net: adding memory barrier to the poll and receive callbacks
[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 /*
330  *      Wait for a TCP event.
331  *
332  *      Note that we don't need to lock the socket, as the upper poll layers
333  *      take care of normal races (between the test and the event) and we don't
334  *      go look at any of the socket buffers directly.
335  */
336 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
337 {
338         unsigned int mask;
339         struct sock *sk = sock->sk;
340         struct tcp_sock *tp = tcp_sk(sk);
341
342         sock_poll_wait(file, sk->sk_sleep, wait);
343         if (sk->sk_state == TCP_LISTEN)
344                 return inet_csk_listen_poll(sk);
345
346         /* Socket is not locked. We are protected from async events
347          * by poll logic and correct handling of state changes
348          * made by other threads is impossible in any case.
349          */
350
351         mask = 0;
352         if (sk->sk_err)
353                 mask = POLLERR;
354
355         /*
356          * POLLHUP is certainly not done right. But poll() doesn't
357          * have a notion of HUP in just one direction, and for a
358          * socket the read side is more interesting.
359          *
360          * Some poll() documentation says that POLLHUP is incompatible
361          * with the POLLOUT/POLLWR flags, so somebody should check this
362          * all. But careful, it tends to be safer to return too many
363          * bits than too few, and you can easily break real applications
364          * if you don't tell them that something has hung up!
365          *
366          * Check-me.
367          *
368          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
369          * our fs/select.c). It means that after we received EOF,
370          * poll always returns immediately, making impossible poll() on write()
371          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
372          * if and only if shutdown has been made in both directions.
373          * Actually, it is interesting to look how Solaris and DUX
374          * solve this dilemma. I would prefer, if POLLHUP were maskable,
375          * then we could set it on SND_SHUTDOWN. BTW examples given
376          * in Stevens' books assume exactly this behaviour, it explains
377          * why POLLHUP is incompatible with POLLOUT.    --ANK
378          *
379          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
380          * blocking on fresh not-connected or disconnected socket. --ANK
381          */
382         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
383                 mask |= POLLHUP;
384         if (sk->sk_shutdown & RCV_SHUTDOWN)
385                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
386
387         /* Connected? */
388         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
389                 int target = sock_rcvlowat(sk, 0, INT_MAX);
390
391                 if (tp->urg_seq == tp->copied_seq &&
392                     !sock_flag(sk, SOCK_URGINLINE) &&
393                     tp->urg_data)
394                         target--;
395
396                 /* Potential race condition. If read of tp below will
397                  * escape above sk->sk_state, we can be illegally awaken
398                  * in SYN_* states. */
399                 if (tp->rcv_nxt - tp->copied_seq >= target)
400                         mask |= POLLIN | POLLRDNORM;
401
402                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
403                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
404                                 mask |= POLLOUT | POLLWRNORM;
405                         } else {  /* send SIGIO later */
406                                 set_bit(SOCK_ASYNC_NOSPACE,
407                                         &sk->sk_socket->flags);
408                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
409
410                                 /* Race breaker. If space is freed after
411                                  * wspace test but before the flags are set,
412                                  * IO signal will be lost.
413                                  */
414                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
415                                         mask |= POLLOUT | POLLWRNORM;
416                         }
417                 }
418
419                 if (tp->urg_data & TCP_URG_VALID)
420                         mask |= POLLPRI;
421         }
422         return mask;
423 }
424
425 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
426 {
427         struct tcp_sock *tp = tcp_sk(sk);
428         int answ;
429
430         switch (cmd) {
431         case SIOCINQ:
432                 if (sk->sk_state == TCP_LISTEN)
433                         return -EINVAL;
434
435                 lock_sock(sk);
436                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
437                         answ = 0;
438                 else if (sock_flag(sk, SOCK_URGINLINE) ||
439                          !tp->urg_data ||
440                          before(tp->urg_seq, tp->copied_seq) ||
441                          !before(tp->urg_seq, tp->rcv_nxt)) {
442                         struct sk_buff *skb;
443
444                         answ = tp->rcv_nxt - tp->copied_seq;
445
446                         /* Subtract 1, if FIN is in queue. */
447                         skb = skb_peek_tail(&sk->sk_receive_queue);
448                         if (answ && skb)
449                                 answ -= tcp_hdr(skb)->fin;
450                 } else
451                         answ = tp->urg_seq - tp->copied_seq;
452                 release_sock(sk);
453                 break;
454         case SIOCATMARK:
455                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
456                 break;
457         case SIOCOUTQ:
458                 if (sk->sk_state == TCP_LISTEN)
459                         return -EINVAL;
460
461                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
462                         answ = 0;
463                 else
464                         answ = tp->write_seq - tp->snd_una;
465                 break;
466         default:
467                 return -ENOIOCTLCMD;
468         }
469
470         return put_user(answ, (int __user *)arg);
471 }
472
473 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
474 {
475         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
476         tp->pushed_seq = tp->write_seq;
477 }
478
479 static inline int forced_push(struct tcp_sock *tp)
480 {
481         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
482 }
483
484 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
485 {
486         struct tcp_sock *tp = tcp_sk(sk);
487         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
488
489         skb->csum    = 0;
490         tcb->seq     = tcb->end_seq = tp->write_seq;
491         tcb->flags   = TCPCB_FLAG_ACK;
492         tcb->sacked  = 0;
493         skb_header_release(skb);
494         tcp_add_write_queue_tail(sk, skb);
495         sk->sk_wmem_queued += skb->truesize;
496         sk_mem_charge(sk, skb->truesize);
497         if (tp->nonagle & TCP_NAGLE_PUSH)
498                 tp->nonagle &= ~TCP_NAGLE_PUSH;
499 }
500
501 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
502                                 struct sk_buff *skb)
503 {
504         if (flags & MSG_OOB)
505                 tp->snd_up = tp->write_seq;
506 }
507
508 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
509                             int nonagle)
510 {
511         struct tcp_sock *tp = tcp_sk(sk);
512
513         if (tcp_send_head(sk)) {
514                 struct sk_buff *skb = tcp_write_queue_tail(sk);
515                 if (!(flags & MSG_MORE) || forced_push(tp))
516                         tcp_mark_push(tp, skb);
517                 tcp_mark_urg(tp, flags, skb);
518                 __tcp_push_pending_frames(sk, mss_now,
519                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
520         }
521 }
522
523 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
524                                 unsigned int offset, size_t len)
525 {
526         struct tcp_splice_state *tss = rd_desc->arg.data;
527         int ret;
528
529         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
530                               tss->flags);
531         if (ret > 0)
532                 rd_desc->count -= ret;
533         return ret;
534 }
535
536 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
537 {
538         /* Store TCP splice context information in read_descriptor_t. */
539         read_descriptor_t rd_desc = {
540                 .arg.data = tss,
541                 .count    = tss->len,
542         };
543
544         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
545 }
546
547 /**
548  *  tcp_splice_read - splice data from TCP socket to a pipe
549  * @sock:       socket to splice from
550  * @ppos:       position (not valid)
551  * @pipe:       pipe to splice to
552  * @len:        number of bytes to splice
553  * @flags:      splice modifier flags
554  *
555  * Description:
556  *    Will read pages from given socket and fill them into a pipe.
557  *
558  **/
559 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
560                         struct pipe_inode_info *pipe, size_t len,
561                         unsigned int flags)
562 {
563         struct sock *sk = sock->sk;
564         struct tcp_splice_state tss = {
565                 .pipe = pipe,
566                 .len = len,
567                 .flags = flags,
568         };
569         long timeo;
570         ssize_t spliced;
571         int ret;
572
573         /*
574          * We can't seek on a socket input
575          */
576         if (unlikely(*ppos))
577                 return -ESPIPE;
578
579         ret = spliced = 0;
580
581         lock_sock(sk);
582
583         timeo = sock_rcvtimeo(sk, flags & SPLICE_F_NONBLOCK);
584         while (tss.len) {
585                 ret = __tcp_splice_read(sk, &tss);
586                 if (ret < 0)
587                         break;
588                 else if (!ret) {
589                         if (spliced)
590                                 break;
591                         if (sock_flag(sk, SOCK_DONE))
592                                 break;
593                         if (sk->sk_err) {
594                                 ret = sock_error(sk);
595                                 break;
596                         }
597                         if (sk->sk_shutdown & RCV_SHUTDOWN)
598                                 break;
599                         if (sk->sk_state == TCP_CLOSE) {
600                                 /*
601                                  * This occurs when user tries to read
602                                  * from never connected socket.
603                                  */
604                                 if (!sock_flag(sk, SOCK_DONE))
605                                         ret = -ENOTCONN;
606                                 break;
607                         }
608                         if (!timeo) {
609                                 ret = -EAGAIN;
610                                 break;
611                         }
612                         sk_wait_data(sk, &timeo);
613                         if (signal_pending(current)) {
614                                 ret = sock_intr_errno(timeo);
615                                 break;
616                         }
617                         continue;
618                 }
619                 tss.len -= ret;
620                 spliced += ret;
621
622                 if (!timeo)
623                         break;
624                 release_sock(sk);
625                 lock_sock(sk);
626
627                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
628                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
629                     signal_pending(current))
630                         break;
631         }
632
633         release_sock(sk);
634
635         if (spliced)
636                 return spliced;
637
638         return ret;
639 }
640
641 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
642 {
643         struct sk_buff *skb;
644
645         /* The TCP header must be at least 32-bit aligned.  */
646         size = ALIGN(size, 4);
647
648         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
649         if (skb) {
650                 if (sk_wmem_schedule(sk, skb->truesize)) {
651                         /*
652                          * Make sure that we have exactly size bytes
653                          * available to the caller, no more, no less.
654                          */
655                         skb_reserve(skb, skb_tailroom(skb) - size);
656                         return skb;
657                 }
658                 __kfree_skb(skb);
659         } else {
660                 sk->sk_prot->enter_memory_pressure(sk);
661                 sk_stream_moderate_sndbuf(sk);
662         }
663         return NULL;
664 }
665
666 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
667                                        int large_allowed)
668 {
669         struct tcp_sock *tp = tcp_sk(sk);
670         u32 xmit_size_goal, old_size_goal;
671
672         xmit_size_goal = mss_now;
673
674         if (large_allowed && sk_can_gso(sk)) {
675                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
676                                   inet_csk(sk)->icsk_af_ops->net_header_len -
677                                   inet_csk(sk)->icsk_ext_hdr_len -
678                                   tp->tcp_header_len);
679
680                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
681
682                 /* We try hard to avoid divides here */
683                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
684
685                 if (likely(old_size_goal <= xmit_size_goal &&
686                            old_size_goal + mss_now > xmit_size_goal)) {
687                         xmit_size_goal = old_size_goal;
688                 } else {
689                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
690                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
691                 }
692         }
693
694         return max(xmit_size_goal, mss_now);
695 }
696
697 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
698 {
699         int mss_now;
700
701         mss_now = tcp_current_mss(sk);
702         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
703
704         return mss_now;
705 }
706
707 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
708                          size_t psize, int flags)
709 {
710         struct tcp_sock *tp = tcp_sk(sk);
711         int mss_now, size_goal;
712         int err;
713         ssize_t copied;
714         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
715
716         /* Wait for a connection to finish. */
717         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
718                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
719                         goto out_err;
720
721         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
722
723         mss_now = tcp_send_mss(sk, &size_goal, flags);
724         copied = 0;
725
726         err = -EPIPE;
727         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
728                 goto out_err;
729
730         while (psize > 0) {
731                 struct sk_buff *skb = tcp_write_queue_tail(sk);
732                 struct page *page = pages[poffset / PAGE_SIZE];
733                 int copy, i, can_coalesce;
734                 int offset = poffset % PAGE_SIZE;
735                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
736
737                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
738 new_segment:
739                         if (!sk_stream_memory_free(sk))
740                                 goto wait_for_sndbuf;
741
742                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
743                         if (!skb)
744                                 goto wait_for_memory;
745
746                         skb_entail(sk, skb);
747                         copy = size_goal;
748                 }
749
750                 if (copy > size)
751                         copy = size;
752
753                 i = skb_shinfo(skb)->nr_frags;
754                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
755                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
756                         tcp_mark_push(tp, skb);
757                         goto new_segment;
758                 }
759                 if (!sk_wmem_schedule(sk, copy))
760                         goto wait_for_memory;
761
762                 if (can_coalesce) {
763                         skb_shinfo(skb)->frags[i - 1].size += copy;
764                 } else {
765                         get_page(page);
766                         skb_fill_page_desc(skb, i, page, offset, copy);
767                 }
768
769                 skb->len += copy;
770                 skb->data_len += copy;
771                 skb->truesize += copy;
772                 sk->sk_wmem_queued += copy;
773                 sk_mem_charge(sk, copy);
774                 skb->ip_summed = CHECKSUM_PARTIAL;
775                 tp->write_seq += copy;
776                 TCP_SKB_CB(skb)->end_seq += copy;
777                 skb_shinfo(skb)->gso_segs = 0;
778
779                 if (!copied)
780                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
781
782                 copied += copy;
783                 poffset += copy;
784                 if (!(psize -= copy))
785                         goto out;
786
787                 if (skb->len < size_goal || (flags & MSG_OOB))
788                         continue;
789
790                 if (forced_push(tp)) {
791                         tcp_mark_push(tp, skb);
792                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
793                 } else if (skb == tcp_send_head(sk))
794                         tcp_push_one(sk, mss_now);
795                 continue;
796
797 wait_for_sndbuf:
798                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
799 wait_for_memory:
800                 if (copied)
801                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
802
803                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
804                         goto do_error;
805
806                 mss_now = tcp_send_mss(sk, &size_goal, flags);
807         }
808
809 out:
810         if (copied)
811                 tcp_push(sk, flags, mss_now, tp->nonagle);
812         return copied;
813
814 do_error:
815         if (copied)
816                 goto out;
817 out_err:
818         return sk_stream_error(sk, flags, err);
819 }
820
821 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
822                      size_t size, int flags)
823 {
824         ssize_t res;
825         struct sock *sk = sock->sk;
826
827         if (!(sk->sk_route_caps & NETIF_F_SG) ||
828             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
829                 return sock_no_sendpage(sock, page, offset, size, flags);
830
831         lock_sock(sk);
832         TCP_CHECK_TIMER(sk);
833         res = do_tcp_sendpages(sk, &page, offset, size, flags);
834         TCP_CHECK_TIMER(sk);
835         release_sock(sk);
836         return res;
837 }
838
839 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
840 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
841
842 static inline int select_size(struct sock *sk)
843 {
844         struct tcp_sock *tp = tcp_sk(sk);
845         int tmp = tp->mss_cache;
846
847         if (sk->sk_route_caps & NETIF_F_SG) {
848                 if (sk_can_gso(sk))
849                         tmp = 0;
850                 else {
851                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
852
853                         if (tmp >= pgbreak &&
854                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
855                                 tmp = pgbreak;
856                 }
857         }
858
859         return tmp;
860 }
861
862 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
863                 size_t size)
864 {
865         struct sock *sk = sock->sk;
866         struct iovec *iov;
867         struct tcp_sock *tp = tcp_sk(sk);
868         struct sk_buff *skb;
869         int iovlen, flags;
870         int mss_now, size_goal;
871         int err, copied;
872         long timeo;
873
874         lock_sock(sk);
875         TCP_CHECK_TIMER(sk);
876
877         flags = msg->msg_flags;
878         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
879
880         /* Wait for a connection to finish. */
881         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
882                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
883                         goto out_err;
884
885         /* This should be in poll */
886         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
887
888         mss_now = tcp_send_mss(sk, &size_goal, flags);
889
890         /* Ok commence sending. */
891         iovlen = msg->msg_iovlen;
892         iov = msg->msg_iov;
893         copied = 0;
894
895         err = -EPIPE;
896         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
897                 goto out_err;
898
899         while (--iovlen >= 0) {
900                 int seglen = iov->iov_len;
901                 unsigned char __user *from = iov->iov_base;
902
903                 iov++;
904
905                 while (seglen > 0) {
906                         int copy = 0;
907                         int max = size_goal;
908
909                         skb = tcp_write_queue_tail(sk);
910                         if (tcp_send_head(sk)) {
911                                 if (skb->ip_summed == CHECKSUM_NONE)
912                                         max = mss_now;
913                                 copy = max - skb->len;
914                         }
915
916                         if (copy <= 0) {
917 new_segment:
918                                 /* Allocate new segment. If the interface is SG,
919                                  * allocate skb fitting to single page.
920                                  */
921                                 if (!sk_stream_memory_free(sk))
922                                         goto wait_for_sndbuf;
923
924                                 skb = sk_stream_alloc_skb(sk, select_size(sk),
925                                                 sk->sk_allocation);
926                                 if (!skb)
927                                         goto wait_for_memory;
928
929                                 /*
930                                  * Check whether we can use HW checksum.
931                                  */
932                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
933                                         skb->ip_summed = CHECKSUM_PARTIAL;
934
935                                 skb_entail(sk, skb);
936                                 copy = size_goal;
937                                 max = size_goal;
938                         }
939
940                         /* Try to append data to the end of skb. */
941                         if (copy > seglen)
942                                 copy = seglen;
943
944                         /* Where to copy to? */
945                         if (skb_tailroom(skb) > 0) {
946                                 /* We have some space in skb head. Superb! */
947                                 if (copy > skb_tailroom(skb))
948                                         copy = skb_tailroom(skb);
949                                 if ((err = skb_add_data(skb, from, copy)) != 0)
950                                         goto do_fault;
951                         } else {
952                                 int merge = 0;
953                                 int i = skb_shinfo(skb)->nr_frags;
954                                 struct page *page = TCP_PAGE(sk);
955                                 int off = TCP_OFF(sk);
956
957                                 if (skb_can_coalesce(skb, i, page, off) &&
958                                     off != PAGE_SIZE) {
959                                         /* We can extend the last page
960                                          * fragment. */
961                                         merge = 1;
962                                 } else if (i == MAX_SKB_FRAGS ||
963                                            (!i &&
964                                            !(sk->sk_route_caps & NETIF_F_SG))) {
965                                         /* Need to add new fragment and cannot
966                                          * do this because interface is non-SG,
967                                          * or because all the page slots are
968                                          * busy. */
969                                         tcp_mark_push(tp, skb);
970                                         goto new_segment;
971                                 } else if (page) {
972                                         if (off == PAGE_SIZE) {
973                                                 put_page(page);
974                                                 TCP_PAGE(sk) = page = NULL;
975                                                 off = 0;
976                                         }
977                                 } else
978                                         off = 0;
979
980                                 if (copy > PAGE_SIZE - off)
981                                         copy = PAGE_SIZE - off;
982
983                                 if (!sk_wmem_schedule(sk, copy))
984                                         goto wait_for_memory;
985
986                                 if (!page) {
987                                         /* Allocate new cache page. */
988                                         if (!(page = sk_stream_alloc_page(sk)))
989                                                 goto wait_for_memory;
990                                 }
991
992                                 /* Time to copy data. We are close to
993                                  * the end! */
994                                 err = skb_copy_to_page(sk, from, skb, page,
995                                                        off, copy);
996                                 if (err) {
997                                         /* If this page was new, give it to the
998                                          * socket so it does not get leaked.
999                                          */
1000                                         if (!TCP_PAGE(sk)) {
1001                                                 TCP_PAGE(sk) = page;
1002                                                 TCP_OFF(sk) = 0;
1003                                         }
1004                                         goto do_error;
1005                                 }
1006
1007                                 /* Update the skb. */
1008                                 if (merge) {
1009                                         skb_shinfo(skb)->frags[i - 1].size +=
1010                                                                         copy;
1011                                 } else {
1012                                         skb_fill_page_desc(skb, i, page, off, copy);
1013                                         if (TCP_PAGE(sk)) {
1014                                                 get_page(page);
1015                                         } else if (off + copy < PAGE_SIZE) {
1016                                                 get_page(page);
1017                                                 TCP_PAGE(sk) = page;
1018                                         }
1019                                 }
1020
1021                                 TCP_OFF(sk) = off + copy;
1022                         }
1023
1024                         if (!copied)
1025                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
1026
1027                         tp->write_seq += copy;
1028                         TCP_SKB_CB(skb)->end_seq += copy;
1029                         skb_shinfo(skb)->gso_segs = 0;
1030
1031                         from += copy;
1032                         copied += copy;
1033                         if ((seglen -= copy) == 0 && iovlen == 0)
1034                                 goto out;
1035
1036                         if (skb->len < max || (flags & MSG_OOB))
1037                                 continue;
1038
1039                         if (forced_push(tp)) {
1040                                 tcp_mark_push(tp, skb);
1041                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1042                         } else if (skb == tcp_send_head(sk))
1043                                 tcp_push_one(sk, mss_now);
1044                         continue;
1045
1046 wait_for_sndbuf:
1047                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1048 wait_for_memory:
1049                         if (copied)
1050                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1051
1052                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1053                                 goto do_error;
1054
1055                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1056                 }
1057         }
1058
1059 out:
1060         if (copied)
1061                 tcp_push(sk, flags, mss_now, tp->nonagle);
1062         TCP_CHECK_TIMER(sk);
1063         release_sock(sk);
1064         return copied;
1065
1066 do_fault:
1067         if (!skb->len) {
1068                 tcp_unlink_write_queue(skb, sk);
1069                 /* It is the one place in all of TCP, except connection
1070                  * reset, where we can be unlinking the send_head.
1071                  */
1072                 tcp_check_send_head(sk, skb);
1073                 sk_wmem_free_skb(sk, skb);
1074         }
1075
1076 do_error:
1077         if (copied)
1078                 goto out;
1079 out_err:
1080         err = sk_stream_error(sk, flags, err);
1081         TCP_CHECK_TIMER(sk);
1082         release_sock(sk);
1083         return err;
1084 }
1085
1086 /*
1087  *      Handle reading urgent data. BSD has very simple semantics for
1088  *      this, no blocking and very strange errors 8)
1089  */
1090
1091 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1092 {
1093         struct tcp_sock *tp = tcp_sk(sk);
1094
1095         /* No URG data to read. */
1096         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1097             tp->urg_data == TCP_URG_READ)
1098                 return -EINVAL; /* Yes this is right ! */
1099
1100         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1101                 return -ENOTCONN;
1102
1103         if (tp->urg_data & TCP_URG_VALID) {
1104                 int err = 0;
1105                 char c = tp->urg_data;
1106
1107                 if (!(flags & MSG_PEEK))
1108                         tp->urg_data = TCP_URG_READ;
1109
1110                 /* Read urgent data. */
1111                 msg->msg_flags |= MSG_OOB;
1112
1113                 if (len > 0) {
1114                         if (!(flags & MSG_TRUNC))
1115                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1116                         len = 1;
1117                 } else
1118                         msg->msg_flags |= MSG_TRUNC;
1119
1120                 return err ? -EFAULT : len;
1121         }
1122
1123         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1124                 return 0;
1125
1126         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1127          * the available implementations agree in this case:
1128          * this call should never block, independent of the
1129          * blocking state of the socket.
1130          * Mike <pall@rz.uni-karlsruhe.de>
1131          */
1132         return -EAGAIN;
1133 }
1134
1135 /* Clean up the receive buffer for full frames taken by the user,
1136  * then send an ACK if necessary.  COPIED is the number of bytes
1137  * tcp_recvmsg has given to the user so far, it speeds up the
1138  * calculation of whether or not we must ACK for the sake of
1139  * a window update.
1140  */
1141 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1142 {
1143         struct tcp_sock *tp = tcp_sk(sk);
1144         int time_to_ack = 0;
1145
1146 #if TCP_DEBUG
1147         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1148
1149         WARN_ON(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1150 #endif
1151
1152         if (inet_csk_ack_scheduled(sk)) {
1153                 const struct inet_connection_sock *icsk = inet_csk(sk);
1154                    /* Delayed ACKs frequently hit locked sockets during bulk
1155                     * receive. */
1156                 if (icsk->icsk_ack.blocked ||
1157                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1158                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1159                     /*
1160                      * If this read emptied read buffer, we send ACK, if
1161                      * connection is not bidirectional, user drained
1162                      * receive buffer and there was a small segment
1163                      * in queue.
1164                      */
1165                     (copied > 0 &&
1166                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1167                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1168                        !icsk->icsk_ack.pingpong)) &&
1169                       !atomic_read(&sk->sk_rmem_alloc)))
1170                         time_to_ack = 1;
1171         }
1172
1173         /* We send an ACK if we can now advertise a non-zero window
1174          * which has been raised "significantly".
1175          *
1176          * Even if window raised up to infinity, do not send window open ACK
1177          * in states, where we will not receive more. It is useless.
1178          */
1179         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1180                 __u32 rcv_window_now = tcp_receive_window(tp);
1181
1182                 /* Optimize, __tcp_select_window() is not cheap. */
1183                 if (2*rcv_window_now <= tp->window_clamp) {
1184                         __u32 new_window = __tcp_select_window(sk);
1185
1186                         /* Send ACK now, if this read freed lots of space
1187                          * in our buffer. Certainly, new_window is new window.
1188                          * We can advertise it now, if it is not less than current one.
1189                          * "Lots" means "at least twice" here.
1190                          */
1191                         if (new_window && new_window >= 2 * rcv_window_now)
1192                                 time_to_ack = 1;
1193                 }
1194         }
1195         if (time_to_ack)
1196                 tcp_send_ack(sk);
1197 }
1198
1199 static void tcp_prequeue_process(struct sock *sk)
1200 {
1201         struct sk_buff *skb;
1202         struct tcp_sock *tp = tcp_sk(sk);
1203
1204         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1205
1206         /* RX process wants to run with disabled BHs, though it is not
1207          * necessary */
1208         local_bh_disable();
1209         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1210                 sk_backlog_rcv(sk, skb);
1211         local_bh_enable();
1212
1213         /* Clear memory counter. */
1214         tp->ucopy.memory = 0;
1215 }
1216
1217 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1218 {
1219         struct sk_buff *skb;
1220         u32 offset;
1221
1222         skb_queue_walk(&sk->sk_receive_queue, skb) {
1223                 offset = seq - TCP_SKB_CB(skb)->seq;
1224                 if (tcp_hdr(skb)->syn)
1225                         offset--;
1226                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1227                         *off = offset;
1228                         return skb;
1229                 }
1230         }
1231         return NULL;
1232 }
1233
1234 /*
1235  * This routine provides an alternative to tcp_recvmsg() for routines
1236  * that would like to handle copying from skbuffs directly in 'sendfile'
1237  * fashion.
1238  * Note:
1239  *      - It is assumed that the socket was locked by the caller.
1240  *      - The routine does not block.
1241  *      - At present, there is no support for reading OOB data
1242  *        or for 'peeking' the socket using this routine
1243  *        (although both would be easy to implement).
1244  */
1245 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1246                   sk_read_actor_t recv_actor)
1247 {
1248         struct sk_buff *skb;
1249         struct tcp_sock *tp = tcp_sk(sk);
1250         u32 seq = tp->copied_seq;
1251         u32 offset;
1252         int copied = 0;
1253
1254         if (sk->sk_state == TCP_LISTEN)
1255                 return -ENOTCONN;
1256         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1257                 if (offset < skb->len) {
1258                         int used;
1259                         size_t len;
1260
1261                         len = skb->len - offset;
1262                         /* Stop reading if we hit a patch of urgent data */
1263                         if (tp->urg_data) {
1264                                 u32 urg_offset = tp->urg_seq - seq;
1265                                 if (urg_offset < len)
1266                                         len = urg_offset;
1267                                 if (!len)
1268                                         break;
1269                         }
1270                         used = recv_actor(desc, skb, offset, len);
1271                         if (used < 0) {
1272                                 if (!copied)
1273                                         copied = used;
1274                                 break;
1275                         } else if (used <= len) {
1276                                 seq += used;
1277                                 copied += used;
1278                                 offset += used;
1279                         }
1280                         /*
1281                          * If recv_actor drops the lock (e.g. TCP splice
1282                          * receive) the skb pointer might be invalid when
1283                          * getting here: tcp_collapse might have deleted it
1284                          * while aggregating skbs from the socket queue.
1285                          */
1286                         skb = tcp_recv_skb(sk, seq-1, &offset);
1287                         if (!skb || (offset+1 != skb->len))
1288                                 break;
1289                 }
1290                 if (tcp_hdr(skb)->fin) {
1291                         sk_eat_skb(sk, skb, 0);
1292                         ++seq;
1293                         break;
1294                 }
1295                 sk_eat_skb(sk, skb, 0);
1296                 if (!desc->count)
1297                         break;
1298         }
1299         tp->copied_seq = seq;
1300
1301         tcp_rcv_space_adjust(sk);
1302
1303         /* Clean up data we have read: This will do ACK frames. */
1304         if (copied > 0)
1305                 tcp_cleanup_rbuf(sk, copied);
1306         return copied;
1307 }
1308
1309 /*
1310  *      This routine copies from a sock struct into the user buffer.
1311  *
1312  *      Technical note: in 2.3 we work on _locked_ socket, so that
1313  *      tricks with *seq access order and skb->users are not required.
1314  *      Probably, code can be easily improved even more.
1315  */
1316
1317 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1318                 size_t len, int nonblock, int flags, int *addr_len)
1319 {
1320         struct tcp_sock *tp = tcp_sk(sk);
1321         int copied = 0;
1322         u32 peek_seq;
1323         u32 *seq;
1324         unsigned long used;
1325         int err;
1326         int target;             /* Read at least this many bytes */
1327         long timeo;
1328         struct task_struct *user_recv = NULL;
1329         int copied_early = 0;
1330         struct sk_buff *skb;
1331         u32 urg_hole = 0;
1332
1333         lock_sock(sk);
1334
1335         TCP_CHECK_TIMER(sk);
1336
1337         err = -ENOTCONN;
1338         if (sk->sk_state == TCP_LISTEN)
1339                 goto out;
1340
1341         timeo = sock_rcvtimeo(sk, nonblock);
1342
1343         /* Urgent data needs to be handled specially. */
1344         if (flags & MSG_OOB)
1345                 goto recv_urg;
1346
1347         seq = &tp->copied_seq;
1348         if (flags & MSG_PEEK) {
1349                 peek_seq = tp->copied_seq;
1350                 seq = &peek_seq;
1351         }
1352
1353         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1354
1355 #ifdef CONFIG_NET_DMA
1356         tp->ucopy.dma_chan = NULL;
1357         preempt_disable();
1358         skb = skb_peek_tail(&sk->sk_receive_queue);
1359         {
1360                 int available = 0;
1361
1362                 if (skb)
1363                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1364                 if ((available < target) &&
1365                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1366                     !sysctl_tcp_low_latency &&
1367                     dma_find_channel(DMA_MEMCPY)) {
1368                         preempt_enable_no_resched();
1369                         tp->ucopy.pinned_list =
1370                                         dma_pin_iovec_pages(msg->msg_iov, len);
1371                 } else {
1372                         preempt_enable_no_resched();
1373                 }
1374         }
1375 #endif
1376
1377         do {
1378                 u32 offset;
1379
1380                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1381                 if (tp->urg_data && tp->urg_seq == *seq) {
1382                         if (copied)
1383                                 break;
1384                         if (signal_pending(current)) {
1385                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1386                                 break;
1387                         }
1388                 }
1389
1390                 /* Next get a buffer. */
1391
1392                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1393                         /* Now that we have two receive queues this
1394                          * shouldn't happen.
1395                          */
1396                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1397                                 printk(KERN_INFO "recvmsg bug: copied %X "
1398                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1399                                 break;
1400                         }
1401                         offset = *seq - TCP_SKB_CB(skb)->seq;
1402                         if (tcp_hdr(skb)->syn)
1403                                 offset--;
1404                         if (offset < skb->len)
1405                                 goto found_ok_skb;
1406                         if (tcp_hdr(skb)->fin)
1407                                 goto found_fin_ok;
1408                         WARN_ON(!(flags & MSG_PEEK));
1409                 }
1410
1411                 /* Well, if we have backlog, try to process it now yet. */
1412
1413                 if (copied >= target && !sk->sk_backlog.tail)
1414                         break;
1415
1416                 if (copied) {
1417                         if (sk->sk_err ||
1418                             sk->sk_state == TCP_CLOSE ||
1419                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1420                             !timeo ||
1421                             signal_pending(current))
1422                                 break;
1423                 } else {
1424                         if (sock_flag(sk, SOCK_DONE))
1425                                 break;
1426
1427                         if (sk->sk_err) {
1428                                 copied = sock_error(sk);
1429                                 break;
1430                         }
1431
1432                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1433                                 break;
1434
1435                         if (sk->sk_state == TCP_CLOSE) {
1436                                 if (!sock_flag(sk, SOCK_DONE)) {
1437                                         /* This occurs when user tries to read
1438                                          * from never connected socket.
1439                                          */
1440                                         copied = -ENOTCONN;
1441                                         break;
1442                                 }
1443                                 break;
1444                         }
1445
1446                         if (!timeo) {
1447                                 copied = -EAGAIN;
1448                                 break;
1449                         }
1450
1451                         if (signal_pending(current)) {
1452                                 copied = sock_intr_errno(timeo);
1453                                 break;
1454                         }
1455                 }
1456
1457                 tcp_cleanup_rbuf(sk, copied);
1458
1459                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1460                         /* Install new reader */
1461                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1462                                 user_recv = current;
1463                                 tp->ucopy.task = user_recv;
1464                                 tp->ucopy.iov = msg->msg_iov;
1465                         }
1466
1467                         tp->ucopy.len = len;
1468
1469                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1470                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1471
1472                         /* Ugly... If prequeue is not empty, we have to
1473                          * process it before releasing socket, otherwise
1474                          * order will be broken at second iteration.
1475                          * More elegant solution is required!!!
1476                          *
1477                          * Look: we have the following (pseudo)queues:
1478                          *
1479                          * 1. packets in flight
1480                          * 2. backlog
1481                          * 3. prequeue
1482                          * 4. receive_queue
1483                          *
1484                          * Each queue can be processed only if the next ones
1485                          * are empty. At this point we have empty receive_queue.
1486                          * But prequeue _can_ be not empty after 2nd iteration,
1487                          * when we jumped to start of loop because backlog
1488                          * processing added something to receive_queue.
1489                          * We cannot release_sock(), because backlog contains
1490                          * packets arrived _after_ prequeued ones.
1491                          *
1492                          * Shortly, algorithm is clear --- to process all
1493                          * the queues in order. We could make it more directly,
1494                          * requeueing packets from backlog to prequeue, if
1495                          * is not empty. It is more elegant, but eats cycles,
1496                          * unfortunately.
1497                          */
1498                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1499                                 goto do_prequeue;
1500
1501                         /* __ Set realtime policy in scheduler __ */
1502                 }
1503
1504                 if (copied >= target) {
1505                         /* Do not sleep, just process backlog. */
1506                         release_sock(sk);
1507                         lock_sock(sk);
1508                 } else
1509                         sk_wait_data(sk, &timeo);
1510
1511 #ifdef CONFIG_NET_DMA
1512                 tp->ucopy.wakeup = 0;
1513 #endif
1514
1515                 if (user_recv) {
1516                         int chunk;
1517
1518                         /* __ Restore normal policy in scheduler __ */
1519
1520                         if ((chunk = len - tp->ucopy.len) != 0) {
1521                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1522                                 len -= chunk;
1523                                 copied += chunk;
1524                         }
1525
1526                         if (tp->rcv_nxt == tp->copied_seq &&
1527                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1528 do_prequeue:
1529                                 tcp_prequeue_process(sk);
1530
1531                                 if ((chunk = len - tp->ucopy.len) != 0) {
1532                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1533                                         len -= chunk;
1534                                         copied += chunk;
1535                                 }
1536                         }
1537                 }
1538                 if ((flags & MSG_PEEK) &&
1539                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1540                         if (net_ratelimit())
1541                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1542                                        current->comm, task_pid_nr(current));
1543                         peek_seq = tp->copied_seq;
1544                 }
1545                 continue;
1546
1547         found_ok_skb:
1548                 /* Ok so how much can we use? */
1549                 used = skb->len - offset;
1550                 if (len < used)
1551                         used = len;
1552
1553                 /* Do we have urgent data here? */
1554                 if (tp->urg_data) {
1555                         u32 urg_offset = tp->urg_seq - *seq;
1556                         if (urg_offset < used) {
1557                                 if (!urg_offset) {
1558                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1559                                                 ++*seq;
1560                                                 urg_hole++;
1561                                                 offset++;
1562                                                 used--;
1563                                                 if (!used)
1564                                                         goto skip_copy;
1565                                         }
1566                                 } else
1567                                         used = urg_offset;
1568                         }
1569                 }
1570
1571                 if (!(flags & MSG_TRUNC)) {
1572 #ifdef CONFIG_NET_DMA
1573                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1574                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1575
1576                         if (tp->ucopy.dma_chan) {
1577                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1578                                         tp->ucopy.dma_chan, skb, offset,
1579                                         msg->msg_iov, used,
1580                                         tp->ucopy.pinned_list);
1581
1582                                 if (tp->ucopy.dma_cookie < 0) {
1583
1584                                         printk(KERN_ALERT "dma_cookie < 0\n");
1585
1586                                         /* Exception. Bailout! */
1587                                         if (!copied)
1588                                                 copied = -EFAULT;
1589                                         break;
1590                                 }
1591                                 if ((offset + used) == skb->len)
1592                                         copied_early = 1;
1593
1594                         } else
1595 #endif
1596                         {
1597                                 err = skb_copy_datagram_iovec(skb, offset,
1598                                                 msg->msg_iov, used);
1599                                 if (err) {
1600                                         /* Exception. Bailout! */
1601                                         if (!copied)
1602                                                 copied = -EFAULT;
1603                                         break;
1604                                 }
1605                         }
1606                 }
1607
1608                 *seq += used;
1609                 copied += used;
1610                 len -= used;
1611
1612                 tcp_rcv_space_adjust(sk);
1613
1614 skip_copy:
1615                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1616                         tp->urg_data = 0;
1617                         tcp_fast_path_check(sk);
1618                 }
1619                 if (used + offset < skb->len)
1620                         continue;
1621
1622                 if (tcp_hdr(skb)->fin)
1623                         goto found_fin_ok;
1624                 if (!(flags & MSG_PEEK)) {
1625                         sk_eat_skb(sk, skb, copied_early);
1626                         copied_early = 0;
1627                 }
1628                 continue;
1629
1630         found_fin_ok:
1631                 /* Process the FIN. */
1632                 ++*seq;
1633                 if (!(flags & MSG_PEEK)) {
1634                         sk_eat_skb(sk, skb, copied_early);
1635                         copied_early = 0;
1636                 }
1637                 break;
1638         } while (len > 0);
1639
1640         if (user_recv) {
1641                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1642                         int chunk;
1643
1644                         tp->ucopy.len = copied > 0 ? len : 0;
1645
1646                         tcp_prequeue_process(sk);
1647
1648                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1649                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1650                                 len -= chunk;
1651                                 copied += chunk;
1652                         }
1653                 }
1654
1655                 tp->ucopy.task = NULL;
1656                 tp->ucopy.len = 0;
1657         }
1658
1659 #ifdef CONFIG_NET_DMA
1660         if (tp->ucopy.dma_chan) {
1661                 dma_cookie_t done, used;
1662
1663                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1664
1665                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1666                                                  tp->ucopy.dma_cookie, &done,
1667                                                  &used) == DMA_IN_PROGRESS) {
1668                         /* do partial cleanup of sk_async_wait_queue */
1669                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1670                                (dma_async_is_complete(skb->dma_cookie, done,
1671                                                       used) == DMA_SUCCESS)) {
1672                                 __skb_dequeue(&sk->sk_async_wait_queue);
1673                                 kfree_skb(skb);
1674                         }
1675                 }
1676
1677                 /* Safe to free early-copied skbs now */
1678                 __skb_queue_purge(&sk->sk_async_wait_queue);
1679                 tp->ucopy.dma_chan = NULL;
1680         }
1681         if (tp->ucopy.pinned_list) {
1682                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1683                 tp->ucopy.pinned_list = NULL;
1684         }
1685 #endif
1686
1687         /* According to UNIX98, msg_name/msg_namelen are ignored
1688          * on connected socket. I was just happy when found this 8) --ANK
1689          */
1690
1691         /* Clean up data we have read: This will do ACK frames. */
1692         tcp_cleanup_rbuf(sk, copied);
1693
1694         TCP_CHECK_TIMER(sk);
1695         release_sock(sk);
1696         return copied;
1697
1698 out:
1699         TCP_CHECK_TIMER(sk);
1700         release_sock(sk);
1701         return err;
1702
1703 recv_urg:
1704         err = tcp_recv_urg(sk, msg, len, flags);
1705         goto out;
1706 }
1707
1708 void tcp_set_state(struct sock *sk, int state)
1709 {
1710         int oldstate = sk->sk_state;
1711
1712         switch (state) {
1713         case TCP_ESTABLISHED:
1714                 if (oldstate != TCP_ESTABLISHED)
1715                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1716                 break;
1717
1718         case TCP_CLOSE:
1719                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1720                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1721
1722                 sk->sk_prot->unhash(sk);
1723                 if (inet_csk(sk)->icsk_bind_hash &&
1724                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1725                         inet_put_port(sk);
1726                 /* fall through */
1727         default:
1728                 if (oldstate == TCP_ESTABLISHED)
1729                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1730         }
1731
1732         /* Change state AFTER socket is unhashed to avoid closed
1733          * socket sitting in hash tables.
1734          */
1735         sk->sk_state = state;
1736
1737 #ifdef STATE_TRACE
1738         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1739 #endif
1740 }
1741 EXPORT_SYMBOL_GPL(tcp_set_state);
1742
1743 /*
1744  *      State processing on a close. This implements the state shift for
1745  *      sending our FIN frame. Note that we only send a FIN for some
1746  *      states. A shutdown() may have already sent the FIN, or we may be
1747  *      closed.
1748  */
1749
1750 static const unsigned char new_state[16] = {
1751   /* current state:        new state:      action:      */
1752   /* (Invalid)          */ TCP_CLOSE,
1753   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1754   /* TCP_SYN_SENT       */ TCP_CLOSE,
1755   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1756   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1757   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1758   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1759   /* TCP_CLOSE          */ TCP_CLOSE,
1760   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1761   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1762   /* TCP_LISTEN         */ TCP_CLOSE,
1763   /* TCP_CLOSING        */ TCP_CLOSING,
1764 };
1765
1766 static int tcp_close_state(struct sock *sk)
1767 {
1768         int next = (int)new_state[sk->sk_state];
1769         int ns = next & TCP_STATE_MASK;
1770
1771         tcp_set_state(sk, ns);
1772
1773         return next & TCP_ACTION_FIN;
1774 }
1775
1776 /*
1777  *      Shutdown the sending side of a connection. Much like close except
1778  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1779  */
1780
1781 void tcp_shutdown(struct sock *sk, int how)
1782 {
1783         /*      We need to grab some memory, and put together a FIN,
1784          *      and then put it into the queue to be sent.
1785          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1786          */
1787         if (!(how & SEND_SHUTDOWN))
1788                 return;
1789
1790         /* If we've already sent a FIN, or it's a closed state, skip this. */
1791         if ((1 << sk->sk_state) &
1792             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1793              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1794                 /* Clear out any half completed packets.  FIN if needed. */
1795                 if (tcp_close_state(sk))
1796                         tcp_send_fin(sk);
1797         }
1798 }
1799
1800 void tcp_close(struct sock *sk, long timeout)
1801 {
1802         struct sk_buff *skb;
1803         int data_was_unread = 0;
1804         int state;
1805
1806         lock_sock(sk);
1807         sk->sk_shutdown = SHUTDOWN_MASK;
1808
1809         if (sk->sk_state == TCP_LISTEN) {
1810                 tcp_set_state(sk, TCP_CLOSE);
1811
1812                 /* Special case. */
1813                 inet_csk_listen_stop(sk);
1814
1815                 goto adjudge_to_death;
1816         }
1817
1818         /*  We need to flush the recv. buffs.  We do this only on the
1819          *  descriptor close, not protocol-sourced closes, because the
1820          *  reader process may not have drained the data yet!
1821          */
1822         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1823                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1824                           tcp_hdr(skb)->fin;
1825                 data_was_unread += len;
1826                 __kfree_skb(skb);
1827         }
1828
1829         sk_mem_reclaim(sk);
1830
1831         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1832          * data was lost. To witness the awful effects of the old behavior of
1833          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1834          * GET in an FTP client, suspend the process, wait for the client to
1835          * advertise a zero window, then kill -9 the FTP client, wheee...
1836          * Note: timeout is always zero in such a case.
1837          */
1838         if (data_was_unread) {
1839                 /* Unread data was tossed, zap the connection. */
1840                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1841                 tcp_set_state(sk, TCP_CLOSE);
1842                 tcp_send_active_reset(sk, GFP_KERNEL);
1843         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1844                 /* Check zero linger _after_ checking for unread data. */
1845                 sk->sk_prot->disconnect(sk, 0);
1846                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1847         } else if (tcp_close_state(sk)) {
1848                 /* We FIN if the application ate all the data before
1849                  * zapping the connection.
1850                  */
1851
1852                 /* RED-PEN. Formally speaking, we have broken TCP state
1853                  * machine. State transitions:
1854                  *
1855                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1856                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1857                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1858                  *
1859                  * are legal only when FIN has been sent (i.e. in window),
1860                  * rather than queued out of window. Purists blame.
1861                  *
1862                  * F.e. "RFC state" is ESTABLISHED,
1863                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1864                  *
1865                  * The visible declinations are that sometimes
1866                  * we enter time-wait state, when it is not required really
1867                  * (harmless), do not send active resets, when they are
1868                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1869                  * they look as CLOSING or LAST_ACK for Linux)
1870                  * Probably, I missed some more holelets.
1871                  *                                              --ANK
1872                  */
1873                 tcp_send_fin(sk);
1874         }
1875
1876         sk_stream_wait_close(sk, timeout);
1877
1878 adjudge_to_death:
1879         state = sk->sk_state;
1880         sock_hold(sk);
1881         sock_orphan(sk);
1882
1883         /* It is the last release_sock in its life. It will remove backlog. */
1884         release_sock(sk);
1885
1886
1887         /* Now socket is owned by kernel and we acquire BH lock
1888            to finish close. No need to check for user refs.
1889          */
1890         local_bh_disable();
1891         bh_lock_sock(sk);
1892         WARN_ON(sock_owned_by_user(sk));
1893
1894         percpu_counter_inc(sk->sk_prot->orphan_count);
1895
1896         /* Have we already been destroyed by a softirq or backlog? */
1897         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1898                 goto out;
1899
1900         /*      This is a (useful) BSD violating of the RFC. There is a
1901          *      problem with TCP as specified in that the other end could
1902          *      keep a socket open forever with no application left this end.
1903          *      We use a 3 minute timeout (about the same as BSD) then kill
1904          *      our end. If they send after that then tough - BUT: long enough
1905          *      that we won't make the old 4*rto = almost no time - whoops
1906          *      reset mistake.
1907          *
1908          *      Nope, it was not mistake. It is really desired behaviour
1909          *      f.e. on http servers, when such sockets are useless, but
1910          *      consume significant resources. Let's do it with special
1911          *      linger2 option.                                 --ANK
1912          */
1913
1914         if (sk->sk_state == TCP_FIN_WAIT2) {
1915                 struct tcp_sock *tp = tcp_sk(sk);
1916                 if (tp->linger2 < 0) {
1917                         tcp_set_state(sk, TCP_CLOSE);
1918                         tcp_send_active_reset(sk, GFP_ATOMIC);
1919                         NET_INC_STATS_BH(sock_net(sk),
1920                                         LINUX_MIB_TCPABORTONLINGER);
1921                 } else {
1922                         const int tmo = tcp_fin_time(sk);
1923
1924                         if (tmo > TCP_TIMEWAIT_LEN) {
1925                                 inet_csk_reset_keepalive_timer(sk,
1926                                                 tmo - TCP_TIMEWAIT_LEN);
1927                         } else {
1928                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1929                                 goto out;
1930                         }
1931                 }
1932         }
1933         if (sk->sk_state != TCP_CLOSE) {
1934                 int orphan_count = percpu_counter_read_positive(
1935                                                 sk->sk_prot->orphan_count);
1936
1937                 sk_mem_reclaim(sk);
1938                 if (tcp_too_many_orphans(sk, orphan_count)) {
1939                         if (net_ratelimit())
1940                                 printk(KERN_INFO "TCP: too many of orphaned "
1941                                        "sockets\n");
1942                         tcp_set_state(sk, TCP_CLOSE);
1943                         tcp_send_active_reset(sk, GFP_ATOMIC);
1944                         NET_INC_STATS_BH(sock_net(sk),
1945                                         LINUX_MIB_TCPABORTONMEMORY);
1946                 }
1947         }
1948
1949         if (sk->sk_state == TCP_CLOSE)
1950                 inet_csk_destroy_sock(sk);
1951         /* Otherwise, socket is reprieved until protocol close. */
1952
1953 out:
1954         bh_unlock_sock(sk);
1955         local_bh_enable();
1956         sock_put(sk);
1957 }
1958
1959 /* These states need RST on ABORT according to RFC793 */
1960
1961 static inline int tcp_need_reset(int state)
1962 {
1963         return (1 << state) &
1964                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1965                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1966 }
1967
1968 int tcp_disconnect(struct sock *sk, int flags)
1969 {
1970         struct inet_sock *inet = inet_sk(sk);
1971         struct inet_connection_sock *icsk = inet_csk(sk);
1972         struct tcp_sock *tp = tcp_sk(sk);
1973         int err = 0;
1974         int old_state = sk->sk_state;
1975
1976         if (old_state != TCP_CLOSE)
1977                 tcp_set_state(sk, TCP_CLOSE);
1978
1979         /* ABORT function of RFC793 */
1980         if (old_state == TCP_LISTEN) {
1981                 inet_csk_listen_stop(sk);
1982         } else if (tcp_need_reset(old_state) ||
1983                    (tp->snd_nxt != tp->write_seq &&
1984                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1985                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1986                  * states
1987                  */
1988                 tcp_send_active_reset(sk, gfp_any());
1989                 sk->sk_err = ECONNRESET;
1990         } else if (old_state == TCP_SYN_SENT)
1991                 sk->sk_err = ECONNRESET;
1992
1993         tcp_clear_xmit_timers(sk);
1994         __skb_queue_purge(&sk->sk_receive_queue);
1995         tcp_write_queue_purge(sk);
1996         __skb_queue_purge(&tp->out_of_order_queue);
1997 #ifdef CONFIG_NET_DMA
1998         __skb_queue_purge(&sk->sk_async_wait_queue);
1999 #endif
2000
2001         inet->dport = 0;
2002
2003         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2004                 inet_reset_saddr(sk);
2005
2006         sk->sk_shutdown = 0;
2007         sock_reset_flag(sk, SOCK_DONE);
2008         tp->srtt = 0;
2009         if ((tp->write_seq += tp->max_window + 2) == 0)
2010                 tp->write_seq = 1;
2011         icsk->icsk_backoff = 0;
2012         tp->snd_cwnd = 2;
2013         icsk->icsk_probes_out = 0;
2014         tp->packets_out = 0;
2015         tp->snd_ssthresh = 0x7fffffff;
2016         tp->snd_cwnd_cnt = 0;
2017         tp->bytes_acked = 0;
2018         tcp_set_ca_state(sk, TCP_CA_Open);
2019         tcp_clear_retrans(tp);
2020         inet_csk_delack_init(sk);
2021         tcp_init_send_head(sk);
2022         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2023         __sk_dst_reset(sk);
2024
2025         WARN_ON(inet->num && !icsk->icsk_bind_hash);
2026
2027         sk->sk_error_report(sk);
2028         return err;
2029 }
2030
2031 /*
2032  *      Socket option code for TCP.
2033  */
2034 static int do_tcp_setsockopt(struct sock *sk, int level,
2035                 int optname, char __user *optval, int optlen)
2036 {
2037         struct tcp_sock *tp = tcp_sk(sk);
2038         struct inet_connection_sock *icsk = inet_csk(sk);
2039         int val;
2040         int err = 0;
2041
2042         /* This is a string value all the others are int's */
2043         if (optname == TCP_CONGESTION) {
2044                 char name[TCP_CA_NAME_MAX];
2045
2046                 if (optlen < 1)
2047                         return -EINVAL;
2048
2049                 val = strncpy_from_user(name, optval,
2050                                         min(TCP_CA_NAME_MAX-1, optlen));
2051                 if (val < 0)
2052                         return -EFAULT;
2053                 name[val] = 0;
2054
2055                 lock_sock(sk);
2056                 err = tcp_set_congestion_control(sk, name);
2057                 release_sock(sk);
2058                 return err;
2059         }
2060
2061         if (optlen < sizeof(int))
2062                 return -EINVAL;
2063
2064         if (get_user(val, (int __user *)optval))
2065                 return -EFAULT;
2066
2067         lock_sock(sk);
2068
2069         switch (optname) {
2070         case TCP_MAXSEG:
2071                 /* Values greater than interface MTU won't take effect. However
2072                  * at the point when this call is done we typically don't yet
2073                  * know which interface is going to be used */
2074                 if (val < 8 || val > MAX_TCP_WINDOW) {
2075                         err = -EINVAL;
2076                         break;
2077                 }
2078                 tp->rx_opt.user_mss = val;
2079                 break;
2080
2081         case TCP_NODELAY:
2082                 if (val) {
2083                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2084                          * this option on corked socket is remembered, but
2085                          * it is not activated until cork is cleared.
2086                          *
2087                          * However, when TCP_NODELAY is set we make
2088                          * an explicit push, which overrides even TCP_CORK
2089                          * for currently queued segments.
2090                          */
2091                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2092                         tcp_push_pending_frames(sk);
2093                 } else {
2094                         tp->nonagle &= ~TCP_NAGLE_OFF;
2095                 }
2096                 break;
2097
2098         case TCP_CORK:
2099                 /* When set indicates to always queue non-full frames.
2100                  * Later the user clears this option and we transmit
2101                  * any pending partial frames in the queue.  This is
2102                  * meant to be used alongside sendfile() to get properly
2103                  * filled frames when the user (for example) must write
2104                  * out headers with a write() call first and then use
2105                  * sendfile to send out the data parts.
2106                  *
2107                  * TCP_CORK can be set together with TCP_NODELAY and it is
2108                  * stronger than TCP_NODELAY.
2109                  */
2110                 if (val) {
2111                         tp->nonagle |= TCP_NAGLE_CORK;
2112                 } else {
2113                         tp->nonagle &= ~TCP_NAGLE_CORK;
2114                         if (tp->nonagle&TCP_NAGLE_OFF)
2115                                 tp->nonagle |= TCP_NAGLE_PUSH;
2116                         tcp_push_pending_frames(sk);
2117                 }
2118                 break;
2119
2120         case TCP_KEEPIDLE:
2121                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2122                         err = -EINVAL;
2123                 else {
2124                         tp->keepalive_time = val * HZ;
2125                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2126                             !((1 << sk->sk_state) &
2127                               (TCPF_CLOSE | TCPF_LISTEN))) {
2128                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2129                                 if (tp->keepalive_time > elapsed)
2130                                         elapsed = tp->keepalive_time - elapsed;
2131                                 else
2132                                         elapsed = 0;
2133                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2134                         }
2135                 }
2136                 break;
2137         case TCP_KEEPINTVL:
2138                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2139                         err = -EINVAL;
2140                 else
2141                         tp->keepalive_intvl = val * HZ;
2142                 break;
2143         case TCP_KEEPCNT:
2144                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2145                         err = -EINVAL;
2146                 else
2147                         tp->keepalive_probes = val;
2148                 break;
2149         case TCP_SYNCNT:
2150                 if (val < 1 || val > MAX_TCP_SYNCNT)
2151                         err = -EINVAL;
2152                 else
2153                         icsk->icsk_syn_retries = val;
2154                 break;
2155
2156         case TCP_LINGER2:
2157                 if (val < 0)
2158                         tp->linger2 = -1;
2159                 else if (val > sysctl_tcp_fin_timeout / HZ)
2160                         tp->linger2 = 0;
2161                 else
2162                         tp->linger2 = val * HZ;
2163                 break;
2164
2165         case TCP_DEFER_ACCEPT:
2166                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
2167                 if (val > 0) {
2168                         /* Translate value in seconds to number of
2169                          * retransmits */
2170                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
2171                                val > ((TCP_TIMEOUT_INIT / HZ) <<
2172                                        icsk->icsk_accept_queue.rskq_defer_accept))
2173                                 icsk->icsk_accept_queue.rskq_defer_accept++;
2174                         icsk->icsk_accept_queue.rskq_defer_accept++;
2175                 }
2176                 break;
2177
2178         case TCP_WINDOW_CLAMP:
2179                 if (!val) {
2180                         if (sk->sk_state != TCP_CLOSE) {
2181                                 err = -EINVAL;
2182                                 break;
2183                         }
2184                         tp->window_clamp = 0;
2185                 } else
2186                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2187                                                 SOCK_MIN_RCVBUF / 2 : val;
2188                 break;
2189
2190         case TCP_QUICKACK:
2191                 if (!val) {
2192                         icsk->icsk_ack.pingpong = 1;
2193                 } else {
2194                         icsk->icsk_ack.pingpong = 0;
2195                         if ((1 << sk->sk_state) &
2196                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2197                             inet_csk_ack_scheduled(sk)) {
2198                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2199                                 tcp_cleanup_rbuf(sk, 1);
2200                                 if (!(val & 1))
2201                                         icsk->icsk_ack.pingpong = 1;
2202                         }
2203                 }
2204                 break;
2205
2206 #ifdef CONFIG_TCP_MD5SIG
2207         case TCP_MD5SIG:
2208                 /* Read the IP->Key mappings from userspace */
2209                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2210                 break;
2211 #endif
2212
2213         default:
2214                 err = -ENOPROTOOPT;
2215                 break;
2216         }
2217
2218         release_sock(sk);
2219         return err;
2220 }
2221
2222 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2223                    int optlen)
2224 {
2225         struct inet_connection_sock *icsk = inet_csk(sk);
2226
2227         if (level != SOL_TCP)
2228                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2229                                                      optval, optlen);
2230         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2231 }
2232
2233 #ifdef CONFIG_COMPAT
2234 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2235                           char __user *optval, int optlen)
2236 {
2237         if (level != SOL_TCP)
2238                 return inet_csk_compat_setsockopt(sk, level, optname,
2239                                                   optval, optlen);
2240         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2241 }
2242
2243 EXPORT_SYMBOL(compat_tcp_setsockopt);
2244 #endif
2245
2246 /* Return information about state of tcp endpoint in API format. */
2247 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2248 {
2249         struct tcp_sock *tp = tcp_sk(sk);
2250         const struct inet_connection_sock *icsk = inet_csk(sk);
2251         u32 now = tcp_time_stamp;
2252
2253         memset(info, 0, sizeof(*info));
2254
2255         info->tcpi_state = sk->sk_state;
2256         info->tcpi_ca_state = icsk->icsk_ca_state;
2257         info->tcpi_retransmits = icsk->icsk_retransmits;
2258         info->tcpi_probes = icsk->icsk_probes_out;
2259         info->tcpi_backoff = icsk->icsk_backoff;
2260
2261         if (tp->rx_opt.tstamp_ok)
2262                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2263         if (tcp_is_sack(tp))
2264                 info->tcpi_options |= TCPI_OPT_SACK;
2265         if (tp->rx_opt.wscale_ok) {
2266                 info->tcpi_options |= TCPI_OPT_WSCALE;
2267                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2268                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2269         }
2270
2271         if (tp->ecn_flags&TCP_ECN_OK)
2272                 info->tcpi_options |= TCPI_OPT_ECN;
2273
2274         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2275         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2276         info->tcpi_snd_mss = tp->mss_cache;
2277         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2278
2279         if (sk->sk_state == TCP_LISTEN) {
2280                 info->tcpi_unacked = sk->sk_ack_backlog;
2281                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2282         } else {
2283                 info->tcpi_unacked = tp->packets_out;
2284                 info->tcpi_sacked = tp->sacked_out;
2285         }
2286         info->tcpi_lost = tp->lost_out;
2287         info->tcpi_retrans = tp->retrans_out;
2288         info->tcpi_fackets = tp->fackets_out;
2289
2290         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2291         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2292         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2293
2294         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2295         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2296         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2297         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2298         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2299         info->tcpi_snd_cwnd = tp->snd_cwnd;
2300         info->tcpi_advmss = tp->advmss;
2301         info->tcpi_reordering = tp->reordering;
2302
2303         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2304         info->tcpi_rcv_space = tp->rcvq_space.space;
2305
2306         info->tcpi_total_retrans = tp->total_retrans;
2307 }
2308
2309 EXPORT_SYMBOL_GPL(tcp_get_info);
2310
2311 static int do_tcp_getsockopt(struct sock *sk, int level,
2312                 int optname, char __user *optval, int __user *optlen)
2313 {
2314         struct inet_connection_sock *icsk = inet_csk(sk);
2315         struct tcp_sock *tp = tcp_sk(sk);
2316         int val, len;
2317
2318         if (get_user(len, optlen))
2319                 return -EFAULT;
2320
2321         len = min_t(unsigned int, len, sizeof(int));
2322
2323         if (len < 0)
2324                 return -EINVAL;
2325
2326         switch (optname) {
2327         case TCP_MAXSEG:
2328                 val = tp->mss_cache;
2329                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2330                         val = tp->rx_opt.user_mss;
2331                 break;
2332         case TCP_NODELAY:
2333                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2334                 break;
2335         case TCP_CORK:
2336                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2337                 break;
2338         case TCP_KEEPIDLE:
2339                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2340                 break;
2341         case TCP_KEEPINTVL:
2342                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2343                 break;
2344         case TCP_KEEPCNT:
2345                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2346                 break;
2347         case TCP_SYNCNT:
2348                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2349                 break;
2350         case TCP_LINGER2:
2351                 val = tp->linger2;
2352                 if (val >= 0)
2353                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2354                 break;
2355         case TCP_DEFER_ACCEPT:
2356                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2357                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2358                 break;
2359         case TCP_WINDOW_CLAMP:
2360                 val = tp->window_clamp;
2361                 break;
2362         case TCP_INFO: {
2363                 struct tcp_info info;
2364
2365                 if (get_user(len, optlen))
2366                         return -EFAULT;
2367
2368                 tcp_get_info(sk, &info);
2369
2370                 len = min_t(unsigned int, len, sizeof(info));
2371                 if (put_user(len, optlen))
2372                         return -EFAULT;
2373                 if (copy_to_user(optval, &info, len))
2374                         return -EFAULT;
2375                 return 0;
2376         }
2377         case TCP_QUICKACK:
2378                 val = !icsk->icsk_ack.pingpong;
2379                 break;
2380
2381         case TCP_CONGESTION:
2382                 if (get_user(len, optlen))
2383                         return -EFAULT;
2384                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2385                 if (put_user(len, optlen))
2386                         return -EFAULT;
2387                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2388                         return -EFAULT;
2389                 return 0;
2390         default:
2391                 return -ENOPROTOOPT;
2392         }
2393
2394         if (put_user(len, optlen))
2395                 return -EFAULT;
2396         if (copy_to_user(optval, &val, len))
2397                 return -EFAULT;
2398         return 0;
2399 }
2400
2401 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2402                    int __user *optlen)
2403 {
2404         struct inet_connection_sock *icsk = inet_csk(sk);
2405
2406         if (level != SOL_TCP)
2407                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2408                                                      optval, optlen);
2409         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2410 }
2411
2412 #ifdef CONFIG_COMPAT
2413 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2414                           char __user *optval, int __user *optlen)
2415 {
2416         if (level != SOL_TCP)
2417                 return inet_csk_compat_getsockopt(sk, level, optname,
2418                                                   optval, optlen);
2419         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2420 }
2421
2422 EXPORT_SYMBOL(compat_tcp_getsockopt);
2423 #endif
2424
2425 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2426 {
2427         struct sk_buff *segs = ERR_PTR(-EINVAL);
2428         struct tcphdr *th;
2429         unsigned thlen;
2430         unsigned int seq;
2431         __be32 delta;
2432         unsigned int oldlen;
2433         unsigned int mss;
2434
2435         if (!pskb_may_pull(skb, sizeof(*th)))
2436                 goto out;
2437
2438         th = tcp_hdr(skb);
2439         thlen = th->doff * 4;
2440         if (thlen < sizeof(*th))
2441                 goto out;
2442
2443         if (!pskb_may_pull(skb, thlen))
2444                 goto out;
2445
2446         oldlen = (u16)~skb->len;
2447         __skb_pull(skb, thlen);
2448
2449         mss = skb_shinfo(skb)->gso_size;
2450         if (unlikely(skb->len <= mss))
2451                 goto out;
2452
2453         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2454                 /* Packet is from an untrusted source, reset gso_segs. */
2455                 int type = skb_shinfo(skb)->gso_type;
2456
2457                 if (unlikely(type &
2458                              ~(SKB_GSO_TCPV4 |
2459                                SKB_GSO_DODGY |
2460                                SKB_GSO_TCP_ECN |
2461                                SKB_GSO_TCPV6 |
2462                                0) ||
2463                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2464                         goto out;
2465
2466                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2467
2468                 segs = NULL;
2469                 goto out;
2470         }
2471
2472         segs = skb_segment(skb, features);
2473         if (IS_ERR(segs))
2474                 goto out;
2475
2476         delta = htonl(oldlen + (thlen + mss));
2477
2478         skb = segs;
2479         th = tcp_hdr(skb);
2480         seq = ntohl(th->seq);
2481
2482         do {
2483                 th->fin = th->psh = 0;
2484
2485                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2486                                        (__force u32)delta));
2487                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2488                         th->check =
2489                              csum_fold(csum_partial(skb_transport_header(skb),
2490                                                     thlen, skb->csum));
2491
2492                 seq += mss;
2493                 skb = skb->next;
2494                 th = tcp_hdr(skb);
2495
2496                 th->seq = htonl(seq);
2497                 th->cwr = 0;
2498         } while (skb->next);
2499
2500         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2501                       skb->data_len);
2502         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2503                                 (__force u32)delta));
2504         if (skb->ip_summed != CHECKSUM_PARTIAL)
2505                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2506                                                    thlen, skb->csum));
2507
2508 out:
2509         return segs;
2510 }
2511 EXPORT_SYMBOL(tcp_tso_segment);
2512
2513 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2514 {
2515         struct sk_buff **pp = NULL;
2516         struct sk_buff *p;
2517         struct tcphdr *th;
2518         struct tcphdr *th2;
2519         unsigned int len;
2520         unsigned int thlen;
2521         unsigned int flags;
2522         unsigned int mss = 1;
2523         unsigned int hlen;
2524         unsigned int off;
2525         int flush = 1;
2526         int i;
2527
2528         off = skb_gro_offset(skb);
2529         hlen = off + sizeof(*th);
2530         th = skb_gro_header_fast(skb, off);
2531         if (skb_gro_header_hard(skb, hlen)) {
2532                 th = skb_gro_header_slow(skb, hlen, off);
2533                 if (unlikely(!th))
2534                         goto out;
2535         }
2536
2537         thlen = th->doff * 4;
2538         if (thlen < sizeof(*th))
2539                 goto out;
2540
2541         hlen = off + thlen;
2542         if (skb_gro_header_hard(skb, hlen)) {
2543                 th = skb_gro_header_slow(skb, hlen, off);
2544                 if (unlikely(!th))
2545                         goto out;
2546         }
2547
2548         skb_gro_pull(skb, thlen);
2549
2550         len = skb_gro_len(skb);
2551         flags = tcp_flag_word(th);
2552
2553         for (; (p = *head); head = &p->next) {
2554                 if (!NAPI_GRO_CB(p)->same_flow)
2555                         continue;
2556
2557                 th2 = tcp_hdr(p);
2558
2559                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2560                         NAPI_GRO_CB(p)->same_flow = 0;
2561                         continue;
2562                 }
2563
2564                 goto found;
2565         }
2566
2567         goto out_check_final;
2568
2569 found:
2570         flush = NAPI_GRO_CB(p)->flush;
2571         flush |= flags & TCP_FLAG_CWR;
2572         flush |= (flags ^ tcp_flag_word(th2)) &
2573                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH);
2574         flush |= th->ack_seq ^ th2->ack_seq;
2575         for (i = sizeof(*th); i < thlen; i += 4)
2576                 flush |= *(u32 *)((u8 *)th + i) ^
2577                          *(u32 *)((u8 *)th2 + i);
2578
2579         mss = skb_shinfo(p)->gso_size;
2580
2581         flush |= (len - 1) >= mss;
2582         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2583
2584         if (flush || skb_gro_receive(head, skb)) {
2585                 mss = 1;
2586                 goto out_check_final;
2587         }
2588
2589         p = *head;
2590         th2 = tcp_hdr(p);
2591         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2592
2593 out_check_final:
2594         flush = len < mss;
2595         flush |= flags & (TCP_FLAG_URG | TCP_FLAG_PSH | TCP_FLAG_RST |
2596                           TCP_FLAG_SYN | TCP_FLAG_FIN);
2597
2598         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2599                 pp = head;
2600
2601 out:
2602         NAPI_GRO_CB(skb)->flush |= flush;
2603
2604         return pp;
2605 }
2606 EXPORT_SYMBOL(tcp_gro_receive);
2607
2608 int tcp_gro_complete(struct sk_buff *skb)
2609 {
2610         struct tcphdr *th = tcp_hdr(skb);
2611
2612         skb->csum_start = skb_transport_header(skb) - skb->head;
2613         skb->csum_offset = offsetof(struct tcphdr, check);
2614         skb->ip_summed = CHECKSUM_PARTIAL;
2615
2616         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2617
2618         if (th->cwr)
2619                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2620
2621         return 0;
2622 }
2623 EXPORT_SYMBOL(tcp_gro_complete);
2624
2625 #ifdef CONFIG_TCP_MD5SIG
2626 static unsigned long tcp_md5sig_users;
2627 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2628 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2629
2630 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2631 {
2632         int cpu;
2633         for_each_possible_cpu(cpu) {
2634                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2635                 if (p) {
2636                         if (p->md5_desc.tfm)
2637                                 crypto_free_hash(p->md5_desc.tfm);
2638                         kfree(p);
2639                         p = NULL;
2640                 }
2641         }
2642         free_percpu(pool);
2643 }
2644
2645 void tcp_free_md5sig_pool(void)
2646 {
2647         struct tcp_md5sig_pool **pool = NULL;
2648
2649         spin_lock_bh(&tcp_md5sig_pool_lock);
2650         if (--tcp_md5sig_users == 0) {
2651                 pool = tcp_md5sig_pool;
2652                 tcp_md5sig_pool = NULL;
2653         }
2654         spin_unlock_bh(&tcp_md5sig_pool_lock);
2655         if (pool)
2656                 __tcp_free_md5sig_pool(pool);
2657 }
2658
2659 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2660
2661 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2662 {
2663         int cpu;
2664         struct tcp_md5sig_pool **pool;
2665
2666         pool = alloc_percpu(struct tcp_md5sig_pool *);
2667         if (!pool)
2668                 return NULL;
2669
2670         for_each_possible_cpu(cpu) {
2671                 struct tcp_md5sig_pool *p;
2672                 struct crypto_hash *hash;
2673
2674                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2675                 if (!p)
2676                         goto out_free;
2677                 *per_cpu_ptr(pool, cpu) = p;
2678
2679                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2680                 if (!hash || IS_ERR(hash))
2681                         goto out_free;
2682
2683                 p->md5_desc.tfm = hash;
2684         }
2685         return pool;
2686 out_free:
2687         __tcp_free_md5sig_pool(pool);
2688         return NULL;
2689 }
2690
2691 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2692 {
2693         struct tcp_md5sig_pool **pool;
2694         int alloc = 0;
2695
2696 retry:
2697         spin_lock_bh(&tcp_md5sig_pool_lock);
2698         pool = tcp_md5sig_pool;
2699         if (tcp_md5sig_users++ == 0) {
2700                 alloc = 1;
2701                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2702         } else if (!pool) {
2703                 tcp_md5sig_users--;
2704                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2705                 cpu_relax();
2706                 goto retry;
2707         } else
2708                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2709
2710         if (alloc) {
2711                 /* we cannot hold spinlock here because this may sleep. */
2712                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2713                 spin_lock_bh(&tcp_md5sig_pool_lock);
2714                 if (!p) {
2715                         tcp_md5sig_users--;
2716                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2717                         return NULL;
2718                 }
2719                 pool = tcp_md5sig_pool;
2720                 if (pool) {
2721                         /* oops, it has already been assigned. */
2722                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2723                         __tcp_free_md5sig_pool(p);
2724                 } else {
2725                         tcp_md5sig_pool = pool = p;
2726                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2727                 }
2728         }
2729         return pool;
2730 }
2731
2732 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2733
2734 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2735 {
2736         struct tcp_md5sig_pool **p;
2737         spin_lock_bh(&tcp_md5sig_pool_lock);
2738         p = tcp_md5sig_pool;
2739         if (p)
2740                 tcp_md5sig_users++;
2741         spin_unlock_bh(&tcp_md5sig_pool_lock);
2742         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2743 }
2744
2745 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2746
2747 void __tcp_put_md5sig_pool(void)
2748 {
2749         tcp_free_md5sig_pool();
2750 }
2751
2752 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2753
2754 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2755                         struct tcphdr *th)
2756 {
2757         struct scatterlist sg;
2758         int err;
2759
2760         __sum16 old_checksum = th->check;
2761         th->check = 0;
2762         /* options aren't included in the hash */
2763         sg_init_one(&sg, th, sizeof(struct tcphdr));
2764         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2765         th->check = old_checksum;
2766         return err;
2767 }
2768
2769 EXPORT_SYMBOL(tcp_md5_hash_header);
2770
2771 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2772                           struct sk_buff *skb, unsigned header_len)
2773 {
2774         struct scatterlist sg;
2775         const struct tcphdr *tp = tcp_hdr(skb);
2776         struct hash_desc *desc = &hp->md5_desc;
2777         unsigned i;
2778         const unsigned head_data_len = skb_headlen(skb) > header_len ?
2779                                        skb_headlen(skb) - header_len : 0;
2780         const struct skb_shared_info *shi = skb_shinfo(skb);
2781
2782         sg_init_table(&sg, 1);
2783
2784         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2785         if (crypto_hash_update(desc, &sg, head_data_len))
2786                 return 1;
2787
2788         for (i = 0; i < shi->nr_frags; ++i) {
2789                 const struct skb_frag_struct *f = &shi->frags[i];
2790                 sg_set_page(&sg, f->page, f->size, f->page_offset);
2791                 if (crypto_hash_update(desc, &sg, f->size))
2792                         return 1;
2793         }
2794
2795         return 0;
2796 }
2797
2798 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2799
2800 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
2801 {
2802         struct scatterlist sg;
2803
2804         sg_init_one(&sg, key->key, key->keylen);
2805         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2806 }
2807
2808 EXPORT_SYMBOL(tcp_md5_hash_key);
2809
2810 #endif
2811
2812 void tcp_done(struct sock *sk)
2813 {
2814         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2815                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2816
2817         tcp_set_state(sk, TCP_CLOSE);
2818         tcp_clear_xmit_timers(sk);
2819
2820         sk->sk_shutdown = SHUTDOWN_MASK;
2821
2822         if (!sock_flag(sk, SOCK_DEAD))
2823                 sk->sk_state_change(sk);
2824         else
2825                 inet_csk_destroy_sock(sk);
2826 }
2827 EXPORT_SYMBOL_GPL(tcp_done);
2828
2829 extern struct tcp_congestion_ops tcp_reno;
2830
2831 static __initdata unsigned long thash_entries;
2832 static int __init set_thash_entries(char *str)
2833 {
2834         if (!str)
2835                 return 0;
2836         thash_entries = simple_strtoul(str, &str, 0);
2837         return 1;
2838 }
2839 __setup("thash_entries=", set_thash_entries);
2840
2841 void __init tcp_init(void)
2842 {
2843         struct sk_buff *skb = NULL;
2844         unsigned long nr_pages, limit;
2845         int order, i, max_share;
2846
2847         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2848
2849         percpu_counter_init(&tcp_sockets_allocated, 0);
2850         percpu_counter_init(&tcp_orphan_count, 0);
2851         tcp_hashinfo.bind_bucket_cachep =
2852                 kmem_cache_create("tcp_bind_bucket",
2853                                   sizeof(struct inet_bind_bucket), 0,
2854                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2855
2856         /* Size and allocate the main established and bind bucket
2857          * hash tables.
2858          *
2859          * The methodology is similar to that of the buffer cache.
2860          */
2861         tcp_hashinfo.ehash =
2862                 alloc_large_system_hash("TCP established",
2863                                         sizeof(struct inet_ehash_bucket),
2864                                         thash_entries,
2865                                         (num_physpages >= 128 * 1024) ?
2866                                         13 : 15,
2867                                         0,
2868                                         &tcp_hashinfo.ehash_size,
2869                                         NULL,
2870                                         thash_entries ? 0 : 512 * 1024);
2871         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2872         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2873                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
2874                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
2875         }
2876         if (inet_ehash_locks_alloc(&tcp_hashinfo))
2877                 panic("TCP: failed to alloc ehash_locks");
2878         tcp_hashinfo.bhash =
2879                 alloc_large_system_hash("TCP bind",
2880                                         sizeof(struct inet_bind_hashbucket),
2881                                         tcp_hashinfo.ehash_size,
2882                                         (num_physpages >= 128 * 1024) ?
2883                                         13 : 15,
2884                                         0,
2885                                         &tcp_hashinfo.bhash_size,
2886                                         NULL,
2887                                         64 * 1024);
2888         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2889         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2890                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2891                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2892         }
2893
2894         /* Try to be a bit smarter and adjust defaults depending
2895          * on available memory.
2896          */
2897         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2898                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2899                         order++)
2900                 ;
2901         if (order >= 4) {
2902                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2903                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2904                 sysctl_max_syn_backlog = 1024;
2905         } else if (order < 3) {
2906                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2907                 sysctl_tcp_max_orphans >>= (3 - order);
2908                 sysctl_max_syn_backlog = 128;
2909         }
2910
2911         /* Set the pressure threshold to be a fraction of global memory that
2912          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2913          * memory, with a floor of 128 pages.
2914          */
2915         nr_pages = totalram_pages - totalhigh_pages;
2916         limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2917         limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2918         limit = max(limit, 128UL);
2919         sysctl_tcp_mem[0] = limit / 4 * 3;
2920         sysctl_tcp_mem[1] = limit;
2921         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2922
2923         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2924         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2925         max_share = min(4UL*1024*1024, limit);
2926
2927         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2928         sysctl_tcp_wmem[1] = 16*1024;
2929         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2930
2931         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2932         sysctl_tcp_rmem[1] = 87380;
2933         sysctl_tcp_rmem[2] = max(87380, max_share);
2934
2935         printk(KERN_INFO "TCP: Hash tables configured "
2936                "(established %d bind %d)\n",
2937                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2938
2939         tcp_register_congestion_control(&tcp_reno);
2940 }
2941
2942 EXPORT_SYMBOL(tcp_close);
2943 EXPORT_SYMBOL(tcp_disconnect);
2944 EXPORT_SYMBOL(tcp_getsockopt);
2945 EXPORT_SYMBOL(tcp_ioctl);
2946 EXPORT_SYMBOL(tcp_poll);
2947 EXPORT_SYMBOL(tcp_read_sock);
2948 EXPORT_SYMBOL(tcp_recvmsg);
2949 EXPORT_SYMBOL(tcp_sendmsg);
2950 EXPORT_SYMBOL(tcp_splice_read);
2951 EXPORT_SYMBOL(tcp_sendpage);
2952 EXPORT_SYMBOL(tcp_setsockopt);
2953 EXPORT_SYMBOL(tcp_shutdown);