Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[linux-2.6.git] / net / ipv4 / tcp.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Authors:     Ross Biro
9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
12  *              Florian La Roche, <flla@stud.uni-sb.de>
13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *              Jorge Cwik, <jorge@laser.satlink.net>
19  *
20  * Fixes:
21  *              Alan Cox        :       Numerous verify_area() calls
22  *              Alan Cox        :       Set the ACK bit on a reset
23  *              Alan Cox        :       Stopped it crashing if it closed while
24  *                                      sk->inuse=1 and was trying to connect
25  *                                      (tcp_err()).
26  *              Alan Cox        :       All icmp error handling was broken
27  *                                      pointers passed where wrong and the
28  *                                      socket was looked up backwards. Nobody
29  *                                      tested any icmp error code obviously.
30  *              Alan Cox        :       tcp_err() now handled properly. It
31  *                                      wakes people on errors. poll
32  *                                      behaves and the icmp error race
33  *                                      has gone by moving it into sock.c
34  *              Alan Cox        :       tcp_send_reset() fixed to work for
35  *                                      everything not just packets for
36  *                                      unknown sockets.
37  *              Alan Cox        :       tcp option processing.
38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
39  *                                      syn rule wrong]
40  *              Herp Rosmanith  :       More reset fixes
41  *              Alan Cox        :       No longer acks invalid rst frames.
42  *                                      Acking any kind of RST is right out.
43  *              Alan Cox        :       Sets an ignore me flag on an rst
44  *                                      receive otherwise odd bits of prattle
45  *                                      escape still
46  *              Alan Cox        :       Fixed another acking RST frame bug.
47  *                                      Should stop LAN workplace lockups.
48  *              Alan Cox        :       Some tidyups using the new skb list
49  *                                      facilities
50  *              Alan Cox        :       sk->keepopen now seems to work
51  *              Alan Cox        :       Pulls options out correctly on accepts
52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
54  *                                      bit to skb ops.
55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
56  *                                      nasty.
57  *              Alan Cox        :       Added some better commenting, as the
58  *                                      tcp is hard to follow
59  *              Alan Cox        :       Removed incorrect check for 20 * psh
60  *      Michael O'Reilly        :       ack < copied bug fix.
61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
62  *              Alan Cox        :       FIN with no memory -> CRASH
63  *              Alan Cox        :       Added socket option proto entries.
64  *                                      Also added awareness of them to accept.
65  *              Alan Cox        :       Added TCP options (SOL_TCP)
66  *              Alan Cox        :       Switched wakeup calls to callbacks,
67  *                                      so the kernel can layer network
68  *                                      sockets.
69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
71  *              Alan Cox        :       RST frames sent on unsynchronised
72  *                                      state ack error.
73  *              Alan Cox        :       Put in missing check for SYN bit.
74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
75  *                                      window non shrink trick.
76  *              Alan Cox        :       Added a couple of small NET2E timer
77  *                                      fixes
78  *              Charles Hedrick :       TCP fixes
79  *              Toomas Tamm     :       TCP window fixes
80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
81  *              Charles Hedrick :       Rewrote most of it to actually work
82  *              Linus           :       Rewrote tcp_read() and URG handling
83  *                                      completely
84  *              Gerhard Koerting:       Fixed some missing timer handling
85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
86  *              Gerhard Koerting:       PC/TCP workarounds
87  *              Adam Caldwell   :       Assorted timer/timing errors
88  *              Matthew Dillon  :       Fixed another RST bug
89  *              Alan Cox        :       Move to kernel side addressing changes.
90  *              Alan Cox        :       Beginning work on TCP fastpathing
91  *                                      (not yet usable)
92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
93  *              Alan Cox        :       TCP fast path debugging
94  *              Alan Cox        :       Window clamping
95  *              Michael Riepe   :       Bug in tcp_check()
96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
97  *              Matt Dillon     :       Yet more small nasties remove from the
98  *                                      TCP code (Be very nice to this man if
99  *                                      tcp finally works 100%) 8)
100  *              Alan Cox        :       BSD accept semantics.
101  *              Alan Cox        :       Reset on closedown bug.
102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
103  *              Michael Pall    :       Handle poll() after URG properly in
104  *                                      all cases.
105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
106  *                                      (multi URG PUSH broke rlogin).
107  *              Michael Pall    :       Fix the multi URG PUSH problem in
108  *                                      tcp_readable(), poll() after URG
109  *                                      works now.
110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
111  *                                      BSD api.
112  *              Alan Cox        :       Changed the semantics of sk->socket to
113  *                                      fix a race and a signal problem with
114  *                                      accept() and async I/O.
115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
118  *                                      clients/servers which listen in on
119  *                                      fixed ports.
120  *              Alan Cox        :       Cleaned the above up and shrank it to
121  *                                      a sensible code size.
122  *              Alan Cox        :       Self connect lockup fix.
123  *              Alan Cox        :       No connect to multicast.
124  *              Ross Biro       :       Close unaccepted children on master
125  *                                      socket close.
126  *              Alan Cox        :       Reset tracing code.
127  *              Alan Cox        :       Spurious resets on shutdown.
128  *              Alan Cox        :       Giant 15 minute/60 second timer error
129  *              Alan Cox        :       Small whoops in polling before an
130  *                                      accept.
131  *              Alan Cox        :       Kept the state trace facility since
132  *                                      it's handy for debugging.
133  *              Alan Cox        :       More reset handler fixes.
134  *              Alan Cox        :       Started rewriting the code based on
135  *                                      the RFC's for other useful protocol
136  *                                      references see: Comer, KA9Q NOS, and
137  *                                      for a reference on the difference
138  *                                      between specifications and how BSD
139  *                                      works see the 4.4lite source.
140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
141  *                                      close.
142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
144  *              Alan Cox        :       Reimplemented timers as per the RFC
145  *                                      and using multiple timers for sanity.
146  *              Alan Cox        :       Small bug fixes, and a lot of new
147  *                                      comments.
148  *              Alan Cox        :       Fixed dual reader crash by locking
149  *                                      the buffers (much like datagram.c)
150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
151  *                                      now gets fed up of retrying without
152  *                                      (even a no space) answer.
153  *              Alan Cox        :       Extracted closing code better
154  *              Alan Cox        :       Fixed the closing state machine to
155  *                                      resemble the RFC.
156  *              Alan Cox        :       More 'per spec' fixes.
157  *              Jorge Cwik      :       Even faster checksumming.
158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
159  *                                      only frames. At least one pc tcp stack
160  *                                      generates them.
161  *              Alan Cox        :       Cache last socket.
162  *              Alan Cox        :       Per route irtt.
163  *              Matt Day        :       poll()->select() match BSD precisely on error
164  *              Alan Cox        :       New buffers
165  *              Marc Tamsky     :       Various sk->prot->retransmits and
166  *                                      sk->retransmits misupdating fixed.
167  *                                      Fixed tcp_write_timeout: stuck close,
168  *                                      and TCP syn retries gets used now.
169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
170  *                                      ack if state is TCP_CLOSED.
171  *              Alan Cox        :       Look up device on a retransmit - routes may
172  *                                      change. Doesn't yet cope with MSS shrink right
173  *                                      but it's a start!
174  *              Marc Tamsky     :       Closing in closing fixes.
175  *              Mike Shaver     :       RFC1122 verifications.
176  *              Alan Cox        :       rcv_saddr errors.
177  *              Alan Cox        :       Block double connect().
178  *              Alan Cox        :       Small hooks for enSKIP.
179  *              Alexey Kuznetsov:       Path MTU discovery.
180  *              Alan Cox        :       Support soft errors.
181  *              Alan Cox        :       Fix MTU discovery pathological case
182  *                                      when the remote claims no mtu!
183  *              Marc Tamsky     :       TCP_CLOSE fix.
184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
185  *                                      window but wrong (fixes NT lpd problems)
186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
187  *              Joerg Reuter    :       No modification of locked buffers in
188  *                                      tcp_do_retransmit()
189  *              Eric Schenk     :       Changed receiver side silly window
190  *                                      avoidance algorithm to BSD style
191  *                                      algorithm. This doubles throughput
192  *                                      against machines running Solaris,
193  *                                      and seems to result in general
194  *                                      improvement.
195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
196  *      Willy Konynenberg       :       Transparent proxying support.
197  *      Mike McLagan            :       Routing by source
198  *              Keith Owens     :       Do proper merging with partial SKB's in
199  *                                      tcp_do_sendmsg to avoid burstiness.
200  *              Eric Schenk     :       Fix fast close down bug with
201  *                                      shutdown() followed by close().
202  *              Andi Kleen      :       Make poll agree with SIGIO
203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
204  *                                      lingertime == 0 (RFC 793 ABORT Call)
205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
206  *                                      csum_and_copy_from_user() if possible.
207  *
208  *              This program is free software; you can redistribute it and/or
209  *              modify it under the terms of the GNU General Public License
210  *              as published by the Free Software Foundation; either version
211  *              2 of the License, or(at your option) any later version.
212  *
213  * Description of States:
214  *
215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
216  *
217  *      TCP_SYN_RECV            received a connection request, sent ack,
218  *                              waiting for final ack in three-way handshake.
219  *
220  *      TCP_ESTABLISHED         connection established
221  *
222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
223  *                              transmission of remaining buffered data
224  *
225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
226  *                              to shutdown
227  *
228  *      TCP_CLOSING             both sides have shutdown but we still have
229  *                              data we have to finish sending
230  *
231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
232  *                              closed, can only be entered from FIN_WAIT2
233  *                              or CLOSING.  Required because the other end
234  *                              may not have gotten our last ACK causing it
235  *                              to retransmit the data packet (which we ignore)
236  *
237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
238  *                              us to finish writing our data and to shutdown
239  *                              (we have to close() to move on to LAST_ACK)
240  *
241  *      TCP_LAST_ACK            out side has shutdown after remote has
242  *                              shutdown.  There may still be data in our
243  *                              buffer that we have to finish sending
244  *
245  *      TCP_CLOSE               socket is finished
246  */
247
248 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
267
268 #include <net/icmp.h>
269 #include <net/tcp.h>
270 #include <net/xfrm.h>
271 #include <net/ip.h>
272 #include <net/netdma.h>
273 #include <net/sock.h>
274
275 #include <asm/uaccess.h>
276 #include <asm/ioctls.h>
277
278 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
279
280 struct percpu_counter tcp_orphan_count;
281 EXPORT_SYMBOL_GPL(tcp_orphan_count);
282
283 int sysctl_tcp_mem[3] __read_mostly;
284 int sysctl_tcp_wmem[3] __read_mostly;
285 int sysctl_tcp_rmem[3] __read_mostly;
286
287 EXPORT_SYMBOL(sysctl_tcp_mem);
288 EXPORT_SYMBOL(sysctl_tcp_rmem);
289 EXPORT_SYMBOL(sysctl_tcp_wmem);
290
291 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
292 EXPORT_SYMBOL(tcp_memory_allocated);
293
294 /*
295  * Current number of TCP sockets.
296  */
297 struct percpu_counter tcp_sockets_allocated;
298 EXPORT_SYMBOL(tcp_sockets_allocated);
299
300 /*
301  * TCP splice context
302  */
303 struct tcp_splice_state {
304         struct pipe_inode_info *pipe;
305         size_t len;
306         unsigned int flags;
307 };
308
309 /*
310  * Pressure flag: try to collapse.
311  * Technical note: it is used by multiple contexts non atomically.
312  * All the __sk_mem_schedule() is of this nature: accounting
313  * is strict, actions are advisory and have some latency.
314  */
315 int tcp_memory_pressure __read_mostly;
316
317 EXPORT_SYMBOL(tcp_memory_pressure);
318
319 void tcp_enter_memory_pressure(struct sock *sk)
320 {
321         if (!tcp_memory_pressure) {
322                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
323                 tcp_memory_pressure = 1;
324         }
325 }
326
327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
328
329 /*
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         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                         answ = tp->rcv_nxt - tp->copied_seq;
443
444                         /* Subtract 1, if FIN is in queue. */
445                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
446                                 answ -=
447                        tcp_hdr((struct sk_buff *)sk->sk_receive_queue.prev)->fin;
448                 } else
449                         answ = tp->urg_seq - tp->copied_seq;
450                 release_sock(sk);
451                 break;
452         case SIOCATMARK:
453                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
454                 break;
455         case SIOCOUTQ:
456                 if (sk->sk_state == TCP_LISTEN)
457                         return -EINVAL;
458
459                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
460                         answ = 0;
461                 else
462                         answ = tp->write_seq - tp->snd_una;
463                 break;
464         default:
465                 return -ENOIOCTLCMD;
466         }
467
468         return put_user(answ, (int __user *)arg);
469 }
470
471 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
472 {
473         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
474         tp->pushed_seq = tp->write_seq;
475 }
476
477 static inline int forced_push(struct tcp_sock *tp)
478 {
479         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
480 }
481
482 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
483 {
484         struct tcp_sock *tp = tcp_sk(sk);
485         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
486
487         skb->csum    = 0;
488         tcb->seq     = tcb->end_seq = tp->write_seq;
489         tcb->flags   = TCPCB_FLAG_ACK;
490         tcb->sacked  = 0;
491         skb_header_release(skb);
492         tcp_add_write_queue_tail(sk, skb);
493         sk->sk_wmem_queued += skb->truesize;
494         sk_mem_charge(sk, skb->truesize);
495         if (tp->nonagle & TCP_NAGLE_PUSH)
496                 tp->nonagle &= ~TCP_NAGLE_PUSH;
497 }
498
499 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
500                                 struct sk_buff *skb)
501 {
502         if (flags & MSG_OOB)
503                 tp->snd_up = tp->write_seq;
504 }
505
506 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
507                             int nonagle)
508 {
509         struct tcp_sock *tp = tcp_sk(sk);
510
511         if (tcp_send_head(sk)) {
512                 struct sk_buff *skb = tcp_write_queue_tail(sk);
513                 if (!(flags & MSG_MORE) || forced_push(tp))
514                         tcp_mark_push(tp, skb);
515                 tcp_mark_urg(tp, flags, skb);
516                 __tcp_push_pending_frames(sk, mss_now,
517                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
518         }
519 }
520
521 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
522                                 unsigned int offset, size_t len)
523 {
524         struct tcp_splice_state *tss = rd_desc->arg.data;
525         int ret;
526
527         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
528                               tss->flags);
529         if (ret > 0)
530                 rd_desc->count -= ret;
531         return ret;
532 }
533
534 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
535 {
536         /* Store TCP splice context information in read_descriptor_t. */
537         read_descriptor_t rd_desc = {
538                 .arg.data = tss,
539                 .count    = tss->len,
540         };
541
542         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
543 }
544
545 /**
546  *  tcp_splice_read - splice data from TCP socket to a pipe
547  * @sock:       socket to splice from
548  * @ppos:       position (not valid)
549  * @pipe:       pipe to splice to
550  * @len:        number of bytes to splice
551  * @flags:      splice modifier flags
552  *
553  * Description:
554  *    Will read pages from given socket and fill them into a pipe.
555  *
556  **/
557 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
558                         struct pipe_inode_info *pipe, size_t len,
559                         unsigned int flags)
560 {
561         struct sock *sk = sock->sk;
562         struct tcp_splice_state tss = {
563                 .pipe = pipe,
564                 .len = len,
565                 .flags = flags,
566         };
567         long timeo;
568         ssize_t spliced;
569         int ret;
570
571         /*
572          * We can't seek on a socket input
573          */
574         if (unlikely(*ppos))
575                 return -ESPIPE;
576
577         ret = spliced = 0;
578
579         lock_sock(sk);
580
581         timeo = sock_rcvtimeo(sk, flags & SPLICE_F_NONBLOCK);
582         while (tss.len) {
583                 ret = __tcp_splice_read(sk, &tss);
584                 if (ret < 0)
585                         break;
586                 else if (!ret) {
587                         if (spliced)
588                                 break;
589                         if (sock_flag(sk, SOCK_DONE))
590                                 break;
591                         if (sk->sk_err) {
592                                 ret = sock_error(sk);
593                                 break;
594                         }
595                         if (sk->sk_shutdown & RCV_SHUTDOWN)
596                                 break;
597                         if (sk->sk_state == TCP_CLOSE) {
598                                 /*
599                                  * This occurs when user tries to read
600                                  * from never connected socket.
601                                  */
602                                 if (!sock_flag(sk, SOCK_DONE))
603                                         ret = -ENOTCONN;
604                                 break;
605                         }
606                         if (!timeo) {
607                                 ret = -EAGAIN;
608                                 break;
609                         }
610                         sk_wait_data(sk, &timeo);
611                         if (signal_pending(current)) {
612                                 ret = sock_intr_errno(timeo);
613                                 break;
614                         }
615                         continue;
616                 }
617                 tss.len -= ret;
618                 spliced += ret;
619
620                 if (!timeo)
621                         break;
622                 release_sock(sk);
623                 lock_sock(sk);
624
625                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
626                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
627                     signal_pending(current))
628                         break;
629         }
630
631         release_sock(sk);
632
633         if (spliced)
634                 return spliced;
635
636         return ret;
637 }
638
639 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
640 {
641         struct sk_buff *skb;
642
643         /* The TCP header must be at least 32-bit aligned.  */
644         size = ALIGN(size, 4);
645
646         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
647         if (skb) {
648                 if (sk_wmem_schedule(sk, skb->truesize)) {
649                         /*
650                          * Make sure that we have exactly size bytes
651                          * available to the caller, no more, no less.
652                          */
653                         skb_reserve(skb, skb_tailroom(skb) - size);
654                         return skb;
655                 }
656                 __kfree_skb(skb);
657         } else {
658                 sk->sk_prot->enter_memory_pressure(sk);
659                 sk_stream_moderate_sndbuf(sk);
660         }
661         return NULL;
662 }
663
664 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
665                          size_t psize, int flags)
666 {
667         struct tcp_sock *tp = tcp_sk(sk);
668         int mss_now, size_goal;
669         int err;
670         ssize_t copied;
671         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
672
673         /* Wait for a connection to finish. */
674         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
675                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
676                         goto out_err;
677
678         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
679
680         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
681         size_goal = tp->xmit_size_goal;
682         copied = 0;
683
684         err = -EPIPE;
685         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
686                 goto do_error;
687
688         while (psize > 0) {
689                 struct sk_buff *skb = tcp_write_queue_tail(sk);
690                 struct page *page = pages[poffset / PAGE_SIZE];
691                 int copy, i, can_coalesce;
692                 int offset = poffset % PAGE_SIZE;
693                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
694
695                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
696 new_segment:
697                         if (!sk_stream_memory_free(sk))
698                                 goto wait_for_sndbuf;
699
700                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
701                         if (!skb)
702                                 goto wait_for_memory;
703
704                         skb_entail(sk, skb);
705                         copy = size_goal;
706                 }
707
708                 if (copy > size)
709                         copy = size;
710
711                 i = skb_shinfo(skb)->nr_frags;
712                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
713                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
714                         tcp_mark_push(tp, skb);
715                         goto new_segment;
716                 }
717                 if (!sk_wmem_schedule(sk, copy))
718                         goto wait_for_memory;
719
720                 if (can_coalesce) {
721                         skb_shinfo(skb)->frags[i - 1].size += copy;
722                 } else {
723                         get_page(page);
724                         skb_fill_page_desc(skb, i, page, offset, copy);
725                 }
726
727                 skb->len += copy;
728                 skb->data_len += copy;
729                 skb->truesize += copy;
730                 sk->sk_wmem_queued += copy;
731                 sk_mem_charge(sk, copy);
732                 skb->ip_summed = CHECKSUM_PARTIAL;
733                 tp->write_seq += copy;
734                 TCP_SKB_CB(skb)->end_seq += copy;
735                 skb_shinfo(skb)->gso_segs = 0;
736
737                 if (!copied)
738                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
739
740                 copied += copy;
741                 poffset += copy;
742                 if (!(psize -= copy))
743                         goto out;
744
745                 if (skb->len < size_goal || (flags & MSG_OOB))
746                         continue;
747
748                 if (forced_push(tp)) {
749                         tcp_mark_push(tp, skb);
750                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
751                 } else if (skb == tcp_send_head(sk))
752                         tcp_push_one(sk, mss_now);
753                 continue;
754
755 wait_for_sndbuf:
756                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
757 wait_for_memory:
758                 if (copied)
759                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
760
761                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
762                         goto do_error;
763
764                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
765                 size_goal = tp->xmit_size_goal;
766         }
767
768 out:
769         if (copied)
770                 tcp_push(sk, flags, mss_now, tp->nonagle);
771         return copied;
772
773 do_error:
774         if (copied)
775                 goto out;
776 out_err:
777         return sk_stream_error(sk, flags, err);
778 }
779
780 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
781                      size_t size, int flags)
782 {
783         ssize_t res;
784         struct sock *sk = sock->sk;
785
786         if (!(sk->sk_route_caps & NETIF_F_SG) ||
787             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
788                 return sock_no_sendpage(sock, page, offset, size, flags);
789
790         lock_sock(sk);
791         TCP_CHECK_TIMER(sk);
792         res = do_tcp_sendpages(sk, &page, offset, size, flags);
793         TCP_CHECK_TIMER(sk);
794         release_sock(sk);
795         return res;
796 }
797
798 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
799 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
800
801 static inline int select_size(struct sock *sk)
802 {
803         struct tcp_sock *tp = tcp_sk(sk);
804         int tmp = tp->mss_cache;
805
806         if (sk->sk_route_caps & NETIF_F_SG) {
807                 if (sk_can_gso(sk))
808                         tmp = 0;
809                 else {
810                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
811
812                         if (tmp >= pgbreak &&
813                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
814                                 tmp = pgbreak;
815                 }
816         }
817
818         return tmp;
819 }
820
821 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
822                 size_t size)
823 {
824         struct sock *sk = sock->sk;
825         struct iovec *iov;
826         struct tcp_sock *tp = tcp_sk(sk);
827         struct sk_buff *skb;
828         int iovlen, flags;
829         int mss_now, size_goal;
830         int err, copied;
831         long timeo;
832
833         lock_sock(sk);
834         TCP_CHECK_TIMER(sk);
835
836         flags = msg->msg_flags;
837         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
838
839         /* Wait for a connection to finish. */
840         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
841                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
842                         goto out_err;
843
844         /* This should be in poll */
845         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
846
847         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
848         size_goal = tp->xmit_size_goal;
849
850         /* Ok commence sending. */
851         iovlen = msg->msg_iovlen;
852         iov = msg->msg_iov;
853         copied = 0;
854
855         err = -EPIPE;
856         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
857                 goto do_error;
858
859         while (--iovlen >= 0) {
860                 int seglen = iov->iov_len;
861                 unsigned char __user *from = iov->iov_base;
862
863                 iov++;
864
865                 while (seglen > 0) {
866                         int copy;
867
868                         skb = tcp_write_queue_tail(sk);
869
870                         if (!tcp_send_head(sk) ||
871                             (copy = size_goal - skb->len) <= 0) {
872
873 new_segment:
874                                 /* Allocate new segment. If the interface is SG,
875                                  * allocate skb fitting to single page.
876                                  */
877                                 if (!sk_stream_memory_free(sk))
878                                         goto wait_for_sndbuf;
879
880                                 skb = sk_stream_alloc_skb(sk, select_size(sk),
881                                                 sk->sk_allocation);
882                                 if (!skb)
883                                         goto wait_for_memory;
884
885                                 /*
886                                  * Check whether we can use HW checksum.
887                                  */
888                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
889                                         skb->ip_summed = CHECKSUM_PARTIAL;
890
891                                 skb_entail(sk, skb);
892                                 copy = size_goal;
893                         }
894
895                         /* Try to append data to the end of skb. */
896                         if (copy > seglen)
897                                 copy = seglen;
898
899                         /* Where to copy to? */
900                         if (skb_tailroom(skb) > 0) {
901                                 /* We have some space in skb head. Superb! */
902                                 if (copy > skb_tailroom(skb))
903                                         copy = skb_tailroom(skb);
904                                 if ((err = skb_add_data(skb, from, copy)) != 0)
905                                         goto do_fault;
906                         } else {
907                                 int merge = 0;
908                                 int i = skb_shinfo(skb)->nr_frags;
909                                 struct page *page = TCP_PAGE(sk);
910                                 int off = TCP_OFF(sk);
911
912                                 if (skb_can_coalesce(skb, i, page, off) &&
913                                     off != PAGE_SIZE) {
914                                         /* We can extend the last page
915                                          * fragment. */
916                                         merge = 1;
917                                 } else if (i == MAX_SKB_FRAGS ||
918                                            (!i &&
919                                            !(sk->sk_route_caps & NETIF_F_SG))) {
920                                         /* Need to add new fragment and cannot
921                                          * do this because interface is non-SG,
922                                          * or because all the page slots are
923                                          * busy. */
924                                         tcp_mark_push(tp, skb);
925                                         goto new_segment;
926                                 } else if (page) {
927                                         if (off == PAGE_SIZE) {
928                                                 put_page(page);
929                                                 TCP_PAGE(sk) = page = NULL;
930                                                 off = 0;
931                                         }
932                                 } else
933                                         off = 0;
934
935                                 if (copy > PAGE_SIZE - off)
936                                         copy = PAGE_SIZE - off;
937
938                                 if (!sk_wmem_schedule(sk, copy))
939                                         goto wait_for_memory;
940
941                                 if (!page) {
942                                         /* Allocate new cache page. */
943                                         if (!(page = sk_stream_alloc_page(sk)))
944                                                 goto wait_for_memory;
945                                 }
946
947                                 /* Time to copy data. We are close to
948                                  * the end! */
949                                 err = skb_copy_to_page(sk, from, skb, page,
950                                                        off, copy);
951                                 if (err) {
952                                         /* If this page was new, give it to the
953                                          * socket so it does not get leaked.
954                                          */
955                                         if (!TCP_PAGE(sk)) {
956                                                 TCP_PAGE(sk) = page;
957                                                 TCP_OFF(sk) = 0;
958                                         }
959                                         goto do_error;
960                                 }
961
962                                 /* Update the skb. */
963                                 if (merge) {
964                                         skb_shinfo(skb)->frags[i - 1].size +=
965                                                                         copy;
966                                 } else {
967                                         skb_fill_page_desc(skb, i, page, off, copy);
968                                         if (TCP_PAGE(sk)) {
969                                                 get_page(page);
970                                         } else if (off + copy < PAGE_SIZE) {
971                                                 get_page(page);
972                                                 TCP_PAGE(sk) = page;
973                                         }
974                                 }
975
976                                 TCP_OFF(sk) = off + copy;
977                         }
978
979                         if (!copied)
980                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
981
982                         tp->write_seq += copy;
983                         TCP_SKB_CB(skb)->end_seq += copy;
984                         skb_shinfo(skb)->gso_segs = 0;
985
986                         from += copy;
987                         copied += copy;
988                         if ((seglen -= copy) == 0 && iovlen == 0)
989                                 goto out;
990
991                         if (skb->len < size_goal || (flags & MSG_OOB))
992                                 continue;
993
994                         if (forced_push(tp)) {
995                                 tcp_mark_push(tp, skb);
996                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
997                         } else if (skb == tcp_send_head(sk))
998                                 tcp_push_one(sk, mss_now);
999                         continue;
1000
1001 wait_for_sndbuf:
1002                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1003 wait_for_memory:
1004                         if (copied)
1005                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1006
1007                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1008                                 goto do_error;
1009
1010                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
1011                         size_goal = tp->xmit_size_goal;
1012                 }
1013         }
1014
1015 out:
1016         if (copied)
1017                 tcp_push(sk, flags, mss_now, tp->nonagle);
1018         TCP_CHECK_TIMER(sk);
1019         release_sock(sk);
1020         return copied;
1021
1022 do_fault:
1023         if (!skb->len) {
1024                 tcp_unlink_write_queue(skb, sk);
1025                 /* It is the one place in all of TCP, except connection
1026                  * reset, where we can be unlinking the send_head.
1027                  */
1028                 tcp_check_send_head(sk, skb);
1029                 sk_wmem_free_skb(sk, skb);
1030         }
1031
1032 do_error:
1033         if (copied)
1034                 goto out;
1035 out_err:
1036         err = sk_stream_error(sk, flags, err);
1037         TCP_CHECK_TIMER(sk);
1038         release_sock(sk);
1039         return err;
1040 }
1041
1042 /*
1043  *      Handle reading urgent data. BSD has very simple semantics for
1044  *      this, no blocking and very strange errors 8)
1045  */
1046
1047 static int tcp_recv_urg(struct sock *sk, long timeo,
1048                         struct msghdr *msg, int len, int flags,
1049                         int *addr_len)
1050 {
1051         struct tcp_sock *tp = tcp_sk(sk);
1052
1053         /* No URG data to read. */
1054         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1055             tp->urg_data == TCP_URG_READ)
1056                 return -EINVAL; /* Yes this is right ! */
1057
1058         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1059                 return -ENOTCONN;
1060
1061         if (tp->urg_data & TCP_URG_VALID) {
1062                 int err = 0;
1063                 char c = tp->urg_data;
1064
1065                 if (!(flags & MSG_PEEK))
1066                         tp->urg_data = TCP_URG_READ;
1067
1068                 /* Read urgent data. */
1069                 msg->msg_flags |= MSG_OOB;
1070
1071                 if (len > 0) {
1072                         if (!(flags & MSG_TRUNC))
1073                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1074                         len = 1;
1075                 } else
1076                         msg->msg_flags |= MSG_TRUNC;
1077
1078                 return err ? -EFAULT : len;
1079         }
1080
1081         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1082                 return 0;
1083
1084         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1085          * the available implementations agree in this case:
1086          * this call should never block, independent of the
1087          * blocking state of the socket.
1088          * Mike <pall@rz.uni-karlsruhe.de>
1089          */
1090         return -EAGAIN;
1091 }
1092
1093 /* Clean up the receive buffer for full frames taken by the user,
1094  * then send an ACK if necessary.  COPIED is the number of bytes
1095  * tcp_recvmsg has given to the user so far, it speeds up the
1096  * calculation of whether or not we must ACK for the sake of
1097  * a window update.
1098  */
1099 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1100 {
1101         struct tcp_sock *tp = tcp_sk(sk);
1102         int time_to_ack = 0;
1103
1104 #if TCP_DEBUG
1105         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1106
1107         WARN_ON(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1108 #endif
1109
1110         if (inet_csk_ack_scheduled(sk)) {
1111                 const struct inet_connection_sock *icsk = inet_csk(sk);
1112                    /* Delayed ACKs frequently hit locked sockets during bulk
1113                     * receive. */
1114                 if (icsk->icsk_ack.blocked ||
1115                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1116                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1117                     /*
1118                      * If this read emptied read buffer, we send ACK, if
1119                      * connection is not bidirectional, user drained
1120                      * receive buffer and there was a small segment
1121                      * in queue.
1122                      */
1123                     (copied > 0 &&
1124                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1125                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1126                        !icsk->icsk_ack.pingpong)) &&
1127                       !atomic_read(&sk->sk_rmem_alloc)))
1128                         time_to_ack = 1;
1129         }
1130
1131         /* We send an ACK if we can now advertise a non-zero window
1132          * which has been raised "significantly".
1133          *
1134          * Even if window raised up to infinity, do not send window open ACK
1135          * in states, where we will not receive more. It is useless.
1136          */
1137         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1138                 __u32 rcv_window_now = tcp_receive_window(tp);
1139
1140                 /* Optimize, __tcp_select_window() is not cheap. */
1141                 if (2*rcv_window_now <= tp->window_clamp) {
1142                         __u32 new_window = __tcp_select_window(sk);
1143
1144                         /* Send ACK now, if this read freed lots of space
1145                          * in our buffer. Certainly, new_window is new window.
1146                          * We can advertise it now, if it is not less than current one.
1147                          * "Lots" means "at least twice" here.
1148                          */
1149                         if (new_window && new_window >= 2 * rcv_window_now)
1150                                 time_to_ack = 1;
1151                 }
1152         }
1153         if (time_to_ack)
1154                 tcp_send_ack(sk);
1155 }
1156
1157 static void tcp_prequeue_process(struct sock *sk)
1158 {
1159         struct sk_buff *skb;
1160         struct tcp_sock *tp = tcp_sk(sk);
1161
1162         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1163
1164         /* RX process wants to run with disabled BHs, though it is not
1165          * necessary */
1166         local_bh_disable();
1167         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1168                 sk_backlog_rcv(sk, skb);
1169         local_bh_enable();
1170
1171         /* Clear memory counter. */
1172         tp->ucopy.memory = 0;
1173 }
1174
1175 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1176 {
1177         struct sk_buff *skb;
1178         u32 offset;
1179
1180         skb_queue_walk(&sk->sk_receive_queue, skb) {
1181                 offset = seq - TCP_SKB_CB(skb)->seq;
1182                 if (tcp_hdr(skb)->syn)
1183                         offset--;
1184                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1185                         *off = offset;
1186                         return skb;
1187                 }
1188         }
1189         return NULL;
1190 }
1191
1192 /*
1193  * This routine provides an alternative to tcp_recvmsg() for routines
1194  * that would like to handle copying from skbuffs directly in 'sendfile'
1195  * fashion.
1196  * Note:
1197  *      - It is assumed that the socket was locked by the caller.
1198  *      - The routine does not block.
1199  *      - At present, there is no support for reading OOB data
1200  *        or for 'peeking' the socket using this routine
1201  *        (although both would be easy to implement).
1202  */
1203 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1204                   sk_read_actor_t recv_actor)
1205 {
1206         struct sk_buff *skb;
1207         struct tcp_sock *tp = tcp_sk(sk);
1208         u32 seq = tp->copied_seq;
1209         u32 offset;
1210         int copied = 0;
1211
1212         if (sk->sk_state == TCP_LISTEN)
1213                 return -ENOTCONN;
1214         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1215                 if (offset < skb->len) {
1216                         int used;
1217                         size_t len;
1218
1219                         len = skb->len - offset;
1220                         /* Stop reading if we hit a patch of urgent data */
1221                         if (tp->urg_data) {
1222                                 u32 urg_offset = tp->urg_seq - seq;
1223                                 if (urg_offset < len)
1224                                         len = urg_offset;
1225                                 if (!len)
1226                                         break;
1227                         }
1228                         used = recv_actor(desc, skb, offset, len);
1229                         if (used < 0) {
1230                                 if (!copied)
1231                                         copied = used;
1232                                 break;
1233                         } else if (used <= len) {
1234                                 seq += used;
1235                                 copied += used;
1236                                 offset += used;
1237                         }
1238                         /*
1239                          * If recv_actor drops the lock (e.g. TCP splice
1240                          * receive) the skb pointer might be invalid when
1241                          * getting here: tcp_collapse might have deleted it
1242                          * while aggregating skbs from the socket queue.
1243                          */
1244                         skb = tcp_recv_skb(sk, seq-1, &offset);
1245                         if (!skb || (offset+1 != skb->len))
1246                                 break;
1247                 }
1248                 if (tcp_hdr(skb)->fin) {
1249                         sk_eat_skb(sk, skb, 0);
1250                         ++seq;
1251                         break;
1252                 }
1253                 sk_eat_skb(sk, skb, 0);
1254                 if (!desc->count)
1255                         break;
1256         }
1257         tp->copied_seq = seq;
1258
1259         tcp_rcv_space_adjust(sk);
1260
1261         /* Clean up data we have read: This will do ACK frames. */
1262         if (copied > 0)
1263                 tcp_cleanup_rbuf(sk, copied);
1264         return copied;
1265 }
1266
1267 /*
1268  *      This routine copies from a sock struct into the user buffer.
1269  *
1270  *      Technical note: in 2.3 we work on _locked_ socket, so that
1271  *      tricks with *seq access order and skb->users are not required.
1272  *      Probably, code can be easily improved even more.
1273  */
1274
1275 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1276                 size_t len, int nonblock, int flags, int *addr_len)
1277 {
1278         struct tcp_sock *tp = tcp_sk(sk);
1279         int copied = 0;
1280         u32 peek_seq;
1281         u32 *seq;
1282         unsigned long used;
1283         int err;
1284         int target;             /* Read at least this many bytes */
1285         long timeo;
1286         struct task_struct *user_recv = NULL;
1287         int copied_early = 0;
1288         struct sk_buff *skb;
1289
1290         lock_sock(sk);
1291
1292         TCP_CHECK_TIMER(sk);
1293
1294         err = -ENOTCONN;
1295         if (sk->sk_state == TCP_LISTEN)
1296                 goto out;
1297
1298         timeo = sock_rcvtimeo(sk, nonblock);
1299
1300         /* Urgent data needs to be handled specially. */
1301         if (flags & MSG_OOB)
1302                 goto recv_urg;
1303
1304         seq = &tp->copied_seq;
1305         if (flags & MSG_PEEK) {
1306                 peek_seq = tp->copied_seq;
1307                 seq = &peek_seq;
1308         }
1309
1310         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1311
1312 #ifdef CONFIG_NET_DMA
1313         tp->ucopy.dma_chan = NULL;
1314         preempt_disable();
1315         skb = skb_peek_tail(&sk->sk_receive_queue);
1316         {
1317                 int available = 0;
1318
1319                 if (skb)
1320                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1321                 if ((available < target) &&
1322                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1323                     !sysctl_tcp_low_latency &&
1324                     dma_find_channel(DMA_MEMCPY)) {
1325                         preempt_enable_no_resched();
1326                         tp->ucopy.pinned_list =
1327                                         dma_pin_iovec_pages(msg->msg_iov, len);
1328                 } else {
1329                         preempt_enable_no_resched();
1330                 }
1331         }
1332 #endif
1333
1334         do {
1335                 u32 offset;
1336
1337                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1338                 if (tp->urg_data && tp->urg_seq == *seq) {
1339                         if (copied)
1340                                 break;
1341                         if (signal_pending(current)) {
1342                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1343                                 break;
1344                         }
1345                 }
1346
1347                 /* Next get a buffer. */
1348
1349                 skb = skb_peek(&sk->sk_receive_queue);
1350                 do {
1351                         if (!skb)
1352                                 break;
1353
1354                         /* Now that we have two receive queues this
1355                          * shouldn't happen.
1356                          */
1357                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1358                                 printk(KERN_INFO "recvmsg bug: copied %X "
1359                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1360                                 break;
1361                         }
1362                         offset = *seq - TCP_SKB_CB(skb)->seq;
1363                         if (tcp_hdr(skb)->syn)
1364                                 offset--;
1365                         if (offset < skb->len)
1366                                 goto found_ok_skb;
1367                         if (tcp_hdr(skb)->fin)
1368                                 goto found_fin_ok;
1369                         WARN_ON(!(flags & MSG_PEEK));
1370                         skb = skb->next;
1371                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1372
1373                 /* Well, if we have backlog, try to process it now yet. */
1374
1375                 if (copied >= target && !sk->sk_backlog.tail)
1376                         break;
1377
1378                 if (copied) {
1379                         if (sk->sk_err ||
1380                             sk->sk_state == TCP_CLOSE ||
1381                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1382                             !timeo ||
1383                             signal_pending(current))
1384                                 break;
1385                 } else {
1386                         if (sock_flag(sk, SOCK_DONE))
1387                                 break;
1388
1389                         if (sk->sk_err) {
1390                                 copied = sock_error(sk);
1391                                 break;
1392                         }
1393
1394                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1395                                 break;
1396
1397                         if (sk->sk_state == TCP_CLOSE) {
1398                                 if (!sock_flag(sk, SOCK_DONE)) {
1399                                         /* This occurs when user tries to read
1400                                          * from never connected socket.
1401                                          */
1402                                         copied = -ENOTCONN;
1403                                         break;
1404                                 }
1405                                 break;
1406                         }
1407
1408                         if (!timeo) {
1409                                 copied = -EAGAIN;
1410                                 break;
1411                         }
1412
1413                         if (signal_pending(current)) {
1414                                 copied = sock_intr_errno(timeo);
1415                                 break;
1416                         }
1417                 }
1418
1419                 tcp_cleanup_rbuf(sk, copied);
1420
1421                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1422                         /* Install new reader */
1423                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1424                                 user_recv = current;
1425                                 tp->ucopy.task = user_recv;
1426                                 tp->ucopy.iov = msg->msg_iov;
1427                         }
1428
1429                         tp->ucopy.len = len;
1430
1431                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1432                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1433
1434                         /* Ugly... If prequeue is not empty, we have to
1435                          * process it before releasing socket, otherwise
1436                          * order will be broken at second iteration.
1437                          * More elegant solution is required!!!
1438                          *
1439                          * Look: we have the following (pseudo)queues:
1440                          *
1441                          * 1. packets in flight
1442                          * 2. backlog
1443                          * 3. prequeue
1444                          * 4. receive_queue
1445                          *
1446                          * Each queue can be processed only if the next ones
1447                          * are empty. At this point we have empty receive_queue.
1448                          * But prequeue _can_ be not empty after 2nd iteration,
1449                          * when we jumped to start of loop because backlog
1450                          * processing added something to receive_queue.
1451                          * We cannot release_sock(), because backlog contains
1452                          * packets arrived _after_ prequeued ones.
1453                          *
1454                          * Shortly, algorithm is clear --- to process all
1455                          * the queues in order. We could make it more directly,
1456                          * requeueing packets from backlog to prequeue, if
1457                          * is not empty. It is more elegant, but eats cycles,
1458                          * unfortunately.
1459                          */
1460                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1461                                 goto do_prequeue;
1462
1463                         /* __ Set realtime policy in scheduler __ */
1464                 }
1465
1466                 if (copied >= target) {
1467                         /* Do not sleep, just process backlog. */
1468                         release_sock(sk);
1469                         lock_sock(sk);
1470                 } else
1471                         sk_wait_data(sk, &timeo);
1472
1473 #ifdef CONFIG_NET_DMA
1474                 tp->ucopy.wakeup = 0;
1475 #endif
1476
1477                 if (user_recv) {
1478                         int chunk;
1479
1480                         /* __ Restore normal policy in scheduler __ */
1481
1482                         if ((chunk = len - tp->ucopy.len) != 0) {
1483                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1484                                 len -= chunk;
1485                                 copied += chunk;
1486                         }
1487
1488                         if (tp->rcv_nxt == tp->copied_seq &&
1489                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1490 do_prequeue:
1491                                 tcp_prequeue_process(sk);
1492
1493                                 if ((chunk = len - tp->ucopy.len) != 0) {
1494                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1495                                         len -= chunk;
1496                                         copied += chunk;
1497                                 }
1498                         }
1499                 }
1500                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1501                         if (net_ratelimit())
1502                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1503                                        current->comm, task_pid_nr(current));
1504                         peek_seq = tp->copied_seq;
1505                 }
1506                 continue;
1507
1508         found_ok_skb:
1509                 /* Ok so how much can we use? */
1510                 used = skb->len - offset;
1511                 if (len < used)
1512                         used = len;
1513
1514                 /* Do we have urgent data here? */
1515                 if (tp->urg_data) {
1516                         u32 urg_offset = tp->urg_seq - *seq;
1517                         if (urg_offset < used) {
1518                                 if (!urg_offset) {
1519                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1520                                                 ++*seq;
1521                                                 offset++;
1522                                                 used--;
1523                                                 if (!used)
1524                                                         goto skip_copy;
1525                                         }
1526                                 } else
1527                                         used = urg_offset;
1528                         }
1529                 }
1530
1531                 if (!(flags & MSG_TRUNC)) {
1532 #ifdef CONFIG_NET_DMA
1533                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1534                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1535
1536                         if (tp->ucopy.dma_chan) {
1537                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1538                                         tp->ucopy.dma_chan, skb, offset,
1539                                         msg->msg_iov, used,
1540                                         tp->ucopy.pinned_list);
1541
1542                                 if (tp->ucopy.dma_cookie < 0) {
1543
1544                                         printk(KERN_ALERT "dma_cookie < 0\n");
1545
1546                                         /* Exception. Bailout! */
1547                                         if (!copied)
1548                                                 copied = -EFAULT;
1549                                         break;
1550                                 }
1551                                 if ((offset + used) == skb->len)
1552                                         copied_early = 1;
1553
1554                         } else
1555 #endif
1556                         {
1557                                 err = skb_copy_datagram_iovec(skb, offset,
1558                                                 msg->msg_iov, used);
1559                                 if (err) {
1560                                         /* Exception. Bailout! */
1561                                         if (!copied)
1562                                                 copied = -EFAULT;
1563                                         break;
1564                                 }
1565                         }
1566                 }
1567
1568                 *seq += used;
1569                 copied += used;
1570                 len -= used;
1571
1572                 tcp_rcv_space_adjust(sk);
1573
1574 skip_copy:
1575                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1576                         tp->urg_data = 0;
1577                         tcp_fast_path_check(sk);
1578                 }
1579                 if (used + offset < skb->len)
1580                         continue;
1581
1582                 if (tcp_hdr(skb)->fin)
1583                         goto found_fin_ok;
1584                 if (!(flags & MSG_PEEK)) {
1585                         sk_eat_skb(sk, skb, copied_early);
1586                         copied_early = 0;
1587                 }
1588                 continue;
1589
1590         found_fin_ok:
1591                 /* Process the FIN. */
1592                 ++*seq;
1593                 if (!(flags & MSG_PEEK)) {
1594                         sk_eat_skb(sk, skb, copied_early);
1595                         copied_early = 0;
1596                 }
1597                 break;
1598         } while (len > 0);
1599
1600         if (user_recv) {
1601                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1602                         int chunk;
1603
1604                         tp->ucopy.len = copied > 0 ? len : 0;
1605
1606                         tcp_prequeue_process(sk);
1607
1608                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1609                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1610                                 len -= chunk;
1611                                 copied += chunk;
1612                         }
1613                 }
1614
1615                 tp->ucopy.task = NULL;
1616                 tp->ucopy.len = 0;
1617         }
1618
1619 #ifdef CONFIG_NET_DMA
1620         if (tp->ucopy.dma_chan) {
1621                 dma_cookie_t done, used;
1622
1623                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1624
1625                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1626                                                  tp->ucopy.dma_cookie, &done,
1627                                                  &used) == DMA_IN_PROGRESS) {
1628                         /* do partial cleanup of sk_async_wait_queue */
1629                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1630                                (dma_async_is_complete(skb->dma_cookie, done,
1631                                                       used) == DMA_SUCCESS)) {
1632                                 __skb_dequeue(&sk->sk_async_wait_queue);
1633                                 kfree_skb(skb);
1634                         }
1635                 }
1636
1637                 /* Safe to free early-copied skbs now */
1638                 __skb_queue_purge(&sk->sk_async_wait_queue);
1639                 tp->ucopy.dma_chan = NULL;
1640         }
1641         if (tp->ucopy.pinned_list) {
1642                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1643                 tp->ucopy.pinned_list = NULL;
1644         }
1645 #endif
1646
1647         /* According to UNIX98, msg_name/msg_namelen are ignored
1648          * on connected socket. I was just happy when found this 8) --ANK
1649          */
1650
1651         /* Clean up data we have read: This will do ACK frames. */
1652         tcp_cleanup_rbuf(sk, copied);
1653
1654         TCP_CHECK_TIMER(sk);
1655         release_sock(sk);
1656         return copied;
1657
1658 out:
1659         TCP_CHECK_TIMER(sk);
1660         release_sock(sk);
1661         return err;
1662
1663 recv_urg:
1664         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1665         goto out;
1666 }
1667
1668 void tcp_set_state(struct sock *sk, int state)
1669 {
1670         int oldstate = sk->sk_state;
1671
1672         switch (state) {
1673         case TCP_ESTABLISHED:
1674                 if (oldstate != TCP_ESTABLISHED)
1675                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1676                 break;
1677
1678         case TCP_CLOSE:
1679                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1680                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1681
1682                 sk->sk_prot->unhash(sk);
1683                 if (inet_csk(sk)->icsk_bind_hash &&
1684                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1685                         inet_put_port(sk);
1686                 /* fall through */
1687         default:
1688                 if (oldstate == TCP_ESTABLISHED)
1689                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1690         }
1691
1692         /* Change state AFTER socket is unhashed to avoid closed
1693          * socket sitting in hash tables.
1694          */
1695         sk->sk_state = state;
1696
1697 #ifdef STATE_TRACE
1698         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1699 #endif
1700 }
1701 EXPORT_SYMBOL_GPL(tcp_set_state);
1702
1703 /*
1704  *      State processing on a close. This implements the state shift for
1705  *      sending our FIN frame. Note that we only send a FIN for some
1706  *      states. A shutdown() may have already sent the FIN, or we may be
1707  *      closed.
1708  */
1709
1710 static const unsigned char new_state[16] = {
1711   /* current state:        new state:      action:      */
1712   /* (Invalid)          */ TCP_CLOSE,
1713   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1714   /* TCP_SYN_SENT       */ TCP_CLOSE,
1715   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1716   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1717   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1718   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1719   /* TCP_CLOSE          */ TCP_CLOSE,
1720   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1721   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1722   /* TCP_LISTEN         */ TCP_CLOSE,
1723   /* TCP_CLOSING        */ TCP_CLOSING,
1724 };
1725
1726 static int tcp_close_state(struct sock *sk)
1727 {
1728         int next = (int)new_state[sk->sk_state];
1729         int ns = next & TCP_STATE_MASK;
1730
1731         tcp_set_state(sk, ns);
1732
1733         return next & TCP_ACTION_FIN;
1734 }
1735
1736 /*
1737  *      Shutdown the sending side of a connection. Much like close except
1738  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1739  */
1740
1741 void tcp_shutdown(struct sock *sk, int how)
1742 {
1743         /*      We need to grab some memory, and put together a FIN,
1744          *      and then put it into the queue to be sent.
1745          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1746          */
1747         if (!(how & SEND_SHUTDOWN))
1748                 return;
1749
1750         /* If we've already sent a FIN, or it's a closed state, skip this. */
1751         if ((1 << sk->sk_state) &
1752             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1753              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1754                 /* Clear out any half completed packets.  FIN if needed. */
1755                 if (tcp_close_state(sk))
1756                         tcp_send_fin(sk);
1757         }
1758 }
1759
1760 void tcp_close(struct sock *sk, long timeout)
1761 {
1762         struct sk_buff *skb;
1763         int data_was_unread = 0;
1764         int state;
1765
1766         lock_sock(sk);
1767         sk->sk_shutdown = SHUTDOWN_MASK;
1768
1769         if (sk->sk_state == TCP_LISTEN) {
1770                 tcp_set_state(sk, TCP_CLOSE);
1771
1772                 /* Special case. */
1773                 inet_csk_listen_stop(sk);
1774
1775                 goto adjudge_to_death;
1776         }
1777
1778         /*  We need to flush the recv. buffs.  We do this only on the
1779          *  descriptor close, not protocol-sourced closes, because the
1780          *  reader process may not have drained the data yet!
1781          */
1782         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1783                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1784                           tcp_hdr(skb)->fin;
1785                 data_was_unread += len;
1786                 __kfree_skb(skb);
1787         }
1788
1789         sk_mem_reclaim(sk);
1790
1791         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1792          * data was lost. To witness the awful effects of the old behavior of
1793          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1794          * GET in an FTP client, suspend the process, wait for the client to
1795          * advertise a zero window, then kill -9 the FTP client, wheee...
1796          * Note: timeout is always zero in such a case.
1797          */
1798         if (data_was_unread) {
1799                 /* Unread data was tossed, zap the connection. */
1800                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1801                 tcp_set_state(sk, TCP_CLOSE);
1802                 tcp_send_active_reset(sk, GFP_KERNEL);
1803         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1804                 /* Check zero linger _after_ checking for unread data. */
1805                 sk->sk_prot->disconnect(sk, 0);
1806                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1807         } else if (tcp_close_state(sk)) {
1808                 /* We FIN if the application ate all the data before
1809                  * zapping the connection.
1810                  */
1811
1812                 /* RED-PEN. Formally speaking, we have broken TCP state
1813                  * machine. State transitions:
1814                  *
1815                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1816                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1817                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1818                  *
1819                  * are legal only when FIN has been sent (i.e. in window),
1820                  * rather than queued out of window. Purists blame.
1821                  *
1822                  * F.e. "RFC state" is ESTABLISHED,
1823                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1824                  *
1825                  * The visible declinations are that sometimes
1826                  * we enter time-wait state, when it is not required really
1827                  * (harmless), do not send active resets, when they are
1828                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1829                  * they look as CLOSING or LAST_ACK for Linux)
1830                  * Probably, I missed some more holelets.
1831                  *                                              --ANK
1832                  */
1833                 tcp_send_fin(sk);
1834         }
1835
1836         sk_stream_wait_close(sk, timeout);
1837
1838 adjudge_to_death:
1839         state = sk->sk_state;
1840         sock_hold(sk);
1841         sock_orphan(sk);
1842
1843         /* It is the last release_sock in its life. It will remove backlog. */
1844         release_sock(sk);
1845
1846
1847         /* Now socket is owned by kernel and we acquire BH lock
1848            to finish close. No need to check for user refs.
1849          */
1850         local_bh_disable();
1851         bh_lock_sock(sk);
1852         WARN_ON(sock_owned_by_user(sk));
1853
1854         percpu_counter_inc(sk->sk_prot->orphan_count);
1855
1856         /* Have we already been destroyed by a softirq or backlog? */
1857         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1858                 goto out;
1859
1860         /*      This is a (useful) BSD violating of the RFC. There is a
1861          *      problem with TCP as specified in that the other end could
1862          *      keep a socket open forever with no application left this end.
1863          *      We use a 3 minute timeout (about the same as BSD) then kill
1864          *      our end. If they send after that then tough - BUT: long enough
1865          *      that we won't make the old 4*rto = almost no time - whoops
1866          *      reset mistake.
1867          *
1868          *      Nope, it was not mistake. It is really desired behaviour
1869          *      f.e. on http servers, when such sockets are useless, but
1870          *      consume significant resources. Let's do it with special
1871          *      linger2 option.                                 --ANK
1872          */
1873
1874         if (sk->sk_state == TCP_FIN_WAIT2) {
1875                 struct tcp_sock *tp = tcp_sk(sk);
1876                 if (tp->linger2 < 0) {
1877                         tcp_set_state(sk, TCP_CLOSE);
1878                         tcp_send_active_reset(sk, GFP_ATOMIC);
1879                         NET_INC_STATS_BH(sock_net(sk),
1880                                         LINUX_MIB_TCPABORTONLINGER);
1881                 } else {
1882                         const int tmo = tcp_fin_time(sk);
1883
1884                         if (tmo > TCP_TIMEWAIT_LEN) {
1885                                 inet_csk_reset_keepalive_timer(sk,
1886                                                 tmo - TCP_TIMEWAIT_LEN);
1887                         } else {
1888                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1889                                 goto out;
1890                         }
1891                 }
1892         }
1893         if (sk->sk_state != TCP_CLOSE) {
1894                 int orphan_count = percpu_counter_read_positive(
1895                                                 sk->sk_prot->orphan_count);
1896
1897                 sk_mem_reclaim(sk);
1898                 if (tcp_too_many_orphans(sk, orphan_count)) {
1899                         if (net_ratelimit())
1900                                 printk(KERN_INFO "TCP: too many of orphaned "
1901                                        "sockets\n");
1902                         tcp_set_state(sk, TCP_CLOSE);
1903                         tcp_send_active_reset(sk, GFP_ATOMIC);
1904                         NET_INC_STATS_BH(sock_net(sk),
1905                                         LINUX_MIB_TCPABORTONMEMORY);
1906                 }
1907         }
1908
1909         if (sk->sk_state == TCP_CLOSE)
1910                 inet_csk_destroy_sock(sk);
1911         /* Otherwise, socket is reprieved until protocol close. */
1912
1913 out:
1914         bh_unlock_sock(sk);
1915         local_bh_enable();
1916         sock_put(sk);
1917 }
1918
1919 /* These states need RST on ABORT according to RFC793 */
1920
1921 static inline int tcp_need_reset(int state)
1922 {
1923         return (1 << state) &
1924                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1925                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1926 }
1927
1928 int tcp_disconnect(struct sock *sk, int flags)
1929 {
1930         struct inet_sock *inet = inet_sk(sk);
1931         struct inet_connection_sock *icsk = inet_csk(sk);
1932         struct tcp_sock *tp = tcp_sk(sk);
1933         int err = 0;
1934         int old_state = sk->sk_state;
1935
1936         if (old_state != TCP_CLOSE)
1937                 tcp_set_state(sk, TCP_CLOSE);
1938
1939         /* ABORT function of RFC793 */
1940         if (old_state == TCP_LISTEN) {
1941                 inet_csk_listen_stop(sk);
1942         } else if (tcp_need_reset(old_state) ||
1943                    (tp->snd_nxt != tp->write_seq &&
1944                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1945                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1946                  * states
1947                  */
1948                 tcp_send_active_reset(sk, gfp_any());
1949                 sk->sk_err = ECONNRESET;
1950         } else if (old_state == TCP_SYN_SENT)
1951                 sk->sk_err = ECONNRESET;
1952
1953         tcp_clear_xmit_timers(sk);
1954         __skb_queue_purge(&sk->sk_receive_queue);
1955         tcp_write_queue_purge(sk);
1956         __skb_queue_purge(&tp->out_of_order_queue);
1957 #ifdef CONFIG_NET_DMA
1958         __skb_queue_purge(&sk->sk_async_wait_queue);
1959 #endif
1960
1961         inet->dport = 0;
1962
1963         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1964                 inet_reset_saddr(sk);
1965
1966         sk->sk_shutdown = 0;
1967         sock_reset_flag(sk, SOCK_DONE);
1968         tp->srtt = 0;
1969         if ((tp->write_seq += tp->max_window + 2) == 0)
1970                 tp->write_seq = 1;
1971         icsk->icsk_backoff = 0;
1972         tp->snd_cwnd = 2;
1973         icsk->icsk_probes_out = 0;
1974         tp->packets_out = 0;
1975         tp->snd_ssthresh = 0x7fffffff;
1976         tp->snd_cwnd_cnt = 0;
1977         tp->bytes_acked = 0;
1978         tcp_set_ca_state(sk, TCP_CA_Open);
1979         tcp_clear_retrans(tp);
1980         inet_csk_delack_init(sk);
1981         tcp_init_send_head(sk);
1982         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1983         __sk_dst_reset(sk);
1984
1985         WARN_ON(inet->num && !icsk->icsk_bind_hash);
1986
1987         sk->sk_error_report(sk);
1988         return err;
1989 }
1990
1991 /*
1992  *      Socket option code for TCP.
1993  */
1994 static int do_tcp_setsockopt(struct sock *sk, int level,
1995                 int optname, char __user *optval, int optlen)
1996 {
1997         struct tcp_sock *tp = tcp_sk(sk);
1998         struct inet_connection_sock *icsk = inet_csk(sk);
1999         int val;
2000         int err = 0;
2001
2002         /* This is a string value all the others are int's */
2003         if (optname == TCP_CONGESTION) {
2004                 char name[TCP_CA_NAME_MAX];
2005
2006                 if (optlen < 1)
2007                         return -EINVAL;
2008
2009                 val = strncpy_from_user(name, optval,
2010                                         min(TCP_CA_NAME_MAX-1, optlen));
2011                 if (val < 0)
2012                         return -EFAULT;
2013                 name[val] = 0;
2014
2015                 lock_sock(sk);
2016                 err = tcp_set_congestion_control(sk, name);
2017                 release_sock(sk);
2018                 return err;
2019         }
2020
2021         if (optlen < sizeof(int))
2022                 return -EINVAL;
2023
2024         if (get_user(val, (int __user *)optval))
2025                 return -EFAULT;
2026
2027         lock_sock(sk);
2028
2029         switch (optname) {
2030         case TCP_MAXSEG:
2031                 /* Values greater than interface MTU won't take effect. However
2032                  * at the point when this call is done we typically don't yet
2033                  * know which interface is going to be used */
2034                 if (val < 8 || val > MAX_TCP_WINDOW) {
2035                         err = -EINVAL;
2036                         break;
2037                 }
2038                 tp->rx_opt.user_mss = val;
2039                 break;
2040
2041         case TCP_NODELAY:
2042                 if (val) {
2043                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2044                          * this option on corked socket is remembered, but
2045                          * it is not activated until cork is cleared.
2046                          *
2047                          * However, when TCP_NODELAY is set we make
2048                          * an explicit push, which overrides even TCP_CORK
2049                          * for currently queued segments.
2050                          */
2051                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2052                         tcp_push_pending_frames(sk);
2053                 } else {
2054                         tp->nonagle &= ~TCP_NAGLE_OFF;
2055                 }
2056                 break;
2057
2058         case TCP_CORK:
2059                 /* When set indicates to always queue non-full frames.
2060                  * Later the user clears this option and we transmit
2061                  * any pending partial frames in the queue.  This is
2062                  * meant to be used alongside sendfile() to get properly
2063                  * filled frames when the user (for example) must write
2064                  * out headers with a write() call first and then use
2065                  * sendfile to send out the data parts.
2066                  *
2067                  * TCP_CORK can be set together with TCP_NODELAY and it is
2068                  * stronger than TCP_NODELAY.
2069                  */
2070                 if (val) {
2071                         tp->nonagle |= TCP_NAGLE_CORK;
2072                 } else {
2073                         tp->nonagle &= ~TCP_NAGLE_CORK;
2074                         if (tp->nonagle&TCP_NAGLE_OFF)
2075                                 tp->nonagle |= TCP_NAGLE_PUSH;
2076                         tcp_push_pending_frames(sk);
2077                 }
2078                 break;
2079
2080         case TCP_KEEPIDLE:
2081                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2082                         err = -EINVAL;
2083                 else {
2084                         tp->keepalive_time = val * HZ;
2085                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2086                             !((1 << sk->sk_state) &
2087                               (TCPF_CLOSE | TCPF_LISTEN))) {
2088                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2089                                 if (tp->keepalive_time > elapsed)
2090                                         elapsed = tp->keepalive_time - elapsed;
2091                                 else
2092                                         elapsed = 0;
2093                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2094                         }
2095                 }
2096                 break;
2097         case TCP_KEEPINTVL:
2098                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2099                         err = -EINVAL;
2100                 else
2101                         tp->keepalive_intvl = val * HZ;
2102                 break;
2103         case TCP_KEEPCNT:
2104                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2105                         err = -EINVAL;
2106                 else
2107                         tp->keepalive_probes = val;
2108                 break;
2109         case TCP_SYNCNT:
2110                 if (val < 1 || val > MAX_TCP_SYNCNT)
2111                         err = -EINVAL;
2112                 else
2113                         icsk->icsk_syn_retries = val;
2114                 break;
2115
2116         case TCP_LINGER2:
2117                 if (val < 0)
2118                         tp->linger2 = -1;
2119                 else if (val > sysctl_tcp_fin_timeout / HZ)
2120                         tp->linger2 = 0;
2121                 else
2122                         tp->linger2 = val * HZ;
2123                 break;
2124
2125         case TCP_DEFER_ACCEPT:
2126                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
2127                 if (val > 0) {
2128                         /* Translate value in seconds to number of
2129                          * retransmits */
2130                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
2131                                val > ((TCP_TIMEOUT_INIT / HZ) <<
2132                                        icsk->icsk_accept_queue.rskq_defer_accept))
2133                                 icsk->icsk_accept_queue.rskq_defer_accept++;
2134                         icsk->icsk_accept_queue.rskq_defer_accept++;
2135                 }
2136                 break;
2137
2138         case TCP_WINDOW_CLAMP:
2139                 if (!val) {
2140                         if (sk->sk_state != TCP_CLOSE) {
2141                                 err = -EINVAL;
2142                                 break;
2143                         }
2144                         tp->window_clamp = 0;
2145                 } else
2146                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2147                                                 SOCK_MIN_RCVBUF / 2 : val;
2148                 break;
2149
2150         case TCP_QUICKACK:
2151                 if (!val) {
2152                         icsk->icsk_ack.pingpong = 1;
2153                 } else {
2154                         icsk->icsk_ack.pingpong = 0;
2155                         if ((1 << sk->sk_state) &
2156                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2157                             inet_csk_ack_scheduled(sk)) {
2158                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2159                                 tcp_cleanup_rbuf(sk, 1);
2160                                 if (!(val & 1))
2161                                         icsk->icsk_ack.pingpong = 1;
2162                         }
2163                 }
2164                 break;
2165
2166 #ifdef CONFIG_TCP_MD5SIG
2167         case TCP_MD5SIG:
2168                 /* Read the IP->Key mappings from userspace */
2169                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2170                 break;
2171 #endif
2172
2173         default:
2174                 err = -ENOPROTOOPT;
2175                 break;
2176         }
2177
2178         release_sock(sk);
2179         return err;
2180 }
2181
2182 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2183                    int optlen)
2184 {
2185         struct inet_connection_sock *icsk = inet_csk(sk);
2186
2187         if (level != SOL_TCP)
2188                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2189                                                      optval, optlen);
2190         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2191 }
2192
2193 #ifdef CONFIG_COMPAT
2194 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2195                           char __user *optval, int optlen)
2196 {
2197         if (level != SOL_TCP)
2198                 return inet_csk_compat_setsockopt(sk, level, optname,
2199                                                   optval, optlen);
2200         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2201 }
2202
2203 EXPORT_SYMBOL(compat_tcp_setsockopt);
2204 #endif
2205
2206 /* Return information about state of tcp endpoint in API format. */
2207 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2208 {
2209         struct tcp_sock *tp = tcp_sk(sk);
2210         const struct inet_connection_sock *icsk = inet_csk(sk);
2211         u32 now = tcp_time_stamp;
2212
2213         memset(info, 0, sizeof(*info));
2214
2215         info->tcpi_state = sk->sk_state;
2216         info->tcpi_ca_state = icsk->icsk_ca_state;
2217         info->tcpi_retransmits = icsk->icsk_retransmits;
2218         info->tcpi_probes = icsk->icsk_probes_out;
2219         info->tcpi_backoff = icsk->icsk_backoff;
2220
2221         if (tp->rx_opt.tstamp_ok)
2222                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2223         if (tcp_is_sack(tp))
2224                 info->tcpi_options |= TCPI_OPT_SACK;
2225         if (tp->rx_opt.wscale_ok) {
2226                 info->tcpi_options |= TCPI_OPT_WSCALE;
2227                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2228                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2229         }
2230
2231         if (tp->ecn_flags&TCP_ECN_OK)
2232                 info->tcpi_options |= TCPI_OPT_ECN;
2233
2234         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2235         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2236         info->tcpi_snd_mss = tp->mss_cache;
2237         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2238
2239         if (sk->sk_state == TCP_LISTEN) {
2240                 info->tcpi_unacked = sk->sk_ack_backlog;
2241                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2242         } else {
2243                 info->tcpi_unacked = tp->packets_out;
2244                 info->tcpi_sacked = tp->sacked_out;
2245         }
2246         info->tcpi_lost = tp->lost_out;
2247         info->tcpi_retrans = tp->retrans_out;
2248         info->tcpi_fackets = tp->fackets_out;
2249
2250         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2251         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2252         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2253
2254         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2255         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2256         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2257         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2258         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2259         info->tcpi_snd_cwnd = tp->snd_cwnd;
2260         info->tcpi_advmss = tp->advmss;
2261         info->tcpi_reordering = tp->reordering;
2262
2263         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2264         info->tcpi_rcv_space = tp->rcvq_space.space;
2265
2266         info->tcpi_total_retrans = tp->total_retrans;
2267 }
2268
2269 EXPORT_SYMBOL_GPL(tcp_get_info);
2270
2271 static int do_tcp_getsockopt(struct sock *sk, int level,
2272                 int optname, char __user *optval, int __user *optlen)
2273 {
2274         struct inet_connection_sock *icsk = inet_csk(sk);
2275         struct tcp_sock *tp = tcp_sk(sk);
2276         int val, len;
2277
2278         if (get_user(len, optlen))
2279                 return -EFAULT;
2280
2281         len = min_t(unsigned int, len, sizeof(int));
2282
2283         if (len < 0)
2284                 return -EINVAL;
2285
2286         switch (optname) {
2287         case TCP_MAXSEG:
2288                 val = tp->mss_cache;
2289                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2290                         val = tp->rx_opt.user_mss;
2291                 break;
2292         case TCP_NODELAY:
2293                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2294                 break;
2295         case TCP_CORK:
2296                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2297                 break;
2298         case TCP_KEEPIDLE:
2299                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2300                 break;
2301         case TCP_KEEPINTVL:
2302                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2303                 break;
2304         case TCP_KEEPCNT:
2305                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2306                 break;
2307         case TCP_SYNCNT:
2308                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2309                 break;
2310         case TCP_LINGER2:
2311                 val = tp->linger2;
2312                 if (val >= 0)
2313                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2314                 break;
2315         case TCP_DEFER_ACCEPT:
2316                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2317                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2318                 break;
2319         case TCP_WINDOW_CLAMP:
2320                 val = tp->window_clamp;
2321                 break;
2322         case TCP_INFO: {
2323                 struct tcp_info info;
2324
2325                 if (get_user(len, optlen))
2326                         return -EFAULT;
2327
2328                 tcp_get_info(sk, &info);
2329
2330                 len = min_t(unsigned int, len, sizeof(info));
2331                 if (put_user(len, optlen))
2332                         return -EFAULT;
2333                 if (copy_to_user(optval, &info, len))
2334                         return -EFAULT;
2335                 return 0;
2336         }
2337         case TCP_QUICKACK:
2338                 val = !icsk->icsk_ack.pingpong;
2339                 break;
2340
2341         case TCP_CONGESTION:
2342                 if (get_user(len, optlen))
2343                         return -EFAULT;
2344                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2345                 if (put_user(len, optlen))
2346                         return -EFAULT;
2347                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2348                         return -EFAULT;
2349                 return 0;
2350         default:
2351                 return -ENOPROTOOPT;
2352         }
2353
2354         if (put_user(len, optlen))
2355                 return -EFAULT;
2356         if (copy_to_user(optval, &val, len))
2357                 return -EFAULT;
2358         return 0;
2359 }
2360
2361 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2362                    int __user *optlen)
2363 {
2364         struct inet_connection_sock *icsk = inet_csk(sk);
2365
2366         if (level != SOL_TCP)
2367                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2368                                                      optval, optlen);
2369         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2370 }
2371
2372 #ifdef CONFIG_COMPAT
2373 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2374                           char __user *optval, int __user *optlen)
2375 {
2376         if (level != SOL_TCP)
2377                 return inet_csk_compat_getsockopt(sk, level, optname,
2378                                                   optval, optlen);
2379         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2380 }
2381
2382 EXPORT_SYMBOL(compat_tcp_getsockopt);
2383 #endif
2384
2385 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2386 {
2387         struct sk_buff *segs = ERR_PTR(-EINVAL);
2388         struct tcphdr *th;
2389         unsigned thlen;
2390         unsigned int seq;
2391         __be32 delta;
2392         unsigned int oldlen;
2393         unsigned int mss;
2394
2395         if (!pskb_may_pull(skb, sizeof(*th)))
2396                 goto out;
2397
2398         th = tcp_hdr(skb);
2399         thlen = th->doff * 4;
2400         if (thlen < sizeof(*th))
2401                 goto out;
2402
2403         if (!pskb_may_pull(skb, thlen))
2404                 goto out;
2405
2406         oldlen = (u16)~skb->len;
2407         __skb_pull(skb, thlen);
2408
2409         mss = skb_shinfo(skb)->gso_size;
2410         if (unlikely(skb->len <= mss))
2411                 goto out;
2412
2413         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2414                 /* Packet is from an untrusted source, reset gso_segs. */
2415                 int type = skb_shinfo(skb)->gso_type;
2416
2417                 if (unlikely(type &
2418                              ~(SKB_GSO_TCPV4 |
2419                                SKB_GSO_DODGY |
2420                                SKB_GSO_TCP_ECN |
2421                                SKB_GSO_TCPV6 |
2422                                0) ||
2423                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2424                         goto out;
2425
2426                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2427
2428                 segs = NULL;
2429                 goto out;
2430         }
2431
2432         segs = skb_segment(skb, features);
2433         if (IS_ERR(segs))
2434                 goto out;
2435
2436         delta = htonl(oldlen + (thlen + mss));
2437
2438         skb = segs;
2439         th = tcp_hdr(skb);
2440         seq = ntohl(th->seq);
2441
2442         do {
2443                 th->fin = th->psh = 0;
2444
2445                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2446                                        (__force u32)delta));
2447                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2448                         th->check =
2449                              csum_fold(csum_partial(skb_transport_header(skb),
2450                                                     thlen, skb->csum));
2451
2452                 seq += mss;
2453                 skb = skb->next;
2454                 th = tcp_hdr(skb);
2455
2456                 th->seq = htonl(seq);
2457                 th->cwr = 0;
2458         } while (skb->next);
2459
2460         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2461                       skb->data_len);
2462         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2463                                 (__force u32)delta));
2464         if (skb->ip_summed != CHECKSUM_PARTIAL)
2465                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2466                                                    thlen, skb->csum));
2467
2468 out:
2469         return segs;
2470 }
2471 EXPORT_SYMBOL(tcp_tso_segment);
2472
2473 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2474 {
2475         struct sk_buff **pp = NULL;
2476         struct sk_buff *p;
2477         struct tcphdr *th;
2478         struct tcphdr *th2;
2479         unsigned int thlen;
2480         unsigned int flags;
2481         unsigned int total;
2482         unsigned int mss = 1;
2483         int flush = 1;
2484
2485         th = skb_gro_header(skb, sizeof(*th));
2486         if (unlikely(!th))
2487                 goto out;
2488
2489         thlen = th->doff * 4;
2490         if (thlen < sizeof(*th))
2491                 goto out;
2492
2493         th = skb_gro_header(skb, thlen);
2494         if (unlikely(!th))
2495                 goto out;
2496
2497         skb_gro_pull(skb, thlen);
2498
2499         flags = tcp_flag_word(th);
2500
2501         for (; (p = *head); head = &p->next) {
2502                 if (!NAPI_GRO_CB(p)->same_flow)
2503                         continue;
2504
2505                 th2 = tcp_hdr(p);
2506
2507                 if (th->source != th2->source || th->dest != th2->dest) {
2508                         NAPI_GRO_CB(p)->same_flow = 0;
2509                         continue;
2510                 }
2511
2512                 goto found;
2513         }
2514
2515         goto out_check_final;
2516
2517 found:
2518         flush = NAPI_GRO_CB(p)->flush;
2519         flush |= flags & TCP_FLAG_CWR;
2520         flush |= (flags ^ tcp_flag_word(th2)) &
2521                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH);
2522         flush |= th->ack_seq != th2->ack_seq || th->window != th2->window;
2523         flush |= memcmp(th + 1, th2 + 1, thlen - sizeof(*th));
2524
2525         total = skb_gro_len(p);
2526         mss = skb_shinfo(p)->gso_size;
2527
2528         flush |= skb_gro_len(skb) > mss || !skb_gro_len(skb);
2529         flush |= ntohl(th2->seq) + total != ntohl(th->seq);
2530
2531         if (flush || skb_gro_receive(head, skb)) {
2532                 mss = 1;
2533                 goto out_check_final;
2534         }
2535
2536         p = *head;
2537         th2 = tcp_hdr(p);
2538         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2539
2540 out_check_final:
2541         flush = skb_gro_len(skb) < mss;
2542         flush |= flags & (TCP_FLAG_URG | TCP_FLAG_PSH | TCP_FLAG_RST |
2543                           TCP_FLAG_SYN | TCP_FLAG_FIN);
2544
2545         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2546                 pp = head;
2547
2548 out:
2549         NAPI_GRO_CB(skb)->flush |= flush;
2550
2551         return pp;
2552 }
2553 EXPORT_SYMBOL(tcp_gro_receive);
2554
2555 int tcp_gro_complete(struct sk_buff *skb)
2556 {
2557         struct tcphdr *th = tcp_hdr(skb);
2558
2559         skb->csum_start = skb_transport_header(skb) - skb->head;
2560         skb->csum_offset = offsetof(struct tcphdr, check);
2561         skb->ip_summed = CHECKSUM_PARTIAL;
2562
2563         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2564
2565         if (th->cwr)
2566                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2567
2568         return 0;
2569 }
2570 EXPORT_SYMBOL(tcp_gro_complete);
2571
2572 #ifdef CONFIG_TCP_MD5SIG
2573 static unsigned long tcp_md5sig_users;
2574 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2575 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2576
2577 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2578 {
2579         int cpu;
2580         for_each_possible_cpu(cpu) {
2581                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2582                 if (p) {
2583                         if (p->md5_desc.tfm)
2584                                 crypto_free_hash(p->md5_desc.tfm);
2585                         kfree(p);
2586                         p = NULL;
2587                 }
2588         }
2589         free_percpu(pool);
2590 }
2591
2592 void tcp_free_md5sig_pool(void)
2593 {
2594         struct tcp_md5sig_pool **pool = NULL;
2595
2596         spin_lock_bh(&tcp_md5sig_pool_lock);
2597         if (--tcp_md5sig_users == 0) {
2598                 pool = tcp_md5sig_pool;
2599                 tcp_md5sig_pool = NULL;
2600         }
2601         spin_unlock_bh(&tcp_md5sig_pool_lock);
2602         if (pool)
2603                 __tcp_free_md5sig_pool(pool);
2604 }
2605
2606 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2607
2608 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2609 {
2610         int cpu;
2611         struct tcp_md5sig_pool **pool;
2612
2613         pool = alloc_percpu(struct tcp_md5sig_pool *);
2614         if (!pool)
2615                 return NULL;
2616
2617         for_each_possible_cpu(cpu) {
2618                 struct tcp_md5sig_pool *p;
2619                 struct crypto_hash *hash;
2620
2621                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2622                 if (!p)
2623                         goto out_free;
2624                 *per_cpu_ptr(pool, cpu) = p;
2625
2626                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2627                 if (!hash || IS_ERR(hash))
2628                         goto out_free;
2629
2630                 p->md5_desc.tfm = hash;
2631         }
2632         return pool;
2633 out_free:
2634         __tcp_free_md5sig_pool(pool);
2635         return NULL;
2636 }
2637
2638 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2639 {
2640         struct tcp_md5sig_pool **pool;
2641         int alloc = 0;
2642
2643 retry:
2644         spin_lock_bh(&tcp_md5sig_pool_lock);
2645         pool = tcp_md5sig_pool;
2646         if (tcp_md5sig_users++ == 0) {
2647                 alloc = 1;
2648                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2649         } else if (!pool) {
2650                 tcp_md5sig_users--;
2651                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2652                 cpu_relax();
2653                 goto retry;
2654         } else
2655                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2656
2657         if (alloc) {
2658                 /* we cannot hold spinlock here because this may sleep. */
2659                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2660                 spin_lock_bh(&tcp_md5sig_pool_lock);
2661                 if (!p) {
2662                         tcp_md5sig_users--;
2663                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2664                         return NULL;
2665                 }
2666                 pool = tcp_md5sig_pool;
2667                 if (pool) {
2668                         /* oops, it has already been assigned. */
2669                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2670                         __tcp_free_md5sig_pool(p);
2671                 } else {
2672                         tcp_md5sig_pool = pool = p;
2673                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2674                 }
2675         }
2676         return pool;
2677 }
2678
2679 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2680
2681 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2682 {
2683         struct tcp_md5sig_pool **p;
2684         spin_lock_bh(&tcp_md5sig_pool_lock);
2685         p = tcp_md5sig_pool;
2686         if (p)
2687                 tcp_md5sig_users++;
2688         spin_unlock_bh(&tcp_md5sig_pool_lock);
2689         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2690 }
2691
2692 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2693
2694 void __tcp_put_md5sig_pool(void)
2695 {
2696         tcp_free_md5sig_pool();
2697 }
2698
2699 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2700
2701 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2702                         struct tcphdr *th)
2703 {
2704         struct scatterlist sg;
2705         int err;
2706
2707         __sum16 old_checksum = th->check;
2708         th->check = 0;
2709         /* options aren't included in the hash */
2710         sg_init_one(&sg, th, sizeof(struct tcphdr));
2711         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2712         th->check = old_checksum;
2713         return err;
2714 }
2715
2716 EXPORT_SYMBOL(tcp_md5_hash_header);
2717
2718 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2719                           struct sk_buff *skb, unsigned header_len)
2720 {
2721         struct scatterlist sg;
2722         const struct tcphdr *tp = tcp_hdr(skb);
2723         struct hash_desc *desc = &hp->md5_desc;
2724         unsigned i;
2725         const unsigned head_data_len = skb_headlen(skb) > header_len ?
2726                                        skb_headlen(skb) - header_len : 0;
2727         const struct skb_shared_info *shi = skb_shinfo(skb);
2728
2729         sg_init_table(&sg, 1);
2730
2731         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2732         if (crypto_hash_update(desc, &sg, head_data_len))
2733                 return 1;
2734
2735         for (i = 0; i < shi->nr_frags; ++i) {
2736                 const struct skb_frag_struct *f = &shi->frags[i];
2737                 sg_set_page(&sg, f->page, f->size, f->page_offset);
2738                 if (crypto_hash_update(desc, &sg, f->size))
2739                         return 1;
2740         }
2741
2742         return 0;
2743 }
2744
2745 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2746
2747 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
2748 {
2749         struct scatterlist sg;
2750
2751         sg_init_one(&sg, key->key, key->keylen);
2752         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2753 }
2754
2755 EXPORT_SYMBOL(tcp_md5_hash_key);
2756
2757 #endif
2758
2759 void tcp_done(struct sock *sk)
2760 {
2761         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2762                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2763
2764         tcp_set_state(sk, TCP_CLOSE);
2765         tcp_clear_xmit_timers(sk);
2766
2767         sk->sk_shutdown = SHUTDOWN_MASK;
2768
2769         if (!sock_flag(sk, SOCK_DEAD))
2770                 sk->sk_state_change(sk);
2771         else
2772                 inet_csk_destroy_sock(sk);
2773 }
2774 EXPORT_SYMBOL_GPL(tcp_done);
2775
2776 extern struct tcp_congestion_ops tcp_reno;
2777
2778 static __initdata unsigned long thash_entries;
2779 static int __init set_thash_entries(char *str)
2780 {
2781         if (!str)
2782                 return 0;
2783         thash_entries = simple_strtoul(str, &str, 0);
2784         return 1;
2785 }
2786 __setup("thash_entries=", set_thash_entries);
2787
2788 void __init tcp_init(void)
2789 {
2790         struct sk_buff *skb = NULL;
2791         unsigned long nr_pages, limit;
2792         int order, i, max_share;
2793
2794         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2795
2796         percpu_counter_init(&tcp_sockets_allocated, 0);
2797         percpu_counter_init(&tcp_orphan_count, 0);
2798         tcp_hashinfo.bind_bucket_cachep =
2799                 kmem_cache_create("tcp_bind_bucket",
2800                                   sizeof(struct inet_bind_bucket), 0,
2801                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2802
2803         /* Size and allocate the main established and bind bucket
2804          * hash tables.
2805          *
2806          * The methodology is similar to that of the buffer cache.
2807          */
2808         tcp_hashinfo.ehash =
2809                 alloc_large_system_hash("TCP established",
2810                                         sizeof(struct inet_ehash_bucket),
2811                                         thash_entries,
2812                                         (num_physpages >= 128 * 1024) ?
2813                                         13 : 15,
2814                                         0,
2815                                         &tcp_hashinfo.ehash_size,
2816                                         NULL,
2817                                         thash_entries ? 0 : 512 * 1024);
2818         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2819         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2820                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
2821                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
2822         }
2823         if (inet_ehash_locks_alloc(&tcp_hashinfo))
2824                 panic("TCP: failed to alloc ehash_locks");
2825         tcp_hashinfo.bhash =
2826                 alloc_large_system_hash("TCP bind",
2827                                         sizeof(struct inet_bind_hashbucket),
2828                                         tcp_hashinfo.ehash_size,
2829                                         (num_physpages >= 128 * 1024) ?
2830                                         13 : 15,
2831                                         0,
2832                                         &tcp_hashinfo.bhash_size,
2833                                         NULL,
2834                                         64 * 1024);
2835         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2836         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2837                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2838                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2839         }
2840
2841         /* Try to be a bit smarter and adjust defaults depending
2842          * on available memory.
2843          */
2844         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2845                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2846                         order++)
2847                 ;
2848         if (order >= 4) {
2849                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2850                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2851                 sysctl_max_syn_backlog = 1024;
2852         } else if (order < 3) {
2853                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2854                 sysctl_tcp_max_orphans >>= (3 - order);
2855                 sysctl_max_syn_backlog = 128;
2856         }
2857
2858         /* Set the pressure threshold to be a fraction of global memory that
2859          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2860          * memory, with a floor of 128 pages.
2861          */
2862         nr_pages = totalram_pages - totalhigh_pages;
2863         limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2864         limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2865         limit = max(limit, 128UL);
2866         sysctl_tcp_mem[0] = limit / 4 * 3;
2867         sysctl_tcp_mem[1] = limit;
2868         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2869
2870         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2871         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2872         max_share = min(4UL*1024*1024, limit);
2873
2874         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2875         sysctl_tcp_wmem[1] = 16*1024;
2876         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2877
2878         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2879         sysctl_tcp_rmem[1] = 87380;
2880         sysctl_tcp_rmem[2] = max(87380, max_share);
2881
2882         printk(KERN_INFO "TCP: Hash tables configured "
2883                "(established %d bind %d)\n",
2884                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2885
2886         tcp_register_congestion_control(&tcp_reno);
2887 }
2888
2889 EXPORT_SYMBOL(tcp_close);
2890 EXPORT_SYMBOL(tcp_disconnect);
2891 EXPORT_SYMBOL(tcp_getsockopt);
2892 EXPORT_SYMBOL(tcp_ioctl);
2893 EXPORT_SYMBOL(tcp_poll);
2894 EXPORT_SYMBOL(tcp_read_sock);
2895 EXPORT_SYMBOL(tcp_recvmsg);
2896 EXPORT_SYMBOL(tcp_sendmsg);
2897 EXPORT_SYMBOL(tcp_splice_read);
2898 EXPORT_SYMBOL(tcp_sendpage);
2899 EXPORT_SYMBOL(tcp_setsockopt);
2900 EXPORT_SYMBOL(tcp_shutdown);