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 <crypto/hash.h>
251 #include <linux/kernel.h>
252 #include <linux/module.h>
253 #include <linux/types.h>
254 #include <linux/fcntl.h>
255 #include <linux/poll.h>
256 #include <linux/inet_diag.h>
257 #include <linux/init.h>
258 #include <linux/fs.h>
259 #include <linux/skbuff.h>
260 #include <linux/scatterlist.h>
261 #include <linux/splice.h>
262 #include <linux/net.h>
263 #include <linux/socket.h>
264 #include <linux/random.h>
265 #include <linux/bootmem.h>
266 #include <linux/highmem.h>
267 #include <linux/swap.h>
268 #include <linux/cache.h>
269 #include <linux/err.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 <linux/uaccess.h>
281 #include <asm/ioctls.h>
282 #include <net/busy_poll.h>
284 int sysctl_tcp_min_tso_segs __read_mostly = 2;
286 int sysctl_tcp_autocorking __read_mostly = 1;
288 struct percpu_counter tcp_orphan_count;
289 EXPORT_SYMBOL_GPL(tcp_orphan_count);
291 long sysctl_tcp_mem[3] __read_mostly;
292 int sysctl_tcp_wmem[3] __read_mostly;
293 int sysctl_tcp_rmem[3] __read_mostly;
295 EXPORT_SYMBOL(sysctl_tcp_mem);
296 EXPORT_SYMBOL(sysctl_tcp_rmem);
297 EXPORT_SYMBOL(sysctl_tcp_wmem);
299 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
300 EXPORT_SYMBOL(tcp_memory_allocated);
303 * Current number of TCP sockets.
305 struct percpu_counter tcp_sockets_allocated;
306 EXPORT_SYMBOL(tcp_sockets_allocated);
311 struct tcp_splice_state {
312 struct pipe_inode_info *pipe;
318 * Pressure flag: try to collapse.
319 * Technical note: it is used by multiple contexts non atomically.
320 * All the __sk_mem_schedule() is of this nature: accounting
321 * is strict, actions are advisory and have some latency.
323 int tcp_memory_pressure __read_mostly;
324 EXPORT_SYMBOL(tcp_memory_pressure);
326 void tcp_enter_memory_pressure(struct sock *sk)
328 if (!tcp_memory_pressure) {
329 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
330 tcp_memory_pressure = 1;
333 EXPORT_SYMBOL(tcp_enter_memory_pressure);
335 /* Convert seconds to retransmits based on initial and max timeout */
336 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
341 int period = timeout;
344 while (seconds > period && res < 255) {
347 if (timeout > rto_max)
355 /* Convert retransmits to seconds based on initial and max timeout */
356 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
364 if (timeout > rto_max)
372 /* Address-family independent initialization for a tcp_sock.
374 * NOTE: A lot of things set to zero explicitly by call to
375 * sk_alloc() so need not be done here.
377 void tcp_init_sock(struct sock *sk)
379 struct inet_connection_sock *icsk = inet_csk(sk);
380 struct tcp_sock *tp = tcp_sk(sk);
382 tp->out_of_order_queue = RB_ROOT;
383 tcp_init_xmit_timers(sk);
384 tcp_prequeue_init(tp);
385 INIT_LIST_HEAD(&tp->tsq_node);
387 icsk->icsk_rto = TCP_TIMEOUT_INIT;
388 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
389 minmax_reset(&tp->rtt_min, tcp_time_stamp, ~0U);
391 /* So many TCP implementations out there (incorrectly) count the
392 * initial SYN frame in their delayed-ACK and congestion control
393 * algorithms that we must have the following bandaid to talk
394 * efficiently to them. -DaveM
396 tp->snd_cwnd = TCP_INIT_CWND;
398 /* There's a bubble in the pipe until at least the first ACK. */
399 tp->app_limited = ~0U;
401 /* See draft-stevens-tcpca-spec-01 for discussion of the
402 * initialization of these values.
404 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
405 tp->snd_cwnd_clamp = ~0;
406 tp->mss_cache = TCP_MSS_DEFAULT;
408 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
409 tcp_enable_early_retrans(tp);
410 tcp_assign_congestion_control(sk);
414 sk->sk_state = TCP_CLOSE;
416 sk->sk_write_space = sk_stream_write_space;
417 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
419 icsk->icsk_sync_mss = tcp_sync_mss;
421 sk->sk_sndbuf = sysctl_tcp_wmem[1];
422 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
425 sk_sockets_allocated_inc(sk);
428 EXPORT_SYMBOL(tcp_init_sock);
430 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags, struct sk_buff *skb)
433 struct skb_shared_info *shinfo = skb_shinfo(skb);
434 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
436 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
437 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
438 tcb->txstamp_ack = 1;
439 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
440 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
445 * Wait for a TCP event.
447 * Note that we don't need to lock the socket, as the upper poll layers
448 * take care of normal races (between the test and the event) and we don't
449 * go look at any of the socket buffers directly.
451 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
454 struct sock *sk = sock->sk;
455 const struct tcp_sock *tp = tcp_sk(sk);
458 sock_rps_record_flow(sk);
460 sock_poll_wait(file, sk_sleep(sk), wait);
462 state = sk_state_load(sk);
463 if (state == TCP_LISTEN)
464 return inet_csk_listen_poll(sk);
466 /* Socket is not locked. We are protected from async events
467 * by poll logic and correct handling of state changes
468 * made by other threads is impossible in any case.
474 * POLLHUP is certainly not done right. But poll() doesn't
475 * have a notion of HUP in just one direction, and for a
476 * socket the read side is more interesting.
478 * Some poll() documentation says that POLLHUP is incompatible
479 * with the POLLOUT/POLLWR flags, so somebody should check this
480 * all. But careful, it tends to be safer to return too many
481 * bits than too few, and you can easily break real applications
482 * if you don't tell them that something has hung up!
486 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
487 * our fs/select.c). It means that after we received EOF,
488 * poll always returns immediately, making impossible poll() on write()
489 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
490 * if and only if shutdown has been made in both directions.
491 * Actually, it is interesting to look how Solaris and DUX
492 * solve this dilemma. I would prefer, if POLLHUP were maskable,
493 * then we could set it on SND_SHUTDOWN. BTW examples given
494 * in Stevens' books assume exactly this behaviour, it explains
495 * why POLLHUP is incompatible with POLLOUT. --ANK
497 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
498 * blocking on fresh not-connected or disconnected socket. --ANK
500 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
502 if (sk->sk_shutdown & RCV_SHUTDOWN)
503 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
505 /* Connected or passive Fast Open socket? */
506 if (state != TCP_SYN_SENT &&
507 (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
508 int target = sock_rcvlowat(sk, 0, INT_MAX);
510 if (tp->urg_seq == tp->copied_seq &&
511 !sock_flag(sk, SOCK_URGINLINE) &&
515 if (tp->rcv_nxt - tp->copied_seq >= target)
516 mask |= POLLIN | POLLRDNORM;
518 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
519 if (sk_stream_is_writeable(sk)) {
520 mask |= POLLOUT | POLLWRNORM;
521 } else { /* send SIGIO later */
522 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
523 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
525 /* Race breaker. If space is freed after
526 * wspace test but before the flags are set,
527 * IO signal will be lost. Memory barrier
528 * pairs with the input side.
530 smp_mb__after_atomic();
531 if (sk_stream_is_writeable(sk))
532 mask |= POLLOUT | POLLWRNORM;
535 mask |= POLLOUT | POLLWRNORM;
537 if (tp->urg_data & TCP_URG_VALID)
540 /* This barrier is coupled with smp_wmb() in tcp_reset() */
542 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
547 EXPORT_SYMBOL(tcp_poll);
549 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
551 struct tcp_sock *tp = tcp_sk(sk);
557 if (sk->sk_state == TCP_LISTEN)
560 slow = lock_sock_fast(sk);
562 unlock_sock_fast(sk, slow);
565 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
568 if (sk->sk_state == TCP_LISTEN)
571 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
574 answ = tp->write_seq - tp->snd_una;
577 if (sk->sk_state == TCP_LISTEN)
580 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
583 answ = tp->write_seq - tp->snd_nxt;
589 return put_user(answ, (int __user *)arg);
591 EXPORT_SYMBOL(tcp_ioctl);
593 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
595 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
596 tp->pushed_seq = tp->write_seq;
599 static inline bool forced_push(const struct tcp_sock *tp)
601 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
604 static void skb_entail(struct sock *sk, struct sk_buff *skb)
606 struct tcp_sock *tp = tcp_sk(sk);
607 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
610 tcb->seq = tcb->end_seq = tp->write_seq;
611 tcb->tcp_flags = TCPHDR_ACK;
613 __skb_header_release(skb);
614 tcp_add_write_queue_tail(sk, skb);
615 sk->sk_wmem_queued += skb->truesize;
616 sk_mem_charge(sk, skb->truesize);
617 if (tp->nonagle & TCP_NAGLE_PUSH)
618 tp->nonagle &= ~TCP_NAGLE_PUSH;
620 tcp_slow_start_after_idle_check(sk);
623 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
626 tp->snd_up = tp->write_seq;
629 /* If a not yet filled skb is pushed, do not send it if
630 * we have data packets in Qdisc or NIC queues :
631 * Because TX completion will happen shortly, it gives a chance
632 * to coalesce future sendmsg() payload into this skb, without
633 * need for a timer, and with no latency trade off.
634 * As packets containing data payload have a bigger truesize
635 * than pure acks (dataless) packets, the last checks prevent
636 * autocorking if we only have an ACK in Qdisc/NIC queues,
637 * or if TX completion was delayed after we processed ACK packet.
639 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
642 return skb->len < size_goal &&
643 sysctl_tcp_autocorking &&
644 skb != tcp_write_queue_head(sk) &&
645 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
648 static void tcp_push(struct sock *sk, int flags, int mss_now,
649 int nonagle, int size_goal)
651 struct tcp_sock *tp = tcp_sk(sk);
654 if (!tcp_send_head(sk))
657 skb = tcp_write_queue_tail(sk);
658 if (!(flags & MSG_MORE) || forced_push(tp))
659 tcp_mark_push(tp, skb);
661 tcp_mark_urg(tp, flags);
663 if (tcp_should_autocork(sk, skb, size_goal)) {
665 /* avoid atomic op if TSQ_THROTTLED bit is already set */
666 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
667 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
668 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
670 /* It is possible TX completion already happened
671 * before we set TSQ_THROTTLED.
673 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
677 if (flags & MSG_MORE)
678 nonagle = TCP_NAGLE_CORK;
680 __tcp_push_pending_frames(sk, mss_now, nonagle);
683 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
684 unsigned int offset, size_t len)
686 struct tcp_splice_state *tss = rd_desc->arg.data;
689 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
690 min(rd_desc->count, len), tss->flags);
692 rd_desc->count -= ret;
696 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
698 /* Store TCP splice context information in read_descriptor_t. */
699 read_descriptor_t rd_desc = {
704 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
708 * tcp_splice_read - splice data from TCP socket to a pipe
709 * @sock: socket to splice from
710 * @ppos: position (not valid)
711 * @pipe: pipe to splice to
712 * @len: number of bytes to splice
713 * @flags: splice modifier flags
716 * Will read pages from given socket and fill them into a pipe.
719 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
720 struct pipe_inode_info *pipe, size_t len,
723 struct sock *sk = sock->sk;
724 struct tcp_splice_state tss = {
733 sock_rps_record_flow(sk);
735 * We can't seek on a socket input
744 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
746 ret = __tcp_splice_read(sk, &tss);
752 if (sock_flag(sk, SOCK_DONE))
755 ret = sock_error(sk);
758 if (sk->sk_shutdown & RCV_SHUTDOWN)
760 if (sk->sk_state == TCP_CLOSE) {
762 * This occurs when user tries to read
763 * from never connected socket.
765 if (!sock_flag(sk, SOCK_DONE))
773 /* if __tcp_splice_read() got nothing while we have
774 * an skb in receive queue, we do not want to loop.
775 * This might happen with URG data.
777 if (!skb_queue_empty(&sk->sk_receive_queue))
779 sk_wait_data(sk, &timeo, NULL);
780 if (signal_pending(current)) {
781 ret = sock_intr_errno(timeo);
794 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
795 (sk->sk_shutdown & RCV_SHUTDOWN) ||
796 signal_pending(current))
807 EXPORT_SYMBOL(tcp_splice_read);
809 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
814 /* The TCP header must be at least 32-bit aligned. */
815 size = ALIGN(size, 4);
817 if (unlikely(tcp_under_memory_pressure(sk)))
818 sk_mem_reclaim_partial(sk);
820 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
824 if (force_schedule) {
825 mem_scheduled = true;
826 sk_forced_mem_schedule(sk, skb->truesize);
828 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
830 if (likely(mem_scheduled)) {
831 skb_reserve(skb, sk->sk_prot->max_header);
833 * Make sure that we have exactly size bytes
834 * available to the caller, no more, no less.
836 skb->reserved_tailroom = skb->end - skb->tail - size;
841 sk->sk_prot->enter_memory_pressure(sk);
842 sk_stream_moderate_sndbuf(sk);
847 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
850 struct tcp_sock *tp = tcp_sk(sk);
851 u32 new_size_goal, size_goal;
853 if (!large_allowed || !sk_can_gso(sk))
856 /* Note : tcp_tso_autosize() will eventually split this later */
857 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
858 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
860 /* We try hard to avoid divides here */
861 size_goal = tp->gso_segs * mss_now;
862 if (unlikely(new_size_goal < size_goal ||
863 new_size_goal >= size_goal + mss_now)) {
864 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
865 sk->sk_gso_max_segs);
866 size_goal = tp->gso_segs * mss_now;
869 return max(size_goal, mss_now);
872 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
876 mss_now = tcp_current_mss(sk);
877 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
882 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
883 size_t size, int flags)
885 struct tcp_sock *tp = tcp_sk(sk);
886 int mss_now, size_goal;
889 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
891 /* Wait for a connection to finish. One exception is TCP Fast Open
892 * (passive side) where data is allowed to be sent before a connection
893 * is fully established.
895 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
896 !tcp_passive_fastopen(sk)) {
897 err = sk_stream_wait_connect(sk, &timeo);
902 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
904 mss_now = tcp_send_mss(sk, &size_goal, flags);
908 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
912 struct sk_buff *skb = tcp_write_queue_tail(sk);
916 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0 ||
917 !tcp_skb_can_collapse_to(skb)) {
919 if (!sk_stream_memory_free(sk))
920 goto wait_for_sndbuf;
922 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
923 skb_queue_empty(&sk->sk_write_queue));
925 goto wait_for_memory;
934 i = skb_shinfo(skb)->nr_frags;
935 can_coalesce = skb_can_coalesce(skb, i, page, offset);
936 if (!can_coalesce && i >= sysctl_max_skb_frags) {
937 tcp_mark_push(tp, skb);
940 if (!sk_wmem_schedule(sk, copy))
941 goto wait_for_memory;
944 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
947 skb_fill_page_desc(skb, i, page, offset, copy);
949 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
952 skb->data_len += copy;
953 skb->truesize += copy;
954 sk->sk_wmem_queued += copy;
955 sk_mem_charge(sk, copy);
956 skb->ip_summed = CHECKSUM_PARTIAL;
957 tp->write_seq += copy;
958 TCP_SKB_CB(skb)->end_seq += copy;
959 tcp_skb_pcount_set(skb, 0);
962 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
968 tcp_tx_timestamp(sk, sk->sk_tsflags, skb);
972 if (skb->len < size_goal || (flags & MSG_OOB))
975 if (forced_push(tp)) {
976 tcp_mark_push(tp, skb);
977 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
978 } else if (skb == tcp_send_head(sk))
979 tcp_push_one(sk, mss_now);
983 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
985 tcp_push(sk, flags & ~MSG_MORE, mss_now,
986 TCP_NAGLE_PUSH, size_goal);
988 err = sk_stream_wait_memory(sk, &timeo);
992 mss_now = tcp_send_mss(sk, &size_goal, flags);
996 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
997 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1004 /* make sure we wake any epoll edge trigger waiter */
1005 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1007 sk->sk_write_space(sk);
1008 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1010 return sk_stream_error(sk, flags, err);
1013 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1014 size_t size, int flags)
1018 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1019 !sk_check_csum_caps(sk))
1020 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1025 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1027 res = do_tcp_sendpages(sk, page, offset, size, flags);
1031 EXPORT_SYMBOL(tcp_sendpage);
1033 /* Do not bother using a page frag for very small frames.
1034 * But use this heuristic only for the first skb in write queue.
1036 * Having no payload in skb->head allows better SACK shifting
1037 * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1038 * write queue has less skbs.
1039 * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1040 * This also speeds up tso_fragment(), since it wont fallback
1041 * to tcp_fragment().
1043 static int linear_payload_sz(bool first_skb)
1046 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1050 static int select_size(const struct sock *sk, bool sg, bool first_skb)
1052 const struct tcp_sock *tp = tcp_sk(sk);
1053 int tmp = tp->mss_cache;
1056 if (sk_can_gso(sk)) {
1057 tmp = linear_payload_sz(first_skb);
1059 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1061 if (tmp >= pgbreak &&
1062 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1070 void tcp_free_fastopen_req(struct tcp_sock *tp)
1072 if (tp->fastopen_req) {
1073 kfree(tp->fastopen_req);
1074 tp->fastopen_req = NULL;
1078 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1079 int *copied, size_t size)
1081 struct tcp_sock *tp = tcp_sk(sk);
1084 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1086 if (tp->fastopen_req)
1087 return -EALREADY; /* Another Fast Open is in progress */
1089 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1091 if (unlikely(!tp->fastopen_req))
1093 tp->fastopen_req->data = msg;
1094 tp->fastopen_req->size = size;
1096 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1097 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1098 msg->msg_namelen, flags);
1099 *copied = tp->fastopen_req->copied;
1100 tcp_free_fastopen_req(tp);
1104 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1106 struct tcp_sock *tp = tcp_sk(sk);
1107 struct sk_buff *skb;
1108 struct sockcm_cookie sockc;
1109 int flags, err, copied = 0;
1110 int mss_now = 0, size_goal, copied_syn = 0;
1111 bool process_backlog = false;
1117 flags = msg->msg_flags;
1118 if (flags & MSG_FASTOPEN) {
1119 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1120 if (err == -EINPROGRESS && copied_syn > 0)
1126 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1128 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1130 /* Wait for a connection to finish. One exception is TCP Fast Open
1131 * (passive side) where data is allowed to be sent before a connection
1132 * is fully established.
1134 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1135 !tcp_passive_fastopen(sk)) {
1136 err = sk_stream_wait_connect(sk, &timeo);
1141 if (unlikely(tp->repair)) {
1142 if (tp->repair_queue == TCP_RECV_QUEUE) {
1143 copied = tcp_send_rcvq(sk, msg, size);
1148 if (tp->repair_queue == TCP_NO_QUEUE)
1151 /* 'common' sending to sendq */
1154 sockc.tsflags = sk->sk_tsflags;
1155 if (msg->msg_controllen) {
1156 err = sock_cmsg_send(sk, msg, &sockc);
1157 if (unlikely(err)) {
1163 /* This should be in poll */
1164 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1166 /* Ok commence sending. */
1170 mss_now = tcp_send_mss(sk, &size_goal, flags);
1173 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1176 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1178 while (msg_data_left(msg)) {
1180 int max = size_goal;
1182 skb = tcp_write_queue_tail(sk);
1183 if (tcp_send_head(sk)) {
1184 if (skb->ip_summed == CHECKSUM_NONE)
1186 copy = max - skb->len;
1189 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1193 /* Allocate new segment. If the interface is SG,
1194 * allocate skb fitting to single page.
1196 if (!sk_stream_memory_free(sk))
1197 goto wait_for_sndbuf;
1199 if (process_backlog && sk_flush_backlog(sk)) {
1200 process_backlog = false;
1203 first_skb = skb_queue_empty(&sk->sk_write_queue);
1204 skb = sk_stream_alloc_skb(sk,
1205 select_size(sk, sg, first_skb),
1209 goto wait_for_memory;
1211 process_backlog = true;
1213 * Check whether we can use HW checksum.
1215 if (sk_check_csum_caps(sk))
1216 skb->ip_summed = CHECKSUM_PARTIAL;
1218 skb_entail(sk, skb);
1222 /* All packets are restored as if they have
1223 * already been sent. skb_mstamp isn't set to
1224 * avoid wrong rtt estimation.
1227 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1230 /* Try to append data to the end of skb. */
1231 if (copy > msg_data_left(msg))
1232 copy = msg_data_left(msg);
1234 /* Where to copy to? */
1235 if (skb_availroom(skb) > 0) {
1236 /* We have some space in skb head. Superb! */
1237 copy = min_t(int, copy, skb_availroom(skb));
1238 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1243 int i = skb_shinfo(skb)->nr_frags;
1244 struct page_frag *pfrag = sk_page_frag(sk);
1246 if (!sk_page_frag_refill(sk, pfrag))
1247 goto wait_for_memory;
1249 if (!skb_can_coalesce(skb, i, pfrag->page,
1251 if (i >= sysctl_max_skb_frags || !sg) {
1252 tcp_mark_push(tp, skb);
1258 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1260 if (!sk_wmem_schedule(sk, copy))
1261 goto wait_for_memory;
1263 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1270 /* Update the skb. */
1272 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1274 skb_fill_page_desc(skb, i, pfrag->page,
1275 pfrag->offset, copy);
1276 get_page(pfrag->page);
1278 pfrag->offset += copy;
1282 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1284 tp->write_seq += copy;
1285 TCP_SKB_CB(skb)->end_seq += copy;
1286 tcp_skb_pcount_set(skb, 0);
1289 if (!msg_data_left(msg)) {
1290 tcp_tx_timestamp(sk, sockc.tsflags, skb);
1291 if (unlikely(flags & MSG_EOR))
1292 TCP_SKB_CB(skb)->eor = 1;
1296 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1299 if (forced_push(tp)) {
1300 tcp_mark_push(tp, skb);
1301 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1302 } else if (skb == tcp_send_head(sk))
1303 tcp_push_one(sk, mss_now);
1307 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1310 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1311 TCP_NAGLE_PUSH, size_goal);
1313 err = sk_stream_wait_memory(sk, &timeo);
1317 mss_now = tcp_send_mss(sk, &size_goal, flags);
1322 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1325 return copied + copied_syn;
1329 tcp_unlink_write_queue(skb, sk);
1330 /* It is the one place in all of TCP, except connection
1331 * reset, where we can be unlinking the send_head.
1333 tcp_check_send_head(sk, skb);
1334 sk_wmem_free_skb(sk, skb);
1338 if (copied + copied_syn)
1341 err = sk_stream_error(sk, flags, err);
1342 /* make sure we wake any epoll edge trigger waiter */
1343 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1345 sk->sk_write_space(sk);
1346 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1351 EXPORT_SYMBOL(tcp_sendmsg);
1354 * Handle reading urgent data. BSD has very simple semantics for
1355 * this, no blocking and very strange errors 8)
1358 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1360 struct tcp_sock *tp = tcp_sk(sk);
1362 /* No URG data to read. */
1363 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1364 tp->urg_data == TCP_URG_READ)
1365 return -EINVAL; /* Yes this is right ! */
1367 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1370 if (tp->urg_data & TCP_URG_VALID) {
1372 char c = tp->urg_data;
1374 if (!(flags & MSG_PEEK))
1375 tp->urg_data = TCP_URG_READ;
1377 /* Read urgent data. */
1378 msg->msg_flags |= MSG_OOB;
1381 if (!(flags & MSG_TRUNC))
1382 err = memcpy_to_msg(msg, &c, 1);
1385 msg->msg_flags |= MSG_TRUNC;
1387 return err ? -EFAULT : len;
1390 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1393 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1394 * the available implementations agree in this case:
1395 * this call should never block, independent of the
1396 * blocking state of the socket.
1397 * Mike <pall@rz.uni-karlsruhe.de>
1402 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1404 struct sk_buff *skb;
1405 int copied = 0, err = 0;
1407 /* XXX -- need to support SO_PEEK_OFF */
1409 skb_queue_walk(&sk->sk_write_queue, skb) {
1410 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1417 return err ?: copied;
1420 /* Clean up the receive buffer for full frames taken by the user,
1421 * then send an ACK if necessary. COPIED is the number of bytes
1422 * tcp_recvmsg has given to the user so far, it speeds up the
1423 * calculation of whether or not we must ACK for the sake of
1426 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1428 struct tcp_sock *tp = tcp_sk(sk);
1429 bool time_to_ack = false;
1431 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1433 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1434 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1435 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1437 if (inet_csk_ack_scheduled(sk)) {
1438 const struct inet_connection_sock *icsk = inet_csk(sk);
1439 /* Delayed ACKs frequently hit locked sockets during bulk
1441 if (icsk->icsk_ack.blocked ||
1442 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1443 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1445 * If this read emptied read buffer, we send ACK, if
1446 * connection is not bidirectional, user drained
1447 * receive buffer and there was a small segment
1451 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1452 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1453 !icsk->icsk_ack.pingpong)) &&
1454 !atomic_read(&sk->sk_rmem_alloc)))
1458 /* We send an ACK if we can now advertise a non-zero window
1459 * which has been raised "significantly".
1461 * Even if window raised up to infinity, do not send window open ACK
1462 * in states, where we will not receive more. It is useless.
1464 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1465 __u32 rcv_window_now = tcp_receive_window(tp);
1467 /* Optimize, __tcp_select_window() is not cheap. */
1468 if (2*rcv_window_now <= tp->window_clamp) {
1469 __u32 new_window = __tcp_select_window(sk);
1471 /* Send ACK now, if this read freed lots of space
1472 * in our buffer. Certainly, new_window is new window.
1473 * We can advertise it now, if it is not less than current one.
1474 * "Lots" means "at least twice" here.
1476 if (new_window && new_window >= 2 * rcv_window_now)
1484 static void tcp_prequeue_process(struct sock *sk)
1486 struct sk_buff *skb;
1487 struct tcp_sock *tp = tcp_sk(sk);
1489 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1491 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1492 sk_backlog_rcv(sk, skb);
1494 /* Clear memory counter. */
1495 tp->ucopy.memory = 0;
1498 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1500 struct sk_buff *skb;
1503 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1504 offset = seq - TCP_SKB_CB(skb)->seq;
1505 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1506 pr_err_once("%s: found a SYN, please report !\n", __func__);
1509 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1513 /* This looks weird, but this can happen if TCP collapsing
1514 * splitted a fat GRO packet, while we released socket lock
1515 * in skb_splice_bits()
1517 sk_eat_skb(sk, skb);
1523 * This routine provides an alternative to tcp_recvmsg() for routines
1524 * that would like to handle copying from skbuffs directly in 'sendfile'
1527 * - It is assumed that the socket was locked by the caller.
1528 * - The routine does not block.
1529 * - At present, there is no support for reading OOB data
1530 * or for 'peeking' the socket using this routine
1531 * (although both would be easy to implement).
1533 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1534 sk_read_actor_t recv_actor)
1536 struct sk_buff *skb;
1537 struct tcp_sock *tp = tcp_sk(sk);
1538 u32 seq = tp->copied_seq;
1542 if (sk->sk_state == TCP_LISTEN)
1544 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1545 if (offset < skb->len) {
1549 len = skb->len - offset;
1550 /* Stop reading if we hit a patch of urgent data */
1552 u32 urg_offset = tp->urg_seq - seq;
1553 if (urg_offset < len)
1558 used = recv_actor(desc, skb, offset, len);
1563 } else if (used <= len) {
1568 /* If recv_actor drops the lock (e.g. TCP splice
1569 * receive) the skb pointer might be invalid when
1570 * getting here: tcp_collapse might have deleted it
1571 * while aggregating skbs from the socket queue.
1573 skb = tcp_recv_skb(sk, seq - 1, &offset);
1576 /* TCP coalescing might have appended data to the skb.
1577 * Try to splice more frags
1579 if (offset + 1 != skb->len)
1582 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1583 sk_eat_skb(sk, skb);
1587 sk_eat_skb(sk, skb);
1590 tp->copied_seq = seq;
1592 tp->copied_seq = seq;
1594 tcp_rcv_space_adjust(sk);
1596 /* Clean up data we have read: This will do ACK frames. */
1598 tcp_recv_skb(sk, seq, &offset);
1599 tcp_cleanup_rbuf(sk, copied);
1603 EXPORT_SYMBOL(tcp_read_sock);
1605 int tcp_peek_len(struct socket *sock)
1607 return tcp_inq(sock->sk);
1609 EXPORT_SYMBOL(tcp_peek_len);
1612 * This routine copies from a sock struct into the user buffer.
1614 * Technical note: in 2.3 we work on _locked_ socket, so that
1615 * tricks with *seq access order and skb->users are not required.
1616 * Probably, code can be easily improved even more.
1619 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1620 int flags, int *addr_len)
1622 struct tcp_sock *tp = tcp_sk(sk);
1628 int target; /* Read at least this many bytes */
1630 struct task_struct *user_recv = NULL;
1631 struct sk_buff *skb, *last;
1634 if (unlikely(flags & MSG_ERRQUEUE))
1635 return inet_recv_error(sk, msg, len, addr_len);
1637 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1638 (sk->sk_state == TCP_ESTABLISHED))
1639 sk_busy_loop(sk, nonblock);
1644 if (sk->sk_state == TCP_LISTEN)
1647 timeo = sock_rcvtimeo(sk, nonblock);
1649 /* Urgent data needs to be handled specially. */
1650 if (flags & MSG_OOB)
1653 if (unlikely(tp->repair)) {
1655 if (!(flags & MSG_PEEK))
1658 if (tp->repair_queue == TCP_SEND_QUEUE)
1662 if (tp->repair_queue == TCP_NO_QUEUE)
1665 /* 'common' recv queue MSG_PEEK-ing */
1668 seq = &tp->copied_seq;
1669 if (flags & MSG_PEEK) {
1670 peek_seq = tp->copied_seq;
1674 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1679 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1680 if (tp->urg_data && tp->urg_seq == *seq) {
1683 if (signal_pending(current)) {
1684 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1689 /* Next get a buffer. */
1691 last = skb_peek_tail(&sk->sk_receive_queue);
1692 skb_queue_walk(&sk->sk_receive_queue, skb) {
1694 /* Now that we have two receive queues this
1697 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1698 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1699 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1703 offset = *seq - TCP_SKB_CB(skb)->seq;
1704 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1705 pr_err_once("%s: found a SYN, please report !\n", __func__);
1708 if (offset < skb->len)
1710 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1712 WARN(!(flags & MSG_PEEK),
1713 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1714 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1717 /* Well, if we have backlog, try to process it now yet. */
1719 if (copied >= target && !sk->sk_backlog.tail)
1724 sk->sk_state == TCP_CLOSE ||
1725 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1727 signal_pending(current))
1730 if (sock_flag(sk, SOCK_DONE))
1734 copied = sock_error(sk);
1738 if (sk->sk_shutdown & RCV_SHUTDOWN)
1741 if (sk->sk_state == TCP_CLOSE) {
1742 if (!sock_flag(sk, SOCK_DONE)) {
1743 /* This occurs when user tries to read
1744 * from never connected socket.
1757 if (signal_pending(current)) {
1758 copied = sock_intr_errno(timeo);
1763 tcp_cleanup_rbuf(sk, copied);
1765 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1766 /* Install new reader */
1767 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1768 user_recv = current;
1769 tp->ucopy.task = user_recv;
1770 tp->ucopy.msg = msg;
1773 tp->ucopy.len = len;
1775 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1776 !(flags & (MSG_PEEK | MSG_TRUNC)));
1778 /* Ugly... If prequeue is not empty, we have to
1779 * process it before releasing socket, otherwise
1780 * order will be broken at second iteration.
1781 * More elegant solution is required!!!
1783 * Look: we have the following (pseudo)queues:
1785 * 1. packets in flight
1790 * Each queue can be processed only if the next ones
1791 * are empty. At this point we have empty receive_queue.
1792 * But prequeue _can_ be not empty after 2nd iteration,
1793 * when we jumped to start of loop because backlog
1794 * processing added something to receive_queue.
1795 * We cannot release_sock(), because backlog contains
1796 * packets arrived _after_ prequeued ones.
1798 * Shortly, algorithm is clear --- to process all
1799 * the queues in order. We could make it more directly,
1800 * requeueing packets from backlog to prequeue, if
1801 * is not empty. It is more elegant, but eats cycles,
1804 if (!skb_queue_empty(&tp->ucopy.prequeue))
1807 /* __ Set realtime policy in scheduler __ */
1810 if (copied >= target) {
1811 /* Do not sleep, just process backlog. */
1815 sk_wait_data(sk, &timeo, last);
1821 /* __ Restore normal policy in scheduler __ */
1823 chunk = len - tp->ucopy.len;
1825 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1830 if (tp->rcv_nxt == tp->copied_seq &&
1831 !skb_queue_empty(&tp->ucopy.prequeue)) {
1833 tcp_prequeue_process(sk);
1835 chunk = len - tp->ucopy.len;
1837 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1843 if ((flags & MSG_PEEK) &&
1844 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1845 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1847 task_pid_nr(current));
1848 peek_seq = tp->copied_seq;
1853 /* Ok so how much can we use? */
1854 used = skb->len - offset;
1858 /* Do we have urgent data here? */
1860 u32 urg_offset = tp->urg_seq - *seq;
1861 if (urg_offset < used) {
1863 if (!sock_flag(sk, SOCK_URGINLINE)) {
1876 if (!(flags & MSG_TRUNC)) {
1877 err = skb_copy_datagram_msg(skb, offset, msg, used);
1879 /* Exception. Bailout! */
1890 tcp_rcv_space_adjust(sk);
1893 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1895 tcp_fast_path_check(sk);
1897 if (used + offset < skb->len)
1900 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1902 if (!(flags & MSG_PEEK))
1903 sk_eat_skb(sk, skb);
1907 /* Process the FIN. */
1909 if (!(flags & MSG_PEEK))
1910 sk_eat_skb(sk, skb);
1915 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1918 tp->ucopy.len = copied > 0 ? len : 0;
1920 tcp_prequeue_process(sk);
1922 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1923 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1929 tp->ucopy.task = NULL;
1933 /* According to UNIX98, msg_name/msg_namelen are ignored
1934 * on connected socket. I was just happy when found this 8) --ANK
1937 /* Clean up data we have read: This will do ACK frames. */
1938 tcp_cleanup_rbuf(sk, copied);
1948 err = tcp_recv_urg(sk, msg, len, flags);
1952 err = tcp_peek_sndq(sk, msg, len);
1955 EXPORT_SYMBOL(tcp_recvmsg);
1957 void tcp_set_state(struct sock *sk, int state)
1959 int oldstate = sk->sk_state;
1962 case TCP_ESTABLISHED:
1963 if (oldstate != TCP_ESTABLISHED)
1964 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1968 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1969 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1971 sk->sk_prot->unhash(sk);
1972 if (inet_csk(sk)->icsk_bind_hash &&
1973 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1977 if (oldstate == TCP_ESTABLISHED)
1978 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1981 /* Change state AFTER socket is unhashed to avoid closed
1982 * socket sitting in hash tables.
1984 sk_state_store(sk, state);
1987 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1990 EXPORT_SYMBOL_GPL(tcp_set_state);
1993 * State processing on a close. This implements the state shift for
1994 * sending our FIN frame. Note that we only send a FIN for some
1995 * states. A shutdown() may have already sent the FIN, or we may be
1999 static const unsigned char new_state[16] = {
2000 /* current state: new state: action: */
2001 [0 /* (Invalid) */] = TCP_CLOSE,
2002 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2003 [TCP_SYN_SENT] = TCP_CLOSE,
2004 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2005 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2006 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2007 [TCP_TIME_WAIT] = TCP_CLOSE,
2008 [TCP_CLOSE] = TCP_CLOSE,
2009 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2010 [TCP_LAST_ACK] = TCP_LAST_ACK,
2011 [TCP_LISTEN] = TCP_CLOSE,
2012 [TCP_CLOSING] = TCP_CLOSING,
2013 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2016 static int tcp_close_state(struct sock *sk)
2018 int next = (int)new_state[sk->sk_state];
2019 int ns = next & TCP_STATE_MASK;
2021 tcp_set_state(sk, ns);
2023 return next & TCP_ACTION_FIN;
2027 * Shutdown the sending side of a connection. Much like close except
2028 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2031 void tcp_shutdown(struct sock *sk, int how)
2033 /* We need to grab some memory, and put together a FIN,
2034 * and then put it into the queue to be sent.
2035 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2037 if (!(how & SEND_SHUTDOWN))
2040 /* If we've already sent a FIN, or it's a closed state, skip this. */
2041 if ((1 << sk->sk_state) &
2042 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2043 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2044 /* Clear out any half completed packets. FIN if needed. */
2045 if (tcp_close_state(sk))
2049 EXPORT_SYMBOL(tcp_shutdown);
2051 bool tcp_check_oom(struct sock *sk, int shift)
2053 bool too_many_orphans, out_of_socket_memory;
2055 too_many_orphans = tcp_too_many_orphans(sk, shift);
2056 out_of_socket_memory = tcp_out_of_memory(sk);
2058 if (too_many_orphans)
2059 net_info_ratelimited("too many orphaned sockets\n");
2060 if (out_of_socket_memory)
2061 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2062 return too_many_orphans || out_of_socket_memory;
2065 void tcp_close(struct sock *sk, long timeout)
2067 struct sk_buff *skb;
2068 int data_was_unread = 0;
2072 sk->sk_shutdown = SHUTDOWN_MASK;
2074 if (sk->sk_state == TCP_LISTEN) {
2075 tcp_set_state(sk, TCP_CLOSE);
2078 inet_csk_listen_stop(sk);
2080 goto adjudge_to_death;
2083 /* We need to flush the recv. buffs. We do this only on the
2084 * descriptor close, not protocol-sourced closes, because the
2085 * reader process may not have drained the data yet!
2087 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2088 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2090 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2092 data_was_unread += len;
2098 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2099 if (sk->sk_state == TCP_CLOSE)
2100 goto adjudge_to_death;
2102 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2103 * data was lost. To witness the awful effects of the old behavior of
2104 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2105 * GET in an FTP client, suspend the process, wait for the client to
2106 * advertise a zero window, then kill -9 the FTP client, wheee...
2107 * Note: timeout is always zero in such a case.
2109 if (unlikely(tcp_sk(sk)->repair)) {
2110 sk->sk_prot->disconnect(sk, 0);
2111 } else if (data_was_unread) {
2112 /* Unread data was tossed, zap the connection. */
2113 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2114 tcp_set_state(sk, TCP_CLOSE);
2115 tcp_send_active_reset(sk, sk->sk_allocation);
2116 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2117 /* Check zero linger _after_ checking for unread data. */
2118 sk->sk_prot->disconnect(sk, 0);
2119 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2120 } else if (tcp_close_state(sk)) {
2121 /* We FIN if the application ate all the data before
2122 * zapping the connection.
2125 /* RED-PEN. Formally speaking, we have broken TCP state
2126 * machine. State transitions:
2128 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2129 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2130 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2132 * are legal only when FIN has been sent (i.e. in window),
2133 * rather than queued out of window. Purists blame.
2135 * F.e. "RFC state" is ESTABLISHED,
2136 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2138 * The visible declinations are that sometimes
2139 * we enter time-wait state, when it is not required really
2140 * (harmless), do not send active resets, when they are
2141 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2142 * they look as CLOSING or LAST_ACK for Linux)
2143 * Probably, I missed some more holelets.
2145 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2146 * in a single packet! (May consider it later but will
2147 * probably need API support or TCP_CORK SYN-ACK until
2148 * data is written and socket is closed.)
2153 sk_stream_wait_close(sk, timeout);
2156 state = sk->sk_state;
2160 /* It is the last release_sock in its life. It will remove backlog. */
2164 /* Now socket is owned by kernel and we acquire BH lock
2165 to finish close. No need to check for user refs.
2169 WARN_ON(sock_owned_by_user(sk));
2171 percpu_counter_inc(sk->sk_prot->orphan_count);
2173 /* Have we already been destroyed by a softirq or backlog? */
2174 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2177 /* This is a (useful) BSD violating of the RFC. There is a
2178 * problem with TCP as specified in that the other end could
2179 * keep a socket open forever with no application left this end.
2180 * We use a 1 minute timeout (about the same as BSD) then kill
2181 * our end. If they send after that then tough - BUT: long enough
2182 * that we won't make the old 4*rto = almost no time - whoops
2185 * Nope, it was not mistake. It is really desired behaviour
2186 * f.e. on http servers, when such sockets are useless, but
2187 * consume significant resources. Let's do it with special
2188 * linger2 option. --ANK
2191 if (sk->sk_state == TCP_FIN_WAIT2) {
2192 struct tcp_sock *tp = tcp_sk(sk);
2193 if (tp->linger2 < 0) {
2194 tcp_set_state(sk, TCP_CLOSE);
2195 tcp_send_active_reset(sk, GFP_ATOMIC);
2196 __NET_INC_STATS(sock_net(sk),
2197 LINUX_MIB_TCPABORTONLINGER);
2199 const int tmo = tcp_fin_time(sk);
2201 if (tmo > TCP_TIMEWAIT_LEN) {
2202 inet_csk_reset_keepalive_timer(sk,
2203 tmo - TCP_TIMEWAIT_LEN);
2205 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2210 if (sk->sk_state != TCP_CLOSE) {
2212 if (tcp_check_oom(sk, 0)) {
2213 tcp_set_state(sk, TCP_CLOSE);
2214 tcp_send_active_reset(sk, GFP_ATOMIC);
2215 __NET_INC_STATS(sock_net(sk),
2216 LINUX_MIB_TCPABORTONMEMORY);
2220 if (sk->sk_state == TCP_CLOSE) {
2221 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2222 /* We could get here with a non-NULL req if the socket is
2223 * aborted (e.g., closed with unread data) before 3WHS
2227 reqsk_fastopen_remove(sk, req, false);
2228 inet_csk_destroy_sock(sk);
2230 /* Otherwise, socket is reprieved until protocol close. */
2237 EXPORT_SYMBOL(tcp_close);
2239 /* These states need RST on ABORT according to RFC793 */
2241 static inline bool tcp_need_reset(int state)
2243 return (1 << state) &
2244 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2245 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2248 int tcp_disconnect(struct sock *sk, int flags)
2250 struct inet_sock *inet = inet_sk(sk);
2251 struct inet_connection_sock *icsk = inet_csk(sk);
2252 struct tcp_sock *tp = tcp_sk(sk);
2254 int old_state = sk->sk_state;
2256 if (old_state != TCP_CLOSE)
2257 tcp_set_state(sk, TCP_CLOSE);
2259 /* ABORT function of RFC793 */
2260 if (old_state == TCP_LISTEN) {
2261 inet_csk_listen_stop(sk);
2262 } else if (unlikely(tp->repair)) {
2263 sk->sk_err = ECONNABORTED;
2264 } else if (tcp_need_reset(old_state) ||
2265 (tp->snd_nxt != tp->write_seq &&
2266 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2267 /* The last check adjusts for discrepancy of Linux wrt. RFC
2270 tcp_send_active_reset(sk, gfp_any());
2271 sk->sk_err = ECONNRESET;
2272 } else if (old_state == TCP_SYN_SENT)
2273 sk->sk_err = ECONNRESET;
2275 tcp_clear_xmit_timers(sk);
2276 __skb_queue_purge(&sk->sk_receive_queue);
2277 tcp_write_queue_purge(sk);
2278 skb_rbtree_purge(&tp->out_of_order_queue);
2280 inet->inet_dport = 0;
2282 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2283 inet_reset_saddr(sk);
2285 sk->sk_shutdown = 0;
2286 sock_reset_flag(sk, SOCK_DONE);
2288 tp->write_seq += tp->max_window + 2;
2289 if (tp->write_seq == 0)
2291 icsk->icsk_backoff = 0;
2293 icsk->icsk_probes_out = 0;
2294 tp->packets_out = 0;
2295 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2296 tp->snd_cwnd_cnt = 0;
2297 tp->window_clamp = 0;
2298 tcp_set_ca_state(sk, TCP_CA_Open);
2299 tcp_clear_retrans(tp);
2300 inet_csk_delack_init(sk);
2301 tcp_init_send_head(sk);
2302 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2305 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2307 sk->sk_error_report(sk);
2310 EXPORT_SYMBOL(tcp_disconnect);
2312 static inline bool tcp_can_repair_sock(const struct sock *sk)
2314 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2315 (sk->sk_state != TCP_LISTEN);
2318 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2320 struct tcp_repair_window opt;
2325 if (len != sizeof(opt))
2328 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2331 if (opt.max_window < opt.snd_wnd)
2334 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2337 if (after(opt.rcv_wup, tp->rcv_nxt))
2340 tp->snd_wl1 = opt.snd_wl1;
2341 tp->snd_wnd = opt.snd_wnd;
2342 tp->max_window = opt.max_window;
2344 tp->rcv_wnd = opt.rcv_wnd;
2345 tp->rcv_wup = opt.rcv_wup;
2350 static int tcp_repair_options_est(struct tcp_sock *tp,
2351 struct tcp_repair_opt __user *optbuf, unsigned int len)
2353 struct tcp_repair_opt opt;
2355 while (len >= sizeof(opt)) {
2356 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2362 switch (opt.opt_code) {
2364 tp->rx_opt.mss_clamp = opt.opt_val;
2368 u16 snd_wscale = opt.opt_val & 0xFFFF;
2369 u16 rcv_wscale = opt.opt_val >> 16;
2371 if (snd_wscale > 14 || rcv_wscale > 14)
2374 tp->rx_opt.snd_wscale = snd_wscale;
2375 tp->rx_opt.rcv_wscale = rcv_wscale;
2376 tp->rx_opt.wscale_ok = 1;
2379 case TCPOPT_SACK_PERM:
2380 if (opt.opt_val != 0)
2383 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2384 if (sysctl_tcp_fack)
2385 tcp_enable_fack(tp);
2387 case TCPOPT_TIMESTAMP:
2388 if (opt.opt_val != 0)
2391 tp->rx_opt.tstamp_ok = 1;
2400 * Socket option code for TCP.
2402 static int do_tcp_setsockopt(struct sock *sk, int level,
2403 int optname, char __user *optval, unsigned int optlen)
2405 struct tcp_sock *tp = tcp_sk(sk);
2406 struct inet_connection_sock *icsk = inet_csk(sk);
2407 struct net *net = sock_net(sk);
2411 /* These are data/string values, all the others are ints */
2413 case TCP_CONGESTION: {
2414 char name[TCP_CA_NAME_MAX];
2419 val = strncpy_from_user(name, optval,
2420 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2426 err = tcp_set_congestion_control(sk, name);
2435 if (optlen < sizeof(int))
2438 if (get_user(val, (int __user *)optval))
2445 /* Values greater than interface MTU won't take effect. However
2446 * at the point when this call is done we typically don't yet
2447 * know which interface is going to be used */
2448 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2452 tp->rx_opt.user_mss = val;
2457 /* TCP_NODELAY is weaker than TCP_CORK, so that
2458 * this option on corked socket is remembered, but
2459 * it is not activated until cork is cleared.
2461 * However, when TCP_NODELAY is set we make
2462 * an explicit push, which overrides even TCP_CORK
2463 * for currently queued segments.
2465 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2466 tcp_push_pending_frames(sk);
2468 tp->nonagle &= ~TCP_NAGLE_OFF;
2472 case TCP_THIN_LINEAR_TIMEOUTS:
2473 if (val < 0 || val > 1)
2479 case TCP_THIN_DUPACK:
2480 if (val < 0 || val > 1)
2483 tp->thin_dupack = val;
2484 if (tp->thin_dupack)
2485 tcp_disable_early_retrans(tp);
2490 if (!tcp_can_repair_sock(sk))
2492 else if (val == 1) {
2494 sk->sk_reuse = SK_FORCE_REUSE;
2495 tp->repair_queue = TCP_NO_QUEUE;
2496 } else if (val == 0) {
2498 sk->sk_reuse = SK_NO_REUSE;
2499 tcp_send_window_probe(sk);
2505 case TCP_REPAIR_QUEUE:
2508 else if (val < TCP_QUEUES_NR)
2509 tp->repair_queue = val;
2515 if (sk->sk_state != TCP_CLOSE)
2517 else if (tp->repair_queue == TCP_SEND_QUEUE)
2518 tp->write_seq = val;
2519 else if (tp->repair_queue == TCP_RECV_QUEUE)
2525 case TCP_REPAIR_OPTIONS:
2528 else if (sk->sk_state == TCP_ESTABLISHED)
2529 err = tcp_repair_options_est(tp,
2530 (struct tcp_repair_opt __user *)optval,
2537 /* When set indicates to always queue non-full frames.
2538 * Later the user clears this option and we transmit
2539 * any pending partial frames in the queue. This is
2540 * meant to be used alongside sendfile() to get properly
2541 * filled frames when the user (for example) must write
2542 * out headers with a write() call first and then use
2543 * sendfile to send out the data parts.
2545 * TCP_CORK can be set together with TCP_NODELAY and it is
2546 * stronger than TCP_NODELAY.
2549 tp->nonagle |= TCP_NAGLE_CORK;
2551 tp->nonagle &= ~TCP_NAGLE_CORK;
2552 if (tp->nonagle&TCP_NAGLE_OFF)
2553 tp->nonagle |= TCP_NAGLE_PUSH;
2554 tcp_push_pending_frames(sk);
2559 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2562 tp->keepalive_time = val * HZ;
2563 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2564 !((1 << sk->sk_state) &
2565 (TCPF_CLOSE | TCPF_LISTEN))) {
2566 u32 elapsed = keepalive_time_elapsed(tp);
2567 if (tp->keepalive_time > elapsed)
2568 elapsed = tp->keepalive_time - elapsed;
2571 inet_csk_reset_keepalive_timer(sk, elapsed);
2576 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2579 tp->keepalive_intvl = val * HZ;
2582 if (val < 1 || val > MAX_TCP_KEEPCNT)
2585 tp->keepalive_probes = val;
2588 if (val < 1 || val > MAX_TCP_SYNCNT)
2591 icsk->icsk_syn_retries = val;
2595 if (val < 0 || val > 1)
2604 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2607 tp->linger2 = val * HZ;
2610 case TCP_DEFER_ACCEPT:
2611 /* Translate value in seconds to number of retransmits */
2612 icsk->icsk_accept_queue.rskq_defer_accept =
2613 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2617 case TCP_WINDOW_CLAMP:
2619 if (sk->sk_state != TCP_CLOSE) {
2623 tp->window_clamp = 0;
2625 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2626 SOCK_MIN_RCVBUF / 2 : val;
2631 icsk->icsk_ack.pingpong = 1;
2633 icsk->icsk_ack.pingpong = 0;
2634 if ((1 << sk->sk_state) &
2635 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2636 inet_csk_ack_scheduled(sk)) {
2637 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2638 tcp_cleanup_rbuf(sk, 1);
2640 icsk->icsk_ack.pingpong = 1;
2645 #ifdef CONFIG_TCP_MD5SIG
2647 /* Read the IP->Key mappings from userspace */
2648 err = tp->af_specific->md5_parse(sk, optval, optlen);
2651 case TCP_USER_TIMEOUT:
2652 /* Cap the max time in ms TCP will retry or probe the window
2653 * before giving up and aborting (ETIMEDOUT) a connection.
2658 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2662 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2664 tcp_fastopen_init_key_once(true);
2666 fastopen_queue_tune(sk, val);
2675 tp->tsoffset = val - tcp_time_stamp;
2677 case TCP_REPAIR_WINDOW:
2678 err = tcp_repair_set_window(tp, optval, optlen);
2680 case TCP_NOTSENT_LOWAT:
2681 tp->notsent_lowat = val;
2682 sk->sk_write_space(sk);
2693 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2694 unsigned int optlen)
2696 const struct inet_connection_sock *icsk = inet_csk(sk);
2698 if (level != SOL_TCP)
2699 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2701 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2703 EXPORT_SYMBOL(tcp_setsockopt);
2705 #ifdef CONFIG_COMPAT
2706 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2707 char __user *optval, unsigned int optlen)
2709 if (level != SOL_TCP)
2710 return inet_csk_compat_setsockopt(sk, level, optname,
2712 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2714 EXPORT_SYMBOL(compat_tcp_setsockopt);
2717 static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
2718 struct tcp_info *info)
2720 u64 stats[__TCP_CHRONO_MAX], total = 0;
2723 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
2724 stats[i] = tp->chrono_stat[i - 1];
2725 if (i == tp->chrono_type)
2726 stats[i] += tcp_time_stamp - tp->chrono_start;
2727 stats[i] *= USEC_PER_SEC / HZ;
2731 info->tcpi_busy_time = total;
2732 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
2733 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
2736 /* Return information about state of tcp endpoint in API format. */
2737 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2739 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2740 const struct inet_connection_sock *icsk = inet_csk(sk);
2741 u32 now = tcp_time_stamp, intv;
2746 memset(info, 0, sizeof(*info));
2747 if (sk->sk_type != SOCK_STREAM)
2750 info->tcpi_state = sk_state_load(sk);
2752 /* Report meaningful fields for all TCP states, including listeners */
2753 rate = READ_ONCE(sk->sk_pacing_rate);
2754 rate64 = rate != ~0U ? rate : ~0ULL;
2755 info->tcpi_pacing_rate = rate64;
2757 rate = READ_ONCE(sk->sk_max_pacing_rate);
2758 rate64 = rate != ~0U ? rate : ~0ULL;
2759 info->tcpi_max_pacing_rate = rate64;
2761 info->tcpi_reordering = tp->reordering;
2762 info->tcpi_snd_cwnd = tp->snd_cwnd;
2764 if (info->tcpi_state == TCP_LISTEN) {
2765 /* listeners aliased fields :
2766 * tcpi_unacked -> Number of children ready for accept()
2767 * tcpi_sacked -> max backlog
2769 info->tcpi_unacked = sk->sk_ack_backlog;
2770 info->tcpi_sacked = sk->sk_max_ack_backlog;
2773 info->tcpi_ca_state = icsk->icsk_ca_state;
2774 info->tcpi_retransmits = icsk->icsk_retransmits;
2775 info->tcpi_probes = icsk->icsk_probes_out;
2776 info->tcpi_backoff = icsk->icsk_backoff;
2778 if (tp->rx_opt.tstamp_ok)
2779 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2780 if (tcp_is_sack(tp))
2781 info->tcpi_options |= TCPI_OPT_SACK;
2782 if (tp->rx_opt.wscale_ok) {
2783 info->tcpi_options |= TCPI_OPT_WSCALE;
2784 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2785 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2788 if (tp->ecn_flags & TCP_ECN_OK)
2789 info->tcpi_options |= TCPI_OPT_ECN;
2790 if (tp->ecn_flags & TCP_ECN_SEEN)
2791 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2792 if (tp->syn_data_acked)
2793 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2795 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2796 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2797 info->tcpi_snd_mss = tp->mss_cache;
2798 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2800 info->tcpi_unacked = tp->packets_out;
2801 info->tcpi_sacked = tp->sacked_out;
2803 info->tcpi_lost = tp->lost_out;
2804 info->tcpi_retrans = tp->retrans_out;
2805 info->tcpi_fackets = tp->fackets_out;
2807 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2808 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2809 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2811 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2812 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2813 info->tcpi_rtt = tp->srtt_us >> 3;
2814 info->tcpi_rttvar = tp->mdev_us >> 2;
2815 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2816 info->tcpi_advmss = tp->advmss;
2818 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2819 info->tcpi_rcv_space = tp->rcvq_space.space;
2821 info->tcpi_total_retrans = tp->total_retrans;
2823 slow = lock_sock_fast(sk);
2825 info->tcpi_bytes_acked = tp->bytes_acked;
2826 info->tcpi_bytes_received = tp->bytes_received;
2827 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
2828 tcp_get_info_chrono_stats(tp, info);
2830 unlock_sock_fast(sk, slow);
2832 info->tcpi_segs_out = tp->segs_out;
2833 info->tcpi_segs_in = tp->segs_in;
2835 info->tcpi_min_rtt = tcp_min_rtt(tp);
2836 info->tcpi_data_segs_in = tp->data_segs_in;
2837 info->tcpi_data_segs_out = tp->data_segs_out;
2839 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
2840 rate = READ_ONCE(tp->rate_delivered);
2841 intv = READ_ONCE(tp->rate_interval_us);
2843 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
2844 do_div(rate64, intv);
2845 info->tcpi_delivery_rate = rate64;
2848 EXPORT_SYMBOL_GPL(tcp_get_info);
2850 struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
2852 const struct tcp_sock *tp = tcp_sk(sk);
2853 struct sk_buff *stats;
2854 struct tcp_info info;
2856 stats = alloc_skb(3 * nla_total_size_64bit(sizeof(u64)), GFP_ATOMIC);
2860 tcp_get_info_chrono_stats(tp, &info);
2861 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
2862 info.tcpi_busy_time, TCP_NLA_PAD);
2863 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
2864 info.tcpi_rwnd_limited, TCP_NLA_PAD);
2865 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
2866 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
2870 static int do_tcp_getsockopt(struct sock *sk, int level,
2871 int optname, char __user *optval, int __user *optlen)
2873 struct inet_connection_sock *icsk = inet_csk(sk);
2874 struct tcp_sock *tp = tcp_sk(sk);
2875 struct net *net = sock_net(sk);
2878 if (get_user(len, optlen))
2881 len = min_t(unsigned int, len, sizeof(int));
2888 val = tp->mss_cache;
2889 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2890 val = tp->rx_opt.user_mss;
2892 val = tp->rx_opt.mss_clamp;
2895 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2898 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2901 val = keepalive_time_when(tp) / HZ;
2904 val = keepalive_intvl_when(tp) / HZ;
2907 val = keepalive_probes(tp);
2910 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
2915 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
2917 case TCP_DEFER_ACCEPT:
2918 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2919 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2921 case TCP_WINDOW_CLAMP:
2922 val = tp->window_clamp;
2925 struct tcp_info info;
2927 if (get_user(len, optlen))
2930 tcp_get_info(sk, &info);
2932 len = min_t(unsigned int, len, sizeof(info));
2933 if (put_user(len, optlen))
2935 if (copy_to_user(optval, &info, len))
2940 const struct tcp_congestion_ops *ca_ops;
2941 union tcp_cc_info info;
2945 if (get_user(len, optlen))
2948 ca_ops = icsk->icsk_ca_ops;
2949 if (ca_ops && ca_ops->get_info)
2950 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2952 len = min_t(unsigned int, len, sz);
2953 if (put_user(len, optlen))
2955 if (copy_to_user(optval, &info, len))
2960 val = !icsk->icsk_ack.pingpong;
2963 case TCP_CONGESTION:
2964 if (get_user(len, optlen))
2966 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2967 if (put_user(len, optlen))
2969 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2973 case TCP_THIN_LINEAR_TIMEOUTS:
2976 case TCP_THIN_DUPACK:
2977 val = tp->thin_dupack;
2984 case TCP_REPAIR_QUEUE:
2986 val = tp->repair_queue;
2991 case TCP_REPAIR_WINDOW: {
2992 struct tcp_repair_window opt;
2994 if (get_user(len, optlen))
2997 if (len != sizeof(opt))
3003 opt.snd_wl1 = tp->snd_wl1;
3004 opt.snd_wnd = tp->snd_wnd;
3005 opt.max_window = tp->max_window;
3006 opt.rcv_wnd = tp->rcv_wnd;
3007 opt.rcv_wup = tp->rcv_wup;
3009 if (copy_to_user(optval, &opt, len))
3014 if (tp->repair_queue == TCP_SEND_QUEUE)
3015 val = tp->write_seq;
3016 else if (tp->repair_queue == TCP_RECV_QUEUE)
3022 case TCP_USER_TIMEOUT:
3023 val = jiffies_to_msecs(icsk->icsk_user_timeout);
3027 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3031 val = tcp_time_stamp + tp->tsoffset;
3033 case TCP_NOTSENT_LOWAT:
3034 val = tp->notsent_lowat;
3039 case TCP_SAVED_SYN: {
3040 if (get_user(len, optlen))
3044 if (tp->saved_syn) {
3045 if (len < tp->saved_syn[0]) {
3046 if (put_user(tp->saved_syn[0], optlen)) {
3053 len = tp->saved_syn[0];
3054 if (put_user(len, optlen)) {
3058 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3062 tcp_saved_syn_free(tp);
3067 if (put_user(len, optlen))
3073 return -ENOPROTOOPT;
3076 if (put_user(len, optlen))
3078 if (copy_to_user(optval, &val, len))
3083 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3086 struct inet_connection_sock *icsk = inet_csk(sk);
3088 if (level != SOL_TCP)
3089 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3091 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3093 EXPORT_SYMBOL(tcp_getsockopt);
3095 #ifdef CONFIG_COMPAT
3096 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3097 char __user *optval, int __user *optlen)
3099 if (level != SOL_TCP)
3100 return inet_csk_compat_getsockopt(sk, level, optname,
3102 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3104 EXPORT_SYMBOL(compat_tcp_getsockopt);
3107 #ifdef CONFIG_TCP_MD5SIG
3108 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3109 static DEFINE_MUTEX(tcp_md5sig_mutex);
3110 static bool tcp_md5sig_pool_populated = false;
3112 static void __tcp_alloc_md5sig_pool(void)
3114 struct crypto_ahash *hash;
3117 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3121 for_each_possible_cpu(cpu) {
3122 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3123 struct ahash_request *req;
3126 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3127 sizeof(struct tcphdr),
3132 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3134 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3137 req = ahash_request_alloc(hash, GFP_KERNEL);
3141 ahash_request_set_callback(req, 0, NULL, NULL);
3143 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3145 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3146 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3149 tcp_md5sig_pool_populated = true;
3152 bool tcp_alloc_md5sig_pool(void)
3154 if (unlikely(!tcp_md5sig_pool_populated)) {
3155 mutex_lock(&tcp_md5sig_mutex);
3157 if (!tcp_md5sig_pool_populated)
3158 __tcp_alloc_md5sig_pool();
3160 mutex_unlock(&tcp_md5sig_mutex);
3162 return tcp_md5sig_pool_populated;
3164 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3168 * tcp_get_md5sig_pool - get md5sig_pool for this user
3170 * We use percpu structure, so if we succeed, we exit with preemption
3171 * and BH disabled, to make sure another thread or softirq handling
3172 * wont try to get same context.
3174 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3178 if (tcp_md5sig_pool_populated) {
3179 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3181 return this_cpu_ptr(&tcp_md5sig_pool);
3186 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3188 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3189 const struct sk_buff *skb, unsigned int header_len)
3191 struct scatterlist sg;
3192 const struct tcphdr *tp = tcp_hdr(skb);
3193 struct ahash_request *req = hp->md5_req;
3195 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3196 skb_headlen(skb) - header_len : 0;
3197 const struct skb_shared_info *shi = skb_shinfo(skb);
3198 struct sk_buff *frag_iter;
3200 sg_init_table(&sg, 1);
3202 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3203 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3204 if (crypto_ahash_update(req))
3207 for (i = 0; i < shi->nr_frags; ++i) {
3208 const struct skb_frag_struct *f = &shi->frags[i];
3209 unsigned int offset = f->page_offset;
3210 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3212 sg_set_page(&sg, page, skb_frag_size(f),
3213 offset_in_page(offset));
3214 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3215 if (crypto_ahash_update(req))
3219 skb_walk_frags(skb, frag_iter)
3220 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3225 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3227 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3229 struct scatterlist sg;
3231 sg_init_one(&sg, key->key, key->keylen);
3232 ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3233 return crypto_ahash_update(hp->md5_req);
3235 EXPORT_SYMBOL(tcp_md5_hash_key);
3239 void tcp_done(struct sock *sk)
3241 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3243 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3244 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3246 tcp_set_state(sk, TCP_CLOSE);
3247 tcp_clear_xmit_timers(sk);
3249 reqsk_fastopen_remove(sk, req, false);
3251 sk->sk_shutdown = SHUTDOWN_MASK;
3253 if (!sock_flag(sk, SOCK_DEAD))
3254 sk->sk_state_change(sk);
3256 inet_csk_destroy_sock(sk);
3258 EXPORT_SYMBOL_GPL(tcp_done);
3260 int tcp_abort(struct sock *sk, int err)
3262 if (!sk_fullsock(sk)) {
3263 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3264 struct request_sock *req = inet_reqsk(sk);
3267 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener,
3275 /* Don't race with userspace socket closes such as tcp_close. */
3278 if (sk->sk_state == TCP_LISTEN) {
3279 tcp_set_state(sk, TCP_CLOSE);
3280 inet_csk_listen_stop(sk);
3283 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3287 if (!sock_flag(sk, SOCK_DEAD)) {
3289 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3291 sk->sk_error_report(sk);
3292 if (tcp_need_reset(sk->sk_state))
3293 tcp_send_active_reset(sk, GFP_ATOMIC);
3302 EXPORT_SYMBOL_GPL(tcp_abort);
3304 extern struct tcp_congestion_ops tcp_reno;
3306 static __initdata unsigned long thash_entries;
3307 static int __init set_thash_entries(char *str)
3314 ret = kstrtoul(str, 0, &thash_entries);
3320 __setup("thash_entries=", set_thash_entries);
3322 static void __init tcp_init_mem(void)
3324 unsigned long limit = nr_free_buffer_pages() / 16;
3326 limit = max(limit, 128UL);
3327 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3328 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3329 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3332 void __init tcp_init(void)
3334 int max_rshare, max_wshare, cnt;
3335 unsigned long limit;
3338 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3339 FIELD_SIZEOF(struct sk_buff, cb));
3341 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3342 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3343 tcp_hashinfo.bind_bucket_cachep =
3344 kmem_cache_create("tcp_bind_bucket",
3345 sizeof(struct inet_bind_bucket), 0,
3346 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3348 /* Size and allocate the main established and bind bucket
3351 * The methodology is similar to that of the buffer cache.
3353 tcp_hashinfo.ehash =
3354 alloc_large_system_hash("TCP established",
3355 sizeof(struct inet_ehash_bucket),
3357 17, /* one slot per 128 KB of memory */
3360 &tcp_hashinfo.ehash_mask,
3362 thash_entries ? 0 : 512 * 1024);
3363 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3364 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3366 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3367 panic("TCP: failed to alloc ehash_locks");
3368 tcp_hashinfo.bhash =
3369 alloc_large_system_hash("TCP bind",
3370 sizeof(struct inet_bind_hashbucket),
3371 tcp_hashinfo.ehash_mask + 1,
3372 17, /* one slot per 128 KB of memory */
3374 &tcp_hashinfo.bhash_size,
3378 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3379 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3380 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3381 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3385 cnt = tcp_hashinfo.ehash_mask + 1;
3387 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3388 sysctl_tcp_max_orphans = cnt / 2;
3389 sysctl_max_syn_backlog = max(128, cnt / 256);
3392 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3393 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3394 max_wshare = min(4UL*1024*1024, limit);
3395 max_rshare = min(6UL*1024*1024, limit);
3397 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3398 sysctl_tcp_wmem[1] = 16*1024;
3399 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3401 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3402 sysctl_tcp_rmem[1] = 87380;
3403 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3405 pr_info("Hash tables configured (established %u bind %u)\n",
3406 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3409 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);