2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
21 * Alan Cox : Numerous verify_area() calls
22 * Alan Cox : Set the ACK bit on a reset
23 * Alan Cox : Stopped it crashing if it closed while
24 * sk->inuse=1 and was trying to connect
26 * Alan Cox : All icmp error handling was broken
27 * pointers passed where wrong and the
28 * socket was looked up backwards. Nobody
29 * tested any icmp error code obviously.
30 * Alan Cox : tcp_err() now handled properly. It
31 * wakes people on errors. poll
32 * behaves and the icmp error race
33 * has gone by moving it into sock.c
34 * Alan Cox : tcp_send_reset() fixed to work for
35 * everything not just packets for
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
40 * Herp Rosmanith : More reset fixes
41 * Alan Cox : No longer acks invalid rst frames.
42 * Acking any kind of RST is right out.
43 * Alan Cox : Sets an ignore me flag on an rst
44 * receive otherwise odd bits of prattle
46 * Alan Cox : Fixed another acking RST frame bug.
47 * Should stop LAN workplace lockups.
48 * Alan Cox : Some tidyups using the new skb list
50 * Alan Cox : sk->keepopen now seems to work
51 * Alan Cox : Pulls options out correctly on accepts
52 * Alan Cox : Fixed assorted sk->rqueue->next errors
53 * Alan Cox : PSH doesn't end a TCP read. Switched a
55 * Alan Cox : Tidied tcp_data to avoid a potential
57 * Alan Cox : Added some better commenting, as the
58 * tcp is hard to follow
59 * Alan Cox : Removed incorrect check for 20 * psh
60 * Michael O'Reilly : ack < copied bug fix.
61 * Johannes Stille : Misc tcp fixes (not all in yet).
62 * Alan Cox : FIN with no memory -> CRASH
63 * Alan Cox : Added socket option proto entries.
64 * Also added awareness of them to accept.
65 * Alan Cox : Added TCP options (SOL_TCP)
66 * Alan Cox : Switched wakeup calls to callbacks,
67 * so the kernel can layer network
69 * Alan Cox : Use ip_tos/ip_ttl settings.
70 * Alan Cox : Handle FIN (more) properly (we hope).
71 * Alan Cox : RST frames sent on unsynchronised
73 * Alan Cox : Put in missing check for SYN bit.
74 * Alan Cox : Added tcp_select_window() aka NET2E
75 * window non shrink trick.
76 * Alan Cox : Added a couple of small NET2E timer
78 * Charles Hedrick : TCP fixes
79 * Toomas Tamm : TCP window fixes
80 * Alan Cox : Small URG fix to rlogin ^C ack fight
81 * Charles Hedrick : Rewrote most of it to actually work
82 * Linus : Rewrote tcp_read() and URG handling
84 * Gerhard Koerting: Fixed some missing timer handling
85 * Matthew Dillon : Reworked TCP machine states as per RFC
86 * Gerhard Koerting: PC/TCP workarounds
87 * Adam Caldwell : Assorted timer/timing errors
88 * Matthew Dillon : Fixed another RST bug
89 * Alan Cox : Move to kernel side addressing changes.
90 * Alan Cox : Beginning work on TCP fastpathing
92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
93 * Alan Cox : TCP fast path debugging
94 * Alan Cox : Window clamping
95 * Michael Riepe : Bug in tcp_check()
96 * Matt Dillon : More TCP improvements and RST bug fixes
97 * Matt Dillon : Yet more small nasties remove from the
98 * TCP code (Be very nice to this man if
99 * tcp finally works 100%) 8)
100 * Alan Cox : BSD accept semantics.
101 * Alan Cox : Reset on closedown bug.
102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
103 * Michael Pall : Handle poll() after URG properly in
105 * Michael Pall : Undo the last fix in tcp_read_urg()
106 * (multi URG PUSH broke rlogin).
107 * Michael Pall : Fix the multi URG PUSH problem in
108 * tcp_readable(), poll() after URG
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
112 * Alan Cox : Changed the semantics of sk->socket to
113 * fix a race and a signal problem with
114 * accept() and async I/O.
115 * Alan Cox : Relaxed the rules on tcp_sendto().
116 * Yury Shevchuk : Really fixed accept() blocking problem.
117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
118 * clients/servers which listen in on
120 * Alan Cox : Cleaned the above up and shrank it to
121 * a sensible code size.
122 * Alan Cox : Self connect lockup fix.
123 * Alan Cox : No connect to multicast.
124 * Ross Biro : Close unaccepted children on master
126 * Alan Cox : Reset tracing code.
127 * Alan Cox : Spurious resets on shutdown.
128 * Alan Cox : Giant 15 minute/60 second timer error
129 * Alan Cox : Small whoops in polling before an
131 * Alan Cox : Kept the state trace facility since
132 * it's handy for debugging.
133 * Alan Cox : More reset handler fixes.
134 * Alan Cox : Started rewriting the code based on
135 * the RFC's for other useful protocol
136 * references see: Comer, KA9Q NOS, and
137 * for a reference on the difference
138 * between specifications and how BSD
139 * works see the 4.4lite source.
140 * A.N.Kuznetsov : Don't time wait on completion of tidy
142 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
143 * Linus Torvalds : Fixed BSD port reuse to work first syn
144 * Alan Cox : Reimplemented timers as per the RFC
145 * and using multiple timers for sanity.
146 * Alan Cox : Small bug fixes, and a lot of new
148 * Alan Cox : Fixed dual reader crash by locking
149 * the buffers (much like datagram.c)
150 * Alan Cox : Fixed stuck sockets in probe. A probe
151 * now gets fed up of retrying without
152 * (even a no space) answer.
153 * Alan Cox : Extracted closing code better
154 * Alan Cox : Fixed the closing state machine to
156 * Alan Cox : More 'per spec' fixes.
157 * Jorge Cwik : Even faster checksumming.
158 * Alan Cox : tcp_data() doesn't ack illegal PSH
159 * only frames. At least one pc tcp stack
161 * Alan Cox : Cache last socket.
162 * Alan Cox : Per route irtt.
163 * Matt Day : poll()->select() match BSD precisely on error
164 * Alan Cox : New buffers
165 * Marc Tamsky : Various sk->prot->retransmits and
166 * sk->retransmits misupdating fixed.
167 * Fixed tcp_write_timeout: stuck close,
168 * and TCP syn retries gets used now.
169 * Mark Yarvis : In tcp_read_wakeup(), don't send an
170 * ack if state is TCP_CLOSED.
171 * Alan Cox : Look up device on a retransmit - routes may
172 * change. Doesn't yet cope with MSS shrink right
174 * Marc Tamsky : Closing in closing fixes.
175 * Mike Shaver : RFC1122 verifications.
176 * Alan Cox : rcv_saddr errors.
177 * Alan Cox : Block double connect().
178 * Alan Cox : Small hooks for enSKIP.
179 * Alexey Kuznetsov: Path MTU discovery.
180 * Alan Cox : Support soft errors.
181 * Alan Cox : Fix MTU discovery pathological case
182 * when the remote claims no mtu!
183 * Marc Tamsky : TCP_CLOSE fix.
184 * Colin (G3TNE) : Send a reset on syn ack replies in
185 * window but wrong (fixes NT lpd problems)
186 * Pedro Roque : Better TCP window handling, delayed ack.
187 * Joerg Reuter : No modification of locked buffers in
188 * tcp_do_retransmit()
189 * Eric Schenk : Changed receiver side silly window
190 * avoidance algorithm to BSD style
191 * algorithm. This doubles throughput
192 * against machines running Solaris,
193 * and seems to result in general
195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
196 * Willy Konynenberg : Transparent proxying support.
197 * Mike McLagan : Routing by source
198 * Keith Owens : Do proper merging with partial SKB's in
199 * tcp_do_sendmsg to avoid burstiness.
200 * Eric Schenk : Fix fast close down bug with
201 * shutdown() followed by close().
202 * Andi Kleen : Make poll agree with SIGIO
203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
204 * lingertime == 0 (RFC 793 ABORT Call)
205 * Hirokazu Takahashi : Use copy_from_user() instead of
206 * csum_and_copy_from_user() if possible.
208 * This program is free software; you can redistribute it and/or
209 * modify it under the terms of the GNU General Public License
210 * as published by the Free Software Foundation; either version
211 * 2 of the License, or(at your option) any later version.
213 * Description of States:
215 * TCP_SYN_SENT sent a connection request, waiting for ack
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
220 * TCP_ESTABLISHED connection established
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
231 * TCP_TIME_WAIT timeout to catch resent junk before entering
232 * closed, can only be entered from FIN_WAIT2
233 * or CLOSING. Required because the other end
234 * may not have gotten our last ACK causing it
235 * to retransmit the data packet (which we ignore)
237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
238 * us to finish writing our data and to shutdown
239 * (we have to close() to move on to LAST_ACK)
241 * TCP_LAST_ACK out side has shutdown after remote has
242 * shutdown. There may still be data in our
243 * buffer that we have to finish sending
245 * TCP_CLOSE socket is finished
248 #define pr_fmt(fmt) "TCP: " fmt
250 #include <linux/kernel.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/inet_diag.h>
256 #include <linux/init.h>
257 #include <linux/fs.h>
258 #include <linux/skbuff.h>
259 #include <linux/scatterlist.h>
260 #include <linux/splice.h>
261 #include <linux/net.h>
262 #include <linux/socket.h>
263 #include <linux/random.h>
264 #include <linux/bootmem.h>
265 #include <linux/highmem.h>
266 #include <linux/swap.h>
267 #include <linux/cache.h>
268 #include <linux/err.h>
269 #include <linux/crypto.h>
270 #include <linux/time.h>
271 #include <linux/slab.h>
273 #include <net/icmp.h>
274 #include <net/inet_common.h>
276 #include <net/xfrm.h>
278 #include <net/sock.h>
280 #include <asm/uaccess.h>
281 #include <asm/ioctls.h>
282 #include <net/busy_poll.h>
284 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
286 int sysctl_tcp_min_tso_segs __read_mostly = 2;
288 int sysctl_tcp_autocorking __read_mostly = 1;
290 struct percpu_counter tcp_orphan_count;
291 EXPORT_SYMBOL_GPL(tcp_orphan_count);
293 long sysctl_tcp_mem[3] __read_mostly;
294 int sysctl_tcp_wmem[3] __read_mostly;
295 int sysctl_tcp_rmem[3] __read_mostly;
297 EXPORT_SYMBOL(sysctl_tcp_mem);
298 EXPORT_SYMBOL(sysctl_tcp_rmem);
299 EXPORT_SYMBOL(sysctl_tcp_wmem);
301 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
302 EXPORT_SYMBOL(tcp_memory_allocated);
305 * Current number of TCP sockets.
307 struct percpu_counter tcp_sockets_allocated;
308 EXPORT_SYMBOL(tcp_sockets_allocated);
313 struct tcp_splice_state {
314 struct pipe_inode_info *pipe;
320 * Pressure flag: try to collapse.
321 * Technical note: it is used by multiple contexts non atomically.
322 * All the __sk_mem_schedule() is of this nature: accounting
323 * is strict, actions are advisory and have some latency.
325 int tcp_memory_pressure __read_mostly;
326 EXPORT_SYMBOL(tcp_memory_pressure);
328 void tcp_enter_memory_pressure(struct sock *sk)
330 if (!tcp_memory_pressure) {
331 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
332 tcp_memory_pressure = 1;
335 EXPORT_SYMBOL(tcp_enter_memory_pressure);
337 /* Convert seconds to retransmits based on initial and max timeout */
338 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
343 int period = timeout;
346 while (seconds > period && res < 255) {
349 if (timeout > rto_max)
357 /* Convert retransmits to seconds based on initial and max timeout */
358 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
366 if (timeout > rto_max)
374 /* Address-family independent initialization for a tcp_sock.
376 * NOTE: A lot of things set to zero explicitly by call to
377 * sk_alloc() so need not be done here.
379 void tcp_init_sock(struct sock *sk)
381 struct inet_connection_sock *icsk = inet_csk(sk);
382 struct tcp_sock *tp = tcp_sk(sk);
384 __skb_queue_head_init(&tp->out_of_order_queue);
385 tcp_init_xmit_timers(sk);
386 tcp_prequeue_init(tp);
387 INIT_LIST_HEAD(&tp->tsq_node);
389 icsk->icsk_rto = TCP_TIMEOUT_INIT;
390 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
392 /* So many TCP implementations out there (incorrectly) count the
393 * initial SYN frame in their delayed-ACK and congestion control
394 * algorithms that we must have the following bandaid to talk
395 * efficiently to them. -DaveM
397 tp->snd_cwnd = TCP_INIT_CWND;
399 /* See draft-stevens-tcpca-spec-01 for discussion of the
400 * initialization of these values.
402 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
403 tp->snd_cwnd_clamp = ~0;
404 tp->mss_cache = TCP_MSS_DEFAULT;
406 tp->reordering = sysctl_tcp_reordering;
407 tcp_enable_early_retrans(tp);
408 tcp_assign_congestion_control(sk);
412 sk->sk_state = TCP_CLOSE;
414 sk->sk_write_space = sk_stream_write_space;
415 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
417 icsk->icsk_sync_mss = tcp_sync_mss;
419 sk->sk_sndbuf = sysctl_tcp_wmem[1];
420 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
423 sock_update_memcg(sk);
424 sk_sockets_allocated_inc(sk);
427 EXPORT_SYMBOL(tcp_init_sock);
429 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb)
431 if (sk->sk_tsflags) {
432 struct skb_shared_info *shinfo = skb_shinfo(skb);
434 sock_tx_timestamp(sk, &shinfo->tx_flags);
435 if (shinfo->tx_flags & SKBTX_ANY_TSTAMP)
436 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
441 * Wait for a TCP event.
443 * Note that we don't need to lock the socket, as the upper poll layers
444 * take care of normal races (between the test and the event) and we don't
445 * go look at any of the socket buffers directly.
447 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
450 struct sock *sk = sock->sk;
451 const struct tcp_sock *tp = tcp_sk(sk);
453 sock_rps_record_flow(sk);
455 sock_poll_wait(file, sk_sleep(sk), wait);
456 if (sk->sk_state == TCP_LISTEN)
457 return inet_csk_listen_poll(sk);
459 /* Socket is not locked. We are protected from async events
460 * by poll logic and correct handling of state changes
461 * made by other threads is impossible in any case.
467 * POLLHUP is certainly not done right. But poll() doesn't
468 * have a notion of HUP in just one direction, and for a
469 * socket the read side is more interesting.
471 * Some poll() documentation says that POLLHUP is incompatible
472 * with the POLLOUT/POLLWR flags, so somebody should check this
473 * all. But careful, it tends to be safer to return too many
474 * bits than too few, and you can easily break real applications
475 * if you don't tell them that something has hung up!
479 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
480 * our fs/select.c). It means that after we received EOF,
481 * poll always returns immediately, making impossible poll() on write()
482 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
483 * if and only if shutdown has been made in both directions.
484 * Actually, it is interesting to look how Solaris and DUX
485 * solve this dilemma. I would prefer, if POLLHUP were maskable,
486 * then we could set it on SND_SHUTDOWN. BTW examples given
487 * in Stevens' books assume exactly this behaviour, it explains
488 * why POLLHUP is incompatible with POLLOUT. --ANK
490 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
491 * blocking on fresh not-connected or disconnected socket. --ANK
493 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
495 if (sk->sk_shutdown & RCV_SHUTDOWN)
496 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
498 /* Connected or passive Fast Open socket? */
499 if (sk->sk_state != TCP_SYN_SENT &&
500 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk)) {
501 int target = sock_rcvlowat(sk, 0, INT_MAX);
503 if (tp->urg_seq == tp->copied_seq &&
504 !sock_flag(sk, SOCK_URGINLINE) &&
508 /* Potential race condition. If read of tp below will
509 * escape above sk->sk_state, we can be illegally awaken
510 * in SYN_* states. */
511 if (tp->rcv_nxt - tp->copied_seq >= target)
512 mask |= POLLIN | POLLRDNORM;
514 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
515 if (sk_stream_is_writeable(sk)) {
516 mask |= POLLOUT | POLLWRNORM;
517 } else { /* send SIGIO later */
518 set_bit(SOCK_ASYNC_NOSPACE,
519 &sk->sk_socket->flags);
520 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
522 /* Race breaker. If space is freed after
523 * wspace test but before the flags are set,
524 * IO signal will be lost. Memory barrier
525 * pairs with the input side.
527 smp_mb__after_atomic();
528 if (sk_stream_is_writeable(sk))
529 mask |= POLLOUT | POLLWRNORM;
532 mask |= POLLOUT | POLLWRNORM;
534 if (tp->urg_data & TCP_URG_VALID)
537 /* This barrier is coupled with smp_wmb() in tcp_reset() */
539 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
544 EXPORT_SYMBOL(tcp_poll);
546 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
548 struct tcp_sock *tp = tcp_sk(sk);
554 if (sk->sk_state == TCP_LISTEN)
557 slow = lock_sock_fast(sk);
558 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
560 else if (sock_flag(sk, SOCK_URGINLINE) ||
562 before(tp->urg_seq, tp->copied_seq) ||
563 !before(tp->urg_seq, tp->rcv_nxt)) {
565 answ = tp->rcv_nxt - tp->copied_seq;
567 /* Subtract 1, if FIN was received */
568 if (answ && sock_flag(sk, SOCK_DONE))
571 answ = tp->urg_seq - tp->copied_seq;
572 unlock_sock_fast(sk, slow);
575 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
578 if (sk->sk_state == TCP_LISTEN)
581 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
584 answ = tp->write_seq - tp->snd_una;
587 if (sk->sk_state == TCP_LISTEN)
590 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
593 answ = tp->write_seq - tp->snd_nxt;
599 return put_user(answ, (int __user *)arg);
601 EXPORT_SYMBOL(tcp_ioctl);
603 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
605 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
606 tp->pushed_seq = tp->write_seq;
609 static inline bool forced_push(const struct tcp_sock *tp)
611 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
614 static void skb_entail(struct sock *sk, struct sk_buff *skb)
616 struct tcp_sock *tp = tcp_sk(sk);
617 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
620 tcb->seq = tcb->end_seq = tp->write_seq;
621 tcb->tcp_flags = TCPHDR_ACK;
623 __skb_header_release(skb);
624 tcp_add_write_queue_tail(sk, skb);
625 sk->sk_wmem_queued += skb->truesize;
626 sk_mem_charge(sk, skb->truesize);
627 if (tp->nonagle & TCP_NAGLE_PUSH)
628 tp->nonagle &= ~TCP_NAGLE_PUSH;
631 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
634 tp->snd_up = tp->write_seq;
637 /* If a not yet filled skb is pushed, do not send it if
638 * we have data packets in Qdisc or NIC queues :
639 * Because TX completion will happen shortly, it gives a chance
640 * to coalesce future sendmsg() payload into this skb, without
641 * need for a timer, and with no latency trade off.
642 * As packets containing data payload have a bigger truesize
643 * than pure acks (dataless) packets, the last checks prevent
644 * autocorking if we only have an ACK in Qdisc/NIC queues,
645 * or if TX completion was delayed after we processed ACK packet.
647 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
650 return skb->len < size_goal &&
651 sysctl_tcp_autocorking &&
652 skb != tcp_write_queue_head(sk) &&
653 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
656 static void tcp_push(struct sock *sk, int flags, int mss_now,
657 int nonagle, int size_goal)
659 struct tcp_sock *tp = tcp_sk(sk);
662 if (!tcp_send_head(sk))
665 skb = tcp_write_queue_tail(sk);
666 if (!(flags & MSG_MORE) || forced_push(tp))
667 tcp_mark_push(tp, skb);
669 tcp_mark_urg(tp, flags);
671 if (tcp_should_autocork(sk, skb, size_goal)) {
673 /* avoid atomic op if TSQ_THROTTLED bit is already set */
674 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
675 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
676 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
678 /* It is possible TX completion already happened
679 * before we set TSQ_THROTTLED.
681 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
685 if (flags & MSG_MORE)
686 nonagle = TCP_NAGLE_CORK;
688 __tcp_push_pending_frames(sk, mss_now, nonagle);
691 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
692 unsigned int offset, size_t len)
694 struct tcp_splice_state *tss = rd_desc->arg.data;
697 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
700 rd_desc->count -= ret;
704 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
706 /* Store TCP splice context information in read_descriptor_t. */
707 read_descriptor_t rd_desc = {
712 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
716 * tcp_splice_read - splice data from TCP socket to a pipe
717 * @sock: socket to splice from
718 * @ppos: position (not valid)
719 * @pipe: pipe to splice to
720 * @len: number of bytes to splice
721 * @flags: splice modifier flags
724 * Will read pages from given socket and fill them into a pipe.
727 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
728 struct pipe_inode_info *pipe, size_t len,
731 struct sock *sk = sock->sk;
732 struct tcp_splice_state tss = {
741 sock_rps_record_flow(sk);
743 * We can't seek on a socket input
752 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
754 ret = __tcp_splice_read(sk, &tss);
760 if (sock_flag(sk, SOCK_DONE))
763 ret = sock_error(sk);
766 if (sk->sk_shutdown & RCV_SHUTDOWN)
768 if (sk->sk_state == TCP_CLOSE) {
770 * This occurs when user tries to read
771 * from never connected socket.
773 if (!sock_flag(sk, SOCK_DONE))
781 sk_wait_data(sk, &timeo);
782 if (signal_pending(current)) {
783 ret = sock_intr_errno(timeo);
796 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
797 (sk->sk_shutdown & RCV_SHUTDOWN) ||
798 signal_pending(current))
809 EXPORT_SYMBOL(tcp_splice_read);
811 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
816 /* The TCP header must be at least 32-bit aligned. */
817 size = ALIGN(size, 4);
819 if (unlikely(tcp_under_memory_pressure(sk)))
820 sk_mem_reclaim_partial(sk);
822 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
826 if (force_schedule) {
827 mem_scheduled = true;
828 sk_forced_mem_schedule(sk, skb->truesize);
830 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
832 if (likely(mem_scheduled)) {
833 skb_reserve(skb, sk->sk_prot->max_header);
835 * Make sure that we have exactly size bytes
836 * available to the caller, no more, no less.
838 skb->reserved_tailroom = skb->end - skb->tail - size;
843 sk->sk_prot->enter_memory_pressure(sk);
844 sk_stream_moderate_sndbuf(sk);
849 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
852 struct tcp_sock *tp = tcp_sk(sk);
853 u32 new_size_goal, size_goal;
855 if (!large_allowed || !sk_can_gso(sk))
858 /* Note : tcp_tso_autosize() will eventually split this later */
859 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
860 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
862 /* We try hard to avoid divides here */
863 size_goal = tp->gso_segs * mss_now;
864 if (unlikely(new_size_goal < size_goal ||
865 new_size_goal >= size_goal + mss_now)) {
866 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
867 sk->sk_gso_max_segs);
868 size_goal = tp->gso_segs * mss_now;
871 return max(size_goal, mss_now);
874 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
878 mss_now = tcp_current_mss(sk);
879 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
884 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
885 size_t size, int flags)
887 struct tcp_sock *tp = tcp_sk(sk);
888 int mss_now, size_goal;
891 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
893 /* Wait for a connection to finish. One exception is TCP Fast Open
894 * (passive side) where data is allowed to be sent before a connection
895 * is fully established.
897 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
898 !tcp_passive_fastopen(sk)) {
899 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
903 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
905 mss_now = tcp_send_mss(sk, &size_goal, flags);
909 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
913 struct sk_buff *skb = tcp_write_queue_tail(sk);
917 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
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 >= 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;
966 if (!(size -= copy)) {
967 tcp_tx_timestamp(sk, skb);
971 if (skb->len < size_goal || (flags & MSG_OOB))
974 if (forced_push(tp)) {
975 tcp_mark_push(tp, skb);
976 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
977 } else if (skb == tcp_send_head(sk))
978 tcp_push_one(sk, mss_now);
982 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
984 tcp_push(sk, flags & ~MSG_MORE, mss_now,
985 TCP_NAGLE_PUSH, size_goal);
987 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
990 mss_now = tcp_send_mss(sk, &size_goal, flags);
994 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
995 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1002 /* make sure we wake any epoll edge trigger waiter */
1003 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1004 sk->sk_write_space(sk);
1005 return sk_stream_error(sk, flags, err);
1008 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1009 size_t size, int flags)
1013 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1014 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1015 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1019 res = do_tcp_sendpages(sk, page, offset, size, flags);
1023 EXPORT_SYMBOL(tcp_sendpage);
1025 static inline int select_size(const struct sock *sk, bool sg)
1027 const struct tcp_sock *tp = tcp_sk(sk);
1028 int tmp = tp->mss_cache;
1031 if (sk_can_gso(sk)) {
1032 /* Small frames wont use a full page:
1033 * Payload will immediately follow tcp header.
1035 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1037 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1039 if (tmp >= pgbreak &&
1040 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1048 void tcp_free_fastopen_req(struct tcp_sock *tp)
1050 if (tp->fastopen_req) {
1051 kfree(tp->fastopen_req);
1052 tp->fastopen_req = NULL;
1056 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1057 int *copied, size_t size)
1059 struct tcp_sock *tp = tcp_sk(sk);
1062 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1064 if (tp->fastopen_req)
1065 return -EALREADY; /* Another Fast Open is in progress */
1067 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1069 if (unlikely(!tp->fastopen_req))
1071 tp->fastopen_req->data = msg;
1072 tp->fastopen_req->size = size;
1074 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1075 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1076 msg->msg_namelen, flags);
1077 *copied = tp->fastopen_req->copied;
1078 tcp_free_fastopen_req(tp);
1082 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1084 struct tcp_sock *tp = tcp_sk(sk);
1085 struct sk_buff *skb;
1086 int flags, err, copied = 0;
1087 int mss_now = 0, size_goal, copied_syn = 0;
1093 flags = msg->msg_flags;
1094 if (flags & MSG_FASTOPEN) {
1095 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1096 if (err == -EINPROGRESS && copied_syn > 0)
1102 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1104 /* Wait for a connection to finish. One exception is TCP Fast Open
1105 * (passive side) where data is allowed to be sent before a connection
1106 * is fully established.
1108 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1109 !tcp_passive_fastopen(sk)) {
1110 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1114 if (unlikely(tp->repair)) {
1115 if (tp->repair_queue == TCP_RECV_QUEUE) {
1116 copied = tcp_send_rcvq(sk, msg, size);
1121 if (tp->repair_queue == TCP_NO_QUEUE)
1124 /* 'common' sending to sendq */
1127 /* This should be in poll */
1128 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1130 mss_now = tcp_send_mss(sk, &size_goal, flags);
1132 /* Ok commence sending. */
1136 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1139 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1141 while (msg_data_left(msg)) {
1143 int max = size_goal;
1145 skb = tcp_write_queue_tail(sk);
1146 if (tcp_send_head(sk)) {
1147 if (skb->ip_summed == CHECKSUM_NONE)
1149 copy = max - skb->len;
1154 /* Allocate new segment. If the interface is SG,
1155 * allocate skb fitting to single page.
1157 if (!sk_stream_memory_free(sk))
1158 goto wait_for_sndbuf;
1160 skb = sk_stream_alloc_skb(sk,
1161 select_size(sk, sg),
1163 skb_queue_empty(&sk->sk_write_queue));
1165 goto wait_for_memory;
1168 * Check whether we can use HW checksum.
1170 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1171 skb->ip_summed = CHECKSUM_PARTIAL;
1173 skb_entail(sk, skb);
1177 /* All packets are restored as if they have
1178 * already been sent. skb_mstamp isn't set to
1179 * avoid wrong rtt estimation.
1182 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1185 /* Try to append data to the end of skb. */
1186 if (copy > msg_data_left(msg))
1187 copy = msg_data_left(msg);
1189 /* Where to copy to? */
1190 if (skb_availroom(skb) > 0) {
1191 /* We have some space in skb head. Superb! */
1192 copy = min_t(int, copy, skb_availroom(skb));
1193 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1198 int i = skb_shinfo(skb)->nr_frags;
1199 struct page_frag *pfrag = sk_page_frag(sk);
1201 if (!sk_page_frag_refill(sk, pfrag))
1202 goto wait_for_memory;
1204 if (!skb_can_coalesce(skb, i, pfrag->page,
1206 if (i == MAX_SKB_FRAGS || !sg) {
1207 tcp_mark_push(tp, skb);
1213 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1215 if (!sk_wmem_schedule(sk, copy))
1216 goto wait_for_memory;
1218 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1225 /* Update the skb. */
1227 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1229 skb_fill_page_desc(skb, i, pfrag->page,
1230 pfrag->offset, copy);
1231 get_page(pfrag->page);
1233 pfrag->offset += copy;
1237 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1239 tp->write_seq += copy;
1240 TCP_SKB_CB(skb)->end_seq += copy;
1241 tcp_skb_pcount_set(skb, 0);
1244 if (!msg_data_left(msg)) {
1245 tcp_tx_timestamp(sk, skb);
1249 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1252 if (forced_push(tp)) {
1253 tcp_mark_push(tp, skb);
1254 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1255 } else if (skb == tcp_send_head(sk))
1256 tcp_push_one(sk, mss_now);
1260 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1263 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1264 TCP_NAGLE_PUSH, size_goal);
1266 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1269 mss_now = tcp_send_mss(sk, &size_goal, flags);
1274 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1277 return copied + copied_syn;
1281 tcp_unlink_write_queue(skb, sk);
1282 /* It is the one place in all of TCP, except connection
1283 * reset, where we can be unlinking the send_head.
1285 tcp_check_send_head(sk, skb);
1286 sk_wmem_free_skb(sk, skb);
1290 if (copied + copied_syn)
1293 err = sk_stream_error(sk, flags, err);
1294 /* make sure we wake any epoll edge trigger waiter */
1295 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1296 sk->sk_write_space(sk);
1300 EXPORT_SYMBOL(tcp_sendmsg);
1303 * Handle reading urgent data. BSD has very simple semantics for
1304 * this, no blocking and very strange errors 8)
1307 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1309 struct tcp_sock *tp = tcp_sk(sk);
1311 /* No URG data to read. */
1312 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1313 tp->urg_data == TCP_URG_READ)
1314 return -EINVAL; /* Yes this is right ! */
1316 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1319 if (tp->urg_data & TCP_URG_VALID) {
1321 char c = tp->urg_data;
1323 if (!(flags & MSG_PEEK))
1324 tp->urg_data = TCP_URG_READ;
1326 /* Read urgent data. */
1327 msg->msg_flags |= MSG_OOB;
1330 if (!(flags & MSG_TRUNC))
1331 err = memcpy_to_msg(msg, &c, 1);
1334 msg->msg_flags |= MSG_TRUNC;
1336 return err ? -EFAULT : len;
1339 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1342 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1343 * the available implementations agree in this case:
1344 * this call should never block, independent of the
1345 * blocking state of the socket.
1346 * Mike <pall@rz.uni-karlsruhe.de>
1351 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1353 struct sk_buff *skb;
1354 int copied = 0, err = 0;
1356 /* XXX -- need to support SO_PEEK_OFF */
1358 skb_queue_walk(&sk->sk_write_queue, skb) {
1359 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1366 return err ?: copied;
1369 /* Clean up the receive buffer for full frames taken by the user,
1370 * then send an ACK if necessary. COPIED is the number of bytes
1371 * tcp_recvmsg has given to the user so far, it speeds up the
1372 * calculation of whether or not we must ACK for the sake of
1375 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1377 struct tcp_sock *tp = tcp_sk(sk);
1378 bool time_to_ack = false;
1380 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1382 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1383 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1384 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1386 if (inet_csk_ack_scheduled(sk)) {
1387 const struct inet_connection_sock *icsk = inet_csk(sk);
1388 /* Delayed ACKs frequently hit locked sockets during bulk
1390 if (icsk->icsk_ack.blocked ||
1391 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1392 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1394 * If this read emptied read buffer, we send ACK, if
1395 * connection is not bidirectional, user drained
1396 * receive buffer and there was a small segment
1400 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1401 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1402 !icsk->icsk_ack.pingpong)) &&
1403 !atomic_read(&sk->sk_rmem_alloc)))
1407 /* We send an ACK if we can now advertise a non-zero window
1408 * which has been raised "significantly".
1410 * Even if window raised up to infinity, do not send window open ACK
1411 * in states, where we will not receive more. It is useless.
1413 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1414 __u32 rcv_window_now = tcp_receive_window(tp);
1416 /* Optimize, __tcp_select_window() is not cheap. */
1417 if (2*rcv_window_now <= tp->window_clamp) {
1418 __u32 new_window = __tcp_select_window(sk);
1420 /* Send ACK now, if this read freed lots of space
1421 * in our buffer. Certainly, new_window is new window.
1422 * We can advertise it now, if it is not less than current one.
1423 * "Lots" means "at least twice" here.
1425 if (new_window && new_window >= 2 * rcv_window_now)
1433 static void tcp_prequeue_process(struct sock *sk)
1435 struct sk_buff *skb;
1436 struct tcp_sock *tp = tcp_sk(sk);
1438 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1440 /* RX process wants to run with disabled BHs, though it is not
1443 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1444 sk_backlog_rcv(sk, skb);
1447 /* Clear memory counter. */
1448 tp->ucopy.memory = 0;
1451 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1453 struct sk_buff *skb;
1456 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1457 offset = seq - TCP_SKB_CB(skb)->seq;
1458 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1460 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1464 /* This looks weird, but this can happen if TCP collapsing
1465 * splitted a fat GRO packet, while we released socket lock
1466 * in skb_splice_bits()
1468 sk_eat_skb(sk, skb);
1474 * This routine provides an alternative to tcp_recvmsg() for routines
1475 * that would like to handle copying from skbuffs directly in 'sendfile'
1478 * - It is assumed that the socket was locked by the caller.
1479 * - The routine does not block.
1480 * - At present, there is no support for reading OOB data
1481 * or for 'peeking' the socket using this routine
1482 * (although both would be easy to implement).
1484 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1485 sk_read_actor_t recv_actor)
1487 struct sk_buff *skb;
1488 struct tcp_sock *tp = tcp_sk(sk);
1489 u32 seq = tp->copied_seq;
1493 if (sk->sk_state == TCP_LISTEN)
1495 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1496 if (offset < skb->len) {
1500 len = skb->len - offset;
1501 /* Stop reading if we hit a patch of urgent data */
1503 u32 urg_offset = tp->urg_seq - seq;
1504 if (urg_offset < len)
1509 used = recv_actor(desc, skb, offset, len);
1514 } else if (used <= len) {
1519 /* If recv_actor drops the lock (e.g. TCP splice
1520 * receive) the skb pointer might be invalid when
1521 * getting here: tcp_collapse might have deleted it
1522 * while aggregating skbs from the socket queue.
1524 skb = tcp_recv_skb(sk, seq - 1, &offset);
1527 /* TCP coalescing might have appended data to the skb.
1528 * Try to splice more frags
1530 if (offset + 1 != skb->len)
1533 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1534 sk_eat_skb(sk, skb);
1538 sk_eat_skb(sk, skb);
1541 tp->copied_seq = seq;
1543 tp->copied_seq = seq;
1545 tcp_rcv_space_adjust(sk);
1547 /* Clean up data we have read: This will do ACK frames. */
1549 tcp_recv_skb(sk, seq, &offset);
1550 tcp_cleanup_rbuf(sk, copied);
1554 EXPORT_SYMBOL(tcp_read_sock);
1557 * This routine copies from a sock struct into the user buffer.
1559 * Technical note: in 2.3 we work on _locked_ socket, so that
1560 * tricks with *seq access order and skb->users are not required.
1561 * Probably, code can be easily improved even more.
1564 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1565 int flags, int *addr_len)
1567 struct tcp_sock *tp = tcp_sk(sk);
1573 int target; /* Read at least this many bytes */
1575 struct task_struct *user_recv = NULL;
1576 struct sk_buff *skb;
1579 if (unlikely(flags & MSG_ERRQUEUE))
1580 return inet_recv_error(sk, msg, len, addr_len);
1582 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1583 (sk->sk_state == TCP_ESTABLISHED))
1584 sk_busy_loop(sk, nonblock);
1589 if (sk->sk_state == TCP_LISTEN)
1592 timeo = sock_rcvtimeo(sk, nonblock);
1594 /* Urgent data needs to be handled specially. */
1595 if (flags & MSG_OOB)
1598 if (unlikely(tp->repair)) {
1600 if (!(flags & MSG_PEEK))
1603 if (tp->repair_queue == TCP_SEND_QUEUE)
1607 if (tp->repair_queue == TCP_NO_QUEUE)
1610 /* 'common' recv queue MSG_PEEK-ing */
1613 seq = &tp->copied_seq;
1614 if (flags & MSG_PEEK) {
1615 peek_seq = tp->copied_seq;
1619 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1624 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1625 if (tp->urg_data && tp->urg_seq == *seq) {
1628 if (signal_pending(current)) {
1629 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1634 /* Next get a buffer. */
1636 skb_queue_walk(&sk->sk_receive_queue, skb) {
1637 /* Now that we have two receive queues this
1640 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1641 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1642 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1646 offset = *seq - TCP_SKB_CB(skb)->seq;
1647 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1649 if (offset < skb->len)
1651 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1653 WARN(!(flags & MSG_PEEK),
1654 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1655 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1658 /* Well, if we have backlog, try to process it now yet. */
1660 if (copied >= target && !sk->sk_backlog.tail)
1665 sk->sk_state == TCP_CLOSE ||
1666 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1668 signal_pending(current))
1671 if (sock_flag(sk, SOCK_DONE))
1675 copied = sock_error(sk);
1679 if (sk->sk_shutdown & RCV_SHUTDOWN)
1682 if (sk->sk_state == TCP_CLOSE) {
1683 if (!sock_flag(sk, SOCK_DONE)) {
1684 /* This occurs when user tries to read
1685 * from never connected socket.
1698 if (signal_pending(current)) {
1699 copied = sock_intr_errno(timeo);
1704 tcp_cleanup_rbuf(sk, copied);
1706 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1707 /* Install new reader */
1708 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1709 user_recv = current;
1710 tp->ucopy.task = user_recv;
1711 tp->ucopy.msg = msg;
1714 tp->ucopy.len = len;
1716 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1717 !(flags & (MSG_PEEK | MSG_TRUNC)));
1719 /* Ugly... If prequeue is not empty, we have to
1720 * process it before releasing socket, otherwise
1721 * order will be broken at second iteration.
1722 * More elegant solution is required!!!
1724 * Look: we have the following (pseudo)queues:
1726 * 1. packets in flight
1731 * Each queue can be processed only if the next ones
1732 * are empty. At this point we have empty receive_queue.
1733 * But prequeue _can_ be not empty after 2nd iteration,
1734 * when we jumped to start of loop because backlog
1735 * processing added something to receive_queue.
1736 * We cannot release_sock(), because backlog contains
1737 * packets arrived _after_ prequeued ones.
1739 * Shortly, algorithm is clear --- to process all
1740 * the queues in order. We could make it more directly,
1741 * requeueing packets from backlog to prequeue, if
1742 * is not empty. It is more elegant, but eats cycles,
1745 if (!skb_queue_empty(&tp->ucopy.prequeue))
1748 /* __ Set realtime policy in scheduler __ */
1751 if (copied >= target) {
1752 /* Do not sleep, just process backlog. */
1756 sk_wait_data(sk, &timeo);
1761 /* __ Restore normal policy in scheduler __ */
1763 if ((chunk = len - tp->ucopy.len) != 0) {
1764 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1769 if (tp->rcv_nxt == tp->copied_seq &&
1770 !skb_queue_empty(&tp->ucopy.prequeue)) {
1772 tcp_prequeue_process(sk);
1774 if ((chunk = len - tp->ucopy.len) != 0) {
1775 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1781 if ((flags & MSG_PEEK) &&
1782 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1783 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1785 task_pid_nr(current));
1786 peek_seq = tp->copied_seq;
1791 /* Ok so how much can we use? */
1792 used = skb->len - offset;
1796 /* Do we have urgent data here? */
1798 u32 urg_offset = tp->urg_seq - *seq;
1799 if (urg_offset < used) {
1801 if (!sock_flag(sk, SOCK_URGINLINE)) {
1814 if (!(flags & MSG_TRUNC)) {
1815 err = skb_copy_datagram_msg(skb, offset, msg, used);
1817 /* Exception. Bailout! */
1828 tcp_rcv_space_adjust(sk);
1831 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1833 tcp_fast_path_check(sk);
1835 if (used + offset < skb->len)
1838 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1840 if (!(flags & MSG_PEEK))
1841 sk_eat_skb(sk, skb);
1845 /* Process the FIN. */
1847 if (!(flags & MSG_PEEK))
1848 sk_eat_skb(sk, skb);
1853 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1856 tp->ucopy.len = copied > 0 ? len : 0;
1858 tcp_prequeue_process(sk);
1860 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1861 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1867 tp->ucopy.task = NULL;
1871 /* According to UNIX98, msg_name/msg_namelen are ignored
1872 * on connected socket. I was just happy when found this 8) --ANK
1875 /* Clean up data we have read: This will do ACK frames. */
1876 tcp_cleanup_rbuf(sk, copied);
1886 err = tcp_recv_urg(sk, msg, len, flags);
1890 err = tcp_peek_sndq(sk, msg, len);
1893 EXPORT_SYMBOL(tcp_recvmsg);
1895 void tcp_set_state(struct sock *sk, int state)
1897 int oldstate = sk->sk_state;
1900 case TCP_ESTABLISHED:
1901 if (oldstate != TCP_ESTABLISHED)
1902 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1906 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1907 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1909 sk->sk_prot->unhash(sk);
1910 if (inet_csk(sk)->icsk_bind_hash &&
1911 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1915 if (oldstate == TCP_ESTABLISHED)
1916 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1919 /* Change state AFTER socket is unhashed to avoid closed
1920 * socket sitting in hash tables.
1922 sk->sk_state = state;
1925 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1928 EXPORT_SYMBOL_GPL(tcp_set_state);
1931 * State processing on a close. This implements the state shift for
1932 * sending our FIN frame. Note that we only send a FIN for some
1933 * states. A shutdown() may have already sent the FIN, or we may be
1937 static const unsigned char new_state[16] = {
1938 /* current state: new state: action: */
1939 [0 /* (Invalid) */] = TCP_CLOSE,
1940 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1941 [TCP_SYN_SENT] = TCP_CLOSE,
1942 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1943 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1944 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1945 [TCP_TIME_WAIT] = TCP_CLOSE,
1946 [TCP_CLOSE] = TCP_CLOSE,
1947 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
1948 [TCP_LAST_ACK] = TCP_LAST_ACK,
1949 [TCP_LISTEN] = TCP_CLOSE,
1950 [TCP_CLOSING] = TCP_CLOSING,
1951 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
1954 static int tcp_close_state(struct sock *sk)
1956 int next = (int)new_state[sk->sk_state];
1957 int ns = next & TCP_STATE_MASK;
1959 tcp_set_state(sk, ns);
1961 return next & TCP_ACTION_FIN;
1965 * Shutdown the sending side of a connection. Much like close except
1966 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1969 void tcp_shutdown(struct sock *sk, int how)
1971 /* We need to grab some memory, and put together a FIN,
1972 * and then put it into the queue to be sent.
1973 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1975 if (!(how & SEND_SHUTDOWN))
1978 /* If we've already sent a FIN, or it's a closed state, skip this. */
1979 if ((1 << sk->sk_state) &
1980 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1981 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1982 /* Clear out any half completed packets. FIN if needed. */
1983 if (tcp_close_state(sk))
1987 EXPORT_SYMBOL(tcp_shutdown);
1989 bool tcp_check_oom(struct sock *sk, int shift)
1991 bool too_many_orphans, out_of_socket_memory;
1993 too_many_orphans = tcp_too_many_orphans(sk, shift);
1994 out_of_socket_memory = tcp_out_of_memory(sk);
1996 if (too_many_orphans)
1997 net_info_ratelimited("too many orphaned sockets\n");
1998 if (out_of_socket_memory)
1999 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2000 return too_many_orphans || out_of_socket_memory;
2003 void tcp_close(struct sock *sk, long timeout)
2005 struct sk_buff *skb;
2006 int data_was_unread = 0;
2010 sk->sk_shutdown = SHUTDOWN_MASK;
2012 if (sk->sk_state == TCP_LISTEN) {
2013 tcp_set_state(sk, TCP_CLOSE);
2016 inet_csk_listen_stop(sk);
2018 goto adjudge_to_death;
2021 /* We need to flush the recv. buffs. We do this only on the
2022 * descriptor close, not protocol-sourced closes, because the
2023 * reader process may not have drained the data yet!
2025 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2026 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2028 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2030 data_was_unread += len;
2036 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2037 if (sk->sk_state == TCP_CLOSE)
2038 goto adjudge_to_death;
2040 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2041 * data was lost. To witness the awful effects of the old behavior of
2042 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2043 * GET in an FTP client, suspend the process, wait for the client to
2044 * advertise a zero window, then kill -9 the FTP client, wheee...
2045 * Note: timeout is always zero in such a case.
2047 if (unlikely(tcp_sk(sk)->repair)) {
2048 sk->sk_prot->disconnect(sk, 0);
2049 } else if (data_was_unread) {
2050 /* Unread data was tossed, zap the connection. */
2051 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2052 tcp_set_state(sk, TCP_CLOSE);
2053 tcp_send_active_reset(sk, sk->sk_allocation);
2054 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2055 /* Check zero linger _after_ checking for unread data. */
2056 sk->sk_prot->disconnect(sk, 0);
2057 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2058 } else if (tcp_close_state(sk)) {
2059 /* We FIN if the application ate all the data before
2060 * zapping the connection.
2063 /* RED-PEN. Formally speaking, we have broken TCP state
2064 * machine. State transitions:
2066 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2067 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2068 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2070 * are legal only when FIN has been sent (i.e. in window),
2071 * rather than queued out of window. Purists blame.
2073 * F.e. "RFC state" is ESTABLISHED,
2074 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2076 * The visible declinations are that sometimes
2077 * we enter time-wait state, when it is not required really
2078 * (harmless), do not send active resets, when they are
2079 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2080 * they look as CLOSING or LAST_ACK for Linux)
2081 * Probably, I missed some more holelets.
2083 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2084 * in a single packet! (May consider it later but will
2085 * probably need API support or TCP_CORK SYN-ACK until
2086 * data is written and socket is closed.)
2091 sk_stream_wait_close(sk, timeout);
2094 state = sk->sk_state;
2098 /* It is the last release_sock in its life. It will remove backlog. */
2102 /* Now socket is owned by kernel and we acquire BH lock
2103 to finish close. No need to check for user refs.
2107 WARN_ON(sock_owned_by_user(sk));
2109 percpu_counter_inc(sk->sk_prot->orphan_count);
2111 /* Have we already been destroyed by a softirq or backlog? */
2112 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2115 /* This is a (useful) BSD violating of the RFC. There is a
2116 * problem with TCP as specified in that the other end could
2117 * keep a socket open forever with no application left this end.
2118 * We use a 1 minute timeout (about the same as BSD) then kill
2119 * our end. If they send after that then tough - BUT: long enough
2120 * that we won't make the old 4*rto = almost no time - whoops
2123 * Nope, it was not mistake. It is really desired behaviour
2124 * f.e. on http servers, when such sockets are useless, but
2125 * consume significant resources. Let's do it with special
2126 * linger2 option. --ANK
2129 if (sk->sk_state == TCP_FIN_WAIT2) {
2130 struct tcp_sock *tp = tcp_sk(sk);
2131 if (tp->linger2 < 0) {
2132 tcp_set_state(sk, TCP_CLOSE);
2133 tcp_send_active_reset(sk, GFP_ATOMIC);
2134 NET_INC_STATS_BH(sock_net(sk),
2135 LINUX_MIB_TCPABORTONLINGER);
2137 const int tmo = tcp_fin_time(sk);
2139 if (tmo > TCP_TIMEWAIT_LEN) {
2140 inet_csk_reset_keepalive_timer(sk,
2141 tmo - TCP_TIMEWAIT_LEN);
2143 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2148 if (sk->sk_state != TCP_CLOSE) {
2150 if (tcp_check_oom(sk, 0)) {
2151 tcp_set_state(sk, TCP_CLOSE);
2152 tcp_send_active_reset(sk, GFP_ATOMIC);
2153 NET_INC_STATS_BH(sock_net(sk),
2154 LINUX_MIB_TCPABORTONMEMORY);
2158 if (sk->sk_state == TCP_CLOSE) {
2159 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2160 /* We could get here with a non-NULL req if the socket is
2161 * aborted (e.g., closed with unread data) before 3WHS
2165 reqsk_fastopen_remove(sk, req, false);
2166 inet_csk_destroy_sock(sk);
2168 /* Otherwise, socket is reprieved until protocol close. */
2175 EXPORT_SYMBOL(tcp_close);
2177 /* These states need RST on ABORT according to RFC793 */
2179 static inline bool tcp_need_reset(int state)
2181 return (1 << state) &
2182 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2183 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2186 int tcp_disconnect(struct sock *sk, int flags)
2188 struct inet_sock *inet = inet_sk(sk);
2189 struct inet_connection_sock *icsk = inet_csk(sk);
2190 struct tcp_sock *tp = tcp_sk(sk);
2192 int old_state = sk->sk_state;
2194 if (old_state != TCP_CLOSE)
2195 tcp_set_state(sk, TCP_CLOSE);
2197 /* ABORT function of RFC793 */
2198 if (old_state == TCP_LISTEN) {
2199 inet_csk_listen_stop(sk);
2200 } else if (unlikely(tp->repair)) {
2201 sk->sk_err = ECONNABORTED;
2202 } else if (tcp_need_reset(old_state) ||
2203 (tp->snd_nxt != tp->write_seq &&
2204 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2205 /* The last check adjusts for discrepancy of Linux wrt. RFC
2208 tcp_send_active_reset(sk, gfp_any());
2209 sk->sk_err = ECONNRESET;
2210 } else if (old_state == TCP_SYN_SENT)
2211 sk->sk_err = ECONNRESET;
2213 tcp_clear_xmit_timers(sk);
2214 __skb_queue_purge(&sk->sk_receive_queue);
2215 tcp_write_queue_purge(sk);
2216 __skb_queue_purge(&tp->out_of_order_queue);
2218 inet->inet_dport = 0;
2220 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2221 inet_reset_saddr(sk);
2223 sk->sk_shutdown = 0;
2224 sock_reset_flag(sk, SOCK_DONE);
2226 if ((tp->write_seq += tp->max_window + 2) == 0)
2228 icsk->icsk_backoff = 0;
2230 icsk->icsk_probes_out = 0;
2231 tp->packets_out = 0;
2232 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2233 tp->snd_cwnd_cnt = 0;
2234 tp->window_clamp = 0;
2235 tcp_set_ca_state(sk, TCP_CA_Open);
2236 tcp_clear_retrans(tp);
2237 inet_csk_delack_init(sk);
2238 tcp_init_send_head(sk);
2239 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2242 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2244 sk->sk_error_report(sk);
2247 EXPORT_SYMBOL(tcp_disconnect);
2249 void tcp_sock_destruct(struct sock *sk)
2251 inet_sock_destruct(sk);
2253 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2256 static inline bool tcp_can_repair_sock(const struct sock *sk)
2258 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2259 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2262 static int tcp_repair_options_est(struct tcp_sock *tp,
2263 struct tcp_repair_opt __user *optbuf, unsigned int len)
2265 struct tcp_repair_opt opt;
2267 while (len >= sizeof(opt)) {
2268 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2274 switch (opt.opt_code) {
2276 tp->rx_opt.mss_clamp = opt.opt_val;
2280 u16 snd_wscale = opt.opt_val & 0xFFFF;
2281 u16 rcv_wscale = opt.opt_val >> 16;
2283 if (snd_wscale > 14 || rcv_wscale > 14)
2286 tp->rx_opt.snd_wscale = snd_wscale;
2287 tp->rx_opt.rcv_wscale = rcv_wscale;
2288 tp->rx_opt.wscale_ok = 1;
2291 case TCPOPT_SACK_PERM:
2292 if (opt.opt_val != 0)
2295 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2296 if (sysctl_tcp_fack)
2297 tcp_enable_fack(tp);
2299 case TCPOPT_TIMESTAMP:
2300 if (opt.opt_val != 0)
2303 tp->rx_opt.tstamp_ok = 1;
2312 * Socket option code for TCP.
2314 static int do_tcp_setsockopt(struct sock *sk, int level,
2315 int optname, char __user *optval, unsigned int optlen)
2317 struct tcp_sock *tp = tcp_sk(sk);
2318 struct inet_connection_sock *icsk = inet_csk(sk);
2322 /* These are data/string values, all the others are ints */
2324 case TCP_CONGESTION: {
2325 char name[TCP_CA_NAME_MAX];
2330 val = strncpy_from_user(name, optval,
2331 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2337 err = tcp_set_congestion_control(sk, name);
2346 if (optlen < sizeof(int))
2349 if (get_user(val, (int __user *)optval))
2356 /* Values greater than interface MTU won't take effect. However
2357 * at the point when this call is done we typically don't yet
2358 * know which interface is going to be used */
2359 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2363 tp->rx_opt.user_mss = val;
2368 /* TCP_NODELAY is weaker than TCP_CORK, so that
2369 * this option on corked socket is remembered, but
2370 * it is not activated until cork is cleared.
2372 * However, when TCP_NODELAY is set we make
2373 * an explicit push, which overrides even TCP_CORK
2374 * for currently queued segments.
2376 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2377 tcp_push_pending_frames(sk);
2379 tp->nonagle &= ~TCP_NAGLE_OFF;
2383 case TCP_THIN_LINEAR_TIMEOUTS:
2384 if (val < 0 || val > 1)
2390 case TCP_THIN_DUPACK:
2391 if (val < 0 || val > 1)
2394 tp->thin_dupack = val;
2395 if (tp->thin_dupack)
2396 tcp_disable_early_retrans(tp);
2401 if (!tcp_can_repair_sock(sk))
2403 else if (val == 1) {
2405 sk->sk_reuse = SK_FORCE_REUSE;
2406 tp->repair_queue = TCP_NO_QUEUE;
2407 } else if (val == 0) {
2409 sk->sk_reuse = SK_NO_REUSE;
2410 tcp_send_window_probe(sk);
2416 case TCP_REPAIR_QUEUE:
2419 else if (val < TCP_QUEUES_NR)
2420 tp->repair_queue = val;
2426 if (sk->sk_state != TCP_CLOSE)
2428 else if (tp->repair_queue == TCP_SEND_QUEUE)
2429 tp->write_seq = val;
2430 else if (tp->repair_queue == TCP_RECV_QUEUE)
2436 case TCP_REPAIR_OPTIONS:
2439 else if (sk->sk_state == TCP_ESTABLISHED)
2440 err = tcp_repair_options_est(tp,
2441 (struct tcp_repair_opt __user *)optval,
2448 /* When set indicates to always queue non-full frames.
2449 * Later the user clears this option and we transmit
2450 * any pending partial frames in the queue. This is
2451 * meant to be used alongside sendfile() to get properly
2452 * filled frames when the user (for example) must write
2453 * out headers with a write() call first and then use
2454 * sendfile to send out the data parts.
2456 * TCP_CORK can be set together with TCP_NODELAY and it is
2457 * stronger than TCP_NODELAY.
2460 tp->nonagle |= TCP_NAGLE_CORK;
2462 tp->nonagle &= ~TCP_NAGLE_CORK;
2463 if (tp->nonagle&TCP_NAGLE_OFF)
2464 tp->nonagle |= TCP_NAGLE_PUSH;
2465 tcp_push_pending_frames(sk);
2470 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2473 tp->keepalive_time = val * HZ;
2474 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2475 !((1 << sk->sk_state) &
2476 (TCPF_CLOSE | TCPF_LISTEN))) {
2477 u32 elapsed = keepalive_time_elapsed(tp);
2478 if (tp->keepalive_time > elapsed)
2479 elapsed = tp->keepalive_time - elapsed;
2482 inet_csk_reset_keepalive_timer(sk, elapsed);
2487 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2490 tp->keepalive_intvl = val * HZ;
2493 if (val < 1 || val > MAX_TCP_KEEPCNT)
2496 tp->keepalive_probes = val;
2499 if (val < 1 || val > MAX_TCP_SYNCNT)
2502 icsk->icsk_syn_retries = val;
2506 if (val < 0 || val > 1)
2515 else if (val > sysctl_tcp_fin_timeout / HZ)
2518 tp->linger2 = val * HZ;
2521 case TCP_DEFER_ACCEPT:
2522 /* Translate value in seconds to number of retransmits */
2523 icsk->icsk_accept_queue.rskq_defer_accept =
2524 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2528 case TCP_WINDOW_CLAMP:
2530 if (sk->sk_state != TCP_CLOSE) {
2534 tp->window_clamp = 0;
2536 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2537 SOCK_MIN_RCVBUF / 2 : val;
2542 icsk->icsk_ack.pingpong = 1;
2544 icsk->icsk_ack.pingpong = 0;
2545 if ((1 << sk->sk_state) &
2546 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2547 inet_csk_ack_scheduled(sk)) {
2548 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2549 tcp_cleanup_rbuf(sk, 1);
2551 icsk->icsk_ack.pingpong = 1;
2556 #ifdef CONFIG_TCP_MD5SIG
2558 /* Read the IP->Key mappings from userspace */
2559 err = tp->af_specific->md5_parse(sk, optval, optlen);
2562 case TCP_USER_TIMEOUT:
2563 /* Cap the max time in ms TCP will retry or probe the window
2564 * before giving up and aborting (ETIMEDOUT) a connection.
2569 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2573 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2575 err = fastopen_init_queue(sk, val);
2583 tp->tsoffset = val - tcp_time_stamp;
2585 case TCP_NOTSENT_LOWAT:
2586 tp->notsent_lowat = val;
2587 sk->sk_write_space(sk);
2598 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2599 unsigned int optlen)
2601 const struct inet_connection_sock *icsk = inet_csk(sk);
2603 if (level != SOL_TCP)
2604 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2606 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2608 EXPORT_SYMBOL(tcp_setsockopt);
2610 #ifdef CONFIG_COMPAT
2611 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2612 char __user *optval, unsigned int optlen)
2614 if (level != SOL_TCP)
2615 return inet_csk_compat_setsockopt(sk, level, optname,
2617 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2619 EXPORT_SYMBOL(compat_tcp_setsockopt);
2622 /* Return information about state of tcp endpoint in API format. */
2623 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2625 const struct tcp_sock *tp = tcp_sk(sk);
2626 const struct inet_connection_sock *icsk = inet_csk(sk);
2627 u32 now = tcp_time_stamp;
2630 memset(info, 0, sizeof(*info));
2632 info->tcpi_state = sk->sk_state;
2633 info->tcpi_ca_state = icsk->icsk_ca_state;
2634 info->tcpi_retransmits = icsk->icsk_retransmits;
2635 info->tcpi_probes = icsk->icsk_probes_out;
2636 info->tcpi_backoff = icsk->icsk_backoff;
2638 if (tp->rx_opt.tstamp_ok)
2639 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2640 if (tcp_is_sack(tp))
2641 info->tcpi_options |= TCPI_OPT_SACK;
2642 if (tp->rx_opt.wscale_ok) {
2643 info->tcpi_options |= TCPI_OPT_WSCALE;
2644 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2645 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2648 if (tp->ecn_flags & TCP_ECN_OK)
2649 info->tcpi_options |= TCPI_OPT_ECN;
2650 if (tp->ecn_flags & TCP_ECN_SEEN)
2651 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2652 if (tp->syn_data_acked)
2653 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2655 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2656 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2657 info->tcpi_snd_mss = tp->mss_cache;
2658 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2660 if (sk->sk_state == TCP_LISTEN) {
2661 info->tcpi_unacked = sk->sk_ack_backlog;
2662 info->tcpi_sacked = sk->sk_max_ack_backlog;
2664 info->tcpi_unacked = tp->packets_out;
2665 info->tcpi_sacked = tp->sacked_out;
2667 info->tcpi_lost = tp->lost_out;
2668 info->tcpi_retrans = tp->retrans_out;
2669 info->tcpi_fackets = tp->fackets_out;
2671 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2672 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2673 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2675 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2676 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2677 info->tcpi_rtt = tp->srtt_us >> 3;
2678 info->tcpi_rttvar = tp->mdev_us >> 2;
2679 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2680 info->tcpi_snd_cwnd = tp->snd_cwnd;
2681 info->tcpi_advmss = tp->advmss;
2682 info->tcpi_reordering = tp->reordering;
2684 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2685 info->tcpi_rcv_space = tp->rcvq_space.space;
2687 info->tcpi_total_retrans = tp->total_retrans;
2689 rate = READ_ONCE(sk->sk_pacing_rate);
2690 info->tcpi_pacing_rate = rate != ~0U ? rate : ~0ULL;
2692 rate = READ_ONCE(sk->sk_max_pacing_rate);
2693 info->tcpi_max_pacing_rate = rate != ~0U ? rate : ~0ULL;
2695 spin_lock_bh(&sk->sk_lock.slock);
2696 info->tcpi_bytes_acked = tp->bytes_acked;
2697 info->tcpi_bytes_received = tp->bytes_received;
2698 info->tcpi_segs_out = tp->segs_out;
2699 info->tcpi_segs_in = tp->segs_in;
2700 spin_unlock_bh(&sk->sk_lock.slock);
2702 EXPORT_SYMBOL_GPL(tcp_get_info);
2704 static int do_tcp_getsockopt(struct sock *sk, int level,
2705 int optname, char __user *optval, int __user *optlen)
2707 struct inet_connection_sock *icsk = inet_csk(sk);
2708 struct tcp_sock *tp = tcp_sk(sk);
2711 if (get_user(len, optlen))
2714 len = min_t(unsigned int, len, sizeof(int));
2721 val = tp->mss_cache;
2722 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2723 val = tp->rx_opt.user_mss;
2725 val = tp->rx_opt.mss_clamp;
2728 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2731 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2734 val = keepalive_time_when(tp) / HZ;
2737 val = keepalive_intvl_when(tp) / HZ;
2740 val = keepalive_probes(tp);
2743 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2748 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2750 case TCP_DEFER_ACCEPT:
2751 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2752 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2754 case TCP_WINDOW_CLAMP:
2755 val = tp->window_clamp;
2758 struct tcp_info info;
2760 if (get_user(len, optlen))
2763 tcp_get_info(sk, &info);
2765 len = min_t(unsigned int, len, sizeof(info));
2766 if (put_user(len, optlen))
2768 if (copy_to_user(optval, &info, len))
2773 const struct tcp_congestion_ops *ca_ops;
2774 union tcp_cc_info info;
2778 if (get_user(len, optlen))
2781 ca_ops = icsk->icsk_ca_ops;
2782 if (ca_ops && ca_ops->get_info)
2783 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2785 len = min_t(unsigned int, len, sz);
2786 if (put_user(len, optlen))
2788 if (copy_to_user(optval, &info, len))
2793 val = !icsk->icsk_ack.pingpong;
2796 case TCP_CONGESTION:
2797 if (get_user(len, optlen))
2799 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2800 if (put_user(len, optlen))
2802 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2806 case TCP_THIN_LINEAR_TIMEOUTS:
2809 case TCP_THIN_DUPACK:
2810 val = tp->thin_dupack;
2817 case TCP_REPAIR_QUEUE:
2819 val = tp->repair_queue;
2825 if (tp->repair_queue == TCP_SEND_QUEUE)
2826 val = tp->write_seq;
2827 else if (tp->repair_queue == TCP_RECV_QUEUE)
2833 case TCP_USER_TIMEOUT:
2834 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2838 if (icsk->icsk_accept_queue.fastopenq)
2839 val = icsk->icsk_accept_queue.fastopenq->max_qlen;
2845 val = tcp_time_stamp + tp->tsoffset;
2847 case TCP_NOTSENT_LOWAT:
2848 val = tp->notsent_lowat;
2853 case TCP_SAVED_SYN: {
2854 if (get_user(len, optlen))
2858 if (tp->saved_syn) {
2859 if (len < tp->saved_syn[0]) {
2860 if (put_user(tp->saved_syn[0], optlen)) {
2867 len = tp->saved_syn[0];
2868 if (put_user(len, optlen)) {
2872 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
2876 tcp_saved_syn_free(tp);
2881 if (put_user(len, optlen))
2887 return -ENOPROTOOPT;
2890 if (put_user(len, optlen))
2892 if (copy_to_user(optval, &val, len))
2897 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2900 struct inet_connection_sock *icsk = inet_csk(sk);
2902 if (level != SOL_TCP)
2903 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2905 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2907 EXPORT_SYMBOL(tcp_getsockopt);
2909 #ifdef CONFIG_COMPAT
2910 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2911 char __user *optval, int __user *optlen)
2913 if (level != SOL_TCP)
2914 return inet_csk_compat_getsockopt(sk, level, optname,
2916 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2918 EXPORT_SYMBOL(compat_tcp_getsockopt);
2921 #ifdef CONFIG_TCP_MD5SIG
2922 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2923 static DEFINE_MUTEX(tcp_md5sig_mutex);
2924 static bool tcp_md5sig_pool_populated = false;
2926 static void __tcp_alloc_md5sig_pool(void)
2930 for_each_possible_cpu(cpu) {
2931 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2932 struct crypto_hash *hash;
2934 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2935 if (IS_ERR_OR_NULL(hash))
2937 per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
2940 /* before setting tcp_md5sig_pool_populated, we must commit all writes
2941 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
2944 tcp_md5sig_pool_populated = true;
2947 bool tcp_alloc_md5sig_pool(void)
2949 if (unlikely(!tcp_md5sig_pool_populated)) {
2950 mutex_lock(&tcp_md5sig_mutex);
2952 if (!tcp_md5sig_pool_populated)
2953 __tcp_alloc_md5sig_pool();
2955 mutex_unlock(&tcp_md5sig_mutex);
2957 return tcp_md5sig_pool_populated;
2959 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2963 * tcp_get_md5sig_pool - get md5sig_pool for this user
2965 * We use percpu structure, so if we succeed, we exit with preemption
2966 * and BH disabled, to make sure another thread or softirq handling
2967 * wont try to get same context.
2969 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2973 if (tcp_md5sig_pool_populated) {
2974 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
2976 return this_cpu_ptr(&tcp_md5sig_pool);
2981 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2983 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2984 const struct tcphdr *th)
2986 struct scatterlist sg;
2990 /* We are not allowed to change tcphdr, make a local copy */
2991 memcpy(&hdr, th, sizeof(hdr));
2994 /* options aren't included in the hash */
2995 sg_init_one(&sg, &hdr, sizeof(hdr));
2996 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
2999 EXPORT_SYMBOL(tcp_md5_hash_header);
3001 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3002 const struct sk_buff *skb, unsigned int header_len)
3004 struct scatterlist sg;
3005 const struct tcphdr *tp = tcp_hdr(skb);
3006 struct hash_desc *desc = &hp->md5_desc;
3008 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3009 skb_headlen(skb) - header_len : 0;
3010 const struct skb_shared_info *shi = skb_shinfo(skb);
3011 struct sk_buff *frag_iter;
3013 sg_init_table(&sg, 1);
3015 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3016 if (crypto_hash_update(desc, &sg, head_data_len))
3019 for (i = 0; i < shi->nr_frags; ++i) {
3020 const struct skb_frag_struct *f = &shi->frags[i];
3021 unsigned int offset = f->page_offset;
3022 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3024 sg_set_page(&sg, page, skb_frag_size(f),
3025 offset_in_page(offset));
3026 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3030 skb_walk_frags(skb, frag_iter)
3031 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3036 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3038 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3040 struct scatterlist sg;
3042 sg_init_one(&sg, key->key, key->keylen);
3043 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3045 EXPORT_SYMBOL(tcp_md5_hash_key);
3049 void tcp_done(struct sock *sk)
3051 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3053 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3054 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3056 tcp_set_state(sk, TCP_CLOSE);
3057 tcp_clear_xmit_timers(sk);
3059 reqsk_fastopen_remove(sk, req, false);
3061 sk->sk_shutdown = SHUTDOWN_MASK;
3063 if (!sock_flag(sk, SOCK_DEAD))
3064 sk->sk_state_change(sk);
3066 inet_csk_destroy_sock(sk);
3068 EXPORT_SYMBOL_GPL(tcp_done);
3070 extern struct tcp_congestion_ops tcp_reno;
3072 static __initdata unsigned long thash_entries;
3073 static int __init set_thash_entries(char *str)
3080 ret = kstrtoul(str, 0, &thash_entries);
3086 __setup("thash_entries=", set_thash_entries);
3088 static void __init tcp_init_mem(void)
3090 unsigned long limit = nr_free_buffer_pages() / 16;
3092 limit = max(limit, 128UL);
3093 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3094 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3095 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3098 void __init tcp_init(void)
3100 unsigned long limit;
3101 int max_rshare, max_wshare, cnt;
3104 sock_skb_cb_check_size(sizeof(struct tcp_skb_cb));
3106 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3107 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3108 tcp_hashinfo.bind_bucket_cachep =
3109 kmem_cache_create("tcp_bind_bucket",
3110 sizeof(struct inet_bind_bucket), 0,
3111 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3113 /* Size and allocate the main established and bind bucket
3116 * The methodology is similar to that of the buffer cache.
3118 tcp_hashinfo.ehash =
3119 alloc_large_system_hash("TCP established",
3120 sizeof(struct inet_ehash_bucket),
3122 17, /* one slot per 128 KB of memory */
3125 &tcp_hashinfo.ehash_mask,
3127 thash_entries ? 0 : 512 * 1024);
3128 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3129 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3131 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3132 panic("TCP: failed to alloc ehash_locks");
3133 tcp_hashinfo.bhash =
3134 alloc_large_system_hash("TCP bind",
3135 sizeof(struct inet_bind_hashbucket),
3136 tcp_hashinfo.ehash_mask + 1,
3137 17, /* one slot per 128 KB of memory */
3139 &tcp_hashinfo.bhash_size,
3143 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3144 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3145 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3146 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3150 cnt = tcp_hashinfo.ehash_mask + 1;
3152 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3153 sysctl_tcp_max_orphans = cnt / 2;
3154 sysctl_max_syn_backlog = max(128, cnt / 256);
3157 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3158 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3159 max_wshare = min(4UL*1024*1024, limit);
3160 max_rshare = min(6UL*1024*1024, limit);
3162 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3163 sysctl_tcp_wmem[1] = 16*1024;
3164 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3166 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3167 sysctl_tcp_rmem[1] = 87380;
3168 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3170 pr_info("Hash tables configured (established %u bind %u)\n",
3171 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3174 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);