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 <asm/unaligned.h>
283 #include <net/busy_poll.h>
285 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
287 int sysctl_tcp_min_tso_segs __read_mostly = 2;
289 int sysctl_tcp_autocorking __read_mostly = 1;
291 struct percpu_counter tcp_orphan_count;
292 EXPORT_SYMBOL_GPL(tcp_orphan_count);
294 long sysctl_tcp_mem[3] __read_mostly;
295 int sysctl_tcp_wmem[3] __read_mostly;
296 int sysctl_tcp_rmem[3] __read_mostly;
298 EXPORT_SYMBOL(sysctl_tcp_mem);
299 EXPORT_SYMBOL(sysctl_tcp_rmem);
300 EXPORT_SYMBOL(sysctl_tcp_wmem);
302 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
303 EXPORT_SYMBOL(tcp_memory_allocated);
305 int sysctl_tcp_delack_seg __read_mostly = TCP_DELACK_SEG;
306 EXPORT_SYMBOL(sysctl_tcp_delack_seg);
308 int sysctl_tcp_use_userconfig __read_mostly;
309 EXPORT_SYMBOL(sysctl_tcp_use_userconfig);
312 * Current number of TCP sockets.
314 struct percpu_counter tcp_sockets_allocated;
315 EXPORT_SYMBOL(tcp_sockets_allocated);
320 struct tcp_splice_state {
321 struct pipe_inode_info *pipe;
327 * Pressure flag: try to collapse.
328 * Technical note: it is used by multiple contexts non atomically.
329 * All the __sk_mem_schedule() is of this nature: accounting
330 * is strict, actions are advisory and have some latency.
332 int tcp_memory_pressure __read_mostly;
333 EXPORT_SYMBOL(tcp_memory_pressure);
335 void tcp_enter_memory_pressure(struct sock *sk)
337 if (!tcp_memory_pressure) {
338 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
339 tcp_memory_pressure = 1;
342 EXPORT_SYMBOL(tcp_enter_memory_pressure);
344 /* Convert seconds to retransmits based on initial and max timeout */
345 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
350 int period = timeout;
353 while (seconds > period && res < 255) {
356 if (timeout > rto_max)
364 /* Convert retransmits to seconds based on initial and max timeout */
365 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
373 if (timeout > rto_max)
381 /* Address-family independent initialization for a tcp_sock.
383 * NOTE: A lot of things set to zero explicitly by call to
384 * sk_alloc() so need not be done here.
386 void tcp_init_sock(struct sock *sk)
388 struct inet_connection_sock *icsk = inet_csk(sk);
389 struct tcp_sock *tp = tcp_sk(sk);
391 tp->out_of_order_queue = RB_ROOT;
392 tcp_init_xmit_timers(sk);
393 tcp_prequeue_init(tp);
394 INIT_LIST_HEAD(&tp->tsq_node);
396 icsk->icsk_rto = TCP_TIMEOUT_INIT;
397 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
398 tp->rtt_min[0].rtt = ~0U;
400 /* So many TCP implementations out there (incorrectly) count the
401 * initial SYN frame in their delayed-ACK and congestion control
402 * algorithms that we must have the following bandaid to talk
403 * efficiently to them. -DaveM
405 tp->snd_cwnd = TCP_INIT_CWND;
407 /* See draft-stevens-tcpca-spec-01 for discussion of the
408 * initialization of these values.
410 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
411 tp->snd_cwnd_clamp = ~0;
412 tp->mss_cache = TCP_MSS_DEFAULT;
413 u64_stats_init(&tp->syncp);
415 tp->reordering = sysctl_tcp_reordering;
416 tcp_enable_early_retrans(tp);
417 tcp_assign_congestion_control(sk);
421 sk->sk_state = TCP_CLOSE;
423 sk->sk_write_space = sk_stream_write_space;
424 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
426 icsk->icsk_sync_mss = tcp_sync_mss;
428 sk->sk_sndbuf = sysctl_tcp_wmem[1];
429 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
432 sock_update_memcg(sk);
433 sk_sockets_allocated_inc(sk);
436 EXPORT_SYMBOL(tcp_init_sock);
438 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb)
440 if (sk->sk_tsflags) {
441 struct skb_shared_info *shinfo = skb_shinfo(skb);
443 sock_tx_timestamp(sk, &shinfo->tx_flags);
444 if (shinfo->tx_flags & SKBTX_ANY_TSTAMP)
445 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
450 * Wait for a TCP event.
452 * Note that we don't need to lock the socket, as the upper poll layers
453 * take care of normal races (between the test and the event) and we don't
454 * go look at any of the socket buffers directly.
456 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
459 struct sock *sk = sock->sk;
460 const struct tcp_sock *tp = tcp_sk(sk);
463 sock_rps_record_flow(sk);
465 sock_poll_wait(file, sk_sleep(sk), wait);
467 state = sk_state_load(sk);
468 if (state == TCP_LISTEN)
469 return inet_csk_listen_poll(sk);
471 /* Socket is not locked. We are protected from async events
472 * by poll logic and correct handling of state changes
473 * made by other threads is impossible in any case.
479 * POLLHUP is certainly not done right. But poll() doesn't
480 * have a notion of HUP in just one direction, and for a
481 * socket the read side is more interesting.
483 * Some poll() documentation says that POLLHUP is incompatible
484 * with the POLLOUT/POLLWR flags, so somebody should check this
485 * all. But careful, it tends to be safer to return too many
486 * bits than too few, and you can easily break real applications
487 * if you don't tell them that something has hung up!
491 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
492 * our fs/select.c). It means that after we received EOF,
493 * poll always returns immediately, making impossible poll() on write()
494 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
495 * if and only if shutdown has been made in both directions.
496 * Actually, it is interesting to look how Solaris and DUX
497 * solve this dilemma. I would prefer, if POLLHUP were maskable,
498 * then we could set it on SND_SHUTDOWN. BTW examples given
499 * in Stevens' books assume exactly this behaviour, it explains
500 * why POLLHUP is incompatible with POLLOUT. --ANK
502 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
503 * blocking on fresh not-connected or disconnected socket. --ANK
505 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
507 if (sk->sk_shutdown & RCV_SHUTDOWN)
508 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
510 /* Connected or passive Fast Open socket? */
511 if (state != TCP_SYN_SENT &&
512 (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
513 int target = sock_rcvlowat(sk, 0, INT_MAX);
515 if (tp->urg_seq == tp->copied_seq &&
516 !sock_flag(sk, SOCK_URGINLINE) &&
520 if (tp->rcv_nxt - tp->copied_seq >= target)
521 mask |= POLLIN | POLLRDNORM;
523 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
524 if (sk_stream_is_writeable(sk)) {
525 mask |= POLLOUT | POLLWRNORM;
526 } else { /* send SIGIO later */
527 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
528 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
530 /* Race breaker. If space is freed after
531 * wspace test but before the flags are set,
532 * IO signal will be lost. Memory barrier
533 * pairs with the input side.
535 smp_mb__after_atomic();
536 if (sk_stream_is_writeable(sk))
537 mask |= POLLOUT | POLLWRNORM;
540 mask |= POLLOUT | POLLWRNORM;
542 if (tp->urg_data & TCP_URG_VALID)
545 /* This barrier is coupled with smp_wmb() in tcp_reset() */
547 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
552 EXPORT_SYMBOL(tcp_poll);
554 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
556 struct tcp_sock *tp = tcp_sk(sk);
562 if (sk->sk_state == TCP_LISTEN)
565 slow = lock_sock_fast(sk);
566 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
568 else if (sock_flag(sk, SOCK_URGINLINE) ||
570 before(tp->urg_seq, tp->copied_seq) ||
571 !before(tp->urg_seq, tp->rcv_nxt)) {
573 answ = tp->rcv_nxt - tp->copied_seq;
575 /* Subtract 1, if FIN was received */
576 if (answ && sock_flag(sk, SOCK_DONE))
579 answ = tp->urg_seq - tp->copied_seq;
580 unlock_sock_fast(sk, slow);
583 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
586 if (sk->sk_state == TCP_LISTEN)
589 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
592 answ = tp->write_seq - tp->snd_una;
595 if (sk->sk_state == TCP_LISTEN)
598 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
601 answ = tp->write_seq - tp->snd_nxt;
607 return put_user(answ, (int __user *)arg);
609 EXPORT_SYMBOL(tcp_ioctl);
611 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
613 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
614 tp->pushed_seq = tp->write_seq;
617 static inline bool forced_push(const struct tcp_sock *tp)
619 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
622 static void skb_entail(struct sock *sk, struct sk_buff *skb)
624 struct tcp_sock *tp = tcp_sk(sk);
625 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
628 tcb->seq = tcb->end_seq = tp->write_seq;
629 tcb->tcp_flags = TCPHDR_ACK;
631 __skb_header_release(skb);
632 tcp_add_write_queue_tail(sk, skb);
633 sk->sk_wmem_queued += skb->truesize;
634 sk_mem_charge(sk, skb->truesize);
635 if (tp->nonagle & TCP_NAGLE_PUSH)
636 tp->nonagle &= ~TCP_NAGLE_PUSH;
638 tcp_slow_start_after_idle_check(sk);
641 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
644 tp->snd_up = tp->write_seq;
647 /* If a not yet filled skb is pushed, do not send it if
648 * we have data packets in Qdisc or NIC queues :
649 * Because TX completion will happen shortly, it gives a chance
650 * to coalesce future sendmsg() payload into this skb, without
651 * need for a timer, and with no latency trade off.
652 * As packets containing data payload have a bigger truesize
653 * than pure acks (dataless) packets, the last checks prevent
654 * autocorking if we only have an ACK in Qdisc/NIC queues,
655 * or if TX completion was delayed after we processed ACK packet.
657 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
660 return skb->len < size_goal &&
661 sysctl_tcp_autocorking &&
662 skb != tcp_write_queue_head(sk) &&
663 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
666 static void tcp_push(struct sock *sk, int flags, int mss_now,
667 int nonagle, int size_goal)
669 struct tcp_sock *tp = tcp_sk(sk);
672 if (!tcp_send_head(sk))
675 skb = tcp_write_queue_tail(sk);
676 if (!(flags & MSG_MORE) || forced_push(tp))
677 tcp_mark_push(tp, skb);
679 tcp_mark_urg(tp, flags);
681 if (tcp_should_autocork(sk, skb, size_goal)) {
683 /* avoid atomic op if TSQ_THROTTLED bit is already set */
684 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
685 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
686 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
688 /* It is possible TX completion already happened
689 * before we set TSQ_THROTTLED.
691 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
695 if (flags & MSG_MORE)
696 nonagle = TCP_NAGLE_CORK;
698 __tcp_push_pending_frames(sk, mss_now, nonagle);
701 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
702 unsigned int offset, size_t len)
704 struct tcp_splice_state *tss = rd_desc->arg.data;
707 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
708 min(rd_desc->count, len), tss->flags,
711 rd_desc->count -= ret;
715 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
717 /* Store TCP splice context information in read_descriptor_t. */
718 read_descriptor_t rd_desc = {
723 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
727 * tcp_splice_read - splice data from TCP socket to a pipe
728 * @sock: socket to splice from
729 * @ppos: position (not valid)
730 * @pipe: pipe to splice to
731 * @len: number of bytes to splice
732 * @flags: splice modifier flags
735 * Will read pages from given socket and fill them into a pipe.
738 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
739 struct pipe_inode_info *pipe, size_t len,
742 struct sock *sk = sock->sk;
743 struct tcp_splice_state tss = {
752 sock_rps_record_flow(sk);
754 * We can't seek on a socket input
763 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
765 ret = __tcp_splice_read(sk, &tss);
771 if (sock_flag(sk, SOCK_DONE))
774 ret = sock_error(sk);
777 if (sk->sk_shutdown & RCV_SHUTDOWN)
779 if (sk->sk_state == TCP_CLOSE) {
781 * This occurs when user tries to read
782 * from never connected socket.
784 if (!sock_flag(sk, SOCK_DONE))
792 /* if __tcp_splice_read() got nothing while we have
793 * an skb in receive queue, we do not want to loop.
794 * This might happen with URG data.
796 if (!skb_queue_empty(&sk->sk_receive_queue))
798 sk_wait_data(sk, &timeo, NULL);
799 if (signal_pending(current)) {
800 ret = sock_intr_errno(timeo);
813 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
814 (sk->sk_shutdown & RCV_SHUTDOWN) ||
815 signal_pending(current))
826 EXPORT_SYMBOL(tcp_splice_read);
828 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
833 /* The TCP header must be at least 32-bit aligned. */
834 size = ALIGN(size, 4);
836 if (unlikely(tcp_under_memory_pressure(sk)))
837 sk_mem_reclaim_partial(sk);
839 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
843 if (force_schedule) {
844 mem_scheduled = true;
845 sk_forced_mem_schedule(sk, skb->truesize);
847 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
849 if (likely(mem_scheduled)) {
850 skb_reserve(skb, sk->sk_prot->max_header);
852 * Make sure that we have exactly size bytes
853 * available to the caller, no more, no less.
855 skb->reserved_tailroom = skb->end - skb->tail - size;
860 sk->sk_prot->enter_memory_pressure(sk);
861 sk_stream_moderate_sndbuf(sk);
866 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
869 struct tcp_sock *tp = tcp_sk(sk);
870 u32 new_size_goal, size_goal;
872 if (!large_allowed || !sk_can_gso(sk))
875 /* Note : tcp_tso_autosize() will eventually split this later */
876 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
877 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
879 /* We try hard to avoid divides here */
880 size_goal = tp->gso_segs * mss_now;
881 if (unlikely(new_size_goal < size_goal ||
882 new_size_goal >= size_goal + mss_now)) {
883 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
884 sk->sk_gso_max_segs);
885 size_goal = tp->gso_segs * mss_now;
888 return max(size_goal, mss_now);
891 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
895 mss_now = tcp_current_mss(sk);
896 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
901 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
902 size_t size, int flags)
904 struct tcp_sock *tp = tcp_sk(sk);
905 int mss_now, size_goal;
908 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
910 /* Wait for a connection to finish. One exception is TCP Fast Open
911 * (passive side) where data is allowed to be sent before a connection
912 * is fully established.
914 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
915 !tcp_passive_fastopen(sk)) {
916 err = sk_stream_wait_connect(sk, &timeo);
921 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
923 mss_now = tcp_send_mss(sk, &size_goal, flags);
927 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
931 struct sk_buff *skb = tcp_write_queue_tail(sk);
935 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
937 if (!sk_stream_memory_free(sk))
938 goto wait_for_sndbuf;
940 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
941 skb_queue_empty(&sk->sk_write_queue));
943 goto wait_for_memory;
952 i = skb_shinfo(skb)->nr_frags;
953 can_coalesce = skb_can_coalesce(skb, i, page, offset);
954 if (!can_coalesce && i >= sysctl_max_skb_frags) {
955 tcp_mark_push(tp, skb);
958 if (!sk_wmem_schedule(sk, copy))
959 goto wait_for_memory;
962 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
965 skb_fill_page_desc(skb, i, page, offset, copy);
967 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
970 skb->data_len += copy;
971 skb->truesize += copy;
972 sk->sk_wmem_queued += copy;
973 sk_mem_charge(sk, copy);
974 skb->ip_summed = CHECKSUM_PARTIAL;
975 tp->write_seq += copy;
976 TCP_SKB_CB(skb)->end_seq += copy;
977 tcp_skb_pcount_set(skb, 0);
980 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
986 tcp_tx_timestamp(sk, skb);
990 if (skb->len < size_goal || (flags & MSG_OOB))
993 if (forced_push(tp)) {
994 tcp_mark_push(tp, skb);
995 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
996 } else if (skb == tcp_send_head(sk))
997 tcp_push_one(sk, mss_now);
1001 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1003 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1004 TCP_NAGLE_PUSH, size_goal);
1006 err = sk_stream_wait_memory(sk, &timeo);
1010 mss_now = tcp_send_mss(sk, &size_goal, flags);
1014 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
1015 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1022 /* make sure we wake any epoll edge trigger waiter */
1023 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1024 sk->sk_write_space(sk);
1025 return sk_stream_error(sk, flags, err);
1028 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1029 size_t size, int flags)
1033 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1034 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1035 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1039 res = do_tcp_sendpages(sk, page, offset, size, flags);
1043 EXPORT_SYMBOL(tcp_sendpage);
1045 static inline int select_size(const struct sock *sk, bool sg)
1047 const struct tcp_sock *tp = tcp_sk(sk);
1048 int tmp = tp->mss_cache;
1051 if (sk_can_gso(sk)) {
1052 /* Small frames wont use a full page:
1053 * Payload will immediately follow tcp header.
1055 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1057 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1059 if (tmp >= pgbreak &&
1060 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1068 void tcp_free_fastopen_req(struct tcp_sock *tp)
1070 if (tp->fastopen_req) {
1071 kfree(tp->fastopen_req);
1072 tp->fastopen_req = NULL;
1076 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1077 int *copied, size_t size)
1079 struct tcp_sock *tp = tcp_sk(sk);
1080 struct sockaddr *uaddr = msg->msg_name;
1083 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1084 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1085 uaddr->sa_family == AF_UNSPEC))
1087 if (tp->fastopen_req)
1088 return -EALREADY; /* Another Fast Open is in progress */
1090 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1092 if (unlikely(!tp->fastopen_req))
1094 tp->fastopen_req->data = msg;
1095 tp->fastopen_req->size = size;
1097 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1098 err = __inet_stream_connect(sk->sk_socket, uaddr,
1099 msg->msg_namelen, flags);
1100 *copied = tp->fastopen_req->copied;
1101 tcp_free_fastopen_req(tp);
1105 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1107 struct tcp_sock *tp = tcp_sk(sk);
1108 struct sk_buff *skb;
1109 int flags, err, copied = 0;
1110 int mss_now = 0, size_goal, copied_syn = 0;
1116 flags = msg->msg_flags;
1117 if ((flags & MSG_FASTOPEN) && !tp->repair) {
1118 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1119 if (err == -EINPROGRESS && copied_syn > 0)
1125 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1127 /* Wait for a connection to finish. One exception is TCP Fast Open
1128 * (passive side) where data is allowed to be sent before a connection
1129 * is fully established.
1131 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1132 !tcp_passive_fastopen(sk)) {
1133 err = sk_stream_wait_connect(sk, &timeo);
1138 if (unlikely(tp->repair)) {
1139 if (tp->repair_queue == TCP_RECV_QUEUE) {
1140 copied = tcp_send_rcvq(sk, msg, size);
1145 if (tp->repair_queue == TCP_NO_QUEUE)
1148 /* 'common' sending to sendq */
1151 /* This should be in poll */
1152 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1154 mss_now = tcp_send_mss(sk, &size_goal, flags);
1156 /* Ok commence sending. */
1160 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1163 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1165 while (msg_data_left(msg)) {
1167 int max = size_goal;
1169 skb = tcp_write_queue_tail(sk);
1170 if (tcp_send_head(sk)) {
1171 if (skb->ip_summed == CHECKSUM_NONE)
1173 copy = max - skb->len;
1178 /* Allocate new segment. If the interface is SG,
1179 * allocate skb fitting to single page.
1181 if (!sk_stream_memory_free(sk))
1182 goto wait_for_sndbuf;
1184 skb = sk_stream_alloc_skb(sk,
1185 select_size(sk, sg),
1187 skb_queue_empty(&sk->sk_write_queue));
1189 goto wait_for_memory;
1192 * Check whether we can use HW checksum.
1194 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1195 skb->ip_summed = CHECKSUM_PARTIAL;
1197 skb_entail(sk, skb);
1201 /* All packets are restored as if they have
1202 * already been sent. skb_mstamp isn't set to
1203 * avoid wrong rtt estimation.
1206 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1209 /* Try to append data to the end of skb. */
1210 if (copy > msg_data_left(msg))
1211 copy = msg_data_left(msg);
1213 /* Where to copy to? */
1214 if (skb_availroom(skb) > 0) {
1215 /* We have some space in skb head. Superb! */
1216 copy = min_t(int, copy, skb_availroom(skb));
1217 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1222 int i = skb_shinfo(skb)->nr_frags;
1223 struct page_frag *pfrag = sk_page_frag(sk);
1225 if (!sk_page_frag_refill(sk, pfrag))
1226 goto wait_for_memory;
1228 if (!skb_can_coalesce(skb, i, pfrag->page,
1230 if (i >= sysctl_max_skb_frags || !sg) {
1231 tcp_mark_push(tp, skb);
1237 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1239 if (!sk_wmem_schedule(sk, copy))
1240 goto wait_for_memory;
1242 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1249 /* Update the skb. */
1251 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1253 skb_fill_page_desc(skb, i, pfrag->page,
1254 pfrag->offset, copy);
1255 get_page(pfrag->page);
1257 pfrag->offset += copy;
1261 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1263 tp->write_seq += copy;
1264 TCP_SKB_CB(skb)->end_seq += copy;
1265 tcp_skb_pcount_set(skb, 0);
1268 if (!msg_data_left(msg)) {
1269 tcp_tx_timestamp(sk, skb);
1273 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1276 if (forced_push(tp)) {
1277 tcp_mark_push(tp, skb);
1278 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1279 } else if (skb == tcp_send_head(sk))
1280 tcp_push_one(sk, mss_now);
1284 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1287 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1288 TCP_NAGLE_PUSH, size_goal);
1290 err = sk_stream_wait_memory(sk, &timeo);
1294 mss_now = tcp_send_mss(sk, &size_goal, flags);
1299 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1302 return copied + copied_syn;
1306 tcp_unlink_write_queue(skb, sk);
1307 /* It is the one place in all of TCP, except connection
1308 * reset, where we can be unlinking the send_head.
1310 tcp_check_send_head(sk, skb);
1311 sk_wmem_free_skb(sk, skb);
1315 if (copied + copied_syn)
1318 err = sk_stream_error(sk, flags, err);
1319 /* make sure we wake any epoll edge trigger waiter */
1320 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1321 sk->sk_write_space(sk);
1325 EXPORT_SYMBOL(tcp_sendmsg);
1328 * Handle reading urgent data. BSD has very simple semantics for
1329 * this, no blocking and very strange errors 8)
1332 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1334 struct tcp_sock *tp = tcp_sk(sk);
1336 /* No URG data to read. */
1337 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1338 tp->urg_data == TCP_URG_READ)
1339 return -EINVAL; /* Yes this is right ! */
1341 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1344 if (tp->urg_data & TCP_URG_VALID) {
1346 char c = tp->urg_data;
1348 if (!(flags & MSG_PEEK))
1349 tp->urg_data = TCP_URG_READ;
1351 /* Read urgent data. */
1352 msg->msg_flags |= MSG_OOB;
1355 if (!(flags & MSG_TRUNC))
1356 err = memcpy_to_msg(msg, &c, 1);
1359 msg->msg_flags |= MSG_TRUNC;
1361 return err ? -EFAULT : len;
1364 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1367 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1368 * the available implementations agree in this case:
1369 * this call should never block, independent of the
1370 * blocking state of the socket.
1371 * Mike <pall@rz.uni-karlsruhe.de>
1376 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1378 struct sk_buff *skb;
1379 int copied = 0, err = 0;
1381 /* XXX -- need to support SO_PEEK_OFF */
1383 skb_queue_walk(&sk->sk_write_queue, skb) {
1384 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1391 return err ?: copied;
1394 /* Clean up the receive buffer for full frames taken by the user,
1395 * then send an ACK if necessary. COPIED is the number of bytes
1396 * tcp_recvmsg has given to the user so far, it speeds up the
1397 * calculation of whether or not we must ACK for the sake of
1400 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1402 struct tcp_sock *tp = tcp_sk(sk);
1403 bool time_to_ack = false;
1405 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1407 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1408 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1409 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1411 if (inet_csk_ack_scheduled(sk)) {
1412 const struct inet_connection_sock *icsk = inet_csk(sk);
1413 /* Delayed ACKs frequently hit locked sockets during bulk
1415 if (icsk->icsk_ack.blocked ||
1416 /* Once-per-sysctl_tcp_delack_seg segments
1417 * ACK was not sent by tcp_input.c
1419 tp->rcv_nxt - tp->rcv_wup > (icsk->icsk_ack.rcv_mss) *
1420 sysctl_tcp_delack_seg ||
1422 * If this read emptied read buffer, we send ACK, if
1423 * connection is not bidirectional, user drained
1424 * receive buffer and there was a small segment
1428 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1429 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1430 !icsk->icsk_ack.pingpong)) &&
1431 !atomic_read(&sk->sk_rmem_alloc)))
1435 /* We send an ACK if we can now advertise a non-zero window
1436 * which has been raised "significantly".
1438 * Even if window raised up to infinity, do not send window open ACK
1439 * in states, where we will not receive more. It is useless.
1441 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1442 __u32 rcv_window_now = tcp_receive_window(tp);
1444 /* Optimize, __tcp_select_window() is not cheap. */
1445 if (2*rcv_window_now <= tp->window_clamp) {
1446 __u32 new_window = __tcp_select_window(sk);
1448 /* Send ACK now, if this read freed lots of space
1449 * in our buffer. Certainly, new_window is new window.
1450 * We can advertise it now, if it is not less than current one.
1451 * "Lots" means "at least twice" here.
1453 if (new_window && new_window >= 2 * rcv_window_now)
1461 static void tcp_prequeue_process(struct sock *sk)
1463 struct sk_buff *skb;
1464 struct tcp_sock *tp = tcp_sk(sk);
1466 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1468 /* RX process wants to run with disabled BHs, though it is not
1471 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1472 sk_backlog_rcv(sk, skb);
1475 /* Clear memory counter. */
1476 tp->ucopy.memory = 0;
1479 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1481 struct sk_buff *skb;
1484 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1485 offset = seq - TCP_SKB_CB(skb)->seq;
1486 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1488 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1492 /* This looks weird, but this can happen if TCP collapsing
1493 * splitted a fat GRO packet, while we released socket lock
1494 * in skb_splice_bits()
1496 sk_eat_skb(sk, skb);
1502 * This routine provides an alternative to tcp_recvmsg() for routines
1503 * that would like to handle copying from skbuffs directly in 'sendfile'
1506 * - It is assumed that the socket was locked by the caller.
1507 * - The routine does not block.
1508 * - At present, there is no support for reading OOB data
1509 * or for 'peeking' the socket using this routine
1510 * (although both would be easy to implement).
1512 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1513 sk_read_actor_t recv_actor)
1515 struct sk_buff *skb;
1516 struct tcp_sock *tp = tcp_sk(sk);
1517 u32 seq = tp->copied_seq;
1521 if (sk->sk_state == TCP_LISTEN)
1523 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1524 if (offset < skb->len) {
1528 len = skb->len - offset;
1529 /* Stop reading if we hit a patch of urgent data */
1531 u32 urg_offset = tp->urg_seq - seq;
1532 if (urg_offset < len)
1537 used = recv_actor(desc, skb, offset, len);
1542 } else if (used <= len) {
1547 /* If recv_actor drops the lock (e.g. TCP splice
1548 * receive) the skb pointer might be invalid when
1549 * getting here: tcp_collapse might have deleted it
1550 * while aggregating skbs from the socket queue.
1552 skb = tcp_recv_skb(sk, seq - 1, &offset);
1555 /* TCP coalescing might have appended data to the skb.
1556 * Try to splice more frags
1558 if (offset + 1 != skb->len)
1561 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1562 sk_eat_skb(sk, skb);
1566 sk_eat_skb(sk, skb);
1569 tp->copied_seq = seq;
1571 tp->copied_seq = seq;
1573 tcp_rcv_space_adjust(sk);
1575 /* Clean up data we have read: This will do ACK frames. */
1577 tcp_recv_skb(sk, seq, &offset);
1578 tcp_cleanup_rbuf(sk, copied);
1582 EXPORT_SYMBOL(tcp_read_sock);
1585 * This routine copies from a sock struct into the user buffer.
1587 * Technical note: in 2.3 we work on _locked_ socket, so that
1588 * tricks with *seq access order and skb->users are not required.
1589 * Probably, code can be easily improved even more.
1592 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1593 int flags, int *addr_len)
1595 struct tcp_sock *tp = tcp_sk(sk);
1601 int target; /* Read at least this many bytes */
1603 struct task_struct *user_recv = NULL;
1604 struct sk_buff *skb, *last;
1607 if (unlikely(flags & MSG_ERRQUEUE))
1608 return inet_recv_error(sk, msg, len, addr_len);
1610 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1611 (sk->sk_state == TCP_ESTABLISHED))
1612 sk_busy_loop(sk, nonblock);
1617 if (sk->sk_state == TCP_LISTEN)
1620 timeo = sock_rcvtimeo(sk, nonblock);
1622 /* Urgent data needs to be handled specially. */
1623 if (flags & MSG_OOB)
1626 if (unlikely(tp->repair)) {
1628 if (!(flags & MSG_PEEK))
1631 if (tp->repair_queue == TCP_SEND_QUEUE)
1635 if (tp->repair_queue == TCP_NO_QUEUE)
1638 /* 'common' recv queue MSG_PEEK-ing */
1641 seq = &tp->copied_seq;
1642 if (flags & MSG_PEEK) {
1643 peek_seq = tp->copied_seq;
1647 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1652 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1653 if (tp->urg_data && tp->urg_seq == *seq) {
1656 if (signal_pending(current)) {
1657 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1662 /* Next get a buffer. */
1664 last = skb_peek_tail(&sk->sk_receive_queue);
1665 skb_queue_walk(&sk->sk_receive_queue, skb) {
1667 /* Now that we have two receive queues this
1670 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1671 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
1672 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1676 offset = *seq - TCP_SKB_CB(skb)->seq;
1677 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1679 if (offset < skb->len)
1681 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1683 WARN(!(flags & MSG_PEEK),
1684 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
1685 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1688 /* Well, if we have backlog, try to process it now yet. */
1690 if (copied >= target && !sk->sk_backlog.tail)
1695 sk->sk_state == TCP_CLOSE ||
1696 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1698 signal_pending(current))
1701 if (sock_flag(sk, SOCK_DONE))
1705 copied = sock_error(sk);
1709 if (sk->sk_shutdown & RCV_SHUTDOWN)
1712 if (sk->sk_state == TCP_CLOSE) {
1713 if (!sock_flag(sk, SOCK_DONE)) {
1714 /* This occurs when user tries to read
1715 * from never connected socket.
1728 if (signal_pending(current)) {
1729 copied = sock_intr_errno(timeo);
1734 tcp_cleanup_rbuf(sk, copied);
1736 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1737 /* Install new reader */
1738 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1739 user_recv = current;
1740 tp->ucopy.task = user_recv;
1741 tp->ucopy.msg = msg;
1744 tp->ucopy.len = len;
1746 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1747 !(flags & (MSG_PEEK | MSG_TRUNC)));
1749 /* Ugly... If prequeue is not empty, we have to
1750 * process it before releasing socket, otherwise
1751 * order will be broken at second iteration.
1752 * More elegant solution is required!!!
1754 * Look: we have the following (pseudo)queues:
1756 * 1. packets in flight
1761 * Each queue can be processed only if the next ones
1762 * are empty. At this point we have empty receive_queue.
1763 * But prequeue _can_ be not empty after 2nd iteration,
1764 * when we jumped to start of loop because backlog
1765 * processing added something to receive_queue.
1766 * We cannot release_sock(), because backlog contains
1767 * packets arrived _after_ prequeued ones.
1769 * Shortly, algorithm is clear --- to process all
1770 * the queues in order. We could make it more directly,
1771 * requeueing packets from backlog to prequeue, if
1772 * is not empty. It is more elegant, but eats cycles,
1775 if (!skb_queue_empty(&tp->ucopy.prequeue))
1778 /* __ Set realtime policy in scheduler __ */
1781 if (copied >= target) {
1782 /* Do not sleep, just process backlog. */
1786 sk_wait_data(sk, &timeo, last);
1792 /* __ Restore normal policy in scheduler __ */
1794 chunk = len - tp->ucopy.len;
1796 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1801 if (tp->rcv_nxt == tp->copied_seq &&
1802 !skb_queue_empty(&tp->ucopy.prequeue)) {
1804 tcp_prequeue_process(sk);
1806 chunk = len - tp->ucopy.len;
1808 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1814 if ((flags & MSG_PEEK) &&
1815 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1816 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1818 task_pid_nr(current));
1819 peek_seq = tp->copied_seq;
1824 /* Ok so how much can we use? */
1825 used = skb->len - offset;
1829 /* Do we have urgent data here? */
1831 u32 urg_offset = tp->urg_seq - *seq;
1832 if (urg_offset < used) {
1834 if (!sock_flag(sk, SOCK_URGINLINE)) {
1847 if (!(flags & MSG_TRUNC)) {
1848 err = skb_copy_datagram_msg(skb, offset, msg, used);
1850 /* Exception. Bailout! */
1861 tcp_rcv_space_adjust(sk);
1864 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1866 tcp_fast_path_check(sk);
1868 if (used + offset < skb->len)
1871 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1873 if (!(flags & MSG_PEEK))
1874 sk_eat_skb(sk, skb);
1878 /* Process the FIN. */
1880 if (!(flags & MSG_PEEK))
1881 sk_eat_skb(sk, skb);
1886 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1889 tp->ucopy.len = copied > 0 ? len : 0;
1891 tcp_prequeue_process(sk);
1893 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1894 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1900 tp->ucopy.task = NULL;
1904 /* According to UNIX98, msg_name/msg_namelen are ignored
1905 * on connected socket. I was just happy when found this 8) --ANK
1908 /* Clean up data we have read: This will do ACK frames. */
1909 tcp_cleanup_rbuf(sk, copied);
1919 err = tcp_recv_urg(sk, msg, len, flags);
1923 err = tcp_peek_sndq(sk, msg, len);
1926 EXPORT_SYMBOL(tcp_recvmsg);
1928 void tcp_set_state(struct sock *sk, int state)
1930 int oldstate = sk->sk_state;
1933 case TCP_ESTABLISHED:
1934 if (oldstate != TCP_ESTABLISHED)
1935 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1939 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1940 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1942 sk->sk_prot->unhash(sk);
1943 if (inet_csk(sk)->icsk_bind_hash &&
1944 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1948 if (oldstate == TCP_ESTABLISHED)
1949 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1952 /* Change state AFTER socket is unhashed to avoid closed
1953 * socket sitting in hash tables.
1955 sk_state_store(sk, state);
1958 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1961 EXPORT_SYMBOL_GPL(tcp_set_state);
1964 * State processing on a close. This implements the state shift for
1965 * sending our FIN frame. Note that we only send a FIN for some
1966 * states. A shutdown() may have already sent the FIN, or we may be
1970 static const unsigned char new_state[16] = {
1971 /* current state: new state: action: */
1972 [0 /* (Invalid) */] = TCP_CLOSE,
1973 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1974 [TCP_SYN_SENT] = TCP_CLOSE,
1975 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1976 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1977 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1978 [TCP_TIME_WAIT] = TCP_CLOSE,
1979 [TCP_CLOSE] = TCP_CLOSE,
1980 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
1981 [TCP_LAST_ACK] = TCP_LAST_ACK,
1982 [TCP_LISTEN] = TCP_CLOSE,
1983 [TCP_CLOSING] = TCP_CLOSING,
1984 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
1987 static int tcp_close_state(struct sock *sk)
1989 int next = (int)new_state[sk->sk_state];
1990 int ns = next & TCP_STATE_MASK;
1992 tcp_set_state(sk, ns);
1994 return next & TCP_ACTION_FIN;
1998 * Shutdown the sending side of a connection. Much like close except
1999 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2002 void tcp_shutdown(struct sock *sk, int how)
2004 /* We need to grab some memory, and put together a FIN,
2005 * and then put it into the queue to be sent.
2006 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2008 if (!(how & SEND_SHUTDOWN))
2011 /* If we've already sent a FIN, or it's a closed state, skip this. */
2012 if ((1 << sk->sk_state) &
2013 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2014 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2015 /* Clear out any half completed packets. FIN if needed. */
2016 if (tcp_close_state(sk))
2020 EXPORT_SYMBOL(tcp_shutdown);
2022 bool tcp_check_oom(struct sock *sk, int shift)
2024 bool too_many_orphans, out_of_socket_memory;
2026 too_many_orphans = tcp_too_many_orphans(sk, shift);
2027 out_of_socket_memory = tcp_out_of_memory(sk);
2029 if (too_many_orphans)
2030 net_info_ratelimited("too many orphaned sockets\n");
2031 if (out_of_socket_memory)
2032 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2033 return too_many_orphans || out_of_socket_memory;
2036 void tcp_close(struct sock *sk, long timeout)
2038 struct sk_buff *skb;
2039 int data_was_unread = 0;
2043 sk->sk_shutdown = SHUTDOWN_MASK;
2045 if (sk->sk_state == TCP_LISTEN) {
2046 tcp_set_state(sk, TCP_CLOSE);
2049 inet_csk_listen_stop(sk);
2051 goto adjudge_to_death;
2054 /* We need to flush the recv. buffs. We do this only on the
2055 * descriptor close, not protocol-sourced closes, because the
2056 * reader process may not have drained the data yet!
2058 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2059 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2061 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2063 data_was_unread += len;
2069 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2070 if (sk->sk_state == TCP_CLOSE)
2071 goto adjudge_to_death;
2073 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2074 * data was lost. To witness the awful effects of the old behavior of
2075 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2076 * GET in an FTP client, suspend the process, wait for the client to
2077 * advertise a zero window, then kill -9 the FTP client, wheee...
2078 * Note: timeout is always zero in such a case.
2080 if (unlikely(tcp_sk(sk)->repair)) {
2081 sk->sk_prot->disconnect(sk, 0);
2082 } else if (data_was_unread) {
2083 /* Unread data was tossed, zap the connection. */
2084 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2085 tcp_set_state(sk, TCP_CLOSE);
2086 tcp_send_active_reset(sk, sk->sk_allocation);
2087 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2088 /* Check zero linger _after_ checking for unread data. */
2089 sk->sk_prot->disconnect(sk, 0);
2090 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2091 } else if (tcp_close_state(sk)) {
2092 /* We FIN if the application ate all the data before
2093 * zapping the connection.
2096 /* RED-PEN. Formally speaking, we have broken TCP state
2097 * machine. State transitions:
2099 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2100 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2101 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2103 * are legal only when FIN has been sent (i.e. in window),
2104 * rather than queued out of window. Purists blame.
2106 * F.e. "RFC state" is ESTABLISHED,
2107 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2109 * The visible declinations are that sometimes
2110 * we enter time-wait state, when it is not required really
2111 * (harmless), do not send active resets, when they are
2112 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2113 * they look as CLOSING or LAST_ACK for Linux)
2114 * Probably, I missed some more holelets.
2116 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2117 * in a single packet! (May consider it later but will
2118 * probably need API support or TCP_CORK SYN-ACK until
2119 * data is written and socket is closed.)
2124 sk_stream_wait_close(sk, timeout);
2127 state = sk->sk_state;
2131 /* It is the last release_sock in its life. It will remove backlog. */
2135 /* Now socket is owned by kernel and we acquire BH lock
2136 to finish close. No need to check for user refs.
2140 WARN_ON(sock_owned_by_user(sk));
2142 percpu_counter_inc(sk->sk_prot->orphan_count);
2144 /* Have we already been destroyed by a softirq or backlog? */
2145 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2148 /* This is a (useful) BSD violating of the RFC. There is a
2149 * problem with TCP as specified in that the other end could
2150 * keep a socket open forever with no application left this end.
2151 * We use a 1 minute timeout (about the same as BSD) then kill
2152 * our end. If they send after that then tough - BUT: long enough
2153 * that we won't make the old 4*rto = almost no time - whoops
2156 * Nope, it was not mistake. It is really desired behaviour
2157 * f.e. on http servers, when such sockets are useless, but
2158 * consume significant resources. Let's do it with special
2159 * linger2 option. --ANK
2162 if (sk->sk_state == TCP_FIN_WAIT2) {
2163 struct tcp_sock *tp = tcp_sk(sk);
2164 if (tp->linger2 < 0) {
2165 tcp_set_state(sk, TCP_CLOSE);
2166 tcp_send_active_reset(sk, GFP_ATOMIC);
2167 NET_INC_STATS_BH(sock_net(sk),
2168 LINUX_MIB_TCPABORTONLINGER);
2170 const int tmo = tcp_fin_time(sk);
2172 if (tmo > TCP_TIMEWAIT_LEN) {
2173 inet_csk_reset_keepalive_timer(sk,
2174 tmo - TCP_TIMEWAIT_LEN);
2176 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2181 if (sk->sk_state != TCP_CLOSE) {
2183 if (tcp_check_oom(sk, 0)) {
2184 tcp_set_state(sk, TCP_CLOSE);
2185 tcp_send_active_reset(sk, GFP_ATOMIC);
2186 NET_INC_STATS_BH(sock_net(sk),
2187 LINUX_MIB_TCPABORTONMEMORY);
2188 } else if (!check_net(sock_net(sk))) {
2189 /* Not possible to send reset; just close */
2190 tcp_set_state(sk, TCP_CLOSE);
2194 if (sk->sk_state == TCP_CLOSE) {
2195 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2196 /* We could get here with a non-NULL req if the socket is
2197 * aborted (e.g., closed with unread data) before 3WHS
2201 reqsk_fastopen_remove(sk, req, false);
2202 inet_csk_destroy_sock(sk);
2204 /* Otherwise, socket is reprieved until protocol close. */
2211 EXPORT_SYMBOL(tcp_close);
2213 /* These states need RST on ABORT according to RFC793 */
2215 static inline bool tcp_need_reset(int state)
2217 return (1 << state) &
2218 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2219 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2222 int tcp_disconnect(struct sock *sk, int flags)
2224 struct inet_sock *inet = inet_sk(sk);
2225 struct inet_connection_sock *icsk = inet_csk(sk);
2226 struct tcp_sock *tp = tcp_sk(sk);
2228 int old_state = sk->sk_state;
2230 if (old_state != TCP_CLOSE)
2231 tcp_set_state(sk, TCP_CLOSE);
2233 /* ABORT function of RFC793 */
2234 if (old_state == TCP_LISTEN) {
2235 inet_csk_listen_stop(sk);
2236 } else if (unlikely(tp->repair)) {
2237 sk->sk_err = ECONNABORTED;
2238 } else if (tcp_need_reset(old_state) ||
2239 (tp->snd_nxt != tp->write_seq &&
2240 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2241 /* The last check adjusts for discrepancy of Linux wrt. RFC
2244 tcp_send_active_reset(sk, gfp_any());
2245 sk->sk_err = ECONNRESET;
2246 } else if (old_state == TCP_SYN_SENT)
2247 sk->sk_err = ECONNRESET;
2249 tcp_clear_xmit_timers(sk);
2250 __skb_queue_purge(&sk->sk_receive_queue);
2251 tcp_write_queue_purge(sk);
2252 skb_rbtree_purge(&tp->out_of_order_queue);
2254 inet->inet_dport = 0;
2256 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2257 inet_reset_saddr(sk);
2259 sk->sk_shutdown = 0;
2260 sock_reset_flag(sk, SOCK_DONE);
2262 tp->write_seq += tp->max_window + 2;
2263 if (tp->write_seq == 0)
2266 icsk->icsk_probes_out = 0;
2267 tp->packets_out = 0;
2268 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2269 tp->snd_cwnd_cnt = 0;
2270 tp->window_clamp = 0;
2271 tcp_set_ca_state(sk, TCP_CA_Open);
2272 tcp_clear_retrans(tp);
2273 tp->total_retrans = 0;
2274 inet_csk_delack_init(sk);
2275 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2276 * issue in __tcp_select_window()
2278 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2279 tcp_init_send_head(sk);
2280 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2282 dst_release(sk->sk_rx_dst);
2283 sk->sk_rx_dst = NULL;
2284 tcp_saved_syn_free(tp);
2287 tp->bytes_acked = 0;
2288 tp->bytes_received = 0;
2290 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2292 if (sk->sk_frag.page) {
2293 put_page(sk->sk_frag.page);
2294 sk->sk_frag.page = NULL;
2295 sk->sk_frag.offset = 0;
2298 sk->sk_error_report(sk);
2301 EXPORT_SYMBOL(tcp_disconnect);
2303 static inline bool tcp_can_repair_sock(const struct sock *sk)
2305 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2306 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2309 static int tcp_repair_options_est(struct tcp_sock *tp,
2310 struct tcp_repair_opt __user *optbuf, unsigned int len)
2312 struct tcp_repair_opt opt;
2314 while (len >= sizeof(opt)) {
2315 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2321 switch (opt.opt_code) {
2323 tp->rx_opt.mss_clamp = opt.opt_val;
2327 u16 snd_wscale = opt.opt_val & 0xFFFF;
2328 u16 rcv_wscale = opt.opt_val >> 16;
2330 if (snd_wscale > 14 || rcv_wscale > 14)
2333 tp->rx_opt.snd_wscale = snd_wscale;
2334 tp->rx_opt.rcv_wscale = rcv_wscale;
2335 tp->rx_opt.wscale_ok = 1;
2338 case TCPOPT_SACK_PERM:
2339 if (opt.opt_val != 0)
2342 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2343 if (sysctl_tcp_fack)
2344 tcp_enable_fack(tp);
2346 case TCPOPT_TIMESTAMP:
2347 if (opt.opt_val != 0)
2350 tp->rx_opt.tstamp_ok = 1;
2359 * Socket option code for TCP.
2361 static int do_tcp_setsockopt(struct sock *sk, int level,
2362 int optname, char __user *optval, unsigned int optlen)
2364 struct tcp_sock *tp = tcp_sk(sk);
2365 struct inet_connection_sock *icsk = inet_csk(sk);
2369 /* These are data/string values, all the others are ints */
2371 case TCP_CONGESTION: {
2372 char name[TCP_CA_NAME_MAX];
2377 val = strncpy_from_user(name, optval,
2378 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2384 err = tcp_set_congestion_control(sk, name);
2393 if (optlen < sizeof(int))
2396 if (get_user(val, (int __user *)optval))
2403 /* Values greater than interface MTU won't take effect. However
2404 * at the point when this call is done we typically don't yet
2405 * know which interface is going to be used */
2406 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2410 tp->rx_opt.user_mss = val;
2415 /* TCP_NODELAY is weaker than TCP_CORK, so that
2416 * this option on corked socket is remembered, but
2417 * it is not activated until cork is cleared.
2419 * However, when TCP_NODELAY is set we make
2420 * an explicit push, which overrides even TCP_CORK
2421 * for currently queued segments.
2423 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2424 tcp_push_pending_frames(sk);
2426 tp->nonagle &= ~TCP_NAGLE_OFF;
2430 case TCP_THIN_LINEAR_TIMEOUTS:
2431 if (val < 0 || val > 1)
2437 case TCP_THIN_DUPACK:
2438 if (val < 0 || val > 1)
2441 tp->thin_dupack = val;
2442 if (tp->thin_dupack)
2443 tcp_disable_early_retrans(tp);
2448 if (!tcp_can_repair_sock(sk))
2450 else if (val == 1) {
2452 sk->sk_reuse = SK_FORCE_REUSE;
2453 tp->repair_queue = TCP_NO_QUEUE;
2454 } else if (val == 0) {
2456 sk->sk_reuse = SK_NO_REUSE;
2457 tcp_send_window_probe(sk);
2463 case TCP_REPAIR_QUEUE:
2466 else if ((unsigned int)val < TCP_QUEUES_NR)
2467 tp->repair_queue = val;
2473 if (sk->sk_state != TCP_CLOSE)
2475 else if (tp->repair_queue == TCP_SEND_QUEUE)
2476 tp->write_seq = val;
2477 else if (tp->repair_queue == TCP_RECV_QUEUE)
2483 case TCP_REPAIR_OPTIONS:
2486 else if (sk->sk_state == TCP_ESTABLISHED)
2487 err = tcp_repair_options_est(tp,
2488 (struct tcp_repair_opt __user *)optval,
2495 /* When set indicates to always queue non-full frames.
2496 * Later the user clears this option and we transmit
2497 * any pending partial frames in the queue. This is
2498 * meant to be used alongside sendfile() to get properly
2499 * filled frames when the user (for example) must write
2500 * out headers with a write() call first and then use
2501 * sendfile to send out the data parts.
2503 * TCP_CORK can be set together with TCP_NODELAY and it is
2504 * stronger than TCP_NODELAY.
2507 tp->nonagle |= TCP_NAGLE_CORK;
2509 tp->nonagle &= ~TCP_NAGLE_CORK;
2510 if (tp->nonagle&TCP_NAGLE_OFF)
2511 tp->nonagle |= TCP_NAGLE_PUSH;
2512 tcp_push_pending_frames(sk);
2517 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2520 tp->keepalive_time = val * HZ;
2521 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2522 !((1 << sk->sk_state) &
2523 (TCPF_CLOSE | TCPF_LISTEN))) {
2524 u32 elapsed = keepalive_time_elapsed(tp);
2525 if (tp->keepalive_time > elapsed)
2526 elapsed = tp->keepalive_time - elapsed;
2529 inet_csk_reset_keepalive_timer(sk, elapsed);
2534 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2537 tp->keepalive_intvl = val * HZ;
2540 if (val < 1 || val > MAX_TCP_KEEPCNT)
2543 tp->keepalive_probes = val;
2546 if (val < 1 || val > MAX_TCP_SYNCNT)
2549 icsk->icsk_syn_retries = val;
2553 if (val < 0 || val > 1)
2562 else if (val > sysctl_tcp_fin_timeout / HZ)
2565 tp->linger2 = val * HZ;
2568 case TCP_DEFER_ACCEPT:
2569 /* Translate value in seconds to number of retransmits */
2570 icsk->icsk_accept_queue.rskq_defer_accept =
2571 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2575 case TCP_WINDOW_CLAMP:
2577 if (sk->sk_state != TCP_CLOSE) {
2581 tp->window_clamp = 0;
2583 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2584 SOCK_MIN_RCVBUF / 2 : val;
2589 icsk->icsk_ack.pingpong = 1;
2591 icsk->icsk_ack.pingpong = 0;
2592 if ((1 << sk->sk_state) &
2593 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2594 inet_csk_ack_scheduled(sk)) {
2595 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2596 tcp_cleanup_rbuf(sk, 1);
2598 icsk->icsk_ack.pingpong = 1;
2603 #ifdef CONFIG_TCP_MD5SIG
2605 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
2606 err = tp->af_specific->md5_parse(sk, optval, optlen);
2611 case TCP_USER_TIMEOUT:
2612 /* Cap the max time in ms TCP will retry or probe the window
2613 * before giving up and aborting (ETIMEDOUT) a connection.
2618 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2622 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2624 tcp_fastopen_init_key_once(true);
2626 fastopen_queue_tune(sk, val);
2635 tp->tsoffset = val - tcp_time_stamp;
2637 case TCP_NOTSENT_LOWAT:
2638 tp->notsent_lowat = val;
2639 sk->sk_write_space(sk);
2650 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2651 unsigned int optlen)
2653 const struct inet_connection_sock *icsk = inet_csk(sk);
2655 if (level != SOL_TCP)
2656 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2658 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2660 EXPORT_SYMBOL(tcp_setsockopt);
2662 #ifdef CONFIG_COMPAT
2663 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2664 char __user *optval, unsigned int optlen)
2666 if (level != SOL_TCP)
2667 return inet_csk_compat_setsockopt(sk, level, optname,
2669 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2671 EXPORT_SYMBOL(compat_tcp_setsockopt);
2674 /* Return information about state of tcp endpoint in API format. */
2675 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2677 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2678 const struct inet_connection_sock *icsk = inet_csk(sk);
2679 u32 now = tcp_time_stamp;
2684 memset(info, 0, sizeof(*info));
2685 if (sk->sk_type != SOCK_STREAM)
2688 info->tcpi_state = sk_state_load(sk);
2690 info->tcpi_ca_state = icsk->icsk_ca_state;
2691 info->tcpi_retransmits = icsk->icsk_retransmits;
2692 info->tcpi_probes = icsk->icsk_probes_out;
2693 info->tcpi_backoff = icsk->icsk_backoff;
2695 if (tp->rx_opt.tstamp_ok)
2696 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2697 if (tcp_is_sack(tp))
2698 info->tcpi_options |= TCPI_OPT_SACK;
2699 if (tp->rx_opt.wscale_ok) {
2700 info->tcpi_options |= TCPI_OPT_WSCALE;
2701 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2702 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2705 if (tp->ecn_flags & TCP_ECN_OK)
2706 info->tcpi_options |= TCPI_OPT_ECN;
2707 if (tp->ecn_flags & TCP_ECN_SEEN)
2708 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2709 if (tp->syn_data_acked)
2710 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2712 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2713 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2714 info->tcpi_snd_mss = tp->mss_cache;
2715 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2717 if (info->tcpi_state == TCP_LISTEN) {
2718 info->tcpi_unacked = sk->sk_ack_backlog;
2719 info->tcpi_sacked = sk->sk_max_ack_backlog;
2721 info->tcpi_unacked = tp->packets_out;
2722 info->tcpi_sacked = tp->sacked_out;
2724 info->tcpi_lost = tp->lost_out;
2725 info->tcpi_retrans = tp->retrans_out;
2726 info->tcpi_fackets = tp->fackets_out;
2728 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2729 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2730 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2732 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2733 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2734 info->tcpi_rtt = tp->srtt_us >> 3;
2735 info->tcpi_rttvar = tp->mdev_us >> 2;
2736 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2737 info->tcpi_snd_cwnd = tp->snd_cwnd;
2738 info->tcpi_advmss = tp->advmss;
2739 info->tcpi_reordering = tp->reordering;
2741 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2742 info->tcpi_rcv_space = tp->rcvq_space.space;
2744 info->tcpi_total_retrans = tp->total_retrans;
2746 rate = READ_ONCE(sk->sk_pacing_rate);
2747 rate64 = rate != ~0U ? rate : ~0ULL;
2748 put_unaligned(rate64, &info->tcpi_pacing_rate);
2750 rate = READ_ONCE(sk->sk_max_pacing_rate);
2751 rate64 = rate != ~0U ? rate : ~0ULL;
2752 put_unaligned(rate64, &info->tcpi_max_pacing_rate);
2754 /* Expose reference count for socket */
2755 if (sk->sk_socket) {
2756 struct file *filep = sk->sk_socket->file;
2759 info->tcpi_count = file_count(filep);
2763 start = u64_stats_fetch_begin_irq(&tp->syncp);
2764 put_unaligned(tp->bytes_acked, &info->tcpi_bytes_acked);
2765 put_unaligned(tp->bytes_received, &info->tcpi_bytes_received);
2766 } while (u64_stats_fetch_retry_irq(&tp->syncp, start));
2767 info->tcpi_segs_out = tp->segs_out;
2768 info->tcpi_segs_in = tp->segs_in;
2770 EXPORT_SYMBOL_GPL(tcp_get_info);
2772 static int do_tcp_getsockopt(struct sock *sk, int level,
2773 int optname, char __user *optval, int __user *optlen)
2775 struct inet_connection_sock *icsk = inet_csk(sk);
2776 struct tcp_sock *tp = tcp_sk(sk);
2779 if (get_user(len, optlen))
2782 len = min_t(unsigned int, len, sizeof(int));
2789 val = tp->mss_cache;
2790 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2791 val = tp->rx_opt.user_mss;
2793 val = tp->rx_opt.mss_clamp;
2796 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2799 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2802 val = keepalive_time_when(tp) / HZ;
2805 val = keepalive_intvl_when(tp) / HZ;
2808 val = keepalive_probes(tp);
2811 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2816 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2818 case TCP_DEFER_ACCEPT:
2819 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2820 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2822 case TCP_WINDOW_CLAMP:
2823 val = tp->window_clamp;
2826 struct tcp_info info;
2828 if (get_user(len, optlen))
2831 tcp_get_info(sk, &info);
2833 len = min_t(unsigned int, len, sizeof(info));
2834 if (put_user(len, optlen))
2836 if (copy_to_user(optval, &info, len))
2841 const struct tcp_congestion_ops *ca_ops;
2842 union tcp_cc_info info;
2846 if (get_user(len, optlen))
2849 ca_ops = icsk->icsk_ca_ops;
2850 if (ca_ops && ca_ops->get_info)
2851 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2853 len = min_t(unsigned int, len, sz);
2854 if (put_user(len, optlen))
2856 if (copy_to_user(optval, &info, len))
2861 val = !icsk->icsk_ack.pingpong;
2864 case TCP_CONGESTION:
2865 if (get_user(len, optlen))
2867 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2868 if (put_user(len, optlen))
2870 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2874 case TCP_THIN_LINEAR_TIMEOUTS:
2877 case TCP_THIN_DUPACK:
2878 val = tp->thin_dupack;
2885 case TCP_REPAIR_QUEUE:
2887 val = tp->repair_queue;
2893 if (tp->repair_queue == TCP_SEND_QUEUE)
2894 val = tp->write_seq;
2895 else if (tp->repair_queue == TCP_RECV_QUEUE)
2901 case TCP_USER_TIMEOUT:
2902 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2906 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
2910 val = tcp_time_stamp + tp->tsoffset;
2912 case TCP_NOTSENT_LOWAT:
2913 val = tp->notsent_lowat;
2918 case TCP_SAVED_SYN: {
2919 if (get_user(len, optlen))
2923 if (tp->saved_syn) {
2924 if (len < tp->saved_syn[0]) {
2925 if (put_user(tp->saved_syn[0], optlen)) {
2932 len = tp->saved_syn[0];
2933 if (put_user(len, optlen)) {
2937 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
2941 tcp_saved_syn_free(tp);
2946 if (put_user(len, optlen))
2952 return -ENOPROTOOPT;
2955 if (put_user(len, optlen))
2957 if (copy_to_user(optval, &val, len))
2962 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2965 struct inet_connection_sock *icsk = inet_csk(sk);
2967 if (level != SOL_TCP)
2968 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2970 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2972 EXPORT_SYMBOL(tcp_getsockopt);
2974 #ifdef CONFIG_COMPAT
2975 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2976 char __user *optval, int __user *optlen)
2978 if (level != SOL_TCP)
2979 return inet_csk_compat_getsockopt(sk, level, optname,
2981 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2983 EXPORT_SYMBOL(compat_tcp_getsockopt);
2986 #ifdef CONFIG_TCP_MD5SIG
2987 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2988 static DEFINE_MUTEX(tcp_md5sig_mutex);
2989 static bool tcp_md5sig_pool_populated = false;
2991 static void __tcp_alloc_md5sig_pool(void)
2995 for_each_possible_cpu(cpu) {
2996 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2997 struct crypto_hash *hash;
2999 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3000 if (IS_ERR_OR_NULL(hash))
3002 per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
3005 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3006 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3009 tcp_md5sig_pool_populated = true;
3012 bool tcp_alloc_md5sig_pool(void)
3014 if (unlikely(!tcp_md5sig_pool_populated)) {
3015 mutex_lock(&tcp_md5sig_mutex);
3017 if (!tcp_md5sig_pool_populated)
3018 __tcp_alloc_md5sig_pool();
3020 mutex_unlock(&tcp_md5sig_mutex);
3022 return tcp_md5sig_pool_populated;
3024 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3028 * tcp_get_md5sig_pool - get md5sig_pool for this user
3030 * We use percpu structure, so if we succeed, we exit with preemption
3031 * and BH disabled, to make sure another thread or softirq handling
3032 * wont try to get same context.
3034 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3038 if (tcp_md5sig_pool_populated) {
3039 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3041 return this_cpu_ptr(&tcp_md5sig_pool);
3046 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3048 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3049 const struct tcphdr *th)
3051 struct scatterlist sg;
3055 /* We are not allowed to change tcphdr, make a local copy */
3056 memcpy(&hdr, th, sizeof(hdr));
3059 /* options aren't included in the hash */
3060 sg_init_one(&sg, &hdr, sizeof(hdr));
3061 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3064 EXPORT_SYMBOL(tcp_md5_hash_header);
3066 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3067 const struct sk_buff *skb, unsigned int header_len)
3069 struct scatterlist sg;
3070 const struct tcphdr *tp = tcp_hdr(skb);
3071 struct hash_desc *desc = &hp->md5_desc;
3073 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3074 skb_headlen(skb) - header_len : 0;
3075 const struct skb_shared_info *shi = skb_shinfo(skb);
3076 struct sk_buff *frag_iter;
3078 sg_init_table(&sg, 1);
3080 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3081 if (crypto_hash_update(desc, &sg, head_data_len))
3084 for (i = 0; i < shi->nr_frags; ++i) {
3085 const struct skb_frag_struct *f = &shi->frags[i];
3086 unsigned int offset = f->page_offset;
3087 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3089 sg_set_page(&sg, page, skb_frag_size(f),
3090 offset_in_page(offset));
3091 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3095 skb_walk_frags(skb, frag_iter)
3096 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3101 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3103 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3105 struct scatterlist sg;
3107 sg_init_one(&sg, key->key, key->keylen);
3108 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3110 EXPORT_SYMBOL(tcp_md5_hash_key);
3114 void tcp_done(struct sock *sk)
3116 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3118 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3119 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3121 tcp_set_state(sk, TCP_CLOSE);
3122 tcp_clear_xmit_timers(sk);
3124 reqsk_fastopen_remove(sk, req, false);
3126 sk->sk_shutdown = SHUTDOWN_MASK;
3128 if (!sock_flag(sk, SOCK_DEAD))
3129 sk->sk_state_change(sk);
3131 inet_csk_destroy_sock(sk);
3133 EXPORT_SYMBOL_GPL(tcp_done);
3135 int tcp_abort(struct sock *sk, int err)
3137 if (!sk_fullsock(sk)) {
3138 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3139 struct request_sock *req = inet_reqsk(sk);
3142 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener,
3151 /* Don't race with userspace socket closes such as tcp_close. */
3154 if (sk->sk_state == TCP_LISTEN) {
3155 tcp_set_state(sk, TCP_CLOSE);
3156 inet_csk_listen_stop(sk);
3159 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3163 if (!sock_flag(sk, SOCK_DEAD)) {
3165 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3167 sk->sk_error_report(sk);
3168 if (tcp_need_reset(sk->sk_state))
3169 tcp_send_active_reset(sk, GFP_ATOMIC);
3179 EXPORT_SYMBOL_GPL(tcp_abort);
3181 extern struct tcp_congestion_ops tcp_reno;
3183 static __initdata unsigned long thash_entries;
3184 static int __init set_thash_entries(char *str)
3191 ret = kstrtoul(str, 0, &thash_entries);
3197 __setup("thash_entries=", set_thash_entries);
3199 static void __init tcp_init_mem(void)
3201 unsigned long limit = nr_free_buffer_pages() / 16;
3203 limit = max(limit, 128UL);
3204 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3205 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3206 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3209 void __init tcp_init(void)
3211 unsigned long limit;
3212 int max_rshare, max_wshare, cnt;
3215 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
3216 sock_skb_cb_check_size(sizeof(struct tcp_skb_cb));
3218 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3219 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3220 tcp_hashinfo.bind_bucket_cachep =
3221 kmem_cache_create("tcp_bind_bucket",
3222 sizeof(struct inet_bind_bucket), 0,
3223 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3225 /* Size and allocate the main established and bind bucket
3228 * The methodology is similar to that of the buffer cache.
3230 tcp_hashinfo.ehash =
3231 alloc_large_system_hash("TCP established",
3232 sizeof(struct inet_ehash_bucket),
3234 17, /* one slot per 128 KB of memory */
3237 &tcp_hashinfo.ehash_mask,
3239 thash_entries ? 0 : 512 * 1024);
3240 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3241 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3243 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3244 panic("TCP: failed to alloc ehash_locks");
3245 tcp_hashinfo.bhash =
3246 alloc_large_system_hash("TCP bind",
3247 sizeof(struct inet_bind_hashbucket),
3248 tcp_hashinfo.ehash_mask + 1,
3249 17, /* one slot per 128 KB of memory */
3251 &tcp_hashinfo.bhash_size,
3255 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3256 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3257 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3258 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3262 cnt = tcp_hashinfo.ehash_mask + 1;
3264 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3265 sysctl_tcp_max_orphans = cnt / 2;
3266 sysctl_max_syn_backlog = max(128, cnt / 256);
3269 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3270 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3271 max_wshare = min(4UL*1024*1024, limit);
3272 max_rshare = min(6UL*1024*1024, limit);
3274 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3275 sysctl_tcp_wmem[1] = 16*1024;
3276 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3278 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3279 sysctl_tcp_rmem[1] = 87380;
3280 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3282 pr_info("Hash tables configured (established %u bind %u)\n",
3283 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3286 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);