Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[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  * Authors:     Ross Biro
9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
12  *              Florian La Roche, <flla@stud.uni-sb.de>
13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *              Jorge Cwik, <jorge@laser.satlink.net>
19  *
20  * Fixes:
21  *              Alan Cox        :       Numerous verify_area() calls
22  *              Alan Cox        :       Set the ACK bit on a reset
23  *              Alan Cox        :       Stopped it crashing if it closed while
24  *                                      sk->inuse=1 and was trying to connect
25  *                                      (tcp_err()).
26  *              Alan Cox        :       All icmp error handling was broken
27  *                                      pointers passed where wrong and the
28  *                                      socket was looked up backwards. Nobody
29  *                                      tested any icmp error code obviously.
30  *              Alan Cox        :       tcp_err() now handled properly. It
31  *                                      wakes people on errors. poll
32  *                                      behaves and the icmp error race
33  *                                      has gone by moving it into sock.c
34  *              Alan Cox        :       tcp_send_reset() fixed to work for
35  *                                      everything not just packets for
36  *                                      unknown sockets.
37  *              Alan Cox        :       tcp option processing.
38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
39  *                                      syn rule wrong]
40  *              Herp Rosmanith  :       More reset fixes
41  *              Alan Cox        :       No longer acks invalid rst frames.
42  *                                      Acking any kind of RST is right out.
43  *              Alan Cox        :       Sets an ignore me flag on an rst
44  *                                      receive otherwise odd bits of prattle
45  *                                      escape still
46  *              Alan Cox        :       Fixed another acking RST frame bug.
47  *                                      Should stop LAN workplace lockups.
48  *              Alan Cox        :       Some tidyups using the new skb list
49  *                                      facilities
50  *              Alan Cox        :       sk->keepopen now seems to work
51  *              Alan Cox        :       Pulls options out correctly on accepts
52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
54  *                                      bit to skb ops.
55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
56  *                                      nasty.
57  *              Alan Cox        :       Added some better commenting, as the
58  *                                      tcp is hard to follow
59  *              Alan Cox        :       Removed incorrect check for 20 * psh
60  *      Michael O'Reilly        :       ack < copied bug fix.
61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
62  *              Alan Cox        :       FIN with no memory -> CRASH
63  *              Alan Cox        :       Added socket option proto entries.
64  *                                      Also added awareness of them to accept.
65  *              Alan Cox        :       Added TCP options (SOL_TCP)
66  *              Alan Cox        :       Switched wakeup calls to callbacks,
67  *                                      so the kernel can layer network
68  *                                      sockets.
69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
71  *              Alan Cox        :       RST frames sent on unsynchronised
72  *                                      state ack error.
73  *              Alan Cox        :       Put in missing check for SYN bit.
74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
75  *                                      window non shrink trick.
76  *              Alan Cox        :       Added a couple of small NET2E timer
77  *                                      fixes
78  *              Charles Hedrick :       TCP fixes
79  *              Toomas Tamm     :       TCP window fixes
80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
81  *              Charles Hedrick :       Rewrote most of it to actually work
82  *              Linus           :       Rewrote tcp_read() and URG handling
83  *                                      completely
84  *              Gerhard Koerting:       Fixed some missing timer handling
85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
86  *              Gerhard Koerting:       PC/TCP workarounds
87  *              Adam Caldwell   :       Assorted timer/timing errors
88  *              Matthew Dillon  :       Fixed another RST bug
89  *              Alan Cox        :       Move to kernel side addressing changes.
90  *              Alan Cox        :       Beginning work on TCP fastpathing
91  *                                      (not yet usable)
92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
93  *              Alan Cox        :       TCP fast path debugging
94  *              Alan Cox        :       Window clamping
95  *              Michael Riepe   :       Bug in tcp_check()
96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
97  *              Matt Dillon     :       Yet more small nasties remove from the
98  *                                      TCP code (Be very nice to this man if
99  *                                      tcp finally works 100%) 8)
100  *              Alan Cox        :       BSD accept semantics.
101  *              Alan Cox        :       Reset on closedown bug.
102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
103  *              Michael Pall    :       Handle poll() after URG properly in
104  *                                      all cases.
105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
106  *                                      (multi URG PUSH broke rlogin).
107  *              Michael Pall    :       Fix the multi URG PUSH problem in
108  *                                      tcp_readable(), poll() after URG
109  *                                      works now.
110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
111  *                                      BSD api.
112  *              Alan Cox        :       Changed the semantics of sk->socket to
113  *                                      fix a race and a signal problem with
114  *                                      accept() and async I/O.
115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
118  *                                      clients/servers which listen in on
119  *                                      fixed ports.
120  *              Alan Cox        :       Cleaned the above up and shrank it to
121  *                                      a sensible code size.
122  *              Alan Cox        :       Self connect lockup fix.
123  *              Alan Cox        :       No connect to multicast.
124  *              Ross Biro       :       Close unaccepted children on master
125  *                                      socket close.
126  *              Alan Cox        :       Reset tracing code.
127  *              Alan Cox        :       Spurious resets on shutdown.
128  *              Alan Cox        :       Giant 15 minute/60 second timer error
129  *              Alan Cox        :       Small whoops in polling before an
130  *                                      accept.
131  *              Alan Cox        :       Kept the state trace facility since
132  *                                      it's handy for debugging.
133  *              Alan Cox        :       More reset handler fixes.
134  *              Alan Cox        :       Started rewriting the code based on
135  *                                      the RFC's for other useful protocol
136  *                                      references see: Comer, KA9Q NOS, and
137  *                                      for a reference on the difference
138  *                                      between specifications and how BSD
139  *                                      works see the 4.4lite source.
140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
141  *                                      close.
142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
144  *              Alan Cox        :       Reimplemented timers as per the RFC
145  *                                      and using multiple timers for sanity.
146  *              Alan Cox        :       Small bug fixes, and a lot of new
147  *                                      comments.
148  *              Alan Cox        :       Fixed dual reader crash by locking
149  *                                      the buffers (much like datagram.c)
150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
151  *                                      now gets fed up of retrying without
152  *                                      (even a no space) answer.
153  *              Alan Cox        :       Extracted closing code better
154  *              Alan Cox        :       Fixed the closing state machine to
155  *                                      resemble the RFC.
156  *              Alan Cox        :       More 'per spec' fixes.
157  *              Jorge Cwik      :       Even faster checksumming.
158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
159  *                                      only frames. At least one pc tcp stack
160  *                                      generates them.
161  *              Alan Cox        :       Cache last socket.
162  *              Alan Cox        :       Per route irtt.
163  *              Matt Day        :       poll()->select() match BSD precisely on error
164  *              Alan Cox        :       New buffers
165  *              Marc Tamsky     :       Various sk->prot->retransmits and
166  *                                      sk->retransmits misupdating fixed.
167  *                                      Fixed tcp_write_timeout: stuck close,
168  *                                      and TCP syn retries gets used now.
169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
170  *                                      ack if state is TCP_CLOSED.
171  *              Alan Cox        :       Look up device on a retransmit - routes may
172  *                                      change. Doesn't yet cope with MSS shrink right
173  *                                      but it's a start!
174  *              Marc Tamsky     :       Closing in closing fixes.
175  *              Mike Shaver     :       RFC1122 verifications.
176  *              Alan Cox        :       rcv_saddr errors.
177  *              Alan Cox        :       Block double connect().
178  *              Alan Cox        :       Small hooks for enSKIP.
179  *              Alexey Kuznetsov:       Path MTU discovery.
180  *              Alan Cox        :       Support soft errors.
181  *              Alan Cox        :       Fix MTU discovery pathological case
182  *                                      when the remote claims no mtu!
183  *              Marc Tamsky     :       TCP_CLOSE fix.
184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
185  *                                      window but wrong (fixes NT lpd problems)
186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
187  *              Joerg Reuter    :       No modification of locked buffers in
188  *                                      tcp_do_retransmit()
189  *              Eric Schenk     :       Changed receiver side silly window
190  *                                      avoidance algorithm to BSD style
191  *                                      algorithm. This doubles throughput
192  *                                      against machines running Solaris,
193  *                                      and seems to result in general
194  *                                      improvement.
195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
196  *      Willy Konynenberg       :       Transparent proxying support.
197  *      Mike McLagan            :       Routing by source
198  *              Keith Owens     :       Do proper merging with partial SKB's in
199  *                                      tcp_do_sendmsg to avoid burstiness.
200  *              Eric Schenk     :       Fix fast close down bug with
201  *                                      shutdown() followed by close().
202  *              Andi Kleen      :       Make poll agree with SIGIO
203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
204  *                                      lingertime == 0 (RFC 793 ABORT Call)
205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
206  *                                      csum_and_copy_from_user() if possible.
207  *
208  *              This program is free software; you can redistribute it and/or
209  *              modify it under the terms of the GNU General Public License
210  *              as published by the Free Software Foundation; either version
211  *              2 of the License, or(at your option) any later version.
212  *
213  * Description of States:
214  *
215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
216  *
217  *      TCP_SYN_RECV            received a connection request, sent ack,
218  *                              waiting for final ack in three-way handshake.
219  *
220  *      TCP_ESTABLISHED         connection established
221  *
222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
223  *                              transmission of remaining buffered data
224  *
225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
226  *                              to shutdown
227  *
228  *      TCP_CLOSING             both sides have shutdown but we still have
229  *                              data we have to finish sending
230  *
231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
232  *                              closed, can only be entered from FIN_WAIT2
233  *                              or CLOSING.  Required because the other end
234  *                              may not have gotten our last ACK causing it
235  *                              to retransmit the data packet (which we ignore)
236  *
237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
238  *                              us to finish writing our data and to shutdown
239  *                              (we have to close() to move on to LAST_ACK)
240  *
241  *      TCP_LAST_ACK            out side has shutdown after remote has
242  *                              shutdown.  There may still be data in our
243  *                              buffer that we have to finish sending
244  *
245  *      TCP_CLOSE               socket is finished
246  */
247
248 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
267
268 #include <net/icmp.h>
269 #include <net/tcp.h>
270 #include <net/xfrm.h>
271 #include <net/ip.h>
272 #include <net/netdma.h>
273 #include <net/sock.h>
274
275 #include <asm/uaccess.h>
276 #include <asm/ioctls.h>
277
278 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
279
280 struct percpu_counter tcp_orphan_count;
281 EXPORT_SYMBOL_GPL(tcp_orphan_count);
282
283 int sysctl_tcp_mem[3] __read_mostly;
284 int sysctl_tcp_wmem[3] __read_mostly;
285 int sysctl_tcp_rmem[3] __read_mostly;
286
287 EXPORT_SYMBOL(sysctl_tcp_mem);
288 EXPORT_SYMBOL(sysctl_tcp_rmem);
289 EXPORT_SYMBOL(sysctl_tcp_wmem);
290
291 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
292 EXPORT_SYMBOL(tcp_memory_allocated);
293
294 /*
295  * Current number of TCP sockets.
296  */
297 struct percpu_counter tcp_sockets_allocated;
298 EXPORT_SYMBOL(tcp_sockets_allocated);
299
300 /*
301  * TCP splice context
302  */
303 struct tcp_splice_state {
304         struct pipe_inode_info *pipe;
305         size_t len;
306         unsigned int flags;
307 };
308
309 /*
310  * Pressure flag: try to collapse.
311  * Technical note: it is used by multiple contexts non atomically.
312  * All the __sk_mem_schedule() is of this nature: accounting
313  * is strict, actions are advisory and have some latency.
314  */
315 int tcp_memory_pressure __read_mostly;
316
317 EXPORT_SYMBOL(tcp_memory_pressure);
318
319 void tcp_enter_memory_pressure(struct sock *sk)
320 {
321         if (!tcp_memory_pressure) {
322                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
323                 tcp_memory_pressure = 1;
324         }
325 }
326
327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
328
329 /* Convert seconds to retransmits based on initial and max timeout */
330 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
331 {
332         u8 res = 0;
333
334         if (seconds > 0) {
335                 int period = timeout;
336
337                 res = 1;
338                 while (seconds > period && res < 255) {
339                         res++;
340                         timeout <<= 1;
341                         if (timeout > rto_max)
342                                 timeout = rto_max;
343                         period += timeout;
344                 }
345         }
346         return res;
347 }
348
349 /* Convert retransmits to seconds based on initial and max timeout */
350 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
351 {
352         int period = 0;
353
354         if (retrans > 0) {
355                 period = timeout;
356                 while (--retrans) {
357                         timeout <<= 1;
358                         if (timeout > rto_max)
359                                 timeout = rto_max;
360                         period += timeout;
361                 }
362         }
363         return period;
364 }
365
366 /*
367  *      Wait for a TCP event.
368  *
369  *      Note that we don't need to lock the socket, as the upper poll layers
370  *      take care of normal races (between the test and the event) and we don't
371  *      go look at any of the socket buffers directly.
372  */
373 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
374 {
375         unsigned int mask;
376         struct sock *sk = sock->sk;
377         struct tcp_sock *tp = tcp_sk(sk);
378
379         sock_poll_wait(file, sk->sk_sleep, wait);
380         if (sk->sk_state == TCP_LISTEN)
381                 return inet_csk_listen_poll(sk);
382
383         /* Socket is not locked. We are protected from async events
384          * by poll logic and correct handling of state changes
385          * made by other threads is impossible in any case.
386          */
387
388         mask = 0;
389         if (sk->sk_err)
390                 mask = POLLERR;
391
392         /*
393          * POLLHUP is certainly not done right. But poll() doesn't
394          * have a notion of HUP in just one direction, and for a
395          * socket the read side is more interesting.
396          *
397          * Some poll() documentation says that POLLHUP is incompatible
398          * with the POLLOUT/POLLWR flags, so somebody should check this
399          * all. But careful, it tends to be safer to return too many
400          * bits than too few, and you can easily break real applications
401          * if you don't tell them that something has hung up!
402          *
403          * Check-me.
404          *
405          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
406          * our fs/select.c). It means that after we received EOF,
407          * poll always returns immediately, making impossible poll() on write()
408          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
409          * if and only if shutdown has been made in both directions.
410          * Actually, it is interesting to look how Solaris and DUX
411          * solve this dilemma. I would prefer, if POLLHUP were maskable,
412          * then we could set it on SND_SHUTDOWN. BTW examples given
413          * in Stevens' books assume exactly this behaviour, it explains
414          * why POLLHUP is incompatible with POLLOUT.    --ANK
415          *
416          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
417          * blocking on fresh not-connected or disconnected socket. --ANK
418          */
419         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
420                 mask |= POLLHUP;
421         if (sk->sk_shutdown & RCV_SHUTDOWN)
422                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
423
424         /* Connected? */
425         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
426                 int target = sock_rcvlowat(sk, 0, INT_MAX);
427
428                 if (tp->urg_seq == tp->copied_seq &&
429                     !sock_flag(sk, SOCK_URGINLINE) &&
430                     tp->urg_data)
431                         target--;
432
433                 /* Potential race condition. If read of tp below will
434                  * escape above sk->sk_state, we can be illegally awaken
435                  * in SYN_* states. */
436                 if (tp->rcv_nxt - tp->copied_seq >= target)
437                         mask |= POLLIN | POLLRDNORM;
438
439                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
440                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
441                                 mask |= POLLOUT | POLLWRNORM;
442                         } else {  /* send SIGIO later */
443                                 set_bit(SOCK_ASYNC_NOSPACE,
444                                         &sk->sk_socket->flags);
445                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
446
447                                 /* Race breaker. If space is freed after
448                                  * wspace test but before the flags are set,
449                                  * IO signal will be lost.
450                                  */
451                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
452                                         mask |= POLLOUT | POLLWRNORM;
453                         }
454                 }
455
456                 if (tp->urg_data & TCP_URG_VALID)
457                         mask |= POLLPRI;
458         }
459         return mask;
460 }
461
462 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
463 {
464         struct tcp_sock *tp = tcp_sk(sk);
465         int answ;
466
467         switch (cmd) {
468         case SIOCINQ:
469                 if (sk->sk_state == TCP_LISTEN)
470                         return -EINVAL;
471
472                 lock_sock(sk);
473                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
474                         answ = 0;
475                 else if (sock_flag(sk, SOCK_URGINLINE) ||
476                          !tp->urg_data ||
477                          before(tp->urg_seq, tp->copied_seq) ||
478                          !before(tp->urg_seq, tp->rcv_nxt)) {
479                         struct sk_buff *skb;
480
481                         answ = tp->rcv_nxt - tp->copied_seq;
482
483                         /* Subtract 1, if FIN is in queue. */
484                         skb = skb_peek_tail(&sk->sk_receive_queue);
485                         if (answ && skb)
486                                 answ -= tcp_hdr(skb)->fin;
487                 } else
488                         answ = tp->urg_seq - tp->copied_seq;
489                 release_sock(sk);
490                 break;
491         case SIOCATMARK:
492                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
493                 break;
494         case SIOCOUTQ:
495                 if (sk->sk_state == TCP_LISTEN)
496                         return -EINVAL;
497
498                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
499                         answ = 0;
500                 else
501                         answ = tp->write_seq - tp->snd_una;
502                 break;
503         default:
504                 return -ENOIOCTLCMD;
505         }
506
507         return put_user(answ, (int __user *)arg);
508 }
509
510 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
511 {
512         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
513         tp->pushed_seq = tp->write_seq;
514 }
515
516 static inline int forced_push(struct tcp_sock *tp)
517 {
518         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
519 }
520
521 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
522 {
523         struct tcp_sock *tp = tcp_sk(sk);
524         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
525
526         skb->csum    = 0;
527         tcb->seq     = tcb->end_seq = tp->write_seq;
528         tcb->flags   = TCPCB_FLAG_ACK;
529         tcb->sacked  = 0;
530         skb_header_release(skb);
531         tcp_add_write_queue_tail(sk, skb);
532         sk->sk_wmem_queued += skb->truesize;
533         sk_mem_charge(sk, skb->truesize);
534         if (tp->nonagle & TCP_NAGLE_PUSH)
535                 tp->nonagle &= ~TCP_NAGLE_PUSH;
536 }
537
538 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
539                                 struct sk_buff *skb)
540 {
541         if (flags & MSG_OOB)
542                 tp->snd_up = tp->write_seq;
543 }
544
545 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
546                             int nonagle)
547 {
548         struct tcp_sock *tp = tcp_sk(sk);
549
550         if (tcp_send_head(sk)) {
551                 struct sk_buff *skb = tcp_write_queue_tail(sk);
552                 if (!(flags & MSG_MORE) || forced_push(tp))
553                         tcp_mark_push(tp, skb);
554                 tcp_mark_urg(tp, flags, skb);
555                 __tcp_push_pending_frames(sk, mss_now,
556                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
557         }
558 }
559
560 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
561                                 unsigned int offset, size_t len)
562 {
563         struct tcp_splice_state *tss = rd_desc->arg.data;
564         int ret;
565
566         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
567                               tss->flags);
568         if (ret > 0)
569                 rd_desc->count -= ret;
570         return ret;
571 }
572
573 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
574 {
575         /* Store TCP splice context information in read_descriptor_t. */
576         read_descriptor_t rd_desc = {
577                 .arg.data = tss,
578                 .count    = tss->len,
579         };
580
581         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
582 }
583
584 /**
585  *  tcp_splice_read - splice data from TCP socket to a pipe
586  * @sock:       socket to splice from
587  * @ppos:       position (not valid)
588  * @pipe:       pipe to splice to
589  * @len:        number of bytes to splice
590  * @flags:      splice modifier flags
591  *
592  * Description:
593  *    Will read pages from given socket and fill them into a pipe.
594  *
595  **/
596 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
597                         struct pipe_inode_info *pipe, size_t len,
598                         unsigned int flags)
599 {
600         struct sock *sk = sock->sk;
601         struct tcp_splice_state tss = {
602                 .pipe = pipe,
603                 .len = len,
604                 .flags = flags,
605         };
606         long timeo;
607         ssize_t spliced;
608         int ret;
609
610         /*
611          * We can't seek on a socket input
612          */
613         if (unlikely(*ppos))
614                 return -ESPIPE;
615
616         ret = spliced = 0;
617
618         lock_sock(sk);
619
620         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
621         while (tss.len) {
622                 ret = __tcp_splice_read(sk, &tss);
623                 if (ret < 0)
624                         break;
625                 else if (!ret) {
626                         if (spliced)
627                                 break;
628                         if (sock_flag(sk, SOCK_DONE))
629                                 break;
630                         if (sk->sk_err) {
631                                 ret = sock_error(sk);
632                                 break;
633                         }
634                         if (sk->sk_shutdown & RCV_SHUTDOWN)
635                                 break;
636                         if (sk->sk_state == TCP_CLOSE) {
637                                 /*
638                                  * This occurs when user tries to read
639                                  * from never connected socket.
640                                  */
641                                 if (!sock_flag(sk, SOCK_DONE))
642                                         ret = -ENOTCONN;
643                                 break;
644                         }
645                         if (!timeo) {
646                                 ret = -EAGAIN;
647                                 break;
648                         }
649                         sk_wait_data(sk, &timeo);
650                         if (signal_pending(current)) {
651                                 ret = sock_intr_errno(timeo);
652                                 break;
653                         }
654                         continue;
655                 }
656                 tss.len -= ret;
657                 spliced += ret;
658
659                 if (!timeo)
660                         break;
661                 release_sock(sk);
662                 lock_sock(sk);
663
664                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
665                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
666                     signal_pending(current))
667                         break;
668         }
669
670         release_sock(sk);
671
672         if (spliced)
673                 return spliced;
674
675         return ret;
676 }
677
678 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
679 {
680         struct sk_buff *skb;
681
682         /* The TCP header must be at least 32-bit aligned.  */
683         size = ALIGN(size, 4);
684
685         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
686         if (skb) {
687                 if (sk_wmem_schedule(sk, skb->truesize)) {
688                         /*
689                          * Make sure that we have exactly size bytes
690                          * available to the caller, no more, no less.
691                          */
692                         skb_reserve(skb, skb_tailroom(skb) - size);
693                         return skb;
694                 }
695                 __kfree_skb(skb);
696         } else {
697                 sk->sk_prot->enter_memory_pressure(sk);
698                 sk_stream_moderate_sndbuf(sk);
699         }
700         return NULL;
701 }
702
703 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
704                                        int large_allowed)
705 {
706         struct tcp_sock *tp = tcp_sk(sk);
707         u32 xmit_size_goal, old_size_goal;
708
709         xmit_size_goal = mss_now;
710
711         if (large_allowed && sk_can_gso(sk)) {
712                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
713                                   inet_csk(sk)->icsk_af_ops->net_header_len -
714                                   inet_csk(sk)->icsk_ext_hdr_len -
715                                   tp->tcp_header_len);
716
717                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
718
719                 /* We try hard to avoid divides here */
720                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
721
722                 if (likely(old_size_goal <= xmit_size_goal &&
723                            old_size_goal + mss_now > xmit_size_goal)) {
724                         xmit_size_goal = old_size_goal;
725                 } else {
726                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
727                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
728                 }
729         }
730
731         return max(xmit_size_goal, mss_now);
732 }
733
734 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
735 {
736         int mss_now;
737
738         mss_now = tcp_current_mss(sk);
739         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
740
741         return mss_now;
742 }
743
744 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
745                          size_t psize, int flags)
746 {
747         struct tcp_sock *tp = tcp_sk(sk);
748         int mss_now, size_goal;
749         int err;
750         ssize_t copied;
751         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
752
753         /* Wait for a connection to finish. */
754         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
755                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
756                         goto out_err;
757
758         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
759
760         mss_now = tcp_send_mss(sk, &size_goal, flags);
761         copied = 0;
762
763         err = -EPIPE;
764         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
765                 goto out_err;
766
767         while (psize > 0) {
768                 struct sk_buff *skb = tcp_write_queue_tail(sk);
769                 struct page *page = pages[poffset / PAGE_SIZE];
770                 int copy, i, can_coalesce;
771                 int offset = poffset % PAGE_SIZE;
772                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
773
774                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
775 new_segment:
776                         if (!sk_stream_memory_free(sk))
777                                 goto wait_for_sndbuf;
778
779                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
780                         if (!skb)
781                                 goto wait_for_memory;
782
783                         skb_entail(sk, skb);
784                         copy = size_goal;
785                 }
786
787                 if (copy > size)
788                         copy = size;
789
790                 i = skb_shinfo(skb)->nr_frags;
791                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
792                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
793                         tcp_mark_push(tp, skb);
794                         goto new_segment;
795                 }
796                 if (!sk_wmem_schedule(sk, copy))
797                         goto wait_for_memory;
798
799                 if (can_coalesce) {
800                         skb_shinfo(skb)->frags[i - 1].size += copy;
801                 } else {
802                         get_page(page);
803                         skb_fill_page_desc(skb, i, page, offset, copy);
804                 }
805
806                 skb->len += copy;
807                 skb->data_len += copy;
808                 skb->truesize += copy;
809                 sk->sk_wmem_queued += copy;
810                 sk_mem_charge(sk, copy);
811                 skb->ip_summed = CHECKSUM_PARTIAL;
812                 tp->write_seq += copy;
813                 TCP_SKB_CB(skb)->end_seq += copy;
814                 skb_shinfo(skb)->gso_segs = 0;
815
816                 if (!copied)
817                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
818
819                 copied += copy;
820                 poffset += copy;
821                 if (!(psize -= copy))
822                         goto out;
823
824                 if (skb->len < size_goal || (flags & MSG_OOB))
825                         continue;
826
827                 if (forced_push(tp)) {
828                         tcp_mark_push(tp, skb);
829                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
830                 } else if (skb == tcp_send_head(sk))
831                         tcp_push_one(sk, mss_now);
832                 continue;
833
834 wait_for_sndbuf:
835                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
836 wait_for_memory:
837                 if (copied)
838                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
839
840                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
841                         goto do_error;
842
843                 mss_now = tcp_send_mss(sk, &size_goal, flags);
844         }
845
846 out:
847         if (copied)
848                 tcp_push(sk, flags, mss_now, tp->nonagle);
849         return copied;
850
851 do_error:
852         if (copied)
853                 goto out;
854 out_err:
855         return sk_stream_error(sk, flags, err);
856 }
857
858 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
859                      size_t size, int flags)
860 {
861         ssize_t res;
862         struct sock *sk = sock->sk;
863
864         if (!(sk->sk_route_caps & NETIF_F_SG) ||
865             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
866                 return sock_no_sendpage(sock, page, offset, size, flags);
867
868         lock_sock(sk);
869         TCP_CHECK_TIMER(sk);
870         res = do_tcp_sendpages(sk, &page, offset, size, flags);
871         TCP_CHECK_TIMER(sk);
872         release_sock(sk);
873         return res;
874 }
875
876 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
877 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
878
879 static inline int select_size(struct sock *sk)
880 {
881         struct tcp_sock *tp = tcp_sk(sk);
882         int tmp = tp->mss_cache;
883
884         if (sk->sk_route_caps & NETIF_F_SG) {
885                 if (sk_can_gso(sk))
886                         tmp = 0;
887                 else {
888                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
889
890                         if (tmp >= pgbreak &&
891                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
892                                 tmp = pgbreak;
893                 }
894         }
895
896         return tmp;
897 }
898
899 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
900                 size_t size)
901 {
902         struct sock *sk = sock->sk;
903         struct iovec *iov;
904         struct tcp_sock *tp = tcp_sk(sk);
905         struct sk_buff *skb;
906         int iovlen, flags;
907         int mss_now, size_goal;
908         int err, copied;
909         long timeo;
910
911         lock_sock(sk);
912         TCP_CHECK_TIMER(sk);
913
914         flags = msg->msg_flags;
915         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
916
917         /* Wait for a connection to finish. */
918         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
919                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
920                         goto out_err;
921
922         /* This should be in poll */
923         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
924
925         mss_now = tcp_send_mss(sk, &size_goal, flags);
926
927         /* Ok commence sending. */
928         iovlen = msg->msg_iovlen;
929         iov = msg->msg_iov;
930         copied = 0;
931
932         err = -EPIPE;
933         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
934                 goto out_err;
935
936         while (--iovlen >= 0) {
937                 int seglen = iov->iov_len;
938                 unsigned char __user *from = iov->iov_base;
939
940                 iov++;
941
942                 while (seglen > 0) {
943                         int copy = 0;
944                         int max = size_goal;
945
946                         skb = tcp_write_queue_tail(sk);
947                         if (tcp_send_head(sk)) {
948                                 if (skb->ip_summed == CHECKSUM_NONE)
949                                         max = mss_now;
950                                 copy = max - skb->len;
951                         }
952
953                         if (copy <= 0) {
954 new_segment:
955                                 /* Allocate new segment. If the interface is SG,
956                                  * allocate skb fitting to single page.
957                                  */
958                                 if (!sk_stream_memory_free(sk))
959                                         goto wait_for_sndbuf;
960
961                                 skb = sk_stream_alloc_skb(sk, select_size(sk),
962                                                 sk->sk_allocation);
963                                 if (!skb)
964                                         goto wait_for_memory;
965
966                                 /*
967                                  * Check whether we can use HW checksum.
968                                  */
969                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
970                                         skb->ip_summed = CHECKSUM_PARTIAL;
971
972                                 skb_entail(sk, skb);
973                                 copy = size_goal;
974                                 max = size_goal;
975                         }
976
977                         /* Try to append data to the end of skb. */
978                         if (copy > seglen)
979                                 copy = seglen;
980
981                         /* Where to copy to? */
982                         if (skb_tailroom(skb) > 0) {
983                                 /* We have some space in skb head. Superb! */
984                                 if (copy > skb_tailroom(skb))
985                                         copy = skb_tailroom(skb);
986                                 if ((err = skb_add_data(skb, from, copy)) != 0)
987                                         goto do_fault;
988                         } else {
989                                 int merge = 0;
990                                 int i = skb_shinfo(skb)->nr_frags;
991                                 struct page *page = TCP_PAGE(sk);
992                                 int off = TCP_OFF(sk);
993
994                                 if (skb_can_coalesce(skb, i, page, off) &&
995                                     off != PAGE_SIZE) {
996                                         /* We can extend the last page
997                                          * fragment. */
998                                         merge = 1;
999                                 } else if (i == MAX_SKB_FRAGS ||
1000                                            (!i &&
1001                                            !(sk->sk_route_caps & NETIF_F_SG))) {
1002                                         /* Need to add new fragment and cannot
1003                                          * do this because interface is non-SG,
1004                                          * or because all the page slots are
1005                                          * busy. */
1006                                         tcp_mark_push(tp, skb);
1007                                         goto new_segment;
1008                                 } else if (page) {
1009                                         if (off == PAGE_SIZE) {
1010                                                 put_page(page);
1011                                                 TCP_PAGE(sk) = page = NULL;
1012                                                 off = 0;
1013                                         }
1014                                 } else
1015                                         off = 0;
1016
1017                                 if (copy > PAGE_SIZE - off)
1018                                         copy = PAGE_SIZE - off;
1019
1020                                 if (!sk_wmem_schedule(sk, copy))
1021                                         goto wait_for_memory;
1022
1023                                 if (!page) {
1024                                         /* Allocate new cache page. */
1025                                         if (!(page = sk_stream_alloc_page(sk)))
1026                                                 goto wait_for_memory;
1027                                 }
1028
1029                                 /* Time to copy data. We are close to
1030                                  * the end! */
1031                                 err = skb_copy_to_page(sk, from, skb, page,
1032                                                        off, copy);
1033                                 if (err) {
1034                                         /* If this page was new, give it to the
1035                                          * socket so it does not get leaked.
1036                                          */
1037                                         if (!TCP_PAGE(sk)) {
1038                                                 TCP_PAGE(sk) = page;
1039                                                 TCP_OFF(sk) = 0;
1040                                         }
1041                                         goto do_error;
1042                                 }
1043
1044                                 /* Update the skb. */
1045                                 if (merge) {
1046                                         skb_shinfo(skb)->frags[i - 1].size +=
1047                                                                         copy;
1048                                 } else {
1049                                         skb_fill_page_desc(skb, i, page, off, copy);
1050                                         if (TCP_PAGE(sk)) {
1051                                                 get_page(page);
1052                                         } else if (off + copy < PAGE_SIZE) {
1053                                                 get_page(page);
1054                                                 TCP_PAGE(sk) = page;
1055                                         }
1056                                 }
1057
1058                                 TCP_OFF(sk) = off + copy;
1059                         }
1060
1061                         if (!copied)
1062                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
1063
1064                         tp->write_seq += copy;
1065                         TCP_SKB_CB(skb)->end_seq += copy;
1066                         skb_shinfo(skb)->gso_segs = 0;
1067
1068                         from += copy;
1069                         copied += copy;
1070                         if ((seglen -= copy) == 0 && iovlen == 0)
1071                                 goto out;
1072
1073                         if (skb->len < max || (flags & MSG_OOB))
1074                                 continue;
1075
1076                         if (forced_push(tp)) {
1077                                 tcp_mark_push(tp, skb);
1078                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1079                         } else if (skb == tcp_send_head(sk))
1080                                 tcp_push_one(sk, mss_now);
1081                         continue;
1082
1083 wait_for_sndbuf:
1084                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1085 wait_for_memory:
1086                         if (copied)
1087                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1088
1089                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1090                                 goto do_error;
1091
1092                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1093                 }
1094         }
1095
1096 out:
1097         if (copied)
1098                 tcp_push(sk, flags, mss_now, tp->nonagle);
1099         TCP_CHECK_TIMER(sk);
1100         release_sock(sk);
1101         return copied;
1102
1103 do_fault:
1104         if (!skb->len) {
1105                 tcp_unlink_write_queue(skb, sk);
1106                 /* It is the one place in all of TCP, except connection
1107                  * reset, where we can be unlinking the send_head.
1108                  */
1109                 tcp_check_send_head(sk, skb);
1110                 sk_wmem_free_skb(sk, skb);
1111         }
1112
1113 do_error:
1114         if (copied)
1115                 goto out;
1116 out_err:
1117         err = sk_stream_error(sk, flags, err);
1118         TCP_CHECK_TIMER(sk);
1119         release_sock(sk);
1120         return err;
1121 }
1122
1123 /*
1124  *      Handle reading urgent data. BSD has very simple semantics for
1125  *      this, no blocking and very strange errors 8)
1126  */
1127
1128 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1129 {
1130         struct tcp_sock *tp = tcp_sk(sk);
1131
1132         /* No URG data to read. */
1133         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1134             tp->urg_data == TCP_URG_READ)
1135                 return -EINVAL; /* Yes this is right ! */
1136
1137         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1138                 return -ENOTCONN;
1139
1140         if (tp->urg_data & TCP_URG_VALID) {
1141                 int err = 0;
1142                 char c = tp->urg_data;
1143
1144                 if (!(flags & MSG_PEEK))
1145                         tp->urg_data = TCP_URG_READ;
1146
1147                 /* Read urgent data. */
1148                 msg->msg_flags |= MSG_OOB;
1149
1150                 if (len > 0) {
1151                         if (!(flags & MSG_TRUNC))
1152                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1153                         len = 1;
1154                 } else
1155                         msg->msg_flags |= MSG_TRUNC;
1156
1157                 return err ? -EFAULT : len;
1158         }
1159
1160         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1161                 return 0;
1162
1163         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1164          * the available implementations agree in this case:
1165          * this call should never block, independent of the
1166          * blocking state of the socket.
1167          * Mike <pall@rz.uni-karlsruhe.de>
1168          */
1169         return -EAGAIN;
1170 }
1171
1172 /* Clean up the receive buffer for full frames taken by the user,
1173  * then send an ACK if necessary.  COPIED is the number of bytes
1174  * tcp_recvmsg has given to the user so far, it speeds up the
1175  * calculation of whether or not we must ACK for the sake of
1176  * a window update.
1177  */
1178 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1179 {
1180         struct tcp_sock *tp = tcp_sk(sk);
1181         int time_to_ack = 0;
1182
1183 #if TCP_DEBUG
1184         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1185
1186         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1187              KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1188              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1189 #endif
1190
1191         if (inet_csk_ack_scheduled(sk)) {
1192                 const struct inet_connection_sock *icsk = inet_csk(sk);
1193                    /* Delayed ACKs frequently hit locked sockets during bulk
1194                     * receive. */
1195                 if (icsk->icsk_ack.blocked ||
1196                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1197                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1198                     /*
1199                      * If this read emptied read buffer, we send ACK, if
1200                      * connection is not bidirectional, user drained
1201                      * receive buffer and there was a small segment
1202                      * in queue.
1203                      */
1204                     (copied > 0 &&
1205                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1206                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1207                        !icsk->icsk_ack.pingpong)) &&
1208                       !atomic_read(&sk->sk_rmem_alloc)))
1209                         time_to_ack = 1;
1210         }
1211
1212         /* We send an ACK if we can now advertise a non-zero window
1213          * which has been raised "significantly".
1214          *
1215          * Even if window raised up to infinity, do not send window open ACK
1216          * in states, where we will not receive more. It is useless.
1217          */
1218         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1219                 __u32 rcv_window_now = tcp_receive_window(tp);
1220
1221                 /* Optimize, __tcp_select_window() is not cheap. */
1222                 if (2*rcv_window_now <= tp->window_clamp) {
1223                         __u32 new_window = __tcp_select_window(sk);
1224
1225                         /* Send ACK now, if this read freed lots of space
1226                          * in our buffer. Certainly, new_window is new window.
1227                          * We can advertise it now, if it is not less than current one.
1228                          * "Lots" means "at least twice" here.
1229                          */
1230                         if (new_window && new_window >= 2 * rcv_window_now)
1231                                 time_to_ack = 1;
1232                 }
1233         }
1234         if (time_to_ack)
1235                 tcp_send_ack(sk);
1236 }
1237
1238 static void tcp_prequeue_process(struct sock *sk)
1239 {
1240         struct sk_buff *skb;
1241         struct tcp_sock *tp = tcp_sk(sk);
1242
1243         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1244
1245         /* RX process wants to run with disabled BHs, though it is not
1246          * necessary */
1247         local_bh_disable();
1248         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1249                 sk_backlog_rcv(sk, skb);
1250         local_bh_enable();
1251
1252         /* Clear memory counter. */
1253         tp->ucopy.memory = 0;
1254 }
1255
1256 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1257 {
1258         struct sk_buff *skb;
1259         u32 offset;
1260
1261         skb_queue_walk(&sk->sk_receive_queue, skb) {
1262                 offset = seq - TCP_SKB_CB(skb)->seq;
1263                 if (tcp_hdr(skb)->syn)
1264                         offset--;
1265                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1266                         *off = offset;
1267                         return skb;
1268                 }
1269         }
1270         return NULL;
1271 }
1272
1273 /*
1274  * This routine provides an alternative to tcp_recvmsg() for routines
1275  * that would like to handle copying from skbuffs directly in 'sendfile'
1276  * fashion.
1277  * Note:
1278  *      - It is assumed that the socket was locked by the caller.
1279  *      - The routine does not block.
1280  *      - At present, there is no support for reading OOB data
1281  *        or for 'peeking' the socket using this routine
1282  *        (although both would be easy to implement).
1283  */
1284 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1285                   sk_read_actor_t recv_actor)
1286 {
1287         struct sk_buff *skb;
1288         struct tcp_sock *tp = tcp_sk(sk);
1289         u32 seq = tp->copied_seq;
1290         u32 offset;
1291         int copied = 0;
1292
1293         if (sk->sk_state == TCP_LISTEN)
1294                 return -ENOTCONN;
1295         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1296                 if (offset < skb->len) {
1297                         int used;
1298                         size_t len;
1299
1300                         len = skb->len - offset;
1301                         /* Stop reading if we hit a patch of urgent data */
1302                         if (tp->urg_data) {
1303                                 u32 urg_offset = tp->urg_seq - seq;
1304                                 if (urg_offset < len)
1305                                         len = urg_offset;
1306                                 if (!len)
1307                                         break;
1308                         }
1309                         used = recv_actor(desc, skb, offset, len);
1310                         if (used < 0) {
1311                                 if (!copied)
1312                                         copied = used;
1313                                 break;
1314                         } else if (used <= len) {
1315                                 seq += used;
1316                                 copied += used;
1317                                 offset += used;
1318                         }
1319                         /*
1320                          * If recv_actor drops the lock (e.g. TCP splice
1321                          * receive) the skb pointer might be invalid when
1322                          * getting here: tcp_collapse might have deleted it
1323                          * while aggregating skbs from the socket queue.
1324                          */
1325                         skb = tcp_recv_skb(sk, seq-1, &offset);
1326                         if (!skb || (offset+1 != skb->len))
1327                                 break;
1328                 }
1329                 if (tcp_hdr(skb)->fin) {
1330                         sk_eat_skb(sk, skb, 0);
1331                         ++seq;
1332                         break;
1333                 }
1334                 sk_eat_skb(sk, skb, 0);
1335                 if (!desc->count)
1336                         break;
1337         }
1338         tp->copied_seq = seq;
1339
1340         tcp_rcv_space_adjust(sk);
1341
1342         /* Clean up data we have read: This will do ACK frames. */
1343         if (copied > 0)
1344                 tcp_cleanup_rbuf(sk, copied);
1345         return copied;
1346 }
1347
1348 /*
1349  *      This routine copies from a sock struct into the user buffer.
1350  *
1351  *      Technical note: in 2.3 we work on _locked_ socket, so that
1352  *      tricks with *seq access order and skb->users are not required.
1353  *      Probably, code can be easily improved even more.
1354  */
1355
1356 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1357                 size_t len, int nonblock, int flags, int *addr_len)
1358 {
1359         struct tcp_sock *tp = tcp_sk(sk);
1360         int copied = 0;
1361         u32 peek_seq;
1362         u32 *seq;
1363         unsigned long used;
1364         int err;
1365         int target;             /* Read at least this many bytes */
1366         long timeo;
1367         struct task_struct *user_recv = NULL;
1368         int copied_early = 0;
1369         struct sk_buff *skb;
1370         u32 urg_hole = 0;
1371
1372         lock_sock(sk);
1373
1374         TCP_CHECK_TIMER(sk);
1375
1376         err = -ENOTCONN;
1377         if (sk->sk_state == TCP_LISTEN)
1378                 goto out;
1379
1380         timeo = sock_rcvtimeo(sk, nonblock);
1381
1382         /* Urgent data needs to be handled specially. */
1383         if (flags & MSG_OOB)
1384                 goto recv_urg;
1385
1386         seq = &tp->copied_seq;
1387         if (flags & MSG_PEEK) {
1388                 peek_seq = tp->copied_seq;
1389                 seq = &peek_seq;
1390         }
1391
1392         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1393
1394 #ifdef CONFIG_NET_DMA
1395         tp->ucopy.dma_chan = NULL;
1396         preempt_disable();
1397         skb = skb_peek_tail(&sk->sk_receive_queue);
1398         {
1399                 int available = 0;
1400
1401                 if (skb)
1402                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1403                 if ((available < target) &&
1404                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1405                     !sysctl_tcp_low_latency &&
1406                     dma_find_channel(DMA_MEMCPY)) {
1407                         preempt_enable_no_resched();
1408                         tp->ucopy.pinned_list =
1409                                         dma_pin_iovec_pages(msg->msg_iov, len);
1410                 } else {
1411                         preempt_enable_no_resched();
1412                 }
1413         }
1414 #endif
1415
1416         do {
1417                 u32 offset;
1418
1419                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1420                 if (tp->urg_data && tp->urg_seq == *seq) {
1421                         if (copied)
1422                                 break;
1423                         if (signal_pending(current)) {
1424                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1425                                 break;
1426                         }
1427                 }
1428
1429                 /* Next get a buffer. */
1430
1431                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1432                         /* Now that we have two receive queues this
1433                          * shouldn't happen.
1434                          */
1435                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1436                              KERN_INFO "recvmsg bug: copied %X "
1437                                        "seq %X rcvnxt %X fl %X\n", *seq,
1438                                        TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1439                                        flags))
1440                                 break;
1441
1442                         offset = *seq - TCP_SKB_CB(skb)->seq;
1443                         if (tcp_hdr(skb)->syn)
1444                                 offset--;
1445                         if (offset < skb->len)
1446                                 goto found_ok_skb;
1447                         if (tcp_hdr(skb)->fin)
1448                                 goto found_fin_ok;
1449                         WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: "
1450                                         "copied %X seq %X rcvnxt %X fl %X\n",
1451                                         *seq, TCP_SKB_CB(skb)->seq,
1452                                         tp->rcv_nxt, flags);
1453                 }
1454
1455                 /* Well, if we have backlog, try to process it now yet. */
1456
1457                 if (copied >= target && !sk->sk_backlog.tail)
1458                         break;
1459
1460                 if (copied) {
1461                         if (sk->sk_err ||
1462                             sk->sk_state == TCP_CLOSE ||
1463                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1464                             !timeo ||
1465                             signal_pending(current))
1466                                 break;
1467                 } else {
1468                         if (sock_flag(sk, SOCK_DONE))
1469                                 break;
1470
1471                         if (sk->sk_err) {
1472                                 copied = sock_error(sk);
1473                                 break;
1474                         }
1475
1476                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1477                                 break;
1478
1479                         if (sk->sk_state == TCP_CLOSE) {
1480                                 if (!sock_flag(sk, SOCK_DONE)) {
1481                                         /* This occurs when user tries to read
1482                                          * from never connected socket.
1483                                          */
1484                                         copied = -ENOTCONN;
1485                                         break;
1486                                 }
1487                                 break;
1488                         }
1489
1490                         if (!timeo) {
1491                                 copied = -EAGAIN;
1492                                 break;
1493                         }
1494
1495                         if (signal_pending(current)) {
1496                                 copied = sock_intr_errno(timeo);
1497                                 break;
1498                         }
1499                 }
1500
1501                 tcp_cleanup_rbuf(sk, copied);
1502
1503                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1504                         /* Install new reader */
1505                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1506                                 user_recv = current;
1507                                 tp->ucopy.task = user_recv;
1508                                 tp->ucopy.iov = msg->msg_iov;
1509                         }
1510
1511                         tp->ucopy.len = len;
1512
1513                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1514                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1515
1516                         /* Ugly... If prequeue is not empty, we have to
1517                          * process it before releasing socket, otherwise
1518                          * order will be broken at second iteration.
1519                          * More elegant solution is required!!!
1520                          *
1521                          * Look: we have the following (pseudo)queues:
1522                          *
1523                          * 1. packets in flight
1524                          * 2. backlog
1525                          * 3. prequeue
1526                          * 4. receive_queue
1527                          *
1528                          * Each queue can be processed only if the next ones
1529                          * are empty. At this point we have empty receive_queue.
1530                          * But prequeue _can_ be not empty after 2nd iteration,
1531                          * when we jumped to start of loop because backlog
1532                          * processing added something to receive_queue.
1533                          * We cannot release_sock(), because backlog contains
1534                          * packets arrived _after_ prequeued ones.
1535                          *
1536                          * Shortly, algorithm is clear --- to process all
1537                          * the queues in order. We could make it more directly,
1538                          * requeueing packets from backlog to prequeue, if
1539                          * is not empty. It is more elegant, but eats cycles,
1540                          * unfortunately.
1541                          */
1542                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1543                                 goto do_prequeue;
1544
1545                         /* __ Set realtime policy in scheduler __ */
1546                 }
1547
1548                 if (copied >= target) {
1549                         /* Do not sleep, just process backlog. */
1550                         release_sock(sk);
1551                         lock_sock(sk);
1552                 } else
1553                         sk_wait_data(sk, &timeo);
1554
1555 #ifdef CONFIG_NET_DMA
1556                 tp->ucopy.wakeup = 0;
1557 #endif
1558
1559                 if (user_recv) {
1560                         int chunk;
1561
1562                         /* __ Restore normal policy in scheduler __ */
1563
1564                         if ((chunk = len - tp->ucopy.len) != 0) {
1565                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1566                                 len -= chunk;
1567                                 copied += chunk;
1568                         }
1569
1570                         if (tp->rcv_nxt == tp->copied_seq &&
1571                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1572 do_prequeue:
1573                                 tcp_prequeue_process(sk);
1574
1575                                 if ((chunk = len - tp->ucopy.len) != 0) {
1576                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1577                                         len -= chunk;
1578                                         copied += chunk;
1579                                 }
1580                         }
1581                 }
1582                 if ((flags & MSG_PEEK) &&
1583                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1584                         if (net_ratelimit())
1585                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1586                                        current->comm, task_pid_nr(current));
1587                         peek_seq = tp->copied_seq;
1588                 }
1589                 continue;
1590
1591         found_ok_skb:
1592                 /* Ok so how much can we use? */
1593                 used = skb->len - offset;
1594                 if (len < used)
1595                         used = len;
1596
1597                 /* Do we have urgent data here? */
1598                 if (tp->urg_data) {
1599                         u32 urg_offset = tp->urg_seq - *seq;
1600                         if (urg_offset < used) {
1601                                 if (!urg_offset) {
1602                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1603                                                 ++*seq;
1604                                                 urg_hole++;
1605                                                 offset++;
1606                                                 used--;
1607                                                 if (!used)
1608                                                         goto skip_copy;
1609                                         }
1610                                 } else
1611                                         used = urg_offset;
1612                         }
1613                 }
1614
1615                 if (!(flags & MSG_TRUNC)) {
1616 #ifdef CONFIG_NET_DMA
1617                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1618                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1619
1620                         if (tp->ucopy.dma_chan) {
1621                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1622                                         tp->ucopy.dma_chan, skb, offset,
1623                                         msg->msg_iov, used,
1624                                         tp->ucopy.pinned_list);
1625
1626                                 if (tp->ucopy.dma_cookie < 0) {
1627
1628                                         printk(KERN_ALERT "dma_cookie < 0\n");
1629
1630                                         /* Exception. Bailout! */
1631                                         if (!copied)
1632                                                 copied = -EFAULT;
1633                                         break;
1634                                 }
1635                                 if ((offset + used) == skb->len)
1636                                         copied_early = 1;
1637
1638                         } else
1639 #endif
1640                         {
1641                                 err = skb_copy_datagram_iovec(skb, offset,
1642                                                 msg->msg_iov, used);
1643                                 if (err) {
1644                                         /* Exception. Bailout! */
1645                                         if (!copied)
1646                                                 copied = -EFAULT;
1647                                         break;
1648                                 }
1649                         }
1650                 }
1651
1652                 *seq += used;
1653                 copied += used;
1654                 len -= used;
1655
1656                 tcp_rcv_space_adjust(sk);
1657
1658 skip_copy:
1659                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1660                         tp->urg_data = 0;
1661                         tcp_fast_path_check(sk);
1662                 }
1663                 if (used + offset < skb->len)
1664                         continue;
1665
1666                 if (tcp_hdr(skb)->fin)
1667                         goto found_fin_ok;
1668                 if (!(flags & MSG_PEEK)) {
1669                         sk_eat_skb(sk, skb, copied_early);
1670                         copied_early = 0;
1671                 }
1672                 continue;
1673
1674         found_fin_ok:
1675                 /* Process the FIN. */
1676                 ++*seq;
1677                 if (!(flags & MSG_PEEK)) {
1678                         sk_eat_skb(sk, skb, copied_early);
1679                         copied_early = 0;
1680                 }
1681                 break;
1682         } while (len > 0);
1683
1684         if (user_recv) {
1685                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1686                         int chunk;
1687
1688                         tp->ucopy.len = copied > 0 ? len : 0;
1689
1690                         tcp_prequeue_process(sk);
1691
1692                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1693                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1694                                 len -= chunk;
1695                                 copied += chunk;
1696                         }
1697                 }
1698
1699                 tp->ucopy.task = NULL;
1700                 tp->ucopy.len = 0;
1701         }
1702
1703 #ifdef CONFIG_NET_DMA
1704         if (tp->ucopy.dma_chan) {
1705                 dma_cookie_t done, used;
1706
1707                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1708
1709                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1710                                                  tp->ucopy.dma_cookie, &done,
1711                                                  &used) == DMA_IN_PROGRESS) {
1712                         /* do partial cleanup of sk_async_wait_queue */
1713                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1714                                (dma_async_is_complete(skb->dma_cookie, done,
1715                                                       used) == DMA_SUCCESS)) {
1716                                 __skb_dequeue(&sk->sk_async_wait_queue);
1717                                 kfree_skb(skb);
1718                         }
1719                 }
1720
1721                 /* Safe to free early-copied skbs now */
1722                 __skb_queue_purge(&sk->sk_async_wait_queue);
1723                 tp->ucopy.dma_chan = NULL;
1724         }
1725         if (tp->ucopy.pinned_list) {
1726                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1727                 tp->ucopy.pinned_list = NULL;
1728         }
1729 #endif
1730
1731         /* According to UNIX98, msg_name/msg_namelen are ignored
1732          * on connected socket. I was just happy when found this 8) --ANK
1733          */
1734
1735         /* Clean up data we have read: This will do ACK frames. */
1736         tcp_cleanup_rbuf(sk, copied);
1737
1738         TCP_CHECK_TIMER(sk);
1739         release_sock(sk);
1740         return copied;
1741
1742 out:
1743         TCP_CHECK_TIMER(sk);
1744         release_sock(sk);
1745         return err;
1746
1747 recv_urg:
1748         err = tcp_recv_urg(sk, msg, len, flags);
1749         goto out;
1750 }
1751
1752 void tcp_set_state(struct sock *sk, int state)
1753 {
1754         int oldstate = sk->sk_state;
1755
1756         switch (state) {
1757         case TCP_ESTABLISHED:
1758                 if (oldstate != TCP_ESTABLISHED)
1759                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1760                 break;
1761
1762         case TCP_CLOSE:
1763                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1764                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1765
1766                 sk->sk_prot->unhash(sk);
1767                 if (inet_csk(sk)->icsk_bind_hash &&
1768                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1769                         inet_put_port(sk);
1770                 /* fall through */
1771         default:
1772                 if (oldstate == TCP_ESTABLISHED)
1773                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1774         }
1775
1776         /* Change state AFTER socket is unhashed to avoid closed
1777          * socket sitting in hash tables.
1778          */
1779         sk->sk_state = state;
1780
1781 #ifdef STATE_TRACE
1782         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1783 #endif
1784 }
1785 EXPORT_SYMBOL_GPL(tcp_set_state);
1786
1787 /*
1788  *      State processing on a close. This implements the state shift for
1789  *      sending our FIN frame. Note that we only send a FIN for some
1790  *      states. A shutdown() may have already sent the FIN, or we may be
1791  *      closed.
1792  */
1793
1794 static const unsigned char new_state[16] = {
1795   /* current state:        new state:      action:      */
1796   /* (Invalid)          */ TCP_CLOSE,
1797   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1798   /* TCP_SYN_SENT       */ TCP_CLOSE,
1799   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1800   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1801   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1802   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1803   /* TCP_CLOSE          */ TCP_CLOSE,
1804   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1805   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1806   /* TCP_LISTEN         */ TCP_CLOSE,
1807   /* TCP_CLOSING        */ TCP_CLOSING,
1808 };
1809
1810 static int tcp_close_state(struct sock *sk)
1811 {
1812         int next = (int)new_state[sk->sk_state];
1813         int ns = next & TCP_STATE_MASK;
1814
1815         tcp_set_state(sk, ns);
1816
1817         return next & TCP_ACTION_FIN;
1818 }
1819
1820 /*
1821  *      Shutdown the sending side of a connection. Much like close except
1822  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1823  */
1824
1825 void tcp_shutdown(struct sock *sk, int how)
1826 {
1827         /*      We need to grab some memory, and put together a FIN,
1828          *      and then put it into the queue to be sent.
1829          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1830          */
1831         if (!(how & SEND_SHUTDOWN))
1832                 return;
1833
1834         /* If we've already sent a FIN, or it's a closed state, skip this. */
1835         if ((1 << sk->sk_state) &
1836             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1837              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1838                 /* Clear out any half completed packets.  FIN if needed. */
1839                 if (tcp_close_state(sk))
1840                         tcp_send_fin(sk);
1841         }
1842 }
1843
1844 void tcp_close(struct sock *sk, long timeout)
1845 {
1846         struct sk_buff *skb;
1847         int data_was_unread = 0;
1848         int state;
1849
1850         lock_sock(sk);
1851         sk->sk_shutdown = SHUTDOWN_MASK;
1852
1853         if (sk->sk_state == TCP_LISTEN) {
1854                 tcp_set_state(sk, TCP_CLOSE);
1855
1856                 /* Special case. */
1857                 inet_csk_listen_stop(sk);
1858
1859                 goto adjudge_to_death;
1860         }
1861
1862         /*  We need to flush the recv. buffs.  We do this only on the
1863          *  descriptor close, not protocol-sourced closes, because the
1864          *  reader process may not have drained the data yet!
1865          */
1866         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1867                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1868                           tcp_hdr(skb)->fin;
1869                 data_was_unread += len;
1870                 __kfree_skb(skb);
1871         }
1872
1873         sk_mem_reclaim(sk);
1874
1875         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1876          * data was lost. To witness the awful effects of the old behavior of
1877          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1878          * GET in an FTP client, suspend the process, wait for the client to
1879          * advertise a zero window, then kill -9 the FTP client, wheee...
1880          * Note: timeout is always zero in such a case.
1881          */
1882         if (data_was_unread) {
1883                 /* Unread data was tossed, zap the connection. */
1884                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1885                 tcp_set_state(sk, TCP_CLOSE);
1886                 tcp_send_active_reset(sk, sk->sk_allocation);
1887         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1888                 /* Check zero linger _after_ checking for unread data. */
1889                 sk->sk_prot->disconnect(sk, 0);
1890                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1891         } else if (tcp_close_state(sk)) {
1892                 /* We FIN if the application ate all the data before
1893                  * zapping the connection.
1894                  */
1895
1896                 /* RED-PEN. Formally speaking, we have broken TCP state
1897                  * machine. State transitions:
1898                  *
1899                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1900                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1901                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1902                  *
1903                  * are legal only when FIN has been sent (i.e. in window),
1904                  * rather than queued out of window. Purists blame.
1905                  *
1906                  * F.e. "RFC state" is ESTABLISHED,
1907                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1908                  *
1909                  * The visible declinations are that sometimes
1910                  * we enter time-wait state, when it is not required really
1911                  * (harmless), do not send active resets, when they are
1912                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1913                  * they look as CLOSING or LAST_ACK for Linux)
1914                  * Probably, I missed some more holelets.
1915                  *                                              --ANK
1916                  */
1917                 tcp_send_fin(sk);
1918         }
1919
1920         sk_stream_wait_close(sk, timeout);
1921
1922 adjudge_to_death:
1923         state = sk->sk_state;
1924         sock_hold(sk);
1925         sock_orphan(sk);
1926
1927         /* It is the last release_sock in its life. It will remove backlog. */
1928         release_sock(sk);
1929
1930
1931         /* Now socket is owned by kernel and we acquire BH lock
1932            to finish close. No need to check for user refs.
1933          */
1934         local_bh_disable();
1935         bh_lock_sock(sk);
1936         WARN_ON(sock_owned_by_user(sk));
1937
1938         percpu_counter_inc(sk->sk_prot->orphan_count);
1939
1940         /* Have we already been destroyed by a softirq or backlog? */
1941         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1942                 goto out;
1943
1944         /*      This is a (useful) BSD violating of the RFC. There is a
1945          *      problem with TCP as specified in that the other end could
1946          *      keep a socket open forever with no application left this end.
1947          *      We use a 3 minute timeout (about the same as BSD) then kill
1948          *      our end. If they send after that then tough - BUT: long enough
1949          *      that we won't make the old 4*rto = almost no time - whoops
1950          *      reset mistake.
1951          *
1952          *      Nope, it was not mistake. It is really desired behaviour
1953          *      f.e. on http servers, when such sockets are useless, but
1954          *      consume significant resources. Let's do it with special
1955          *      linger2 option.                                 --ANK
1956          */
1957
1958         if (sk->sk_state == TCP_FIN_WAIT2) {
1959                 struct tcp_sock *tp = tcp_sk(sk);
1960                 if (tp->linger2 < 0) {
1961                         tcp_set_state(sk, TCP_CLOSE);
1962                         tcp_send_active_reset(sk, GFP_ATOMIC);
1963                         NET_INC_STATS_BH(sock_net(sk),
1964                                         LINUX_MIB_TCPABORTONLINGER);
1965                 } else {
1966                         const int tmo = tcp_fin_time(sk);
1967
1968                         if (tmo > TCP_TIMEWAIT_LEN) {
1969                                 inet_csk_reset_keepalive_timer(sk,
1970                                                 tmo - TCP_TIMEWAIT_LEN);
1971                         } else {
1972                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1973                                 goto out;
1974                         }
1975                 }
1976         }
1977         if (sk->sk_state != TCP_CLOSE) {
1978                 int orphan_count = percpu_counter_read_positive(
1979                                                 sk->sk_prot->orphan_count);
1980
1981                 sk_mem_reclaim(sk);
1982                 if (tcp_too_many_orphans(sk, orphan_count)) {
1983                         if (net_ratelimit())
1984                                 printk(KERN_INFO "TCP: too many of orphaned "
1985                                        "sockets\n");
1986                         tcp_set_state(sk, TCP_CLOSE);
1987                         tcp_send_active_reset(sk, GFP_ATOMIC);
1988                         NET_INC_STATS_BH(sock_net(sk),
1989                                         LINUX_MIB_TCPABORTONMEMORY);
1990                 }
1991         }
1992
1993         if (sk->sk_state == TCP_CLOSE)
1994                 inet_csk_destroy_sock(sk);
1995         /* Otherwise, socket is reprieved until protocol close. */
1996
1997 out:
1998         bh_unlock_sock(sk);
1999         local_bh_enable();
2000         sock_put(sk);
2001 }
2002
2003 /* These states need RST on ABORT according to RFC793 */
2004
2005 static inline int tcp_need_reset(int state)
2006 {
2007         return (1 << state) &
2008                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2009                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2010 }
2011
2012 int tcp_disconnect(struct sock *sk, int flags)
2013 {
2014         struct inet_sock *inet = inet_sk(sk);
2015         struct inet_connection_sock *icsk = inet_csk(sk);
2016         struct tcp_sock *tp = tcp_sk(sk);
2017         int err = 0;
2018         int old_state = sk->sk_state;
2019
2020         if (old_state != TCP_CLOSE)
2021                 tcp_set_state(sk, TCP_CLOSE);
2022
2023         /* ABORT function of RFC793 */
2024         if (old_state == TCP_LISTEN) {
2025                 inet_csk_listen_stop(sk);
2026         } else if (tcp_need_reset(old_state) ||
2027                    (tp->snd_nxt != tp->write_seq &&
2028                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2029                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2030                  * states
2031                  */
2032                 tcp_send_active_reset(sk, gfp_any());
2033                 sk->sk_err = ECONNRESET;
2034         } else if (old_state == TCP_SYN_SENT)
2035                 sk->sk_err = ECONNRESET;
2036
2037         tcp_clear_xmit_timers(sk);
2038         __skb_queue_purge(&sk->sk_receive_queue);
2039         tcp_write_queue_purge(sk);
2040         __skb_queue_purge(&tp->out_of_order_queue);
2041 #ifdef CONFIG_NET_DMA
2042         __skb_queue_purge(&sk->sk_async_wait_queue);
2043 #endif
2044
2045         inet->inet_dport = 0;
2046
2047         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2048                 inet_reset_saddr(sk);
2049
2050         sk->sk_shutdown = 0;
2051         sock_reset_flag(sk, SOCK_DONE);
2052         tp->srtt = 0;
2053         if ((tp->write_seq += tp->max_window + 2) == 0)
2054                 tp->write_seq = 1;
2055         icsk->icsk_backoff = 0;
2056         tp->snd_cwnd = 2;
2057         icsk->icsk_probes_out = 0;
2058         tp->packets_out = 0;
2059         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2060         tp->snd_cwnd_cnt = 0;
2061         tp->bytes_acked = 0;
2062         tcp_set_ca_state(sk, TCP_CA_Open);
2063         tcp_clear_retrans(tp);
2064         inet_csk_delack_init(sk);
2065         tcp_init_send_head(sk);
2066         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2067         __sk_dst_reset(sk);
2068
2069         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2070
2071         sk->sk_error_report(sk);
2072         return err;
2073 }
2074
2075 /*
2076  *      Socket option code for TCP.
2077  */
2078 static int do_tcp_setsockopt(struct sock *sk, int level,
2079                 int optname, char __user *optval, unsigned int optlen)
2080 {
2081         struct tcp_sock *tp = tcp_sk(sk);
2082         struct inet_connection_sock *icsk = inet_csk(sk);
2083         int val;
2084         int err = 0;
2085
2086         /* This is a string value all the others are int's */
2087         if (optname == TCP_CONGESTION) {
2088                 char name[TCP_CA_NAME_MAX];
2089
2090                 if (optlen < 1)
2091                         return -EINVAL;
2092
2093                 val = strncpy_from_user(name, optval,
2094                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2095                 if (val < 0)
2096                         return -EFAULT;
2097                 name[val] = 0;
2098
2099                 lock_sock(sk);
2100                 err = tcp_set_congestion_control(sk, name);
2101                 release_sock(sk);
2102                 return err;
2103         }
2104
2105         if (optlen < sizeof(int))
2106                 return -EINVAL;
2107
2108         if (get_user(val, (int __user *)optval))
2109                 return -EFAULT;
2110
2111         lock_sock(sk);
2112
2113         switch (optname) {
2114         case TCP_MAXSEG:
2115                 /* Values greater than interface MTU won't take effect. However
2116                  * at the point when this call is done we typically don't yet
2117                  * know which interface is going to be used */
2118                 if (val < 8 || val > MAX_TCP_WINDOW) {
2119                         err = -EINVAL;
2120                         break;
2121                 }
2122                 tp->rx_opt.user_mss = val;
2123                 break;
2124
2125         case TCP_NODELAY:
2126                 if (val) {
2127                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2128                          * this option on corked socket is remembered, but
2129                          * it is not activated until cork is cleared.
2130                          *
2131                          * However, when TCP_NODELAY is set we make
2132                          * an explicit push, which overrides even TCP_CORK
2133                          * for currently queued segments.
2134                          */
2135                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2136                         tcp_push_pending_frames(sk);
2137                 } else {
2138                         tp->nonagle &= ~TCP_NAGLE_OFF;
2139                 }
2140                 break;
2141
2142         case TCP_CORK:
2143                 /* When set indicates to always queue non-full frames.
2144                  * Later the user clears this option and we transmit
2145                  * any pending partial frames in the queue.  This is
2146                  * meant to be used alongside sendfile() to get properly
2147                  * filled frames when the user (for example) must write
2148                  * out headers with a write() call first and then use
2149                  * sendfile to send out the data parts.
2150                  *
2151                  * TCP_CORK can be set together with TCP_NODELAY and it is
2152                  * stronger than TCP_NODELAY.
2153                  */
2154                 if (val) {
2155                         tp->nonagle |= TCP_NAGLE_CORK;
2156                 } else {
2157                         tp->nonagle &= ~TCP_NAGLE_CORK;
2158                         if (tp->nonagle&TCP_NAGLE_OFF)
2159                                 tp->nonagle |= TCP_NAGLE_PUSH;
2160                         tcp_push_pending_frames(sk);
2161                 }
2162                 break;
2163
2164         case TCP_KEEPIDLE:
2165                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2166                         err = -EINVAL;
2167                 else {
2168                         tp->keepalive_time = val * HZ;
2169                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2170                             !((1 << sk->sk_state) &
2171                               (TCPF_CLOSE | TCPF_LISTEN))) {
2172                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2173                                 if (tp->keepalive_time > elapsed)
2174                                         elapsed = tp->keepalive_time - elapsed;
2175                                 else
2176                                         elapsed = 0;
2177                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2178                         }
2179                 }
2180                 break;
2181         case TCP_KEEPINTVL:
2182                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2183                         err = -EINVAL;
2184                 else
2185                         tp->keepalive_intvl = val * HZ;
2186                 break;
2187         case TCP_KEEPCNT:
2188                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2189                         err = -EINVAL;
2190                 else
2191                         tp->keepalive_probes = val;
2192                 break;
2193         case TCP_SYNCNT:
2194                 if (val < 1 || val > MAX_TCP_SYNCNT)
2195                         err = -EINVAL;
2196                 else
2197                         icsk->icsk_syn_retries = val;
2198                 break;
2199
2200         case TCP_LINGER2:
2201                 if (val < 0)
2202                         tp->linger2 = -1;
2203                 else if (val > sysctl_tcp_fin_timeout / HZ)
2204                         tp->linger2 = 0;
2205                 else
2206                         tp->linger2 = val * HZ;
2207                 break;
2208
2209         case TCP_DEFER_ACCEPT:
2210                 /* Translate value in seconds to number of retransmits */
2211                 icsk->icsk_accept_queue.rskq_defer_accept =
2212                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2213                                         TCP_RTO_MAX / HZ);
2214                 break;
2215
2216         case TCP_WINDOW_CLAMP:
2217                 if (!val) {
2218                         if (sk->sk_state != TCP_CLOSE) {
2219                                 err = -EINVAL;
2220                                 break;
2221                         }
2222                         tp->window_clamp = 0;
2223                 } else
2224                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2225                                                 SOCK_MIN_RCVBUF / 2 : val;
2226                 break;
2227
2228         case TCP_QUICKACK:
2229                 if (!val) {
2230                         icsk->icsk_ack.pingpong = 1;
2231                 } else {
2232                         icsk->icsk_ack.pingpong = 0;
2233                         if ((1 << sk->sk_state) &
2234                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2235                             inet_csk_ack_scheduled(sk)) {
2236                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2237                                 tcp_cleanup_rbuf(sk, 1);
2238                                 if (!(val & 1))
2239                                         icsk->icsk_ack.pingpong = 1;
2240                         }
2241                 }
2242                 break;
2243
2244 #ifdef CONFIG_TCP_MD5SIG
2245         case TCP_MD5SIG:
2246                 /* Read the IP->Key mappings from userspace */
2247                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2248                 break;
2249 #endif
2250
2251         default:
2252                 err = -ENOPROTOOPT;
2253                 break;
2254         }
2255
2256         release_sock(sk);
2257         return err;
2258 }
2259
2260 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2261                    unsigned int optlen)
2262 {
2263         struct inet_connection_sock *icsk = inet_csk(sk);
2264
2265         if (level != SOL_TCP)
2266                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2267                                                      optval, optlen);
2268         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2269 }
2270
2271 #ifdef CONFIG_COMPAT
2272 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2273                           char __user *optval, unsigned int optlen)
2274 {
2275         if (level != SOL_TCP)
2276                 return inet_csk_compat_setsockopt(sk, level, optname,
2277                                                   optval, optlen);
2278         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2279 }
2280
2281 EXPORT_SYMBOL(compat_tcp_setsockopt);
2282 #endif
2283
2284 /* Return information about state of tcp endpoint in API format. */
2285 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2286 {
2287         struct tcp_sock *tp = tcp_sk(sk);
2288         const struct inet_connection_sock *icsk = inet_csk(sk);
2289         u32 now = tcp_time_stamp;
2290
2291         memset(info, 0, sizeof(*info));
2292
2293         info->tcpi_state = sk->sk_state;
2294         info->tcpi_ca_state = icsk->icsk_ca_state;
2295         info->tcpi_retransmits = icsk->icsk_retransmits;
2296         info->tcpi_probes = icsk->icsk_probes_out;
2297         info->tcpi_backoff = icsk->icsk_backoff;
2298
2299         if (tp->rx_opt.tstamp_ok)
2300                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2301         if (tcp_is_sack(tp))
2302                 info->tcpi_options |= TCPI_OPT_SACK;
2303         if (tp->rx_opt.wscale_ok) {
2304                 info->tcpi_options |= TCPI_OPT_WSCALE;
2305                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2306                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2307         }
2308
2309         if (tp->ecn_flags&TCP_ECN_OK)
2310                 info->tcpi_options |= TCPI_OPT_ECN;
2311
2312         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2313         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2314         info->tcpi_snd_mss = tp->mss_cache;
2315         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2316
2317         if (sk->sk_state == TCP_LISTEN) {
2318                 info->tcpi_unacked = sk->sk_ack_backlog;
2319                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2320         } else {
2321                 info->tcpi_unacked = tp->packets_out;
2322                 info->tcpi_sacked = tp->sacked_out;
2323         }
2324         info->tcpi_lost = tp->lost_out;
2325         info->tcpi_retrans = tp->retrans_out;
2326         info->tcpi_fackets = tp->fackets_out;
2327
2328         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2329         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2330         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2331
2332         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2333         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2334         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2335         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2336         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2337         info->tcpi_snd_cwnd = tp->snd_cwnd;
2338         info->tcpi_advmss = tp->advmss;
2339         info->tcpi_reordering = tp->reordering;
2340
2341         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2342         info->tcpi_rcv_space = tp->rcvq_space.space;
2343
2344         info->tcpi_total_retrans = tp->total_retrans;
2345 }
2346
2347 EXPORT_SYMBOL_GPL(tcp_get_info);
2348
2349 static int do_tcp_getsockopt(struct sock *sk, int level,
2350                 int optname, char __user *optval, int __user *optlen)
2351 {
2352         struct inet_connection_sock *icsk = inet_csk(sk);
2353         struct tcp_sock *tp = tcp_sk(sk);
2354         int val, len;
2355
2356         if (get_user(len, optlen))
2357                 return -EFAULT;
2358
2359         len = min_t(unsigned int, len, sizeof(int));
2360
2361         if (len < 0)
2362                 return -EINVAL;
2363
2364         switch (optname) {
2365         case TCP_MAXSEG:
2366                 val = tp->mss_cache;
2367                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2368                         val = tp->rx_opt.user_mss;
2369                 break;
2370         case TCP_NODELAY:
2371                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2372                 break;
2373         case TCP_CORK:
2374                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2375                 break;
2376         case TCP_KEEPIDLE:
2377                 val = keepalive_time_when(tp) / HZ;
2378                 break;
2379         case TCP_KEEPINTVL:
2380                 val = keepalive_intvl_when(tp) / HZ;
2381                 break;
2382         case TCP_KEEPCNT:
2383                 val = keepalive_probes(tp);
2384                 break;
2385         case TCP_SYNCNT:
2386                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2387                 break;
2388         case TCP_LINGER2:
2389                 val = tp->linger2;
2390                 if (val >= 0)
2391                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2392                 break;
2393         case TCP_DEFER_ACCEPT:
2394                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2395                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2396                 break;
2397         case TCP_WINDOW_CLAMP:
2398                 val = tp->window_clamp;
2399                 break;
2400         case TCP_INFO: {
2401                 struct tcp_info info;
2402
2403                 if (get_user(len, optlen))
2404                         return -EFAULT;
2405
2406                 tcp_get_info(sk, &info);
2407
2408                 len = min_t(unsigned int, len, sizeof(info));
2409                 if (put_user(len, optlen))
2410                         return -EFAULT;
2411                 if (copy_to_user(optval, &info, len))
2412                         return -EFAULT;
2413                 return 0;
2414         }
2415         case TCP_QUICKACK:
2416                 val = !icsk->icsk_ack.pingpong;
2417                 break;
2418
2419         case TCP_CONGESTION:
2420                 if (get_user(len, optlen))
2421                         return -EFAULT;
2422                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2423                 if (put_user(len, optlen))
2424                         return -EFAULT;
2425                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2426                         return -EFAULT;
2427                 return 0;
2428         default:
2429                 return -ENOPROTOOPT;
2430         }
2431
2432         if (put_user(len, optlen))
2433                 return -EFAULT;
2434         if (copy_to_user(optval, &val, len))
2435                 return -EFAULT;
2436         return 0;
2437 }
2438
2439 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2440                    int __user *optlen)
2441 {
2442         struct inet_connection_sock *icsk = inet_csk(sk);
2443
2444         if (level != SOL_TCP)
2445                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2446                                                      optval, optlen);
2447         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2448 }
2449
2450 #ifdef CONFIG_COMPAT
2451 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2452                           char __user *optval, int __user *optlen)
2453 {
2454         if (level != SOL_TCP)
2455                 return inet_csk_compat_getsockopt(sk, level, optname,
2456                                                   optval, optlen);
2457         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2458 }
2459
2460 EXPORT_SYMBOL(compat_tcp_getsockopt);
2461 #endif
2462
2463 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2464 {
2465         struct sk_buff *segs = ERR_PTR(-EINVAL);
2466         struct tcphdr *th;
2467         unsigned thlen;
2468         unsigned int seq;
2469         __be32 delta;
2470         unsigned int oldlen;
2471         unsigned int mss;
2472
2473         if (!pskb_may_pull(skb, sizeof(*th)))
2474                 goto out;
2475
2476         th = tcp_hdr(skb);
2477         thlen = th->doff * 4;
2478         if (thlen < sizeof(*th))
2479                 goto out;
2480
2481         if (!pskb_may_pull(skb, thlen))
2482                 goto out;
2483
2484         oldlen = (u16)~skb->len;
2485         __skb_pull(skb, thlen);
2486
2487         mss = skb_shinfo(skb)->gso_size;
2488         if (unlikely(skb->len <= mss))
2489                 goto out;
2490
2491         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2492                 /* Packet is from an untrusted source, reset gso_segs. */
2493                 int type = skb_shinfo(skb)->gso_type;
2494
2495                 if (unlikely(type &
2496                              ~(SKB_GSO_TCPV4 |
2497                                SKB_GSO_DODGY |
2498                                SKB_GSO_TCP_ECN |
2499                                SKB_GSO_TCPV6 |
2500                                0) ||
2501                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2502                         goto out;
2503
2504                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2505
2506                 segs = NULL;
2507                 goto out;
2508         }
2509
2510         segs = skb_segment(skb, features);
2511         if (IS_ERR(segs))
2512                 goto out;
2513
2514         delta = htonl(oldlen + (thlen + mss));
2515
2516         skb = segs;
2517         th = tcp_hdr(skb);
2518         seq = ntohl(th->seq);
2519
2520         do {
2521                 th->fin = th->psh = 0;
2522
2523                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2524                                        (__force u32)delta));
2525                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2526                         th->check =
2527                              csum_fold(csum_partial(skb_transport_header(skb),
2528                                                     thlen, skb->csum));
2529
2530                 seq += mss;
2531                 skb = skb->next;
2532                 th = tcp_hdr(skb);
2533
2534                 th->seq = htonl(seq);
2535                 th->cwr = 0;
2536         } while (skb->next);
2537
2538         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2539                       skb->data_len);
2540         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2541                                 (__force u32)delta));
2542         if (skb->ip_summed != CHECKSUM_PARTIAL)
2543                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2544                                                    thlen, skb->csum));
2545
2546 out:
2547         return segs;
2548 }
2549 EXPORT_SYMBOL(tcp_tso_segment);
2550
2551 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2552 {
2553         struct sk_buff **pp = NULL;
2554         struct sk_buff *p;
2555         struct tcphdr *th;
2556         struct tcphdr *th2;
2557         unsigned int len;
2558         unsigned int thlen;
2559         unsigned int flags;
2560         unsigned int mss = 1;
2561         unsigned int hlen;
2562         unsigned int off;
2563         int flush = 1;
2564         int i;
2565
2566         off = skb_gro_offset(skb);
2567         hlen = off + sizeof(*th);
2568         th = skb_gro_header_fast(skb, off);
2569         if (skb_gro_header_hard(skb, hlen)) {
2570                 th = skb_gro_header_slow(skb, hlen, off);
2571                 if (unlikely(!th))
2572                         goto out;
2573         }
2574
2575         thlen = th->doff * 4;
2576         if (thlen < sizeof(*th))
2577                 goto out;
2578
2579         hlen = off + thlen;
2580         if (skb_gro_header_hard(skb, hlen)) {
2581                 th = skb_gro_header_slow(skb, hlen, off);
2582                 if (unlikely(!th))
2583                         goto out;
2584         }
2585
2586         skb_gro_pull(skb, thlen);
2587
2588         len = skb_gro_len(skb);
2589         flags = tcp_flag_word(th);
2590
2591         for (; (p = *head); head = &p->next) {
2592                 if (!NAPI_GRO_CB(p)->same_flow)
2593                         continue;
2594
2595                 th2 = tcp_hdr(p);
2596
2597                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2598                         NAPI_GRO_CB(p)->same_flow = 0;
2599                         continue;
2600                 }
2601
2602                 goto found;
2603         }
2604
2605         goto out_check_final;
2606
2607 found:
2608         flush = NAPI_GRO_CB(p)->flush;
2609         flush |= flags & TCP_FLAG_CWR;
2610         flush |= (flags ^ tcp_flag_word(th2)) &
2611                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH);
2612         flush |= th->ack_seq ^ th2->ack_seq;
2613         for (i = sizeof(*th); i < thlen; i += 4)
2614                 flush |= *(u32 *)((u8 *)th + i) ^
2615                          *(u32 *)((u8 *)th2 + i);
2616
2617         mss = skb_shinfo(p)->gso_size;
2618
2619         flush |= (len - 1) >= mss;
2620         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2621
2622         if (flush || skb_gro_receive(head, skb)) {
2623                 mss = 1;
2624                 goto out_check_final;
2625         }
2626
2627         p = *head;
2628         th2 = tcp_hdr(p);
2629         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2630
2631 out_check_final:
2632         flush = len < mss;
2633         flush |= flags & (TCP_FLAG_URG | TCP_FLAG_PSH | TCP_FLAG_RST |
2634                           TCP_FLAG_SYN | TCP_FLAG_FIN);
2635
2636         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2637                 pp = head;
2638
2639 out:
2640         NAPI_GRO_CB(skb)->flush |= flush;
2641
2642         return pp;
2643 }
2644 EXPORT_SYMBOL(tcp_gro_receive);
2645
2646 int tcp_gro_complete(struct sk_buff *skb)
2647 {
2648         struct tcphdr *th = tcp_hdr(skb);
2649
2650         skb->csum_start = skb_transport_header(skb) - skb->head;
2651         skb->csum_offset = offsetof(struct tcphdr, check);
2652         skb->ip_summed = CHECKSUM_PARTIAL;
2653
2654         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2655
2656         if (th->cwr)
2657                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2658
2659         return 0;
2660 }
2661 EXPORT_SYMBOL(tcp_gro_complete);
2662
2663 #ifdef CONFIG_TCP_MD5SIG
2664 static unsigned long tcp_md5sig_users;
2665 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2666 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2667
2668 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2669 {
2670         int cpu;
2671         for_each_possible_cpu(cpu) {
2672                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2673                 if (p) {
2674                         if (p->md5_desc.tfm)
2675                                 crypto_free_hash(p->md5_desc.tfm);
2676                         kfree(p);
2677                         p = NULL;
2678                 }
2679         }
2680         free_percpu(pool);
2681 }
2682
2683 void tcp_free_md5sig_pool(void)
2684 {
2685         struct tcp_md5sig_pool **pool = NULL;
2686
2687         spin_lock_bh(&tcp_md5sig_pool_lock);
2688         if (--tcp_md5sig_users == 0) {
2689                 pool = tcp_md5sig_pool;
2690                 tcp_md5sig_pool = NULL;
2691         }
2692         spin_unlock_bh(&tcp_md5sig_pool_lock);
2693         if (pool)
2694                 __tcp_free_md5sig_pool(pool);
2695 }
2696
2697 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2698
2699 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(struct sock *sk)
2700 {
2701         int cpu;
2702         struct tcp_md5sig_pool **pool;
2703
2704         pool = alloc_percpu(struct tcp_md5sig_pool *);
2705         if (!pool)
2706                 return NULL;
2707
2708         for_each_possible_cpu(cpu) {
2709                 struct tcp_md5sig_pool *p;
2710                 struct crypto_hash *hash;
2711
2712                 p = kzalloc(sizeof(*p), sk->sk_allocation);
2713                 if (!p)
2714                         goto out_free;
2715                 *per_cpu_ptr(pool, cpu) = p;
2716
2717                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2718                 if (!hash || IS_ERR(hash))
2719                         goto out_free;
2720
2721                 p->md5_desc.tfm = hash;
2722         }
2723         return pool;
2724 out_free:
2725         __tcp_free_md5sig_pool(pool);
2726         return NULL;
2727 }
2728
2729 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(struct sock *sk)
2730 {
2731         struct tcp_md5sig_pool **pool;
2732         int alloc = 0;
2733
2734 retry:
2735         spin_lock_bh(&tcp_md5sig_pool_lock);
2736         pool = tcp_md5sig_pool;
2737         if (tcp_md5sig_users++ == 0) {
2738                 alloc = 1;
2739                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2740         } else if (!pool) {
2741                 tcp_md5sig_users--;
2742                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2743                 cpu_relax();
2744                 goto retry;
2745         } else
2746                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2747
2748         if (alloc) {
2749                 /* we cannot hold spinlock here because this may sleep. */
2750                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool(sk);
2751                 spin_lock_bh(&tcp_md5sig_pool_lock);
2752                 if (!p) {
2753                         tcp_md5sig_users--;
2754                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2755                         return NULL;
2756                 }
2757                 pool = tcp_md5sig_pool;
2758                 if (pool) {
2759                         /* oops, it has already been assigned. */
2760                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2761                         __tcp_free_md5sig_pool(p);
2762                 } else {
2763                         tcp_md5sig_pool = pool = p;
2764                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2765                 }
2766         }
2767         return pool;
2768 }
2769
2770 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2771
2772 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2773 {
2774         struct tcp_md5sig_pool **p;
2775         spin_lock_bh(&tcp_md5sig_pool_lock);
2776         p = tcp_md5sig_pool;
2777         if (p)
2778                 tcp_md5sig_users++;
2779         spin_unlock_bh(&tcp_md5sig_pool_lock);
2780         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2781 }
2782
2783 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2784
2785 void __tcp_put_md5sig_pool(void)
2786 {
2787         tcp_free_md5sig_pool();
2788 }
2789
2790 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2791
2792 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2793                         struct tcphdr *th)
2794 {
2795         struct scatterlist sg;
2796         int err;
2797
2798         __sum16 old_checksum = th->check;
2799         th->check = 0;
2800         /* options aren't included in the hash */
2801         sg_init_one(&sg, th, sizeof(struct tcphdr));
2802         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2803         th->check = old_checksum;
2804         return err;
2805 }
2806
2807 EXPORT_SYMBOL(tcp_md5_hash_header);
2808
2809 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2810                           struct sk_buff *skb, unsigned header_len)
2811 {
2812         struct scatterlist sg;
2813         const struct tcphdr *tp = tcp_hdr(skb);
2814         struct hash_desc *desc = &hp->md5_desc;
2815         unsigned i;
2816         const unsigned head_data_len = skb_headlen(skb) > header_len ?
2817                                        skb_headlen(skb) - header_len : 0;
2818         const struct skb_shared_info *shi = skb_shinfo(skb);
2819
2820         sg_init_table(&sg, 1);
2821
2822         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2823         if (crypto_hash_update(desc, &sg, head_data_len))
2824                 return 1;
2825
2826         for (i = 0; i < shi->nr_frags; ++i) {
2827                 const struct skb_frag_struct *f = &shi->frags[i];
2828                 sg_set_page(&sg, f->page, f->size, f->page_offset);
2829                 if (crypto_hash_update(desc, &sg, f->size))
2830                         return 1;
2831         }
2832
2833         return 0;
2834 }
2835
2836 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2837
2838 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
2839 {
2840         struct scatterlist sg;
2841
2842         sg_init_one(&sg, key->key, key->keylen);
2843         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2844 }
2845
2846 EXPORT_SYMBOL(tcp_md5_hash_key);
2847
2848 #endif
2849
2850 void tcp_done(struct sock *sk)
2851 {
2852         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2853                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2854
2855         tcp_set_state(sk, TCP_CLOSE);
2856         tcp_clear_xmit_timers(sk);
2857
2858         sk->sk_shutdown = SHUTDOWN_MASK;
2859
2860         if (!sock_flag(sk, SOCK_DEAD))
2861                 sk->sk_state_change(sk);
2862         else
2863                 inet_csk_destroy_sock(sk);
2864 }
2865 EXPORT_SYMBOL_GPL(tcp_done);
2866
2867 extern struct tcp_congestion_ops tcp_reno;
2868
2869 static __initdata unsigned long thash_entries;
2870 static int __init set_thash_entries(char *str)
2871 {
2872         if (!str)
2873                 return 0;
2874         thash_entries = simple_strtoul(str, &str, 0);
2875         return 1;
2876 }
2877 __setup("thash_entries=", set_thash_entries);
2878
2879 void __init tcp_init(void)
2880 {
2881         struct sk_buff *skb = NULL;
2882         unsigned long nr_pages, limit;
2883         int order, i, max_share;
2884
2885         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2886
2887         percpu_counter_init(&tcp_sockets_allocated, 0);
2888         percpu_counter_init(&tcp_orphan_count, 0);
2889         tcp_hashinfo.bind_bucket_cachep =
2890                 kmem_cache_create("tcp_bind_bucket",
2891                                   sizeof(struct inet_bind_bucket), 0,
2892                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2893
2894         /* Size and allocate the main established and bind bucket
2895          * hash tables.
2896          *
2897          * The methodology is similar to that of the buffer cache.
2898          */
2899         tcp_hashinfo.ehash =
2900                 alloc_large_system_hash("TCP established",
2901                                         sizeof(struct inet_ehash_bucket),
2902                                         thash_entries,
2903                                         (totalram_pages >= 128 * 1024) ?
2904                                         13 : 15,
2905                                         0,
2906                                         NULL,
2907                                         &tcp_hashinfo.ehash_mask,
2908                                         thash_entries ? 0 : 512 * 1024);
2909         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
2910                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
2911                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
2912         }
2913         if (inet_ehash_locks_alloc(&tcp_hashinfo))
2914                 panic("TCP: failed to alloc ehash_locks");
2915         tcp_hashinfo.bhash =
2916                 alloc_large_system_hash("TCP bind",
2917                                         sizeof(struct inet_bind_hashbucket),
2918                                         tcp_hashinfo.ehash_mask + 1,
2919                                         (totalram_pages >= 128 * 1024) ?
2920                                         13 : 15,
2921                                         0,
2922                                         &tcp_hashinfo.bhash_size,
2923                                         NULL,
2924                                         64 * 1024);
2925         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2926         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2927                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2928                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2929         }
2930
2931         /* Try to be a bit smarter and adjust defaults depending
2932          * on available memory.
2933          */
2934         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2935                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2936                         order++)
2937                 ;
2938         if (order >= 4) {
2939                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2940                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2941                 sysctl_max_syn_backlog = 1024;
2942         } else if (order < 3) {
2943                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2944                 sysctl_tcp_max_orphans >>= (3 - order);
2945                 sysctl_max_syn_backlog = 128;
2946         }
2947
2948         /* Set the pressure threshold to be a fraction of global memory that
2949          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2950          * memory, with a floor of 128 pages.
2951          */
2952         nr_pages = totalram_pages - totalhigh_pages;
2953         limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2954         limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2955         limit = max(limit, 128UL);
2956         sysctl_tcp_mem[0] = limit / 4 * 3;
2957         sysctl_tcp_mem[1] = limit;
2958         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2959
2960         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2961         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2962         max_share = min(4UL*1024*1024, limit);
2963
2964         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2965         sysctl_tcp_wmem[1] = 16*1024;
2966         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2967
2968         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2969         sysctl_tcp_rmem[1] = 87380;
2970         sysctl_tcp_rmem[2] = max(87380, max_share);
2971
2972         printk(KERN_INFO "TCP: Hash tables configured "
2973                "(established %u bind %u)\n",
2974                tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
2975
2976         tcp_register_congestion_control(&tcp_reno);
2977 }
2978
2979 EXPORT_SYMBOL(tcp_close);
2980 EXPORT_SYMBOL(tcp_disconnect);
2981 EXPORT_SYMBOL(tcp_getsockopt);
2982 EXPORT_SYMBOL(tcp_ioctl);
2983 EXPORT_SYMBOL(tcp_poll);
2984 EXPORT_SYMBOL(tcp_read_sock);
2985 EXPORT_SYMBOL(tcp_recvmsg);
2986 EXPORT_SYMBOL(tcp_sendmsg);
2987 EXPORT_SYMBOL(tcp_splice_read);
2988 EXPORT_SYMBOL(tcp_sendpage);
2989 EXPORT_SYMBOL(tcp_setsockopt);
2990 EXPORT_SYMBOL(tcp_shutdown);