gma500: Fix clashes with DRM updates
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / net / ipv4 / tcp.c
blob46febcacb729f0657c9c325ee18f007c2f9fe417
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 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
267 #include <linux/time.h>
268 #include <linux/slab.h>
270 #include <net/icmp.h>
271 #include <net/tcp.h>
272 #include <net/xfrm.h>
273 #include <net/ip.h>
274 #include <net/netdma.h>
275 #include <net/sock.h>
277 #include <asm/uaccess.h>
278 #include <asm/ioctls.h>
280 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
282 struct percpu_counter tcp_orphan_count;
283 EXPORT_SYMBOL_GPL(tcp_orphan_count);
285 long sysctl_tcp_mem[3] __read_mostly;
286 int sysctl_tcp_wmem[3] __read_mostly;
287 int sysctl_tcp_rmem[3] __read_mostly;
289 EXPORT_SYMBOL(sysctl_tcp_mem);
290 EXPORT_SYMBOL(sysctl_tcp_rmem);
291 EXPORT_SYMBOL(sysctl_tcp_wmem);
293 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
294 EXPORT_SYMBOL(tcp_memory_allocated);
297 * Current number of TCP sockets.
299 struct percpu_counter tcp_sockets_allocated;
300 EXPORT_SYMBOL(tcp_sockets_allocated);
303 * TCP splice context
305 struct tcp_splice_state {
306 struct pipe_inode_info *pipe;
307 size_t len;
308 unsigned int flags;
312 * Pressure flag: try to collapse.
313 * Technical note: it is used by multiple contexts non atomically.
314 * All the __sk_mem_schedule() is of this nature: accounting
315 * is strict, actions are advisory and have some latency.
317 int tcp_memory_pressure __read_mostly;
318 EXPORT_SYMBOL(tcp_memory_pressure);
320 void tcp_enter_memory_pressure(struct sock *sk)
322 if (!tcp_memory_pressure) {
323 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
324 tcp_memory_pressure = 1;
327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
329 /* Convert seconds to retransmits based on initial and max timeout */
330 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
332 u8 res = 0;
334 if (seconds > 0) {
335 int period = timeout;
337 res = 1;
338 while (seconds > period && res < 255) {
339 res++;
340 timeout <<= 1;
341 if (timeout > rto_max)
342 timeout = rto_max;
343 period += timeout;
346 return res;
349 /* Convert retransmits to seconds based on initial and max timeout */
350 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
352 int period = 0;
354 if (retrans > 0) {
355 period = timeout;
356 while (--retrans) {
357 timeout <<= 1;
358 if (timeout > rto_max)
359 timeout = rto_max;
360 period += timeout;
363 return period;
367 * Wait for a TCP event.
369 * Note that we don't need to lock the socket, as the upper poll layers
370 * take care of normal races (between the test and the event) and we don't
371 * go look at any of the socket buffers directly.
373 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
375 unsigned int mask;
376 struct sock *sk = sock->sk;
377 struct tcp_sock *tp = tcp_sk(sk);
379 sock_poll_wait(file, sk_sleep(sk), wait);
380 if (sk->sk_state == TCP_LISTEN)
381 return inet_csk_listen_poll(sk);
383 /* Socket is not locked. We are protected from async events
384 * by poll logic and correct handling of state changes
385 * made by other threads is impossible in any case.
388 mask = 0;
391 * POLLHUP is certainly not done right. But poll() doesn't
392 * have a notion of HUP in just one direction, and for a
393 * socket the read side is more interesting.
395 * Some poll() documentation says that POLLHUP is incompatible
396 * with the POLLOUT/POLLWR flags, so somebody should check this
397 * all. But careful, it tends to be safer to return too many
398 * bits than too few, and you can easily break real applications
399 * if you don't tell them that something has hung up!
401 * Check-me.
403 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
404 * our fs/select.c). It means that after we received EOF,
405 * poll always returns immediately, making impossible poll() on write()
406 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
407 * if and only if shutdown has been made in both directions.
408 * Actually, it is interesting to look how Solaris and DUX
409 * solve this dilemma. I would prefer, if POLLHUP were maskable,
410 * then we could set it on SND_SHUTDOWN. BTW examples given
411 * in Stevens' books assume exactly this behaviour, it explains
412 * why POLLHUP is incompatible with POLLOUT. --ANK
414 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
415 * blocking on fresh not-connected or disconnected socket. --ANK
417 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
418 mask |= POLLHUP;
419 if (sk->sk_shutdown & RCV_SHUTDOWN)
420 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
422 /* Connected? */
423 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
424 int target = sock_rcvlowat(sk, 0, INT_MAX);
426 if (tp->urg_seq == tp->copied_seq &&
427 !sock_flag(sk, SOCK_URGINLINE) &&
428 tp->urg_data)
429 target++;
431 /* Potential race condition. If read of tp below will
432 * escape above sk->sk_state, we can be illegally awaken
433 * in SYN_* states. */
434 if (tp->rcv_nxt - tp->copied_seq >= target)
435 mask |= POLLIN | POLLRDNORM;
437 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
438 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
439 mask |= POLLOUT | POLLWRNORM;
440 } else { /* send SIGIO later */
441 set_bit(SOCK_ASYNC_NOSPACE,
442 &sk->sk_socket->flags);
443 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
445 /* Race breaker. If space is freed after
446 * wspace test but before the flags are set,
447 * IO signal will be lost.
449 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
450 mask |= POLLOUT | POLLWRNORM;
452 } else
453 mask |= POLLOUT | POLLWRNORM;
455 if (tp->urg_data & TCP_URG_VALID)
456 mask |= POLLPRI;
458 /* This barrier is coupled with smp_wmb() in tcp_reset() */
459 smp_rmb();
460 if (sk->sk_err)
461 mask |= POLLERR;
463 return mask;
465 EXPORT_SYMBOL(tcp_poll);
467 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
469 struct tcp_sock *tp = tcp_sk(sk);
470 int answ;
472 switch (cmd) {
473 case SIOCINQ:
474 if (sk->sk_state == TCP_LISTEN)
475 return -EINVAL;
477 lock_sock(sk);
478 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
479 answ = 0;
480 else if (sock_flag(sk, SOCK_URGINLINE) ||
481 !tp->urg_data ||
482 before(tp->urg_seq, tp->copied_seq) ||
483 !before(tp->urg_seq, tp->rcv_nxt)) {
484 struct sk_buff *skb;
486 answ = tp->rcv_nxt - tp->copied_seq;
488 /* Subtract 1, if FIN is in queue. */
489 skb = skb_peek_tail(&sk->sk_receive_queue);
490 if (answ && skb)
491 answ -= tcp_hdr(skb)->fin;
492 } else
493 answ = tp->urg_seq - tp->copied_seq;
494 release_sock(sk);
495 break;
496 case SIOCATMARK:
497 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
498 break;
499 case SIOCOUTQ:
500 if (sk->sk_state == TCP_LISTEN)
501 return -EINVAL;
503 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
504 answ = 0;
505 else
506 answ = tp->write_seq - tp->snd_una;
507 break;
508 case SIOCOUTQNSD:
509 if (sk->sk_state == TCP_LISTEN)
510 return -EINVAL;
512 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
513 answ = 0;
514 else
515 answ = tp->write_seq - tp->snd_nxt;
516 break;
517 default:
518 return -ENOIOCTLCMD;
521 return put_user(answ, (int __user *)arg);
523 EXPORT_SYMBOL(tcp_ioctl);
525 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
527 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
528 tp->pushed_seq = tp->write_seq;
531 static inline int forced_push(struct tcp_sock *tp)
533 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
536 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
538 struct tcp_sock *tp = tcp_sk(sk);
539 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
541 skb->csum = 0;
542 tcb->seq = tcb->end_seq = tp->write_seq;
543 tcb->flags = TCPHDR_ACK;
544 tcb->sacked = 0;
545 skb_header_release(skb);
546 tcp_add_write_queue_tail(sk, skb);
547 sk->sk_wmem_queued += skb->truesize;
548 sk_mem_charge(sk, skb->truesize);
549 if (tp->nonagle & TCP_NAGLE_PUSH)
550 tp->nonagle &= ~TCP_NAGLE_PUSH;
553 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
555 if (flags & MSG_OOB)
556 tp->snd_up = tp->write_seq;
559 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
560 int nonagle)
562 if (tcp_send_head(sk)) {
563 struct tcp_sock *tp = tcp_sk(sk);
565 if (!(flags & MSG_MORE) || forced_push(tp))
566 tcp_mark_push(tp, tcp_write_queue_tail(sk));
568 tcp_mark_urg(tp, flags);
569 __tcp_push_pending_frames(sk, mss_now,
570 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
574 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
575 unsigned int offset, size_t len)
577 struct tcp_splice_state *tss = rd_desc->arg.data;
578 int ret;
580 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
581 tss->flags);
582 if (ret > 0)
583 rd_desc->count -= ret;
584 return ret;
587 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
589 /* Store TCP splice context information in read_descriptor_t. */
590 read_descriptor_t rd_desc = {
591 .arg.data = tss,
592 .count = tss->len,
595 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
599 * tcp_splice_read - splice data from TCP socket to a pipe
600 * @sock: socket to splice from
601 * @ppos: position (not valid)
602 * @pipe: pipe to splice to
603 * @len: number of bytes to splice
604 * @flags: splice modifier flags
606 * Description:
607 * Will read pages from given socket and fill them into a pipe.
610 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
611 struct pipe_inode_info *pipe, size_t len,
612 unsigned int flags)
614 struct sock *sk = sock->sk;
615 struct tcp_splice_state tss = {
616 .pipe = pipe,
617 .len = len,
618 .flags = flags,
620 long timeo;
621 ssize_t spliced;
622 int ret;
624 sock_rps_record_flow(sk);
626 * We can't seek on a socket input
628 if (unlikely(*ppos))
629 return -ESPIPE;
631 ret = spliced = 0;
633 lock_sock(sk);
635 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
636 while (tss.len) {
637 ret = __tcp_splice_read(sk, &tss);
638 if (ret < 0)
639 break;
640 else if (!ret) {
641 if (spliced)
642 break;
643 if (sock_flag(sk, SOCK_DONE))
644 break;
645 if (sk->sk_err) {
646 ret = sock_error(sk);
647 break;
649 if (sk->sk_shutdown & RCV_SHUTDOWN)
650 break;
651 if (sk->sk_state == TCP_CLOSE) {
653 * This occurs when user tries to read
654 * from never connected socket.
656 if (!sock_flag(sk, SOCK_DONE))
657 ret = -ENOTCONN;
658 break;
660 if (!timeo) {
661 ret = -EAGAIN;
662 break;
664 sk_wait_data(sk, &timeo);
665 if (signal_pending(current)) {
666 ret = sock_intr_errno(timeo);
667 break;
669 continue;
671 tss.len -= ret;
672 spliced += ret;
674 if (!timeo)
675 break;
676 release_sock(sk);
677 lock_sock(sk);
679 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
680 (sk->sk_shutdown & RCV_SHUTDOWN) ||
681 signal_pending(current))
682 break;
685 release_sock(sk);
687 if (spliced)
688 return spliced;
690 return ret;
692 EXPORT_SYMBOL(tcp_splice_read);
694 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
696 struct sk_buff *skb;
698 /* The TCP header must be at least 32-bit aligned. */
699 size = ALIGN(size, 4);
701 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
702 if (skb) {
703 if (sk_wmem_schedule(sk, skb->truesize)) {
705 * Make sure that we have exactly size bytes
706 * available to the caller, no more, no less.
708 skb_reserve(skb, skb_tailroom(skb) - size);
709 return skb;
711 __kfree_skb(skb);
712 } else {
713 sk->sk_prot->enter_memory_pressure(sk);
714 sk_stream_moderate_sndbuf(sk);
716 return NULL;
719 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
720 int large_allowed)
722 struct tcp_sock *tp = tcp_sk(sk);
723 u32 xmit_size_goal, old_size_goal;
725 xmit_size_goal = mss_now;
727 if (large_allowed && sk_can_gso(sk)) {
728 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
729 inet_csk(sk)->icsk_af_ops->net_header_len -
730 inet_csk(sk)->icsk_ext_hdr_len -
731 tp->tcp_header_len);
733 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
735 /* We try hard to avoid divides here */
736 old_size_goal = tp->xmit_size_goal_segs * mss_now;
738 if (likely(old_size_goal <= xmit_size_goal &&
739 old_size_goal + mss_now > xmit_size_goal)) {
740 xmit_size_goal = old_size_goal;
741 } else {
742 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
743 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
747 return max(xmit_size_goal, mss_now);
750 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
752 int mss_now;
754 mss_now = tcp_current_mss(sk);
755 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
757 return mss_now;
760 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
761 size_t psize, int flags)
763 struct tcp_sock *tp = tcp_sk(sk);
764 int mss_now, size_goal;
765 int err;
766 ssize_t copied;
767 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
769 /* Wait for a connection to finish. */
770 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
771 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
772 goto out_err;
774 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
776 mss_now = tcp_send_mss(sk, &size_goal, flags);
777 copied = 0;
779 err = -EPIPE;
780 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
781 goto out_err;
783 while (psize > 0) {
784 struct sk_buff *skb = tcp_write_queue_tail(sk);
785 struct page *page = pages[poffset / PAGE_SIZE];
786 int copy, i, can_coalesce;
787 int offset = poffset % PAGE_SIZE;
788 int size = min_t(size_t, psize, PAGE_SIZE - offset);
790 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
791 new_segment:
792 if (!sk_stream_memory_free(sk))
793 goto wait_for_sndbuf;
795 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
796 if (!skb)
797 goto wait_for_memory;
799 skb_entail(sk, skb);
800 copy = size_goal;
803 if (copy > size)
804 copy = size;
806 i = skb_shinfo(skb)->nr_frags;
807 can_coalesce = skb_can_coalesce(skb, i, page, offset);
808 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
809 tcp_mark_push(tp, skb);
810 goto new_segment;
812 if (!sk_wmem_schedule(sk, copy))
813 goto wait_for_memory;
815 if (can_coalesce) {
816 skb_shinfo(skb)->frags[i - 1].size += copy;
817 } else {
818 get_page(page);
819 skb_fill_page_desc(skb, i, page, offset, copy);
822 skb->len += copy;
823 skb->data_len += copy;
824 skb->truesize += copy;
825 sk->sk_wmem_queued += copy;
826 sk_mem_charge(sk, copy);
827 skb->ip_summed = CHECKSUM_PARTIAL;
828 tp->write_seq += copy;
829 TCP_SKB_CB(skb)->end_seq += copy;
830 skb_shinfo(skb)->gso_segs = 0;
832 if (!copied)
833 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
835 copied += copy;
836 poffset += copy;
837 if (!(psize -= copy))
838 goto out;
840 if (skb->len < size_goal || (flags & MSG_OOB))
841 continue;
843 if (forced_push(tp)) {
844 tcp_mark_push(tp, skb);
845 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
846 } else if (skb == tcp_send_head(sk))
847 tcp_push_one(sk, mss_now);
848 continue;
850 wait_for_sndbuf:
851 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
852 wait_for_memory:
853 if (copied)
854 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
856 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
857 goto do_error;
859 mss_now = tcp_send_mss(sk, &size_goal, flags);
862 out:
863 if (copied)
864 tcp_push(sk, flags, mss_now, tp->nonagle);
865 return copied;
867 do_error:
868 if (copied)
869 goto out;
870 out_err:
871 return sk_stream_error(sk, flags, err);
874 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
875 size_t size, int flags)
877 ssize_t res;
879 if (!(sk->sk_route_caps & NETIF_F_SG) ||
880 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
881 return sock_no_sendpage(sk->sk_socket, page, offset, size,
882 flags);
884 lock_sock(sk);
885 res = do_tcp_sendpages(sk, &page, offset, size, flags);
886 release_sock(sk);
887 return res;
889 EXPORT_SYMBOL(tcp_sendpage);
891 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
892 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
894 static inline int select_size(struct sock *sk, int sg)
896 struct tcp_sock *tp = tcp_sk(sk);
897 int tmp = tp->mss_cache;
899 if (sg) {
900 if (sk_can_gso(sk))
901 tmp = 0;
902 else {
903 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
905 if (tmp >= pgbreak &&
906 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
907 tmp = pgbreak;
911 return tmp;
914 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
915 size_t size)
917 struct iovec *iov;
918 struct tcp_sock *tp = tcp_sk(sk);
919 struct sk_buff *skb;
920 int iovlen, flags;
921 int mss_now, size_goal;
922 int sg, err, copied;
923 long timeo;
925 lock_sock(sk);
927 flags = msg->msg_flags;
928 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
930 /* Wait for a connection to finish. */
931 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
932 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
933 goto out_err;
935 /* This should be in poll */
936 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
938 mss_now = tcp_send_mss(sk, &size_goal, flags);
940 /* Ok commence sending. */
941 iovlen = msg->msg_iovlen;
942 iov = msg->msg_iov;
943 copied = 0;
945 err = -EPIPE;
946 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
947 goto out_err;
949 sg = sk->sk_route_caps & NETIF_F_SG;
951 while (--iovlen >= 0) {
952 size_t seglen = iov->iov_len;
953 unsigned char __user *from = iov->iov_base;
955 iov++;
957 while (seglen > 0) {
958 int copy = 0;
959 int max = size_goal;
961 skb = tcp_write_queue_tail(sk);
962 if (tcp_send_head(sk)) {
963 if (skb->ip_summed == CHECKSUM_NONE)
964 max = mss_now;
965 copy = max - skb->len;
968 if (copy <= 0) {
969 new_segment:
970 /* Allocate new segment. If the interface is SG,
971 * allocate skb fitting to single page.
973 if (!sk_stream_memory_free(sk))
974 goto wait_for_sndbuf;
976 skb = sk_stream_alloc_skb(sk,
977 select_size(sk, sg),
978 sk->sk_allocation);
979 if (!skb)
980 goto wait_for_memory;
983 * Check whether we can use HW checksum.
985 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
986 skb->ip_summed = CHECKSUM_PARTIAL;
988 skb_entail(sk, skb);
989 copy = size_goal;
990 max = size_goal;
993 /* Try to append data to the end of skb. */
994 if (copy > seglen)
995 copy = seglen;
997 /* Where to copy to? */
998 if (skb_tailroom(skb) > 0) {
999 /* We have some space in skb head. Superb! */
1000 if (copy > skb_tailroom(skb))
1001 copy = skb_tailroom(skb);
1002 err = skb_add_data_nocache(sk, skb, from, copy);
1003 if (err)
1004 goto do_fault;
1005 } else {
1006 int merge = 0;
1007 int i = skb_shinfo(skb)->nr_frags;
1008 struct page *page = TCP_PAGE(sk);
1009 int off = TCP_OFF(sk);
1011 if (skb_can_coalesce(skb, i, page, off) &&
1012 off != PAGE_SIZE) {
1013 /* We can extend the last page
1014 * fragment. */
1015 merge = 1;
1016 } else if (i == MAX_SKB_FRAGS || !sg) {
1017 /* Need to add new fragment and cannot
1018 * do this because interface is non-SG,
1019 * or because all the page slots are
1020 * busy. */
1021 tcp_mark_push(tp, skb);
1022 goto new_segment;
1023 } else if (page) {
1024 if (off == PAGE_SIZE) {
1025 put_page(page);
1026 TCP_PAGE(sk) = page = NULL;
1027 off = 0;
1029 } else
1030 off = 0;
1032 if (copy > PAGE_SIZE - off)
1033 copy = PAGE_SIZE - off;
1035 if (!sk_wmem_schedule(sk, copy))
1036 goto wait_for_memory;
1038 if (!page) {
1039 /* Allocate new cache page. */
1040 if (!(page = sk_stream_alloc_page(sk)))
1041 goto wait_for_memory;
1044 /* Time to copy data. We are close to
1045 * the end! */
1046 err = skb_copy_to_page_nocache(sk, from, skb,
1047 page, off, copy);
1048 if (err) {
1049 /* If this page was new, give it to the
1050 * socket so it does not get leaked.
1052 if (!TCP_PAGE(sk)) {
1053 TCP_PAGE(sk) = page;
1054 TCP_OFF(sk) = 0;
1056 goto do_error;
1059 /* Update the skb. */
1060 if (merge) {
1061 skb_shinfo(skb)->frags[i - 1].size +=
1062 copy;
1063 } else {
1064 skb_fill_page_desc(skb, i, page, off, copy);
1065 if (TCP_PAGE(sk)) {
1066 get_page(page);
1067 } else if (off + copy < PAGE_SIZE) {
1068 get_page(page);
1069 TCP_PAGE(sk) = page;
1073 TCP_OFF(sk) = off + copy;
1076 if (!copied)
1077 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
1079 tp->write_seq += copy;
1080 TCP_SKB_CB(skb)->end_seq += copy;
1081 skb_shinfo(skb)->gso_segs = 0;
1083 from += copy;
1084 copied += copy;
1085 if ((seglen -= copy) == 0 && iovlen == 0)
1086 goto out;
1088 if (skb->len < max || (flags & MSG_OOB))
1089 continue;
1091 if (forced_push(tp)) {
1092 tcp_mark_push(tp, skb);
1093 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1094 } else if (skb == tcp_send_head(sk))
1095 tcp_push_one(sk, mss_now);
1096 continue;
1098 wait_for_sndbuf:
1099 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1100 wait_for_memory:
1101 if (copied)
1102 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1104 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1105 goto do_error;
1107 mss_now = tcp_send_mss(sk, &size_goal, flags);
1111 out:
1112 if (copied)
1113 tcp_push(sk, flags, mss_now, tp->nonagle);
1114 release_sock(sk);
1115 return copied;
1117 do_fault:
1118 if (!skb->len) {
1119 tcp_unlink_write_queue(skb, sk);
1120 /* It is the one place in all of TCP, except connection
1121 * reset, where we can be unlinking the send_head.
1123 tcp_check_send_head(sk, skb);
1124 sk_wmem_free_skb(sk, skb);
1127 do_error:
1128 if (copied)
1129 goto out;
1130 out_err:
1131 err = sk_stream_error(sk, flags, err);
1132 release_sock(sk);
1133 return err;
1135 EXPORT_SYMBOL(tcp_sendmsg);
1138 * Handle reading urgent data. BSD has very simple semantics for
1139 * this, no blocking and very strange errors 8)
1142 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1144 struct tcp_sock *tp = tcp_sk(sk);
1146 /* No URG data to read. */
1147 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1148 tp->urg_data == TCP_URG_READ)
1149 return -EINVAL; /* Yes this is right ! */
1151 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1152 return -ENOTCONN;
1154 if (tp->urg_data & TCP_URG_VALID) {
1155 int err = 0;
1156 char c = tp->urg_data;
1158 if (!(flags & MSG_PEEK))
1159 tp->urg_data = TCP_URG_READ;
1161 /* Read urgent data. */
1162 msg->msg_flags |= MSG_OOB;
1164 if (len > 0) {
1165 if (!(flags & MSG_TRUNC))
1166 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1167 len = 1;
1168 } else
1169 msg->msg_flags |= MSG_TRUNC;
1171 return err ? -EFAULT : len;
1174 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1175 return 0;
1177 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1178 * the available implementations agree in this case:
1179 * this call should never block, independent of the
1180 * blocking state of the socket.
1181 * Mike <pall@rz.uni-karlsruhe.de>
1183 return -EAGAIN;
1186 /* Clean up the receive buffer for full frames taken by the user,
1187 * then send an ACK if necessary. COPIED is the number of bytes
1188 * tcp_recvmsg has given to the user so far, it speeds up the
1189 * calculation of whether or not we must ACK for the sake of
1190 * a window update.
1192 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1194 struct tcp_sock *tp = tcp_sk(sk);
1195 int time_to_ack = 0;
1197 #if TCP_DEBUG
1198 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1200 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1201 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1202 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1203 #endif
1205 if (inet_csk_ack_scheduled(sk)) {
1206 const struct inet_connection_sock *icsk = inet_csk(sk);
1207 /* Delayed ACKs frequently hit locked sockets during bulk
1208 * receive. */
1209 if (icsk->icsk_ack.blocked ||
1210 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1211 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1213 * If this read emptied read buffer, we send ACK, if
1214 * connection is not bidirectional, user drained
1215 * receive buffer and there was a small segment
1216 * in queue.
1218 (copied > 0 &&
1219 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1220 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1221 !icsk->icsk_ack.pingpong)) &&
1222 !atomic_read(&sk->sk_rmem_alloc)))
1223 time_to_ack = 1;
1226 /* We send an ACK if we can now advertise a non-zero window
1227 * which has been raised "significantly".
1229 * Even if window raised up to infinity, do not send window open ACK
1230 * in states, where we will not receive more. It is useless.
1232 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1233 __u32 rcv_window_now = tcp_receive_window(tp);
1235 /* Optimize, __tcp_select_window() is not cheap. */
1236 if (2*rcv_window_now <= tp->window_clamp) {
1237 __u32 new_window = __tcp_select_window(sk);
1239 /* Send ACK now, if this read freed lots of space
1240 * in our buffer. Certainly, new_window is new window.
1241 * We can advertise it now, if it is not less than current one.
1242 * "Lots" means "at least twice" here.
1244 if (new_window && new_window >= 2 * rcv_window_now)
1245 time_to_ack = 1;
1248 if (time_to_ack)
1249 tcp_send_ack(sk);
1252 static void tcp_prequeue_process(struct sock *sk)
1254 struct sk_buff *skb;
1255 struct tcp_sock *tp = tcp_sk(sk);
1257 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1259 /* RX process wants to run with disabled BHs, though it is not
1260 * necessary */
1261 local_bh_disable();
1262 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1263 sk_backlog_rcv(sk, skb);
1264 local_bh_enable();
1266 /* Clear memory counter. */
1267 tp->ucopy.memory = 0;
1270 #ifdef CONFIG_NET_DMA
1271 static void tcp_service_net_dma(struct sock *sk, bool wait)
1273 dma_cookie_t done, used;
1274 dma_cookie_t last_issued;
1275 struct tcp_sock *tp = tcp_sk(sk);
1277 if (!tp->ucopy.dma_chan)
1278 return;
1280 last_issued = tp->ucopy.dma_cookie;
1281 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1283 do {
1284 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1285 last_issued, &done,
1286 &used) == DMA_SUCCESS) {
1287 /* Safe to free early-copied skbs now */
1288 __skb_queue_purge(&sk->sk_async_wait_queue);
1289 break;
1290 } else {
1291 struct sk_buff *skb;
1292 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1293 (dma_async_is_complete(skb->dma_cookie, done,
1294 used) == DMA_SUCCESS)) {
1295 __skb_dequeue(&sk->sk_async_wait_queue);
1296 kfree_skb(skb);
1299 } while (wait);
1301 #endif
1303 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1305 struct sk_buff *skb;
1306 u32 offset;
1308 skb_queue_walk(&sk->sk_receive_queue, skb) {
1309 offset = seq - TCP_SKB_CB(skb)->seq;
1310 if (tcp_hdr(skb)->syn)
1311 offset--;
1312 if (offset < skb->len || tcp_hdr(skb)->fin) {
1313 *off = offset;
1314 return skb;
1317 return NULL;
1321 * This routine provides an alternative to tcp_recvmsg() for routines
1322 * that would like to handle copying from skbuffs directly in 'sendfile'
1323 * fashion.
1324 * Note:
1325 * - It is assumed that the socket was locked by the caller.
1326 * - The routine does not block.
1327 * - At present, there is no support for reading OOB data
1328 * or for 'peeking' the socket using this routine
1329 * (although both would be easy to implement).
1331 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1332 sk_read_actor_t recv_actor)
1334 struct sk_buff *skb;
1335 struct tcp_sock *tp = tcp_sk(sk);
1336 u32 seq = tp->copied_seq;
1337 u32 offset;
1338 int copied = 0;
1340 if (sk->sk_state == TCP_LISTEN)
1341 return -ENOTCONN;
1342 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1343 if (offset < skb->len) {
1344 int used;
1345 size_t len;
1347 len = skb->len - offset;
1348 /* Stop reading if we hit a patch of urgent data */
1349 if (tp->urg_data) {
1350 u32 urg_offset = tp->urg_seq - seq;
1351 if (urg_offset < len)
1352 len = urg_offset;
1353 if (!len)
1354 break;
1356 used = recv_actor(desc, skb, offset, len);
1357 if (used < 0) {
1358 if (!copied)
1359 copied = used;
1360 break;
1361 } else if (used <= len) {
1362 seq += used;
1363 copied += used;
1364 offset += used;
1367 * If recv_actor drops the lock (e.g. TCP splice
1368 * receive) the skb pointer might be invalid when
1369 * getting here: tcp_collapse might have deleted it
1370 * while aggregating skbs from the socket queue.
1372 skb = tcp_recv_skb(sk, seq-1, &offset);
1373 if (!skb || (offset+1 != skb->len))
1374 break;
1376 if (tcp_hdr(skb)->fin) {
1377 sk_eat_skb(sk, skb, 0);
1378 ++seq;
1379 break;
1381 sk_eat_skb(sk, skb, 0);
1382 if (!desc->count)
1383 break;
1384 tp->copied_seq = seq;
1386 tp->copied_seq = seq;
1388 tcp_rcv_space_adjust(sk);
1390 /* Clean up data we have read: This will do ACK frames. */
1391 if (copied > 0)
1392 tcp_cleanup_rbuf(sk, copied);
1393 return copied;
1395 EXPORT_SYMBOL(tcp_read_sock);
1398 * This routine copies from a sock struct into the user buffer.
1400 * Technical note: in 2.3 we work on _locked_ socket, so that
1401 * tricks with *seq access order and skb->users are not required.
1402 * Probably, code can be easily improved even more.
1405 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1406 size_t len, int nonblock, int flags, int *addr_len)
1408 struct tcp_sock *tp = tcp_sk(sk);
1409 int copied = 0;
1410 u32 peek_seq;
1411 u32 *seq;
1412 unsigned long used;
1413 int err;
1414 int target; /* Read at least this many bytes */
1415 long timeo;
1416 struct task_struct *user_recv = NULL;
1417 int copied_early = 0;
1418 struct sk_buff *skb;
1419 u32 urg_hole = 0;
1421 lock_sock(sk);
1423 err = -ENOTCONN;
1424 if (sk->sk_state == TCP_LISTEN)
1425 goto out;
1427 timeo = sock_rcvtimeo(sk, nonblock);
1429 /* Urgent data needs to be handled specially. */
1430 if (flags & MSG_OOB)
1431 goto recv_urg;
1433 seq = &tp->copied_seq;
1434 if (flags & MSG_PEEK) {
1435 peek_seq = tp->copied_seq;
1436 seq = &peek_seq;
1439 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1441 #ifdef CONFIG_NET_DMA
1442 tp->ucopy.dma_chan = NULL;
1443 preempt_disable();
1444 skb = skb_peek_tail(&sk->sk_receive_queue);
1446 int available = 0;
1448 if (skb)
1449 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1450 if ((available < target) &&
1451 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1452 !sysctl_tcp_low_latency &&
1453 dma_find_channel(DMA_MEMCPY)) {
1454 preempt_enable_no_resched();
1455 tp->ucopy.pinned_list =
1456 dma_pin_iovec_pages(msg->msg_iov, len);
1457 } else {
1458 preempt_enable_no_resched();
1461 #endif
1463 do {
1464 u32 offset;
1466 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1467 if (tp->urg_data && tp->urg_seq == *seq) {
1468 if (copied)
1469 break;
1470 if (signal_pending(current)) {
1471 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1472 break;
1476 /* Next get a buffer. */
1478 skb_queue_walk(&sk->sk_receive_queue, skb) {
1479 /* Now that we have two receive queues this
1480 * shouldn't happen.
1482 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1483 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1484 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1485 flags))
1486 break;
1488 offset = *seq - TCP_SKB_CB(skb)->seq;
1489 if (tcp_hdr(skb)->syn)
1490 offset--;
1491 if (offset < skb->len)
1492 goto found_ok_skb;
1493 if (tcp_hdr(skb)->fin)
1494 goto found_fin_ok;
1495 WARN(!(flags & MSG_PEEK),
1496 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1497 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1500 /* Well, if we have backlog, try to process it now yet. */
1502 if (copied >= target && !sk->sk_backlog.tail)
1503 break;
1505 if (copied) {
1506 if (sk->sk_err ||
1507 sk->sk_state == TCP_CLOSE ||
1508 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1509 !timeo ||
1510 signal_pending(current))
1511 break;
1512 } else {
1513 if (sock_flag(sk, SOCK_DONE))
1514 break;
1516 if (sk->sk_err) {
1517 copied = sock_error(sk);
1518 break;
1521 if (sk->sk_shutdown & RCV_SHUTDOWN)
1522 break;
1524 if (sk->sk_state == TCP_CLOSE) {
1525 if (!sock_flag(sk, SOCK_DONE)) {
1526 /* This occurs when user tries to read
1527 * from never connected socket.
1529 copied = -ENOTCONN;
1530 break;
1532 break;
1535 if (!timeo) {
1536 copied = -EAGAIN;
1537 break;
1540 if (signal_pending(current)) {
1541 copied = sock_intr_errno(timeo);
1542 break;
1546 tcp_cleanup_rbuf(sk, copied);
1548 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1549 /* Install new reader */
1550 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1551 user_recv = current;
1552 tp->ucopy.task = user_recv;
1553 tp->ucopy.iov = msg->msg_iov;
1556 tp->ucopy.len = len;
1558 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1559 !(flags & (MSG_PEEK | MSG_TRUNC)));
1561 /* Ugly... If prequeue is not empty, we have to
1562 * process it before releasing socket, otherwise
1563 * order will be broken at second iteration.
1564 * More elegant solution is required!!!
1566 * Look: we have the following (pseudo)queues:
1568 * 1. packets in flight
1569 * 2. backlog
1570 * 3. prequeue
1571 * 4. receive_queue
1573 * Each queue can be processed only if the next ones
1574 * are empty. At this point we have empty receive_queue.
1575 * But prequeue _can_ be not empty after 2nd iteration,
1576 * when we jumped to start of loop because backlog
1577 * processing added something to receive_queue.
1578 * We cannot release_sock(), because backlog contains
1579 * packets arrived _after_ prequeued ones.
1581 * Shortly, algorithm is clear --- to process all
1582 * the queues in order. We could make it more directly,
1583 * requeueing packets from backlog to prequeue, if
1584 * is not empty. It is more elegant, but eats cycles,
1585 * unfortunately.
1587 if (!skb_queue_empty(&tp->ucopy.prequeue))
1588 goto do_prequeue;
1590 /* __ Set realtime policy in scheduler __ */
1593 #ifdef CONFIG_NET_DMA
1594 if (tp->ucopy.dma_chan)
1595 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1596 #endif
1597 if (copied >= target) {
1598 /* Do not sleep, just process backlog. */
1599 release_sock(sk);
1600 lock_sock(sk);
1601 } else
1602 sk_wait_data(sk, &timeo);
1604 #ifdef CONFIG_NET_DMA
1605 tcp_service_net_dma(sk, false); /* Don't block */
1606 tp->ucopy.wakeup = 0;
1607 #endif
1609 if (user_recv) {
1610 int chunk;
1612 /* __ Restore normal policy in scheduler __ */
1614 if ((chunk = len - tp->ucopy.len) != 0) {
1615 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1616 len -= chunk;
1617 copied += chunk;
1620 if (tp->rcv_nxt == tp->copied_seq &&
1621 !skb_queue_empty(&tp->ucopy.prequeue)) {
1622 do_prequeue:
1623 tcp_prequeue_process(sk);
1625 if ((chunk = len - tp->ucopy.len) != 0) {
1626 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1627 len -= chunk;
1628 copied += chunk;
1632 if ((flags & MSG_PEEK) &&
1633 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1634 if (net_ratelimit())
1635 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1636 current->comm, task_pid_nr(current));
1637 peek_seq = tp->copied_seq;
1639 continue;
1641 found_ok_skb:
1642 /* Ok so how much can we use? */
1643 used = skb->len - offset;
1644 if (len < used)
1645 used = len;
1647 /* Do we have urgent data here? */
1648 if (tp->urg_data) {
1649 u32 urg_offset = tp->urg_seq - *seq;
1650 if (urg_offset < used) {
1651 if (!urg_offset) {
1652 if (!sock_flag(sk, SOCK_URGINLINE)) {
1653 ++*seq;
1654 urg_hole++;
1655 offset++;
1656 used--;
1657 if (!used)
1658 goto skip_copy;
1660 } else
1661 used = urg_offset;
1665 if (!(flags & MSG_TRUNC)) {
1666 #ifdef CONFIG_NET_DMA
1667 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1668 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1670 if (tp->ucopy.dma_chan) {
1671 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1672 tp->ucopy.dma_chan, skb, offset,
1673 msg->msg_iov, used,
1674 tp->ucopy.pinned_list);
1676 if (tp->ucopy.dma_cookie < 0) {
1678 printk(KERN_ALERT "dma_cookie < 0\n");
1680 /* Exception. Bailout! */
1681 if (!copied)
1682 copied = -EFAULT;
1683 break;
1686 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1688 if ((offset + used) == skb->len)
1689 copied_early = 1;
1691 } else
1692 #endif
1694 err = skb_copy_datagram_iovec(skb, offset,
1695 msg->msg_iov, used);
1696 if (err) {
1697 /* Exception. Bailout! */
1698 if (!copied)
1699 copied = -EFAULT;
1700 break;
1705 *seq += used;
1706 copied += used;
1707 len -= used;
1709 tcp_rcv_space_adjust(sk);
1711 skip_copy:
1712 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1713 tp->urg_data = 0;
1714 tcp_fast_path_check(sk);
1716 if (used + offset < skb->len)
1717 continue;
1719 if (tcp_hdr(skb)->fin)
1720 goto found_fin_ok;
1721 if (!(flags & MSG_PEEK)) {
1722 sk_eat_skb(sk, skb, copied_early);
1723 copied_early = 0;
1725 continue;
1727 found_fin_ok:
1728 /* Process the FIN. */
1729 ++*seq;
1730 if (!(flags & MSG_PEEK)) {
1731 sk_eat_skb(sk, skb, copied_early);
1732 copied_early = 0;
1734 break;
1735 } while (len > 0);
1737 if (user_recv) {
1738 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1739 int chunk;
1741 tp->ucopy.len = copied > 0 ? len : 0;
1743 tcp_prequeue_process(sk);
1745 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1746 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1747 len -= chunk;
1748 copied += chunk;
1752 tp->ucopy.task = NULL;
1753 tp->ucopy.len = 0;
1756 #ifdef CONFIG_NET_DMA
1757 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1758 tp->ucopy.dma_chan = NULL;
1760 if (tp->ucopy.pinned_list) {
1761 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1762 tp->ucopy.pinned_list = NULL;
1764 #endif
1766 /* According to UNIX98, msg_name/msg_namelen are ignored
1767 * on connected socket. I was just happy when found this 8) --ANK
1770 /* Clean up data we have read: This will do ACK frames. */
1771 tcp_cleanup_rbuf(sk, copied);
1773 release_sock(sk);
1774 return copied;
1776 out:
1777 release_sock(sk);
1778 return err;
1780 recv_urg:
1781 err = tcp_recv_urg(sk, msg, len, flags);
1782 goto out;
1784 EXPORT_SYMBOL(tcp_recvmsg);
1786 void tcp_set_state(struct sock *sk, int state)
1788 int oldstate = sk->sk_state;
1790 switch (state) {
1791 case TCP_ESTABLISHED:
1792 if (oldstate != TCP_ESTABLISHED)
1793 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1794 break;
1796 case TCP_CLOSE:
1797 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1798 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1800 sk->sk_prot->unhash(sk);
1801 if (inet_csk(sk)->icsk_bind_hash &&
1802 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1803 inet_put_port(sk);
1804 /* fall through */
1805 default:
1806 if (oldstate == TCP_ESTABLISHED)
1807 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1810 /* Change state AFTER socket is unhashed to avoid closed
1811 * socket sitting in hash tables.
1813 sk->sk_state = state;
1815 #ifdef STATE_TRACE
1816 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1817 #endif
1819 EXPORT_SYMBOL_GPL(tcp_set_state);
1822 * State processing on a close. This implements the state shift for
1823 * sending our FIN frame. Note that we only send a FIN for some
1824 * states. A shutdown() may have already sent the FIN, or we may be
1825 * closed.
1828 static const unsigned char new_state[16] = {
1829 /* current state: new state: action: */
1830 /* (Invalid) */ TCP_CLOSE,
1831 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1832 /* TCP_SYN_SENT */ TCP_CLOSE,
1833 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1834 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1835 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1836 /* TCP_TIME_WAIT */ TCP_CLOSE,
1837 /* TCP_CLOSE */ TCP_CLOSE,
1838 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1839 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1840 /* TCP_LISTEN */ TCP_CLOSE,
1841 /* TCP_CLOSING */ TCP_CLOSING,
1844 static int tcp_close_state(struct sock *sk)
1846 int next = (int)new_state[sk->sk_state];
1847 int ns = next & TCP_STATE_MASK;
1849 tcp_set_state(sk, ns);
1851 return next & TCP_ACTION_FIN;
1855 * Shutdown the sending side of a connection. Much like close except
1856 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1859 void tcp_shutdown(struct sock *sk, int how)
1861 /* We need to grab some memory, and put together a FIN,
1862 * and then put it into the queue to be sent.
1863 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1865 if (!(how & SEND_SHUTDOWN))
1866 return;
1868 /* If we've already sent a FIN, or it's a closed state, skip this. */
1869 if ((1 << sk->sk_state) &
1870 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1871 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1872 /* Clear out any half completed packets. FIN if needed. */
1873 if (tcp_close_state(sk))
1874 tcp_send_fin(sk);
1877 EXPORT_SYMBOL(tcp_shutdown);
1879 void tcp_close(struct sock *sk, long timeout)
1881 struct sk_buff *skb;
1882 int data_was_unread = 0;
1883 int state;
1885 lock_sock(sk);
1886 sk->sk_shutdown = SHUTDOWN_MASK;
1888 if (sk->sk_state == TCP_LISTEN) {
1889 tcp_set_state(sk, TCP_CLOSE);
1891 /* Special case. */
1892 inet_csk_listen_stop(sk);
1894 goto adjudge_to_death;
1897 /* We need to flush the recv. buffs. We do this only on the
1898 * descriptor close, not protocol-sourced closes, because the
1899 * reader process may not have drained the data yet!
1901 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1902 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1903 tcp_hdr(skb)->fin;
1904 data_was_unread += len;
1905 __kfree_skb(skb);
1908 sk_mem_reclaim(sk);
1910 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1911 if (sk->sk_state == TCP_CLOSE)
1912 goto adjudge_to_death;
1914 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1915 * data was lost. To witness the awful effects of the old behavior of
1916 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1917 * GET in an FTP client, suspend the process, wait for the client to
1918 * advertise a zero window, then kill -9 the FTP client, wheee...
1919 * Note: timeout is always zero in such a case.
1921 if (data_was_unread) {
1922 /* Unread data was tossed, zap the connection. */
1923 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1924 tcp_set_state(sk, TCP_CLOSE);
1925 tcp_send_active_reset(sk, sk->sk_allocation);
1926 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1927 /* Check zero linger _after_ checking for unread data. */
1928 sk->sk_prot->disconnect(sk, 0);
1929 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1930 } else if (tcp_close_state(sk)) {
1931 /* We FIN if the application ate all the data before
1932 * zapping the connection.
1935 /* RED-PEN. Formally speaking, we have broken TCP state
1936 * machine. State transitions:
1938 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1939 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1940 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1942 * are legal only when FIN has been sent (i.e. in window),
1943 * rather than queued out of window. Purists blame.
1945 * F.e. "RFC state" is ESTABLISHED,
1946 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1948 * The visible declinations are that sometimes
1949 * we enter time-wait state, when it is not required really
1950 * (harmless), do not send active resets, when they are
1951 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1952 * they look as CLOSING or LAST_ACK for Linux)
1953 * Probably, I missed some more holelets.
1954 * --ANK
1956 tcp_send_fin(sk);
1959 sk_stream_wait_close(sk, timeout);
1961 adjudge_to_death:
1962 state = sk->sk_state;
1963 sock_hold(sk);
1964 sock_orphan(sk);
1966 /* It is the last release_sock in its life. It will remove backlog. */
1967 release_sock(sk);
1970 /* Now socket is owned by kernel and we acquire BH lock
1971 to finish close. No need to check for user refs.
1973 local_bh_disable();
1974 bh_lock_sock(sk);
1975 WARN_ON(sock_owned_by_user(sk));
1977 percpu_counter_inc(sk->sk_prot->orphan_count);
1979 /* Have we already been destroyed by a softirq or backlog? */
1980 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1981 goto out;
1983 /* This is a (useful) BSD violating of the RFC. There is a
1984 * problem with TCP as specified in that the other end could
1985 * keep a socket open forever with no application left this end.
1986 * We use a 3 minute timeout (about the same as BSD) then kill
1987 * our end. If they send after that then tough - BUT: long enough
1988 * that we won't make the old 4*rto = almost no time - whoops
1989 * reset mistake.
1991 * Nope, it was not mistake. It is really desired behaviour
1992 * f.e. on http servers, when such sockets are useless, but
1993 * consume significant resources. Let's do it with special
1994 * linger2 option. --ANK
1997 if (sk->sk_state == TCP_FIN_WAIT2) {
1998 struct tcp_sock *tp = tcp_sk(sk);
1999 if (tp->linger2 < 0) {
2000 tcp_set_state(sk, TCP_CLOSE);
2001 tcp_send_active_reset(sk, GFP_ATOMIC);
2002 NET_INC_STATS_BH(sock_net(sk),
2003 LINUX_MIB_TCPABORTONLINGER);
2004 } else {
2005 const int tmo = tcp_fin_time(sk);
2007 if (tmo > TCP_TIMEWAIT_LEN) {
2008 inet_csk_reset_keepalive_timer(sk,
2009 tmo - TCP_TIMEWAIT_LEN);
2010 } else {
2011 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2012 goto out;
2016 if (sk->sk_state != TCP_CLOSE) {
2017 sk_mem_reclaim(sk);
2018 if (tcp_too_many_orphans(sk, 0)) {
2019 if (net_ratelimit())
2020 printk(KERN_INFO "TCP: too many of orphaned "
2021 "sockets\n");
2022 tcp_set_state(sk, TCP_CLOSE);
2023 tcp_send_active_reset(sk, GFP_ATOMIC);
2024 NET_INC_STATS_BH(sock_net(sk),
2025 LINUX_MIB_TCPABORTONMEMORY);
2029 if (sk->sk_state == TCP_CLOSE)
2030 inet_csk_destroy_sock(sk);
2031 /* Otherwise, socket is reprieved until protocol close. */
2033 out:
2034 bh_unlock_sock(sk);
2035 local_bh_enable();
2036 sock_put(sk);
2038 EXPORT_SYMBOL(tcp_close);
2040 /* These states need RST on ABORT according to RFC793 */
2042 static inline int tcp_need_reset(int state)
2044 return (1 << state) &
2045 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2046 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2049 int tcp_disconnect(struct sock *sk, int flags)
2051 struct inet_sock *inet = inet_sk(sk);
2052 struct inet_connection_sock *icsk = inet_csk(sk);
2053 struct tcp_sock *tp = tcp_sk(sk);
2054 int err = 0;
2055 int old_state = sk->sk_state;
2057 if (old_state != TCP_CLOSE)
2058 tcp_set_state(sk, TCP_CLOSE);
2060 /* ABORT function of RFC793 */
2061 if (old_state == TCP_LISTEN) {
2062 inet_csk_listen_stop(sk);
2063 } else if (tcp_need_reset(old_state) ||
2064 (tp->snd_nxt != tp->write_seq &&
2065 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2066 /* The last check adjusts for discrepancy of Linux wrt. RFC
2067 * states
2069 tcp_send_active_reset(sk, gfp_any());
2070 sk->sk_err = ECONNRESET;
2071 } else if (old_state == TCP_SYN_SENT)
2072 sk->sk_err = ECONNRESET;
2074 tcp_clear_xmit_timers(sk);
2075 __skb_queue_purge(&sk->sk_receive_queue);
2076 tcp_write_queue_purge(sk);
2077 __skb_queue_purge(&tp->out_of_order_queue);
2078 #ifdef CONFIG_NET_DMA
2079 __skb_queue_purge(&sk->sk_async_wait_queue);
2080 #endif
2082 inet->inet_dport = 0;
2084 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2085 inet_reset_saddr(sk);
2087 sk->sk_shutdown = 0;
2088 sock_reset_flag(sk, SOCK_DONE);
2089 tp->srtt = 0;
2090 if ((tp->write_seq += tp->max_window + 2) == 0)
2091 tp->write_seq = 1;
2092 icsk->icsk_backoff = 0;
2093 tp->snd_cwnd = 2;
2094 icsk->icsk_probes_out = 0;
2095 tp->packets_out = 0;
2096 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2097 tp->snd_cwnd_cnt = 0;
2098 tp->bytes_acked = 0;
2099 tp->window_clamp = 0;
2100 tcp_set_ca_state(sk, TCP_CA_Open);
2101 tcp_clear_retrans(tp);
2102 inet_csk_delack_init(sk);
2103 tcp_init_send_head(sk);
2104 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2105 __sk_dst_reset(sk);
2107 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2109 sk->sk_error_report(sk);
2110 return err;
2112 EXPORT_SYMBOL(tcp_disconnect);
2115 * Socket option code for TCP.
2117 static int do_tcp_setsockopt(struct sock *sk, int level,
2118 int optname, char __user *optval, unsigned int optlen)
2120 struct tcp_sock *tp = tcp_sk(sk);
2121 struct inet_connection_sock *icsk = inet_csk(sk);
2122 int val;
2123 int err = 0;
2125 /* These are data/string values, all the others are ints */
2126 switch (optname) {
2127 case TCP_CONGESTION: {
2128 char name[TCP_CA_NAME_MAX];
2130 if (optlen < 1)
2131 return -EINVAL;
2133 val = strncpy_from_user(name, optval,
2134 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2135 if (val < 0)
2136 return -EFAULT;
2137 name[val] = 0;
2139 lock_sock(sk);
2140 err = tcp_set_congestion_control(sk, name);
2141 release_sock(sk);
2142 return err;
2144 case TCP_COOKIE_TRANSACTIONS: {
2145 struct tcp_cookie_transactions ctd;
2146 struct tcp_cookie_values *cvp = NULL;
2148 if (sizeof(ctd) > optlen)
2149 return -EINVAL;
2150 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2151 return -EFAULT;
2153 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2154 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2155 return -EINVAL;
2157 if (ctd.tcpct_cookie_desired == 0) {
2158 /* default to global value */
2159 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2160 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2161 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2162 return -EINVAL;
2165 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2166 /* Supercedes all other values */
2167 lock_sock(sk);
2168 if (tp->cookie_values != NULL) {
2169 kref_put(&tp->cookie_values->kref,
2170 tcp_cookie_values_release);
2171 tp->cookie_values = NULL;
2173 tp->rx_opt.cookie_in_always = 0; /* false */
2174 tp->rx_opt.cookie_out_never = 1; /* true */
2175 release_sock(sk);
2176 return err;
2179 /* Allocate ancillary memory before locking.
2181 if (ctd.tcpct_used > 0 ||
2182 (tp->cookie_values == NULL &&
2183 (sysctl_tcp_cookie_size > 0 ||
2184 ctd.tcpct_cookie_desired > 0 ||
2185 ctd.tcpct_s_data_desired > 0))) {
2186 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2187 GFP_KERNEL);
2188 if (cvp == NULL)
2189 return -ENOMEM;
2191 kref_init(&cvp->kref);
2193 lock_sock(sk);
2194 tp->rx_opt.cookie_in_always =
2195 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2196 tp->rx_opt.cookie_out_never = 0; /* false */
2198 if (tp->cookie_values != NULL) {
2199 if (cvp != NULL) {
2200 /* Changed values are recorded by a changed
2201 * pointer, ensuring the cookie will differ,
2202 * without separately hashing each value later.
2204 kref_put(&tp->cookie_values->kref,
2205 tcp_cookie_values_release);
2206 } else {
2207 cvp = tp->cookie_values;
2211 if (cvp != NULL) {
2212 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2214 if (ctd.tcpct_used > 0) {
2215 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2216 ctd.tcpct_used);
2217 cvp->s_data_desired = ctd.tcpct_used;
2218 cvp->s_data_constant = 1; /* true */
2219 } else {
2220 /* No constant payload data. */
2221 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2222 cvp->s_data_constant = 0; /* false */
2225 tp->cookie_values = cvp;
2227 release_sock(sk);
2228 return err;
2230 default:
2231 /* fallthru */
2232 break;
2235 if (optlen < sizeof(int))
2236 return -EINVAL;
2238 if (get_user(val, (int __user *)optval))
2239 return -EFAULT;
2241 lock_sock(sk);
2243 switch (optname) {
2244 case TCP_MAXSEG:
2245 /* Values greater than interface MTU won't take effect. However
2246 * at the point when this call is done we typically don't yet
2247 * know which interface is going to be used */
2248 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2249 err = -EINVAL;
2250 break;
2252 tp->rx_opt.user_mss = val;
2253 break;
2255 case TCP_NODELAY:
2256 if (val) {
2257 /* TCP_NODELAY is weaker than TCP_CORK, so that
2258 * this option on corked socket is remembered, but
2259 * it is not activated until cork is cleared.
2261 * However, when TCP_NODELAY is set we make
2262 * an explicit push, which overrides even TCP_CORK
2263 * for currently queued segments.
2265 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2266 tcp_push_pending_frames(sk);
2267 } else {
2268 tp->nonagle &= ~TCP_NAGLE_OFF;
2270 break;
2272 case TCP_THIN_LINEAR_TIMEOUTS:
2273 if (val < 0 || val > 1)
2274 err = -EINVAL;
2275 else
2276 tp->thin_lto = val;
2277 break;
2279 case TCP_THIN_DUPACK:
2280 if (val < 0 || val > 1)
2281 err = -EINVAL;
2282 else
2283 tp->thin_dupack = val;
2284 break;
2286 case TCP_CORK:
2287 /* When set indicates to always queue non-full frames.
2288 * Later the user clears this option and we transmit
2289 * any pending partial frames in the queue. This is
2290 * meant to be used alongside sendfile() to get properly
2291 * filled frames when the user (for example) must write
2292 * out headers with a write() call first and then use
2293 * sendfile to send out the data parts.
2295 * TCP_CORK can be set together with TCP_NODELAY and it is
2296 * stronger than TCP_NODELAY.
2298 if (val) {
2299 tp->nonagle |= TCP_NAGLE_CORK;
2300 } else {
2301 tp->nonagle &= ~TCP_NAGLE_CORK;
2302 if (tp->nonagle&TCP_NAGLE_OFF)
2303 tp->nonagle |= TCP_NAGLE_PUSH;
2304 tcp_push_pending_frames(sk);
2306 break;
2308 case TCP_KEEPIDLE:
2309 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2310 err = -EINVAL;
2311 else {
2312 tp->keepalive_time = val * HZ;
2313 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2314 !((1 << sk->sk_state) &
2315 (TCPF_CLOSE | TCPF_LISTEN))) {
2316 u32 elapsed = keepalive_time_elapsed(tp);
2317 if (tp->keepalive_time > elapsed)
2318 elapsed = tp->keepalive_time - elapsed;
2319 else
2320 elapsed = 0;
2321 inet_csk_reset_keepalive_timer(sk, elapsed);
2324 break;
2325 case TCP_KEEPINTVL:
2326 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2327 err = -EINVAL;
2328 else
2329 tp->keepalive_intvl = val * HZ;
2330 break;
2331 case TCP_KEEPCNT:
2332 if (val < 1 || val > MAX_TCP_KEEPCNT)
2333 err = -EINVAL;
2334 else
2335 tp->keepalive_probes = val;
2336 break;
2337 case TCP_SYNCNT:
2338 if (val < 1 || val > MAX_TCP_SYNCNT)
2339 err = -EINVAL;
2340 else
2341 icsk->icsk_syn_retries = val;
2342 break;
2344 case TCP_LINGER2:
2345 if (val < 0)
2346 tp->linger2 = -1;
2347 else if (val > sysctl_tcp_fin_timeout / HZ)
2348 tp->linger2 = 0;
2349 else
2350 tp->linger2 = val * HZ;
2351 break;
2353 case TCP_DEFER_ACCEPT:
2354 /* Translate value in seconds to number of retransmits */
2355 icsk->icsk_accept_queue.rskq_defer_accept =
2356 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2357 TCP_RTO_MAX / HZ);
2358 break;
2360 case TCP_WINDOW_CLAMP:
2361 if (!val) {
2362 if (sk->sk_state != TCP_CLOSE) {
2363 err = -EINVAL;
2364 break;
2366 tp->window_clamp = 0;
2367 } else
2368 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2369 SOCK_MIN_RCVBUF / 2 : val;
2370 break;
2372 case TCP_QUICKACK:
2373 if (!val) {
2374 icsk->icsk_ack.pingpong = 1;
2375 } else {
2376 icsk->icsk_ack.pingpong = 0;
2377 if ((1 << sk->sk_state) &
2378 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2379 inet_csk_ack_scheduled(sk)) {
2380 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2381 tcp_cleanup_rbuf(sk, 1);
2382 if (!(val & 1))
2383 icsk->icsk_ack.pingpong = 1;
2386 break;
2388 #ifdef CONFIG_TCP_MD5SIG
2389 case TCP_MD5SIG:
2390 /* Read the IP->Key mappings from userspace */
2391 err = tp->af_specific->md5_parse(sk, optval, optlen);
2392 break;
2393 #endif
2394 case TCP_USER_TIMEOUT:
2395 /* Cap the max timeout in ms TCP will retry/retrans
2396 * before giving up and aborting (ETIMEDOUT) a connection.
2398 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2399 break;
2400 default:
2401 err = -ENOPROTOOPT;
2402 break;
2405 release_sock(sk);
2406 return err;
2409 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2410 unsigned int optlen)
2412 struct inet_connection_sock *icsk = inet_csk(sk);
2414 if (level != SOL_TCP)
2415 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2416 optval, optlen);
2417 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2419 EXPORT_SYMBOL(tcp_setsockopt);
2421 #ifdef CONFIG_COMPAT
2422 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2423 char __user *optval, unsigned int optlen)
2425 if (level != SOL_TCP)
2426 return inet_csk_compat_setsockopt(sk, level, optname,
2427 optval, optlen);
2428 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2430 EXPORT_SYMBOL(compat_tcp_setsockopt);
2431 #endif
2433 /* Return information about state of tcp endpoint in API format. */
2434 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2436 struct tcp_sock *tp = tcp_sk(sk);
2437 const struct inet_connection_sock *icsk = inet_csk(sk);
2438 u32 now = tcp_time_stamp;
2440 memset(info, 0, sizeof(*info));
2442 info->tcpi_state = sk->sk_state;
2443 info->tcpi_ca_state = icsk->icsk_ca_state;
2444 info->tcpi_retransmits = icsk->icsk_retransmits;
2445 info->tcpi_probes = icsk->icsk_probes_out;
2446 info->tcpi_backoff = icsk->icsk_backoff;
2448 if (tp->rx_opt.tstamp_ok)
2449 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2450 if (tcp_is_sack(tp))
2451 info->tcpi_options |= TCPI_OPT_SACK;
2452 if (tp->rx_opt.wscale_ok) {
2453 info->tcpi_options |= TCPI_OPT_WSCALE;
2454 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2455 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2458 if (tp->ecn_flags&TCP_ECN_OK)
2459 info->tcpi_options |= TCPI_OPT_ECN;
2461 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2462 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2463 info->tcpi_snd_mss = tp->mss_cache;
2464 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2466 if (sk->sk_state == TCP_LISTEN) {
2467 info->tcpi_unacked = sk->sk_ack_backlog;
2468 info->tcpi_sacked = sk->sk_max_ack_backlog;
2469 } else {
2470 info->tcpi_unacked = tp->packets_out;
2471 info->tcpi_sacked = tp->sacked_out;
2473 info->tcpi_lost = tp->lost_out;
2474 info->tcpi_retrans = tp->retrans_out;
2475 info->tcpi_fackets = tp->fackets_out;
2477 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2478 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2479 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2481 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2482 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2483 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2484 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2485 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2486 info->tcpi_snd_cwnd = tp->snd_cwnd;
2487 info->tcpi_advmss = tp->advmss;
2488 info->tcpi_reordering = tp->reordering;
2490 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2491 info->tcpi_rcv_space = tp->rcvq_space.space;
2493 info->tcpi_total_retrans = tp->total_retrans;
2495 EXPORT_SYMBOL_GPL(tcp_get_info);
2497 static int do_tcp_getsockopt(struct sock *sk, int level,
2498 int optname, char __user *optval, int __user *optlen)
2500 struct inet_connection_sock *icsk = inet_csk(sk);
2501 struct tcp_sock *tp = tcp_sk(sk);
2502 int val, len;
2504 if (get_user(len, optlen))
2505 return -EFAULT;
2507 len = min_t(unsigned int, len, sizeof(int));
2509 if (len < 0)
2510 return -EINVAL;
2512 switch (optname) {
2513 case TCP_MAXSEG:
2514 val = tp->mss_cache;
2515 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2516 val = tp->rx_opt.user_mss;
2517 break;
2518 case TCP_NODELAY:
2519 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2520 break;
2521 case TCP_CORK:
2522 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2523 break;
2524 case TCP_KEEPIDLE:
2525 val = keepalive_time_when(tp) / HZ;
2526 break;
2527 case TCP_KEEPINTVL:
2528 val = keepalive_intvl_when(tp) / HZ;
2529 break;
2530 case TCP_KEEPCNT:
2531 val = keepalive_probes(tp);
2532 break;
2533 case TCP_SYNCNT:
2534 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2535 break;
2536 case TCP_LINGER2:
2537 val = tp->linger2;
2538 if (val >= 0)
2539 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2540 break;
2541 case TCP_DEFER_ACCEPT:
2542 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2543 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2544 break;
2545 case TCP_WINDOW_CLAMP:
2546 val = tp->window_clamp;
2547 break;
2548 case TCP_INFO: {
2549 struct tcp_info info;
2551 if (get_user(len, optlen))
2552 return -EFAULT;
2554 tcp_get_info(sk, &info);
2556 len = min_t(unsigned int, len, sizeof(info));
2557 if (put_user(len, optlen))
2558 return -EFAULT;
2559 if (copy_to_user(optval, &info, len))
2560 return -EFAULT;
2561 return 0;
2563 case TCP_QUICKACK:
2564 val = !icsk->icsk_ack.pingpong;
2565 break;
2567 case TCP_CONGESTION:
2568 if (get_user(len, optlen))
2569 return -EFAULT;
2570 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2571 if (put_user(len, optlen))
2572 return -EFAULT;
2573 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2574 return -EFAULT;
2575 return 0;
2577 case TCP_COOKIE_TRANSACTIONS: {
2578 struct tcp_cookie_transactions ctd;
2579 struct tcp_cookie_values *cvp = tp->cookie_values;
2581 if (get_user(len, optlen))
2582 return -EFAULT;
2583 if (len < sizeof(ctd))
2584 return -EINVAL;
2586 memset(&ctd, 0, sizeof(ctd));
2587 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2588 TCP_COOKIE_IN_ALWAYS : 0)
2589 | (tp->rx_opt.cookie_out_never ?
2590 TCP_COOKIE_OUT_NEVER : 0);
2592 if (cvp != NULL) {
2593 ctd.tcpct_flags |= (cvp->s_data_in ?
2594 TCP_S_DATA_IN : 0)
2595 | (cvp->s_data_out ?
2596 TCP_S_DATA_OUT : 0);
2598 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2599 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2601 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2602 cvp->cookie_pair_size);
2603 ctd.tcpct_used = cvp->cookie_pair_size;
2606 if (put_user(sizeof(ctd), optlen))
2607 return -EFAULT;
2608 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2609 return -EFAULT;
2610 return 0;
2612 case TCP_THIN_LINEAR_TIMEOUTS:
2613 val = tp->thin_lto;
2614 break;
2615 case TCP_THIN_DUPACK:
2616 val = tp->thin_dupack;
2617 break;
2619 case TCP_USER_TIMEOUT:
2620 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2621 break;
2622 default:
2623 return -ENOPROTOOPT;
2626 if (put_user(len, optlen))
2627 return -EFAULT;
2628 if (copy_to_user(optval, &val, len))
2629 return -EFAULT;
2630 return 0;
2633 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2634 int __user *optlen)
2636 struct inet_connection_sock *icsk = inet_csk(sk);
2638 if (level != SOL_TCP)
2639 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2640 optval, optlen);
2641 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2643 EXPORT_SYMBOL(tcp_getsockopt);
2645 #ifdef CONFIG_COMPAT
2646 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2647 char __user *optval, int __user *optlen)
2649 if (level != SOL_TCP)
2650 return inet_csk_compat_getsockopt(sk, level, optname,
2651 optval, optlen);
2652 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2654 EXPORT_SYMBOL(compat_tcp_getsockopt);
2655 #endif
2657 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, u32 features)
2659 struct sk_buff *segs = ERR_PTR(-EINVAL);
2660 struct tcphdr *th;
2661 unsigned thlen;
2662 unsigned int seq;
2663 __be32 delta;
2664 unsigned int oldlen;
2665 unsigned int mss;
2667 if (!pskb_may_pull(skb, sizeof(*th)))
2668 goto out;
2670 th = tcp_hdr(skb);
2671 thlen = th->doff * 4;
2672 if (thlen < sizeof(*th))
2673 goto out;
2675 if (!pskb_may_pull(skb, thlen))
2676 goto out;
2678 oldlen = (u16)~skb->len;
2679 __skb_pull(skb, thlen);
2681 mss = skb_shinfo(skb)->gso_size;
2682 if (unlikely(skb->len <= mss))
2683 goto out;
2685 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2686 /* Packet is from an untrusted source, reset gso_segs. */
2687 int type = skb_shinfo(skb)->gso_type;
2689 if (unlikely(type &
2690 ~(SKB_GSO_TCPV4 |
2691 SKB_GSO_DODGY |
2692 SKB_GSO_TCP_ECN |
2693 SKB_GSO_TCPV6 |
2694 0) ||
2695 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2696 goto out;
2698 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2700 segs = NULL;
2701 goto out;
2704 segs = skb_segment(skb, features);
2705 if (IS_ERR(segs))
2706 goto out;
2708 delta = htonl(oldlen + (thlen + mss));
2710 skb = segs;
2711 th = tcp_hdr(skb);
2712 seq = ntohl(th->seq);
2714 do {
2715 th->fin = th->psh = 0;
2717 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2718 (__force u32)delta));
2719 if (skb->ip_summed != CHECKSUM_PARTIAL)
2720 th->check =
2721 csum_fold(csum_partial(skb_transport_header(skb),
2722 thlen, skb->csum));
2724 seq += mss;
2725 skb = skb->next;
2726 th = tcp_hdr(skb);
2728 th->seq = htonl(seq);
2729 th->cwr = 0;
2730 } while (skb->next);
2732 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2733 skb->data_len);
2734 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2735 (__force u32)delta));
2736 if (skb->ip_summed != CHECKSUM_PARTIAL)
2737 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2738 thlen, skb->csum));
2740 out:
2741 return segs;
2743 EXPORT_SYMBOL(tcp_tso_segment);
2745 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2747 struct sk_buff **pp = NULL;
2748 struct sk_buff *p;
2749 struct tcphdr *th;
2750 struct tcphdr *th2;
2751 unsigned int len;
2752 unsigned int thlen;
2753 __be32 flags;
2754 unsigned int mss = 1;
2755 unsigned int hlen;
2756 unsigned int off;
2757 int flush = 1;
2758 int i;
2760 off = skb_gro_offset(skb);
2761 hlen = off + sizeof(*th);
2762 th = skb_gro_header_fast(skb, off);
2763 if (skb_gro_header_hard(skb, hlen)) {
2764 th = skb_gro_header_slow(skb, hlen, off);
2765 if (unlikely(!th))
2766 goto out;
2769 thlen = th->doff * 4;
2770 if (thlen < sizeof(*th))
2771 goto out;
2773 hlen = off + thlen;
2774 if (skb_gro_header_hard(skb, hlen)) {
2775 th = skb_gro_header_slow(skb, hlen, off);
2776 if (unlikely(!th))
2777 goto out;
2780 skb_gro_pull(skb, thlen);
2782 len = skb_gro_len(skb);
2783 flags = tcp_flag_word(th);
2785 for (; (p = *head); head = &p->next) {
2786 if (!NAPI_GRO_CB(p)->same_flow)
2787 continue;
2789 th2 = tcp_hdr(p);
2791 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2792 NAPI_GRO_CB(p)->same_flow = 0;
2793 continue;
2796 goto found;
2799 goto out_check_final;
2801 found:
2802 flush = NAPI_GRO_CB(p)->flush;
2803 flush |= (__force int)(flags & TCP_FLAG_CWR);
2804 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2805 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2806 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2807 for (i = sizeof(*th); i < thlen; i += 4)
2808 flush |= *(u32 *)((u8 *)th + i) ^
2809 *(u32 *)((u8 *)th2 + i);
2811 mss = skb_shinfo(p)->gso_size;
2813 flush |= (len - 1) >= mss;
2814 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2816 if (flush || skb_gro_receive(head, skb)) {
2817 mss = 1;
2818 goto out_check_final;
2821 p = *head;
2822 th2 = tcp_hdr(p);
2823 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2825 out_check_final:
2826 flush = len < mss;
2827 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2828 TCP_FLAG_RST | TCP_FLAG_SYN |
2829 TCP_FLAG_FIN));
2831 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2832 pp = head;
2834 out:
2835 NAPI_GRO_CB(skb)->flush |= flush;
2837 return pp;
2839 EXPORT_SYMBOL(tcp_gro_receive);
2841 int tcp_gro_complete(struct sk_buff *skb)
2843 struct tcphdr *th = tcp_hdr(skb);
2845 skb->csum_start = skb_transport_header(skb) - skb->head;
2846 skb->csum_offset = offsetof(struct tcphdr, check);
2847 skb->ip_summed = CHECKSUM_PARTIAL;
2849 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2851 if (th->cwr)
2852 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2854 return 0;
2856 EXPORT_SYMBOL(tcp_gro_complete);
2858 #ifdef CONFIG_TCP_MD5SIG
2859 static unsigned long tcp_md5sig_users;
2860 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool;
2861 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2863 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool)
2865 int cpu;
2866 for_each_possible_cpu(cpu) {
2867 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2868 if (p) {
2869 if (p->md5_desc.tfm)
2870 crypto_free_hash(p->md5_desc.tfm);
2871 kfree(p);
2874 free_percpu(pool);
2877 void tcp_free_md5sig_pool(void)
2879 struct tcp_md5sig_pool * __percpu *pool = NULL;
2881 spin_lock_bh(&tcp_md5sig_pool_lock);
2882 if (--tcp_md5sig_users == 0) {
2883 pool = tcp_md5sig_pool;
2884 tcp_md5sig_pool = NULL;
2886 spin_unlock_bh(&tcp_md5sig_pool_lock);
2887 if (pool)
2888 __tcp_free_md5sig_pool(pool);
2890 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2892 static struct tcp_md5sig_pool * __percpu *
2893 __tcp_alloc_md5sig_pool(struct sock *sk)
2895 int cpu;
2896 struct tcp_md5sig_pool * __percpu *pool;
2898 pool = alloc_percpu(struct tcp_md5sig_pool *);
2899 if (!pool)
2900 return NULL;
2902 for_each_possible_cpu(cpu) {
2903 struct tcp_md5sig_pool *p;
2904 struct crypto_hash *hash;
2906 p = kzalloc(sizeof(*p), sk->sk_allocation);
2907 if (!p)
2908 goto out_free;
2909 *per_cpu_ptr(pool, cpu) = p;
2911 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2912 if (!hash || IS_ERR(hash))
2913 goto out_free;
2915 p->md5_desc.tfm = hash;
2917 return pool;
2918 out_free:
2919 __tcp_free_md5sig_pool(pool);
2920 return NULL;
2923 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2925 struct tcp_md5sig_pool * __percpu *pool;
2926 int alloc = 0;
2928 retry:
2929 spin_lock_bh(&tcp_md5sig_pool_lock);
2930 pool = tcp_md5sig_pool;
2931 if (tcp_md5sig_users++ == 0) {
2932 alloc = 1;
2933 spin_unlock_bh(&tcp_md5sig_pool_lock);
2934 } else if (!pool) {
2935 tcp_md5sig_users--;
2936 spin_unlock_bh(&tcp_md5sig_pool_lock);
2937 cpu_relax();
2938 goto retry;
2939 } else
2940 spin_unlock_bh(&tcp_md5sig_pool_lock);
2942 if (alloc) {
2943 /* we cannot hold spinlock here because this may sleep. */
2944 struct tcp_md5sig_pool * __percpu *p;
2946 p = __tcp_alloc_md5sig_pool(sk);
2947 spin_lock_bh(&tcp_md5sig_pool_lock);
2948 if (!p) {
2949 tcp_md5sig_users--;
2950 spin_unlock_bh(&tcp_md5sig_pool_lock);
2951 return NULL;
2953 pool = tcp_md5sig_pool;
2954 if (pool) {
2955 /* oops, it has already been assigned. */
2956 spin_unlock_bh(&tcp_md5sig_pool_lock);
2957 __tcp_free_md5sig_pool(p);
2958 } else {
2959 tcp_md5sig_pool = pool = p;
2960 spin_unlock_bh(&tcp_md5sig_pool_lock);
2963 return pool;
2965 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2969 * tcp_get_md5sig_pool - get md5sig_pool for this user
2971 * We use percpu structure, so if we succeed, we exit with preemption
2972 * and BH disabled, to make sure another thread or softirq handling
2973 * wont try to get same context.
2975 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2977 struct tcp_md5sig_pool * __percpu *p;
2979 local_bh_disable();
2981 spin_lock(&tcp_md5sig_pool_lock);
2982 p = tcp_md5sig_pool;
2983 if (p)
2984 tcp_md5sig_users++;
2985 spin_unlock(&tcp_md5sig_pool_lock);
2987 if (p)
2988 return *this_cpu_ptr(p);
2990 local_bh_enable();
2991 return NULL;
2993 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2995 void tcp_put_md5sig_pool(void)
2997 local_bh_enable();
2998 tcp_free_md5sig_pool();
3000 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3002 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3003 struct tcphdr *th)
3005 struct scatterlist sg;
3006 int err;
3008 __sum16 old_checksum = th->check;
3009 th->check = 0;
3010 /* options aren't included in the hash */
3011 sg_init_one(&sg, th, sizeof(struct tcphdr));
3012 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
3013 th->check = old_checksum;
3014 return err;
3016 EXPORT_SYMBOL(tcp_md5_hash_header);
3018 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3019 struct sk_buff *skb, unsigned header_len)
3021 struct scatterlist sg;
3022 const struct tcphdr *tp = tcp_hdr(skb);
3023 struct hash_desc *desc = &hp->md5_desc;
3024 unsigned i;
3025 const unsigned head_data_len = skb_headlen(skb) > header_len ?
3026 skb_headlen(skb) - header_len : 0;
3027 const struct skb_shared_info *shi = skb_shinfo(skb);
3028 struct sk_buff *frag_iter;
3030 sg_init_table(&sg, 1);
3032 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3033 if (crypto_hash_update(desc, &sg, head_data_len))
3034 return 1;
3036 for (i = 0; i < shi->nr_frags; ++i) {
3037 const struct skb_frag_struct *f = &shi->frags[i];
3038 sg_set_page(&sg, f->page, f->size, f->page_offset);
3039 if (crypto_hash_update(desc, &sg, f->size))
3040 return 1;
3043 skb_walk_frags(skb, frag_iter)
3044 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3045 return 1;
3047 return 0;
3049 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3051 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
3053 struct scatterlist sg;
3055 sg_init_one(&sg, key->key, key->keylen);
3056 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3058 EXPORT_SYMBOL(tcp_md5_hash_key);
3060 #endif
3063 * Each Responder maintains up to two secret values concurrently for
3064 * efficient secret rollover. Each secret value has 4 states:
3066 * Generating. (tcp_secret_generating != tcp_secret_primary)
3067 * Generates new Responder-Cookies, but not yet used for primary
3068 * verification. This is a short-term state, typically lasting only
3069 * one round trip time (RTT).
3071 * Primary. (tcp_secret_generating == tcp_secret_primary)
3072 * Used both for generation and primary verification.
3074 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3075 * Used for verification, until the first failure that can be
3076 * verified by the newer Generating secret. At that time, this
3077 * cookie's state is changed to Secondary, and the Generating
3078 * cookie's state is changed to Primary. This is a short-term state,
3079 * typically lasting only one round trip time (RTT).
3081 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3082 * Used for secondary verification, after primary verification
3083 * failures. This state lasts no more than twice the Maximum Segment
3084 * Lifetime (2MSL). Then, the secret is discarded.
3086 struct tcp_cookie_secret {
3087 /* The secret is divided into two parts. The digest part is the
3088 * equivalent of previously hashing a secret and saving the state,
3089 * and serves as an initialization vector (IV). The message part
3090 * serves as the trailing secret.
3092 u32 secrets[COOKIE_WORKSPACE_WORDS];
3093 unsigned long expires;
3096 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3097 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3098 #define TCP_SECRET_LIFE (HZ * 600)
3100 static struct tcp_cookie_secret tcp_secret_one;
3101 static struct tcp_cookie_secret tcp_secret_two;
3103 /* Essentially a circular list, without dynamic allocation. */
3104 static struct tcp_cookie_secret *tcp_secret_generating;
3105 static struct tcp_cookie_secret *tcp_secret_primary;
3106 static struct tcp_cookie_secret *tcp_secret_retiring;
3107 static struct tcp_cookie_secret *tcp_secret_secondary;
3109 static DEFINE_SPINLOCK(tcp_secret_locker);
3111 /* Select a pseudo-random word in the cookie workspace.
3113 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3115 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3118 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3119 * Called in softirq context.
3120 * Returns: 0 for success.
3122 int tcp_cookie_generator(u32 *bakery)
3124 unsigned long jiffy = jiffies;
3126 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3127 spin_lock_bh(&tcp_secret_locker);
3128 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3129 /* refreshed by another */
3130 memcpy(bakery,
3131 &tcp_secret_generating->secrets[0],
3132 COOKIE_WORKSPACE_WORDS);
3133 } else {
3134 /* still needs refreshing */
3135 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3137 /* The first time, paranoia assumes that the
3138 * randomization function isn't as strong. But,
3139 * this secret initialization is delayed until
3140 * the last possible moment (packet arrival).
3141 * Although that time is observable, it is
3142 * unpredictably variable. Mash in the most
3143 * volatile clock bits available, and expire the
3144 * secret extra quickly.
3146 if (unlikely(tcp_secret_primary->expires ==
3147 tcp_secret_secondary->expires)) {
3148 struct timespec tv;
3150 getnstimeofday(&tv);
3151 bakery[COOKIE_DIGEST_WORDS+0] ^=
3152 (u32)tv.tv_nsec;
3154 tcp_secret_secondary->expires = jiffy
3155 + TCP_SECRET_1MSL
3156 + (0x0f & tcp_cookie_work(bakery, 0));
3157 } else {
3158 tcp_secret_secondary->expires = jiffy
3159 + TCP_SECRET_LIFE
3160 + (0xff & tcp_cookie_work(bakery, 1));
3161 tcp_secret_primary->expires = jiffy
3162 + TCP_SECRET_2MSL
3163 + (0x1f & tcp_cookie_work(bakery, 2));
3165 memcpy(&tcp_secret_secondary->secrets[0],
3166 bakery, COOKIE_WORKSPACE_WORDS);
3168 rcu_assign_pointer(tcp_secret_generating,
3169 tcp_secret_secondary);
3170 rcu_assign_pointer(tcp_secret_retiring,
3171 tcp_secret_primary);
3173 * Neither call_rcu() nor synchronize_rcu() needed.
3174 * Retiring data is not freed. It is replaced after
3175 * further (locked) pointer updates, and a quiet time
3176 * (minimum 1MSL, maximum LIFE - 2MSL).
3179 spin_unlock_bh(&tcp_secret_locker);
3180 } else {
3181 rcu_read_lock_bh();
3182 memcpy(bakery,
3183 &rcu_dereference(tcp_secret_generating)->secrets[0],
3184 COOKIE_WORKSPACE_WORDS);
3185 rcu_read_unlock_bh();
3187 return 0;
3189 EXPORT_SYMBOL(tcp_cookie_generator);
3191 void tcp_done(struct sock *sk)
3193 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3194 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3196 tcp_set_state(sk, TCP_CLOSE);
3197 tcp_clear_xmit_timers(sk);
3199 sk->sk_shutdown = SHUTDOWN_MASK;
3201 if (!sock_flag(sk, SOCK_DEAD))
3202 sk->sk_state_change(sk);
3203 else
3204 inet_csk_destroy_sock(sk);
3206 EXPORT_SYMBOL_GPL(tcp_done);
3208 extern struct tcp_congestion_ops tcp_reno;
3210 static __initdata unsigned long thash_entries;
3211 static int __init set_thash_entries(char *str)
3213 if (!str)
3214 return 0;
3215 thash_entries = simple_strtoul(str, &str, 0);
3216 return 1;
3218 __setup("thash_entries=", set_thash_entries);
3220 void __init tcp_init(void)
3222 struct sk_buff *skb = NULL;
3223 unsigned long limit;
3224 int i, max_share, cnt;
3225 unsigned long jiffy = jiffies;
3227 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3229 percpu_counter_init(&tcp_sockets_allocated, 0);
3230 percpu_counter_init(&tcp_orphan_count, 0);
3231 tcp_hashinfo.bind_bucket_cachep =
3232 kmem_cache_create("tcp_bind_bucket",
3233 sizeof(struct inet_bind_bucket), 0,
3234 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3236 /* Size and allocate the main established and bind bucket
3237 * hash tables.
3239 * The methodology is similar to that of the buffer cache.
3241 tcp_hashinfo.ehash =
3242 alloc_large_system_hash("TCP established",
3243 sizeof(struct inet_ehash_bucket),
3244 thash_entries,
3245 (totalram_pages >= 128 * 1024) ?
3246 13 : 15,
3248 NULL,
3249 &tcp_hashinfo.ehash_mask,
3250 thash_entries ? 0 : 512 * 1024);
3251 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3252 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3253 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3255 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3256 panic("TCP: failed to alloc ehash_locks");
3257 tcp_hashinfo.bhash =
3258 alloc_large_system_hash("TCP bind",
3259 sizeof(struct inet_bind_hashbucket),
3260 tcp_hashinfo.ehash_mask + 1,
3261 (totalram_pages >= 128 * 1024) ?
3262 13 : 15,
3264 &tcp_hashinfo.bhash_size,
3265 NULL,
3266 64 * 1024);
3267 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3268 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3269 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3270 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3274 cnt = tcp_hashinfo.ehash_mask + 1;
3276 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3277 sysctl_tcp_max_orphans = cnt / 2;
3278 sysctl_max_syn_backlog = max(128, cnt / 256);
3280 limit = nr_free_buffer_pages() / 8;
3281 limit = max(limit, 128UL);
3282 sysctl_tcp_mem[0] = limit / 4 * 3;
3283 sysctl_tcp_mem[1] = limit;
3284 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3286 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3287 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3288 max_share = min(4UL*1024*1024, limit);
3290 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3291 sysctl_tcp_wmem[1] = 16*1024;
3292 sysctl_tcp_wmem[2] = max(64*1024, max_share);
3294 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3295 sysctl_tcp_rmem[1] = 87380;
3296 sysctl_tcp_rmem[2] = max(87380, max_share);
3298 printk(KERN_INFO "TCP: Hash tables configured "
3299 "(established %u bind %u)\n",
3300 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3302 tcp_register_congestion_control(&tcp_reno);
3304 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3305 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3306 tcp_secret_one.expires = jiffy; /* past due */
3307 tcp_secret_two.expires = jiffy; /* past due */
3308 tcp_secret_generating = &tcp_secret_one;
3309 tcp_secret_primary = &tcp_secret_one;
3310 tcp_secret_retiring = &tcp_secret_two;
3311 tcp_secret_secondary = &tcp_secret_two;