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/sock.h>
280 #include <asm/uaccess.h>
281 #include <asm/ioctls.h>
282 #include <net/busy_poll.h>
284 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
286 int sysctl_tcp_min_tso_segs __read_mostly = 2;
288 int sysctl_tcp_autocorking __read_mostly = 1;
290 struct percpu_counter tcp_orphan_count;
291 EXPORT_SYMBOL_GPL(tcp_orphan_count);
293 long sysctl_tcp_mem[3] __read_mostly;
294 int sysctl_tcp_wmem[3] __read_mostly;
295 int sysctl_tcp_rmem[3] __read_mostly;
297 EXPORT_SYMBOL(sysctl_tcp_mem);
298 EXPORT_SYMBOL(sysctl_tcp_rmem);
299 EXPORT_SYMBOL(sysctl_tcp_wmem);
301 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
302 EXPORT_SYMBOL(tcp_memory_allocated);
305 * Current number of TCP sockets.
307 struct percpu_counter tcp_sockets_allocated;
308 EXPORT_SYMBOL(tcp_sockets_allocated);
313 struct tcp_splice_state {
314 struct pipe_inode_info *pipe;
320 * Pressure flag: try to collapse.
321 * Technical note: it is used by multiple contexts non atomically.
322 * All the __sk_mem_schedule() is of this nature: accounting
323 * is strict, actions are advisory and have some latency.
325 int tcp_memory_pressure __read_mostly;
326 EXPORT_SYMBOL(tcp_memory_pressure);
328 void tcp_enter_memory_pressure(struct sock *sk)
330 if (!tcp_memory_pressure) {
331 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
332 tcp_memory_pressure = 1;
335 EXPORT_SYMBOL(tcp_enter_memory_pressure);
337 /* Convert seconds to retransmits based on initial and max timeout */
338 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
343 int period = timeout;
346 while (seconds > period && res < 255) {
349 if (timeout > rto_max)
357 /* Convert retransmits to seconds based on initial and max timeout */
358 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
366 if (timeout > rto_max)
374 /* Address-family independent initialization for a tcp_sock.
376 * NOTE: A lot of things set to zero explicitly by call to
377 * sk_alloc() so need not be done here.
379 void tcp_init_sock(struct sock *sk)
381 struct inet_connection_sock *icsk = inet_csk(sk);
382 struct tcp_sock *tp = tcp_sk(sk);
384 __skb_queue_head_init(&tp->out_of_order_queue);
385 tcp_init_xmit_timers(sk);
386 tcp_prequeue_init(tp);
387 INIT_LIST_HEAD(&tp->tsq_node);
389 icsk->icsk_rto = TCP_TIMEOUT_INIT;
390 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
392 /* So many TCP implementations out there (incorrectly) count the
393 * initial SYN frame in their delayed-ACK and congestion control
394 * algorithms that we must have the following bandaid to talk
395 * efficiently to them. -DaveM
397 tp->snd_cwnd = TCP_INIT_CWND;
399 /* See draft-stevens-tcpca-spec-01 for discussion of the
400 * initialization of these values.
402 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
403 tp->snd_cwnd_clamp = ~0;
404 tp->mss_cache = TCP_MSS_DEFAULT;
405 u64_stats_init(&tp->syncp);
407 tp->reordering = sysctl_tcp_reordering;
408 tcp_enable_early_retrans(tp);
409 tcp_assign_congestion_control(sk);
413 sk->sk_state = TCP_CLOSE;
415 sk->sk_write_space = sk_stream_write_space;
416 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
418 icsk->icsk_sync_mss = tcp_sync_mss;
420 sk->sk_sndbuf = sysctl_tcp_wmem[1];
421 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
424 sock_update_memcg(sk);
425 sk_sockets_allocated_inc(sk);
428 EXPORT_SYMBOL(tcp_init_sock);
430 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb)
432 if (sk->sk_tsflags) {
433 struct skb_shared_info *shinfo = skb_shinfo(skb);
435 sock_tx_timestamp(sk, &shinfo->tx_flags);
436 if (shinfo->tx_flags & SKBTX_ANY_TSTAMP)
437 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
442 * Wait for a TCP event.
444 * Note that we don't need to lock the socket, as the upper poll layers
445 * take care of normal races (between the test and the event) and we don't
446 * go look at any of the socket buffers directly.
448 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
451 struct sock *sk = sock->sk;
452 const struct tcp_sock *tp = tcp_sk(sk);
454 sock_rps_record_flow(sk);
456 sock_poll_wait(file, sk_sleep(sk), wait);
457 if (sk->sk_state == TCP_LISTEN)
458 return inet_csk_listen_poll(sk);
460 /* Socket is not locked. We are protected from async events
461 * by poll logic and correct handling of state changes
462 * made by other threads is impossible in any case.
468 * POLLHUP is certainly not done right. But poll() doesn't
469 * have a notion of HUP in just one direction, and for a
470 * socket the read side is more interesting.
472 * Some poll() documentation says that POLLHUP is incompatible
473 * with the POLLOUT/POLLWR flags, so somebody should check this
474 * all. But careful, it tends to be safer to return too many
475 * bits than too few, and you can easily break real applications
476 * if you don't tell them that something has hung up!
480 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
481 * our fs/select.c). It means that after we received EOF,
482 * poll always returns immediately, making impossible poll() on write()
483 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
484 * if and only if shutdown has been made in both directions.
485 * Actually, it is interesting to look how Solaris and DUX
486 * solve this dilemma. I would prefer, if POLLHUP were maskable,
487 * then we could set it on SND_SHUTDOWN. BTW examples given
488 * in Stevens' books assume exactly this behaviour, it explains
489 * why POLLHUP is incompatible with POLLOUT. --ANK
491 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
492 * blocking on fresh not-connected or disconnected socket. --ANK
494 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
496 if (sk->sk_shutdown & RCV_SHUTDOWN)
497 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
499 /* Connected or passive Fast Open socket? */
500 if (sk->sk_state != TCP_SYN_SENT &&
501 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk)) {
502 int target = sock_rcvlowat(sk, 0, INT_MAX);
504 if (tp->urg_seq == tp->copied_seq &&
505 !sock_flag(sk, SOCK_URGINLINE) &&
509 /* Potential race condition. If read of tp below will
510 * escape above sk->sk_state, we can be illegally awaken
511 * in SYN_* states. */
512 if (tp->rcv_nxt - tp->copied_seq >= target)
513 mask |= POLLIN | POLLRDNORM;
515 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
516 if (sk_stream_is_writeable(sk)) {
517 mask |= POLLOUT | POLLWRNORM;
518 } else { /* send SIGIO later */
519 set_bit(SOCK_ASYNC_NOSPACE,
520 &sk->sk_socket->flags);
521 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
523 /* Race breaker. If space is freed after
524 * wspace test but before the flags are set,
525 * IO signal will be lost. Memory barrier
526 * pairs with the input side.
528 smp_mb__after_atomic();
529 if (sk_stream_is_writeable(sk))
530 mask |= POLLOUT | POLLWRNORM;
533 mask |= POLLOUT | POLLWRNORM;
535 if (tp->urg_data & TCP_URG_VALID)
538 /* This barrier is coupled with smp_wmb() in tcp_reset() */
540 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
545 EXPORT_SYMBOL(tcp_poll);
547 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
549 struct tcp_sock *tp = tcp_sk(sk);
555 if (sk->sk_state == TCP_LISTEN)
558 slow = lock_sock_fast(sk);
559 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
561 else if (sock_flag(sk, SOCK_URGINLINE) ||
563 before(tp->urg_seq, tp->copied_seq) ||
564 !before(tp->urg_seq, tp->rcv_nxt)) {
566 answ = tp->rcv_nxt - tp->copied_seq;
568 /* Subtract 1, if FIN was received */
569 if (answ && sock_flag(sk, SOCK_DONE))
572 answ = tp->urg_seq - tp->copied_seq;
573 unlock_sock_fast(sk, slow);
576 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
579 if (sk->sk_state == TCP_LISTEN)
582 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
585 answ = tp->write_seq - tp->snd_una;
588 if (sk->sk_state == TCP_LISTEN)
591 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
594 answ = tp->write_seq - tp->snd_nxt;
600 return put_user(answ, (int __user *)arg);
602 EXPORT_SYMBOL(tcp_ioctl);
604 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
606 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
607 tp->pushed_seq = tp->write_seq;
610 static inline bool forced_push(const struct tcp_sock *tp)
612 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
615 static void skb_entail(struct sock *sk, struct sk_buff *skb)
617 struct tcp_sock *tp = tcp_sk(sk);
618 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
621 tcb->seq = tcb->end_seq = tp->write_seq;
622 tcb->tcp_flags = TCPHDR_ACK;
624 __skb_header_release(skb);
625 tcp_add_write_queue_tail(sk, skb);
626 sk->sk_wmem_queued += skb->truesize;
627 sk_mem_charge(sk, skb->truesize);
628 if (tp->nonagle & TCP_NAGLE_PUSH)
629 tp->nonagle &= ~TCP_NAGLE_PUSH;
631 tcp_slow_start_after_idle_check(sk);
634 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
637 tp->snd_up = tp->write_seq;
640 /* If a not yet filled skb is pushed, do not send it if
641 * we have data packets in Qdisc or NIC queues :
642 * Because TX completion will happen shortly, it gives a chance
643 * to coalesce future sendmsg() payload into this skb, without
644 * need for a timer, and with no latency trade off.
645 * As packets containing data payload have a bigger truesize
646 * than pure acks (dataless) packets, the last checks prevent
647 * autocorking if we only have an ACK in Qdisc/NIC queues,
648 * or if TX completion was delayed after we processed ACK packet.
650 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
653 return skb->len < size_goal &&
654 sysctl_tcp_autocorking &&
655 skb != tcp_write_queue_head(sk) &&
656 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
659 static void tcp_push(struct sock *sk, int flags, int mss_now,
660 int nonagle, int size_goal)
662 struct tcp_sock *tp = tcp_sk(sk);
665 if (!tcp_send_head(sk))
668 skb = tcp_write_queue_tail(sk);
669 if (!(flags & MSG_MORE) || forced_push(tp))
670 tcp_mark_push(tp, skb);
672 tcp_mark_urg(tp, flags);
674 if (tcp_should_autocork(sk, skb, size_goal)) {
676 /* avoid atomic op if TSQ_THROTTLED bit is already set */
677 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
678 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
679 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
681 /* It is possible TX completion already happened
682 * before we set TSQ_THROTTLED.
684 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
688 if (flags & MSG_MORE)
689 nonagle = TCP_NAGLE_CORK;
691 __tcp_push_pending_frames(sk, mss_now, nonagle);
694 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
695 unsigned int offset, size_t len)
697 struct tcp_splice_state *tss = rd_desc->arg.data;
700 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
701 min(rd_desc->count, len), tss->flags,
704 rd_desc->count -= ret;
708 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
710 /* Store TCP splice context information in read_descriptor_t. */
711 read_descriptor_t rd_desc = {
716 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
720 * tcp_splice_read - splice data from TCP socket to a pipe
721 * @sock: socket to splice from
722 * @ppos: position (not valid)
723 * @pipe: pipe to splice to
724 * @len: number of bytes to splice
725 * @flags: splice modifier flags
728 * Will read pages from given socket and fill them into a pipe.
731 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
732 struct pipe_inode_info *pipe, size_t len,
735 struct sock *sk = sock->sk;
736 struct tcp_splice_state tss = {
745 sock_rps_record_flow(sk);
747 * We can't seek on a socket input
756 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
758 ret = __tcp_splice_read(sk, &tss);
764 if (sock_flag(sk, SOCK_DONE))
767 ret = sock_error(sk);
770 if (sk->sk_shutdown & RCV_SHUTDOWN)
772 if (sk->sk_state == TCP_CLOSE) {
774 * This occurs when user tries to read
775 * from never connected socket.
777 if (!sock_flag(sk, SOCK_DONE))
785 sk_wait_data(sk, &timeo, NULL);
786 if (signal_pending(current)) {
787 ret = sock_intr_errno(timeo);
800 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
801 (sk->sk_shutdown & RCV_SHUTDOWN) ||
802 signal_pending(current))
813 EXPORT_SYMBOL(tcp_splice_read);
815 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
820 /* The TCP header must be at least 32-bit aligned. */
821 size = ALIGN(size, 4);
823 if (unlikely(tcp_under_memory_pressure(sk)))
824 sk_mem_reclaim_partial(sk);
826 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
830 if (force_schedule) {
831 mem_scheduled = true;
832 sk_forced_mem_schedule(sk, skb->truesize);
834 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
836 if (likely(mem_scheduled)) {
837 skb_reserve(skb, sk->sk_prot->max_header);
839 * Make sure that we have exactly size bytes
840 * available to the caller, no more, no less.
842 skb->reserved_tailroom = skb->end - skb->tail - size;
847 sk->sk_prot->enter_memory_pressure(sk);
848 sk_stream_moderate_sndbuf(sk);
853 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
856 struct tcp_sock *tp = tcp_sk(sk);
857 u32 new_size_goal, size_goal;
859 if (!large_allowed || !sk_can_gso(sk))
862 /* Note : tcp_tso_autosize() will eventually split this later */
863 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
864 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
866 /* We try hard to avoid divides here */
867 size_goal = tp->gso_segs * mss_now;
868 if (unlikely(new_size_goal < size_goal ||
869 new_size_goal >= size_goal + mss_now)) {
870 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
871 sk->sk_gso_max_segs);
872 size_goal = tp->gso_segs * mss_now;
875 return max(size_goal, mss_now);
878 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
882 mss_now = tcp_current_mss(sk);
883 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
888 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
889 size_t size, int flags)
891 struct tcp_sock *tp = tcp_sk(sk);
892 int mss_now, size_goal;
895 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
897 /* Wait for a connection to finish. One exception is TCP Fast Open
898 * (passive side) where data is allowed to be sent before a connection
899 * is fully established.
901 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
902 !tcp_passive_fastopen(sk)) {
903 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
907 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
909 mss_now = tcp_send_mss(sk, &size_goal, flags);
913 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
917 struct sk_buff *skb = tcp_write_queue_tail(sk);
921 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
923 if (!sk_stream_memory_free(sk))
924 goto wait_for_sndbuf;
926 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
927 skb_queue_empty(&sk->sk_write_queue));
929 goto wait_for_memory;
938 i = skb_shinfo(skb)->nr_frags;
939 can_coalesce = skb_can_coalesce(skb, i, page, offset);
940 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
941 tcp_mark_push(tp, skb);
944 if (!sk_wmem_schedule(sk, copy))
945 goto wait_for_memory;
948 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
951 skb_fill_page_desc(skb, i, page, offset, copy);
953 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
956 skb->data_len += copy;
957 skb->truesize += copy;
958 sk->sk_wmem_queued += copy;
959 sk_mem_charge(sk, copy);
960 skb->ip_summed = CHECKSUM_PARTIAL;
961 tp->write_seq += copy;
962 TCP_SKB_CB(skb)->end_seq += copy;
963 tcp_skb_pcount_set(skb, 0);
966 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
970 if (!(size -= copy)) {
971 tcp_tx_timestamp(sk, skb);
975 if (skb->len < size_goal || (flags & MSG_OOB))
978 if (forced_push(tp)) {
979 tcp_mark_push(tp, skb);
980 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
981 } else if (skb == tcp_send_head(sk))
982 tcp_push_one(sk, mss_now);
986 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
988 tcp_push(sk, flags & ~MSG_MORE, mss_now,
989 TCP_NAGLE_PUSH, size_goal);
991 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
994 mss_now = tcp_send_mss(sk, &size_goal, flags);
998 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
999 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1006 /* make sure we wake any epoll edge trigger waiter */
1007 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1008 sk->sk_write_space(sk);
1009 return sk_stream_error(sk, flags, err);
1012 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1013 size_t size, int flags)
1017 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1018 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1019 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1023 res = do_tcp_sendpages(sk, page, offset, size, flags);
1027 EXPORT_SYMBOL(tcp_sendpage);
1029 static inline int select_size(const struct sock *sk, bool sg)
1031 const struct tcp_sock *tp = tcp_sk(sk);
1032 int tmp = tp->mss_cache;
1035 if (sk_can_gso(sk)) {
1036 /* Small frames wont use a full page:
1037 * Payload will immediately follow tcp header.
1039 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1041 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1043 if (tmp >= pgbreak &&
1044 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1052 void tcp_free_fastopen_req(struct tcp_sock *tp)
1054 if (tp->fastopen_req) {
1055 kfree(tp->fastopen_req);
1056 tp->fastopen_req = NULL;
1060 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1061 int *copied, size_t size)
1063 struct tcp_sock *tp = tcp_sk(sk);
1066 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1068 if (tp->fastopen_req)
1069 return -EALREADY; /* Another Fast Open is in progress */
1071 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1073 if (unlikely(!tp->fastopen_req))
1075 tp->fastopen_req->data = msg;
1076 tp->fastopen_req->size = size;
1078 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1079 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1080 msg->msg_namelen, flags);
1081 *copied = tp->fastopen_req->copied;
1082 tcp_free_fastopen_req(tp);
1086 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1088 struct tcp_sock *tp = tcp_sk(sk);
1089 struct sk_buff *skb;
1090 int flags, err, copied = 0;
1091 int mss_now = 0, size_goal, copied_syn = 0;
1097 flags = msg->msg_flags;
1098 if (flags & MSG_FASTOPEN) {
1099 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1100 if (err == -EINPROGRESS && copied_syn > 0)
1106 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1108 /* Wait for a connection to finish. One exception is TCP Fast Open
1109 * (passive side) where data is allowed to be sent before a connection
1110 * is fully established.
1112 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1113 !tcp_passive_fastopen(sk)) {
1114 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1118 if (unlikely(tp->repair)) {
1119 if (tp->repair_queue == TCP_RECV_QUEUE) {
1120 copied = tcp_send_rcvq(sk, msg, size);
1125 if (tp->repair_queue == TCP_NO_QUEUE)
1128 /* 'common' sending to sendq */
1131 /* This should be in poll */
1132 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1134 mss_now = tcp_send_mss(sk, &size_goal, flags);
1136 /* Ok commence sending. */
1140 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1143 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1145 while (msg_data_left(msg)) {
1147 int max = size_goal;
1149 skb = tcp_write_queue_tail(sk);
1150 if (tcp_send_head(sk)) {
1151 if (skb->ip_summed == CHECKSUM_NONE)
1153 copy = max - skb->len;
1158 /* Allocate new segment. If the interface is SG,
1159 * allocate skb fitting to single page.
1161 if (!sk_stream_memory_free(sk))
1162 goto wait_for_sndbuf;
1164 skb = sk_stream_alloc_skb(sk,
1165 select_size(sk, sg),
1167 skb_queue_empty(&sk->sk_write_queue));
1169 goto wait_for_memory;
1172 * Check whether we can use HW checksum.
1174 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1175 skb->ip_summed = CHECKSUM_PARTIAL;
1177 skb_entail(sk, skb);
1181 /* All packets are restored as if they have
1182 * already been sent. skb_mstamp isn't set to
1183 * avoid wrong rtt estimation.
1186 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1189 /* Try to append data to the end of skb. */
1190 if (copy > msg_data_left(msg))
1191 copy = msg_data_left(msg);
1193 /* Where to copy to? */
1194 if (skb_availroom(skb) > 0) {
1195 /* We have some space in skb head. Superb! */
1196 copy = min_t(int, copy, skb_availroom(skb));
1197 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1202 int i = skb_shinfo(skb)->nr_frags;
1203 struct page_frag *pfrag = sk_page_frag(sk);
1205 if (!sk_page_frag_refill(sk, pfrag))
1206 goto wait_for_memory;
1208 if (!skb_can_coalesce(skb, i, pfrag->page,
1210 if (i == MAX_SKB_FRAGS || !sg) {
1211 tcp_mark_push(tp, skb);
1217 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1219 if (!sk_wmem_schedule(sk, copy))
1220 goto wait_for_memory;
1222 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1229 /* Update the skb. */
1231 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1233 skb_fill_page_desc(skb, i, pfrag->page,
1234 pfrag->offset, copy);
1235 get_page(pfrag->page);
1237 pfrag->offset += copy;
1241 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1243 tp->write_seq += copy;
1244 TCP_SKB_CB(skb)->end_seq += copy;
1245 tcp_skb_pcount_set(skb, 0);
1248 if (!msg_data_left(msg)) {
1249 tcp_tx_timestamp(sk, skb);
1253 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1256 if (forced_push(tp)) {
1257 tcp_mark_push(tp, skb);
1258 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1259 } else if (skb == tcp_send_head(sk))
1260 tcp_push_one(sk, mss_now);
1264 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1267 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1268 TCP_NAGLE_PUSH, size_goal);
1270 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1273 mss_now = tcp_send_mss(sk, &size_goal, flags);
1278 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1281 return copied + copied_syn;
1285 tcp_unlink_write_queue(skb, sk);
1286 /* It is the one place in all of TCP, except connection
1287 * reset, where we can be unlinking the send_head.
1289 tcp_check_send_head(sk, skb);
1290 sk_wmem_free_skb(sk, skb);
1294 if (copied + copied_syn)
1297 err = sk_stream_error(sk, flags, err);
1298 /* make sure we wake any epoll edge trigger waiter */
1299 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1300 sk->sk_write_space(sk);
1304 EXPORT_SYMBOL(tcp_sendmsg);
1307 * Handle reading urgent data. BSD has very simple semantics for
1308 * this, no blocking and very strange errors 8)
1311 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1313 struct tcp_sock *tp = tcp_sk(sk);
1315 /* No URG data to read. */
1316 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1317 tp->urg_data == TCP_URG_READ)
1318 return -EINVAL; /* Yes this is right ! */
1320 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1323 if (tp->urg_data & TCP_URG_VALID) {
1325 char c = tp->urg_data;
1327 if (!(flags & MSG_PEEK))
1328 tp->urg_data = TCP_URG_READ;
1330 /* Read urgent data. */
1331 msg->msg_flags |= MSG_OOB;
1334 if (!(flags & MSG_TRUNC))
1335 err = memcpy_to_msg(msg, &c, 1);
1338 msg->msg_flags |= MSG_TRUNC;
1340 return err ? -EFAULT : len;
1343 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1346 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1347 * the available implementations agree in this case:
1348 * this call should never block, independent of the
1349 * blocking state of the socket.
1350 * Mike <pall@rz.uni-karlsruhe.de>
1355 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1357 struct sk_buff *skb;
1358 int copied = 0, err = 0;
1360 /* XXX -- need to support SO_PEEK_OFF */
1362 skb_queue_walk(&sk->sk_write_queue, skb) {
1363 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1370 return err ?: copied;
1373 /* Clean up the receive buffer for full frames taken by the user,
1374 * then send an ACK if necessary. COPIED is the number of bytes
1375 * tcp_recvmsg has given to the user so far, it speeds up the
1376 * calculation of whether or not we must ACK for the sake of
1379 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1381 struct tcp_sock *tp = tcp_sk(sk);
1382 bool time_to_ack = false;
1384 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1386 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1387 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1388 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1390 if (inet_csk_ack_scheduled(sk)) {
1391 const struct inet_connection_sock *icsk = inet_csk(sk);
1392 /* Delayed ACKs frequently hit locked sockets during bulk
1394 if (icsk->icsk_ack.blocked ||
1395 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1396 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1398 * If this read emptied read buffer, we send ACK, if
1399 * connection is not bidirectional, user drained
1400 * receive buffer and there was a small segment
1404 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1405 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1406 !icsk->icsk_ack.pingpong)) &&
1407 !atomic_read(&sk->sk_rmem_alloc)))
1411 /* We send an ACK if we can now advertise a non-zero window
1412 * which has been raised "significantly".
1414 * Even if window raised up to infinity, do not send window open ACK
1415 * in states, where we will not receive more. It is useless.
1417 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1418 __u32 rcv_window_now = tcp_receive_window(tp);
1420 /* Optimize, __tcp_select_window() is not cheap. */
1421 if (2*rcv_window_now <= tp->window_clamp) {
1422 __u32 new_window = __tcp_select_window(sk);
1424 /* Send ACK now, if this read freed lots of space
1425 * in our buffer. Certainly, new_window is new window.
1426 * We can advertise it now, if it is not less than current one.
1427 * "Lots" means "at least twice" here.
1429 if (new_window && new_window >= 2 * rcv_window_now)
1437 static void tcp_prequeue_process(struct sock *sk)
1439 struct sk_buff *skb;
1440 struct tcp_sock *tp = tcp_sk(sk);
1442 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1444 /* RX process wants to run with disabled BHs, though it is not
1447 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1448 sk_backlog_rcv(sk, skb);
1451 /* Clear memory counter. */
1452 tp->ucopy.memory = 0;
1455 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1457 struct sk_buff *skb;
1460 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1461 offset = seq - TCP_SKB_CB(skb)->seq;
1462 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1464 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1468 /* This looks weird, but this can happen if TCP collapsing
1469 * splitted a fat GRO packet, while we released socket lock
1470 * in skb_splice_bits()
1472 sk_eat_skb(sk, skb);
1478 * This routine provides an alternative to tcp_recvmsg() for routines
1479 * that would like to handle copying from skbuffs directly in 'sendfile'
1482 * - It is assumed that the socket was locked by the caller.
1483 * - The routine does not block.
1484 * - At present, there is no support for reading OOB data
1485 * or for 'peeking' the socket using this routine
1486 * (although both would be easy to implement).
1488 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1489 sk_read_actor_t recv_actor)
1491 struct sk_buff *skb;
1492 struct tcp_sock *tp = tcp_sk(sk);
1493 u32 seq = tp->copied_seq;
1497 if (sk->sk_state == TCP_LISTEN)
1499 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1500 if (offset < skb->len) {
1504 len = skb->len - offset;
1505 /* Stop reading if we hit a patch of urgent data */
1507 u32 urg_offset = tp->urg_seq - seq;
1508 if (urg_offset < len)
1513 used = recv_actor(desc, skb, offset, len);
1518 } else if (used <= len) {
1523 /* If recv_actor drops the lock (e.g. TCP splice
1524 * receive) the skb pointer might be invalid when
1525 * getting here: tcp_collapse might have deleted it
1526 * while aggregating skbs from the socket queue.
1528 skb = tcp_recv_skb(sk, seq - 1, &offset);
1531 /* TCP coalescing might have appended data to the skb.
1532 * Try to splice more frags
1534 if (offset + 1 != skb->len)
1537 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1538 sk_eat_skb(sk, skb);
1542 sk_eat_skb(sk, skb);
1545 tp->copied_seq = seq;
1547 tp->copied_seq = seq;
1549 tcp_rcv_space_adjust(sk);
1551 /* Clean up data we have read: This will do ACK frames. */
1553 tcp_recv_skb(sk, seq, &offset);
1554 tcp_cleanup_rbuf(sk, copied);
1558 EXPORT_SYMBOL(tcp_read_sock);
1561 * This routine copies from a sock struct into the user buffer.
1563 * Technical note: in 2.3 we work on _locked_ socket, so that
1564 * tricks with *seq access order and skb->users are not required.
1565 * Probably, code can be easily improved even more.
1568 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1569 int flags, int *addr_len)
1571 struct tcp_sock *tp = tcp_sk(sk);
1577 int target; /* Read at least this many bytes */
1579 struct task_struct *user_recv = NULL;
1580 struct sk_buff *skb, *last;
1583 if (unlikely(flags & MSG_ERRQUEUE))
1584 return inet_recv_error(sk, msg, len, addr_len);
1586 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1587 (sk->sk_state == TCP_ESTABLISHED))
1588 sk_busy_loop(sk, nonblock);
1593 if (sk->sk_state == TCP_LISTEN)
1596 timeo = sock_rcvtimeo(sk, nonblock);
1598 /* Urgent data needs to be handled specially. */
1599 if (flags & MSG_OOB)
1602 if (unlikely(tp->repair)) {
1604 if (!(flags & MSG_PEEK))
1607 if (tp->repair_queue == TCP_SEND_QUEUE)
1611 if (tp->repair_queue == TCP_NO_QUEUE)
1614 /* 'common' recv queue MSG_PEEK-ing */
1617 seq = &tp->copied_seq;
1618 if (flags & MSG_PEEK) {
1619 peek_seq = tp->copied_seq;
1623 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1628 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1629 if (tp->urg_data && tp->urg_seq == *seq) {
1632 if (signal_pending(current)) {
1633 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1638 /* Next get a buffer. */
1640 last = skb_peek_tail(&sk->sk_receive_queue);
1641 skb_queue_walk(&sk->sk_receive_queue, skb) {
1643 /* Now that we have two receive queues this
1646 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1647 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1648 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1652 offset = *seq - TCP_SKB_CB(skb)->seq;
1653 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1655 if (offset < skb->len)
1657 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1659 WARN(!(flags & MSG_PEEK),
1660 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1661 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1664 /* Well, if we have backlog, try to process it now yet. */
1666 if (copied >= target && !sk->sk_backlog.tail)
1671 sk->sk_state == TCP_CLOSE ||
1672 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1674 signal_pending(current))
1677 if (sock_flag(sk, SOCK_DONE))
1681 copied = sock_error(sk);
1685 if (sk->sk_shutdown & RCV_SHUTDOWN)
1688 if (sk->sk_state == TCP_CLOSE) {
1689 if (!sock_flag(sk, SOCK_DONE)) {
1690 /* This occurs when user tries to read
1691 * from never connected socket.
1704 if (signal_pending(current)) {
1705 copied = sock_intr_errno(timeo);
1710 tcp_cleanup_rbuf(sk, copied);
1712 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1713 /* Install new reader */
1714 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1715 user_recv = current;
1716 tp->ucopy.task = user_recv;
1717 tp->ucopy.msg = msg;
1720 tp->ucopy.len = len;
1722 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1723 !(flags & (MSG_PEEK | MSG_TRUNC)));
1725 /* Ugly... If prequeue is not empty, we have to
1726 * process it before releasing socket, otherwise
1727 * order will be broken at second iteration.
1728 * More elegant solution is required!!!
1730 * Look: we have the following (pseudo)queues:
1732 * 1. packets in flight
1737 * Each queue can be processed only if the next ones
1738 * are empty. At this point we have empty receive_queue.
1739 * But prequeue _can_ be not empty after 2nd iteration,
1740 * when we jumped to start of loop because backlog
1741 * processing added something to receive_queue.
1742 * We cannot release_sock(), because backlog contains
1743 * packets arrived _after_ prequeued ones.
1745 * Shortly, algorithm is clear --- to process all
1746 * the queues in order. We could make it more directly,
1747 * requeueing packets from backlog to prequeue, if
1748 * is not empty. It is more elegant, but eats cycles,
1751 if (!skb_queue_empty(&tp->ucopy.prequeue))
1754 /* __ Set realtime policy in scheduler __ */
1757 if (copied >= target) {
1758 /* Do not sleep, just process backlog. */
1762 sk_wait_data(sk, &timeo, last);
1768 /* __ Restore normal policy in scheduler __ */
1770 if ((chunk = len - tp->ucopy.len) != 0) {
1771 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1776 if (tp->rcv_nxt == tp->copied_seq &&
1777 !skb_queue_empty(&tp->ucopy.prequeue)) {
1779 tcp_prequeue_process(sk);
1781 if ((chunk = len - tp->ucopy.len) != 0) {
1782 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1788 if ((flags & MSG_PEEK) &&
1789 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1790 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1792 task_pid_nr(current));
1793 peek_seq = tp->copied_seq;
1798 /* Ok so how much can we use? */
1799 used = skb->len - offset;
1803 /* Do we have urgent data here? */
1805 u32 urg_offset = tp->urg_seq - *seq;
1806 if (urg_offset < used) {
1808 if (!sock_flag(sk, SOCK_URGINLINE)) {
1821 if (!(flags & MSG_TRUNC)) {
1822 err = skb_copy_datagram_msg(skb, offset, msg, used);
1824 /* Exception. Bailout! */
1835 tcp_rcv_space_adjust(sk);
1838 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1840 tcp_fast_path_check(sk);
1842 if (used + offset < skb->len)
1845 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1847 if (!(flags & MSG_PEEK))
1848 sk_eat_skb(sk, skb);
1852 /* Process the FIN. */
1854 if (!(flags & MSG_PEEK))
1855 sk_eat_skb(sk, skb);
1860 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1863 tp->ucopy.len = copied > 0 ? len : 0;
1865 tcp_prequeue_process(sk);
1867 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1868 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1874 tp->ucopy.task = NULL;
1878 /* According to UNIX98, msg_name/msg_namelen are ignored
1879 * on connected socket. I was just happy when found this 8) --ANK
1882 /* Clean up data we have read: This will do ACK frames. */
1883 tcp_cleanup_rbuf(sk, copied);
1893 err = tcp_recv_urg(sk, msg, len, flags);
1897 err = tcp_peek_sndq(sk, msg, len);
1900 EXPORT_SYMBOL(tcp_recvmsg);
1902 void tcp_set_state(struct sock *sk, int state)
1904 int oldstate = sk->sk_state;
1907 case TCP_ESTABLISHED:
1908 if (oldstate != TCP_ESTABLISHED)
1909 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1913 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1914 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1916 sk->sk_prot->unhash(sk);
1917 if (inet_csk(sk)->icsk_bind_hash &&
1918 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1922 if (oldstate == TCP_ESTABLISHED)
1923 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1926 /* Change state AFTER socket is unhashed to avoid closed
1927 * socket sitting in hash tables.
1929 sk->sk_state = state;
1932 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1935 EXPORT_SYMBOL_GPL(tcp_set_state);
1938 * State processing on a close. This implements the state shift for
1939 * sending our FIN frame. Note that we only send a FIN for some
1940 * states. A shutdown() may have already sent the FIN, or we may be
1944 static const unsigned char new_state[16] = {
1945 /* current state: new state: action: */
1946 [0 /* (Invalid) */] = TCP_CLOSE,
1947 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1948 [TCP_SYN_SENT] = TCP_CLOSE,
1949 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1950 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1951 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1952 [TCP_TIME_WAIT] = TCP_CLOSE,
1953 [TCP_CLOSE] = TCP_CLOSE,
1954 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
1955 [TCP_LAST_ACK] = TCP_LAST_ACK,
1956 [TCP_LISTEN] = TCP_CLOSE,
1957 [TCP_CLOSING] = TCP_CLOSING,
1958 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
1961 static int tcp_close_state(struct sock *sk)
1963 int next = (int)new_state[sk->sk_state];
1964 int ns = next & TCP_STATE_MASK;
1966 tcp_set_state(sk, ns);
1968 return next & TCP_ACTION_FIN;
1972 * Shutdown the sending side of a connection. Much like close except
1973 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1976 void tcp_shutdown(struct sock *sk, int how)
1978 /* We need to grab some memory, and put together a FIN,
1979 * and then put it into the queue to be sent.
1980 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1982 if (!(how & SEND_SHUTDOWN))
1985 /* If we've already sent a FIN, or it's a closed state, skip this. */
1986 if ((1 << sk->sk_state) &
1987 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1988 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1989 /* Clear out any half completed packets. FIN if needed. */
1990 if (tcp_close_state(sk))
1994 EXPORT_SYMBOL(tcp_shutdown);
1996 bool tcp_check_oom(struct sock *sk, int shift)
1998 bool too_many_orphans, out_of_socket_memory;
2000 too_many_orphans = tcp_too_many_orphans(sk, shift);
2001 out_of_socket_memory = tcp_out_of_memory(sk);
2003 if (too_many_orphans)
2004 net_info_ratelimited("too many orphaned sockets\n");
2005 if (out_of_socket_memory)
2006 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2007 return too_many_orphans || out_of_socket_memory;
2010 void tcp_close(struct sock *sk, long timeout)
2012 struct sk_buff *skb;
2013 int data_was_unread = 0;
2017 sk->sk_shutdown = SHUTDOWN_MASK;
2019 if (sk->sk_state == TCP_LISTEN) {
2020 tcp_set_state(sk, TCP_CLOSE);
2023 inet_csk_listen_stop(sk);
2025 goto adjudge_to_death;
2028 /* We need to flush the recv. buffs. We do this only on the
2029 * descriptor close, not protocol-sourced closes, because the
2030 * reader process may not have drained the data yet!
2032 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2033 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2035 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2037 data_was_unread += len;
2043 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2044 if (sk->sk_state == TCP_CLOSE)
2045 goto adjudge_to_death;
2047 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2048 * data was lost. To witness the awful effects of the old behavior of
2049 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2050 * GET in an FTP client, suspend the process, wait for the client to
2051 * advertise a zero window, then kill -9 the FTP client, wheee...
2052 * Note: timeout is always zero in such a case.
2054 if (unlikely(tcp_sk(sk)->repair)) {
2055 sk->sk_prot->disconnect(sk, 0);
2056 } else if (data_was_unread) {
2057 /* Unread data was tossed, zap the connection. */
2058 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2059 tcp_set_state(sk, TCP_CLOSE);
2060 tcp_send_active_reset(sk, sk->sk_allocation);
2061 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2062 /* Check zero linger _after_ checking for unread data. */
2063 sk->sk_prot->disconnect(sk, 0);
2064 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2065 } else if (tcp_close_state(sk)) {
2066 /* We FIN if the application ate all the data before
2067 * zapping the connection.
2070 /* RED-PEN. Formally speaking, we have broken TCP state
2071 * machine. State transitions:
2073 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2074 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2075 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2077 * are legal only when FIN has been sent (i.e. in window),
2078 * rather than queued out of window. Purists blame.
2080 * F.e. "RFC state" is ESTABLISHED,
2081 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2083 * The visible declinations are that sometimes
2084 * we enter time-wait state, when it is not required really
2085 * (harmless), do not send active resets, when they are
2086 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2087 * they look as CLOSING or LAST_ACK for Linux)
2088 * Probably, I missed some more holelets.
2090 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2091 * in a single packet! (May consider it later but will
2092 * probably need API support or TCP_CORK SYN-ACK until
2093 * data is written and socket is closed.)
2098 sk_stream_wait_close(sk, timeout);
2101 state = sk->sk_state;
2105 /* It is the last release_sock in its life. It will remove backlog. */
2109 /* Now socket is owned by kernel and we acquire BH lock
2110 to finish close. No need to check for user refs.
2114 WARN_ON(sock_owned_by_user(sk));
2116 percpu_counter_inc(sk->sk_prot->orphan_count);
2118 /* Have we already been destroyed by a softirq or backlog? */
2119 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2122 /* This is a (useful) BSD violating of the RFC. There is a
2123 * problem with TCP as specified in that the other end could
2124 * keep a socket open forever with no application left this end.
2125 * We use a 1 minute timeout (about the same as BSD) then kill
2126 * our end. If they send after that then tough - BUT: long enough
2127 * that we won't make the old 4*rto = almost no time - whoops
2130 * Nope, it was not mistake. It is really desired behaviour
2131 * f.e. on http servers, when such sockets are useless, but
2132 * consume significant resources. Let's do it with special
2133 * linger2 option. --ANK
2136 if (sk->sk_state == TCP_FIN_WAIT2) {
2137 struct tcp_sock *tp = tcp_sk(sk);
2138 if (tp->linger2 < 0) {
2139 tcp_set_state(sk, TCP_CLOSE);
2140 tcp_send_active_reset(sk, GFP_ATOMIC);
2141 NET_INC_STATS_BH(sock_net(sk),
2142 LINUX_MIB_TCPABORTONLINGER);
2144 const int tmo = tcp_fin_time(sk);
2146 if (tmo > TCP_TIMEWAIT_LEN) {
2147 inet_csk_reset_keepalive_timer(sk,
2148 tmo - TCP_TIMEWAIT_LEN);
2150 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2155 if (sk->sk_state != TCP_CLOSE) {
2157 if (tcp_check_oom(sk, 0)) {
2158 tcp_set_state(sk, TCP_CLOSE);
2159 tcp_send_active_reset(sk, GFP_ATOMIC);
2160 NET_INC_STATS_BH(sock_net(sk),
2161 LINUX_MIB_TCPABORTONMEMORY);
2165 if (sk->sk_state == TCP_CLOSE) {
2166 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2167 /* We could get here with a non-NULL req if the socket is
2168 * aborted (e.g., closed with unread data) before 3WHS
2172 reqsk_fastopen_remove(sk, req, false);
2173 inet_csk_destroy_sock(sk);
2175 /* Otherwise, socket is reprieved until protocol close. */
2182 EXPORT_SYMBOL(tcp_close);
2184 /* These states need RST on ABORT according to RFC793 */
2186 static inline bool tcp_need_reset(int state)
2188 return (1 << state) &
2189 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2190 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2193 int tcp_disconnect(struct sock *sk, int flags)
2195 struct inet_sock *inet = inet_sk(sk);
2196 struct inet_connection_sock *icsk = inet_csk(sk);
2197 struct tcp_sock *tp = tcp_sk(sk);
2199 int old_state = sk->sk_state;
2201 if (old_state != TCP_CLOSE)
2202 tcp_set_state(sk, TCP_CLOSE);
2204 /* ABORT function of RFC793 */
2205 if (old_state == TCP_LISTEN) {
2206 inet_csk_listen_stop(sk);
2207 } else if (unlikely(tp->repair)) {
2208 sk->sk_err = ECONNABORTED;
2209 } else if (tcp_need_reset(old_state) ||
2210 (tp->snd_nxt != tp->write_seq &&
2211 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2212 /* The last check adjusts for discrepancy of Linux wrt. RFC
2215 tcp_send_active_reset(sk, gfp_any());
2216 sk->sk_err = ECONNRESET;
2217 } else if (old_state == TCP_SYN_SENT)
2218 sk->sk_err = ECONNRESET;
2220 tcp_clear_xmit_timers(sk);
2221 __skb_queue_purge(&sk->sk_receive_queue);
2222 tcp_write_queue_purge(sk);
2223 __skb_queue_purge(&tp->out_of_order_queue);
2225 inet->inet_dport = 0;
2227 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2228 inet_reset_saddr(sk);
2230 sk->sk_shutdown = 0;
2231 sock_reset_flag(sk, SOCK_DONE);
2233 if ((tp->write_seq += tp->max_window + 2) == 0)
2235 icsk->icsk_backoff = 0;
2237 icsk->icsk_probes_out = 0;
2238 tp->packets_out = 0;
2239 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2240 tp->snd_cwnd_cnt = 0;
2241 tp->window_clamp = 0;
2242 tcp_set_ca_state(sk, TCP_CA_Open);
2243 tcp_clear_retrans(tp);
2244 inet_csk_delack_init(sk);
2245 tcp_init_send_head(sk);
2246 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2249 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2251 sk->sk_error_report(sk);
2254 EXPORT_SYMBOL(tcp_disconnect);
2256 static inline bool tcp_can_repair_sock(const struct sock *sk)
2258 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2259 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2262 static int tcp_repair_options_est(struct tcp_sock *tp,
2263 struct tcp_repair_opt __user *optbuf, unsigned int len)
2265 struct tcp_repair_opt opt;
2267 while (len >= sizeof(opt)) {
2268 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2274 switch (opt.opt_code) {
2276 tp->rx_opt.mss_clamp = opt.opt_val;
2280 u16 snd_wscale = opt.opt_val & 0xFFFF;
2281 u16 rcv_wscale = opt.opt_val >> 16;
2283 if (snd_wscale > 14 || rcv_wscale > 14)
2286 tp->rx_opt.snd_wscale = snd_wscale;
2287 tp->rx_opt.rcv_wscale = rcv_wscale;
2288 tp->rx_opt.wscale_ok = 1;
2291 case TCPOPT_SACK_PERM:
2292 if (opt.opt_val != 0)
2295 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2296 if (sysctl_tcp_fack)
2297 tcp_enable_fack(tp);
2299 case TCPOPT_TIMESTAMP:
2300 if (opt.opt_val != 0)
2303 tp->rx_opt.tstamp_ok = 1;
2312 * Socket option code for TCP.
2314 static int do_tcp_setsockopt(struct sock *sk, int level,
2315 int optname, char __user *optval, unsigned int optlen)
2317 struct tcp_sock *tp = tcp_sk(sk);
2318 struct inet_connection_sock *icsk = inet_csk(sk);
2322 /* These are data/string values, all the others are ints */
2324 case TCP_CONGESTION: {
2325 char name[TCP_CA_NAME_MAX];
2330 val = strncpy_from_user(name, optval,
2331 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2337 err = tcp_set_congestion_control(sk, name);
2346 if (optlen < sizeof(int))
2349 if (get_user(val, (int __user *)optval))
2356 /* Values greater than interface MTU won't take effect. However
2357 * at the point when this call is done we typically don't yet
2358 * know which interface is going to be used */
2359 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2363 tp->rx_opt.user_mss = val;
2368 /* TCP_NODELAY is weaker than TCP_CORK, so that
2369 * this option on corked socket is remembered, but
2370 * it is not activated until cork is cleared.
2372 * However, when TCP_NODELAY is set we make
2373 * an explicit push, which overrides even TCP_CORK
2374 * for currently queued segments.
2376 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2377 tcp_push_pending_frames(sk);
2379 tp->nonagle &= ~TCP_NAGLE_OFF;
2383 case TCP_THIN_LINEAR_TIMEOUTS:
2384 if (val < 0 || val > 1)
2390 case TCP_THIN_DUPACK:
2391 if (val < 0 || val > 1)
2394 tp->thin_dupack = val;
2395 if (tp->thin_dupack)
2396 tcp_disable_early_retrans(tp);
2401 if (!tcp_can_repair_sock(sk))
2403 else if (val == 1) {
2405 sk->sk_reuse = SK_FORCE_REUSE;
2406 tp->repair_queue = TCP_NO_QUEUE;
2407 } else if (val == 0) {
2409 sk->sk_reuse = SK_NO_REUSE;
2410 tcp_send_window_probe(sk);
2416 case TCP_REPAIR_QUEUE:
2419 else if (val < TCP_QUEUES_NR)
2420 tp->repair_queue = val;
2426 if (sk->sk_state != TCP_CLOSE)
2428 else if (tp->repair_queue == TCP_SEND_QUEUE)
2429 tp->write_seq = val;
2430 else if (tp->repair_queue == TCP_RECV_QUEUE)
2436 case TCP_REPAIR_OPTIONS:
2439 else if (sk->sk_state == TCP_ESTABLISHED)
2440 err = tcp_repair_options_est(tp,
2441 (struct tcp_repair_opt __user *)optval,
2448 /* When set indicates to always queue non-full frames.
2449 * Later the user clears this option and we transmit
2450 * any pending partial frames in the queue. This is
2451 * meant to be used alongside sendfile() to get properly
2452 * filled frames when the user (for example) must write
2453 * out headers with a write() call first and then use
2454 * sendfile to send out the data parts.
2456 * TCP_CORK can be set together with TCP_NODELAY and it is
2457 * stronger than TCP_NODELAY.
2460 tp->nonagle |= TCP_NAGLE_CORK;
2462 tp->nonagle &= ~TCP_NAGLE_CORK;
2463 if (tp->nonagle&TCP_NAGLE_OFF)
2464 tp->nonagle |= TCP_NAGLE_PUSH;
2465 tcp_push_pending_frames(sk);
2470 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2473 tp->keepalive_time = val * HZ;
2474 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2475 !((1 << sk->sk_state) &
2476 (TCPF_CLOSE | TCPF_LISTEN))) {
2477 u32 elapsed = keepalive_time_elapsed(tp);
2478 if (tp->keepalive_time > elapsed)
2479 elapsed = tp->keepalive_time - elapsed;
2482 inet_csk_reset_keepalive_timer(sk, elapsed);
2487 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2490 tp->keepalive_intvl = val * HZ;
2493 if (val < 1 || val > MAX_TCP_KEEPCNT)
2496 tp->keepalive_probes = val;
2499 if (val < 1 || val > MAX_TCP_SYNCNT)
2502 icsk->icsk_syn_retries = val;
2506 if (val < 0 || val > 1)
2515 else if (val > sysctl_tcp_fin_timeout / HZ)
2518 tp->linger2 = val * HZ;
2521 case TCP_DEFER_ACCEPT:
2522 /* Translate value in seconds to number of retransmits */
2523 icsk->icsk_accept_queue.rskq_defer_accept =
2524 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2528 case TCP_WINDOW_CLAMP:
2530 if (sk->sk_state != TCP_CLOSE) {
2534 tp->window_clamp = 0;
2536 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2537 SOCK_MIN_RCVBUF / 2 : val;
2542 icsk->icsk_ack.pingpong = 1;
2544 icsk->icsk_ack.pingpong = 0;
2545 if ((1 << sk->sk_state) &
2546 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2547 inet_csk_ack_scheduled(sk)) {
2548 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2549 tcp_cleanup_rbuf(sk, 1);
2551 icsk->icsk_ack.pingpong = 1;
2556 #ifdef CONFIG_TCP_MD5SIG
2558 /* Read the IP->Key mappings from userspace */
2559 err = tp->af_specific->md5_parse(sk, optval, optlen);
2562 case TCP_USER_TIMEOUT:
2563 /* Cap the max time in ms TCP will retry or probe the window
2564 * before giving up and aborting (ETIMEDOUT) a connection.
2569 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2573 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2575 tcp_fastopen_init_key_once(true);
2577 fastopen_queue_tune(sk, val);
2586 tp->tsoffset = val - tcp_time_stamp;
2588 case TCP_NOTSENT_LOWAT:
2589 tp->notsent_lowat = val;
2590 sk->sk_write_space(sk);
2601 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2602 unsigned int optlen)
2604 const struct inet_connection_sock *icsk = inet_csk(sk);
2606 if (level != SOL_TCP)
2607 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2609 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2611 EXPORT_SYMBOL(tcp_setsockopt);
2613 #ifdef CONFIG_COMPAT
2614 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2615 char __user *optval, unsigned int optlen)
2617 if (level != SOL_TCP)
2618 return inet_csk_compat_setsockopt(sk, level, optname,
2620 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2622 EXPORT_SYMBOL(compat_tcp_setsockopt);
2625 /* Return information about state of tcp endpoint in API format. */
2626 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2628 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2629 const struct inet_connection_sock *icsk = inet_csk(sk);
2630 u32 now = tcp_time_stamp;
2634 memset(info, 0, sizeof(*info));
2635 if (sk->sk_type != SOCK_STREAM)
2638 info->tcpi_state = sk->sk_state;
2639 info->tcpi_ca_state = icsk->icsk_ca_state;
2640 info->tcpi_retransmits = icsk->icsk_retransmits;
2641 info->tcpi_probes = icsk->icsk_probes_out;
2642 info->tcpi_backoff = icsk->icsk_backoff;
2644 if (tp->rx_opt.tstamp_ok)
2645 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2646 if (tcp_is_sack(tp))
2647 info->tcpi_options |= TCPI_OPT_SACK;
2648 if (tp->rx_opt.wscale_ok) {
2649 info->tcpi_options |= TCPI_OPT_WSCALE;
2650 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2651 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2654 if (tp->ecn_flags & TCP_ECN_OK)
2655 info->tcpi_options |= TCPI_OPT_ECN;
2656 if (tp->ecn_flags & TCP_ECN_SEEN)
2657 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2658 if (tp->syn_data_acked)
2659 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2661 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2662 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2663 info->tcpi_snd_mss = tp->mss_cache;
2664 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2666 if (sk->sk_state == TCP_LISTEN) {
2667 info->tcpi_unacked = sk->sk_ack_backlog;
2668 info->tcpi_sacked = sk->sk_max_ack_backlog;
2670 info->tcpi_unacked = tp->packets_out;
2671 info->tcpi_sacked = tp->sacked_out;
2673 info->tcpi_lost = tp->lost_out;
2674 info->tcpi_retrans = tp->retrans_out;
2675 info->tcpi_fackets = tp->fackets_out;
2677 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2678 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2679 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2681 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2682 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2683 info->tcpi_rtt = tp->srtt_us >> 3;
2684 info->tcpi_rttvar = tp->mdev_us >> 2;
2685 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2686 info->tcpi_snd_cwnd = tp->snd_cwnd;
2687 info->tcpi_advmss = tp->advmss;
2688 info->tcpi_reordering = tp->reordering;
2690 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2691 info->tcpi_rcv_space = tp->rcvq_space.space;
2693 info->tcpi_total_retrans = tp->total_retrans;
2695 rate = READ_ONCE(sk->sk_pacing_rate);
2696 info->tcpi_pacing_rate = rate != ~0U ? rate : ~0ULL;
2698 rate = READ_ONCE(sk->sk_max_pacing_rate);
2699 info->tcpi_max_pacing_rate = rate != ~0U ? rate : ~0ULL;
2702 start = u64_stats_fetch_begin_irq(&tp->syncp);
2703 info->tcpi_bytes_acked = tp->bytes_acked;
2704 info->tcpi_bytes_received = tp->bytes_received;
2705 } while (u64_stats_fetch_retry_irq(&tp->syncp, start));
2706 info->tcpi_segs_out = tp->segs_out;
2707 info->tcpi_segs_in = tp->segs_in;
2709 EXPORT_SYMBOL_GPL(tcp_get_info);
2711 static int do_tcp_getsockopt(struct sock *sk, int level,
2712 int optname, char __user *optval, int __user *optlen)
2714 struct inet_connection_sock *icsk = inet_csk(sk);
2715 struct tcp_sock *tp = tcp_sk(sk);
2718 if (get_user(len, optlen))
2721 len = min_t(unsigned int, len, sizeof(int));
2728 val = tp->mss_cache;
2729 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2730 val = tp->rx_opt.user_mss;
2732 val = tp->rx_opt.mss_clamp;
2735 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2738 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2741 val = keepalive_time_when(tp) / HZ;
2744 val = keepalive_intvl_when(tp) / HZ;
2747 val = keepalive_probes(tp);
2750 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2755 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2757 case TCP_DEFER_ACCEPT:
2758 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2759 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2761 case TCP_WINDOW_CLAMP:
2762 val = tp->window_clamp;
2765 struct tcp_info info;
2767 if (get_user(len, optlen))
2770 tcp_get_info(sk, &info);
2772 len = min_t(unsigned int, len, sizeof(info));
2773 if (put_user(len, optlen))
2775 if (copy_to_user(optval, &info, len))
2780 const struct tcp_congestion_ops *ca_ops;
2781 union tcp_cc_info info;
2785 if (get_user(len, optlen))
2788 ca_ops = icsk->icsk_ca_ops;
2789 if (ca_ops && ca_ops->get_info)
2790 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2792 len = min_t(unsigned int, len, sz);
2793 if (put_user(len, optlen))
2795 if (copy_to_user(optval, &info, len))
2800 val = !icsk->icsk_ack.pingpong;
2803 case TCP_CONGESTION:
2804 if (get_user(len, optlen))
2806 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2807 if (put_user(len, optlen))
2809 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2813 case TCP_THIN_LINEAR_TIMEOUTS:
2816 case TCP_THIN_DUPACK:
2817 val = tp->thin_dupack;
2824 case TCP_REPAIR_QUEUE:
2826 val = tp->repair_queue;
2832 if (tp->repair_queue == TCP_SEND_QUEUE)
2833 val = tp->write_seq;
2834 else if (tp->repair_queue == TCP_RECV_QUEUE)
2840 case TCP_USER_TIMEOUT:
2841 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2845 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
2849 val = tcp_time_stamp + tp->tsoffset;
2851 case TCP_NOTSENT_LOWAT:
2852 val = tp->notsent_lowat;
2857 case TCP_SAVED_SYN: {
2858 if (get_user(len, optlen))
2862 if (tp->saved_syn) {
2863 if (len < tp->saved_syn[0]) {
2864 if (put_user(tp->saved_syn[0], optlen)) {
2871 len = tp->saved_syn[0];
2872 if (put_user(len, optlen)) {
2876 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
2880 tcp_saved_syn_free(tp);
2885 if (put_user(len, optlen))
2891 return -ENOPROTOOPT;
2894 if (put_user(len, optlen))
2896 if (copy_to_user(optval, &val, len))
2901 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2904 struct inet_connection_sock *icsk = inet_csk(sk);
2906 if (level != SOL_TCP)
2907 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2909 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2911 EXPORT_SYMBOL(tcp_getsockopt);
2913 #ifdef CONFIG_COMPAT
2914 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2915 char __user *optval, int __user *optlen)
2917 if (level != SOL_TCP)
2918 return inet_csk_compat_getsockopt(sk, level, optname,
2920 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2922 EXPORT_SYMBOL(compat_tcp_getsockopt);
2925 #ifdef CONFIG_TCP_MD5SIG
2926 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2927 static DEFINE_MUTEX(tcp_md5sig_mutex);
2928 static bool tcp_md5sig_pool_populated = false;
2930 static void __tcp_alloc_md5sig_pool(void)
2934 for_each_possible_cpu(cpu) {
2935 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2936 struct crypto_hash *hash;
2938 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2939 if (IS_ERR_OR_NULL(hash))
2941 per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
2944 /* before setting tcp_md5sig_pool_populated, we must commit all writes
2945 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
2948 tcp_md5sig_pool_populated = true;
2951 bool tcp_alloc_md5sig_pool(void)
2953 if (unlikely(!tcp_md5sig_pool_populated)) {
2954 mutex_lock(&tcp_md5sig_mutex);
2956 if (!tcp_md5sig_pool_populated)
2957 __tcp_alloc_md5sig_pool();
2959 mutex_unlock(&tcp_md5sig_mutex);
2961 return tcp_md5sig_pool_populated;
2963 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2967 * tcp_get_md5sig_pool - get md5sig_pool for this user
2969 * We use percpu structure, so if we succeed, we exit with preemption
2970 * and BH disabled, to make sure another thread or softirq handling
2971 * wont try to get same context.
2973 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2977 if (tcp_md5sig_pool_populated) {
2978 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
2980 return this_cpu_ptr(&tcp_md5sig_pool);
2985 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2987 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2988 const struct tcphdr *th)
2990 struct scatterlist sg;
2994 /* We are not allowed to change tcphdr, make a local copy */
2995 memcpy(&hdr, th, sizeof(hdr));
2998 /* options aren't included in the hash */
2999 sg_init_one(&sg, &hdr, sizeof(hdr));
3000 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3003 EXPORT_SYMBOL(tcp_md5_hash_header);
3005 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3006 const struct sk_buff *skb, unsigned int header_len)
3008 struct scatterlist sg;
3009 const struct tcphdr *tp = tcp_hdr(skb);
3010 struct hash_desc *desc = &hp->md5_desc;
3012 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3013 skb_headlen(skb) - header_len : 0;
3014 const struct skb_shared_info *shi = skb_shinfo(skb);
3015 struct sk_buff *frag_iter;
3017 sg_init_table(&sg, 1);
3019 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3020 if (crypto_hash_update(desc, &sg, head_data_len))
3023 for (i = 0; i < shi->nr_frags; ++i) {
3024 const struct skb_frag_struct *f = &shi->frags[i];
3025 unsigned int offset = f->page_offset;
3026 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3028 sg_set_page(&sg, page, skb_frag_size(f),
3029 offset_in_page(offset));
3030 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3034 skb_walk_frags(skb, frag_iter)
3035 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3040 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3042 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3044 struct scatterlist sg;
3046 sg_init_one(&sg, key->key, key->keylen);
3047 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3049 EXPORT_SYMBOL(tcp_md5_hash_key);
3053 void tcp_done(struct sock *sk)
3055 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3057 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3058 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3060 tcp_set_state(sk, TCP_CLOSE);
3061 tcp_clear_xmit_timers(sk);
3063 reqsk_fastopen_remove(sk, req, false);
3065 sk->sk_shutdown = SHUTDOWN_MASK;
3067 if (!sock_flag(sk, SOCK_DEAD))
3068 sk->sk_state_change(sk);
3070 inet_csk_destroy_sock(sk);
3072 EXPORT_SYMBOL_GPL(tcp_done);
3074 extern struct tcp_congestion_ops tcp_reno;
3076 static __initdata unsigned long thash_entries;
3077 static int __init set_thash_entries(char *str)
3084 ret = kstrtoul(str, 0, &thash_entries);
3090 __setup("thash_entries=", set_thash_entries);
3092 static void __init tcp_init_mem(void)
3094 unsigned long limit = nr_free_buffer_pages() / 16;
3096 limit = max(limit, 128UL);
3097 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3098 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3099 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3102 void __init tcp_init(void)
3104 unsigned long limit;
3105 int max_rshare, max_wshare, cnt;
3108 sock_skb_cb_check_size(sizeof(struct tcp_skb_cb));
3110 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3111 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3112 tcp_hashinfo.bind_bucket_cachep =
3113 kmem_cache_create("tcp_bind_bucket",
3114 sizeof(struct inet_bind_bucket), 0,
3115 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3117 /* Size and allocate the main established and bind bucket
3120 * The methodology is similar to that of the buffer cache.
3122 tcp_hashinfo.ehash =
3123 alloc_large_system_hash("TCP established",
3124 sizeof(struct inet_ehash_bucket),
3126 17, /* one slot per 128 KB of memory */
3129 &tcp_hashinfo.ehash_mask,
3131 thash_entries ? 0 : 512 * 1024);
3132 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3133 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3135 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3136 panic("TCP: failed to alloc ehash_locks");
3137 tcp_hashinfo.bhash =
3138 alloc_large_system_hash("TCP bind",
3139 sizeof(struct inet_bind_hashbucket),
3140 tcp_hashinfo.ehash_mask + 1,
3141 17, /* one slot per 128 KB of memory */
3143 &tcp_hashinfo.bhash_size,
3147 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3148 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3149 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3150 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3154 cnt = tcp_hashinfo.ehash_mask + 1;
3156 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3157 sysctl_tcp_max_orphans = cnt / 2;
3158 sysctl_max_syn_backlog = max(128, cnt / 256);
3161 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3162 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3163 max_wshare = min(4UL*1024*1024, limit);
3164 max_rshare = min(6UL*1024*1024, limit);
3166 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3167 sysctl_tcp_wmem[1] = 16*1024;
3168 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3170 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3171 sysctl_tcp_rmem[1] = 87380;
3172 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3174 pr_info("Hash tables configured (established %u bind %u)\n",
3175 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3178 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);