net: Convert TCP & DCCP hash tables to use RCU / hlist_nulls
[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 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
281
282 EXPORT_SYMBOL_GPL(tcp_orphan_count);
283
284 int sysctl_tcp_mem[3] __read_mostly;
285 int sysctl_tcp_wmem[3] __read_mostly;
286 int sysctl_tcp_rmem[3] __read_mostly;
287
288 EXPORT_SYMBOL(sysctl_tcp_mem);
289 EXPORT_SYMBOL(sysctl_tcp_rmem);
290 EXPORT_SYMBOL(sysctl_tcp_wmem);
291
292 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
293 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
294
295 EXPORT_SYMBOL(tcp_memory_allocated);
296 EXPORT_SYMBOL(tcp_sockets_allocated);
297
298 /*
299  * TCP splice context
300  */
301 struct tcp_splice_state {
302         struct pipe_inode_info *pipe;
303         size_t len;
304         unsigned int flags;
305 };
306
307 /*
308  * Pressure flag: try to collapse.
309  * Technical note: it is used by multiple contexts non atomically.
310  * All the __sk_mem_schedule() is of this nature: accounting
311  * is strict, actions are advisory and have some latency.
312  */
313 int tcp_memory_pressure __read_mostly;
314
315 EXPORT_SYMBOL(tcp_memory_pressure);
316
317 void tcp_enter_memory_pressure(struct sock *sk)
318 {
319         if (!tcp_memory_pressure) {
320                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
321                 tcp_memory_pressure = 1;
322         }
323 }
324
325 EXPORT_SYMBOL(tcp_enter_memory_pressure);
326
327 /*
328  *      Wait for a TCP event.
329  *
330  *      Note that we don't need to lock the socket, as the upper poll layers
331  *      take care of normal races (between the test and the event) and we don't
332  *      go look at any of the socket buffers directly.
333  */
334 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
335 {
336         unsigned int mask;
337         struct sock *sk = sock->sk;
338         struct tcp_sock *tp = tcp_sk(sk);
339
340         poll_wait(file, sk->sk_sleep, wait);
341         if (sk->sk_state == TCP_LISTEN)
342                 return inet_csk_listen_poll(sk);
343
344         /* Socket is not locked. We are protected from async events
345          * by poll logic and correct handling of state changes
346          * made by other threads is impossible in any case.
347          */
348
349         mask = 0;
350         if (sk->sk_err)
351                 mask = POLLERR;
352
353         /*
354          * POLLHUP is certainly not done right. But poll() doesn't
355          * have a notion of HUP in just one direction, and for a
356          * socket the read side is more interesting.
357          *
358          * Some poll() documentation says that POLLHUP is incompatible
359          * with the POLLOUT/POLLWR flags, so somebody should check this
360          * all. But careful, it tends to be safer to return too many
361          * bits than too few, and you can easily break real applications
362          * if you don't tell them that something has hung up!
363          *
364          * Check-me.
365          *
366          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
367          * our fs/select.c). It means that after we received EOF,
368          * poll always returns immediately, making impossible poll() on write()
369          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
370          * if and only if shutdown has been made in both directions.
371          * Actually, it is interesting to look how Solaris and DUX
372          * solve this dilemma. I would prefer, if POLLHUP were maskable,
373          * then we could set it on SND_SHUTDOWN. BTW examples given
374          * in Stevens' books assume exactly this behaviour, it explains
375          * why POLLHUP is incompatible with POLLOUT.    --ANK
376          *
377          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
378          * blocking on fresh not-connected or disconnected socket. --ANK
379          */
380         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
381                 mask |= POLLHUP;
382         if (sk->sk_shutdown & RCV_SHUTDOWN)
383                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
384
385         /* Connected? */
386         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
387                 int target = sock_rcvlowat(sk, 0, INT_MAX);
388
389                 if (tp->urg_seq == tp->copied_seq &&
390                     !sock_flag(sk, SOCK_URGINLINE) &&
391                     tp->urg_data)
392                         target--;
393
394                 /* Potential race condition. If read of tp below will
395                  * escape above sk->sk_state, we can be illegally awaken
396                  * in SYN_* states. */
397                 if (tp->rcv_nxt - tp->copied_seq >= target)
398                         mask |= POLLIN | POLLRDNORM;
399
400                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
401                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
402                                 mask |= POLLOUT | POLLWRNORM;
403                         } else {  /* send SIGIO later */
404                                 set_bit(SOCK_ASYNC_NOSPACE,
405                                         &sk->sk_socket->flags);
406                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
407
408                                 /* Race breaker. If space is freed after
409                                  * wspace test but before the flags are set,
410                                  * IO signal will be lost.
411                                  */
412                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
413                                         mask |= POLLOUT | POLLWRNORM;
414                         }
415                 }
416
417                 if (tp->urg_data & TCP_URG_VALID)
418                         mask |= POLLPRI;
419         }
420         return mask;
421 }
422
423 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
424 {
425         struct tcp_sock *tp = tcp_sk(sk);
426         int answ;
427
428         switch (cmd) {
429         case SIOCINQ:
430                 if (sk->sk_state == TCP_LISTEN)
431                         return -EINVAL;
432
433                 lock_sock(sk);
434                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
435                         answ = 0;
436                 else if (sock_flag(sk, SOCK_URGINLINE) ||
437                          !tp->urg_data ||
438                          before(tp->urg_seq, tp->copied_seq) ||
439                          !before(tp->urg_seq, tp->rcv_nxt)) {
440                         answ = tp->rcv_nxt - tp->copied_seq;
441
442                         /* Subtract 1, if FIN is in queue. */
443                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
444                                 answ -=
445                        tcp_hdr((struct sk_buff *)sk->sk_receive_queue.prev)->fin;
446                 } else
447                         answ = tp->urg_seq - tp->copied_seq;
448                 release_sock(sk);
449                 break;
450         case SIOCATMARK:
451                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
452                 break;
453         case SIOCOUTQ:
454                 if (sk->sk_state == TCP_LISTEN)
455                         return -EINVAL;
456
457                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
458                         answ = 0;
459                 else
460                         answ = tp->write_seq - tp->snd_una;
461                 break;
462         default:
463                 return -ENOIOCTLCMD;
464         }
465
466         return put_user(answ, (int __user *)arg);
467 }
468
469 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
470 {
471         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
472         tp->pushed_seq = tp->write_seq;
473 }
474
475 static inline int forced_push(struct tcp_sock *tp)
476 {
477         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
478 }
479
480 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
481 {
482         struct tcp_sock *tp = tcp_sk(sk);
483         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
484
485         skb->csum    = 0;
486         tcb->seq     = tcb->end_seq = tp->write_seq;
487         tcb->flags   = TCPCB_FLAG_ACK;
488         tcb->sacked  = 0;
489         skb_header_release(skb);
490         tcp_add_write_queue_tail(sk, skb);
491         sk->sk_wmem_queued += skb->truesize;
492         sk_mem_charge(sk, skb->truesize);
493         if (tp->nonagle & TCP_NAGLE_PUSH)
494                 tp->nonagle &= ~TCP_NAGLE_PUSH;
495 }
496
497 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
498                                 struct sk_buff *skb)
499 {
500         if (flags & MSG_OOB)
501                 tp->snd_up = tp->write_seq;
502 }
503
504 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
505                             int nonagle)
506 {
507         struct tcp_sock *tp = tcp_sk(sk);
508
509         if (tcp_send_head(sk)) {
510                 struct sk_buff *skb = tcp_write_queue_tail(sk);
511                 if (!(flags & MSG_MORE) || forced_push(tp))
512                         tcp_mark_push(tp, skb);
513                 tcp_mark_urg(tp, flags, skb);
514                 __tcp_push_pending_frames(sk, mss_now,
515                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
516         }
517 }
518
519 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
520                                 unsigned int offset, size_t len)
521 {
522         struct tcp_splice_state *tss = rd_desc->arg.data;
523
524         return skb_splice_bits(skb, offset, tss->pipe, tss->len, tss->flags);
525 }
526
527 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
528 {
529         /* Store TCP splice context information in read_descriptor_t. */
530         read_descriptor_t rd_desc = {
531                 .arg.data = tss,
532         };
533
534         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
535 }
536
537 /**
538  *  tcp_splice_read - splice data from TCP socket to a pipe
539  * @sock:       socket to splice from
540  * @ppos:       position (not valid)
541  * @pipe:       pipe to splice to
542  * @len:        number of bytes to splice
543  * @flags:      splice modifier flags
544  *
545  * Description:
546  *    Will read pages from given socket and fill them into a pipe.
547  *
548  **/
549 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
550                         struct pipe_inode_info *pipe, size_t len,
551                         unsigned int flags)
552 {
553         struct sock *sk = sock->sk;
554         struct tcp_splice_state tss = {
555                 .pipe = pipe,
556                 .len = len,
557                 .flags = flags,
558         };
559         long timeo;
560         ssize_t spliced;
561         int ret;
562
563         /*
564          * We can't seek on a socket input
565          */
566         if (unlikely(*ppos))
567                 return -ESPIPE;
568
569         ret = spliced = 0;
570
571         lock_sock(sk);
572
573         timeo = sock_rcvtimeo(sk, flags & SPLICE_F_NONBLOCK);
574         while (tss.len) {
575                 ret = __tcp_splice_read(sk, &tss);
576                 if (ret < 0)
577                         break;
578                 else if (!ret) {
579                         if (spliced)
580                                 break;
581                         if (flags & SPLICE_F_NONBLOCK) {
582                                 ret = -EAGAIN;
583                                 break;
584                         }
585                         if (sock_flag(sk, SOCK_DONE))
586                                 break;
587                         if (sk->sk_err) {
588                                 ret = sock_error(sk);
589                                 break;
590                         }
591                         if (sk->sk_shutdown & RCV_SHUTDOWN)
592                                 break;
593                         if (sk->sk_state == TCP_CLOSE) {
594                                 /*
595                                  * This occurs when user tries to read
596                                  * from never connected socket.
597                                  */
598                                 if (!sock_flag(sk, SOCK_DONE))
599                                         ret = -ENOTCONN;
600                                 break;
601                         }
602                         if (!timeo) {
603                                 ret = -EAGAIN;
604                                 break;
605                         }
606                         sk_wait_data(sk, &timeo);
607                         if (signal_pending(current)) {
608                                 ret = sock_intr_errno(timeo);
609                                 break;
610                         }
611                         continue;
612                 }
613                 tss.len -= ret;
614                 spliced += ret;
615
616                 release_sock(sk);
617                 lock_sock(sk);
618
619                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
620                     (sk->sk_shutdown & RCV_SHUTDOWN) || !timeo ||
621                     signal_pending(current))
622                         break;
623         }
624
625         release_sock(sk);
626
627         if (spliced)
628                 return spliced;
629
630         return ret;
631 }
632
633 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
634 {
635         struct sk_buff *skb;
636
637         /* The TCP header must be at least 32-bit aligned.  */
638         size = ALIGN(size, 4);
639
640         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
641         if (skb) {
642                 if (sk_wmem_schedule(sk, skb->truesize)) {
643                         /*
644                          * Make sure that we have exactly size bytes
645                          * available to the caller, no more, no less.
646                          */
647                         skb_reserve(skb, skb_tailroom(skb) - size);
648                         return skb;
649                 }
650                 __kfree_skb(skb);
651         } else {
652                 sk->sk_prot->enter_memory_pressure(sk);
653                 sk_stream_moderate_sndbuf(sk);
654         }
655         return NULL;
656 }
657
658 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
659                          size_t psize, int flags)
660 {
661         struct tcp_sock *tp = tcp_sk(sk);
662         int mss_now, size_goal;
663         int err;
664         ssize_t copied;
665         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
666
667         /* Wait for a connection to finish. */
668         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
669                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
670                         goto out_err;
671
672         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
673
674         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
675         size_goal = tp->xmit_size_goal;
676         copied = 0;
677
678         err = -EPIPE;
679         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
680                 goto do_error;
681
682         while (psize > 0) {
683                 struct sk_buff *skb = tcp_write_queue_tail(sk);
684                 struct page *page = pages[poffset / PAGE_SIZE];
685                 int copy, i, can_coalesce;
686                 int offset = poffset % PAGE_SIZE;
687                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
688
689                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
690 new_segment:
691                         if (!sk_stream_memory_free(sk))
692                                 goto wait_for_sndbuf;
693
694                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
695                         if (!skb)
696                                 goto wait_for_memory;
697
698                         skb_entail(sk, skb);
699                         copy = size_goal;
700                 }
701
702                 if (copy > size)
703                         copy = size;
704
705                 i = skb_shinfo(skb)->nr_frags;
706                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
707                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
708                         tcp_mark_push(tp, skb);
709                         goto new_segment;
710                 }
711                 if (!sk_wmem_schedule(sk, copy))
712                         goto wait_for_memory;
713
714                 if (can_coalesce) {
715                         skb_shinfo(skb)->frags[i - 1].size += copy;
716                 } else {
717                         get_page(page);
718                         skb_fill_page_desc(skb, i, page, offset, copy);
719                 }
720
721                 skb->len += copy;
722                 skb->data_len += copy;
723                 skb->truesize += copy;
724                 sk->sk_wmem_queued += copy;
725                 sk_mem_charge(sk, copy);
726                 skb->ip_summed = CHECKSUM_PARTIAL;
727                 tp->write_seq += copy;
728                 TCP_SKB_CB(skb)->end_seq += copy;
729                 skb_shinfo(skb)->gso_segs = 0;
730
731                 if (!copied)
732                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
733
734                 copied += copy;
735                 poffset += copy;
736                 if (!(psize -= copy))
737                         goto out;
738
739                 if (skb->len < size_goal || (flags & MSG_OOB))
740                         continue;
741
742                 if (forced_push(tp)) {
743                         tcp_mark_push(tp, skb);
744                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
745                 } else if (skb == tcp_send_head(sk))
746                         tcp_push_one(sk, mss_now);
747                 continue;
748
749 wait_for_sndbuf:
750                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
751 wait_for_memory:
752                 if (copied)
753                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
754
755                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
756                         goto do_error;
757
758                 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
759                 size_goal = tp->xmit_size_goal;
760         }
761
762 out:
763         if (copied)
764                 tcp_push(sk, flags, mss_now, tp->nonagle);
765         return copied;
766
767 do_error:
768         if (copied)
769                 goto out;
770 out_err:
771         return sk_stream_error(sk, flags, err);
772 }
773
774 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
775                      size_t size, int flags)
776 {
777         ssize_t res;
778         struct sock *sk = sock->sk;
779
780         if (!(sk->sk_route_caps & NETIF_F_SG) ||
781             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
782                 return sock_no_sendpage(sock, page, offset, size, flags);
783
784         lock_sock(sk);
785         TCP_CHECK_TIMER(sk);
786         res = do_tcp_sendpages(sk, &page, offset, size, flags);
787         TCP_CHECK_TIMER(sk);
788         release_sock(sk);
789         return res;
790 }
791
792 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
793 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
794
795 static inline int select_size(struct sock *sk)
796 {
797         struct tcp_sock *tp = tcp_sk(sk);
798         int tmp = tp->mss_cache;
799
800         if (sk->sk_route_caps & NETIF_F_SG) {
801                 if (sk_can_gso(sk))
802                         tmp = 0;
803                 else {
804                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
805
806                         if (tmp >= pgbreak &&
807                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
808                                 tmp = pgbreak;
809                 }
810         }
811
812         return tmp;
813 }
814
815 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
816                 size_t size)
817 {
818         struct sock *sk = sock->sk;
819         struct iovec *iov;
820         struct tcp_sock *tp = tcp_sk(sk);
821         struct sk_buff *skb;
822         int iovlen, flags;
823         int mss_now, size_goal;
824         int err, copied;
825         long timeo;
826
827         lock_sock(sk);
828         TCP_CHECK_TIMER(sk);
829
830         flags = msg->msg_flags;
831         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
832
833         /* Wait for a connection to finish. */
834         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
835                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
836                         goto out_err;
837
838         /* This should be in poll */
839         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
840
841         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
842         size_goal = tp->xmit_size_goal;
843
844         /* Ok commence sending. */
845         iovlen = msg->msg_iovlen;
846         iov = msg->msg_iov;
847         copied = 0;
848
849         err = -EPIPE;
850         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
851                 goto do_error;
852
853         while (--iovlen >= 0) {
854                 int seglen = iov->iov_len;
855                 unsigned char __user *from = iov->iov_base;
856
857                 iov++;
858
859                 while (seglen > 0) {
860                         int copy;
861
862                         skb = tcp_write_queue_tail(sk);
863
864                         if (!tcp_send_head(sk) ||
865                             (copy = size_goal - skb->len) <= 0) {
866
867 new_segment:
868                                 /* Allocate new segment. If the interface is SG,
869                                  * allocate skb fitting to single page.
870                                  */
871                                 if (!sk_stream_memory_free(sk))
872                                         goto wait_for_sndbuf;
873
874                                 skb = sk_stream_alloc_skb(sk, select_size(sk),
875                                                 sk->sk_allocation);
876                                 if (!skb)
877                                         goto wait_for_memory;
878
879                                 /*
880                                  * Check whether we can use HW checksum.
881                                  */
882                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
883                                         skb->ip_summed = CHECKSUM_PARTIAL;
884
885                                 skb_entail(sk, skb);
886                                 copy = size_goal;
887                         }
888
889                         /* Try to append data to the end of skb. */
890                         if (copy > seglen)
891                                 copy = seglen;
892
893                         /* Where to copy to? */
894                         if (skb_tailroom(skb) > 0) {
895                                 /* We have some space in skb head. Superb! */
896                                 if (copy > skb_tailroom(skb))
897                                         copy = skb_tailroom(skb);
898                                 if ((err = skb_add_data(skb, from, copy)) != 0)
899                                         goto do_fault;
900                         } else {
901                                 int merge = 0;
902                                 int i = skb_shinfo(skb)->nr_frags;
903                                 struct page *page = TCP_PAGE(sk);
904                                 int off = TCP_OFF(sk);
905
906                                 if (skb_can_coalesce(skb, i, page, off) &&
907                                     off != PAGE_SIZE) {
908                                         /* We can extend the last page
909                                          * fragment. */
910                                         merge = 1;
911                                 } else if (i == MAX_SKB_FRAGS ||
912                                            (!i &&
913                                            !(sk->sk_route_caps & NETIF_F_SG))) {
914                                         /* Need to add new fragment and cannot
915                                          * do this because interface is non-SG,
916                                          * or because all the page slots are
917                                          * busy. */
918                                         tcp_mark_push(tp, skb);
919                                         goto new_segment;
920                                 } else if (page) {
921                                         if (off == PAGE_SIZE) {
922                                                 put_page(page);
923                                                 TCP_PAGE(sk) = page = NULL;
924                                                 off = 0;
925                                         }
926                                 } else
927                                         off = 0;
928
929                                 if (copy > PAGE_SIZE - off)
930                                         copy = PAGE_SIZE - off;
931
932                                 if (!sk_wmem_schedule(sk, copy))
933                                         goto wait_for_memory;
934
935                                 if (!page) {
936                                         /* Allocate new cache page. */
937                                         if (!(page = sk_stream_alloc_page(sk)))
938                                                 goto wait_for_memory;
939                                 }
940
941                                 /* Time to copy data. We are close to
942                                  * the end! */
943                                 err = skb_copy_to_page(sk, from, skb, page,
944                                                        off, copy);
945                                 if (err) {
946                                         /* If this page was new, give it to the
947                                          * socket so it does not get leaked.
948                                          */
949                                         if (!TCP_PAGE(sk)) {
950                                                 TCP_PAGE(sk) = page;
951                                                 TCP_OFF(sk) = 0;
952                                         }
953                                         goto do_error;
954                                 }
955
956                                 /* Update the skb. */
957                                 if (merge) {
958                                         skb_shinfo(skb)->frags[i - 1].size +=
959                                                                         copy;
960                                 } else {
961                                         skb_fill_page_desc(skb, i, page, off, copy);
962                                         if (TCP_PAGE(sk)) {
963                                                 get_page(page);
964                                         } else if (off + copy < PAGE_SIZE) {
965                                                 get_page(page);
966                                                 TCP_PAGE(sk) = page;
967                                         }
968                                 }
969
970                                 TCP_OFF(sk) = off + copy;
971                         }
972
973                         if (!copied)
974                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
975
976                         tp->write_seq += copy;
977                         TCP_SKB_CB(skb)->end_seq += copy;
978                         skb_shinfo(skb)->gso_segs = 0;
979
980                         from += copy;
981                         copied += copy;
982                         if ((seglen -= copy) == 0 && iovlen == 0)
983                                 goto out;
984
985                         if (skb->len < size_goal || (flags & MSG_OOB))
986                                 continue;
987
988                         if (forced_push(tp)) {
989                                 tcp_mark_push(tp, skb);
990                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
991                         } else if (skb == tcp_send_head(sk))
992                                 tcp_push_one(sk, mss_now);
993                         continue;
994
995 wait_for_sndbuf:
996                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
997 wait_for_memory:
998                         if (copied)
999                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1000
1001                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1002                                 goto do_error;
1003
1004                         mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
1005                         size_goal = tp->xmit_size_goal;
1006                 }
1007         }
1008
1009 out:
1010         if (copied)
1011                 tcp_push(sk, flags, mss_now, tp->nonagle);
1012         TCP_CHECK_TIMER(sk);
1013         release_sock(sk);
1014         return copied;
1015
1016 do_fault:
1017         if (!skb->len) {
1018                 tcp_unlink_write_queue(skb, sk);
1019                 /* It is the one place in all of TCP, except connection
1020                  * reset, where we can be unlinking the send_head.
1021                  */
1022                 tcp_check_send_head(sk, skb);
1023                 sk_wmem_free_skb(sk, skb);
1024         }
1025
1026 do_error:
1027         if (copied)
1028                 goto out;
1029 out_err:
1030         err = sk_stream_error(sk, flags, err);
1031         TCP_CHECK_TIMER(sk);
1032         release_sock(sk);
1033         return err;
1034 }
1035
1036 /*
1037  *      Handle reading urgent data. BSD has very simple semantics for
1038  *      this, no blocking and very strange errors 8)
1039  */
1040
1041 static int tcp_recv_urg(struct sock *sk, long timeo,
1042                         struct msghdr *msg, int len, int flags,
1043                         int *addr_len)
1044 {
1045         struct tcp_sock *tp = tcp_sk(sk);
1046
1047         /* No URG data to read. */
1048         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1049             tp->urg_data == TCP_URG_READ)
1050                 return -EINVAL; /* Yes this is right ! */
1051
1052         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1053                 return -ENOTCONN;
1054
1055         if (tp->urg_data & TCP_URG_VALID) {
1056                 int err = 0;
1057                 char c = tp->urg_data;
1058
1059                 if (!(flags & MSG_PEEK))
1060                         tp->urg_data = TCP_URG_READ;
1061
1062                 /* Read urgent data. */
1063                 msg->msg_flags |= MSG_OOB;
1064
1065                 if (len > 0) {
1066                         if (!(flags & MSG_TRUNC))
1067                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1068                         len = 1;
1069                 } else
1070                         msg->msg_flags |= MSG_TRUNC;
1071
1072                 return err ? -EFAULT : len;
1073         }
1074
1075         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1076                 return 0;
1077
1078         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1079          * the available implementations agree in this case:
1080          * this call should never block, independent of the
1081          * blocking state of the socket.
1082          * Mike <pall@rz.uni-karlsruhe.de>
1083          */
1084         return -EAGAIN;
1085 }
1086
1087 /* Clean up the receive buffer for full frames taken by the user,
1088  * then send an ACK if necessary.  COPIED is the number of bytes
1089  * tcp_recvmsg has given to the user so far, it speeds up the
1090  * calculation of whether or not we must ACK for the sake of
1091  * a window update.
1092  */
1093 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1094 {
1095         struct tcp_sock *tp = tcp_sk(sk);
1096         int time_to_ack = 0;
1097
1098 #if TCP_DEBUG
1099         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1100
1101         WARN_ON(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1102 #endif
1103
1104         if (inet_csk_ack_scheduled(sk)) {
1105                 const struct inet_connection_sock *icsk = inet_csk(sk);
1106                    /* Delayed ACKs frequently hit locked sockets during bulk
1107                     * receive. */
1108                 if (icsk->icsk_ack.blocked ||
1109                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1110                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1111                     /*
1112                      * If this read emptied read buffer, we send ACK, if
1113                      * connection is not bidirectional, user drained
1114                      * receive buffer and there was a small segment
1115                      * in queue.
1116                      */
1117                     (copied > 0 &&
1118                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1119                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1120                        !icsk->icsk_ack.pingpong)) &&
1121                       !atomic_read(&sk->sk_rmem_alloc)))
1122                         time_to_ack = 1;
1123         }
1124
1125         /* We send an ACK if we can now advertise a non-zero window
1126          * which has been raised "significantly".
1127          *
1128          * Even if window raised up to infinity, do not send window open ACK
1129          * in states, where we will not receive more. It is useless.
1130          */
1131         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1132                 __u32 rcv_window_now = tcp_receive_window(tp);
1133
1134                 /* Optimize, __tcp_select_window() is not cheap. */
1135                 if (2*rcv_window_now <= tp->window_clamp) {
1136                         __u32 new_window = __tcp_select_window(sk);
1137
1138                         /* Send ACK now, if this read freed lots of space
1139                          * in our buffer. Certainly, new_window is new window.
1140                          * We can advertise it now, if it is not less than current one.
1141                          * "Lots" means "at least twice" here.
1142                          */
1143                         if (new_window && new_window >= 2 * rcv_window_now)
1144                                 time_to_ack = 1;
1145                 }
1146         }
1147         if (time_to_ack)
1148                 tcp_send_ack(sk);
1149 }
1150
1151 static void tcp_prequeue_process(struct sock *sk)
1152 {
1153         struct sk_buff *skb;
1154         struct tcp_sock *tp = tcp_sk(sk);
1155
1156         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1157
1158         /* RX process wants to run with disabled BHs, though it is not
1159          * necessary */
1160         local_bh_disable();
1161         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1162                 sk_backlog_rcv(sk, skb);
1163         local_bh_enable();
1164
1165         /* Clear memory counter. */
1166         tp->ucopy.memory = 0;
1167 }
1168
1169 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1170 {
1171         struct sk_buff *skb;
1172         u32 offset;
1173
1174         skb_queue_walk(&sk->sk_receive_queue, skb) {
1175                 offset = seq - TCP_SKB_CB(skb)->seq;
1176                 if (tcp_hdr(skb)->syn)
1177                         offset--;
1178                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1179                         *off = offset;
1180                         return skb;
1181                 }
1182         }
1183         return NULL;
1184 }
1185
1186 /*
1187  * This routine provides an alternative to tcp_recvmsg() for routines
1188  * that would like to handle copying from skbuffs directly in 'sendfile'
1189  * fashion.
1190  * Note:
1191  *      - It is assumed that the socket was locked by the caller.
1192  *      - The routine does not block.
1193  *      - At present, there is no support for reading OOB data
1194  *        or for 'peeking' the socket using this routine
1195  *        (although both would be easy to implement).
1196  */
1197 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1198                   sk_read_actor_t recv_actor)
1199 {
1200         struct sk_buff *skb;
1201         struct tcp_sock *tp = tcp_sk(sk);
1202         u32 seq = tp->copied_seq;
1203         u32 offset;
1204         int copied = 0;
1205
1206         if (sk->sk_state == TCP_LISTEN)
1207                 return -ENOTCONN;
1208         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1209                 if (offset < skb->len) {
1210                         int used;
1211                         size_t len;
1212
1213                         len = skb->len - offset;
1214                         /* Stop reading if we hit a patch of urgent data */
1215                         if (tp->urg_data) {
1216                                 u32 urg_offset = tp->urg_seq - seq;
1217                                 if (urg_offset < len)
1218                                         len = urg_offset;
1219                                 if (!len)
1220                                         break;
1221                         }
1222                         used = recv_actor(desc, skb, offset, len);
1223                         if (used < 0) {
1224                                 if (!copied)
1225                                         copied = used;
1226                                 break;
1227                         } else if (used <= len) {
1228                                 seq += used;
1229                                 copied += used;
1230                                 offset += used;
1231                         }
1232                         /*
1233                          * If recv_actor drops the lock (e.g. TCP splice
1234                          * receive) the skb pointer might be invalid when
1235                          * getting here: tcp_collapse might have deleted it
1236                          * while aggregating skbs from the socket queue.
1237                          */
1238                         skb = tcp_recv_skb(sk, seq-1, &offset);
1239                         if (!skb || (offset+1 != skb->len))
1240                                 break;
1241                 }
1242                 if (tcp_hdr(skb)->fin) {
1243                         sk_eat_skb(sk, skb, 0);
1244                         ++seq;
1245                         break;
1246                 }
1247                 sk_eat_skb(sk, skb, 0);
1248                 if (!desc->count)
1249                         break;
1250         }
1251         tp->copied_seq = seq;
1252
1253         tcp_rcv_space_adjust(sk);
1254
1255         /* Clean up data we have read: This will do ACK frames. */
1256         if (copied > 0)
1257                 tcp_cleanup_rbuf(sk, copied);
1258         return copied;
1259 }
1260
1261 /*
1262  *      This routine copies from a sock struct into the user buffer.
1263  *
1264  *      Technical note: in 2.3 we work on _locked_ socket, so that
1265  *      tricks with *seq access order and skb->users are not required.
1266  *      Probably, code can be easily improved even more.
1267  */
1268
1269 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1270                 size_t len, int nonblock, int flags, int *addr_len)
1271 {
1272         struct tcp_sock *tp = tcp_sk(sk);
1273         int copied = 0;
1274         u32 peek_seq;
1275         u32 *seq;
1276         unsigned long used;
1277         int err;
1278         int target;             /* Read at least this many bytes */
1279         long timeo;
1280         struct task_struct *user_recv = NULL;
1281         int copied_early = 0;
1282         struct sk_buff *skb;
1283
1284         lock_sock(sk);
1285
1286         TCP_CHECK_TIMER(sk);
1287
1288         err = -ENOTCONN;
1289         if (sk->sk_state == TCP_LISTEN)
1290                 goto out;
1291
1292         timeo = sock_rcvtimeo(sk, nonblock);
1293
1294         /* Urgent data needs to be handled specially. */
1295         if (flags & MSG_OOB)
1296                 goto recv_urg;
1297
1298         seq = &tp->copied_seq;
1299         if (flags & MSG_PEEK) {
1300                 peek_seq = tp->copied_seq;
1301                 seq = &peek_seq;
1302         }
1303
1304         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1305
1306 #ifdef CONFIG_NET_DMA
1307         tp->ucopy.dma_chan = NULL;
1308         preempt_disable();
1309         skb = skb_peek_tail(&sk->sk_receive_queue);
1310         {
1311                 int available = 0;
1312
1313                 if (skb)
1314                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1315                 if ((available < target) &&
1316                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1317                     !sysctl_tcp_low_latency &&
1318                     __get_cpu_var(softnet_data).net_dma) {
1319                         preempt_enable_no_resched();
1320                         tp->ucopy.pinned_list =
1321                                         dma_pin_iovec_pages(msg->msg_iov, len);
1322                 } else {
1323                         preempt_enable_no_resched();
1324                 }
1325         }
1326 #endif
1327
1328         do {
1329                 u32 offset;
1330
1331                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1332                 if (tp->urg_data && tp->urg_seq == *seq) {
1333                         if (copied)
1334                                 break;
1335                         if (signal_pending(current)) {
1336                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1337                                 break;
1338                         }
1339                 }
1340
1341                 /* Next get a buffer. */
1342
1343                 skb = skb_peek(&sk->sk_receive_queue);
1344                 do {
1345                         if (!skb)
1346                                 break;
1347
1348                         /* Now that we have two receive queues this
1349                          * shouldn't happen.
1350                          */
1351                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1352                                 printk(KERN_INFO "recvmsg bug: copied %X "
1353                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1354                                 break;
1355                         }
1356                         offset = *seq - TCP_SKB_CB(skb)->seq;
1357                         if (tcp_hdr(skb)->syn)
1358                                 offset--;
1359                         if (offset < skb->len)
1360                                 goto found_ok_skb;
1361                         if (tcp_hdr(skb)->fin)
1362                                 goto found_fin_ok;
1363                         WARN_ON(!(flags & MSG_PEEK));
1364                         skb = skb->next;
1365                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1366
1367                 /* Well, if we have backlog, try to process it now yet. */
1368
1369                 if (copied >= target && !sk->sk_backlog.tail)
1370                         break;
1371
1372                 if (copied) {
1373                         if (sk->sk_err ||
1374                             sk->sk_state == TCP_CLOSE ||
1375                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1376                             !timeo ||
1377                             signal_pending(current))
1378                                 break;
1379                 } else {
1380                         if (sock_flag(sk, SOCK_DONE))
1381                                 break;
1382
1383                         if (sk->sk_err) {
1384                                 copied = sock_error(sk);
1385                                 break;
1386                         }
1387
1388                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1389                                 break;
1390
1391                         if (sk->sk_state == TCP_CLOSE) {
1392                                 if (!sock_flag(sk, SOCK_DONE)) {
1393                                         /* This occurs when user tries to read
1394                                          * from never connected socket.
1395                                          */
1396                                         copied = -ENOTCONN;
1397                                         break;
1398                                 }
1399                                 break;
1400                         }
1401
1402                         if (!timeo) {
1403                                 copied = -EAGAIN;
1404                                 break;
1405                         }
1406
1407                         if (signal_pending(current)) {
1408                                 copied = sock_intr_errno(timeo);
1409                                 break;
1410                         }
1411                 }
1412
1413                 tcp_cleanup_rbuf(sk, copied);
1414
1415                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1416                         /* Install new reader */
1417                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1418                                 user_recv = current;
1419                                 tp->ucopy.task = user_recv;
1420                                 tp->ucopy.iov = msg->msg_iov;
1421                         }
1422
1423                         tp->ucopy.len = len;
1424
1425                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1426                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1427
1428                         /* Ugly... If prequeue is not empty, we have to
1429                          * process it before releasing socket, otherwise
1430                          * order will be broken at second iteration.
1431                          * More elegant solution is required!!!
1432                          *
1433                          * Look: we have the following (pseudo)queues:
1434                          *
1435                          * 1. packets in flight
1436                          * 2. backlog
1437                          * 3. prequeue
1438                          * 4. receive_queue
1439                          *
1440                          * Each queue can be processed only if the next ones
1441                          * are empty. At this point we have empty receive_queue.
1442                          * But prequeue _can_ be not empty after 2nd iteration,
1443                          * when we jumped to start of loop because backlog
1444                          * processing added something to receive_queue.
1445                          * We cannot release_sock(), because backlog contains
1446                          * packets arrived _after_ prequeued ones.
1447                          *
1448                          * Shortly, algorithm is clear --- to process all
1449                          * the queues in order. We could make it more directly,
1450                          * requeueing packets from backlog to prequeue, if
1451                          * is not empty. It is more elegant, but eats cycles,
1452                          * unfortunately.
1453                          */
1454                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1455                                 goto do_prequeue;
1456
1457                         /* __ Set realtime policy in scheduler __ */
1458                 }
1459
1460                 if (copied >= target) {
1461                         /* Do not sleep, just process backlog. */
1462                         release_sock(sk);
1463                         lock_sock(sk);
1464                 } else
1465                         sk_wait_data(sk, &timeo);
1466
1467 #ifdef CONFIG_NET_DMA
1468                 tp->ucopy.wakeup = 0;
1469 #endif
1470
1471                 if (user_recv) {
1472                         int chunk;
1473
1474                         /* __ Restore normal policy in scheduler __ */
1475
1476                         if ((chunk = len - tp->ucopy.len) != 0) {
1477                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1478                                 len -= chunk;
1479                                 copied += chunk;
1480                         }
1481
1482                         if (tp->rcv_nxt == tp->copied_seq &&
1483                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1484 do_prequeue:
1485                                 tcp_prequeue_process(sk);
1486
1487                                 if ((chunk = len - tp->ucopy.len) != 0) {
1488                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1489                                         len -= chunk;
1490                                         copied += chunk;
1491                                 }
1492                         }
1493                 }
1494                 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1495                         if (net_ratelimit())
1496                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1497                                        current->comm, task_pid_nr(current));
1498                         peek_seq = tp->copied_seq;
1499                 }
1500                 continue;
1501
1502         found_ok_skb:
1503                 /* Ok so how much can we use? */
1504                 used = skb->len - offset;
1505                 if (len < used)
1506                         used = len;
1507
1508                 /* Do we have urgent data here? */
1509                 if (tp->urg_data) {
1510                         u32 urg_offset = tp->urg_seq - *seq;
1511                         if (urg_offset < used) {
1512                                 if (!urg_offset) {
1513                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1514                                                 ++*seq;
1515                                                 offset++;
1516                                                 used--;
1517                                                 if (!used)
1518                                                         goto skip_copy;
1519                                         }
1520                                 } else
1521                                         used = urg_offset;
1522                         }
1523                 }
1524
1525                 if (!(flags & MSG_TRUNC)) {
1526 #ifdef CONFIG_NET_DMA
1527                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1528                                 tp->ucopy.dma_chan = get_softnet_dma();
1529
1530                         if (tp->ucopy.dma_chan) {
1531                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1532                                         tp->ucopy.dma_chan, skb, offset,
1533                                         msg->msg_iov, used,
1534                                         tp->ucopy.pinned_list);
1535
1536                                 if (tp->ucopy.dma_cookie < 0) {
1537
1538                                         printk(KERN_ALERT "dma_cookie < 0\n");
1539
1540                                         /* Exception. Bailout! */
1541                                         if (!copied)
1542                                                 copied = -EFAULT;
1543                                         break;
1544                                 }
1545                                 if ((offset + used) == skb->len)
1546                                         copied_early = 1;
1547
1548                         } else
1549 #endif
1550                         {
1551                                 err = skb_copy_datagram_iovec(skb, offset,
1552                                                 msg->msg_iov, used);
1553                                 if (err) {
1554                                         /* Exception. Bailout! */
1555                                         if (!copied)
1556                                                 copied = -EFAULT;
1557                                         break;
1558                                 }
1559                         }
1560                 }
1561
1562                 *seq += used;
1563                 copied += used;
1564                 len -= used;
1565
1566                 tcp_rcv_space_adjust(sk);
1567
1568 skip_copy:
1569                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1570                         tp->urg_data = 0;
1571                         tcp_fast_path_check(sk);
1572                 }
1573                 if (used + offset < skb->len)
1574                         continue;
1575
1576                 if (tcp_hdr(skb)->fin)
1577                         goto found_fin_ok;
1578                 if (!(flags & MSG_PEEK)) {
1579                         sk_eat_skb(sk, skb, copied_early);
1580                         copied_early = 0;
1581                 }
1582                 continue;
1583
1584         found_fin_ok:
1585                 /* Process the FIN. */
1586                 ++*seq;
1587                 if (!(flags & MSG_PEEK)) {
1588                         sk_eat_skb(sk, skb, copied_early);
1589                         copied_early = 0;
1590                 }
1591                 break;
1592         } while (len > 0);
1593
1594         if (user_recv) {
1595                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1596                         int chunk;
1597
1598                         tp->ucopy.len = copied > 0 ? len : 0;
1599
1600                         tcp_prequeue_process(sk);
1601
1602                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1603                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1604                                 len -= chunk;
1605                                 copied += chunk;
1606                         }
1607                 }
1608
1609                 tp->ucopy.task = NULL;
1610                 tp->ucopy.len = 0;
1611         }
1612
1613 #ifdef CONFIG_NET_DMA
1614         if (tp->ucopy.dma_chan) {
1615                 dma_cookie_t done, used;
1616
1617                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1618
1619                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1620                                                  tp->ucopy.dma_cookie, &done,
1621                                                  &used) == DMA_IN_PROGRESS) {
1622                         /* do partial cleanup of sk_async_wait_queue */
1623                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1624                                (dma_async_is_complete(skb->dma_cookie, done,
1625                                                       used) == DMA_SUCCESS)) {
1626                                 __skb_dequeue(&sk->sk_async_wait_queue);
1627                                 kfree_skb(skb);
1628                         }
1629                 }
1630
1631                 /* Safe to free early-copied skbs now */
1632                 __skb_queue_purge(&sk->sk_async_wait_queue);
1633                 dma_chan_put(tp->ucopy.dma_chan);
1634                 tp->ucopy.dma_chan = NULL;
1635         }
1636         if (tp->ucopy.pinned_list) {
1637                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1638                 tp->ucopy.pinned_list = NULL;
1639         }
1640 #endif
1641
1642         /* According to UNIX98, msg_name/msg_namelen are ignored
1643          * on connected socket. I was just happy when found this 8) --ANK
1644          */
1645
1646         /* Clean up data we have read: This will do ACK frames. */
1647         tcp_cleanup_rbuf(sk, copied);
1648
1649         TCP_CHECK_TIMER(sk);
1650         release_sock(sk);
1651         return copied;
1652
1653 out:
1654         TCP_CHECK_TIMER(sk);
1655         release_sock(sk);
1656         return err;
1657
1658 recv_urg:
1659         err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1660         goto out;
1661 }
1662
1663 void tcp_set_state(struct sock *sk, int state)
1664 {
1665         int oldstate = sk->sk_state;
1666
1667         switch (state) {
1668         case TCP_ESTABLISHED:
1669                 if (oldstate != TCP_ESTABLISHED)
1670                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1671                 break;
1672
1673         case TCP_CLOSE:
1674                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1675                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1676
1677                 sk->sk_prot->unhash(sk);
1678                 if (inet_csk(sk)->icsk_bind_hash &&
1679                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1680                         inet_put_port(sk);
1681                 /* fall through */
1682         default:
1683                 if (oldstate == TCP_ESTABLISHED)
1684                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1685         }
1686
1687         /* Change state AFTER socket is unhashed to avoid closed
1688          * socket sitting in hash tables.
1689          */
1690         sk->sk_state = state;
1691
1692 #ifdef STATE_TRACE
1693         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1694 #endif
1695 }
1696 EXPORT_SYMBOL_GPL(tcp_set_state);
1697
1698 /*
1699  *      State processing on a close. This implements the state shift for
1700  *      sending our FIN frame. Note that we only send a FIN for some
1701  *      states. A shutdown() may have already sent the FIN, or we may be
1702  *      closed.
1703  */
1704
1705 static const unsigned char new_state[16] = {
1706   /* current state:        new state:      action:      */
1707   /* (Invalid)          */ TCP_CLOSE,
1708   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1709   /* TCP_SYN_SENT       */ TCP_CLOSE,
1710   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1711   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1712   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1713   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1714   /* TCP_CLOSE          */ TCP_CLOSE,
1715   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1716   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1717   /* TCP_LISTEN         */ TCP_CLOSE,
1718   /* TCP_CLOSING        */ TCP_CLOSING,
1719 };
1720
1721 static int tcp_close_state(struct sock *sk)
1722 {
1723         int next = (int)new_state[sk->sk_state];
1724         int ns = next & TCP_STATE_MASK;
1725
1726         tcp_set_state(sk, ns);
1727
1728         return next & TCP_ACTION_FIN;
1729 }
1730
1731 /*
1732  *      Shutdown the sending side of a connection. Much like close except
1733  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1734  */
1735
1736 void tcp_shutdown(struct sock *sk, int how)
1737 {
1738         /*      We need to grab some memory, and put together a FIN,
1739          *      and then put it into the queue to be sent.
1740          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1741          */
1742         if (!(how & SEND_SHUTDOWN))
1743                 return;
1744
1745         /* If we've already sent a FIN, or it's a closed state, skip this. */
1746         if ((1 << sk->sk_state) &
1747             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1748              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1749                 /* Clear out any half completed packets.  FIN if needed. */
1750                 if (tcp_close_state(sk))
1751                         tcp_send_fin(sk);
1752         }
1753 }
1754
1755 void tcp_close(struct sock *sk, long timeout)
1756 {
1757         struct sk_buff *skb;
1758         int data_was_unread = 0;
1759         int state;
1760
1761         lock_sock(sk);
1762         sk->sk_shutdown = SHUTDOWN_MASK;
1763
1764         if (sk->sk_state == TCP_LISTEN) {
1765                 tcp_set_state(sk, TCP_CLOSE);
1766
1767                 /* Special case. */
1768                 inet_csk_listen_stop(sk);
1769
1770                 goto adjudge_to_death;
1771         }
1772
1773         /*  We need to flush the recv. buffs.  We do this only on the
1774          *  descriptor close, not protocol-sourced closes, because the
1775          *  reader process may not have drained the data yet!
1776          */
1777         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1778                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1779                           tcp_hdr(skb)->fin;
1780                 data_was_unread += len;
1781                 __kfree_skb(skb);
1782         }
1783
1784         sk_mem_reclaim(sk);
1785
1786         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1787          * data was lost. To witness the awful effects of the old behavior of
1788          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1789          * GET in an FTP client, suspend the process, wait for the client to
1790          * advertise a zero window, then kill -9 the FTP client, wheee...
1791          * Note: timeout is always zero in such a case.
1792          */
1793         if (data_was_unread) {
1794                 /* Unread data was tossed, zap the connection. */
1795                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1796                 tcp_set_state(sk, TCP_CLOSE);
1797                 tcp_send_active_reset(sk, GFP_KERNEL);
1798         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1799                 /* Check zero linger _after_ checking for unread data. */
1800                 sk->sk_prot->disconnect(sk, 0);
1801                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1802         } else if (tcp_close_state(sk)) {
1803                 /* We FIN if the application ate all the data before
1804                  * zapping the connection.
1805                  */
1806
1807                 /* RED-PEN. Formally speaking, we have broken TCP state
1808                  * machine. State transitions:
1809                  *
1810                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1811                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1812                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1813                  *
1814                  * are legal only when FIN has been sent (i.e. in window),
1815                  * rather than queued out of window. Purists blame.
1816                  *
1817                  * F.e. "RFC state" is ESTABLISHED,
1818                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1819                  *
1820                  * The visible declinations are that sometimes
1821                  * we enter time-wait state, when it is not required really
1822                  * (harmless), do not send active resets, when they are
1823                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1824                  * they look as CLOSING or LAST_ACK for Linux)
1825                  * Probably, I missed some more holelets.
1826                  *                                              --ANK
1827                  */
1828                 tcp_send_fin(sk);
1829         }
1830
1831         sk_stream_wait_close(sk, timeout);
1832
1833 adjudge_to_death:
1834         state = sk->sk_state;
1835         sock_hold(sk);
1836         sock_orphan(sk);
1837         atomic_inc(sk->sk_prot->orphan_count);
1838
1839         /* It is the last release_sock in its life. It will remove backlog. */
1840         release_sock(sk);
1841
1842
1843         /* Now socket is owned by kernel and we acquire BH lock
1844            to finish close. No need to check for user refs.
1845          */
1846         local_bh_disable();
1847         bh_lock_sock(sk);
1848         WARN_ON(sock_owned_by_user(sk));
1849
1850         /* Have we already been destroyed by a softirq or backlog? */
1851         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1852                 goto out;
1853
1854         /*      This is a (useful) BSD violating of the RFC. There is a
1855          *      problem with TCP as specified in that the other end could
1856          *      keep a socket open forever with no application left this end.
1857          *      We use a 3 minute timeout (about the same as BSD) then kill
1858          *      our end. If they send after that then tough - BUT: long enough
1859          *      that we won't make the old 4*rto = almost no time - whoops
1860          *      reset mistake.
1861          *
1862          *      Nope, it was not mistake. It is really desired behaviour
1863          *      f.e. on http servers, when such sockets are useless, but
1864          *      consume significant resources. Let's do it with special
1865          *      linger2 option.                                 --ANK
1866          */
1867
1868         if (sk->sk_state == TCP_FIN_WAIT2) {
1869                 struct tcp_sock *tp = tcp_sk(sk);
1870                 if (tp->linger2 < 0) {
1871                         tcp_set_state(sk, TCP_CLOSE);
1872                         tcp_send_active_reset(sk, GFP_ATOMIC);
1873                         NET_INC_STATS_BH(sock_net(sk),
1874                                         LINUX_MIB_TCPABORTONLINGER);
1875                 } else {
1876                         const int tmo = tcp_fin_time(sk);
1877
1878                         if (tmo > TCP_TIMEWAIT_LEN) {
1879                                 inet_csk_reset_keepalive_timer(sk,
1880                                                 tmo - TCP_TIMEWAIT_LEN);
1881                         } else {
1882                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1883                                 goto out;
1884                         }
1885                 }
1886         }
1887         if (sk->sk_state != TCP_CLOSE) {
1888                 sk_mem_reclaim(sk);
1889                 if (tcp_too_many_orphans(sk,
1890                                 atomic_read(sk->sk_prot->orphan_count))) {
1891                         if (net_ratelimit())
1892                                 printk(KERN_INFO "TCP: too many of orphaned "
1893                                        "sockets\n");
1894                         tcp_set_state(sk, TCP_CLOSE);
1895                         tcp_send_active_reset(sk, GFP_ATOMIC);
1896                         NET_INC_STATS_BH(sock_net(sk),
1897                                         LINUX_MIB_TCPABORTONMEMORY);
1898                 }
1899         }
1900
1901         if (sk->sk_state == TCP_CLOSE)
1902                 inet_csk_destroy_sock(sk);
1903         /* Otherwise, socket is reprieved until protocol close. */
1904
1905 out:
1906         bh_unlock_sock(sk);
1907         local_bh_enable();
1908         sock_put(sk);
1909 }
1910
1911 /* These states need RST on ABORT according to RFC793 */
1912
1913 static inline int tcp_need_reset(int state)
1914 {
1915         return (1 << state) &
1916                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1917                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1918 }
1919
1920 int tcp_disconnect(struct sock *sk, int flags)
1921 {
1922         struct inet_sock *inet = inet_sk(sk);
1923         struct inet_connection_sock *icsk = inet_csk(sk);
1924         struct tcp_sock *tp = tcp_sk(sk);
1925         int err = 0;
1926         int old_state = sk->sk_state;
1927
1928         if (old_state != TCP_CLOSE)
1929                 tcp_set_state(sk, TCP_CLOSE);
1930
1931         /* ABORT function of RFC793 */
1932         if (old_state == TCP_LISTEN) {
1933                 inet_csk_listen_stop(sk);
1934         } else if (tcp_need_reset(old_state) ||
1935                    (tp->snd_nxt != tp->write_seq &&
1936                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1937                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1938                  * states
1939                  */
1940                 tcp_send_active_reset(sk, gfp_any());
1941                 sk->sk_err = ECONNRESET;
1942         } else if (old_state == TCP_SYN_SENT)
1943                 sk->sk_err = ECONNRESET;
1944
1945         tcp_clear_xmit_timers(sk);
1946         __skb_queue_purge(&sk->sk_receive_queue);
1947         tcp_write_queue_purge(sk);
1948         __skb_queue_purge(&tp->out_of_order_queue);
1949 #ifdef CONFIG_NET_DMA
1950         __skb_queue_purge(&sk->sk_async_wait_queue);
1951 #endif
1952
1953         inet->dport = 0;
1954
1955         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1956                 inet_reset_saddr(sk);
1957
1958         sk->sk_shutdown = 0;
1959         sock_reset_flag(sk, SOCK_DONE);
1960         tp->srtt = 0;
1961         if ((tp->write_seq += tp->max_window + 2) == 0)
1962                 tp->write_seq = 1;
1963         icsk->icsk_backoff = 0;
1964         tp->snd_cwnd = 2;
1965         icsk->icsk_probes_out = 0;
1966         tp->packets_out = 0;
1967         tp->snd_ssthresh = 0x7fffffff;
1968         tp->snd_cwnd_cnt = 0;
1969         tp->bytes_acked = 0;
1970         tcp_set_ca_state(sk, TCP_CA_Open);
1971         tcp_clear_retrans(tp);
1972         inet_csk_delack_init(sk);
1973         tcp_init_send_head(sk);
1974         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1975         __sk_dst_reset(sk);
1976
1977         WARN_ON(inet->num && !icsk->icsk_bind_hash);
1978
1979         sk->sk_error_report(sk);
1980         return err;
1981 }
1982
1983 /*
1984  *      Socket option code for TCP.
1985  */
1986 static int do_tcp_setsockopt(struct sock *sk, int level,
1987                 int optname, char __user *optval, int optlen)
1988 {
1989         struct tcp_sock *tp = tcp_sk(sk);
1990         struct inet_connection_sock *icsk = inet_csk(sk);
1991         int val;
1992         int err = 0;
1993
1994         /* This is a string value all the others are int's */
1995         if (optname == TCP_CONGESTION) {
1996                 char name[TCP_CA_NAME_MAX];
1997
1998                 if (optlen < 1)
1999                         return -EINVAL;
2000
2001                 val = strncpy_from_user(name, optval,
2002                                         min(TCP_CA_NAME_MAX-1, optlen));
2003                 if (val < 0)
2004                         return -EFAULT;
2005                 name[val] = 0;
2006
2007                 lock_sock(sk);
2008                 err = tcp_set_congestion_control(sk, name);
2009                 release_sock(sk);
2010                 return err;
2011         }
2012
2013         if (optlen < sizeof(int))
2014                 return -EINVAL;
2015
2016         if (get_user(val, (int __user *)optval))
2017                 return -EFAULT;
2018
2019         lock_sock(sk);
2020
2021         switch (optname) {
2022         case TCP_MAXSEG:
2023                 /* Values greater than interface MTU won't take effect. However
2024                  * at the point when this call is done we typically don't yet
2025                  * know which interface is going to be used */
2026                 if (val < 8 || val > MAX_TCP_WINDOW) {
2027                         err = -EINVAL;
2028                         break;
2029                 }
2030                 tp->rx_opt.user_mss = val;
2031                 break;
2032
2033         case TCP_NODELAY:
2034                 if (val) {
2035                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2036                          * this option on corked socket is remembered, but
2037                          * it is not activated until cork is cleared.
2038                          *
2039                          * However, when TCP_NODELAY is set we make
2040                          * an explicit push, which overrides even TCP_CORK
2041                          * for currently queued segments.
2042                          */
2043                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2044                         tcp_push_pending_frames(sk);
2045                 } else {
2046                         tp->nonagle &= ~TCP_NAGLE_OFF;
2047                 }
2048                 break;
2049
2050         case TCP_CORK:
2051                 /* When set indicates to always queue non-full frames.
2052                  * Later the user clears this option and we transmit
2053                  * any pending partial frames in the queue.  This is
2054                  * meant to be used alongside sendfile() to get properly
2055                  * filled frames when the user (for example) must write
2056                  * out headers with a write() call first and then use
2057                  * sendfile to send out the data parts.
2058                  *
2059                  * TCP_CORK can be set together with TCP_NODELAY and it is
2060                  * stronger than TCP_NODELAY.
2061                  */
2062                 if (val) {
2063                         tp->nonagle |= TCP_NAGLE_CORK;
2064                 } else {
2065                         tp->nonagle &= ~TCP_NAGLE_CORK;
2066                         if (tp->nonagle&TCP_NAGLE_OFF)
2067                                 tp->nonagle |= TCP_NAGLE_PUSH;
2068                         tcp_push_pending_frames(sk);
2069                 }
2070                 break;
2071
2072         case TCP_KEEPIDLE:
2073                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2074                         err = -EINVAL;
2075                 else {
2076                         tp->keepalive_time = val * HZ;
2077                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2078                             !((1 << sk->sk_state) &
2079                               (TCPF_CLOSE | TCPF_LISTEN))) {
2080                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2081                                 if (tp->keepalive_time > elapsed)
2082                                         elapsed = tp->keepalive_time - elapsed;
2083                                 else
2084                                         elapsed = 0;
2085                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2086                         }
2087                 }
2088                 break;
2089         case TCP_KEEPINTVL:
2090                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2091                         err = -EINVAL;
2092                 else
2093                         tp->keepalive_intvl = val * HZ;
2094                 break;
2095         case TCP_KEEPCNT:
2096                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2097                         err = -EINVAL;
2098                 else
2099                         tp->keepalive_probes = val;
2100                 break;
2101         case TCP_SYNCNT:
2102                 if (val < 1 || val > MAX_TCP_SYNCNT)
2103                         err = -EINVAL;
2104                 else
2105                         icsk->icsk_syn_retries = val;
2106                 break;
2107
2108         case TCP_LINGER2:
2109                 if (val < 0)
2110                         tp->linger2 = -1;
2111                 else if (val > sysctl_tcp_fin_timeout / HZ)
2112                         tp->linger2 = 0;
2113                 else
2114                         tp->linger2 = val * HZ;
2115                 break;
2116
2117         case TCP_DEFER_ACCEPT:
2118                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
2119                 if (val > 0) {
2120                         /* Translate value in seconds to number of
2121                          * retransmits */
2122                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
2123                                val > ((TCP_TIMEOUT_INIT / HZ) <<
2124                                        icsk->icsk_accept_queue.rskq_defer_accept))
2125                                 icsk->icsk_accept_queue.rskq_defer_accept++;
2126                         icsk->icsk_accept_queue.rskq_defer_accept++;
2127                 }
2128                 break;
2129
2130         case TCP_WINDOW_CLAMP:
2131                 if (!val) {
2132                         if (sk->sk_state != TCP_CLOSE) {
2133                                 err = -EINVAL;
2134                                 break;
2135                         }
2136                         tp->window_clamp = 0;
2137                 } else
2138                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2139                                                 SOCK_MIN_RCVBUF / 2 : val;
2140                 break;
2141
2142         case TCP_QUICKACK:
2143                 if (!val) {
2144                         icsk->icsk_ack.pingpong = 1;
2145                 } else {
2146                         icsk->icsk_ack.pingpong = 0;
2147                         if ((1 << sk->sk_state) &
2148                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2149                             inet_csk_ack_scheduled(sk)) {
2150                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2151                                 tcp_cleanup_rbuf(sk, 1);
2152                                 if (!(val & 1))
2153                                         icsk->icsk_ack.pingpong = 1;
2154                         }
2155                 }
2156                 break;
2157
2158 #ifdef CONFIG_TCP_MD5SIG
2159         case TCP_MD5SIG:
2160                 /* Read the IP->Key mappings from userspace */
2161                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2162                 break;
2163 #endif
2164
2165         default:
2166                 err = -ENOPROTOOPT;
2167                 break;
2168         }
2169
2170         release_sock(sk);
2171         return err;
2172 }
2173
2174 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2175                    int optlen)
2176 {
2177         struct inet_connection_sock *icsk = inet_csk(sk);
2178
2179         if (level != SOL_TCP)
2180                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2181                                                      optval, optlen);
2182         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2183 }
2184
2185 #ifdef CONFIG_COMPAT
2186 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2187                           char __user *optval, int optlen)
2188 {
2189         if (level != SOL_TCP)
2190                 return inet_csk_compat_setsockopt(sk, level, optname,
2191                                                   optval, optlen);
2192         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2193 }
2194
2195 EXPORT_SYMBOL(compat_tcp_setsockopt);
2196 #endif
2197
2198 /* Return information about state of tcp endpoint in API format. */
2199 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2200 {
2201         struct tcp_sock *tp = tcp_sk(sk);
2202         const struct inet_connection_sock *icsk = inet_csk(sk);
2203         u32 now = tcp_time_stamp;
2204
2205         memset(info, 0, sizeof(*info));
2206
2207         info->tcpi_state = sk->sk_state;
2208         info->tcpi_ca_state = icsk->icsk_ca_state;
2209         info->tcpi_retransmits = icsk->icsk_retransmits;
2210         info->tcpi_probes = icsk->icsk_probes_out;
2211         info->tcpi_backoff = icsk->icsk_backoff;
2212
2213         if (tp->rx_opt.tstamp_ok)
2214                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2215         if (tcp_is_sack(tp))
2216                 info->tcpi_options |= TCPI_OPT_SACK;
2217         if (tp->rx_opt.wscale_ok) {
2218                 info->tcpi_options |= TCPI_OPT_WSCALE;
2219                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2220                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2221         }
2222
2223         if (tp->ecn_flags&TCP_ECN_OK)
2224                 info->tcpi_options |= TCPI_OPT_ECN;
2225
2226         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2227         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2228         info->tcpi_snd_mss = tp->mss_cache;
2229         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2230
2231         if (sk->sk_state == TCP_LISTEN) {
2232                 info->tcpi_unacked = sk->sk_ack_backlog;
2233                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2234         } else {
2235                 info->tcpi_unacked = tp->packets_out;
2236                 info->tcpi_sacked = tp->sacked_out;
2237         }
2238         info->tcpi_lost = tp->lost_out;
2239         info->tcpi_retrans = tp->retrans_out;
2240         info->tcpi_fackets = tp->fackets_out;
2241
2242         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2243         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2244         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2245
2246         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2247         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2248         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2249         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2250         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2251         info->tcpi_snd_cwnd = tp->snd_cwnd;
2252         info->tcpi_advmss = tp->advmss;
2253         info->tcpi_reordering = tp->reordering;
2254
2255         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2256         info->tcpi_rcv_space = tp->rcvq_space.space;
2257
2258         info->tcpi_total_retrans = tp->total_retrans;
2259 }
2260
2261 EXPORT_SYMBOL_GPL(tcp_get_info);
2262
2263 static int do_tcp_getsockopt(struct sock *sk, int level,
2264                 int optname, char __user *optval, int __user *optlen)
2265 {
2266         struct inet_connection_sock *icsk = inet_csk(sk);
2267         struct tcp_sock *tp = tcp_sk(sk);
2268         int val, len;
2269
2270         if (get_user(len, optlen))
2271                 return -EFAULT;
2272
2273         len = min_t(unsigned int, len, sizeof(int));
2274
2275         if (len < 0)
2276                 return -EINVAL;
2277
2278         switch (optname) {
2279         case TCP_MAXSEG:
2280                 val = tp->mss_cache;
2281                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2282                         val = tp->rx_opt.user_mss;
2283                 break;
2284         case TCP_NODELAY:
2285                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2286                 break;
2287         case TCP_CORK:
2288                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2289                 break;
2290         case TCP_KEEPIDLE:
2291                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2292                 break;
2293         case TCP_KEEPINTVL:
2294                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2295                 break;
2296         case TCP_KEEPCNT:
2297                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2298                 break;
2299         case TCP_SYNCNT:
2300                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2301                 break;
2302         case TCP_LINGER2:
2303                 val = tp->linger2;
2304                 if (val >= 0)
2305                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2306                 break;
2307         case TCP_DEFER_ACCEPT:
2308                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2309                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2310                 break;
2311         case TCP_WINDOW_CLAMP:
2312                 val = tp->window_clamp;
2313                 break;
2314         case TCP_INFO: {
2315                 struct tcp_info info;
2316
2317                 if (get_user(len, optlen))
2318                         return -EFAULT;
2319
2320                 tcp_get_info(sk, &info);
2321
2322                 len = min_t(unsigned int, len, sizeof(info));
2323                 if (put_user(len, optlen))
2324                         return -EFAULT;
2325                 if (copy_to_user(optval, &info, len))
2326                         return -EFAULT;
2327                 return 0;
2328         }
2329         case TCP_QUICKACK:
2330                 val = !icsk->icsk_ack.pingpong;
2331                 break;
2332
2333         case TCP_CONGESTION:
2334                 if (get_user(len, optlen))
2335                         return -EFAULT;
2336                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2337                 if (put_user(len, optlen))
2338                         return -EFAULT;
2339                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2340                         return -EFAULT;
2341                 return 0;
2342         default:
2343                 return -ENOPROTOOPT;
2344         }
2345
2346         if (put_user(len, optlen))
2347                 return -EFAULT;
2348         if (copy_to_user(optval, &val, len))
2349                 return -EFAULT;
2350         return 0;
2351 }
2352
2353 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2354                    int __user *optlen)
2355 {
2356         struct inet_connection_sock *icsk = inet_csk(sk);
2357
2358         if (level != SOL_TCP)
2359                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2360                                                      optval, optlen);
2361         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2362 }
2363
2364 #ifdef CONFIG_COMPAT
2365 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2366                           char __user *optval, int __user *optlen)
2367 {
2368         if (level != SOL_TCP)
2369                 return inet_csk_compat_getsockopt(sk, level, optname,
2370                                                   optval, optlen);
2371         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2372 }
2373
2374 EXPORT_SYMBOL(compat_tcp_getsockopt);
2375 #endif
2376
2377 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2378 {
2379         struct sk_buff *segs = ERR_PTR(-EINVAL);
2380         struct tcphdr *th;
2381         unsigned thlen;
2382         unsigned int seq;
2383         __be32 delta;
2384         unsigned int oldlen;
2385         unsigned int len;
2386
2387         if (!pskb_may_pull(skb, sizeof(*th)))
2388                 goto out;
2389
2390         th = tcp_hdr(skb);
2391         thlen = th->doff * 4;
2392         if (thlen < sizeof(*th))
2393                 goto out;
2394
2395         if (!pskb_may_pull(skb, thlen))
2396                 goto out;
2397
2398         oldlen = (u16)~skb->len;
2399         __skb_pull(skb, thlen);
2400
2401         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2402                 /* Packet is from an untrusted source, reset gso_segs. */
2403                 int type = skb_shinfo(skb)->gso_type;
2404                 int mss;
2405
2406                 if (unlikely(type &
2407                              ~(SKB_GSO_TCPV4 |
2408                                SKB_GSO_DODGY |
2409                                SKB_GSO_TCP_ECN |
2410                                SKB_GSO_TCPV6 |
2411                                0) ||
2412                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2413                         goto out;
2414
2415                 mss = skb_shinfo(skb)->gso_size;
2416                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2417
2418                 segs = NULL;
2419                 goto out;
2420         }
2421
2422         segs = skb_segment(skb, features);
2423         if (IS_ERR(segs))
2424                 goto out;
2425
2426         len = skb_shinfo(skb)->gso_size;
2427         delta = htonl(oldlen + (thlen + len));
2428
2429         skb = segs;
2430         th = tcp_hdr(skb);
2431         seq = ntohl(th->seq);
2432
2433         do {
2434                 th->fin = th->psh = 0;
2435
2436                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2437                                        (__force u32)delta));
2438                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2439                         th->check =
2440                              csum_fold(csum_partial(skb_transport_header(skb),
2441                                                     thlen, skb->csum));
2442
2443                 seq += len;
2444                 skb = skb->next;
2445                 th = tcp_hdr(skb);
2446
2447                 th->seq = htonl(seq);
2448                 th->cwr = 0;
2449         } while (skb->next);
2450
2451         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2452                       skb->data_len);
2453         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2454                                 (__force u32)delta));
2455         if (skb->ip_summed != CHECKSUM_PARTIAL)
2456                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2457                                                    thlen, skb->csum));
2458
2459 out:
2460         return segs;
2461 }
2462 EXPORT_SYMBOL(tcp_tso_segment);
2463
2464 #ifdef CONFIG_TCP_MD5SIG
2465 static unsigned long tcp_md5sig_users;
2466 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2467 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2468
2469 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2470 {
2471         int cpu;
2472         for_each_possible_cpu(cpu) {
2473                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2474                 if (p) {
2475                         if (p->md5_desc.tfm)
2476                                 crypto_free_hash(p->md5_desc.tfm);
2477                         kfree(p);
2478                         p = NULL;
2479                 }
2480         }
2481         free_percpu(pool);
2482 }
2483
2484 void tcp_free_md5sig_pool(void)
2485 {
2486         struct tcp_md5sig_pool **pool = NULL;
2487
2488         spin_lock_bh(&tcp_md5sig_pool_lock);
2489         if (--tcp_md5sig_users == 0) {
2490                 pool = tcp_md5sig_pool;
2491                 tcp_md5sig_pool = NULL;
2492         }
2493         spin_unlock_bh(&tcp_md5sig_pool_lock);
2494         if (pool)
2495                 __tcp_free_md5sig_pool(pool);
2496 }
2497
2498 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2499
2500 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2501 {
2502         int cpu;
2503         struct tcp_md5sig_pool **pool;
2504
2505         pool = alloc_percpu(struct tcp_md5sig_pool *);
2506         if (!pool)
2507                 return NULL;
2508
2509         for_each_possible_cpu(cpu) {
2510                 struct tcp_md5sig_pool *p;
2511                 struct crypto_hash *hash;
2512
2513                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2514                 if (!p)
2515                         goto out_free;
2516                 *per_cpu_ptr(pool, cpu) = p;
2517
2518                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2519                 if (!hash || IS_ERR(hash))
2520                         goto out_free;
2521
2522                 p->md5_desc.tfm = hash;
2523         }
2524         return pool;
2525 out_free:
2526         __tcp_free_md5sig_pool(pool);
2527         return NULL;
2528 }
2529
2530 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2531 {
2532         struct tcp_md5sig_pool **pool;
2533         int alloc = 0;
2534
2535 retry:
2536         spin_lock_bh(&tcp_md5sig_pool_lock);
2537         pool = tcp_md5sig_pool;
2538         if (tcp_md5sig_users++ == 0) {
2539                 alloc = 1;
2540                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2541         } else if (!pool) {
2542                 tcp_md5sig_users--;
2543                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2544                 cpu_relax();
2545                 goto retry;
2546         } else
2547                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2548
2549         if (alloc) {
2550                 /* we cannot hold spinlock here because this may sleep. */
2551                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2552                 spin_lock_bh(&tcp_md5sig_pool_lock);
2553                 if (!p) {
2554                         tcp_md5sig_users--;
2555                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2556                         return NULL;
2557                 }
2558                 pool = tcp_md5sig_pool;
2559                 if (pool) {
2560                         /* oops, it has already been assigned. */
2561                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2562                         __tcp_free_md5sig_pool(p);
2563                 } else {
2564                         tcp_md5sig_pool = pool = p;
2565                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2566                 }
2567         }
2568         return pool;
2569 }
2570
2571 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2572
2573 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2574 {
2575         struct tcp_md5sig_pool **p;
2576         spin_lock_bh(&tcp_md5sig_pool_lock);
2577         p = tcp_md5sig_pool;
2578         if (p)
2579                 tcp_md5sig_users++;
2580         spin_unlock_bh(&tcp_md5sig_pool_lock);
2581         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2582 }
2583
2584 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2585
2586 void __tcp_put_md5sig_pool(void)
2587 {
2588         tcp_free_md5sig_pool();
2589 }
2590
2591 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2592
2593 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2594                         struct tcphdr *th)
2595 {
2596         struct scatterlist sg;
2597         int err;
2598
2599         __sum16 old_checksum = th->check;
2600         th->check = 0;
2601         /* options aren't included in the hash */
2602         sg_init_one(&sg, th, sizeof(struct tcphdr));
2603         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2604         th->check = old_checksum;
2605         return err;
2606 }
2607
2608 EXPORT_SYMBOL(tcp_md5_hash_header);
2609
2610 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2611                           struct sk_buff *skb, unsigned header_len)
2612 {
2613         struct scatterlist sg;
2614         const struct tcphdr *tp = tcp_hdr(skb);
2615         struct hash_desc *desc = &hp->md5_desc;
2616         unsigned i;
2617         const unsigned head_data_len = skb_headlen(skb) > header_len ?
2618                                        skb_headlen(skb) - header_len : 0;
2619         const struct skb_shared_info *shi = skb_shinfo(skb);
2620
2621         sg_init_table(&sg, 1);
2622
2623         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2624         if (crypto_hash_update(desc, &sg, head_data_len))
2625                 return 1;
2626
2627         for (i = 0; i < shi->nr_frags; ++i) {
2628                 const struct skb_frag_struct *f = &shi->frags[i];
2629                 sg_set_page(&sg, f->page, f->size, f->page_offset);
2630                 if (crypto_hash_update(desc, &sg, f->size))
2631                         return 1;
2632         }
2633
2634         return 0;
2635 }
2636
2637 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2638
2639 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
2640 {
2641         struct scatterlist sg;
2642
2643         sg_init_one(&sg, key->key, key->keylen);
2644         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2645 }
2646
2647 EXPORT_SYMBOL(tcp_md5_hash_key);
2648
2649 #endif
2650
2651 void tcp_done(struct sock *sk)
2652 {
2653         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2654                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2655
2656         tcp_set_state(sk, TCP_CLOSE);
2657         tcp_clear_xmit_timers(sk);
2658
2659         sk->sk_shutdown = SHUTDOWN_MASK;
2660
2661         if (!sock_flag(sk, SOCK_DEAD))
2662                 sk->sk_state_change(sk);
2663         else
2664                 inet_csk_destroy_sock(sk);
2665 }
2666 EXPORT_SYMBOL_GPL(tcp_done);
2667
2668 extern struct tcp_congestion_ops tcp_reno;
2669
2670 static __initdata unsigned long thash_entries;
2671 static int __init set_thash_entries(char *str)
2672 {
2673         if (!str)
2674                 return 0;
2675         thash_entries = simple_strtoul(str, &str, 0);
2676         return 1;
2677 }
2678 __setup("thash_entries=", set_thash_entries);
2679
2680 void __init tcp_init(void)
2681 {
2682         struct sk_buff *skb = NULL;
2683         unsigned long nr_pages, limit;
2684         int order, i, max_share;
2685
2686         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2687
2688         tcp_hashinfo.bind_bucket_cachep =
2689                 kmem_cache_create("tcp_bind_bucket",
2690                                   sizeof(struct inet_bind_bucket), 0,
2691                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2692
2693         /* Size and allocate the main established and bind bucket
2694          * hash tables.
2695          *
2696          * The methodology is similar to that of the buffer cache.
2697          */
2698         tcp_hashinfo.ehash =
2699                 alloc_large_system_hash("TCP established",
2700                                         sizeof(struct inet_ehash_bucket),
2701                                         thash_entries,
2702                                         (num_physpages >= 128 * 1024) ?
2703                                         13 : 15,
2704                                         0,
2705                                         &tcp_hashinfo.ehash_size,
2706                                         NULL,
2707                                         thash_entries ? 0 : 512 * 1024);
2708         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2709         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2710                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
2711                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
2712         }
2713         if (inet_ehash_locks_alloc(&tcp_hashinfo))
2714                 panic("TCP: failed to alloc ehash_locks");
2715         tcp_hashinfo.bhash =
2716                 alloc_large_system_hash("TCP bind",
2717                                         sizeof(struct inet_bind_hashbucket),
2718                                         tcp_hashinfo.ehash_size,
2719                                         (num_physpages >= 128 * 1024) ?
2720                                         13 : 15,
2721                                         0,
2722                                         &tcp_hashinfo.bhash_size,
2723                                         NULL,
2724                                         64 * 1024);
2725         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2726         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2727                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2728                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2729         }
2730
2731         /* Try to be a bit smarter and adjust defaults depending
2732          * on available memory.
2733          */
2734         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2735                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2736                         order++)
2737                 ;
2738         if (order >= 4) {
2739                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2740                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2741                 sysctl_max_syn_backlog = 1024;
2742         } else if (order < 3) {
2743                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2744                 sysctl_tcp_max_orphans >>= (3 - order);
2745                 sysctl_max_syn_backlog = 128;
2746         }
2747
2748         /* Set the pressure threshold to be a fraction of global memory that
2749          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2750          * memory, with a floor of 128 pages.
2751          */
2752         nr_pages = totalram_pages - totalhigh_pages;
2753         limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2754         limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2755         limit = max(limit, 128UL);
2756         sysctl_tcp_mem[0] = limit / 4 * 3;
2757         sysctl_tcp_mem[1] = limit;
2758         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2759
2760         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2761         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2762         max_share = min(4UL*1024*1024, limit);
2763
2764         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2765         sysctl_tcp_wmem[1] = 16*1024;
2766         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2767
2768         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2769         sysctl_tcp_rmem[1] = 87380;
2770         sysctl_tcp_rmem[2] = max(87380, max_share);
2771
2772         printk(KERN_INFO "TCP: Hash tables configured "
2773                "(established %d bind %d)\n",
2774                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2775
2776         tcp_register_congestion_control(&tcp_reno);
2777 }
2778
2779 EXPORT_SYMBOL(tcp_close);
2780 EXPORT_SYMBOL(tcp_disconnect);
2781 EXPORT_SYMBOL(tcp_getsockopt);
2782 EXPORT_SYMBOL(tcp_ioctl);
2783 EXPORT_SYMBOL(tcp_poll);
2784 EXPORT_SYMBOL(tcp_read_sock);
2785 EXPORT_SYMBOL(tcp_recvmsg);
2786 EXPORT_SYMBOL(tcp_sendmsg);
2787 EXPORT_SYMBOL(tcp_splice_read);
2788 EXPORT_SYMBOL(tcp_sendpage);
2789 EXPORT_SYMBOL(tcp_setsockopt);
2790 EXPORT_SYMBOL(tcp_shutdown);