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 <asm/uaccess.h>
281 #include <asm/ioctls.h>
282 #include <asm/unaligned.h>
283 #include <net/busy_poll.h>
285 int sysctl_tcp_min_tso_segs __read_mostly = 2;
287 int sysctl_tcp_autocorking __read_mostly = 1;
289 struct percpu_counter tcp_orphan_count;
290 EXPORT_SYMBOL_GPL(tcp_orphan_count);
292 long sysctl_tcp_mem[3] __read_mostly;
293 int sysctl_tcp_wmem[3] __read_mostly;
294 int sysctl_tcp_rmem[3] __read_mostly;
296 EXPORT_SYMBOL(sysctl_tcp_mem);
297 EXPORT_SYMBOL(sysctl_tcp_rmem);
298 EXPORT_SYMBOL(sysctl_tcp_wmem);
300 int sysctl_tcp_delack_seg __read_mostly = TCP_DELACK_SEG;
301 EXPORT_SYMBOL(sysctl_tcp_delack_seg);
303 int sysctl_tcp_use_userconfig __read_mostly;
304 EXPORT_SYMBOL(sysctl_tcp_use_userconfig);
306 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
307 EXPORT_SYMBOL(tcp_memory_allocated);
310 * Current number of TCP sockets.
312 struct percpu_counter tcp_sockets_allocated;
313 EXPORT_SYMBOL(tcp_sockets_allocated);
318 struct tcp_splice_state {
319 struct pipe_inode_info *pipe;
325 * Pressure flag: try to collapse.
326 * Technical note: it is used by multiple contexts non atomically.
327 * All the __sk_mem_schedule() is of this nature: accounting
328 * is strict, actions are advisory and have some latency.
330 int tcp_memory_pressure __read_mostly;
331 EXPORT_SYMBOL(tcp_memory_pressure);
333 void tcp_enter_memory_pressure(struct sock *sk)
335 if (!tcp_memory_pressure) {
336 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
337 tcp_memory_pressure = 1;
340 EXPORT_SYMBOL(tcp_enter_memory_pressure);
342 /* Convert seconds to retransmits based on initial and max timeout */
343 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
348 int period = timeout;
351 while (seconds > period && res < 255) {
354 if (timeout > rto_max)
362 /* Convert retransmits to seconds based on initial and max timeout */
363 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
371 if (timeout > rto_max)
379 /* Address-family independent initialization for a tcp_sock.
381 * NOTE: A lot of things set to zero explicitly by call to
382 * sk_alloc() so need not be done here.
384 void tcp_init_sock(struct sock *sk)
386 struct inet_connection_sock *icsk = inet_csk(sk);
387 struct tcp_sock *tp = tcp_sk(sk);
389 tp->out_of_order_queue = RB_ROOT;
390 tcp_init_xmit_timers(sk);
391 tcp_prequeue_init(tp);
392 INIT_LIST_HEAD(&tp->tsq_node);
394 icsk->icsk_rto = TCP_TIMEOUT_INIT;
395 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
396 minmax_reset(&tp->rtt_min, tcp_time_stamp, ~0U);
398 /* So many TCP implementations out there (incorrectly) count the
399 * initial SYN frame in their delayed-ACK and congestion control
400 * algorithms that we must have the following bandaid to talk
401 * efficiently to them. -DaveM
403 tp->snd_cwnd = TCP_INIT_CWND;
405 /* There's a bubble in the pipe until at least the first ACK. */
406 tp->app_limited = ~0U;
408 /* See draft-stevens-tcpca-spec-01 for discussion of the
409 * initialization of these values.
411 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
412 tp->snd_cwnd_clamp = ~0;
413 tp->mss_cache = TCP_MSS_DEFAULT;
414 u64_stats_init(&tp->syncp);
416 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
417 tcp_enable_early_retrans(tp);
418 tcp_assign_congestion_control(sk);
422 sk->sk_state = TCP_CLOSE;
424 sk->sk_write_space = sk_stream_write_space;
425 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
427 icsk->icsk_sync_mss = tcp_sync_mss;
429 sk->sk_sndbuf = sysctl_tcp_wmem[1];
430 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
433 sk_sockets_allocated_inc(sk);
436 EXPORT_SYMBOL(tcp_init_sock);
438 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags, struct sk_buff *skb)
440 if (tsflags && skb) {
441 struct skb_shared_info *shinfo = skb_shinfo(skb);
442 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
444 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
445 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
446 tcb->txstamp_ack = 1;
447 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
448 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
453 * Wait for a TCP event.
455 * Note that we don't need to lock the socket, as the upper poll layers
456 * take care of normal races (between the test and the event) and we don't
457 * go look at any of the socket buffers directly.
459 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
462 struct sock *sk = sock->sk;
463 const struct tcp_sock *tp = tcp_sk(sk);
466 sock_rps_record_flow(sk);
468 sock_poll_wait(file, sk_sleep(sk), wait);
470 state = sk_state_load(sk);
471 if (state == TCP_LISTEN)
472 return inet_csk_listen_poll(sk);
474 /* Socket is not locked. We are protected from async events
475 * by poll logic and correct handling of state changes
476 * made by other threads is impossible in any case.
482 * POLLHUP is certainly not done right. But poll() doesn't
483 * have a notion of HUP in just one direction, and for a
484 * socket the read side is more interesting.
486 * Some poll() documentation says that POLLHUP is incompatible
487 * with the POLLOUT/POLLWR flags, so somebody should check this
488 * all. But careful, it tends to be safer to return too many
489 * bits than too few, and you can easily break real applications
490 * if you don't tell them that something has hung up!
494 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
495 * our fs/select.c). It means that after we received EOF,
496 * poll always returns immediately, making impossible poll() on write()
497 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
498 * if and only if shutdown has been made in both directions.
499 * Actually, it is interesting to look how Solaris and DUX
500 * solve this dilemma. I would prefer, if POLLHUP were maskable,
501 * then we could set it on SND_SHUTDOWN. BTW examples given
502 * in Stevens' books assume exactly this behaviour, it explains
503 * why POLLHUP is incompatible with POLLOUT. --ANK
505 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
506 * blocking on fresh not-connected or disconnected socket. --ANK
508 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
510 if (sk->sk_shutdown & RCV_SHUTDOWN)
511 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
513 /* Connected or passive Fast Open socket? */
514 if (state != TCP_SYN_SENT &&
515 (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
516 int target = sock_rcvlowat(sk, 0, INT_MAX);
518 if (tp->urg_seq == tp->copied_seq &&
519 !sock_flag(sk, SOCK_URGINLINE) &&
523 if (tp->rcv_nxt - tp->copied_seq >= target)
524 mask |= POLLIN | POLLRDNORM;
526 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
527 if (sk_stream_is_writeable(sk)) {
528 mask |= POLLOUT | POLLWRNORM;
529 } else { /* send SIGIO later */
530 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
531 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
533 /* Race breaker. If space is freed after
534 * wspace test but before the flags are set,
535 * IO signal will be lost. Memory barrier
536 * pairs with the input side.
538 smp_mb__after_atomic();
539 if (sk_stream_is_writeable(sk))
540 mask |= POLLOUT | POLLWRNORM;
543 mask |= POLLOUT | POLLWRNORM;
545 if (tp->urg_data & TCP_URG_VALID)
547 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
548 /* Active TCP fastopen socket with defer_connect
549 * Return POLLOUT so application can call write()
550 * in order for kernel to generate SYN+data
552 mask |= POLLOUT | POLLWRNORM;
554 /* This barrier is coupled with smp_wmb() in tcp_reset() */
556 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
561 EXPORT_SYMBOL(tcp_poll);
563 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
565 struct tcp_sock *tp = tcp_sk(sk);
571 if (sk->sk_state == TCP_LISTEN)
574 slow = lock_sock_fast(sk);
576 unlock_sock_fast(sk, slow);
579 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
582 if (sk->sk_state == TCP_LISTEN)
585 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
588 answ = tp->write_seq - tp->snd_una;
591 if (sk->sk_state == TCP_LISTEN)
594 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
597 answ = tp->write_seq - tp->snd_nxt;
603 return put_user(answ, (int __user *)arg);
605 EXPORT_SYMBOL(tcp_ioctl);
607 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
609 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
610 tp->pushed_seq = tp->write_seq;
613 static inline bool forced_push(const struct tcp_sock *tp)
615 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
618 static void skb_entail(struct sock *sk, struct sk_buff *skb)
620 struct tcp_sock *tp = tcp_sk(sk);
621 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
624 tcb->seq = tcb->end_seq = tp->write_seq;
625 tcb->tcp_flags = TCPHDR_ACK;
627 __skb_header_release(skb);
628 tcp_add_write_queue_tail(sk, skb);
629 sk->sk_wmem_queued += skb->truesize;
630 sk_mem_charge(sk, skb->truesize);
631 if (tp->nonagle & TCP_NAGLE_PUSH)
632 tp->nonagle &= ~TCP_NAGLE_PUSH;
634 tcp_slow_start_after_idle_check(sk);
637 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
640 tp->snd_up = tp->write_seq;
643 /* If a not yet filled skb is pushed, do not send it if
644 * we have data packets in Qdisc or NIC queues :
645 * Because TX completion will happen shortly, it gives a chance
646 * to coalesce future sendmsg() payload into this skb, without
647 * need for a timer, and with no latency trade off.
648 * As packets containing data payload have a bigger truesize
649 * than pure acks (dataless) packets, the last checks prevent
650 * autocorking if we only have an ACK in Qdisc/NIC queues,
651 * or if TX completion was delayed after we processed ACK packet.
653 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
656 return skb->len < size_goal &&
657 sysctl_tcp_autocorking &&
658 skb != tcp_write_queue_head(sk) &&
659 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
662 static void tcp_push(struct sock *sk, int flags, int mss_now,
663 int nonagle, int size_goal)
665 struct tcp_sock *tp = tcp_sk(sk);
668 if (!tcp_send_head(sk))
671 skb = tcp_write_queue_tail(sk);
672 if (!(flags & MSG_MORE) || forced_push(tp))
673 tcp_mark_push(tp, skb);
675 tcp_mark_urg(tp, flags);
677 if (tcp_should_autocork(sk, skb, size_goal)) {
679 /* avoid atomic op if TSQ_THROTTLED bit is already set */
680 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
681 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
682 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
684 /* It is possible TX completion already happened
685 * before we set TSQ_THROTTLED.
687 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
691 if (flags & MSG_MORE)
692 nonagle = TCP_NAGLE_CORK;
694 __tcp_push_pending_frames(sk, mss_now, nonagle);
697 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
698 unsigned int offset, size_t len)
700 struct tcp_splice_state *tss = rd_desc->arg.data;
703 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
704 min(rd_desc->count, len), tss->flags);
706 rd_desc->count -= ret;
710 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
712 /* Store TCP splice context information in read_descriptor_t. */
713 read_descriptor_t rd_desc = {
718 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
722 * tcp_splice_read - splice data from TCP socket to a pipe
723 * @sock: socket to splice from
724 * @ppos: position (not valid)
725 * @pipe: pipe to splice to
726 * @len: number of bytes to splice
727 * @flags: splice modifier flags
730 * Will read pages from given socket and fill them into a pipe.
733 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
734 struct pipe_inode_info *pipe, size_t len,
737 struct sock *sk = sock->sk;
738 struct tcp_splice_state tss = {
747 sock_rps_record_flow(sk);
749 * We can't seek on a socket input
758 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
760 ret = __tcp_splice_read(sk, &tss);
766 if (sock_flag(sk, SOCK_DONE))
769 ret = sock_error(sk);
772 if (sk->sk_shutdown & RCV_SHUTDOWN)
774 if (sk->sk_state == TCP_CLOSE) {
776 * This occurs when user tries to read
777 * from never connected socket.
779 if (!sock_flag(sk, SOCK_DONE))
787 /* if __tcp_splice_read() got nothing while we have
788 * an skb in receive queue, we do not want to loop.
789 * This might happen with URG data.
791 if (!skb_queue_empty(&sk->sk_receive_queue))
793 sk_wait_data(sk, &timeo, NULL);
794 if (signal_pending(current)) {
795 ret = sock_intr_errno(timeo);
808 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
809 (sk->sk_shutdown & RCV_SHUTDOWN) ||
810 signal_pending(current))
821 EXPORT_SYMBOL(tcp_splice_read);
823 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
828 /* The TCP header must be at least 32-bit aligned. */
829 size = ALIGN(size, 4);
831 if (unlikely(tcp_under_memory_pressure(sk)))
832 sk_mem_reclaim_partial(sk);
834 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
838 if (force_schedule) {
839 mem_scheduled = true;
840 sk_forced_mem_schedule(sk, skb->truesize);
842 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
844 if (likely(mem_scheduled)) {
845 skb_reserve(skb, sk->sk_prot->max_header);
847 * Make sure that we have exactly size bytes
848 * available to the caller, no more, no less.
850 skb->reserved_tailroom = skb->end - skb->tail - size;
855 sk->sk_prot->enter_memory_pressure(sk);
856 sk_stream_moderate_sndbuf(sk);
861 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
864 struct tcp_sock *tp = tcp_sk(sk);
865 u32 new_size_goal, size_goal;
867 if (!large_allowed || !sk_can_gso(sk))
870 /* Note : tcp_tso_autosize() will eventually split this later */
871 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
872 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
874 /* We try hard to avoid divides here */
875 size_goal = tp->gso_segs * mss_now;
876 if (unlikely(new_size_goal < size_goal ||
877 new_size_goal >= size_goal + mss_now)) {
878 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
879 sk->sk_gso_max_segs);
880 size_goal = tp->gso_segs * mss_now;
883 return max(size_goal, mss_now);
886 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
890 mss_now = tcp_current_mss(sk);
891 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
896 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
897 size_t size, int flags)
899 struct tcp_sock *tp = tcp_sk(sk);
900 int mss_now, size_goal;
903 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
905 /* Wait for a connection to finish. One exception is TCP Fast Open
906 * (passive side) where data is allowed to be sent before a connection
907 * is fully established.
909 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
910 !tcp_passive_fastopen(sk)) {
911 err = sk_stream_wait_connect(sk, &timeo);
916 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
918 mss_now = tcp_send_mss(sk, &size_goal, flags);
922 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
926 struct sk_buff *skb = tcp_write_queue_tail(sk);
930 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0 ||
931 !tcp_skb_can_collapse_to(skb)) {
933 if (!sk_stream_memory_free(sk))
934 goto wait_for_sndbuf;
936 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
937 skb_queue_empty(&sk->sk_write_queue));
939 goto wait_for_memory;
948 i = skb_shinfo(skb)->nr_frags;
949 can_coalesce = skb_can_coalesce(skb, i, page, offset);
950 if (!can_coalesce && i >= sysctl_max_skb_frags) {
951 tcp_mark_push(tp, skb);
954 if (!sk_wmem_schedule(sk, copy))
955 goto wait_for_memory;
958 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
961 skb_fill_page_desc(skb, i, page, offset, copy);
963 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
966 skb->data_len += copy;
967 skb->truesize += copy;
968 sk->sk_wmem_queued += copy;
969 sk_mem_charge(sk, copy);
970 skb->ip_summed = CHECKSUM_PARTIAL;
971 tp->write_seq += copy;
972 TCP_SKB_CB(skb)->end_seq += copy;
973 tcp_skb_pcount_set(skb, 0);
976 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
984 if (skb->len < size_goal || (flags & MSG_OOB))
987 if (forced_push(tp)) {
988 tcp_mark_push(tp, skb);
989 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
990 } else if (skb == tcp_send_head(sk))
991 tcp_push_one(sk, mss_now);
995 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
997 tcp_push(sk, flags & ~MSG_MORE, mss_now,
998 TCP_NAGLE_PUSH, size_goal);
1000 err = sk_stream_wait_memory(sk, &timeo);
1004 mss_now = tcp_send_mss(sk, &size_goal, flags);
1009 tcp_tx_timestamp(sk, sk->sk_tsflags, tcp_write_queue_tail(sk));
1010 if (!(flags & MSG_SENDPAGE_NOTLAST))
1011 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1019 /* make sure we wake any epoll edge trigger waiter */
1020 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1021 sk->sk_write_space(sk);
1022 return sk_stream_error(sk, flags, err);
1025 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1026 size_t size, int flags)
1030 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1031 !sk_check_csum_caps(sk))
1032 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1037 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1039 res = do_tcp_sendpages(sk, page, offset, size, flags);
1043 EXPORT_SYMBOL(tcp_sendpage);
1045 /* Do not bother using a page frag for very small frames.
1046 * But use this heuristic only for the first skb in write queue.
1048 * Having no payload in skb->head allows better SACK shifting
1049 * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1050 * write queue has less skbs.
1051 * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1052 * This also speeds up tso_fragment(), since it wont fallback
1053 * to tcp_fragment().
1055 static int linear_payload_sz(bool first_skb)
1058 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1062 static int select_size(const struct sock *sk, bool sg, bool first_skb)
1064 const struct tcp_sock *tp = tcp_sk(sk);
1065 int tmp = tp->mss_cache;
1068 if (sk_can_gso(sk)) {
1069 tmp = linear_payload_sz(first_skb);
1071 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1073 if (tmp >= pgbreak &&
1074 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1082 void tcp_free_fastopen_req(struct tcp_sock *tp)
1084 if (tp->fastopen_req) {
1085 kfree(tp->fastopen_req);
1086 tp->fastopen_req = NULL;
1090 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1091 int *copied, size_t size)
1093 struct tcp_sock *tp = tcp_sk(sk);
1094 struct inet_sock *inet = inet_sk(sk);
1095 struct sockaddr *uaddr = msg->msg_name;
1098 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1099 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1100 uaddr->sa_family == AF_UNSPEC))
1102 if (tp->fastopen_req)
1103 return -EALREADY; /* Another Fast Open is in progress */
1105 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1107 if (unlikely(!tp->fastopen_req))
1109 tp->fastopen_req->data = msg;
1110 tp->fastopen_req->size = size;
1112 if (inet->defer_connect) {
1113 err = tcp_connect(sk);
1114 /* Same failure procedure as in tcp_v4/6_connect */
1116 tcp_set_state(sk, TCP_CLOSE);
1117 inet->inet_dport = 0;
1118 sk->sk_route_caps = 0;
1121 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1122 err = __inet_stream_connect(sk->sk_socket, uaddr,
1123 msg->msg_namelen, flags);
1124 /* fastopen_req could already be freed in __inet_stream_connect
1125 * if the connection times out or gets rst
1127 if (tp->fastopen_req) {
1128 *copied = tp->fastopen_req->copied;
1129 tcp_free_fastopen_req(tp);
1130 inet->defer_connect = 0;
1135 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1137 struct tcp_sock *tp = tcp_sk(sk);
1138 struct sk_buff *skb;
1139 struct sockcm_cookie sockc;
1140 int flags, err, copied = 0;
1141 int mss_now = 0, size_goal, copied_syn = 0;
1142 bool process_backlog = false;
1148 flags = msg->msg_flags;
1149 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1151 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1152 if (err == -EINPROGRESS && copied_syn > 0)
1158 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1160 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1162 /* Wait for a connection to finish. One exception is TCP Fast Open
1163 * (passive side) where data is allowed to be sent before a connection
1164 * is fully established.
1166 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1167 !tcp_passive_fastopen(sk)) {
1168 err = sk_stream_wait_connect(sk, &timeo);
1173 if (unlikely(tp->repair)) {
1174 if (tp->repair_queue == TCP_RECV_QUEUE) {
1175 copied = tcp_send_rcvq(sk, msg, size);
1180 if (tp->repair_queue == TCP_NO_QUEUE)
1183 /* 'common' sending to sendq */
1186 sockc.tsflags = sk->sk_tsflags;
1187 if (msg->msg_controllen) {
1188 err = sock_cmsg_send(sk, msg, &sockc);
1189 if (unlikely(err)) {
1195 /* This should be in poll */
1196 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1198 /* Ok commence sending. */
1202 mss_now = tcp_send_mss(sk, &size_goal, flags);
1205 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1208 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1210 while (msg_data_left(msg)) {
1212 int max = size_goal;
1214 skb = tcp_write_queue_tail(sk);
1215 if (tcp_send_head(sk)) {
1216 if (skb->ip_summed == CHECKSUM_NONE)
1218 copy = max - skb->len;
1221 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1225 /* Allocate new segment. If the interface is SG,
1226 * allocate skb fitting to single page.
1228 if (!sk_stream_memory_free(sk))
1229 goto wait_for_sndbuf;
1231 if (process_backlog && sk_flush_backlog(sk)) {
1232 process_backlog = false;
1235 first_skb = skb_queue_empty(&sk->sk_write_queue);
1236 skb = sk_stream_alloc_skb(sk,
1237 select_size(sk, sg, first_skb),
1241 goto wait_for_memory;
1243 process_backlog = true;
1245 * Check whether we can use HW checksum.
1247 if (sk_check_csum_caps(sk))
1248 skb->ip_summed = CHECKSUM_PARTIAL;
1250 skb_entail(sk, skb);
1254 /* All packets are restored as if they have
1255 * already been sent. skb_mstamp isn't set to
1256 * avoid wrong rtt estimation.
1259 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1262 /* Try to append data to the end of skb. */
1263 if (copy > msg_data_left(msg))
1264 copy = msg_data_left(msg);
1266 /* Where to copy to? */
1267 if (skb_availroom(skb) > 0) {
1268 /* We have some space in skb head. Superb! */
1269 copy = min_t(int, copy, skb_availroom(skb));
1270 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1275 int i = skb_shinfo(skb)->nr_frags;
1276 struct page_frag *pfrag = sk_page_frag(sk);
1278 if (!sk_page_frag_refill(sk, pfrag))
1279 goto wait_for_memory;
1281 if (!skb_can_coalesce(skb, i, pfrag->page,
1283 if (i >= sysctl_max_skb_frags || !sg) {
1284 tcp_mark_push(tp, skb);
1290 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1292 if (!sk_wmem_schedule(sk, copy))
1293 goto wait_for_memory;
1295 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1302 /* Update the skb. */
1304 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1306 skb_fill_page_desc(skb, i, pfrag->page,
1307 pfrag->offset, copy);
1308 get_page(pfrag->page);
1310 pfrag->offset += copy;
1314 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1316 tp->write_seq += copy;
1317 TCP_SKB_CB(skb)->end_seq += copy;
1318 tcp_skb_pcount_set(skb, 0);
1321 if (!msg_data_left(msg)) {
1322 if (unlikely(flags & MSG_EOR))
1323 TCP_SKB_CB(skb)->eor = 1;
1327 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1330 if (forced_push(tp)) {
1331 tcp_mark_push(tp, skb);
1332 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1333 } else if (skb == tcp_send_head(sk))
1334 tcp_push_one(sk, mss_now);
1338 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1341 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1342 TCP_NAGLE_PUSH, size_goal);
1344 err = sk_stream_wait_memory(sk, &timeo);
1348 mss_now = tcp_send_mss(sk, &size_goal, flags);
1353 tcp_tx_timestamp(sk, sockc.tsflags, tcp_write_queue_tail(sk));
1354 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1358 return copied + copied_syn;
1362 tcp_unlink_write_queue(skb, sk);
1363 /* It is the one place in all of TCP, except connection
1364 * reset, where we can be unlinking the send_head.
1366 tcp_check_send_head(sk, skb);
1367 sk_wmem_free_skb(sk, skb);
1371 if (copied + copied_syn)
1374 err = sk_stream_error(sk, flags, err);
1375 /* make sure we wake any epoll edge trigger waiter */
1376 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1377 sk->sk_write_space(sk);
1381 EXPORT_SYMBOL(tcp_sendmsg);
1384 * Handle reading urgent data. BSD has very simple semantics for
1385 * this, no blocking and very strange errors 8)
1388 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1390 struct tcp_sock *tp = tcp_sk(sk);
1392 /* No URG data to read. */
1393 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1394 tp->urg_data == TCP_URG_READ)
1395 return -EINVAL; /* Yes this is right ! */
1397 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1400 if (tp->urg_data & TCP_URG_VALID) {
1402 char c = tp->urg_data;
1404 if (!(flags & MSG_PEEK))
1405 tp->urg_data = TCP_URG_READ;
1407 /* Read urgent data. */
1408 msg->msg_flags |= MSG_OOB;
1411 if (!(flags & MSG_TRUNC))
1412 err = memcpy_to_msg(msg, &c, 1);
1415 msg->msg_flags |= MSG_TRUNC;
1417 return err ? -EFAULT : len;
1420 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1423 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1424 * the available implementations agree in this case:
1425 * this call should never block, independent of the
1426 * blocking state of the socket.
1427 * Mike <pall@rz.uni-karlsruhe.de>
1432 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1434 struct sk_buff *skb;
1435 int copied = 0, err = 0;
1437 /* XXX -- need to support SO_PEEK_OFF */
1439 skb_queue_walk(&sk->sk_write_queue, skb) {
1440 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1447 return err ?: copied;
1450 /* Clean up the receive buffer for full frames taken by the user,
1451 * then send an ACK if necessary. COPIED is the number of bytes
1452 * tcp_recvmsg has given to the user so far, it speeds up the
1453 * calculation of whether or not we must ACK for the sake of
1456 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1458 struct tcp_sock *tp = tcp_sk(sk);
1459 bool time_to_ack = false;
1461 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1463 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1464 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1465 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1467 if (inet_csk_ack_scheduled(sk)) {
1468 const struct inet_connection_sock *icsk = inet_csk(sk);
1469 /* Delayed ACKs frequently hit locked sockets during bulk
1471 if (icsk->icsk_ack.blocked ||
1472 /* Once-per-sysctl_tcp_delack_seg segments
1473 * ACK was not sent by tcp_input.c
1475 tp->rcv_nxt - tp->rcv_wup > (icsk->icsk_ack.rcv_mss) *
1476 sysctl_tcp_delack_seg ||
1479 * If this read emptied read buffer, we send ACK, if
1480 * connection is not bidirectional, user drained
1481 * receive buffer and there was a small segment
1485 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1486 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1487 !icsk->icsk_ack.pingpong)) &&
1488 !atomic_read(&sk->sk_rmem_alloc)))
1492 /* We send an ACK if we can now advertise a non-zero window
1493 * which has been raised "significantly".
1495 * Even if window raised up to infinity, do not send window open ACK
1496 * in states, where we will not receive more. It is useless.
1498 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1499 __u32 rcv_window_now = tcp_receive_window(tp);
1501 /* Optimize, __tcp_select_window() is not cheap. */
1502 if (2*rcv_window_now <= tp->window_clamp) {
1503 __u32 new_window = __tcp_select_window(sk);
1505 /* Send ACK now, if this read freed lots of space
1506 * in our buffer. Certainly, new_window is new window.
1507 * We can advertise it now, if it is not less than current one.
1508 * "Lots" means "at least twice" here.
1510 if (new_window && new_window >= 2 * rcv_window_now)
1518 static void tcp_prequeue_process(struct sock *sk)
1520 struct sk_buff *skb;
1521 struct tcp_sock *tp = tcp_sk(sk);
1523 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1525 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1526 sk_backlog_rcv(sk, skb);
1528 /* Clear memory counter. */
1529 tp->ucopy.memory = 0;
1532 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1534 struct sk_buff *skb;
1537 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1538 offset = seq - TCP_SKB_CB(skb)->seq;
1539 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1540 pr_err_once("%s: found a SYN, please report !\n", __func__);
1543 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1547 /* This looks weird, but this can happen if TCP collapsing
1548 * splitted a fat GRO packet, while we released socket lock
1549 * in skb_splice_bits()
1551 sk_eat_skb(sk, skb);
1557 * This routine provides an alternative to tcp_recvmsg() for routines
1558 * that would like to handle copying from skbuffs directly in 'sendfile'
1561 * - It is assumed that the socket was locked by the caller.
1562 * - The routine does not block.
1563 * - At present, there is no support for reading OOB data
1564 * or for 'peeking' the socket using this routine
1565 * (although both would be easy to implement).
1567 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1568 sk_read_actor_t recv_actor)
1570 struct sk_buff *skb;
1571 struct tcp_sock *tp = tcp_sk(sk);
1572 u32 seq = tp->copied_seq;
1576 if (sk->sk_state == TCP_LISTEN)
1578 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1579 if (offset < skb->len) {
1583 len = skb->len - offset;
1584 /* Stop reading if we hit a patch of urgent data */
1586 u32 urg_offset = tp->urg_seq - seq;
1587 if (urg_offset < len)
1592 used = recv_actor(desc, skb, offset, len);
1597 } else if (used <= len) {
1602 /* If recv_actor drops the lock (e.g. TCP splice
1603 * receive) the skb pointer might be invalid when
1604 * getting here: tcp_collapse might have deleted it
1605 * while aggregating skbs from the socket queue.
1607 skb = tcp_recv_skb(sk, seq - 1, &offset);
1610 /* TCP coalescing might have appended data to the skb.
1611 * Try to splice more frags
1613 if (offset + 1 != skb->len)
1616 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1617 sk_eat_skb(sk, skb);
1621 sk_eat_skb(sk, skb);
1624 tp->copied_seq = seq;
1626 tp->copied_seq = seq;
1628 tcp_rcv_space_adjust(sk);
1630 /* Clean up data we have read: This will do ACK frames. */
1632 tcp_recv_skb(sk, seq, &offset);
1633 tcp_cleanup_rbuf(sk, copied);
1637 EXPORT_SYMBOL(tcp_read_sock);
1639 int tcp_peek_len(struct socket *sock)
1641 return tcp_inq(sock->sk);
1643 EXPORT_SYMBOL(tcp_peek_len);
1646 * This routine copies from a sock struct into the user buffer.
1648 * Technical note: in 2.3 we work on _locked_ socket, so that
1649 * tricks with *seq access order and skb->users are not required.
1650 * Probably, code can be easily improved even more.
1653 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1654 int flags, int *addr_len)
1656 struct tcp_sock *tp = tcp_sk(sk);
1662 int target; /* Read at least this many bytes */
1664 struct task_struct *user_recv = NULL;
1665 struct sk_buff *skb, *last;
1668 if (unlikely(flags & MSG_ERRQUEUE))
1669 return inet_recv_error(sk, msg, len, addr_len);
1671 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1672 (sk->sk_state == TCP_ESTABLISHED))
1673 sk_busy_loop(sk, nonblock);
1678 if (sk->sk_state == TCP_LISTEN)
1681 timeo = sock_rcvtimeo(sk, nonblock);
1683 /* Urgent data needs to be handled specially. */
1684 if (flags & MSG_OOB)
1687 if (unlikely(tp->repair)) {
1689 if (!(flags & MSG_PEEK))
1692 if (tp->repair_queue == TCP_SEND_QUEUE)
1696 if (tp->repair_queue == TCP_NO_QUEUE)
1699 /* 'common' recv queue MSG_PEEK-ing */
1702 seq = &tp->copied_seq;
1703 if (flags & MSG_PEEK) {
1704 peek_seq = tp->copied_seq;
1708 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1713 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1714 if (tp->urg_data && tp->urg_seq == *seq) {
1717 if (signal_pending(current)) {
1718 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1723 /* Next get a buffer. */
1725 last = skb_peek_tail(&sk->sk_receive_queue);
1726 skb_queue_walk(&sk->sk_receive_queue, skb) {
1728 /* Now that we have two receive queues this
1731 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1732 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
1733 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1737 offset = *seq - TCP_SKB_CB(skb)->seq;
1738 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1739 pr_err_once("%s: found a SYN, please report !\n", __func__);
1742 if (offset < skb->len)
1744 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1746 WARN(!(flags & MSG_PEEK),
1747 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
1748 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1751 /* Well, if we have backlog, try to process it now yet. */
1753 if (copied >= target && !sk->sk_backlog.tail)
1758 sk->sk_state == TCP_CLOSE ||
1759 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1761 signal_pending(current))
1764 if (sock_flag(sk, SOCK_DONE))
1768 copied = sock_error(sk);
1772 if (sk->sk_shutdown & RCV_SHUTDOWN)
1775 if (sk->sk_state == TCP_CLOSE) {
1776 if (!sock_flag(sk, SOCK_DONE)) {
1777 /* This occurs when user tries to read
1778 * from never connected socket.
1791 if (signal_pending(current)) {
1792 copied = sock_intr_errno(timeo);
1797 tcp_cleanup_rbuf(sk, copied);
1799 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1800 /* Install new reader */
1801 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1802 user_recv = current;
1803 tp->ucopy.task = user_recv;
1804 tp->ucopy.msg = msg;
1807 tp->ucopy.len = len;
1809 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1810 !(flags & (MSG_PEEK | MSG_TRUNC)));
1812 /* Ugly... If prequeue is not empty, we have to
1813 * process it before releasing socket, otherwise
1814 * order will be broken at second iteration.
1815 * More elegant solution is required!!!
1817 * Look: we have the following (pseudo)queues:
1819 * 1. packets in flight
1824 * Each queue can be processed only if the next ones
1825 * are empty. At this point we have empty receive_queue.
1826 * But prequeue _can_ be not empty after 2nd iteration,
1827 * when we jumped to start of loop because backlog
1828 * processing added something to receive_queue.
1829 * We cannot release_sock(), because backlog contains
1830 * packets arrived _after_ prequeued ones.
1832 * Shortly, algorithm is clear --- to process all
1833 * the queues in order. We could make it more directly,
1834 * requeueing packets from backlog to prequeue, if
1835 * is not empty. It is more elegant, but eats cycles,
1838 if (!skb_queue_empty(&tp->ucopy.prequeue))
1841 /* __ Set realtime policy in scheduler __ */
1844 if (copied >= target) {
1845 /* Do not sleep, just process backlog. */
1849 sk_wait_data(sk, &timeo, last);
1855 /* __ Restore normal policy in scheduler __ */
1857 chunk = len - tp->ucopy.len;
1859 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1864 if (tp->rcv_nxt == tp->copied_seq &&
1865 !skb_queue_empty(&tp->ucopy.prequeue)) {
1867 tcp_prequeue_process(sk);
1869 chunk = len - tp->ucopy.len;
1871 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1877 if ((flags & MSG_PEEK) &&
1878 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1879 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1881 task_pid_nr(current));
1882 peek_seq = tp->copied_seq;
1887 /* Ok so how much can we use? */
1888 used = skb->len - offset;
1892 /* Do we have urgent data here? */
1894 u32 urg_offset = tp->urg_seq - *seq;
1895 if (urg_offset < used) {
1897 if (!sock_flag(sk, SOCK_URGINLINE)) {
1910 if (!(flags & MSG_TRUNC)) {
1911 err = skb_copy_datagram_msg(skb, offset, msg, used);
1913 /* Exception. Bailout! */
1924 tcp_rcv_space_adjust(sk);
1927 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1929 tcp_fast_path_check(sk);
1931 if (used + offset < skb->len)
1934 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1936 if (!(flags & MSG_PEEK))
1937 sk_eat_skb(sk, skb);
1941 /* Process the FIN. */
1943 if (!(flags & MSG_PEEK))
1944 sk_eat_skb(sk, skb);
1949 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1952 tp->ucopy.len = copied > 0 ? len : 0;
1954 tcp_prequeue_process(sk);
1956 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1957 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1963 tp->ucopy.task = NULL;
1967 /* According to UNIX98, msg_name/msg_namelen are ignored
1968 * on connected socket. I was just happy when found this 8) --ANK
1971 /* Clean up data we have read: This will do ACK frames. */
1972 tcp_cleanup_rbuf(sk, copied);
1982 err = tcp_recv_urg(sk, msg, len, flags);
1986 err = tcp_peek_sndq(sk, msg, len);
1989 EXPORT_SYMBOL(tcp_recvmsg);
1991 void tcp_set_state(struct sock *sk, int state)
1993 int oldstate = sk->sk_state;
1996 case TCP_ESTABLISHED:
1997 if (oldstate != TCP_ESTABLISHED)
1998 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2002 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2003 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2005 sk->sk_prot->unhash(sk);
2006 if (inet_csk(sk)->icsk_bind_hash &&
2007 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2011 if (oldstate == TCP_ESTABLISHED)
2012 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2015 /* Change state AFTER socket is unhashed to avoid closed
2016 * socket sitting in hash tables.
2018 sk_state_store(sk, state);
2021 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2024 EXPORT_SYMBOL_GPL(tcp_set_state);
2027 * State processing on a close. This implements the state shift for
2028 * sending our FIN frame. Note that we only send a FIN for some
2029 * states. A shutdown() may have already sent the FIN, or we may be
2033 static const unsigned char new_state[16] = {
2034 /* current state: new state: action: */
2035 [0 /* (Invalid) */] = TCP_CLOSE,
2036 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2037 [TCP_SYN_SENT] = TCP_CLOSE,
2038 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2039 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2040 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2041 [TCP_TIME_WAIT] = TCP_CLOSE,
2042 [TCP_CLOSE] = TCP_CLOSE,
2043 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2044 [TCP_LAST_ACK] = TCP_LAST_ACK,
2045 [TCP_LISTEN] = TCP_CLOSE,
2046 [TCP_CLOSING] = TCP_CLOSING,
2047 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2050 static int tcp_close_state(struct sock *sk)
2052 int next = (int)new_state[sk->sk_state];
2053 int ns = next & TCP_STATE_MASK;
2055 tcp_set_state(sk, ns);
2057 return next & TCP_ACTION_FIN;
2061 * Shutdown the sending side of a connection. Much like close except
2062 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2065 void tcp_shutdown(struct sock *sk, int how)
2067 /* We need to grab some memory, and put together a FIN,
2068 * and then put it into the queue to be sent.
2069 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2071 if (!(how & SEND_SHUTDOWN))
2074 /* If we've already sent a FIN, or it's a closed state, skip this. */
2075 if ((1 << sk->sk_state) &
2076 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2077 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2078 /* Clear out any half completed packets. FIN if needed. */
2079 if (tcp_close_state(sk))
2083 EXPORT_SYMBOL(tcp_shutdown);
2085 bool tcp_check_oom(struct sock *sk, int shift)
2087 bool too_many_orphans, out_of_socket_memory;
2089 too_many_orphans = tcp_too_many_orphans(sk, shift);
2090 out_of_socket_memory = tcp_out_of_memory(sk);
2092 if (too_many_orphans)
2093 net_info_ratelimited("too many orphaned sockets\n");
2094 if (out_of_socket_memory)
2095 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2096 return too_many_orphans || out_of_socket_memory;
2099 void tcp_close(struct sock *sk, long timeout)
2101 struct sk_buff *skb;
2102 int data_was_unread = 0;
2106 sk->sk_shutdown = SHUTDOWN_MASK;
2108 if (sk->sk_state == TCP_LISTEN) {
2109 tcp_set_state(sk, TCP_CLOSE);
2112 inet_csk_listen_stop(sk);
2114 goto adjudge_to_death;
2117 /* We need to flush the recv. buffs. We do this only on the
2118 * descriptor close, not protocol-sourced closes, because the
2119 * reader process may not have drained the data yet!
2121 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2122 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2124 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2126 data_was_unread += len;
2132 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2133 if (sk->sk_state == TCP_CLOSE)
2134 goto adjudge_to_death;
2136 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2137 * data was lost. To witness the awful effects of the old behavior of
2138 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2139 * GET in an FTP client, suspend the process, wait for the client to
2140 * advertise a zero window, then kill -9 the FTP client, wheee...
2141 * Note: timeout is always zero in such a case.
2143 if (unlikely(tcp_sk(sk)->repair)) {
2144 sk->sk_prot->disconnect(sk, 0);
2145 } else if (data_was_unread) {
2146 /* Unread data was tossed, zap the connection. */
2147 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2148 tcp_set_state(sk, TCP_CLOSE);
2149 tcp_send_active_reset(sk, sk->sk_allocation);
2150 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2151 /* Check zero linger _after_ checking for unread data. */
2152 sk->sk_prot->disconnect(sk, 0);
2153 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2154 } else if (tcp_close_state(sk)) {
2155 /* We FIN if the application ate all the data before
2156 * zapping the connection.
2159 /* RED-PEN. Formally speaking, we have broken TCP state
2160 * machine. State transitions:
2162 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2163 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2164 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2166 * are legal only when FIN has been sent (i.e. in window),
2167 * rather than queued out of window. Purists blame.
2169 * F.e. "RFC state" is ESTABLISHED,
2170 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2172 * The visible declinations are that sometimes
2173 * we enter time-wait state, when it is not required really
2174 * (harmless), do not send active resets, when they are
2175 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2176 * they look as CLOSING or LAST_ACK for Linux)
2177 * Probably, I missed some more holelets.
2179 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2180 * in a single packet! (May consider it later but will
2181 * probably need API support or TCP_CORK SYN-ACK until
2182 * data is written and socket is closed.)
2187 sk_stream_wait_close(sk, timeout);
2190 state = sk->sk_state;
2194 /* It is the last release_sock in its life. It will remove backlog. */
2198 /* Now socket is owned by kernel and we acquire BH lock
2199 to finish close. No need to check for user refs.
2203 WARN_ON(sock_owned_by_user(sk));
2205 percpu_counter_inc(sk->sk_prot->orphan_count);
2207 /* Have we already been destroyed by a softirq or backlog? */
2208 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2211 /* This is a (useful) BSD violating of the RFC. There is a
2212 * problem with TCP as specified in that the other end could
2213 * keep a socket open forever with no application left this end.
2214 * We use a 1 minute timeout (about the same as BSD) then kill
2215 * our end. If they send after that then tough - BUT: long enough
2216 * that we won't make the old 4*rto = almost no time - whoops
2219 * Nope, it was not mistake. It is really desired behaviour
2220 * f.e. on http servers, when such sockets are useless, but
2221 * consume significant resources. Let's do it with special
2222 * linger2 option. --ANK
2225 if (sk->sk_state == TCP_FIN_WAIT2) {
2226 struct tcp_sock *tp = tcp_sk(sk);
2227 if (tp->linger2 < 0) {
2228 tcp_set_state(sk, TCP_CLOSE);
2229 tcp_send_active_reset(sk, GFP_ATOMIC);
2230 __NET_INC_STATS(sock_net(sk),
2231 LINUX_MIB_TCPABORTONLINGER);
2233 const int tmo = tcp_fin_time(sk);
2235 if (tmo > TCP_TIMEWAIT_LEN) {
2236 inet_csk_reset_keepalive_timer(sk,
2237 tmo - TCP_TIMEWAIT_LEN);
2239 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2244 if (sk->sk_state != TCP_CLOSE) {
2246 if (tcp_check_oom(sk, 0)) {
2247 tcp_set_state(sk, TCP_CLOSE);
2248 tcp_send_active_reset(sk, GFP_ATOMIC);
2249 __NET_INC_STATS(sock_net(sk),
2250 LINUX_MIB_TCPABORTONMEMORY);
2251 } else if (!check_net(sock_net(sk))) {
2252 /* Not possible to send reset; just close */
2253 tcp_set_state(sk, TCP_CLOSE);
2257 if (sk->sk_state == TCP_CLOSE) {
2258 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2259 /* We could get here with a non-NULL req if the socket is
2260 * aborted (e.g., closed with unread data) before 3WHS
2264 reqsk_fastopen_remove(sk, req, false);
2265 inet_csk_destroy_sock(sk);
2267 /* Otherwise, socket is reprieved until protocol close. */
2274 EXPORT_SYMBOL(tcp_close);
2276 /* These states need RST on ABORT according to RFC793 */
2278 static inline bool tcp_need_reset(int state)
2280 return (1 << state) &
2281 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2282 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2285 int tcp_disconnect(struct sock *sk, int flags)
2287 struct inet_sock *inet = inet_sk(sk);
2288 struct inet_connection_sock *icsk = inet_csk(sk);
2289 struct tcp_sock *tp = tcp_sk(sk);
2291 int old_state = sk->sk_state;
2293 if (old_state != TCP_CLOSE)
2294 tcp_set_state(sk, TCP_CLOSE);
2296 /* ABORT function of RFC793 */
2297 if (old_state == TCP_LISTEN) {
2298 inet_csk_listen_stop(sk);
2299 } else if (unlikely(tp->repair)) {
2300 sk->sk_err = ECONNABORTED;
2301 } else if (tcp_need_reset(old_state) ||
2302 (tp->snd_nxt != tp->write_seq &&
2303 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2304 /* The last check adjusts for discrepancy of Linux wrt. RFC
2307 tcp_send_active_reset(sk, gfp_any());
2308 sk->sk_err = ECONNRESET;
2309 } else if (old_state == TCP_SYN_SENT)
2310 sk->sk_err = ECONNRESET;
2312 tcp_clear_xmit_timers(sk);
2313 __skb_queue_purge(&sk->sk_receive_queue);
2314 tcp_write_queue_purge(sk);
2315 skb_rbtree_purge(&tp->out_of_order_queue);
2317 inet->inet_dport = 0;
2319 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2320 inet_reset_saddr(sk);
2322 sk->sk_shutdown = 0;
2323 sock_reset_flag(sk, SOCK_DONE);
2325 tp->write_seq += tp->max_window + 2;
2326 if (tp->write_seq == 0)
2329 icsk->icsk_probes_out = 0;
2330 tp->packets_out = 0;
2331 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2332 tp->snd_cwnd_cnt = 0;
2333 tp->window_clamp = 0;
2335 tcp_set_ca_state(sk, TCP_CA_Open);
2336 tp->is_sack_reneg = 0;
2337 tcp_clear_retrans(tp);
2338 tp->total_retrans = 0;
2339 inet_csk_delack_init(sk);
2340 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2341 * issue in __tcp_select_window()
2343 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2344 tcp_init_send_head(sk);
2345 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2347 dst_release(sk->sk_rx_dst);
2348 sk->sk_rx_dst = NULL;
2349 tcp_saved_syn_free(tp);
2352 tp->bytes_acked = 0;
2353 tp->bytes_received = 0;
2354 tp->data_segs_in = 0;
2355 tp->data_segs_out = 0;
2357 /* Clean up fastopen related fields */
2358 tcp_free_fastopen_req(tp);
2359 inet->defer_connect = 0;
2361 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2363 if (sk->sk_frag.page) {
2364 put_page(sk->sk_frag.page);
2365 sk->sk_frag.page = NULL;
2366 sk->sk_frag.offset = 0;
2369 sk->sk_error_report(sk);
2372 EXPORT_SYMBOL(tcp_disconnect);
2374 static inline bool tcp_can_repair_sock(const struct sock *sk)
2376 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2377 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2380 static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2382 struct tcp_repair_window opt;
2387 if (len != sizeof(opt))
2390 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2393 if (opt.max_window < opt.snd_wnd)
2396 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2399 if (after(opt.rcv_wup, tp->rcv_nxt))
2402 tp->snd_wl1 = opt.snd_wl1;
2403 tp->snd_wnd = opt.snd_wnd;
2404 tp->max_window = opt.max_window;
2406 tp->rcv_wnd = opt.rcv_wnd;
2407 tp->rcv_wup = opt.rcv_wup;
2412 static int tcp_repair_options_est(struct tcp_sock *tp,
2413 struct tcp_repair_opt __user *optbuf, unsigned int len)
2415 struct tcp_repair_opt opt;
2417 while (len >= sizeof(opt)) {
2418 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2424 switch (opt.opt_code) {
2426 tp->rx_opt.mss_clamp = opt.opt_val;
2430 u16 snd_wscale = opt.opt_val & 0xFFFF;
2431 u16 rcv_wscale = opt.opt_val >> 16;
2433 if (snd_wscale > 14 || rcv_wscale > 14)
2436 tp->rx_opt.snd_wscale = snd_wscale;
2437 tp->rx_opt.rcv_wscale = rcv_wscale;
2438 tp->rx_opt.wscale_ok = 1;
2441 case TCPOPT_SACK_PERM:
2442 if (opt.opt_val != 0)
2445 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2446 if (sysctl_tcp_fack)
2447 tcp_enable_fack(tp);
2449 case TCPOPT_TIMESTAMP:
2450 if (opt.opt_val != 0)
2453 tp->rx_opt.tstamp_ok = 1;
2462 * Socket option code for TCP.
2464 static int do_tcp_setsockopt(struct sock *sk, int level,
2465 int optname, char __user *optval, unsigned int optlen)
2467 struct tcp_sock *tp = tcp_sk(sk);
2468 struct inet_connection_sock *icsk = inet_csk(sk);
2469 struct net *net = sock_net(sk);
2473 /* These are data/string values, all the others are ints */
2475 case TCP_CONGESTION: {
2476 char name[TCP_CA_NAME_MAX];
2481 val = strncpy_from_user(name, optval,
2482 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2488 err = tcp_set_congestion_control(sk, name);
2497 if (optlen < sizeof(int))
2500 if (get_user(val, (int __user *)optval))
2507 /* Values greater than interface MTU won't take effect. However
2508 * at the point when this call is done we typically don't yet
2509 * know which interface is going to be used */
2510 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2514 tp->rx_opt.user_mss = val;
2519 /* TCP_NODELAY is weaker than TCP_CORK, so that
2520 * this option on corked socket is remembered, but
2521 * it is not activated until cork is cleared.
2523 * However, when TCP_NODELAY is set we make
2524 * an explicit push, which overrides even TCP_CORK
2525 * for currently queued segments.
2527 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2528 tcp_push_pending_frames(sk);
2530 tp->nonagle &= ~TCP_NAGLE_OFF;
2534 case TCP_THIN_LINEAR_TIMEOUTS:
2535 if (val < 0 || val > 1)
2541 case TCP_THIN_DUPACK:
2542 if (val < 0 || val > 1)
2545 tp->thin_dupack = val;
2546 if (tp->thin_dupack)
2547 tcp_disable_early_retrans(tp);
2552 if (!tcp_can_repair_sock(sk))
2554 else if (val == 1) {
2556 sk->sk_reuse = SK_FORCE_REUSE;
2557 tp->repair_queue = TCP_NO_QUEUE;
2558 } else if (val == 0) {
2560 sk->sk_reuse = SK_NO_REUSE;
2561 tcp_send_window_probe(sk);
2567 case TCP_REPAIR_QUEUE:
2570 else if ((unsigned int)val < TCP_QUEUES_NR)
2571 tp->repair_queue = val;
2577 if (sk->sk_state != TCP_CLOSE)
2579 else if (tp->repair_queue == TCP_SEND_QUEUE)
2580 tp->write_seq = val;
2581 else if (tp->repair_queue == TCP_RECV_QUEUE)
2587 case TCP_REPAIR_OPTIONS:
2590 else if (sk->sk_state == TCP_ESTABLISHED)
2591 err = tcp_repair_options_est(tp,
2592 (struct tcp_repair_opt __user *)optval,
2599 /* When set indicates to always queue non-full frames.
2600 * Later the user clears this option and we transmit
2601 * any pending partial frames in the queue. This is
2602 * meant to be used alongside sendfile() to get properly
2603 * filled frames when the user (for example) must write
2604 * out headers with a write() call first and then use
2605 * sendfile to send out the data parts.
2607 * TCP_CORK can be set together with TCP_NODELAY and it is
2608 * stronger than TCP_NODELAY.
2611 tp->nonagle |= TCP_NAGLE_CORK;
2613 tp->nonagle &= ~TCP_NAGLE_CORK;
2614 if (tp->nonagle&TCP_NAGLE_OFF)
2615 tp->nonagle |= TCP_NAGLE_PUSH;
2616 tcp_push_pending_frames(sk);
2621 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2624 tp->keepalive_time = val * HZ;
2625 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2626 !((1 << sk->sk_state) &
2627 (TCPF_CLOSE | TCPF_LISTEN))) {
2628 u32 elapsed = keepalive_time_elapsed(tp);
2629 if (tp->keepalive_time > elapsed)
2630 elapsed = tp->keepalive_time - elapsed;
2633 inet_csk_reset_keepalive_timer(sk, elapsed);
2638 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2641 tp->keepalive_intvl = val * HZ;
2644 if (val < 1 || val > MAX_TCP_KEEPCNT)
2647 tp->keepalive_probes = val;
2650 if (val < 1 || val > MAX_TCP_SYNCNT)
2653 icsk->icsk_syn_retries = val;
2657 if (val < 0 || val > 1)
2666 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2669 tp->linger2 = val * HZ;
2672 case TCP_DEFER_ACCEPT:
2673 /* Translate value in seconds to number of retransmits */
2674 icsk->icsk_accept_queue.rskq_defer_accept =
2675 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2679 case TCP_WINDOW_CLAMP:
2681 if (sk->sk_state != TCP_CLOSE) {
2685 tp->window_clamp = 0;
2687 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2688 SOCK_MIN_RCVBUF / 2 : val;
2693 icsk->icsk_ack.pingpong = 1;
2695 icsk->icsk_ack.pingpong = 0;
2696 if ((1 << sk->sk_state) &
2697 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2698 inet_csk_ack_scheduled(sk)) {
2699 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2700 tcp_cleanup_rbuf(sk, 1);
2702 icsk->icsk_ack.pingpong = 1;
2707 #ifdef CONFIG_TCP_MD5SIG
2709 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
2710 err = tp->af_specific->md5_parse(sk, optval, optlen);
2715 case TCP_USER_TIMEOUT:
2716 /* Cap the max time in ms TCP will retry or probe the window
2717 * before giving up and aborting (ETIMEDOUT) a connection.
2722 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2726 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2728 tcp_fastopen_init_key_once(true);
2730 fastopen_queue_tune(sk, val);
2735 case TCP_FASTOPEN_CONNECT:
2736 if (val > 1 || val < 0) {
2738 } else if (sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
2739 if (sk->sk_state == TCP_CLOSE)
2740 tp->fastopen_connect = val;
2751 tp->tsoffset = val - tcp_time_stamp;
2753 case TCP_REPAIR_WINDOW:
2754 err = tcp_repair_set_window(tp, optval, optlen);
2756 case TCP_NOTSENT_LOWAT:
2757 tp->notsent_lowat = val;
2758 sk->sk_write_space(sk);
2769 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2770 unsigned int optlen)
2772 const struct inet_connection_sock *icsk = inet_csk(sk);
2774 if (level != SOL_TCP)
2775 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2777 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2779 EXPORT_SYMBOL(tcp_setsockopt);
2781 #ifdef CONFIG_COMPAT
2782 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2783 char __user *optval, unsigned int optlen)
2785 if (level != SOL_TCP)
2786 return inet_csk_compat_setsockopt(sk, level, optname,
2788 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2790 EXPORT_SYMBOL(compat_tcp_setsockopt);
2793 /* Return information about state of tcp endpoint in API format. */
2794 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2796 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2797 const struct inet_connection_sock *icsk = inet_csk(sk);
2798 u32 now = tcp_time_stamp, intv;
2804 memset(info, 0, sizeof(*info));
2805 if (sk->sk_type != SOCK_STREAM)
2808 info->tcpi_state = sk_state_load(sk);
2810 info->tcpi_ca_state = icsk->icsk_ca_state;
2811 info->tcpi_retransmits = icsk->icsk_retransmits;
2812 info->tcpi_probes = icsk->icsk_probes_out;
2813 info->tcpi_backoff = icsk->icsk_backoff;
2815 if (tp->rx_opt.tstamp_ok)
2816 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2817 if (tcp_is_sack(tp))
2818 info->tcpi_options |= TCPI_OPT_SACK;
2819 if (tp->rx_opt.wscale_ok) {
2820 info->tcpi_options |= TCPI_OPT_WSCALE;
2821 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2822 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2825 if (tp->ecn_flags & TCP_ECN_OK)
2826 info->tcpi_options |= TCPI_OPT_ECN;
2827 if (tp->ecn_flags & TCP_ECN_SEEN)
2828 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2829 if (tp->syn_data_acked)
2830 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2832 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2833 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2834 info->tcpi_snd_mss = tp->mss_cache;
2835 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2837 if (info->tcpi_state == TCP_LISTEN) {
2838 info->tcpi_unacked = sk->sk_ack_backlog;
2839 info->tcpi_sacked = sk->sk_max_ack_backlog;
2841 info->tcpi_unacked = tp->packets_out;
2842 info->tcpi_sacked = tp->sacked_out;
2844 info->tcpi_lost = tp->lost_out;
2845 info->tcpi_retrans = tp->retrans_out;
2846 info->tcpi_fackets = tp->fackets_out;
2848 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2849 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2850 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2852 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2853 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2854 info->tcpi_rtt = tp->srtt_us >> 3;
2855 info->tcpi_rttvar = tp->mdev_us >> 2;
2856 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2857 info->tcpi_snd_cwnd = tp->snd_cwnd;
2858 info->tcpi_advmss = tp->advmss;
2859 info->tcpi_reordering = tp->reordering;
2861 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2862 info->tcpi_rcv_space = tp->rcvq_space.space;
2864 info->tcpi_total_retrans = tp->total_retrans;
2866 rate = READ_ONCE(sk->sk_pacing_rate);
2867 rate64 = rate != ~0U ? rate : ~0ULL;
2868 put_unaligned(rate64, &info->tcpi_pacing_rate);
2870 rate = READ_ONCE(sk->sk_max_pacing_rate);
2871 rate64 = rate != ~0U ? rate : ~0ULL;
2872 put_unaligned(rate64, &info->tcpi_max_pacing_rate);
2875 start = u64_stats_fetch_begin_irq(&tp->syncp);
2876 put_unaligned(tp->bytes_acked, &info->tcpi_bytes_acked);
2877 put_unaligned(tp->bytes_received, &info->tcpi_bytes_received);
2878 } while (u64_stats_fetch_retry_irq(&tp->syncp, start));
2879 info->tcpi_segs_out = tp->segs_out;
2880 info->tcpi_segs_in = tp->segs_in;
2882 notsent_bytes = READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_nxt);
2883 info->tcpi_notsent_bytes = max(0, notsent_bytes);
2885 info->tcpi_min_rtt = tcp_min_rtt(tp);
2886 info->tcpi_data_segs_in = tp->data_segs_in;
2887 info->tcpi_data_segs_out = tp->data_segs_out;
2889 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
2890 rate = READ_ONCE(tp->rate_delivered);
2891 intv = READ_ONCE(tp->rate_interval_us);
2893 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
2894 do_div(rate64, intv);
2895 put_unaligned(rate64, &info->tcpi_delivery_rate);
2898 EXPORT_SYMBOL_GPL(tcp_get_info);
2900 static int do_tcp_getsockopt(struct sock *sk, int level,
2901 int optname, char __user *optval, int __user *optlen)
2903 struct inet_connection_sock *icsk = inet_csk(sk);
2904 struct tcp_sock *tp = tcp_sk(sk);
2905 struct net *net = sock_net(sk);
2908 if (get_user(len, optlen))
2911 len = min_t(unsigned int, len, sizeof(int));
2918 val = tp->mss_cache;
2919 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2920 val = tp->rx_opt.user_mss;
2922 val = tp->rx_opt.mss_clamp;
2925 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2928 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2931 val = keepalive_time_when(tp) / HZ;
2934 val = keepalive_intvl_when(tp) / HZ;
2937 val = keepalive_probes(tp);
2940 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
2945 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
2947 case TCP_DEFER_ACCEPT:
2948 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2949 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2951 case TCP_WINDOW_CLAMP:
2952 val = tp->window_clamp;
2955 struct tcp_info info;
2957 if (get_user(len, optlen))
2960 tcp_get_info(sk, &info);
2962 len = min_t(unsigned int, len, sizeof(info));
2963 if (put_user(len, optlen))
2965 if (copy_to_user(optval, &info, len))
2970 const struct tcp_congestion_ops *ca_ops;
2971 union tcp_cc_info info;
2975 if (get_user(len, optlen))
2978 ca_ops = icsk->icsk_ca_ops;
2979 if (ca_ops && ca_ops->get_info)
2980 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2982 len = min_t(unsigned int, len, sz);
2983 if (put_user(len, optlen))
2985 if (copy_to_user(optval, &info, len))
2990 val = !icsk->icsk_ack.pingpong;
2993 case TCP_CONGESTION:
2994 if (get_user(len, optlen))
2996 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2997 if (put_user(len, optlen))
2999 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3003 case TCP_THIN_LINEAR_TIMEOUTS:
3006 case TCP_THIN_DUPACK:
3007 val = tp->thin_dupack;
3014 case TCP_REPAIR_QUEUE:
3016 val = tp->repair_queue;
3021 case TCP_REPAIR_WINDOW: {
3022 struct tcp_repair_window opt;
3024 if (get_user(len, optlen))
3027 if (len != sizeof(opt))
3033 opt.snd_wl1 = tp->snd_wl1;
3034 opt.snd_wnd = tp->snd_wnd;
3035 opt.max_window = tp->max_window;
3036 opt.rcv_wnd = tp->rcv_wnd;
3037 opt.rcv_wup = tp->rcv_wup;
3039 if (copy_to_user(optval, &opt, len))
3044 if (tp->repair_queue == TCP_SEND_QUEUE)
3045 val = tp->write_seq;
3046 else if (tp->repair_queue == TCP_RECV_QUEUE)
3052 case TCP_USER_TIMEOUT:
3053 val = jiffies_to_msecs(icsk->icsk_user_timeout);
3057 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3060 case TCP_FASTOPEN_CONNECT:
3061 val = tp->fastopen_connect;
3065 val = tcp_time_stamp + tp->tsoffset;
3067 case TCP_NOTSENT_LOWAT:
3068 val = tp->notsent_lowat;
3073 case TCP_SAVED_SYN: {
3074 if (get_user(len, optlen))
3078 if (tp->saved_syn) {
3079 if (len < tp->saved_syn[0]) {
3080 if (put_user(tp->saved_syn[0], optlen)) {
3087 len = tp->saved_syn[0];
3088 if (put_user(len, optlen)) {
3092 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3096 tcp_saved_syn_free(tp);
3101 if (put_user(len, optlen))
3107 return -ENOPROTOOPT;
3110 if (put_user(len, optlen))
3112 if (copy_to_user(optval, &val, len))
3117 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3120 struct inet_connection_sock *icsk = inet_csk(sk);
3122 if (level != SOL_TCP)
3123 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3125 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3127 EXPORT_SYMBOL(tcp_getsockopt);
3129 #ifdef CONFIG_COMPAT
3130 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3131 char __user *optval, int __user *optlen)
3133 if (level != SOL_TCP)
3134 return inet_csk_compat_getsockopt(sk, level, optname,
3136 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3138 EXPORT_SYMBOL(compat_tcp_getsockopt);
3141 #ifdef CONFIG_TCP_MD5SIG
3142 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3143 static DEFINE_MUTEX(tcp_md5sig_mutex);
3144 static bool tcp_md5sig_pool_populated = false;
3146 static void __tcp_alloc_md5sig_pool(void)
3148 struct crypto_ahash *hash;
3151 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3155 for_each_possible_cpu(cpu) {
3156 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3157 struct ahash_request *req;
3160 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3161 sizeof(struct tcphdr),
3166 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3168 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3171 req = ahash_request_alloc(hash, GFP_KERNEL);
3175 ahash_request_set_callback(req, 0, NULL, NULL);
3177 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3179 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3180 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3183 tcp_md5sig_pool_populated = true;
3186 bool tcp_alloc_md5sig_pool(void)
3188 if (unlikely(!tcp_md5sig_pool_populated)) {
3189 mutex_lock(&tcp_md5sig_mutex);
3191 if (!tcp_md5sig_pool_populated)
3192 __tcp_alloc_md5sig_pool();
3194 mutex_unlock(&tcp_md5sig_mutex);
3196 return tcp_md5sig_pool_populated;
3198 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3202 * tcp_get_md5sig_pool - get md5sig_pool for this user
3204 * We use percpu structure, so if we succeed, we exit with preemption
3205 * and BH disabled, to make sure another thread or softirq handling
3206 * wont try to get same context.
3208 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3212 if (tcp_md5sig_pool_populated) {
3213 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3215 return this_cpu_ptr(&tcp_md5sig_pool);
3220 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3222 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3223 const struct sk_buff *skb, unsigned int header_len)
3225 struct scatterlist sg;
3226 const struct tcphdr *tp = tcp_hdr(skb);
3227 struct ahash_request *req = hp->md5_req;
3229 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3230 skb_headlen(skb) - header_len : 0;
3231 const struct skb_shared_info *shi = skb_shinfo(skb);
3232 struct sk_buff *frag_iter;
3234 sg_init_table(&sg, 1);
3236 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3237 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3238 if (crypto_ahash_update(req))
3241 for (i = 0; i < shi->nr_frags; ++i) {
3242 const struct skb_frag_struct *f = &shi->frags[i];
3243 unsigned int offset = f->page_offset;
3244 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3246 sg_set_page(&sg, page, skb_frag_size(f),
3247 offset_in_page(offset));
3248 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3249 if (crypto_ahash_update(req))
3253 skb_walk_frags(skb, frag_iter)
3254 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3259 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3261 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3263 struct scatterlist sg;
3265 sg_init_one(&sg, key->key, key->keylen);
3266 ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3267 return crypto_ahash_update(hp->md5_req);
3269 EXPORT_SYMBOL(tcp_md5_hash_key);
3273 void tcp_done(struct sock *sk)
3275 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3277 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3278 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3280 tcp_set_state(sk, TCP_CLOSE);
3281 tcp_clear_xmit_timers(sk);
3283 reqsk_fastopen_remove(sk, req, false);
3285 sk->sk_shutdown = SHUTDOWN_MASK;
3287 if (!sock_flag(sk, SOCK_DEAD))
3288 sk->sk_state_change(sk);
3290 inet_csk_destroy_sock(sk);
3292 EXPORT_SYMBOL_GPL(tcp_done);
3294 int tcp_abort(struct sock *sk, int err)
3296 if (!sk_fullsock(sk)) {
3297 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3298 struct request_sock *req = inet_reqsk(sk);
3301 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3308 /* Don't race with userspace socket closes such as tcp_close. */
3311 if (sk->sk_state == TCP_LISTEN) {
3312 tcp_set_state(sk, TCP_CLOSE);
3313 inet_csk_listen_stop(sk);
3316 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3320 if (!sock_flag(sk, SOCK_DEAD)) {
3322 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3324 sk->sk_error_report(sk);
3325 if (tcp_need_reset(sk->sk_state))
3326 tcp_send_active_reset(sk, GFP_ATOMIC);
3335 EXPORT_SYMBOL_GPL(tcp_abort);
3337 extern struct tcp_congestion_ops tcp_reno;
3339 static __initdata unsigned long thash_entries;
3340 static int __init set_thash_entries(char *str)
3347 ret = kstrtoul(str, 0, &thash_entries);
3353 __setup("thash_entries=", set_thash_entries);
3355 static void __init tcp_init_mem(void)
3357 unsigned long limit = nr_free_buffer_pages() / 16;
3359 limit = max(limit, 128UL);
3360 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3361 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3362 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3365 void __init tcp_init(void)
3367 int max_rshare, max_wshare, cnt;
3368 unsigned long limit;
3371 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
3372 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3373 FIELD_SIZEOF(struct sk_buff, cb));
3375 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3376 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3377 tcp_hashinfo.bind_bucket_cachep =
3378 kmem_cache_create("tcp_bind_bucket",
3379 sizeof(struct inet_bind_bucket), 0,
3380 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3382 /* Size and allocate the main established and bind bucket
3385 * The methodology is similar to that of the buffer cache.
3387 tcp_hashinfo.ehash =
3388 alloc_large_system_hash("TCP established",
3389 sizeof(struct inet_ehash_bucket),
3391 17, /* one slot per 128 KB of memory */
3394 &tcp_hashinfo.ehash_mask,
3396 thash_entries ? 0 : 512 * 1024);
3397 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3398 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3400 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3401 panic("TCP: failed to alloc ehash_locks");
3402 tcp_hashinfo.bhash =
3403 alloc_large_system_hash("TCP bind",
3404 sizeof(struct inet_bind_hashbucket),
3405 tcp_hashinfo.ehash_mask + 1,
3406 17, /* one slot per 128 KB of memory */
3408 &tcp_hashinfo.bhash_size,
3412 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3413 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3414 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3415 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3419 cnt = tcp_hashinfo.ehash_mask + 1;
3421 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3422 sysctl_tcp_max_orphans = cnt / 2;
3423 sysctl_max_syn_backlog = max(128, cnt / 256);
3426 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3427 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3428 max_wshare = min(4UL*1024*1024, limit);
3429 max_rshare = min(6UL*1024*1024, limit);
3431 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3432 sysctl_tcp_wmem[1] = 16*1024;
3433 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3435 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3436 sysctl_tcp_rmem[1] = 87380;
3437 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3439 pr_info("Hash tables configured (established %u bind %u)\n",
3440 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3443 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);