1359 * the available implementations agree in this case:
1360 * this call should never block, independent of the
1361 * blocking state of the socket.
1362 * Mike <pall@rz.uni-karlsruhe.de>
1367 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1369 struct sk_buff *skb;
1370 int copied = 0, err = 0;
1372 /* XXX -- need to support SO_PEEK_OFF */
1374 skb_queue_walk(&sk->sk_write_queue, skb) {
1375 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1382 return err ?: copied;
1385 /* Clean up the receive buffer for full frames taken by the user,
1386 * then send an ACK if necessary. COPIED is the number of bytes
1387 * tcp_recvmsg has given to the user so far, it speeds up the
1388 * calculation of whether or not we must ACK for the sake of
1391 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1393 struct tcp_sock *tp = tcp_sk(sk);
1394 bool time_to_ack = false;
1396 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1398 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1399 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1400 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1402 if (inet_csk_ack_scheduled(sk)) {
1403 const struct inet_connection_sock *icsk = inet_csk(sk);
1404 /* Delayed ACKs frequently hit locked sockets during bulk
1406 if (icsk->icsk_ack.blocked ||
1407 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1408 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1410 * If this read emptied read buffer, we send ACK, if
1411 * connection is not bidirectional, user drained
1412 * receive buffer and there was a small segment
1416 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1417 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1418 !icsk->icsk_ack.pingpong)) &&
1419 !atomic_read(&sk->sk_rmem_alloc)))
1423 /* We send an ACK if we can now advertise a non-zero window
1424 * which has been raised "significantly".
1426 * Even if window raised up to infinity, do not send window open ACK
1427 * in states, where we will not receive more. It is useless.
1429 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1430 __u32 rcv_window_now = tcp_receive_window(tp);
1432 /* Optimize, __tcp_select_window() is not cheap. */
1433 if (2*rcv_window_now <= tp->window_clamp) {
1434 __u32 new_window = __tcp_select_window(sk);
1436 /* Send ACK now, if this read freed lots of space
1437 * in our buffer. Certainly, new_window is new window.
1438 * We can advertise it now, if it is not less than current one.
1439 * "Lots" means "at least twice" here.
1441 if (new_window && new_window >= 2 * rcv_window_now)
1449 static void tcp_prequeue_process(struct sock *sk)
1451 struct sk_buff *skb;
1452 struct tcp_sock *tp = tcp_sk(sk);
1454 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1456 /* RX process wants to run with disabled BHs, though it is not
1459 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1460 sk_backlog_rcv(sk, skb);
1463 /* Clear memory counter. */
1464 tp->ucopy.memory = 0;
1467 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1469 struct sk_buff *skb;
1472 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1473 offset = seq - TCP_SKB_CB(skb)->seq;
1474 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1476 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1480 /* This looks weird, but this can happen if TCP collapsing
1481 * splitted a fat GRO packet, while we released socket lock
1482 * in skb_splice_bits()
1484 sk_eat_skb(sk, skb);
1490 * This routine provides an alternative to tcp_recvmsg() for routines
1491 * that would like to handle copying from skbuffs directly in 'sendfile'
1494 * - It is assumed that the socket was locked by the caller.
1495 * - The routine does not block.
1496 * - At present, there is no support for reading OOB data
1497 * or for 'peeking' the socket using this routine
1498 * (although both would be easy to implement).
1500 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1501 sk_read_actor_t recv_actor)
1503 struct sk_buff *skb;
1504 struct tcp_sock *tp = tcp_sk(sk);
1505 u32 seq = tp->copied_seq;
1509 if (sk->sk_state == TCP_LISTEN)
1511 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1512 if (offset < skb->len) {
1516 len = skb->len - offset;
1517 /* Stop reading if we hit a patch of urgent data */
1519 u32 urg_offset = tp->urg_seq - seq;
1520 if (urg_offset < len)
1525 used = recv_actor(desc, skb, offset, len);
1530 } else if (used <= len) {
1535 /* If recv_actor drops the lock (e.g. TCP splice
1536 * receive) the skb pointer might be invalid when
1537 * getting here: tcp_collapse might have deleted it
1538 * while aggregating skbs from the socket queue.
1540 skb = tcp_recv_skb(sk, seq - 1, &offset);
1543 /* TCP coalescing might have appended data to the skb.
1544 * Try to splice more frags
1546 if (offset + 1 != skb->len)
1549 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1550 sk_eat_skb(sk, skb);
1554 sk_eat_skb(sk, skb);
1557 tp->copied_seq = seq;
1559 tp->copied_seq = seq;
1561 tcp_rcv_space_adjust(sk);
1563 /* Clean up data we have read: This will do ACK frames. */
1565 tcp_recv_skb(sk, seq, &offset);
1566 tcp_cleanup_rbuf(sk, copied);
1567 uid_stat_tcp_rcv(current_uid(), copied);
1571 EXPORT_SYMBOL(tcp_read_sock);
1574 * This routine copies from a sock struct into the user buffer.
1576 * Technical note: in 2.3 we work on _locked_ socket, so that
1577 * tricks with *seq access order and skb->users are not required.
1578 * Probably, code can be easily improved even more.
1581 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1582 int flags, int *addr_len)
1584 struct tcp_sock *tp = tcp_sk(sk);
1590 int target; /* Read at least this many bytes */
1592 struct task_struct *user_recv = NULL;
1593 struct sk_buff *skb, *last;
1596 if (unlikely(flags & MSG_ERRQUEUE))
1597 return inet_recv_error(sk, msg, len, addr_len);
1599 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1600 (sk->sk_state == TCP_ESTABLISHED))
1601 sk_busy_loop(sk, nonblock);
1606 if (sk->sk_state == TCP_LISTEN)
1609 timeo = sock_rcvtimeo(sk, nonblock);
1611 /* Urgent data needs to be handled specially. */
1612 if (flags & MSG_OOB)
1615 if (unlikely(tp->repair)) {
1617 if (!(flags & MSG_PEEK))
1620 if (tp->repair_queue == TCP_SEND_QUEUE)
1624 if (tp->repair_queue == TCP_NO_QUEUE)
1627 /* 'common' recv queue MSG_PEEK-ing */
1630 seq = &tp->copied_seq;
1631 if (flags & MSG_PEEK) {
1632 peek_seq = tp->copied_seq;
1636 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1641 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1642 if (tp->urg_data && tp->urg_seq == *seq) {
1645 if (signal_pending(current)) {
1646 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1651 /* Next get a buffer. */
1653 last = skb_peek_tail(&sk->sk_receive_queue);
1654 skb_queue_walk(&sk->sk_receive_queue, skb) {
1656 /* Now that we have two receive queues this
1659 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1660 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1661 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1665 offset = *seq - TCP_SKB_CB(skb)->seq;
1666 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1668 if (offset < skb->len)
1670 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1672 WARN(!(flags & MSG_PEEK),
1673 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1674 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1677 /* Well, if we have backlog, try to process it now yet. */
1679 if (copied >= target && !sk->sk_backlog.tail)
1684 sk->sk_state == TCP_CLOSE ||
1685 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1687 signal_pending(current))
1690 if (sock_flag(sk, SOCK_DONE))
1694 copied = sock_error(sk);
1698 if (sk->sk_shutdown & RCV_SHUTDOWN)
1701 if (sk->sk_state == TCP_CLOSE) {
1702 if (!sock_flag(sk, SOCK_DONE)) {
1703 /* This occurs when user tries to read
1704 * from never connected socket.
1717 if (signal_pending(current)) {
1718 copied = sock_intr_errno(timeo);
1723 tcp_cleanup_rbuf(sk, copied);
1725 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1726 /* Install new reader */
1727 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1728 user_recv = current;
1729 tp->ucopy.task = user_recv;
1730 tp->ucopy.msg = msg;
1733 tp->ucopy.len = len;
1735 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1736 !(flags & (MSG_PEEK | MSG_TRUNC)));
1738 /* Ugly... If prequeue is not empty, we have to
1739 * process it before releasing socket, otherwise
1740 * order will be broken at second iteration.
1741 * More elegant solution is required!!!
1743 * Look: we have the following (pseudo)queues:
1745 * 1. packets in flight
1750 * Each queue can be processed only if the next ones
1751 * are empty. At this point we have empty receive_queue.
1752 * But prequeue _can_ be not empty after 2nd iteration,
1753 * when we jumped to start of loop because backlog
1754 * processing added something to receive_queue.
1755 * We cannot release_sock(), because backlog contains
1756 * packets arrived _after_ prequeued ones.
1758 * Shortly, algorithm is clear --- to process all
1759 * the queues in order. We could make it more directly,
1760 * requeueing packets from backlog to prequeue, if
1761 * is not empty. It is more elegant, but eats cycles,
1764 if (!skb_queue_empty(&tp->ucopy.prequeue))
1767 /* __ Set realtime policy in scheduler __ */
1770 if (copied >= target) {
1771 /* Do not sleep, just process backlog. */
1775 sk_wait_data(sk, &timeo, last);
1781 /* __ Restore normal policy in scheduler __ */
1783 chunk = len - tp->ucopy.len;
1785 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1790 if (tp->rcv_nxt == tp->copied_seq &&
1791 !skb_queue_empty(&tp->ucopy.prequeue)) {
1793 tcp_prequeue_process(sk);
1795 chunk = len - tp->ucopy.len;
1797 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1803 if ((flags & MSG_PEEK) &&
1804 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1805 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1807 task_pid_nr(current));
1808 peek_seq = tp->copied_seq;
1813 /* Ok so how much can we use? */
1814 used = skb->len - offset;
1818 /* Do we have urgent data here? */
1820 u32 urg_offset = tp->urg_seq - *seq;
1821 if (urg_offset < used) {
1823 if (!sock_flag(sk, SOCK_URGINLINE)) {
1836 if (!(flags & MSG_TRUNC)) {
1837 err = skb_copy_datagram_msg(skb, offset, msg, used);
1839 /* Exception. Bailout! */
1850 tcp_rcv_space_adjust(sk);
1853 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1855 tcp_fast_path_check(sk);
1857 if (used + offset < skb->len)
1860 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1862 if (!(flags & MSG_PEEK))
1863 sk_eat_skb(sk, skb);
1867 /* Process the FIN. */
1869 if (!(flags & MSG_PEEK))
1870 sk_eat_skb(sk, skb);
1875 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1878 tp->ucopy.len = copied > 0 ? len : 0;
1880 tcp_prequeue_process(sk);
1882 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1883 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1889 tp->ucopy.task = NULL;
1893 /* According to UNIX98, msg_name/msg_namelen are ignored
1894 * on connected socket. I was just happy when found this 8) --ANK
1897 /* Clean up data we have read: This will do ACK frames. */
1898 tcp_cleanup_rbuf(sk, copied);
1903 uid_stat_tcp_rcv(current_uid(), copied);
1911 err = tcp_recv_urg(sk, msg, len, flags);
1913 uid_stat_tcp_rcv(current_uid(), err);
1917 err = tcp_peek_sndq(sk, msg, len);
1920 EXPORT_SYMBOL(tcp_recvmsg);
1922 void tcp_set_state(struct sock *sk, int state)
1924 int oldstate = sk->sk_state;
1927 case TCP_ESTABLISHED:
1928 if (oldstate != TCP_ESTABLISHED)
1929 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1933 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1934 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1936 sk->sk_prot->unhash(sk);
1937 if (inet_csk(sk)->icsk_bind_hash &&
1938 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1942 if (oldstate == TCP_ESTABLISHED)
1943 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1946 /* Change state AFTER socket is unhashed to avoid closed
1947 * socket sitting in hash tables.
1949 sk_state_store(sk, state);
1952 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1955 EXPORT_SYMBOL_GPL(tcp_set_state);
1958 * State processing on a close. This implements the state shift for
1959 * sending our FIN frame. Note that we only send a FIN for some
1960 * states. A shutdown() may have already sent the FIN, or we may be
1964 static const unsigned char new_state[16] = {
1965 /* current state: new state: action: */
1966 [0 /* (Invalid) */] = TCP_CLOSE,
1967 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1968 [TCP_SYN_SENT] = TCP_CLOSE,
1969 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1970 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1971 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1972 [TCP_TIME_WAIT] = TCP_CLOSE,
1973 [TCP_CLOSE] = TCP_CLOSE,
1974 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
1975 [TCP_LAST_ACK] = TCP_LAST_ACK,
1976 [TCP_LISTEN] = TCP_CLOSE,
1977 [TCP_CLOSING] = TCP_CLOSING,
1978 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
1981 static int tcp_close_state(struct sock *sk)
1983 int next = (int)new_state[sk->sk_state];
1984 int ns = next & TCP_STATE_MASK;
1986 tcp_set_state(sk, ns);
1988 return next & TCP_ACTION_FIN;
1992 * Shutdown the sending side of a connection. Much like close except
1993 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1996 void tcp_shutdown(struct sock *sk, int how)
1998 /* We need to grab some memory, and put together a FIN,
1999 * and then put it into the queue to be sent.
2000 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2002 if (!(how & SEND_SHUTDOWN))
2005 /* If we've already sent a FIN, or it's a closed state, skip this. */
2006 if ((1 << sk->sk_state) &
2007 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2008 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2009 /* Clear out any half completed packets. FIN if needed. */
2010 if (tcp_close_state(sk))
2014 EXPORT_SYMBOL(tcp_shutdown);
2016 bool tcp_check_oom(struct sock *sk, int shift)
2018 bool too_many_orphans, out_of_socket_memory;
2020 too_many_orphans = tcp_too_many_orphans(sk, shift);
2021 out_of_socket_memory = tcp_out_of_memory(sk);
2023 if (too_many_orphans)
2024 net_info_ratelimited("too many orphaned sockets\n");
2025 if (out_of_socket_memory)
2026 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2027 return too_many_orphans || out_of_socket_memory;
2030 void tcp_close(struct sock *sk, long timeout)
2032 struct sk_buff *skb;
2033 int data_was_unread = 0;
2037 sk->sk_shutdown = SHUTDOWN_MASK;
2039 if (sk->sk_state == TCP_LISTEN) {
2040 tcp_set_state(sk, TCP_CLOSE);
2043 inet_csk_listen_stop(sk);
2045 goto adjudge_to_death;
2048 /* We need to flush the recv. buffs. We do this only on the
2049 * descriptor close, not protocol-sourced closes, because the
2050 * reader process may not have drained the data yet!
2052 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2053 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2055 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2057 data_was_unread += len;
2063 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2064 if (sk->sk_state == TCP_CLOSE)
2065 goto adjudge_to_death;
2067 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2068 * data was lost. To witness the awful effects of the old behavior of
2069 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2070 * GET in an FTP client, suspend the process, wait for the client to
2071 * advertise a zero window, then kill -9 the FTP client, wheee...
2072 * Note: timeout is always zero in such a case.
2074 if (unlikely(tcp_sk(sk)->repair)) {
2075 sk->sk_prot->disconnect(sk, 0);
2076 } else if (data_was_unread) {
2077 /* Unread data was tossed, zap the connection. */
2078 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2079 tcp_set_state(sk, TCP_CLOSE);
2080 tcp_send_active_reset(sk, sk->sk_allocation);
2081 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2082 /* Check zero linger _after_ checking for unread data. */
2083 sk->sk_prot->disconnect(sk, 0);
2084 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2085 } else if (tcp_close_state(sk)) {
2086 /* We FIN if the application ate all the data before
2087 * zapping the connection.
2090 /* RED-PEN. Formally speaking, we have broken TCP state
2091 * machine. State transitions:
2093 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2094 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2095 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2097 * are legal only when FIN has been sent (i.e. in window),
2098 * rather than queued out of window. Purists blame.
2100 * F.e. "RFC state" is ESTABLISHED,
2101 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2103 * The visible declinations are that sometimes
2104 * we enter time-wait state, when it is not required really
2105 * (harmless), do not send active resets, when they are
2106 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2107 * they look as CLOSING or LAST_ACK for Linux)
2108 * Probably, I missed some more holelets.
2110 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2111 * in a single packet! (May consider it later but will
2112 * probably need API support or TCP_CORK SYN-ACK until
2113 * data is written and socket is closed.)
2118 sk_stream_wait_close(sk, timeout);
2121 state = sk->sk_state;
2125 /* It is the last release_sock in its life. It will remove backlog. */
2129 /* Now socket is owned by kernel and we acquire BH lock
2130 to finish close. No need to check for user refs.
2134 WARN_ON(sock_owned_by_user(sk));
2136 percpu_counter_inc(sk->sk_prot->orphan_count);
2138 /* Have we already been destroyed by a softirq or backlog? */
2139 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2142 /* This is a (useful) BSD violating of the RFC. There is a
2143 * problem with TCP as specified in that the other end could
2144 * keep a socket open forever with no application left this end.
2145 * We use a 1 minute timeout (about the same as BSD) then kill
2146 * our end. If they send after that then tough - BUT: long enough
2147 * that we won't make the old 4*rto = almost no time - whoops
2150 * Nope, it was not mistake. It is really desired behaviour
2151 * f.e. on http servers, when such sockets are useless, but
2152 * consume significant resources. Let's do it with special
2153 * linger2 option. --ANK
2156 if (sk->sk_state == TCP_FIN_WAIT2) {
2157 struct tcp_sock *tp = tcp_sk(sk);
2158 if (tp->linger2 < 0) {
2159 tcp_set_state(sk, TCP_CLOSE);
2160 tcp_send_active_reset(sk, GFP_ATOMIC);
2161 NET_INC_STATS_BH(sock_net(sk),
2162 LINUX_MIB_TCPABORTONLINGER);
2164 const int tmo = tcp_fin_time(sk);
2166 if (tmo > TCP_TIMEWAIT_LEN) {
2167 inet_csk_reset_keepalive_timer(sk,
2168 tmo - TCP_TIMEWAIT_LEN);
2170 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2175 if (sk->sk_state != TCP_CLOSE) {
2177 if (tcp_check_oom(sk, 0)) {
2178 tcp_set_state(sk, TCP_CLOSE);
2179 tcp_send_active_reset(sk, GFP_ATOMIC);
2180 NET_INC_STATS_BH(sock_net(sk),
2181 LINUX_MIB_TCPABORTONMEMORY);
2185 if (sk->sk_state == TCP_CLOSE) {
2186 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2187 /* We could get here with a non-NULL req if the socket is
2188 * aborted (e.g., closed with unread data) before 3WHS
2192 reqsk_fastopen_remove(sk, req, false);
2193 inet_csk_destroy_sock(sk);
2195 /* Otherwise, socket is reprieved until protocol close. */
2202 EXPORT_SYMBOL(tcp_close);
2204 /* These states need RST on ABORT according to RFC793 */
2206 static inline bool tcp_need_reset(int state)
2208 return (1 << state) &
2209 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2210 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2213 int tcp_disconnect(struct sock *sk, int flags)
2215 struct inet_sock *inet = inet_sk(sk);
2216 struct inet_connection_sock *icsk = inet_csk(sk);
2217 struct tcp_sock *tp = tcp_sk(sk);
2219 int old_state = sk->sk_state;
2221 if (old_state != TCP_CLOSE)
2222 tcp_set_state(sk, TCP_CLOSE);
2224 /* ABORT function of RFC793 */
2225 if (old_state == TCP_LISTEN) {
2226 inet_csk_listen_stop(sk);
2227 } else if (unlikely(tp->repair)) {
2228 sk->sk_err = ECONNABORTED;
2229 } else if (tcp_need_reset(old_state) ||
2230 (tp->snd_nxt != tp->write_seq &&
2231 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2232 /* The last check adjusts for discrepancy of Linux wrt. RFC
2235 tcp_send_active_reset(sk, gfp_any());
2236 sk->sk_err = ECONNRESET;
2237 } else if (old_state == TCP_SYN_SENT)
2238 sk->sk_err = ECONNRESET;
2240 tcp_clear_xmit_timers(sk);
2241 __skb_queue_purge(&sk->sk_receive_queue);
2242 tcp_write_queue_purge(sk);
2243 __skb_queue_purge(&tp->out_of_order_queue);
2245 inet->inet_dport = 0;
2247 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2248 inet_reset_saddr(sk);
2250 sk->sk_shutdown = 0;
2251 sock_reset_flag(sk, SOCK_DONE);
2253 tp->write_seq += tp->max_window + 2;
2254 if (tp->write_seq == 0)
2256 icsk->icsk_backoff = 0;
2258 icsk->icsk_probes_out = 0;
2259 tp->packets_out = 0;
2260 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2261 tp->snd_cwnd_cnt = 0;
2262 tp->window_clamp = 0;
2263 tcp_set_ca_state(sk, TCP_CA_Open);
2264 tcp_clear_retrans(tp);
2265 inet_csk_delack_init(sk);
2266 tcp_init_send_head(sk);
2267 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2270 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2272 sk->sk_error_report(sk);
2275 EXPORT_SYMBOL(tcp_disconnect);
2277 static inline bool tcp_can_repair_sock(const struct sock *sk)
2279 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2280 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2283 static int tcp_repair_options_est(struct tcp_sock *tp,
2284 struct tcp_repair_opt __user *optbuf, unsigned int len)
2286 struct tcp_repair_opt opt;
2288 while (len >= sizeof(opt)) {
2289 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2295 switch (opt.opt_code) {
2297 tp->rx_opt.mss_clamp = opt.opt_val;
2301 u16 snd_wscale = opt.opt_val & 0xFFFF;
2302 u16 rcv_wscale = opt.opt_val >> 16;
2304 if (snd_wscale > 14 || rcv_wscale > 14)
2307 tp->rx_opt.snd_wscale = snd_wscale;
2308 tp->rx_opt.rcv_wscale = rcv_wscale;
2309 tp->rx_opt.wscale_ok = 1;
2312 case TCPOPT_SACK_PERM:
2313 if (opt.opt_val != 0)
2316 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2317 if (sysctl_tcp_fack)
2318 tcp_enable_fack(tp);
2320 case TCPOPT_TIMESTAMP:
2321 if (opt.opt_val != 0)
2324 tp->rx_opt.tstamp_ok = 1;
2333 * Socket option code for TCP.
2335 static int do_tcp_setsockopt(struct sock *sk, int level,
2336 int optname, char __user *optval, unsigned int optlen)
2338 struct tcp_sock *tp = tcp_sk(sk);
2339 struct inet_connection_sock *icsk = inet_csk(sk);
2343 /* These are data/string values, all the others are ints */
2345 case TCP_CONGESTION: {
2346 char name[TCP_CA_NAME_MAX];
2351 val = strncpy_from_user(name, optval,
2352 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2358 err = tcp_set_congestion_control(sk, name);
2367 if (optlen < sizeof(int))
2370 if (get_user(val, (int __user *)optval))
2377 /* Values greater than interface MTU won't take effect. However
2378 * at the point when this call is done we typically don't yet
2379 * know which interface is going to be used */
2380 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2384 tp->rx_opt.user_mss = val;
2389 /* TCP_NODELAY is weaker than TCP_CORK, so that
2390 * this option on corked socket is remembered, but
2391 * it is not activated until cork is cleared.
2393 * However, when TCP_NODELAY is set we make
2394 * an explicit push, which overrides even TCP_CORK
2395 * for currently queued segments.
2397 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2398 tcp_push_pending_frames(sk);
2400 tp->nonagle &= ~TCP_NAGLE_OFF;
2404 case TCP_THIN_LINEAR_TIMEOUTS:
2405 if (val < 0 || val > 1)
2411 case TCP_THIN_DUPACK:
2412 if (val < 0 || val > 1)
2415 tp->thin_dupack = val;
2416 if (tp->thin_dupack)
2417 tcp_disable_early_retrans(tp);
2422 if (!tcp_can_repair_sock(sk))
2424 else if (val == 1) {
2426 sk->sk_reuse = SK_FORCE_REUSE;
2427 tp->repair_queue = TCP_NO_QUEUE;
2428 } else if (val == 0) {
2430 sk->sk_reuse = SK_NO_REUSE;
2431 tcp_send_window_probe(sk);
2437 case TCP_REPAIR_QUEUE:
2440 else if (val < TCP_QUEUES_NR)
2441 tp->repair_queue = val;
2447 if (sk->sk_state != TCP_CLOSE)
2449 else if (tp->repair_queue == TCP_SEND_QUEUE)
2450 tp->write_seq = val;
2451 else if (tp->repair_queue == TCP_RECV_QUEUE)
2457 case TCP_REPAIR_OPTIONS:
2460 else if (sk->sk_state == TCP_ESTABLISHED)
2461 err = tcp_repair_options_est(tp,
2462 (struct tcp_repair_opt __user *)optval,
2469 /* When set indicates to always queue non-full frames.
2470 * Later the user clears this option and we transmit
2471 * any pending partial frames in the queue. This is
2472 * meant to be used alongside sendfile() to get properly
2473 * filled frames when the user (for example) must write
2474 * out headers with a write() call first and then use
2475 * sendfile to send out the data parts.
2477 * TCP_CORK can be set together with TCP_NODELAY and it is
2478 * stronger than TCP_NODELAY.
2481 tp->nonagle |= TCP_NAGLE_CORK;
2483 tp->nonagle &= ~TCP_NAGLE_CORK;
2484 if (tp->nonagle&TCP_NAGLE_OFF)
2485 tp->nonagle |= TCP_NAGLE_PUSH;
2486 tcp_push_pending_frames(sk);
2491 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2494 tp->keepalive_time = val * HZ;
2495 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2496 !((1 << sk->sk_state) &
2497 (TCPF_CLOSE | TCPF_LISTEN))) {
2498 u32 elapsed = keepalive_time_elapsed(tp);
2499 if (tp->keepalive_time > elapsed)
2500 elapsed = tp->keepalive_time - elapsed;
2503 inet_csk_reset_keepalive_timer(sk, elapsed);
2508 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2511 tp->keepalive_intvl = val * HZ;
2514 if (val < 1 || val > MAX_TCP_KEEPCNT)
2517 tp->keepalive_probes = val;
2520 if (val < 1 || val > MAX_TCP_SYNCNT)
2523 icsk->icsk_syn_retries = val;
2527 if (val < 0 || val > 1)
2536 else if (val > sysctl_tcp_fin_timeout / HZ)
2539 tp->linger2 = val * HZ;
2542 case TCP_DEFER_ACCEPT:
2543 /* Translate value in seconds to number of retransmits */
2544 icsk->icsk_accept_queue.rskq_defer_accept =
2545 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2549 case TCP_WINDOW_CLAMP:
2551 if (sk->sk_state != TCP_CLOSE) {
2555 tp->window_clamp = 0;
2557 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2558 SOCK_MIN_RCVBUF / 2 : val;
2563 icsk->icsk_ack.pingpong = 1;
2565 icsk->icsk_ack.pingpong = 0;
2566 if ((1 << sk->sk_state) &
2567 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2568 inet_csk_ack_scheduled(sk)) {
2569 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2570 tcp_cleanup_rbuf(sk, 1);
2572 icsk->icsk_ack.pingpong = 1;
2577 #ifdef CONFIG_TCP_MD5SIG
2579 /* Read the IP->Key mappings from userspace */
2580 err = tp->af_specific->md5_parse(sk, optval, optlen);
2583 case TCP_USER_TIMEOUT:
2584 /* Cap the max time in ms TCP will retry or probe the window
2585 * before giving up and aborting (ETIMEDOUT) a connection.
2590 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2594 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2596 tcp_fastopen_init_key_once(true);
2598 fastopen_queue_tune(sk, val);
2607 tp->tsoffset = val - tcp_time_stamp;
2609 case TCP_NOTSENT_LOWAT:
2610 tp->notsent_lowat = val;
2611 sk->sk_write_space(sk);
2622 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2623 unsigned int optlen)
2625 const struct inet_connection_sock *icsk = inet_csk(sk);
2627 if (level != SOL_TCP)
2628 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2630 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2632 EXPORT_SYMBOL(tcp_setsockopt);
2634 #ifdef CONFIG_COMPAT
2635 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2636 char __user *optval, unsigned int optlen)
2638 if (level != SOL_TCP)
2639 return inet_csk_compat_setsockopt(sk, level, optname,
2641 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2643 EXPORT_SYMBOL(compat_tcp_setsockopt);
2646 /* Return information about state of tcp endpoint in API format. */
2647 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2649 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2650 const struct inet_connection_sock *icsk = inet_csk(sk);
2651 u32 now = tcp_time_stamp;
2655 memset(info, 0, sizeof(*info));
2656 if (sk->sk_type != SOCK_STREAM)
2659 info->tcpi_state = sk_state_load(sk);
2661 info->tcpi_ca_state = icsk->icsk_ca_state;
2662 info->tcpi_retransmits = icsk->icsk_retransmits;
2663 info->tcpi_probes = icsk->icsk_probes_out;
2664 info->tcpi_backoff = icsk->icsk_backoff;
2666 if (tp->rx_opt.tstamp_ok)
2667 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2668 if (tcp_is_sack(tp))
2669 info->tcpi_options |= TCPI_OPT_SACK;
2670 if (tp->rx_opt.wscale_ok) {
2671 info->tcpi_options |= TCPI_OPT_WSCALE;
2672 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2673 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2676 if (tp->ecn_flags & TCP_ECN_OK)
2677 info->tcpi_options |= TCPI_OPT_ECN;
2678 if (tp->ecn_flags & TCP_ECN_SEEN)
2679 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2680 if (tp->syn_data_acked)
2681 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2683 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2684 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2685 info->tcpi_snd_mss = tp->mss_cache;
2686 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2688 if (info->tcpi_state == TCP_LISTEN) {
2689 info->tcpi_unacked = sk->sk_ack_backlog;
2690 info->tcpi_sacked = sk->sk_max_ack_backlog;
2692 info->tcpi_unacked = tp->packets_out;
2693 info->tcpi_sacked = tp->sacked_out;
2695 info->tcpi_lost = tp->lost_out;
2696 info->tcpi_retrans = tp->retrans_out;
2697 info->tcpi_fackets = tp->fackets_out;
2699 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2700 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2701 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2703 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2704 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2705 info->tcpi_rtt = tp->srtt_us >> 3;
2706 info->tcpi_rttvar = tp->mdev_us >> 2;
2707 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2708 info->tcpi_snd_cwnd = tp->snd_cwnd;
2709 info->tcpi_advmss = tp->advmss;
2710 info->tcpi_reordering = tp->reordering;
2712 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2713 info->tcpi_rcv_space = tp->rcvq_space.space;
2715 info->tcpi_total_retrans = tp->total_retrans;
2717 rate = READ_ONCE(sk->sk_pacing_rate);
2718 info->tcpi_pacing_rate = rate != ~0U ? rate : ~0ULL;
2720 rate = READ_ONCE(sk->sk_max_pacing_rate);
2721 info->tcpi_max_pacing_rate = rate != ~0U ? rate : ~0ULL;
2724 start = u64_stats_fetch_begin_irq(&tp->syncp);
2725 info->tcpi_bytes_acked = tp->bytes_acked;
2726 info->tcpi_bytes_received = tp->bytes_received;
2727 } while (u64_stats_fetch_retry_irq(&tp->syncp, start));
2728 info->tcpi_segs_out = tp->segs_out;
2729 info->tcpi_segs_in = tp->segs_in;
2731 EXPORT_SYMBOL_GPL(tcp_get_info);
2733 static int do_tcp_getsockopt(struct sock *sk, int level,
2734 int optname, char __user *optval, int __user *optlen)
2736 struct inet_connection_sock *icsk = inet_csk(sk);
2737 struct tcp_sock *tp = tcp_sk(sk);
2740 if (get_user(len, optlen))
2743 len = min_t(unsigned int, len, sizeof(int));
2750 val = tp->mss_cache;
2751 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2752 val = tp->rx_opt.user_mss;
2754 val = tp->rx_opt.mss_clamp;
2757 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2760 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2763 val = keepalive_time_when(tp) / HZ;
2766 val = keepalive_intvl_when(tp) / HZ;
2769 val = keepalive_probes(tp);
2772 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2777 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2779 case TCP_DEFER_ACCEPT:
2780 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2781 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2783 case TCP_WINDOW_CLAMP:
2784 val = tp->window_clamp;
2787 struct tcp_info info;
2789 if (get_user(len, optlen))
2792 tcp_get_info(sk, &info);
2794 len = min_t(unsigned int, len, sizeof(info));
2795 if (put_user(len, optlen))
2797 if (copy_to_user(optval, &info, len))
2802 const struct tcp_congestion_ops *ca_ops;
2803 union tcp_cc_info info;
2807 if (get_user(len, optlen))
2810 ca_ops = icsk->icsk_ca_ops;
2811 if (ca_ops && ca_ops->get_info)
2812 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2814 len = min_t(unsigned int, len, sz);
2815 if (put_user(len, optlen))
2817 if (copy_to_user(optval, &info, len))
2822 val = !icsk->icsk_ack.pingpong;
2825 case TCP_CONGESTION:
2826 if (get_user(len, optlen))
2828 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2829 if (put_user(len, optlen))
2831 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2835 case TCP_THIN_LINEAR_TIMEOUTS:
2838 case TCP_THIN_DUPACK:
2839 val = tp->thin_dupack;
2846 case TCP_REPAIR_QUEUE:
2848 val = tp->repair_queue;
2854 if (tp->repair_queue == TCP_SEND_QUEUE)
2855 val = tp->write_seq;
2856 else if (tp->repair_queue == TCP_RECV_QUEUE)
2862 case TCP_USER_TIMEOUT:
2863 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2867 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
2871 val = tcp_time_stamp + tp->tsoffset;
2873 case TCP_NOTSENT_LOWAT:
2874 val = tp->notsent_lowat;
2879 case TCP_SAVED_SYN: {
2880 if (get_user(len, optlen))
2884 if (tp->saved_syn) {
2885 if (len < tp->saved_syn[0]) {
2886 if (put_user(tp->saved_syn[0], optlen)) {
2893 len = tp->saved_syn[0];
2894 if (put_user(len, optlen)) {
2898 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
2902 tcp_saved_syn_free(tp);
2907 if (put_user(len, optlen))
2913 return -ENOPROTOOPT;
2916 if (put_user(len, optlen))
2918 if (copy_to_user(optval, &val, len))
2923 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2926 struct inet_connection_sock *icsk = inet_csk(sk);
2928 if (level != SOL_TCP)
2929 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2931 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2933 EXPORT_SYMBOL(tcp_getsockopt);
2935 #ifdef CONFIG_COMPAT
2936 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2937 char __user *optval, int __user *optlen)
2939 if (level != SOL_TCP)
2940 return inet_csk_compat_getsockopt(sk, level, optname,
2942 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2944 EXPORT_SYMBOL(compat_tcp_getsockopt);
2947 #ifdef CONFIG_TCP_MD5SIG
2948 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2949 static DEFINE_MUTEX(tcp_md5sig_mutex);
2950 static bool tcp_md5sig_pool_populated = false;
2952 static void __tcp_alloc_md5sig_pool(void)
2956 for_each_possible_cpu(cpu) {
2957 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2958 struct crypto_hash *hash;
2960 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2961 if (IS_ERR_OR_NULL(hash))
2963 per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
2966 /* before setting tcp_md5sig_pool_populated, we must commit all writes
2967 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
2970 tcp_md5sig_pool_populated = true;
2973 bool tcp_alloc_md5sig_pool(void)
2975 if (unlikely(!tcp_md5sig_pool_populated)) {
2976 mutex_lock(&tcp_md5sig_mutex);
2978 if (!tcp_md5sig_pool_populated)
2979 __tcp_alloc_md5sig_pool();
2981 mutex_unlock(&tcp_md5sig_mutex);
2983 return tcp_md5sig_pool_populated;
2985 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2989 * tcp_get_md5sig_pool - get md5sig_pool for this user
2991 * We use percpu structure, so if we succeed, we exit with preemption
2992 * and BH disabled, to make sure another thread or softirq handling
2993 * wont try to get same context.
2995 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2999 if (tcp_md5sig_pool_populated) {
3000 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3002 return this_cpu_ptr(&tcp_md5sig_pool);
3007 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3009 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3010 const struct tcphdr *th)
3012 struct scatterlist sg;
3016 /* We are not allowed to change tcphdr, make a local copy */
3017 memcpy(&hdr, th, sizeof(hdr));
3020 /* options aren't included in the hash */
3021 sg_init_one(&sg, &hdr, sizeof(hdr));
3022 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3025 EXPORT_SYMBOL(tcp_md5_hash_header);
3027 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3028 const struct sk_buff *skb, unsigned int header_len)
3030 struct scatterlist sg;
3031 const struct tcphdr *tp = tcp_hdr(skb);
3032 struct hash_desc *desc = &hp->md5_desc;
3034 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3035 skb_headlen(skb) - header_len : 0;
3036 const struct skb_shared_info *shi = skb_shinfo(skb);
3037 struct sk_buff *frag_iter;
3039 sg_init_table(&sg, 1);
3041 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3042 if (crypto_hash_update(desc, &sg, head_data_len))
3045 for (i = 0; i < shi->nr_frags; ++i) {
3046 const struct skb_frag_struct *f = &shi->frags[i];
3047 unsigned int offset = f->page_offset;
3048 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3050 sg_set_page(&sg, page, skb_frag_size(f),
3051 offset_in_page(offset));
3052 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3056 skb_walk_frags(skb, frag_iter)
3057 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3062 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3064 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3066 struct scatterlist sg;
3068 sg_init_one(&sg, key->key, key->keylen);
3069 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3071 EXPORT_SYMBOL(tcp_md5_hash_key);
3075 void tcp_done(struct sock *sk)
3077 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3079 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3080 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3082 tcp_set_state(sk, TCP_CLOSE);
3083 tcp_clear_xmit_timers(sk);
3085 reqsk_fastopen_remove(sk, req, false);
3087 sk->sk_shutdown = SHUTDOWN_MASK;
3089 if (!sock_flag(sk, SOCK_DEAD))
3090 sk->sk_state_change(sk);
3092 inet_csk_destroy_sock(sk);
3094 EXPORT_SYMBOL_GPL(tcp_done);
3096 extern struct tcp_congestion_ops tcp_reno;
3098 static __initdata unsigned long thash_entries;
3099 static int __init set_thash_entries(char *str)
3106 ret = kstrtoul(str, 0, &thash_entries);
3112 __setup("thash_entries=", set_thash_entries);
3114 static void __init tcp_init_mem(void)
3116 unsigned long limit = nr_free_buffer_pages() / 16;
3118 limit = max(limit, 128UL);
3119 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3120 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3121 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3124 void __init tcp_init(void)
3126 unsigned long limit;
3127 int max_rshare, max_wshare, cnt;
3130 sock_skb_cb_check_size(sizeof(struct tcp_skb_cb));
3132 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3133 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3134 tcp_hashinfo.bind_bucket_cachep =
3135 kmem_cache_create("tcp_bind_bucket",
3136 sizeof(struct inet_bind_bucket), 0,
3137 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3139 /* Size and allocate the main established and bind bucket
3142 * The methodology is similar to that of the buffer cache.
3144 tcp_hashinfo.ehash =
3145 alloc_large_system_hash("TCP established",
3146 sizeof(struct inet_ehash_bucket),
3148 17, /* one slot per 128 KB of memory */
3151 &tcp_hashinfo.ehash_mask,
3153 thash_entries ? 0 : 512 * 1024);
3154 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3155 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3157 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3158 panic("TCP: failed to alloc ehash_locks");
3159 tcp_hashinfo.bhash =
3160 alloc_large_system_hash("TCP bind",
3161 sizeof(struct inet_bind_hashbucket),
3162 tcp_hashinfo.ehash_mask + 1,
3163 17, /* one slot per 128 KB of memory */
3165 &tcp_hashinfo.bhash_size,
3169 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3170 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3171 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3172 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3176 cnt = tcp_hashinfo.ehash_mask + 1;
3178 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3179 sysctl_tcp_max_orphans = cnt / 2;
3180 sysctl_max_syn_backlog = max(128, cnt / 256);
3183 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3184 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3185 max_wshare = min(4UL*1024*1024, limit);
3186 max_rshare = min(6UL*1024*1024, limit);
3188 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3189 sysctl_tcp_wmem[1] = 16*1024;
3190 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3192 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3193 sysctl_tcp_rmem[1] = 87380;
3194 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3196 pr_info("Hash tables configured (established %u bind %u)\n",
3197 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3200 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3204 static int tcp_is_local(struct net *net, __be32 addr) {
3206 struct flowi4 fl4 = { .daddr = addr };
3207 rt = ip_route_output_key(net, &fl4);
3208 if (IS_ERR_OR_NULL(rt))
3210 return rt->dst.dev && (rt->dst.dev->flags & IFF_LOOPBACK);
3213 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3214 static int tcp_is_local6(struct net *net, struct in6_addr *addr) {
3215 struct rt6_info *rt6 = rt6_lookup(net, addr, addr, 0, 0);
3216 return rt6 && rt6->dst.dev && (rt6->dst.dev->flags & IFF_LOOPBACK);
3221 * tcp_nuke_addr - destroy all sockets on the given local address
3222 * if local address is the unspecified address (0.0.0.0 or ::), destroy all
3223 * sockets with local addresses that are not configured.
3225 int tcp_nuke_addr(struct net *net, struct sockaddr *addr)
3227 int family = addr->sa_family;
3228 unsigned int bucket;
3231 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3232 struct in6_addr *in6;
3234 if (family == AF_INET) {
3235 in = &((struct sockaddr_in *)addr)->sin_addr;
3236 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3237 } else if (family == AF_INET6) {
3238 in6 = &((struct sockaddr_in6 *)addr)->sin6_addr;
3241 return -EAFNOSUPPORT;
3244 for (bucket = 0; bucket <= tcp_hashinfo.ehash_mask; bucket++) {
3245 struct hlist_nulls_node *node;
3247 spinlock_t *lock = inet_ehash_lockp(&tcp_hashinfo, bucket);
3251 sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[bucket].chain) {
3252 struct inet_sock *inet = inet_sk(sk);
3254 if (sk->sk_state == TCP_TIME_WAIT) {
3256 * Sockets that are in TIME_WAIT state are
3257 * instances of lightweight inet_timewait_sock,
3258 * we should simply skip them (or we'll try to
3259 * access non-existing fields and crash).
3264 if (sysctl_ip_dynaddr && sk->sk_state == TCP_SYN_SENT)
3267 if (sock_flag(sk, SOCK_DEAD))
3270 if (family == AF_INET) {
3271 __be32 s4 = inet->inet_rcv_saddr;
3272 if (s4 == LOOPBACK4_IPV6)
3275 if (in->s_addr != s4 &&
3276 !(in->s_addr == INADDR_ANY &&
3277 !tcp_is_local(net, s4)))
3281 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3282 if (family == AF_INET6) {
3283 struct in6_addr *s6;
3287 s6 = &inet->pinet6->rcv_saddr;
3288 if (ipv6_addr_type(s6) == IPV6_ADDR_MAPPED)
3291 if (!ipv6_addr_equal(in6, s6) &&
3292 !(ipv6_addr_equal(in6, &in6addr_any) &&
3293 !tcp_is_local6(net, s6)))
3299 spin_unlock_bh(lock);
3303 sk->sk_err = ETIMEDOUT;
3304 sk->sk_error_report(sk);
3313 spin_unlock_bh(lock);