fix a kmap leak in virtio_console
[linux/fpc-iii.git] / net / ipv4 / tcp.c
blob4475b3bb494d5d126dd844b0159629cd672afbdc
1 /*
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).
8 * Authors: Ross Biro
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>
20 * Fixes:
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
25 * (tcp_err()).
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
36 * unknown sockets.
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
39 * syn rule wrong]
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
45 * escape still
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
49 * facilities
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
54 * bit to skb ops.
55 * Alan Cox : Tidied tcp_data to avoid a potential
56 * nasty.
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
68 * sockets.
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
72 * state ack error.
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
77 * fixes
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
83 * completely
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
91 * (not yet usable)
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
104 * all cases.
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
109 * works now.
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
111 * BSD api.
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
119 * fixed ports.
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
125 * socket close.
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
130 * accept.
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
141 * close.
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
147 * comments.
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
155 * resemble the RFC.
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
160 * generates them.
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
173 * but it's a start!
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
194 * improvement.
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
226 * to shutdown
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/init.h>
256 #include <linux/fs.h>
257 #include <linux/skbuff.h>
258 #include <linux/scatterlist.h>
259 #include <linux/splice.h>
260 #include <linux/net.h>
261 #include <linux/socket.h>
262 #include <linux/random.h>
263 #include <linux/bootmem.h>
264 #include <linux/highmem.h>
265 #include <linux/swap.h>
266 #include <linux/cache.h>
267 #include <linux/err.h>
268 #include <linux/crypto.h>
269 #include <linux/time.h>
270 #include <linux/slab.h>
272 #include <net/icmp.h>
273 #include <net/inet_common.h>
274 #include <net/tcp.h>
275 #include <net/xfrm.h>
276 #include <net/ip.h>
277 #include <net/netdma.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);
311 * TCP splice context
313 struct tcp_splice_state {
314 struct pipe_inode_info *pipe;
315 size_t len;
316 unsigned int flags;
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)
340 u8 res = 0;
342 if (seconds > 0) {
343 int period = timeout;
345 res = 1;
346 while (seconds > period && res < 255) {
347 res++;
348 timeout <<= 1;
349 if (timeout > rto_max)
350 timeout = rto_max;
351 period += timeout;
354 return res;
357 /* Convert retransmits to seconds based on initial and max timeout */
358 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
360 int period = 0;
362 if (retrans > 0) {
363 period = timeout;
364 while (--retrans) {
365 timeout <<= 1;
366 if (timeout > rto_max)
367 timeout = rto_max;
368 period += timeout;
371 return period;
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 = 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 icsk->icsk_ca_ops = &tcp_init_congestion_ops;
410 tp->tsoffset = 0;
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];
422 local_bh_disable();
423 sock_update_memcg(sk);
424 sk_sockets_allocated_inc(sk);
425 local_bh_enable();
427 EXPORT_SYMBOL(tcp_init_sock);
430 * Wait for a TCP event.
432 * Note that we don't need to lock the socket, as the upper poll layers
433 * take care of normal races (between the test and the event) and we don't
434 * go look at any of the socket buffers directly.
436 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
438 unsigned int mask;
439 struct sock *sk = sock->sk;
440 const struct tcp_sock *tp = tcp_sk(sk);
442 sock_rps_record_flow(sk);
444 sock_poll_wait(file, sk_sleep(sk), wait);
445 if (sk->sk_state == TCP_LISTEN)
446 return inet_csk_listen_poll(sk);
448 /* Socket is not locked. We are protected from async events
449 * by poll logic and correct handling of state changes
450 * made by other threads is impossible in any case.
453 mask = 0;
456 * POLLHUP is certainly not done right. But poll() doesn't
457 * have a notion of HUP in just one direction, and for a
458 * socket the read side is more interesting.
460 * Some poll() documentation says that POLLHUP is incompatible
461 * with the POLLOUT/POLLWR flags, so somebody should check this
462 * all. But careful, it tends to be safer to return too many
463 * bits than too few, and you can easily break real applications
464 * if you don't tell them that something has hung up!
466 * Check-me.
468 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
469 * our fs/select.c). It means that after we received EOF,
470 * poll always returns immediately, making impossible poll() on write()
471 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
472 * if and only if shutdown has been made in both directions.
473 * Actually, it is interesting to look how Solaris and DUX
474 * solve this dilemma. I would prefer, if POLLHUP were maskable,
475 * then we could set it on SND_SHUTDOWN. BTW examples given
476 * in Stevens' books assume exactly this behaviour, it explains
477 * why POLLHUP is incompatible with POLLOUT. --ANK
479 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
480 * blocking on fresh not-connected or disconnected socket. --ANK
482 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
483 mask |= POLLHUP;
484 if (sk->sk_shutdown & RCV_SHUTDOWN)
485 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
487 /* Connected or passive Fast Open socket? */
488 if (sk->sk_state != TCP_SYN_SENT &&
489 (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
490 int target = sock_rcvlowat(sk, 0, INT_MAX);
492 if (tp->urg_seq == tp->copied_seq &&
493 !sock_flag(sk, SOCK_URGINLINE) &&
494 tp->urg_data)
495 target++;
497 /* Potential race condition. If read of tp below will
498 * escape above sk->sk_state, we can be illegally awaken
499 * in SYN_* states. */
500 if (tp->rcv_nxt - tp->copied_seq >= target)
501 mask |= POLLIN | POLLRDNORM;
503 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
504 if (sk_stream_is_writeable(sk)) {
505 mask |= POLLOUT | POLLWRNORM;
506 } else { /* send SIGIO later */
507 set_bit(SOCK_ASYNC_NOSPACE,
508 &sk->sk_socket->flags);
509 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
511 /* Race breaker. If space is freed after
512 * wspace test but before the flags are set,
513 * IO signal will be lost.
515 if (sk_stream_is_writeable(sk))
516 mask |= POLLOUT | POLLWRNORM;
518 } else
519 mask |= POLLOUT | POLLWRNORM;
521 if (tp->urg_data & TCP_URG_VALID)
522 mask |= POLLPRI;
524 /* This barrier is coupled with smp_wmb() in tcp_reset() */
525 smp_rmb();
526 if (sk->sk_err)
527 mask |= POLLERR;
529 return mask;
531 EXPORT_SYMBOL(tcp_poll);
533 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
535 struct tcp_sock *tp = tcp_sk(sk);
536 int answ;
537 bool slow;
539 switch (cmd) {
540 case SIOCINQ:
541 if (sk->sk_state == TCP_LISTEN)
542 return -EINVAL;
544 slow = lock_sock_fast(sk);
545 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
546 answ = 0;
547 else if (sock_flag(sk, SOCK_URGINLINE) ||
548 !tp->urg_data ||
549 before(tp->urg_seq, tp->copied_seq) ||
550 !before(tp->urg_seq, tp->rcv_nxt)) {
552 answ = tp->rcv_nxt - tp->copied_seq;
554 /* Subtract 1, if FIN was received */
555 if (answ && sock_flag(sk, SOCK_DONE))
556 answ--;
557 } else
558 answ = tp->urg_seq - tp->copied_seq;
559 unlock_sock_fast(sk, slow);
560 break;
561 case SIOCATMARK:
562 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
563 break;
564 case SIOCOUTQ:
565 if (sk->sk_state == TCP_LISTEN)
566 return -EINVAL;
568 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
569 answ = 0;
570 else
571 answ = tp->write_seq - tp->snd_una;
572 break;
573 case SIOCOUTQNSD:
574 if (sk->sk_state == TCP_LISTEN)
575 return -EINVAL;
577 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
578 answ = 0;
579 else
580 answ = tp->write_seq - tp->snd_nxt;
581 break;
582 default:
583 return -ENOIOCTLCMD;
586 return put_user(answ, (int __user *)arg);
588 EXPORT_SYMBOL(tcp_ioctl);
590 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
592 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
593 tp->pushed_seq = tp->write_seq;
596 static inline bool forced_push(const struct tcp_sock *tp)
598 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
601 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
603 struct tcp_sock *tp = tcp_sk(sk);
604 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
606 skb->csum = 0;
607 tcb->seq = tcb->end_seq = tp->write_seq;
608 tcb->tcp_flags = TCPHDR_ACK;
609 tcb->sacked = 0;
610 skb_header_release(skb);
611 tcp_add_write_queue_tail(sk, skb);
612 sk->sk_wmem_queued += skb->truesize;
613 sk_mem_charge(sk, skb->truesize);
614 if (tp->nonagle & TCP_NAGLE_PUSH)
615 tp->nonagle &= ~TCP_NAGLE_PUSH;
618 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
620 if (flags & MSG_OOB)
621 tp->snd_up = tp->write_seq;
624 /* If a not yet filled skb is pushed, do not send it if
625 * we have data packets in Qdisc or NIC queues :
626 * Because TX completion will happen shortly, it gives a chance
627 * to coalesce future sendmsg() payload into this skb, without
628 * need for a timer, and with no latency trade off.
629 * As packets containing data payload have a bigger truesize
630 * than pure acks (dataless) packets, the last checks prevent
631 * autocorking if we only have an ACK in Qdisc/NIC queues,
632 * or if TX completion was delayed after we processed ACK packet.
634 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
635 int size_goal)
637 return skb->len < size_goal &&
638 sysctl_tcp_autocorking &&
639 skb != tcp_write_queue_head(sk) &&
640 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
643 static void tcp_push(struct sock *sk, int flags, int mss_now,
644 int nonagle, int size_goal)
646 struct tcp_sock *tp = tcp_sk(sk);
647 struct sk_buff *skb;
649 if (!tcp_send_head(sk))
650 return;
652 skb = tcp_write_queue_tail(sk);
653 if (!(flags & MSG_MORE) || forced_push(tp))
654 tcp_mark_push(tp, skb);
656 tcp_mark_urg(tp, flags);
658 if (tcp_should_autocork(sk, skb, size_goal)) {
660 /* avoid atomic op if TSQ_THROTTLED bit is already set */
661 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
662 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
663 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
665 /* It is possible TX completion already happened
666 * before we set TSQ_THROTTLED.
668 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
669 return;
672 if (flags & MSG_MORE)
673 nonagle = TCP_NAGLE_CORK;
675 __tcp_push_pending_frames(sk, mss_now, nonagle);
678 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
679 unsigned int offset, size_t len)
681 struct tcp_splice_state *tss = rd_desc->arg.data;
682 int ret;
684 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
685 tss->flags);
686 if (ret > 0)
687 rd_desc->count -= ret;
688 return ret;
691 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
693 /* Store TCP splice context information in read_descriptor_t. */
694 read_descriptor_t rd_desc = {
695 .arg.data = tss,
696 .count = tss->len,
699 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
703 * tcp_splice_read - splice data from TCP socket to a pipe
704 * @sock: socket to splice from
705 * @ppos: position (not valid)
706 * @pipe: pipe to splice to
707 * @len: number of bytes to splice
708 * @flags: splice modifier flags
710 * Description:
711 * Will read pages from given socket and fill them into a pipe.
714 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
715 struct pipe_inode_info *pipe, size_t len,
716 unsigned int flags)
718 struct sock *sk = sock->sk;
719 struct tcp_splice_state tss = {
720 .pipe = pipe,
721 .len = len,
722 .flags = flags,
724 long timeo;
725 ssize_t spliced;
726 int ret;
728 sock_rps_record_flow(sk);
730 * We can't seek on a socket input
732 if (unlikely(*ppos))
733 return -ESPIPE;
735 ret = spliced = 0;
737 lock_sock(sk);
739 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
740 while (tss.len) {
741 ret = __tcp_splice_read(sk, &tss);
742 if (ret < 0)
743 break;
744 else if (!ret) {
745 if (spliced)
746 break;
747 if (sock_flag(sk, SOCK_DONE))
748 break;
749 if (sk->sk_err) {
750 ret = sock_error(sk);
751 break;
753 if (sk->sk_shutdown & RCV_SHUTDOWN)
754 break;
755 if (sk->sk_state == TCP_CLOSE) {
757 * This occurs when user tries to read
758 * from never connected socket.
760 if (!sock_flag(sk, SOCK_DONE))
761 ret = -ENOTCONN;
762 break;
764 if (!timeo) {
765 ret = -EAGAIN;
766 break;
768 sk_wait_data(sk, &timeo);
769 if (signal_pending(current)) {
770 ret = sock_intr_errno(timeo);
771 break;
773 continue;
775 tss.len -= ret;
776 spliced += ret;
778 if (!timeo)
779 break;
780 release_sock(sk);
781 lock_sock(sk);
783 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
784 (sk->sk_shutdown & RCV_SHUTDOWN) ||
785 signal_pending(current))
786 break;
789 release_sock(sk);
791 if (spliced)
792 return spliced;
794 return ret;
796 EXPORT_SYMBOL(tcp_splice_read);
798 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
800 struct sk_buff *skb;
802 /* The TCP header must be at least 32-bit aligned. */
803 size = ALIGN(size, 4);
805 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
806 if (skb) {
807 if (sk_wmem_schedule(sk, skb->truesize)) {
808 skb_reserve(skb, sk->sk_prot->max_header);
810 * Make sure that we have exactly size bytes
811 * available to the caller, no more, no less.
813 skb->reserved_tailroom = skb->end - skb->tail - size;
814 return skb;
816 __kfree_skb(skb);
817 } else {
818 sk->sk_prot->enter_memory_pressure(sk);
819 sk_stream_moderate_sndbuf(sk);
821 return NULL;
824 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
825 int large_allowed)
827 struct tcp_sock *tp = tcp_sk(sk);
828 u32 xmit_size_goal, old_size_goal;
830 xmit_size_goal = mss_now;
832 if (large_allowed && sk_can_gso(sk)) {
833 u32 gso_size, hlen;
835 /* Maybe we should/could use sk->sk_prot->max_header here ? */
836 hlen = inet_csk(sk)->icsk_af_ops->net_header_len +
837 inet_csk(sk)->icsk_ext_hdr_len +
838 tp->tcp_header_len;
840 /* Goal is to send at least one packet per ms,
841 * not one big TSO packet every 100 ms.
842 * This preserves ACK clocking and is consistent
843 * with tcp_tso_should_defer() heuristic.
845 gso_size = sk->sk_pacing_rate / (2 * MSEC_PER_SEC);
846 gso_size = max_t(u32, gso_size,
847 sysctl_tcp_min_tso_segs * mss_now);
849 xmit_size_goal = min_t(u32, gso_size,
850 sk->sk_gso_max_size - 1 - hlen);
852 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
854 /* We try hard to avoid divides here */
855 old_size_goal = tp->xmit_size_goal_segs * mss_now;
857 if (likely(old_size_goal <= xmit_size_goal &&
858 old_size_goal + mss_now > xmit_size_goal)) {
859 xmit_size_goal = old_size_goal;
860 } else {
861 tp->xmit_size_goal_segs =
862 min_t(u16, xmit_size_goal / mss_now,
863 sk->sk_gso_max_segs);
864 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
868 return max(xmit_size_goal, mss_now);
871 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
873 int mss_now;
875 mss_now = tcp_current_mss(sk);
876 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
878 return mss_now;
881 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
882 size_t size, int flags)
884 struct tcp_sock *tp = tcp_sk(sk);
885 int mss_now, size_goal;
886 int err;
887 ssize_t copied;
888 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
890 /* Wait for a connection to finish. One exception is TCP Fast Open
891 * (passive side) where data is allowed to be sent before a connection
892 * is fully established.
894 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
895 !tcp_passive_fastopen(sk)) {
896 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
897 goto out_err;
900 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
902 mss_now = tcp_send_mss(sk, &size_goal, flags);
903 copied = 0;
905 err = -EPIPE;
906 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
907 goto out_err;
909 while (size > 0) {
910 struct sk_buff *skb = tcp_write_queue_tail(sk);
911 int copy, i;
912 bool can_coalesce;
914 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
915 new_segment:
916 if (!sk_stream_memory_free(sk))
917 goto wait_for_sndbuf;
919 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
920 if (!skb)
921 goto wait_for_memory;
923 skb_entail(sk, skb);
924 copy = size_goal;
927 if (copy > size)
928 copy = size;
930 i = skb_shinfo(skb)->nr_frags;
931 can_coalesce = skb_can_coalesce(skb, i, page, offset);
932 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
933 tcp_mark_push(tp, skb);
934 goto new_segment;
936 if (!sk_wmem_schedule(sk, copy))
937 goto wait_for_memory;
939 if (can_coalesce) {
940 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
941 } else {
942 get_page(page);
943 skb_fill_page_desc(skb, i, page, offset, copy);
945 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
947 skb->len += copy;
948 skb->data_len += copy;
949 skb->truesize += copy;
950 sk->sk_wmem_queued += copy;
951 sk_mem_charge(sk, copy);
952 skb->ip_summed = CHECKSUM_PARTIAL;
953 tp->write_seq += copy;
954 TCP_SKB_CB(skb)->end_seq += copy;
955 skb_shinfo(skb)->gso_segs = 0;
957 if (!copied)
958 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
960 copied += copy;
961 offset += copy;
962 if (!(size -= copy))
963 goto out;
965 if (skb->len < size_goal || (flags & MSG_OOB))
966 continue;
968 if (forced_push(tp)) {
969 tcp_mark_push(tp, skb);
970 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
971 } else if (skb == tcp_send_head(sk))
972 tcp_push_one(sk, mss_now);
973 continue;
975 wait_for_sndbuf:
976 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
977 wait_for_memory:
978 tcp_push(sk, flags & ~MSG_MORE, mss_now,
979 TCP_NAGLE_PUSH, size_goal);
981 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
982 goto do_error;
984 mss_now = tcp_send_mss(sk, &size_goal, flags);
987 out:
988 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
989 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
990 return copied;
992 do_error:
993 if (copied)
994 goto out;
995 out_err:
996 return sk_stream_error(sk, flags, err);
999 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1000 size_t size, int flags)
1002 ssize_t res;
1004 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1005 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
1006 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1007 flags);
1009 lock_sock(sk);
1010 res = do_tcp_sendpages(sk, page, offset, size, flags);
1011 release_sock(sk);
1012 return res;
1014 EXPORT_SYMBOL(tcp_sendpage);
1016 static inline int select_size(const struct sock *sk, bool sg)
1018 const struct tcp_sock *tp = tcp_sk(sk);
1019 int tmp = tp->mss_cache;
1021 if (sg) {
1022 if (sk_can_gso(sk)) {
1023 /* Small frames wont use a full page:
1024 * Payload will immediately follow tcp header.
1026 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1027 } else {
1028 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1030 if (tmp >= pgbreak &&
1031 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1032 tmp = pgbreak;
1036 return tmp;
1039 void tcp_free_fastopen_req(struct tcp_sock *tp)
1041 if (tp->fastopen_req != NULL) {
1042 kfree(tp->fastopen_req);
1043 tp->fastopen_req = NULL;
1047 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size)
1049 struct tcp_sock *tp = tcp_sk(sk);
1050 int err, flags;
1052 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1053 return -EOPNOTSUPP;
1054 if (tp->fastopen_req != NULL)
1055 return -EALREADY; /* Another Fast Open is in progress */
1057 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1058 sk->sk_allocation);
1059 if (unlikely(tp->fastopen_req == NULL))
1060 return -ENOBUFS;
1061 tp->fastopen_req->data = msg;
1063 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1064 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1065 msg->msg_namelen, flags);
1066 *size = tp->fastopen_req->copied;
1067 tcp_free_fastopen_req(tp);
1068 return err;
1071 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1072 size_t size)
1074 struct iovec *iov;
1075 struct tcp_sock *tp = tcp_sk(sk);
1076 struct sk_buff *skb;
1077 int iovlen, flags, err, copied = 0;
1078 int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1079 bool sg;
1080 long timeo;
1082 lock_sock(sk);
1084 flags = msg->msg_flags;
1085 if (flags & MSG_FASTOPEN) {
1086 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn);
1087 if (err == -EINPROGRESS && copied_syn > 0)
1088 goto out;
1089 else if (err)
1090 goto out_err;
1091 offset = copied_syn;
1094 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1096 /* Wait for a connection to finish. One exception is TCP Fast Open
1097 * (passive side) where data is allowed to be sent before a connection
1098 * is fully established.
1100 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1101 !tcp_passive_fastopen(sk)) {
1102 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1103 goto do_error;
1106 if (unlikely(tp->repair)) {
1107 if (tp->repair_queue == TCP_RECV_QUEUE) {
1108 copied = tcp_send_rcvq(sk, msg, size);
1109 goto out;
1112 err = -EINVAL;
1113 if (tp->repair_queue == TCP_NO_QUEUE)
1114 goto out_err;
1116 /* 'common' sending to sendq */
1119 /* This should be in poll */
1120 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1122 mss_now = tcp_send_mss(sk, &size_goal, flags);
1124 /* Ok commence sending. */
1125 iovlen = msg->msg_iovlen;
1126 iov = msg->msg_iov;
1127 copied = 0;
1129 err = -EPIPE;
1130 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1131 goto out_err;
1133 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1135 while (--iovlen >= 0) {
1136 size_t seglen = iov->iov_len;
1137 unsigned char __user *from = iov->iov_base;
1139 iov++;
1140 if (unlikely(offset > 0)) { /* Skip bytes copied in SYN */
1141 if (offset >= seglen) {
1142 offset -= seglen;
1143 continue;
1145 seglen -= offset;
1146 from += offset;
1147 offset = 0;
1150 while (seglen > 0) {
1151 int copy = 0;
1152 int max = size_goal;
1154 skb = tcp_write_queue_tail(sk);
1155 if (tcp_send_head(sk)) {
1156 if (skb->ip_summed == CHECKSUM_NONE)
1157 max = mss_now;
1158 copy = max - skb->len;
1161 if (copy <= 0) {
1162 new_segment:
1163 /* Allocate new segment. If the interface is SG,
1164 * allocate skb fitting to single page.
1166 if (!sk_stream_memory_free(sk))
1167 goto wait_for_sndbuf;
1169 skb = sk_stream_alloc_skb(sk,
1170 select_size(sk, sg),
1171 sk->sk_allocation);
1172 if (!skb)
1173 goto wait_for_memory;
1176 * All packets are restored as if they have
1177 * already been sent.
1179 if (tp->repair)
1180 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1183 * Check whether we can use HW checksum.
1185 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1186 skb->ip_summed = CHECKSUM_PARTIAL;
1188 skb_entail(sk, skb);
1189 copy = size_goal;
1190 max = size_goal;
1193 /* Try to append data to the end of skb. */
1194 if (copy > seglen)
1195 copy = seglen;
1197 /* Where to copy to? */
1198 if (skb_availroom(skb) > 0) {
1199 /* We have some space in skb head. Superb! */
1200 copy = min_t(int, copy, skb_availroom(skb));
1201 err = skb_add_data_nocache(sk, skb, from, copy);
1202 if (err)
1203 goto do_fault;
1204 } else {
1205 bool merge = true;
1206 int i = skb_shinfo(skb)->nr_frags;
1207 struct page_frag *pfrag = sk_page_frag(sk);
1209 if (!sk_page_frag_refill(sk, pfrag))
1210 goto wait_for_memory;
1212 if (!skb_can_coalesce(skb, i, pfrag->page,
1213 pfrag->offset)) {
1214 if (i == MAX_SKB_FRAGS || !sg) {
1215 tcp_mark_push(tp, skb);
1216 goto new_segment;
1218 merge = false;
1221 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1223 if (!sk_wmem_schedule(sk, copy))
1224 goto wait_for_memory;
1226 err = skb_copy_to_page_nocache(sk, from, skb,
1227 pfrag->page,
1228 pfrag->offset,
1229 copy);
1230 if (err)
1231 goto do_error;
1233 /* Update the skb. */
1234 if (merge) {
1235 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1236 } else {
1237 skb_fill_page_desc(skb, i, pfrag->page,
1238 pfrag->offset, copy);
1239 get_page(pfrag->page);
1241 pfrag->offset += copy;
1244 if (!copied)
1245 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1247 tp->write_seq += copy;
1248 TCP_SKB_CB(skb)->end_seq += copy;
1249 skb_shinfo(skb)->gso_segs = 0;
1251 from += copy;
1252 copied += copy;
1253 if ((seglen -= copy) == 0 && iovlen == 0)
1254 goto out;
1256 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1257 continue;
1259 if (forced_push(tp)) {
1260 tcp_mark_push(tp, skb);
1261 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1262 } else if (skb == tcp_send_head(sk))
1263 tcp_push_one(sk, mss_now);
1264 continue;
1266 wait_for_sndbuf:
1267 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1268 wait_for_memory:
1269 if (copied)
1270 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1271 TCP_NAGLE_PUSH, size_goal);
1273 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1274 goto do_error;
1276 mss_now = tcp_send_mss(sk, &size_goal, flags);
1280 out:
1281 if (copied)
1282 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1283 release_sock(sk);
1284 return copied + copied_syn;
1286 do_fault:
1287 if (!skb->len) {
1288 tcp_unlink_write_queue(skb, sk);
1289 /* It is the one place in all of TCP, except connection
1290 * reset, where we can be unlinking the send_head.
1292 tcp_check_send_head(sk, skb);
1293 sk_wmem_free_skb(sk, skb);
1296 do_error:
1297 if (copied + copied_syn)
1298 goto out;
1299 out_err:
1300 err = sk_stream_error(sk, flags, err);
1301 release_sock(sk);
1302 return err;
1304 EXPORT_SYMBOL(tcp_sendmsg);
1307 * Handle reading urgent data. BSD has very simple semantics for
1308 * this, no blocking and very strange errors 8)
1311 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1313 struct tcp_sock *tp = tcp_sk(sk);
1315 /* No URG data to read. */
1316 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1317 tp->urg_data == TCP_URG_READ)
1318 return -EINVAL; /* Yes this is right ! */
1320 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1321 return -ENOTCONN;
1323 if (tp->urg_data & TCP_URG_VALID) {
1324 int err = 0;
1325 char c = tp->urg_data;
1327 if (!(flags & MSG_PEEK))
1328 tp->urg_data = TCP_URG_READ;
1330 /* Read urgent data. */
1331 msg->msg_flags |= MSG_OOB;
1333 if (len > 0) {
1334 if (!(flags & MSG_TRUNC))
1335 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1336 len = 1;
1337 } else
1338 msg->msg_flags |= MSG_TRUNC;
1340 return err ? -EFAULT : len;
1343 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1344 return 0;
1346 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1347 * the available implementations agree in this case:
1348 * this call should never block, independent of the
1349 * blocking state of the socket.
1350 * Mike <pall@rz.uni-karlsruhe.de>
1352 return -EAGAIN;
1355 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1357 struct sk_buff *skb;
1358 int copied = 0, err = 0;
1360 /* XXX -- need to support SO_PEEK_OFF */
1362 skb_queue_walk(&sk->sk_write_queue, skb) {
1363 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1364 if (err)
1365 break;
1367 copied += skb->len;
1370 return err ?: copied;
1373 /* Clean up the receive buffer for full frames taken by the user,
1374 * then send an ACK if necessary. COPIED is the number of bytes
1375 * tcp_recvmsg has given to the user so far, it speeds up the
1376 * calculation of whether or not we must ACK for the sake of
1377 * a window update.
1379 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1381 struct tcp_sock *tp = tcp_sk(sk);
1382 bool time_to_ack = false;
1384 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1386 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1387 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1388 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1390 if (inet_csk_ack_scheduled(sk)) {
1391 const struct inet_connection_sock *icsk = inet_csk(sk);
1392 /* Delayed ACKs frequently hit locked sockets during bulk
1393 * receive. */
1394 if (icsk->icsk_ack.blocked ||
1395 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1396 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1398 * If this read emptied read buffer, we send ACK, if
1399 * connection is not bidirectional, user drained
1400 * receive buffer and there was a small segment
1401 * in queue.
1403 (copied > 0 &&
1404 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1405 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1406 !icsk->icsk_ack.pingpong)) &&
1407 !atomic_read(&sk->sk_rmem_alloc)))
1408 time_to_ack = true;
1411 /* We send an ACK if we can now advertise a non-zero window
1412 * which has been raised "significantly".
1414 * Even if window raised up to infinity, do not send window open ACK
1415 * in states, where we will not receive more. It is useless.
1417 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1418 __u32 rcv_window_now = tcp_receive_window(tp);
1420 /* Optimize, __tcp_select_window() is not cheap. */
1421 if (2*rcv_window_now <= tp->window_clamp) {
1422 __u32 new_window = __tcp_select_window(sk);
1424 /* Send ACK now, if this read freed lots of space
1425 * in our buffer. Certainly, new_window is new window.
1426 * We can advertise it now, if it is not less than current one.
1427 * "Lots" means "at least twice" here.
1429 if (new_window && new_window >= 2 * rcv_window_now)
1430 time_to_ack = true;
1433 if (time_to_ack)
1434 tcp_send_ack(sk);
1437 static void tcp_prequeue_process(struct sock *sk)
1439 struct sk_buff *skb;
1440 struct tcp_sock *tp = tcp_sk(sk);
1442 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1444 /* RX process wants to run with disabled BHs, though it is not
1445 * necessary */
1446 local_bh_disable();
1447 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1448 sk_backlog_rcv(sk, skb);
1449 local_bh_enable();
1451 /* Clear memory counter. */
1452 tp->ucopy.memory = 0;
1455 #ifdef CONFIG_NET_DMA
1456 static void tcp_service_net_dma(struct sock *sk, bool wait)
1458 dma_cookie_t done, used;
1459 dma_cookie_t last_issued;
1460 struct tcp_sock *tp = tcp_sk(sk);
1462 if (!tp->ucopy.dma_chan)
1463 return;
1465 last_issued = tp->ucopy.dma_cookie;
1466 dma_async_issue_pending(tp->ucopy.dma_chan);
1468 do {
1469 if (dma_async_is_tx_complete(tp->ucopy.dma_chan,
1470 last_issued, &done,
1471 &used) == DMA_COMPLETE) {
1472 /* Safe to free early-copied skbs now */
1473 __skb_queue_purge(&sk->sk_async_wait_queue);
1474 break;
1475 } else {
1476 struct sk_buff *skb;
1477 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1478 (dma_async_is_complete(skb->dma_cookie, done,
1479 used) == DMA_COMPLETE)) {
1480 __skb_dequeue(&sk->sk_async_wait_queue);
1481 kfree_skb(skb);
1484 } while (wait);
1486 #endif
1488 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1490 struct sk_buff *skb;
1491 u32 offset;
1493 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1494 offset = seq - TCP_SKB_CB(skb)->seq;
1495 if (tcp_hdr(skb)->syn)
1496 offset--;
1497 if (offset < skb->len || tcp_hdr(skb)->fin) {
1498 *off = offset;
1499 return skb;
1501 /* This looks weird, but this can happen if TCP collapsing
1502 * splitted a fat GRO packet, while we released socket lock
1503 * in skb_splice_bits()
1505 sk_eat_skb(sk, skb, false);
1507 return NULL;
1511 * This routine provides an alternative to tcp_recvmsg() for routines
1512 * that would like to handle copying from skbuffs directly in 'sendfile'
1513 * fashion.
1514 * Note:
1515 * - It is assumed that the socket was locked by the caller.
1516 * - The routine does not block.
1517 * - At present, there is no support for reading OOB data
1518 * or for 'peeking' the socket using this routine
1519 * (although both would be easy to implement).
1521 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1522 sk_read_actor_t recv_actor)
1524 struct sk_buff *skb;
1525 struct tcp_sock *tp = tcp_sk(sk);
1526 u32 seq = tp->copied_seq;
1527 u32 offset;
1528 int copied = 0;
1530 if (sk->sk_state == TCP_LISTEN)
1531 return -ENOTCONN;
1532 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1533 if (offset < skb->len) {
1534 int used;
1535 size_t len;
1537 len = skb->len - offset;
1538 /* Stop reading if we hit a patch of urgent data */
1539 if (tp->urg_data) {
1540 u32 urg_offset = tp->urg_seq - seq;
1541 if (urg_offset < len)
1542 len = urg_offset;
1543 if (!len)
1544 break;
1546 used = recv_actor(desc, skb, offset, len);
1547 if (used <= 0) {
1548 if (!copied)
1549 copied = used;
1550 break;
1551 } else if (used <= len) {
1552 seq += used;
1553 copied += used;
1554 offset += used;
1556 /* If recv_actor drops the lock (e.g. TCP splice
1557 * receive) the skb pointer might be invalid when
1558 * getting here: tcp_collapse might have deleted it
1559 * while aggregating skbs from the socket queue.
1561 skb = tcp_recv_skb(sk, seq - 1, &offset);
1562 if (!skb)
1563 break;
1564 /* TCP coalescing might have appended data to the skb.
1565 * Try to splice more frags
1567 if (offset + 1 != skb->len)
1568 continue;
1570 if (tcp_hdr(skb)->fin) {
1571 sk_eat_skb(sk, skb, false);
1572 ++seq;
1573 break;
1575 sk_eat_skb(sk, skb, false);
1576 if (!desc->count)
1577 break;
1578 tp->copied_seq = seq;
1580 tp->copied_seq = seq;
1582 tcp_rcv_space_adjust(sk);
1584 /* Clean up data we have read: This will do ACK frames. */
1585 if (copied > 0) {
1586 tcp_recv_skb(sk, seq, &offset);
1587 tcp_cleanup_rbuf(sk, copied);
1589 return copied;
1591 EXPORT_SYMBOL(tcp_read_sock);
1594 * This routine copies from a sock struct into the user buffer.
1596 * Technical note: in 2.3 we work on _locked_ socket, so that
1597 * tricks with *seq access order and skb->users are not required.
1598 * Probably, code can be easily improved even more.
1601 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1602 size_t len, int nonblock, int flags, int *addr_len)
1604 struct tcp_sock *tp = tcp_sk(sk);
1605 int copied = 0;
1606 u32 peek_seq;
1607 u32 *seq;
1608 unsigned long used;
1609 int err;
1610 int target; /* Read at least this many bytes */
1611 long timeo;
1612 struct task_struct *user_recv = NULL;
1613 bool copied_early = false;
1614 struct sk_buff *skb;
1615 u32 urg_hole = 0;
1617 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1618 (sk->sk_state == TCP_ESTABLISHED))
1619 sk_busy_loop(sk, nonblock);
1621 lock_sock(sk);
1623 err = -ENOTCONN;
1624 if (sk->sk_state == TCP_LISTEN)
1625 goto out;
1627 timeo = sock_rcvtimeo(sk, nonblock);
1629 /* Urgent data needs to be handled specially. */
1630 if (flags & MSG_OOB)
1631 goto recv_urg;
1633 if (unlikely(tp->repair)) {
1634 err = -EPERM;
1635 if (!(flags & MSG_PEEK))
1636 goto out;
1638 if (tp->repair_queue == TCP_SEND_QUEUE)
1639 goto recv_sndq;
1641 err = -EINVAL;
1642 if (tp->repair_queue == TCP_NO_QUEUE)
1643 goto out;
1645 /* 'common' recv queue MSG_PEEK-ing */
1648 seq = &tp->copied_seq;
1649 if (flags & MSG_PEEK) {
1650 peek_seq = tp->copied_seq;
1651 seq = &peek_seq;
1654 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1656 #ifdef CONFIG_NET_DMA
1657 tp->ucopy.dma_chan = NULL;
1658 preempt_disable();
1659 skb = skb_peek_tail(&sk->sk_receive_queue);
1661 int available = 0;
1663 if (skb)
1664 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1665 if ((available < target) &&
1666 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1667 !sysctl_tcp_low_latency &&
1668 net_dma_find_channel()) {
1669 preempt_enable();
1670 tp->ucopy.pinned_list =
1671 dma_pin_iovec_pages(msg->msg_iov, len);
1672 } else {
1673 preempt_enable();
1676 #endif
1678 do {
1679 u32 offset;
1681 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1682 if (tp->urg_data && tp->urg_seq == *seq) {
1683 if (copied)
1684 break;
1685 if (signal_pending(current)) {
1686 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1687 break;
1691 /* Next get a buffer. */
1693 skb_queue_walk(&sk->sk_receive_queue, skb) {
1694 /* Now that we have two receive queues this
1695 * shouldn't happen.
1697 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1698 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1699 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1700 flags))
1701 break;
1703 offset = *seq - TCP_SKB_CB(skb)->seq;
1704 if (tcp_hdr(skb)->syn)
1705 offset--;
1706 if (offset < skb->len)
1707 goto found_ok_skb;
1708 if (tcp_hdr(skb)->fin)
1709 goto found_fin_ok;
1710 WARN(!(flags & MSG_PEEK),
1711 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1712 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1715 /* Well, if we have backlog, try to process it now yet. */
1717 if (copied >= target && !sk->sk_backlog.tail)
1718 break;
1720 if (copied) {
1721 if (sk->sk_err ||
1722 sk->sk_state == TCP_CLOSE ||
1723 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1724 !timeo ||
1725 signal_pending(current))
1726 break;
1727 } else {
1728 if (sock_flag(sk, SOCK_DONE))
1729 break;
1731 if (sk->sk_err) {
1732 copied = sock_error(sk);
1733 break;
1736 if (sk->sk_shutdown & RCV_SHUTDOWN)
1737 break;
1739 if (sk->sk_state == TCP_CLOSE) {
1740 if (!sock_flag(sk, SOCK_DONE)) {
1741 /* This occurs when user tries to read
1742 * from never connected socket.
1744 copied = -ENOTCONN;
1745 break;
1747 break;
1750 if (!timeo) {
1751 copied = -EAGAIN;
1752 break;
1755 if (signal_pending(current)) {
1756 copied = sock_intr_errno(timeo);
1757 break;
1761 tcp_cleanup_rbuf(sk, copied);
1763 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1764 /* Install new reader */
1765 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1766 user_recv = current;
1767 tp->ucopy.task = user_recv;
1768 tp->ucopy.iov = msg->msg_iov;
1771 tp->ucopy.len = len;
1773 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1774 !(flags & (MSG_PEEK | MSG_TRUNC)));
1776 /* Ugly... If prequeue is not empty, we have to
1777 * process it before releasing socket, otherwise
1778 * order will be broken at second iteration.
1779 * More elegant solution is required!!!
1781 * Look: we have the following (pseudo)queues:
1783 * 1. packets in flight
1784 * 2. backlog
1785 * 3. prequeue
1786 * 4. receive_queue
1788 * Each queue can be processed only if the next ones
1789 * are empty. At this point we have empty receive_queue.
1790 * But prequeue _can_ be not empty after 2nd iteration,
1791 * when we jumped to start of loop because backlog
1792 * processing added something to receive_queue.
1793 * We cannot release_sock(), because backlog contains
1794 * packets arrived _after_ prequeued ones.
1796 * Shortly, algorithm is clear --- to process all
1797 * the queues in order. We could make it more directly,
1798 * requeueing packets from backlog to prequeue, if
1799 * is not empty. It is more elegant, but eats cycles,
1800 * unfortunately.
1802 if (!skb_queue_empty(&tp->ucopy.prequeue))
1803 goto do_prequeue;
1805 /* __ Set realtime policy in scheduler __ */
1808 #ifdef CONFIG_NET_DMA
1809 if (tp->ucopy.dma_chan) {
1810 if (tp->rcv_wnd == 0 &&
1811 !skb_queue_empty(&sk->sk_async_wait_queue)) {
1812 tcp_service_net_dma(sk, true);
1813 tcp_cleanup_rbuf(sk, copied);
1814 } else
1815 dma_async_issue_pending(tp->ucopy.dma_chan);
1817 #endif
1818 if (copied >= target) {
1819 /* Do not sleep, just process backlog. */
1820 release_sock(sk);
1821 lock_sock(sk);
1822 } else
1823 sk_wait_data(sk, &timeo);
1825 #ifdef CONFIG_NET_DMA
1826 tcp_service_net_dma(sk, false); /* Don't block */
1827 tp->ucopy.wakeup = 0;
1828 #endif
1830 if (user_recv) {
1831 int chunk;
1833 /* __ Restore normal policy in scheduler __ */
1835 if ((chunk = len - tp->ucopy.len) != 0) {
1836 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1837 len -= chunk;
1838 copied += chunk;
1841 if (tp->rcv_nxt == tp->copied_seq &&
1842 !skb_queue_empty(&tp->ucopy.prequeue)) {
1843 do_prequeue:
1844 tcp_prequeue_process(sk);
1846 if ((chunk = len - tp->ucopy.len) != 0) {
1847 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1848 len -= chunk;
1849 copied += chunk;
1853 if ((flags & MSG_PEEK) &&
1854 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1855 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1856 current->comm,
1857 task_pid_nr(current));
1858 peek_seq = tp->copied_seq;
1860 continue;
1862 found_ok_skb:
1863 /* Ok so how much can we use? */
1864 used = skb->len - offset;
1865 if (len < used)
1866 used = len;
1868 /* Do we have urgent data here? */
1869 if (tp->urg_data) {
1870 u32 urg_offset = tp->urg_seq - *seq;
1871 if (urg_offset < used) {
1872 if (!urg_offset) {
1873 if (!sock_flag(sk, SOCK_URGINLINE)) {
1874 ++*seq;
1875 urg_hole++;
1876 offset++;
1877 used--;
1878 if (!used)
1879 goto skip_copy;
1881 } else
1882 used = urg_offset;
1886 if (!(flags & MSG_TRUNC)) {
1887 #ifdef CONFIG_NET_DMA
1888 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1889 tp->ucopy.dma_chan = net_dma_find_channel();
1891 if (tp->ucopy.dma_chan) {
1892 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1893 tp->ucopy.dma_chan, skb, offset,
1894 msg->msg_iov, used,
1895 tp->ucopy.pinned_list);
1897 if (tp->ucopy.dma_cookie < 0) {
1899 pr_alert("%s: dma_cookie < 0\n",
1900 __func__);
1902 /* Exception. Bailout! */
1903 if (!copied)
1904 copied = -EFAULT;
1905 break;
1908 dma_async_issue_pending(tp->ucopy.dma_chan);
1910 if ((offset + used) == skb->len)
1911 copied_early = true;
1913 } else
1914 #endif
1916 err = skb_copy_datagram_iovec(skb, offset,
1917 msg->msg_iov, used);
1918 if (err) {
1919 /* Exception. Bailout! */
1920 if (!copied)
1921 copied = -EFAULT;
1922 break;
1927 *seq += used;
1928 copied += used;
1929 len -= used;
1931 tcp_rcv_space_adjust(sk);
1933 skip_copy:
1934 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1935 tp->urg_data = 0;
1936 tcp_fast_path_check(sk);
1938 if (used + offset < skb->len)
1939 continue;
1941 if (tcp_hdr(skb)->fin)
1942 goto found_fin_ok;
1943 if (!(flags & MSG_PEEK)) {
1944 sk_eat_skb(sk, skb, copied_early);
1945 copied_early = false;
1947 continue;
1949 found_fin_ok:
1950 /* Process the FIN. */
1951 ++*seq;
1952 if (!(flags & MSG_PEEK)) {
1953 sk_eat_skb(sk, skb, copied_early);
1954 copied_early = false;
1956 break;
1957 } while (len > 0);
1959 if (user_recv) {
1960 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1961 int chunk;
1963 tp->ucopy.len = copied > 0 ? len : 0;
1965 tcp_prequeue_process(sk);
1967 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1968 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1969 len -= chunk;
1970 copied += chunk;
1974 tp->ucopy.task = NULL;
1975 tp->ucopy.len = 0;
1978 #ifdef CONFIG_NET_DMA
1979 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1980 tp->ucopy.dma_chan = NULL;
1982 if (tp->ucopy.pinned_list) {
1983 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1984 tp->ucopy.pinned_list = NULL;
1986 #endif
1988 /* According to UNIX98, msg_name/msg_namelen are ignored
1989 * on connected socket. I was just happy when found this 8) --ANK
1992 /* Clean up data we have read: This will do ACK frames. */
1993 tcp_cleanup_rbuf(sk, copied);
1995 release_sock(sk);
1996 return copied;
1998 out:
1999 release_sock(sk);
2000 return err;
2002 recv_urg:
2003 err = tcp_recv_urg(sk, msg, len, flags);
2004 goto out;
2006 recv_sndq:
2007 err = tcp_peek_sndq(sk, msg, len);
2008 goto out;
2010 EXPORT_SYMBOL(tcp_recvmsg);
2012 void tcp_set_state(struct sock *sk, int state)
2014 int oldstate = sk->sk_state;
2016 switch (state) {
2017 case TCP_ESTABLISHED:
2018 if (oldstate != TCP_ESTABLISHED)
2019 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2020 break;
2022 case TCP_CLOSE:
2023 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2024 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2026 sk->sk_prot->unhash(sk);
2027 if (inet_csk(sk)->icsk_bind_hash &&
2028 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2029 inet_put_port(sk);
2030 /* fall through */
2031 default:
2032 if (oldstate == TCP_ESTABLISHED)
2033 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2036 /* Change state AFTER socket is unhashed to avoid closed
2037 * socket sitting in hash tables.
2039 sk->sk_state = state;
2041 #ifdef STATE_TRACE
2042 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
2043 #endif
2045 EXPORT_SYMBOL_GPL(tcp_set_state);
2048 * State processing on a close. This implements the state shift for
2049 * sending our FIN frame. Note that we only send a FIN for some
2050 * states. A shutdown() may have already sent the FIN, or we may be
2051 * closed.
2054 static const unsigned char new_state[16] = {
2055 /* current state: new state: action: */
2056 /* (Invalid) */ TCP_CLOSE,
2057 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2058 /* TCP_SYN_SENT */ TCP_CLOSE,
2059 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2060 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
2061 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
2062 /* TCP_TIME_WAIT */ TCP_CLOSE,
2063 /* TCP_CLOSE */ TCP_CLOSE,
2064 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
2065 /* TCP_LAST_ACK */ TCP_LAST_ACK,
2066 /* TCP_LISTEN */ TCP_CLOSE,
2067 /* TCP_CLOSING */ TCP_CLOSING,
2070 static int tcp_close_state(struct sock *sk)
2072 int next = (int)new_state[sk->sk_state];
2073 int ns = next & TCP_STATE_MASK;
2075 tcp_set_state(sk, ns);
2077 return next & TCP_ACTION_FIN;
2081 * Shutdown the sending side of a connection. Much like close except
2082 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2085 void tcp_shutdown(struct sock *sk, int how)
2087 /* We need to grab some memory, and put together a FIN,
2088 * and then put it into the queue to be sent.
2089 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2091 if (!(how & SEND_SHUTDOWN))
2092 return;
2094 /* If we've already sent a FIN, or it's a closed state, skip this. */
2095 if ((1 << sk->sk_state) &
2096 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2097 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2098 /* Clear out any half completed packets. FIN if needed. */
2099 if (tcp_close_state(sk))
2100 tcp_send_fin(sk);
2103 EXPORT_SYMBOL(tcp_shutdown);
2105 bool tcp_check_oom(struct sock *sk, int shift)
2107 bool too_many_orphans, out_of_socket_memory;
2109 too_many_orphans = tcp_too_many_orphans(sk, shift);
2110 out_of_socket_memory = tcp_out_of_memory(sk);
2112 if (too_many_orphans)
2113 net_info_ratelimited("too many orphaned sockets\n");
2114 if (out_of_socket_memory)
2115 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2116 return too_many_orphans || out_of_socket_memory;
2119 void tcp_close(struct sock *sk, long timeout)
2121 struct sk_buff *skb;
2122 int data_was_unread = 0;
2123 int state;
2125 lock_sock(sk);
2126 sk->sk_shutdown = SHUTDOWN_MASK;
2128 if (sk->sk_state == TCP_LISTEN) {
2129 tcp_set_state(sk, TCP_CLOSE);
2131 /* Special case. */
2132 inet_csk_listen_stop(sk);
2134 goto adjudge_to_death;
2137 /* We need to flush the recv. buffs. We do this only on the
2138 * descriptor close, not protocol-sourced closes, because the
2139 * reader process may not have drained the data yet!
2141 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2142 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2143 tcp_hdr(skb)->fin;
2144 data_was_unread += len;
2145 __kfree_skb(skb);
2148 sk_mem_reclaim(sk);
2150 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2151 if (sk->sk_state == TCP_CLOSE)
2152 goto adjudge_to_death;
2154 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2155 * data was lost. To witness the awful effects of the old behavior of
2156 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2157 * GET in an FTP client, suspend the process, wait for the client to
2158 * advertise a zero window, then kill -9 the FTP client, wheee...
2159 * Note: timeout is always zero in such a case.
2161 if (unlikely(tcp_sk(sk)->repair)) {
2162 sk->sk_prot->disconnect(sk, 0);
2163 } else if (data_was_unread) {
2164 /* Unread data was tossed, zap the connection. */
2165 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2166 tcp_set_state(sk, TCP_CLOSE);
2167 tcp_send_active_reset(sk, sk->sk_allocation);
2168 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2169 /* Check zero linger _after_ checking for unread data. */
2170 sk->sk_prot->disconnect(sk, 0);
2171 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2172 } else if (tcp_close_state(sk)) {
2173 /* We FIN if the application ate all the data before
2174 * zapping the connection.
2177 /* RED-PEN. Formally speaking, we have broken TCP state
2178 * machine. State transitions:
2180 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2181 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2182 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2184 * are legal only when FIN has been sent (i.e. in window),
2185 * rather than queued out of window. Purists blame.
2187 * F.e. "RFC state" is ESTABLISHED,
2188 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2190 * The visible declinations are that sometimes
2191 * we enter time-wait state, when it is not required really
2192 * (harmless), do not send active resets, when they are
2193 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2194 * they look as CLOSING or LAST_ACK for Linux)
2195 * Probably, I missed some more holelets.
2196 * --ANK
2197 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2198 * in a single packet! (May consider it later but will
2199 * probably need API support or TCP_CORK SYN-ACK until
2200 * data is written and socket is closed.)
2202 tcp_send_fin(sk);
2205 sk_stream_wait_close(sk, timeout);
2207 adjudge_to_death:
2208 state = sk->sk_state;
2209 sock_hold(sk);
2210 sock_orphan(sk);
2212 /* It is the last release_sock in its life. It will remove backlog. */
2213 release_sock(sk);
2216 /* Now socket is owned by kernel and we acquire BH lock
2217 to finish close. No need to check for user refs.
2219 local_bh_disable();
2220 bh_lock_sock(sk);
2221 WARN_ON(sock_owned_by_user(sk));
2223 percpu_counter_inc(sk->sk_prot->orphan_count);
2225 /* Have we already been destroyed by a softirq or backlog? */
2226 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2227 goto out;
2229 /* This is a (useful) BSD violating of the RFC. There is a
2230 * problem with TCP as specified in that the other end could
2231 * keep a socket open forever with no application left this end.
2232 * We use a 3 minute timeout (about the same as BSD) then kill
2233 * our end. If they send after that then tough - BUT: long enough
2234 * that we won't make the old 4*rto = almost no time - whoops
2235 * reset mistake.
2237 * Nope, it was not mistake. It is really desired behaviour
2238 * f.e. on http servers, when such sockets are useless, but
2239 * consume significant resources. Let's do it with special
2240 * linger2 option. --ANK
2243 if (sk->sk_state == TCP_FIN_WAIT2) {
2244 struct tcp_sock *tp = tcp_sk(sk);
2245 if (tp->linger2 < 0) {
2246 tcp_set_state(sk, TCP_CLOSE);
2247 tcp_send_active_reset(sk, GFP_ATOMIC);
2248 NET_INC_STATS_BH(sock_net(sk),
2249 LINUX_MIB_TCPABORTONLINGER);
2250 } else {
2251 const int tmo = tcp_fin_time(sk);
2253 if (tmo > TCP_TIMEWAIT_LEN) {
2254 inet_csk_reset_keepalive_timer(sk,
2255 tmo - TCP_TIMEWAIT_LEN);
2256 } else {
2257 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2258 goto out;
2262 if (sk->sk_state != TCP_CLOSE) {
2263 sk_mem_reclaim(sk);
2264 if (tcp_check_oom(sk, 0)) {
2265 tcp_set_state(sk, TCP_CLOSE);
2266 tcp_send_active_reset(sk, GFP_ATOMIC);
2267 NET_INC_STATS_BH(sock_net(sk),
2268 LINUX_MIB_TCPABORTONMEMORY);
2272 if (sk->sk_state == TCP_CLOSE) {
2273 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2274 /* We could get here with a non-NULL req if the socket is
2275 * aborted (e.g., closed with unread data) before 3WHS
2276 * finishes.
2278 if (req != NULL)
2279 reqsk_fastopen_remove(sk, req, false);
2280 inet_csk_destroy_sock(sk);
2282 /* Otherwise, socket is reprieved until protocol close. */
2284 out:
2285 bh_unlock_sock(sk);
2286 local_bh_enable();
2287 sock_put(sk);
2289 EXPORT_SYMBOL(tcp_close);
2291 /* These states need RST on ABORT according to RFC793 */
2293 static inline bool tcp_need_reset(int state)
2295 return (1 << state) &
2296 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2297 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2300 int tcp_disconnect(struct sock *sk, int flags)
2302 struct inet_sock *inet = inet_sk(sk);
2303 struct inet_connection_sock *icsk = inet_csk(sk);
2304 struct tcp_sock *tp = tcp_sk(sk);
2305 int err = 0;
2306 int old_state = sk->sk_state;
2308 if (old_state != TCP_CLOSE)
2309 tcp_set_state(sk, TCP_CLOSE);
2311 /* ABORT function of RFC793 */
2312 if (old_state == TCP_LISTEN) {
2313 inet_csk_listen_stop(sk);
2314 } else if (unlikely(tp->repair)) {
2315 sk->sk_err = ECONNABORTED;
2316 } else if (tcp_need_reset(old_state) ||
2317 (tp->snd_nxt != tp->write_seq &&
2318 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2319 /* The last check adjusts for discrepancy of Linux wrt. RFC
2320 * states
2322 tcp_send_active_reset(sk, gfp_any());
2323 sk->sk_err = ECONNRESET;
2324 } else if (old_state == TCP_SYN_SENT)
2325 sk->sk_err = ECONNRESET;
2327 tcp_clear_xmit_timers(sk);
2328 __skb_queue_purge(&sk->sk_receive_queue);
2329 tcp_write_queue_purge(sk);
2330 __skb_queue_purge(&tp->out_of_order_queue);
2331 #ifdef CONFIG_NET_DMA
2332 __skb_queue_purge(&sk->sk_async_wait_queue);
2333 #endif
2335 inet->inet_dport = 0;
2337 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2338 inet_reset_saddr(sk);
2340 sk->sk_shutdown = 0;
2341 sock_reset_flag(sk, SOCK_DONE);
2342 tp->srtt = 0;
2343 if ((tp->write_seq += tp->max_window + 2) == 0)
2344 tp->write_seq = 1;
2345 icsk->icsk_backoff = 0;
2346 tp->snd_cwnd = 2;
2347 icsk->icsk_probes_out = 0;
2348 tp->packets_out = 0;
2349 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2350 tp->snd_cwnd_cnt = 0;
2351 tp->window_clamp = 0;
2352 tcp_set_ca_state(sk, TCP_CA_Open);
2353 tcp_clear_retrans(tp);
2354 inet_csk_delack_init(sk);
2355 tcp_init_send_head(sk);
2356 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2357 __sk_dst_reset(sk);
2359 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2361 sk->sk_error_report(sk);
2362 return err;
2364 EXPORT_SYMBOL(tcp_disconnect);
2366 void tcp_sock_destruct(struct sock *sk)
2368 inet_sock_destruct(sk);
2370 kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2373 static inline bool tcp_can_repair_sock(const struct sock *sk)
2375 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2376 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2379 static int tcp_repair_options_est(struct tcp_sock *tp,
2380 struct tcp_repair_opt __user *optbuf, unsigned int len)
2382 struct tcp_repair_opt opt;
2384 while (len >= sizeof(opt)) {
2385 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2386 return -EFAULT;
2388 optbuf++;
2389 len -= sizeof(opt);
2391 switch (opt.opt_code) {
2392 case TCPOPT_MSS:
2393 tp->rx_opt.mss_clamp = opt.opt_val;
2394 break;
2395 case TCPOPT_WINDOW:
2397 u16 snd_wscale = opt.opt_val & 0xFFFF;
2398 u16 rcv_wscale = opt.opt_val >> 16;
2400 if (snd_wscale > 14 || rcv_wscale > 14)
2401 return -EFBIG;
2403 tp->rx_opt.snd_wscale = snd_wscale;
2404 tp->rx_opt.rcv_wscale = rcv_wscale;
2405 tp->rx_opt.wscale_ok = 1;
2407 break;
2408 case TCPOPT_SACK_PERM:
2409 if (opt.opt_val != 0)
2410 return -EINVAL;
2412 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2413 if (sysctl_tcp_fack)
2414 tcp_enable_fack(tp);
2415 break;
2416 case TCPOPT_TIMESTAMP:
2417 if (opt.opt_val != 0)
2418 return -EINVAL;
2420 tp->rx_opt.tstamp_ok = 1;
2421 break;
2425 return 0;
2429 * Socket option code for TCP.
2431 static int do_tcp_setsockopt(struct sock *sk, int level,
2432 int optname, char __user *optval, unsigned int optlen)
2434 struct tcp_sock *tp = tcp_sk(sk);
2435 struct inet_connection_sock *icsk = inet_csk(sk);
2436 int val;
2437 int err = 0;
2439 /* These are data/string values, all the others are ints */
2440 switch (optname) {
2441 case TCP_CONGESTION: {
2442 char name[TCP_CA_NAME_MAX];
2444 if (optlen < 1)
2445 return -EINVAL;
2447 val = strncpy_from_user(name, optval,
2448 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2449 if (val < 0)
2450 return -EFAULT;
2451 name[val] = 0;
2453 lock_sock(sk);
2454 err = tcp_set_congestion_control(sk, name);
2455 release_sock(sk);
2456 return err;
2458 default:
2459 /* fallthru */
2460 break;
2463 if (optlen < sizeof(int))
2464 return -EINVAL;
2466 if (get_user(val, (int __user *)optval))
2467 return -EFAULT;
2469 lock_sock(sk);
2471 switch (optname) {
2472 case TCP_MAXSEG:
2473 /* Values greater than interface MTU won't take effect. However
2474 * at the point when this call is done we typically don't yet
2475 * know which interface is going to be used */
2476 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2477 err = -EINVAL;
2478 break;
2480 tp->rx_opt.user_mss = val;
2481 break;
2483 case TCP_NODELAY:
2484 if (val) {
2485 /* TCP_NODELAY is weaker than TCP_CORK, so that
2486 * this option on corked socket is remembered, but
2487 * it is not activated until cork is cleared.
2489 * However, when TCP_NODELAY is set we make
2490 * an explicit push, which overrides even TCP_CORK
2491 * for currently queued segments.
2493 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2494 tcp_push_pending_frames(sk);
2495 } else {
2496 tp->nonagle &= ~TCP_NAGLE_OFF;
2498 break;
2500 case TCP_THIN_LINEAR_TIMEOUTS:
2501 if (val < 0 || val > 1)
2502 err = -EINVAL;
2503 else
2504 tp->thin_lto = val;
2505 break;
2507 case TCP_THIN_DUPACK:
2508 if (val < 0 || val > 1)
2509 err = -EINVAL;
2510 else {
2511 tp->thin_dupack = val;
2512 if (tp->thin_dupack)
2513 tcp_disable_early_retrans(tp);
2515 break;
2517 case TCP_REPAIR:
2518 if (!tcp_can_repair_sock(sk))
2519 err = -EPERM;
2520 else if (val == 1) {
2521 tp->repair = 1;
2522 sk->sk_reuse = SK_FORCE_REUSE;
2523 tp->repair_queue = TCP_NO_QUEUE;
2524 } else if (val == 0) {
2525 tp->repair = 0;
2526 sk->sk_reuse = SK_NO_REUSE;
2527 tcp_send_window_probe(sk);
2528 } else
2529 err = -EINVAL;
2531 break;
2533 case TCP_REPAIR_QUEUE:
2534 if (!tp->repair)
2535 err = -EPERM;
2536 else if (val < TCP_QUEUES_NR)
2537 tp->repair_queue = val;
2538 else
2539 err = -EINVAL;
2540 break;
2542 case TCP_QUEUE_SEQ:
2543 if (sk->sk_state != TCP_CLOSE)
2544 err = -EPERM;
2545 else if (tp->repair_queue == TCP_SEND_QUEUE)
2546 tp->write_seq = val;
2547 else if (tp->repair_queue == TCP_RECV_QUEUE)
2548 tp->rcv_nxt = val;
2549 else
2550 err = -EINVAL;
2551 break;
2553 case TCP_REPAIR_OPTIONS:
2554 if (!tp->repair)
2555 err = -EINVAL;
2556 else if (sk->sk_state == TCP_ESTABLISHED)
2557 err = tcp_repair_options_est(tp,
2558 (struct tcp_repair_opt __user *)optval,
2559 optlen);
2560 else
2561 err = -EPERM;
2562 break;
2564 case TCP_CORK:
2565 /* When set indicates to always queue non-full frames.
2566 * Later the user clears this option and we transmit
2567 * any pending partial frames in the queue. This is
2568 * meant to be used alongside sendfile() to get properly
2569 * filled frames when the user (for example) must write
2570 * out headers with a write() call first and then use
2571 * sendfile to send out the data parts.
2573 * TCP_CORK can be set together with TCP_NODELAY and it is
2574 * stronger than TCP_NODELAY.
2576 if (val) {
2577 tp->nonagle |= TCP_NAGLE_CORK;
2578 } else {
2579 tp->nonagle &= ~TCP_NAGLE_CORK;
2580 if (tp->nonagle&TCP_NAGLE_OFF)
2581 tp->nonagle |= TCP_NAGLE_PUSH;
2582 tcp_push_pending_frames(sk);
2584 break;
2586 case TCP_KEEPIDLE:
2587 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2588 err = -EINVAL;
2589 else {
2590 tp->keepalive_time = val * HZ;
2591 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2592 !((1 << sk->sk_state) &
2593 (TCPF_CLOSE | TCPF_LISTEN))) {
2594 u32 elapsed = keepalive_time_elapsed(tp);
2595 if (tp->keepalive_time > elapsed)
2596 elapsed = tp->keepalive_time - elapsed;
2597 else
2598 elapsed = 0;
2599 inet_csk_reset_keepalive_timer(sk, elapsed);
2602 break;
2603 case TCP_KEEPINTVL:
2604 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2605 err = -EINVAL;
2606 else
2607 tp->keepalive_intvl = val * HZ;
2608 break;
2609 case TCP_KEEPCNT:
2610 if (val < 1 || val > MAX_TCP_KEEPCNT)
2611 err = -EINVAL;
2612 else
2613 tp->keepalive_probes = val;
2614 break;
2615 case TCP_SYNCNT:
2616 if (val < 1 || val > MAX_TCP_SYNCNT)
2617 err = -EINVAL;
2618 else
2619 icsk->icsk_syn_retries = val;
2620 break;
2622 case TCP_LINGER2:
2623 if (val < 0)
2624 tp->linger2 = -1;
2625 else if (val > sysctl_tcp_fin_timeout / HZ)
2626 tp->linger2 = 0;
2627 else
2628 tp->linger2 = val * HZ;
2629 break;
2631 case TCP_DEFER_ACCEPT:
2632 /* Translate value in seconds to number of retransmits */
2633 icsk->icsk_accept_queue.rskq_defer_accept =
2634 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2635 TCP_RTO_MAX / HZ);
2636 break;
2638 case TCP_WINDOW_CLAMP:
2639 if (!val) {
2640 if (sk->sk_state != TCP_CLOSE) {
2641 err = -EINVAL;
2642 break;
2644 tp->window_clamp = 0;
2645 } else
2646 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2647 SOCK_MIN_RCVBUF / 2 : val;
2648 break;
2650 case TCP_QUICKACK:
2651 if (!val) {
2652 icsk->icsk_ack.pingpong = 1;
2653 } else {
2654 icsk->icsk_ack.pingpong = 0;
2655 if ((1 << sk->sk_state) &
2656 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2657 inet_csk_ack_scheduled(sk)) {
2658 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2659 tcp_cleanup_rbuf(sk, 1);
2660 if (!(val & 1))
2661 icsk->icsk_ack.pingpong = 1;
2664 break;
2666 #ifdef CONFIG_TCP_MD5SIG
2667 case TCP_MD5SIG:
2668 /* Read the IP->Key mappings from userspace */
2669 err = tp->af_specific->md5_parse(sk, optval, optlen);
2670 break;
2671 #endif
2672 case TCP_USER_TIMEOUT:
2673 /* Cap the max timeout in ms TCP will retry/retrans
2674 * before giving up and aborting (ETIMEDOUT) a connection.
2676 if (val < 0)
2677 err = -EINVAL;
2678 else
2679 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2680 break;
2682 case TCP_FASTOPEN:
2683 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2684 TCPF_LISTEN)))
2685 err = fastopen_init_queue(sk, val);
2686 else
2687 err = -EINVAL;
2688 break;
2689 case TCP_TIMESTAMP:
2690 if (!tp->repair)
2691 err = -EPERM;
2692 else
2693 tp->tsoffset = val - tcp_time_stamp;
2694 break;
2695 case TCP_NOTSENT_LOWAT:
2696 tp->notsent_lowat = val;
2697 sk->sk_write_space(sk);
2698 break;
2699 default:
2700 err = -ENOPROTOOPT;
2701 break;
2704 release_sock(sk);
2705 return err;
2708 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2709 unsigned int optlen)
2711 const struct inet_connection_sock *icsk = inet_csk(sk);
2713 if (level != SOL_TCP)
2714 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2715 optval, optlen);
2716 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2718 EXPORT_SYMBOL(tcp_setsockopt);
2720 #ifdef CONFIG_COMPAT
2721 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2722 char __user *optval, unsigned int optlen)
2724 if (level != SOL_TCP)
2725 return inet_csk_compat_setsockopt(sk, level, optname,
2726 optval, optlen);
2727 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2729 EXPORT_SYMBOL(compat_tcp_setsockopt);
2730 #endif
2732 /* Return information about state of tcp endpoint in API format. */
2733 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2735 const struct tcp_sock *tp = tcp_sk(sk);
2736 const struct inet_connection_sock *icsk = inet_csk(sk);
2737 u32 now = tcp_time_stamp;
2739 memset(info, 0, sizeof(*info));
2741 info->tcpi_state = sk->sk_state;
2742 info->tcpi_ca_state = icsk->icsk_ca_state;
2743 info->tcpi_retransmits = icsk->icsk_retransmits;
2744 info->tcpi_probes = icsk->icsk_probes_out;
2745 info->tcpi_backoff = icsk->icsk_backoff;
2747 if (tp->rx_opt.tstamp_ok)
2748 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2749 if (tcp_is_sack(tp))
2750 info->tcpi_options |= TCPI_OPT_SACK;
2751 if (tp->rx_opt.wscale_ok) {
2752 info->tcpi_options |= TCPI_OPT_WSCALE;
2753 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2754 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2757 if (tp->ecn_flags & TCP_ECN_OK)
2758 info->tcpi_options |= TCPI_OPT_ECN;
2759 if (tp->ecn_flags & TCP_ECN_SEEN)
2760 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2761 if (tp->syn_data_acked)
2762 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2764 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2765 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2766 info->tcpi_snd_mss = tp->mss_cache;
2767 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2769 if (sk->sk_state == TCP_LISTEN) {
2770 info->tcpi_unacked = sk->sk_ack_backlog;
2771 info->tcpi_sacked = sk->sk_max_ack_backlog;
2772 } else {
2773 info->tcpi_unacked = tp->packets_out;
2774 info->tcpi_sacked = tp->sacked_out;
2776 info->tcpi_lost = tp->lost_out;
2777 info->tcpi_retrans = tp->retrans_out;
2778 info->tcpi_fackets = tp->fackets_out;
2780 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2781 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2782 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2784 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2785 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2786 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2787 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2788 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2789 info->tcpi_snd_cwnd = tp->snd_cwnd;
2790 info->tcpi_advmss = tp->advmss;
2791 info->tcpi_reordering = tp->reordering;
2793 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2794 info->tcpi_rcv_space = tp->rcvq_space.space;
2796 info->tcpi_total_retrans = tp->total_retrans;
2798 EXPORT_SYMBOL_GPL(tcp_get_info);
2800 static int do_tcp_getsockopt(struct sock *sk, int level,
2801 int optname, char __user *optval, int __user *optlen)
2803 struct inet_connection_sock *icsk = inet_csk(sk);
2804 struct tcp_sock *tp = tcp_sk(sk);
2805 int val, len;
2807 if (get_user(len, optlen))
2808 return -EFAULT;
2810 len = min_t(unsigned int, len, sizeof(int));
2812 if (len < 0)
2813 return -EINVAL;
2815 switch (optname) {
2816 case TCP_MAXSEG:
2817 val = tp->mss_cache;
2818 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2819 val = tp->rx_opt.user_mss;
2820 if (tp->repair)
2821 val = tp->rx_opt.mss_clamp;
2822 break;
2823 case TCP_NODELAY:
2824 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2825 break;
2826 case TCP_CORK:
2827 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2828 break;
2829 case TCP_KEEPIDLE:
2830 val = keepalive_time_when(tp) / HZ;
2831 break;
2832 case TCP_KEEPINTVL:
2833 val = keepalive_intvl_when(tp) / HZ;
2834 break;
2835 case TCP_KEEPCNT:
2836 val = keepalive_probes(tp);
2837 break;
2838 case TCP_SYNCNT:
2839 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2840 break;
2841 case TCP_LINGER2:
2842 val = tp->linger2;
2843 if (val >= 0)
2844 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2845 break;
2846 case TCP_DEFER_ACCEPT:
2847 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2848 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2849 break;
2850 case TCP_WINDOW_CLAMP:
2851 val = tp->window_clamp;
2852 break;
2853 case TCP_INFO: {
2854 struct tcp_info info;
2856 if (get_user(len, optlen))
2857 return -EFAULT;
2859 tcp_get_info(sk, &info);
2861 len = min_t(unsigned int, len, sizeof(info));
2862 if (put_user(len, optlen))
2863 return -EFAULT;
2864 if (copy_to_user(optval, &info, len))
2865 return -EFAULT;
2866 return 0;
2868 case TCP_QUICKACK:
2869 val = !icsk->icsk_ack.pingpong;
2870 break;
2872 case TCP_CONGESTION:
2873 if (get_user(len, optlen))
2874 return -EFAULT;
2875 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2876 if (put_user(len, optlen))
2877 return -EFAULT;
2878 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2879 return -EFAULT;
2880 return 0;
2882 case TCP_THIN_LINEAR_TIMEOUTS:
2883 val = tp->thin_lto;
2884 break;
2885 case TCP_THIN_DUPACK:
2886 val = tp->thin_dupack;
2887 break;
2889 case TCP_REPAIR:
2890 val = tp->repair;
2891 break;
2893 case TCP_REPAIR_QUEUE:
2894 if (tp->repair)
2895 val = tp->repair_queue;
2896 else
2897 return -EINVAL;
2898 break;
2900 case TCP_QUEUE_SEQ:
2901 if (tp->repair_queue == TCP_SEND_QUEUE)
2902 val = tp->write_seq;
2903 else if (tp->repair_queue == TCP_RECV_QUEUE)
2904 val = tp->rcv_nxt;
2905 else
2906 return -EINVAL;
2907 break;
2909 case TCP_USER_TIMEOUT:
2910 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2911 break;
2912 case TCP_TIMESTAMP:
2913 val = tcp_time_stamp + tp->tsoffset;
2914 break;
2915 case TCP_NOTSENT_LOWAT:
2916 val = tp->notsent_lowat;
2917 break;
2918 default:
2919 return -ENOPROTOOPT;
2922 if (put_user(len, optlen))
2923 return -EFAULT;
2924 if (copy_to_user(optval, &val, len))
2925 return -EFAULT;
2926 return 0;
2929 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2930 int __user *optlen)
2932 struct inet_connection_sock *icsk = inet_csk(sk);
2934 if (level != SOL_TCP)
2935 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2936 optval, optlen);
2937 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2939 EXPORT_SYMBOL(tcp_getsockopt);
2941 #ifdef CONFIG_COMPAT
2942 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2943 char __user *optval, int __user *optlen)
2945 if (level != SOL_TCP)
2946 return inet_csk_compat_getsockopt(sk, level, optname,
2947 optval, optlen);
2948 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2950 EXPORT_SYMBOL(compat_tcp_getsockopt);
2951 #endif
2953 #ifdef CONFIG_TCP_MD5SIG
2954 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool __read_mostly;
2955 static DEFINE_MUTEX(tcp_md5sig_mutex);
2957 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
2959 int cpu;
2961 for_each_possible_cpu(cpu) {
2962 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
2964 if (p->md5_desc.tfm)
2965 crypto_free_hash(p->md5_desc.tfm);
2967 free_percpu(pool);
2970 static void __tcp_alloc_md5sig_pool(void)
2972 int cpu;
2973 struct tcp_md5sig_pool __percpu *pool;
2975 pool = alloc_percpu(struct tcp_md5sig_pool);
2976 if (!pool)
2977 return;
2979 for_each_possible_cpu(cpu) {
2980 struct crypto_hash *hash;
2982 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2983 if (IS_ERR_OR_NULL(hash))
2984 goto out_free;
2986 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
2988 /* before setting tcp_md5sig_pool, we must commit all writes
2989 * to memory. See ACCESS_ONCE() in tcp_get_md5sig_pool()
2991 smp_wmb();
2992 tcp_md5sig_pool = pool;
2993 return;
2994 out_free:
2995 __tcp_free_md5sig_pool(pool);
2998 bool tcp_alloc_md5sig_pool(void)
3000 if (unlikely(!tcp_md5sig_pool)) {
3001 mutex_lock(&tcp_md5sig_mutex);
3003 if (!tcp_md5sig_pool)
3004 __tcp_alloc_md5sig_pool();
3006 mutex_unlock(&tcp_md5sig_mutex);
3008 return tcp_md5sig_pool != NULL;
3010 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3014 * tcp_get_md5sig_pool - get md5sig_pool for this user
3016 * We use percpu structure, so if we succeed, we exit with preemption
3017 * and BH disabled, to make sure another thread or softirq handling
3018 * wont try to get same context.
3020 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3022 struct tcp_md5sig_pool __percpu *p;
3024 local_bh_disable();
3025 p = ACCESS_ONCE(tcp_md5sig_pool);
3026 if (p)
3027 return __this_cpu_ptr(p);
3029 local_bh_enable();
3030 return NULL;
3032 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3034 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3035 const struct tcphdr *th)
3037 struct scatterlist sg;
3038 struct tcphdr hdr;
3039 int err;
3041 /* We are not allowed to change tcphdr, make a local copy */
3042 memcpy(&hdr, th, sizeof(hdr));
3043 hdr.check = 0;
3045 /* options aren't included in the hash */
3046 sg_init_one(&sg, &hdr, sizeof(hdr));
3047 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3048 return err;
3050 EXPORT_SYMBOL(tcp_md5_hash_header);
3052 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3053 const struct sk_buff *skb, unsigned int header_len)
3055 struct scatterlist sg;
3056 const struct tcphdr *tp = tcp_hdr(skb);
3057 struct hash_desc *desc = &hp->md5_desc;
3058 unsigned int i;
3059 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3060 skb_headlen(skb) - header_len : 0;
3061 const struct skb_shared_info *shi = skb_shinfo(skb);
3062 struct sk_buff *frag_iter;
3064 sg_init_table(&sg, 1);
3066 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3067 if (crypto_hash_update(desc, &sg, head_data_len))
3068 return 1;
3070 for (i = 0; i < shi->nr_frags; ++i) {
3071 const struct skb_frag_struct *f = &shi->frags[i];
3072 unsigned int offset = f->page_offset;
3073 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3075 sg_set_page(&sg, page, skb_frag_size(f),
3076 offset_in_page(offset));
3077 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3078 return 1;
3081 skb_walk_frags(skb, frag_iter)
3082 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3083 return 1;
3085 return 0;
3087 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3089 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3091 struct scatterlist sg;
3093 sg_init_one(&sg, key->key, key->keylen);
3094 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3096 EXPORT_SYMBOL(tcp_md5_hash_key);
3098 #endif
3100 void tcp_done(struct sock *sk)
3102 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3104 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3105 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3107 tcp_set_state(sk, TCP_CLOSE);
3108 tcp_clear_xmit_timers(sk);
3109 if (req != NULL)
3110 reqsk_fastopen_remove(sk, req, false);
3112 sk->sk_shutdown = SHUTDOWN_MASK;
3114 if (!sock_flag(sk, SOCK_DEAD))
3115 sk->sk_state_change(sk);
3116 else
3117 inet_csk_destroy_sock(sk);
3119 EXPORT_SYMBOL_GPL(tcp_done);
3121 extern struct tcp_congestion_ops tcp_reno;
3123 static __initdata unsigned long thash_entries;
3124 static int __init set_thash_entries(char *str)
3126 ssize_t ret;
3128 if (!str)
3129 return 0;
3131 ret = kstrtoul(str, 0, &thash_entries);
3132 if (ret)
3133 return 0;
3135 return 1;
3137 __setup("thash_entries=", set_thash_entries);
3139 static void tcp_init_mem(void)
3141 unsigned long limit = nr_free_buffer_pages() / 8;
3142 limit = max(limit, 128UL);
3143 sysctl_tcp_mem[0] = limit / 4 * 3;
3144 sysctl_tcp_mem[1] = limit;
3145 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3148 void __init tcp_init(void)
3150 struct sk_buff *skb = NULL;
3151 unsigned long limit;
3152 int max_rshare, max_wshare, cnt;
3153 unsigned int i;
3155 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3157 percpu_counter_init(&tcp_sockets_allocated, 0);
3158 percpu_counter_init(&tcp_orphan_count, 0);
3159 tcp_hashinfo.bind_bucket_cachep =
3160 kmem_cache_create("tcp_bind_bucket",
3161 sizeof(struct inet_bind_bucket), 0,
3162 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3164 /* Size and allocate the main established and bind bucket
3165 * hash tables.
3167 * The methodology is similar to that of the buffer cache.
3169 tcp_hashinfo.ehash =
3170 alloc_large_system_hash("TCP established",
3171 sizeof(struct inet_ehash_bucket),
3172 thash_entries,
3173 17, /* one slot per 128 KB of memory */
3175 NULL,
3176 &tcp_hashinfo.ehash_mask,
3178 thash_entries ? 0 : 512 * 1024);
3179 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3180 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3182 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3183 panic("TCP: failed to alloc ehash_locks");
3184 tcp_hashinfo.bhash =
3185 alloc_large_system_hash("TCP bind",
3186 sizeof(struct inet_bind_hashbucket),
3187 tcp_hashinfo.ehash_mask + 1,
3188 17, /* one slot per 128 KB of memory */
3190 &tcp_hashinfo.bhash_size,
3191 NULL,
3193 64 * 1024);
3194 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3195 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3196 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3197 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3201 cnt = tcp_hashinfo.ehash_mask + 1;
3203 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3204 sysctl_tcp_max_orphans = cnt / 2;
3205 sysctl_max_syn_backlog = max(128, cnt / 256);
3207 tcp_init_mem();
3208 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3209 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3210 max_wshare = min(4UL*1024*1024, limit);
3211 max_rshare = min(6UL*1024*1024, limit);
3213 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3214 sysctl_tcp_wmem[1] = 16*1024;
3215 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3217 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3218 sysctl_tcp_rmem[1] = 87380;
3219 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3221 pr_info("Hash tables configured (established %u bind %u)\n",
3222 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3224 tcp_metrics_init();
3226 tcp_register_congestion_control(&tcp_reno);
3228 tcp_tasklet_init();