[IPV4]: ip_build_and_send_pkt() annotations
[linux-2.6.git] / net / ipv4 / tcp.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp.c,v 1.216 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  *
22  * Fixes:
23  *              Alan Cox        :       Numerous verify_area() calls
24  *              Alan Cox        :       Set the ACK bit on a reset
25  *              Alan Cox        :       Stopped it crashing if it closed while
26  *                                      sk->inuse=1 and was trying to connect
27  *                                      (tcp_err()).
28  *              Alan Cox        :       All icmp error handling was broken
29  *                                      pointers passed where wrong and the
30  *                                      socket was looked up backwards. Nobody
31  *                                      tested any icmp error code obviously.
32  *              Alan Cox        :       tcp_err() now handled properly. It
33  *                                      wakes people on errors. poll
34  *                                      behaves and the icmp error race
35  *                                      has gone by moving it into sock.c
36  *              Alan Cox        :       tcp_send_reset() fixed to work for
37  *                                      everything not just packets for
38  *                                      unknown sockets.
39  *              Alan Cox        :       tcp option processing.
40  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
41  *                                      syn rule wrong]
42  *              Herp Rosmanith  :       More reset fixes
43  *              Alan Cox        :       No longer acks invalid rst frames.
44  *                                      Acking any kind of RST is right out.
45  *              Alan Cox        :       Sets an ignore me flag on an rst
46  *                                      receive otherwise odd bits of prattle
47  *                                      escape still
48  *              Alan Cox        :       Fixed another acking RST frame bug.
49  *                                      Should stop LAN workplace lockups.
50  *              Alan Cox        :       Some tidyups using the new skb list
51  *                                      facilities
52  *              Alan Cox        :       sk->keepopen now seems to work
53  *              Alan Cox        :       Pulls options out correctly on accepts
54  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
55  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
56  *                                      bit to skb ops.
57  *              Alan Cox        :       Tidied tcp_data to avoid a potential
58  *                                      nasty.
59  *              Alan Cox        :       Added some better commenting, as the
60  *                                      tcp is hard to follow
61  *              Alan Cox        :       Removed incorrect check for 20 * psh
62  *      Michael O'Reilly        :       ack < copied bug fix.
63  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
64  *              Alan Cox        :       FIN with no memory -> CRASH
65  *              Alan Cox        :       Added socket option proto entries.
66  *                                      Also added awareness of them to accept.
67  *              Alan Cox        :       Added TCP options (SOL_TCP)
68  *              Alan Cox        :       Switched wakeup calls to callbacks,
69  *                                      so the kernel can layer network
70  *                                      sockets.
71  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
72  *              Alan Cox        :       Handle FIN (more) properly (we hope).
73  *              Alan Cox        :       RST frames sent on unsynchronised
74  *                                      state ack error.
75  *              Alan Cox        :       Put in missing check for SYN bit.
76  *              Alan Cox        :       Added tcp_select_window() aka NET2E
77  *                                      window non shrink trick.
78  *              Alan Cox        :       Added a couple of small NET2E timer
79  *                                      fixes
80  *              Charles Hedrick :       TCP fixes
81  *              Toomas Tamm     :       TCP window fixes
82  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
83  *              Charles Hedrick :       Rewrote most of it to actually work
84  *              Linus           :       Rewrote tcp_read() and URG handling
85  *                                      completely
86  *              Gerhard Koerting:       Fixed some missing timer handling
87  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
88  *              Gerhard Koerting:       PC/TCP workarounds
89  *              Adam Caldwell   :       Assorted timer/timing errors
90  *              Matthew Dillon  :       Fixed another RST bug
91  *              Alan Cox        :       Move to kernel side addressing changes.
92  *              Alan Cox        :       Beginning work on TCP fastpathing
93  *                                      (not yet usable)
94  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
95  *              Alan Cox        :       TCP fast path debugging
96  *              Alan Cox        :       Window clamping
97  *              Michael Riepe   :       Bug in tcp_check()
98  *              Matt Dillon     :       More TCP improvements and RST bug fixes
99  *              Matt Dillon     :       Yet more small nasties remove from the
100  *                                      TCP code (Be very nice to this man if
101  *                                      tcp finally works 100%) 8)
102  *              Alan Cox        :       BSD accept semantics.
103  *              Alan Cox        :       Reset on closedown bug.
104  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
105  *              Michael Pall    :       Handle poll() after URG properly in
106  *                                      all cases.
107  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
108  *                                      (multi URG PUSH broke rlogin).
109  *              Michael Pall    :       Fix the multi URG PUSH problem in
110  *                                      tcp_readable(), poll() after URG
111  *                                      works now.
112  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
113  *                                      BSD api.
114  *              Alan Cox        :       Changed the semantics of sk->socket to
115  *                                      fix a race and a signal problem with
116  *                                      accept() and async I/O.
117  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
118  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
119  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
120  *                                      clients/servers which listen in on
121  *                                      fixed ports.
122  *              Alan Cox        :       Cleaned the above up and shrank it to
123  *                                      a sensible code size.
124  *              Alan Cox        :       Self connect lockup fix.
125  *              Alan Cox        :       No connect to multicast.
126  *              Ross Biro       :       Close unaccepted children on master
127  *                                      socket close.
128  *              Alan Cox        :       Reset tracing code.
129  *              Alan Cox        :       Spurious resets on shutdown.
130  *              Alan Cox        :       Giant 15 minute/60 second timer error
131  *              Alan Cox        :       Small whoops in polling before an
132  *                                      accept.
133  *              Alan Cox        :       Kept the state trace facility since
134  *                                      it's handy for debugging.
135  *              Alan Cox        :       More reset handler fixes.
136  *              Alan Cox        :       Started rewriting the code based on
137  *                                      the RFC's for other useful protocol
138  *                                      references see: Comer, KA9Q NOS, and
139  *                                      for a reference on the difference
140  *                                      between specifications and how BSD
141  *                                      works see the 4.4lite source.
142  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
143  *                                      close.
144  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
145  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
146  *              Alan Cox        :       Reimplemented timers as per the RFC
147  *                                      and using multiple timers for sanity.
148  *              Alan Cox        :       Small bug fixes, and a lot of new
149  *                                      comments.
150  *              Alan Cox        :       Fixed dual reader crash by locking
151  *                                      the buffers (much like datagram.c)
152  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
153  *                                      now gets fed up of retrying without
154  *                                      (even a no space) answer.
155  *              Alan Cox        :       Extracted closing code better
156  *              Alan Cox        :       Fixed the closing state machine to
157  *                                      resemble the RFC.
158  *              Alan Cox        :       More 'per spec' fixes.
159  *              Jorge Cwik      :       Even faster checksumming.
160  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
161  *                                      only frames. At least one pc tcp stack
162  *                                      generates them.
163  *              Alan Cox        :       Cache last socket.
164  *              Alan Cox        :       Per route irtt.
165  *              Matt Day        :       poll()->select() match BSD precisely on error
166  *              Alan Cox        :       New buffers
167  *              Marc Tamsky     :       Various sk->prot->retransmits and
168  *                                      sk->retransmits misupdating fixed.
169  *                                      Fixed tcp_write_timeout: stuck close,
170  *                                      and TCP syn retries gets used now.
171  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
172  *                                      ack if state is TCP_CLOSED.
173  *              Alan Cox        :       Look up device on a retransmit - routes may
174  *                                      change. Doesn't yet cope with MSS shrink right
175  *                                      but it's a start!
176  *              Marc Tamsky     :       Closing in closing fixes.
177  *              Mike Shaver     :       RFC1122 verifications.
178  *              Alan Cox        :       rcv_saddr errors.
179  *              Alan Cox        :       Block double connect().
180  *              Alan Cox        :       Small hooks for enSKIP.
181  *              Alexey Kuznetsov:       Path MTU discovery.
182  *              Alan Cox        :       Support soft errors.
183  *              Alan Cox        :       Fix MTU discovery pathological case
184  *                                      when the remote claims no mtu!
185  *              Marc Tamsky     :       TCP_CLOSE fix.
186  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
187  *                                      window but wrong (fixes NT lpd problems)
188  *              Pedro Roque     :       Better TCP window handling, delayed ack.
189  *              Joerg Reuter    :       No modification of locked buffers in
190  *                                      tcp_do_retransmit()
191  *              Eric Schenk     :       Changed receiver side silly window
192  *                                      avoidance algorithm to BSD style
193  *                                      algorithm. This doubles throughput
194  *                                      against machines running Solaris,
195  *                                      and seems to result in general
196  *                                      improvement.
197  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
198  *      Willy Konynenberg       :       Transparent proxying support.
199  *      Mike McLagan            :       Routing by source
200  *              Keith Owens     :       Do proper merging with partial SKB's in
201  *                                      tcp_do_sendmsg to avoid burstiness.
202  *              Eric Schenk     :       Fix fast close down bug with
203  *                                      shutdown() followed by close().
204  *              Andi Kleen      :       Make poll agree with SIGIO
205  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
206  *                                      lingertime == 0 (RFC 793 ABORT Call)
207  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
208  *                                      csum_and_copy_from_user() if possible.
209  *
210  *              This program is free software; you can redistribute it and/or
211  *              modify it under the terms of the GNU General Public License
212  *              as published by the Free Software Foundation; either version
213  *              2 of the License, or(at your option) any later version.
214  *
215  * Description of States:
216  *
217  *      TCP_SYN_SENT            sent a connection request, waiting for ack
218  *
219  *      TCP_SYN_RECV            received a connection request, sent ack,
220  *                              waiting for final ack in three-way handshake.
221  *
222  *      TCP_ESTABLISHED         connection established
223  *
224  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
225  *                              transmission of remaining buffered data
226  *
227  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
228  *                              to shutdown
229  *
230  *      TCP_CLOSING             both sides have shutdown but we still have
231  *                              data we have to finish sending
232  *
233  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
234  *                              closed, can only be entered from FIN_WAIT2
235  *                              or CLOSING.  Required because the other end
236  *                              may not have gotten our last ACK causing it
237  *                              to retransmit the data packet (which we ignore)
238  *
239  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
240  *                              us to finish writing our data and to shutdown
241  *                              (we have to close() to move on to LAST_ACK)
242  *
243  *      TCP_LAST_ACK            out side has shutdown after remote has
244  *                              shutdown.  There may still be data in our
245  *                              buffer that we have to finish sending
246  *
247  *      TCP_CLOSE               socket is finished
248  */
249
250 #include <linux/module.h>
251 #include <linux/types.h>
252 #include <linux/fcntl.h>
253 #include <linux/poll.h>
254 #include <linux/init.h>
255 #include <linux/smp_lock.h>
256 #include <linux/fs.h>
257 #include <linux/random.h>
258 #include <linux/bootmem.h>
259 #include <linux/cache.h>
260 #include <linux/err.h>
261
262 #include <net/icmp.h>
263 #include <net/tcp.h>
264 #include <net/xfrm.h>
265 #include <net/ip.h>
266 #include <net/netdma.h>
267
268 #include <asm/uaccess.h>
269 #include <asm/ioctls.h>
270
271 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
272
273 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
274
275 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
276
277 EXPORT_SYMBOL_GPL(tcp_orphan_count);
278
279 int sysctl_tcp_mem[3] __read_mostly;
280 int sysctl_tcp_wmem[3] __read_mostly;
281 int sysctl_tcp_rmem[3] __read_mostly;
282
283 EXPORT_SYMBOL(sysctl_tcp_mem);
284 EXPORT_SYMBOL(sysctl_tcp_rmem);
285 EXPORT_SYMBOL(sysctl_tcp_wmem);
286
287 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
288 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
289
290 EXPORT_SYMBOL(tcp_memory_allocated);
291 EXPORT_SYMBOL(tcp_sockets_allocated);
292
293 /*
294  * Pressure flag: try to collapse.
295  * Technical note: it is used by multiple contexts non atomically.
296  * All the sk_stream_mem_schedule() is of this nature: accounting
297  * is strict, actions are advisory and have some latency.
298  */
299 int tcp_memory_pressure;
300
301 EXPORT_SYMBOL(tcp_memory_pressure);
302
303 void tcp_enter_memory_pressure(void)
304 {
305         if (!tcp_memory_pressure) {
306                 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
307                 tcp_memory_pressure = 1;
308         }
309 }
310
311 EXPORT_SYMBOL(tcp_enter_memory_pressure);
312
313 /*
314  *      Wait for a TCP event.
315  *
316  *      Note that we don't need to lock the socket, as the upper poll layers
317  *      take care of normal races (between the test and the event) and we don't
318  *      go look at any of the socket buffers directly.
319  */
320 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
321 {
322         unsigned int mask;
323         struct sock *sk = sock->sk;
324         struct tcp_sock *tp = tcp_sk(sk);
325
326         poll_wait(file, sk->sk_sleep, wait);
327         if (sk->sk_state == TCP_LISTEN)
328                 return inet_csk_listen_poll(sk);
329
330         /* Socket is not locked. We are protected from async events
331            by poll logic and correct handling of state changes
332            made by another threads is impossible in any case.
333          */
334
335         mask = 0;
336         if (sk->sk_err)
337                 mask = POLLERR;
338
339         /*
340          * POLLHUP is certainly not done right. But poll() doesn't
341          * have a notion of HUP in just one direction, and for a
342          * socket the read side is more interesting.
343          *
344          * Some poll() documentation says that POLLHUP is incompatible
345          * with the POLLOUT/POLLWR flags, so somebody should check this
346          * all. But careful, it tends to be safer to return too many
347          * bits than too few, and you can easily break real applications
348          * if you don't tell them that something has hung up!
349          *
350          * Check-me.
351          *
352          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
353          * our fs/select.c). It means that after we received EOF,
354          * poll always returns immediately, making impossible poll() on write()
355          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
356          * if and only if shutdown has been made in both directions.
357          * Actually, it is interesting to look how Solaris and DUX
358          * solve this dilemma. I would prefer, if PULLHUP were maskable,
359          * then we could set it on SND_SHUTDOWN. BTW examples given
360          * in Stevens' books assume exactly this behaviour, it explains
361          * why PULLHUP is incompatible with POLLOUT.    --ANK
362          *
363          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
364          * blocking on fresh not-connected or disconnected socket. --ANK
365          */
366         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
367                 mask |= POLLHUP;
368         if (sk->sk_shutdown & RCV_SHUTDOWN)
369                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
370
371         /* Connected? */
372         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
373                 /* Potential race condition. If read of tp below will
374                  * escape above sk->sk_state, we can be illegally awaken
375                  * in SYN_* states. */
376                 if ((tp->rcv_nxt != tp->copied_seq) &&
377                     (tp->urg_seq != tp->copied_seq ||
378                      tp->rcv_nxt != tp->copied_seq + 1 ||
379                      sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
380                         mask |= POLLIN | POLLRDNORM;
381
382                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
383                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
384                                 mask |= POLLOUT | POLLWRNORM;
385                         } else {  /* send SIGIO later */
386                                 set_bit(SOCK_ASYNC_NOSPACE,
387                                         &sk->sk_socket->flags);
388                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
389
390                                 /* Race breaker. If space is freed after
391                                  * wspace test but before the flags are set,
392                                  * IO signal will be lost.
393                                  */
394                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
395                                         mask |= POLLOUT | POLLWRNORM;
396                         }
397                 }
398
399                 if (tp->urg_data & TCP_URG_VALID)
400                         mask |= POLLPRI;
401         }
402         return mask;
403 }
404
405 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
406 {
407         struct tcp_sock *tp = tcp_sk(sk);
408         int answ;
409
410         switch (cmd) {
411         case SIOCINQ:
412                 if (sk->sk_state == TCP_LISTEN)
413                         return -EINVAL;
414
415                 lock_sock(sk);
416                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
417                         answ = 0;
418                 else if (sock_flag(sk, SOCK_URGINLINE) ||
419                          !tp->urg_data ||
420                          before(tp->urg_seq, tp->copied_seq) ||
421                          !before(tp->urg_seq, tp->rcv_nxt)) {
422                         answ = tp->rcv_nxt - tp->copied_seq;
423
424                         /* Subtract 1, if FIN is in queue. */
425                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
426                                 answ -=
427                        ((struct sk_buff *)sk->sk_receive_queue.prev)->h.th->fin;
428                 } else
429                         answ = tp->urg_seq - tp->copied_seq;
430                 release_sock(sk);
431                 break;
432         case SIOCATMARK:
433                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
434                 break;
435         case SIOCOUTQ:
436                 if (sk->sk_state == TCP_LISTEN)
437                         return -EINVAL;
438
439                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
440                         answ = 0;
441                 else
442                         answ = tp->write_seq - tp->snd_una;
443                 break;
444         default:
445                 return -ENOIOCTLCMD;
446         };
447
448         return put_user(answ, (int __user *)arg);
449 }
450
451 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
452 {
453         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
454         tp->pushed_seq = tp->write_seq;
455 }
456
457 static inline int forced_push(struct tcp_sock *tp)
458 {
459         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
460 }
461
462 static inline void skb_entail(struct sock *sk, struct tcp_sock *tp,
463                               struct sk_buff *skb)
464 {
465         skb->csum = 0;
466         TCP_SKB_CB(skb)->seq = tp->write_seq;
467         TCP_SKB_CB(skb)->end_seq = tp->write_seq;
468         TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
469         TCP_SKB_CB(skb)->sacked = 0;
470         skb_header_release(skb);
471         __skb_queue_tail(&sk->sk_write_queue, skb);
472         sk_charge_skb(sk, skb);
473         if (!sk->sk_send_head)
474                 sk->sk_send_head = skb;
475         if (tp->nonagle & TCP_NAGLE_PUSH)
476                 tp->nonagle &= ~TCP_NAGLE_PUSH; 
477 }
478
479 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
480                                 struct sk_buff *skb)
481 {
482         if (flags & MSG_OOB) {
483                 tp->urg_mode = 1;
484                 tp->snd_up = tp->write_seq;
485                 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
486         }
487 }
488
489 static inline void tcp_push(struct sock *sk, struct tcp_sock *tp, int flags,
490                             int mss_now, int nonagle)
491 {
492         if (sk->sk_send_head) {
493                 struct sk_buff *skb = sk->sk_write_queue.prev;
494                 if (!(flags & MSG_MORE) || forced_push(tp))
495                         tcp_mark_push(tp, skb);
496                 tcp_mark_urg(tp, flags, skb);
497                 __tcp_push_pending_frames(sk, tp, mss_now,
498                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
499         }
500 }
501
502 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
503                          size_t psize, int flags)
504 {
505         struct tcp_sock *tp = tcp_sk(sk);
506         int mss_now, size_goal;
507         int err;
508         ssize_t copied;
509         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
510
511         /* Wait for a connection to finish. */
512         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
513                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
514                         goto out_err;
515
516         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
517
518         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
519         size_goal = tp->xmit_size_goal;
520         copied = 0;
521
522         err = -EPIPE;
523         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
524                 goto do_error;
525
526         while (psize > 0) {
527                 struct sk_buff *skb = sk->sk_write_queue.prev;
528                 struct page *page = pages[poffset / PAGE_SIZE];
529                 int copy, i, can_coalesce;
530                 int offset = poffset % PAGE_SIZE;
531                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
532
533                 if (!sk->sk_send_head || (copy = size_goal - skb->len) <= 0) {
534 new_segment:
535                         if (!sk_stream_memory_free(sk))
536                                 goto wait_for_sndbuf;
537
538                         skb = sk_stream_alloc_pskb(sk, 0, 0,
539                                                    sk->sk_allocation);
540                         if (!skb)
541                                 goto wait_for_memory;
542
543                         skb_entail(sk, tp, skb);
544                         copy = size_goal;
545                 }
546
547                 if (copy > size)
548                         copy = size;
549
550                 i = skb_shinfo(skb)->nr_frags;
551                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
552                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
553                         tcp_mark_push(tp, skb);
554                         goto new_segment;
555                 }
556                 if (!sk_stream_wmem_schedule(sk, copy))
557                         goto wait_for_memory;
558                 
559                 if (can_coalesce) {
560                         skb_shinfo(skb)->frags[i - 1].size += copy;
561                 } else {
562                         get_page(page);
563                         skb_fill_page_desc(skb, i, page, offset, copy);
564                 }
565
566                 skb->len += copy;
567                 skb->data_len += copy;
568                 skb->truesize += copy;
569                 sk->sk_wmem_queued += copy;
570                 sk->sk_forward_alloc -= copy;
571                 skb->ip_summed = CHECKSUM_PARTIAL;
572                 tp->write_seq += copy;
573                 TCP_SKB_CB(skb)->end_seq += copy;
574                 skb_shinfo(skb)->gso_segs = 0;
575
576                 if (!copied)
577                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
578
579                 copied += copy;
580                 poffset += copy;
581                 if (!(psize -= copy))
582                         goto out;
583
584                 if (skb->len < mss_now || (flags & MSG_OOB))
585                         continue;
586
587                 if (forced_push(tp)) {
588                         tcp_mark_push(tp, skb);
589                         __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
590                 } else if (skb == sk->sk_send_head)
591                         tcp_push_one(sk, mss_now);
592                 continue;
593
594 wait_for_sndbuf:
595                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
596 wait_for_memory:
597                 if (copied)
598                         tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
599
600                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
601                         goto do_error;
602
603                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
604                 size_goal = tp->xmit_size_goal;
605         }
606
607 out:
608         if (copied)
609                 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
610         return copied;
611
612 do_error:
613         if (copied)
614                 goto out;
615 out_err:
616         return sk_stream_error(sk, flags, err);
617 }
618
619 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
620                      size_t size, int flags)
621 {
622         ssize_t res;
623         struct sock *sk = sock->sk;
624
625         if (!(sk->sk_route_caps & NETIF_F_SG) ||
626             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
627                 return sock_no_sendpage(sock, page, offset, size, flags);
628
629         lock_sock(sk);
630         TCP_CHECK_TIMER(sk);
631         res = do_tcp_sendpages(sk, &page, offset, size, flags);
632         TCP_CHECK_TIMER(sk);
633         release_sock(sk);
634         return res;
635 }
636
637 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
638 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
639
640 static inline int select_size(struct sock *sk, struct tcp_sock *tp)
641 {
642         int tmp = tp->mss_cache;
643
644         if (sk->sk_route_caps & NETIF_F_SG) {
645                 if (sk_can_gso(sk))
646                         tmp = 0;
647                 else {
648                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
649
650                         if (tmp >= pgbreak &&
651                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
652                                 tmp = pgbreak;
653                 }
654         }
655
656         return tmp;
657 }
658
659 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
660                 size_t size)
661 {
662         struct iovec *iov;
663         struct tcp_sock *tp = tcp_sk(sk);
664         struct sk_buff *skb;
665         int iovlen, flags;
666         int mss_now, size_goal;
667         int err, copied;
668         long timeo;
669
670         lock_sock(sk);
671         TCP_CHECK_TIMER(sk);
672
673         flags = msg->msg_flags;
674         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
675
676         /* Wait for a connection to finish. */
677         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
678                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
679                         goto out_err;
680
681         /* This should be in poll */
682         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
683
684         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
685         size_goal = tp->xmit_size_goal;
686
687         /* Ok commence sending. */
688         iovlen = msg->msg_iovlen;
689         iov = msg->msg_iov;
690         copied = 0;
691
692         err = -EPIPE;
693         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
694                 goto do_error;
695
696         while (--iovlen >= 0) {
697                 int seglen = iov->iov_len;
698                 unsigned char __user *from = iov->iov_base;
699
700                 iov++;
701
702                 while (seglen > 0) {
703                         int copy;
704
705                         skb = sk->sk_write_queue.prev;
706
707                         if (!sk->sk_send_head ||
708                             (copy = size_goal - skb->len) <= 0) {
709
710 new_segment:
711                                 /* Allocate new segment. If the interface is SG,
712                                  * allocate skb fitting to single page.
713                                  */
714                                 if (!sk_stream_memory_free(sk))
715                                         goto wait_for_sndbuf;
716
717                                 skb = sk_stream_alloc_pskb(sk, select_size(sk, tp),
718                                                            0, sk->sk_allocation);
719                                 if (!skb)
720                                         goto wait_for_memory;
721
722                                 /*
723                                  * Check whether we can use HW checksum.
724                                  */
725                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
726                                         skb->ip_summed = CHECKSUM_PARTIAL;
727
728                                 skb_entail(sk, tp, skb);
729                                 copy = size_goal;
730                         }
731
732                         /* Try to append data to the end of skb. */
733                         if (copy > seglen)
734                                 copy = seglen;
735
736                         /* Where to copy to? */
737                         if (skb_tailroom(skb) > 0) {
738                                 /* We have some space in skb head. Superb! */
739                                 if (copy > skb_tailroom(skb))
740                                         copy = skb_tailroom(skb);
741                                 if ((err = skb_add_data(skb, from, copy)) != 0)
742                                         goto do_fault;
743                         } else {
744                                 int merge = 0;
745                                 int i = skb_shinfo(skb)->nr_frags;
746                                 struct page *page = TCP_PAGE(sk);
747                                 int off = TCP_OFF(sk);
748
749                                 if (skb_can_coalesce(skb, i, page, off) &&
750                                     off != PAGE_SIZE) {
751                                         /* We can extend the last page
752                                          * fragment. */
753                                         merge = 1;
754                                 } else if (i == MAX_SKB_FRAGS ||
755                                            (!i &&
756                                            !(sk->sk_route_caps & NETIF_F_SG))) {
757                                         /* Need to add new fragment and cannot
758                                          * do this because interface is non-SG,
759                                          * or because all the page slots are
760                                          * busy. */
761                                         tcp_mark_push(tp, skb);
762                                         goto new_segment;
763                                 } else if (page) {
764                                         if (off == PAGE_SIZE) {
765                                                 put_page(page);
766                                                 TCP_PAGE(sk) = page = NULL;
767                                                 off = 0;
768                                         }
769                                 } else
770                                         off = 0;
771
772                                 if (copy > PAGE_SIZE - off)
773                                         copy = PAGE_SIZE - off;
774
775                                 if (!sk_stream_wmem_schedule(sk, copy))
776                                         goto wait_for_memory;
777
778                                 if (!page) {
779                                         /* Allocate new cache page. */
780                                         if (!(page = sk_stream_alloc_page(sk)))
781                                                 goto wait_for_memory;
782                                 }
783
784                                 /* Time to copy data. We are close to
785                                  * the end! */
786                                 err = skb_copy_to_page(sk, from, skb, page,
787                                                        off, copy);
788                                 if (err) {
789                                         /* If this page was new, give it to the
790                                          * socket so it does not get leaked.
791                                          */
792                                         if (!TCP_PAGE(sk)) {
793                                                 TCP_PAGE(sk) = page;
794                                                 TCP_OFF(sk) = 0;
795                                         }
796                                         goto do_error;
797                                 }
798
799                                 /* Update the skb. */
800                                 if (merge) {
801                                         skb_shinfo(skb)->frags[i - 1].size +=
802                                                                         copy;
803                                 } else {
804                                         skb_fill_page_desc(skb, i, page, off, copy);
805                                         if (TCP_PAGE(sk)) {
806                                                 get_page(page);
807                                         } else if (off + copy < PAGE_SIZE) {
808                                                 get_page(page);
809                                                 TCP_PAGE(sk) = page;
810                                         }
811                                 }
812
813                                 TCP_OFF(sk) = off + copy;
814                         }
815
816                         if (!copied)
817                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
818
819                         tp->write_seq += copy;
820                         TCP_SKB_CB(skb)->end_seq += copy;
821                         skb_shinfo(skb)->gso_segs = 0;
822
823                         from += copy;
824                         copied += copy;
825                         if ((seglen -= copy) == 0 && iovlen == 0)
826                                 goto out;
827
828                         if (skb->len < mss_now || (flags & MSG_OOB))
829                                 continue;
830
831                         if (forced_push(tp)) {
832                                 tcp_mark_push(tp, skb);
833                                 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
834                         } else if (skb == sk->sk_send_head)
835                                 tcp_push_one(sk, mss_now);
836                         continue;
837
838 wait_for_sndbuf:
839                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
840 wait_for_memory:
841                         if (copied)
842                                 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
843
844                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
845                                 goto do_error;
846
847                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
848                         size_goal = tp->xmit_size_goal;
849                 }
850         }
851
852 out:
853         if (copied)
854                 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
855         TCP_CHECK_TIMER(sk);
856         release_sock(sk);
857         return copied;
858
859 do_fault:
860         if (!skb->len) {
861                 if (sk->sk_send_head == skb)
862                         sk->sk_send_head = NULL;
863                 __skb_unlink(skb, &sk->sk_write_queue);
864                 sk_stream_free_skb(sk, skb);
865         }
866
867 do_error:
868         if (copied)
869                 goto out;
870 out_err:
871         err = sk_stream_error(sk, flags, err);
872         TCP_CHECK_TIMER(sk);
873         release_sock(sk);
874         return err;
875 }
876
877 /*
878  *      Handle reading urgent data. BSD has very simple semantics for
879  *      this, no blocking and very strange errors 8)
880  */
881
882 static int tcp_recv_urg(struct sock *sk, long timeo,
883                         struct msghdr *msg, int len, int flags,
884                         int *addr_len)
885 {
886         struct tcp_sock *tp = tcp_sk(sk);
887
888         /* No URG data to read. */
889         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
890             tp->urg_data == TCP_URG_READ)
891                 return -EINVAL; /* Yes this is right ! */
892
893         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
894                 return -ENOTCONN;
895
896         if (tp->urg_data & TCP_URG_VALID) {
897                 int err = 0;
898                 char c = tp->urg_data;
899
900                 if (!(flags & MSG_PEEK))
901                         tp->urg_data = TCP_URG_READ;
902
903                 /* Read urgent data. */
904                 msg->msg_flags |= MSG_OOB;
905
906                 if (len > 0) {
907                         if (!(flags & MSG_TRUNC))
908                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
909                         len = 1;
910                 } else
911                         msg->msg_flags |= MSG_TRUNC;
912
913                 return err ? -EFAULT : len;
914         }
915
916         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
917                 return 0;
918
919         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
920          * the available implementations agree in this case:
921          * this call should never block, independent of the
922          * blocking state of the socket.
923          * Mike <pall@rz.uni-karlsruhe.de>
924          */
925         return -EAGAIN;
926 }
927
928 /* Clean up the receive buffer for full frames taken by the user,
929  * then send an ACK if necessary.  COPIED is the number of bytes
930  * tcp_recvmsg has given to the user so far, it speeds up the
931  * calculation of whether or not we must ACK for the sake of
932  * a window update.
933  */
934 void tcp_cleanup_rbuf(struct sock *sk, int copied)
935 {
936         struct tcp_sock *tp = tcp_sk(sk);
937         int time_to_ack = 0;
938
939 #if TCP_DEBUG
940         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
941
942         BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
943 #endif
944
945         if (inet_csk_ack_scheduled(sk)) {
946                 const struct inet_connection_sock *icsk = inet_csk(sk);
947                    /* Delayed ACKs frequently hit locked sockets during bulk
948                     * receive. */
949                 if (icsk->icsk_ack.blocked ||
950                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
951                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
952                     /*
953                      * If this read emptied read buffer, we send ACK, if
954                      * connection is not bidirectional, user drained
955                      * receive buffer and there was a small segment
956                      * in queue.
957                      */
958                     (copied > 0 &&
959                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
960                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
961                        !icsk->icsk_ack.pingpong)) &&
962                       !atomic_read(&sk->sk_rmem_alloc)))
963                         time_to_ack = 1;
964         }
965
966         /* We send an ACK if we can now advertise a non-zero window
967          * which has been raised "significantly".
968          *
969          * Even if window raised up to infinity, do not send window open ACK
970          * in states, where we will not receive more. It is useless.
971          */
972         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
973                 __u32 rcv_window_now = tcp_receive_window(tp);
974
975                 /* Optimize, __tcp_select_window() is not cheap. */
976                 if (2*rcv_window_now <= tp->window_clamp) {
977                         __u32 new_window = __tcp_select_window(sk);
978
979                         /* Send ACK now, if this read freed lots of space
980                          * in our buffer. Certainly, new_window is new window.
981                          * We can advertise it now, if it is not less than current one.
982                          * "Lots" means "at least twice" here.
983                          */
984                         if (new_window && new_window >= 2 * rcv_window_now)
985                                 time_to_ack = 1;
986                 }
987         }
988         if (time_to_ack)
989                 tcp_send_ack(sk);
990 }
991
992 static void tcp_prequeue_process(struct sock *sk)
993 {
994         struct sk_buff *skb;
995         struct tcp_sock *tp = tcp_sk(sk);
996
997         NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
998
999         /* RX process wants to run with disabled BHs, though it is not
1000          * necessary */
1001         local_bh_disable();
1002         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1003                 sk->sk_backlog_rcv(sk, skb);
1004         local_bh_enable();
1005
1006         /* Clear memory counter. */
1007         tp->ucopy.memory = 0;
1008 }
1009
1010 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1011 {
1012         struct sk_buff *skb;
1013         u32 offset;
1014
1015         skb_queue_walk(&sk->sk_receive_queue, skb) {
1016                 offset = seq - TCP_SKB_CB(skb)->seq;
1017                 if (skb->h.th->syn)
1018                         offset--;
1019                 if (offset < skb->len || skb->h.th->fin) {
1020                         *off = offset;
1021                         return skb;
1022                 }
1023         }
1024         return NULL;
1025 }
1026
1027 /*
1028  * This routine provides an alternative to tcp_recvmsg() for routines
1029  * that would like to handle copying from skbuffs directly in 'sendfile'
1030  * fashion.
1031  * Note:
1032  *      - It is assumed that the socket was locked by the caller.
1033  *      - The routine does not block.
1034  *      - At present, there is no support for reading OOB data
1035  *        or for 'peeking' the socket using this routine
1036  *        (although both would be easy to implement).
1037  */
1038 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1039                   sk_read_actor_t recv_actor)
1040 {
1041         struct sk_buff *skb;
1042         struct tcp_sock *tp = tcp_sk(sk);
1043         u32 seq = tp->copied_seq;
1044         u32 offset;
1045         int copied = 0;
1046
1047         if (sk->sk_state == TCP_LISTEN)
1048                 return -ENOTCONN;
1049         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1050                 if (offset < skb->len) {
1051                         size_t used, len;
1052
1053                         len = skb->len - offset;
1054                         /* Stop reading if we hit a patch of urgent data */
1055                         if (tp->urg_data) {
1056                                 u32 urg_offset = tp->urg_seq - seq;
1057                                 if (urg_offset < len)
1058                                         len = urg_offset;
1059                                 if (!len)
1060                                         break;
1061                         }
1062                         used = recv_actor(desc, skb, offset, len);
1063                         if (used <= len) {
1064                                 seq += used;
1065                                 copied += used;
1066                                 offset += used;
1067                         }
1068                         if (offset != skb->len)
1069                                 break;
1070                 }
1071                 if (skb->h.th->fin) {
1072                         sk_eat_skb(sk, skb, 0);
1073                         ++seq;
1074                         break;
1075                 }
1076                 sk_eat_skb(sk, skb, 0);
1077                 if (!desc->count)
1078                         break;
1079         }
1080         tp->copied_seq = seq;
1081
1082         tcp_rcv_space_adjust(sk);
1083
1084         /* Clean up data we have read: This will do ACK frames. */
1085         if (copied)
1086                 tcp_cleanup_rbuf(sk, copied);
1087         return copied;
1088 }
1089
1090 /*
1091  *      This routine copies from a sock struct into the user buffer.
1092  *
1093  *      Technical note: in 2.3 we work on _locked_ socket, so that
1094  *      tricks with *seq access order and skb->users are not required.
1095  *      Probably, code can be easily improved even more.
1096  */
1097
1098 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1099                 size_t len, int nonblock, int flags, int *addr_len)
1100 {
1101         struct tcp_sock *tp = tcp_sk(sk);
1102         int copied = 0;
1103         u32 peek_seq;
1104         u32 *seq;
1105         unsigned long used;
1106         int err;
1107         int target;             /* Read at least this many bytes */
1108         long timeo;
1109         struct task_struct *user_recv = NULL;
1110         int copied_early = 0;
1111
1112         lock_sock(sk);
1113
1114         TCP_CHECK_TIMER(sk);
1115
1116         err = -ENOTCONN;
1117         if (sk->sk_state == TCP_LISTEN)
1118                 goto out;
1119
1120         timeo = sock_rcvtimeo(sk, nonblock);
1121
1122         /* Urgent data needs to be handled specially. */
1123         if (flags & MSG_OOB)
1124                 goto recv_urg;
1125
1126         seq = &tp->copied_seq;
1127         if (flags & MSG_PEEK) {
1128                 peek_seq = tp->copied_seq;
1129                 seq = &peek_seq;
1130         }
1131
1132         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1133
1134 #ifdef CONFIG_NET_DMA
1135         tp->ucopy.dma_chan = NULL;
1136         preempt_disable();
1137         if ((len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1138             !sysctl_tcp_low_latency && __get_cpu_var(softnet_data).net_dma) {
1139                 preempt_enable_no_resched();
1140                 tp->ucopy.pinned_list = dma_pin_iovec_pages(msg->msg_iov, len);
1141         } else
1142                 preempt_enable_no_resched();
1143 #endif
1144
1145         do {
1146                 struct sk_buff *skb;
1147                 u32 offset;
1148
1149                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1150                 if (tp->urg_data && tp->urg_seq == *seq) {
1151                         if (copied)
1152                                 break;
1153                         if (signal_pending(current)) {
1154                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1155                                 break;
1156                         }
1157                 }
1158
1159                 /* Next get a buffer. */
1160
1161                 skb = skb_peek(&sk->sk_receive_queue);
1162                 do {
1163                         if (!skb)
1164                                 break;
1165
1166                         /* Now that we have two receive queues this
1167                          * shouldn't happen.
1168                          */
1169                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1170                                 printk(KERN_INFO "recvmsg bug: copied %X "
1171                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1172                                 break;
1173                         }
1174                         offset = *seq - TCP_SKB_CB(skb)->seq;
1175                         if (skb->h.th->syn)
1176                                 offset--;
1177                         if (offset < skb->len)
1178                                 goto found_ok_skb;
1179                         if (skb->h.th->fin)
1180                                 goto found_fin_ok;
1181                         BUG_TRAP(flags & MSG_PEEK);
1182                         skb = skb->next;
1183                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1184
1185                 /* Well, if we have backlog, try to process it now yet. */
1186
1187                 if (copied >= target && !sk->sk_backlog.tail)
1188                         break;
1189
1190                 if (copied) {
1191                         if (sk->sk_err ||
1192                             sk->sk_state == TCP_CLOSE ||
1193                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1194                             !timeo ||
1195                             signal_pending(current) ||
1196                             (flags & MSG_PEEK))
1197                                 break;
1198                 } else {
1199                         if (sock_flag(sk, SOCK_DONE))
1200                                 break;
1201
1202                         if (sk->sk_err) {
1203                                 copied = sock_error(sk);
1204                                 break;
1205                         }
1206
1207                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1208                                 break;
1209
1210                         if (sk->sk_state == TCP_CLOSE) {
1211                                 if (!sock_flag(sk, SOCK_DONE)) {
1212                                         /* This occurs when user tries to read
1213                                          * from never connected socket.
1214                                          */
1215                                         copied = -ENOTCONN;
1216                                         break;
1217                                 }
1218                                 break;
1219                         }
1220
1221                         if (!timeo) {
1222                                 copied = -EAGAIN;
1223                                 break;
1224                         }
1225
1226                         if (signal_pending(current)) {
1227                                 copied = sock_intr_errno(timeo);
1228                                 break;
1229                         }
1230                 }
1231
1232                 tcp_cleanup_rbuf(sk, copied);
1233
1234                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1235                         /* Install new reader */
1236                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1237                                 user_recv = current;
1238                                 tp->ucopy.task = user_recv;
1239                                 tp->ucopy.iov = msg->msg_iov;
1240                         }
1241
1242                         tp->ucopy.len = len;
1243
1244                         BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1245                                  (flags & (MSG_PEEK | MSG_TRUNC)));
1246
1247                         /* Ugly... If prequeue is not empty, we have to
1248                          * process it before releasing socket, otherwise
1249                          * order will be broken at second iteration.
1250                          * More elegant solution is required!!!
1251                          *
1252                          * Look: we have the following (pseudo)queues:
1253                          *
1254                          * 1. packets in flight
1255                          * 2. backlog
1256                          * 3. prequeue
1257                          * 4. receive_queue
1258                          *
1259                          * Each queue can be processed only if the next ones
1260                          * are empty. At this point we have empty receive_queue.
1261                          * But prequeue _can_ be not empty after 2nd iteration,
1262                          * when we jumped to start of loop because backlog
1263                          * processing added something to receive_queue.
1264                          * We cannot release_sock(), because backlog contains
1265                          * packets arrived _after_ prequeued ones.
1266                          *
1267                          * Shortly, algorithm is clear --- to process all
1268                          * the queues in order. We could make it more directly,
1269                          * requeueing packets from backlog to prequeue, if
1270                          * is not empty. It is more elegant, but eats cycles,
1271                          * unfortunately.
1272                          */
1273                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1274                                 goto do_prequeue;
1275
1276                         /* __ Set realtime policy in scheduler __ */
1277                 }
1278
1279                 if (copied >= target) {
1280                         /* Do not sleep, just process backlog. */
1281                         release_sock(sk);
1282                         lock_sock(sk);
1283                 } else
1284                         sk_wait_data(sk, &timeo);
1285
1286 #ifdef CONFIG_NET_DMA
1287                 tp->ucopy.wakeup = 0;
1288 #endif
1289
1290                 if (user_recv) {
1291                         int chunk;
1292
1293                         /* __ Restore normal policy in scheduler __ */
1294
1295                         if ((chunk = len - tp->ucopy.len) != 0) {
1296                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1297                                 len -= chunk;
1298                                 copied += chunk;
1299                         }
1300
1301                         if (tp->rcv_nxt == tp->copied_seq &&
1302                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1303 do_prequeue:
1304                                 tcp_prequeue_process(sk);
1305
1306                                 if ((chunk = len - tp->ucopy.len) != 0) {
1307                                         NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1308                                         len -= chunk;
1309                                         copied += chunk;
1310                                 }
1311                         }
1312                 }
1313                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1314                         if (net_ratelimit())
1315                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1316                                        current->comm, current->pid);
1317                         peek_seq = tp->copied_seq;
1318                 }
1319                 continue;
1320
1321         found_ok_skb:
1322                 /* Ok so how much can we use? */
1323                 used = skb->len - offset;
1324                 if (len < used)
1325                         used = len;
1326
1327                 /* Do we have urgent data here? */
1328                 if (tp->urg_data) {
1329                         u32 urg_offset = tp->urg_seq - *seq;
1330                         if (urg_offset < used) {
1331                                 if (!urg_offset) {
1332                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1333                                                 ++*seq;
1334                                                 offset++;
1335                                                 used--;
1336                                                 if (!used)
1337                                                         goto skip_copy;
1338                                         }
1339                                 } else
1340                                         used = urg_offset;
1341                         }
1342                 }
1343
1344                 if (!(flags & MSG_TRUNC)) {
1345 #ifdef CONFIG_NET_DMA
1346                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1347                                 tp->ucopy.dma_chan = get_softnet_dma();
1348
1349                         if (tp->ucopy.dma_chan) {
1350                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1351                                         tp->ucopy.dma_chan, skb, offset,
1352                                         msg->msg_iov, used,
1353                                         tp->ucopy.pinned_list);
1354
1355                                 if (tp->ucopy.dma_cookie < 0) {
1356
1357                                         printk(KERN_ALERT "dma_cookie < 0\n");
1358
1359                                         /* Exception. Bailout! */
1360                                         if (!copied)
1361                                                 copied = -EFAULT;
1362                                         break;
1363                                 }
1364                                 if ((offset + used) == skb->len)
1365                                         copied_early = 1;
1366
1367                         } else
1368 #endif
1369                         {
1370                                 err = skb_copy_datagram_iovec(skb, offset,
1371                                                 msg->msg_iov, used);
1372                                 if (err) {
1373                                         /* Exception. Bailout! */
1374                                         if (!copied)
1375                                                 copied = -EFAULT;
1376                                         break;
1377                                 }
1378                         }
1379                 }
1380
1381                 *seq += used;
1382                 copied += used;
1383                 len -= used;
1384
1385                 tcp_rcv_space_adjust(sk);
1386
1387 skip_copy:
1388                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1389                         tp->urg_data = 0;
1390                         tcp_fast_path_check(sk, tp);
1391                 }
1392                 if (used + offset < skb->len)
1393                         continue;
1394
1395                 if (skb->h.th->fin)
1396                         goto found_fin_ok;
1397                 if (!(flags & MSG_PEEK)) {
1398                         sk_eat_skb(sk, skb, copied_early);
1399                         copied_early = 0;
1400                 }
1401                 continue;
1402
1403         found_fin_ok:
1404                 /* Process the FIN. */
1405                 ++*seq;
1406                 if (!(flags & MSG_PEEK)) {
1407                         sk_eat_skb(sk, skb, copied_early);
1408                         copied_early = 0;
1409                 }
1410                 break;
1411         } while (len > 0);
1412
1413         if (user_recv) {
1414                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1415                         int chunk;
1416
1417                         tp->ucopy.len = copied > 0 ? len : 0;
1418
1419                         tcp_prequeue_process(sk);
1420
1421                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1422                                 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1423                                 len -= chunk;
1424                                 copied += chunk;
1425                         }
1426                 }
1427
1428                 tp->ucopy.task = NULL;
1429                 tp->ucopy.len = 0;
1430         }
1431
1432 #ifdef CONFIG_NET_DMA
1433         if (tp->ucopy.dma_chan) {
1434                 struct sk_buff *skb;
1435                 dma_cookie_t done, used;
1436
1437                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1438
1439                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1440                                                  tp->ucopy.dma_cookie, &done,
1441                                                  &used) == DMA_IN_PROGRESS) {
1442                         /* do partial cleanup of sk_async_wait_queue */
1443                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1444                                (dma_async_is_complete(skb->dma_cookie, done,
1445                                                       used) == DMA_SUCCESS)) {
1446                                 __skb_dequeue(&sk->sk_async_wait_queue);
1447                                 kfree_skb(skb);
1448                         }
1449                 }
1450
1451                 /* Safe to free early-copied skbs now */
1452                 __skb_queue_purge(&sk->sk_async_wait_queue);
1453                 dma_chan_put(tp->ucopy.dma_chan);
1454                 tp->ucopy.dma_chan = NULL;
1455         }
1456         if (tp->ucopy.pinned_list) {
1457                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1458                 tp->ucopy.pinned_list = NULL;
1459         }
1460 #endif
1461
1462         /* According to UNIX98, msg_name/msg_namelen are ignored
1463          * on connected socket. I was just happy when found this 8) --ANK
1464          */
1465
1466         /* Clean up data we have read: This will do ACK frames. */
1467         tcp_cleanup_rbuf(sk, copied);
1468
1469         TCP_CHECK_TIMER(sk);
1470         release_sock(sk);
1471         return copied;
1472
1473 out:
1474         TCP_CHECK_TIMER(sk);
1475         release_sock(sk);
1476         return err;
1477
1478 recv_urg:
1479         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1480         goto out;
1481 }
1482
1483 /*
1484  *      State processing on a close. This implements the state shift for
1485  *      sending our FIN frame. Note that we only send a FIN for some
1486  *      states. A shutdown() may have already sent the FIN, or we may be
1487  *      closed.
1488  */
1489
1490 static const unsigned char new_state[16] = {
1491   /* current state:        new state:      action:      */
1492   /* (Invalid)          */ TCP_CLOSE,
1493   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1494   /* TCP_SYN_SENT       */ TCP_CLOSE,
1495   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1496   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1497   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1498   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1499   /* TCP_CLOSE          */ TCP_CLOSE,
1500   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1501   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1502   /* TCP_LISTEN         */ TCP_CLOSE,
1503   /* TCP_CLOSING        */ TCP_CLOSING,
1504 };
1505
1506 static int tcp_close_state(struct sock *sk)
1507 {
1508         int next = (int)new_state[sk->sk_state];
1509         int ns = next & TCP_STATE_MASK;
1510
1511         tcp_set_state(sk, ns);
1512
1513         return next & TCP_ACTION_FIN;
1514 }
1515
1516 /*
1517  *      Shutdown the sending side of a connection. Much like close except
1518  *      that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1519  */
1520
1521 void tcp_shutdown(struct sock *sk, int how)
1522 {
1523         /*      We need to grab some memory, and put together a FIN,
1524          *      and then put it into the queue to be sent.
1525          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1526          */
1527         if (!(how & SEND_SHUTDOWN))
1528                 return;
1529
1530         /* If we've already sent a FIN, or it's a closed state, skip this. */
1531         if ((1 << sk->sk_state) &
1532             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1533              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1534                 /* Clear out any half completed packets.  FIN if needed. */
1535                 if (tcp_close_state(sk))
1536                         tcp_send_fin(sk);
1537         }
1538 }
1539
1540 void tcp_close(struct sock *sk, long timeout)
1541 {
1542         struct sk_buff *skb;
1543         int data_was_unread = 0;
1544         int state;
1545
1546         lock_sock(sk);
1547         sk->sk_shutdown = SHUTDOWN_MASK;
1548
1549         if (sk->sk_state == TCP_LISTEN) {
1550                 tcp_set_state(sk, TCP_CLOSE);
1551
1552                 /* Special case. */
1553                 inet_csk_listen_stop(sk);
1554
1555                 goto adjudge_to_death;
1556         }
1557
1558         /*  We need to flush the recv. buffs.  We do this only on the
1559          *  descriptor close, not protocol-sourced closes, because the
1560          *  reader process may not have drained the data yet!
1561          */
1562         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1563                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1564                           skb->h.th->fin;
1565                 data_was_unread += len;
1566                 __kfree_skb(skb);
1567         }
1568
1569         sk_stream_mem_reclaim(sk);
1570
1571         /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1572          * 3.10, we send a RST here because data was lost.  To
1573          * witness the awful effects of the old behavior of always
1574          * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1575          * a bulk GET in an FTP client, suspend the process, wait
1576          * for the client to advertise a zero window, then kill -9
1577          * the FTP client, wheee...  Note: timeout is always zero
1578          * in such a case.
1579          */
1580         if (data_was_unread) {
1581                 /* Unread data was tossed, zap the connection. */
1582                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1583                 tcp_set_state(sk, TCP_CLOSE);
1584                 tcp_send_active_reset(sk, GFP_KERNEL);
1585         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1586                 /* Check zero linger _after_ checking for unread data. */
1587                 sk->sk_prot->disconnect(sk, 0);
1588                 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1589         } else if (tcp_close_state(sk)) {
1590                 /* We FIN if the application ate all the data before
1591                  * zapping the connection.
1592                  */
1593
1594                 /* RED-PEN. Formally speaking, we have broken TCP state
1595                  * machine. State transitions:
1596                  *
1597                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1598                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1599                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1600                  *
1601                  * are legal only when FIN has been sent (i.e. in window),
1602                  * rather than queued out of window. Purists blame.
1603                  *
1604                  * F.e. "RFC state" is ESTABLISHED,
1605                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1606                  *
1607                  * The visible declinations are that sometimes
1608                  * we enter time-wait state, when it is not required really
1609                  * (harmless), do not send active resets, when they are
1610                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1611                  * they look as CLOSING or LAST_ACK for Linux)
1612                  * Probably, I missed some more holelets.
1613                  *                                              --ANK
1614                  */
1615                 tcp_send_fin(sk);
1616         }
1617
1618         sk_stream_wait_close(sk, timeout);
1619
1620 adjudge_to_death:
1621         state = sk->sk_state;
1622         sock_hold(sk);
1623         sock_orphan(sk);
1624         atomic_inc(sk->sk_prot->orphan_count);
1625
1626         /* It is the last release_sock in its life. It will remove backlog. */
1627         release_sock(sk);
1628
1629
1630         /* Now socket is owned by kernel and we acquire BH lock
1631            to finish close. No need to check for user refs.
1632          */
1633         local_bh_disable();
1634         bh_lock_sock(sk);
1635         BUG_TRAP(!sock_owned_by_user(sk));
1636
1637         /* Have we already been destroyed by a softirq or backlog? */
1638         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1639                 goto out;
1640
1641         /*      This is a (useful) BSD violating of the RFC. There is a
1642          *      problem with TCP as specified in that the other end could
1643          *      keep a socket open forever with no application left this end.
1644          *      We use a 3 minute timeout (about the same as BSD) then kill
1645          *      our end. If they send after that then tough - BUT: long enough
1646          *      that we won't make the old 4*rto = almost no time - whoops
1647          *      reset mistake.
1648          *
1649          *      Nope, it was not mistake. It is really desired behaviour
1650          *      f.e. on http servers, when such sockets are useless, but
1651          *      consume significant resources. Let's do it with special
1652          *      linger2 option.                                 --ANK
1653          */
1654
1655         if (sk->sk_state == TCP_FIN_WAIT2) {
1656                 struct tcp_sock *tp = tcp_sk(sk);
1657                 if (tp->linger2 < 0) {
1658                         tcp_set_state(sk, TCP_CLOSE);
1659                         tcp_send_active_reset(sk, GFP_ATOMIC);
1660                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1661                 } else {
1662                         const int tmo = tcp_fin_time(sk);
1663
1664                         if (tmo > TCP_TIMEWAIT_LEN) {
1665                                 inet_csk_reset_keepalive_timer(sk,
1666                                                 tmo - TCP_TIMEWAIT_LEN);
1667                         } else {
1668                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1669                                 goto out;
1670                         }
1671                 }
1672         }
1673         if (sk->sk_state != TCP_CLOSE) {
1674                 sk_stream_mem_reclaim(sk);
1675                 if (atomic_read(sk->sk_prot->orphan_count) > sysctl_tcp_max_orphans ||
1676                     (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1677                      atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1678                         if (net_ratelimit())
1679                                 printk(KERN_INFO "TCP: too many of orphaned "
1680                                        "sockets\n");
1681                         tcp_set_state(sk, TCP_CLOSE);
1682                         tcp_send_active_reset(sk, GFP_ATOMIC);
1683                         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1684                 }
1685         }
1686
1687         if (sk->sk_state == TCP_CLOSE)
1688                 inet_csk_destroy_sock(sk);
1689         /* Otherwise, socket is reprieved until protocol close. */
1690
1691 out:
1692         bh_unlock_sock(sk);
1693         local_bh_enable();
1694         sock_put(sk);
1695 }
1696
1697 /* These states need RST on ABORT according to RFC793 */
1698
1699 static inline int tcp_need_reset(int state)
1700 {
1701         return (1 << state) &
1702                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1703                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1704 }
1705
1706 int tcp_disconnect(struct sock *sk, int flags)
1707 {
1708         struct inet_sock *inet = inet_sk(sk);
1709         struct inet_connection_sock *icsk = inet_csk(sk);
1710         struct tcp_sock *tp = tcp_sk(sk);
1711         int err = 0;
1712         int old_state = sk->sk_state;
1713
1714         if (old_state != TCP_CLOSE)
1715                 tcp_set_state(sk, TCP_CLOSE);
1716
1717         /* ABORT function of RFC793 */
1718         if (old_state == TCP_LISTEN) {
1719                 inet_csk_listen_stop(sk);
1720         } else if (tcp_need_reset(old_state) ||
1721                    (tp->snd_nxt != tp->write_seq &&
1722                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1723                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1724                  * states
1725                  */
1726                 tcp_send_active_reset(sk, gfp_any());
1727                 sk->sk_err = ECONNRESET;
1728         } else if (old_state == TCP_SYN_SENT)
1729                 sk->sk_err = ECONNRESET;
1730
1731         tcp_clear_xmit_timers(sk);
1732         __skb_queue_purge(&sk->sk_receive_queue);
1733         sk_stream_writequeue_purge(sk);
1734         __skb_queue_purge(&tp->out_of_order_queue);
1735 #ifdef CONFIG_NET_DMA
1736         __skb_queue_purge(&sk->sk_async_wait_queue);
1737 #endif
1738
1739         inet->dport = 0;
1740
1741         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1742                 inet_reset_saddr(sk);
1743
1744         sk->sk_shutdown = 0;
1745         sock_reset_flag(sk, SOCK_DONE);
1746         tp->srtt = 0;
1747         if ((tp->write_seq += tp->max_window + 2) == 0)
1748                 tp->write_seq = 1;
1749         icsk->icsk_backoff = 0;
1750         tp->snd_cwnd = 2;
1751         icsk->icsk_probes_out = 0;
1752         tp->packets_out = 0;
1753         tp->snd_ssthresh = 0x7fffffff;
1754         tp->snd_cwnd_cnt = 0;
1755         tp->bytes_acked = 0;
1756         tcp_set_ca_state(sk, TCP_CA_Open);
1757         tcp_clear_retrans(tp);
1758         inet_csk_delack_init(sk);
1759         sk->sk_send_head = NULL;
1760         tp->rx_opt.saw_tstamp = 0;
1761         tcp_sack_reset(&tp->rx_opt);
1762         __sk_dst_reset(sk);
1763
1764         BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1765
1766         sk->sk_error_report(sk);
1767         return err;
1768 }
1769
1770 /*
1771  *      Socket option code for TCP.
1772  */
1773 static int do_tcp_setsockopt(struct sock *sk, int level,
1774                 int optname, char __user *optval, int optlen)
1775 {
1776         struct tcp_sock *tp = tcp_sk(sk);
1777         struct inet_connection_sock *icsk = inet_csk(sk);
1778         int val;
1779         int err = 0;
1780
1781         /* This is a string value all the others are int's */
1782         if (optname == TCP_CONGESTION) {
1783                 char name[TCP_CA_NAME_MAX];
1784
1785                 if (optlen < 1)
1786                         return -EINVAL;
1787
1788                 val = strncpy_from_user(name, optval,
1789                                         min(TCP_CA_NAME_MAX-1, optlen));
1790                 if (val < 0)
1791                         return -EFAULT;
1792                 name[val] = 0;
1793
1794                 lock_sock(sk);
1795                 err = tcp_set_congestion_control(sk, name);
1796                 release_sock(sk);
1797                 return err;
1798         }
1799
1800         if (optlen < sizeof(int))
1801                 return -EINVAL;
1802
1803         if (get_user(val, (int __user *)optval))
1804                 return -EFAULT;
1805
1806         lock_sock(sk);
1807
1808         switch (optname) {
1809         case TCP_MAXSEG:
1810                 /* Values greater than interface MTU won't take effect. However
1811                  * at the point when this call is done we typically don't yet
1812                  * know which interface is going to be used */
1813                 if (val < 8 || val > MAX_TCP_WINDOW) {
1814                         err = -EINVAL;
1815                         break;
1816                 }
1817                 tp->rx_opt.user_mss = val;
1818                 break;
1819
1820         case TCP_NODELAY:
1821                 if (val) {
1822                         /* TCP_NODELAY is weaker than TCP_CORK, so that
1823                          * this option on corked socket is remembered, but
1824                          * it is not activated until cork is cleared.
1825                          *
1826                          * However, when TCP_NODELAY is set we make
1827                          * an explicit push, which overrides even TCP_CORK
1828                          * for currently queued segments.
1829                          */
1830                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1831                         tcp_push_pending_frames(sk, tp);
1832                 } else {
1833                         tp->nonagle &= ~TCP_NAGLE_OFF;
1834                 }
1835                 break;
1836
1837         case TCP_CORK:
1838                 /* When set indicates to always queue non-full frames.
1839                  * Later the user clears this option and we transmit
1840                  * any pending partial frames in the queue.  This is
1841                  * meant to be used alongside sendfile() to get properly
1842                  * filled frames when the user (for example) must write
1843                  * out headers with a write() call first and then use
1844                  * sendfile to send out the data parts.
1845                  *
1846                  * TCP_CORK can be set together with TCP_NODELAY and it is
1847                  * stronger than TCP_NODELAY.
1848                  */
1849                 if (val) {
1850                         tp->nonagle |= TCP_NAGLE_CORK;
1851                 } else {
1852                         tp->nonagle &= ~TCP_NAGLE_CORK;
1853                         if (tp->nonagle&TCP_NAGLE_OFF)
1854                                 tp->nonagle |= TCP_NAGLE_PUSH;
1855                         tcp_push_pending_frames(sk, tp);
1856                 }
1857                 break;
1858
1859         case TCP_KEEPIDLE:
1860                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1861                         err = -EINVAL;
1862                 else {
1863                         tp->keepalive_time = val * HZ;
1864                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
1865                             !((1 << sk->sk_state) &
1866                               (TCPF_CLOSE | TCPF_LISTEN))) {
1867                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1868                                 if (tp->keepalive_time > elapsed)
1869                                         elapsed = tp->keepalive_time - elapsed;
1870                                 else
1871                                         elapsed = 0;
1872                                 inet_csk_reset_keepalive_timer(sk, elapsed);
1873                         }
1874                 }
1875                 break;
1876         case TCP_KEEPINTVL:
1877                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1878                         err = -EINVAL;
1879                 else
1880                         tp->keepalive_intvl = val * HZ;
1881                 break;
1882         case TCP_KEEPCNT:
1883                 if (val < 1 || val > MAX_TCP_KEEPCNT)
1884                         err = -EINVAL;
1885                 else
1886                         tp->keepalive_probes = val;
1887                 break;
1888         case TCP_SYNCNT:
1889                 if (val < 1 || val > MAX_TCP_SYNCNT)
1890                         err = -EINVAL;
1891                 else
1892                         icsk->icsk_syn_retries = val;
1893                 break;
1894
1895         case TCP_LINGER2:
1896                 if (val < 0)
1897                         tp->linger2 = -1;
1898                 else if (val > sysctl_tcp_fin_timeout / HZ)
1899                         tp->linger2 = 0;
1900                 else
1901                         tp->linger2 = val * HZ;
1902                 break;
1903
1904         case TCP_DEFER_ACCEPT:
1905                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1906                 if (val > 0) {
1907                         /* Translate value in seconds to number of
1908                          * retransmits */
1909                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1910                                val > ((TCP_TIMEOUT_INIT / HZ) <<
1911                                        icsk->icsk_accept_queue.rskq_defer_accept))
1912                                 icsk->icsk_accept_queue.rskq_defer_accept++;
1913                         icsk->icsk_accept_queue.rskq_defer_accept++;
1914                 }
1915                 break;
1916
1917         case TCP_WINDOW_CLAMP:
1918                 if (!val) {
1919                         if (sk->sk_state != TCP_CLOSE) {
1920                                 err = -EINVAL;
1921                                 break;
1922                         }
1923                         tp->window_clamp = 0;
1924                 } else
1925                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1926                                                 SOCK_MIN_RCVBUF / 2 : val;
1927                 break;
1928
1929         case TCP_QUICKACK:
1930                 if (!val) {
1931                         icsk->icsk_ack.pingpong = 1;
1932                 } else {
1933                         icsk->icsk_ack.pingpong = 0;
1934                         if ((1 << sk->sk_state) &
1935                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1936                             inet_csk_ack_scheduled(sk)) {
1937                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1938                                 tcp_cleanup_rbuf(sk, 1);
1939                                 if (!(val & 1))
1940                                         icsk->icsk_ack.pingpong = 1;
1941                         }
1942                 }
1943                 break;
1944
1945         default:
1946                 err = -ENOPROTOOPT;
1947                 break;
1948         };
1949         release_sock(sk);
1950         return err;
1951 }
1952
1953 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1954                    int optlen)
1955 {
1956         struct inet_connection_sock *icsk = inet_csk(sk);
1957
1958         if (level != SOL_TCP)
1959                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1960                                                      optval, optlen);
1961         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1962 }
1963
1964 #ifdef CONFIG_COMPAT
1965 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1966                           char __user *optval, int optlen)
1967 {
1968         if (level != SOL_TCP)
1969                 return inet_csk_compat_setsockopt(sk, level, optname,
1970                                                   optval, optlen);
1971         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1972 }
1973
1974 EXPORT_SYMBOL(compat_tcp_setsockopt);
1975 #endif
1976
1977 /* Return information about state of tcp endpoint in API format. */
1978 void tcp_get_info(struct sock *sk, struct tcp_info *info)
1979 {
1980         struct tcp_sock *tp = tcp_sk(sk);
1981         const struct inet_connection_sock *icsk = inet_csk(sk);
1982         u32 now = tcp_time_stamp;
1983
1984         memset(info, 0, sizeof(*info));
1985
1986         info->tcpi_state = sk->sk_state;
1987         info->tcpi_ca_state = icsk->icsk_ca_state;
1988         info->tcpi_retransmits = icsk->icsk_retransmits;
1989         info->tcpi_probes = icsk->icsk_probes_out;
1990         info->tcpi_backoff = icsk->icsk_backoff;
1991
1992         if (tp->rx_opt.tstamp_ok)
1993                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
1994         if (tp->rx_opt.sack_ok)
1995                 info->tcpi_options |= TCPI_OPT_SACK;
1996         if (tp->rx_opt.wscale_ok) {
1997                 info->tcpi_options |= TCPI_OPT_WSCALE;
1998                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
1999                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2000         } 
2001
2002         if (tp->ecn_flags&TCP_ECN_OK)
2003                 info->tcpi_options |= TCPI_OPT_ECN;
2004
2005         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2006         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2007         info->tcpi_snd_mss = tp->mss_cache;
2008         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2009
2010         info->tcpi_unacked = tp->packets_out;
2011         info->tcpi_sacked = tp->sacked_out;
2012         info->tcpi_lost = tp->lost_out;
2013         info->tcpi_retrans = tp->retrans_out;
2014         info->tcpi_fackets = tp->fackets_out;
2015
2016         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2017         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2018         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2019
2020         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2021         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2022         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2023         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2024         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2025         info->tcpi_snd_cwnd = tp->snd_cwnd;
2026         info->tcpi_advmss = tp->advmss;
2027         info->tcpi_reordering = tp->reordering;
2028
2029         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2030         info->tcpi_rcv_space = tp->rcvq_space.space;
2031
2032         info->tcpi_total_retrans = tp->total_retrans;
2033 }
2034
2035 EXPORT_SYMBOL_GPL(tcp_get_info);
2036
2037 static int do_tcp_getsockopt(struct sock *sk, int level,
2038                 int optname, char __user *optval, int __user *optlen)
2039 {
2040         struct inet_connection_sock *icsk = inet_csk(sk);
2041         struct tcp_sock *tp = tcp_sk(sk);
2042         int val, len;
2043
2044         if (get_user(len, optlen))
2045                 return -EFAULT;
2046
2047         len = min_t(unsigned int, len, sizeof(int));
2048
2049         if (len < 0)
2050                 return -EINVAL;
2051
2052         switch (optname) {
2053         case TCP_MAXSEG:
2054                 val = tp->mss_cache;
2055                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2056                         val = tp->rx_opt.user_mss;
2057                 break;
2058         case TCP_NODELAY:
2059                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2060                 break;
2061         case TCP_CORK:
2062                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2063                 break;
2064         case TCP_KEEPIDLE:
2065                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2066                 break;
2067         case TCP_KEEPINTVL:
2068                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2069                 break;
2070         case TCP_KEEPCNT:
2071                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2072                 break;
2073         case TCP_SYNCNT:
2074                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2075                 break;
2076         case TCP_LINGER2:
2077                 val = tp->linger2;
2078                 if (val >= 0)
2079                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2080                 break;
2081         case TCP_DEFER_ACCEPT:
2082                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2083                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2084                 break;
2085         case TCP_WINDOW_CLAMP:
2086                 val = tp->window_clamp;
2087                 break;
2088         case TCP_INFO: {
2089                 struct tcp_info info;
2090
2091                 if (get_user(len, optlen))
2092                         return -EFAULT;
2093
2094                 tcp_get_info(sk, &info);
2095
2096                 len = min_t(unsigned int, len, sizeof(info));
2097                 if (put_user(len, optlen))
2098                         return -EFAULT;
2099                 if (copy_to_user(optval, &info, len))
2100                         return -EFAULT;
2101                 return 0;
2102         }
2103         case TCP_QUICKACK:
2104                 val = !icsk->icsk_ack.pingpong;
2105                 break;
2106
2107         case TCP_CONGESTION:
2108                 if (get_user(len, optlen))
2109                         return -EFAULT;
2110                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2111                 if (put_user(len, optlen))
2112                         return -EFAULT;
2113                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2114                         return -EFAULT;
2115                 return 0;
2116         default:
2117                 return -ENOPROTOOPT;
2118         };
2119
2120         if (put_user(len, optlen))
2121                 return -EFAULT;
2122         if (copy_to_user(optval, &val, len))
2123                 return -EFAULT;
2124         return 0;
2125 }
2126
2127 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2128                    int __user *optlen)
2129 {
2130         struct inet_connection_sock *icsk = inet_csk(sk);
2131
2132         if (level != SOL_TCP)
2133                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2134                                                      optval, optlen);
2135         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2136 }
2137
2138 #ifdef CONFIG_COMPAT
2139 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2140                           char __user *optval, int __user *optlen)
2141 {
2142         if (level != SOL_TCP)
2143                 return inet_csk_compat_getsockopt(sk, level, optname,
2144                                                   optval, optlen);
2145         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2146 }
2147
2148 EXPORT_SYMBOL(compat_tcp_getsockopt);
2149 #endif
2150
2151 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2152 {
2153         struct sk_buff *segs = ERR_PTR(-EINVAL);
2154         struct tcphdr *th;
2155         unsigned thlen;
2156         unsigned int seq;
2157         unsigned int delta;
2158         unsigned int oldlen;
2159         unsigned int len;
2160
2161         if (!pskb_may_pull(skb, sizeof(*th)))
2162                 goto out;
2163
2164         th = skb->h.th;
2165         thlen = th->doff * 4;
2166         if (thlen < sizeof(*th))
2167                 goto out;
2168
2169         if (!pskb_may_pull(skb, thlen))
2170                 goto out;
2171
2172         oldlen = (u16)~skb->len;
2173         __skb_pull(skb, thlen);
2174
2175         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2176                 /* Packet is from an untrusted source, reset gso_segs. */
2177                 int type = skb_shinfo(skb)->gso_type;
2178                 int mss;
2179
2180                 if (unlikely(type &
2181                              ~(SKB_GSO_TCPV4 |
2182                                SKB_GSO_DODGY |
2183                                SKB_GSO_TCP_ECN |
2184                                SKB_GSO_TCPV6 |
2185                                0) ||
2186                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2187                         goto out;
2188
2189                 mss = skb_shinfo(skb)->gso_size;
2190                 skb_shinfo(skb)->gso_segs = (skb->len + mss - 1) / mss;
2191
2192                 segs = NULL;
2193                 goto out;
2194         }
2195
2196         segs = skb_segment(skb, features);
2197         if (IS_ERR(segs))
2198                 goto out;
2199
2200         len = skb_shinfo(skb)->gso_size;
2201         delta = htonl(oldlen + (thlen + len));
2202
2203         skb = segs;
2204         th = skb->h.th;
2205         seq = ntohl(th->seq);
2206
2207         do {
2208                 th->fin = th->psh = 0;
2209
2210                 th->check = ~csum_fold(th->check + delta);
2211                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2212                         th->check = csum_fold(csum_partial(skb->h.raw, thlen,
2213                                                            skb->csum));
2214
2215                 seq += len;
2216                 skb = skb->next;
2217                 th = skb->h.th;
2218
2219                 th->seq = htonl(seq);
2220                 th->cwr = 0;
2221         } while (skb->next);
2222
2223         delta = htonl(oldlen + (skb->tail - skb->h.raw) + skb->data_len);
2224         th->check = ~csum_fold(th->check + delta);
2225         if (skb->ip_summed != CHECKSUM_PARTIAL)
2226                 th->check = csum_fold(csum_partial(skb->h.raw, thlen,
2227                                                    skb->csum));
2228
2229 out:
2230         return segs;
2231 }
2232 EXPORT_SYMBOL(tcp_tso_segment);
2233
2234 extern void __skb_cb_too_small_for_tcp(int, int);
2235 extern struct tcp_congestion_ops tcp_reno;
2236
2237 static __initdata unsigned long thash_entries;
2238 static int __init set_thash_entries(char *str)
2239 {
2240         if (!str)
2241                 return 0;
2242         thash_entries = simple_strtoul(str, &str, 0);
2243         return 1;
2244 }
2245 __setup("thash_entries=", set_thash_entries);
2246
2247 void __init tcp_init(void)
2248 {
2249         struct sk_buff *skb = NULL;
2250         unsigned long limit;
2251         int order, i, max_share;
2252
2253         if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2254                 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2255                                            sizeof(skb->cb));
2256
2257         tcp_hashinfo.bind_bucket_cachep =
2258                 kmem_cache_create("tcp_bind_bucket",
2259                                   sizeof(struct inet_bind_bucket), 0,
2260                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
2261
2262         /* Size and allocate the main established and bind bucket
2263          * hash tables.
2264          *
2265          * The methodology is similar to that of the buffer cache.
2266          */
2267         tcp_hashinfo.ehash =
2268                 alloc_large_system_hash("TCP established",
2269                                         sizeof(struct inet_ehash_bucket),
2270                                         thash_entries,
2271                                         (num_physpages >= 128 * 1024) ?
2272                                         13 : 15,
2273                                         HASH_HIGHMEM,
2274                                         &tcp_hashinfo.ehash_size,
2275                                         NULL,
2276                                         0);
2277         tcp_hashinfo.ehash_size = (1 << tcp_hashinfo.ehash_size) >> 1;
2278         for (i = 0; i < (tcp_hashinfo.ehash_size << 1); i++) {
2279                 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2280                 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2281         }
2282
2283         tcp_hashinfo.bhash =
2284                 alloc_large_system_hash("TCP bind",
2285                                         sizeof(struct inet_bind_hashbucket),
2286                                         tcp_hashinfo.ehash_size,
2287                                         (num_physpages >= 128 * 1024) ?
2288                                         13 : 15,
2289                                         HASH_HIGHMEM,
2290                                         &tcp_hashinfo.bhash_size,
2291                                         NULL,
2292                                         64 * 1024);
2293         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2294         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2295                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2296                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2297         }
2298
2299         /* Try to be a bit smarter and adjust defaults depending
2300          * on available memory.
2301          */
2302         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2303                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2304                         order++)
2305                 ;
2306         if (order >= 4) {
2307                 sysctl_local_port_range[0] = 32768;
2308                 sysctl_local_port_range[1] = 61000;
2309                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2310                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2311                 sysctl_max_syn_backlog = 1024;
2312         } else if (order < 3) {
2313                 sysctl_local_port_range[0] = 1024 * (3 - order);
2314                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2315                 sysctl_tcp_max_orphans >>= (3 - order);
2316                 sysctl_max_syn_backlog = 128;
2317         }
2318
2319         sysctl_tcp_mem[0] =  768 << order;
2320         sysctl_tcp_mem[1] = 1024 << order;
2321         sysctl_tcp_mem[2] = 1536 << order;
2322
2323         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2324         max_share = min(4UL*1024*1024, limit);
2325
2326         sysctl_tcp_wmem[0] = SK_STREAM_MEM_QUANTUM;
2327         sysctl_tcp_wmem[1] = 16*1024;
2328         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2329
2330         sysctl_tcp_rmem[0] = SK_STREAM_MEM_QUANTUM;
2331         sysctl_tcp_rmem[1] = 87380;
2332         sysctl_tcp_rmem[2] = max(87380, max_share);
2333
2334         printk(KERN_INFO "TCP: Hash tables configured "
2335                "(established %d bind %d)\n",
2336                tcp_hashinfo.ehash_size << 1, tcp_hashinfo.bhash_size);
2337
2338         tcp_register_congestion_control(&tcp_reno);
2339 }
2340
2341 EXPORT_SYMBOL(tcp_close);
2342 EXPORT_SYMBOL(tcp_disconnect);
2343 EXPORT_SYMBOL(tcp_getsockopt);
2344 EXPORT_SYMBOL(tcp_ioctl);
2345 EXPORT_SYMBOL(tcp_poll);
2346 EXPORT_SYMBOL(tcp_read_sock);
2347 EXPORT_SYMBOL(tcp_recvmsg);
2348 EXPORT_SYMBOL(tcp_sendmsg);
2349 EXPORT_SYMBOL(tcp_sendpage);
2350 EXPORT_SYMBOL(tcp_setsockopt);
2351 EXPORT_SYMBOL(tcp_shutdown);
2352 EXPORT_SYMBOL(tcp_statistics);