Merge davem@master.kernel.org:/pub/scm/linux/kernel/git/vxy/lksctp-dev
[linux-2.6] / 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 socket *sock, struct msghdr *msg,
662                 size_t size)
663 {
664         struct sock *sk = sock->sk;
665         struct iovec *iov;
666         struct tcp_sock *tp = tcp_sk(sk);
667         struct sk_buff *skb;
668         int iovlen, flags;
669         int mss_now, size_goal;
670         int err, copied;
671         long timeo;
672
673         lock_sock(sk);
674         TCP_CHECK_TIMER(sk);
675
676         flags = msg->msg_flags;
677         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
678
679         /* Wait for a connection to finish. */
680         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
681                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
682                         goto out_err;
683
684         /* This should be in poll */
685         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
686
687         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
688         size_goal = tp->xmit_size_goal;
689
690         /* Ok commence sending. */
691         iovlen = msg->msg_iovlen;
692         iov = msg->msg_iov;
693         copied = 0;
694
695         err = -EPIPE;
696         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
697                 goto do_error;
698
699         while (--iovlen >= 0) {
700                 int seglen = iov->iov_len;
701                 unsigned char __user *from = iov->iov_base;
702
703                 iov++;
704
705                 while (seglen > 0) {
706                         int copy;
707
708                         skb = tcp_write_queue_tail(sk);
709
710                         if (!tcp_send_head(sk) ||
711                             (copy = size_goal - skb->len) <= 0) {
712
713 new_segment:
714                                 /* Allocate new segment. If the interface is SG,
715                                  * allocate skb fitting to single page.
716                                  */
717                                 if (!sk_stream_memory_free(sk))
718                                         goto wait_for_sndbuf;
719
720                                 skb = sk_stream_alloc_pskb(sk, select_size(sk),
721                                                            0, sk->sk_allocation);
722                                 if (!skb)
723                                         goto wait_for_memory;
724
725                                 /*
726                                  * Check whether we can use HW checksum.
727                                  */
728                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
729                                         skb->ip_summed = CHECKSUM_PARTIAL;
730
731                                 skb_entail(sk, skb);
732                                 copy = size_goal;
733                         }
734
735                         /* Try to append data to the end of skb. */
736                         if (copy > seglen)
737                                 copy = seglen;
738
739                         /* Where to copy to? */
740                         if (skb_tailroom(skb) > 0) {
741                                 /* We have some space in skb head. Superb! */
742                                 if (copy > skb_tailroom(skb))
743                                         copy = skb_tailroom(skb);
744                                 if ((err = skb_add_data(skb, from, copy)) != 0)
745                                         goto do_fault;
746                         } else {
747                                 int merge = 0;
748                                 int i = skb_shinfo(skb)->nr_frags;
749                                 struct page *page = TCP_PAGE(sk);
750                                 int off = TCP_OFF(sk);
751
752                                 if (skb_can_coalesce(skb, i, page, off) &&
753                                     off != PAGE_SIZE) {
754                                         /* We can extend the last page
755                                          * fragment. */
756                                         merge = 1;
757                                 } else if (i == MAX_SKB_FRAGS ||
758                                            (!i &&
759                                            !(sk->sk_route_caps & NETIF_F_SG))) {
760                                         /* Need to add new fragment and cannot
761                                          * do this because interface is non-SG,
762                                          * or because all the page slots are
763                                          * busy. */
764                                         tcp_mark_push(tp, skb);
765                                         goto new_segment;
766                                 } else if (page) {
767                                         if (off == PAGE_SIZE) {
768                                                 put_page(page);
769                                                 TCP_PAGE(sk) = page = NULL;
770                                                 off = 0;
771                                         }
772                                 } else
773                                         off = 0;
774
775                                 if (copy > PAGE_SIZE - off)
776                                         copy = PAGE_SIZE - off;
777
778                                 if (!sk_stream_wmem_schedule(sk, copy))
779                                         goto wait_for_memory;
780
781                                 if (!page) {
782                                         /* Allocate new cache page. */
783                                         if (!(page = sk_stream_alloc_page(sk)))
784                                                 goto wait_for_memory;
785                                 }
786
787                                 /* Time to copy data. We are close to
788                                  * the end! */
789                                 err = skb_copy_to_page(sk, from, skb, page,
790                                                        off, copy);
791                                 if (err) {
792                                         /* If this page was new, give it to the
793                                          * socket so it does not get leaked.
794                                          */
795                                         if (!TCP_PAGE(sk)) {
796                                                 TCP_PAGE(sk) = page;
797                                                 TCP_OFF(sk) = 0;
798                                         }
799                                         goto do_error;
800                                 }
801
802                                 /* Update the skb. */
803                                 if (merge) {
804                                         skb_shinfo(skb)->frags[i - 1].size +=
805                                                                         copy;
806                                 } else {
807                                         skb_fill_page_desc(skb, i, page, off, copy);
808                                         if (TCP_PAGE(sk)) {
809                                                 get_page(page);
810                                         } else if (off + copy < PAGE_SIZE) {
811                                                 get_page(page);
812                                                 TCP_PAGE(sk) = page;
813                                         }
814                                 }
815
816                                 TCP_OFF(sk) = off + copy;
817                         }
818
819                         if (!copied)
820                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
821
822                         tp->write_seq += copy;
823                         TCP_SKB_CB(skb)->end_seq += copy;
824                         skb_shinfo(skb)->gso_segs = 0;
825
826                         from += copy;
827                         copied += copy;
828                         if ((seglen -= copy) == 0 && iovlen == 0)
829                                 goto out;
830
831                         if (skb->len < mss_now || (flags & MSG_OOB))
832                                 continue;
833
834                         if (forced_push(tp)) {
835                                 tcp_mark_push(tp, skb);
836                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
837                         } else if (skb == tcp_send_head(sk))
838                                 tcp_push_one(sk, mss_now);
839                         continue;
840
841 wait_for_sndbuf:
842                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
843 wait_for_memory:
844                         if (copied)
845                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
846
847                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
848                                 goto do_error;
849
850                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
851                         size_goal = tp->xmit_size_goal;
852                 }
853         }
854
855 out:
856         if (copied)
857                 tcp_push(sk, flags, mss_now, tp->nonagle);
858         TCP_CHECK_TIMER(sk);
859         release_sock(sk);
860         return copied;
861
862 do_fault:
863         if (!skb->len) {
864                 tcp_unlink_write_queue(skb, sk);
865                 /* It is the one place in all of TCP, except connection
866                  * reset, where we can be unlinking the send_head.
867                  */
868                 tcp_check_send_head(sk, skb);
869                 sk_stream_free_skb(sk, skb);
870         }
871
872 do_error:
873         if (copied)
874                 goto out;
875 out_err:
876         err = sk_stream_error(sk, flags, err);
877         TCP_CHECK_TIMER(sk);
878         release_sock(sk);
879         return err;
880 }
881
882 /*
883  *      Handle reading urgent data. BSD has very simple semantics for
884  *      this, no blocking and very strange errors 8)
885  */
886
887 static int tcp_recv_urg(struct sock *sk, long timeo,
888                         struct msghdr *msg, int len, int flags,
889                         int *addr_len)
890 {
891         struct tcp_sock *tp = tcp_sk(sk);
892
893         /* No URG data to read. */
894         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
895             tp->urg_data == TCP_URG_READ)
896                 return -EINVAL; /* Yes this is right ! */
897
898         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
899                 return -ENOTCONN;
900
901         if (tp->urg_data & TCP_URG_VALID) {
902                 int err = 0;
903                 char c = tp->urg_data;
904
905                 if (!(flags & MSG_PEEK))
906                         tp->urg_data = TCP_URG_READ;
907
908                 /* Read urgent data. */
909                 msg->msg_flags |= MSG_OOB;
910
911                 if (len > 0) {
912                         if (!(flags & MSG_TRUNC))
913                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
914                         len = 1;
915                 } else
916                         msg->msg_flags |= MSG_TRUNC;
917
918                 return err ? -EFAULT : len;
919         }
920
921         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
922                 return 0;
923
924         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
925          * the available implementations agree in this case:
926          * this call should never block, independent of the
927          * blocking state of the socket.
928          * Mike <pall@rz.uni-karlsruhe.de>
929          */
930         return -EAGAIN;
931 }
932
933 /* Clean up the receive buffer for full frames taken by the user,
934  * then send an ACK if necessary.  COPIED is the number of bytes
935  * tcp_recvmsg has given to the user so far, it speeds up the
936  * calculation of whether or not we must ACK for the sake of
937  * a window update.
938  */
939 void tcp_cleanup_rbuf(struct sock *sk, int copied)
940 {
941         struct tcp_sock *tp = tcp_sk(sk);
942         int time_to_ack = 0;
943
944 #if TCP_DEBUG
945         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
946
947         BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
948 #endif
949
950         if (inet_csk_ack_scheduled(sk)) {
951                 const struct inet_connection_sock *icsk = inet_csk(sk);
952                    /* Delayed ACKs frequently hit locked sockets during bulk
953                     * receive. */
954                 if (icsk->icsk_ack.blocked ||
955                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
956                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
957                     /*
958                      * If this read emptied read buffer, we send ACK, if
959                      * connection is not bidirectional, user drained
960                      * receive buffer and there was a small segment
961                      * in queue.
962                      */
963                     (copied > 0 &&
964                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
965                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
966                        !icsk->icsk_ack.pingpong)) &&
967                       !atomic_read(&sk->sk_rmem_alloc)))
968                         time_to_ack = 1;
969         }
970
971         /* We send an ACK if we can now advertise a non-zero window
972          * which has been raised "significantly".
973          *
974          * Even if window raised up to infinity, do not send window open ACK
975          * in states, where we will not receive more. It is useless.
976          */
977         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
978                 __u32 rcv_window_now = tcp_receive_window(tp);
979
980                 /* Optimize, __tcp_select_window() is not cheap. */
981                 if (2*rcv_window_now <= tp->window_clamp) {
982                         __u32 new_window = __tcp_select_window(sk);
983
984                         /* Send ACK now, if this read freed lots of space
985                          * in our buffer. Certainly, new_window is new window.
986                          * We can advertise it now, if it is not less than current one.
987                          * "Lots" means "at least twice" here.
988                          */
989                         if (new_window && new_window >= 2 * rcv_window_now)
990                                 time_to_ack = 1;
991                 }
992         }
993         if (time_to_ack)
994                 tcp_send_ack(sk);
995 }
996
997 static void tcp_prequeue_process(struct sock *sk)
998 {
999         struct sk_buff *skb;
1000         struct tcp_sock *tp = tcp_sk(sk);
1001
1002         NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
1003
1004         /* RX process wants to run with disabled BHs, though it is not
1005          * necessary */
1006         local_bh_disable();
1007         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1008                 sk->sk_backlog_rcv(sk, skb);
1009         local_bh_enable();
1010
1011         /* Clear memory counter. */
1012         tp->ucopy.memory = 0;
1013 }
1014
1015 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1016 {
1017         struct sk_buff *skb;
1018         u32 offset;
1019
1020         skb_queue_walk(&sk->sk_receive_queue, skb) {
1021                 offset = seq - TCP_SKB_CB(skb)->seq;
1022                 if (tcp_hdr(skb)->syn)
1023                         offset--;
1024                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1025                         *off = offset;
1026                         return skb;
1027                 }
1028         }
1029         return NULL;
1030 }
1031
1032 /*
1033  * This routine provides an alternative to tcp_recvmsg() for routines
1034  * that would like to handle copying from skbuffs directly in 'sendfile'
1035  * fashion.
1036  * Note:
1037  *      - It is assumed that the socket was locked by the caller.
1038  *      - The routine does not block.
1039  *      - At present, there is no support for reading OOB data
1040  *        or for 'peeking' the socket using this routine
1041  *        (although both would be easy to implement).
1042  */
1043 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1044                   sk_read_actor_t recv_actor)
1045 {
1046         struct sk_buff *skb;
1047         struct tcp_sock *tp = tcp_sk(sk);
1048         u32 seq = tp->copied_seq;
1049         u32 offset;
1050         int copied = 0;
1051
1052         if (sk->sk_state == TCP_LISTEN)
1053                 return -ENOTCONN;
1054         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1055                 if (offset < skb->len) {
1056                         size_t used, len;
1057
1058                         len = skb->len - offset;
1059                         /* Stop reading if we hit a patch of urgent data */
1060                         if (tp->urg_data) {
1061                                 u32 urg_offset = tp->urg_seq - seq;
1062                                 if (urg_offset < len)
1063                                         len = urg_offset;
1064                                 if (!len)
1065                                         break;
1066                         }
1067                         used = recv_actor(desc, skb, offset, len);
1068                         if (used < 0) {
1069                                 if (!copied)
1070                                         copied = used;
1071                                 break;
1072                         } else if (used <= len) {
1073                                 seq += used;
1074                                 copied += used;
1075                                 offset += used;
1076                         }
1077                         if (offset != skb->len)
1078                                 break;
1079                 }
1080                 if (tcp_hdr(skb)->fin) {
1081                         sk_eat_skb(sk, skb, 0);
1082                         ++seq;
1083                         break;
1084                 }
1085                 sk_eat_skb(sk, skb, 0);
1086                 if (!desc->count)
1087                         break;
1088         }
1089         tp->copied_seq = seq;
1090
1091         tcp_rcv_space_adjust(sk);
1092
1093         /* Clean up data we have read: This will do ACK frames. */
1094         if (copied > 0)
1095                 tcp_cleanup_rbuf(sk, copied);
1096         return copied;
1097 }
1098
1099 /*
1100  *      This routine copies from a sock struct into the user buffer.
1101  *
1102  *      Technical note: in 2.3 we work on _locked_ socket, so that
1103  *      tricks with *seq access order and skb->users are not required.
1104  *      Probably, code can be easily improved even more.
1105  */
1106
1107 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1108                 size_t len, int nonblock, int flags, int *addr_len)
1109 {
1110         struct tcp_sock *tp = tcp_sk(sk);
1111         int copied = 0;
1112         u32 peek_seq;
1113         u32 *seq;
1114         unsigned long used;
1115         int err;
1116         int target;             /* Read at least this many bytes */
1117         long timeo;
1118         struct task_struct *user_recv = NULL;
1119         int copied_early = 0;
1120         struct sk_buff *skb;
1121
1122         lock_sock(sk);
1123
1124         TCP_CHECK_TIMER(sk);
1125
1126         err = -ENOTCONN;
1127         if (sk->sk_state == TCP_LISTEN)
1128                 goto out;
1129
1130         timeo = sock_rcvtimeo(sk, nonblock);
1131
1132         /* Urgent data needs to be handled specially. */
1133         if (flags & MSG_OOB)
1134                 goto recv_urg;
1135
1136         seq = &tp->copied_seq;
1137         if (flags & MSG_PEEK) {
1138                 peek_seq = tp->copied_seq;
1139                 seq = &peek_seq;
1140         }
1141
1142         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1143
1144 #ifdef CONFIG_NET_DMA
1145         tp->ucopy.dma_chan = NULL;
1146         preempt_disable();
1147         skb = skb_peek_tail(&sk->sk_receive_queue);
1148         {
1149                 int available = 0;
1150
1151                 if (skb)
1152                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1153                 if ((available < target) &&
1154                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1155                     !sysctl_tcp_low_latency &&
1156                     __get_cpu_var(softnet_data).net_dma) {
1157                         preempt_enable_no_resched();
1158                         tp->ucopy.pinned_list =
1159                                         dma_pin_iovec_pages(msg->msg_iov, len);
1160                 } else {
1161                         preempt_enable_no_resched();
1162                 }
1163         }
1164 #endif
1165
1166         do {
1167                 u32 offset;
1168
1169                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1170                 if (tp->urg_data && tp->urg_seq == *seq) {
1171                         if (copied)
1172                                 break;
1173                         if (signal_pending(current)) {
1174                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1175                                 break;
1176                         }
1177                 }
1178
1179                 /* Next get a buffer. */
1180
1181                 skb = skb_peek(&sk->sk_receive_queue);
1182                 do {
1183                         if (!skb)
1184                                 break;
1185
1186                         /* Now that we have two receive queues this
1187                          * shouldn't happen.
1188                          */
1189                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1190                                 printk(KERN_INFO "recvmsg bug: copied %X "
1191                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1192                                 break;
1193                         }
1194                         offset = *seq - TCP_SKB_CB(skb)->seq;
1195                         if (tcp_hdr(skb)->syn)
1196                                 offset--;
1197                         if (offset < skb->len)
1198                                 goto found_ok_skb;
1199                         if (tcp_hdr(skb)->fin)
1200                                 goto found_fin_ok;
1201                         BUG_TRAP(flags & MSG_PEEK);
1202                         skb = skb->next;
1203                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1204
1205                 /* Well, if we have backlog, try to process it now yet. */
1206
1207                 if (copied >= target && !sk->sk_backlog.tail)
1208                         break;
1209
1210                 if (copied) {
1211                         if (sk->sk_err ||
1212                             sk->sk_state == TCP_CLOSE ||
1213                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1214                             !timeo ||
1215                             signal_pending(current) ||
1216                             (flags & MSG_PEEK))
1217                                 break;
1218                 } else {
1219                         if (sock_flag(sk, SOCK_DONE))
1220                                 break;
1221
1222                         if (sk->sk_err) {
1223                                 copied = sock_error(sk);
1224                                 break;
1225                         }
1226
1227                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1228                                 break;
1229
1230                         if (sk->sk_state == TCP_CLOSE) {
1231                                 if (!sock_flag(sk, SOCK_DONE)) {
1232                                         /* This occurs when user tries to read
1233                                          * from never connected socket.
1234                                          */
1235                                         copied = -ENOTCONN;
1236                                         break;
1237                                 }
1238                                 break;
1239                         }
1240
1241                         if (!timeo) {
1242                                 copied = -EAGAIN;
1243                                 break;
1244                         }
1245
1246                         if (signal_pending(current)) {
1247                                 copied = sock_intr_errno(timeo);
1248                                 break;
1249                         }
1250                 }
1251
1252                 tcp_cleanup_rbuf(sk, copied);
1253
1254                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1255                         /* Install new reader */
1256                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1257                                 user_recv = current;
1258                                 tp->ucopy.task = user_recv;
1259                                 tp->ucopy.iov = msg->msg_iov;
1260                         }
1261
1262                         tp->ucopy.len = len;
1263
1264                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1265                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1266
1267                         /* Ugly... If prequeue is not empty, we have to
1268                          * process it before releasing socket, otherwise
1269                          * order will be broken at second iteration.
1270                          * More elegant solution is required!!!
1271                          *
1272                          * Look: we have the following (pseudo)queues:
1273                          *
1274                          * 1. packets in flight
1275                          * 2. backlog
1276                          * 3. prequeue
1277                          * 4. receive_queue
1278                          *
1279                          * Each queue can be processed only if the next ones
1280                          * are empty. At this point we have empty receive_queue.
1281                          * But prequeue _can_ be not empty after 2nd iteration,
1282                          * when we jumped to start of loop because backlog
1283                          * processing added something to receive_queue.
1284                          * We cannot release_sock(), because backlog contains
1285                          * packets arrived _after_ prequeued ones.
1286                          *
1287                          * Shortly, algorithm is clear --- to process all
1288                          * the queues in order. We could make it more directly,
1289                          * requeueing packets from backlog to prequeue, if
1290                          * is not empty. It is more elegant, but eats cycles,
1291                          * unfortunately.
1292                          */
1293                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1294                                 goto do_prequeue;
1295
1296                         /* __ Set realtime policy in scheduler __ */
1297                 }
1298
1299                 if (copied >= target) {
1300                         /* Do not sleep, just process backlog. */
1301                         release_sock(sk);
1302                         lock_sock(sk);
1303                 } else
1304                         sk_wait_data(sk, &timeo);
1305
1306 #ifdef CONFIG_NET_DMA
1307                 tp->ucopy.wakeup = 0;
1308 #endif
1309
1310                 if (user_recv) {
1311                         int chunk;
1312
1313                         /* __ Restore normal policy in scheduler __ */
1314
1315                         if ((chunk = len - tp->ucopy.len) != 0) {
1316                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1317                                 len -= chunk;
1318                                 copied += chunk;
1319                         }
1320
1321                         if (tp->rcv_nxt == tp->copied_seq &&
1322                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1323 do_prequeue:
1324                                 tcp_prequeue_process(sk);
1325
1326                                 if ((chunk = len - tp->ucopy.len) != 0) {
1327                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1328                                         len -= chunk;
1329                                         copied += chunk;
1330                                 }
1331                         }
1332                 }
1333                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1334                         if (net_ratelimit())
1335                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1336                                        current->comm, current->pid);
1337                         peek_seq = tp->copied_seq;
1338                 }
1339                 continue;
1340
1341         found_ok_skb:
1342                 /* Ok so how much can we use? */
1343                 used = skb->len - offset;
1344                 if (len < used)
1345                         used = len;
1346
1347                 /* Do we have urgent data here? */
1348                 if (tp->urg_data) {
1349                         u32 urg_offset = tp->urg_seq - *seq;
1350                         if (urg_offset < used) {
1351                                 if (!urg_offset) {
1352                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1353                                                 ++*seq;
1354                                                 offset++;
1355                                                 used--;
1356                                                 if (!used)
1357                                                         goto skip_copy;
1358                                         }
1359                                 } else
1360                                         used = urg_offset;
1361                         }
1362                 }
1363
1364                 if (!(flags & MSG_TRUNC)) {
1365 #ifdef CONFIG_NET_DMA
1366                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1367                                 tp->ucopy.dma_chan = get_softnet_dma();
1368
1369                         if (tp->ucopy.dma_chan) {
1370                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1371                                         tp->ucopy.dma_chan, skb, offset,
1372                                         msg->msg_iov, used,
1373                                         tp->ucopy.pinned_list);
1374
1375                                 if (tp->ucopy.dma_cookie < 0) {
1376
1377                                         printk(KERN_ALERT "dma_cookie < 0\n");
1378
1379                                         /* Exception. Bailout! */
1380                                         if (!copied)
1381                                                 copied = -EFAULT;
1382                                         break;
1383                                 }
1384                                 if ((offset + used) == skb->len)
1385                                         copied_early = 1;
1386
1387                         } else
1388 #endif
1389                         {
1390                                 err = skb_copy_datagram_iovec(skb, offset,
1391                                                 msg->msg_iov, used);
1392                                 if (err) {
1393                                         /* Exception. Bailout! */
1394                                         if (!copied)
1395                                                 copied = -EFAULT;
1396                                         break;
1397                                 }
1398                         }
1399                 }
1400
1401                 *seq += used;
1402                 copied += used;
1403                 len -= used;
1404
1405                 tcp_rcv_space_adjust(sk);
1406
1407 skip_copy:
1408                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1409                         tp->urg_data = 0;
1410                         tcp_fast_path_check(sk);
1411                 }
1412                 if (used + offset < skb->len)
1413                         continue;
1414
1415                 if (tcp_hdr(skb)->fin)
1416                         goto found_fin_ok;
1417                 if (!(flags & MSG_PEEK)) {
1418                         sk_eat_skb(sk, skb, copied_early);
1419                         copied_early = 0;
1420                 }
1421                 continue;
1422
1423         found_fin_ok:
1424                 /* Process the FIN. */
1425                 ++*seq;
1426                 if (!(flags & MSG_PEEK)) {
1427                         sk_eat_skb(sk, skb, copied_early);
1428                         copied_early = 0;
1429                 }
1430                 break;
1431         } while (len > 0);
1432
1433         if (user_recv) {
1434                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1435                         int chunk;
1436
1437                         tp->ucopy.len = copied > 0 ? len : 0;
1438
1439                         tcp_prequeue_process(sk);
1440
1441                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1442                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1443                                 len -= chunk;
1444                                 copied += chunk;
1445                         }
1446                 }
1447
1448                 tp->ucopy.task = NULL;
1449                 tp->ucopy.len = 0;
1450         }
1451
1452 #ifdef CONFIG_NET_DMA
1453         if (tp->ucopy.dma_chan) {
1454                 dma_cookie_t done, used;
1455
1456                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1457
1458                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1459                                                  tp->ucopy.dma_cookie, &done,
1460                                                  &used) == DMA_IN_PROGRESS) {
1461                         /* do partial cleanup of sk_async_wait_queue */
1462                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1463                                (dma_async_is_complete(skb->dma_cookie, done,
1464                                                       used) == DMA_SUCCESS)) {
1465                                 __skb_dequeue(&sk->sk_async_wait_queue);
1466                                 kfree_skb(skb);
1467                         }
1468                 }
1469
1470                 /* Safe to free early-copied skbs now */
1471                 __skb_queue_purge(&sk->sk_async_wait_queue);
1472                 dma_chan_put(tp->ucopy.dma_chan);
1473                 tp->ucopy.dma_chan = NULL;
1474         }
1475         if (tp->ucopy.pinned_list) {
1476                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1477                 tp->ucopy.pinned_list = NULL;
1478         }
1479 #endif
1480
1481         /* According to UNIX98, msg_name/msg_namelen are ignored
1482          * on connected socket. I was just happy when found this 8) --ANK
1483          */
1484
1485         /* Clean up data we have read: This will do ACK frames. */
1486         tcp_cleanup_rbuf(sk, copied);
1487
1488         TCP_CHECK_TIMER(sk);
1489         release_sock(sk);
1490         return copied;
1491
1492 out:
1493         TCP_CHECK_TIMER(sk);
1494         release_sock(sk);
1495         return err;
1496
1497 recv_urg:
1498         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1499         goto out;
1500 }
1501
1502 /*
1503  *      State processing on a close. This implements the state shift for
1504  *      sending our FIN frame. Note that we only send a FIN for some
1505  *      states. A shutdown() may have already sent the FIN, or we may be
1506  *      closed.
1507  */
1508
1509 static const unsigned char new_state[16] = {
1510   /* current state:        new state:      action:      */
1511   /* (Invalid)          */ TCP_CLOSE,
1512   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1513   /* TCP_SYN_SENT       */ TCP_CLOSE,
1514   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1515   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1516   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1517   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1518   /* TCP_CLOSE          */ TCP_CLOSE,
1519   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1520   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1521   /* TCP_LISTEN         */ TCP_CLOSE,
1522   /* TCP_CLOSING        */ TCP_CLOSING,
1523 };
1524
1525 static int tcp_close_state(struct sock *sk)
1526 {
1527         int next = (int)new_state[sk->sk_state];
1528         int ns = next & TCP_STATE_MASK;
1529
1530         tcp_set_state(sk, ns);
1531
1532         return next & TCP_ACTION_FIN;
1533 }
1534
1535 /*
1536  *      Shutdown the sending side of a connection. Much like close except
1537  *      that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1538  */
1539
1540 void tcp_shutdown(struct sock *sk, int how)
1541 {
1542         /*      We need to grab some memory, and put together a FIN,
1543          *      and then put it into the queue to be sent.
1544          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1545          */
1546         if (!(how & SEND_SHUTDOWN))
1547                 return;
1548
1549         /* If we've already sent a FIN, or it's a closed state, skip this. */
1550         if ((1 << sk->sk_state) &
1551             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1552              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1553                 /* Clear out any half completed packets.  FIN if needed. */
1554                 if (tcp_close_state(sk))
1555                         tcp_send_fin(sk);
1556         }
1557 }
1558
1559 void tcp_close(struct sock *sk, long timeout)
1560 {
1561         struct sk_buff *skb;
1562         int data_was_unread = 0;
1563         int state;
1564
1565         lock_sock(sk);
1566         sk->sk_shutdown = SHUTDOWN_MASK;
1567
1568         if (sk->sk_state == TCP_LISTEN) {
1569                 tcp_set_state(sk, TCP_CLOSE);
1570
1571                 /* Special case. */
1572                 inet_csk_listen_stop(sk);
1573
1574                 goto adjudge_to_death;
1575         }
1576
1577         /*  We need to flush the recv. buffs.  We do this only on the
1578          *  descriptor close, not protocol-sourced closes, because the
1579          *  reader process may not have drained the data yet!
1580          */
1581         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1582                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1583                           tcp_hdr(skb)->fin;
1584                 data_was_unread += len;
1585                 __kfree_skb(skb);
1586         }
1587
1588         sk_stream_mem_reclaim(sk);
1589
1590         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1591          * data was lost. To witness the awful effects of the old behavior of
1592          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1593          * GET in an FTP client, suspend the process, wait for the client to
1594          * advertise a zero window, then kill -9 the FTP client, wheee...
1595          * Note: timeout is always zero in such a case.
1596          */
1597         if (data_was_unread) {
1598                 /* Unread data was tossed, zap the connection. */
1599                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1600                 tcp_set_state(sk, TCP_CLOSE);
1601                 tcp_send_active_reset(sk, GFP_KERNEL);
1602         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1603                 /* Check zero linger _after_ checking for unread data. */
1604                 sk->sk_prot->disconnect(sk, 0);
1605                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1606         } else if (tcp_close_state(sk)) {
1607                 /* We FIN if the application ate all the data before
1608                  * zapping the connection.
1609                  */
1610
1611                 /* RED-PEN. Formally speaking, we have broken TCP state
1612                  * machine. State transitions:
1613                  *
1614                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1615                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1616                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1617                  *
1618                  * are legal only when FIN has been sent (i.e. in window),
1619                  * rather than queued out of window. Purists blame.
1620                  *
1621                  * F.e. "RFC state" is ESTABLISHED,
1622                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1623                  *
1624                  * The visible declinations are that sometimes
1625                  * we enter time-wait state, when it is not required really
1626                  * (harmless), do not send active resets, when they are
1627                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1628                  * they look as CLOSING or LAST_ACK for Linux)
1629                  * Probably, I missed some more holelets.
1630                  *                                              --ANK
1631                  */
1632                 tcp_send_fin(sk);
1633         }
1634
1635         sk_stream_wait_close(sk, timeout);
1636
1637 adjudge_to_death:
1638         state = sk->sk_state;
1639         sock_hold(sk);
1640         sock_orphan(sk);
1641         atomic_inc(sk->sk_prot->orphan_count);
1642
1643         /* It is the last release_sock in its life. It will remove backlog. */
1644         release_sock(sk);
1645
1646
1647         /* Now socket is owned by kernel and we acquire BH lock
1648            to finish close. No need to check for user refs.
1649          */
1650         local_bh_disable();
1651         bh_lock_sock(sk);
1652         BUG_TRAP(!sock_owned_by_user(sk));
1653
1654         /* Have we already been destroyed by a softirq or backlog? */
1655         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1656                 goto out;
1657
1658         /*      This is a (useful) BSD violating of the RFC. There is a
1659          *      problem with TCP as specified in that the other end could
1660          *      keep a socket open forever with no application left this end.
1661          *      We use a 3 minute timeout (about the same as BSD) then kill
1662          *      our end. If they send after that then tough - BUT: long enough
1663          *      that we won't make the old 4*rto = almost no time - whoops
1664          *      reset mistake.
1665          *
1666          *      Nope, it was not mistake. It is really desired behaviour
1667          *      f.e. on http servers, when such sockets are useless, but
1668          *      consume significant resources. Let's do it with special
1669          *      linger2 option.                                 --ANK
1670          */
1671
1672         if (sk->sk_state == TCP_FIN_WAIT2) {
1673                 struct tcp_sock *tp = tcp_sk(sk);
1674                 if (tp->linger2 < 0) {
1675                         tcp_set_state(sk, TCP_CLOSE);
1676                         tcp_send_active_reset(sk, GFP_ATOMIC);
1677                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1678                 } else {
1679                         const int tmo = tcp_fin_time(sk);
1680
1681                         if (tmo > TCP_TIMEWAIT_LEN) {
1682                                 inet_csk_reset_keepalive_timer(sk,
1683                                                 tmo - TCP_TIMEWAIT_LEN);
1684                         } else {
1685                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1686                                 goto out;
1687                         }
1688                 }
1689         }
1690         if (sk->sk_state != TCP_CLOSE) {
1691                 sk_stream_mem_reclaim(sk);
1692                 if (tcp_too_many_orphans(sk,
1693                                 atomic_read(sk->sk_prot->orphan_count))) {
1694                         if (net_ratelimit())
1695                                 printk(KERN_INFO "TCP: too many of orphaned "
1696                                        "sockets\n");
1697                         tcp_set_state(sk, TCP_CLOSE);
1698                         tcp_send_active_reset(sk, GFP_ATOMIC);
1699                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1700                 }
1701         }
1702
1703         if (sk->sk_state == TCP_CLOSE)
1704                 inet_csk_destroy_sock(sk);
1705         /* Otherwise, socket is reprieved until protocol close. */
1706
1707 out:
1708         bh_unlock_sock(sk);
1709         local_bh_enable();
1710         sock_put(sk);
1711 }
1712
1713 /* These states need RST on ABORT according to RFC793 */
1714
1715 static inline int tcp_need_reset(int state)
1716 {
1717         return (1 << state) &
1718                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1719                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1720 }
1721
1722 int tcp_disconnect(struct sock *sk, int flags)
1723 {
1724         struct inet_sock *inet = inet_sk(sk);
1725         struct inet_connection_sock *icsk = inet_csk(sk);
1726         struct tcp_sock *tp = tcp_sk(sk);
1727         int err = 0;
1728         int old_state = sk->sk_state;
1729
1730         if (old_state != TCP_CLOSE)
1731                 tcp_set_state(sk, TCP_CLOSE);
1732
1733         /* ABORT function of RFC793 */
1734         if (old_state == TCP_LISTEN) {
1735                 inet_csk_listen_stop(sk);
1736         } else if (tcp_need_reset(old_state) ||
1737                    (tp->snd_nxt != tp->write_seq &&
1738                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1739                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1740                  * states
1741                  */
1742                 tcp_send_active_reset(sk, gfp_any());
1743                 sk->sk_err = ECONNRESET;
1744         } else if (old_state == TCP_SYN_SENT)
1745                 sk->sk_err = ECONNRESET;
1746
1747         tcp_clear_xmit_timers(sk);
1748         __skb_queue_purge(&sk->sk_receive_queue);
1749         tcp_write_queue_purge(sk);
1750         __skb_queue_purge(&tp->out_of_order_queue);
1751 #ifdef CONFIG_NET_DMA
1752         __skb_queue_purge(&sk->sk_async_wait_queue);
1753 #endif
1754
1755         inet->dport = 0;
1756
1757         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1758                 inet_reset_saddr(sk);
1759
1760         sk->sk_shutdown = 0;
1761         sock_reset_flag(sk, SOCK_DONE);
1762         tp->srtt = 0;
1763         if ((tp->write_seq += tp->max_window + 2) == 0)
1764                 tp->write_seq = 1;
1765         icsk->icsk_backoff = 0;
1766         tp->snd_cwnd = 2;
1767         icsk->icsk_probes_out = 0;
1768         tp->packets_out = 0;
1769         tp->snd_ssthresh = 0x7fffffff;
1770         tp->snd_cwnd_cnt = 0;
1771         tp->bytes_acked = 0;
1772         tcp_set_ca_state(sk, TCP_CA_Open);
1773         tcp_clear_retrans(tp);
1774         inet_csk_delack_init(sk);
1775         tcp_init_send_head(sk);
1776         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1777         __sk_dst_reset(sk);
1778
1779         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1780
1781         sk->sk_error_report(sk);
1782         return err;
1783 }
1784
1785 /*
1786  *      Socket option code for TCP.
1787  */
1788 static int do_tcp_setsockopt(struct sock *sk, int level,
1789                 int optname, char __user *optval, int optlen)
1790 {
1791         struct tcp_sock *tp = tcp_sk(sk);
1792         struct inet_connection_sock *icsk = inet_csk(sk);
1793         int val;
1794         int err = 0;
1795
1796         /* This is a string value all the others are int's */
1797         if (optname == TCP_CONGESTION) {
1798                 char name[TCP_CA_NAME_MAX];
1799
1800                 if (optlen < 1)
1801                         return -EINVAL;
1802
1803                 val = strncpy_from_user(name, optval,
1804                                         min(TCP_CA_NAME_MAX-1, optlen));
1805                 if (val < 0)
1806                         return -EFAULT;
1807                 name[val] = 0;
1808
1809                 lock_sock(sk);
1810                 err = tcp_set_congestion_control(sk, name);
1811                 release_sock(sk);
1812                 return err;
1813         }
1814
1815         if (optlen < sizeof(int))
1816                 return -EINVAL;
1817
1818         if (get_user(val, (int __user *)optval))
1819                 return -EFAULT;
1820
1821         lock_sock(sk);
1822
1823         switch (optname) {
1824         case TCP_MAXSEG:
1825                 /* Values greater than interface MTU won't take effect. However
1826                  * at the point when this call is done we typically don't yet
1827                  * know which interface is going to be used */
1828                 if (val < 8 || val > MAX_TCP_WINDOW) {
1829                         err = -EINVAL;
1830                         break;
1831                 }
1832                 tp->rx_opt.user_mss = val;
1833                 break;
1834
1835         case TCP_NODELAY:
1836                 if (val) {
1837                         /* TCP_NODELAY is weaker than TCP_CORK, so that
1838                          * this option on corked socket is remembered, but
1839                          * it is not activated until cork is cleared.
1840                          *
1841                          * However, when TCP_NODELAY is set we make
1842                          * an explicit push, which overrides even TCP_CORK
1843                          * for currently queued segments.
1844                          */
1845                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1846                         tcp_push_pending_frames(sk);
1847                 } else {
1848                         tp->nonagle &= ~TCP_NAGLE_OFF;
1849                 }
1850                 break;
1851
1852         case TCP_CORK:
1853                 /* When set indicates to always queue non-full frames.
1854                  * Later the user clears this option and we transmit
1855                  * any pending partial frames in the queue.  This is
1856                  * meant to be used alongside sendfile() to get properly
1857                  * filled frames when the user (for example) must write
1858                  * out headers with a write() call first and then use
1859                  * sendfile to send out the data parts.
1860                  *
1861                  * TCP_CORK can be set together with TCP_NODELAY and it is
1862                  * stronger than TCP_NODELAY.
1863                  */
1864                 if (val) {
1865                         tp->nonagle |= TCP_NAGLE_CORK;
1866                 } else {
1867                         tp->nonagle &= ~TCP_NAGLE_CORK;
1868                         if (tp->nonagle&TCP_NAGLE_OFF)
1869                                 tp->nonagle |= TCP_NAGLE_PUSH;
1870                         tcp_push_pending_frames(sk);
1871                 }
1872                 break;
1873
1874         case TCP_KEEPIDLE:
1875                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1876                         err = -EINVAL;
1877                 else {
1878                         tp->keepalive_time = val * HZ;
1879                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
1880                             !((1 << sk->sk_state) &
1881                               (TCPF_CLOSE | TCPF_LISTEN))) {
1882                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1883                                 if (tp->keepalive_time > elapsed)
1884                                         elapsed = tp->keepalive_time - elapsed;
1885                                 else
1886                                         elapsed = 0;
1887                                 inet_csk_reset_keepalive_timer(sk, elapsed);
1888                         }
1889                 }
1890                 break;
1891         case TCP_KEEPINTVL:
1892                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1893                         err = -EINVAL;
1894                 else
1895                         tp->keepalive_intvl = val * HZ;
1896                 break;
1897         case TCP_KEEPCNT:
1898                 if (val < 1 || val > MAX_TCP_KEEPCNT)
1899                         err = -EINVAL;
1900                 else
1901                         tp->keepalive_probes = val;
1902                 break;
1903         case TCP_SYNCNT:
1904                 if (val < 1 || val > MAX_TCP_SYNCNT)
1905                         err = -EINVAL;
1906                 else
1907                         icsk->icsk_syn_retries = val;
1908                 break;
1909
1910         case TCP_LINGER2:
1911                 if (val < 0)
1912                         tp->linger2 = -1;
1913                 else if (val > sysctl_tcp_fin_timeout / HZ)
1914                         tp->linger2 = 0;
1915                 else
1916                         tp->linger2 = val * HZ;
1917                 break;
1918
1919         case TCP_DEFER_ACCEPT:
1920                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1921                 if (val > 0) {
1922                         /* Translate value in seconds to number of
1923                          * retransmits */
1924                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1925                                val > ((TCP_TIMEOUT_INIT / HZ) <<
1926                                        icsk->icsk_accept_queue.rskq_defer_accept))
1927                                 icsk->icsk_accept_queue.rskq_defer_accept++;
1928                         icsk->icsk_accept_queue.rskq_defer_accept++;
1929                 }
1930                 break;
1931
1932         case TCP_WINDOW_CLAMP:
1933                 if (!val) {
1934                         if (sk->sk_state != TCP_CLOSE) {
1935                                 err = -EINVAL;
1936                                 break;
1937                         }
1938                         tp->window_clamp = 0;
1939                 } else
1940                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1941                                                 SOCK_MIN_RCVBUF / 2 : val;
1942                 break;
1943
1944         case TCP_QUICKACK:
1945                 if (!val) {
1946                         icsk->icsk_ack.pingpong = 1;
1947                 } else {
1948                         icsk->icsk_ack.pingpong = 0;
1949                         if ((1 << sk->sk_state) &
1950                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1951                             inet_csk_ack_scheduled(sk)) {
1952                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1953                                 tcp_cleanup_rbuf(sk, 1);
1954                                 if (!(val & 1))
1955                                         icsk->icsk_ack.pingpong = 1;
1956                         }
1957                 }
1958                 break;
1959
1960 #ifdef CONFIG_TCP_MD5SIG
1961         case TCP_MD5SIG:
1962                 /* Read the IP->Key mappings from userspace */
1963                 err = tp->af_specific->md5_parse(sk, optval, optlen);
1964                 break;
1965 #endif
1966
1967         default:
1968                 err = -ENOPROTOOPT;
1969                 break;
1970         }
1971
1972         release_sock(sk);
1973         return err;
1974 }
1975
1976 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1977                    int optlen)
1978 {
1979         struct inet_connection_sock *icsk = inet_csk(sk);
1980
1981         if (level != SOL_TCP)
1982                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1983                                                      optval, optlen);
1984         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1985 }
1986
1987 #ifdef CONFIG_COMPAT
1988 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1989                           char __user *optval, int optlen)
1990 {
1991         if (level != SOL_TCP)
1992                 return inet_csk_compat_setsockopt(sk, level, optname,
1993                                                   optval, optlen);
1994         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1995 }
1996
1997 EXPORT_SYMBOL(compat_tcp_setsockopt);
1998 #endif
1999
2000 /* Return information about state of tcp endpoint in API format. */
2001 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2002 {
2003         struct tcp_sock *tp = tcp_sk(sk);
2004         const struct inet_connection_sock *icsk = inet_csk(sk);
2005         u32 now = tcp_time_stamp;
2006
2007         memset(info, 0, sizeof(*info));
2008
2009         info->tcpi_state = sk->sk_state;
2010         info->tcpi_ca_state = icsk->icsk_ca_state;
2011         info->tcpi_retransmits = icsk->icsk_retransmits;
2012         info->tcpi_probes = icsk->icsk_probes_out;
2013         info->tcpi_backoff = icsk->icsk_backoff;
2014
2015         if (tp->rx_opt.tstamp_ok)
2016                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2017         if (tp->rx_opt.sack_ok)
2018                 info->tcpi_options |= TCPI_OPT_SACK;
2019         if (tp->rx_opt.wscale_ok) {
2020                 info->tcpi_options |= TCPI_OPT_WSCALE;
2021                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2022                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2023         }
2024
2025         if (tp->ecn_flags&TCP_ECN_OK)
2026                 info->tcpi_options |= TCPI_OPT_ECN;
2027
2028         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2029         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2030         info->tcpi_snd_mss = tp->mss_cache;
2031         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2032
2033         info->tcpi_unacked = tp->packets_out;
2034         info->tcpi_sacked = tp->sacked_out;
2035         info->tcpi_lost = tp->lost_out;
2036         info->tcpi_retrans = tp->retrans_out;
2037         info->tcpi_fackets = tp->fackets_out;
2038
2039         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2040         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2041         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2042
2043         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2044         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2045         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2046         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2047         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2048         info->tcpi_snd_cwnd = tp->snd_cwnd;
2049         info->tcpi_advmss = tp->advmss;
2050         info->tcpi_reordering = tp->reordering;
2051
2052         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2053         info->tcpi_rcv_space = tp->rcvq_space.space;
2054
2055         info->tcpi_total_retrans = tp->total_retrans;
2056 }
2057
2058 EXPORT_SYMBOL_GPL(tcp_get_info);
2059
2060 static int do_tcp_getsockopt(struct sock *sk, int level,
2061                 int optname, char __user *optval, int __user *optlen)
2062 {
2063         struct inet_connection_sock *icsk = inet_csk(sk);
2064         struct tcp_sock *tp = tcp_sk(sk);
2065         int val, len;
2066
2067         if (get_user(len, optlen))
2068                 return -EFAULT;
2069
2070         len = min_t(unsigned int, len, sizeof(int));
2071
2072         if (len < 0)
2073                 return -EINVAL;
2074
2075         switch (optname) {
2076         case TCP_MAXSEG:
2077                 val = tp->mss_cache;
2078                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2079                         val = tp->rx_opt.user_mss;
2080                 break;
2081         case TCP_NODELAY:
2082                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2083                 break;
2084         case TCP_CORK:
2085                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2086                 break;
2087         case TCP_KEEPIDLE:
2088                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2089                 break;
2090         case TCP_KEEPINTVL:
2091                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2092                 break;
2093         case TCP_KEEPCNT:
2094                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2095                 break;
2096         case TCP_SYNCNT:
2097                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2098                 break;
2099         case TCP_LINGER2:
2100                 val = tp->linger2;
2101                 if (val >= 0)
2102                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2103                 break;
2104         case TCP_DEFER_ACCEPT:
2105                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2106                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2107                 break;
2108         case TCP_WINDOW_CLAMP:
2109                 val = tp->window_clamp;
2110                 break;
2111         case TCP_INFO: {
2112                 struct tcp_info info;
2113
2114                 if (get_user(len, optlen))
2115                         return -EFAULT;
2116
2117                 tcp_get_info(sk, &info);
2118
2119                 len = min_t(unsigned int, len, sizeof(info));
2120                 if (put_user(len, optlen))
2121                         return -EFAULT;
2122                 if (copy_to_user(optval, &info, len))
2123                         return -EFAULT;
2124                 return 0;
2125         }
2126         case TCP_QUICKACK:
2127                 val = !icsk->icsk_ack.pingpong;
2128                 break;
2129
2130         case TCP_CONGESTION:
2131                 if (get_user(len, optlen))
2132                         return -EFAULT;
2133                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2134                 if (put_user(len, optlen))
2135                         return -EFAULT;
2136                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2137                         return -EFAULT;
2138                 return 0;
2139         default:
2140                 return -ENOPROTOOPT;
2141         }
2142
2143         if (put_user(len, optlen))
2144                 return -EFAULT;
2145         if (copy_to_user(optval, &val, len))
2146                 return -EFAULT;
2147         return 0;
2148 }
2149
2150 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2151                    int __user *optlen)
2152 {
2153         struct inet_connection_sock *icsk = inet_csk(sk);
2154
2155         if (level != SOL_TCP)
2156                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2157                                                      optval, optlen);
2158         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2159 }
2160
2161 #ifdef CONFIG_COMPAT
2162 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2163                           char __user *optval, int __user *optlen)
2164 {
2165         if (level != SOL_TCP)
2166                 return inet_csk_compat_getsockopt(sk, level, optname,
2167                                                   optval, optlen);
2168         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2169 }
2170
2171 EXPORT_SYMBOL(compat_tcp_getsockopt);
2172 #endif
2173
2174 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2175 {
2176         struct sk_buff *segs = ERR_PTR(-EINVAL);
2177         struct tcphdr *th;
2178         unsigned thlen;
2179         unsigned int seq;
2180         __be32 delta;
2181         unsigned int oldlen;
2182         unsigned int len;
2183
2184         if (!pskb_may_pull(skb, sizeof(*th)))
2185                 goto out;
2186
2187         th = tcp_hdr(skb);
2188         thlen = th->doff * 4;
2189         if (thlen < sizeof(*th))
2190                 goto out;
2191
2192         if (!pskb_may_pull(skb, thlen))
2193                 goto out;
2194
2195         oldlen = (u16)~skb->len;
2196         __skb_pull(skb, thlen);
2197
2198         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2199                 /* Packet is from an untrusted source, reset gso_segs. */
2200                 int type = skb_shinfo(skb)->gso_type;
2201                 int mss;
2202
2203                 if (unlikely(type &
2204                              ~(SKB_GSO_TCPV4 |
2205                                SKB_GSO_DODGY |
2206                                SKB_GSO_TCP_ECN |
2207                                SKB_GSO_TCPV6 |
2208                                0) ||
2209                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2210                         goto out;
2211
2212                 mss = skb_shinfo(skb)->gso_size;
2213                 skb_shinfo(skb)->gso_segs = (skb->len + mss - 1) / mss;
2214
2215                 segs = NULL;
2216                 goto out;
2217         }
2218
2219         segs = skb_segment(skb, features);
2220         if (IS_ERR(segs))
2221                 goto out;
2222
2223         len = skb_shinfo(skb)->gso_size;
2224         delta = htonl(oldlen + (thlen + len));
2225
2226         skb = segs;
2227         th = tcp_hdr(skb);
2228         seq = ntohl(th->seq);
2229
2230         do {
2231                 th->fin = th->psh = 0;
2232
2233                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2234                                        (__force u32)delta));
2235                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2236                         th->check =
2237                              csum_fold(csum_partial(skb_transport_header(skb),
2238                                                     thlen, skb->csum));
2239
2240                 seq += len;
2241                 skb = skb->next;
2242                 th = tcp_hdr(skb);
2243
2244                 th->seq = htonl(seq);
2245                 th->cwr = 0;
2246         } while (skb->next);
2247
2248         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2249                       skb->data_len);
2250         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2251                                 (__force u32)delta));
2252         if (skb->ip_summed != CHECKSUM_PARTIAL)
2253                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2254                                                    thlen, skb->csum));
2255
2256 out:
2257         return segs;
2258 }
2259 EXPORT_SYMBOL(tcp_tso_segment);
2260
2261 #ifdef CONFIG_TCP_MD5SIG
2262 static unsigned long tcp_md5sig_users;
2263 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2264 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2265
2266 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2267 {
2268         int cpu;
2269         for_each_possible_cpu(cpu) {
2270                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2271                 if (p) {
2272                         if (p->md5_desc.tfm)
2273                                 crypto_free_hash(p->md5_desc.tfm);
2274                         kfree(p);
2275                         p = NULL;
2276                 }
2277         }
2278         free_percpu(pool);
2279 }
2280
2281 void tcp_free_md5sig_pool(void)
2282 {
2283         struct tcp_md5sig_pool **pool = NULL;
2284
2285         spin_lock_bh(&tcp_md5sig_pool_lock);
2286         if (--tcp_md5sig_users == 0) {
2287                 pool = tcp_md5sig_pool;
2288                 tcp_md5sig_pool = NULL;
2289         }
2290         spin_unlock_bh(&tcp_md5sig_pool_lock);
2291         if (pool)
2292                 __tcp_free_md5sig_pool(pool);
2293 }
2294
2295 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2296
2297 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2298 {
2299         int cpu;
2300         struct tcp_md5sig_pool **pool;
2301
2302         pool = alloc_percpu(struct tcp_md5sig_pool *);
2303         if (!pool)
2304                 return NULL;
2305
2306         for_each_possible_cpu(cpu) {
2307                 struct tcp_md5sig_pool *p;
2308                 struct crypto_hash *hash;
2309
2310                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2311                 if (!p)
2312                         goto out_free;
2313                 *per_cpu_ptr(pool, cpu) = p;
2314
2315                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2316                 if (!hash || IS_ERR(hash))
2317                         goto out_free;
2318
2319                 p->md5_desc.tfm = hash;
2320         }
2321         return pool;
2322 out_free:
2323         __tcp_free_md5sig_pool(pool);
2324         return NULL;
2325 }
2326
2327 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2328 {
2329         struct tcp_md5sig_pool **pool;
2330         int alloc = 0;
2331
2332 retry:
2333         spin_lock_bh(&tcp_md5sig_pool_lock);
2334         pool = tcp_md5sig_pool;
2335         if (tcp_md5sig_users++ == 0) {
2336                 alloc = 1;
2337                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2338         } else if (!pool) {
2339                 tcp_md5sig_users--;
2340                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2341                 cpu_relax();
2342                 goto retry;
2343         } else
2344                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2345
2346         if (alloc) {
2347                 /* we cannot hold spinlock here because this may sleep. */
2348                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2349                 spin_lock_bh(&tcp_md5sig_pool_lock);
2350                 if (!p) {
2351                         tcp_md5sig_users--;
2352                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2353                         return NULL;
2354                 }
2355                 pool = tcp_md5sig_pool;
2356                 if (pool) {
2357                         /* oops, it has already been assigned. */
2358                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2359                         __tcp_free_md5sig_pool(p);
2360                 } else {
2361                         tcp_md5sig_pool = pool = p;
2362                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2363                 }
2364         }
2365         return pool;
2366 }
2367
2368 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2369
2370 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2371 {
2372         struct tcp_md5sig_pool **p;
2373         spin_lock_bh(&tcp_md5sig_pool_lock);
2374         p = tcp_md5sig_pool;
2375         if (p)
2376                 tcp_md5sig_users++;
2377         spin_unlock_bh(&tcp_md5sig_pool_lock);
2378         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2379 }
2380
2381 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2382
2383 void __tcp_put_md5sig_pool(void)
2384 {
2385         tcp_free_md5sig_pool();
2386 }
2387
2388 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2389 #endif
2390
2391 void tcp_done(struct sock *sk)
2392 {
2393         if(sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2394                 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
2395
2396         tcp_set_state(sk, TCP_CLOSE);
2397         tcp_clear_xmit_timers(sk);
2398
2399         sk->sk_shutdown = SHUTDOWN_MASK;
2400
2401         if (!sock_flag(sk, SOCK_DEAD))
2402                 sk->sk_state_change(sk);
2403         else
2404                 inet_csk_destroy_sock(sk);
2405 }
2406 EXPORT_SYMBOL_GPL(tcp_done);
2407
2408 extern void __skb_cb_too_small_for_tcp(int, int);
2409 extern struct tcp_congestion_ops tcp_reno;
2410
2411 static __initdata unsigned long thash_entries;
2412 static int __init set_thash_entries(char *str)
2413 {
2414         if (!str)
2415                 return 0;
2416         thash_entries = simple_strtoul(str, &str, 0);
2417         return 1;
2418 }
2419 __setup("thash_entries=", set_thash_entries);
2420
2421 void __init tcp_init(void)
2422 {
2423         struct sk_buff *skb = NULL;
2424         unsigned long limit;
2425         int order, i, max_share;
2426
2427         if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2428                 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2429                                            sizeof(skb->cb));
2430
2431         tcp_hashinfo.bind_bucket_cachep =
2432                 kmem_cache_create("tcp_bind_bucket",
2433                                   sizeof(struct inet_bind_bucket), 0,
2434                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2435
2436         /* Size and allocate the main established and bind bucket
2437          * hash tables.
2438          *
2439          * The methodology is similar to that of the buffer cache.
2440          */
2441         tcp_hashinfo.ehash =
2442                 alloc_large_system_hash("TCP established",
2443                                         sizeof(struct inet_ehash_bucket),
2444                                         thash_entries,
2445                                         (num_physpages >= 128 * 1024) ?
2446                                         13 : 15,
2447                                         0,
2448                                         &tcp_hashinfo.ehash_size,
2449                                         NULL,
2450                                         0);
2451         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2452         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2453                 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2454                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2455                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
2456         }
2457
2458         tcp_hashinfo.bhash =
2459                 alloc_large_system_hash("TCP bind",
2460                                         sizeof(struct inet_bind_hashbucket),
2461                                         tcp_hashinfo.ehash_size,
2462                                         (num_physpages >= 128 * 1024) ?
2463                                         13 : 15,
2464                                         0,
2465                                         &tcp_hashinfo.bhash_size,
2466                                         NULL,
2467                                         64 * 1024);
2468         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2469         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2470                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2471                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2472         }
2473
2474         /* Try to be a bit smarter and adjust defaults depending
2475          * on available memory.
2476          */
2477         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2478                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2479                         order++)
2480                 ;
2481         if (order >= 4) {
2482                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2483                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2484                 sysctl_max_syn_backlog = 1024;
2485         } else if (order < 3) {
2486                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2487                 sysctl_tcp_max_orphans >>= (3 - order);
2488                 sysctl_max_syn_backlog = 128;
2489         }
2490
2491         /* Set the pressure threshold to be a fraction of global memory that
2492          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2493          * memory, with a floor of 128 pages.
2494          */
2495         limit = min(nr_all_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2496         limit = (limit * (nr_all_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2497         limit = max(limit, 128UL);
2498         sysctl_tcp_mem[0] = limit / 4 * 3;
2499         sysctl_tcp_mem[1] = limit;
2500         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2501
2502         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2503         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2504         max_share = min(4UL*1024*1024, limit);
2505
2506         sysctl_tcp_wmem[0] = SK_STREAM_MEM_QUANTUM;
2507         sysctl_tcp_wmem[1] = 16*1024;
2508         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2509
2510         sysctl_tcp_rmem[0] = SK_STREAM_MEM_QUANTUM;
2511         sysctl_tcp_rmem[1] = 87380;
2512         sysctl_tcp_rmem[2] = max(87380, max_share);
2513
2514         printk(KERN_INFO "TCP: Hash tables configured "
2515                "(established %d bind %d)\n",
2516                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2517
2518         tcp_register_congestion_control(&tcp_reno);
2519 }
2520
2521 EXPORT_SYMBOL(tcp_close);
2522 EXPORT_SYMBOL(tcp_disconnect);
2523 EXPORT_SYMBOL(tcp_getsockopt);
2524 EXPORT_SYMBOL(tcp_ioctl);
2525 EXPORT_SYMBOL(tcp_poll);
2526 EXPORT_SYMBOL(tcp_read_sock);
2527 EXPORT_SYMBOL(tcp_recvmsg);
2528 EXPORT_SYMBOL(tcp_sendmsg);
2529 EXPORT_SYMBOL(tcp_sendpage);
2530 EXPORT_SYMBOL(tcp_setsockopt);
2531 EXPORT_SYMBOL(tcp_shutdown);
2532 EXPORT_SYMBOL(tcp_statistics);