Merge branch 'server-cluster-locking-api' of git://linux-nfs.org/~bfields/linux
[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/smp_lock.h>
256 #include <linux/fs.h>
257 #include <linux/random.h>
258 #include <linux/bootmem.h>
259 #include <linux/cache.h>
260 #include <linux/err.h>
261 #include <linux/crypto.h>
262
263 #include <net/icmp.h>
264 #include <net/tcp.h>
265 #include <net/xfrm.h>
266 #include <net/ip.h>
267 #include <net/netdma.h>
268
269 #include <asm/uaccess.h>
270 #include <asm/ioctls.h>
271
272 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
273
274 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
275
276 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
277
278 EXPORT_SYMBOL_GPL(tcp_orphan_count);
279
280 int sysctl_tcp_mem[3] __read_mostly;
281 int sysctl_tcp_wmem[3] __read_mostly;
282 int sysctl_tcp_rmem[3] __read_mostly;
283
284 EXPORT_SYMBOL(sysctl_tcp_mem);
285 EXPORT_SYMBOL(sysctl_tcp_rmem);
286 EXPORT_SYMBOL(sysctl_tcp_wmem);
287
288 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
289 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
290
291 EXPORT_SYMBOL(tcp_memory_allocated);
292 EXPORT_SYMBOL(tcp_sockets_allocated);
293
294 /*
295  * Pressure flag: try to collapse.
296  * Technical note: it is used by multiple contexts non atomically.
297  * All the sk_stream_mem_schedule() is of this nature: accounting
298  * is strict, actions are advisory and have some latency.
299  */
300 int tcp_memory_pressure __read_mostly;
301
302 EXPORT_SYMBOL(tcp_memory_pressure);
303
304 void tcp_enter_memory_pressure(void)
305 {
306         if (!tcp_memory_pressure) {
307                 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
308                 tcp_memory_pressure = 1;
309         }
310 }
311
312 EXPORT_SYMBOL(tcp_enter_memory_pressure);
313
314 /*
315  *      Wait for a TCP event.
316  *
317  *      Note that we don't need to lock the socket, as the upper poll layers
318  *      take care of normal races (between the test and the event) and we don't
319  *      go look at any of the socket buffers directly.
320  */
321 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
322 {
323         unsigned int mask;
324         struct sock *sk = sock->sk;
325         struct tcp_sock *tp = tcp_sk(sk);
326
327         poll_wait(file, sk->sk_sleep, wait);
328         if (sk->sk_state == TCP_LISTEN)
329                 return inet_csk_listen_poll(sk);
330
331         /* Socket is not locked. We are protected from async events
332            by poll logic and correct handling of state changes
333            made by another threads is impossible in any case.
334          */
335
336         mask = 0;
337         if (sk->sk_err)
338                 mask = POLLERR;
339
340         /*
341          * POLLHUP is certainly not done right. But poll() doesn't
342          * have a notion of HUP in just one direction, and for a
343          * socket the read side is more interesting.
344          *
345          * Some poll() documentation says that POLLHUP is incompatible
346          * with the POLLOUT/POLLWR flags, so somebody should check this
347          * all. But careful, it tends to be safer to return too many
348          * bits than too few, and you can easily break real applications
349          * if you don't tell them that something has hung up!
350          *
351          * Check-me.
352          *
353          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
354          * our fs/select.c). It means that after we received EOF,
355          * poll always returns immediately, making impossible poll() on write()
356          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
357          * if and only if shutdown has been made in both directions.
358          * Actually, it is interesting to look how Solaris and DUX
359          * solve this dilemma. I would prefer, if PULLHUP were maskable,
360          * then we could set it on SND_SHUTDOWN. BTW examples given
361          * in Stevens' books assume exactly this behaviour, it explains
362          * why PULLHUP is incompatible with POLLOUT.    --ANK
363          *
364          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
365          * blocking on fresh not-connected or disconnected socket. --ANK
366          */
367         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
368                 mask |= POLLHUP;
369         if (sk->sk_shutdown & RCV_SHUTDOWN)
370                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
371
372         /* Connected? */
373         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
374                 /* Potential race condition. If read of tp below will
375                  * escape above sk->sk_state, we can be illegally awaken
376                  * in SYN_* states. */
377                 if ((tp->rcv_nxt != tp->copied_seq) &&
378                     (tp->urg_seq != tp->copied_seq ||
379                      tp->rcv_nxt != tp->copied_seq + 1 ||
380                      sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
381                         mask |= POLLIN | POLLRDNORM;
382
383                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
384                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
385                                 mask |= POLLOUT | POLLWRNORM;
386                         } else {  /* send SIGIO later */
387                                 set_bit(SOCK_ASYNC_NOSPACE,
388                                         &sk->sk_socket->flags);
389                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
390
391                                 /* Race breaker. If space is freed after
392                                  * wspace test but before the flags are set,
393                                  * IO signal will be lost.
394                                  */
395                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
396                                         mask |= POLLOUT | POLLWRNORM;
397                         }
398                 }
399
400                 if (tp->urg_data & TCP_URG_VALID)
401                         mask |= POLLPRI;
402         }
403         return mask;
404 }
405
406 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
407 {
408         struct tcp_sock *tp = tcp_sk(sk);
409         int answ;
410
411         switch (cmd) {
412         case SIOCINQ:
413                 if (sk->sk_state == TCP_LISTEN)
414                         return -EINVAL;
415
416                 lock_sock(sk);
417                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
418                         answ = 0;
419                 else if (sock_flag(sk, SOCK_URGINLINE) ||
420                          !tp->urg_data ||
421                          before(tp->urg_seq, tp->copied_seq) ||
422                          !before(tp->urg_seq, tp->rcv_nxt)) {
423                         answ = tp->rcv_nxt - tp->copied_seq;
424
425                         /* Subtract 1, if FIN is in queue. */
426                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
427                                 answ -=
428                        tcp_hdr((struct sk_buff *)sk->sk_receive_queue.prev)->fin;
429                 } else
430                         answ = tp->urg_seq - tp->copied_seq;
431                 release_sock(sk);
432                 break;
433         case SIOCATMARK:
434                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
435                 break;
436         case SIOCOUTQ:
437                 if (sk->sk_state == TCP_LISTEN)
438                         return -EINVAL;
439
440                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
441                         answ = 0;
442                 else
443                         answ = tp->write_seq - tp->snd_una;
444                 break;
445         default:
446                 return -ENOIOCTLCMD;
447         }
448
449         return put_user(answ, (int __user *)arg);
450 }
451
452 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
453 {
454         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
455         tp->pushed_seq = tp->write_seq;
456 }
457
458 static inline int forced_push(struct tcp_sock *tp)
459 {
460         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
461 }
462
463 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
464 {
465         struct tcp_sock *tp = tcp_sk(sk);
466         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
467
468         skb->csum    = 0;
469         tcb->seq     = tcb->end_seq = tp->write_seq;
470         tcb->flags   = TCPCB_FLAG_ACK;
471         tcb->sacked  = 0;
472         skb_header_release(skb);
473         tcp_add_write_queue_tail(sk, skb);
474         sk_charge_skb(sk, skb);
475         if (tp->nonagle & TCP_NAGLE_PUSH)
476                 tp->nonagle &= ~TCP_NAGLE_PUSH;
477 }
478
479 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
480                                 struct sk_buff *skb)
481 {
482         if (flags & MSG_OOB) {
483                 tp->urg_mode = 1;
484                 tp->snd_up = tp->write_seq;
485                 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
486         }
487 }
488
489 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
490                             int nonagle)
491 {
492         struct tcp_sock *tp = tcp_sk(sk);
493
494         if (tcp_send_head(sk)) {
495                 struct sk_buff *skb = tcp_write_queue_tail(sk);
496                 if (!(flags & MSG_MORE) || forced_push(tp))
497                         tcp_mark_push(tp, skb);
498                 tcp_mark_urg(tp, flags, skb);
499                 __tcp_push_pending_frames(sk, mss_now,
500                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
501         }
502 }
503
504 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
505                          size_t psize, int flags)
506 {
507         struct tcp_sock *tp = tcp_sk(sk);
508         int mss_now, size_goal;
509         int err;
510         ssize_t copied;
511         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
512
513         /* Wait for a connection to finish. */
514         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
515                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
516                         goto out_err;
517
518         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
519
520         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
521         size_goal = tp->xmit_size_goal;
522         copied = 0;
523
524         err = -EPIPE;
525         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
526                 goto do_error;
527
528         while (psize > 0) {
529                 struct sk_buff *skb = tcp_write_queue_tail(sk);
530                 struct page *page = pages[poffset / PAGE_SIZE];
531                 int copy, i, can_coalesce;
532                 int offset = poffset % PAGE_SIZE;
533                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
534
535                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
536 new_segment:
537                         if (!sk_stream_memory_free(sk))
538                                 goto wait_for_sndbuf;
539
540                         skb = sk_stream_alloc_pskb(sk, 0, 0,
541                                                    sk->sk_allocation);
542                         if (!skb)
543                                 goto wait_for_memory;
544
545                         skb_entail(sk, skb);
546                         copy = size_goal;
547                 }
548
549                 if (copy > size)
550                         copy = size;
551
552                 i = skb_shinfo(skb)->nr_frags;
553                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
554                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
555                         tcp_mark_push(tp, skb);
556                         goto new_segment;
557                 }
558                 if (!sk_stream_wmem_schedule(sk, copy))
559                         goto wait_for_memory;
560
561                 if (can_coalesce) {
562                         skb_shinfo(skb)->frags[i - 1].size += copy;
563                 } else {
564                         get_page(page);
565                         skb_fill_page_desc(skb, i, page, offset, copy);
566                 }
567
568                 skb->len += copy;
569                 skb->data_len += copy;
570                 skb->truesize += copy;
571                 sk->sk_wmem_queued += copy;
572                 sk->sk_forward_alloc -= copy;
573                 skb->ip_summed = CHECKSUM_PARTIAL;
574                 tp->write_seq += copy;
575                 TCP_SKB_CB(skb)->end_seq += copy;
576                 skb_shinfo(skb)->gso_segs = 0;
577
578                 if (!copied)
579                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
580
581                 copied += copy;
582                 poffset += copy;
583                 if (!(psize -= copy))
584                         goto out;
585
586                 if (skb->len < mss_now || (flags & MSG_OOB))
587                         continue;
588
589                 if (forced_push(tp)) {
590                         tcp_mark_push(tp, skb);
591                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
592                 } else if (skb == tcp_send_head(sk))
593                         tcp_push_one(sk, mss_now);
594                 continue;
595
596 wait_for_sndbuf:
597                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
598 wait_for_memory:
599                 if (copied)
600                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
601
602                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
603                         goto do_error;
604
605                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
606                 size_goal = tp->xmit_size_goal;
607         }
608
609 out:
610         if (copied)
611                 tcp_push(sk, flags, mss_now, tp->nonagle);
612         return copied;
613
614 do_error:
615         if (copied)
616                 goto out;
617 out_err:
618         return sk_stream_error(sk, flags, err);
619 }
620
621 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
622                      size_t size, int flags)
623 {
624         ssize_t res;
625         struct sock *sk = sock->sk;
626
627         if (!(sk->sk_route_caps & NETIF_F_SG) ||
628             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
629                 return sock_no_sendpage(sock, page, offset, size, flags);
630
631         lock_sock(sk);
632         TCP_CHECK_TIMER(sk);
633         res = do_tcp_sendpages(sk, &page, offset, size, flags);
634         TCP_CHECK_TIMER(sk);
635         release_sock(sk);
636         return res;
637 }
638
639 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
640 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
641
642 static inline int select_size(struct sock *sk)
643 {
644         struct tcp_sock *tp = tcp_sk(sk);
645         int tmp = tp->mss_cache;
646
647         if (sk->sk_route_caps & NETIF_F_SG) {
648                 if (sk_can_gso(sk))
649                         tmp = 0;
650                 else {
651                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
652
653                         if (tmp >= pgbreak &&
654                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
655                                 tmp = pgbreak;
656                 }
657         }
658
659         return tmp;
660 }
661
662 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
663                 size_t size)
664 {
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 <= len) {
1069                                 seq += used;
1070                                 copied += used;
1071                                 offset += used;
1072                         }
1073                         if (offset != skb->len)
1074                                 break;
1075                 }
1076                 if (tcp_hdr(skb)->fin) {
1077                         sk_eat_skb(sk, skb, 0);
1078                         ++seq;
1079                         break;
1080                 }
1081                 sk_eat_skb(sk, skb, 0);
1082                 if (!desc->count)
1083                         break;
1084         }
1085         tp->copied_seq = seq;
1086
1087         tcp_rcv_space_adjust(sk);
1088
1089         /* Clean up data we have read: This will do ACK frames. */
1090         if (copied)
1091                 tcp_cleanup_rbuf(sk, copied);
1092         return copied;
1093 }
1094
1095 /*
1096  *      This routine copies from a sock struct into the user buffer.
1097  *
1098  *      Technical note: in 2.3 we work on _locked_ socket, so that
1099  *      tricks with *seq access order and skb->users are not required.
1100  *      Probably, code can be easily improved even more.
1101  */
1102
1103 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1104                 size_t len, int nonblock, int flags, int *addr_len)
1105 {
1106         struct tcp_sock *tp = tcp_sk(sk);
1107         int copied = 0;
1108         u32 peek_seq;
1109         u32 *seq;
1110         unsigned long used;
1111         int err;
1112         int target;             /* Read at least this many bytes */
1113         long timeo;
1114         struct task_struct *user_recv = NULL;
1115         int copied_early = 0;
1116
1117         lock_sock(sk);
1118
1119         TCP_CHECK_TIMER(sk);
1120
1121         err = -ENOTCONN;
1122         if (sk->sk_state == TCP_LISTEN)
1123                 goto out;
1124
1125         timeo = sock_rcvtimeo(sk, nonblock);
1126
1127         /* Urgent data needs to be handled specially. */
1128         if (flags & MSG_OOB)
1129                 goto recv_urg;
1130
1131         seq = &tp->copied_seq;
1132         if (flags & MSG_PEEK) {
1133                 peek_seq = tp->copied_seq;
1134                 seq = &peek_seq;
1135         }
1136
1137         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1138
1139 #ifdef CONFIG_NET_DMA
1140         tp->ucopy.dma_chan = NULL;
1141         preempt_disable();
1142         if ((len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1143             !sysctl_tcp_low_latency && __get_cpu_var(softnet_data).net_dma) {
1144                 preempt_enable_no_resched();
1145                 tp->ucopy.pinned_list = dma_pin_iovec_pages(msg->msg_iov, len);
1146         } else
1147                 preempt_enable_no_resched();
1148 #endif
1149
1150         do {
1151                 struct sk_buff *skb;
1152                 u32 offset;
1153
1154                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1155                 if (tp->urg_data && tp->urg_seq == *seq) {
1156                         if (copied)
1157                                 break;
1158                         if (signal_pending(current)) {
1159                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1160                                 break;
1161                         }
1162                 }
1163
1164                 /* Next get a buffer. */
1165
1166                 skb = skb_peek(&sk->sk_receive_queue);
1167                 do {
1168                         if (!skb)
1169                                 break;
1170
1171                         /* Now that we have two receive queues this
1172                          * shouldn't happen.
1173                          */
1174                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1175                                 printk(KERN_INFO "recvmsg bug: copied %X "
1176                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1177                                 break;
1178                         }
1179                         offset = *seq - TCP_SKB_CB(skb)->seq;
1180                         if (tcp_hdr(skb)->syn)
1181                                 offset--;
1182                         if (offset < skb->len)
1183                                 goto found_ok_skb;
1184                         if (tcp_hdr(skb)->fin)
1185                                 goto found_fin_ok;
1186                         BUG_TRAP(flags & MSG_PEEK);
1187                         skb = skb->next;
1188                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1189
1190                 /* Well, if we have backlog, try to process it now yet. */
1191
1192                 if (copied >= target && !sk->sk_backlog.tail)
1193                         break;
1194
1195                 if (copied) {
1196                         if (sk->sk_err ||
1197                             sk->sk_state == TCP_CLOSE ||
1198                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1199                             !timeo ||
1200                             signal_pending(current) ||
1201                             (flags & MSG_PEEK))
1202                                 break;
1203                 } else {
1204                         if (sock_flag(sk, SOCK_DONE))
1205                                 break;
1206
1207                         if (sk->sk_err) {
1208                                 copied = sock_error(sk);
1209                                 break;
1210                         }
1211
1212                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1213                                 break;
1214
1215                         if (sk->sk_state == TCP_CLOSE) {
1216                                 if (!sock_flag(sk, SOCK_DONE)) {
1217                                         /* This occurs when user tries to read
1218                                          * from never connected socket.
1219                                          */
1220                                         copied = -ENOTCONN;
1221                                         break;
1222                                 }
1223                                 break;
1224                         }
1225
1226                         if (!timeo) {
1227                                 copied = -EAGAIN;
1228                                 break;
1229                         }
1230
1231                         if (signal_pending(current)) {
1232                                 copied = sock_intr_errno(timeo);
1233                                 break;
1234                         }
1235                 }
1236
1237                 tcp_cleanup_rbuf(sk, copied);
1238
1239                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1240                         /* Install new reader */
1241                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1242                                 user_recv = current;
1243                                 tp->ucopy.task = user_recv;
1244                                 tp->ucopy.iov = msg->msg_iov;
1245                         }
1246
1247                         tp->ucopy.len = len;
1248
1249                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1250                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1251
1252                         /* Ugly... If prequeue is not empty, we have to
1253                          * process it before releasing socket, otherwise
1254                          * order will be broken at second iteration.
1255                          * More elegant solution is required!!!
1256                          *
1257                          * Look: we have the following (pseudo)queues:
1258                          *
1259                          * 1. packets in flight
1260                          * 2. backlog
1261                          * 3. prequeue
1262                          * 4. receive_queue
1263                          *
1264                          * Each queue can be processed only if the next ones
1265                          * are empty. At this point we have empty receive_queue.
1266                          * But prequeue _can_ be not empty after 2nd iteration,
1267                          * when we jumped to start of loop because backlog
1268                          * processing added something to receive_queue.
1269                          * We cannot release_sock(), because backlog contains
1270                          * packets arrived _after_ prequeued ones.
1271                          *
1272                          * Shortly, algorithm is clear --- to process all
1273                          * the queues in order. We could make it more directly,
1274                          * requeueing packets from backlog to prequeue, if
1275                          * is not empty. It is more elegant, but eats cycles,
1276                          * unfortunately.
1277                          */
1278                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1279                                 goto do_prequeue;
1280
1281                         /* __ Set realtime policy in scheduler __ */
1282                 }
1283
1284                 if (copied >= target) {
1285                         /* Do not sleep, just process backlog. */
1286                         release_sock(sk);
1287                         lock_sock(sk);
1288                 } else
1289                         sk_wait_data(sk, &timeo);
1290
1291 #ifdef CONFIG_NET_DMA
1292                 tp->ucopy.wakeup = 0;
1293 #endif
1294
1295                 if (user_recv) {
1296                         int chunk;
1297
1298                         /* __ Restore normal policy in scheduler __ */
1299
1300                         if ((chunk = len - tp->ucopy.len) != 0) {
1301                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1302                                 len -= chunk;
1303                                 copied += chunk;
1304                         }
1305
1306                         if (tp->rcv_nxt == tp->copied_seq &&
1307                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1308 do_prequeue:
1309                                 tcp_prequeue_process(sk);
1310
1311                                 if ((chunk = len - tp->ucopy.len) != 0) {
1312                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1313                                         len -= chunk;
1314                                         copied += chunk;
1315                                 }
1316                         }
1317                 }
1318                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1319                         if (net_ratelimit())
1320                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1321                                        current->comm, current->pid);
1322                         peek_seq = tp->copied_seq;
1323                 }
1324                 continue;
1325
1326         found_ok_skb:
1327                 /* Ok so how much can we use? */
1328                 used = skb->len - offset;
1329                 if (len < used)
1330                         used = len;
1331
1332                 /* Do we have urgent data here? */
1333                 if (tp->urg_data) {
1334                         u32 urg_offset = tp->urg_seq - *seq;
1335                         if (urg_offset < used) {
1336                                 if (!urg_offset) {
1337                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1338                                                 ++*seq;
1339                                                 offset++;
1340                                                 used--;
1341                                                 if (!used)
1342                                                         goto skip_copy;
1343                                         }
1344                                 } else
1345                                         used = urg_offset;
1346                         }
1347                 }
1348
1349                 if (!(flags & MSG_TRUNC)) {
1350 #ifdef CONFIG_NET_DMA
1351                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1352                                 tp->ucopy.dma_chan = get_softnet_dma();
1353
1354                         if (tp->ucopy.dma_chan) {
1355                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1356                                         tp->ucopy.dma_chan, skb, offset,
1357                                         msg->msg_iov, used,
1358                                         tp->ucopy.pinned_list);
1359
1360                                 if (tp->ucopy.dma_cookie < 0) {
1361
1362                                         printk(KERN_ALERT "dma_cookie < 0\n");
1363
1364                                         /* Exception. Bailout! */
1365                                         if (!copied)
1366                                                 copied = -EFAULT;
1367                                         break;
1368                                 }
1369                                 if ((offset + used) == skb->len)
1370                                         copied_early = 1;
1371
1372                         } else
1373 #endif
1374                         {
1375                                 err = skb_copy_datagram_iovec(skb, offset,
1376                                                 msg->msg_iov, used);
1377                                 if (err) {
1378                                         /* Exception. Bailout! */
1379                                         if (!copied)
1380                                                 copied = -EFAULT;
1381                                         break;
1382                                 }
1383                         }
1384                 }
1385
1386                 *seq += used;
1387                 copied += used;
1388                 len -= used;
1389
1390                 tcp_rcv_space_adjust(sk);
1391
1392 skip_copy:
1393                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1394                         tp->urg_data = 0;
1395                         tcp_fast_path_check(sk);
1396                 }
1397                 if (used + offset < skb->len)
1398                         continue;
1399
1400                 if (tcp_hdr(skb)->fin)
1401                         goto found_fin_ok;
1402                 if (!(flags & MSG_PEEK)) {
1403                         sk_eat_skb(sk, skb, copied_early);
1404                         copied_early = 0;
1405                 }
1406                 continue;
1407
1408         found_fin_ok:
1409                 /* Process the FIN. */
1410                 ++*seq;
1411                 if (!(flags & MSG_PEEK)) {
1412                         sk_eat_skb(sk, skb, copied_early);
1413                         copied_early = 0;
1414                 }
1415                 break;
1416         } while (len > 0);
1417
1418         if (user_recv) {
1419                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1420                         int chunk;
1421
1422                         tp->ucopy.len = copied > 0 ? len : 0;
1423
1424                         tcp_prequeue_process(sk);
1425
1426                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1427                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1428                                 len -= chunk;
1429                                 copied += chunk;
1430                         }
1431                 }
1432
1433                 tp->ucopy.task = NULL;
1434                 tp->ucopy.len = 0;
1435         }
1436
1437 #ifdef CONFIG_NET_DMA
1438         if (tp->ucopy.dma_chan) {
1439                 struct sk_buff *skb;
1440                 dma_cookie_t done, used;
1441
1442                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1443
1444                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1445                                                  tp->ucopy.dma_cookie, &done,
1446                                                  &used) == DMA_IN_PROGRESS) {
1447                         /* do partial cleanup of sk_async_wait_queue */
1448                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1449                                (dma_async_is_complete(skb->dma_cookie, done,
1450                                                       used) == DMA_SUCCESS)) {
1451                                 __skb_dequeue(&sk->sk_async_wait_queue);
1452                                 kfree_skb(skb);
1453                         }
1454                 }
1455
1456                 /* Safe to free early-copied skbs now */
1457                 __skb_queue_purge(&sk->sk_async_wait_queue);
1458                 dma_chan_put(tp->ucopy.dma_chan);
1459                 tp->ucopy.dma_chan = NULL;
1460         }
1461         if (tp->ucopy.pinned_list) {
1462                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1463                 tp->ucopy.pinned_list = NULL;
1464         }
1465 #endif
1466
1467         /* According to UNIX98, msg_name/msg_namelen are ignored
1468          * on connected socket. I was just happy when found this 8) --ANK
1469          */
1470
1471         /* Clean up data we have read: This will do ACK frames. */
1472         tcp_cleanup_rbuf(sk, copied);
1473
1474         TCP_CHECK_TIMER(sk);
1475         release_sock(sk);
1476         return copied;
1477
1478 out:
1479         TCP_CHECK_TIMER(sk);
1480         release_sock(sk);
1481         return err;
1482
1483 recv_urg:
1484         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1485         goto out;
1486 }
1487
1488 /*
1489  *      State processing on a close. This implements the state shift for
1490  *      sending our FIN frame. Note that we only send a FIN for some
1491  *      states. A shutdown() may have already sent the FIN, or we may be
1492  *      closed.
1493  */
1494
1495 static const unsigned char new_state[16] = {
1496   /* current state:        new state:      action:      */
1497   /* (Invalid)          */ TCP_CLOSE,
1498   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1499   /* TCP_SYN_SENT       */ TCP_CLOSE,
1500   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1501   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1502   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1503   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1504   /* TCP_CLOSE          */ TCP_CLOSE,
1505   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1506   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1507   /* TCP_LISTEN         */ TCP_CLOSE,
1508   /* TCP_CLOSING        */ TCP_CLOSING,
1509 };
1510
1511 static int tcp_close_state(struct sock *sk)
1512 {
1513         int next = (int)new_state[sk->sk_state];
1514         int ns = next & TCP_STATE_MASK;
1515
1516         tcp_set_state(sk, ns);
1517
1518         return next & TCP_ACTION_FIN;
1519 }
1520
1521 /*
1522  *      Shutdown the sending side of a connection. Much like close except
1523  *      that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1524  */
1525
1526 void tcp_shutdown(struct sock *sk, int how)
1527 {
1528         /*      We need to grab some memory, and put together a FIN,
1529          *      and then put it into the queue to be sent.
1530          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1531          */
1532         if (!(how & SEND_SHUTDOWN))
1533                 return;
1534
1535         /* If we've already sent a FIN, or it's a closed state, skip this. */
1536         if ((1 << sk->sk_state) &
1537             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1538              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1539                 /* Clear out any half completed packets.  FIN if needed. */
1540                 if (tcp_close_state(sk))
1541                         tcp_send_fin(sk);
1542         }
1543 }
1544
1545 void tcp_close(struct sock *sk, long timeout)
1546 {
1547         struct sk_buff *skb;
1548         int data_was_unread = 0;
1549         int state;
1550
1551         lock_sock(sk);
1552         sk->sk_shutdown = SHUTDOWN_MASK;
1553
1554         if (sk->sk_state == TCP_LISTEN) {
1555                 tcp_set_state(sk, TCP_CLOSE);
1556
1557                 /* Special case. */
1558                 inet_csk_listen_stop(sk);
1559
1560                 goto adjudge_to_death;
1561         }
1562
1563         /*  We need to flush the recv. buffs.  We do this only on the
1564          *  descriptor close, not protocol-sourced closes, because the
1565          *  reader process may not have drained the data yet!
1566          */
1567         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1568                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1569                           tcp_hdr(skb)->fin;
1570                 data_was_unread += len;
1571                 __kfree_skb(skb);
1572         }
1573
1574         sk_stream_mem_reclaim(sk);
1575
1576         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1577          * data was lost. To witness the awful effects of the old behavior of
1578          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1579          * GET in an FTP client, suspend the process, wait for the client to
1580          * advertise a zero window, then kill -9 the FTP client, wheee...
1581          * Note: timeout is always zero in such a case.
1582          */
1583         if (data_was_unread) {
1584                 /* Unread data was tossed, zap the connection. */
1585                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1586                 tcp_set_state(sk, TCP_CLOSE);
1587                 tcp_send_active_reset(sk, GFP_KERNEL);
1588         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1589                 /* Check zero linger _after_ checking for unread data. */
1590                 sk->sk_prot->disconnect(sk, 0);
1591                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1592         } else if (tcp_close_state(sk)) {
1593                 /* We FIN if the application ate all the data before
1594                  * zapping the connection.
1595                  */
1596
1597                 /* RED-PEN. Formally speaking, we have broken TCP state
1598                  * machine. State transitions:
1599                  *
1600                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1601                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1602                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1603                  *
1604                  * are legal only when FIN has been sent (i.e. in window),
1605                  * rather than queued out of window. Purists blame.
1606                  *
1607                  * F.e. "RFC state" is ESTABLISHED,
1608                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1609                  *
1610                  * The visible declinations are that sometimes
1611                  * we enter time-wait state, when it is not required really
1612                  * (harmless), do not send active resets, when they are
1613                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1614                  * they look as CLOSING or LAST_ACK for Linux)
1615                  * Probably, I missed some more holelets.
1616                  *                                              --ANK
1617                  */
1618                 tcp_send_fin(sk);
1619         }
1620
1621         sk_stream_wait_close(sk, timeout);
1622
1623 adjudge_to_death:
1624         state = sk->sk_state;
1625         sock_hold(sk);
1626         sock_orphan(sk);
1627         atomic_inc(sk->sk_prot->orphan_count);
1628
1629         /* It is the last release_sock in its life. It will remove backlog. */
1630         release_sock(sk);
1631
1632
1633         /* Now socket is owned by kernel and we acquire BH lock
1634            to finish close. No need to check for user refs.
1635          */
1636         local_bh_disable();
1637         bh_lock_sock(sk);
1638         BUG_TRAP(!sock_owned_by_user(sk));
1639
1640         /* Have we already been destroyed by a softirq or backlog? */
1641         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1642                 goto out;
1643
1644         /*      This is a (useful) BSD violating of the RFC. There is a
1645          *      problem with TCP as specified in that the other end could
1646          *      keep a socket open forever with no application left this end.
1647          *      We use a 3 minute timeout (about the same as BSD) then kill
1648          *      our end. If they send after that then tough - BUT: long enough
1649          *      that we won't make the old 4*rto = almost no time - whoops
1650          *      reset mistake.
1651          *
1652          *      Nope, it was not mistake. It is really desired behaviour
1653          *      f.e. on http servers, when such sockets are useless, but
1654          *      consume significant resources. Let's do it with special
1655          *      linger2 option.                                 --ANK
1656          */
1657
1658         if (sk->sk_state == TCP_FIN_WAIT2) {
1659                 struct tcp_sock *tp = tcp_sk(sk);
1660                 if (tp->linger2 < 0) {
1661                         tcp_set_state(sk, TCP_CLOSE);
1662                         tcp_send_active_reset(sk, GFP_ATOMIC);
1663                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1664                 } else {
1665                         const int tmo = tcp_fin_time(sk);
1666
1667                         if (tmo > TCP_TIMEWAIT_LEN) {
1668                                 inet_csk_reset_keepalive_timer(sk,
1669                                                 tmo - TCP_TIMEWAIT_LEN);
1670                         } else {
1671                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1672                                 goto out;
1673                         }
1674                 }
1675         }
1676         if (sk->sk_state != TCP_CLOSE) {
1677                 sk_stream_mem_reclaim(sk);
1678                 if (atomic_read(sk->sk_prot->orphan_count) > sysctl_tcp_max_orphans ||
1679                     (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1680                      atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1681                         if (net_ratelimit())
1682                                 printk(KERN_INFO "TCP: too many of orphaned "
1683                                        "sockets\n");
1684                         tcp_set_state(sk, TCP_CLOSE);
1685                         tcp_send_active_reset(sk, GFP_ATOMIC);
1686                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1687                 }
1688         }
1689
1690         if (sk->sk_state == TCP_CLOSE)
1691                 inet_csk_destroy_sock(sk);
1692         /* Otherwise, socket is reprieved until protocol close. */
1693
1694 out:
1695         bh_unlock_sock(sk);
1696         local_bh_enable();
1697         sock_put(sk);
1698 }
1699
1700 /* These states need RST on ABORT according to RFC793 */
1701
1702 static inline int tcp_need_reset(int state)
1703 {
1704         return (1 << state) &
1705                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1706                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1707 }
1708
1709 int tcp_disconnect(struct sock *sk, int flags)
1710 {
1711         struct inet_sock *inet = inet_sk(sk);
1712         struct inet_connection_sock *icsk = inet_csk(sk);
1713         struct tcp_sock *tp = tcp_sk(sk);
1714         int err = 0;
1715         int old_state = sk->sk_state;
1716
1717         if (old_state != TCP_CLOSE)
1718                 tcp_set_state(sk, TCP_CLOSE);
1719
1720         /* ABORT function of RFC793 */
1721         if (old_state == TCP_LISTEN) {
1722                 inet_csk_listen_stop(sk);
1723         } else if (tcp_need_reset(old_state) ||
1724                    (tp->snd_nxt != tp->write_seq &&
1725                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1726                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1727                  * states
1728                  */
1729                 tcp_send_active_reset(sk, gfp_any());
1730                 sk->sk_err = ECONNRESET;
1731         } else if (old_state == TCP_SYN_SENT)
1732                 sk->sk_err = ECONNRESET;
1733
1734         tcp_clear_xmit_timers(sk);
1735         __skb_queue_purge(&sk->sk_receive_queue);
1736         tcp_write_queue_purge(sk);
1737         __skb_queue_purge(&tp->out_of_order_queue);
1738 #ifdef CONFIG_NET_DMA
1739         __skb_queue_purge(&sk->sk_async_wait_queue);
1740 #endif
1741
1742         inet->dport = 0;
1743
1744         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1745                 inet_reset_saddr(sk);
1746
1747         sk->sk_shutdown = 0;
1748         sock_reset_flag(sk, SOCK_DONE);
1749         tp->srtt = 0;
1750         if ((tp->write_seq += tp->max_window + 2) == 0)
1751                 tp->write_seq = 1;
1752         icsk->icsk_backoff = 0;
1753         tp->snd_cwnd = 2;
1754         icsk->icsk_probes_out = 0;
1755         tp->packets_out = 0;
1756         tp->snd_ssthresh = 0x7fffffff;
1757         tp->snd_cwnd_cnt = 0;
1758         tp->bytes_acked = 0;
1759         tcp_set_ca_state(sk, TCP_CA_Open);
1760         tcp_clear_retrans(tp);
1761         inet_csk_delack_init(sk);
1762         tcp_init_send_head(sk);
1763         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1764         __sk_dst_reset(sk);
1765
1766         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1767
1768         sk->sk_error_report(sk);
1769         return err;
1770 }
1771
1772 /*
1773  *      Socket option code for TCP.
1774  */
1775 static int do_tcp_setsockopt(struct sock *sk, int level,
1776                 int optname, char __user *optval, int optlen)
1777 {
1778         struct tcp_sock *tp = tcp_sk(sk);
1779         struct inet_connection_sock *icsk = inet_csk(sk);
1780         int val;
1781         int err = 0;
1782
1783         /* This is a string value all the others are int's */
1784         if (optname == TCP_CONGESTION) {
1785                 char name[TCP_CA_NAME_MAX];
1786
1787                 if (optlen < 1)
1788                         return -EINVAL;
1789
1790                 val = strncpy_from_user(name, optval,
1791                                         min(TCP_CA_NAME_MAX-1, optlen));
1792                 if (val < 0)
1793                         return -EFAULT;
1794                 name[val] = 0;
1795
1796                 lock_sock(sk);
1797                 err = tcp_set_congestion_control(sk, name);
1798                 release_sock(sk);
1799                 return err;
1800         }
1801
1802         if (optlen < sizeof(int))
1803                 return -EINVAL;
1804
1805         if (get_user(val, (int __user *)optval))
1806                 return -EFAULT;
1807
1808         lock_sock(sk);
1809
1810         switch (optname) {
1811         case TCP_MAXSEG:
1812                 /* Values greater than interface MTU won't take effect. However
1813                  * at the point when this call is done we typically don't yet
1814                  * know which interface is going to be used */
1815                 if (val < 8 || val > MAX_TCP_WINDOW) {
1816                         err = -EINVAL;
1817                         break;
1818                 }
1819                 tp->rx_opt.user_mss = val;
1820                 break;
1821
1822         case TCP_NODELAY:
1823                 if (val) {
1824                         /* TCP_NODELAY is weaker than TCP_CORK, so that
1825                          * this option on corked socket is remembered, but
1826                          * it is not activated until cork is cleared.
1827                          *
1828                          * However, when TCP_NODELAY is set we make
1829                          * an explicit push, which overrides even TCP_CORK
1830                          * for currently queued segments.
1831                          */
1832                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1833                         tcp_push_pending_frames(sk);
1834                 } else {
1835                         tp->nonagle &= ~TCP_NAGLE_OFF;
1836                 }
1837                 break;
1838
1839         case TCP_CORK:
1840                 /* When set indicates to always queue non-full frames.
1841                  * Later the user clears this option and we transmit
1842                  * any pending partial frames in the queue.  This is
1843                  * meant to be used alongside sendfile() to get properly
1844                  * filled frames when the user (for example) must write
1845                  * out headers with a write() call first and then use
1846                  * sendfile to send out the data parts.
1847                  *
1848                  * TCP_CORK can be set together with TCP_NODELAY and it is
1849                  * stronger than TCP_NODELAY.
1850                  */
1851                 if (val) {
1852                         tp->nonagle |= TCP_NAGLE_CORK;
1853                 } else {
1854                         tp->nonagle &= ~TCP_NAGLE_CORK;
1855                         if (tp->nonagle&TCP_NAGLE_OFF)
1856                                 tp->nonagle |= TCP_NAGLE_PUSH;
1857                         tcp_push_pending_frames(sk);
1858                 }
1859                 break;
1860
1861         case TCP_KEEPIDLE:
1862                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1863                         err = -EINVAL;
1864                 else {
1865                         tp->keepalive_time = val * HZ;
1866                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
1867                             !((1 << sk->sk_state) &
1868                               (TCPF_CLOSE | TCPF_LISTEN))) {
1869                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1870                                 if (tp->keepalive_time > elapsed)
1871                                         elapsed = tp->keepalive_time - elapsed;
1872                                 else
1873                                         elapsed = 0;
1874                                 inet_csk_reset_keepalive_timer(sk, elapsed);
1875                         }
1876                 }
1877                 break;
1878         case TCP_KEEPINTVL:
1879                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1880                         err = -EINVAL;
1881                 else
1882                         tp->keepalive_intvl = val * HZ;
1883                 break;
1884         case TCP_KEEPCNT:
1885                 if (val < 1 || val > MAX_TCP_KEEPCNT)
1886                         err = -EINVAL;
1887                 else
1888                         tp->keepalive_probes = val;
1889                 break;
1890         case TCP_SYNCNT:
1891                 if (val < 1 || val > MAX_TCP_SYNCNT)
1892                         err = -EINVAL;
1893                 else
1894                         icsk->icsk_syn_retries = val;
1895                 break;
1896
1897         case TCP_LINGER2:
1898                 if (val < 0)
1899                         tp->linger2 = -1;
1900                 else if (val > sysctl_tcp_fin_timeout / HZ)
1901                         tp->linger2 = 0;
1902                 else
1903                         tp->linger2 = val * HZ;
1904                 break;
1905
1906         case TCP_DEFER_ACCEPT:
1907                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1908                 if (val > 0) {
1909                         /* Translate value in seconds to number of
1910                          * retransmits */
1911                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1912                                val > ((TCP_TIMEOUT_INIT / HZ) <<
1913                                        icsk->icsk_accept_queue.rskq_defer_accept))
1914                                 icsk->icsk_accept_queue.rskq_defer_accept++;
1915                         icsk->icsk_accept_queue.rskq_defer_accept++;
1916                 }
1917                 break;
1918
1919         case TCP_WINDOW_CLAMP:
1920                 if (!val) {
1921                         if (sk->sk_state != TCP_CLOSE) {
1922                                 err = -EINVAL;
1923                                 break;
1924                         }
1925                         tp->window_clamp = 0;
1926                 } else
1927                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1928                                                 SOCK_MIN_RCVBUF / 2 : val;
1929                 break;
1930
1931         case TCP_QUICKACK:
1932                 if (!val) {
1933                         icsk->icsk_ack.pingpong = 1;
1934                 } else {
1935                         icsk->icsk_ack.pingpong = 0;
1936                         if ((1 << sk->sk_state) &
1937                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1938                             inet_csk_ack_scheduled(sk)) {
1939                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1940                                 tcp_cleanup_rbuf(sk, 1);
1941                                 if (!(val & 1))
1942                                         icsk->icsk_ack.pingpong = 1;
1943                         }
1944                 }
1945                 break;
1946
1947 #ifdef CONFIG_TCP_MD5SIG
1948         case TCP_MD5SIG:
1949                 /* Read the IP->Key mappings from userspace */
1950                 err = tp->af_specific->md5_parse(sk, optval, optlen);
1951                 break;
1952 #endif
1953
1954         default:
1955                 err = -ENOPROTOOPT;
1956                 break;
1957         }
1958
1959         release_sock(sk);
1960         return err;
1961 }
1962
1963 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1964                    int optlen)
1965 {
1966         struct inet_connection_sock *icsk = inet_csk(sk);
1967
1968         if (level != SOL_TCP)
1969                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1970                                                      optval, optlen);
1971         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1972 }
1973
1974 #ifdef CONFIG_COMPAT
1975 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1976                           char __user *optval, int optlen)
1977 {
1978         if (level != SOL_TCP)
1979                 return inet_csk_compat_setsockopt(sk, level, optname,
1980                                                   optval, optlen);
1981         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1982 }
1983
1984 EXPORT_SYMBOL(compat_tcp_setsockopt);
1985 #endif
1986
1987 /* Return information about state of tcp endpoint in API format. */
1988 void tcp_get_info(struct sock *sk, struct tcp_info *info)
1989 {
1990         struct tcp_sock *tp = tcp_sk(sk);
1991         const struct inet_connection_sock *icsk = inet_csk(sk);
1992         u32 now = tcp_time_stamp;
1993
1994         memset(info, 0, sizeof(*info));
1995
1996         info->tcpi_state = sk->sk_state;
1997         info->tcpi_ca_state = icsk->icsk_ca_state;
1998         info->tcpi_retransmits = icsk->icsk_retransmits;
1999         info->tcpi_probes = icsk->icsk_probes_out;
2000         info->tcpi_backoff = icsk->icsk_backoff;
2001
2002         if (tp->rx_opt.tstamp_ok)
2003                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2004         if (tp->rx_opt.sack_ok)
2005                 info->tcpi_options |= TCPI_OPT_SACK;
2006         if (tp->rx_opt.wscale_ok) {
2007                 info->tcpi_options |= TCPI_OPT_WSCALE;
2008                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2009                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2010         }
2011
2012         if (tp->ecn_flags&TCP_ECN_OK)
2013                 info->tcpi_options |= TCPI_OPT_ECN;
2014
2015         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2016         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2017         info->tcpi_snd_mss = tp->mss_cache;
2018         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2019
2020         info->tcpi_unacked = tp->packets_out;
2021         info->tcpi_sacked = tp->sacked_out;
2022         info->tcpi_lost = tp->lost_out;
2023         info->tcpi_retrans = tp->retrans_out;
2024         info->tcpi_fackets = tp->fackets_out;
2025
2026         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2027         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2028         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2029
2030         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2031         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2032         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2033         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2034         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2035         info->tcpi_snd_cwnd = tp->snd_cwnd;
2036         info->tcpi_advmss = tp->advmss;
2037         info->tcpi_reordering = tp->reordering;
2038
2039         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2040         info->tcpi_rcv_space = tp->rcvq_space.space;
2041
2042         info->tcpi_total_retrans = tp->total_retrans;
2043 }
2044
2045 EXPORT_SYMBOL_GPL(tcp_get_info);
2046
2047 static int do_tcp_getsockopt(struct sock *sk, int level,
2048                 int optname, char __user *optval, int __user *optlen)
2049 {
2050         struct inet_connection_sock *icsk = inet_csk(sk);
2051         struct tcp_sock *tp = tcp_sk(sk);
2052         int val, len;
2053
2054         if (get_user(len, optlen))
2055                 return -EFAULT;
2056
2057         len = min_t(unsigned int, len, sizeof(int));
2058
2059         if (len < 0)
2060                 return -EINVAL;
2061
2062         switch (optname) {
2063         case TCP_MAXSEG:
2064                 val = tp->mss_cache;
2065                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2066                         val = tp->rx_opt.user_mss;
2067                 break;
2068         case TCP_NODELAY:
2069                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2070                 break;
2071         case TCP_CORK:
2072                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2073                 break;
2074         case TCP_KEEPIDLE:
2075                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2076                 break;
2077         case TCP_KEEPINTVL:
2078                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2079                 break;
2080         case TCP_KEEPCNT:
2081                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2082                 break;
2083         case TCP_SYNCNT:
2084                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2085                 break;
2086         case TCP_LINGER2:
2087                 val = tp->linger2;
2088                 if (val >= 0)
2089                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2090                 break;
2091         case TCP_DEFER_ACCEPT:
2092                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2093                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2094                 break;
2095         case TCP_WINDOW_CLAMP:
2096                 val = tp->window_clamp;
2097                 break;
2098         case TCP_INFO: {
2099                 struct tcp_info info;
2100
2101                 if (get_user(len, optlen))
2102                         return -EFAULT;
2103
2104                 tcp_get_info(sk, &info);
2105
2106                 len = min_t(unsigned int, len, sizeof(info));
2107                 if (put_user(len, optlen))
2108                         return -EFAULT;
2109                 if (copy_to_user(optval, &info, len))
2110                         return -EFAULT;
2111                 return 0;
2112         }
2113         case TCP_QUICKACK:
2114                 val = !icsk->icsk_ack.pingpong;
2115                 break;
2116
2117         case TCP_CONGESTION:
2118                 if (get_user(len, optlen))
2119                         return -EFAULT;
2120                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2121                 if (put_user(len, optlen))
2122                         return -EFAULT;
2123                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2124                         return -EFAULT;
2125                 return 0;
2126         default:
2127                 return -ENOPROTOOPT;
2128         }
2129
2130         if (put_user(len, optlen))
2131                 return -EFAULT;
2132         if (copy_to_user(optval, &val, len))
2133                 return -EFAULT;
2134         return 0;
2135 }
2136
2137 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2138                    int __user *optlen)
2139 {
2140         struct inet_connection_sock *icsk = inet_csk(sk);
2141
2142         if (level != SOL_TCP)
2143                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2144                                                      optval, optlen);
2145         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2146 }
2147
2148 #ifdef CONFIG_COMPAT
2149 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2150                           char __user *optval, int __user *optlen)
2151 {
2152         if (level != SOL_TCP)
2153                 return inet_csk_compat_getsockopt(sk, level, optname,
2154                                                   optval, optlen);
2155         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2156 }
2157
2158 EXPORT_SYMBOL(compat_tcp_getsockopt);
2159 #endif
2160
2161 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2162 {
2163         struct sk_buff *segs = ERR_PTR(-EINVAL);
2164         struct tcphdr *th;
2165         unsigned thlen;
2166         unsigned int seq;
2167         __be32 delta;
2168         unsigned int oldlen;
2169         unsigned int len;
2170
2171         if (!pskb_may_pull(skb, sizeof(*th)))
2172                 goto out;
2173
2174         th = tcp_hdr(skb);
2175         thlen = th->doff * 4;
2176         if (thlen < sizeof(*th))
2177                 goto out;
2178
2179         if (!pskb_may_pull(skb, thlen))
2180                 goto out;
2181
2182         oldlen = (u16)~skb->len;
2183         __skb_pull(skb, thlen);
2184
2185         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2186                 /* Packet is from an untrusted source, reset gso_segs. */
2187                 int type = skb_shinfo(skb)->gso_type;
2188                 int mss;
2189
2190                 if (unlikely(type &
2191                              ~(SKB_GSO_TCPV4 |
2192                                SKB_GSO_DODGY |
2193                                SKB_GSO_TCP_ECN |
2194                                SKB_GSO_TCPV6 |
2195                                0) ||
2196                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2197                         goto out;
2198
2199                 mss = skb_shinfo(skb)->gso_size;
2200                 skb_shinfo(skb)->gso_segs = (skb->len + mss - 1) / mss;
2201
2202                 segs = NULL;
2203                 goto out;
2204         }
2205
2206         segs = skb_segment(skb, features);
2207         if (IS_ERR(segs))
2208                 goto out;
2209
2210         len = skb_shinfo(skb)->gso_size;
2211         delta = htonl(oldlen + (thlen + len));
2212
2213         skb = segs;
2214         th = tcp_hdr(skb);
2215         seq = ntohl(th->seq);
2216
2217         do {
2218                 th->fin = th->psh = 0;
2219
2220                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2221                                        (__force u32)delta));
2222                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2223                         th->check =
2224                              csum_fold(csum_partial(skb_transport_header(skb),
2225                                                     thlen, skb->csum));
2226
2227                 seq += len;
2228                 skb = skb->next;
2229                 th = tcp_hdr(skb);
2230
2231                 th->seq = htonl(seq);
2232                 th->cwr = 0;
2233         } while (skb->next);
2234
2235         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2236                       skb->data_len);
2237         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2238                                 (__force u32)delta));
2239         if (skb->ip_summed != CHECKSUM_PARTIAL)
2240                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2241                                                    thlen, skb->csum));
2242
2243 out:
2244         return segs;
2245 }
2246 EXPORT_SYMBOL(tcp_tso_segment);
2247
2248 #ifdef CONFIG_TCP_MD5SIG
2249 static unsigned long tcp_md5sig_users;
2250 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2251 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2252
2253 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2254 {
2255         int cpu;
2256         for_each_possible_cpu(cpu) {
2257                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2258                 if (p) {
2259                         if (p->md5_desc.tfm)
2260                                 crypto_free_hash(p->md5_desc.tfm);
2261                         kfree(p);
2262                         p = NULL;
2263                 }
2264         }
2265         free_percpu(pool);
2266 }
2267
2268 void tcp_free_md5sig_pool(void)
2269 {
2270         struct tcp_md5sig_pool **pool = NULL;
2271
2272         spin_lock_bh(&tcp_md5sig_pool_lock);
2273         if (--tcp_md5sig_users == 0) {
2274                 pool = tcp_md5sig_pool;
2275                 tcp_md5sig_pool = NULL;
2276         }
2277         spin_unlock_bh(&tcp_md5sig_pool_lock);
2278         if (pool)
2279                 __tcp_free_md5sig_pool(pool);
2280 }
2281
2282 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2283
2284 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2285 {
2286         int cpu;
2287         struct tcp_md5sig_pool **pool;
2288
2289         pool = alloc_percpu(struct tcp_md5sig_pool *);
2290         if (!pool)
2291                 return NULL;
2292
2293         for_each_possible_cpu(cpu) {
2294                 struct tcp_md5sig_pool *p;
2295                 struct crypto_hash *hash;
2296
2297                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2298                 if (!p)
2299                         goto out_free;
2300                 *per_cpu_ptr(pool, cpu) = p;
2301
2302                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2303                 if (!hash || IS_ERR(hash))
2304                         goto out_free;
2305
2306                 p->md5_desc.tfm = hash;
2307         }
2308         return pool;
2309 out_free:
2310         __tcp_free_md5sig_pool(pool);
2311         return NULL;
2312 }
2313
2314 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2315 {
2316         struct tcp_md5sig_pool **pool;
2317         int alloc = 0;
2318
2319 retry:
2320         spin_lock_bh(&tcp_md5sig_pool_lock);
2321         pool = tcp_md5sig_pool;
2322         if (tcp_md5sig_users++ == 0) {
2323                 alloc = 1;
2324                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2325         } else if (!pool) {
2326                 tcp_md5sig_users--;
2327                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2328                 cpu_relax();
2329                 goto retry;
2330         } else
2331                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2332
2333         if (alloc) {
2334                 /* we cannot hold spinlock here because this may sleep. */
2335                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2336                 spin_lock_bh(&tcp_md5sig_pool_lock);
2337                 if (!p) {
2338                         tcp_md5sig_users--;
2339                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2340                         return NULL;
2341                 }
2342                 pool = tcp_md5sig_pool;
2343                 if (pool) {
2344                         /* oops, it has already been assigned. */
2345                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2346                         __tcp_free_md5sig_pool(p);
2347                 } else {
2348                         tcp_md5sig_pool = pool = p;
2349                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2350                 }
2351         }
2352         return pool;
2353 }
2354
2355 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2356
2357 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2358 {
2359         struct tcp_md5sig_pool **p;
2360         spin_lock_bh(&tcp_md5sig_pool_lock);
2361         p = tcp_md5sig_pool;
2362         if (p)
2363                 tcp_md5sig_users++;
2364         spin_unlock_bh(&tcp_md5sig_pool_lock);
2365         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2366 }
2367
2368 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2369
2370 void __tcp_put_md5sig_pool(void)
2371 {
2372         tcp_free_md5sig_pool();
2373 }
2374
2375 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2376 #endif
2377
2378 void tcp_done(struct sock *sk)
2379 {
2380         if(sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2381                 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
2382
2383         tcp_set_state(sk, TCP_CLOSE);
2384         tcp_clear_xmit_timers(sk);
2385
2386         sk->sk_shutdown = SHUTDOWN_MASK;
2387
2388         if (!sock_flag(sk, SOCK_DEAD))
2389                 sk->sk_state_change(sk);
2390         else
2391                 inet_csk_destroy_sock(sk);
2392 }
2393 EXPORT_SYMBOL_GPL(tcp_done);
2394
2395 extern void __skb_cb_too_small_for_tcp(int, int);
2396 extern struct tcp_congestion_ops tcp_reno;
2397
2398 static __initdata unsigned long thash_entries;
2399 static int __init set_thash_entries(char *str)
2400 {
2401         if (!str)
2402                 return 0;
2403         thash_entries = simple_strtoul(str, &str, 0);
2404         return 1;
2405 }
2406 __setup("thash_entries=", set_thash_entries);
2407
2408 void __init tcp_init(void)
2409 {
2410         struct sk_buff *skb = NULL;
2411         unsigned long limit;
2412         int order, i, max_share;
2413
2414         if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2415                 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2416                                            sizeof(skb->cb));
2417
2418         tcp_hashinfo.bind_bucket_cachep =
2419                 kmem_cache_create("tcp_bind_bucket",
2420                                   sizeof(struct inet_bind_bucket), 0,
2421                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
2422
2423         /* Size and allocate the main established and bind bucket
2424          * hash tables.
2425          *
2426          * The methodology is similar to that of the buffer cache.
2427          */
2428         tcp_hashinfo.ehash =
2429                 alloc_large_system_hash("TCP established",
2430                                         sizeof(struct inet_ehash_bucket),
2431                                         thash_entries,
2432                                         (num_physpages >= 128 * 1024) ?
2433                                         13 : 15,
2434                                         0,
2435                                         &tcp_hashinfo.ehash_size,
2436                                         NULL,
2437                                         0);
2438         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2439         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2440                 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2441                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2442                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
2443         }
2444
2445         tcp_hashinfo.bhash =
2446                 alloc_large_system_hash("TCP bind",
2447                                         sizeof(struct inet_bind_hashbucket),
2448                                         tcp_hashinfo.ehash_size,
2449                                         (num_physpages >= 128 * 1024) ?
2450                                         13 : 15,
2451                                         0,
2452                                         &tcp_hashinfo.bhash_size,
2453                                         NULL,
2454                                         64 * 1024);
2455         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2456         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2457                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2458                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2459         }
2460
2461         /* Try to be a bit smarter and adjust defaults depending
2462          * on available memory.
2463          */
2464         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2465                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2466                         order++)
2467                 ;
2468         if (order >= 4) {
2469                 sysctl_local_port_range[0] = 32768;
2470                 sysctl_local_port_range[1] = 61000;
2471                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2472                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2473                 sysctl_max_syn_backlog = 1024;
2474         } else if (order < 3) {
2475                 sysctl_local_port_range[0] = 1024 * (3 - order);
2476                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2477                 sysctl_tcp_max_orphans >>= (3 - order);
2478                 sysctl_max_syn_backlog = 128;
2479         }
2480
2481         /* Set the pressure threshold to be a fraction of global memory that
2482          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2483          * memory, with a floor of 128 pages.
2484          */
2485         limit = min(nr_all_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2486         limit = (limit * (nr_all_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2487         limit = max(limit, 128UL);
2488         sysctl_tcp_mem[0] = limit / 4 * 3;
2489         sysctl_tcp_mem[1] = limit;
2490         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2491
2492         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2493         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2494         max_share = min(4UL*1024*1024, limit);
2495
2496         sysctl_tcp_wmem[0] = SK_STREAM_MEM_QUANTUM;
2497         sysctl_tcp_wmem[1] = 16*1024;
2498         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2499
2500         sysctl_tcp_rmem[0] = SK_STREAM_MEM_QUANTUM;
2501         sysctl_tcp_rmem[1] = 87380;
2502         sysctl_tcp_rmem[2] = max(87380, max_share);
2503
2504         printk(KERN_INFO "TCP: Hash tables configured "
2505                "(established %d bind %d)\n",
2506                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2507
2508         tcp_register_congestion_control(&tcp_reno);
2509 }
2510
2511 EXPORT_SYMBOL(tcp_close);
2512 EXPORT_SYMBOL(tcp_disconnect);
2513 EXPORT_SYMBOL(tcp_getsockopt);
2514 EXPORT_SYMBOL(tcp_ioctl);
2515 EXPORT_SYMBOL(tcp_poll);
2516 EXPORT_SYMBOL(tcp_read_sock);
2517 EXPORT_SYMBOL(tcp_recvmsg);
2518 EXPORT_SYMBOL(tcp_sendmsg);
2519 EXPORT_SYMBOL(tcp_sendpage);
2520 EXPORT_SYMBOL(tcp_setsockopt);
2521 EXPORT_SYMBOL(tcp_shutdown);
2522 EXPORT_SYMBOL(tcp_statistics);