Merge branch 'fixes-davem' of git://git.kernel.org/pub/scm/linux/kernel/git/linville...
[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/kernel.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/init.h>
256 #include <linux/fs.h>
257 #include <linux/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 socket *sock, struct msghdr *msg,
663                 size_t size)
664 {
665         struct sock *sk = sock->sk;
666         struct iovec *iov;
667         struct tcp_sock *tp = tcp_sk(sk);
668         struct sk_buff *skb;
669         int iovlen, flags;
670         int mss_now, size_goal;
671         int err, copied;
672         long timeo;
673
674         lock_sock(sk);
675         TCP_CHECK_TIMER(sk);
676
677         flags = msg->msg_flags;
678         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
679
680         /* Wait for a connection to finish. */
681         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
682                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
683                         goto out_err;
684
685         /* This should be in poll */
686         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
687
688         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
689         size_goal = tp->xmit_size_goal;
690
691         /* Ok commence sending. */
692         iovlen = msg->msg_iovlen;
693         iov = msg->msg_iov;
694         copied = 0;
695
696         err = -EPIPE;
697         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
698                 goto do_error;
699
700         while (--iovlen >= 0) {
701                 int seglen = iov->iov_len;
702                 unsigned char __user *from = iov->iov_base;
703
704                 iov++;
705
706                 while (seglen > 0) {
707                         int copy;
708
709                         skb = tcp_write_queue_tail(sk);
710
711                         if (!tcp_send_head(sk) ||
712                             (copy = size_goal - skb->len) <= 0) {
713
714 new_segment:
715                                 /* Allocate new segment. If the interface is SG,
716                                  * allocate skb fitting to single page.
717                                  */
718                                 if (!sk_stream_memory_free(sk))
719                                         goto wait_for_sndbuf;
720
721                                 skb = sk_stream_alloc_pskb(sk, select_size(sk),
722                                                            0, sk->sk_allocation);
723                                 if (!skb)
724                                         goto wait_for_memory;
725
726                                 /*
727                                  * Check whether we can use HW checksum.
728                                  */
729                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
730                                         skb->ip_summed = CHECKSUM_PARTIAL;
731
732                                 skb_entail(sk, skb);
733                                 copy = size_goal;
734                         }
735
736                         /* Try to append data to the end of skb. */
737                         if (copy > seglen)
738                                 copy = seglen;
739
740                         /* Where to copy to? */
741                         if (skb_tailroom(skb) > 0) {
742                                 /* We have some space in skb head. Superb! */
743                                 if (copy > skb_tailroom(skb))
744                                         copy = skb_tailroom(skb);
745                                 if ((err = skb_add_data(skb, from, copy)) != 0)
746                                         goto do_fault;
747                         } else {
748                                 int merge = 0;
749                                 int i = skb_shinfo(skb)->nr_frags;
750                                 struct page *page = TCP_PAGE(sk);
751                                 int off = TCP_OFF(sk);
752
753                                 if (skb_can_coalesce(skb, i, page, off) &&
754                                     off != PAGE_SIZE) {
755                                         /* We can extend the last page
756                                          * fragment. */
757                                         merge = 1;
758                                 } else if (i == MAX_SKB_FRAGS ||
759                                            (!i &&
760                                            !(sk->sk_route_caps & NETIF_F_SG))) {
761                                         /* Need to add new fragment and cannot
762                                          * do this because interface is non-SG,
763                                          * or because all the page slots are
764                                          * busy. */
765                                         tcp_mark_push(tp, skb);
766                                         goto new_segment;
767                                 } else if (page) {
768                                         if (off == PAGE_SIZE) {
769                                                 put_page(page);
770                                                 TCP_PAGE(sk) = page = NULL;
771                                                 off = 0;
772                                         }
773                                 } else
774                                         off = 0;
775
776                                 if (copy > PAGE_SIZE - off)
777                                         copy = PAGE_SIZE - off;
778
779                                 if (!sk_stream_wmem_schedule(sk, copy))
780                                         goto wait_for_memory;
781
782                                 if (!page) {
783                                         /* Allocate new cache page. */
784                                         if (!(page = sk_stream_alloc_page(sk)))
785                                                 goto wait_for_memory;
786                                 }
787
788                                 /* Time to copy data. We are close to
789                                  * the end! */
790                                 err = skb_copy_to_page(sk, from, skb, page,
791                                                        off, copy);
792                                 if (err) {
793                                         /* If this page was new, give it to the
794                                          * socket so it does not get leaked.
795                                          */
796                                         if (!TCP_PAGE(sk)) {
797                                                 TCP_PAGE(sk) = page;
798                                                 TCP_OFF(sk) = 0;
799                                         }
800                                         goto do_error;
801                                 }
802
803                                 /* Update the skb. */
804                                 if (merge) {
805                                         skb_shinfo(skb)->frags[i - 1].size +=
806                                                                         copy;
807                                 } else {
808                                         skb_fill_page_desc(skb, i, page, off, copy);
809                                         if (TCP_PAGE(sk)) {
810                                                 get_page(page);
811                                         } else if (off + copy < PAGE_SIZE) {
812                                                 get_page(page);
813                                                 TCP_PAGE(sk) = page;
814                                         }
815                                 }
816
817                                 TCP_OFF(sk) = off + copy;
818                         }
819
820                         if (!copied)
821                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
822
823                         tp->write_seq += copy;
824                         TCP_SKB_CB(skb)->end_seq += copy;
825                         skb_shinfo(skb)->gso_segs = 0;
826
827                         from += copy;
828                         copied += copy;
829                         if ((seglen -= copy) == 0 && iovlen == 0)
830                                 goto out;
831
832                         if (skb->len < mss_now || (flags & MSG_OOB))
833                                 continue;
834
835                         if (forced_push(tp)) {
836                                 tcp_mark_push(tp, skb);
837                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
838                         } else if (skb == tcp_send_head(sk))
839                                 tcp_push_one(sk, mss_now);
840                         continue;
841
842 wait_for_sndbuf:
843                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
844 wait_for_memory:
845                         if (copied)
846                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
847
848                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
849                                 goto do_error;
850
851                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
852                         size_goal = tp->xmit_size_goal;
853                 }
854         }
855
856 out:
857         if (copied)
858                 tcp_push(sk, flags, mss_now, tp->nonagle);
859         TCP_CHECK_TIMER(sk);
860         release_sock(sk);
861         return copied;
862
863 do_fault:
864         if (!skb->len) {
865                 tcp_unlink_write_queue(skb, sk);
866                 /* It is the one place in all of TCP, except connection
867                  * reset, where we can be unlinking the send_head.
868                  */
869                 tcp_check_send_head(sk, skb);
870                 sk_stream_free_skb(sk, skb);
871         }
872
873 do_error:
874         if (copied)
875                 goto out;
876 out_err:
877         err = sk_stream_error(sk, flags, err);
878         TCP_CHECK_TIMER(sk);
879         release_sock(sk);
880         return err;
881 }
882
883 /*
884  *      Handle reading urgent data. BSD has very simple semantics for
885  *      this, no blocking and very strange errors 8)
886  */
887
888 static int tcp_recv_urg(struct sock *sk, long timeo,
889                         struct msghdr *msg, int len, int flags,
890                         int *addr_len)
891 {
892         struct tcp_sock *tp = tcp_sk(sk);
893
894         /* No URG data to read. */
895         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
896             tp->urg_data == TCP_URG_READ)
897                 return -EINVAL; /* Yes this is right ! */
898
899         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
900                 return -ENOTCONN;
901
902         if (tp->urg_data & TCP_URG_VALID) {
903                 int err = 0;
904                 char c = tp->urg_data;
905
906                 if (!(flags & MSG_PEEK))
907                         tp->urg_data = TCP_URG_READ;
908
909                 /* Read urgent data. */
910                 msg->msg_flags |= MSG_OOB;
911
912                 if (len > 0) {
913                         if (!(flags & MSG_TRUNC))
914                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
915                         len = 1;
916                 } else
917                         msg->msg_flags |= MSG_TRUNC;
918
919                 return err ? -EFAULT : len;
920         }
921
922         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
923                 return 0;
924
925         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
926          * the available implementations agree in this case:
927          * this call should never block, independent of the
928          * blocking state of the socket.
929          * Mike <pall@rz.uni-karlsruhe.de>
930          */
931         return -EAGAIN;
932 }
933
934 /* Clean up the receive buffer for full frames taken by the user,
935  * then send an ACK if necessary.  COPIED is the number of bytes
936  * tcp_recvmsg has given to the user so far, it speeds up the
937  * calculation of whether or not we must ACK for the sake of
938  * a window update.
939  */
940 void tcp_cleanup_rbuf(struct sock *sk, int copied)
941 {
942         struct tcp_sock *tp = tcp_sk(sk);
943         int time_to_ack = 0;
944
945 #if TCP_DEBUG
946         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
947
948         BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
949 #endif
950
951         if (inet_csk_ack_scheduled(sk)) {
952                 const struct inet_connection_sock *icsk = inet_csk(sk);
953                    /* Delayed ACKs frequently hit locked sockets during bulk
954                     * receive. */
955                 if (icsk->icsk_ack.blocked ||
956                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
957                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
958                     /*
959                      * If this read emptied read buffer, we send ACK, if
960                      * connection is not bidirectional, user drained
961                      * receive buffer and there was a small segment
962                      * in queue.
963                      */
964                     (copied > 0 &&
965                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
966                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
967                        !icsk->icsk_ack.pingpong)) &&
968                       !atomic_read(&sk->sk_rmem_alloc)))
969                         time_to_ack = 1;
970         }
971
972         /* We send an ACK if we can now advertise a non-zero window
973          * which has been raised "significantly".
974          *
975          * Even if window raised up to infinity, do not send window open ACK
976          * in states, where we will not receive more. It is useless.
977          */
978         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
979                 __u32 rcv_window_now = tcp_receive_window(tp);
980
981                 /* Optimize, __tcp_select_window() is not cheap. */
982                 if (2*rcv_window_now <= tp->window_clamp) {
983                         __u32 new_window = __tcp_select_window(sk);
984
985                         /* Send ACK now, if this read freed lots of space
986                          * in our buffer. Certainly, new_window is new window.
987                          * We can advertise it now, if it is not less than current one.
988                          * "Lots" means "at least twice" here.
989                          */
990                         if (new_window && new_window >= 2 * rcv_window_now)
991                                 time_to_ack = 1;
992                 }
993         }
994         if (time_to_ack)
995                 tcp_send_ack(sk);
996 }
997
998 static void tcp_prequeue_process(struct sock *sk)
999 {
1000         struct sk_buff *skb;
1001         struct tcp_sock *tp = tcp_sk(sk);
1002
1003         NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
1004
1005         /* RX process wants to run with disabled BHs, though it is not
1006          * necessary */
1007         local_bh_disable();
1008         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1009                 sk->sk_backlog_rcv(sk, skb);
1010         local_bh_enable();
1011
1012         /* Clear memory counter. */
1013         tp->ucopy.memory = 0;
1014 }
1015
1016 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1017 {
1018         struct sk_buff *skb;
1019         u32 offset;
1020
1021         skb_queue_walk(&sk->sk_receive_queue, skb) {
1022                 offset = seq - TCP_SKB_CB(skb)->seq;
1023                 if (tcp_hdr(skb)->syn)
1024                         offset--;
1025                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1026                         *off = offset;
1027                         return skb;
1028                 }
1029         }
1030         return NULL;
1031 }
1032
1033 /*
1034  * This routine provides an alternative to tcp_recvmsg() for routines
1035  * that would like to handle copying from skbuffs directly in 'sendfile'
1036  * fashion.
1037  * Note:
1038  *      - It is assumed that the socket was locked by the caller.
1039  *      - The routine does not block.
1040  *      - At present, there is no support for reading OOB data
1041  *        or for 'peeking' the socket using this routine
1042  *        (although both would be easy to implement).
1043  */
1044 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1045                   sk_read_actor_t recv_actor)
1046 {
1047         struct sk_buff *skb;
1048         struct tcp_sock *tp = tcp_sk(sk);
1049         u32 seq = tp->copied_seq;
1050         u32 offset;
1051         int copied = 0;
1052
1053         if (sk->sk_state == TCP_LISTEN)
1054                 return -ENOTCONN;
1055         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1056                 if (offset < skb->len) {
1057                         size_t used, len;
1058
1059                         len = skb->len - offset;
1060                         /* Stop reading if we hit a patch of urgent data */
1061                         if (tp->urg_data) {
1062                                 u32 urg_offset = tp->urg_seq - seq;
1063                                 if (urg_offset < len)
1064                                         len = urg_offset;
1065                                 if (!len)
1066                                         break;
1067                         }
1068                         used = recv_actor(desc, skb, offset, len);
1069                         if (used < 0) {
1070                                 if (!copied)
1071                                         copied = used;
1072                                 break;
1073                         } else if (used <= len) {
1074                                 seq += used;
1075                                 copied += used;
1076                                 offset += used;
1077                         }
1078                         if (offset != skb->len)
1079                                 break;
1080                 }
1081                 if (tcp_hdr(skb)->fin) {
1082                         sk_eat_skb(sk, skb, 0);
1083                         ++seq;
1084                         break;
1085                 }
1086                 sk_eat_skb(sk, skb, 0);
1087                 if (!desc->count)
1088                         break;
1089         }
1090         tp->copied_seq = seq;
1091
1092         tcp_rcv_space_adjust(sk);
1093
1094         /* Clean up data we have read: This will do ACK frames. */
1095         if (copied > 0)
1096                 tcp_cleanup_rbuf(sk, copied);
1097         return copied;
1098 }
1099
1100 /*
1101  *      This routine copies from a sock struct into the user buffer.
1102  *
1103  *      Technical note: in 2.3 we work on _locked_ socket, so that
1104  *      tricks with *seq access order and skb->users are not required.
1105  *      Probably, code can be easily improved even more.
1106  */
1107
1108 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1109                 size_t len, int nonblock, int flags, int *addr_len)
1110 {
1111         struct tcp_sock *tp = tcp_sk(sk);
1112         int copied = 0;
1113         u32 peek_seq;
1114         u32 *seq;
1115         unsigned long used;
1116         int err;
1117         int target;             /* Read at least this many bytes */
1118         long timeo;
1119         struct task_struct *user_recv = NULL;
1120         int copied_early = 0;
1121         struct sk_buff *skb;
1122
1123         lock_sock(sk);
1124
1125         TCP_CHECK_TIMER(sk);
1126
1127         err = -ENOTCONN;
1128         if (sk->sk_state == TCP_LISTEN)
1129                 goto out;
1130
1131         timeo = sock_rcvtimeo(sk, nonblock);
1132
1133         /* Urgent data needs to be handled specially. */
1134         if (flags & MSG_OOB)
1135                 goto recv_urg;
1136
1137         seq = &tp->copied_seq;
1138         if (flags & MSG_PEEK) {
1139                 peek_seq = tp->copied_seq;
1140                 seq = &peek_seq;
1141         }
1142
1143         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1144
1145 #ifdef CONFIG_NET_DMA
1146         tp->ucopy.dma_chan = NULL;
1147         preempt_disable();
1148         skb = skb_peek_tail(&sk->sk_receive_queue);
1149         {
1150                 int available = 0;
1151
1152                 if (skb)
1153                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1154                 if ((available < target) &&
1155                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1156                     !sysctl_tcp_low_latency &&
1157                     __get_cpu_var(softnet_data).net_dma) {
1158                         preempt_enable_no_resched();
1159                         tp->ucopy.pinned_list =
1160                                         dma_pin_iovec_pages(msg->msg_iov, len);
1161                 } else {
1162                         preempt_enable_no_resched();
1163                 }
1164         }
1165 #endif
1166
1167         do {
1168                 u32 offset;
1169
1170                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1171                 if (tp->urg_data && tp->urg_seq == *seq) {
1172                         if (copied)
1173                                 break;
1174                         if (signal_pending(current)) {
1175                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1176                                 break;
1177                         }
1178                 }
1179
1180                 /* Next get a buffer. */
1181
1182                 skb = skb_peek(&sk->sk_receive_queue);
1183                 do {
1184                         if (!skb)
1185                                 break;
1186
1187                         /* Now that we have two receive queues this
1188                          * shouldn't happen.
1189                          */
1190                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1191                                 printk(KERN_INFO "recvmsg bug: copied %X "
1192                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1193                                 break;
1194                         }
1195                         offset = *seq - TCP_SKB_CB(skb)->seq;
1196                         if (tcp_hdr(skb)->syn)
1197                                 offset--;
1198                         if (offset < skb->len)
1199                                 goto found_ok_skb;
1200                         if (tcp_hdr(skb)->fin)
1201                                 goto found_fin_ok;
1202                         BUG_TRAP(flags & MSG_PEEK);
1203                         skb = skb->next;
1204                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1205
1206                 /* Well, if we have backlog, try to process it now yet. */
1207
1208                 if (copied >= target && !sk->sk_backlog.tail)
1209                         break;
1210
1211                 if (copied) {
1212                         if (sk->sk_err ||
1213                             sk->sk_state == TCP_CLOSE ||
1214                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1215                             !timeo ||
1216                             signal_pending(current) ||
1217                             (flags & MSG_PEEK))
1218                                 break;
1219                 } else {
1220                         if (sock_flag(sk, SOCK_DONE))
1221                                 break;
1222
1223                         if (sk->sk_err) {
1224                                 copied = sock_error(sk);
1225                                 break;
1226                         }
1227
1228                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1229                                 break;
1230
1231                         if (sk->sk_state == TCP_CLOSE) {
1232                                 if (!sock_flag(sk, SOCK_DONE)) {
1233                                         /* This occurs when user tries to read
1234                                          * from never connected socket.
1235                                          */
1236                                         copied = -ENOTCONN;
1237                                         break;
1238                                 }
1239                                 break;
1240                         }
1241
1242                         if (!timeo) {
1243                                 copied = -EAGAIN;
1244                                 break;
1245                         }
1246
1247                         if (signal_pending(current)) {
1248                                 copied = sock_intr_errno(timeo);
1249                                 break;
1250                         }
1251                 }
1252
1253                 tcp_cleanup_rbuf(sk, copied);
1254
1255                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1256                         /* Install new reader */
1257                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1258                                 user_recv = current;
1259                                 tp->ucopy.task = user_recv;
1260                                 tp->ucopy.iov = msg->msg_iov;
1261                         }
1262
1263                         tp->ucopy.len = len;
1264
1265                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1266                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1267
1268                         /* Ugly... If prequeue is not empty, we have to
1269                          * process it before releasing socket, otherwise
1270                          * order will be broken at second iteration.
1271                          * More elegant solution is required!!!
1272                          *
1273                          * Look: we have the following (pseudo)queues:
1274                          *
1275                          * 1. packets in flight
1276                          * 2. backlog
1277                          * 3. prequeue
1278                          * 4. receive_queue
1279                          *
1280                          * Each queue can be processed only if the next ones
1281                          * are empty. At this point we have empty receive_queue.
1282                          * But prequeue _can_ be not empty after 2nd iteration,
1283                          * when we jumped to start of loop because backlog
1284                          * processing added something to receive_queue.
1285                          * We cannot release_sock(), because backlog contains
1286                          * packets arrived _after_ prequeued ones.
1287                          *
1288                          * Shortly, algorithm is clear --- to process all
1289                          * the queues in order. We could make it more directly,
1290                          * requeueing packets from backlog to prequeue, if
1291                          * is not empty. It is more elegant, but eats cycles,
1292                          * unfortunately.
1293                          */
1294                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1295                                 goto do_prequeue;
1296
1297                         /* __ Set realtime policy in scheduler __ */
1298                 }
1299
1300                 if (copied >= target) {
1301                         /* Do not sleep, just process backlog. */
1302                         release_sock(sk);
1303                         lock_sock(sk);
1304                 } else
1305                         sk_wait_data(sk, &timeo);
1306
1307 #ifdef CONFIG_NET_DMA
1308                 tp->ucopy.wakeup = 0;
1309 #endif
1310
1311                 if (user_recv) {
1312                         int chunk;
1313
1314                         /* __ Restore normal policy in scheduler __ */
1315
1316                         if ((chunk = len - tp->ucopy.len) != 0) {
1317                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1318                                 len -= chunk;
1319                                 copied += chunk;
1320                         }
1321
1322                         if (tp->rcv_nxt == tp->copied_seq &&
1323                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1324 do_prequeue:
1325                                 tcp_prequeue_process(sk);
1326
1327                                 if ((chunk = len - tp->ucopy.len) != 0) {
1328                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1329                                         len -= chunk;
1330                                         copied += chunk;
1331                                 }
1332                         }
1333                 }
1334                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1335                         if (net_ratelimit())
1336                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1337                                        current->comm, task_pid_nr(current));
1338                         peek_seq = tp->copied_seq;
1339                 }
1340                 continue;
1341
1342         found_ok_skb:
1343                 /* Ok so how much can we use? */
1344                 used = skb->len - offset;
1345                 if (len < used)
1346                         used = len;
1347
1348                 /* Do we have urgent data here? */
1349                 if (tp->urg_data) {
1350                         u32 urg_offset = tp->urg_seq - *seq;
1351                         if (urg_offset < used) {
1352                                 if (!urg_offset) {
1353                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1354                                                 ++*seq;
1355                                                 offset++;
1356                                                 used--;
1357                                                 if (!used)
1358                                                         goto skip_copy;
1359                                         }
1360                                 } else
1361                                         used = urg_offset;
1362                         }
1363                 }
1364
1365                 if (!(flags & MSG_TRUNC)) {
1366 #ifdef CONFIG_NET_DMA
1367                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1368                                 tp->ucopy.dma_chan = get_softnet_dma();
1369
1370                         if (tp->ucopy.dma_chan) {
1371                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1372                                         tp->ucopy.dma_chan, skb, offset,
1373                                         msg->msg_iov, used,
1374                                         tp->ucopy.pinned_list);
1375
1376                                 if (tp->ucopy.dma_cookie < 0) {
1377
1378                                         printk(KERN_ALERT "dma_cookie < 0\n");
1379
1380                                         /* Exception. Bailout! */
1381                                         if (!copied)
1382                                                 copied = -EFAULT;
1383                                         break;
1384                                 }
1385                                 if ((offset + used) == skb->len)
1386                                         copied_early = 1;
1387
1388                         } else
1389 #endif
1390                         {
1391                                 err = skb_copy_datagram_iovec(skb, offset,
1392                                                 msg->msg_iov, used);
1393                                 if (err) {
1394                                         /* Exception. Bailout! */
1395                                         if (!copied)
1396                                                 copied = -EFAULT;
1397                                         break;
1398                                 }
1399                         }
1400                 }
1401
1402                 *seq += used;
1403                 copied += used;
1404                 len -= used;
1405
1406                 tcp_rcv_space_adjust(sk);
1407
1408 skip_copy:
1409                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1410                         tp->urg_data = 0;
1411                         tcp_fast_path_check(sk);
1412                 }
1413                 if (used + offset < skb->len)
1414                         continue;
1415
1416                 if (tcp_hdr(skb)->fin)
1417                         goto found_fin_ok;
1418                 if (!(flags & MSG_PEEK)) {
1419                         sk_eat_skb(sk, skb, copied_early);
1420                         copied_early = 0;
1421                 }
1422                 continue;
1423
1424         found_fin_ok:
1425                 /* Process the FIN. */
1426                 ++*seq;
1427                 if (!(flags & MSG_PEEK)) {
1428                         sk_eat_skb(sk, skb, copied_early);
1429                         copied_early = 0;
1430                 }
1431                 break;
1432         } while (len > 0);
1433
1434         if (user_recv) {
1435                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1436                         int chunk;
1437
1438                         tp->ucopy.len = copied > 0 ? len : 0;
1439
1440                         tcp_prequeue_process(sk);
1441
1442                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1443                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1444                                 len -= chunk;
1445                                 copied += chunk;
1446                         }
1447                 }
1448
1449                 tp->ucopy.task = NULL;
1450                 tp->ucopy.len = 0;
1451         }
1452
1453 #ifdef CONFIG_NET_DMA
1454         if (tp->ucopy.dma_chan) {
1455                 dma_cookie_t done, used;
1456
1457                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1458
1459                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1460                                                  tp->ucopy.dma_cookie, &done,
1461                                                  &used) == DMA_IN_PROGRESS) {
1462                         /* do partial cleanup of sk_async_wait_queue */
1463                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1464                                (dma_async_is_complete(skb->dma_cookie, done,
1465                                                       used) == DMA_SUCCESS)) {
1466                                 __skb_dequeue(&sk->sk_async_wait_queue);
1467                                 kfree_skb(skb);
1468                         }
1469                 }
1470
1471                 /* Safe to free early-copied skbs now */
1472                 __skb_queue_purge(&sk->sk_async_wait_queue);
1473                 dma_chan_put(tp->ucopy.dma_chan);
1474                 tp->ucopy.dma_chan = NULL;
1475         }
1476         if (tp->ucopy.pinned_list) {
1477                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1478                 tp->ucopy.pinned_list = NULL;
1479         }
1480 #endif
1481
1482         /* According to UNIX98, msg_name/msg_namelen are ignored
1483          * on connected socket. I was just happy when found this 8) --ANK
1484          */
1485
1486         /* Clean up data we have read: This will do ACK frames. */
1487         tcp_cleanup_rbuf(sk, copied);
1488
1489         TCP_CHECK_TIMER(sk);
1490         release_sock(sk);
1491         return copied;
1492
1493 out:
1494         TCP_CHECK_TIMER(sk);
1495         release_sock(sk);
1496         return err;
1497
1498 recv_urg:
1499         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1500         goto out;
1501 }
1502
1503 /*
1504  *      State processing on a close. This implements the state shift for
1505  *      sending our FIN frame. Note that we only send a FIN for some
1506  *      states. A shutdown() may have already sent the FIN, or we may be
1507  *      closed.
1508  */
1509
1510 static const unsigned char new_state[16] = {
1511   /* current state:        new state:      action:      */
1512   /* (Invalid)          */ TCP_CLOSE,
1513   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1514   /* TCP_SYN_SENT       */ TCP_CLOSE,
1515   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1516   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1517   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1518   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1519   /* TCP_CLOSE          */ TCP_CLOSE,
1520   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1521   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1522   /* TCP_LISTEN         */ TCP_CLOSE,
1523   /* TCP_CLOSING        */ TCP_CLOSING,
1524 };
1525
1526 static int tcp_close_state(struct sock *sk)
1527 {
1528         int next = (int)new_state[sk->sk_state];
1529         int ns = next & TCP_STATE_MASK;
1530
1531         tcp_set_state(sk, ns);
1532
1533         return next & TCP_ACTION_FIN;
1534 }
1535
1536 /*
1537  *      Shutdown the sending side of a connection. Much like close except
1538  *      that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1539  */
1540
1541 void tcp_shutdown(struct sock *sk, int how)
1542 {
1543         /*      We need to grab some memory, and put together a FIN,
1544          *      and then put it into the queue to be sent.
1545          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1546          */
1547         if (!(how & SEND_SHUTDOWN))
1548                 return;
1549
1550         /* If we've already sent a FIN, or it's a closed state, skip this. */
1551         if ((1 << sk->sk_state) &
1552             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1553              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1554                 /* Clear out any half completed packets.  FIN if needed. */
1555                 if (tcp_close_state(sk))
1556                         tcp_send_fin(sk);
1557         }
1558 }
1559
1560 void tcp_close(struct sock *sk, long timeout)
1561 {
1562         struct sk_buff *skb;
1563         int data_was_unread = 0;
1564         int state;
1565
1566         lock_sock(sk);
1567         sk->sk_shutdown = SHUTDOWN_MASK;
1568
1569         if (sk->sk_state == TCP_LISTEN) {
1570                 tcp_set_state(sk, TCP_CLOSE);
1571
1572                 /* Special case. */
1573                 inet_csk_listen_stop(sk);
1574
1575                 goto adjudge_to_death;
1576         }
1577
1578         /*  We need to flush the recv. buffs.  We do this only on the
1579          *  descriptor close, not protocol-sourced closes, because the
1580          *  reader process may not have drained the data yet!
1581          */
1582         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1583                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1584                           tcp_hdr(skb)->fin;
1585                 data_was_unread += len;
1586                 __kfree_skb(skb);
1587         }
1588
1589         sk_stream_mem_reclaim(sk);
1590
1591         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1592          * data was lost. To witness the awful effects of the old behavior of
1593          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1594          * GET in an FTP client, suspend the process, wait for the client to
1595          * advertise a zero window, then kill -9 the FTP client, wheee...
1596          * Note: timeout is always zero in such a case.
1597          */
1598         if (data_was_unread) {
1599                 /* Unread data was tossed, zap the connection. */
1600                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1601                 tcp_set_state(sk, TCP_CLOSE);
1602                 tcp_send_active_reset(sk, GFP_KERNEL);
1603         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1604                 /* Check zero linger _after_ checking for unread data. */
1605                 sk->sk_prot->disconnect(sk, 0);
1606                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1607         } else if (tcp_close_state(sk)) {
1608                 /* We FIN if the application ate all the data before
1609                  * zapping the connection.
1610                  */
1611
1612                 /* RED-PEN. Formally speaking, we have broken TCP state
1613                  * machine. State transitions:
1614                  *
1615                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1616                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1617                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1618                  *
1619                  * are legal only when FIN has been sent (i.e. in window),
1620                  * rather than queued out of window. Purists blame.
1621                  *
1622                  * F.e. "RFC state" is ESTABLISHED,
1623                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1624                  *
1625                  * The visible declinations are that sometimes
1626                  * we enter time-wait state, when it is not required really
1627                  * (harmless), do not send active resets, when they are
1628                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1629                  * they look as CLOSING or LAST_ACK for Linux)
1630                  * Probably, I missed some more holelets.
1631                  *                                              --ANK
1632                  */
1633                 tcp_send_fin(sk);
1634         }
1635
1636         sk_stream_wait_close(sk, timeout);
1637
1638 adjudge_to_death:
1639         state = sk->sk_state;
1640         sock_hold(sk);
1641         sock_orphan(sk);
1642         atomic_inc(sk->sk_prot->orphan_count);
1643
1644         /* It is the last release_sock in its life. It will remove backlog. */
1645         release_sock(sk);
1646
1647
1648         /* Now socket is owned by kernel and we acquire BH lock
1649            to finish close. No need to check for user refs.
1650          */
1651         local_bh_disable();
1652         bh_lock_sock(sk);
1653         BUG_TRAP(!sock_owned_by_user(sk));
1654
1655         /* Have we already been destroyed by a softirq or backlog? */
1656         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1657                 goto out;
1658
1659         /*      This is a (useful) BSD violating of the RFC. There is a
1660          *      problem with TCP as specified in that the other end could
1661          *      keep a socket open forever with no application left this end.
1662          *      We use a 3 minute timeout (about the same as BSD) then kill
1663          *      our end. If they send after that then tough - BUT: long enough
1664          *      that we won't make the old 4*rto = almost no time - whoops
1665          *      reset mistake.
1666          *
1667          *      Nope, it was not mistake. It is really desired behaviour
1668          *      f.e. on http servers, when such sockets are useless, but
1669          *      consume significant resources. Let's do it with special
1670          *      linger2 option.                                 --ANK
1671          */
1672
1673         if (sk->sk_state == TCP_FIN_WAIT2) {
1674                 struct tcp_sock *tp = tcp_sk(sk);
1675                 if (tp->linger2 < 0) {
1676                         tcp_set_state(sk, TCP_CLOSE);
1677                         tcp_send_active_reset(sk, GFP_ATOMIC);
1678                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1679                 } else {
1680                         const int tmo = tcp_fin_time(sk);
1681
1682                         if (tmo > TCP_TIMEWAIT_LEN) {
1683                                 inet_csk_reset_keepalive_timer(sk,
1684                                                 tmo - TCP_TIMEWAIT_LEN);
1685                         } else {
1686                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1687                                 goto out;
1688                         }
1689                 }
1690         }
1691         if (sk->sk_state != TCP_CLOSE) {
1692                 sk_stream_mem_reclaim(sk);
1693                 if (tcp_too_many_orphans(sk,
1694                                 atomic_read(sk->sk_prot->orphan_count))) {
1695                         if (net_ratelimit())
1696                                 printk(KERN_INFO "TCP: too many of orphaned "
1697                                        "sockets\n");
1698                         tcp_set_state(sk, TCP_CLOSE);
1699                         tcp_send_active_reset(sk, GFP_ATOMIC);
1700                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1701                 }
1702         }
1703
1704         if (sk->sk_state == TCP_CLOSE)
1705                 inet_csk_destroy_sock(sk);
1706         /* Otherwise, socket is reprieved until protocol close. */
1707
1708 out:
1709         bh_unlock_sock(sk);
1710         local_bh_enable();
1711         sock_put(sk);
1712 }
1713
1714 /* These states need RST on ABORT according to RFC793 */
1715
1716 static inline int tcp_need_reset(int state)
1717 {
1718         return (1 << state) &
1719                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1720                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1721 }
1722
1723 int tcp_disconnect(struct sock *sk, int flags)
1724 {
1725         struct inet_sock *inet = inet_sk(sk);
1726         struct inet_connection_sock *icsk = inet_csk(sk);
1727         struct tcp_sock *tp = tcp_sk(sk);
1728         int err = 0;
1729         int old_state = sk->sk_state;
1730
1731         if (old_state != TCP_CLOSE)
1732                 tcp_set_state(sk, TCP_CLOSE);
1733
1734         /* ABORT function of RFC793 */
1735         if (old_state == TCP_LISTEN) {
1736                 inet_csk_listen_stop(sk);
1737         } else if (tcp_need_reset(old_state) ||
1738                    (tp->snd_nxt != tp->write_seq &&
1739                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1740                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1741                  * states
1742                  */
1743                 tcp_send_active_reset(sk, gfp_any());
1744                 sk->sk_err = ECONNRESET;
1745         } else if (old_state == TCP_SYN_SENT)
1746                 sk->sk_err = ECONNRESET;
1747
1748         tcp_clear_xmit_timers(sk);
1749         __skb_queue_purge(&sk->sk_receive_queue);
1750         tcp_write_queue_purge(sk);
1751         __skb_queue_purge(&tp->out_of_order_queue);
1752 #ifdef CONFIG_NET_DMA
1753         __skb_queue_purge(&sk->sk_async_wait_queue);
1754 #endif
1755
1756         inet->dport = 0;
1757
1758         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1759                 inet_reset_saddr(sk);
1760
1761         sk->sk_shutdown = 0;
1762         sock_reset_flag(sk, SOCK_DONE);
1763         tp->srtt = 0;
1764         if ((tp->write_seq += tp->max_window + 2) == 0)
1765                 tp->write_seq = 1;
1766         icsk->icsk_backoff = 0;
1767         tp->snd_cwnd = 2;
1768         icsk->icsk_probes_out = 0;
1769         tp->packets_out = 0;
1770         tp->snd_ssthresh = 0x7fffffff;
1771         tp->snd_cwnd_cnt = 0;
1772         tp->bytes_acked = 0;
1773         tcp_set_ca_state(sk, TCP_CA_Open);
1774         tcp_clear_retrans(tp);
1775         inet_csk_delack_init(sk);
1776         tcp_init_send_head(sk);
1777         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1778         __sk_dst_reset(sk);
1779
1780         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1781
1782         sk->sk_error_report(sk);
1783         return err;
1784 }
1785
1786 /*
1787  *      Socket option code for TCP.
1788  */
1789 static int do_tcp_setsockopt(struct sock *sk, int level,
1790                 int optname, char __user *optval, int optlen)
1791 {
1792         struct tcp_sock *tp = tcp_sk(sk);
1793         struct inet_connection_sock *icsk = inet_csk(sk);
1794         int val;
1795         int err = 0;
1796
1797         /* This is a string value all the others are int's */
1798         if (optname == TCP_CONGESTION) {
1799                 char name[TCP_CA_NAME_MAX];
1800
1801                 if (optlen < 1)
1802                         return -EINVAL;
1803
1804                 val = strncpy_from_user(name, optval,
1805                                         min(TCP_CA_NAME_MAX-1, optlen));
1806                 if (val < 0)
1807                         return -EFAULT;
1808                 name[val] = 0;
1809
1810                 lock_sock(sk);
1811                 err = tcp_set_congestion_control(sk, name);
1812                 release_sock(sk);
1813                 return err;
1814         }
1815
1816         if (optlen < sizeof(int))
1817                 return -EINVAL;
1818
1819         if (get_user(val, (int __user *)optval))
1820                 return -EFAULT;
1821
1822         lock_sock(sk);
1823
1824         switch (optname) {
1825         case TCP_MAXSEG:
1826                 /* Values greater than interface MTU won't take effect. However
1827                  * at the point when this call is done we typically don't yet
1828                  * know which interface is going to be used */
1829                 if (val < 8 || val > MAX_TCP_WINDOW) {
1830                         err = -EINVAL;
1831                         break;
1832                 }
1833                 tp->rx_opt.user_mss = val;
1834                 break;
1835
1836         case TCP_NODELAY:
1837                 if (val) {
1838                         /* TCP_NODELAY is weaker than TCP_CORK, so that
1839                          * this option on corked socket is remembered, but
1840                          * it is not activated until cork is cleared.
1841                          *
1842                          * However, when TCP_NODELAY is set we make
1843                          * an explicit push, which overrides even TCP_CORK
1844                          * for currently queued segments.
1845                          */
1846                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1847                         tcp_push_pending_frames(sk);
1848                 } else {
1849                         tp->nonagle &= ~TCP_NAGLE_OFF;
1850                 }
1851                 break;
1852
1853         case TCP_CORK:
1854                 /* When set indicates to always queue non-full frames.
1855                  * Later the user clears this option and we transmit
1856                  * any pending partial frames in the queue.  This is
1857                  * meant to be used alongside sendfile() to get properly
1858                  * filled frames when the user (for example) must write
1859                  * out headers with a write() call first and then use
1860                  * sendfile to send out the data parts.
1861                  *
1862                  * TCP_CORK can be set together with TCP_NODELAY and it is
1863                  * stronger than TCP_NODELAY.
1864                  */
1865                 if (val) {
1866                         tp->nonagle |= TCP_NAGLE_CORK;
1867                 } else {
1868                         tp->nonagle &= ~TCP_NAGLE_CORK;
1869                         if (tp->nonagle&TCP_NAGLE_OFF)
1870                                 tp->nonagle |= TCP_NAGLE_PUSH;
1871                         tcp_push_pending_frames(sk);
1872                 }
1873                 break;
1874
1875         case TCP_KEEPIDLE:
1876                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1877                         err = -EINVAL;
1878                 else {
1879                         tp->keepalive_time = val * HZ;
1880                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
1881                             !((1 << sk->sk_state) &
1882                               (TCPF_CLOSE | TCPF_LISTEN))) {
1883                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1884                                 if (tp->keepalive_time > elapsed)
1885                                         elapsed = tp->keepalive_time - elapsed;
1886                                 else
1887                                         elapsed = 0;
1888                                 inet_csk_reset_keepalive_timer(sk, elapsed);
1889                         }
1890                 }
1891                 break;
1892         case TCP_KEEPINTVL:
1893                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1894                         err = -EINVAL;
1895                 else
1896                         tp->keepalive_intvl = val * HZ;
1897                 break;
1898         case TCP_KEEPCNT:
1899                 if (val < 1 || val > MAX_TCP_KEEPCNT)
1900                         err = -EINVAL;
1901                 else
1902                         tp->keepalive_probes = val;
1903                 break;
1904         case TCP_SYNCNT:
1905                 if (val < 1 || val > MAX_TCP_SYNCNT)
1906                         err = -EINVAL;
1907                 else
1908                         icsk->icsk_syn_retries = val;
1909                 break;
1910
1911         case TCP_LINGER2:
1912                 if (val < 0)
1913                         tp->linger2 = -1;
1914                 else if (val > sysctl_tcp_fin_timeout / HZ)
1915                         tp->linger2 = 0;
1916                 else
1917                         tp->linger2 = val * HZ;
1918                 break;
1919
1920         case TCP_DEFER_ACCEPT:
1921                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1922                 if (val > 0) {
1923                         /* Translate value in seconds to number of
1924                          * retransmits */
1925                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1926                                val > ((TCP_TIMEOUT_INIT / HZ) <<
1927                                        icsk->icsk_accept_queue.rskq_defer_accept))
1928                                 icsk->icsk_accept_queue.rskq_defer_accept++;
1929                         icsk->icsk_accept_queue.rskq_defer_accept++;
1930                 }
1931                 break;
1932
1933         case TCP_WINDOW_CLAMP:
1934                 if (!val) {
1935                         if (sk->sk_state != TCP_CLOSE) {
1936                                 err = -EINVAL;
1937                                 break;
1938                         }
1939                         tp->window_clamp = 0;
1940                 } else
1941                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1942                                                 SOCK_MIN_RCVBUF / 2 : val;
1943                 break;
1944
1945         case TCP_QUICKACK:
1946                 if (!val) {
1947                         icsk->icsk_ack.pingpong = 1;
1948                 } else {
1949                         icsk->icsk_ack.pingpong = 0;
1950                         if ((1 << sk->sk_state) &
1951                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1952                             inet_csk_ack_scheduled(sk)) {
1953                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1954                                 tcp_cleanup_rbuf(sk, 1);
1955                                 if (!(val & 1))
1956                                         icsk->icsk_ack.pingpong = 1;
1957                         }
1958                 }
1959                 break;
1960
1961 #ifdef CONFIG_TCP_MD5SIG
1962         case TCP_MD5SIG:
1963                 /* Read the IP->Key mappings from userspace */
1964                 err = tp->af_specific->md5_parse(sk, optval, optlen);
1965                 break;
1966 #endif
1967
1968         default:
1969                 err = -ENOPROTOOPT;
1970                 break;
1971         }
1972
1973         release_sock(sk);
1974         return err;
1975 }
1976
1977 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1978                    int optlen)
1979 {
1980         struct inet_connection_sock *icsk = inet_csk(sk);
1981
1982         if (level != SOL_TCP)
1983                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1984                                                      optval, optlen);
1985         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1986 }
1987
1988 #ifdef CONFIG_COMPAT
1989 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1990                           char __user *optval, int optlen)
1991 {
1992         if (level != SOL_TCP)
1993                 return inet_csk_compat_setsockopt(sk, level, optname,
1994                                                   optval, optlen);
1995         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1996 }
1997
1998 EXPORT_SYMBOL(compat_tcp_setsockopt);
1999 #endif
2000
2001 /* Return information about state of tcp endpoint in API format. */
2002 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2003 {
2004         struct tcp_sock *tp = tcp_sk(sk);
2005         const struct inet_connection_sock *icsk = inet_csk(sk);
2006         u32 now = tcp_time_stamp;
2007
2008         memset(info, 0, sizeof(*info));
2009
2010         info->tcpi_state = sk->sk_state;
2011         info->tcpi_ca_state = icsk->icsk_ca_state;
2012         info->tcpi_retransmits = icsk->icsk_retransmits;
2013         info->tcpi_probes = icsk->icsk_probes_out;
2014         info->tcpi_backoff = icsk->icsk_backoff;
2015
2016         if (tp->rx_opt.tstamp_ok)
2017                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2018         if (tcp_is_sack(tp))
2019                 info->tcpi_options |= TCPI_OPT_SACK;
2020         if (tp->rx_opt.wscale_ok) {
2021                 info->tcpi_options |= TCPI_OPT_WSCALE;
2022                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2023                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2024         }
2025
2026         if (tp->ecn_flags&TCP_ECN_OK)
2027                 info->tcpi_options |= TCPI_OPT_ECN;
2028
2029         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2030         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2031         info->tcpi_snd_mss = tp->mss_cache;
2032         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2033
2034         if (sk->sk_state == TCP_LISTEN) {
2035                 info->tcpi_unacked = sk->sk_ack_backlog;
2036                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2037         } else {
2038                 info->tcpi_unacked = tp->packets_out;
2039                 info->tcpi_sacked = tp->sacked_out;
2040         }
2041         info->tcpi_lost = tp->lost_out;
2042         info->tcpi_retrans = tp->retrans_out;
2043         info->tcpi_fackets = tp->fackets_out;
2044
2045         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2046         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2047         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2048
2049         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2050         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2051         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2052         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2053         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2054         info->tcpi_snd_cwnd = tp->snd_cwnd;
2055         info->tcpi_advmss = tp->advmss;
2056         info->tcpi_reordering = tp->reordering;
2057
2058         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2059         info->tcpi_rcv_space = tp->rcvq_space.space;
2060
2061         info->tcpi_total_retrans = tp->total_retrans;
2062 }
2063
2064 EXPORT_SYMBOL_GPL(tcp_get_info);
2065
2066 static int do_tcp_getsockopt(struct sock *sk, int level,
2067                 int optname, char __user *optval, int __user *optlen)
2068 {
2069         struct inet_connection_sock *icsk = inet_csk(sk);
2070         struct tcp_sock *tp = tcp_sk(sk);
2071         int val, len;
2072
2073         if (get_user(len, optlen))
2074                 return -EFAULT;
2075
2076         len = min_t(unsigned int, len, sizeof(int));
2077
2078         if (len < 0)
2079                 return -EINVAL;
2080
2081         switch (optname) {
2082         case TCP_MAXSEG:
2083                 val = tp->mss_cache;
2084                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2085                         val = tp->rx_opt.user_mss;
2086                 break;
2087         case TCP_NODELAY:
2088                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2089                 break;
2090         case TCP_CORK:
2091                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2092                 break;
2093         case TCP_KEEPIDLE:
2094                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2095                 break;
2096         case TCP_KEEPINTVL:
2097                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2098                 break;
2099         case TCP_KEEPCNT:
2100                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2101                 break;
2102         case TCP_SYNCNT:
2103                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2104                 break;
2105         case TCP_LINGER2:
2106                 val = tp->linger2;
2107                 if (val >= 0)
2108                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2109                 break;
2110         case TCP_DEFER_ACCEPT:
2111                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2112                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2113                 break;
2114         case TCP_WINDOW_CLAMP:
2115                 val = tp->window_clamp;
2116                 break;
2117         case TCP_INFO: {
2118                 struct tcp_info info;
2119
2120                 if (get_user(len, optlen))
2121                         return -EFAULT;
2122
2123                 tcp_get_info(sk, &info);
2124
2125                 len = min_t(unsigned int, len, sizeof(info));
2126                 if (put_user(len, optlen))
2127                         return -EFAULT;
2128                 if (copy_to_user(optval, &info, len))
2129                         return -EFAULT;
2130                 return 0;
2131         }
2132         case TCP_QUICKACK:
2133                 val = !icsk->icsk_ack.pingpong;
2134                 break;
2135
2136         case TCP_CONGESTION:
2137                 if (get_user(len, optlen))
2138                         return -EFAULT;
2139                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2140                 if (put_user(len, optlen))
2141                         return -EFAULT;
2142                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2143                         return -EFAULT;
2144                 return 0;
2145         default:
2146                 return -ENOPROTOOPT;
2147         }
2148
2149         if (put_user(len, optlen))
2150                 return -EFAULT;
2151         if (copy_to_user(optval, &val, len))
2152                 return -EFAULT;
2153         return 0;
2154 }
2155
2156 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2157                    int __user *optlen)
2158 {
2159         struct inet_connection_sock *icsk = inet_csk(sk);
2160
2161         if (level != SOL_TCP)
2162                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2163                                                      optval, optlen);
2164         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2165 }
2166
2167 #ifdef CONFIG_COMPAT
2168 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2169                           char __user *optval, int __user *optlen)
2170 {
2171         if (level != SOL_TCP)
2172                 return inet_csk_compat_getsockopt(sk, level, optname,
2173                                                   optval, optlen);
2174         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2175 }
2176
2177 EXPORT_SYMBOL(compat_tcp_getsockopt);
2178 #endif
2179
2180 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2181 {
2182         struct sk_buff *segs = ERR_PTR(-EINVAL);
2183         struct tcphdr *th;
2184         unsigned thlen;
2185         unsigned int seq;
2186         __be32 delta;
2187         unsigned int oldlen;
2188         unsigned int len;
2189
2190         if (!pskb_may_pull(skb, sizeof(*th)))
2191                 goto out;
2192
2193         th = tcp_hdr(skb);
2194         thlen = th->doff * 4;
2195         if (thlen < sizeof(*th))
2196                 goto out;
2197
2198         if (!pskb_may_pull(skb, thlen))
2199                 goto out;
2200
2201         oldlen = (u16)~skb->len;
2202         __skb_pull(skb, thlen);
2203
2204         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2205                 /* Packet is from an untrusted source, reset gso_segs. */
2206                 int type = skb_shinfo(skb)->gso_type;
2207                 int mss;
2208
2209                 if (unlikely(type &
2210                              ~(SKB_GSO_TCPV4 |
2211                                SKB_GSO_DODGY |
2212                                SKB_GSO_TCP_ECN |
2213                                SKB_GSO_TCPV6 |
2214                                0) ||
2215                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2216                         goto out;
2217
2218                 mss = skb_shinfo(skb)->gso_size;
2219                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2220
2221                 segs = NULL;
2222                 goto out;
2223         }
2224
2225         segs = skb_segment(skb, features);
2226         if (IS_ERR(segs))
2227                 goto out;
2228
2229         len = skb_shinfo(skb)->gso_size;
2230         delta = htonl(oldlen + (thlen + len));
2231
2232         skb = segs;
2233         th = tcp_hdr(skb);
2234         seq = ntohl(th->seq);
2235
2236         do {
2237                 th->fin = th->psh = 0;
2238
2239                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2240                                        (__force u32)delta));
2241                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2242                         th->check =
2243                              csum_fold(csum_partial(skb_transport_header(skb),
2244                                                     thlen, skb->csum));
2245
2246                 seq += len;
2247                 skb = skb->next;
2248                 th = tcp_hdr(skb);
2249
2250                 th->seq = htonl(seq);
2251                 th->cwr = 0;
2252         } while (skb->next);
2253
2254         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2255                       skb->data_len);
2256         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2257                                 (__force u32)delta));
2258         if (skb->ip_summed != CHECKSUM_PARTIAL)
2259                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2260                                                    thlen, skb->csum));
2261
2262 out:
2263         return segs;
2264 }
2265 EXPORT_SYMBOL(tcp_tso_segment);
2266
2267 #ifdef CONFIG_TCP_MD5SIG
2268 static unsigned long tcp_md5sig_users;
2269 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2270 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2271
2272 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2273 {
2274         int cpu;
2275         for_each_possible_cpu(cpu) {
2276                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2277                 if (p) {
2278                         if (p->md5_desc.tfm)
2279                                 crypto_free_hash(p->md5_desc.tfm);
2280                         kfree(p);
2281                         p = NULL;
2282                 }
2283         }
2284         free_percpu(pool);
2285 }
2286
2287 void tcp_free_md5sig_pool(void)
2288 {
2289         struct tcp_md5sig_pool **pool = NULL;
2290
2291         spin_lock_bh(&tcp_md5sig_pool_lock);
2292         if (--tcp_md5sig_users == 0) {
2293                 pool = tcp_md5sig_pool;
2294                 tcp_md5sig_pool = NULL;
2295         }
2296         spin_unlock_bh(&tcp_md5sig_pool_lock);
2297         if (pool)
2298                 __tcp_free_md5sig_pool(pool);
2299 }
2300
2301 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2302
2303 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2304 {
2305         int cpu;
2306         struct tcp_md5sig_pool **pool;
2307
2308         pool = alloc_percpu(struct tcp_md5sig_pool *);
2309         if (!pool)
2310                 return NULL;
2311
2312         for_each_possible_cpu(cpu) {
2313                 struct tcp_md5sig_pool *p;
2314                 struct crypto_hash *hash;
2315
2316                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2317                 if (!p)
2318                         goto out_free;
2319                 *per_cpu_ptr(pool, cpu) = p;
2320
2321                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2322                 if (!hash || IS_ERR(hash))
2323                         goto out_free;
2324
2325                 p->md5_desc.tfm = hash;
2326         }
2327         return pool;
2328 out_free:
2329         __tcp_free_md5sig_pool(pool);
2330         return NULL;
2331 }
2332
2333 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2334 {
2335         struct tcp_md5sig_pool **pool;
2336         int alloc = 0;
2337
2338 retry:
2339         spin_lock_bh(&tcp_md5sig_pool_lock);
2340         pool = tcp_md5sig_pool;
2341         if (tcp_md5sig_users++ == 0) {
2342                 alloc = 1;
2343                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2344         } else if (!pool) {
2345                 tcp_md5sig_users--;
2346                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2347                 cpu_relax();
2348                 goto retry;
2349         } else
2350                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2351
2352         if (alloc) {
2353                 /* we cannot hold spinlock here because this may sleep. */
2354                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2355                 spin_lock_bh(&tcp_md5sig_pool_lock);
2356                 if (!p) {
2357                         tcp_md5sig_users--;
2358                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2359                         return NULL;
2360                 }
2361                 pool = tcp_md5sig_pool;
2362                 if (pool) {
2363                         /* oops, it has already been assigned. */
2364                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2365                         __tcp_free_md5sig_pool(p);
2366                 } else {
2367                         tcp_md5sig_pool = pool = p;
2368                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2369                 }
2370         }
2371         return pool;
2372 }
2373
2374 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2375
2376 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2377 {
2378         struct tcp_md5sig_pool **p;
2379         spin_lock_bh(&tcp_md5sig_pool_lock);
2380         p = tcp_md5sig_pool;
2381         if (p)
2382                 tcp_md5sig_users++;
2383         spin_unlock_bh(&tcp_md5sig_pool_lock);
2384         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2385 }
2386
2387 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2388
2389 void __tcp_put_md5sig_pool(void)
2390 {
2391         tcp_free_md5sig_pool();
2392 }
2393
2394 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2395 #endif
2396
2397 void tcp_done(struct sock *sk)
2398 {
2399         if(sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2400                 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
2401
2402         tcp_set_state(sk, TCP_CLOSE);
2403         tcp_clear_xmit_timers(sk);
2404
2405         sk->sk_shutdown = SHUTDOWN_MASK;
2406
2407         if (!sock_flag(sk, SOCK_DEAD))
2408                 sk->sk_state_change(sk);
2409         else
2410                 inet_csk_destroy_sock(sk);
2411 }
2412 EXPORT_SYMBOL_GPL(tcp_done);
2413
2414 extern void __skb_cb_too_small_for_tcp(int, int);
2415 extern struct tcp_congestion_ops tcp_reno;
2416
2417 static __initdata unsigned long thash_entries;
2418 static int __init set_thash_entries(char *str)
2419 {
2420         if (!str)
2421                 return 0;
2422         thash_entries = simple_strtoul(str, &str, 0);
2423         return 1;
2424 }
2425 __setup("thash_entries=", set_thash_entries);
2426
2427 void __init tcp_init(void)
2428 {
2429         struct sk_buff *skb = NULL;
2430         unsigned long limit;
2431         int order, i, max_share;
2432
2433         if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2434                 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2435                                            sizeof(skb->cb));
2436
2437         tcp_hashinfo.bind_bucket_cachep =
2438                 kmem_cache_create("tcp_bind_bucket",
2439                                   sizeof(struct inet_bind_bucket), 0,
2440                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2441
2442         /* Size and allocate the main established and bind bucket
2443          * hash tables.
2444          *
2445          * The methodology is similar to that of the buffer cache.
2446          */
2447         tcp_hashinfo.ehash =
2448                 alloc_large_system_hash("TCP established",
2449                                         sizeof(struct inet_ehash_bucket),
2450                                         thash_entries,
2451                                         (num_physpages >= 128 * 1024) ?
2452                                         13 : 15,
2453                                         0,
2454                                         &tcp_hashinfo.ehash_size,
2455                                         NULL,
2456                                         thash_entries ? 0 : 512 * 1024);
2457         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2458         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2459                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2460                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
2461         }
2462         if (inet_ehash_locks_alloc(&tcp_hashinfo))
2463                 panic("TCP: failed to alloc ehash_locks");
2464         tcp_hashinfo.bhash =
2465                 alloc_large_system_hash("TCP bind",
2466                                         sizeof(struct inet_bind_hashbucket),
2467                                         tcp_hashinfo.ehash_size,
2468                                         (num_physpages >= 128 * 1024) ?
2469                                         13 : 15,
2470                                         0,
2471                                         &tcp_hashinfo.bhash_size,
2472                                         NULL,
2473                                         64 * 1024);
2474         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2475         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2476                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2477                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2478         }
2479
2480         /* Try to be a bit smarter and adjust defaults depending
2481          * on available memory.
2482          */
2483         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2484                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2485                         order++)
2486                 ;
2487         if (order >= 4) {
2488                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2489                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2490                 sysctl_max_syn_backlog = 1024;
2491         } else if (order < 3) {
2492                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2493                 sysctl_tcp_max_orphans >>= (3 - order);
2494                 sysctl_max_syn_backlog = 128;
2495         }
2496
2497         /* Set the pressure threshold to be a fraction of global memory that
2498          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2499          * memory, with a floor of 128 pages.
2500          */
2501         limit = min(nr_all_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2502         limit = (limit * (nr_all_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2503         limit = max(limit, 128UL);
2504         sysctl_tcp_mem[0] = limit / 4 * 3;
2505         sysctl_tcp_mem[1] = limit;
2506         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2507
2508         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2509         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2510         max_share = min(4UL*1024*1024, limit);
2511
2512         sysctl_tcp_wmem[0] = SK_STREAM_MEM_QUANTUM;
2513         sysctl_tcp_wmem[1] = 16*1024;
2514         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2515
2516         sysctl_tcp_rmem[0] = SK_STREAM_MEM_QUANTUM;
2517         sysctl_tcp_rmem[1] = 87380;
2518         sysctl_tcp_rmem[2] = max(87380, max_share);
2519
2520         printk(KERN_INFO "TCP: Hash tables configured "
2521                "(established %d bind %d)\n",
2522                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2523
2524         tcp_register_congestion_control(&tcp_reno);
2525 }
2526
2527 EXPORT_SYMBOL(tcp_close);
2528 EXPORT_SYMBOL(tcp_disconnect);
2529 EXPORT_SYMBOL(tcp_getsockopt);
2530 EXPORT_SYMBOL(tcp_ioctl);
2531 EXPORT_SYMBOL(tcp_poll);
2532 EXPORT_SYMBOL(tcp_read_sock);
2533 EXPORT_SYMBOL(tcp_recvmsg);
2534 EXPORT_SYMBOL(tcp_sendmsg);
2535 EXPORT_SYMBOL(tcp_sendpage);
2536 EXPORT_SYMBOL(tcp_setsockopt);
2537 EXPORT_SYMBOL(tcp_shutdown);
2538 EXPORT_SYMBOL(tcp_statistics);