Merge branch 'release' of git://git.kernel.org/pub/scm/linux/kernel/git/lenb/linux...
[linux-2.6] / net / ipv4 / tcp.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp.c,v 1.216 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  *
22  * Fixes:
23  *              Alan Cox        :       Numerous verify_area() calls
24  *              Alan Cox        :       Set the ACK bit on a reset
25  *              Alan Cox        :       Stopped it crashing if it closed while
26  *                                      sk->inuse=1 and was trying to connect
27  *                                      (tcp_err()).
28  *              Alan Cox        :       All icmp error handling was broken
29  *                                      pointers passed where wrong and the
30  *                                      socket was looked up backwards. Nobody
31  *                                      tested any icmp error code obviously.
32  *              Alan Cox        :       tcp_err() now handled properly. It
33  *                                      wakes people on errors. poll
34  *                                      behaves and the icmp error race
35  *                                      has gone by moving it into sock.c
36  *              Alan Cox        :       tcp_send_reset() fixed to work for
37  *                                      everything not just packets for
38  *                                      unknown sockets.
39  *              Alan Cox        :       tcp option processing.
40  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
41  *                                      syn rule wrong]
42  *              Herp Rosmanith  :       More reset fixes
43  *              Alan Cox        :       No longer acks invalid rst frames.
44  *                                      Acking any kind of RST is right out.
45  *              Alan Cox        :       Sets an ignore me flag on an rst
46  *                                      receive otherwise odd bits of prattle
47  *                                      escape still
48  *              Alan Cox        :       Fixed another acking RST frame bug.
49  *                                      Should stop LAN workplace lockups.
50  *              Alan Cox        :       Some tidyups using the new skb list
51  *                                      facilities
52  *              Alan Cox        :       sk->keepopen now seems to work
53  *              Alan Cox        :       Pulls options out correctly on accepts
54  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
55  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
56  *                                      bit to skb ops.
57  *              Alan Cox        :       Tidied tcp_data to avoid a potential
58  *                                      nasty.
59  *              Alan Cox        :       Added some better commenting, as the
60  *                                      tcp is hard to follow
61  *              Alan Cox        :       Removed incorrect check for 20 * psh
62  *      Michael O'Reilly        :       ack < copied bug fix.
63  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
64  *              Alan Cox        :       FIN with no memory -> CRASH
65  *              Alan Cox        :       Added socket option proto entries.
66  *                                      Also added awareness of them to accept.
67  *              Alan Cox        :       Added TCP options (SOL_TCP)
68  *              Alan Cox        :       Switched wakeup calls to callbacks,
69  *                                      so the kernel can layer network
70  *                                      sockets.
71  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
72  *              Alan Cox        :       Handle FIN (more) properly (we hope).
73  *              Alan Cox        :       RST frames sent on unsynchronised
74  *                                      state ack error.
75  *              Alan Cox        :       Put in missing check for SYN bit.
76  *              Alan Cox        :       Added tcp_select_window() aka NET2E
77  *                                      window non shrink trick.
78  *              Alan Cox        :       Added a couple of small NET2E timer
79  *                                      fixes
80  *              Charles Hedrick :       TCP fixes
81  *              Toomas Tamm     :       TCP window fixes
82  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
83  *              Charles Hedrick :       Rewrote most of it to actually work
84  *              Linus           :       Rewrote tcp_read() and URG handling
85  *                                      completely
86  *              Gerhard Koerting:       Fixed some missing timer handling
87  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
88  *              Gerhard Koerting:       PC/TCP workarounds
89  *              Adam Caldwell   :       Assorted timer/timing errors
90  *              Matthew Dillon  :       Fixed another RST bug
91  *              Alan Cox        :       Move to kernel side addressing changes.
92  *              Alan Cox        :       Beginning work on TCP fastpathing
93  *                                      (not yet usable)
94  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
95  *              Alan Cox        :       TCP fast path debugging
96  *              Alan Cox        :       Window clamping
97  *              Michael Riepe   :       Bug in tcp_check()
98  *              Matt Dillon     :       More TCP improvements and RST bug fixes
99  *              Matt Dillon     :       Yet more small nasties remove from the
100  *                                      TCP code (Be very nice to this man if
101  *                                      tcp finally works 100%) 8)
102  *              Alan Cox        :       BSD accept semantics.
103  *              Alan Cox        :       Reset on closedown bug.
104  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
105  *              Michael Pall    :       Handle poll() after URG properly in
106  *                                      all cases.
107  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
108  *                                      (multi URG PUSH broke rlogin).
109  *              Michael Pall    :       Fix the multi URG PUSH problem in
110  *                                      tcp_readable(), poll() after URG
111  *                                      works now.
112  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
113  *                                      BSD api.
114  *              Alan Cox        :       Changed the semantics of sk->socket to
115  *                                      fix a race and a signal problem with
116  *                                      accept() and async I/O.
117  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
118  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
119  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
120  *                                      clients/servers which listen in on
121  *                                      fixed ports.
122  *              Alan Cox        :       Cleaned the above up and shrank it to
123  *                                      a sensible code size.
124  *              Alan Cox        :       Self connect lockup fix.
125  *              Alan Cox        :       No connect to multicast.
126  *              Ross Biro       :       Close unaccepted children on master
127  *                                      socket close.
128  *              Alan Cox        :       Reset tracing code.
129  *              Alan Cox        :       Spurious resets on shutdown.
130  *              Alan Cox        :       Giant 15 minute/60 second timer error
131  *              Alan Cox        :       Small whoops in polling before an
132  *                                      accept.
133  *              Alan Cox        :       Kept the state trace facility since
134  *                                      it's handy for debugging.
135  *              Alan Cox        :       More reset handler fixes.
136  *              Alan Cox        :       Started rewriting the code based on
137  *                                      the RFC's for other useful protocol
138  *                                      references see: Comer, KA9Q NOS, and
139  *                                      for a reference on the difference
140  *                                      between specifications and how BSD
141  *                                      works see the 4.4lite source.
142  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
143  *                                      close.
144  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
145  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
146  *              Alan Cox        :       Reimplemented timers as per the RFC
147  *                                      and using multiple timers for sanity.
148  *              Alan Cox        :       Small bug fixes, and a lot of new
149  *                                      comments.
150  *              Alan Cox        :       Fixed dual reader crash by locking
151  *                                      the buffers (much like datagram.c)
152  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
153  *                                      now gets fed up of retrying without
154  *                                      (even a no space) answer.
155  *              Alan Cox        :       Extracted closing code better
156  *              Alan Cox        :       Fixed the closing state machine to
157  *                                      resemble the RFC.
158  *              Alan Cox        :       More 'per spec' fixes.
159  *              Jorge Cwik      :       Even faster checksumming.
160  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
161  *                                      only frames. At least one pc tcp stack
162  *                                      generates them.
163  *              Alan Cox        :       Cache last socket.
164  *              Alan Cox        :       Per route irtt.
165  *              Matt Day        :       poll()->select() match BSD precisely on error
166  *              Alan Cox        :       New buffers
167  *              Marc Tamsky     :       Various sk->prot->retransmits and
168  *                                      sk->retransmits misupdating fixed.
169  *                                      Fixed tcp_write_timeout: stuck close,
170  *                                      and TCP syn retries gets used now.
171  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
172  *                                      ack if state is TCP_CLOSED.
173  *              Alan Cox        :       Look up device on a retransmit - routes may
174  *                                      change. Doesn't yet cope with MSS shrink right
175  *                                      but it's a start!
176  *              Marc Tamsky     :       Closing in closing fixes.
177  *              Mike Shaver     :       RFC1122 verifications.
178  *              Alan Cox        :       rcv_saddr errors.
179  *              Alan Cox        :       Block double connect().
180  *              Alan Cox        :       Small hooks for enSKIP.
181  *              Alexey Kuznetsov:       Path MTU discovery.
182  *              Alan Cox        :       Support soft errors.
183  *              Alan Cox        :       Fix MTU discovery pathological case
184  *                                      when the remote claims no mtu!
185  *              Marc Tamsky     :       TCP_CLOSE fix.
186  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
187  *                                      window but wrong (fixes NT lpd problems)
188  *              Pedro Roque     :       Better TCP window handling, delayed ack.
189  *              Joerg Reuter    :       No modification of locked buffers in
190  *                                      tcp_do_retransmit()
191  *              Eric Schenk     :       Changed receiver side silly window
192  *                                      avoidance algorithm to BSD style
193  *                                      algorithm. This doubles throughput
194  *                                      against machines running Solaris,
195  *                                      and seems to result in general
196  *                                      improvement.
197  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
198  *      Willy Konynenberg       :       Transparent proxying support.
199  *      Mike McLagan            :       Routing by source
200  *              Keith Owens     :       Do proper merging with partial SKB's in
201  *                                      tcp_do_sendmsg to avoid burstiness.
202  *              Eric Schenk     :       Fix fast close down bug with
203  *                                      shutdown() followed by close().
204  *              Andi Kleen      :       Make poll agree with SIGIO
205  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
206  *                                      lingertime == 0 (RFC 793 ABORT Call)
207  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
208  *                                      csum_and_copy_from_user() if possible.
209  *
210  *              This program is free software; you can redistribute it and/or
211  *              modify it under the terms of the GNU General Public License
212  *              as published by the Free Software Foundation; either version
213  *              2 of the License, or(at your option) any later version.
214  *
215  * Description of States:
216  *
217  *      TCP_SYN_SENT            sent a connection request, waiting for ack
218  *
219  *      TCP_SYN_RECV            received a connection request, sent ack,
220  *                              waiting for final ack in three-way handshake.
221  *
222  *      TCP_ESTABLISHED         connection established
223  *
224  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
225  *                              transmission of remaining buffered data
226  *
227  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
228  *                              to shutdown
229  *
230  *      TCP_CLOSING             both sides have shutdown but we still have
231  *                              data we have to finish sending
232  *
233  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
234  *                              closed, can only be entered from FIN_WAIT2
235  *                              or CLOSING.  Required because the other end
236  *                              may not have gotten our last ACK causing it
237  *                              to retransmit the data packet (which we ignore)
238  *
239  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
240  *                              us to finish writing our data and to shutdown
241  *                              (we have to close() to move on to LAST_ACK)
242  *
243  *      TCP_LAST_ACK            out side has shutdown after remote has
244  *                              shutdown.  There may still be data in our
245  *                              buffer that we have to finish sending
246  *
247  *      TCP_CLOSE               socket is finished
248  */
249
250 #include <linux/config.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/init.h>
256 #include <linux/smp_lock.h>
257 #include <linux/fs.h>
258 #include <linux/random.h>
259 #include <linux/bootmem.h>
260 #include <linux/cache.h>
261 #include <linux/err.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 = 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_HW;
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->sk_route_caps & NETIF_F_TSO)
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_HW;
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 && (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
960                      !icsk->icsk_ack.pingpong && !atomic_read(&sk->sk_rmem_alloc)))
961                         time_to_ack = 1;
962         }
963
964         /* We send an ACK if we can now advertise a non-zero window
965          * which has been raised "significantly".
966          *
967          * Even if window raised up to infinity, do not send window open ACK
968          * in states, where we will not receive more. It is useless.
969          */
970         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
971                 __u32 rcv_window_now = tcp_receive_window(tp);
972
973                 /* Optimize, __tcp_select_window() is not cheap. */
974                 if (2*rcv_window_now <= tp->window_clamp) {
975                         __u32 new_window = __tcp_select_window(sk);
976
977                         /* Send ACK now, if this read freed lots of space
978                          * in our buffer. Certainly, new_window is new window.
979                          * We can advertise it now, if it is not less than current one.
980                          * "Lots" means "at least twice" here.
981                          */
982                         if (new_window && new_window >= 2 * rcv_window_now)
983                                 time_to_ack = 1;
984                 }
985         }
986         if (time_to_ack)
987                 tcp_send_ack(sk);
988 }
989
990 static void tcp_prequeue_process(struct sock *sk)
991 {
992         struct sk_buff *skb;
993         struct tcp_sock *tp = tcp_sk(sk);
994
995         NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
996
997         /* RX process wants to run with disabled BHs, though it is not
998          * necessary */
999         local_bh_disable();
1000         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1001                 sk->sk_backlog_rcv(sk, skb);
1002         local_bh_enable();
1003
1004         /* Clear memory counter. */
1005         tp->ucopy.memory = 0;
1006 }
1007
1008 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1009 {
1010         struct sk_buff *skb;
1011         u32 offset;
1012
1013         skb_queue_walk(&sk->sk_receive_queue, skb) {
1014                 offset = seq - TCP_SKB_CB(skb)->seq;
1015                 if (skb->h.th->syn)
1016                         offset--;
1017                 if (offset < skb->len || skb->h.th->fin) {
1018                         *off = offset;
1019                         return skb;
1020                 }
1021         }
1022         return NULL;
1023 }
1024
1025 /*
1026  * This routine provides an alternative to tcp_recvmsg() for routines
1027  * that would like to handle copying from skbuffs directly in 'sendfile'
1028  * fashion.
1029  * Note:
1030  *      - It is assumed that the socket was locked by the caller.
1031  *      - The routine does not block.
1032  *      - At present, there is no support for reading OOB data
1033  *        or for 'peeking' the socket using this routine
1034  *        (although both would be easy to implement).
1035  */
1036 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1037                   sk_read_actor_t recv_actor)
1038 {
1039         struct sk_buff *skb;
1040         struct tcp_sock *tp = tcp_sk(sk);
1041         u32 seq = tp->copied_seq;
1042         u32 offset;
1043         int copied = 0;
1044
1045         if (sk->sk_state == TCP_LISTEN)
1046                 return -ENOTCONN;
1047         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1048                 if (offset < skb->len) {
1049                         size_t used, len;
1050
1051                         len = skb->len - offset;
1052                         /* Stop reading if we hit a patch of urgent data */
1053                         if (tp->urg_data) {
1054                                 u32 urg_offset = tp->urg_seq - seq;
1055                                 if (urg_offset < len)
1056                                         len = urg_offset;
1057                                 if (!len)
1058                                         break;
1059                         }
1060                         used = recv_actor(desc, skb, offset, len);
1061                         if (used <= len) {
1062                                 seq += used;
1063                                 copied += used;
1064                                 offset += used;
1065                         }
1066                         if (offset != skb->len)
1067                                 break;
1068                 }
1069                 if (skb->h.th->fin) {
1070                         sk_eat_skb(sk, skb, 0);
1071                         ++seq;
1072                         break;
1073                 }
1074                 sk_eat_skb(sk, skb, 0);
1075                 if (!desc->count)
1076                         break;
1077         }
1078         tp->copied_seq = seq;
1079
1080         tcp_rcv_space_adjust(sk);
1081
1082         /* Clean up data we have read: This will do ACK frames. */
1083         if (copied)
1084                 tcp_cleanup_rbuf(sk, copied);
1085         return copied;
1086 }
1087
1088 /*
1089  *      This routine copies from a sock struct into the user buffer.
1090  *
1091  *      Technical note: in 2.3 we work on _locked_ socket, so that
1092  *      tricks with *seq access order and skb->users are not required.
1093  *      Probably, code can be easily improved even more.
1094  */
1095
1096 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1097                 size_t len, int nonblock, int flags, int *addr_len)
1098 {
1099         struct tcp_sock *tp = tcp_sk(sk);
1100         int copied = 0;
1101         u32 peek_seq;
1102         u32 *seq;
1103         unsigned long used;
1104         int err;
1105         int target;             /* Read at least this many bytes */
1106         long timeo;
1107         struct task_struct *user_recv = NULL;
1108         int copied_early = 0;
1109
1110         lock_sock(sk);
1111
1112         TCP_CHECK_TIMER(sk);
1113
1114         err = -ENOTCONN;
1115         if (sk->sk_state == TCP_LISTEN)
1116                 goto out;
1117
1118         timeo = sock_rcvtimeo(sk, nonblock);
1119
1120         /* Urgent data needs to be handled specially. */
1121         if (flags & MSG_OOB)
1122                 goto recv_urg;
1123
1124         seq = &tp->copied_seq;
1125         if (flags & MSG_PEEK) {
1126                 peek_seq = tp->copied_seq;
1127                 seq = &peek_seq;
1128         }
1129
1130         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1131
1132 #ifdef CONFIG_NET_DMA
1133         tp->ucopy.dma_chan = NULL;
1134         preempt_disable();
1135         if ((len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1136             !sysctl_tcp_low_latency && __get_cpu_var(softnet_data.net_dma)) {
1137                 preempt_enable_no_resched();
1138                 tp->ucopy.pinned_list = dma_pin_iovec_pages(msg->msg_iov, len);
1139         } else
1140                 preempt_enable_no_resched();
1141 #endif
1142
1143         do {
1144                 struct sk_buff *skb;
1145                 u32 offset;
1146
1147                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1148                 if (tp->urg_data && tp->urg_seq == *seq) {
1149                         if (copied)
1150                                 break;
1151                         if (signal_pending(current)) {
1152                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1153                                 break;
1154                         }
1155                 }
1156
1157                 /* Next get a buffer. */
1158
1159                 skb = skb_peek(&sk->sk_receive_queue);
1160                 do {
1161                         if (!skb)
1162                                 break;
1163
1164                         /* Now that we have two receive queues this
1165                          * shouldn't happen.
1166                          */
1167                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1168                                 printk(KERN_INFO "recvmsg bug: copied %X "
1169                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1170                                 break;
1171                         }
1172                         offset = *seq - TCP_SKB_CB(skb)->seq;
1173                         if (skb->h.th->syn)
1174                                 offset--;
1175                         if (offset < skb->len)
1176                                 goto found_ok_skb;
1177                         if (skb->h.th->fin)
1178                                 goto found_fin_ok;
1179                         BUG_TRAP(flags & MSG_PEEK);
1180                         skb = skb->next;
1181                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1182
1183                 /* Well, if we have backlog, try to process it now yet. */
1184
1185                 if (copied >= target && !sk->sk_backlog.tail)
1186                         break;
1187
1188                 if (copied) {
1189                         if (sk->sk_err ||
1190                             sk->sk_state == TCP_CLOSE ||
1191                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1192                             !timeo ||
1193                             signal_pending(current) ||
1194                             (flags & MSG_PEEK))
1195                                 break;
1196                 } else {
1197                         if (sock_flag(sk, SOCK_DONE))
1198                                 break;
1199
1200                         if (sk->sk_err) {
1201                                 copied = sock_error(sk);
1202                                 break;
1203                         }
1204
1205                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1206                                 break;
1207
1208                         if (sk->sk_state == TCP_CLOSE) {
1209                                 if (!sock_flag(sk, SOCK_DONE)) {
1210                                         /* This occurs when user tries to read
1211                                          * from never connected socket.
1212                                          */
1213                                         copied = -ENOTCONN;
1214                                         break;
1215                                 }
1216                                 break;
1217                         }
1218
1219                         if (!timeo) {
1220                                 copied = -EAGAIN;
1221                                 break;
1222                         }
1223
1224                         if (signal_pending(current)) {
1225                                 copied = sock_intr_errno(timeo);
1226                                 break;
1227                         }
1228                 }
1229
1230                 tcp_cleanup_rbuf(sk, copied);
1231
1232                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1233                         /* Install new reader */
1234                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1235                                 user_recv = current;
1236                                 tp->ucopy.task = user_recv;
1237                                 tp->ucopy.iov = msg->msg_iov;
1238                         }
1239
1240                         tp->ucopy.len = len;
1241
1242                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1243                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1244
1245                         /* Ugly... If prequeue is not empty, we have to
1246                          * process it before releasing socket, otherwise
1247                          * order will be broken at second iteration.
1248                          * More elegant solution is required!!!
1249                          *
1250                          * Look: we have the following (pseudo)queues:
1251                          *
1252                          * 1. packets in flight
1253                          * 2. backlog
1254                          * 3. prequeue
1255                          * 4. receive_queue
1256                          *
1257                          * Each queue can be processed only if the next ones
1258                          * are empty. At this point we have empty receive_queue.
1259                          * But prequeue _can_ be not empty after 2nd iteration,
1260                          * when we jumped to start of loop because backlog
1261                          * processing added something to receive_queue.
1262                          * We cannot release_sock(), because backlog contains
1263                          * packets arrived _after_ prequeued ones.
1264                          *
1265                          * Shortly, algorithm is clear --- to process all
1266                          * the queues in order. We could make it more directly,
1267                          * requeueing packets from backlog to prequeue, if
1268                          * is not empty. It is more elegant, but eats cycles,
1269                          * unfortunately.
1270                          */
1271                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1272                                 goto do_prequeue;
1273
1274                         /* __ Set realtime policy in scheduler __ */
1275                 }
1276
1277                 if (copied >= target) {
1278                         /* Do not sleep, just process backlog. */
1279                         release_sock(sk);
1280                         lock_sock(sk);
1281                 } else
1282                         sk_wait_data(sk, &timeo);
1283
1284 #ifdef CONFIG_NET_DMA
1285                 tp->ucopy.wakeup = 0;
1286 #endif
1287
1288                 if (user_recv) {
1289                         int chunk;
1290
1291                         /* __ Restore normal policy in scheduler __ */
1292
1293                         if ((chunk = len - tp->ucopy.len) != 0) {
1294                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1295                                 len -= chunk;
1296                                 copied += chunk;
1297                         }
1298
1299                         if (tp->rcv_nxt == tp->copied_seq &&
1300                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1301 do_prequeue:
1302                                 tcp_prequeue_process(sk);
1303
1304                                 if ((chunk = len - tp->ucopy.len) != 0) {
1305                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1306                                         len -= chunk;
1307                                         copied += chunk;
1308                                 }
1309                         }
1310                 }
1311                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1312                         if (net_ratelimit())
1313                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1314                                        current->comm, current->pid);
1315                         peek_seq = tp->copied_seq;
1316                 }
1317                 continue;
1318
1319         found_ok_skb:
1320                 /* Ok so how much can we use? */
1321                 used = skb->len - offset;
1322                 if (len < used)
1323                         used = len;
1324
1325                 /* Do we have urgent data here? */
1326                 if (tp->urg_data) {
1327                         u32 urg_offset = tp->urg_seq - *seq;
1328                         if (urg_offset < used) {
1329                                 if (!urg_offset) {
1330                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1331                                                 ++*seq;
1332                                                 offset++;
1333                                                 used--;
1334                                                 if (!used)
1335                                                         goto skip_copy;
1336                                         }
1337                                 } else
1338                                         used = urg_offset;
1339                         }
1340                 }
1341
1342                 if (!(flags & MSG_TRUNC)) {
1343 #ifdef CONFIG_NET_DMA
1344                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1345                                 tp->ucopy.dma_chan = get_softnet_dma();
1346
1347                         if (tp->ucopy.dma_chan) {
1348                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1349                                         tp->ucopy.dma_chan, skb, offset,
1350                                         msg->msg_iov, used,
1351                                         tp->ucopy.pinned_list);
1352
1353                                 if (tp->ucopy.dma_cookie < 0) {
1354
1355                                         printk(KERN_ALERT "dma_cookie < 0\n");
1356
1357                                         /* Exception. Bailout! */
1358                                         if (!copied)
1359                                                 copied = -EFAULT;
1360                                         break;
1361                                 }
1362                                 if ((offset + used) == skb->len)
1363                                         copied_early = 1;
1364
1365                         } else
1366 #endif
1367                         {
1368                                 err = skb_copy_datagram_iovec(skb, offset,
1369                                                 msg->msg_iov, used);
1370                                 if (err) {
1371                                         /* Exception. Bailout! */
1372                                         if (!copied)
1373                                                 copied = -EFAULT;
1374                                         break;
1375                                 }
1376                         }
1377                 }
1378
1379                 *seq += used;
1380                 copied += used;
1381                 len -= used;
1382
1383                 tcp_rcv_space_adjust(sk);
1384
1385 skip_copy:
1386                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1387                         tp->urg_data = 0;
1388                         tcp_fast_path_check(sk, tp);
1389                 }
1390                 if (used + offset < skb->len)
1391                         continue;
1392
1393                 if (skb->h.th->fin)
1394                         goto found_fin_ok;
1395                 if (!(flags & MSG_PEEK)) {
1396                         sk_eat_skb(sk, skb, copied_early);
1397                         copied_early = 0;
1398                 }
1399                 continue;
1400
1401         found_fin_ok:
1402                 /* Process the FIN. */
1403                 ++*seq;
1404                 if (!(flags & MSG_PEEK)) {
1405                         sk_eat_skb(sk, skb, copied_early);
1406                         copied_early = 0;
1407                 }
1408                 break;
1409         } while (len > 0);
1410
1411         if (user_recv) {
1412                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1413                         int chunk;
1414
1415                         tp->ucopy.len = copied > 0 ? len : 0;
1416
1417                         tcp_prequeue_process(sk);
1418
1419                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1420                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1421                                 len -= chunk;
1422                                 copied += chunk;
1423                         }
1424                 }
1425
1426                 tp->ucopy.task = NULL;
1427                 tp->ucopy.len = 0;
1428         }
1429
1430 #ifdef CONFIG_NET_DMA
1431         if (tp->ucopy.dma_chan) {
1432                 struct sk_buff *skb;
1433                 dma_cookie_t done, used;
1434
1435                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1436
1437                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1438                                                  tp->ucopy.dma_cookie, &done,
1439                                                  &used) == DMA_IN_PROGRESS) {
1440                         /* do partial cleanup of sk_async_wait_queue */
1441                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1442                                (dma_async_is_complete(skb->dma_cookie, done,
1443                                                       used) == DMA_SUCCESS)) {
1444                                 __skb_dequeue(&sk->sk_async_wait_queue);
1445                                 kfree_skb(skb);
1446                         }
1447                 }
1448
1449                 /* Safe to free early-copied skbs now */
1450                 __skb_queue_purge(&sk->sk_async_wait_queue);
1451                 dma_chan_put(tp->ucopy.dma_chan);
1452                 tp->ucopy.dma_chan = NULL;
1453         }
1454         if (tp->ucopy.pinned_list) {
1455                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1456                 tp->ucopy.pinned_list = NULL;
1457         }
1458 #endif
1459
1460         /* According to UNIX98, msg_name/msg_namelen are ignored
1461          * on connected socket. I was just happy when found this 8) --ANK
1462          */
1463
1464         /* Clean up data we have read: This will do ACK frames. */
1465         tcp_cleanup_rbuf(sk, copied);
1466
1467         TCP_CHECK_TIMER(sk);
1468         release_sock(sk);
1469         return copied;
1470
1471 out:
1472         TCP_CHECK_TIMER(sk);
1473         release_sock(sk);
1474         return err;
1475
1476 recv_urg:
1477         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1478         goto out;
1479 }
1480
1481 /*
1482  *      State processing on a close. This implements the state shift for
1483  *      sending our FIN frame. Note that we only send a FIN for some
1484  *      states. A shutdown() may have already sent the FIN, or we may be
1485  *      closed.
1486  */
1487
1488 static const unsigned char new_state[16] = {
1489   /* current state:        new state:      action:      */
1490   /* (Invalid)          */ TCP_CLOSE,
1491   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1492   /* TCP_SYN_SENT       */ TCP_CLOSE,
1493   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1494   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1495   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1496   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1497   /* TCP_CLOSE          */ TCP_CLOSE,
1498   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1499   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1500   /* TCP_LISTEN         */ TCP_CLOSE,
1501   /* TCP_CLOSING        */ TCP_CLOSING,
1502 };
1503
1504 static int tcp_close_state(struct sock *sk)
1505 {
1506         int next = (int)new_state[sk->sk_state];
1507         int ns = next & TCP_STATE_MASK;
1508
1509         tcp_set_state(sk, ns);
1510
1511         return next & TCP_ACTION_FIN;
1512 }
1513
1514 /*
1515  *      Shutdown the sending side of a connection. Much like close except
1516  *      that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1517  */
1518
1519 void tcp_shutdown(struct sock *sk, int how)
1520 {
1521         /*      We need to grab some memory, and put together a FIN,
1522          *      and then put it into the queue to be sent.
1523          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1524          */
1525         if (!(how & SEND_SHUTDOWN))
1526                 return;
1527
1528         /* If we've already sent a FIN, or it's a closed state, skip this. */
1529         if ((1 << sk->sk_state) &
1530             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1531              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1532                 /* Clear out any half completed packets.  FIN if needed. */
1533                 if (tcp_close_state(sk))
1534                         tcp_send_fin(sk);
1535         }
1536 }
1537
1538 void tcp_close(struct sock *sk, long timeout)
1539 {
1540         struct sk_buff *skb;
1541         int data_was_unread = 0;
1542         int state;
1543
1544         lock_sock(sk);
1545         sk->sk_shutdown = SHUTDOWN_MASK;
1546
1547         if (sk->sk_state == TCP_LISTEN) {
1548                 tcp_set_state(sk, TCP_CLOSE);
1549
1550                 /* Special case. */
1551                 inet_csk_listen_stop(sk);
1552
1553                 goto adjudge_to_death;
1554         }
1555
1556         /*  We need to flush the recv. buffs.  We do this only on the
1557          *  descriptor close, not protocol-sourced closes, because the
1558          *  reader process may not have drained the data yet!
1559          */
1560         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1561                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1562                           skb->h.th->fin;
1563                 data_was_unread += len;
1564                 __kfree_skb(skb);
1565         }
1566
1567         sk_stream_mem_reclaim(sk);
1568
1569         /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1570          * 3.10, we send a RST here because data was lost.  To
1571          * witness the awful effects of the old behavior of always
1572          * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1573          * a bulk GET in an FTP client, suspend the process, wait
1574          * for the client to advertise a zero window, then kill -9
1575          * the FTP client, wheee...  Note: timeout is always zero
1576          * in such a case.
1577          */
1578         if (data_was_unread) {
1579                 /* Unread data was tossed, zap the connection. */
1580                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1581                 tcp_set_state(sk, TCP_CLOSE);
1582                 tcp_send_active_reset(sk, GFP_KERNEL);
1583         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1584                 /* Check zero linger _after_ checking for unread data. */
1585                 sk->sk_prot->disconnect(sk, 0);
1586                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1587         } else if (tcp_close_state(sk)) {
1588                 /* We FIN if the application ate all the data before
1589                  * zapping the connection.
1590                  */
1591
1592                 /* RED-PEN. Formally speaking, we have broken TCP state
1593                  * machine. State transitions:
1594                  *
1595                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1596                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1597                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1598                  *
1599                  * are legal only when FIN has been sent (i.e. in window),
1600                  * rather than queued out of window. Purists blame.
1601                  *
1602                  * F.e. "RFC state" is ESTABLISHED,
1603                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1604                  *
1605                  * The visible declinations are that sometimes
1606                  * we enter time-wait state, when it is not required really
1607                  * (harmless), do not send active resets, when they are
1608                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1609                  * they look as CLOSING or LAST_ACK for Linux)
1610                  * Probably, I missed some more holelets.
1611                  *                                              --ANK
1612                  */
1613                 tcp_send_fin(sk);
1614         }
1615
1616         sk_stream_wait_close(sk, timeout);
1617
1618 adjudge_to_death:
1619         state = sk->sk_state;
1620         sock_hold(sk);
1621         sock_orphan(sk);
1622         atomic_inc(sk->sk_prot->orphan_count);
1623
1624         /* It is the last release_sock in its life. It will remove backlog. */
1625         release_sock(sk);
1626
1627
1628         /* Now socket is owned by kernel and we acquire BH lock
1629            to finish close. No need to check for user refs.
1630          */
1631         local_bh_disable();
1632         bh_lock_sock(sk);
1633         BUG_TRAP(!sock_owned_by_user(sk));
1634
1635         /* Have we already been destroyed by a softirq or backlog? */
1636         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1637                 goto out;
1638
1639         /*      This is a (useful) BSD violating of the RFC. There is a
1640          *      problem with TCP as specified in that the other end could
1641          *      keep a socket open forever with no application left this end.
1642          *      We use a 3 minute timeout (about the same as BSD) then kill
1643          *      our end. If they send after that then tough - BUT: long enough
1644          *      that we won't make the old 4*rto = almost no time - whoops
1645          *      reset mistake.
1646          *
1647          *      Nope, it was not mistake. It is really desired behaviour
1648          *      f.e. on http servers, when such sockets are useless, but
1649          *      consume significant resources. Let's do it with special
1650          *      linger2 option.                                 --ANK
1651          */
1652
1653         if (sk->sk_state == TCP_FIN_WAIT2) {
1654                 struct tcp_sock *tp = tcp_sk(sk);
1655                 if (tp->linger2 < 0) {
1656                         tcp_set_state(sk, TCP_CLOSE);
1657                         tcp_send_active_reset(sk, GFP_ATOMIC);
1658                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1659                 } else {
1660                         const int tmo = tcp_fin_time(sk);
1661
1662                         if (tmo > TCP_TIMEWAIT_LEN) {
1663                                 inet_csk_reset_keepalive_timer(sk, tcp_fin_time(sk));
1664                         } else {
1665                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1666                                 goto out;
1667                         }
1668                 }
1669         }
1670         if (sk->sk_state != TCP_CLOSE) {
1671                 sk_stream_mem_reclaim(sk);
1672                 if (atomic_read(sk->sk_prot->orphan_count) > sysctl_tcp_max_orphans ||
1673                     (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1674                      atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1675                         if (net_ratelimit())
1676                                 printk(KERN_INFO "TCP: too many of orphaned "
1677                                        "sockets\n");
1678                         tcp_set_state(sk, TCP_CLOSE);
1679                         tcp_send_active_reset(sk, GFP_ATOMIC);
1680                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1681                 }
1682         }
1683
1684         if (sk->sk_state == TCP_CLOSE)
1685                 inet_csk_destroy_sock(sk);
1686         /* Otherwise, socket is reprieved until protocol close. */
1687
1688 out:
1689         bh_unlock_sock(sk);
1690         local_bh_enable();
1691         sock_put(sk);
1692 }
1693
1694 /* These states need RST on ABORT according to RFC793 */
1695
1696 static inline int tcp_need_reset(int state)
1697 {
1698         return (1 << state) &
1699                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1700                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1701 }
1702
1703 int tcp_disconnect(struct sock *sk, int flags)
1704 {
1705         struct inet_sock *inet = inet_sk(sk);
1706         struct inet_connection_sock *icsk = inet_csk(sk);
1707         struct tcp_sock *tp = tcp_sk(sk);
1708         int err = 0;
1709         int old_state = sk->sk_state;
1710
1711         if (old_state != TCP_CLOSE)
1712                 tcp_set_state(sk, TCP_CLOSE);
1713
1714         /* ABORT function of RFC793 */
1715         if (old_state == TCP_LISTEN) {
1716                 inet_csk_listen_stop(sk);
1717         } else if (tcp_need_reset(old_state) ||
1718                    (tp->snd_nxt != tp->write_seq &&
1719                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1720                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1721                  * states
1722                  */
1723                 tcp_send_active_reset(sk, gfp_any());
1724                 sk->sk_err = ECONNRESET;
1725         } else if (old_state == TCP_SYN_SENT)
1726                 sk->sk_err = ECONNRESET;
1727
1728         tcp_clear_xmit_timers(sk);
1729         __skb_queue_purge(&sk->sk_receive_queue);
1730         sk_stream_writequeue_purge(sk);
1731         __skb_queue_purge(&tp->out_of_order_queue);
1732 #ifdef CONFIG_NET_DMA
1733         __skb_queue_purge(&sk->sk_async_wait_queue);
1734 #endif
1735
1736         inet->dport = 0;
1737
1738         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1739                 inet_reset_saddr(sk);
1740
1741         sk->sk_shutdown = 0;
1742         sock_reset_flag(sk, SOCK_DONE);
1743         tp->srtt = 0;
1744         if ((tp->write_seq += tp->max_window + 2) == 0)
1745                 tp->write_seq = 1;
1746         icsk->icsk_backoff = 0;
1747         tp->snd_cwnd = 2;
1748         icsk->icsk_probes_out = 0;
1749         tp->packets_out = 0;
1750         tp->snd_ssthresh = 0x7fffffff;
1751         tp->snd_cwnd_cnt = 0;
1752         tp->bytes_acked = 0;
1753         tcp_set_ca_state(sk, TCP_CA_Open);
1754         tcp_clear_retrans(tp);
1755         inet_csk_delack_init(sk);
1756         sk->sk_send_head = NULL;
1757         tp->rx_opt.saw_tstamp = 0;
1758         tcp_sack_reset(&tp->rx_opt);
1759         __sk_dst_reset(sk);
1760
1761         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1762
1763         sk->sk_error_report(sk);
1764         return err;
1765 }
1766
1767 /*
1768  *      Socket option code for TCP.
1769  */
1770 static int do_tcp_setsockopt(struct sock *sk, int level,
1771                 int optname, char __user *optval, int optlen)
1772 {
1773         struct tcp_sock *tp = tcp_sk(sk);
1774         struct inet_connection_sock *icsk = inet_csk(sk);
1775         int val;
1776         int err = 0;
1777
1778         /* This is a string value all the others are int's */
1779         if (optname == TCP_CONGESTION) {
1780                 char name[TCP_CA_NAME_MAX];
1781
1782                 if (optlen < 1)
1783                         return -EINVAL;
1784
1785                 val = strncpy_from_user(name, optval,
1786                                         min(TCP_CA_NAME_MAX-1, optlen));
1787                 if (val < 0)
1788                         return -EFAULT;
1789                 name[val] = 0;
1790
1791                 lock_sock(sk);
1792                 err = tcp_set_congestion_control(sk, name);
1793                 release_sock(sk);
1794                 return err;
1795         }
1796
1797         if (optlen < sizeof(int))
1798                 return -EINVAL;
1799
1800         if (get_user(val, (int __user *)optval))
1801                 return -EFAULT;
1802
1803         lock_sock(sk);
1804
1805         switch (optname) {
1806         case TCP_MAXSEG:
1807                 /* Values greater than interface MTU won't take effect. However
1808                  * at the point when this call is done we typically don't yet
1809                  * know which interface is going to be used */
1810                 if (val < 8 || val > MAX_TCP_WINDOW) {
1811                         err = -EINVAL;
1812                         break;
1813                 }
1814                 tp->rx_opt.user_mss = val;
1815                 break;
1816
1817         case TCP_NODELAY:
1818                 if (val) {
1819                         /* TCP_NODELAY is weaker than TCP_CORK, so that
1820                          * this option on corked socket is remembered, but
1821                          * it is not activated until cork is cleared.
1822                          *
1823                          * However, when TCP_NODELAY is set we make
1824                          * an explicit push, which overrides even TCP_CORK
1825                          * for currently queued segments.
1826                          */
1827                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1828                         tcp_push_pending_frames(sk, tp);
1829                 } else {
1830                         tp->nonagle &= ~TCP_NAGLE_OFF;
1831                 }
1832                 break;
1833
1834         case TCP_CORK:
1835                 /* When set indicates to always queue non-full frames.
1836                  * Later the user clears this option and we transmit
1837                  * any pending partial frames in the queue.  This is
1838                  * meant to be used alongside sendfile() to get properly
1839                  * filled frames when the user (for example) must write
1840                  * out headers with a write() call first and then use
1841                  * sendfile to send out the data parts.
1842                  *
1843                  * TCP_CORK can be set together with TCP_NODELAY and it is
1844                  * stronger than TCP_NODELAY.
1845                  */
1846                 if (val) {
1847                         tp->nonagle |= TCP_NAGLE_CORK;
1848                 } else {
1849                         tp->nonagle &= ~TCP_NAGLE_CORK;
1850                         if (tp->nonagle&TCP_NAGLE_OFF)
1851                                 tp->nonagle |= TCP_NAGLE_PUSH;
1852                         tcp_push_pending_frames(sk, tp);
1853                 }
1854                 break;
1855
1856         case TCP_KEEPIDLE:
1857                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1858                         err = -EINVAL;
1859                 else {
1860                         tp->keepalive_time = val * HZ;
1861                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
1862                             !((1 << sk->sk_state) &
1863                               (TCPF_CLOSE | TCPF_LISTEN))) {
1864                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1865                                 if (tp->keepalive_time > elapsed)
1866                                         elapsed = tp->keepalive_time - elapsed;
1867                                 else
1868                                         elapsed = 0;
1869                                 inet_csk_reset_keepalive_timer(sk, elapsed);
1870                         }
1871                 }
1872                 break;
1873         case TCP_KEEPINTVL:
1874                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1875                         err = -EINVAL;
1876                 else
1877                         tp->keepalive_intvl = val * HZ;
1878                 break;
1879         case TCP_KEEPCNT:
1880                 if (val < 1 || val > MAX_TCP_KEEPCNT)
1881                         err = -EINVAL;
1882                 else
1883                         tp->keepalive_probes = val;
1884                 break;
1885         case TCP_SYNCNT:
1886                 if (val < 1 || val > MAX_TCP_SYNCNT)
1887                         err = -EINVAL;
1888                 else
1889                         icsk->icsk_syn_retries = val;
1890                 break;
1891
1892         case TCP_LINGER2:
1893                 if (val < 0)
1894                         tp->linger2 = -1;
1895                 else if (val > sysctl_tcp_fin_timeout / HZ)
1896                         tp->linger2 = 0;
1897                 else
1898                         tp->linger2 = val * HZ;
1899                 break;
1900
1901         case TCP_DEFER_ACCEPT:
1902                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1903                 if (val > 0) {
1904                         /* Translate value in seconds to number of
1905                          * retransmits */
1906                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1907                                val > ((TCP_TIMEOUT_INIT / HZ) <<
1908                                        icsk->icsk_accept_queue.rskq_defer_accept))
1909                                 icsk->icsk_accept_queue.rskq_defer_accept++;
1910                         icsk->icsk_accept_queue.rskq_defer_accept++;
1911                 }
1912                 break;
1913
1914         case TCP_WINDOW_CLAMP:
1915                 if (!val) {
1916                         if (sk->sk_state != TCP_CLOSE) {
1917                                 err = -EINVAL;
1918                                 break;
1919                         }
1920                         tp->window_clamp = 0;
1921                 } else
1922                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1923                                                 SOCK_MIN_RCVBUF / 2 : val;
1924                 break;
1925
1926         case TCP_QUICKACK:
1927                 if (!val) {
1928                         icsk->icsk_ack.pingpong = 1;
1929                 } else {
1930                         icsk->icsk_ack.pingpong = 0;
1931                         if ((1 << sk->sk_state) &
1932                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1933                             inet_csk_ack_scheduled(sk)) {
1934                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1935                                 tcp_cleanup_rbuf(sk, 1);
1936                                 if (!(val & 1))
1937                                         icsk->icsk_ack.pingpong = 1;
1938                         }
1939                 }
1940                 break;
1941
1942         default:
1943                 err = -ENOPROTOOPT;
1944                 break;
1945         };
1946         release_sock(sk);
1947         return err;
1948 }
1949
1950 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1951                    int optlen)
1952 {
1953         struct inet_connection_sock *icsk = inet_csk(sk);
1954
1955         if (level != SOL_TCP)
1956                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1957                                                      optval, optlen);
1958         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1959 }
1960
1961 #ifdef CONFIG_COMPAT
1962 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1963                           char __user *optval, int optlen)
1964 {
1965         if (level != SOL_TCP)
1966                 return inet_csk_compat_setsockopt(sk, level, optname,
1967                                                   optval, optlen);
1968         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1969 }
1970
1971 EXPORT_SYMBOL(compat_tcp_setsockopt);
1972 #endif
1973
1974 /* Return information about state of tcp endpoint in API format. */
1975 void tcp_get_info(struct sock *sk, struct tcp_info *info)
1976 {
1977         struct tcp_sock *tp = tcp_sk(sk);
1978         const struct inet_connection_sock *icsk = inet_csk(sk);
1979         u32 now = tcp_time_stamp;
1980
1981         memset(info, 0, sizeof(*info));
1982
1983         info->tcpi_state = sk->sk_state;
1984         info->tcpi_ca_state = icsk->icsk_ca_state;
1985         info->tcpi_retransmits = icsk->icsk_retransmits;
1986         info->tcpi_probes = icsk->icsk_probes_out;
1987         info->tcpi_backoff = icsk->icsk_backoff;
1988
1989         if (tp->rx_opt.tstamp_ok)
1990                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
1991         if (tp->rx_opt.sack_ok)
1992                 info->tcpi_options |= TCPI_OPT_SACK;
1993         if (tp->rx_opt.wscale_ok) {
1994                 info->tcpi_options |= TCPI_OPT_WSCALE;
1995                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
1996                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
1997         } 
1998
1999         if (tp->ecn_flags&TCP_ECN_OK)
2000                 info->tcpi_options |= TCPI_OPT_ECN;
2001
2002         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2003         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2004         info->tcpi_snd_mss = tp->mss_cache;
2005         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2006
2007         info->tcpi_unacked = tp->packets_out;
2008         info->tcpi_sacked = tp->sacked_out;
2009         info->tcpi_lost = tp->lost_out;
2010         info->tcpi_retrans = tp->retrans_out;
2011         info->tcpi_fackets = tp->fackets_out;
2012
2013         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2014         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2015         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2016
2017         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2018         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2019         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2020         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2021         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2022         info->tcpi_snd_cwnd = tp->snd_cwnd;
2023         info->tcpi_advmss = tp->advmss;
2024         info->tcpi_reordering = tp->reordering;
2025
2026         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2027         info->tcpi_rcv_space = tp->rcvq_space.space;
2028
2029         info->tcpi_total_retrans = tp->total_retrans;
2030 }
2031
2032 EXPORT_SYMBOL_GPL(tcp_get_info);
2033
2034 static int do_tcp_getsockopt(struct sock *sk, int level,
2035                 int optname, char __user *optval, int __user *optlen)
2036 {
2037         struct inet_connection_sock *icsk = inet_csk(sk);
2038         struct tcp_sock *tp = tcp_sk(sk);
2039         int val, len;
2040
2041         if (get_user(len, optlen))
2042                 return -EFAULT;
2043
2044         len = min_t(unsigned int, len, sizeof(int));
2045
2046         if (len < 0)
2047                 return -EINVAL;
2048
2049         switch (optname) {
2050         case TCP_MAXSEG:
2051                 val = tp->mss_cache;
2052                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2053                         val = tp->rx_opt.user_mss;
2054                 break;
2055         case TCP_NODELAY:
2056                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2057                 break;
2058         case TCP_CORK:
2059                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2060                 break;
2061         case TCP_KEEPIDLE:
2062                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2063                 break;
2064         case TCP_KEEPINTVL:
2065                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2066                 break;
2067         case TCP_KEEPCNT:
2068                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2069                 break;
2070         case TCP_SYNCNT:
2071                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2072                 break;
2073         case TCP_LINGER2:
2074                 val = tp->linger2;
2075                 if (val >= 0)
2076                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2077                 break;
2078         case TCP_DEFER_ACCEPT:
2079                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2080                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2081                 break;
2082         case TCP_WINDOW_CLAMP:
2083                 val = tp->window_clamp;
2084                 break;
2085         case TCP_INFO: {
2086                 struct tcp_info info;
2087
2088                 if (get_user(len, optlen))
2089                         return -EFAULT;
2090
2091                 tcp_get_info(sk, &info);
2092
2093                 len = min_t(unsigned int, len, sizeof(info));
2094                 if (put_user(len, optlen))
2095                         return -EFAULT;
2096                 if (copy_to_user(optval, &info, len))
2097                         return -EFAULT;
2098                 return 0;
2099         }
2100         case TCP_QUICKACK:
2101                 val = !icsk->icsk_ack.pingpong;
2102                 break;
2103
2104         case TCP_CONGESTION:
2105                 if (get_user(len, optlen))
2106                         return -EFAULT;
2107                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2108                 if (put_user(len, optlen))
2109                         return -EFAULT;
2110                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2111                         return -EFAULT;
2112                 return 0;
2113         default:
2114                 return -ENOPROTOOPT;
2115         };
2116
2117         if (put_user(len, optlen))
2118                 return -EFAULT;
2119         if (copy_to_user(optval, &val, len))
2120                 return -EFAULT;
2121         return 0;
2122 }
2123
2124 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2125                    int __user *optlen)
2126 {
2127         struct inet_connection_sock *icsk = inet_csk(sk);
2128
2129         if (level != SOL_TCP)
2130                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2131                                                      optval, optlen);
2132         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2133 }
2134
2135 #ifdef CONFIG_COMPAT
2136 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2137                           char __user *optval, int __user *optlen)
2138 {
2139         if (level != SOL_TCP)
2140                 return inet_csk_compat_getsockopt(sk, level, optname,
2141                                                   optval, optlen);
2142         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2143 }
2144
2145 EXPORT_SYMBOL(compat_tcp_getsockopt);
2146 #endif
2147
2148 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2149 {
2150         struct sk_buff *segs = ERR_PTR(-EINVAL);
2151         struct tcphdr *th;
2152         unsigned thlen;
2153         unsigned int seq;
2154         unsigned int delta;
2155         unsigned int oldlen;
2156         unsigned int len;
2157
2158         if (!pskb_may_pull(skb, sizeof(*th)))
2159                 goto out;
2160
2161         th = skb->h.th;
2162         thlen = th->doff * 4;
2163         if (thlen < sizeof(*th))
2164                 goto out;
2165
2166         if (!pskb_may_pull(skb, thlen))
2167                 goto out;
2168
2169         segs = NULL;
2170         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST))
2171                 goto out;
2172
2173         oldlen = (u16)~skb->len;
2174         __skb_pull(skb, thlen);
2175
2176         segs = skb_segment(skb, features);
2177         if (IS_ERR(segs))
2178                 goto out;
2179
2180         len = skb_shinfo(skb)->gso_size;
2181         delta = htonl(oldlen + (thlen + len));
2182
2183         skb = segs;
2184         th = skb->h.th;
2185         seq = ntohl(th->seq);
2186
2187         do {
2188                 th->fin = th->psh = 0;
2189
2190                 th->check = ~csum_fold(th->check + delta);
2191                 if (skb->ip_summed != CHECKSUM_HW)
2192                         th->check = csum_fold(csum_partial(skb->h.raw, thlen,
2193                                                            skb->csum));
2194
2195                 seq += len;
2196                 skb = skb->next;
2197                 th = skb->h.th;
2198
2199                 th->seq = htonl(seq);
2200                 th->cwr = 0;
2201         } while (skb->next);
2202
2203         delta = htonl(oldlen + (skb->tail - skb->h.raw) + skb->data_len);
2204         th->check = ~csum_fold(th->check + delta);
2205         if (skb->ip_summed != CHECKSUM_HW)
2206                 th->check = csum_fold(csum_partial(skb->h.raw, thlen,
2207                                                    skb->csum));
2208
2209 out:
2210         return segs;
2211 }
2212
2213 extern void __skb_cb_too_small_for_tcp(int, int);
2214 extern struct tcp_congestion_ops tcp_reno;
2215
2216 static __initdata unsigned long thash_entries;
2217 static int __init set_thash_entries(char *str)
2218 {
2219         if (!str)
2220                 return 0;
2221         thash_entries = simple_strtoul(str, &str, 0);
2222         return 1;
2223 }
2224 __setup("thash_entries=", set_thash_entries);
2225
2226 void __init tcp_init(void)
2227 {
2228         struct sk_buff *skb = NULL;
2229         unsigned long limit;
2230         int order, i, max_share;
2231
2232         if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2233                 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2234                                            sizeof(skb->cb));
2235
2236         tcp_hashinfo.bind_bucket_cachep =
2237                 kmem_cache_create("tcp_bind_bucket",
2238                                   sizeof(struct inet_bind_bucket), 0,
2239                                   SLAB_HWCACHE_ALIGN, NULL, NULL);
2240         if (!tcp_hashinfo.bind_bucket_cachep)
2241                 panic("tcp_init: Cannot alloc tcp_bind_bucket cache.");
2242
2243         /* Size and allocate the main established and bind bucket
2244          * hash tables.
2245          *
2246          * The methodology is similar to that of the buffer cache.
2247          */
2248         tcp_hashinfo.ehash =
2249                 alloc_large_system_hash("TCP established",
2250                                         sizeof(struct inet_ehash_bucket),
2251                                         thash_entries,
2252                                         (num_physpages >= 128 * 1024) ?
2253                                         13 : 15,
2254                                         HASH_HIGHMEM,
2255                                         &tcp_hashinfo.ehash_size,
2256                                         NULL,
2257                                         0);
2258         tcp_hashinfo.ehash_size = (1 << tcp_hashinfo.ehash_size) >> 1;
2259         for (i = 0; i < (tcp_hashinfo.ehash_size << 1); i++) {
2260                 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2261                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2262         }
2263
2264         tcp_hashinfo.bhash =
2265                 alloc_large_system_hash("TCP bind",
2266                                         sizeof(struct inet_bind_hashbucket),
2267                                         tcp_hashinfo.ehash_size,
2268                                         (num_physpages >= 128 * 1024) ?
2269                                         13 : 15,
2270                                         HASH_HIGHMEM,
2271                                         &tcp_hashinfo.bhash_size,
2272                                         NULL,
2273                                         64 * 1024);
2274         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2275         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2276                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2277                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2278         }
2279
2280         /* Try to be a bit smarter and adjust defaults depending
2281          * on available memory.
2282          */
2283         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2284                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2285                         order++)
2286                 ;
2287         if (order >= 4) {
2288                 sysctl_local_port_range[0] = 32768;
2289                 sysctl_local_port_range[1] = 61000;
2290                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2291                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2292                 sysctl_max_syn_backlog = 1024;
2293         } else if (order < 3) {
2294                 sysctl_local_port_range[0] = 1024 * (3 - order);
2295                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2296                 sysctl_tcp_max_orphans >>= (3 - order);
2297                 sysctl_max_syn_backlog = 128;
2298         }
2299
2300         sysctl_tcp_mem[0] =  768 << order;
2301         sysctl_tcp_mem[1] = 1024 << order;
2302         sysctl_tcp_mem[2] = 1536 << order;
2303
2304         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2305         max_share = min(4UL*1024*1024, limit);
2306
2307         sysctl_tcp_wmem[0] = SK_STREAM_MEM_QUANTUM;
2308         sysctl_tcp_wmem[1] = 16*1024;
2309         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2310
2311         sysctl_tcp_rmem[0] = SK_STREAM_MEM_QUANTUM;
2312         sysctl_tcp_rmem[1] = 87380;
2313         sysctl_tcp_rmem[2] = max(87380, max_share);
2314
2315         printk(KERN_INFO "TCP: Hash tables configured "
2316                "(established %d bind %d)\n",
2317                tcp_hashinfo.ehash_size << 1, tcp_hashinfo.bhash_size);
2318
2319         tcp_register_congestion_control(&tcp_reno);
2320 }
2321
2322 EXPORT_SYMBOL(tcp_close);
2323 EXPORT_SYMBOL(tcp_disconnect);
2324 EXPORT_SYMBOL(tcp_getsockopt);
2325 EXPORT_SYMBOL(tcp_ioctl);
2326 EXPORT_SYMBOL(tcp_poll);
2327 EXPORT_SYMBOL(tcp_read_sock);
2328 EXPORT_SYMBOL(tcp_recvmsg);
2329 EXPORT_SYMBOL(tcp_sendmsg);
2330 EXPORT_SYMBOL(tcp_sendpage);
2331 EXPORT_SYMBOL(tcp_setsockopt);
2332 EXPORT_SYMBOL(tcp_shutdown);
2333 EXPORT_SYMBOL(tcp_statistics);