[NET]: netfilter checksum annotations
[linux-2.6] / net / ipv4 / tcp.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp.c,v 1.216 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  *
22  * Fixes:
23  *              Alan Cox        :       Numerous verify_area() calls
24  *              Alan Cox        :       Set the ACK bit on a reset
25  *              Alan Cox        :       Stopped it crashing if it closed while
26  *                                      sk->inuse=1 and was trying to connect
27  *                                      (tcp_err()).
28  *              Alan Cox        :       All icmp error handling was broken
29  *                                      pointers passed where wrong and the
30  *                                      socket was looked up backwards. Nobody
31  *                                      tested any icmp error code obviously.
32  *              Alan Cox        :       tcp_err() now handled properly. It
33  *                                      wakes people on errors. poll
34  *                                      behaves and the icmp error race
35  *                                      has gone by moving it into sock.c
36  *              Alan Cox        :       tcp_send_reset() fixed to work for
37  *                                      everything not just packets for
38  *                                      unknown sockets.
39  *              Alan Cox        :       tcp option processing.
40  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
41  *                                      syn rule wrong]
42  *              Herp Rosmanith  :       More reset fixes
43  *              Alan Cox        :       No longer acks invalid rst frames.
44  *                                      Acking any kind of RST is right out.
45  *              Alan Cox        :       Sets an ignore me flag on an rst
46  *                                      receive otherwise odd bits of prattle
47  *                                      escape still
48  *              Alan Cox        :       Fixed another acking RST frame bug.
49  *                                      Should stop LAN workplace lockups.
50  *              Alan Cox        :       Some tidyups using the new skb list
51  *                                      facilities
52  *              Alan Cox        :       sk->keepopen now seems to work
53  *              Alan Cox        :       Pulls options out correctly on accepts
54  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
55  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
56  *                                      bit to skb ops.
57  *              Alan Cox        :       Tidied tcp_data to avoid a potential
58  *                                      nasty.
59  *              Alan Cox        :       Added some better commenting, as the
60  *                                      tcp is hard to follow
61  *              Alan Cox        :       Removed incorrect check for 20 * psh
62  *      Michael O'Reilly        :       ack < copied bug fix.
63  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
64  *              Alan Cox        :       FIN with no memory -> CRASH
65  *              Alan Cox        :       Added socket option proto entries.
66  *                                      Also added awareness of them to accept.
67  *              Alan Cox        :       Added TCP options (SOL_TCP)
68  *              Alan Cox        :       Switched wakeup calls to callbacks,
69  *                                      so the kernel can layer network
70  *                                      sockets.
71  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
72  *              Alan Cox        :       Handle FIN (more) properly (we hope).
73  *              Alan Cox        :       RST frames sent on unsynchronised
74  *                                      state ack error.
75  *              Alan Cox        :       Put in missing check for SYN bit.
76  *              Alan Cox        :       Added tcp_select_window() aka NET2E
77  *                                      window non shrink trick.
78  *              Alan Cox        :       Added a couple of small NET2E timer
79  *                                      fixes
80  *              Charles Hedrick :       TCP fixes
81  *              Toomas Tamm     :       TCP window fixes
82  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
83  *              Charles Hedrick :       Rewrote most of it to actually work
84  *              Linus           :       Rewrote tcp_read() and URG handling
85  *                                      completely
86  *              Gerhard Koerting:       Fixed some missing timer handling
87  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
88  *              Gerhard Koerting:       PC/TCP workarounds
89  *              Adam Caldwell   :       Assorted timer/timing errors
90  *              Matthew Dillon  :       Fixed another RST bug
91  *              Alan Cox        :       Move to kernel side addressing changes.
92  *              Alan Cox        :       Beginning work on TCP fastpathing
93  *                                      (not yet usable)
94  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
95  *              Alan Cox        :       TCP fast path debugging
96  *              Alan Cox        :       Window clamping
97  *              Michael Riepe   :       Bug in tcp_check()
98  *              Matt Dillon     :       More TCP improvements and RST bug fixes
99  *              Matt Dillon     :       Yet more small nasties remove from the
100  *                                      TCP code (Be very nice to this man if
101  *                                      tcp finally works 100%) 8)
102  *              Alan Cox        :       BSD accept semantics.
103  *              Alan Cox        :       Reset on closedown bug.
104  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
105  *              Michael Pall    :       Handle poll() after URG properly in
106  *                                      all cases.
107  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
108  *                                      (multi URG PUSH broke rlogin).
109  *              Michael Pall    :       Fix the multi URG PUSH problem in
110  *                                      tcp_readable(), poll() after URG
111  *                                      works now.
112  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
113  *                                      BSD api.
114  *              Alan Cox        :       Changed the semantics of sk->socket to
115  *                                      fix a race and a signal problem with
116  *                                      accept() and async I/O.
117  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
118  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
119  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
120  *                                      clients/servers which listen in on
121  *                                      fixed ports.
122  *              Alan Cox        :       Cleaned the above up and shrank it to
123  *                                      a sensible code size.
124  *              Alan Cox        :       Self connect lockup fix.
125  *              Alan Cox        :       No connect to multicast.
126  *              Ross Biro       :       Close unaccepted children on master
127  *                                      socket close.
128  *              Alan Cox        :       Reset tracing code.
129  *              Alan Cox        :       Spurious resets on shutdown.
130  *              Alan Cox        :       Giant 15 minute/60 second timer error
131  *              Alan Cox        :       Small whoops in polling before an
132  *                                      accept.
133  *              Alan Cox        :       Kept the state trace facility since
134  *                                      it's handy for debugging.
135  *              Alan Cox        :       More reset handler fixes.
136  *              Alan Cox        :       Started rewriting the code based on
137  *                                      the RFC's for other useful protocol
138  *                                      references see: Comer, KA9Q NOS, and
139  *                                      for a reference on the difference
140  *                                      between specifications and how BSD
141  *                                      works see the 4.4lite source.
142  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
143  *                                      close.
144  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
145  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
146  *              Alan Cox        :       Reimplemented timers as per the RFC
147  *                                      and using multiple timers for sanity.
148  *              Alan Cox        :       Small bug fixes, and a lot of new
149  *                                      comments.
150  *              Alan Cox        :       Fixed dual reader crash by locking
151  *                                      the buffers (much like datagram.c)
152  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
153  *                                      now gets fed up of retrying without
154  *                                      (even a no space) answer.
155  *              Alan Cox        :       Extracted closing code better
156  *              Alan Cox        :       Fixed the closing state machine to
157  *                                      resemble the RFC.
158  *              Alan Cox        :       More 'per spec' fixes.
159  *              Jorge Cwik      :       Even faster checksumming.
160  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
161  *                                      only frames. At least one pc tcp stack
162  *                                      generates them.
163  *              Alan Cox        :       Cache last socket.
164  *              Alan Cox        :       Per route irtt.
165  *              Matt Day        :       poll()->select() match BSD precisely on error
166  *              Alan Cox        :       New buffers
167  *              Marc Tamsky     :       Various sk->prot->retransmits and
168  *                                      sk->retransmits misupdating fixed.
169  *                                      Fixed tcp_write_timeout: stuck close,
170  *                                      and TCP syn retries gets used now.
171  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
172  *                                      ack if state is TCP_CLOSED.
173  *              Alan Cox        :       Look up device on a retransmit - routes may
174  *                                      change. Doesn't yet cope with MSS shrink right
175  *                                      but it's a start!
176  *              Marc Tamsky     :       Closing in closing fixes.
177  *              Mike Shaver     :       RFC1122 verifications.
178  *              Alan Cox        :       rcv_saddr errors.
179  *              Alan Cox        :       Block double connect().
180  *              Alan Cox        :       Small hooks for enSKIP.
181  *              Alexey Kuznetsov:       Path MTU discovery.
182  *              Alan Cox        :       Support soft errors.
183  *              Alan Cox        :       Fix MTU discovery pathological case
184  *                                      when the remote claims no mtu!
185  *              Marc Tamsky     :       TCP_CLOSE fix.
186  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
187  *                                      window but wrong (fixes NT lpd problems)
188  *              Pedro Roque     :       Better TCP window handling, delayed ack.
189  *              Joerg Reuter    :       No modification of locked buffers in
190  *                                      tcp_do_retransmit()
191  *              Eric Schenk     :       Changed receiver side silly window
192  *                                      avoidance algorithm to BSD style
193  *                                      algorithm. This doubles throughput
194  *                                      against machines running Solaris,
195  *                                      and seems to result in general
196  *                                      improvement.
197  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
198  *      Willy Konynenberg       :       Transparent proxying support.
199  *      Mike McLagan            :       Routing by source
200  *              Keith Owens     :       Do proper merging with partial SKB's in
201  *                                      tcp_do_sendmsg to avoid burstiness.
202  *              Eric Schenk     :       Fix fast close down bug with
203  *                                      shutdown() followed by close().
204  *              Andi Kleen      :       Make poll agree with SIGIO
205  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
206  *                                      lingertime == 0 (RFC 793 ABORT Call)
207  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
208  *                                      csum_and_copy_from_user() if possible.
209  *
210  *              This program is free software; you can redistribute it and/or
211  *              modify it under the terms of the GNU General Public License
212  *              as published by the Free Software Foundation; either version
213  *              2 of the License, or(at your option) any later version.
214  *
215  * Description of States:
216  *
217  *      TCP_SYN_SENT            sent a connection request, waiting for ack
218  *
219  *      TCP_SYN_RECV            received a connection request, sent ack,
220  *                              waiting for final ack in three-way handshake.
221  *
222  *      TCP_ESTABLISHED         connection established
223  *
224  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
225  *                              transmission of remaining buffered data
226  *
227  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
228  *                              to shutdown
229  *
230  *      TCP_CLOSING             both sides have shutdown but we still have
231  *                              data we have to finish sending
232  *
233  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
234  *                              closed, can only be entered from FIN_WAIT2
235  *                              or CLOSING.  Required because the other end
236  *                              may not have gotten our last ACK causing it
237  *                              to retransmit the data packet (which we ignore)
238  *
239  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
240  *                              us to finish writing our data and to shutdown
241  *                              (we have to close() to move on to LAST_ACK)
242  *
243  *      TCP_LAST_ACK            out side has shutdown after remote has
244  *                              shutdown.  There may still be data in our
245  *                              buffer that we have to finish sending
246  *
247  *      TCP_CLOSE               socket is finished
248  */
249
250 #include <linux/module.h>
251 #include <linux/types.h>
252 #include <linux/fcntl.h>
253 #include <linux/poll.h>
254 #include <linux/init.h>
255 #include <linux/smp_lock.h>
256 #include <linux/fs.h>
257 #include <linux/random.h>
258 #include <linux/bootmem.h>
259 #include <linux/cache.h>
260 #include <linux/err.h>
261 #include <linux/crypto.h>
262
263 #include <net/icmp.h>
264 #include <net/tcp.h>
265 #include <net/xfrm.h>
266 #include <net/ip.h>
267 #include <net/netdma.h>
268
269 #include <asm/uaccess.h>
270 #include <asm/ioctls.h>
271
272 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
273
274 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
275
276 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
277
278 EXPORT_SYMBOL_GPL(tcp_orphan_count);
279
280 int sysctl_tcp_mem[3] __read_mostly;
281 int sysctl_tcp_wmem[3] __read_mostly;
282 int sysctl_tcp_rmem[3] __read_mostly;
283
284 EXPORT_SYMBOL(sysctl_tcp_mem);
285 EXPORT_SYMBOL(sysctl_tcp_rmem);
286 EXPORT_SYMBOL(sysctl_tcp_wmem);
287
288 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
289 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
290
291 EXPORT_SYMBOL(tcp_memory_allocated);
292 EXPORT_SYMBOL(tcp_sockets_allocated);
293
294 /*
295  * Pressure flag: try to collapse.
296  * Technical note: it is used by multiple contexts non atomically.
297  * All the sk_stream_mem_schedule() is of this nature: accounting
298  * is strict, actions are advisory and have some latency.
299  */
300 int tcp_memory_pressure;
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                        ((struct sk_buff *)sk->sk_receive_queue.prev)->h.th->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 tcp_sock *tp,
464                               struct sk_buff *skb)
465 {
466         skb->csum = 0;
467         TCP_SKB_CB(skb)->seq = tp->write_seq;
468         TCP_SKB_CB(skb)->end_seq = tp->write_seq;
469         TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
470         TCP_SKB_CB(skb)->sacked = 0;
471         skb_header_release(skb);
472         __skb_queue_tail(&sk->sk_write_queue, skb);
473         sk_charge_skb(sk, skb);
474         if (!sk->sk_send_head)
475                 sk->sk_send_head = skb;
476         if (tp->nonagle & TCP_NAGLE_PUSH)
477                 tp->nonagle &= ~TCP_NAGLE_PUSH; 
478 }
479
480 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
481                                 struct sk_buff *skb)
482 {
483         if (flags & MSG_OOB) {
484                 tp->urg_mode = 1;
485                 tp->snd_up = tp->write_seq;
486                 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
487         }
488 }
489
490 static inline void tcp_push(struct sock *sk, struct tcp_sock *tp, int flags,
491                             int mss_now, int nonagle)
492 {
493         if (sk->sk_send_head) {
494                 struct sk_buff *skb = sk->sk_write_queue.prev;
495                 if (!(flags & MSG_MORE) || forced_push(tp))
496                         tcp_mark_push(tp, skb);
497                 tcp_mark_urg(tp, flags, skb);
498                 __tcp_push_pending_frames(sk, tp, mss_now,
499                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
500         }
501 }
502
503 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
504                          size_t psize, int flags)
505 {
506         struct tcp_sock *tp = tcp_sk(sk);
507         int mss_now, size_goal;
508         int err;
509         ssize_t copied;
510         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
511
512         /* Wait for a connection to finish. */
513         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
514                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
515                         goto out_err;
516
517         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
518
519         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
520         size_goal = tp->xmit_size_goal;
521         copied = 0;
522
523         err = -EPIPE;
524         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
525                 goto do_error;
526
527         while (psize > 0) {
528                 struct sk_buff *skb = sk->sk_write_queue.prev;
529                 struct page *page = pages[poffset / PAGE_SIZE];
530                 int copy, i, can_coalesce;
531                 int offset = poffset % PAGE_SIZE;
532                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
533
534                 if (!sk->sk_send_head || (copy = size_goal - skb->len) <= 0) {
535 new_segment:
536                         if (!sk_stream_memory_free(sk))
537                                 goto wait_for_sndbuf;
538
539                         skb = sk_stream_alloc_pskb(sk, 0, 0,
540                                                    sk->sk_allocation);
541                         if (!skb)
542                                 goto wait_for_memory;
543
544                         skb_entail(sk, tp, skb);
545                         copy = size_goal;
546                 }
547
548                 if (copy > size)
549                         copy = size;
550
551                 i = skb_shinfo(skb)->nr_frags;
552                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
553                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
554                         tcp_mark_push(tp, skb);
555                         goto new_segment;
556                 }
557                 if (!sk_stream_wmem_schedule(sk, copy))
558                         goto wait_for_memory;
559                 
560                 if (can_coalesce) {
561                         skb_shinfo(skb)->frags[i - 1].size += copy;
562                 } else {
563                         get_page(page);
564                         skb_fill_page_desc(skb, i, page, offset, copy);
565                 }
566
567                 skb->len += copy;
568                 skb->data_len += copy;
569                 skb->truesize += copy;
570                 sk->sk_wmem_queued += copy;
571                 sk->sk_forward_alloc -= copy;
572                 skb->ip_summed = CHECKSUM_PARTIAL;
573                 tp->write_seq += copy;
574                 TCP_SKB_CB(skb)->end_seq += copy;
575                 skb_shinfo(skb)->gso_segs = 0;
576
577                 if (!copied)
578                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
579
580                 copied += copy;
581                 poffset += copy;
582                 if (!(psize -= copy))
583                         goto out;
584
585                 if (skb->len < mss_now || (flags & MSG_OOB))
586                         continue;
587
588                 if (forced_push(tp)) {
589                         tcp_mark_push(tp, skb);
590                         __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
591                 } else if (skb == sk->sk_send_head)
592                         tcp_push_one(sk, mss_now);
593                 continue;
594
595 wait_for_sndbuf:
596                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
597 wait_for_memory:
598                 if (copied)
599                         tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
600
601                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
602                         goto do_error;
603
604                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
605                 size_goal = tp->xmit_size_goal;
606         }
607
608 out:
609         if (copied)
610                 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
611         return copied;
612
613 do_error:
614         if (copied)
615                 goto out;
616 out_err:
617         return sk_stream_error(sk, flags, err);
618 }
619
620 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
621                      size_t size, int flags)
622 {
623         ssize_t res;
624         struct sock *sk = sock->sk;
625
626         if (!(sk->sk_route_caps & NETIF_F_SG) ||
627             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
628                 return sock_no_sendpage(sock, page, offset, size, flags);
629
630         lock_sock(sk);
631         TCP_CHECK_TIMER(sk);
632         res = do_tcp_sendpages(sk, &page, offset, size, flags);
633         TCP_CHECK_TIMER(sk);
634         release_sock(sk);
635         return res;
636 }
637
638 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
639 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
640
641 static inline int select_size(struct sock *sk, struct tcp_sock *tp)
642 {
643         int tmp = tp->mss_cache;
644
645         if (sk->sk_route_caps & NETIF_F_SG) {
646                 if (sk_can_gso(sk))
647                         tmp = 0;
648                 else {
649                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
650
651                         if (tmp >= pgbreak &&
652                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
653                                 tmp = pgbreak;
654                 }
655         }
656
657         return tmp;
658 }
659
660 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
661                 size_t size)
662 {
663         struct iovec *iov;
664         struct tcp_sock *tp = tcp_sk(sk);
665         struct sk_buff *skb;
666         int iovlen, flags;
667         int mss_now, size_goal;
668         int err, copied;
669         long timeo;
670
671         lock_sock(sk);
672         TCP_CHECK_TIMER(sk);
673
674         flags = msg->msg_flags;
675         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
676
677         /* Wait for a connection to finish. */
678         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
679                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
680                         goto out_err;
681
682         /* This should be in poll */
683         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
684
685         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
686         size_goal = tp->xmit_size_goal;
687
688         /* Ok commence sending. */
689         iovlen = msg->msg_iovlen;
690         iov = msg->msg_iov;
691         copied = 0;
692
693         err = -EPIPE;
694         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
695                 goto do_error;
696
697         while (--iovlen >= 0) {
698                 int seglen = iov->iov_len;
699                 unsigned char __user *from = iov->iov_base;
700
701                 iov++;
702
703                 while (seglen > 0) {
704                         int copy;
705
706                         skb = sk->sk_write_queue.prev;
707
708                         if (!sk->sk_send_head ||
709                             (copy = size_goal - skb->len) <= 0) {
710
711 new_segment:
712                                 /* Allocate new segment. If the interface is SG,
713                                  * allocate skb fitting to single page.
714                                  */
715                                 if (!sk_stream_memory_free(sk))
716                                         goto wait_for_sndbuf;
717
718                                 skb = sk_stream_alloc_pskb(sk, select_size(sk, tp),
719                                                            0, sk->sk_allocation);
720                                 if (!skb)
721                                         goto wait_for_memory;
722
723                                 /*
724                                  * Check whether we can use HW checksum.
725                                  */
726                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
727                                         skb->ip_summed = CHECKSUM_PARTIAL;
728
729                                 skb_entail(sk, tp, skb);
730                                 copy = size_goal;
731                         }
732
733                         /* Try to append data to the end of skb. */
734                         if (copy > seglen)
735                                 copy = seglen;
736
737                         /* Where to copy to? */
738                         if (skb_tailroom(skb) > 0) {
739                                 /* We have some space in skb head. Superb! */
740                                 if (copy > skb_tailroom(skb))
741                                         copy = skb_tailroom(skb);
742                                 if ((err = skb_add_data(skb, from, copy)) != 0)
743                                         goto do_fault;
744                         } else {
745                                 int merge = 0;
746                                 int i = skb_shinfo(skb)->nr_frags;
747                                 struct page *page = TCP_PAGE(sk);
748                                 int off = TCP_OFF(sk);
749
750                                 if (skb_can_coalesce(skb, i, page, off) &&
751                                     off != PAGE_SIZE) {
752                                         /* We can extend the last page
753                                          * fragment. */
754                                         merge = 1;
755                                 } else if (i == MAX_SKB_FRAGS ||
756                                            (!i &&
757                                            !(sk->sk_route_caps & NETIF_F_SG))) {
758                                         /* Need to add new fragment and cannot
759                                          * do this because interface is non-SG,
760                                          * or because all the page slots are
761                                          * busy. */
762                                         tcp_mark_push(tp, skb);
763                                         goto new_segment;
764                                 } else if (page) {
765                                         if (off == PAGE_SIZE) {
766                                                 put_page(page);
767                                                 TCP_PAGE(sk) = page = NULL;
768                                                 off = 0;
769                                         }
770                                 } else
771                                         off = 0;
772
773                                 if (copy > PAGE_SIZE - off)
774                                         copy = PAGE_SIZE - off;
775
776                                 if (!sk_stream_wmem_schedule(sk, copy))
777                                         goto wait_for_memory;
778
779                                 if (!page) {
780                                         /* Allocate new cache page. */
781                                         if (!(page = sk_stream_alloc_page(sk)))
782                                                 goto wait_for_memory;
783                                 }
784
785                                 /* Time to copy data. We are close to
786                                  * the end! */
787                                 err = skb_copy_to_page(sk, from, skb, page,
788                                                        off, copy);
789                                 if (err) {
790                                         /* If this page was new, give it to the
791                                          * socket so it does not get leaked.
792                                          */
793                                         if (!TCP_PAGE(sk)) {
794                                                 TCP_PAGE(sk) = page;
795                                                 TCP_OFF(sk) = 0;
796                                         }
797                                         goto do_error;
798                                 }
799
800                                 /* Update the skb. */
801                                 if (merge) {
802                                         skb_shinfo(skb)->frags[i - 1].size +=
803                                                                         copy;
804                                 } else {
805                                         skb_fill_page_desc(skb, i, page, off, copy);
806                                         if (TCP_PAGE(sk)) {
807                                                 get_page(page);
808                                         } else if (off + copy < PAGE_SIZE) {
809                                                 get_page(page);
810                                                 TCP_PAGE(sk) = page;
811                                         }
812                                 }
813
814                                 TCP_OFF(sk) = off + copy;
815                         }
816
817                         if (!copied)
818                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
819
820                         tp->write_seq += copy;
821                         TCP_SKB_CB(skb)->end_seq += copy;
822                         skb_shinfo(skb)->gso_segs = 0;
823
824                         from += copy;
825                         copied += copy;
826                         if ((seglen -= copy) == 0 && iovlen == 0)
827                                 goto out;
828
829                         if (skb->len < mss_now || (flags & MSG_OOB))
830                                 continue;
831
832                         if (forced_push(tp)) {
833                                 tcp_mark_push(tp, skb);
834                                 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
835                         } else if (skb == sk->sk_send_head)
836                                 tcp_push_one(sk, mss_now);
837                         continue;
838
839 wait_for_sndbuf:
840                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
841 wait_for_memory:
842                         if (copied)
843                                 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
844
845                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
846                                 goto do_error;
847
848                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
849                         size_goal = tp->xmit_size_goal;
850                 }
851         }
852
853 out:
854         if (copied)
855                 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
856         TCP_CHECK_TIMER(sk);
857         release_sock(sk);
858         return copied;
859
860 do_fault:
861         if (!skb->len) {
862                 if (sk->sk_send_head == skb)
863                         sk->sk_send_head = NULL;
864                 __skb_unlink(skb, &sk->sk_write_queue);
865                 sk_stream_free_skb(sk, skb);
866         }
867
868 do_error:
869         if (copied)
870                 goto out;
871 out_err:
872         err = sk_stream_error(sk, flags, err);
873         TCP_CHECK_TIMER(sk);
874         release_sock(sk);
875         return err;
876 }
877
878 /*
879  *      Handle reading urgent data. BSD has very simple semantics for
880  *      this, no blocking and very strange errors 8)
881  */
882
883 static int tcp_recv_urg(struct sock *sk, long timeo,
884                         struct msghdr *msg, int len, int flags,
885                         int *addr_len)
886 {
887         struct tcp_sock *tp = tcp_sk(sk);
888
889         /* No URG data to read. */
890         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
891             tp->urg_data == TCP_URG_READ)
892                 return -EINVAL; /* Yes this is right ! */
893
894         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
895                 return -ENOTCONN;
896
897         if (tp->urg_data & TCP_URG_VALID) {
898                 int err = 0;
899                 char c = tp->urg_data;
900
901                 if (!(flags & MSG_PEEK))
902                         tp->urg_data = TCP_URG_READ;
903
904                 /* Read urgent data. */
905                 msg->msg_flags |= MSG_OOB;
906
907                 if (len > 0) {
908                         if (!(flags & MSG_TRUNC))
909                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
910                         len = 1;
911                 } else
912                         msg->msg_flags |= MSG_TRUNC;
913
914                 return err ? -EFAULT : len;
915         }
916
917         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
918                 return 0;
919
920         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
921          * the available implementations agree in this case:
922          * this call should never block, independent of the
923          * blocking state of the socket.
924          * Mike <pall@rz.uni-karlsruhe.de>
925          */
926         return -EAGAIN;
927 }
928
929 /* Clean up the receive buffer for full frames taken by the user,
930  * then send an ACK if necessary.  COPIED is the number of bytes
931  * tcp_recvmsg has given to the user so far, it speeds up the
932  * calculation of whether or not we must ACK for the sake of
933  * a window update.
934  */
935 void tcp_cleanup_rbuf(struct sock *sk, int copied)
936 {
937         struct tcp_sock *tp = tcp_sk(sk);
938         int time_to_ack = 0;
939
940 #if TCP_DEBUG
941         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
942
943         BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
944 #endif
945
946         if (inet_csk_ack_scheduled(sk)) {
947                 const struct inet_connection_sock *icsk = inet_csk(sk);
948                    /* Delayed ACKs frequently hit locked sockets during bulk
949                     * receive. */
950                 if (icsk->icsk_ack.blocked ||
951                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
952                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
953                     /*
954                      * If this read emptied read buffer, we send ACK, if
955                      * connection is not bidirectional, user drained
956                      * receive buffer and there was a small segment
957                      * in queue.
958                      */
959                     (copied > 0 &&
960                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
961                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
962                        !icsk->icsk_ack.pingpong)) &&
963                       !atomic_read(&sk->sk_rmem_alloc)))
964                         time_to_ack = 1;
965         }
966
967         /* We send an ACK if we can now advertise a non-zero window
968          * which has been raised "significantly".
969          *
970          * Even if window raised up to infinity, do not send window open ACK
971          * in states, where we will not receive more. It is useless.
972          */
973         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
974                 __u32 rcv_window_now = tcp_receive_window(tp);
975
976                 /* Optimize, __tcp_select_window() is not cheap. */
977                 if (2*rcv_window_now <= tp->window_clamp) {
978                         __u32 new_window = __tcp_select_window(sk);
979
980                         /* Send ACK now, if this read freed lots of space
981                          * in our buffer. Certainly, new_window is new window.
982                          * We can advertise it now, if it is not less than current one.
983                          * "Lots" means "at least twice" here.
984                          */
985                         if (new_window && new_window >= 2 * rcv_window_now)
986                                 time_to_ack = 1;
987                 }
988         }
989         if (time_to_ack)
990                 tcp_send_ack(sk);
991 }
992
993 static void tcp_prequeue_process(struct sock *sk)
994 {
995         struct sk_buff *skb;
996         struct tcp_sock *tp = tcp_sk(sk);
997
998         NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
999
1000         /* RX process wants to run with disabled BHs, though it is not
1001          * necessary */
1002         local_bh_disable();
1003         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1004                 sk->sk_backlog_rcv(sk, skb);
1005         local_bh_enable();
1006
1007         /* Clear memory counter. */
1008         tp->ucopy.memory = 0;
1009 }
1010
1011 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1012 {
1013         struct sk_buff *skb;
1014         u32 offset;
1015
1016         skb_queue_walk(&sk->sk_receive_queue, skb) {
1017                 offset = seq - TCP_SKB_CB(skb)->seq;
1018                 if (skb->h.th->syn)
1019                         offset--;
1020                 if (offset < skb->len || skb->h.th->fin) {
1021                         *off = offset;
1022                         return skb;
1023                 }
1024         }
1025         return NULL;
1026 }
1027
1028 /*
1029  * This routine provides an alternative to tcp_recvmsg() for routines
1030  * that would like to handle copying from skbuffs directly in 'sendfile'
1031  * fashion.
1032  * Note:
1033  *      - It is assumed that the socket was locked by the caller.
1034  *      - The routine does not block.
1035  *      - At present, there is no support for reading OOB data
1036  *        or for 'peeking' the socket using this routine
1037  *        (although both would be easy to implement).
1038  */
1039 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1040                   sk_read_actor_t recv_actor)
1041 {
1042         struct sk_buff *skb;
1043         struct tcp_sock *tp = tcp_sk(sk);
1044         u32 seq = tp->copied_seq;
1045         u32 offset;
1046         int copied = 0;
1047
1048         if (sk->sk_state == TCP_LISTEN)
1049                 return -ENOTCONN;
1050         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1051                 if (offset < skb->len) {
1052                         size_t used, len;
1053
1054                         len = skb->len - offset;
1055                         /* Stop reading if we hit a patch of urgent data */
1056                         if (tp->urg_data) {
1057                                 u32 urg_offset = tp->urg_seq - seq;
1058                                 if (urg_offset < len)
1059                                         len = urg_offset;
1060                                 if (!len)
1061                                         break;
1062                         }
1063                         used = recv_actor(desc, skb, offset, len);
1064                         if (used <= len) {
1065                                 seq += used;
1066                                 copied += used;
1067                                 offset += used;
1068                         }
1069                         if (offset != skb->len)
1070                                 break;
1071                 }
1072                 if (skb->h.th->fin) {
1073                         sk_eat_skb(sk, skb, 0);
1074                         ++seq;
1075                         break;
1076                 }
1077                 sk_eat_skb(sk, skb, 0);
1078                 if (!desc->count)
1079                         break;
1080         }
1081         tp->copied_seq = seq;
1082
1083         tcp_rcv_space_adjust(sk);
1084
1085         /* Clean up data we have read: This will do ACK frames. */
1086         if (copied)
1087                 tcp_cleanup_rbuf(sk, copied);
1088         return copied;
1089 }
1090
1091 /*
1092  *      This routine copies from a sock struct into the user buffer.
1093  *
1094  *      Technical note: in 2.3 we work on _locked_ socket, so that
1095  *      tricks with *seq access order and skb->users are not required.
1096  *      Probably, code can be easily improved even more.
1097  */
1098
1099 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1100                 size_t len, int nonblock, int flags, int *addr_len)
1101 {
1102         struct tcp_sock *tp = tcp_sk(sk);
1103         int copied = 0;
1104         u32 peek_seq;
1105         u32 *seq;
1106         unsigned long used;
1107         int err;
1108         int target;             /* Read at least this many bytes */
1109         long timeo;
1110         struct task_struct *user_recv = NULL;
1111         int copied_early = 0;
1112
1113         lock_sock(sk);
1114
1115         TCP_CHECK_TIMER(sk);
1116
1117         err = -ENOTCONN;
1118         if (sk->sk_state == TCP_LISTEN)
1119                 goto out;
1120
1121         timeo = sock_rcvtimeo(sk, nonblock);
1122
1123         /* Urgent data needs to be handled specially. */
1124         if (flags & MSG_OOB)
1125                 goto recv_urg;
1126
1127         seq = &tp->copied_seq;
1128         if (flags & MSG_PEEK) {
1129                 peek_seq = tp->copied_seq;
1130                 seq = &peek_seq;
1131         }
1132
1133         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1134
1135 #ifdef CONFIG_NET_DMA
1136         tp->ucopy.dma_chan = NULL;
1137         preempt_disable();
1138         if ((len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1139             !sysctl_tcp_low_latency && __get_cpu_var(softnet_data).net_dma) {
1140                 preempt_enable_no_resched();
1141                 tp->ucopy.pinned_list = dma_pin_iovec_pages(msg->msg_iov, len);
1142         } else
1143                 preempt_enable_no_resched();
1144 #endif
1145
1146         do {
1147                 struct sk_buff *skb;
1148                 u32 offset;
1149
1150                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1151                 if (tp->urg_data && tp->urg_seq == *seq) {
1152                         if (copied)
1153                                 break;
1154                         if (signal_pending(current)) {
1155                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1156                                 break;
1157                         }
1158                 }
1159
1160                 /* Next get a buffer. */
1161
1162                 skb = skb_peek(&sk->sk_receive_queue);
1163                 do {
1164                         if (!skb)
1165                                 break;
1166
1167                         /* Now that we have two receive queues this
1168                          * shouldn't happen.
1169                          */
1170                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1171                                 printk(KERN_INFO "recvmsg bug: copied %X "
1172                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1173                                 break;
1174                         }
1175                         offset = *seq - TCP_SKB_CB(skb)->seq;
1176                         if (skb->h.th->syn)
1177                                 offset--;
1178                         if (offset < skb->len)
1179                                 goto found_ok_skb;
1180                         if (skb->h.th->fin)
1181                                 goto found_fin_ok;
1182                         BUG_TRAP(flags & MSG_PEEK);
1183                         skb = skb->next;
1184                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1185
1186                 /* Well, if we have backlog, try to process it now yet. */
1187
1188                 if (copied >= target && !sk->sk_backlog.tail)
1189                         break;
1190
1191                 if (copied) {
1192                         if (sk->sk_err ||
1193                             sk->sk_state == TCP_CLOSE ||
1194                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1195                             !timeo ||
1196                             signal_pending(current) ||
1197                             (flags & MSG_PEEK))
1198                                 break;
1199                 } else {
1200                         if (sock_flag(sk, SOCK_DONE))
1201                                 break;
1202
1203                         if (sk->sk_err) {
1204                                 copied = sock_error(sk);
1205                                 break;
1206                         }
1207
1208                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1209                                 break;
1210
1211                         if (sk->sk_state == TCP_CLOSE) {
1212                                 if (!sock_flag(sk, SOCK_DONE)) {
1213                                         /* This occurs when user tries to read
1214                                          * from never connected socket.
1215                                          */
1216                                         copied = -ENOTCONN;
1217                                         break;
1218                                 }
1219                                 break;
1220                         }
1221
1222                         if (!timeo) {
1223                                 copied = -EAGAIN;
1224                                 break;
1225                         }
1226
1227                         if (signal_pending(current)) {
1228                                 copied = sock_intr_errno(timeo);
1229                                 break;
1230                         }
1231                 }
1232
1233                 tcp_cleanup_rbuf(sk, copied);
1234
1235                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1236                         /* Install new reader */
1237                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1238                                 user_recv = current;
1239                                 tp->ucopy.task = user_recv;
1240                                 tp->ucopy.iov = msg->msg_iov;
1241                         }
1242
1243                         tp->ucopy.len = len;
1244
1245                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1246                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1247
1248                         /* Ugly... If prequeue is not empty, we have to
1249                          * process it before releasing socket, otherwise
1250                          * order will be broken at second iteration.
1251                          * More elegant solution is required!!!
1252                          *
1253                          * Look: we have the following (pseudo)queues:
1254                          *
1255                          * 1. packets in flight
1256                          * 2. backlog
1257                          * 3. prequeue
1258                          * 4. receive_queue
1259                          *
1260                          * Each queue can be processed only if the next ones
1261                          * are empty. At this point we have empty receive_queue.
1262                          * But prequeue _can_ be not empty after 2nd iteration,
1263                          * when we jumped to start of loop because backlog
1264                          * processing added something to receive_queue.
1265                          * We cannot release_sock(), because backlog contains
1266                          * packets arrived _after_ prequeued ones.
1267                          *
1268                          * Shortly, algorithm is clear --- to process all
1269                          * the queues in order. We could make it more directly,
1270                          * requeueing packets from backlog to prequeue, if
1271                          * is not empty. It is more elegant, but eats cycles,
1272                          * unfortunately.
1273                          */
1274                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1275                                 goto do_prequeue;
1276
1277                         /* __ Set realtime policy in scheduler __ */
1278                 }
1279
1280                 if (copied >= target) {
1281                         /* Do not sleep, just process backlog. */
1282                         release_sock(sk);
1283                         lock_sock(sk);
1284                 } else
1285                         sk_wait_data(sk, &timeo);
1286
1287 #ifdef CONFIG_NET_DMA
1288                 tp->ucopy.wakeup = 0;
1289 #endif
1290
1291                 if (user_recv) {
1292                         int chunk;
1293
1294                         /* __ Restore normal policy in scheduler __ */
1295
1296                         if ((chunk = len - tp->ucopy.len) != 0) {
1297                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1298                                 len -= chunk;
1299                                 copied += chunk;
1300                         }
1301
1302                         if (tp->rcv_nxt == tp->copied_seq &&
1303                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1304 do_prequeue:
1305                                 tcp_prequeue_process(sk);
1306
1307                                 if ((chunk = len - tp->ucopy.len) != 0) {
1308                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1309                                         len -= chunk;
1310                                         copied += chunk;
1311                                 }
1312                         }
1313                 }
1314                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1315                         if (net_ratelimit())
1316                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1317                                        current->comm, current->pid);
1318                         peek_seq = tp->copied_seq;
1319                 }
1320                 continue;
1321
1322         found_ok_skb:
1323                 /* Ok so how much can we use? */
1324                 used = skb->len - offset;
1325                 if (len < used)
1326                         used = len;
1327
1328                 /* Do we have urgent data here? */
1329                 if (tp->urg_data) {
1330                         u32 urg_offset = tp->urg_seq - *seq;
1331                         if (urg_offset < used) {
1332                                 if (!urg_offset) {
1333                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1334                                                 ++*seq;
1335                                                 offset++;
1336                                                 used--;
1337                                                 if (!used)
1338                                                         goto skip_copy;
1339                                         }
1340                                 } else
1341                                         used = urg_offset;
1342                         }
1343                 }
1344
1345                 if (!(flags & MSG_TRUNC)) {
1346 #ifdef CONFIG_NET_DMA
1347                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1348                                 tp->ucopy.dma_chan = get_softnet_dma();
1349
1350                         if (tp->ucopy.dma_chan) {
1351                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1352                                         tp->ucopy.dma_chan, skb, offset,
1353                                         msg->msg_iov, used,
1354                                         tp->ucopy.pinned_list);
1355
1356                                 if (tp->ucopy.dma_cookie < 0) {
1357
1358                                         printk(KERN_ALERT "dma_cookie < 0\n");
1359
1360                                         /* Exception. Bailout! */
1361                                         if (!copied)
1362                                                 copied = -EFAULT;
1363                                         break;
1364                                 }
1365                                 if ((offset + used) == skb->len)
1366                                         copied_early = 1;
1367
1368                         } else
1369 #endif
1370                         {
1371                                 err = skb_copy_datagram_iovec(skb, offset,
1372                                                 msg->msg_iov, used);
1373                                 if (err) {
1374                                         /* Exception. Bailout! */
1375                                         if (!copied)
1376                                                 copied = -EFAULT;
1377                                         break;
1378                                 }
1379                         }
1380                 }
1381
1382                 *seq += used;
1383                 copied += used;
1384                 len -= used;
1385
1386                 tcp_rcv_space_adjust(sk);
1387
1388 skip_copy:
1389                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1390                         tp->urg_data = 0;
1391                         tcp_fast_path_check(sk, tp);
1392                 }
1393                 if (used + offset < skb->len)
1394                         continue;
1395
1396                 if (skb->h.th->fin)
1397                         goto found_fin_ok;
1398                 if (!(flags & MSG_PEEK)) {
1399                         sk_eat_skb(sk, skb, copied_early);
1400                         copied_early = 0;
1401                 }
1402                 continue;
1403
1404         found_fin_ok:
1405                 /* Process the FIN. */
1406                 ++*seq;
1407                 if (!(flags & MSG_PEEK)) {
1408                         sk_eat_skb(sk, skb, copied_early);
1409                         copied_early = 0;
1410                 }
1411                 break;
1412         } while (len > 0);
1413
1414         if (user_recv) {
1415                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1416                         int chunk;
1417
1418                         tp->ucopy.len = copied > 0 ? len : 0;
1419
1420                         tcp_prequeue_process(sk);
1421
1422                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1423                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1424                                 len -= chunk;
1425                                 copied += chunk;
1426                         }
1427                 }
1428
1429                 tp->ucopy.task = NULL;
1430                 tp->ucopy.len = 0;
1431         }
1432
1433 #ifdef CONFIG_NET_DMA
1434         if (tp->ucopy.dma_chan) {
1435                 struct sk_buff *skb;
1436                 dma_cookie_t done, used;
1437
1438                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1439
1440                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1441                                                  tp->ucopy.dma_cookie, &done,
1442                                                  &used) == DMA_IN_PROGRESS) {
1443                         /* do partial cleanup of sk_async_wait_queue */
1444                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1445                                (dma_async_is_complete(skb->dma_cookie, done,
1446                                                       used) == DMA_SUCCESS)) {
1447                                 __skb_dequeue(&sk->sk_async_wait_queue);
1448                                 kfree_skb(skb);
1449                         }
1450                 }
1451
1452                 /* Safe to free early-copied skbs now */
1453                 __skb_queue_purge(&sk->sk_async_wait_queue);
1454                 dma_chan_put(tp->ucopy.dma_chan);
1455                 tp->ucopy.dma_chan = NULL;
1456         }
1457         if (tp->ucopy.pinned_list) {
1458                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1459                 tp->ucopy.pinned_list = NULL;
1460         }
1461 #endif
1462
1463         /* According to UNIX98, msg_name/msg_namelen are ignored
1464          * on connected socket. I was just happy when found this 8) --ANK
1465          */
1466
1467         /* Clean up data we have read: This will do ACK frames. */
1468         tcp_cleanup_rbuf(sk, copied);
1469
1470         TCP_CHECK_TIMER(sk);
1471         release_sock(sk);
1472         return copied;
1473
1474 out:
1475         TCP_CHECK_TIMER(sk);
1476         release_sock(sk);
1477         return err;
1478
1479 recv_urg:
1480         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1481         goto out;
1482 }
1483
1484 /*
1485  *      State processing on a close. This implements the state shift for
1486  *      sending our FIN frame. Note that we only send a FIN for some
1487  *      states. A shutdown() may have already sent the FIN, or we may be
1488  *      closed.
1489  */
1490
1491 static const unsigned char new_state[16] = {
1492   /* current state:        new state:      action:      */
1493   /* (Invalid)          */ TCP_CLOSE,
1494   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1495   /* TCP_SYN_SENT       */ TCP_CLOSE,
1496   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1497   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1498   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1499   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1500   /* TCP_CLOSE          */ TCP_CLOSE,
1501   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1502   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1503   /* TCP_LISTEN         */ TCP_CLOSE,
1504   /* TCP_CLOSING        */ TCP_CLOSING,
1505 };
1506
1507 static int tcp_close_state(struct sock *sk)
1508 {
1509         int next = (int)new_state[sk->sk_state];
1510         int ns = next & TCP_STATE_MASK;
1511
1512         tcp_set_state(sk, ns);
1513
1514         return next & TCP_ACTION_FIN;
1515 }
1516
1517 /*
1518  *      Shutdown the sending side of a connection. Much like close except
1519  *      that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1520  */
1521
1522 void tcp_shutdown(struct sock *sk, int how)
1523 {
1524         /*      We need to grab some memory, and put together a FIN,
1525          *      and then put it into the queue to be sent.
1526          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1527          */
1528         if (!(how & SEND_SHUTDOWN))
1529                 return;
1530
1531         /* If we've already sent a FIN, or it's a closed state, skip this. */
1532         if ((1 << sk->sk_state) &
1533             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1534              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1535                 /* Clear out any half completed packets.  FIN if needed. */
1536                 if (tcp_close_state(sk))
1537                         tcp_send_fin(sk);
1538         }
1539 }
1540
1541 void tcp_close(struct sock *sk, long timeout)
1542 {
1543         struct sk_buff *skb;
1544         int data_was_unread = 0;
1545         int state;
1546
1547         lock_sock(sk);
1548         sk->sk_shutdown = SHUTDOWN_MASK;
1549
1550         if (sk->sk_state == TCP_LISTEN) {
1551                 tcp_set_state(sk, TCP_CLOSE);
1552
1553                 /* Special case. */
1554                 inet_csk_listen_stop(sk);
1555
1556                 goto adjudge_to_death;
1557         }
1558
1559         /*  We need to flush the recv. buffs.  We do this only on the
1560          *  descriptor close, not protocol-sourced closes, because the
1561          *  reader process may not have drained the data yet!
1562          */
1563         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1564                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1565                           skb->h.th->fin;
1566                 data_was_unread += len;
1567                 __kfree_skb(skb);
1568         }
1569
1570         sk_stream_mem_reclaim(sk);
1571
1572         /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1573          * 3.10, we send a RST here because data was lost.  To
1574          * witness the awful effects of the old behavior of always
1575          * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1576          * a bulk GET in an FTP client, suspend the process, wait
1577          * for the client to advertise a zero window, then kill -9
1578          * the FTP client, wheee...  Note: timeout is always zero
1579          * in such a case.
1580          */
1581         if (data_was_unread) {
1582                 /* Unread data was tossed, zap the connection. */
1583                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1584                 tcp_set_state(sk, TCP_CLOSE);
1585                 tcp_send_active_reset(sk, GFP_KERNEL);
1586         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1587                 /* Check zero linger _after_ checking for unread data. */
1588                 sk->sk_prot->disconnect(sk, 0);
1589                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1590         } else if (tcp_close_state(sk)) {
1591                 /* We FIN if the application ate all the data before
1592                  * zapping the connection.
1593                  */
1594
1595                 /* RED-PEN. Formally speaking, we have broken TCP state
1596                  * machine. State transitions:
1597                  *
1598                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1599                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1600                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1601                  *
1602                  * are legal only when FIN has been sent (i.e. in window),
1603                  * rather than queued out of window. Purists blame.
1604                  *
1605                  * F.e. "RFC state" is ESTABLISHED,
1606                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1607                  *
1608                  * The visible declinations are that sometimes
1609                  * we enter time-wait state, when it is not required really
1610                  * (harmless), do not send active resets, when they are
1611                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1612                  * they look as CLOSING or LAST_ACK for Linux)
1613                  * Probably, I missed some more holelets.
1614                  *                                              --ANK
1615                  */
1616                 tcp_send_fin(sk);
1617         }
1618
1619         sk_stream_wait_close(sk, timeout);
1620
1621 adjudge_to_death:
1622         state = sk->sk_state;
1623         sock_hold(sk);
1624         sock_orphan(sk);
1625         atomic_inc(sk->sk_prot->orphan_count);
1626
1627         /* It is the last release_sock in its life. It will remove backlog. */
1628         release_sock(sk);
1629
1630
1631         /* Now socket is owned by kernel and we acquire BH lock
1632            to finish close. No need to check for user refs.
1633          */
1634         local_bh_disable();
1635         bh_lock_sock(sk);
1636         BUG_TRAP(!sock_owned_by_user(sk));
1637
1638         /* Have we already been destroyed by a softirq or backlog? */
1639         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1640                 goto out;
1641
1642         /*      This is a (useful) BSD violating of the RFC. There is a
1643          *      problem with TCP as specified in that the other end could
1644          *      keep a socket open forever with no application left this end.
1645          *      We use a 3 minute timeout (about the same as BSD) then kill
1646          *      our end. If they send after that then tough - BUT: long enough
1647          *      that we won't make the old 4*rto = almost no time - whoops
1648          *      reset mistake.
1649          *
1650          *      Nope, it was not mistake. It is really desired behaviour
1651          *      f.e. on http servers, when such sockets are useless, but
1652          *      consume significant resources. Let's do it with special
1653          *      linger2 option.                                 --ANK
1654          */
1655
1656         if (sk->sk_state == TCP_FIN_WAIT2) {
1657                 struct tcp_sock *tp = tcp_sk(sk);
1658                 if (tp->linger2 < 0) {
1659                         tcp_set_state(sk, TCP_CLOSE);
1660                         tcp_send_active_reset(sk, GFP_ATOMIC);
1661                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1662                 } else {
1663                         const int tmo = tcp_fin_time(sk);
1664
1665                         if (tmo > TCP_TIMEWAIT_LEN) {
1666                                 inet_csk_reset_keepalive_timer(sk,
1667                                                 tmo - TCP_TIMEWAIT_LEN);
1668                         } else {
1669                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1670                                 goto out;
1671                         }
1672                 }
1673         }
1674         if (sk->sk_state != TCP_CLOSE) {
1675                 sk_stream_mem_reclaim(sk);
1676                 if (atomic_read(sk->sk_prot->orphan_count) > sysctl_tcp_max_orphans ||
1677                     (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1678                      atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1679                         if (net_ratelimit())
1680                                 printk(KERN_INFO "TCP: too many of orphaned "
1681                                        "sockets\n");
1682                         tcp_set_state(sk, TCP_CLOSE);
1683                         tcp_send_active_reset(sk, GFP_ATOMIC);
1684                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1685                 }
1686         }
1687
1688         if (sk->sk_state == TCP_CLOSE)
1689                 inet_csk_destroy_sock(sk);
1690         /* Otherwise, socket is reprieved until protocol close. */
1691
1692 out:
1693         bh_unlock_sock(sk);
1694         local_bh_enable();
1695         sock_put(sk);
1696 }
1697
1698 /* These states need RST on ABORT according to RFC793 */
1699
1700 static inline int tcp_need_reset(int state)
1701 {
1702         return (1 << state) &
1703                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1704                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1705 }
1706
1707 int tcp_disconnect(struct sock *sk, int flags)
1708 {
1709         struct inet_sock *inet = inet_sk(sk);
1710         struct inet_connection_sock *icsk = inet_csk(sk);
1711         struct tcp_sock *tp = tcp_sk(sk);
1712         int err = 0;
1713         int old_state = sk->sk_state;
1714
1715         if (old_state != TCP_CLOSE)
1716                 tcp_set_state(sk, TCP_CLOSE);
1717
1718         /* ABORT function of RFC793 */
1719         if (old_state == TCP_LISTEN) {
1720                 inet_csk_listen_stop(sk);
1721         } else if (tcp_need_reset(old_state) ||
1722                    (tp->snd_nxt != tp->write_seq &&
1723                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1724                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1725                  * states
1726                  */
1727                 tcp_send_active_reset(sk, gfp_any());
1728                 sk->sk_err = ECONNRESET;
1729         } else if (old_state == TCP_SYN_SENT)
1730                 sk->sk_err = ECONNRESET;
1731
1732         tcp_clear_xmit_timers(sk);
1733         __skb_queue_purge(&sk->sk_receive_queue);
1734         sk_stream_writequeue_purge(sk);
1735         __skb_queue_purge(&tp->out_of_order_queue);
1736 #ifdef CONFIG_NET_DMA
1737         __skb_queue_purge(&sk->sk_async_wait_queue);
1738 #endif
1739
1740         inet->dport = 0;
1741
1742         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1743                 inet_reset_saddr(sk);
1744
1745         sk->sk_shutdown = 0;
1746         sock_reset_flag(sk, SOCK_DONE);
1747         tp->srtt = 0;
1748         if ((tp->write_seq += tp->max_window + 2) == 0)
1749                 tp->write_seq = 1;
1750         icsk->icsk_backoff = 0;
1751         tp->snd_cwnd = 2;
1752         icsk->icsk_probes_out = 0;
1753         tp->packets_out = 0;
1754         tp->snd_ssthresh = 0x7fffffff;
1755         tp->snd_cwnd_cnt = 0;
1756         tp->bytes_acked = 0;
1757         tcp_set_ca_state(sk, TCP_CA_Open);
1758         tcp_clear_retrans(tp);
1759         inet_csk_delack_init(sk);
1760         sk->sk_send_head = NULL;
1761         tp->rx_opt.saw_tstamp = 0;
1762         tcp_sack_reset(&tp->rx_opt);
1763         __sk_dst_reset(sk);
1764
1765         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1766
1767         sk->sk_error_report(sk);
1768         return err;
1769 }
1770
1771 /*
1772  *      Socket option code for TCP.
1773  */
1774 static int do_tcp_setsockopt(struct sock *sk, int level,
1775                 int optname, char __user *optval, int optlen)
1776 {
1777         struct tcp_sock *tp = tcp_sk(sk);
1778         struct inet_connection_sock *icsk = inet_csk(sk);
1779         int val;
1780         int err = 0;
1781
1782         /* This is a string value all the others are int's */
1783         if (optname == TCP_CONGESTION) {
1784                 char name[TCP_CA_NAME_MAX];
1785
1786                 if (optlen < 1)
1787                         return -EINVAL;
1788
1789                 val = strncpy_from_user(name, optval,
1790                                         min(TCP_CA_NAME_MAX-1, optlen));
1791                 if (val < 0)
1792                         return -EFAULT;
1793                 name[val] = 0;
1794
1795                 lock_sock(sk);
1796                 err = tcp_set_congestion_control(sk, name);
1797                 release_sock(sk);
1798                 return err;
1799         }
1800
1801         if (optlen < sizeof(int))
1802                 return -EINVAL;
1803
1804         if (get_user(val, (int __user *)optval))
1805                 return -EFAULT;
1806
1807         lock_sock(sk);
1808
1809         switch (optname) {
1810         case TCP_MAXSEG:
1811                 /* Values greater than interface MTU won't take effect. However
1812                  * at the point when this call is done we typically don't yet
1813                  * know which interface is going to be used */
1814                 if (val < 8 || val > MAX_TCP_WINDOW) {
1815                         err = -EINVAL;
1816                         break;
1817                 }
1818                 tp->rx_opt.user_mss = val;
1819                 break;
1820
1821         case TCP_NODELAY:
1822                 if (val) {
1823                         /* TCP_NODELAY is weaker than TCP_CORK, so that
1824                          * this option on corked socket is remembered, but
1825                          * it is not activated until cork is cleared.
1826                          *
1827                          * However, when TCP_NODELAY is set we make
1828                          * an explicit push, which overrides even TCP_CORK
1829                          * for currently queued segments.
1830                          */
1831                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1832                         tcp_push_pending_frames(sk, tp);
1833                 } else {
1834                         tp->nonagle &= ~TCP_NAGLE_OFF;
1835                 }
1836                 break;
1837
1838         case TCP_CORK:
1839                 /* When set indicates to always queue non-full frames.
1840                  * Later the user clears this option and we transmit
1841                  * any pending partial frames in the queue.  This is
1842                  * meant to be used alongside sendfile() to get properly
1843                  * filled frames when the user (for example) must write
1844                  * out headers with a write() call first and then use
1845                  * sendfile to send out the data parts.
1846                  *
1847                  * TCP_CORK can be set together with TCP_NODELAY and it is
1848                  * stronger than TCP_NODELAY.
1849                  */
1850                 if (val) {
1851                         tp->nonagle |= TCP_NAGLE_CORK;
1852                 } else {
1853                         tp->nonagle &= ~TCP_NAGLE_CORK;
1854                         if (tp->nonagle&TCP_NAGLE_OFF)
1855                                 tp->nonagle |= TCP_NAGLE_PUSH;
1856                         tcp_push_pending_frames(sk, tp);
1857                 }
1858                 break;
1859
1860         case TCP_KEEPIDLE:
1861                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1862                         err = -EINVAL;
1863                 else {
1864                         tp->keepalive_time = val * HZ;
1865                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
1866                             !((1 << sk->sk_state) &
1867                               (TCPF_CLOSE | TCPF_LISTEN))) {
1868                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1869                                 if (tp->keepalive_time > elapsed)
1870                                         elapsed = tp->keepalive_time - elapsed;
1871                                 else
1872                                         elapsed = 0;
1873                                 inet_csk_reset_keepalive_timer(sk, elapsed);
1874                         }
1875                 }
1876                 break;
1877         case TCP_KEEPINTVL:
1878                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1879                         err = -EINVAL;
1880                 else
1881                         tp->keepalive_intvl = val * HZ;
1882                 break;
1883         case TCP_KEEPCNT:
1884                 if (val < 1 || val > MAX_TCP_KEEPCNT)
1885                         err = -EINVAL;
1886                 else
1887                         tp->keepalive_probes = val;
1888                 break;
1889         case TCP_SYNCNT:
1890                 if (val < 1 || val > MAX_TCP_SYNCNT)
1891                         err = -EINVAL;
1892                 else
1893                         icsk->icsk_syn_retries = val;
1894                 break;
1895
1896         case TCP_LINGER2:
1897                 if (val < 0)
1898                         tp->linger2 = -1;
1899                 else if (val > sysctl_tcp_fin_timeout / HZ)
1900                         tp->linger2 = 0;
1901                 else
1902                         tp->linger2 = val * HZ;
1903                 break;
1904
1905         case TCP_DEFER_ACCEPT:
1906                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1907                 if (val > 0) {
1908                         /* Translate value in seconds to number of
1909                          * retransmits */
1910                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1911                                val > ((TCP_TIMEOUT_INIT / HZ) <<
1912                                        icsk->icsk_accept_queue.rskq_defer_accept))
1913                                 icsk->icsk_accept_queue.rskq_defer_accept++;
1914                         icsk->icsk_accept_queue.rskq_defer_accept++;
1915                 }
1916                 break;
1917
1918         case TCP_WINDOW_CLAMP:
1919                 if (!val) {
1920                         if (sk->sk_state != TCP_CLOSE) {
1921                                 err = -EINVAL;
1922                                 break;
1923                         }
1924                         tp->window_clamp = 0;
1925                 } else
1926                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1927                                                 SOCK_MIN_RCVBUF / 2 : val;
1928                 break;
1929
1930         case TCP_QUICKACK:
1931                 if (!val) {
1932                         icsk->icsk_ack.pingpong = 1;
1933                 } else {
1934                         icsk->icsk_ack.pingpong = 0;
1935                         if ((1 << sk->sk_state) &
1936                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1937                             inet_csk_ack_scheduled(sk)) {
1938                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1939                                 tcp_cleanup_rbuf(sk, 1);
1940                                 if (!(val & 1))
1941                                         icsk->icsk_ack.pingpong = 1;
1942                         }
1943                 }
1944                 break;
1945
1946 #ifdef CONFIG_TCP_MD5SIG
1947         case TCP_MD5SIG:
1948                 /* Read the IP->Key mappings from userspace */
1949                 err = tp->af_specific->md5_parse(sk, optval, optlen);
1950                 break;
1951 #endif
1952
1953         default:
1954                 err = -ENOPROTOOPT;
1955                 break;
1956         };
1957         release_sock(sk);
1958         return err;
1959 }
1960
1961 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1962                    int optlen)
1963 {
1964         struct inet_connection_sock *icsk = inet_csk(sk);
1965
1966         if (level != SOL_TCP)
1967                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1968                                                      optval, optlen);
1969         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1970 }
1971
1972 #ifdef CONFIG_COMPAT
1973 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1974                           char __user *optval, int optlen)
1975 {
1976         if (level != SOL_TCP)
1977                 return inet_csk_compat_setsockopt(sk, level, optname,
1978                                                   optval, optlen);
1979         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1980 }
1981
1982 EXPORT_SYMBOL(compat_tcp_setsockopt);
1983 #endif
1984
1985 /* Return information about state of tcp endpoint in API format. */
1986 void tcp_get_info(struct sock *sk, struct tcp_info *info)
1987 {
1988         struct tcp_sock *tp = tcp_sk(sk);
1989         const struct inet_connection_sock *icsk = inet_csk(sk);
1990         u32 now = tcp_time_stamp;
1991
1992         memset(info, 0, sizeof(*info));
1993
1994         info->tcpi_state = sk->sk_state;
1995         info->tcpi_ca_state = icsk->icsk_ca_state;
1996         info->tcpi_retransmits = icsk->icsk_retransmits;
1997         info->tcpi_probes = icsk->icsk_probes_out;
1998         info->tcpi_backoff = icsk->icsk_backoff;
1999
2000         if (tp->rx_opt.tstamp_ok)
2001                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2002         if (tp->rx_opt.sack_ok)
2003                 info->tcpi_options |= TCPI_OPT_SACK;
2004         if (tp->rx_opt.wscale_ok) {
2005                 info->tcpi_options |= TCPI_OPT_WSCALE;
2006                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2007                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2008         } 
2009
2010         if (tp->ecn_flags&TCP_ECN_OK)
2011                 info->tcpi_options |= TCPI_OPT_ECN;
2012
2013         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2014         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2015         info->tcpi_snd_mss = tp->mss_cache;
2016         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2017
2018         info->tcpi_unacked = tp->packets_out;
2019         info->tcpi_sacked = tp->sacked_out;
2020         info->tcpi_lost = tp->lost_out;
2021         info->tcpi_retrans = tp->retrans_out;
2022         info->tcpi_fackets = tp->fackets_out;
2023
2024         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2025         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2026         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2027
2028         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2029         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2030         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2031         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2032         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2033         info->tcpi_snd_cwnd = tp->snd_cwnd;
2034         info->tcpi_advmss = tp->advmss;
2035         info->tcpi_reordering = tp->reordering;
2036
2037         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2038         info->tcpi_rcv_space = tp->rcvq_space.space;
2039
2040         info->tcpi_total_retrans = tp->total_retrans;
2041 }
2042
2043 EXPORT_SYMBOL_GPL(tcp_get_info);
2044
2045 static int do_tcp_getsockopt(struct sock *sk, int level,
2046                 int optname, char __user *optval, int __user *optlen)
2047 {
2048         struct inet_connection_sock *icsk = inet_csk(sk);
2049         struct tcp_sock *tp = tcp_sk(sk);
2050         int val, len;
2051
2052         if (get_user(len, optlen))
2053                 return -EFAULT;
2054
2055         len = min_t(unsigned int, len, sizeof(int));
2056
2057         if (len < 0)
2058                 return -EINVAL;
2059
2060         switch (optname) {
2061         case TCP_MAXSEG:
2062                 val = tp->mss_cache;
2063                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2064                         val = tp->rx_opt.user_mss;
2065                 break;
2066         case TCP_NODELAY:
2067                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2068                 break;
2069         case TCP_CORK:
2070                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2071                 break;
2072         case TCP_KEEPIDLE:
2073                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2074                 break;
2075         case TCP_KEEPINTVL:
2076                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2077                 break;
2078         case TCP_KEEPCNT:
2079                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2080                 break;
2081         case TCP_SYNCNT:
2082                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2083                 break;
2084         case TCP_LINGER2:
2085                 val = tp->linger2;
2086                 if (val >= 0)
2087                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2088                 break;
2089         case TCP_DEFER_ACCEPT:
2090                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2091                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2092                 break;
2093         case TCP_WINDOW_CLAMP:
2094                 val = tp->window_clamp;
2095                 break;
2096         case TCP_INFO: {
2097                 struct tcp_info info;
2098
2099                 if (get_user(len, optlen))
2100                         return -EFAULT;
2101
2102                 tcp_get_info(sk, &info);
2103
2104                 len = min_t(unsigned int, len, sizeof(info));
2105                 if (put_user(len, optlen))
2106                         return -EFAULT;
2107                 if (copy_to_user(optval, &info, len))
2108                         return -EFAULT;
2109                 return 0;
2110         }
2111         case TCP_QUICKACK:
2112                 val = !icsk->icsk_ack.pingpong;
2113                 break;
2114
2115         case TCP_CONGESTION:
2116                 if (get_user(len, optlen))
2117                         return -EFAULT;
2118                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2119                 if (put_user(len, optlen))
2120                         return -EFAULT;
2121                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2122                         return -EFAULT;
2123                 return 0;
2124         default:
2125                 return -ENOPROTOOPT;
2126         };
2127
2128         if (put_user(len, optlen))
2129                 return -EFAULT;
2130         if (copy_to_user(optval, &val, len))
2131                 return -EFAULT;
2132         return 0;
2133 }
2134
2135 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2136                    int __user *optlen)
2137 {
2138         struct inet_connection_sock *icsk = inet_csk(sk);
2139
2140         if (level != SOL_TCP)
2141                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2142                                                      optval, optlen);
2143         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2144 }
2145
2146 #ifdef CONFIG_COMPAT
2147 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2148                           char __user *optval, int __user *optlen)
2149 {
2150         if (level != SOL_TCP)
2151                 return inet_csk_compat_getsockopt(sk, level, optname,
2152                                                   optval, optlen);
2153         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2154 }
2155
2156 EXPORT_SYMBOL(compat_tcp_getsockopt);
2157 #endif
2158
2159 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2160 {
2161         struct sk_buff *segs = ERR_PTR(-EINVAL);
2162         struct tcphdr *th;
2163         unsigned thlen;
2164         unsigned int seq;
2165         __be32 delta;
2166         unsigned int oldlen;
2167         unsigned int len;
2168
2169         if (!pskb_may_pull(skb, sizeof(*th)))
2170                 goto out;
2171
2172         th = skb->h.th;
2173         thlen = th->doff * 4;
2174         if (thlen < sizeof(*th))
2175                 goto out;
2176
2177         if (!pskb_may_pull(skb, thlen))
2178                 goto out;
2179
2180         oldlen = (u16)~skb->len;
2181         __skb_pull(skb, thlen);
2182
2183         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2184                 /* Packet is from an untrusted source, reset gso_segs. */
2185                 int type = skb_shinfo(skb)->gso_type;
2186                 int mss;
2187
2188                 if (unlikely(type &
2189                              ~(SKB_GSO_TCPV4 |
2190                                SKB_GSO_DODGY |
2191                                SKB_GSO_TCP_ECN |
2192                                SKB_GSO_TCPV6 |
2193                                0) ||
2194                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2195                         goto out;
2196
2197                 mss = skb_shinfo(skb)->gso_size;
2198                 skb_shinfo(skb)->gso_segs = (skb->len + mss - 1) / mss;
2199
2200                 segs = NULL;
2201                 goto out;
2202         }
2203
2204         segs = skb_segment(skb, features);
2205         if (IS_ERR(segs))
2206                 goto out;
2207
2208         len = skb_shinfo(skb)->gso_size;
2209         delta = htonl(oldlen + (thlen + len));
2210
2211         skb = segs;
2212         th = skb->h.th;
2213         seq = ntohl(th->seq);
2214
2215         do {
2216                 th->fin = th->psh = 0;
2217
2218                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2219                                        (__force u32)delta));
2220                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2221                         th->check = csum_fold(csum_partial(skb->h.raw, thlen,
2222                                                            skb->csum));
2223
2224                 seq += len;
2225                 skb = skb->next;
2226                 th = skb->h.th;
2227
2228                 th->seq = htonl(seq);
2229                 th->cwr = 0;
2230         } while (skb->next);
2231
2232         delta = htonl(oldlen + (skb->tail - skb->h.raw) + skb->data_len);
2233         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2234                                 (__force u32)delta));
2235         if (skb->ip_summed != CHECKSUM_PARTIAL)
2236                 th->check = csum_fold(csum_partial(skb->h.raw, thlen,
2237                                                    skb->csum));
2238
2239 out:
2240         return segs;
2241 }
2242 EXPORT_SYMBOL(tcp_tso_segment);
2243
2244 #ifdef CONFIG_TCP_MD5SIG
2245 static unsigned long tcp_md5sig_users;
2246 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2247 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2248
2249 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2250 {
2251         int cpu;
2252         for_each_possible_cpu(cpu) {
2253                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2254                 if (p) {
2255                         if (p->md5_desc.tfm)
2256                                 crypto_free_hash(p->md5_desc.tfm);
2257                         kfree(p);
2258                         p = NULL;
2259                 }
2260         }
2261         free_percpu(pool);
2262 }
2263
2264 void tcp_free_md5sig_pool(void)
2265 {
2266         struct tcp_md5sig_pool **pool = NULL;
2267
2268         spin_lock(&tcp_md5sig_pool_lock);
2269         if (--tcp_md5sig_users == 0) {
2270                 pool = tcp_md5sig_pool;
2271                 tcp_md5sig_pool = NULL;
2272         }
2273         spin_unlock(&tcp_md5sig_pool_lock);
2274         if (pool)
2275                 __tcp_free_md5sig_pool(pool);
2276 }
2277
2278 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2279
2280 struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2281 {
2282         int cpu;
2283         struct tcp_md5sig_pool **pool;
2284
2285         pool = alloc_percpu(struct tcp_md5sig_pool *);
2286         if (!pool)
2287                 return NULL;
2288
2289         for_each_possible_cpu(cpu) {
2290                 struct tcp_md5sig_pool *p;
2291                 struct crypto_hash *hash;
2292
2293                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2294                 if (!p)
2295                         goto out_free;
2296                 *per_cpu_ptr(pool, cpu) = p;
2297
2298                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2299                 if (!hash || IS_ERR(hash))
2300                         goto out_free;
2301
2302                 p->md5_desc.tfm = hash;
2303         }
2304         return pool;
2305 out_free:
2306         __tcp_free_md5sig_pool(pool);
2307         return NULL;
2308 }
2309
2310 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2311 {
2312         struct tcp_md5sig_pool **pool;
2313         int alloc = 0;
2314
2315 retry:
2316         spin_lock(&tcp_md5sig_pool_lock);
2317         pool = tcp_md5sig_pool;
2318         if (tcp_md5sig_users++ == 0) {
2319                 alloc = 1;
2320                 spin_unlock(&tcp_md5sig_pool_lock);
2321         } else if (!pool) {
2322                 tcp_md5sig_users--;
2323                 spin_unlock(&tcp_md5sig_pool_lock);
2324                 cpu_relax();
2325                 goto retry;
2326         } else
2327                 spin_unlock(&tcp_md5sig_pool_lock);
2328
2329         if (alloc) {
2330                 /* we cannot hold spinlock here because this may sleep. */
2331                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2332                 spin_lock(&tcp_md5sig_pool_lock);
2333                 if (!p) {
2334                         tcp_md5sig_users--;
2335                         spin_unlock(&tcp_md5sig_pool_lock);
2336                         return NULL;
2337                 }
2338                 pool = tcp_md5sig_pool;
2339                 if (pool) {
2340                         /* oops, it has already been assigned. */
2341                         spin_unlock(&tcp_md5sig_pool_lock);
2342                         __tcp_free_md5sig_pool(p);
2343                 } else {
2344                         tcp_md5sig_pool = pool = p;
2345                         spin_unlock(&tcp_md5sig_pool_lock);
2346                 }
2347         }
2348         return pool;
2349 }
2350
2351 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2352
2353 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2354 {
2355         struct tcp_md5sig_pool **p;
2356         spin_lock(&tcp_md5sig_pool_lock);
2357         p = tcp_md5sig_pool;
2358         if (p)
2359                 tcp_md5sig_users++;
2360         spin_unlock(&tcp_md5sig_pool_lock);
2361         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2362 }
2363
2364 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2365
2366 void __tcp_put_md5sig_pool(void) {
2367         __tcp_free_md5sig_pool(tcp_md5sig_pool);
2368 }
2369
2370 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2371 #endif
2372
2373 extern void __skb_cb_too_small_for_tcp(int, int);
2374 extern struct tcp_congestion_ops tcp_reno;
2375
2376 static __initdata unsigned long thash_entries;
2377 static int __init set_thash_entries(char *str)
2378 {
2379         if (!str)
2380                 return 0;
2381         thash_entries = simple_strtoul(str, &str, 0);
2382         return 1;
2383 }
2384 __setup("thash_entries=", set_thash_entries);
2385
2386 void __init tcp_init(void)
2387 {
2388         struct sk_buff *skb = NULL;
2389         unsigned long limit;
2390         int order, i, max_share;
2391
2392         if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2393                 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2394                                            sizeof(skb->cb));
2395
2396         tcp_hashinfo.bind_bucket_cachep =
2397                 kmem_cache_create("tcp_bind_bucket",
2398                                   sizeof(struct inet_bind_bucket), 0,
2399                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
2400
2401         /* Size and allocate the main established and bind bucket
2402          * hash tables.
2403          *
2404          * The methodology is similar to that of the buffer cache.
2405          */
2406         tcp_hashinfo.ehash =
2407                 alloc_large_system_hash("TCP established",
2408                                         sizeof(struct inet_ehash_bucket),
2409                                         thash_entries,
2410                                         (num_physpages >= 128 * 1024) ?
2411                                         13 : 15,
2412                                         0,
2413                                         &tcp_hashinfo.ehash_size,
2414                                         NULL,
2415                                         0);
2416         tcp_hashinfo.ehash_size = (1 << tcp_hashinfo.ehash_size) >> 1;
2417         for (i = 0; i < (tcp_hashinfo.ehash_size << 1); i++) {
2418                 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2419                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2420         }
2421
2422         tcp_hashinfo.bhash =
2423                 alloc_large_system_hash("TCP bind",
2424                                         sizeof(struct inet_bind_hashbucket),
2425                                         tcp_hashinfo.ehash_size,
2426                                         (num_physpages >= 128 * 1024) ?
2427                                         13 : 15,
2428                                         0,
2429                                         &tcp_hashinfo.bhash_size,
2430                                         NULL,
2431                                         64 * 1024);
2432         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2433         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2434                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2435                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2436         }
2437
2438         /* Try to be a bit smarter and adjust defaults depending
2439          * on available memory.
2440          */
2441         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2442                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2443                         order++)
2444                 ;
2445         if (order >= 4) {
2446                 sysctl_local_port_range[0] = 32768;
2447                 sysctl_local_port_range[1] = 61000;
2448                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2449                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2450                 sysctl_max_syn_backlog = 1024;
2451         } else if (order < 3) {
2452                 sysctl_local_port_range[0] = 1024 * (3 - order);
2453                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2454                 sysctl_tcp_max_orphans >>= (3 - order);
2455                 sysctl_max_syn_backlog = 128;
2456         }
2457
2458         /* Allow no more than 3/4 kernel memory (usually less) allocated to TCP */
2459         sysctl_tcp_mem[0] = (1536 / sizeof (struct inet_bind_hashbucket)) << order;
2460         sysctl_tcp_mem[1] = sysctl_tcp_mem[0] * 4 / 3;
2461         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2462
2463         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2464         max_share = min(4UL*1024*1024, limit);
2465
2466         sysctl_tcp_wmem[0] = SK_STREAM_MEM_QUANTUM;
2467         sysctl_tcp_wmem[1] = 16*1024;
2468         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2469
2470         sysctl_tcp_rmem[0] = SK_STREAM_MEM_QUANTUM;
2471         sysctl_tcp_rmem[1] = 87380;
2472         sysctl_tcp_rmem[2] = max(87380, max_share);
2473
2474         printk(KERN_INFO "TCP: Hash tables configured "
2475                "(established %d bind %d)\n",
2476                tcp_hashinfo.ehash_size << 1, tcp_hashinfo.bhash_size);
2477
2478         tcp_register_congestion_control(&tcp_reno);
2479 }
2480
2481 EXPORT_SYMBOL(tcp_close);
2482 EXPORT_SYMBOL(tcp_disconnect);
2483 EXPORT_SYMBOL(tcp_getsockopt);
2484 EXPORT_SYMBOL(tcp_ioctl);
2485 EXPORT_SYMBOL(tcp_poll);
2486 EXPORT_SYMBOL(tcp_read_sock);
2487 EXPORT_SYMBOL(tcp_recvmsg);
2488 EXPORT_SYMBOL(tcp_sendmsg);
2489 EXPORT_SYMBOL(tcp_sendpage);
2490 EXPORT_SYMBOL(tcp_setsockopt);
2491 EXPORT_SYMBOL(tcp_shutdown);
2492 EXPORT_SYMBOL(tcp_statistics);