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>
273 #include <net/icmp.h>
274 #include <net/inet_common.h>
276 #include <net/xfrm.h>
278 #include <net/ip6_route.h>
279 #include <net/ipv6.h>
280 #include <net/transp_v6.h>
281 #include <net/sock.h>
283 #include <asm/uaccess.h>
284 #include <asm/ioctls.h>
285 #include <net/busy_poll.h>
287 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
289 int sysctl_tcp_min_tso_segs __read_mostly = 2;
291 int sysctl_tcp_autocorking __read_mostly = 1;
293 struct percpu_counter tcp_orphan_count;
294 EXPORT_SYMBOL_GPL(tcp_orphan_count);
296 long sysctl_tcp_mem[3] __read_mostly;
297 int sysctl_tcp_wmem[3] __read_mostly;
298 int sysctl_tcp_rmem[3] __read_mostly;
300 EXPORT_SYMBOL(sysctl_tcp_mem);
301 EXPORT_SYMBOL(sysctl_tcp_rmem);
302 EXPORT_SYMBOL(sysctl_tcp_wmem);
304 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
305 EXPORT_SYMBOL(tcp_memory_allocated);
308 * Current number of TCP sockets.
310 struct percpu_counter tcp_sockets_allocated;
311 EXPORT_SYMBOL(tcp_sockets_allocated);
316 struct tcp_splice_state {
317 struct pipe_inode_info *pipe;
323 * Pressure flag: try to collapse.
324 * Technical note: it is used by multiple contexts non atomically.
325 * All the __sk_mem_schedule() is of this nature: accounting
326 * is strict, actions are advisory and have some latency.
328 int tcp_memory_pressure __read_mostly;
329 EXPORT_SYMBOL(tcp_memory_pressure);
331 void tcp_enter_memory_pressure(struct sock *sk)
333 if (!tcp_memory_pressure) {
334 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
335 tcp_memory_pressure = 1;
338 EXPORT_SYMBOL(tcp_enter_memory_pressure);
340 /* Convert seconds to retransmits based on initial and max timeout */
341 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
346 int period = timeout;
349 while (seconds > period && res < 255) {
352 if (timeout > rto_max)
360 /* Convert retransmits to seconds based on initial and max timeout */
361 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
369 if (timeout > rto_max)
377 /* Address-family independent initialization for a tcp_sock.
379 * NOTE: A lot of things set to zero explicitly by call to
380 * sk_alloc() so need not be done here.
382 void tcp_init_sock(struct sock *sk)
384 struct inet_connection_sock *icsk = inet_csk(sk);
385 struct tcp_sock *tp = tcp_sk(sk);
387 __skb_queue_head_init(&tp->out_of_order_queue);
388 tcp_init_xmit_timers(sk);
389 tcp_prequeue_init(tp);
390 INIT_LIST_HEAD(&tp->tsq_node);
392 icsk->icsk_rto = TCP_TIMEOUT_INIT;
393 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
394 tp->rtt_min[0].rtt = ~0U;
396 /* So many TCP implementations out there (incorrectly) count the
397 * initial SYN frame in their delayed-ACK and congestion control
398 * algorithms that we must have the following bandaid to talk
399 * efficiently to them. -DaveM
401 tp->snd_cwnd = TCP_INIT_CWND;
403 /* See draft-stevens-tcpca-spec-01 for discussion of the
404 * initialization of these values.
406 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
407 tp->snd_cwnd_clamp = ~0;
408 tp->mss_cache = TCP_MSS_DEFAULT;
409 u64_stats_init(&tp->syncp);
411 tp->reordering = sysctl_tcp_reordering;
412 tcp_enable_early_retrans(tp);
413 tcp_assign_congestion_control(sk);
417 sk->sk_state = TCP_CLOSE;
419 sk->sk_write_space = sk_stream_write_space;
420 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
422 icsk->icsk_sync_mss = tcp_sync_mss;
424 sk->sk_sndbuf = sysctl_tcp_wmem[1];
425 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
428 sock_update_memcg(sk);
429 sk_sockets_allocated_inc(sk);
432 EXPORT_SYMBOL(tcp_init_sock);
434 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb)
436 if (sk->sk_tsflags) {
437 struct skb_shared_info *shinfo = skb_shinfo(skb);
439 sock_tx_timestamp(sk, &shinfo->tx_flags);
440 if (shinfo->tx_flags & SKBTX_ANY_TSTAMP)
441 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
446 * Wait for a TCP event.
448 * Note that we don't need to lock the socket, as the upper poll layers
449 * take care of normal races (between the test and the event) and we don't
450 * go look at any of the socket buffers directly.
452 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
455 struct sock *sk = sock->sk;
456 const struct tcp_sock *tp = tcp_sk(sk);
459 sock_rps_record_flow(sk);
461 sock_poll_wait(file, sk_sleep(sk), wait);
463 state = sk_state_load(sk);
464 if (state == TCP_LISTEN)
465 return inet_csk_listen_poll(sk);
467 /* Socket is not locked. We are protected from async events
468 * by poll logic and correct handling of state changes
469 * made by other threads is impossible in any case.
475 * POLLHUP is certainly not done right. But poll() doesn't
476 * have a notion of HUP in just one direction, and for a
477 * socket the read side is more interesting.
479 * Some poll() documentation says that POLLHUP is incompatible
480 * with the POLLOUT/POLLWR flags, so somebody should check this
481 * all. But careful, it tends to be safer to return too many
482 * bits than too few, and you can easily break real applications
483 * if you don't tell them that something has hung up!
487 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
488 * our fs/select.c). It means that after we received EOF,
489 * poll always returns immediately, making impossible poll() on write()
490 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
491 * if and only if shutdown has been made in both directions.
492 * Actually, it is interesting to look how Solaris and DUX
493 * solve this dilemma. I would prefer, if POLLHUP were maskable,
494 * then we could set it on SND_SHUTDOWN. BTW examples given
495 * in Stevens' books assume exactly this behaviour, it explains
496 * why POLLHUP is incompatible with POLLOUT. --ANK
498 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
499 * blocking on fresh not-connected or disconnected socket. --ANK
501 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
503 if (sk->sk_shutdown & RCV_SHUTDOWN)
504 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
506 /* Connected or passive Fast Open socket? */
507 if (state != TCP_SYN_SENT &&
508 (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
509 int target = sock_rcvlowat(sk, 0, INT_MAX);
511 if (tp->urg_seq == tp->copied_seq &&
512 !sock_flag(sk, SOCK_URGINLINE) &&
516 if (tp->rcv_nxt - tp->copied_seq >= target)
517 mask |= POLLIN | POLLRDNORM;
519 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
520 if (sk_stream_is_writeable(sk)) {
521 mask |= POLLOUT | POLLWRNORM;
522 } else { /* send SIGIO later */
523 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
524 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
526 /* Race breaker. If space is freed after
527 * wspace test but before the flags are set,
528 * IO signal will be lost. Memory barrier
529 * pairs with the input side.
531 smp_mb__after_atomic();
532 if (sk_stream_is_writeable(sk))
533 mask |= POLLOUT | POLLWRNORM;
536 mask |= POLLOUT | POLLWRNORM;
538 if (tp->urg_data & TCP_URG_VALID)
541 /* This barrier is coupled with smp_wmb() in tcp_reset() */
543 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
548 EXPORT_SYMBOL(tcp_poll);
550 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
552 struct tcp_sock *tp = tcp_sk(sk);
558 if (sk->sk_state == TCP_LISTEN)
561 slow = lock_sock_fast(sk);
562 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
564 else if (sock_flag(sk, SOCK_URGINLINE) ||
566 before(tp->urg_seq, tp->copied_seq) ||
567 !before(tp->urg_seq, tp->rcv_nxt)) {
569 answ = tp->rcv_nxt - tp->copied_seq;
571 /* Subtract 1, if FIN was received */
572 if (answ && sock_flag(sk, SOCK_DONE))
575 answ = tp->urg_seq - tp->copied_seq;
576 unlock_sock_fast(sk, slow);
579 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
582 if (sk->sk_state == TCP_LISTEN)
585 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
588 answ = tp->write_seq - tp->snd_una;
591 if (sk->sk_state == TCP_LISTEN)
594 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
597 answ = tp->write_seq - tp->snd_nxt;
603 return put_user(answ, (int __user *)arg);
605 EXPORT_SYMBOL(tcp_ioctl);
607 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
609 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
610 tp->pushed_seq = tp->write_seq;
613 static inline bool forced_push(const struct tcp_sock *tp)
615 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
618 static void skb_entail(struct sock *sk, struct sk_buff *skb)
620 struct tcp_sock *tp = tcp_sk(sk);
621 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
624 tcb->seq = tcb->end_seq = tp->write_seq;
625 tcb->tcp_flags = TCPHDR_ACK;
627 __skb_header_release(skb);
628 tcp_add_write_queue_tail(sk, skb);
629 sk->sk_wmem_queued += skb->truesize;
630 sk_mem_charge(sk, skb->truesize);
631 if (tp->nonagle & TCP_NAGLE_PUSH)
632 tp->nonagle &= ~TCP_NAGLE_PUSH;
634 tcp_slow_start_after_idle_check(sk);
637 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
640 tp->snd_up = tp->write_seq;
643 /* If a not yet filled skb is pushed, do not send it if
644 * we have data packets in Qdisc or NIC queues :
645 * Because TX completion will happen shortly, it gives a chance
646 * to coalesce future sendmsg() payload into this skb, without
647 * need for a timer, and with no latency trade off.
648 * As packets containing data payload have a bigger truesize
649 * than pure acks (dataless) packets, the last checks prevent
650 * autocorking if we only have an ACK in Qdisc/NIC queues,
651 * or if TX completion was delayed after we processed ACK packet.
653 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
656 return skb->len < size_goal &&
657 sysctl_tcp_autocorking &&
658 skb != tcp_write_queue_head(sk) &&
659 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
662 static void tcp_push(struct sock *sk, int flags, int mss_now,
663 int nonagle, int size_goal)
665 struct tcp_sock *tp = tcp_sk(sk);
668 if (!tcp_send_head(sk))
671 skb = tcp_write_queue_tail(sk);
672 if (!(flags & MSG_MORE) || forced_push(tp))
673 tcp_mark_push(tp, skb);
675 tcp_mark_urg(tp, flags);
677 if (tcp_should_autocork(sk, skb, size_goal)) {
679 /* avoid atomic op if TSQ_THROTTLED bit is already set */
680 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
681 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
682 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
684 /* It is possible TX completion already happened
685 * before we set TSQ_THROTTLED.
687 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
691 if (flags & MSG_MORE)
692 nonagle = TCP_NAGLE_CORK;
694 __tcp_push_pending_frames(sk, mss_now, nonagle);
697 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
698 unsigned int offset, size_t len)
700 struct tcp_splice_state *tss = rd_desc->arg.data;
703 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
704 min(rd_desc->count, len), tss->flags,
707 rd_desc->count -= ret;
711 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
713 /* Store TCP splice context information in read_descriptor_t. */
714 read_descriptor_t rd_desc = {
719 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
723 * tcp_splice_read - splice data from TCP socket to a pipe
724 * @sock: socket to splice from
725 * @ppos: position (not valid)
726 * @pipe: pipe to splice to
727 * @len: number of bytes to splice
728 * @flags: splice modifier flags
731 * Will read pages from given socket and fill them into a pipe.
734 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
735 struct pipe_inode_info *pipe, size_t len,
738 struct sock *sk = sock->sk;
739 struct tcp_splice_state tss = {
748 sock_rps_record_flow(sk);
750 * We can't seek on a socket input
759 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
761 ret = __tcp_splice_read(sk, &tss);
767 if (sock_flag(sk, SOCK_DONE))
770 ret = sock_error(sk);
773 if (sk->sk_shutdown & RCV_SHUTDOWN)
775 if (sk->sk_state == TCP_CLOSE) {
777 * This occurs when user tries to read
778 * from never connected socket.
780 if (!sock_flag(sk, SOCK_DONE))
788 sk_wait_data(sk, &timeo, NULL);
789 if (signal_pending(current)) {
790 ret = sock_intr_errno(timeo);
803 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
804 (sk->sk_shutdown & RCV_SHUTDOWN) ||
805 signal_pending(current))
816 EXPORT_SYMBOL(tcp_splice_read);
818 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
823 /* The TCP header must be at least 32-bit aligned. */
824 size = ALIGN(size, 4);
826 if (unlikely(tcp_under_memory_pressure(sk)))
827 sk_mem_reclaim_partial(sk);
829 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
833 if (force_schedule) {
834 mem_scheduled = true;
835 sk_forced_mem_schedule(sk, skb->truesize);
837 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
839 if (likely(mem_scheduled)) {
840 skb_reserve(skb, sk->sk_prot->max_header);
842 * Make sure that we have exactly size bytes
843 * available to the caller, no more, no less.
845 skb->reserved_tailroom = skb->end - skb->tail - size;
850 sk->sk_prot->enter_memory_pressure(sk);
851 sk_stream_moderate_sndbuf(sk);
856 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
859 struct tcp_sock *tp = tcp_sk(sk);
860 u32 new_size_goal, size_goal;
862 if (!large_allowed || !sk_can_gso(sk))
865 /* Note : tcp_tso_autosize() will eventually split this later */
866 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
867 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
869 /* We try hard to avoid divides here */
870 size_goal = tp->gso_segs * mss_now;
871 if (unlikely(new_size_goal < size_goal ||
872 new_size_goal >= size_goal + mss_now)) {
873 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
874 sk->sk_gso_max_segs);
875 size_goal = tp->gso_segs * mss_now;
878 return max(size_goal, mss_now);
881 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
885 mss_now = tcp_current_mss(sk);
886 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
891 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
892 size_t size, int flags)
894 struct tcp_sock *tp = tcp_sk(sk);
895 int mss_now, size_goal;
898 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
900 /* Wait for a connection to finish. One exception is TCP Fast Open
901 * (passive side) where data is allowed to be sent before a connection
902 * is fully established.
904 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
905 !tcp_passive_fastopen(sk)) {
906 err = sk_stream_wait_connect(sk, &timeo);
911 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
913 mss_now = tcp_send_mss(sk, &size_goal, flags);
917 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
921 struct sk_buff *skb = tcp_write_queue_tail(sk);
925 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
927 if (!sk_stream_memory_free(sk))
928 goto wait_for_sndbuf;
930 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
931 skb_queue_empty(&sk->sk_write_queue));
933 goto wait_for_memory;
942 i = skb_shinfo(skb)->nr_frags;
943 can_coalesce = skb_can_coalesce(skb, i, page, offset);
944 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
945 tcp_mark_push(tp, skb);
948 if (!sk_wmem_schedule(sk, copy))
949 goto wait_for_memory;
952 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
955 skb_fill_page_desc(skb, i, page, offset, copy);
957 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
960 skb->data_len += copy;
961 skb->truesize += copy;
962 sk->sk_wmem_queued += copy;
963 sk_mem_charge(sk, copy);
964 skb->ip_summed = CHECKSUM_PARTIAL;
965 tp->write_seq += copy;
966 TCP_SKB_CB(skb)->end_seq += copy;
967 tcp_skb_pcount_set(skb, 0);
970 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
976 tcp_tx_timestamp(sk, skb);
980 if (skb->len < size_goal || (flags & MSG_OOB))
983 if (forced_push(tp)) {
984 tcp_mark_push(tp, skb);
985 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
986 } else if (skb == tcp_send_head(sk))
987 tcp_push_one(sk, mss_now);
991 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
993 tcp_push(sk, flags & ~MSG_MORE, mss_now,
994 TCP_NAGLE_PUSH, size_goal);
996 err = sk_stream_wait_memory(sk, &timeo);
1000 mss_now = tcp_send_mss(sk, &size_goal, flags);
1004 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
1005 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1012 /* make sure we wake any epoll edge trigger waiter */
1013 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1014 sk->sk_write_space(sk);
1015 return sk_stream_error(sk, flags, err);
1018 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1019 size_t size, int flags)
1023 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1024 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1025 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1029 res = do_tcp_sendpages(sk, page, offset, size, flags);
1033 EXPORT_SYMBOL(tcp_sendpage);
1035 static inline int select_size(const struct sock *sk, bool sg)
1037 const struct tcp_sock *tp = tcp_sk(sk);
1038 int tmp = tp->mss_cache;
1041 if (sk_can_gso(sk)) {
1042 /* Small frames wont use a full page:
1043 * Payload will immediately follow tcp header.
1045 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1047 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1049 if (tmp >= pgbreak &&
1050 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1058 void tcp_free_fastopen_req(struct tcp_sock *tp)
1060 if (tp->fastopen_req) {
1061 kfree(tp->fastopen_req);
1062 tp->fastopen_req = NULL;
1066 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1067 int *copied, size_t size)
1069 struct tcp_sock *tp = tcp_sk(sk);
1072 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1074 if (tp->fastopen_req)
1075 return -EALREADY; /* Another Fast Open is in progress */
1077 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1079 if (unlikely(!tp->fastopen_req))
1081 tp->fastopen_req->data = msg;
1082 tp->fastopen_req->size = size;
1084 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1085 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1086 msg->msg_namelen, flags);
1087 *copied = tp->fastopen_req->copied;
1088 tcp_free_fastopen_req(tp);
1092 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1094 struct tcp_sock *tp = tcp_sk(sk);
1095 struct sk_buff *skb;
1096 int flags, err, copied = 0;
1097 int mss_now = 0, size_goal, copied_syn = 0;
1103 flags = msg->msg_flags;
1104 if (flags & MSG_FASTOPEN) {
1105 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1106 if (err == -EINPROGRESS && copied_syn > 0)
1112 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1114 /* Wait for a connection to finish. One exception is TCP Fast Open
1115 * (passive side) where data is allowed to be sent before a connection
1116 * is fully established.
1118 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1119 !tcp_passive_fastopen(sk)) {
1120 err = sk_stream_wait_connect(sk, &timeo);
1125 if (unlikely(tp->repair)) {
1126 if (tp->repair_queue == TCP_RECV_QUEUE) {
1127 copied = tcp_send_rcvq(sk, msg, size);
1132 if (tp->repair_queue == TCP_NO_QUEUE)
1135 /* 'common' sending to sendq */
1138 /* This should be in poll */
1139 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1141 mss_now = tcp_send_mss(sk, &size_goal, flags);
1143 /* Ok commence sending. */
1147 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1150 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1152 while (msg_data_left(msg)) {
1154 int max = size_goal;
1156 skb = tcp_write_queue_tail(sk);
1157 if (tcp_send_head(sk)) {
1158 if (skb->ip_summed == CHECKSUM_NONE)
1160 copy = max - skb->len;
1165 /* Allocate new segment. If the interface is SG,
1166 * allocate skb fitting to single page.
1168 if (!sk_stream_memory_free(sk))
1169 goto wait_for_sndbuf;
1171 skb = sk_stream_alloc_skb(sk,
1172 select_size(sk, sg),
1174 skb_queue_empty(&sk->sk_write_queue));
1176 goto wait_for_memory;
1179 * Check whether we can use HW checksum.
1181 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1182 skb->ip_summed = CHECKSUM_PARTIAL;
1184 skb_entail(sk, skb);
1188 /* All packets are restored as if they have
1189 * already been sent. skb_mstamp isn't set to
1190 * avoid wrong rtt estimation.
1193 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1196 /* Try to append data to the end of skb. */
1197 if (copy > msg_data_left(msg))
1198 copy = msg_data_left(msg);
1200 /* Where to copy to? */
1201 if (skb_availroom(skb) > 0) {
1202 /* We have some space in skb head. Superb! */
1203 copy = min_t(int, copy, skb_availroom(skb));
1204 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1209 int i = skb_shinfo(skb)->nr_frags;
1210 struct page_frag *pfrag = sk_page_frag(sk);
1212 if (!sk_page_frag_refill(sk, pfrag))
1213 goto wait_for_memory;
1215 if (!skb_can_coalesce(skb, i, pfrag->page,
1217 if (i == MAX_SKB_FRAGS || !sg) {
1218 tcp_mark_push(tp, skb);
1224 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1226 if (!sk_wmem_schedule(sk, copy))
1227 goto wait_for_memory;
1229 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1236 /* Update the skb. */
1238 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1240 skb_fill_page_desc(skb, i, pfrag->page,
1241 pfrag->offset, copy);
1242 get_page(pfrag->page);
1244 pfrag->offset += copy;
1248 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1250 tp->write_seq += copy;
1251 TCP_SKB_CB(skb)->end_seq += copy;
1252 tcp_skb_pcount_set(skb, 0);
1255 if (!msg_data_left(msg)) {
1256 tcp_tx_timestamp(sk, skb);
1260 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1263 if (forced_push(tp)) {
1264 tcp_mark_push(tp, skb);
1265 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1266 } else if (skb == tcp_send_head(sk))
1267 tcp_push_one(sk, mss_now);
1271 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1274 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1275 TCP_NAGLE_PUSH, size_goal);
1277 err = sk_stream_wait_memory(sk, &timeo);
1281 mss_now = tcp_send_mss(sk, &size_goal, flags);
1286 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1289 return copied + copied_syn;
1293 tcp_unlink_write_queue(skb, sk);
1294 /* It is the one place in all of TCP, except connection
1295 * reset, where we can be unlinking the send_head.
1297 tcp_check_send_head(sk, skb);
1298 sk_wmem_free_skb(sk, skb);
1302 if (copied + copied_syn)
1305 err = sk_stream_error(sk, flags, err);
1306 /* make sure we wake any epoll edge trigger waiter */
1307 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1308 sk->sk_write_space(sk);
1312 EXPORT_SYMBOL(tcp_sendmsg);
1315 * Handle reading urgent data. BSD has very simple semantics for
1316 * this, no blocking and very strange errors 8)
1319 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1321 struct tcp_sock *tp = tcp_sk(sk);
1323 /* No URG data to read. */
1324 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1325 tp->urg_data == TCP_URG_READ)
1326 return -EINVAL; /* Yes this is right ! */
1328 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1331 if (tp->urg_data & TCP_URG_VALID) {
1333 char c = tp->urg_data;
1335 if (!(flags & MSG_PEEK))
1336 tp->urg_data = TCP_URG_READ;
1338 /* Read urgent data. */
1339 msg->msg_flags |= MSG_OOB;
1342 if (!(flags & MSG_TRUNC))
1343 err = memcpy_to_msg(msg, &c, 1);
1346 msg->msg_flags |= MSG_TRUNC;
1348 return err ? -EFAULT : len;
1351 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1354 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1355 * the available implementations agree in this case:
1356 * this call should never block, independent of the
1357 * blocking state of the socket.
1358 * Mike <pall@rz.uni-karlsruhe.de>
1363 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1365 struct sk_buff *skb;
1366 int copied = 0, err = 0;
1368 /* XXX -- need to support SO_PEEK_OFF */
1370 skb_queue_walk(&sk->sk_write_queue, skb) {
1371 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1378 return err ?: copied;
1381 /* Clean up the receive buffer for full frames taken by the user,
1382 * then send an ACK if necessary. COPIED is the number of bytes
1383 * tcp_recvmsg has given to the user so far, it speeds up the
1384 * calculation of whether or not we must ACK for the sake of
1387 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1389 struct tcp_sock *tp = tcp_sk(sk);
1390 bool time_to_ack = false;
1392 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1394 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1395 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1396 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1398 if (inet_csk_ack_scheduled(sk)) {
1399 const struct inet_connection_sock *icsk = inet_csk(sk);
1400 /* Delayed ACKs frequently hit locked sockets during bulk
1402 if (icsk->icsk_ack.blocked ||
1403 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1404 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1406 * If this read emptied read buffer, we send ACK, if
1407 * connection is not bidirectional, user drained
1408 * receive buffer and there was a small segment
1412 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1413 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1414 !icsk->icsk_ack.pingpong)) &&
1415 !atomic_read(&sk->sk_rmem_alloc)))
1419 /* We send an ACK if we can now advertise a non-zero window
1420 * which has been raised "significantly".
1422 * Even if window raised up to infinity, do not send window open ACK
1423 * in states, where we will not receive more. It is useless.
1425 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1426 __u32 rcv_window_now = tcp_receive_window(tp);
1428 /* Optimize, __tcp_select_window() is not cheap. */
1429 if (2*rcv_window_now <= tp->window_clamp) {
1430 __u32 new_window = __tcp_select_window(sk);
1432 /* Send ACK now, if this read freed lots of space
1433 * in our buffer. Certainly, new_window is new window.
1434 * We can advertise it now, if it is not less than current one.
1435 * "Lots" means "at least twice" here.
1437 if (new_window && new_window >= 2 * rcv_window_now)
1445 static void tcp_prequeue_process(struct sock *sk)
1447 struct sk_buff *skb;
1448 struct tcp_sock *tp = tcp_sk(sk);
1450 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1452 /* RX process wants to run with disabled BHs, though it is not
1455 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1456 sk_backlog_rcv(sk, skb);
1459 /* Clear memory counter. */
1460 tp->ucopy.memory = 0;
1463 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1465 struct sk_buff *skb;
1468 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1469 offset = seq - TCP_SKB_CB(skb)->seq;
1470 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1472 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1476 /* This looks weird, but this can happen if TCP collapsing
1477 * splitted a fat GRO packet, while we released socket lock
1478 * in skb_splice_bits()
1480 sk_eat_skb(sk, skb);
1486 * This routine provides an alternative to tcp_recvmsg() for routines
1487 * that would like to handle copying from skbuffs directly in 'sendfile'
1490 * - It is assumed that the socket was locked by the caller.
1491 * - The routine does not block.
1492 * - At present, there is no support for reading OOB data
1493 * or for 'peeking' the socket using this routine
1494 * (although both would be easy to implement).
1496 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1497 sk_read_actor_t recv_actor)
1499 struct sk_buff *skb;
1500 struct tcp_sock *tp = tcp_sk(sk);
1501 u32 seq = tp->copied_seq;
1505 if (sk->sk_state == TCP_LISTEN)
1507 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1508 if (offset < skb->len) {
1512 len = skb->len - offset;
1513 /* Stop reading if we hit a patch of urgent data */
1515 u32 urg_offset = tp->urg_seq - seq;
1516 if (urg_offset < len)
1521 used = recv_actor(desc, skb, offset, len);
1526 } else if (used <= len) {
1531 /* If recv_actor drops the lock (e.g. TCP splice
1532 * receive) the skb pointer might be invalid when
1533 * getting here: tcp_collapse might have deleted it
1534 * while aggregating skbs from the socket queue.
1536 skb = tcp_recv_skb(sk, seq - 1, &offset);
1539 /* TCP coalescing might have appended data to the skb.
1540 * Try to splice more frags
1542 if (offset + 1 != skb->len)
1545 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1546 sk_eat_skb(sk, skb);
1550 sk_eat_skb(sk, skb);
1553 tp->copied_seq = seq;
1555 tp->copied_seq = seq;
1557 tcp_rcv_space_adjust(sk);
1559 /* Clean up data we have read: This will do ACK frames. */
1561 tcp_recv_skb(sk, seq, &offset);
1562 tcp_cleanup_rbuf(sk, copied);
1566 EXPORT_SYMBOL(tcp_read_sock);
1569 * This routine copies from a sock struct into the user buffer.
1571 * Technical note: in 2.3 we work on _locked_ socket, so that
1572 * tricks with *seq access order and skb->users are not required.
1573 * Probably, code can be easily improved even more.
1576 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1577 int flags, int *addr_len)
1579 struct tcp_sock *tp = tcp_sk(sk);
1585 int target; /* Read at least this many bytes */
1587 struct task_struct *user_recv = NULL;
1588 struct sk_buff *skb, *last;
1591 if (unlikely(flags & MSG_ERRQUEUE))
1592 return inet_recv_error(sk, msg, len, addr_len);
1594 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1595 (sk->sk_state == TCP_ESTABLISHED))
1596 sk_busy_loop(sk, nonblock);
1601 if (sk->sk_state == TCP_LISTEN)
1604 timeo = sock_rcvtimeo(sk, nonblock);
1606 /* Urgent data needs to be handled specially. */
1607 if (flags & MSG_OOB)
1610 if (unlikely(tp->repair)) {
1612 if (!(flags & MSG_PEEK))
1615 if (tp->repair_queue == TCP_SEND_QUEUE)
1619 if (tp->repair_queue == TCP_NO_QUEUE)
1622 /* 'common' recv queue MSG_PEEK-ing */
1625 seq = &tp->copied_seq;
1626 if (flags & MSG_PEEK) {
1627 peek_seq = tp->copied_seq;
1631 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
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 last = skb_peek_tail(&sk->sk_receive_queue);
1649 skb_queue_walk(&sk->sk_receive_queue, skb) {
1651 /* Now that we have two receive queues this
1654 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1655 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1656 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1660 offset = *seq - TCP_SKB_CB(skb)->seq;
1661 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1663 if (offset < skb->len)
1665 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1667 WARN(!(flags & MSG_PEEK),
1668 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1669 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1672 /* Well, if we have backlog, try to process it now yet. */
1674 if (copied >= target && !sk->sk_backlog.tail)
1679 sk->sk_state == TCP_CLOSE ||
1680 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1682 signal_pending(current))
1685 if (sock_flag(sk, SOCK_DONE))
1689 copied = sock_error(sk);
1693 if (sk->sk_shutdown & RCV_SHUTDOWN)
1696 if (sk->sk_state == TCP_CLOSE) {
1697 if (!sock_flag(sk, SOCK_DONE)) {
1698 /* This occurs when user tries to read
1699 * from never connected socket.
1712 if (signal_pending(current)) {
1713 copied = sock_intr_errno(timeo);
1718 tcp_cleanup_rbuf(sk, copied);
1720 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1721 /* Install new reader */
1722 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1723 user_recv = current;
1724 tp->ucopy.task = user_recv;
1725 tp->ucopy.msg = msg;
1728 tp->ucopy.len = len;
1730 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1731 !(flags & (MSG_PEEK | MSG_TRUNC)));
1733 /* Ugly... If prequeue is not empty, we have to
1734 * process it before releasing socket, otherwise
1735 * order will be broken at second iteration.
1736 * More elegant solution is required!!!
1738 * Look: we have the following (pseudo)queues:
1740 * 1. packets in flight
1745 * Each queue can be processed only if the next ones
1746 * are empty. At this point we have empty receive_queue.
1747 * But prequeue _can_ be not empty after 2nd iteration,
1748 * when we jumped to start of loop because backlog
1749 * processing added something to receive_queue.
1750 * We cannot release_sock(), because backlog contains
1751 * packets arrived _after_ prequeued ones.
1753 * Shortly, algorithm is clear --- to process all
1754 * the queues in order. We could make it more directly,
1755 * requeueing packets from backlog to prequeue, if
1756 * is not empty. It is more elegant, but eats cycles,
1759 if (!skb_queue_empty(&tp->ucopy.prequeue))
1762 /* __ Set realtime policy in scheduler __ */
1765 if (copied >= target) {
1766 /* Do not sleep, just process backlog. */
1770 sk_wait_data(sk, &timeo, last);
1776 /* __ Restore normal policy in scheduler __ */
1778 chunk = len - tp->ucopy.len;
1780 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1785 if (tp->rcv_nxt == tp->copied_seq &&
1786 !skb_queue_empty(&tp->ucopy.prequeue)) {
1788 tcp_prequeue_process(sk);
1790 chunk = len - tp->ucopy.len;
1792 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1798 if ((flags & MSG_PEEK) &&
1799 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1800 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1802 task_pid_nr(current));
1803 peek_seq = tp->copied_seq;
1808 /* Ok so how much can we use? */
1809 used = skb->len - offset;
1813 /* Do we have urgent data here? */
1815 u32 urg_offset = tp->urg_seq - *seq;
1816 if (urg_offset < used) {
1818 if (!sock_flag(sk, SOCK_URGINLINE)) {
1831 if (!(flags & MSG_TRUNC)) {
1832 err = skb_copy_datagram_msg(skb, offset, msg, used);
1834 /* Exception. Bailout! */
1845 tcp_rcv_space_adjust(sk);
1848 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1850 tcp_fast_path_check(sk);
1852 if (used + offset < skb->len)
1855 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1857 if (!(flags & MSG_PEEK))
1858 sk_eat_skb(sk, skb);
1862 /* Process the FIN. */
1864 if (!(flags & MSG_PEEK))
1865 sk_eat_skb(sk, skb);
1870 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1873 tp->ucopy.len = copied > 0 ? len : 0;
1875 tcp_prequeue_process(sk);
1877 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1878 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1884 tp->ucopy.task = NULL;
1888 /* According to UNIX98, msg_name/msg_namelen are ignored
1889 * on connected socket. I was just happy when found this 8) --ANK
1892 /* Clean up data we have read: This will do ACK frames. */
1893 tcp_cleanup_rbuf(sk, copied);
1903 err = tcp_recv_urg(sk, msg, len, flags);
1907 err = tcp_peek_sndq(sk, msg, len);
1910 EXPORT_SYMBOL(tcp_recvmsg);
1912 void tcp_set_state(struct sock *sk, int state)
1914 int oldstate = sk->sk_state;
1917 case TCP_ESTABLISHED:
1918 if (oldstate != TCP_ESTABLISHED)
1919 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1923 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1924 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1926 sk->sk_prot->unhash(sk);
1927 if (inet_csk(sk)->icsk_bind_hash &&
1928 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1932 if (oldstate == TCP_ESTABLISHED)
1933 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1936 /* Change state AFTER socket is unhashed to avoid closed
1937 * socket sitting in hash tables.
1939 sk_state_store(sk, state);
1942 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1945 EXPORT_SYMBOL_GPL(tcp_set_state);
1948 * State processing on a close. This implements the state shift for
1949 * sending our FIN frame. Note that we only send a FIN for some
1950 * states. A shutdown() may have already sent the FIN, or we may be
1954 static const unsigned char new_state[16] = {
1955 /* current state: new state: action: */
1956 [0 /* (Invalid) */] = TCP_CLOSE,
1957 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1958 [TCP_SYN_SENT] = TCP_CLOSE,
1959 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1960 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1961 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1962 [TCP_TIME_WAIT] = TCP_CLOSE,
1963 [TCP_CLOSE] = TCP_CLOSE,
1964 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
1965 [TCP_LAST_ACK] = TCP_LAST_ACK,
1966 [TCP_LISTEN] = TCP_CLOSE,
1967 [TCP_CLOSING] = TCP_CLOSING,
1968 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
1971 static int tcp_close_state(struct sock *sk)
1973 int next = (int)new_state[sk->sk_state];
1974 int ns = next & TCP_STATE_MASK;
1976 tcp_set_state(sk, ns);
1978 return next & TCP_ACTION_FIN;
1982 * Shutdown the sending side of a connection. Much like close except
1983 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1986 void tcp_shutdown(struct sock *sk, int how)
1988 /* We need to grab some memory, and put together a FIN,
1989 * and then put it into the queue to be sent.
1990 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1992 if (!(how & SEND_SHUTDOWN))
1995 /* If we've already sent a FIN, or it's a closed state, skip this. */
1996 if ((1 << sk->sk_state) &
1997 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1998 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1999 /* Clear out any half completed packets. FIN if needed. */
2000 if (tcp_close_state(sk))
2004 EXPORT_SYMBOL(tcp_shutdown);
2006 bool tcp_check_oom(struct sock *sk, int shift)
2008 bool too_many_orphans, out_of_socket_memory;
2010 too_many_orphans = tcp_too_many_orphans(sk, shift);
2011 out_of_socket_memory = tcp_out_of_memory(sk);
2013 if (too_many_orphans)
2014 net_info_ratelimited("too many orphaned sockets\n");
2015 if (out_of_socket_memory)
2016 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2017 return too_many_orphans || out_of_socket_memory;
2020 void tcp_close(struct sock *sk, long timeout)
2022 struct sk_buff *skb;
2023 int data_was_unread = 0;
2027 sk->sk_shutdown = SHUTDOWN_MASK;
2029 if (sk->sk_state == TCP_LISTEN) {
2030 tcp_set_state(sk, TCP_CLOSE);
2033 inet_csk_listen_stop(sk);
2035 goto adjudge_to_death;
2038 /* We need to flush the recv. buffs. We do this only on the
2039 * descriptor close, not protocol-sourced closes, because the
2040 * reader process may not have drained the data yet!
2042 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2043 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2045 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2047 data_was_unread += len;
2053 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2054 if (sk->sk_state == TCP_CLOSE)
2055 goto adjudge_to_death;
2057 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2058 * data was lost. To witness the awful effects of the old behavior of
2059 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2060 * GET in an FTP client, suspend the process, wait for the client to
2061 * advertise a zero window, then kill -9 the FTP client, wheee...
2062 * Note: timeout is always zero in such a case.
2064 if (unlikely(tcp_sk(sk)->repair)) {
2065 sk->sk_prot->disconnect(sk, 0);
2066 } else if (data_was_unread) {
2067 /* Unread data was tossed, zap the connection. */
2068 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2069 tcp_set_state(sk, TCP_CLOSE);
2070 tcp_send_active_reset(sk, sk->sk_allocation);
2071 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2072 /* Check zero linger _after_ checking for unread data. */
2073 sk->sk_prot->disconnect(sk, 0);
2074 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2075 } else if (tcp_close_state(sk)) {
2076 /* We FIN if the application ate all the data before
2077 * zapping the connection.
2080 /* RED-PEN. Formally speaking, we have broken TCP state
2081 * machine. State transitions:
2083 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2084 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2085 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2087 * are legal only when FIN has been sent (i.e. in window),
2088 * rather than queued out of window. Purists blame.
2090 * F.e. "RFC state" is ESTABLISHED,
2091 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2093 * The visible declinations are that sometimes
2094 * we enter time-wait state, when it is not required really
2095 * (harmless), do not send active resets, when they are
2096 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2097 * they look as CLOSING or LAST_ACK for Linux)
2098 * Probably, I missed some more holelets.
2100 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2101 * in a single packet! (May consider it later but will
2102 * probably need API support or TCP_CORK SYN-ACK until
2103 * data is written and socket is closed.)
2108 sk_stream_wait_close(sk, timeout);
2111 state = sk->sk_state;
2115 /* It is the last release_sock in its life. It will remove backlog. */
2119 /* Now socket is owned by kernel and we acquire BH lock
2120 to finish close. No need to check for user refs.
2124 WARN_ON(sock_owned_by_user(sk));
2126 percpu_counter_inc(sk->sk_prot->orphan_count);
2128 /* Have we already been destroyed by a softirq or backlog? */
2129 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2132 /* This is a (useful) BSD violating of the RFC. There is a
2133 * problem with TCP as specified in that the other end could
2134 * keep a socket open forever with no application left this end.
2135 * We use a 1 minute timeout (about the same as BSD) then kill
2136 * our end. If they send after that then tough - BUT: long enough
2137 * that we won't make the old 4*rto = almost no time - whoops
2140 * Nope, it was not mistake. It is really desired behaviour
2141 * f.e. on http servers, when such sockets are useless, but
2142 * consume significant resources. Let's do it with special
2143 * linger2 option. --ANK
2146 if (sk->sk_state == TCP_FIN_WAIT2) {
2147 struct tcp_sock *tp = tcp_sk(sk);
2148 if (tp->linger2 < 0) {
2149 tcp_set_state(sk, TCP_CLOSE);
2150 tcp_send_active_reset(sk, GFP_ATOMIC);
2151 NET_INC_STATS_BH(sock_net(sk),
2152 LINUX_MIB_TCPABORTONLINGER);
2154 const int tmo = tcp_fin_time(sk);
2156 if (tmo > TCP_TIMEWAIT_LEN) {
2157 inet_csk_reset_keepalive_timer(sk,
2158 tmo - TCP_TIMEWAIT_LEN);
2160 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2165 if (sk->sk_state != TCP_CLOSE) {
2167 if (tcp_check_oom(sk, 0)) {
2168 tcp_set_state(sk, TCP_CLOSE);
2169 tcp_send_active_reset(sk, GFP_ATOMIC);
2170 NET_INC_STATS_BH(sock_net(sk),
2171 LINUX_MIB_TCPABORTONMEMORY);
2175 if (sk->sk_state == TCP_CLOSE) {
2176 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2177 /* We could get here with a non-NULL req if the socket is
2178 * aborted (e.g., closed with unread data) before 3WHS
2182 reqsk_fastopen_remove(sk, req, false);
2183 inet_csk_destroy_sock(sk);
2185 /* Otherwise, socket is reprieved until protocol close. */
2192 EXPORT_SYMBOL(tcp_close);
2194 /* These states need RST on ABORT according to RFC793 */
2196 static inline bool tcp_need_reset(int state)
2198 return (1 << state) &
2199 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2200 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2203 int tcp_disconnect(struct sock *sk, int flags)
2205 struct inet_sock *inet = inet_sk(sk);
2206 struct inet_connection_sock *icsk = inet_csk(sk);
2207 struct tcp_sock *tp = tcp_sk(sk);
2209 int old_state = sk->sk_state;
2211 if (old_state != TCP_CLOSE)
2212 tcp_set_state(sk, TCP_CLOSE);
2214 /* ABORT function of RFC793 */
2215 if (old_state == TCP_LISTEN) {
2216 inet_csk_listen_stop(sk);
2217 } else if (unlikely(tp->repair)) {
2218 sk->sk_err = ECONNABORTED;
2219 } else if (tcp_need_reset(old_state) ||
2220 (tp->snd_nxt != tp->write_seq &&
2221 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2222 /* The last check adjusts for discrepancy of Linux wrt. RFC
2225 tcp_send_active_reset(sk, gfp_any());
2226 sk->sk_err = ECONNRESET;
2227 } else if (old_state == TCP_SYN_SENT)
2228 sk->sk_err = ECONNRESET;
2230 tcp_clear_xmit_timers(sk);
2231 __skb_queue_purge(&sk->sk_receive_queue);
2232 tcp_write_queue_purge(sk);
2233 __skb_queue_purge(&tp->out_of_order_queue);
2235 inet->inet_dport = 0;
2237 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2238 inet_reset_saddr(sk);
2240 sk->sk_shutdown = 0;
2241 sock_reset_flag(sk, SOCK_DONE);
2243 tp->write_seq += tp->max_window + 2;
2244 if (tp->write_seq == 0)
2246 icsk->icsk_backoff = 0;
2248 icsk->icsk_probes_out = 0;
2249 tp->packets_out = 0;
2250 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2251 tp->snd_cwnd_cnt = 0;
2252 tp->window_clamp = 0;
2253 tcp_set_ca_state(sk, TCP_CA_Open);
2254 tcp_clear_retrans(tp);
2255 inet_csk_delack_init(sk);
2256 tcp_init_send_head(sk);
2257 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2260 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2262 sk->sk_error_report(sk);
2265 EXPORT_SYMBOL(tcp_disconnect);
2267 static inline bool tcp_can_repair_sock(const struct sock *sk)
2269 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2270 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2273 static int tcp_repair_options_est(struct tcp_sock *tp,
2274 struct tcp_repair_opt __user *optbuf, unsigned int len)
2276 struct tcp_repair_opt opt;
2278 while (len >= sizeof(opt)) {
2279 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2285 switch (opt.opt_code) {
2287 tp->rx_opt.mss_clamp = opt.opt_val;
2291 u16 snd_wscale = opt.opt_val & 0xFFFF;
2292 u16 rcv_wscale = opt.opt_val >> 16;
2294 if (snd_wscale > 14 || rcv_wscale > 14)
2297 tp->rx_opt.snd_wscale = snd_wscale;
2298 tp->rx_opt.rcv_wscale = rcv_wscale;
2299 tp->rx_opt.wscale_ok = 1;
2302 case TCPOPT_SACK_PERM:
2303 if (opt.opt_val != 0)
2306 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2307 if (sysctl_tcp_fack)
2308 tcp_enable_fack(tp);
2310 case TCPOPT_TIMESTAMP:
2311 if (opt.opt_val != 0)
2314 tp->rx_opt.tstamp_ok = 1;
2323 * Socket option code for TCP.
2325 static int do_tcp_setsockopt(struct sock *sk, int level,
2326 int optname, char __user *optval, unsigned int optlen)
2328 struct tcp_sock *tp = tcp_sk(sk);
2329 struct inet_connection_sock *icsk = inet_csk(sk);
2333 /* These are data/string values, all the others are ints */
2335 case TCP_CONGESTION: {
2336 char name[TCP_CA_NAME_MAX];
2341 val = strncpy_from_user(name, optval,
2342 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2348 err = tcp_set_congestion_control(sk, name);
2357 if (optlen < sizeof(int))
2360 if (get_user(val, (int __user *)optval))
2367 /* Values greater than interface MTU won't take effect. However
2368 * at the point when this call is done we typically don't yet
2369 * know which interface is going to be used */
2370 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2374 tp->rx_opt.user_mss = val;
2379 /* TCP_NODELAY is weaker than TCP_CORK, so that
2380 * this option on corked socket is remembered, but
2381 * it is not activated until cork is cleared.
2383 * However, when TCP_NODELAY is set we make
2384 * an explicit push, which overrides even TCP_CORK
2385 * for currently queued segments.
2387 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2388 tcp_push_pending_frames(sk);
2390 tp->nonagle &= ~TCP_NAGLE_OFF;
2394 case TCP_THIN_LINEAR_TIMEOUTS:
2395 if (val < 0 || val > 1)
2401 case TCP_THIN_DUPACK:
2402 if (val < 0 || val > 1)
2405 tp->thin_dupack = val;
2406 if (tp->thin_dupack)
2407 tcp_disable_early_retrans(tp);
2412 if (!tcp_can_repair_sock(sk))
2414 else if (val == 1) {
2416 sk->sk_reuse = SK_FORCE_REUSE;
2417 tp->repair_queue = TCP_NO_QUEUE;
2418 } else if (val == 0) {
2420 sk->sk_reuse = SK_NO_REUSE;
2421 tcp_send_window_probe(sk);
2427 case TCP_REPAIR_QUEUE:
2430 else if (val < TCP_QUEUES_NR)
2431 tp->repair_queue = val;
2437 if (sk->sk_state != TCP_CLOSE)
2439 else if (tp->repair_queue == TCP_SEND_QUEUE)
2440 tp->write_seq = val;
2441 else if (tp->repair_queue == TCP_RECV_QUEUE)
2447 case TCP_REPAIR_OPTIONS:
2450 else if (sk->sk_state == TCP_ESTABLISHED)
2451 err = tcp_repair_options_est(tp,
2452 (struct tcp_repair_opt __user *)optval,
2459 /* When set indicates to always queue non-full frames.
2460 * Later the user clears this option and we transmit
2461 * any pending partial frames in the queue. This is
2462 * meant to be used alongside sendfile() to get properly
2463 * filled frames when the user (for example) must write
2464 * out headers with a write() call first and then use
2465 * sendfile to send out the data parts.
2467 * TCP_CORK can be set together with TCP_NODELAY and it is
2468 * stronger than TCP_NODELAY.
2471 tp->nonagle |= TCP_NAGLE_CORK;
2473 tp->nonagle &= ~TCP_NAGLE_CORK;
2474 if (tp->nonagle&TCP_NAGLE_OFF)
2475 tp->nonagle |= TCP_NAGLE_PUSH;
2476 tcp_push_pending_frames(sk);
2481 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2484 tp->keepalive_time = val * HZ;
2485 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2486 !((1 << sk->sk_state) &
2487 (TCPF_CLOSE | TCPF_LISTEN))) {
2488 u32 elapsed = keepalive_time_elapsed(tp);
2489 if (tp->keepalive_time > elapsed)
2490 elapsed = tp->keepalive_time - elapsed;
2493 inet_csk_reset_keepalive_timer(sk, elapsed);
2498 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2501 tp->keepalive_intvl = val * HZ;
2504 if (val < 1 || val > MAX_TCP_KEEPCNT)
2507 tp->keepalive_probes = val;
2510 if (val < 1 || val > MAX_TCP_SYNCNT)
2513 icsk->icsk_syn_retries = val;
2517 if (val < 0 || val > 1)
2526 else if (val > sysctl_tcp_fin_timeout / HZ)
2529 tp->linger2 = val * HZ;
2532 case TCP_DEFER_ACCEPT:
2533 /* Translate value in seconds to number of retransmits */
2534 icsk->icsk_accept_queue.rskq_defer_accept =
2535 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2539 case TCP_WINDOW_CLAMP:
2541 if (sk->sk_state != TCP_CLOSE) {
2545 tp->window_clamp = 0;
2547 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2548 SOCK_MIN_RCVBUF / 2 : val;
2553 icsk->icsk_ack.pingpong = 1;
2555 icsk->icsk_ack.pingpong = 0;
2556 if ((1 << sk->sk_state) &
2557 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2558 inet_csk_ack_scheduled(sk)) {
2559 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2560 tcp_cleanup_rbuf(sk, 1);
2562 icsk->icsk_ack.pingpong = 1;
2567 #ifdef CONFIG_TCP_MD5SIG
2569 /* Read the IP->Key mappings from userspace */
2570 err = tp->af_specific->md5_parse(sk, optval, optlen);
2573 case TCP_USER_TIMEOUT:
2574 /* Cap the max time in ms TCP will retry or probe the window
2575 * before giving up and aborting (ETIMEDOUT) a connection.
2580 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2584 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2586 tcp_fastopen_init_key_once(true);
2588 fastopen_queue_tune(sk, val);
2597 tp->tsoffset = val - tcp_time_stamp;
2599 case TCP_NOTSENT_LOWAT:
2600 tp->notsent_lowat = val;
2601 sk->sk_write_space(sk);
2612 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2613 unsigned int optlen)
2615 const struct inet_connection_sock *icsk = inet_csk(sk);
2617 if (level != SOL_TCP)
2618 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2620 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2622 EXPORT_SYMBOL(tcp_setsockopt);
2624 #ifdef CONFIG_COMPAT
2625 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2626 char __user *optval, unsigned int optlen)
2628 if (level != SOL_TCP)
2629 return inet_csk_compat_setsockopt(sk, level, optname,
2631 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2633 EXPORT_SYMBOL(compat_tcp_setsockopt);
2636 /* Return information about state of tcp endpoint in API format. */
2637 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2639 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2640 const struct inet_connection_sock *icsk = inet_csk(sk);
2641 u32 now = tcp_time_stamp;
2645 memset(info, 0, sizeof(*info));
2646 if (sk->sk_type != SOCK_STREAM)
2649 info->tcpi_state = sk_state_load(sk);
2651 info->tcpi_ca_state = icsk->icsk_ca_state;
2652 info->tcpi_retransmits = icsk->icsk_retransmits;
2653 info->tcpi_probes = icsk->icsk_probes_out;
2654 info->tcpi_backoff = icsk->icsk_backoff;
2656 if (tp->rx_opt.tstamp_ok)
2657 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2658 if (tcp_is_sack(tp))
2659 info->tcpi_options |= TCPI_OPT_SACK;
2660 if (tp->rx_opt.wscale_ok) {
2661 info->tcpi_options |= TCPI_OPT_WSCALE;
2662 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2663 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2666 if (tp->ecn_flags & TCP_ECN_OK)
2667 info->tcpi_options |= TCPI_OPT_ECN;
2668 if (tp->ecn_flags & TCP_ECN_SEEN)
2669 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2670 if (tp->syn_data_acked)
2671 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2673 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2674 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2675 info->tcpi_snd_mss = tp->mss_cache;
2676 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2678 if (info->tcpi_state == TCP_LISTEN) {
2679 info->tcpi_unacked = sk->sk_ack_backlog;
2680 info->tcpi_sacked = sk->sk_max_ack_backlog;
2682 info->tcpi_unacked = tp->packets_out;
2683 info->tcpi_sacked = tp->sacked_out;
2685 info->tcpi_lost = tp->lost_out;
2686 info->tcpi_retrans = tp->retrans_out;
2687 info->tcpi_fackets = tp->fackets_out;
2689 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2690 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2691 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2693 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2694 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2695 info->tcpi_rtt = tp->srtt_us >> 3;
2696 info->tcpi_rttvar = tp->mdev_us >> 2;
2697 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2698 info->tcpi_snd_cwnd = tp->snd_cwnd;
2699 info->tcpi_advmss = tp->advmss;
2700 info->tcpi_reordering = tp->reordering;
2702 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2703 info->tcpi_rcv_space = tp->rcvq_space.space;
2705 info->tcpi_total_retrans = tp->total_retrans;
2707 rate = READ_ONCE(sk->sk_pacing_rate);
2708 info->tcpi_pacing_rate = rate != ~0U ? rate : ~0ULL;
2710 rate = READ_ONCE(sk->sk_max_pacing_rate);
2711 info->tcpi_max_pacing_rate = rate != ~0U ? rate : ~0ULL;
2714 start = u64_stats_fetch_begin_irq(&tp->syncp);
2715 info->tcpi_bytes_acked = tp->bytes_acked;
2716 info->tcpi_bytes_received = tp->bytes_received;
2717 } while (u64_stats_fetch_retry_irq(&tp->syncp, start));
2718 info->tcpi_segs_out = tp->segs_out;
2719 info->tcpi_segs_in = tp->segs_in;
2721 EXPORT_SYMBOL_GPL(tcp_get_info);
2723 static int do_tcp_getsockopt(struct sock *sk, int level,
2724 int optname, char __user *optval, int __user *optlen)
2726 struct inet_connection_sock *icsk = inet_csk(sk);
2727 struct tcp_sock *tp = tcp_sk(sk);
2730 if (get_user(len, optlen))
2733 len = min_t(unsigned int, len, sizeof(int));
2740 val = tp->mss_cache;
2741 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2742 val = tp->rx_opt.user_mss;
2744 val = tp->rx_opt.mss_clamp;
2747 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2750 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2753 val = keepalive_time_when(tp) / HZ;
2756 val = keepalive_intvl_when(tp) / HZ;
2759 val = keepalive_probes(tp);
2762 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2767 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2769 case TCP_DEFER_ACCEPT:
2770 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2771 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2773 case TCP_WINDOW_CLAMP:
2774 val = tp->window_clamp;
2777 struct tcp_info info;
2779 if (get_user(len, optlen))
2782 tcp_get_info(sk, &info);
2784 len = min_t(unsigned int, len, sizeof(info));
2785 if (put_user(len, optlen))
2787 if (copy_to_user(optval, &info, len))
2792 const struct tcp_congestion_ops *ca_ops;
2793 union tcp_cc_info info;
2797 if (get_user(len, optlen))
2800 ca_ops = icsk->icsk_ca_ops;
2801 if (ca_ops && ca_ops->get_info)
2802 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2804 len = min_t(unsigned int, len, sz);
2805 if (put_user(len, optlen))
2807 if (copy_to_user(optval, &info, len))
2812 val = !icsk->icsk_ack.pingpong;
2815 case TCP_CONGESTION:
2816 if (get_user(len, optlen))
2818 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2819 if (put_user(len, optlen))
2821 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2825 case TCP_THIN_LINEAR_TIMEOUTS:
2828 case TCP_THIN_DUPACK:
2829 val = tp->thin_dupack;
2836 case TCP_REPAIR_QUEUE:
2838 val = tp->repair_queue;
2844 if (tp->repair_queue == TCP_SEND_QUEUE)
2845 val = tp->write_seq;
2846 else if (tp->repair_queue == TCP_RECV_QUEUE)
2852 case TCP_USER_TIMEOUT:
2853 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2857 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
2861 val = tcp_time_stamp + tp->tsoffset;
2863 case TCP_NOTSENT_LOWAT:
2864 val = tp->notsent_lowat;
2869 case TCP_SAVED_SYN: {
2870 if (get_user(len, optlen))
2874 if (tp->saved_syn) {
2875 if (len < tp->saved_syn[0]) {
2876 if (put_user(tp->saved_syn[0], optlen)) {
2883 len = tp->saved_syn[0];
2884 if (put_user(len, optlen)) {
2888 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
2892 tcp_saved_syn_free(tp);
2897 if (put_user(len, optlen))
2903 return -ENOPROTOOPT;
2906 if (put_user(len, optlen))
2908 if (copy_to_user(optval, &val, len))
2913 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2916 struct inet_connection_sock *icsk = inet_csk(sk);
2918 if (level != SOL_TCP)
2919 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2921 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2923 EXPORT_SYMBOL(tcp_getsockopt);
2925 #ifdef CONFIG_COMPAT
2926 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2927 char __user *optval, int __user *optlen)
2929 if (level != SOL_TCP)
2930 return inet_csk_compat_getsockopt(sk, level, optname,
2932 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2934 EXPORT_SYMBOL(compat_tcp_getsockopt);
2937 #ifdef CONFIG_TCP_MD5SIG
2938 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2939 static DEFINE_MUTEX(tcp_md5sig_mutex);
2940 static bool tcp_md5sig_pool_populated = false;
2942 static void __tcp_alloc_md5sig_pool(void)
2946 for_each_possible_cpu(cpu) {
2947 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2948 struct crypto_hash *hash;
2950 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2951 if (IS_ERR_OR_NULL(hash))
2953 per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
2956 /* before setting tcp_md5sig_pool_populated, we must commit all writes
2957 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
2960 tcp_md5sig_pool_populated = true;
2963 bool tcp_alloc_md5sig_pool(void)
2965 if (unlikely(!tcp_md5sig_pool_populated)) {
2966 mutex_lock(&tcp_md5sig_mutex);
2968 if (!tcp_md5sig_pool_populated)
2969 __tcp_alloc_md5sig_pool();
2971 mutex_unlock(&tcp_md5sig_mutex);
2973 return tcp_md5sig_pool_populated;
2975 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2979 * tcp_get_md5sig_pool - get md5sig_pool for this user
2981 * We use percpu structure, so if we succeed, we exit with preemption
2982 * and BH disabled, to make sure another thread or softirq handling
2983 * wont try to get same context.
2985 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2989 if (tcp_md5sig_pool_populated) {
2990 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
2992 return this_cpu_ptr(&tcp_md5sig_pool);
2997 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2999 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3000 const struct tcphdr *th)
3002 struct scatterlist sg;
3006 /* We are not allowed to change tcphdr, make a local copy */
3007 memcpy(&hdr, th, sizeof(hdr));
3010 /* options aren't included in the hash */
3011 sg_init_one(&sg, &hdr, sizeof(hdr));
3012 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3015 EXPORT_SYMBOL(tcp_md5_hash_header);
3017 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3018 const struct sk_buff *skb, unsigned int header_len)
3020 struct scatterlist sg;
3021 const struct tcphdr *tp = tcp_hdr(skb);
3022 struct hash_desc *desc = &hp->md5_desc;
3024 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3025 skb_headlen(skb) - header_len : 0;
3026 const struct skb_shared_info *shi = skb_shinfo(skb);
3027 struct sk_buff *frag_iter;
3029 sg_init_table(&sg, 1);
3031 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3032 if (crypto_hash_update(desc, &sg, head_data_len))
3035 for (i = 0; i < shi->nr_frags; ++i) {
3036 const struct skb_frag_struct *f = &shi->frags[i];
3037 unsigned int offset = f->page_offset;
3038 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3040 sg_set_page(&sg, page, skb_frag_size(f),
3041 offset_in_page(offset));
3042 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3046 skb_walk_frags(skb, frag_iter)
3047 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3052 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3054 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3056 struct scatterlist sg;
3058 sg_init_one(&sg, key->key, key->keylen);
3059 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3061 EXPORT_SYMBOL(tcp_md5_hash_key);
3065 void tcp_done(struct sock *sk)
3067 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3069 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3070 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3072 tcp_set_state(sk, TCP_CLOSE);
3073 tcp_clear_xmit_timers(sk);
3075 reqsk_fastopen_remove(sk, req, false);
3077 sk->sk_shutdown = SHUTDOWN_MASK;
3079 if (!sock_flag(sk, SOCK_DEAD))
3080 sk->sk_state_change(sk);
3082 inet_csk_destroy_sock(sk);
3084 EXPORT_SYMBOL_GPL(tcp_done);
3086 extern struct tcp_congestion_ops tcp_reno;
3088 static __initdata unsigned long thash_entries;
3089 static int __init set_thash_entries(char *str)
3096 ret = kstrtoul(str, 0, &thash_entries);
3102 __setup("thash_entries=", set_thash_entries);
3104 static void __init tcp_init_mem(void)
3106 unsigned long limit = nr_free_buffer_pages() / 16;
3108 limit = max(limit, 128UL);
3109 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3110 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3111 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3114 void __init tcp_init(void)
3116 unsigned long limit;
3117 int max_rshare, max_wshare, cnt;
3120 sock_skb_cb_check_size(sizeof(struct tcp_skb_cb));
3122 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3123 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3124 tcp_hashinfo.bind_bucket_cachep =
3125 kmem_cache_create("tcp_bind_bucket",
3126 sizeof(struct inet_bind_bucket), 0,
3127 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3129 /* Size and allocate the main established and bind bucket
3132 * The methodology is similar to that of the buffer cache.
3134 tcp_hashinfo.ehash =
3135 alloc_large_system_hash("TCP established",
3136 sizeof(struct inet_ehash_bucket),
3138 17, /* one slot per 128 KB of memory */
3141 &tcp_hashinfo.ehash_mask,
3143 thash_entries ? 0 : 512 * 1024);
3144 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3145 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3147 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3148 panic("TCP: failed to alloc ehash_locks");
3149 tcp_hashinfo.bhash =
3150 alloc_large_system_hash("TCP bind",
3151 sizeof(struct inet_bind_hashbucket),
3152 tcp_hashinfo.ehash_mask + 1,
3153 17, /* one slot per 128 KB of memory */
3155 &tcp_hashinfo.bhash_size,
3159 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3160 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3161 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3162 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3166 cnt = tcp_hashinfo.ehash_mask + 1;
3168 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3169 sysctl_tcp_max_orphans = cnt / 2;
3170 sysctl_max_syn_backlog = max(128, cnt / 256);
3173 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3174 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3175 max_wshare = min(4UL*1024*1024, limit);
3176 max_rshare = min(6UL*1024*1024, limit);
3178 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3179 sysctl_tcp_wmem[1] = 16*1024;
3180 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3182 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3183 sysctl_tcp_rmem[1] = 87380;
3184 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3186 pr_info("Hash tables configured (established %u bind %u)\n",
3187 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3190 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3194 static int tcp_is_local(struct net *net, __be32 addr) {
3196 struct flowi4 fl4 = { .daddr = addr };
3197 rt = ip_route_output_key(net, &fl4);
3198 if (IS_ERR_OR_NULL(rt))
3200 return rt->dst.dev && (rt->dst.dev->flags & IFF_LOOPBACK);
3203 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3204 static int tcp_is_local6(struct net *net, struct in6_addr *addr) {
3205 struct rt6_info *rt6 = rt6_lookup(net, addr, addr, 0, 0);
3206 return rt6 && rt6->dst.dev && (rt6->dst.dev->flags & IFF_LOOPBACK);
3211 * tcp_nuke_addr - destroy all sockets on the given local address
3212 * if local address is the unspecified address (0.0.0.0 or ::), destroy all
3213 * sockets with local addresses that are not configured.
3215 int tcp_nuke_addr(struct net *net, struct sockaddr *addr)
3217 int family = addr->sa_family;
3218 unsigned int bucket;
3221 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3222 struct in6_addr *in6;
3224 if (family == AF_INET) {
3225 in = &((struct sockaddr_in *)addr)->sin_addr;
3226 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3227 } else if (family == AF_INET6) {
3228 in6 = &((struct sockaddr_in6 *)addr)->sin6_addr;
3231 return -EAFNOSUPPORT;
3234 for (bucket = 0; bucket <= tcp_hashinfo.ehash_mask; bucket++) {
3235 struct hlist_nulls_node *node;
3237 spinlock_t *lock = inet_ehash_lockp(&tcp_hashinfo, bucket);
3241 sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[bucket].chain) {
3242 struct inet_sock *inet = inet_sk(sk);
3244 if (sk->sk_state == TCP_TIME_WAIT) {
3246 * Sockets that are in TIME_WAIT state are
3247 * instances of lightweight inet_timewait_sock,
3248 * we should simply skip them (or we'll try to
3249 * access non-existing fields and crash).
3254 if (sysctl_ip_dynaddr && sk->sk_state == TCP_SYN_SENT)
3257 if (sock_flag(sk, SOCK_DEAD))
3260 if (family == AF_INET) {
3261 __be32 s4 = inet->inet_rcv_saddr;
3262 if (s4 == LOOPBACK4_IPV6)
3265 if (in->s_addr != s4 &&
3266 !(in->s_addr == INADDR_ANY &&
3267 !tcp_is_local(net, s4)))
3271 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
3272 if (family == AF_INET6) {
3273 struct in6_addr *s6;
3277 s6 = &inet->pinet6->rcv_saddr;
3278 if (ipv6_addr_type(s6) == IPV6_ADDR_MAPPED)
3281 if (!ipv6_addr_equal(in6, s6) &&
3282 !(ipv6_addr_equal(in6, &in6addr_any) &&
3283 !tcp_is_local6(net, s6)))
3289 spin_unlock_bh(lock);
3293 sk->sk_err = ETIMEDOUT;
3294 sk->sk_error_report(sk);
3303 spin_unlock_bh(lock);