987b94403be55be814fb5245300236fb0b863558
[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/module.h>
251 #include <linux/types.h>
252 #include <linux/fcntl.h>
253 #include <linux/poll.h>
254 #include <linux/init.h>
255 #include <linux/fs.h>
256 #include <linux/random.h>
257 #include <linux/bootmem.h>
258 #include <linux/cache.h>
259 #include <linux/err.h>
260 #include <linux/crypto.h>
261
262 #include <net/icmp.h>
263 #include <net/tcp.h>
264 #include <net/xfrm.h>
265 #include <net/ip.h>
266 #include <net/netdma.h>
267
268 #include <asm/uaccess.h>
269 #include <asm/ioctls.h>
270
271 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
272
273 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
274
275 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
276
277 EXPORT_SYMBOL_GPL(tcp_orphan_count);
278
279 int sysctl_tcp_mem[3] __read_mostly;
280 int sysctl_tcp_wmem[3] __read_mostly;
281 int sysctl_tcp_rmem[3] __read_mostly;
282
283 EXPORT_SYMBOL(sysctl_tcp_mem);
284 EXPORT_SYMBOL(sysctl_tcp_rmem);
285 EXPORT_SYMBOL(sysctl_tcp_wmem);
286
287 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
288 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
289
290 EXPORT_SYMBOL(tcp_memory_allocated);
291 EXPORT_SYMBOL(tcp_sockets_allocated);
292
293 /*
294  * Pressure flag: try to collapse.
295  * Technical note: it is used by multiple contexts non atomically.
296  * All the sk_stream_mem_schedule() is of this nature: accounting
297  * is strict, actions are advisory and have some latency.
298  */
299 int tcp_memory_pressure __read_mostly;
300
301 EXPORT_SYMBOL(tcp_memory_pressure);
302
303 void tcp_enter_memory_pressure(void)
304 {
305         if (!tcp_memory_pressure) {
306                 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
307                 tcp_memory_pressure = 1;
308         }
309 }
310
311 EXPORT_SYMBOL(tcp_enter_memory_pressure);
312
313 /*
314  *      Wait for a TCP event.
315  *
316  *      Note that we don't need to lock the socket, as the upper poll layers
317  *      take care of normal races (between the test and the event) and we don't
318  *      go look at any of the socket buffers directly.
319  */
320 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
321 {
322         unsigned int mask;
323         struct sock *sk = sock->sk;
324         struct tcp_sock *tp = tcp_sk(sk);
325
326         poll_wait(file, sk->sk_sleep, wait);
327         if (sk->sk_state == TCP_LISTEN)
328                 return inet_csk_listen_poll(sk);
329
330         /* Socket is not locked. We are protected from async events
331            by poll logic and correct handling of state changes
332            made by another threads is impossible in any case.
333          */
334
335         mask = 0;
336         if (sk->sk_err)
337                 mask = POLLERR;
338
339         /*
340          * POLLHUP is certainly not done right. But poll() doesn't
341          * have a notion of HUP in just one direction, and for a
342          * socket the read side is more interesting.
343          *
344          * Some poll() documentation says that POLLHUP is incompatible
345          * with the POLLOUT/POLLWR flags, so somebody should check this
346          * all. But careful, it tends to be safer to return too many
347          * bits than too few, and you can easily break real applications
348          * if you don't tell them that something has hung up!
349          *
350          * Check-me.
351          *
352          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
353          * our fs/select.c). It means that after we received EOF,
354          * poll always returns immediately, making impossible poll() on write()
355          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
356          * if and only if shutdown has been made in both directions.
357          * Actually, it is interesting to look how Solaris and DUX
358          * solve this dilemma. I would prefer, if PULLHUP were maskable,
359          * then we could set it on SND_SHUTDOWN. BTW examples given
360          * in Stevens' books assume exactly this behaviour, it explains
361          * why PULLHUP is incompatible with POLLOUT.    --ANK
362          *
363          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
364          * blocking on fresh not-connected or disconnected socket. --ANK
365          */
366         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
367                 mask |= POLLHUP;
368         if (sk->sk_shutdown & RCV_SHUTDOWN)
369                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
370
371         /* Connected? */
372         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
373                 /* Potential race condition. If read of tp below will
374                  * escape above sk->sk_state, we can be illegally awaken
375                  * in SYN_* states. */
376                 if ((tp->rcv_nxt != tp->copied_seq) &&
377                     (tp->urg_seq != tp->copied_seq ||
378                      tp->rcv_nxt != tp->copied_seq + 1 ||
379                      sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
380                         mask |= POLLIN | POLLRDNORM;
381
382                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
383                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
384                                 mask |= POLLOUT | POLLWRNORM;
385                         } else {  /* send SIGIO later */
386                                 set_bit(SOCK_ASYNC_NOSPACE,
387                                         &sk->sk_socket->flags);
388                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
389
390                                 /* Race breaker. If space is freed after
391                                  * wspace test but before the flags are set,
392                                  * IO signal will be lost.
393                                  */
394                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
395                                         mask |= POLLOUT | POLLWRNORM;
396                         }
397                 }
398
399                 if (tp->urg_data & TCP_URG_VALID)
400                         mask |= POLLPRI;
401         }
402         return mask;
403 }
404
405 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
406 {
407         struct tcp_sock *tp = tcp_sk(sk);
408         int answ;
409
410         switch (cmd) {
411         case SIOCINQ:
412                 if (sk->sk_state == TCP_LISTEN)
413                         return -EINVAL;
414
415                 lock_sock(sk);
416                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
417                         answ = 0;
418                 else if (sock_flag(sk, SOCK_URGINLINE) ||
419                          !tp->urg_data ||
420                          before(tp->urg_seq, tp->copied_seq) ||
421                          !before(tp->urg_seq, tp->rcv_nxt)) {
422                         answ = tp->rcv_nxt - tp->copied_seq;
423
424                         /* Subtract 1, if FIN is in queue. */
425                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
426                                 answ -=
427                        tcp_hdr((struct sk_buff *)sk->sk_receive_queue.prev)->fin;
428                 } else
429                         answ = tp->urg_seq - tp->copied_seq;
430                 release_sock(sk);
431                 break;
432         case SIOCATMARK:
433                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
434                 break;
435         case SIOCOUTQ:
436                 if (sk->sk_state == TCP_LISTEN)
437                         return -EINVAL;
438
439                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
440                         answ = 0;
441                 else
442                         answ = tp->write_seq - tp->snd_una;
443                 break;
444         default:
445                 return -ENOIOCTLCMD;
446         }
447
448         return put_user(answ, (int __user *)arg);
449 }
450
451 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
452 {
453         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
454         tp->pushed_seq = tp->write_seq;
455 }
456
457 static inline int forced_push(struct tcp_sock *tp)
458 {
459         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
460 }
461
462 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
463 {
464         struct tcp_sock *tp = tcp_sk(sk);
465         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
466
467         skb->csum    = 0;
468         tcb->seq     = tcb->end_seq = tp->write_seq;
469         tcb->flags   = TCPCB_FLAG_ACK;
470         tcb->sacked  = 0;
471         skb_header_release(skb);
472         tcp_add_write_queue_tail(sk, skb);
473         sk_charge_skb(sk, skb);
474         if (tp->nonagle & TCP_NAGLE_PUSH)
475                 tp->nonagle &= ~TCP_NAGLE_PUSH;
476 }
477
478 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
479                                 struct sk_buff *skb)
480 {
481         if (flags & MSG_OOB) {
482                 tp->urg_mode = 1;
483                 tp->snd_up = tp->write_seq;
484                 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
485         }
486 }
487
488 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
489                             int nonagle)
490 {
491         struct tcp_sock *tp = tcp_sk(sk);
492
493         if (tcp_send_head(sk)) {
494                 struct sk_buff *skb = tcp_write_queue_tail(sk);
495                 if (!(flags & MSG_MORE) || forced_push(tp))
496                         tcp_mark_push(tp, skb);
497                 tcp_mark_urg(tp, flags, skb);
498                 __tcp_push_pending_frames(sk, mss_now,
499                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
500         }
501 }
502
503 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
504                          size_t psize, int flags)
505 {
506         struct tcp_sock *tp = tcp_sk(sk);
507         int mss_now, size_goal;
508         int err;
509         ssize_t copied;
510         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
511
512         /* Wait for a connection to finish. */
513         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
514                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
515                         goto out_err;
516
517         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
518
519         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
520         size_goal = tp->xmit_size_goal;
521         copied = 0;
522
523         err = -EPIPE;
524         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
525                 goto do_error;
526
527         while (psize > 0) {
528                 struct sk_buff *skb = tcp_write_queue_tail(sk);
529                 struct page *page = pages[poffset / PAGE_SIZE];
530                 int copy, i, can_coalesce;
531                 int offset = poffset % PAGE_SIZE;
532                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
533
534                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
535 new_segment:
536                         if (!sk_stream_memory_free(sk))
537                                 goto wait_for_sndbuf;
538
539                         skb = sk_stream_alloc_pskb(sk, 0, 0,
540                                                    sk->sk_allocation);
541                         if (!skb)
542                                 goto wait_for_memory;
543
544                         skb_entail(sk, skb);
545                         copy = size_goal;
546                 }
547
548                 if (copy > size)
549                         copy = size;
550
551                 i = skb_shinfo(skb)->nr_frags;
552                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
553                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
554                         tcp_mark_push(tp, skb);
555                         goto new_segment;
556                 }
557                 if (!sk_stream_wmem_schedule(sk, copy))
558                         goto wait_for_memory;
559
560                 if (can_coalesce) {
561                         skb_shinfo(skb)->frags[i - 1].size += copy;
562                 } else {
563                         get_page(page);
564                         skb_fill_page_desc(skb, i, page, offset, copy);
565                 }
566
567                 skb->len += copy;
568                 skb->data_len += copy;
569                 skb->truesize += copy;
570                 sk->sk_wmem_queued += copy;
571                 sk->sk_forward_alloc -= copy;
572                 skb->ip_summed = CHECKSUM_PARTIAL;
573                 tp->write_seq += copy;
574                 TCP_SKB_CB(skb)->end_seq += copy;
575                 skb_shinfo(skb)->gso_segs = 0;
576
577                 if (!copied)
578                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
579
580                 copied += copy;
581                 poffset += copy;
582                 if (!(psize -= copy))
583                         goto out;
584
585                 if (skb->len < mss_now || (flags & MSG_OOB))
586                         continue;
587
588                 if (forced_push(tp)) {
589                         tcp_mark_push(tp, skb);
590                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
591                 } else if (skb == tcp_send_head(sk))
592                         tcp_push_one(sk, mss_now);
593                 continue;
594
595 wait_for_sndbuf:
596                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
597 wait_for_memory:
598                 if (copied)
599                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
600
601                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
602                         goto do_error;
603
604                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
605                 size_goal = tp->xmit_size_goal;
606         }
607
608 out:
609         if (copied)
610                 tcp_push(sk, flags, mss_now, tp->nonagle);
611         return copied;
612
613 do_error:
614         if (copied)
615                 goto out;
616 out_err:
617         return sk_stream_error(sk, flags, err);
618 }
619
620 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
621                      size_t size, int flags)
622 {
623         ssize_t res;
624         struct sock *sk = sock->sk;
625
626         if (!(sk->sk_route_caps & NETIF_F_SG) ||
627             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
628                 return sock_no_sendpage(sock, page, offset, size, flags);
629
630         lock_sock(sk);
631         TCP_CHECK_TIMER(sk);
632         res = do_tcp_sendpages(sk, &page, offset, size, flags);
633         TCP_CHECK_TIMER(sk);
634         release_sock(sk);
635         return res;
636 }
637
638 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
639 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
640
641 static inline int select_size(struct sock *sk)
642 {
643         struct tcp_sock *tp = tcp_sk(sk);
644         int tmp = tp->mss_cache;
645
646         if (sk->sk_route_caps & NETIF_F_SG) {
647                 if (sk_can_gso(sk))
648                         tmp = 0;
649                 else {
650                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
651
652                         if (tmp >= pgbreak &&
653                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
654                                 tmp = pgbreak;
655                 }
656         }
657
658         return tmp;
659 }
660
661 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
662                 size_t size)
663 {
664         struct iovec *iov;
665         struct tcp_sock *tp = tcp_sk(sk);
666         struct sk_buff *skb;
667         int iovlen, flags;
668         int mss_now, size_goal;
669         int err, copied;
670         long timeo;
671
672         lock_sock(sk);
673         TCP_CHECK_TIMER(sk);
674
675         flags = msg->msg_flags;
676         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
677
678         /* Wait for a connection to finish. */
679         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
680                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
681                         goto out_err;
682
683         /* This should be in poll */
684         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
685
686         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
687         size_goal = tp->xmit_size_goal;
688
689         /* Ok commence sending. */
690         iovlen = msg->msg_iovlen;
691         iov = msg->msg_iov;
692         copied = 0;
693
694         err = -EPIPE;
695         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
696                 goto do_error;
697
698         while (--iovlen >= 0) {
699                 int seglen = iov->iov_len;
700                 unsigned char __user *from = iov->iov_base;
701
702                 iov++;
703
704                 while (seglen > 0) {
705                         int copy;
706
707                         skb = tcp_write_queue_tail(sk);
708
709                         if (!tcp_send_head(sk) ||
710                             (copy = size_goal - skb->len) <= 0) {
711
712 new_segment:
713                                 /* Allocate new segment. If the interface is SG,
714                                  * allocate skb fitting to single page.
715                                  */
716                                 if (!sk_stream_memory_free(sk))
717                                         goto wait_for_sndbuf;
718
719                                 skb = sk_stream_alloc_pskb(sk, select_size(sk),
720                                                            0, sk->sk_allocation);
721                                 if (!skb)
722                                         goto wait_for_memory;
723
724                                 /*
725                                  * Check whether we can use HW checksum.
726                                  */
727                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
728                                         skb->ip_summed = CHECKSUM_PARTIAL;
729
730                                 skb_entail(sk, skb);
731                                 copy = size_goal;
732                         }
733
734                         /* Try to append data to the end of skb. */
735                         if (copy > seglen)
736                                 copy = seglen;
737
738                         /* Where to copy to? */
739                         if (skb_tailroom(skb) > 0) {
740                                 /* We have some space in skb head. Superb! */
741                                 if (copy > skb_tailroom(skb))
742                                         copy = skb_tailroom(skb);
743                                 if ((err = skb_add_data(skb, from, copy)) != 0)
744                                         goto do_fault;
745                         } else {
746                                 int merge = 0;
747                                 int i = skb_shinfo(skb)->nr_frags;
748                                 struct page *page = TCP_PAGE(sk);
749                                 int off = TCP_OFF(sk);
750
751                                 if (skb_can_coalesce(skb, i, page, off) &&
752                                     off != PAGE_SIZE) {
753                                         /* We can extend the last page
754                                          * fragment. */
755                                         merge = 1;
756                                 } else if (i == MAX_SKB_FRAGS ||
757                                            (!i &&
758                                            !(sk->sk_route_caps & NETIF_F_SG))) {
759                                         /* Need to add new fragment and cannot
760                                          * do this because interface is non-SG,
761                                          * or because all the page slots are
762                                          * busy. */
763                                         tcp_mark_push(tp, skb);
764                                         goto new_segment;
765                                 } else if (page) {
766                                         if (off == PAGE_SIZE) {
767                                                 put_page(page);
768                                                 TCP_PAGE(sk) = page = NULL;
769                                                 off = 0;
770                                         }
771                                 } else
772                                         off = 0;
773
774                                 if (copy > PAGE_SIZE - off)
775                                         copy = PAGE_SIZE - off;
776
777                                 if (!sk_stream_wmem_schedule(sk, copy))
778                                         goto wait_for_memory;
779
780                                 if (!page) {
781                                         /* Allocate new cache page. */
782                                         if (!(page = sk_stream_alloc_page(sk)))
783                                                 goto wait_for_memory;
784                                 }
785
786                                 /* Time to copy data. We are close to
787                                  * the end! */
788                                 err = skb_copy_to_page(sk, from, skb, page,
789                                                        off, copy);
790                                 if (err) {
791                                         /* If this page was new, give it to the
792                                          * socket so it does not get leaked.
793                                          */
794                                         if (!TCP_PAGE(sk)) {
795                                                 TCP_PAGE(sk) = page;
796                                                 TCP_OFF(sk) = 0;
797                                         }
798                                         goto do_error;
799                                 }
800
801                                 /* Update the skb. */
802                                 if (merge) {
803                                         skb_shinfo(skb)->frags[i - 1].size +=
804                                                                         copy;
805                                 } else {
806                                         skb_fill_page_desc(skb, i, page, off, copy);
807                                         if (TCP_PAGE(sk)) {
808                                                 get_page(page);
809                                         } else if (off + copy < PAGE_SIZE) {
810                                                 get_page(page);
811                                                 TCP_PAGE(sk) = page;
812                                         }
813                                 }
814
815                                 TCP_OFF(sk) = off + copy;
816                         }
817
818                         if (!copied)
819                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
820
821                         tp->write_seq += copy;
822                         TCP_SKB_CB(skb)->end_seq += copy;
823                         skb_shinfo(skb)->gso_segs = 0;
824
825                         from += copy;
826                         copied += copy;
827                         if ((seglen -= copy) == 0 && iovlen == 0)
828                                 goto out;
829
830                         if (skb->len < mss_now || (flags & MSG_OOB))
831                                 continue;
832
833                         if (forced_push(tp)) {
834                                 tcp_mark_push(tp, skb);
835                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
836                         } else if (skb == tcp_send_head(sk))
837                                 tcp_push_one(sk, mss_now);
838                         continue;
839
840 wait_for_sndbuf:
841                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
842 wait_for_memory:
843                         if (copied)
844                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
845
846                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
847                                 goto do_error;
848
849                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
850                         size_goal = tp->xmit_size_goal;
851                 }
852         }
853
854 out:
855         if (copied)
856                 tcp_push(sk, flags, mss_now, tp->nonagle);
857         TCP_CHECK_TIMER(sk);
858         release_sock(sk);
859         return copied;
860
861 do_fault:
862         if (!skb->len) {
863                 tcp_unlink_write_queue(skb, sk);
864                 /* It is the one place in all of TCP, except connection
865                  * reset, where we can be unlinking the send_head.
866                  */
867                 tcp_check_send_head(sk, skb);
868                 sk_stream_free_skb(sk, skb);
869         }
870
871 do_error:
872         if (copied)
873                 goto out;
874 out_err:
875         err = sk_stream_error(sk, flags, err);
876         TCP_CHECK_TIMER(sk);
877         release_sock(sk);
878         return err;
879 }
880
881 /*
882  *      Handle reading urgent data. BSD has very simple semantics for
883  *      this, no blocking and very strange errors 8)
884  */
885
886 static int tcp_recv_urg(struct sock *sk, long timeo,
887                         struct msghdr *msg, int len, int flags,
888                         int *addr_len)
889 {
890         struct tcp_sock *tp = tcp_sk(sk);
891
892         /* No URG data to read. */
893         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
894             tp->urg_data == TCP_URG_READ)
895                 return -EINVAL; /* Yes this is right ! */
896
897         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
898                 return -ENOTCONN;
899
900         if (tp->urg_data & TCP_URG_VALID) {
901                 int err = 0;
902                 char c = tp->urg_data;
903
904                 if (!(flags & MSG_PEEK))
905                         tp->urg_data = TCP_URG_READ;
906
907                 /* Read urgent data. */
908                 msg->msg_flags |= MSG_OOB;
909
910                 if (len > 0) {
911                         if (!(flags & MSG_TRUNC))
912                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
913                         len = 1;
914                 } else
915                         msg->msg_flags |= MSG_TRUNC;
916
917                 return err ? -EFAULT : len;
918         }
919
920         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
921                 return 0;
922
923         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
924          * the available implementations agree in this case:
925          * this call should never block, independent of the
926          * blocking state of the socket.
927          * Mike <pall@rz.uni-karlsruhe.de>
928          */
929         return -EAGAIN;
930 }
931
932 /* Clean up the receive buffer for full frames taken by the user,
933  * then send an ACK if necessary.  COPIED is the number of bytes
934  * tcp_recvmsg has given to the user so far, it speeds up the
935  * calculation of whether or not we must ACK for the sake of
936  * a window update.
937  */
938 void tcp_cleanup_rbuf(struct sock *sk, int copied)
939 {
940         struct tcp_sock *tp = tcp_sk(sk);
941         int time_to_ack = 0;
942
943 #if TCP_DEBUG
944         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
945
946         BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
947 #endif
948
949         if (inet_csk_ack_scheduled(sk)) {
950                 const struct inet_connection_sock *icsk = inet_csk(sk);
951                    /* Delayed ACKs frequently hit locked sockets during bulk
952                     * receive. */
953                 if (icsk->icsk_ack.blocked ||
954                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
955                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
956                     /*
957                      * If this read emptied read buffer, we send ACK, if
958                      * connection is not bidirectional, user drained
959                      * receive buffer and there was a small segment
960                      * in queue.
961                      */
962                     (copied > 0 &&
963                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
964                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
965                        !icsk->icsk_ack.pingpong)) &&
966                       !atomic_read(&sk->sk_rmem_alloc)))
967                         time_to_ack = 1;
968         }
969
970         /* We send an ACK if we can now advertise a non-zero window
971          * which has been raised "significantly".
972          *
973          * Even if window raised up to infinity, do not send window open ACK
974          * in states, where we will not receive more. It is useless.
975          */
976         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
977                 __u32 rcv_window_now = tcp_receive_window(tp);
978
979                 /* Optimize, __tcp_select_window() is not cheap. */
980                 if (2*rcv_window_now <= tp->window_clamp) {
981                         __u32 new_window = __tcp_select_window(sk);
982
983                         /* Send ACK now, if this read freed lots of space
984                          * in our buffer. Certainly, new_window is new window.
985                          * We can advertise it now, if it is not less than current one.
986                          * "Lots" means "at least twice" here.
987                          */
988                         if (new_window && new_window >= 2 * rcv_window_now)
989                                 time_to_ack = 1;
990                 }
991         }
992         if (time_to_ack)
993                 tcp_send_ack(sk);
994 }
995
996 static void tcp_prequeue_process(struct sock *sk)
997 {
998         struct sk_buff *skb;
999         struct tcp_sock *tp = tcp_sk(sk);
1000
1001         NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
1002
1003         /* RX process wants to run with disabled BHs, though it is not
1004          * necessary */
1005         local_bh_disable();
1006         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1007                 sk->sk_backlog_rcv(sk, skb);
1008         local_bh_enable();
1009
1010         /* Clear memory counter. */
1011         tp->ucopy.memory = 0;
1012 }
1013
1014 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1015 {
1016         struct sk_buff *skb;
1017         u32 offset;
1018
1019         skb_queue_walk(&sk->sk_receive_queue, skb) {
1020                 offset = seq - TCP_SKB_CB(skb)->seq;
1021                 if (tcp_hdr(skb)->syn)
1022                         offset--;
1023                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1024                         *off = offset;
1025                         return skb;
1026                 }
1027         }
1028         return NULL;
1029 }
1030
1031 /*
1032  * This routine provides an alternative to tcp_recvmsg() for routines
1033  * that would like to handle copying from skbuffs directly in 'sendfile'
1034  * fashion.
1035  * Note:
1036  *      - It is assumed that the socket was locked by the caller.
1037  *      - The routine does not block.
1038  *      - At present, there is no support for reading OOB data
1039  *        or for 'peeking' the socket using this routine
1040  *        (although both would be easy to implement).
1041  */
1042 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1043                   sk_read_actor_t recv_actor)
1044 {
1045         struct sk_buff *skb;
1046         struct tcp_sock *tp = tcp_sk(sk);
1047         u32 seq = tp->copied_seq;
1048         u32 offset;
1049         int copied = 0;
1050
1051         if (sk->sk_state == TCP_LISTEN)
1052                 return -ENOTCONN;
1053         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1054                 if (offset < skb->len) {
1055                         size_t used, len;
1056
1057                         len = skb->len - offset;
1058                         /* Stop reading if we hit a patch of urgent data */
1059                         if (tp->urg_data) {
1060                                 u32 urg_offset = tp->urg_seq - seq;
1061                                 if (urg_offset < len)
1062                                         len = urg_offset;
1063                                 if (!len)
1064                                         break;
1065                         }
1066                         used = recv_actor(desc, skb, offset, len);
1067                         if (used < 0) {
1068                                 if (!copied)
1069                                         copied = used;
1070                                 break;
1071                         } else if (used <= len) {
1072                                 seq += used;
1073                                 copied += used;
1074                                 offset += used;
1075                         }
1076                         if (offset != skb->len)
1077                                 break;
1078                 }
1079                 if (tcp_hdr(skb)->fin) {
1080                         sk_eat_skb(sk, skb, 0);
1081                         ++seq;
1082                         break;
1083                 }
1084                 sk_eat_skb(sk, skb, 0);
1085                 if (!desc->count)
1086                         break;
1087         }
1088         tp->copied_seq = seq;
1089
1090         tcp_rcv_space_adjust(sk);
1091
1092         /* Clean up data we have read: This will do ACK frames. */
1093         if (copied > 0)
1094                 tcp_cleanup_rbuf(sk, copied);
1095         return copied;
1096 }
1097
1098 /*
1099  *      This routine copies from a sock struct into the user buffer.
1100  *
1101  *      Technical note: in 2.3 we work on _locked_ socket, so that
1102  *      tricks with *seq access order and skb->users are not required.
1103  *      Probably, code can be easily improved even more.
1104  */
1105
1106 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1107                 size_t len, int nonblock, int flags, int *addr_len)
1108 {
1109         struct tcp_sock *tp = tcp_sk(sk);
1110         int copied = 0;
1111         u32 peek_seq;
1112         u32 *seq;
1113         unsigned long used;
1114         int err;
1115         int target;             /* Read at least this many bytes */
1116         long timeo;
1117         struct task_struct *user_recv = NULL;
1118         int copied_early = 0;
1119         struct sk_buff *skb;
1120
1121         lock_sock(sk);
1122
1123         TCP_CHECK_TIMER(sk);
1124
1125         err = -ENOTCONN;
1126         if (sk->sk_state == TCP_LISTEN)
1127                 goto out;
1128
1129         timeo = sock_rcvtimeo(sk, nonblock);
1130
1131         /* Urgent data needs to be handled specially. */
1132         if (flags & MSG_OOB)
1133                 goto recv_urg;
1134
1135         seq = &tp->copied_seq;
1136         if (flags & MSG_PEEK) {
1137                 peek_seq = tp->copied_seq;
1138                 seq = &peek_seq;
1139         }
1140
1141         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1142
1143 #ifdef CONFIG_NET_DMA
1144         tp->ucopy.dma_chan = NULL;
1145         preempt_disable();
1146         skb = skb_peek_tail(&sk->sk_receive_queue);
1147         {
1148                 int available = 0;
1149
1150                 if (skb)
1151                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1152                 if ((available < target) &&
1153                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1154                     !sysctl_tcp_low_latency &&
1155                     __get_cpu_var(softnet_data).net_dma) {
1156                         preempt_enable_no_resched();
1157                         tp->ucopy.pinned_list =
1158                                         dma_pin_iovec_pages(msg->msg_iov, len);
1159                 } else {
1160                         preempt_enable_no_resched();
1161                 }
1162         }
1163 #endif
1164
1165         do {
1166                 u32 offset;
1167
1168                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1169                 if (tp->urg_data && tp->urg_seq == *seq) {
1170                         if (copied)
1171                                 break;
1172                         if (signal_pending(current)) {
1173                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1174                                 break;
1175                         }
1176                 }
1177
1178                 /* Next get a buffer. */
1179
1180                 skb = skb_peek(&sk->sk_receive_queue);
1181                 do {
1182                         if (!skb)
1183                                 break;
1184
1185                         /* Now that we have two receive queues this
1186                          * shouldn't happen.
1187                          */
1188                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1189                                 printk(KERN_INFO "recvmsg bug: copied %X "
1190                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1191                                 break;
1192                         }
1193                         offset = *seq - TCP_SKB_CB(skb)->seq;
1194                         if (tcp_hdr(skb)->syn)
1195                                 offset--;
1196                         if (offset < skb->len)
1197                                 goto found_ok_skb;
1198                         if (tcp_hdr(skb)->fin)
1199                                 goto found_fin_ok;
1200                         BUG_TRAP(flags & MSG_PEEK);
1201                         skb = skb->next;
1202                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1203
1204                 /* Well, if we have backlog, try to process it now yet. */
1205
1206                 if (copied >= target && !sk->sk_backlog.tail)
1207                         break;
1208
1209                 if (copied) {
1210                         if (sk->sk_err ||
1211                             sk->sk_state == TCP_CLOSE ||
1212                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1213                             !timeo ||
1214                             signal_pending(current) ||
1215                             (flags & MSG_PEEK))
1216                                 break;
1217                 } else {
1218                         if (sock_flag(sk, SOCK_DONE))
1219                                 break;
1220
1221                         if (sk->sk_err) {
1222                                 copied = sock_error(sk);
1223                                 break;
1224                         }
1225
1226                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1227                                 break;
1228
1229                         if (sk->sk_state == TCP_CLOSE) {
1230                                 if (!sock_flag(sk, SOCK_DONE)) {
1231                                         /* This occurs when user tries to read
1232                                          * from never connected socket.
1233                                          */
1234                                         copied = -ENOTCONN;
1235                                         break;
1236                                 }
1237                                 break;
1238                         }
1239
1240                         if (!timeo) {
1241                                 copied = -EAGAIN;
1242                                 break;
1243                         }
1244
1245                         if (signal_pending(current)) {
1246                                 copied = sock_intr_errno(timeo);
1247                                 break;
1248                         }
1249                 }
1250
1251                 tcp_cleanup_rbuf(sk, copied);
1252
1253                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1254                         /* Install new reader */
1255                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1256                                 user_recv = current;
1257                                 tp->ucopy.task = user_recv;
1258                                 tp->ucopy.iov = msg->msg_iov;
1259                         }
1260
1261                         tp->ucopy.len = len;
1262
1263                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1264                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1265
1266                         /* Ugly... If prequeue is not empty, we have to
1267                          * process it before releasing socket, otherwise
1268                          * order will be broken at second iteration.
1269                          * More elegant solution is required!!!
1270                          *
1271                          * Look: we have the following (pseudo)queues:
1272                          *
1273                          * 1. packets in flight
1274                          * 2. backlog
1275                          * 3. prequeue
1276                          * 4. receive_queue
1277                          *
1278                          * Each queue can be processed only if the next ones
1279                          * are empty. At this point we have empty receive_queue.
1280                          * But prequeue _can_ be not empty after 2nd iteration,
1281                          * when we jumped to start of loop because backlog
1282                          * processing added something to receive_queue.
1283                          * We cannot release_sock(), because backlog contains
1284                          * packets arrived _after_ prequeued ones.
1285                          *
1286                          * Shortly, algorithm is clear --- to process all
1287                          * the queues in order. We could make it more directly,
1288                          * requeueing packets from backlog to prequeue, if
1289                          * is not empty. It is more elegant, but eats cycles,
1290                          * unfortunately.
1291                          */
1292                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1293                                 goto do_prequeue;
1294
1295                         /* __ Set realtime policy in scheduler __ */
1296                 }
1297
1298                 if (copied >= target) {
1299                         /* Do not sleep, just process backlog. */
1300                         release_sock(sk);
1301                         lock_sock(sk);
1302                 } else
1303                         sk_wait_data(sk, &timeo);
1304
1305 #ifdef CONFIG_NET_DMA
1306                 tp->ucopy.wakeup = 0;
1307 #endif
1308
1309                 if (user_recv) {
1310                         int chunk;
1311
1312                         /* __ Restore normal policy in scheduler __ */
1313
1314                         if ((chunk = len - tp->ucopy.len) != 0) {
1315                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1316                                 len -= chunk;
1317                                 copied += chunk;
1318                         }
1319
1320                         if (tp->rcv_nxt == tp->copied_seq &&
1321                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1322 do_prequeue:
1323                                 tcp_prequeue_process(sk);
1324
1325                                 if ((chunk = len - tp->ucopy.len) != 0) {
1326                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1327                                         len -= chunk;
1328                                         copied += chunk;
1329                                 }
1330                         }
1331                 }
1332                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1333                         if (net_ratelimit())
1334                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1335                                        current->comm, current->pid);
1336                         peek_seq = tp->copied_seq;
1337                 }
1338                 continue;
1339
1340         found_ok_skb:
1341                 /* Ok so how much can we use? */
1342                 used = skb->len - offset;
1343                 if (len < used)
1344                         used = len;
1345
1346                 /* Do we have urgent data here? */
1347                 if (tp->urg_data) {
1348                         u32 urg_offset = tp->urg_seq - *seq;
1349                         if (urg_offset < used) {
1350                                 if (!urg_offset) {
1351                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1352                                                 ++*seq;
1353                                                 offset++;
1354                                                 used--;
1355                                                 if (!used)
1356                                                         goto skip_copy;
1357                                         }
1358                                 } else
1359                                         used = urg_offset;
1360                         }
1361                 }
1362
1363                 if (!(flags & MSG_TRUNC)) {
1364 #ifdef CONFIG_NET_DMA
1365                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1366                                 tp->ucopy.dma_chan = get_softnet_dma();
1367
1368                         if (tp->ucopy.dma_chan) {
1369                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1370                                         tp->ucopy.dma_chan, skb, offset,
1371                                         msg->msg_iov, used,
1372                                         tp->ucopy.pinned_list);
1373
1374                                 if (tp->ucopy.dma_cookie < 0) {
1375
1376                                         printk(KERN_ALERT "dma_cookie < 0\n");
1377
1378                                         /* Exception. Bailout! */
1379                                         if (!copied)
1380                                                 copied = -EFAULT;
1381                                         break;
1382                                 }
1383                                 if ((offset + used) == skb->len)
1384                                         copied_early = 1;
1385
1386                         } else
1387 #endif
1388                         {
1389                                 err = skb_copy_datagram_iovec(skb, offset,
1390                                                 msg->msg_iov, used);
1391                                 if (err) {
1392                                         /* Exception. Bailout! */
1393                                         if (!copied)
1394                                                 copied = -EFAULT;
1395                                         break;
1396                                 }
1397                         }
1398                 }
1399
1400                 *seq += used;
1401                 copied += used;
1402                 len -= used;
1403
1404                 tcp_rcv_space_adjust(sk);
1405
1406 skip_copy:
1407                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1408                         tp->urg_data = 0;
1409                         tcp_fast_path_check(sk);
1410                 }
1411                 if (used + offset < skb->len)
1412                         continue;
1413
1414                 if (tcp_hdr(skb)->fin)
1415                         goto found_fin_ok;
1416                 if (!(flags & MSG_PEEK)) {
1417                         sk_eat_skb(sk, skb, copied_early);
1418                         copied_early = 0;
1419                 }
1420                 continue;
1421
1422         found_fin_ok:
1423                 /* Process the FIN. */
1424                 ++*seq;
1425                 if (!(flags & MSG_PEEK)) {
1426                         sk_eat_skb(sk, skb, copied_early);
1427                         copied_early = 0;
1428                 }
1429                 break;
1430         } while (len > 0);
1431
1432         if (user_recv) {
1433                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1434                         int chunk;
1435
1436                         tp->ucopy.len = copied > 0 ? len : 0;
1437
1438                         tcp_prequeue_process(sk);
1439
1440                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1441                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1442                                 len -= chunk;
1443                                 copied += chunk;
1444                         }
1445                 }
1446
1447                 tp->ucopy.task = NULL;
1448                 tp->ucopy.len = 0;
1449         }
1450
1451 #ifdef CONFIG_NET_DMA
1452         if (tp->ucopy.dma_chan) {
1453                 dma_cookie_t done, used;
1454
1455                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1456
1457                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1458                                                  tp->ucopy.dma_cookie, &done,
1459                                                  &used) == DMA_IN_PROGRESS) {
1460                         /* do partial cleanup of sk_async_wait_queue */
1461                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1462                                (dma_async_is_complete(skb->dma_cookie, done,
1463                                                       used) == DMA_SUCCESS)) {
1464                                 __skb_dequeue(&sk->sk_async_wait_queue);
1465                                 kfree_skb(skb);
1466                         }
1467                 }
1468
1469                 /* Safe to free early-copied skbs now */
1470                 __skb_queue_purge(&sk->sk_async_wait_queue);
1471                 dma_chan_put(tp->ucopy.dma_chan);
1472                 tp->ucopy.dma_chan = NULL;
1473         }
1474         if (tp->ucopy.pinned_list) {
1475                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1476                 tp->ucopy.pinned_list = NULL;
1477         }
1478 #endif
1479
1480         /* According to UNIX98, msg_name/msg_namelen are ignored
1481          * on connected socket. I was just happy when found this 8) --ANK
1482          */
1483
1484         /* Clean up data we have read: This will do ACK frames. */
1485         tcp_cleanup_rbuf(sk, copied);
1486
1487         TCP_CHECK_TIMER(sk);
1488         release_sock(sk);
1489         return copied;
1490
1491 out:
1492         TCP_CHECK_TIMER(sk);
1493         release_sock(sk);
1494         return err;
1495
1496 recv_urg:
1497         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1498         goto out;
1499 }
1500
1501 /*
1502  *      State processing on a close. This implements the state shift for
1503  *      sending our FIN frame. Note that we only send a FIN for some
1504  *      states. A shutdown() may have already sent the FIN, or we may be
1505  *      closed.
1506  */
1507
1508 static const unsigned char new_state[16] = {
1509   /* current state:        new state:      action:      */
1510   /* (Invalid)          */ TCP_CLOSE,
1511   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1512   /* TCP_SYN_SENT       */ TCP_CLOSE,
1513   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1514   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1515   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1516   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1517   /* TCP_CLOSE          */ TCP_CLOSE,
1518   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1519   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1520   /* TCP_LISTEN         */ TCP_CLOSE,
1521   /* TCP_CLOSING        */ TCP_CLOSING,
1522 };
1523
1524 static int tcp_close_state(struct sock *sk)
1525 {
1526         int next = (int)new_state[sk->sk_state];
1527         int ns = next & TCP_STATE_MASK;
1528
1529         tcp_set_state(sk, ns);
1530
1531         return next & TCP_ACTION_FIN;
1532 }
1533
1534 /*
1535  *      Shutdown the sending side of a connection. Much like close except
1536  *      that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1537  */
1538
1539 void tcp_shutdown(struct sock *sk, int how)
1540 {
1541         /*      We need to grab some memory, and put together a FIN,
1542          *      and then put it into the queue to be sent.
1543          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1544          */
1545         if (!(how & SEND_SHUTDOWN))
1546                 return;
1547
1548         /* If we've already sent a FIN, or it's a closed state, skip this. */
1549         if ((1 << sk->sk_state) &
1550             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1551              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1552                 /* Clear out any half completed packets.  FIN if needed. */
1553                 if (tcp_close_state(sk))
1554                         tcp_send_fin(sk);
1555         }
1556 }
1557
1558 void tcp_close(struct sock *sk, long timeout)
1559 {
1560         struct sk_buff *skb;
1561         int data_was_unread = 0;
1562         int state;
1563
1564         lock_sock(sk);
1565         sk->sk_shutdown = SHUTDOWN_MASK;
1566
1567         if (sk->sk_state == TCP_LISTEN) {
1568                 tcp_set_state(sk, TCP_CLOSE);
1569
1570                 /* Special case. */
1571                 inet_csk_listen_stop(sk);
1572
1573                 goto adjudge_to_death;
1574         }
1575
1576         /*  We need to flush the recv. buffs.  We do this only on the
1577          *  descriptor close, not protocol-sourced closes, because the
1578          *  reader process may not have drained the data yet!
1579          */
1580         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1581                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1582                           tcp_hdr(skb)->fin;
1583                 data_was_unread += len;
1584                 __kfree_skb(skb);
1585         }
1586
1587         sk_stream_mem_reclaim(sk);
1588
1589         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1590          * data was lost. To witness the awful effects of the old behavior of
1591          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1592          * GET in an FTP client, suspend the process, wait for the client to
1593          * advertise a zero window, then kill -9 the FTP client, wheee...
1594          * Note: timeout is always zero in such a case.
1595          */
1596         if (data_was_unread) {
1597                 /* Unread data was tossed, zap the connection. */
1598                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1599                 tcp_set_state(sk, TCP_CLOSE);
1600                 tcp_send_active_reset(sk, GFP_KERNEL);
1601         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1602                 /* Check zero linger _after_ checking for unread data. */
1603                 sk->sk_prot->disconnect(sk, 0);
1604                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1605         } else if (tcp_close_state(sk)) {
1606                 /* We FIN if the application ate all the data before
1607                  * zapping the connection.
1608                  */
1609
1610                 /* RED-PEN. Formally speaking, we have broken TCP state
1611                  * machine. State transitions:
1612                  *
1613                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1614                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1615                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1616                  *
1617                  * are legal only when FIN has been sent (i.e. in window),
1618                  * rather than queued out of window. Purists blame.
1619                  *
1620                  * F.e. "RFC state" is ESTABLISHED,
1621                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1622                  *
1623                  * The visible declinations are that sometimes
1624                  * we enter time-wait state, when it is not required really
1625                  * (harmless), do not send active resets, when they are
1626                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1627                  * they look as CLOSING or LAST_ACK for Linux)
1628                  * Probably, I missed some more holelets.
1629                  *                                              --ANK
1630                  */
1631                 tcp_send_fin(sk);
1632         }
1633
1634         sk_stream_wait_close(sk, timeout);
1635
1636 adjudge_to_death:
1637         state = sk->sk_state;
1638         sock_hold(sk);
1639         sock_orphan(sk);
1640         atomic_inc(sk->sk_prot->orphan_count);
1641
1642         /* It is the last release_sock in its life. It will remove backlog. */
1643         release_sock(sk);
1644
1645
1646         /* Now socket is owned by kernel and we acquire BH lock
1647            to finish close. No need to check for user refs.
1648          */
1649         local_bh_disable();
1650         bh_lock_sock(sk);
1651         BUG_TRAP(!sock_owned_by_user(sk));
1652
1653         /* Have we already been destroyed by a softirq or backlog? */
1654         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1655                 goto out;
1656
1657         /*      This is a (useful) BSD violating of the RFC. There is a
1658          *      problem with TCP as specified in that the other end could
1659          *      keep a socket open forever with no application left this end.
1660          *      We use a 3 minute timeout (about the same as BSD) then kill
1661          *      our end. If they send after that then tough - BUT: long enough
1662          *      that we won't make the old 4*rto = almost no time - whoops
1663          *      reset mistake.
1664          *
1665          *      Nope, it was not mistake. It is really desired behaviour
1666          *      f.e. on http servers, when such sockets are useless, but
1667          *      consume significant resources. Let's do it with special
1668          *      linger2 option.                                 --ANK
1669          */
1670
1671         if (sk->sk_state == TCP_FIN_WAIT2) {
1672                 struct tcp_sock *tp = tcp_sk(sk);
1673                 if (tp->linger2 < 0) {
1674                         tcp_set_state(sk, TCP_CLOSE);
1675                         tcp_send_active_reset(sk, GFP_ATOMIC);
1676                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1677                 } else {
1678                         const int tmo = tcp_fin_time(sk);
1679
1680                         if (tmo > TCP_TIMEWAIT_LEN) {
1681                                 inet_csk_reset_keepalive_timer(sk,
1682                                                 tmo - TCP_TIMEWAIT_LEN);
1683                         } else {
1684                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1685                                 goto out;
1686                         }
1687                 }
1688         }
1689         if (sk->sk_state != TCP_CLOSE) {
1690                 sk_stream_mem_reclaim(sk);
1691                 if (tcp_too_many_orphans(sk,
1692                                 atomic_read(sk->sk_prot->orphan_count))) {
1693                         if (net_ratelimit())
1694                                 printk(KERN_INFO "TCP: too many of orphaned "
1695                                        "sockets\n");
1696                         tcp_set_state(sk, TCP_CLOSE);
1697                         tcp_send_active_reset(sk, GFP_ATOMIC);
1698                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1699                 }
1700         }
1701
1702         if (sk->sk_state == TCP_CLOSE)
1703                 inet_csk_destroy_sock(sk);
1704         /* Otherwise, socket is reprieved until protocol close. */
1705
1706 out:
1707         bh_unlock_sock(sk);
1708         local_bh_enable();
1709         sock_put(sk);
1710 }
1711
1712 /* These states need RST on ABORT according to RFC793 */
1713
1714 static inline int tcp_need_reset(int state)
1715 {
1716         return (1 << state) &
1717                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1718                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1719 }
1720
1721 int tcp_disconnect(struct sock *sk, int flags)
1722 {
1723         struct inet_sock *inet = inet_sk(sk);
1724         struct inet_connection_sock *icsk = inet_csk(sk);
1725         struct tcp_sock *tp = tcp_sk(sk);
1726         int err = 0;
1727         int old_state = sk->sk_state;
1728
1729         if (old_state != TCP_CLOSE)
1730                 tcp_set_state(sk, TCP_CLOSE);
1731
1732         /* ABORT function of RFC793 */
1733         if (old_state == TCP_LISTEN) {
1734                 inet_csk_listen_stop(sk);
1735         } else if (tcp_need_reset(old_state) ||
1736                    (tp->snd_nxt != tp->write_seq &&
1737                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1738                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1739                  * states
1740                  */
1741                 tcp_send_active_reset(sk, gfp_any());
1742                 sk->sk_err = ECONNRESET;
1743         } else if (old_state == TCP_SYN_SENT)
1744                 sk->sk_err = ECONNRESET;
1745
1746         tcp_clear_xmit_timers(sk);
1747         __skb_queue_purge(&sk->sk_receive_queue);
1748         tcp_write_queue_purge(sk);
1749         __skb_queue_purge(&tp->out_of_order_queue);
1750 #ifdef CONFIG_NET_DMA
1751         __skb_queue_purge(&sk->sk_async_wait_queue);
1752 #endif
1753
1754         inet->dport = 0;
1755
1756         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1757                 inet_reset_saddr(sk);
1758
1759         sk->sk_shutdown = 0;
1760         sock_reset_flag(sk, SOCK_DONE);
1761         tp->srtt = 0;
1762         if ((tp->write_seq += tp->max_window + 2) == 0)
1763                 tp->write_seq = 1;
1764         icsk->icsk_backoff = 0;
1765         tp->snd_cwnd = 2;
1766         icsk->icsk_probes_out = 0;
1767         tp->packets_out = 0;
1768         tp->snd_ssthresh = 0x7fffffff;
1769         tp->snd_cwnd_cnt = 0;
1770         tp->bytes_acked = 0;
1771         tcp_set_ca_state(sk, TCP_CA_Open);
1772         tcp_clear_retrans(tp);
1773         inet_csk_delack_init(sk);
1774         tcp_init_send_head(sk);
1775         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1776         __sk_dst_reset(sk);
1777
1778         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1779
1780         sk->sk_error_report(sk);
1781         return err;
1782 }
1783
1784 /*
1785  *      Socket option code for TCP.
1786  */
1787 static int do_tcp_setsockopt(struct sock *sk, int level,
1788                 int optname, char __user *optval, int optlen)
1789 {
1790         struct tcp_sock *tp = tcp_sk(sk);
1791         struct inet_connection_sock *icsk = inet_csk(sk);
1792         int val;
1793         int err = 0;
1794
1795         /* This is a string value all the others are int's */
1796         if (optname == TCP_CONGESTION) {
1797                 char name[TCP_CA_NAME_MAX];
1798
1799                 if (optlen < 1)
1800                         return -EINVAL;
1801
1802                 val = strncpy_from_user(name, optval,
1803                                         min(TCP_CA_NAME_MAX-1, optlen));
1804                 if (val < 0)
1805                         return -EFAULT;
1806                 name[val] = 0;
1807
1808                 lock_sock(sk);
1809                 err = tcp_set_congestion_control(sk, name);
1810                 release_sock(sk);
1811                 return err;
1812         }
1813
1814         if (optlen < sizeof(int))
1815                 return -EINVAL;
1816
1817         if (get_user(val, (int __user *)optval))
1818                 return -EFAULT;
1819
1820         lock_sock(sk);
1821
1822         switch (optname) {
1823         case TCP_MAXSEG:
1824                 /* Values greater than interface MTU won't take effect. However
1825                  * at the point when this call is done we typically don't yet
1826                  * know which interface is going to be used */
1827                 if (val < 8 || val > MAX_TCP_WINDOW) {
1828                         err = -EINVAL;
1829                         break;
1830                 }
1831                 tp->rx_opt.user_mss = val;
1832                 break;
1833
1834         case TCP_NODELAY:
1835                 if (val) {
1836                         /* TCP_NODELAY is weaker than TCP_CORK, so that
1837                          * this option on corked socket is remembered, but
1838                          * it is not activated until cork is cleared.
1839                          *
1840                          * However, when TCP_NODELAY is set we make
1841                          * an explicit push, which overrides even TCP_CORK
1842                          * for currently queued segments.
1843                          */
1844                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1845                         tcp_push_pending_frames(sk);
1846                 } else {
1847                         tp->nonagle &= ~TCP_NAGLE_OFF;
1848                 }
1849                 break;
1850
1851         case TCP_CORK:
1852                 /* When set indicates to always queue non-full frames.
1853                  * Later the user clears this option and we transmit
1854                  * any pending partial frames in the queue.  This is
1855                  * meant to be used alongside sendfile() to get properly
1856                  * filled frames when the user (for example) must write
1857                  * out headers with a write() call first and then use
1858                  * sendfile to send out the data parts.
1859                  *
1860                  * TCP_CORK can be set together with TCP_NODELAY and it is
1861                  * stronger than TCP_NODELAY.
1862                  */
1863                 if (val) {
1864                         tp->nonagle |= TCP_NAGLE_CORK;
1865                 } else {
1866                         tp->nonagle &= ~TCP_NAGLE_CORK;
1867                         if (tp->nonagle&TCP_NAGLE_OFF)
1868                                 tp->nonagle |= TCP_NAGLE_PUSH;
1869                         tcp_push_pending_frames(sk);
1870                 }
1871                 break;
1872
1873         case TCP_KEEPIDLE:
1874                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1875                         err = -EINVAL;
1876                 else {
1877                         tp->keepalive_time = val * HZ;
1878                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
1879                             !((1 << sk->sk_state) &
1880                               (TCPF_CLOSE | TCPF_LISTEN))) {
1881                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1882                                 if (tp->keepalive_time > elapsed)
1883                                         elapsed = tp->keepalive_time - elapsed;
1884                                 else
1885                                         elapsed = 0;
1886                                 inet_csk_reset_keepalive_timer(sk, elapsed);
1887                         }
1888                 }
1889                 break;
1890         case TCP_KEEPINTVL:
1891                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1892                         err = -EINVAL;
1893                 else
1894                         tp->keepalive_intvl = val * HZ;
1895                 break;
1896         case TCP_KEEPCNT:
1897                 if (val < 1 || val > MAX_TCP_KEEPCNT)
1898                         err = -EINVAL;
1899                 else
1900                         tp->keepalive_probes = val;
1901                 break;
1902         case TCP_SYNCNT:
1903                 if (val < 1 || val > MAX_TCP_SYNCNT)
1904                         err = -EINVAL;
1905                 else
1906                         icsk->icsk_syn_retries = val;
1907                 break;
1908
1909         case TCP_LINGER2:
1910                 if (val < 0)
1911                         tp->linger2 = -1;
1912                 else if (val > sysctl_tcp_fin_timeout / HZ)
1913                         tp->linger2 = 0;
1914                 else
1915                         tp->linger2 = val * HZ;
1916                 break;
1917
1918         case TCP_DEFER_ACCEPT:
1919                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1920                 if (val > 0) {
1921                         /* Translate value in seconds to number of
1922                          * retransmits */
1923                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1924                                val > ((TCP_TIMEOUT_INIT / HZ) <<
1925                                        icsk->icsk_accept_queue.rskq_defer_accept))
1926                                 icsk->icsk_accept_queue.rskq_defer_accept++;
1927                         icsk->icsk_accept_queue.rskq_defer_accept++;
1928                 }
1929                 break;
1930
1931         case TCP_WINDOW_CLAMP:
1932                 if (!val) {
1933                         if (sk->sk_state != TCP_CLOSE) {
1934                                 err = -EINVAL;
1935                                 break;
1936                         }
1937                         tp->window_clamp = 0;
1938                 } else
1939                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1940                                                 SOCK_MIN_RCVBUF / 2 : val;
1941                 break;
1942
1943         case TCP_QUICKACK:
1944                 if (!val) {
1945                         icsk->icsk_ack.pingpong = 1;
1946                 } else {
1947                         icsk->icsk_ack.pingpong = 0;
1948                         if ((1 << sk->sk_state) &
1949                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1950                             inet_csk_ack_scheduled(sk)) {
1951                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1952                                 tcp_cleanup_rbuf(sk, 1);
1953                                 if (!(val & 1))
1954                                         icsk->icsk_ack.pingpong = 1;
1955                         }
1956                 }
1957                 break;
1958
1959 #ifdef CONFIG_TCP_MD5SIG
1960         case TCP_MD5SIG:
1961                 /* Read the IP->Key mappings from userspace */
1962                 err = tp->af_specific->md5_parse(sk, optval, optlen);
1963                 break;
1964 #endif
1965
1966         default:
1967                 err = -ENOPROTOOPT;
1968                 break;
1969         }
1970
1971         release_sock(sk);
1972         return err;
1973 }
1974
1975 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1976                    int optlen)
1977 {
1978         struct inet_connection_sock *icsk = inet_csk(sk);
1979
1980         if (level != SOL_TCP)
1981                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1982                                                      optval, optlen);
1983         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1984 }
1985
1986 #ifdef CONFIG_COMPAT
1987 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1988                           char __user *optval, int optlen)
1989 {
1990         if (level != SOL_TCP)
1991                 return inet_csk_compat_setsockopt(sk, level, optname,
1992                                                   optval, optlen);
1993         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1994 }
1995
1996 EXPORT_SYMBOL(compat_tcp_setsockopt);
1997 #endif
1998
1999 /* Return information about state of tcp endpoint in API format. */
2000 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2001 {
2002         struct tcp_sock *tp = tcp_sk(sk);
2003         const struct inet_connection_sock *icsk = inet_csk(sk);
2004         u32 now = tcp_time_stamp;
2005
2006         memset(info, 0, sizeof(*info));
2007
2008         info->tcpi_state = sk->sk_state;
2009         info->tcpi_ca_state = icsk->icsk_ca_state;
2010         info->tcpi_retransmits = icsk->icsk_retransmits;
2011         info->tcpi_probes = icsk->icsk_probes_out;
2012         info->tcpi_backoff = icsk->icsk_backoff;
2013
2014         if (tp->rx_opt.tstamp_ok)
2015                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2016         if (tp->rx_opt.sack_ok)
2017                 info->tcpi_options |= TCPI_OPT_SACK;
2018         if (tp->rx_opt.wscale_ok) {
2019                 info->tcpi_options |= TCPI_OPT_WSCALE;
2020                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2021                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2022         }
2023
2024         if (tp->ecn_flags&TCP_ECN_OK)
2025                 info->tcpi_options |= TCPI_OPT_ECN;
2026
2027         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2028         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2029         info->tcpi_snd_mss = tp->mss_cache;
2030         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2031
2032         info->tcpi_unacked = tp->packets_out;
2033         info->tcpi_sacked = tp->sacked_out;
2034         info->tcpi_lost = tp->lost_out;
2035         info->tcpi_retrans = tp->retrans_out;
2036         info->tcpi_fackets = tp->fackets_out;
2037
2038         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2039         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2040         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2041
2042         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2043         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2044         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2045         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2046         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2047         info->tcpi_snd_cwnd = tp->snd_cwnd;
2048         info->tcpi_advmss = tp->advmss;
2049         info->tcpi_reordering = tp->reordering;
2050
2051         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2052         info->tcpi_rcv_space = tp->rcvq_space.space;
2053
2054         info->tcpi_total_retrans = tp->total_retrans;
2055 }
2056
2057 EXPORT_SYMBOL_GPL(tcp_get_info);
2058
2059 static int do_tcp_getsockopt(struct sock *sk, int level,
2060                 int optname, char __user *optval, int __user *optlen)
2061 {
2062         struct inet_connection_sock *icsk = inet_csk(sk);
2063         struct tcp_sock *tp = tcp_sk(sk);
2064         int val, len;
2065
2066         if (get_user(len, optlen))
2067                 return -EFAULT;
2068
2069         len = min_t(unsigned int, len, sizeof(int));
2070
2071         if (len < 0)
2072                 return -EINVAL;
2073
2074         switch (optname) {
2075         case TCP_MAXSEG:
2076                 val = tp->mss_cache;
2077                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2078                         val = tp->rx_opt.user_mss;
2079                 break;
2080         case TCP_NODELAY:
2081                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2082                 break;
2083         case TCP_CORK:
2084                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2085                 break;
2086         case TCP_KEEPIDLE:
2087                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2088                 break;
2089         case TCP_KEEPINTVL:
2090                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2091                 break;
2092         case TCP_KEEPCNT:
2093                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2094                 break;
2095         case TCP_SYNCNT:
2096                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2097                 break;
2098         case TCP_LINGER2:
2099                 val = tp->linger2;
2100                 if (val >= 0)
2101                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2102                 break;
2103         case TCP_DEFER_ACCEPT:
2104                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2105                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2106                 break;
2107         case TCP_WINDOW_CLAMP:
2108                 val = tp->window_clamp;
2109                 break;
2110         case TCP_INFO: {
2111                 struct tcp_info info;
2112
2113                 if (get_user(len, optlen))
2114                         return -EFAULT;
2115
2116                 tcp_get_info(sk, &info);
2117
2118                 len = min_t(unsigned int, len, sizeof(info));
2119                 if (put_user(len, optlen))
2120                         return -EFAULT;
2121                 if (copy_to_user(optval, &info, len))
2122                         return -EFAULT;
2123                 return 0;
2124         }
2125         case TCP_QUICKACK:
2126                 val = !icsk->icsk_ack.pingpong;
2127                 break;
2128
2129         case TCP_CONGESTION:
2130                 if (get_user(len, optlen))
2131                         return -EFAULT;
2132                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2133                 if (put_user(len, optlen))
2134                         return -EFAULT;
2135                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2136                         return -EFAULT;
2137                 return 0;
2138         default:
2139                 return -ENOPROTOOPT;
2140         }
2141
2142         if (put_user(len, optlen))
2143                 return -EFAULT;
2144         if (copy_to_user(optval, &val, len))
2145                 return -EFAULT;
2146         return 0;
2147 }
2148
2149 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2150                    int __user *optlen)
2151 {
2152         struct inet_connection_sock *icsk = inet_csk(sk);
2153
2154         if (level != SOL_TCP)
2155                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2156                                                      optval, optlen);
2157         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2158 }
2159
2160 #ifdef CONFIG_COMPAT
2161 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2162                           char __user *optval, int __user *optlen)
2163 {
2164         if (level != SOL_TCP)
2165                 return inet_csk_compat_getsockopt(sk, level, optname,
2166                                                   optval, optlen);
2167         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2168 }
2169
2170 EXPORT_SYMBOL(compat_tcp_getsockopt);
2171 #endif
2172
2173 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2174 {
2175         struct sk_buff *segs = ERR_PTR(-EINVAL);
2176         struct tcphdr *th;
2177         unsigned thlen;
2178         unsigned int seq;
2179         __be32 delta;
2180         unsigned int oldlen;
2181         unsigned int len;
2182
2183         if (!pskb_may_pull(skb, sizeof(*th)))
2184                 goto out;
2185
2186         th = tcp_hdr(skb);
2187         thlen = th->doff * 4;
2188         if (thlen < sizeof(*th))
2189                 goto out;
2190
2191         if (!pskb_may_pull(skb, thlen))
2192                 goto out;
2193
2194         oldlen = (u16)~skb->len;
2195         __skb_pull(skb, thlen);
2196
2197         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2198                 /* Packet is from an untrusted source, reset gso_segs. */
2199                 int type = skb_shinfo(skb)->gso_type;
2200                 int mss;
2201
2202                 if (unlikely(type &
2203                              ~(SKB_GSO_TCPV4 |
2204                                SKB_GSO_DODGY |
2205                                SKB_GSO_TCP_ECN |
2206                                SKB_GSO_TCPV6 |
2207                                0) ||
2208                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2209                         goto out;
2210
2211                 mss = skb_shinfo(skb)->gso_size;
2212                 skb_shinfo(skb)->gso_segs = (skb->len + mss - 1) / mss;
2213
2214                 segs = NULL;
2215                 goto out;
2216         }
2217
2218         segs = skb_segment(skb, features);
2219         if (IS_ERR(segs))
2220                 goto out;
2221
2222         len = skb_shinfo(skb)->gso_size;
2223         delta = htonl(oldlen + (thlen + len));
2224
2225         skb = segs;
2226         th = tcp_hdr(skb);
2227         seq = ntohl(th->seq);
2228
2229         do {
2230                 th->fin = th->psh = 0;
2231
2232                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2233                                        (__force u32)delta));
2234                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2235                         th->check =
2236                              csum_fold(csum_partial(skb_transport_header(skb),
2237                                                     thlen, skb->csum));
2238
2239                 seq += len;
2240                 skb = skb->next;
2241                 th = tcp_hdr(skb);
2242
2243                 th->seq = htonl(seq);
2244                 th->cwr = 0;
2245         } while (skb->next);
2246
2247         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2248                       skb->data_len);
2249         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2250                                 (__force u32)delta));
2251         if (skb->ip_summed != CHECKSUM_PARTIAL)
2252                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2253                                                    thlen, skb->csum));
2254
2255 out:
2256         return segs;
2257 }
2258 EXPORT_SYMBOL(tcp_tso_segment);
2259
2260 #ifdef CONFIG_TCP_MD5SIG
2261 static unsigned long tcp_md5sig_users;
2262 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2263 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2264
2265 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2266 {
2267         int cpu;
2268         for_each_possible_cpu(cpu) {
2269                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2270                 if (p) {
2271                         if (p->md5_desc.tfm)
2272                                 crypto_free_hash(p->md5_desc.tfm);
2273                         kfree(p);
2274                         p = NULL;
2275                 }
2276         }
2277         free_percpu(pool);
2278 }
2279
2280 void tcp_free_md5sig_pool(void)
2281 {
2282         struct tcp_md5sig_pool **pool = NULL;
2283
2284         spin_lock_bh(&tcp_md5sig_pool_lock);
2285         if (--tcp_md5sig_users == 0) {
2286                 pool = tcp_md5sig_pool;
2287                 tcp_md5sig_pool = NULL;
2288         }
2289         spin_unlock_bh(&tcp_md5sig_pool_lock);
2290         if (pool)
2291                 __tcp_free_md5sig_pool(pool);
2292 }
2293
2294 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2295
2296 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2297 {
2298         int cpu;
2299         struct tcp_md5sig_pool **pool;
2300
2301         pool = alloc_percpu(struct tcp_md5sig_pool *);
2302         if (!pool)
2303                 return NULL;
2304
2305         for_each_possible_cpu(cpu) {
2306                 struct tcp_md5sig_pool *p;
2307                 struct crypto_hash *hash;
2308
2309                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2310                 if (!p)
2311                         goto out_free;
2312                 *per_cpu_ptr(pool, cpu) = p;
2313
2314                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2315                 if (!hash || IS_ERR(hash))
2316                         goto out_free;
2317
2318                 p->md5_desc.tfm = hash;
2319         }
2320         return pool;
2321 out_free:
2322         __tcp_free_md5sig_pool(pool);
2323         return NULL;
2324 }
2325
2326 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2327 {
2328         struct tcp_md5sig_pool **pool;
2329         int alloc = 0;
2330
2331 retry:
2332         spin_lock_bh(&tcp_md5sig_pool_lock);
2333         pool = tcp_md5sig_pool;
2334         if (tcp_md5sig_users++ == 0) {
2335                 alloc = 1;
2336                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2337         } else if (!pool) {
2338                 tcp_md5sig_users--;
2339                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2340                 cpu_relax();
2341                 goto retry;
2342         } else
2343                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2344
2345         if (alloc) {
2346                 /* we cannot hold spinlock here because this may sleep. */
2347                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2348                 spin_lock_bh(&tcp_md5sig_pool_lock);
2349                 if (!p) {
2350                         tcp_md5sig_users--;
2351                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2352                         return NULL;
2353                 }
2354                 pool = tcp_md5sig_pool;
2355                 if (pool) {
2356                         /* oops, it has already been assigned. */
2357                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2358                         __tcp_free_md5sig_pool(p);
2359                 } else {
2360                         tcp_md5sig_pool = pool = p;
2361                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2362                 }
2363         }
2364         return pool;
2365 }
2366
2367 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2368
2369 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2370 {
2371         struct tcp_md5sig_pool **p;
2372         spin_lock_bh(&tcp_md5sig_pool_lock);
2373         p = tcp_md5sig_pool;
2374         if (p)
2375                 tcp_md5sig_users++;
2376         spin_unlock_bh(&tcp_md5sig_pool_lock);
2377         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2378 }
2379
2380 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2381
2382 void __tcp_put_md5sig_pool(void)
2383 {
2384         tcp_free_md5sig_pool();
2385 }
2386
2387 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2388 #endif
2389
2390 void tcp_done(struct sock *sk)
2391 {
2392         if(sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2393                 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
2394
2395         tcp_set_state(sk, TCP_CLOSE);
2396         tcp_clear_xmit_timers(sk);
2397
2398         sk->sk_shutdown = SHUTDOWN_MASK;
2399
2400         if (!sock_flag(sk, SOCK_DEAD))
2401                 sk->sk_state_change(sk);
2402         else
2403                 inet_csk_destroy_sock(sk);
2404 }
2405 EXPORT_SYMBOL_GPL(tcp_done);
2406
2407 extern void __skb_cb_too_small_for_tcp(int, int);
2408 extern struct tcp_congestion_ops tcp_reno;
2409
2410 static __initdata unsigned long thash_entries;
2411 static int __init set_thash_entries(char *str)
2412 {
2413         if (!str)
2414                 return 0;
2415         thash_entries = simple_strtoul(str, &str, 0);
2416         return 1;
2417 }
2418 __setup("thash_entries=", set_thash_entries);
2419
2420 void __init tcp_init(void)
2421 {
2422         struct sk_buff *skb = NULL;
2423         unsigned long limit;
2424         int order, i, max_share;
2425
2426         if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2427                 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2428                                            sizeof(skb->cb));
2429
2430         tcp_hashinfo.bind_bucket_cachep =
2431                 kmem_cache_create("tcp_bind_bucket",
2432                                   sizeof(struct inet_bind_bucket), 0,
2433                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
2434
2435         /* Size and allocate the main established and bind bucket
2436          * hash tables.
2437          *
2438          * The methodology is similar to that of the buffer cache.
2439          */
2440         tcp_hashinfo.ehash =
2441                 alloc_large_system_hash("TCP established",
2442                                         sizeof(struct inet_ehash_bucket),
2443                                         thash_entries,
2444                                         (num_physpages >= 128 * 1024) ?
2445                                         13 : 15,
2446                                         0,
2447                                         &tcp_hashinfo.ehash_size,
2448                                         NULL,
2449                                         0);
2450         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2451         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2452                 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2453                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2454                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
2455         }
2456
2457         tcp_hashinfo.bhash =
2458                 alloc_large_system_hash("TCP bind",
2459                                         sizeof(struct inet_bind_hashbucket),
2460                                         tcp_hashinfo.ehash_size,
2461                                         (num_physpages >= 128 * 1024) ?
2462                                         13 : 15,
2463                                         0,
2464                                         &tcp_hashinfo.bhash_size,
2465                                         NULL,
2466                                         64 * 1024);
2467         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2468         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2469                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2470                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2471         }
2472
2473         /* Try to be a bit smarter and adjust defaults depending
2474          * on available memory.
2475          */
2476         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2477                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2478                         order++)
2479                 ;
2480         if (order >= 4) {
2481                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2482                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2483                 sysctl_max_syn_backlog = 1024;
2484         } else if (order < 3) {
2485                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2486                 sysctl_tcp_max_orphans >>= (3 - order);
2487                 sysctl_max_syn_backlog = 128;
2488         }
2489
2490         /* Set the pressure threshold to be a fraction of global memory that
2491          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2492          * memory, with a floor of 128 pages.
2493          */
2494         limit = min(nr_all_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2495         limit = (limit * (nr_all_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2496         limit = max(limit, 128UL);
2497         sysctl_tcp_mem[0] = limit / 4 * 3;
2498         sysctl_tcp_mem[1] = limit;
2499         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2500
2501         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2502         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2503         max_share = min(4UL*1024*1024, limit);
2504
2505         sysctl_tcp_wmem[0] = SK_STREAM_MEM_QUANTUM;
2506         sysctl_tcp_wmem[1] = 16*1024;
2507         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2508
2509         sysctl_tcp_rmem[0] = SK_STREAM_MEM_QUANTUM;
2510         sysctl_tcp_rmem[1] = 87380;
2511         sysctl_tcp_rmem[2] = max(87380, max_share);
2512
2513         printk(KERN_INFO "TCP: Hash tables configured "
2514                "(established %d bind %d)\n",
2515                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2516
2517         tcp_register_congestion_control(&tcp_reno);
2518 }
2519
2520 EXPORT_SYMBOL(tcp_close);
2521 EXPORT_SYMBOL(tcp_disconnect);
2522 EXPORT_SYMBOL(tcp_getsockopt);
2523 EXPORT_SYMBOL(tcp_ioctl);
2524 EXPORT_SYMBOL(tcp_poll);
2525 EXPORT_SYMBOL(tcp_read_sock);
2526 EXPORT_SYMBOL(tcp_recvmsg);
2527 EXPORT_SYMBOL(tcp_sendmsg);
2528 EXPORT_SYMBOL(tcp_sendpage);
2529 EXPORT_SYMBOL(tcp_setsockopt);
2530 EXPORT_SYMBOL(tcp_shutdown);
2531 EXPORT_SYMBOL(tcp_statistics);