[TCP]: Let skbs grow over a page on fast peers
[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                         if (offset != skb->len)
1231                                 break;
1232                 }
1233                 if (tcp_hdr(skb)->fin) {
1234                         sk_eat_skb(sk, skb, 0);
1235                         ++seq;
1236                         break;
1237                 }
1238                 sk_eat_skb(sk, skb, 0);
1239                 if (!desc->count)
1240                         break;
1241         }
1242         tp->copied_seq = seq;
1243
1244         tcp_rcv_space_adjust(sk);
1245
1246         /* Clean up data we have read: This will do ACK frames. */
1247         if (copied > 0)
1248                 tcp_cleanup_rbuf(sk, copied);
1249         return copied;
1250 }
1251
1252 /*
1253  *      This routine copies from a sock struct into the user buffer.
1254  *
1255  *      Technical note: in 2.3 we work on _locked_ socket, so that
1256  *      tricks with *seq access order and skb->users are not required.
1257  *      Probably, code can be easily improved even more.
1258  */
1259
1260 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1261                 size_t len, int nonblock, int flags, int *addr_len)
1262 {
1263         struct tcp_sock *tp = tcp_sk(sk);
1264         int copied = 0;
1265         u32 peek_seq;
1266         u32 *seq;
1267         unsigned long used;
1268         int err;
1269         int target;             /* Read at least this many bytes */
1270         long timeo;
1271         struct task_struct *user_recv = NULL;
1272         int copied_early = 0;
1273         struct sk_buff *skb;
1274
1275         lock_sock(sk);
1276
1277         TCP_CHECK_TIMER(sk);
1278
1279         err = -ENOTCONN;
1280         if (sk->sk_state == TCP_LISTEN)
1281                 goto out;
1282
1283         timeo = sock_rcvtimeo(sk, nonblock);
1284
1285         /* Urgent data needs to be handled specially. */
1286         if (flags & MSG_OOB)
1287                 goto recv_urg;
1288
1289         seq = &tp->copied_seq;
1290         if (flags & MSG_PEEK) {
1291                 peek_seq = tp->copied_seq;
1292                 seq = &peek_seq;
1293         }
1294
1295         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1296
1297 #ifdef CONFIG_NET_DMA
1298         tp->ucopy.dma_chan = NULL;
1299         preempt_disable();
1300         skb = skb_peek_tail(&sk->sk_receive_queue);
1301         {
1302                 int available = 0;
1303
1304                 if (skb)
1305                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1306                 if ((available < target) &&
1307                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1308                     !sysctl_tcp_low_latency &&
1309                     __get_cpu_var(softnet_data).net_dma) {
1310                         preempt_enable_no_resched();
1311                         tp->ucopy.pinned_list =
1312                                         dma_pin_iovec_pages(msg->msg_iov, len);
1313                 } else {
1314                         preempt_enable_no_resched();
1315                 }
1316         }
1317 #endif
1318
1319         do {
1320                 u32 offset;
1321
1322                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1323                 if (tp->urg_data && tp->urg_seq == *seq) {
1324                         if (copied)
1325                                 break;
1326                         if (signal_pending(current)) {
1327                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1328                                 break;
1329                         }
1330                 }
1331
1332                 /* Next get a buffer. */
1333
1334                 skb = skb_peek(&sk->sk_receive_queue);
1335                 do {
1336                         if (!skb)
1337                                 break;
1338
1339                         /* Now that we have two receive queues this
1340                          * shouldn't happen.
1341                          */
1342                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1343                                 printk(KERN_INFO "recvmsg bug: copied %X "
1344                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1345                                 break;
1346                         }
1347                         offset = *seq - TCP_SKB_CB(skb)->seq;
1348                         if (tcp_hdr(skb)->syn)
1349                                 offset--;
1350                         if (offset < skb->len)
1351                                 goto found_ok_skb;
1352                         if (tcp_hdr(skb)->fin)
1353                                 goto found_fin_ok;
1354                         BUG_TRAP(flags & MSG_PEEK);
1355                         skb = skb->next;
1356                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1357
1358                 /* Well, if we have backlog, try to process it now yet. */
1359
1360                 if (copied >= target && !sk->sk_backlog.tail)
1361                         break;
1362
1363                 if (copied) {
1364                         if (sk->sk_err ||
1365                             sk->sk_state == TCP_CLOSE ||
1366                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1367                             !timeo ||
1368                             signal_pending(current) ||
1369                             (flags & MSG_PEEK))
1370                                 break;
1371                 } else {
1372                         if (sock_flag(sk, SOCK_DONE))
1373                                 break;
1374
1375                         if (sk->sk_err) {
1376                                 copied = sock_error(sk);
1377                                 break;
1378                         }
1379
1380                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1381                                 break;
1382
1383                         if (sk->sk_state == TCP_CLOSE) {
1384                                 if (!sock_flag(sk, SOCK_DONE)) {
1385                                         /* This occurs when user tries to read
1386                                          * from never connected socket.
1387                                          */
1388                                         copied = -ENOTCONN;
1389                                         break;
1390                                 }
1391                                 break;
1392                         }
1393
1394                         if (!timeo) {
1395                                 copied = -EAGAIN;
1396                                 break;
1397                         }
1398
1399                         if (signal_pending(current)) {
1400                                 copied = sock_intr_errno(timeo);
1401                                 break;
1402                         }
1403                 }
1404
1405                 tcp_cleanup_rbuf(sk, copied);
1406
1407                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1408                         /* Install new reader */
1409                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1410                                 user_recv = current;
1411                                 tp->ucopy.task = user_recv;
1412                                 tp->ucopy.iov = msg->msg_iov;
1413                         }
1414
1415                         tp->ucopy.len = len;
1416
1417                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1418                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1419
1420                         /* Ugly... If prequeue is not empty, we have to
1421                          * process it before releasing socket, otherwise
1422                          * order will be broken at second iteration.
1423                          * More elegant solution is required!!!
1424                          *
1425                          * Look: we have the following (pseudo)queues:
1426                          *
1427                          * 1. packets in flight
1428                          * 2. backlog
1429                          * 3. prequeue
1430                          * 4. receive_queue
1431                          *
1432                          * Each queue can be processed only if the next ones
1433                          * are empty. At this point we have empty receive_queue.
1434                          * But prequeue _can_ be not empty after 2nd iteration,
1435                          * when we jumped to start of loop because backlog
1436                          * processing added something to receive_queue.
1437                          * We cannot release_sock(), because backlog contains
1438                          * packets arrived _after_ prequeued ones.
1439                          *
1440                          * Shortly, algorithm is clear --- to process all
1441                          * the queues in order. We could make it more directly,
1442                          * requeueing packets from backlog to prequeue, if
1443                          * is not empty. It is more elegant, but eats cycles,
1444                          * unfortunately.
1445                          */
1446                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1447                                 goto do_prequeue;
1448
1449                         /* __ Set realtime policy in scheduler __ */
1450                 }
1451
1452                 if (copied >= target) {
1453                         /* Do not sleep, just process backlog. */
1454                         release_sock(sk);
1455                         lock_sock(sk);
1456                 } else
1457                         sk_wait_data(sk, &timeo);
1458
1459 #ifdef CONFIG_NET_DMA
1460                 tp->ucopy.wakeup = 0;
1461 #endif
1462
1463                 if (user_recv) {
1464                         int chunk;
1465
1466                         /* __ Restore normal policy in scheduler __ */
1467
1468                         if ((chunk = len - tp->ucopy.len) != 0) {
1469                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1470                                 len -= chunk;
1471                                 copied += chunk;
1472                         }
1473
1474                         if (tp->rcv_nxt == tp->copied_seq &&
1475                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1476 do_prequeue:
1477                                 tcp_prequeue_process(sk);
1478
1479                                 if ((chunk = len - tp->ucopy.len) != 0) {
1480                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1481                                         len -= chunk;
1482                                         copied += chunk;
1483                                 }
1484                         }
1485                 }
1486                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1487                         if (net_ratelimit())
1488                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1489                                        current->comm, task_pid_nr(current));
1490                         peek_seq = tp->copied_seq;
1491                 }
1492                 continue;
1493
1494         found_ok_skb:
1495                 /* Ok so how much can we use? */
1496                 used = skb->len - offset;
1497                 if (len < used)
1498                         used = len;
1499
1500                 /* Do we have urgent data here? */
1501                 if (tp->urg_data) {
1502                         u32 urg_offset = tp->urg_seq - *seq;
1503                         if (urg_offset < used) {
1504                                 if (!urg_offset) {
1505                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1506                                                 ++*seq;
1507                                                 offset++;
1508                                                 used--;
1509                                                 if (!used)
1510                                                         goto skip_copy;
1511                                         }
1512                                 } else
1513                                         used = urg_offset;
1514                         }
1515                 }
1516
1517                 if (!(flags & MSG_TRUNC)) {
1518 #ifdef CONFIG_NET_DMA
1519                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1520                                 tp->ucopy.dma_chan = get_softnet_dma();
1521
1522                         if (tp->ucopy.dma_chan) {
1523                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1524                                         tp->ucopy.dma_chan, skb, offset,
1525                                         msg->msg_iov, used,
1526                                         tp->ucopy.pinned_list);
1527
1528                                 if (tp->ucopy.dma_cookie < 0) {
1529
1530                                         printk(KERN_ALERT "dma_cookie < 0\n");
1531
1532                                         /* Exception. Bailout! */
1533                                         if (!copied)
1534                                                 copied = -EFAULT;
1535                                         break;
1536                                 }
1537                                 if ((offset + used) == skb->len)
1538                                         copied_early = 1;
1539
1540                         } else
1541 #endif
1542                         {
1543                                 err = skb_copy_datagram_iovec(skb, offset,
1544                                                 msg->msg_iov, used);
1545                                 if (err) {
1546                                         /* Exception. Bailout! */
1547                                         if (!copied)
1548                                                 copied = -EFAULT;
1549                                         break;
1550                                 }
1551                         }
1552                 }
1553
1554                 *seq += used;
1555                 copied += used;
1556                 len -= used;
1557
1558                 tcp_rcv_space_adjust(sk);
1559
1560 skip_copy:
1561                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1562                         tp->urg_data = 0;
1563                         tcp_fast_path_check(sk);
1564                 }
1565                 if (used + offset < skb->len)
1566                         continue;
1567
1568                 if (tcp_hdr(skb)->fin)
1569                         goto found_fin_ok;
1570                 if (!(flags & MSG_PEEK)) {
1571                         sk_eat_skb(sk, skb, copied_early);
1572                         copied_early = 0;
1573                 }
1574                 continue;
1575
1576         found_fin_ok:
1577                 /* Process the FIN. */
1578                 ++*seq;
1579                 if (!(flags & MSG_PEEK)) {
1580                         sk_eat_skb(sk, skb, copied_early);
1581                         copied_early = 0;
1582                 }
1583                 break;
1584         } while (len > 0);
1585
1586         if (user_recv) {
1587                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1588                         int chunk;
1589
1590                         tp->ucopy.len = copied > 0 ? len : 0;
1591
1592                         tcp_prequeue_process(sk);
1593
1594                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1595                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1596                                 len -= chunk;
1597                                 copied += chunk;
1598                         }
1599                 }
1600
1601                 tp->ucopy.task = NULL;
1602                 tp->ucopy.len = 0;
1603         }
1604
1605 #ifdef CONFIG_NET_DMA
1606         if (tp->ucopy.dma_chan) {
1607                 dma_cookie_t done, used;
1608
1609                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1610
1611                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1612                                                  tp->ucopy.dma_cookie, &done,
1613                                                  &used) == DMA_IN_PROGRESS) {
1614                         /* do partial cleanup of sk_async_wait_queue */
1615                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1616                                (dma_async_is_complete(skb->dma_cookie, done,
1617                                                       used) == DMA_SUCCESS)) {
1618                                 __skb_dequeue(&sk->sk_async_wait_queue);
1619                                 kfree_skb(skb);
1620                         }
1621                 }
1622
1623                 /* Safe to free early-copied skbs now */
1624                 __skb_queue_purge(&sk->sk_async_wait_queue);
1625                 dma_chan_put(tp->ucopy.dma_chan);
1626                 tp->ucopy.dma_chan = NULL;
1627         }
1628         if (tp->ucopy.pinned_list) {
1629                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1630                 tp->ucopy.pinned_list = NULL;
1631         }
1632 #endif
1633
1634         /* According to UNIX98, msg_name/msg_namelen are ignored
1635          * on connected socket. I was just happy when found this 8) --ANK
1636          */
1637
1638         /* Clean up data we have read: This will do ACK frames. */
1639         tcp_cleanup_rbuf(sk, copied);
1640
1641         TCP_CHECK_TIMER(sk);
1642         release_sock(sk);
1643         return copied;
1644
1645 out:
1646         TCP_CHECK_TIMER(sk);
1647         release_sock(sk);
1648         return err;
1649
1650 recv_urg:
1651         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1652         goto out;
1653 }
1654
1655 void tcp_set_state(struct sock *sk, int state)
1656 {
1657         int oldstate = sk->sk_state;
1658
1659         switch (state) {
1660         case TCP_ESTABLISHED:
1661                 if (oldstate != TCP_ESTABLISHED)
1662                         TCP_INC_STATS(TCP_MIB_CURRESTAB);
1663                 break;
1664
1665         case TCP_CLOSE:
1666                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1667                         TCP_INC_STATS(TCP_MIB_ESTABRESETS);
1668
1669                 sk->sk_prot->unhash(sk);
1670                 if (inet_csk(sk)->icsk_bind_hash &&
1671                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1672                         inet_put_port(sk);
1673                 /* fall through */
1674         default:
1675                 if (oldstate==TCP_ESTABLISHED)
1676                         TCP_DEC_STATS(TCP_MIB_CURRESTAB);
1677         }
1678
1679         /* Change state AFTER socket is unhashed to avoid closed
1680          * socket sitting in hash tables.
1681          */
1682         sk->sk_state = state;
1683
1684 #ifdef STATE_TRACE
1685         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n",sk, statename[oldstate],statename[state]);
1686 #endif
1687 }
1688 EXPORT_SYMBOL_GPL(tcp_set_state);
1689
1690 /*
1691  *      State processing on a close. This implements the state shift for
1692  *      sending our FIN frame. Note that we only send a FIN for some
1693  *      states. A shutdown() may have already sent the FIN, or we may be
1694  *      closed.
1695  */
1696
1697 static const unsigned char new_state[16] = {
1698   /* current state:        new state:      action:      */
1699   /* (Invalid)          */ TCP_CLOSE,
1700   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1701   /* TCP_SYN_SENT       */ TCP_CLOSE,
1702   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1703   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1704   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1705   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1706   /* TCP_CLOSE          */ TCP_CLOSE,
1707   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1708   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1709   /* TCP_LISTEN         */ TCP_CLOSE,
1710   /* TCP_CLOSING        */ TCP_CLOSING,
1711 };
1712
1713 static int tcp_close_state(struct sock *sk)
1714 {
1715         int next = (int)new_state[sk->sk_state];
1716         int ns = next & TCP_STATE_MASK;
1717
1718         tcp_set_state(sk, ns);
1719
1720         return next & TCP_ACTION_FIN;
1721 }
1722
1723 /*
1724  *      Shutdown the sending side of a connection. Much like close except
1725  *      that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1726  */
1727
1728 void tcp_shutdown(struct sock *sk, int how)
1729 {
1730         /*      We need to grab some memory, and put together a FIN,
1731          *      and then put it into the queue to be sent.
1732          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1733          */
1734         if (!(how & SEND_SHUTDOWN))
1735                 return;
1736
1737         /* If we've already sent a FIN, or it's a closed state, skip this. */
1738         if ((1 << sk->sk_state) &
1739             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1740              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1741                 /* Clear out any half completed packets.  FIN if needed. */
1742                 if (tcp_close_state(sk))
1743                         tcp_send_fin(sk);
1744         }
1745 }
1746
1747 void tcp_close(struct sock *sk, long timeout)
1748 {
1749         struct sk_buff *skb;
1750         int data_was_unread = 0;
1751         int state;
1752
1753         lock_sock(sk);
1754         sk->sk_shutdown = SHUTDOWN_MASK;
1755
1756         if (sk->sk_state == TCP_LISTEN) {
1757                 tcp_set_state(sk, TCP_CLOSE);
1758
1759                 /* Special case. */
1760                 inet_csk_listen_stop(sk);
1761
1762                 goto adjudge_to_death;
1763         }
1764
1765         /*  We need to flush the recv. buffs.  We do this only on the
1766          *  descriptor close, not protocol-sourced closes, because the
1767          *  reader process may not have drained the data yet!
1768          */
1769         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1770                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1771                           tcp_hdr(skb)->fin;
1772                 data_was_unread += len;
1773                 __kfree_skb(skb);
1774         }
1775
1776         sk_mem_reclaim(sk);
1777
1778         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1779          * data was lost. To witness the awful effects of the old behavior of
1780          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1781          * GET in an FTP client, suspend the process, wait for the client to
1782          * advertise a zero window, then kill -9 the FTP client, wheee...
1783          * Note: timeout is always zero in such a case.
1784          */
1785         if (data_was_unread) {
1786                 /* Unread data was tossed, zap the connection. */
1787                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1788                 tcp_set_state(sk, TCP_CLOSE);
1789                 tcp_send_active_reset(sk, GFP_KERNEL);
1790         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1791                 /* Check zero linger _after_ checking for unread data. */
1792                 sk->sk_prot->disconnect(sk, 0);
1793                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1794         } else if (tcp_close_state(sk)) {
1795                 /* We FIN if the application ate all the data before
1796                  * zapping the connection.
1797                  */
1798
1799                 /* RED-PEN. Formally speaking, we have broken TCP state
1800                  * machine. State transitions:
1801                  *
1802                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1803                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1804                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1805                  *
1806                  * are legal only when FIN has been sent (i.e. in window),
1807                  * rather than queued out of window. Purists blame.
1808                  *
1809                  * F.e. "RFC state" is ESTABLISHED,
1810                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1811                  *
1812                  * The visible declinations are that sometimes
1813                  * we enter time-wait state, when it is not required really
1814                  * (harmless), do not send active resets, when they are
1815                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1816                  * they look as CLOSING or LAST_ACK for Linux)
1817                  * Probably, I missed some more holelets.
1818                  *                                              --ANK
1819                  */
1820                 tcp_send_fin(sk);
1821         }
1822
1823         sk_stream_wait_close(sk, timeout);
1824
1825 adjudge_to_death:
1826         state = sk->sk_state;
1827         sock_hold(sk);
1828         sock_orphan(sk);
1829         atomic_inc(sk->sk_prot->orphan_count);
1830
1831         /* It is the last release_sock in its life. It will remove backlog. */
1832         release_sock(sk);
1833
1834
1835         /* Now socket is owned by kernel and we acquire BH lock
1836            to finish close. No need to check for user refs.
1837          */
1838         local_bh_disable();
1839         bh_lock_sock(sk);
1840         BUG_TRAP(!sock_owned_by_user(sk));
1841
1842         /* Have we already been destroyed by a softirq or backlog? */
1843         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1844                 goto out;
1845
1846         /*      This is a (useful) BSD violating of the RFC. There is a
1847          *      problem with TCP as specified in that the other end could
1848          *      keep a socket open forever with no application left this end.
1849          *      We use a 3 minute timeout (about the same as BSD) then kill
1850          *      our end. If they send after that then tough - BUT: long enough
1851          *      that we won't make the old 4*rto = almost no time - whoops
1852          *      reset mistake.
1853          *
1854          *      Nope, it was not mistake. It is really desired behaviour
1855          *      f.e. on http servers, when such sockets are useless, but
1856          *      consume significant resources. Let's do it with special
1857          *      linger2 option.                                 --ANK
1858          */
1859
1860         if (sk->sk_state == TCP_FIN_WAIT2) {
1861                 struct tcp_sock *tp = tcp_sk(sk);
1862                 if (tp->linger2 < 0) {
1863                         tcp_set_state(sk, TCP_CLOSE);
1864                         tcp_send_active_reset(sk, GFP_ATOMIC);
1865                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1866                 } else {
1867                         const int tmo = tcp_fin_time(sk);
1868
1869                         if (tmo > TCP_TIMEWAIT_LEN) {
1870                                 inet_csk_reset_keepalive_timer(sk,
1871                                                 tmo - TCP_TIMEWAIT_LEN);
1872                         } else {
1873                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1874                                 goto out;
1875                         }
1876                 }
1877         }
1878         if (sk->sk_state != TCP_CLOSE) {
1879                 sk_mem_reclaim(sk);
1880                 if (tcp_too_many_orphans(sk,
1881                                 atomic_read(sk->sk_prot->orphan_count))) {
1882                         if (net_ratelimit())
1883                                 printk(KERN_INFO "TCP: too many of orphaned "
1884                                        "sockets\n");
1885                         tcp_set_state(sk, TCP_CLOSE);
1886                         tcp_send_active_reset(sk, GFP_ATOMIC);
1887                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1888                 }
1889         }
1890
1891         if (sk->sk_state == TCP_CLOSE)
1892                 inet_csk_destroy_sock(sk);
1893         /* Otherwise, socket is reprieved until protocol close. */
1894
1895 out:
1896         bh_unlock_sock(sk);
1897         local_bh_enable();
1898         sock_put(sk);
1899 }
1900
1901 /* These states need RST on ABORT according to RFC793 */
1902
1903 static inline int tcp_need_reset(int state)
1904 {
1905         return (1 << state) &
1906                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1907                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1908 }
1909
1910 int tcp_disconnect(struct sock *sk, int flags)
1911 {
1912         struct inet_sock *inet = inet_sk(sk);
1913         struct inet_connection_sock *icsk = inet_csk(sk);
1914         struct tcp_sock *tp = tcp_sk(sk);
1915         int err = 0;
1916         int old_state = sk->sk_state;
1917
1918         if (old_state != TCP_CLOSE)
1919                 tcp_set_state(sk, TCP_CLOSE);
1920
1921         /* ABORT function of RFC793 */
1922         if (old_state == TCP_LISTEN) {
1923                 inet_csk_listen_stop(sk);
1924         } else if (tcp_need_reset(old_state) ||
1925                    (tp->snd_nxt != tp->write_seq &&
1926                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1927                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1928                  * states
1929                  */
1930                 tcp_send_active_reset(sk, gfp_any());
1931                 sk->sk_err = ECONNRESET;
1932         } else if (old_state == TCP_SYN_SENT)
1933                 sk->sk_err = ECONNRESET;
1934
1935         tcp_clear_xmit_timers(sk);
1936         __skb_queue_purge(&sk->sk_receive_queue);
1937         tcp_write_queue_purge(sk);
1938         __skb_queue_purge(&tp->out_of_order_queue);
1939 #ifdef CONFIG_NET_DMA
1940         __skb_queue_purge(&sk->sk_async_wait_queue);
1941 #endif
1942
1943         inet->dport = 0;
1944
1945         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1946                 inet_reset_saddr(sk);
1947
1948         sk->sk_shutdown = 0;
1949         sock_reset_flag(sk, SOCK_DONE);
1950         tp->srtt = 0;
1951         if ((tp->write_seq += tp->max_window + 2) == 0)
1952                 tp->write_seq = 1;
1953         icsk->icsk_backoff = 0;
1954         tp->snd_cwnd = 2;
1955         icsk->icsk_probes_out = 0;
1956         tp->packets_out = 0;
1957         tp->snd_ssthresh = 0x7fffffff;
1958         tp->snd_cwnd_cnt = 0;
1959         tp->bytes_acked = 0;
1960         tcp_set_ca_state(sk, TCP_CA_Open);
1961         tcp_clear_retrans(tp);
1962         inet_csk_delack_init(sk);
1963         tcp_init_send_head(sk);
1964         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1965         __sk_dst_reset(sk);
1966
1967         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1968
1969         sk->sk_error_report(sk);
1970         return err;
1971 }
1972
1973 /*
1974  *      Socket option code for TCP.
1975  */
1976 static int do_tcp_setsockopt(struct sock *sk, int level,
1977                 int optname, char __user *optval, int optlen)
1978 {
1979         struct tcp_sock *tp = tcp_sk(sk);
1980         struct inet_connection_sock *icsk = inet_csk(sk);
1981         int val;
1982         int err = 0;
1983
1984         /* This is a string value all the others are int's */
1985         if (optname == TCP_CONGESTION) {
1986                 char name[TCP_CA_NAME_MAX];
1987
1988                 if (optlen < 1)
1989                         return -EINVAL;
1990
1991                 val = strncpy_from_user(name, optval,
1992                                         min(TCP_CA_NAME_MAX-1, optlen));
1993                 if (val < 0)
1994                         return -EFAULT;
1995                 name[val] = 0;
1996
1997                 lock_sock(sk);
1998                 err = tcp_set_congestion_control(sk, name);
1999                 release_sock(sk);
2000                 return err;
2001         }
2002
2003         if (optlen < sizeof(int))
2004                 return -EINVAL;
2005
2006         if (get_user(val, (int __user *)optval))
2007                 return -EFAULT;
2008
2009         lock_sock(sk);
2010
2011         switch (optname) {
2012         case TCP_MAXSEG:
2013                 /* Values greater than interface MTU won't take effect. However
2014                  * at the point when this call is done we typically don't yet
2015                  * know which interface is going to be used */
2016                 if (val < 8 || val > MAX_TCP_WINDOW) {
2017                         err = -EINVAL;
2018                         break;
2019                 }
2020                 tp->rx_opt.user_mss = val;
2021                 break;
2022
2023         case TCP_NODELAY:
2024                 if (val) {
2025                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2026                          * this option on corked socket is remembered, but
2027                          * it is not activated until cork is cleared.
2028                          *
2029                          * However, when TCP_NODELAY is set we make
2030                          * an explicit push, which overrides even TCP_CORK
2031                          * for currently queued segments.
2032                          */
2033                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2034                         tcp_push_pending_frames(sk);
2035                 } else {
2036                         tp->nonagle &= ~TCP_NAGLE_OFF;
2037                 }
2038                 break;
2039
2040         case TCP_CORK:
2041                 /* When set indicates to always queue non-full frames.
2042                  * Later the user clears this option and we transmit
2043                  * any pending partial frames in the queue.  This is
2044                  * meant to be used alongside sendfile() to get properly
2045                  * filled frames when the user (for example) must write
2046                  * out headers with a write() call first and then use
2047                  * sendfile to send out the data parts.
2048                  *
2049                  * TCP_CORK can be set together with TCP_NODELAY and it is
2050                  * stronger than TCP_NODELAY.
2051                  */
2052                 if (val) {
2053                         tp->nonagle |= TCP_NAGLE_CORK;
2054                 } else {
2055                         tp->nonagle &= ~TCP_NAGLE_CORK;
2056                         if (tp->nonagle&TCP_NAGLE_OFF)
2057                                 tp->nonagle |= TCP_NAGLE_PUSH;
2058                         tcp_push_pending_frames(sk);
2059                 }
2060                 break;
2061
2062         case TCP_KEEPIDLE:
2063                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2064                         err = -EINVAL;
2065                 else {
2066                         tp->keepalive_time = val * HZ;
2067                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2068                             !((1 << sk->sk_state) &
2069                               (TCPF_CLOSE | TCPF_LISTEN))) {
2070                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2071                                 if (tp->keepalive_time > elapsed)
2072                                         elapsed = tp->keepalive_time - elapsed;
2073                                 else
2074                                         elapsed = 0;
2075                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2076                         }
2077                 }
2078                 break;
2079         case TCP_KEEPINTVL:
2080                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2081                         err = -EINVAL;
2082                 else
2083                         tp->keepalive_intvl = val * HZ;
2084                 break;
2085         case TCP_KEEPCNT:
2086                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2087                         err = -EINVAL;
2088                 else
2089                         tp->keepalive_probes = val;
2090                 break;
2091         case TCP_SYNCNT:
2092                 if (val < 1 || val > MAX_TCP_SYNCNT)
2093                         err = -EINVAL;
2094                 else
2095                         icsk->icsk_syn_retries = val;
2096                 break;
2097
2098         case TCP_LINGER2:
2099                 if (val < 0)
2100                         tp->linger2 = -1;
2101                 else if (val > sysctl_tcp_fin_timeout / HZ)
2102                         tp->linger2 = 0;
2103                 else
2104                         tp->linger2 = val * HZ;
2105                 break;
2106
2107         case TCP_DEFER_ACCEPT:
2108                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
2109                 if (val > 0) {
2110                         /* Translate value in seconds to number of
2111                          * retransmits */
2112                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
2113                                val > ((TCP_TIMEOUT_INIT / HZ) <<
2114                                        icsk->icsk_accept_queue.rskq_defer_accept))
2115                                 icsk->icsk_accept_queue.rskq_defer_accept++;
2116                         icsk->icsk_accept_queue.rskq_defer_accept++;
2117                 }
2118                 break;
2119
2120         case TCP_WINDOW_CLAMP:
2121                 if (!val) {
2122                         if (sk->sk_state != TCP_CLOSE) {
2123                                 err = -EINVAL;
2124                                 break;
2125                         }
2126                         tp->window_clamp = 0;
2127                 } else
2128                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2129                                                 SOCK_MIN_RCVBUF / 2 : val;
2130                 break;
2131
2132         case TCP_QUICKACK:
2133                 if (!val) {
2134                         icsk->icsk_ack.pingpong = 1;
2135                 } else {
2136                         icsk->icsk_ack.pingpong = 0;
2137                         if ((1 << sk->sk_state) &
2138                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2139                             inet_csk_ack_scheduled(sk)) {
2140                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2141                                 tcp_cleanup_rbuf(sk, 1);
2142                                 if (!(val & 1))
2143                                         icsk->icsk_ack.pingpong = 1;
2144                         }
2145                 }
2146                 break;
2147
2148 #ifdef CONFIG_TCP_MD5SIG
2149         case TCP_MD5SIG:
2150                 /* Read the IP->Key mappings from userspace */
2151                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2152                 break;
2153 #endif
2154
2155         default:
2156                 err = -ENOPROTOOPT;
2157                 break;
2158         }
2159
2160         release_sock(sk);
2161         return err;
2162 }
2163
2164 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2165                    int optlen)
2166 {
2167         struct inet_connection_sock *icsk = inet_csk(sk);
2168
2169         if (level != SOL_TCP)
2170                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2171                                                      optval, optlen);
2172         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2173 }
2174
2175 #ifdef CONFIG_COMPAT
2176 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2177                           char __user *optval, int optlen)
2178 {
2179         if (level != SOL_TCP)
2180                 return inet_csk_compat_setsockopt(sk, level, optname,
2181                                                   optval, optlen);
2182         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2183 }
2184
2185 EXPORT_SYMBOL(compat_tcp_setsockopt);
2186 #endif
2187
2188 /* Return information about state of tcp endpoint in API format. */
2189 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2190 {
2191         struct tcp_sock *tp = tcp_sk(sk);
2192         const struct inet_connection_sock *icsk = inet_csk(sk);
2193         u32 now = tcp_time_stamp;
2194
2195         memset(info, 0, sizeof(*info));
2196
2197         info->tcpi_state = sk->sk_state;
2198         info->tcpi_ca_state = icsk->icsk_ca_state;
2199         info->tcpi_retransmits = icsk->icsk_retransmits;
2200         info->tcpi_probes = icsk->icsk_probes_out;
2201         info->tcpi_backoff = icsk->icsk_backoff;
2202
2203         if (tp->rx_opt.tstamp_ok)
2204                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2205         if (tcp_is_sack(tp))
2206                 info->tcpi_options |= TCPI_OPT_SACK;
2207         if (tp->rx_opt.wscale_ok) {
2208                 info->tcpi_options |= TCPI_OPT_WSCALE;
2209                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2210                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2211         }
2212
2213         if (tp->ecn_flags&TCP_ECN_OK)
2214                 info->tcpi_options |= TCPI_OPT_ECN;
2215
2216         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2217         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2218         info->tcpi_snd_mss = tp->mss_cache;
2219         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2220
2221         if (sk->sk_state == TCP_LISTEN) {
2222                 info->tcpi_unacked = sk->sk_ack_backlog;
2223                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2224         } else {
2225                 info->tcpi_unacked = tp->packets_out;
2226                 info->tcpi_sacked = tp->sacked_out;
2227         }
2228         info->tcpi_lost = tp->lost_out;
2229         info->tcpi_retrans = tp->retrans_out;
2230         info->tcpi_fackets = tp->fackets_out;
2231
2232         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2233         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2234         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2235
2236         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2237         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2238         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2239         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2240         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2241         info->tcpi_snd_cwnd = tp->snd_cwnd;
2242         info->tcpi_advmss = tp->advmss;
2243         info->tcpi_reordering = tp->reordering;
2244
2245         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2246         info->tcpi_rcv_space = tp->rcvq_space.space;
2247
2248         info->tcpi_total_retrans = tp->total_retrans;
2249 }
2250
2251 EXPORT_SYMBOL_GPL(tcp_get_info);
2252
2253 static int do_tcp_getsockopt(struct sock *sk, int level,
2254                 int optname, char __user *optval, int __user *optlen)
2255 {
2256         struct inet_connection_sock *icsk = inet_csk(sk);
2257         struct tcp_sock *tp = tcp_sk(sk);
2258         int val, len;
2259
2260         if (get_user(len, optlen))
2261                 return -EFAULT;
2262
2263         len = min_t(unsigned int, len, sizeof(int));
2264
2265         if (len < 0)
2266                 return -EINVAL;
2267
2268         switch (optname) {
2269         case TCP_MAXSEG:
2270                 val = tp->mss_cache;
2271                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2272                         val = tp->rx_opt.user_mss;
2273                 break;
2274         case TCP_NODELAY:
2275                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2276                 break;
2277         case TCP_CORK:
2278                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2279                 break;
2280         case TCP_KEEPIDLE:
2281                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2282                 break;
2283         case TCP_KEEPINTVL:
2284                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2285                 break;
2286         case TCP_KEEPCNT:
2287                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2288                 break;
2289         case TCP_SYNCNT:
2290                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2291                 break;
2292         case TCP_LINGER2:
2293                 val = tp->linger2;
2294                 if (val >= 0)
2295                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2296                 break;
2297         case TCP_DEFER_ACCEPT:
2298                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2299                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2300                 break;
2301         case TCP_WINDOW_CLAMP:
2302                 val = tp->window_clamp;
2303                 break;
2304         case TCP_INFO: {
2305                 struct tcp_info info;
2306
2307                 if (get_user(len, optlen))
2308                         return -EFAULT;
2309
2310                 tcp_get_info(sk, &info);
2311
2312                 len = min_t(unsigned int, len, sizeof(info));
2313                 if (put_user(len, optlen))
2314                         return -EFAULT;
2315                 if (copy_to_user(optval, &info, len))
2316                         return -EFAULT;
2317                 return 0;
2318         }
2319         case TCP_QUICKACK:
2320                 val = !icsk->icsk_ack.pingpong;
2321                 break;
2322
2323         case TCP_CONGESTION:
2324                 if (get_user(len, optlen))
2325                         return -EFAULT;
2326                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2327                 if (put_user(len, optlen))
2328                         return -EFAULT;
2329                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2330                         return -EFAULT;
2331                 return 0;
2332         default:
2333                 return -ENOPROTOOPT;
2334         }
2335
2336         if (put_user(len, optlen))
2337                 return -EFAULT;
2338         if (copy_to_user(optval, &val, len))
2339                 return -EFAULT;
2340         return 0;
2341 }
2342
2343 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2344                    int __user *optlen)
2345 {
2346         struct inet_connection_sock *icsk = inet_csk(sk);
2347
2348         if (level != SOL_TCP)
2349                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2350                                                      optval, optlen);
2351         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2352 }
2353
2354 #ifdef CONFIG_COMPAT
2355 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2356                           char __user *optval, int __user *optlen)
2357 {
2358         if (level != SOL_TCP)
2359                 return inet_csk_compat_getsockopt(sk, level, optname,
2360                                                   optval, optlen);
2361         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2362 }
2363
2364 EXPORT_SYMBOL(compat_tcp_getsockopt);
2365 #endif
2366
2367 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2368 {
2369         struct sk_buff *segs = ERR_PTR(-EINVAL);
2370         struct tcphdr *th;
2371         unsigned thlen;
2372         unsigned int seq;
2373         __be32 delta;
2374         unsigned int oldlen;
2375         unsigned int len;
2376
2377         if (!pskb_may_pull(skb, sizeof(*th)))
2378                 goto out;
2379
2380         th = tcp_hdr(skb);
2381         thlen = th->doff * 4;
2382         if (thlen < sizeof(*th))
2383                 goto out;
2384
2385         if (!pskb_may_pull(skb, thlen))
2386                 goto out;
2387
2388         oldlen = (u16)~skb->len;
2389         __skb_pull(skb, thlen);
2390
2391         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2392                 /* Packet is from an untrusted source, reset gso_segs. */
2393                 int type = skb_shinfo(skb)->gso_type;
2394                 int mss;
2395
2396                 if (unlikely(type &
2397                              ~(SKB_GSO_TCPV4 |
2398                                SKB_GSO_DODGY |
2399                                SKB_GSO_TCP_ECN |
2400                                SKB_GSO_TCPV6 |
2401                                0) ||
2402                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2403                         goto out;
2404
2405                 mss = skb_shinfo(skb)->gso_size;
2406                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2407
2408                 segs = NULL;
2409                 goto out;
2410         }
2411
2412         segs = skb_segment(skb, features);
2413         if (IS_ERR(segs))
2414                 goto out;
2415
2416         len = skb_shinfo(skb)->gso_size;
2417         delta = htonl(oldlen + (thlen + len));
2418
2419         skb = segs;
2420         th = tcp_hdr(skb);
2421         seq = ntohl(th->seq);
2422
2423         do {
2424                 th->fin = th->psh = 0;
2425
2426                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2427                                        (__force u32)delta));
2428                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2429                         th->check =
2430                              csum_fold(csum_partial(skb_transport_header(skb),
2431                                                     thlen, skb->csum));
2432
2433                 seq += len;
2434                 skb = skb->next;
2435                 th = tcp_hdr(skb);
2436
2437                 th->seq = htonl(seq);
2438                 th->cwr = 0;
2439         } while (skb->next);
2440
2441         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2442                       skb->data_len);
2443         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2444                                 (__force u32)delta));
2445         if (skb->ip_summed != CHECKSUM_PARTIAL)
2446                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2447                                                    thlen, skb->csum));
2448
2449 out:
2450         return segs;
2451 }
2452 EXPORT_SYMBOL(tcp_tso_segment);
2453
2454 #ifdef CONFIG_TCP_MD5SIG
2455 static unsigned long tcp_md5sig_users;
2456 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2457 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2458
2459 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2460 {
2461         int cpu;
2462         for_each_possible_cpu(cpu) {
2463                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2464                 if (p) {
2465                         if (p->md5_desc.tfm)
2466                                 crypto_free_hash(p->md5_desc.tfm);
2467                         kfree(p);
2468                         p = NULL;
2469                 }
2470         }
2471         free_percpu(pool);
2472 }
2473
2474 void tcp_free_md5sig_pool(void)
2475 {
2476         struct tcp_md5sig_pool **pool = NULL;
2477
2478         spin_lock_bh(&tcp_md5sig_pool_lock);
2479         if (--tcp_md5sig_users == 0) {
2480                 pool = tcp_md5sig_pool;
2481                 tcp_md5sig_pool = NULL;
2482         }
2483         spin_unlock_bh(&tcp_md5sig_pool_lock);
2484         if (pool)
2485                 __tcp_free_md5sig_pool(pool);
2486 }
2487
2488 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2489
2490 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2491 {
2492         int cpu;
2493         struct tcp_md5sig_pool **pool;
2494
2495         pool = alloc_percpu(struct tcp_md5sig_pool *);
2496         if (!pool)
2497                 return NULL;
2498
2499         for_each_possible_cpu(cpu) {
2500                 struct tcp_md5sig_pool *p;
2501                 struct crypto_hash *hash;
2502
2503                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2504                 if (!p)
2505                         goto out_free;
2506                 *per_cpu_ptr(pool, cpu) = p;
2507
2508                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2509                 if (!hash || IS_ERR(hash))
2510                         goto out_free;
2511
2512                 p->md5_desc.tfm = hash;
2513         }
2514         return pool;
2515 out_free:
2516         __tcp_free_md5sig_pool(pool);
2517         return NULL;
2518 }
2519
2520 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2521 {
2522         struct tcp_md5sig_pool **pool;
2523         int alloc = 0;
2524
2525 retry:
2526         spin_lock_bh(&tcp_md5sig_pool_lock);
2527         pool = tcp_md5sig_pool;
2528         if (tcp_md5sig_users++ == 0) {
2529                 alloc = 1;
2530                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2531         } else if (!pool) {
2532                 tcp_md5sig_users--;
2533                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2534                 cpu_relax();
2535                 goto retry;
2536         } else
2537                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2538
2539         if (alloc) {
2540                 /* we cannot hold spinlock here because this may sleep. */
2541                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2542                 spin_lock_bh(&tcp_md5sig_pool_lock);
2543                 if (!p) {
2544                         tcp_md5sig_users--;
2545                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2546                         return NULL;
2547                 }
2548                 pool = tcp_md5sig_pool;
2549                 if (pool) {
2550                         /* oops, it has already been assigned. */
2551                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2552                         __tcp_free_md5sig_pool(p);
2553                 } else {
2554                         tcp_md5sig_pool = pool = p;
2555                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2556                 }
2557         }
2558         return pool;
2559 }
2560
2561 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2562
2563 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2564 {
2565         struct tcp_md5sig_pool **p;
2566         spin_lock_bh(&tcp_md5sig_pool_lock);
2567         p = tcp_md5sig_pool;
2568         if (p)
2569                 tcp_md5sig_users++;
2570         spin_unlock_bh(&tcp_md5sig_pool_lock);
2571         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2572 }
2573
2574 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2575
2576 void __tcp_put_md5sig_pool(void)
2577 {
2578         tcp_free_md5sig_pool();
2579 }
2580
2581 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2582 #endif
2583
2584 void tcp_done(struct sock *sk)
2585 {
2586         if(sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2587                 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
2588
2589         tcp_set_state(sk, TCP_CLOSE);
2590         tcp_clear_xmit_timers(sk);
2591
2592         sk->sk_shutdown = SHUTDOWN_MASK;
2593
2594         if (!sock_flag(sk, SOCK_DEAD))
2595                 sk->sk_state_change(sk);
2596         else
2597                 inet_csk_destroy_sock(sk);
2598 }
2599 EXPORT_SYMBOL_GPL(tcp_done);
2600
2601 extern struct tcp_congestion_ops tcp_reno;
2602
2603 static __initdata unsigned long thash_entries;
2604 static int __init set_thash_entries(char *str)
2605 {
2606         if (!str)
2607                 return 0;
2608         thash_entries = simple_strtoul(str, &str, 0);
2609         return 1;
2610 }
2611 __setup("thash_entries=", set_thash_entries);
2612
2613 void __init tcp_init(void)
2614 {
2615         struct sk_buff *skb = NULL;
2616         unsigned long limit;
2617         int order, i, max_share;
2618
2619         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2620
2621         tcp_hashinfo.bind_bucket_cachep =
2622                 kmem_cache_create("tcp_bind_bucket",
2623                                   sizeof(struct inet_bind_bucket), 0,
2624                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2625
2626         /* Size and allocate the main established and bind bucket
2627          * hash tables.
2628          *
2629          * The methodology is similar to that of the buffer cache.
2630          */
2631         tcp_hashinfo.ehash =
2632                 alloc_large_system_hash("TCP established",
2633                                         sizeof(struct inet_ehash_bucket),
2634                                         thash_entries,
2635                                         (num_physpages >= 128 * 1024) ?
2636                                         13 : 15,
2637                                         0,
2638                                         &tcp_hashinfo.ehash_size,
2639                                         NULL,
2640                                         thash_entries ? 0 : 512 * 1024);
2641         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2642         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2643                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2644                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
2645         }
2646         if (inet_ehash_locks_alloc(&tcp_hashinfo))
2647                 panic("TCP: failed to alloc ehash_locks");
2648         tcp_hashinfo.bhash =
2649                 alloc_large_system_hash("TCP bind",
2650                                         sizeof(struct inet_bind_hashbucket),
2651                                         tcp_hashinfo.ehash_size,
2652                                         (num_physpages >= 128 * 1024) ?
2653                                         13 : 15,
2654                                         0,
2655                                         &tcp_hashinfo.bhash_size,
2656                                         NULL,
2657                                         64 * 1024);
2658         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2659         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2660                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2661                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2662         }
2663
2664         /* Try to be a bit smarter and adjust defaults depending
2665          * on available memory.
2666          */
2667         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2668                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2669                         order++)
2670                 ;
2671         if (order >= 4) {
2672                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2673                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2674                 sysctl_max_syn_backlog = 1024;
2675         } else if (order < 3) {
2676                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2677                 sysctl_tcp_max_orphans >>= (3 - order);
2678                 sysctl_max_syn_backlog = 128;
2679         }
2680
2681         /* Set the pressure threshold to be a fraction of global memory that
2682          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2683          * memory, with a floor of 128 pages.
2684          */
2685         limit = min(nr_all_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2686         limit = (limit * (nr_all_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2687         limit = max(limit, 128UL);
2688         sysctl_tcp_mem[0] = limit / 4 * 3;
2689         sysctl_tcp_mem[1] = limit;
2690         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2691
2692         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2693         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2694         max_share = min(4UL*1024*1024, limit);
2695
2696         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2697         sysctl_tcp_wmem[1] = 16*1024;
2698         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2699
2700         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2701         sysctl_tcp_rmem[1] = 87380;
2702         sysctl_tcp_rmem[2] = max(87380, max_share);
2703
2704         printk(KERN_INFO "TCP: Hash tables configured "
2705                "(established %d bind %d)\n",
2706                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2707
2708         tcp_register_congestion_control(&tcp_reno);
2709 }
2710
2711 EXPORT_SYMBOL(tcp_close);
2712 EXPORT_SYMBOL(tcp_disconnect);
2713 EXPORT_SYMBOL(tcp_getsockopt);
2714 EXPORT_SYMBOL(tcp_ioctl);
2715 EXPORT_SYMBOL(tcp_poll);
2716 EXPORT_SYMBOL(tcp_read_sock);
2717 EXPORT_SYMBOL(tcp_recvmsg);
2718 EXPORT_SYMBOL(tcp_sendmsg);
2719 EXPORT_SYMBOL(tcp_splice_read);
2720 EXPORT_SYMBOL(tcp_sendpage);
2721 EXPORT_SYMBOL(tcp_setsockopt);
2722 EXPORT_SYMBOL(tcp_shutdown);
2723 EXPORT_SYMBOL(tcp_statistics);