tcp: Revert 'process defer accept as established' changes.
[linux-2.6.git] / net / ipv4 / tcp.c
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.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp.c,v 1.216 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  *
22  * Fixes:
23  *              Alan Cox        :       Numerous verify_area() calls
24  *              Alan Cox        :       Set the ACK bit on a reset
25  *              Alan Cox        :       Stopped it crashing if it closed while
26  *                                      sk->inuse=1 and was trying to connect
27  *                                      (tcp_err()).
28  *              Alan Cox        :       All icmp error handling was broken
29  *                                      pointers passed where wrong and the
30  *                                      socket was looked up backwards. Nobody
31  *                                      tested any icmp error code obviously.
32  *              Alan Cox        :       tcp_err() now handled properly. It
33  *                                      wakes people on errors. poll
34  *                                      behaves and the icmp error race
35  *                                      has gone by moving it into sock.c
36  *              Alan Cox        :       tcp_send_reset() fixed to work for
37  *                                      everything not just packets for
38  *                                      unknown sockets.
39  *              Alan Cox        :       tcp option processing.
40  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
41  *                                      syn rule wrong]
42  *              Herp Rosmanith  :       More reset fixes
43  *              Alan Cox        :       No longer acks invalid rst frames.
44  *                                      Acking any kind of RST is right out.
45  *              Alan Cox        :       Sets an ignore me flag on an rst
46  *                                      receive otherwise odd bits of prattle
47  *                                      escape still
48  *              Alan Cox        :       Fixed another acking RST frame bug.
49  *                                      Should stop LAN workplace lockups.
50  *              Alan Cox        :       Some tidyups using the new skb list
51  *                                      facilities
52  *              Alan Cox        :       sk->keepopen now seems to work
53  *              Alan Cox        :       Pulls options out correctly on accepts
54  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
55  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
56  *                                      bit to skb ops.
57  *              Alan Cox        :       Tidied tcp_data to avoid a potential
58  *                                      nasty.
59  *              Alan Cox        :       Added some better commenting, as the
60  *                                      tcp is hard to follow
61  *              Alan Cox        :       Removed incorrect check for 20 * psh
62  *      Michael O'Reilly        :       ack < copied bug fix.
63  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
64  *              Alan Cox        :       FIN with no memory -> CRASH
65  *              Alan Cox        :       Added socket option proto entries.
66  *                                      Also added awareness of them to accept.
67  *              Alan Cox        :       Added TCP options (SOL_TCP)
68  *              Alan Cox        :       Switched wakeup calls to callbacks,
69  *                                      so the kernel can layer network
70  *                                      sockets.
71  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
72  *              Alan Cox        :       Handle FIN (more) properly (we hope).
73  *              Alan Cox        :       RST frames sent on unsynchronised
74  *                                      state ack error.
75  *              Alan Cox        :       Put in missing check for SYN bit.
76  *              Alan Cox        :       Added tcp_select_window() aka NET2E
77  *                                      window non shrink trick.
78  *              Alan Cox        :       Added a couple of small NET2E timer
79  *                                      fixes
80  *              Charles Hedrick :       TCP fixes
81  *              Toomas Tamm     :       TCP window fixes
82  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
83  *              Charles Hedrick :       Rewrote most of it to actually work
84  *              Linus           :       Rewrote tcp_read() and URG handling
85  *                                      completely
86  *              Gerhard Koerting:       Fixed some missing timer handling
87  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
88  *              Gerhard Koerting:       PC/TCP workarounds
89  *              Adam Caldwell   :       Assorted timer/timing errors
90  *              Matthew Dillon  :       Fixed another RST bug
91  *              Alan Cox        :       Move to kernel side addressing changes.
92  *              Alan Cox        :       Beginning work on TCP fastpathing
93  *                                      (not yet usable)
94  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
95  *              Alan Cox        :       TCP fast path debugging
96  *              Alan Cox        :       Window clamping
97  *              Michael Riepe   :       Bug in tcp_check()
98  *              Matt Dillon     :       More TCP improvements and RST bug fixes
99  *              Matt Dillon     :       Yet more small nasties remove from the
100  *                                      TCP code (Be very nice to this man if
101  *                                      tcp finally works 100%) 8)
102  *              Alan Cox        :       BSD accept semantics.
103  *              Alan Cox        :       Reset on closedown bug.
104  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
105  *              Michael Pall    :       Handle poll() after URG properly in
106  *                                      all cases.
107  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
108  *                                      (multi URG PUSH broke rlogin).
109  *              Michael Pall    :       Fix the multi URG PUSH problem in
110  *                                      tcp_readable(), poll() after URG
111  *                                      works now.
112  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
113  *                                      BSD api.
114  *              Alan Cox        :       Changed the semantics of sk->socket to
115  *                                      fix a race and a signal problem with
116  *                                      accept() and async I/O.
117  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
118  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
119  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
120  *                                      clients/servers which listen in on
121  *                                      fixed ports.
122  *              Alan Cox        :       Cleaned the above up and shrank it to
123  *                                      a sensible code size.
124  *              Alan Cox        :       Self connect lockup fix.
125  *              Alan Cox        :       No connect to multicast.
126  *              Ross Biro       :       Close unaccepted children on master
127  *                                      socket close.
128  *              Alan Cox        :       Reset tracing code.
129  *              Alan Cox        :       Spurious resets on shutdown.
130  *              Alan Cox        :       Giant 15 minute/60 second timer error
131  *              Alan Cox        :       Small whoops in polling before an
132  *                                      accept.
133  *              Alan Cox        :       Kept the state trace facility since
134  *                                      it's handy for debugging.
135  *              Alan Cox        :       More reset handler fixes.
136  *              Alan Cox        :       Started rewriting the code based on
137  *                                      the RFC's for other useful protocol
138  *                                      references see: Comer, KA9Q NOS, and
139  *                                      for a reference on the difference
140  *                                      between specifications and how BSD
141  *                                      works see the 4.4lite source.
142  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
143  *                                      close.
144  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
145  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
146  *              Alan Cox        :       Reimplemented timers as per the RFC
147  *                                      and using multiple timers for sanity.
148  *              Alan Cox        :       Small bug fixes, and a lot of new
149  *                                      comments.
150  *              Alan Cox        :       Fixed dual reader crash by locking
151  *                                      the buffers (much like datagram.c)
152  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
153  *                                      now gets fed up of retrying without
154  *                                      (even a no space) answer.
155  *              Alan Cox        :       Extracted closing code better
156  *              Alan Cox        :       Fixed the closing state machine to
157  *                                      resemble the RFC.
158  *              Alan Cox        :       More 'per spec' fixes.
159  *              Jorge Cwik      :       Even faster checksumming.
160  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
161  *                                      only frames. At least one pc tcp stack
162  *                                      generates them.
163  *              Alan Cox        :       Cache last socket.
164  *              Alan Cox        :       Per route irtt.
165  *              Matt Day        :       poll()->select() match BSD precisely on error
166  *              Alan Cox        :       New buffers
167  *              Marc Tamsky     :       Various sk->prot->retransmits and
168  *                                      sk->retransmits misupdating fixed.
169  *                                      Fixed tcp_write_timeout: stuck close,
170  *                                      and TCP syn retries gets used now.
171  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
172  *                                      ack if state is TCP_CLOSED.
173  *              Alan Cox        :       Look up device on a retransmit - routes may
174  *                                      change. Doesn't yet cope with MSS shrink right
175  *                                      but it's a start!
176  *              Marc Tamsky     :       Closing in closing fixes.
177  *              Mike Shaver     :       RFC1122 verifications.
178  *              Alan Cox        :       rcv_saddr errors.
179  *              Alan Cox        :       Block double connect().
180  *              Alan Cox        :       Small hooks for enSKIP.
181  *              Alexey Kuznetsov:       Path MTU discovery.
182  *              Alan Cox        :       Support soft errors.
183  *              Alan Cox        :       Fix MTU discovery pathological case
184  *                                      when the remote claims no mtu!
185  *              Marc Tamsky     :       TCP_CLOSE fix.
186  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
187  *                                      window but wrong (fixes NT lpd problems)
188  *              Pedro Roque     :       Better TCP window handling, delayed ack.
189  *              Joerg Reuter    :       No modification of locked buffers in
190  *                                      tcp_do_retransmit()
191  *              Eric Schenk     :       Changed receiver side silly window
192  *                                      avoidance algorithm to BSD style
193  *                                      algorithm. This doubles throughput
194  *                                      against machines running Solaris,
195  *                                      and seems to result in general
196  *                                      improvement.
197  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
198  *      Willy Konynenberg       :       Transparent proxying support.
199  *      Mike McLagan            :       Routing by source
200  *              Keith Owens     :       Do proper merging with partial SKB's in
201  *                                      tcp_do_sendmsg to avoid burstiness.
202  *              Eric Schenk     :       Fix fast close down bug with
203  *                                      shutdown() followed by close().
204  *              Andi Kleen      :       Make poll agree with SIGIO
205  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
206  *                                      lingertime == 0 (RFC 793 ABORT Call)
207  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
208  *                                      csum_and_copy_from_user() if possible.
209  *
210  *              This program is free software; you can redistribute it and/or
211  *              modify it under the terms of the GNU General Public License
212  *              as published by the Free Software Foundation; either version
213  *              2 of the License, or(at your option) any later version.
214  *
215  * Description of States:
216  *
217  *      TCP_SYN_SENT            sent a connection request, waiting for ack
218  *
219  *      TCP_SYN_RECV            received a connection request, sent ack,
220  *                              waiting for final ack in three-way handshake.
221  *
222  *      TCP_ESTABLISHED         connection established
223  *
224  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
225  *                              transmission of remaining buffered data
226  *
227  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
228  *                              to shutdown
229  *
230  *      TCP_CLOSING             both sides have shutdown but we still have
231  *                              data we have to finish sending
232  *
233  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
234  *                              closed, can only be entered from FIN_WAIT2
235  *                              or CLOSING.  Required because the other end
236  *                              may not have gotten our last ACK causing it
237  *                              to retransmit the data packet (which we ignore)
238  *
239  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
240  *                              us to finish writing our data and to shutdown
241  *                              (we have to close() to move on to LAST_ACK)
242  *
243  *      TCP_LAST_ACK            out side has shutdown after remote has
244  *                              shutdown.  There may still be data in our
245  *                              buffer that we have to finish sending
246  *
247  *      TCP_CLOSE               socket is finished
248  */
249
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/splice.h>
259 #include <linux/net.h>
260 #include <linux/socket.h>
261 #include <linux/random.h>
262 #include <linux/bootmem.h>
263 #include <linux/cache.h>
264 #include <linux/err.h>
265 #include <linux/crypto.h>
266
267 #include <net/icmp.h>
268 #include <net/tcp.h>
269 #include <net/xfrm.h>
270 #include <net/ip.h>
271 #include <net/netdma.h>
272 #include <net/sock.h>
273
274 #include <asm/uaccess.h>
275 #include <asm/ioctls.h>
276
277 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
278
279 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
280
281 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
282
283 EXPORT_SYMBOL_GPL(tcp_orphan_count);
284
285 int sysctl_tcp_mem[3] __read_mostly;
286 int sysctl_tcp_wmem[3] __read_mostly;
287 int sysctl_tcp_rmem[3] __read_mostly;
288
289 EXPORT_SYMBOL(sysctl_tcp_mem);
290 EXPORT_SYMBOL(sysctl_tcp_rmem);
291 EXPORT_SYMBOL(sysctl_tcp_wmem);
292
293 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
294 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
295
296 EXPORT_SYMBOL(tcp_memory_allocated);
297 EXPORT_SYMBOL(tcp_sockets_allocated);
298
299 /*
300  * TCP splice context
301  */
302 struct tcp_splice_state {
303         struct pipe_inode_info *pipe;
304         size_t len;
305         unsigned int flags;
306 };
307
308 /*
309  * Pressure flag: try to collapse.
310  * Technical note: it is used by multiple contexts non atomically.
311  * All the __sk_mem_schedule() is of this nature: accounting
312  * is strict, actions are advisory and have some latency.
313  */
314 int tcp_memory_pressure __read_mostly;
315
316 EXPORT_SYMBOL(tcp_memory_pressure);
317
318 void tcp_enter_memory_pressure(void)
319 {
320         if (!tcp_memory_pressure) {
321                 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
322                 tcp_memory_pressure = 1;
323         }
324 }
325
326 EXPORT_SYMBOL(tcp_enter_memory_pressure);
327
328 /*
329  *      Wait for a TCP event.
330  *
331  *      Note that we don't need to lock the socket, as the upper poll layers
332  *      take care of normal races (between the test and the event) and we don't
333  *      go look at any of the socket buffers directly.
334  */
335 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
336 {
337         unsigned int mask;
338         struct sock *sk = sock->sk;
339         struct tcp_sock *tp = tcp_sk(sk);
340
341         poll_wait(file, sk->sk_sleep, wait);
342         if (sk->sk_state == TCP_LISTEN)
343                 return inet_csk_listen_poll(sk);
344
345         /* Socket is not locked. We are protected from async events
346            by poll logic and correct handling of state changes
347            made by another threads is impossible in any case.
348          */
349
350         mask = 0;
351         if (sk->sk_err)
352                 mask = POLLERR;
353
354         /*
355          * POLLHUP is certainly not done right. But poll() doesn't
356          * have a notion of HUP in just one direction, and for a
357          * socket the read side is more interesting.
358          *
359          * Some poll() documentation says that POLLHUP is incompatible
360          * with the POLLOUT/POLLWR flags, so somebody should check this
361          * all. But careful, it tends to be safer to return too many
362          * bits than too few, and you can easily break real applications
363          * if you don't tell them that something has hung up!
364          *
365          * Check-me.
366          *
367          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
368          * our fs/select.c). It means that after we received EOF,
369          * poll always returns immediately, making impossible poll() on write()
370          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
371          * if and only if shutdown has been made in both directions.
372          * Actually, it is interesting to look how Solaris and DUX
373          * solve this dilemma. I would prefer, if PULLHUP were maskable,
374          * then we could set it on SND_SHUTDOWN. BTW examples given
375          * in Stevens' books assume exactly this behaviour, it explains
376          * why PULLHUP is incompatible with POLLOUT.    --ANK
377          *
378          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
379          * blocking on fresh not-connected or disconnected socket. --ANK
380          */
381         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
382                 mask |= POLLHUP;
383         if (sk->sk_shutdown & RCV_SHUTDOWN)
384                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
385
386         /* Connected? */
387         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
388                 /* Potential race condition. If read of tp below will
389                  * escape above sk->sk_state, we can be illegally awaken
390                  * in SYN_* states. */
391                 if ((tp->rcv_nxt != tp->copied_seq) &&
392                     (tp->urg_seq != tp->copied_seq ||
393                      tp->rcv_nxt != tp->copied_seq + 1 ||
394                      sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
395                         mask |= POLLIN | POLLRDNORM;
396
397                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
398                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
399                                 mask |= POLLOUT | POLLWRNORM;
400                         } else {  /* send SIGIO later */
401                                 set_bit(SOCK_ASYNC_NOSPACE,
402                                         &sk->sk_socket->flags);
403                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
404
405                                 /* Race breaker. If space is freed after
406                                  * wspace test but before the flags are set,
407                                  * IO signal will be lost.
408                                  */
409                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
410                                         mask |= POLLOUT | POLLWRNORM;
411                         }
412                 }
413
414                 if (tp->urg_data & TCP_URG_VALID)
415                         mask |= POLLPRI;
416         }
417         return mask;
418 }
419
420 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
421 {
422         struct tcp_sock *tp = tcp_sk(sk);
423         int answ;
424
425         switch (cmd) {
426         case SIOCINQ:
427                 if (sk->sk_state == TCP_LISTEN)
428                         return -EINVAL;
429
430                 lock_sock(sk);
431                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
432                         answ = 0;
433                 else if (sock_flag(sk, SOCK_URGINLINE) ||
434                          !tp->urg_data ||
435                          before(tp->urg_seq, tp->copied_seq) ||
436                          !before(tp->urg_seq, tp->rcv_nxt)) {
437                         answ = tp->rcv_nxt - tp->copied_seq;
438
439                         /* Subtract 1, if FIN is in queue. */
440                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
441                                 answ -=
442                        tcp_hdr((struct sk_buff *)sk->sk_receive_queue.prev)->fin;
443                 } else
444                         answ = tp->urg_seq - tp->copied_seq;
445                 release_sock(sk);
446                 break;
447         case SIOCATMARK:
448                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
449                 break;
450         case SIOCOUTQ:
451                 if (sk->sk_state == TCP_LISTEN)
452                         return -EINVAL;
453
454                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
455                         answ = 0;
456                 else
457                         answ = tp->write_seq - tp->snd_una;
458                 break;
459         default:
460                 return -ENOIOCTLCMD;
461         }
462
463         return put_user(answ, (int __user *)arg);
464 }
465
466 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
467 {
468         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
469         tp->pushed_seq = tp->write_seq;
470 }
471
472 static inline int forced_push(struct tcp_sock *tp)
473 {
474         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
475 }
476
477 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
478 {
479         struct tcp_sock *tp = tcp_sk(sk);
480         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
481
482         skb->csum    = 0;
483         tcb->seq     = tcb->end_seq = tp->write_seq;
484         tcb->flags   = TCPCB_FLAG_ACK;
485         tcb->sacked  = 0;
486         skb_header_release(skb);
487         tcp_add_write_queue_tail(sk, skb);
488         sk->sk_wmem_queued += skb->truesize;
489         sk_mem_charge(sk, skb->truesize);
490         if (tp->nonagle & TCP_NAGLE_PUSH)
491                 tp->nonagle &= ~TCP_NAGLE_PUSH;
492 }
493
494 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
495                                 struct sk_buff *skb)
496 {
497         if (flags & MSG_OOB) {
498                 tp->urg_mode = 1;
499                 tp->snd_up = tp->write_seq;
500         }
501 }
502
503 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
504                             int nonagle)
505 {
506         struct tcp_sock *tp = tcp_sk(sk);
507
508         if (tcp_send_head(sk)) {
509                 struct sk_buff *skb = tcp_write_queue_tail(sk);
510                 if (!(flags & MSG_MORE) || forced_push(tp))
511                         tcp_mark_push(tp, skb);
512                 tcp_mark_urg(tp, flags, skb);
513                 __tcp_push_pending_frames(sk, mss_now,
514                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
515         }
516 }
517
518 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
519                                 unsigned int offset, size_t len)
520 {
521         struct tcp_splice_state *tss = rd_desc->arg.data;
522
523         return skb_splice_bits(skb, offset, tss->pipe, tss->len, tss->flags);
524 }
525
526 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
527 {
528         /* Store TCP splice context information in read_descriptor_t. */
529         read_descriptor_t rd_desc = {
530                 .arg.data = tss,
531         };
532
533         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
534 }
535
536 /**
537  *  tcp_splice_read - splice data from TCP socket to a pipe
538  * @sock:       socket to splice from
539  * @ppos:       position (not valid)
540  * @pipe:       pipe to splice to
541  * @len:        number of bytes to splice
542  * @flags:      splice modifier flags
543  *
544  * Description:
545  *    Will read pages from given socket and fill them into a pipe.
546  *
547  **/
548 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
549                         struct pipe_inode_info *pipe, size_t len,
550                         unsigned int flags)
551 {
552         struct sock *sk = sock->sk;
553         struct tcp_splice_state tss = {
554                 .pipe = pipe,
555                 .len = len,
556                 .flags = flags,
557         };
558         long timeo;
559         ssize_t spliced;
560         int ret;
561
562         /*
563          * We can't seek on a socket input
564          */
565         if (unlikely(*ppos))
566                 return -ESPIPE;
567
568         ret = spliced = 0;
569
570         lock_sock(sk);
571
572         timeo = sock_rcvtimeo(sk, flags & SPLICE_F_NONBLOCK);
573         while (tss.len) {
574                 ret = __tcp_splice_read(sk, &tss);
575                 if (ret < 0)
576                         break;
577                 else if (!ret) {
578                         if (spliced)
579                                 break;
580                         if (flags & SPLICE_F_NONBLOCK) {
581                                 ret = -EAGAIN;
582                                 break;
583                         }
584                         if (sock_flag(sk, SOCK_DONE))
585                                 break;
586                         if (sk->sk_err) {
587                                 ret = sock_error(sk);
588                                 break;
589                         }
590                         if (sk->sk_shutdown & RCV_SHUTDOWN)
591                                 break;
592                         if (sk->sk_state == TCP_CLOSE) {
593                                 /*
594                                  * This occurs when user tries to read
595                                  * from never connected socket.
596                                  */
597                                 if (!sock_flag(sk, SOCK_DONE))
598                                         ret = -ENOTCONN;
599                                 break;
600                         }
601                         if (!timeo) {
602                                 ret = -EAGAIN;
603                                 break;
604                         }
605                         sk_wait_data(sk, &timeo);
606                         if (signal_pending(current)) {
607                                 ret = sock_intr_errno(timeo);
608                                 break;
609                         }
610                         continue;
611                 }
612                 tss.len -= ret;
613                 spliced += ret;
614
615                 release_sock(sk);
616                 lock_sock(sk);
617
618                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
619                     (sk->sk_shutdown & RCV_SHUTDOWN) || !timeo ||
620                     signal_pending(current))
621                         break;
622         }
623
624         release_sock(sk);
625
626         if (spliced)
627                 return spliced;
628
629         return ret;
630 }
631
632 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
633 {
634         struct sk_buff *skb;
635
636         /* The TCP header must be at least 32-bit aligned.  */
637         size = ALIGN(size, 4);
638
639         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
640         if (skb) {
641                 if (sk_wmem_schedule(sk, skb->truesize)) {
642                         /*
643                          * Make sure that we have exactly size bytes
644                          * available to the caller, no more, no less.
645                          */
646                         skb_reserve(skb, skb_tailroom(skb) - size);
647                         return skb;
648                 }
649                 __kfree_skb(skb);
650         } else {
651                 sk->sk_prot->enter_memory_pressure();
652                 sk_stream_moderate_sndbuf(sk);
653         }
654         return NULL;
655 }
656
657 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
658                          size_t psize, int flags)
659 {
660         struct tcp_sock *tp = tcp_sk(sk);
661         int mss_now, size_goal;
662         int err;
663         ssize_t copied;
664         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
665
666         /* Wait for a connection to finish. */
667         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
668                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
669                         goto out_err;
670
671         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
672
673         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
674         size_goal = tp->xmit_size_goal;
675         copied = 0;
676
677         err = -EPIPE;
678         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
679                 goto do_error;
680
681         while (psize > 0) {
682                 struct sk_buff *skb = tcp_write_queue_tail(sk);
683                 struct page *page = pages[poffset / PAGE_SIZE];
684                 int copy, i, can_coalesce;
685                 int offset = poffset % PAGE_SIZE;
686                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
687
688                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
689 new_segment:
690                         if (!sk_stream_memory_free(sk))
691                                 goto wait_for_sndbuf;
692
693                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
694                         if (!skb)
695                                 goto wait_for_memory;
696
697                         skb_entail(sk, skb);
698                         copy = size_goal;
699                 }
700
701                 if (copy > size)
702                         copy = size;
703
704                 i = skb_shinfo(skb)->nr_frags;
705                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
706                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
707                         tcp_mark_push(tp, skb);
708                         goto new_segment;
709                 }
710                 if (!sk_wmem_schedule(sk, copy))
711                         goto wait_for_memory;
712
713                 if (can_coalesce) {
714                         skb_shinfo(skb)->frags[i - 1].size += copy;
715                 } else {
716                         get_page(page);
717                         skb_fill_page_desc(skb, i, page, offset, copy);
718                 }
719
720                 skb->len += copy;
721                 skb->data_len += copy;
722                 skb->truesize += copy;
723                 sk->sk_wmem_queued += copy;
724                 sk_mem_charge(sk, copy);
725                 skb->ip_summed = CHECKSUM_PARTIAL;
726                 tp->write_seq += copy;
727                 TCP_SKB_CB(skb)->end_seq += copy;
728                 skb_shinfo(skb)->gso_segs = 0;
729
730                 if (!copied)
731                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
732
733                 copied += copy;
734                 poffset += copy;
735                 if (!(psize -= copy))
736                         goto out;
737
738                 if (skb->len < size_goal || (flags & MSG_OOB))
739                         continue;
740
741                 if (forced_push(tp)) {
742                         tcp_mark_push(tp, skb);
743                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
744                 } else if (skb == tcp_send_head(sk))
745                         tcp_push_one(sk, mss_now);
746                 continue;
747
748 wait_for_sndbuf:
749                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
750 wait_for_memory:
751                 if (copied)
752                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
753
754                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
755                         goto do_error;
756
757                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
758                 size_goal = tp->xmit_size_goal;
759         }
760
761 out:
762         if (copied)
763                 tcp_push(sk, flags, mss_now, tp->nonagle);
764         return copied;
765
766 do_error:
767         if (copied)
768                 goto out;
769 out_err:
770         return sk_stream_error(sk, flags, err);
771 }
772
773 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
774                      size_t size, int flags)
775 {
776         ssize_t res;
777         struct sock *sk = sock->sk;
778
779         if (!(sk->sk_route_caps & NETIF_F_SG) ||
780             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
781                 return sock_no_sendpage(sock, page, offset, size, flags);
782
783         lock_sock(sk);
784         TCP_CHECK_TIMER(sk);
785         res = do_tcp_sendpages(sk, &page, offset, size, flags);
786         TCP_CHECK_TIMER(sk);
787         release_sock(sk);
788         return res;
789 }
790
791 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
792 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
793
794 static inline int select_size(struct sock *sk)
795 {
796         struct tcp_sock *tp = tcp_sk(sk);
797         int tmp = tp->mss_cache;
798
799         if (sk->sk_route_caps & NETIF_F_SG) {
800                 if (sk_can_gso(sk))
801                         tmp = 0;
802                 else {
803                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
804
805                         if (tmp >= pgbreak &&
806                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
807                                 tmp = pgbreak;
808                 }
809         }
810
811         return tmp;
812 }
813
814 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
815                 size_t size)
816 {
817         struct sock *sk = sock->sk;
818         struct iovec *iov;
819         struct tcp_sock *tp = tcp_sk(sk);
820         struct sk_buff *skb;
821         int iovlen, flags;
822         int mss_now, size_goal;
823         int err, copied;
824         long timeo;
825
826         lock_sock(sk);
827         TCP_CHECK_TIMER(sk);
828
829         flags = msg->msg_flags;
830         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
831
832         /* Wait for a connection to finish. */
833         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
834                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
835                         goto out_err;
836
837         /* This should be in poll */
838         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
839
840         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
841         size_goal = tp->xmit_size_goal;
842
843         /* Ok commence sending. */
844         iovlen = msg->msg_iovlen;
845         iov = msg->msg_iov;
846         copied = 0;
847
848         err = -EPIPE;
849         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
850                 goto do_error;
851
852         while (--iovlen >= 0) {
853                 int seglen = iov->iov_len;
854                 unsigned char __user *from = iov->iov_base;
855
856                 iov++;
857
858                 while (seglen > 0) {
859                         int copy;
860
861                         skb = tcp_write_queue_tail(sk);
862
863                         if (!tcp_send_head(sk) ||
864                             (copy = size_goal - skb->len) <= 0) {
865
866 new_segment:
867                                 /* Allocate new segment. If the interface is SG,
868                                  * allocate skb fitting to single page.
869                                  */
870                                 if (!sk_stream_memory_free(sk))
871                                         goto wait_for_sndbuf;
872
873                                 skb = sk_stream_alloc_skb(sk, select_size(sk),
874                                                 sk->sk_allocation);
875                                 if (!skb)
876                                         goto wait_for_memory;
877
878                                 /*
879                                  * Check whether we can use HW checksum.
880                                  */
881                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
882                                         skb->ip_summed = CHECKSUM_PARTIAL;
883
884                                 skb_entail(sk, skb);
885                                 copy = size_goal;
886                         }
887
888                         /* Try to append data to the end of skb. */
889                         if (copy > seglen)
890                                 copy = seglen;
891
892                         /* Where to copy to? */
893                         if (skb_tailroom(skb) > 0) {
894                                 /* We have some space in skb head. Superb! */
895                                 if (copy > skb_tailroom(skb))
896                                         copy = skb_tailroom(skb);
897                                 if ((err = skb_add_data(skb, from, copy)) != 0)
898                                         goto do_fault;
899                         } else {
900                                 int merge = 0;
901                                 int i = skb_shinfo(skb)->nr_frags;
902                                 struct page *page = TCP_PAGE(sk);
903                                 int off = TCP_OFF(sk);
904
905                                 if (skb_can_coalesce(skb, i, page, off) &&
906                                     off != PAGE_SIZE) {
907                                         /* We can extend the last page
908                                          * fragment. */
909                                         merge = 1;
910                                 } else if (i == MAX_SKB_FRAGS ||
911                                            (!i &&
912                                            !(sk->sk_route_caps & NETIF_F_SG))) {
913                                         /* Need to add new fragment and cannot
914                                          * do this because interface is non-SG,
915                                          * or because all the page slots are
916                                          * busy. */
917                                         tcp_mark_push(tp, skb);
918                                         goto new_segment;
919                                 } else if (page) {
920                                         if (off == PAGE_SIZE) {
921                                                 put_page(page);
922                                                 TCP_PAGE(sk) = page = NULL;
923                                                 off = 0;
924                                         }
925                                 } else
926                                         off = 0;
927
928                                 if (copy > PAGE_SIZE - off)
929                                         copy = PAGE_SIZE - off;
930
931                                 if (!sk_wmem_schedule(sk, copy))
932                                         goto wait_for_memory;
933
934                                 if (!page) {
935                                         /* Allocate new cache page. */
936                                         if (!(page = sk_stream_alloc_page(sk)))
937                                                 goto wait_for_memory;
938                                 }
939
940                                 /* Time to copy data. We are close to
941                                  * the end! */
942                                 err = skb_copy_to_page(sk, from, skb, page,
943                                                        off, copy);
944                                 if (err) {
945                                         /* If this page was new, give it to the
946                                          * socket so it does not get leaked.
947                                          */
948                                         if (!TCP_PAGE(sk)) {
949                                                 TCP_PAGE(sk) = page;
950                                                 TCP_OFF(sk) = 0;
951                                         }
952                                         goto do_error;
953                                 }
954
955                                 /* Update the skb. */
956                                 if (merge) {
957                                         skb_shinfo(skb)->frags[i - 1].size +=
958                                                                         copy;
959                                 } else {
960                                         skb_fill_page_desc(skb, i, page, off, copy);
961                                         if (TCP_PAGE(sk)) {
962                                                 get_page(page);
963                                         } else if (off + copy < PAGE_SIZE) {
964                                                 get_page(page);
965                                                 TCP_PAGE(sk) = page;
966                                         }
967                                 }
968
969                                 TCP_OFF(sk) = off + copy;
970                         }
971
972                         if (!copied)
973                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
974
975                         tp->write_seq += copy;
976                         TCP_SKB_CB(skb)->end_seq += copy;
977                         skb_shinfo(skb)->gso_segs = 0;
978
979                         from += copy;
980                         copied += copy;
981                         if ((seglen -= copy) == 0 && iovlen == 0)
982                                 goto out;
983
984                         if (skb->len < size_goal || (flags & MSG_OOB))
985                                 continue;
986
987                         if (forced_push(tp)) {
988                                 tcp_mark_push(tp, skb);
989                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
990                         } else if (skb == tcp_send_head(sk))
991                                 tcp_push_one(sk, mss_now);
992                         continue;
993
994 wait_for_sndbuf:
995                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
996 wait_for_memory:
997                         if (copied)
998                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
999
1000                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1001                                 goto do_error;
1002
1003                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
1004                         size_goal = tp->xmit_size_goal;
1005                 }
1006         }
1007
1008 out:
1009         if (copied)
1010                 tcp_push(sk, flags, mss_now, tp->nonagle);
1011         TCP_CHECK_TIMER(sk);
1012         release_sock(sk);
1013         return copied;
1014
1015 do_fault:
1016         if (!skb->len) {
1017                 tcp_unlink_write_queue(skb, sk);
1018                 /* It is the one place in all of TCP, except connection
1019                  * reset, where we can be unlinking the send_head.
1020                  */
1021                 tcp_check_send_head(sk, skb);
1022                 sk_wmem_free_skb(sk, skb);
1023         }
1024
1025 do_error:
1026         if (copied)
1027                 goto out;
1028 out_err:
1029         err = sk_stream_error(sk, flags, err);
1030         TCP_CHECK_TIMER(sk);
1031         release_sock(sk);
1032         return err;
1033 }
1034
1035 /*
1036  *      Handle reading urgent data. BSD has very simple semantics for
1037  *      this, no blocking and very strange errors 8)
1038  */
1039
1040 static int tcp_recv_urg(struct sock *sk, long timeo,
1041                         struct msghdr *msg, int len, int flags,
1042                         int *addr_len)
1043 {
1044         struct tcp_sock *tp = tcp_sk(sk);
1045
1046         /* No URG data to read. */
1047         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1048             tp->urg_data == TCP_URG_READ)
1049                 return -EINVAL; /* Yes this is right ! */
1050
1051         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1052                 return -ENOTCONN;
1053
1054         if (tp->urg_data & TCP_URG_VALID) {
1055                 int err = 0;
1056                 char c = tp->urg_data;
1057
1058                 if (!(flags & MSG_PEEK))
1059                         tp->urg_data = TCP_URG_READ;
1060
1061                 /* Read urgent data. */
1062                 msg->msg_flags |= MSG_OOB;
1063
1064                 if (len > 0) {
1065                         if (!(flags & MSG_TRUNC))
1066                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1067                         len = 1;
1068                 } else
1069                         msg->msg_flags |= MSG_TRUNC;
1070
1071                 return err ? -EFAULT : len;
1072         }
1073
1074         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1075                 return 0;
1076
1077         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1078          * the available implementations agree in this case:
1079          * this call should never block, independent of the
1080          * blocking state of the socket.
1081          * Mike <pall@rz.uni-karlsruhe.de>
1082          */
1083         return -EAGAIN;
1084 }
1085
1086 /* Clean up the receive buffer for full frames taken by the user,
1087  * then send an ACK if necessary.  COPIED is the number of bytes
1088  * tcp_recvmsg has given to the user so far, it speeds up the
1089  * calculation of whether or not we must ACK for the sake of
1090  * a window update.
1091  */
1092 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1093 {
1094         struct tcp_sock *tp = tcp_sk(sk);
1095         int time_to_ack = 0;
1096
1097 #if TCP_DEBUG
1098         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1099
1100         BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1101 #endif
1102
1103         if (inet_csk_ack_scheduled(sk)) {
1104                 const struct inet_connection_sock *icsk = inet_csk(sk);
1105                    /* Delayed ACKs frequently hit locked sockets during bulk
1106                     * receive. */
1107                 if (icsk->icsk_ack.blocked ||
1108                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1109                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1110                     /*
1111                      * If this read emptied read buffer, we send ACK, if
1112                      * connection is not bidirectional, user drained
1113                      * receive buffer and there was a small segment
1114                      * in queue.
1115                      */
1116                     (copied > 0 &&
1117                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1118                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1119                        !icsk->icsk_ack.pingpong)) &&
1120                       !atomic_read(&sk->sk_rmem_alloc)))
1121                         time_to_ack = 1;
1122         }
1123
1124         /* We send an ACK if we can now advertise a non-zero window
1125          * which has been raised "significantly".
1126          *
1127          * Even if window raised up to infinity, do not send window open ACK
1128          * in states, where we will not receive more. It is useless.
1129          */
1130         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1131                 __u32 rcv_window_now = tcp_receive_window(tp);
1132
1133                 /* Optimize, __tcp_select_window() is not cheap. */
1134                 if (2*rcv_window_now <= tp->window_clamp) {
1135                         __u32 new_window = __tcp_select_window(sk);
1136
1137                         /* Send ACK now, if this read freed lots of space
1138                          * in our buffer. Certainly, new_window is new window.
1139                          * We can advertise it now, if it is not less than current one.
1140                          * "Lots" means "at least twice" here.
1141                          */
1142                         if (new_window && new_window >= 2 * rcv_window_now)
1143                                 time_to_ack = 1;
1144                 }
1145         }
1146         if (time_to_ack)
1147                 tcp_send_ack(sk);
1148 }
1149
1150 static void tcp_prequeue_process(struct sock *sk)
1151 {
1152         struct sk_buff *skb;
1153         struct tcp_sock *tp = tcp_sk(sk);
1154
1155         NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
1156
1157         /* RX process wants to run with disabled BHs, though it is not
1158          * necessary */
1159         local_bh_disable();
1160         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1161                 sk->sk_backlog_rcv(sk, skb);
1162         local_bh_enable();
1163
1164         /* Clear memory counter. */
1165         tp->ucopy.memory = 0;
1166 }
1167
1168 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1169 {
1170         struct sk_buff *skb;
1171         u32 offset;
1172
1173         skb_queue_walk(&sk->sk_receive_queue, skb) {
1174                 offset = seq - TCP_SKB_CB(skb)->seq;
1175                 if (tcp_hdr(skb)->syn)
1176                         offset--;
1177                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1178                         *off = offset;
1179                         return skb;
1180                 }
1181         }
1182         return NULL;
1183 }
1184
1185 /*
1186  * This routine provides an alternative to tcp_recvmsg() for routines
1187  * that would like to handle copying from skbuffs directly in 'sendfile'
1188  * fashion.
1189  * Note:
1190  *      - It is assumed that the socket was locked by the caller.
1191  *      - The routine does not block.
1192  *      - At present, there is no support for reading OOB data
1193  *        or for 'peeking' the socket using this routine
1194  *        (although both would be easy to implement).
1195  */
1196 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1197                   sk_read_actor_t recv_actor)
1198 {
1199         struct sk_buff *skb;
1200         struct tcp_sock *tp = tcp_sk(sk);
1201         u32 seq = tp->copied_seq;
1202         u32 offset;
1203         int copied = 0;
1204
1205         if (sk->sk_state == TCP_LISTEN)
1206                 return -ENOTCONN;
1207         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1208                 if (offset < skb->len) {
1209                         size_t used, len;
1210
1211                         len = skb->len - offset;
1212                         /* Stop reading if we hit a patch of urgent data */
1213                         if (tp->urg_data) {
1214                                 u32 urg_offset = tp->urg_seq - seq;
1215                                 if (urg_offset < len)
1216                                         len = urg_offset;
1217                                 if (!len)
1218                                         break;
1219                         }
1220                         used = recv_actor(desc, skb, offset, len);
1221                         if (used < 0) {
1222                                 if (!copied)
1223                                         copied = used;
1224                                 break;
1225                         } else if (used <= len) {
1226                                 seq += used;
1227                                 copied += used;
1228                                 offset += used;
1229                         }
1230                         /*
1231                          * If recv_actor drops the lock (e.g. TCP splice
1232                          * receive) the skb pointer might be invalid when
1233                          * getting here: tcp_collapse might have deleted it
1234                          * while aggregating skbs from the socket queue.
1235                          */
1236                         skb = tcp_recv_skb(sk, seq-1, &offset);
1237                         if (!skb || (offset+1 != skb->len))
1238                                 break;
1239                 }
1240                 if (tcp_hdr(skb)->fin) {
1241                         sk_eat_skb(sk, skb, 0);
1242                         ++seq;
1243                         break;
1244                 }
1245                 sk_eat_skb(sk, skb, 0);
1246                 if (!desc->count)
1247                         break;
1248         }
1249         tp->copied_seq = seq;
1250
1251         tcp_rcv_space_adjust(sk);
1252
1253         /* Clean up data we have read: This will do ACK frames. */
1254         if (copied > 0)
1255                 tcp_cleanup_rbuf(sk, copied);
1256         return copied;
1257 }
1258
1259 /*
1260  *      This routine copies from a sock struct into the user buffer.
1261  *
1262  *      Technical note: in 2.3 we work on _locked_ socket, so that
1263  *      tricks with *seq access order and skb->users are not required.
1264  *      Probably, code can be easily improved even more.
1265  */
1266
1267 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1268                 size_t len, int nonblock, int flags, int *addr_len)
1269 {
1270         struct tcp_sock *tp = tcp_sk(sk);
1271         int copied = 0;
1272         u32 peek_seq;
1273         u32 *seq;
1274         unsigned long used;
1275         int err;
1276         int target;             /* Read at least this many bytes */
1277         long timeo;
1278         struct task_struct *user_recv = NULL;
1279         int copied_early = 0;
1280         struct sk_buff *skb;
1281
1282         lock_sock(sk);
1283
1284         TCP_CHECK_TIMER(sk);
1285
1286         err = -ENOTCONN;
1287         if (sk->sk_state == TCP_LISTEN)
1288                 goto out;
1289
1290         timeo = sock_rcvtimeo(sk, nonblock);
1291
1292         /* Urgent data needs to be handled specially. */
1293         if (flags & MSG_OOB)
1294                 goto recv_urg;
1295
1296         seq = &tp->copied_seq;
1297         if (flags & MSG_PEEK) {
1298                 peek_seq = tp->copied_seq;
1299                 seq = &peek_seq;
1300         }
1301
1302         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1303
1304 #ifdef CONFIG_NET_DMA
1305         tp->ucopy.dma_chan = NULL;
1306         preempt_disable();
1307         skb = skb_peek_tail(&sk->sk_receive_queue);
1308         {
1309                 int available = 0;
1310
1311                 if (skb)
1312                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1313                 if ((available < target) &&
1314                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1315                     !sysctl_tcp_low_latency &&
1316                     __get_cpu_var(softnet_data).net_dma) {
1317                         preempt_enable_no_resched();
1318                         tp->ucopy.pinned_list =
1319                                         dma_pin_iovec_pages(msg->msg_iov, len);
1320                 } else {
1321                         preempt_enable_no_resched();
1322                 }
1323         }
1324 #endif
1325
1326         do {
1327                 u32 offset;
1328
1329                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1330                 if (tp->urg_data && tp->urg_seq == *seq) {
1331                         if (copied)
1332                                 break;
1333                         if (signal_pending(current)) {
1334                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1335                                 break;
1336                         }
1337                 }
1338
1339                 /* Next get a buffer. */
1340
1341                 skb = skb_peek(&sk->sk_receive_queue);
1342                 do {
1343                         if (!skb)
1344                                 break;
1345
1346                         /* Now that we have two receive queues this
1347                          * shouldn't happen.
1348                          */
1349                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1350                                 printk(KERN_INFO "recvmsg bug: copied %X "
1351                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1352                                 break;
1353                         }
1354                         offset = *seq - TCP_SKB_CB(skb)->seq;
1355                         if (tcp_hdr(skb)->syn)
1356                                 offset--;
1357                         if (offset < skb->len)
1358                                 goto found_ok_skb;
1359                         if (tcp_hdr(skb)->fin)
1360                                 goto found_fin_ok;
1361                         BUG_TRAP(flags & MSG_PEEK);
1362                         skb = skb->next;
1363                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1364
1365                 /* Well, if we have backlog, try to process it now yet. */
1366
1367                 if (copied >= target && !sk->sk_backlog.tail)
1368                         break;
1369
1370                 if (copied) {
1371                         if (sk->sk_err ||
1372                             sk->sk_state == TCP_CLOSE ||
1373                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1374                             !timeo ||
1375                             signal_pending(current) ||
1376                             (flags & MSG_PEEK))
1377                                 break;
1378                 } else {
1379                         if (sock_flag(sk, SOCK_DONE))
1380                                 break;
1381
1382                         if (sk->sk_err) {
1383                                 copied = sock_error(sk);
1384                                 break;
1385                         }
1386
1387                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1388                                 break;
1389
1390                         if (sk->sk_state == TCP_CLOSE) {
1391                                 if (!sock_flag(sk, SOCK_DONE)) {
1392                                         /* This occurs when user tries to read
1393                                          * from never connected socket.
1394                                          */
1395                                         copied = -ENOTCONN;
1396                                         break;
1397                                 }
1398                                 break;
1399                         }
1400
1401                         if (!timeo) {
1402                                 copied = -EAGAIN;
1403                                 break;
1404                         }
1405
1406                         if (signal_pending(current)) {
1407                                 copied = sock_intr_errno(timeo);
1408                                 break;
1409                         }
1410                 }
1411
1412                 tcp_cleanup_rbuf(sk, copied);
1413
1414                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1415                         /* Install new reader */
1416                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1417                                 user_recv = current;
1418                                 tp->ucopy.task = user_recv;
1419                                 tp->ucopy.iov = msg->msg_iov;
1420                         }
1421
1422                         tp->ucopy.len = len;
1423
1424                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1425                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1426
1427                         /* Ugly... If prequeue is not empty, we have to
1428                          * process it before releasing socket, otherwise
1429                          * order will be broken at second iteration.
1430                          * More elegant solution is required!!!
1431                          *
1432                          * Look: we have the following (pseudo)queues:
1433                          *
1434                          * 1. packets in flight
1435                          * 2. backlog
1436                          * 3. prequeue
1437                          * 4. receive_queue
1438                          *
1439                          * Each queue can be processed only if the next ones
1440                          * are empty. At this point we have empty receive_queue.
1441                          * But prequeue _can_ be not empty after 2nd iteration,
1442                          * when we jumped to start of loop because backlog
1443                          * processing added something to receive_queue.
1444                          * We cannot release_sock(), because backlog contains
1445                          * packets arrived _after_ prequeued ones.
1446                          *
1447                          * Shortly, algorithm is clear --- to process all
1448                          * the queues in order. We could make it more directly,
1449                          * requeueing packets from backlog to prequeue, if
1450                          * is not empty. It is more elegant, but eats cycles,
1451                          * unfortunately.
1452                          */
1453                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1454                                 goto do_prequeue;
1455
1456                         /* __ Set realtime policy in scheduler __ */
1457                 }
1458
1459                 if (copied >= target) {
1460                         /* Do not sleep, just process backlog. */
1461                         release_sock(sk);
1462                         lock_sock(sk);
1463                 } else
1464                         sk_wait_data(sk, &timeo);
1465
1466 #ifdef CONFIG_NET_DMA
1467                 tp->ucopy.wakeup = 0;
1468 #endif
1469
1470                 if (user_recv) {
1471                         int chunk;
1472
1473                         /* __ Restore normal policy in scheduler __ */
1474
1475                         if ((chunk = len - tp->ucopy.len) != 0) {
1476                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1477                                 len -= chunk;
1478                                 copied += chunk;
1479                         }
1480
1481                         if (tp->rcv_nxt == tp->copied_seq &&
1482                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1483 do_prequeue:
1484                                 tcp_prequeue_process(sk);
1485
1486                                 if ((chunk = len - tp->ucopy.len) != 0) {
1487                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1488                                         len -= chunk;
1489                                         copied += chunk;
1490                                 }
1491                         }
1492                 }
1493                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1494                         if (net_ratelimit())
1495                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1496                                        current->comm, task_pid_nr(current));
1497                         peek_seq = tp->copied_seq;
1498                 }
1499                 continue;
1500
1501         found_ok_skb:
1502                 /* Ok so how much can we use? */
1503                 used = skb->len - offset;
1504                 if (len < used)
1505                         used = len;
1506
1507                 /* Do we have urgent data here? */
1508                 if (tp->urg_data) {
1509                         u32 urg_offset = tp->urg_seq - *seq;
1510                         if (urg_offset < used) {
1511                                 if (!urg_offset) {
1512                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1513                                                 ++*seq;
1514                                                 offset++;
1515                                                 used--;
1516                                                 if (!used)
1517                                                         goto skip_copy;
1518                                         }
1519                                 } else
1520                                         used = urg_offset;
1521                         }
1522                 }
1523
1524                 if (!(flags & MSG_TRUNC)) {
1525 #ifdef CONFIG_NET_DMA
1526                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1527                                 tp->ucopy.dma_chan = get_softnet_dma();
1528
1529                         if (tp->ucopy.dma_chan) {
1530                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1531                                         tp->ucopy.dma_chan, skb, offset,
1532                                         msg->msg_iov, used,
1533                                         tp->ucopy.pinned_list);
1534
1535                                 if (tp->ucopy.dma_cookie < 0) {
1536
1537                                         printk(KERN_ALERT "dma_cookie < 0\n");
1538
1539                                         /* Exception. Bailout! */
1540                                         if (!copied)
1541                                                 copied = -EFAULT;
1542                                         break;
1543                                 }
1544                                 if ((offset + used) == skb->len)
1545                                         copied_early = 1;
1546
1547                         } else
1548 #endif
1549                         {
1550                                 err = skb_copy_datagram_iovec(skb, offset,
1551                                                 msg->msg_iov, used);
1552                                 if (err) {
1553                                         /* Exception. Bailout! */
1554                                         if (!copied)
1555                                                 copied = -EFAULT;
1556                                         break;
1557                                 }
1558                         }
1559                 }
1560
1561                 *seq += used;
1562                 copied += used;
1563                 len -= used;
1564
1565                 tcp_rcv_space_adjust(sk);
1566
1567 skip_copy:
1568                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1569                         tp->urg_data = 0;
1570                         tcp_fast_path_check(sk);
1571                 }
1572                 if (used + offset < skb->len)
1573                         continue;
1574
1575                 if (tcp_hdr(skb)->fin)
1576                         goto found_fin_ok;
1577                 if (!(flags & MSG_PEEK)) {
1578                         sk_eat_skb(sk, skb, copied_early);
1579                         copied_early = 0;
1580                 }
1581                 continue;
1582
1583         found_fin_ok:
1584                 /* Process the FIN. */
1585                 ++*seq;
1586                 if (!(flags & MSG_PEEK)) {
1587                         sk_eat_skb(sk, skb, copied_early);
1588                         copied_early = 0;
1589                 }
1590                 break;
1591         } while (len > 0);
1592
1593         if (user_recv) {
1594                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1595                         int chunk;
1596
1597                         tp->ucopy.len = copied > 0 ? len : 0;
1598
1599                         tcp_prequeue_process(sk);
1600
1601                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1602                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1603                                 len -= chunk;
1604                                 copied += chunk;
1605                         }
1606                 }
1607
1608                 tp->ucopy.task = NULL;
1609                 tp->ucopy.len = 0;
1610         }
1611
1612 #ifdef CONFIG_NET_DMA
1613         if (tp->ucopy.dma_chan) {
1614                 dma_cookie_t done, used;
1615
1616                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1617
1618                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1619                                                  tp->ucopy.dma_cookie, &done,
1620                                                  &used) == DMA_IN_PROGRESS) {
1621                         /* do partial cleanup of sk_async_wait_queue */
1622                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1623                                (dma_async_is_complete(skb->dma_cookie, done,
1624                                                       used) == DMA_SUCCESS)) {
1625                                 __skb_dequeue(&sk->sk_async_wait_queue);
1626                                 kfree_skb(skb);
1627                         }
1628                 }
1629
1630                 /* Safe to free early-copied skbs now */
1631                 __skb_queue_purge(&sk->sk_async_wait_queue);
1632                 dma_chan_put(tp->ucopy.dma_chan);
1633                 tp->ucopy.dma_chan = NULL;
1634         }
1635         if (tp->ucopy.pinned_list) {
1636                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1637                 tp->ucopy.pinned_list = NULL;
1638         }
1639 #endif
1640
1641         /* According to UNIX98, msg_name/msg_namelen are ignored
1642          * on connected socket. I was just happy when found this 8) --ANK
1643          */
1644
1645         /* Clean up data we have read: This will do ACK frames. */
1646         tcp_cleanup_rbuf(sk, copied);
1647
1648         TCP_CHECK_TIMER(sk);
1649         release_sock(sk);
1650         return copied;
1651
1652 out:
1653         TCP_CHECK_TIMER(sk);
1654         release_sock(sk);
1655         return err;
1656
1657 recv_urg:
1658         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1659         goto out;
1660 }
1661
1662 void tcp_set_state(struct sock *sk, int state)
1663 {
1664         int oldstate = sk->sk_state;
1665
1666         switch (state) {
1667         case TCP_ESTABLISHED:
1668                 if (oldstate != TCP_ESTABLISHED)
1669                         TCP_INC_STATS(TCP_MIB_CURRESTAB);
1670                 break;
1671
1672         case TCP_CLOSE:
1673                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1674                         TCP_INC_STATS(TCP_MIB_ESTABRESETS);
1675
1676                 sk->sk_prot->unhash(sk);
1677                 if (inet_csk(sk)->icsk_bind_hash &&
1678                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1679                         inet_put_port(sk);
1680                 /* fall through */
1681         default:
1682                 if (oldstate==TCP_ESTABLISHED)
1683                         TCP_DEC_STATS(TCP_MIB_CURRESTAB);
1684         }
1685
1686         /* Change state AFTER socket is unhashed to avoid closed
1687          * socket sitting in hash tables.
1688          */
1689         sk->sk_state = state;
1690
1691 #ifdef STATE_TRACE
1692         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n",sk, statename[oldstate],statename[state]);
1693 #endif
1694 }
1695 EXPORT_SYMBOL_GPL(tcp_set_state);
1696
1697 /*
1698  *      State processing on a close. This implements the state shift for
1699  *      sending our FIN frame. Note that we only send a FIN for some
1700  *      states. A shutdown() may have already sent the FIN, or we may be
1701  *      closed.
1702  */
1703
1704 static const unsigned char new_state[16] = {
1705   /* current state:        new state:      action:      */
1706   /* (Invalid)          */ TCP_CLOSE,
1707   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1708   /* TCP_SYN_SENT       */ TCP_CLOSE,
1709   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1710   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1711   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1712   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1713   /* TCP_CLOSE          */ TCP_CLOSE,
1714   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1715   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1716   /* TCP_LISTEN         */ TCP_CLOSE,
1717   /* TCP_CLOSING        */ TCP_CLOSING,
1718 };
1719
1720 static int tcp_close_state(struct sock *sk)
1721 {
1722         int next = (int)new_state[sk->sk_state];
1723         int ns = next & TCP_STATE_MASK;
1724
1725         tcp_set_state(sk, ns);
1726
1727         return next & TCP_ACTION_FIN;
1728 }
1729
1730 /*
1731  *      Shutdown the sending side of a connection. Much like close except
1732  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1733  */
1734
1735 void tcp_shutdown(struct sock *sk, int how)
1736 {
1737         /*      We need to grab some memory, and put together a FIN,
1738          *      and then put it into the queue to be sent.
1739          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1740          */
1741         if (!(how & SEND_SHUTDOWN))
1742                 return;
1743
1744         /* If we've already sent a FIN, or it's a closed state, skip this. */
1745         if ((1 << sk->sk_state) &
1746             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1747              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1748                 /* Clear out any half completed packets.  FIN if needed. */
1749                 if (tcp_close_state(sk))
1750                         tcp_send_fin(sk);
1751         }
1752 }
1753
1754 void tcp_close(struct sock *sk, long timeout)
1755 {
1756         struct sk_buff *skb;
1757         int data_was_unread = 0;
1758         int state;
1759
1760         lock_sock(sk);
1761         sk->sk_shutdown = SHUTDOWN_MASK;
1762
1763         if (sk->sk_state == TCP_LISTEN) {
1764                 tcp_set_state(sk, TCP_CLOSE);
1765
1766                 /* Special case. */
1767                 inet_csk_listen_stop(sk);
1768
1769                 goto adjudge_to_death;
1770         }
1771
1772         /*  We need to flush the recv. buffs.  We do this only on the
1773          *  descriptor close, not protocol-sourced closes, because the
1774          *  reader process may not have drained the data yet!
1775          */
1776         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1777                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1778                           tcp_hdr(skb)->fin;
1779                 data_was_unread += len;
1780                 __kfree_skb(skb);
1781         }
1782
1783         sk_mem_reclaim(sk);
1784
1785         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1786          * data was lost. To witness the awful effects of the old behavior of
1787          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1788          * GET in an FTP client, suspend the process, wait for the client to
1789          * advertise a zero window, then kill -9 the FTP client, wheee...
1790          * Note: timeout is always zero in such a case.
1791          */
1792         if (data_was_unread) {
1793                 /* Unread data was tossed, zap the connection. */
1794                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1795                 tcp_set_state(sk, TCP_CLOSE);
1796                 tcp_send_active_reset(sk, GFP_KERNEL);
1797         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1798                 /* Check zero linger _after_ checking for unread data. */
1799                 sk->sk_prot->disconnect(sk, 0);
1800                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1801         } else if (tcp_close_state(sk)) {
1802                 /* We FIN if the application ate all the data before
1803                  * zapping the connection.
1804                  */
1805
1806                 /* RED-PEN. Formally speaking, we have broken TCP state
1807                  * machine. State transitions:
1808                  *
1809                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1810                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1811                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1812                  *
1813                  * are legal only when FIN has been sent (i.e. in window),
1814                  * rather than queued out of window. Purists blame.
1815                  *
1816                  * F.e. "RFC state" is ESTABLISHED,
1817                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1818                  *
1819                  * The visible declinations are that sometimes
1820                  * we enter time-wait state, when it is not required really
1821                  * (harmless), do not send active resets, when they are
1822                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1823                  * they look as CLOSING or LAST_ACK for Linux)
1824                  * Probably, I missed some more holelets.
1825                  *                                              --ANK
1826                  */
1827                 tcp_send_fin(sk);
1828         }
1829
1830         sk_stream_wait_close(sk, timeout);
1831
1832 adjudge_to_death:
1833         state = sk->sk_state;
1834         sock_hold(sk);
1835         sock_orphan(sk);
1836         atomic_inc(sk->sk_prot->orphan_count);
1837
1838         /* It is the last release_sock in its life. It will remove backlog. */
1839         release_sock(sk);
1840
1841
1842         /* Now socket is owned by kernel and we acquire BH lock
1843            to finish close. No need to check for user refs.
1844          */
1845         local_bh_disable();
1846         bh_lock_sock(sk);
1847         BUG_TRAP(!sock_owned_by_user(sk));
1848
1849         /* Have we already been destroyed by a softirq or backlog? */
1850         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1851                 goto out;
1852
1853         /*      This is a (useful) BSD violating of the RFC. There is a
1854          *      problem with TCP as specified in that the other end could
1855          *      keep a socket open forever with no application left this end.
1856          *      We use a 3 minute timeout (about the same as BSD) then kill
1857          *      our end. If they send after that then tough - BUT: long enough
1858          *      that we won't make the old 4*rto = almost no time - whoops
1859          *      reset mistake.
1860          *
1861          *      Nope, it was not mistake. It is really desired behaviour
1862          *      f.e. on http servers, when such sockets are useless, but
1863          *      consume significant resources. Let's do it with special
1864          *      linger2 option.                                 --ANK
1865          */
1866
1867         if (sk->sk_state == TCP_FIN_WAIT2) {
1868                 struct tcp_sock *tp = tcp_sk(sk);
1869                 if (tp->linger2 < 0) {
1870                         tcp_set_state(sk, TCP_CLOSE);
1871                         tcp_send_active_reset(sk, GFP_ATOMIC);
1872                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1873                 } else {
1874                         const int tmo = tcp_fin_time(sk);
1875
1876                         if (tmo > TCP_TIMEWAIT_LEN) {
1877                                 inet_csk_reset_keepalive_timer(sk,
1878                                                 tmo - TCP_TIMEWAIT_LEN);
1879                         } else {
1880                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1881                                 goto out;
1882                         }
1883                 }
1884         }
1885         if (sk->sk_state != TCP_CLOSE) {
1886                 sk_mem_reclaim(sk);
1887                 if (tcp_too_many_orphans(sk,
1888                                 atomic_read(sk->sk_prot->orphan_count))) {
1889                         if (net_ratelimit())
1890                                 printk(KERN_INFO "TCP: too many of orphaned "
1891                                        "sockets\n");
1892                         tcp_set_state(sk, TCP_CLOSE);
1893                         tcp_send_active_reset(sk, GFP_ATOMIC);
1894                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1895                 }
1896         }
1897
1898         if (sk->sk_state == TCP_CLOSE)
1899                 inet_csk_destroy_sock(sk);
1900         /* Otherwise, socket is reprieved until protocol close. */
1901
1902 out:
1903         bh_unlock_sock(sk);
1904         local_bh_enable();
1905         sock_put(sk);
1906 }
1907
1908 /* These states need RST on ABORT according to RFC793 */
1909
1910 static inline int tcp_need_reset(int state)
1911 {
1912         return (1 << state) &
1913                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1914                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1915 }
1916
1917 int tcp_disconnect(struct sock *sk, int flags)
1918 {
1919         struct inet_sock *inet = inet_sk(sk);
1920         struct inet_connection_sock *icsk = inet_csk(sk);
1921         struct tcp_sock *tp = tcp_sk(sk);
1922         int err = 0;
1923         int old_state = sk->sk_state;
1924
1925         if (old_state != TCP_CLOSE)
1926                 tcp_set_state(sk, TCP_CLOSE);
1927
1928         /* ABORT function of RFC793 */
1929         if (old_state == TCP_LISTEN) {
1930                 inet_csk_listen_stop(sk);
1931         } else if (tcp_need_reset(old_state) ||
1932                    (tp->snd_nxt != tp->write_seq &&
1933                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1934                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1935                  * states
1936                  */
1937                 tcp_send_active_reset(sk, gfp_any());
1938                 sk->sk_err = ECONNRESET;
1939         } else if (old_state == TCP_SYN_SENT)
1940                 sk->sk_err = ECONNRESET;
1941
1942         tcp_clear_xmit_timers(sk);
1943         __skb_queue_purge(&sk->sk_receive_queue);
1944         tcp_write_queue_purge(sk);
1945         __skb_queue_purge(&tp->out_of_order_queue);
1946 #ifdef CONFIG_NET_DMA
1947         __skb_queue_purge(&sk->sk_async_wait_queue);
1948 #endif
1949
1950         inet->dport = 0;
1951
1952         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1953                 inet_reset_saddr(sk);
1954
1955         sk->sk_shutdown = 0;
1956         sock_reset_flag(sk, SOCK_DONE);
1957         tp->srtt = 0;
1958         if ((tp->write_seq += tp->max_window + 2) == 0)
1959                 tp->write_seq = 1;
1960         icsk->icsk_backoff = 0;
1961         tp->snd_cwnd = 2;
1962         icsk->icsk_probes_out = 0;
1963         tp->packets_out = 0;
1964         tp->snd_ssthresh = 0x7fffffff;
1965         tp->snd_cwnd_cnt = 0;
1966         tp->bytes_acked = 0;
1967         tcp_set_ca_state(sk, TCP_CA_Open);
1968         tcp_clear_retrans(tp);
1969         inet_csk_delack_init(sk);
1970         tcp_init_send_head(sk);
1971         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1972         __sk_dst_reset(sk);
1973
1974         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1975
1976         sk->sk_error_report(sk);
1977         return err;
1978 }
1979
1980 /*
1981  *      Socket option code for TCP.
1982  */
1983 static int do_tcp_setsockopt(struct sock *sk, int level,
1984                 int optname, char __user *optval, int optlen)
1985 {
1986         struct tcp_sock *tp = tcp_sk(sk);
1987         struct inet_connection_sock *icsk = inet_csk(sk);
1988         int val;
1989         int err = 0;
1990
1991         /* This is a string value all the others are int's */
1992         if (optname == TCP_CONGESTION) {
1993                 char name[TCP_CA_NAME_MAX];
1994
1995                 if (optlen < 1)
1996                         return -EINVAL;
1997
1998                 val = strncpy_from_user(name, optval,
1999                                         min(TCP_CA_NAME_MAX-1, optlen));
2000                 if (val < 0)
2001                         return -EFAULT;
2002                 name[val] = 0;
2003
2004                 lock_sock(sk);
2005                 err = tcp_set_congestion_control(sk, name);
2006                 release_sock(sk);
2007                 return err;
2008         }
2009
2010         if (optlen < sizeof(int))
2011                 return -EINVAL;
2012
2013         if (get_user(val, (int __user *)optval))
2014                 return -EFAULT;
2015
2016         lock_sock(sk);
2017
2018         switch (optname) {
2019         case TCP_MAXSEG:
2020                 /* Values greater than interface MTU won't take effect. However
2021                  * at the point when this call is done we typically don't yet
2022                  * know which interface is going to be used */
2023                 if (val < 8 || val > MAX_TCP_WINDOW) {
2024                         err = -EINVAL;
2025                         break;
2026                 }
2027                 tp->rx_opt.user_mss = val;
2028                 break;
2029
2030         case TCP_NODELAY:
2031                 if (val) {
2032                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2033                          * this option on corked socket is remembered, but
2034                          * it is not activated until cork is cleared.
2035                          *
2036                          * However, when TCP_NODELAY is set we make
2037                          * an explicit push, which overrides even TCP_CORK
2038                          * for currently queued segments.
2039                          */
2040                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2041                         tcp_push_pending_frames(sk);
2042                 } else {
2043                         tp->nonagle &= ~TCP_NAGLE_OFF;
2044                 }
2045                 break;
2046
2047         case TCP_CORK:
2048                 /* When set indicates to always queue non-full frames.
2049                  * Later the user clears this option and we transmit
2050                  * any pending partial frames in the queue.  This is
2051                  * meant to be used alongside sendfile() to get properly
2052                  * filled frames when the user (for example) must write
2053                  * out headers with a write() call first and then use
2054                  * sendfile to send out the data parts.
2055                  *
2056                  * TCP_CORK can be set together with TCP_NODELAY and it is
2057                  * stronger than TCP_NODELAY.
2058                  */
2059                 if (val) {
2060                         tp->nonagle |= TCP_NAGLE_CORK;
2061                 } else {
2062                         tp->nonagle &= ~TCP_NAGLE_CORK;
2063                         if (tp->nonagle&TCP_NAGLE_OFF)
2064                                 tp->nonagle |= TCP_NAGLE_PUSH;
2065                         tcp_push_pending_frames(sk);
2066                 }
2067                 break;
2068
2069         case TCP_KEEPIDLE:
2070                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2071                         err = -EINVAL;
2072                 else {
2073                         tp->keepalive_time = val * HZ;
2074                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2075                             !((1 << sk->sk_state) &
2076                               (TCPF_CLOSE | TCPF_LISTEN))) {
2077                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2078                                 if (tp->keepalive_time > elapsed)
2079                                         elapsed = tp->keepalive_time - elapsed;
2080                                 else
2081                                         elapsed = 0;
2082                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2083                         }
2084                 }
2085                 break;
2086         case TCP_KEEPINTVL:
2087                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2088                         err = -EINVAL;
2089                 else
2090                         tp->keepalive_intvl = val * HZ;
2091                 break;
2092         case TCP_KEEPCNT:
2093                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2094                         err = -EINVAL;
2095                 else
2096                         tp->keepalive_probes = val;
2097                 break;
2098         case TCP_SYNCNT:
2099                 if (val < 1 || val > MAX_TCP_SYNCNT)
2100                         err = -EINVAL;
2101                 else
2102                         icsk->icsk_syn_retries = val;
2103                 break;
2104
2105         case TCP_LINGER2:
2106                 if (val < 0)
2107                         tp->linger2 = -1;
2108                 else if (val > sysctl_tcp_fin_timeout / HZ)
2109                         tp->linger2 = 0;
2110                 else
2111                         tp->linger2 = val * HZ;
2112                 break;
2113
2114         case TCP_DEFER_ACCEPT:
2115                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
2116                 if (val > 0) {
2117                         /* Translate value in seconds to number of
2118                          * retransmits */
2119                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
2120                                val > ((TCP_TIMEOUT_INIT / HZ) <<
2121                                        icsk->icsk_accept_queue.rskq_defer_accept))
2122                                 icsk->icsk_accept_queue.rskq_defer_accept++;
2123                         icsk->icsk_accept_queue.rskq_defer_accept++;
2124                 }
2125                 break;
2126
2127         case TCP_WINDOW_CLAMP:
2128                 if (!val) {
2129                         if (sk->sk_state != TCP_CLOSE) {
2130                                 err = -EINVAL;
2131                                 break;
2132                         }
2133                         tp->window_clamp = 0;
2134                 } else
2135                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2136                                                 SOCK_MIN_RCVBUF / 2 : val;
2137                 break;
2138
2139         case TCP_QUICKACK:
2140                 if (!val) {
2141                         icsk->icsk_ack.pingpong = 1;
2142                 } else {
2143                         icsk->icsk_ack.pingpong = 0;
2144                         if ((1 << sk->sk_state) &
2145                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2146                             inet_csk_ack_scheduled(sk)) {
2147                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2148                                 tcp_cleanup_rbuf(sk, 1);
2149                                 if (!(val & 1))
2150                                         icsk->icsk_ack.pingpong = 1;
2151                         }
2152                 }
2153                 break;
2154
2155 #ifdef CONFIG_TCP_MD5SIG
2156         case TCP_MD5SIG:
2157                 /* Read the IP->Key mappings from userspace */
2158                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2159                 break;
2160 #endif
2161
2162         default:
2163                 err = -ENOPROTOOPT;
2164                 break;
2165         }
2166
2167         release_sock(sk);
2168         return err;
2169 }
2170
2171 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2172                    int optlen)
2173 {
2174         struct inet_connection_sock *icsk = inet_csk(sk);
2175
2176         if (level != SOL_TCP)
2177                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2178                                                      optval, optlen);
2179         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2180 }
2181
2182 #ifdef CONFIG_COMPAT
2183 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2184                           char __user *optval, int optlen)
2185 {
2186         if (level != SOL_TCP)
2187                 return inet_csk_compat_setsockopt(sk, level, optname,
2188                                                   optval, optlen);
2189         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2190 }
2191
2192 EXPORT_SYMBOL(compat_tcp_setsockopt);
2193 #endif
2194
2195 /* Return information about state of tcp endpoint in API format. */
2196 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2197 {
2198         struct tcp_sock *tp = tcp_sk(sk);
2199         const struct inet_connection_sock *icsk = inet_csk(sk);
2200         u32 now = tcp_time_stamp;
2201
2202         memset(info, 0, sizeof(*info));
2203
2204         info->tcpi_state = sk->sk_state;
2205         info->tcpi_ca_state = icsk->icsk_ca_state;
2206         info->tcpi_retransmits = icsk->icsk_retransmits;
2207         info->tcpi_probes = icsk->icsk_probes_out;
2208         info->tcpi_backoff = icsk->icsk_backoff;
2209
2210         if (tp->rx_opt.tstamp_ok)
2211                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2212         if (tcp_is_sack(tp))
2213                 info->tcpi_options |= TCPI_OPT_SACK;
2214         if (tp->rx_opt.wscale_ok) {
2215                 info->tcpi_options |= TCPI_OPT_WSCALE;
2216                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2217                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2218         }
2219
2220         if (tp->ecn_flags&TCP_ECN_OK)
2221                 info->tcpi_options |= TCPI_OPT_ECN;
2222
2223         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2224         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2225         info->tcpi_snd_mss = tp->mss_cache;
2226         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2227
2228         if (sk->sk_state == TCP_LISTEN) {
2229                 info->tcpi_unacked = sk->sk_ack_backlog;
2230                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2231         } else {
2232                 info->tcpi_unacked = tp->packets_out;
2233                 info->tcpi_sacked = tp->sacked_out;
2234         }
2235         info->tcpi_lost = tp->lost_out;
2236         info->tcpi_retrans = tp->retrans_out;
2237         info->tcpi_fackets = tp->fackets_out;
2238
2239         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2240         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2241         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2242
2243         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2244         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2245         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2246         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2247         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2248         info->tcpi_snd_cwnd = tp->snd_cwnd;
2249         info->tcpi_advmss = tp->advmss;
2250         info->tcpi_reordering = tp->reordering;
2251
2252         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2253         info->tcpi_rcv_space = tp->rcvq_space.space;
2254
2255         info->tcpi_total_retrans = tp->total_retrans;
2256 }
2257
2258 EXPORT_SYMBOL_GPL(tcp_get_info);
2259
2260 static int do_tcp_getsockopt(struct sock *sk, int level,
2261                 int optname, char __user *optval, int __user *optlen)
2262 {
2263         struct inet_connection_sock *icsk = inet_csk(sk);
2264         struct tcp_sock *tp = tcp_sk(sk);
2265         int val, len;
2266
2267         if (get_user(len, optlen))
2268                 return -EFAULT;
2269
2270         len = min_t(unsigned int, len, sizeof(int));
2271
2272         if (len < 0)
2273                 return -EINVAL;
2274
2275         switch (optname) {
2276         case TCP_MAXSEG:
2277                 val = tp->mss_cache;
2278                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2279                         val = tp->rx_opt.user_mss;
2280                 break;
2281         case TCP_NODELAY:
2282                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2283                 break;
2284         case TCP_CORK:
2285                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2286                 break;
2287         case TCP_KEEPIDLE:
2288                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2289                 break;
2290         case TCP_KEEPINTVL:
2291                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2292                 break;
2293         case TCP_KEEPCNT:
2294                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2295                 break;
2296         case TCP_SYNCNT:
2297                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2298                 break;
2299         case TCP_LINGER2:
2300                 val = tp->linger2;
2301                 if (val >= 0)
2302                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2303                 break;
2304         case TCP_DEFER_ACCEPT:
2305                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2306                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2307                 break;
2308         case TCP_WINDOW_CLAMP:
2309                 val = tp->window_clamp;
2310                 break;
2311         case TCP_INFO: {
2312                 struct tcp_info info;
2313
2314                 if (get_user(len, optlen))
2315                         return -EFAULT;
2316
2317                 tcp_get_info(sk, &info);
2318
2319                 len = min_t(unsigned int, len, sizeof(info));
2320                 if (put_user(len, optlen))
2321                         return -EFAULT;
2322                 if (copy_to_user(optval, &info, len))
2323                         return -EFAULT;
2324                 return 0;
2325         }
2326         case TCP_QUICKACK:
2327                 val = !icsk->icsk_ack.pingpong;
2328                 break;
2329
2330         case TCP_CONGESTION:
2331                 if (get_user(len, optlen))
2332                         return -EFAULT;
2333                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2334                 if (put_user(len, optlen))
2335                         return -EFAULT;
2336                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2337                         return -EFAULT;
2338                 return 0;
2339         default:
2340                 return -ENOPROTOOPT;
2341         }
2342
2343         if (put_user(len, optlen))
2344                 return -EFAULT;
2345         if (copy_to_user(optval, &val, len))
2346                 return -EFAULT;
2347         return 0;
2348 }
2349
2350 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2351                    int __user *optlen)
2352 {
2353         struct inet_connection_sock *icsk = inet_csk(sk);
2354
2355         if (level != SOL_TCP)
2356                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2357                                                      optval, optlen);
2358         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2359 }
2360
2361 #ifdef CONFIG_COMPAT
2362 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2363                           char __user *optval, int __user *optlen)
2364 {
2365         if (level != SOL_TCP)
2366                 return inet_csk_compat_getsockopt(sk, level, optname,
2367                                                   optval, optlen);
2368         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2369 }
2370
2371 EXPORT_SYMBOL(compat_tcp_getsockopt);
2372 #endif
2373
2374 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2375 {
2376         struct sk_buff *segs = ERR_PTR(-EINVAL);
2377         struct tcphdr *th;
2378         unsigned thlen;
2379         unsigned int seq;
2380         __be32 delta;
2381         unsigned int oldlen;
2382         unsigned int len;
2383
2384         if (!pskb_may_pull(skb, sizeof(*th)))
2385                 goto out;
2386
2387         th = tcp_hdr(skb);
2388         thlen = th->doff * 4;
2389         if (thlen < sizeof(*th))
2390                 goto out;
2391
2392         if (!pskb_may_pull(skb, thlen))
2393                 goto out;
2394
2395         oldlen = (u16)~skb->len;
2396         __skb_pull(skb, thlen);
2397
2398         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2399                 /* Packet is from an untrusted source, reset gso_segs. */
2400                 int type = skb_shinfo(skb)->gso_type;
2401                 int mss;
2402
2403                 if (unlikely(type &
2404                              ~(SKB_GSO_TCPV4 |
2405                                SKB_GSO_DODGY |
2406                                SKB_GSO_TCP_ECN |
2407                                SKB_GSO_TCPV6 |
2408                                0) ||
2409                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2410                         goto out;
2411
2412                 mss = skb_shinfo(skb)->gso_size;
2413                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2414
2415                 segs = NULL;
2416                 goto out;
2417         }
2418
2419         segs = skb_segment(skb, features);
2420         if (IS_ERR(segs))
2421                 goto out;
2422
2423         len = skb_shinfo(skb)->gso_size;
2424         delta = htonl(oldlen + (thlen + len));
2425
2426         skb = segs;
2427         th = tcp_hdr(skb);
2428         seq = ntohl(th->seq);
2429
2430         do {
2431                 th->fin = th->psh = 0;
2432
2433                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2434                                        (__force u32)delta));
2435                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2436                         th->check =
2437                              csum_fold(csum_partial(skb_transport_header(skb),
2438                                                     thlen, skb->csum));
2439
2440                 seq += len;
2441                 skb = skb->next;
2442                 th = tcp_hdr(skb);
2443
2444                 th->seq = htonl(seq);
2445                 th->cwr = 0;
2446         } while (skb->next);
2447
2448         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2449                       skb->data_len);
2450         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2451                                 (__force u32)delta));
2452         if (skb->ip_summed != CHECKSUM_PARTIAL)
2453                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2454                                                    thlen, skb->csum));
2455
2456 out:
2457         return segs;
2458 }
2459 EXPORT_SYMBOL(tcp_tso_segment);
2460
2461 #ifdef CONFIG_TCP_MD5SIG
2462 static unsigned long tcp_md5sig_users;
2463 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2464 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2465
2466 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2467 {
2468         int cpu;
2469         for_each_possible_cpu(cpu) {
2470                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2471                 if (p) {
2472                         if (p->md5_desc.tfm)
2473                                 crypto_free_hash(p->md5_desc.tfm);
2474                         kfree(p);
2475                         p = NULL;
2476                 }
2477         }
2478         free_percpu(pool);
2479 }
2480
2481 void tcp_free_md5sig_pool(void)
2482 {
2483         struct tcp_md5sig_pool **pool = NULL;
2484
2485         spin_lock_bh(&tcp_md5sig_pool_lock);
2486         if (--tcp_md5sig_users == 0) {
2487                 pool = tcp_md5sig_pool;
2488                 tcp_md5sig_pool = NULL;
2489         }
2490         spin_unlock_bh(&tcp_md5sig_pool_lock);
2491         if (pool)
2492                 __tcp_free_md5sig_pool(pool);
2493 }
2494
2495 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2496
2497 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2498 {
2499         int cpu;
2500         struct tcp_md5sig_pool **pool;
2501
2502         pool = alloc_percpu(struct tcp_md5sig_pool *);
2503         if (!pool)
2504                 return NULL;
2505
2506         for_each_possible_cpu(cpu) {
2507                 struct tcp_md5sig_pool *p;
2508                 struct crypto_hash *hash;
2509
2510                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2511                 if (!p)
2512                         goto out_free;
2513                 *per_cpu_ptr(pool, cpu) = p;
2514
2515                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2516                 if (!hash || IS_ERR(hash))
2517                         goto out_free;
2518
2519                 p->md5_desc.tfm = hash;
2520         }
2521         return pool;
2522 out_free:
2523         __tcp_free_md5sig_pool(pool);
2524         return NULL;
2525 }
2526
2527 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2528 {
2529         struct tcp_md5sig_pool **pool;
2530         int alloc = 0;
2531
2532 retry:
2533         spin_lock_bh(&tcp_md5sig_pool_lock);
2534         pool = tcp_md5sig_pool;
2535         if (tcp_md5sig_users++ == 0) {
2536                 alloc = 1;
2537                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2538         } else if (!pool) {
2539                 tcp_md5sig_users--;
2540                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2541                 cpu_relax();
2542                 goto retry;
2543         } else
2544                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2545
2546         if (alloc) {
2547                 /* we cannot hold spinlock here because this may sleep. */
2548                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2549                 spin_lock_bh(&tcp_md5sig_pool_lock);
2550                 if (!p) {
2551                         tcp_md5sig_users--;
2552                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2553                         return NULL;
2554                 }
2555                 pool = tcp_md5sig_pool;
2556                 if (pool) {
2557                         /* oops, it has already been assigned. */
2558                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2559                         __tcp_free_md5sig_pool(p);
2560                 } else {
2561                         tcp_md5sig_pool = pool = p;
2562                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2563                 }
2564         }
2565         return pool;
2566 }
2567
2568 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2569
2570 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2571 {
2572         struct tcp_md5sig_pool **p;
2573         spin_lock_bh(&tcp_md5sig_pool_lock);
2574         p = tcp_md5sig_pool;
2575         if (p)
2576                 tcp_md5sig_users++;
2577         spin_unlock_bh(&tcp_md5sig_pool_lock);
2578         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2579 }
2580
2581 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2582
2583 void __tcp_put_md5sig_pool(void)
2584 {
2585         tcp_free_md5sig_pool();
2586 }
2587
2588 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2589 #endif
2590
2591 void tcp_done(struct sock *sk)
2592 {
2593         if(sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2594                 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
2595
2596         tcp_set_state(sk, TCP_CLOSE);
2597         tcp_clear_xmit_timers(sk);
2598
2599         sk->sk_shutdown = SHUTDOWN_MASK;
2600
2601         if (!sock_flag(sk, SOCK_DEAD))
2602                 sk->sk_state_change(sk);
2603         else
2604                 inet_csk_destroy_sock(sk);
2605 }
2606 EXPORT_SYMBOL_GPL(tcp_done);
2607
2608 extern struct tcp_congestion_ops tcp_reno;
2609
2610 static __initdata unsigned long thash_entries;
2611 static int __init set_thash_entries(char *str)
2612 {
2613         if (!str)
2614                 return 0;
2615         thash_entries = simple_strtoul(str, &str, 0);
2616         return 1;
2617 }
2618 __setup("thash_entries=", set_thash_entries);
2619
2620 void __init tcp_init(void)
2621 {
2622         struct sk_buff *skb = NULL;
2623         unsigned long limit;
2624         int order, i, max_share;
2625
2626         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2627
2628         tcp_hashinfo.bind_bucket_cachep =
2629                 kmem_cache_create("tcp_bind_bucket",
2630                                   sizeof(struct inet_bind_bucket), 0,
2631                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2632
2633         /* Size and allocate the main established and bind bucket
2634          * hash tables.
2635          *
2636          * The methodology is similar to that of the buffer cache.
2637          */
2638         tcp_hashinfo.ehash =
2639                 alloc_large_system_hash("TCP established",
2640                                         sizeof(struct inet_ehash_bucket),
2641                                         thash_entries,
2642                                         (num_physpages >= 128 * 1024) ?
2643                                         13 : 15,
2644                                         0,
2645                                         &tcp_hashinfo.ehash_size,
2646                                         NULL,
2647                                         thash_entries ? 0 : 512 * 1024);
2648         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2649         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2650                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2651                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
2652         }
2653         if (inet_ehash_locks_alloc(&tcp_hashinfo))
2654                 panic("TCP: failed to alloc ehash_locks");
2655         tcp_hashinfo.bhash =
2656                 alloc_large_system_hash("TCP bind",
2657                                         sizeof(struct inet_bind_hashbucket),
2658                                         tcp_hashinfo.ehash_size,
2659                                         (num_physpages >= 128 * 1024) ?
2660                                         13 : 15,
2661                                         0,
2662                                         &tcp_hashinfo.bhash_size,
2663                                         NULL,
2664                                         64 * 1024);
2665         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2666         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2667                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2668                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2669         }
2670
2671         /* Try to be a bit smarter and adjust defaults depending
2672          * on available memory.
2673          */
2674         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2675                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2676                         order++)
2677                 ;
2678         if (order >= 4) {
2679                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2680                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2681                 sysctl_max_syn_backlog = 1024;
2682         } else if (order < 3) {
2683                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2684                 sysctl_tcp_max_orphans >>= (3 - order);
2685                 sysctl_max_syn_backlog = 128;
2686         }
2687
2688         /* Set the pressure threshold to be a fraction of global memory that
2689          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2690          * memory, with a floor of 128 pages.
2691          */
2692         limit = min(nr_all_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2693         limit = (limit * (nr_all_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2694         limit = max(limit, 128UL);
2695         sysctl_tcp_mem[0] = limit / 4 * 3;
2696         sysctl_tcp_mem[1] = limit;
2697         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2698
2699         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2700         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2701         max_share = min(4UL*1024*1024, limit);
2702
2703         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2704         sysctl_tcp_wmem[1] = 16*1024;
2705         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2706
2707         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2708         sysctl_tcp_rmem[1] = 87380;
2709         sysctl_tcp_rmem[2] = max(87380, max_share);
2710
2711         printk(KERN_INFO "TCP: Hash tables configured "
2712                "(established %d bind %d)\n",
2713                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2714
2715         tcp_register_congestion_control(&tcp_reno);
2716 }
2717
2718 EXPORT_SYMBOL(tcp_close);
2719 EXPORT_SYMBOL(tcp_disconnect);
2720 EXPORT_SYMBOL(tcp_getsockopt);
2721 EXPORT_SYMBOL(tcp_ioctl);
2722 EXPORT_SYMBOL(tcp_poll);
2723 EXPORT_SYMBOL(tcp_read_sock);
2724 EXPORT_SYMBOL(tcp_recvmsg);
2725 EXPORT_SYMBOL(tcp_sendmsg);
2726 EXPORT_SYMBOL(tcp_splice_read);
2727 EXPORT_SYMBOL(tcp_sendpage);
2728 EXPORT_SYMBOL(tcp_setsockopt);
2729 EXPORT_SYMBOL(tcp_shutdown);
2730 EXPORT_SYMBOL(tcp_statistics);