Merge branch 'nf' of git://1984.lsi.us.es/net
[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 #include <linux/time.h>
268 #include <linux/slab.h>
269
270 #include <net/icmp.h>
271 #include <net/tcp.h>
272 #include <net/xfrm.h>
273 #include <net/ip.h>
274 #include <net/netdma.h>
275 #include <net/sock.h>
276
277 #include <asm/uaccess.h>
278 #include <asm/ioctls.h>
279
280 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
281
282 struct percpu_counter tcp_orphan_count;
283 EXPORT_SYMBOL_GPL(tcp_orphan_count);
284
285 long sysctl_tcp_mem[3] __read_mostly;
286 int sysctl_tcp_wmem[3] __read_mostly;
287 int sysctl_tcp_rmem[3] __read_mostly;
288
289 EXPORT_SYMBOL(sysctl_tcp_mem);
290 EXPORT_SYMBOL(sysctl_tcp_rmem);
291 EXPORT_SYMBOL(sysctl_tcp_wmem);
292
293 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
294 EXPORT_SYMBOL(tcp_memory_allocated);
295
296 /*
297  * Current number of TCP sockets.
298  */
299 struct percpu_counter tcp_sockets_allocated;
300 EXPORT_SYMBOL(tcp_sockets_allocated);
301
302 /*
303  * TCP splice context
304  */
305 struct tcp_splice_state {
306         struct pipe_inode_info *pipe;
307         size_t len;
308         unsigned int flags;
309 };
310
311 /*
312  * Pressure flag: try to collapse.
313  * Technical note: it is used by multiple contexts non atomically.
314  * All the __sk_mem_schedule() is of this nature: accounting
315  * is strict, actions are advisory and have some latency.
316  */
317 int tcp_memory_pressure __read_mostly;
318 EXPORT_SYMBOL(tcp_memory_pressure);
319
320 void tcp_enter_memory_pressure(struct sock *sk)
321 {
322         if (!tcp_memory_pressure) {
323                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
324                 tcp_memory_pressure = 1;
325         }
326 }
327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
328
329 /* Convert seconds to retransmits based on initial and max timeout */
330 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
331 {
332         u8 res = 0;
333
334         if (seconds > 0) {
335                 int period = timeout;
336
337                 res = 1;
338                 while (seconds > period && res < 255) {
339                         res++;
340                         timeout <<= 1;
341                         if (timeout > rto_max)
342                                 timeout = rto_max;
343                         period += timeout;
344                 }
345         }
346         return res;
347 }
348
349 /* Convert retransmits to seconds based on initial and max timeout */
350 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
351 {
352         int period = 0;
353
354         if (retrans > 0) {
355                 period = timeout;
356                 while (--retrans) {
357                         timeout <<= 1;
358                         if (timeout > rto_max)
359                                 timeout = rto_max;
360                         period += timeout;
361                 }
362         }
363         return period;
364 }
365
366 /*
367  *      Wait for a TCP event.
368  *
369  *      Note that we don't need to lock the socket, as the upper poll layers
370  *      take care of normal races (between the test and the event) and we don't
371  *      go look at any of the socket buffers directly.
372  */
373 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
374 {
375         unsigned int mask;
376         struct sock *sk = sock->sk;
377         const struct tcp_sock *tp = tcp_sk(sk);
378
379         sock_poll_wait(file, sk_sleep(sk), wait);
380         if (sk->sk_state == TCP_LISTEN)
381                 return inet_csk_listen_poll(sk);
382
383         /* Socket is not locked. We are protected from async events
384          * by poll logic and correct handling of state changes
385          * made by other threads is impossible in any case.
386          */
387
388         mask = 0;
389
390         /*
391          * POLLHUP is certainly not done right. But poll() doesn't
392          * have a notion of HUP in just one direction, and for a
393          * socket the read side is more interesting.
394          *
395          * Some poll() documentation says that POLLHUP is incompatible
396          * with the POLLOUT/POLLWR flags, so somebody should check this
397          * all. But careful, it tends to be safer to return too many
398          * bits than too few, and you can easily break real applications
399          * if you don't tell them that something has hung up!
400          *
401          * Check-me.
402          *
403          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
404          * our fs/select.c). It means that after we received EOF,
405          * poll always returns immediately, making impossible poll() on write()
406          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
407          * if and only if shutdown has been made in both directions.
408          * Actually, it is interesting to look how Solaris and DUX
409          * solve this dilemma. I would prefer, if POLLHUP were maskable,
410          * then we could set it on SND_SHUTDOWN. BTW examples given
411          * in Stevens' books assume exactly this behaviour, it explains
412          * why POLLHUP is incompatible with POLLOUT.    --ANK
413          *
414          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
415          * blocking on fresh not-connected or disconnected socket. --ANK
416          */
417         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
418                 mask |= POLLHUP;
419         if (sk->sk_shutdown & RCV_SHUTDOWN)
420                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
421
422         /* Connected? */
423         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
424                 int target = sock_rcvlowat(sk, 0, INT_MAX);
425
426                 if (tp->urg_seq == tp->copied_seq &&
427                     !sock_flag(sk, SOCK_URGINLINE) &&
428                     tp->urg_data)
429                         target++;
430
431                 /* Potential race condition. If read of tp below will
432                  * escape above sk->sk_state, we can be illegally awaken
433                  * in SYN_* states. */
434                 if (tp->rcv_nxt - tp->copied_seq >= target)
435                         mask |= POLLIN | POLLRDNORM;
436
437                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
438                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
439                                 mask |= POLLOUT | POLLWRNORM;
440                         } else {  /* send SIGIO later */
441                                 set_bit(SOCK_ASYNC_NOSPACE,
442                                         &sk->sk_socket->flags);
443                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
444
445                                 /* Race breaker. If space is freed after
446                                  * wspace test but before the flags are set,
447                                  * IO signal will be lost.
448                                  */
449                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
450                                         mask |= POLLOUT | POLLWRNORM;
451                         }
452                 } else
453                         mask |= POLLOUT | POLLWRNORM;
454
455                 if (tp->urg_data & TCP_URG_VALID)
456                         mask |= POLLPRI;
457         }
458         /* This barrier is coupled with smp_wmb() in tcp_reset() */
459         smp_rmb();
460         if (sk->sk_err)
461                 mask |= POLLERR;
462
463         return mask;
464 }
465 EXPORT_SYMBOL(tcp_poll);
466
467 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
468 {
469         struct tcp_sock *tp = tcp_sk(sk);
470         int answ;
471
472         switch (cmd) {
473         case SIOCINQ:
474                 if (sk->sk_state == TCP_LISTEN)
475                         return -EINVAL;
476
477                 lock_sock(sk);
478                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
479                         answ = 0;
480                 else if (sock_flag(sk, SOCK_URGINLINE) ||
481                          !tp->urg_data ||
482                          before(tp->urg_seq, tp->copied_seq) ||
483                          !before(tp->urg_seq, tp->rcv_nxt)) {
484                         struct sk_buff *skb;
485
486                         answ = tp->rcv_nxt - tp->copied_seq;
487
488                         /* Subtract 1, if FIN is in queue. */
489                         skb = skb_peek_tail(&sk->sk_receive_queue);
490                         if (answ && skb)
491                                 answ -= tcp_hdr(skb)->fin;
492                 } else
493                         answ = tp->urg_seq - tp->copied_seq;
494                 release_sock(sk);
495                 break;
496         case SIOCATMARK:
497                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
498                 break;
499         case SIOCOUTQ:
500                 if (sk->sk_state == TCP_LISTEN)
501                         return -EINVAL;
502
503                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
504                         answ = 0;
505                 else
506                         answ = tp->write_seq - tp->snd_una;
507                 break;
508         case SIOCOUTQNSD:
509                 if (sk->sk_state == TCP_LISTEN)
510                         return -EINVAL;
511
512                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
513                         answ = 0;
514                 else
515                         answ = tp->write_seq - tp->snd_nxt;
516                 break;
517         default:
518                 return -ENOIOCTLCMD;
519         }
520
521         return put_user(answ, (int __user *)arg);
522 }
523 EXPORT_SYMBOL(tcp_ioctl);
524
525 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
526 {
527         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
528         tp->pushed_seq = tp->write_seq;
529 }
530
531 static inline int forced_push(const struct tcp_sock *tp)
532 {
533         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
534 }
535
536 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
537 {
538         struct tcp_sock *tp = tcp_sk(sk);
539         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
540
541         skb->csum    = 0;
542         tcb->seq     = tcb->end_seq = tp->write_seq;
543         tcb->tcp_flags = TCPHDR_ACK;
544         tcb->sacked  = 0;
545         skb_header_release(skb);
546         tcp_add_write_queue_tail(sk, skb);
547         sk->sk_wmem_queued += skb->truesize;
548         sk_mem_charge(sk, skb->truesize);
549         if (tp->nonagle & TCP_NAGLE_PUSH)
550                 tp->nonagle &= ~TCP_NAGLE_PUSH;
551 }
552
553 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
554 {
555         if (flags & MSG_OOB)
556                 tp->snd_up = tp->write_seq;
557 }
558
559 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
560                             int nonagle)
561 {
562         if (tcp_send_head(sk)) {
563                 struct tcp_sock *tp = tcp_sk(sk);
564
565                 if (!(flags & MSG_MORE) || forced_push(tp))
566                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
567
568                 tcp_mark_urg(tp, flags);
569                 __tcp_push_pending_frames(sk, mss_now,
570                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
571         }
572 }
573
574 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
575                                 unsigned int offset, size_t len)
576 {
577         struct tcp_splice_state *tss = rd_desc->arg.data;
578         int ret;
579
580         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
581                               tss->flags);
582         if (ret > 0)
583                 rd_desc->count -= ret;
584         return ret;
585 }
586
587 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
588 {
589         /* Store TCP splice context information in read_descriptor_t. */
590         read_descriptor_t rd_desc = {
591                 .arg.data = tss,
592                 .count    = tss->len,
593         };
594
595         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
596 }
597
598 /**
599  *  tcp_splice_read - splice data from TCP socket to a pipe
600  * @sock:       socket to splice from
601  * @ppos:       position (not valid)
602  * @pipe:       pipe to splice to
603  * @len:        number of bytes to splice
604  * @flags:      splice modifier flags
605  *
606  * Description:
607  *    Will read pages from given socket and fill them into a pipe.
608  *
609  **/
610 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
611                         struct pipe_inode_info *pipe, size_t len,
612                         unsigned int flags)
613 {
614         struct sock *sk = sock->sk;
615         struct tcp_splice_state tss = {
616                 .pipe = pipe,
617                 .len = len,
618                 .flags = flags,
619         };
620         long timeo;
621         ssize_t spliced;
622         int ret;
623
624         sock_rps_record_flow(sk);
625         /*
626          * We can't seek on a socket input
627          */
628         if (unlikely(*ppos))
629                 return -ESPIPE;
630
631         ret = spliced = 0;
632
633         lock_sock(sk);
634
635         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
636         while (tss.len) {
637                 ret = __tcp_splice_read(sk, &tss);
638                 if (ret < 0)
639                         break;
640                 else if (!ret) {
641                         if (spliced)
642                                 break;
643                         if (sock_flag(sk, SOCK_DONE))
644                                 break;
645                         if (sk->sk_err) {
646                                 ret = sock_error(sk);
647                                 break;
648                         }
649                         if (sk->sk_shutdown & RCV_SHUTDOWN)
650                                 break;
651                         if (sk->sk_state == TCP_CLOSE) {
652                                 /*
653                                  * This occurs when user tries to read
654                                  * from never connected socket.
655                                  */
656                                 if (!sock_flag(sk, SOCK_DONE))
657                                         ret = -ENOTCONN;
658                                 break;
659                         }
660                         if (!timeo) {
661                                 ret = -EAGAIN;
662                                 break;
663                         }
664                         sk_wait_data(sk, &timeo);
665                         if (signal_pending(current)) {
666                                 ret = sock_intr_errno(timeo);
667                                 break;
668                         }
669                         continue;
670                 }
671                 tss.len -= ret;
672                 spliced += ret;
673
674                 if (!timeo)
675                         break;
676                 release_sock(sk);
677                 lock_sock(sk);
678
679                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
680                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
681                     signal_pending(current))
682                         break;
683         }
684
685         release_sock(sk);
686
687         if (spliced)
688                 return spliced;
689
690         return ret;
691 }
692 EXPORT_SYMBOL(tcp_splice_read);
693
694 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
695 {
696         struct sk_buff *skb;
697
698         /* The TCP header must be at least 32-bit aligned.  */
699         size = ALIGN(size, 4);
700
701         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
702         if (skb) {
703                 if (sk_wmem_schedule(sk, skb->truesize)) {
704                         /*
705                          * Make sure that we have exactly size bytes
706                          * available to the caller, no more, no less.
707                          */
708                         skb_reserve(skb, skb_tailroom(skb) - size);
709                         return skb;
710                 }
711                 __kfree_skb(skb);
712         } else {
713                 sk->sk_prot->enter_memory_pressure(sk);
714                 sk_stream_moderate_sndbuf(sk);
715         }
716         return NULL;
717 }
718
719 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
720                                        int large_allowed)
721 {
722         struct tcp_sock *tp = tcp_sk(sk);
723         u32 xmit_size_goal, old_size_goal;
724
725         xmit_size_goal = mss_now;
726
727         if (large_allowed && sk_can_gso(sk)) {
728                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
729                                   inet_csk(sk)->icsk_af_ops->net_header_len -
730                                   inet_csk(sk)->icsk_ext_hdr_len -
731                                   tp->tcp_header_len);
732
733                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
734
735                 /* We try hard to avoid divides here */
736                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
737
738                 if (likely(old_size_goal <= xmit_size_goal &&
739                            old_size_goal + mss_now > xmit_size_goal)) {
740                         xmit_size_goal = old_size_goal;
741                 } else {
742                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
743                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
744                 }
745         }
746
747         return max(xmit_size_goal, mss_now);
748 }
749
750 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
751 {
752         int mss_now;
753
754         mss_now = tcp_current_mss(sk);
755         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
756
757         return mss_now;
758 }
759
760 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
761                          size_t psize, int flags)
762 {
763         struct tcp_sock *tp = tcp_sk(sk);
764         int mss_now, size_goal;
765         int err;
766         ssize_t copied;
767         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
768
769         /* Wait for a connection to finish. */
770         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
771                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
772                         goto out_err;
773
774         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
775
776         mss_now = tcp_send_mss(sk, &size_goal, flags);
777         copied = 0;
778
779         err = -EPIPE;
780         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
781                 goto out_err;
782
783         while (psize > 0) {
784                 struct sk_buff *skb = tcp_write_queue_tail(sk);
785                 struct page *page = pages[poffset / PAGE_SIZE];
786                 int copy, i, can_coalesce;
787                 int offset = poffset % PAGE_SIZE;
788                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
789
790                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
791 new_segment:
792                         if (!sk_stream_memory_free(sk))
793                                 goto wait_for_sndbuf;
794
795                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
796                         if (!skb)
797                                 goto wait_for_memory;
798
799                         skb_entail(sk, skb);
800                         copy = size_goal;
801                 }
802
803                 if (copy > size)
804                         copy = size;
805
806                 i = skb_shinfo(skb)->nr_frags;
807                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
808                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
809                         tcp_mark_push(tp, skb);
810                         goto new_segment;
811                 }
812                 if (!sk_wmem_schedule(sk, copy))
813                         goto wait_for_memory;
814
815                 if (can_coalesce) {
816                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
817                 } else {
818                         get_page(page);
819                         skb_fill_page_desc(skb, i, page, offset, copy);
820                 }
821
822                 skb->len += copy;
823                 skb->data_len += copy;
824                 skb->truesize += copy;
825                 sk->sk_wmem_queued += copy;
826                 sk_mem_charge(sk, copy);
827                 skb->ip_summed = CHECKSUM_PARTIAL;
828                 tp->write_seq += copy;
829                 TCP_SKB_CB(skb)->end_seq += copy;
830                 skb_shinfo(skb)->gso_segs = 0;
831
832                 if (!copied)
833                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
834
835                 copied += copy;
836                 poffset += copy;
837                 if (!(psize -= copy))
838                         goto out;
839
840                 if (skb->len < size_goal || (flags & MSG_OOB))
841                         continue;
842
843                 if (forced_push(tp)) {
844                         tcp_mark_push(tp, skb);
845                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
846                 } else if (skb == tcp_send_head(sk))
847                         tcp_push_one(sk, mss_now);
848                 continue;
849
850 wait_for_sndbuf:
851                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
852 wait_for_memory:
853                 if (copied)
854                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
855
856                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
857                         goto do_error;
858
859                 mss_now = tcp_send_mss(sk, &size_goal, flags);
860         }
861
862 out:
863         if (copied)
864                 tcp_push(sk, flags, mss_now, tp->nonagle);
865         return copied;
866
867 do_error:
868         if (copied)
869                 goto out;
870 out_err:
871         return sk_stream_error(sk, flags, err);
872 }
873
874 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
875                  size_t size, int flags)
876 {
877         ssize_t res;
878
879         if (!(sk->sk_route_caps & NETIF_F_SG) ||
880             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
881                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
882                                         flags);
883
884         lock_sock(sk);
885         res = do_tcp_sendpages(sk, &page, offset, size, flags);
886         release_sock(sk);
887         return res;
888 }
889 EXPORT_SYMBOL(tcp_sendpage);
890
891 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
892 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
893
894 static inline int select_size(const struct sock *sk, int sg)
895 {
896         const struct tcp_sock *tp = tcp_sk(sk);
897         int tmp = tp->mss_cache;
898
899         if (sg) {
900                 if (sk_can_gso(sk))
901                         tmp = 0;
902                 else {
903                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
904
905                         if (tmp >= pgbreak &&
906                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
907                                 tmp = pgbreak;
908                 }
909         }
910
911         return tmp;
912 }
913
914 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
915                 size_t size)
916 {
917         struct iovec *iov;
918         struct tcp_sock *tp = tcp_sk(sk);
919         struct sk_buff *skb;
920         int iovlen, flags;
921         int mss_now, size_goal;
922         int sg, err, copied;
923         long timeo;
924
925         lock_sock(sk);
926
927         flags = msg->msg_flags;
928         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
929
930         /* Wait for a connection to finish. */
931         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
932                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
933                         goto out_err;
934
935         /* This should be in poll */
936         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
937
938         mss_now = tcp_send_mss(sk, &size_goal, flags);
939
940         /* Ok commence sending. */
941         iovlen = msg->msg_iovlen;
942         iov = msg->msg_iov;
943         copied = 0;
944
945         err = -EPIPE;
946         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
947                 goto out_err;
948
949         sg = sk->sk_route_caps & NETIF_F_SG;
950
951         while (--iovlen >= 0) {
952                 size_t seglen = iov->iov_len;
953                 unsigned char __user *from = iov->iov_base;
954
955                 iov++;
956
957                 while (seglen > 0) {
958                         int copy = 0;
959                         int max = size_goal;
960
961                         skb = tcp_write_queue_tail(sk);
962                         if (tcp_send_head(sk)) {
963                                 if (skb->ip_summed == CHECKSUM_NONE)
964                                         max = mss_now;
965                                 copy = max - skb->len;
966                         }
967
968                         if (copy <= 0) {
969 new_segment:
970                                 /* Allocate new segment. If the interface is SG,
971                                  * allocate skb fitting to single page.
972                                  */
973                                 if (!sk_stream_memory_free(sk))
974                                         goto wait_for_sndbuf;
975
976                                 skb = sk_stream_alloc_skb(sk,
977                                                           select_size(sk, sg),
978                                                           sk->sk_allocation);
979                                 if (!skb)
980                                         goto wait_for_memory;
981
982                                 /*
983                                  * Check whether we can use HW checksum.
984                                  */
985                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
986                                         skb->ip_summed = CHECKSUM_PARTIAL;
987
988                                 skb_entail(sk, skb);
989                                 copy = size_goal;
990                                 max = size_goal;
991                         }
992
993                         /* Try to append data to the end of skb. */
994                         if (copy > seglen)
995                                 copy = seglen;
996
997                         /* Where to copy to? */
998                         if (skb_tailroom(skb) > 0) {
999                                 /* We have some space in skb head. Superb! */
1000                                 if (copy > skb_tailroom(skb))
1001                                         copy = skb_tailroom(skb);
1002                                 err = skb_add_data_nocache(sk, skb, from, copy);
1003                                 if (err)
1004                                         goto do_fault;
1005                         } else {
1006                                 int merge = 0;
1007                                 int i = skb_shinfo(skb)->nr_frags;
1008                                 struct page *page = TCP_PAGE(sk);
1009                                 int off = TCP_OFF(sk);
1010
1011                                 if (skb_can_coalesce(skb, i, page, off) &&
1012                                     off != PAGE_SIZE) {
1013                                         /* We can extend the last page
1014                                          * fragment. */
1015                                         merge = 1;
1016                                 } else if (i == MAX_SKB_FRAGS || !sg) {
1017                                         /* Need to add new fragment and cannot
1018                                          * do this because interface is non-SG,
1019                                          * or because all the page slots are
1020                                          * busy. */
1021                                         tcp_mark_push(tp, skb);
1022                                         goto new_segment;
1023                                 } else if (page) {
1024                                         if (off == PAGE_SIZE) {
1025                                                 put_page(page);
1026                                                 TCP_PAGE(sk) = page = NULL;
1027                                                 off = 0;
1028                                         }
1029                                 } else
1030                                         off = 0;
1031
1032                                 if (copy > PAGE_SIZE - off)
1033                                         copy = PAGE_SIZE - off;
1034
1035                                 if (!sk_wmem_schedule(sk, copy))
1036                                         goto wait_for_memory;
1037
1038                                 if (!page) {
1039                                         /* Allocate new cache page. */
1040                                         if (!(page = sk_stream_alloc_page(sk)))
1041                                                 goto wait_for_memory;
1042                                 }
1043
1044                                 /* Time to copy data. We are close to
1045                                  * the end! */
1046                                 err = skb_copy_to_page_nocache(sk, from, skb,
1047                                                                page, off, copy);
1048                                 if (err) {
1049                                         /* If this page was new, give it to the
1050                                          * socket so it does not get leaked.
1051                                          */
1052                                         if (!TCP_PAGE(sk)) {
1053                                                 TCP_PAGE(sk) = page;
1054                                                 TCP_OFF(sk) = 0;
1055                                         }
1056                                         goto do_error;
1057                                 }
1058
1059                                 /* Update the skb. */
1060                                 if (merge) {
1061                                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1062                                 } else {
1063                                         skb_fill_page_desc(skb, i, page, off, copy);
1064                                         if (TCP_PAGE(sk)) {
1065                                                 get_page(page);
1066                                         } else if (off + copy < PAGE_SIZE) {
1067                                                 get_page(page);
1068                                                 TCP_PAGE(sk) = page;
1069                                         }
1070                                 }
1071
1072                                 TCP_OFF(sk) = off + copy;
1073                         }
1074
1075                         if (!copied)
1076                                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1077
1078                         tp->write_seq += copy;
1079                         TCP_SKB_CB(skb)->end_seq += copy;
1080                         skb_shinfo(skb)->gso_segs = 0;
1081
1082                         from += copy;
1083                         copied += copy;
1084                         if ((seglen -= copy) == 0 && iovlen == 0)
1085                                 goto out;
1086
1087                         if (skb->len < max || (flags & MSG_OOB))
1088                                 continue;
1089
1090                         if (forced_push(tp)) {
1091                                 tcp_mark_push(tp, skb);
1092                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1093                         } else if (skb == tcp_send_head(sk))
1094                                 tcp_push_one(sk, mss_now);
1095                         continue;
1096
1097 wait_for_sndbuf:
1098                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1099 wait_for_memory:
1100                         if (copied)
1101                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1102
1103                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1104                                 goto do_error;
1105
1106                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1107                 }
1108         }
1109
1110 out:
1111         if (copied)
1112                 tcp_push(sk, flags, mss_now, tp->nonagle);
1113         release_sock(sk);
1114         return copied;
1115
1116 do_fault:
1117         if (!skb->len) {
1118                 tcp_unlink_write_queue(skb, sk);
1119                 /* It is the one place in all of TCP, except connection
1120                  * reset, where we can be unlinking the send_head.
1121                  */
1122                 tcp_check_send_head(sk, skb);
1123                 sk_wmem_free_skb(sk, skb);
1124         }
1125
1126 do_error:
1127         if (copied)
1128                 goto out;
1129 out_err:
1130         err = sk_stream_error(sk, flags, err);
1131         release_sock(sk);
1132         return err;
1133 }
1134 EXPORT_SYMBOL(tcp_sendmsg);
1135
1136 /*
1137  *      Handle reading urgent data. BSD has very simple semantics for
1138  *      this, no blocking and very strange errors 8)
1139  */
1140
1141 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1142 {
1143         struct tcp_sock *tp = tcp_sk(sk);
1144
1145         /* No URG data to read. */
1146         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1147             tp->urg_data == TCP_URG_READ)
1148                 return -EINVAL; /* Yes this is right ! */
1149
1150         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1151                 return -ENOTCONN;
1152
1153         if (tp->urg_data & TCP_URG_VALID) {
1154                 int err = 0;
1155                 char c = tp->urg_data;
1156
1157                 if (!(flags & MSG_PEEK))
1158                         tp->urg_data = TCP_URG_READ;
1159
1160                 /* Read urgent data. */
1161                 msg->msg_flags |= MSG_OOB;
1162
1163                 if (len > 0) {
1164                         if (!(flags & MSG_TRUNC))
1165                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1166                         len = 1;
1167                 } else
1168                         msg->msg_flags |= MSG_TRUNC;
1169
1170                 return err ? -EFAULT : len;
1171         }
1172
1173         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1174                 return 0;
1175
1176         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1177          * the available implementations agree in this case:
1178          * this call should never block, independent of the
1179          * blocking state of the socket.
1180          * Mike <pall@rz.uni-karlsruhe.de>
1181          */
1182         return -EAGAIN;
1183 }
1184
1185 /* Clean up the receive buffer for full frames taken by the user,
1186  * then send an ACK if necessary.  COPIED is the number of bytes
1187  * tcp_recvmsg has given to the user so far, it speeds up the
1188  * calculation of whether or not we must ACK for the sake of
1189  * a window update.
1190  */
1191 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1192 {
1193         struct tcp_sock *tp = tcp_sk(sk);
1194         int time_to_ack = 0;
1195
1196         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1197
1198         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1199              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1200              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1201
1202         if (inet_csk_ack_scheduled(sk)) {
1203                 const struct inet_connection_sock *icsk = inet_csk(sk);
1204                    /* Delayed ACKs frequently hit locked sockets during bulk
1205                     * receive. */
1206                 if (icsk->icsk_ack.blocked ||
1207                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1208                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1209                     /*
1210                      * If this read emptied read buffer, we send ACK, if
1211                      * connection is not bidirectional, user drained
1212                      * receive buffer and there was a small segment
1213                      * in queue.
1214                      */
1215                     (copied > 0 &&
1216                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1217                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1218                        !icsk->icsk_ack.pingpong)) &&
1219                       !atomic_read(&sk->sk_rmem_alloc)))
1220                         time_to_ack = 1;
1221         }
1222
1223         /* We send an ACK if we can now advertise a non-zero window
1224          * which has been raised "significantly".
1225          *
1226          * Even if window raised up to infinity, do not send window open ACK
1227          * in states, where we will not receive more. It is useless.
1228          */
1229         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1230                 __u32 rcv_window_now = tcp_receive_window(tp);
1231
1232                 /* Optimize, __tcp_select_window() is not cheap. */
1233                 if (2*rcv_window_now <= tp->window_clamp) {
1234                         __u32 new_window = __tcp_select_window(sk);
1235
1236                         /* Send ACK now, if this read freed lots of space
1237                          * in our buffer. Certainly, new_window is new window.
1238                          * We can advertise it now, if it is not less than current one.
1239                          * "Lots" means "at least twice" here.
1240                          */
1241                         if (new_window && new_window >= 2 * rcv_window_now)
1242                                 time_to_ack = 1;
1243                 }
1244         }
1245         if (time_to_ack)
1246                 tcp_send_ack(sk);
1247 }
1248
1249 static void tcp_prequeue_process(struct sock *sk)
1250 {
1251         struct sk_buff *skb;
1252         struct tcp_sock *tp = tcp_sk(sk);
1253
1254         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1255
1256         /* RX process wants to run with disabled BHs, though it is not
1257          * necessary */
1258         local_bh_disable();
1259         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1260                 sk_backlog_rcv(sk, skb);
1261         local_bh_enable();
1262
1263         /* Clear memory counter. */
1264         tp->ucopy.memory = 0;
1265 }
1266
1267 #ifdef CONFIG_NET_DMA
1268 static void tcp_service_net_dma(struct sock *sk, bool wait)
1269 {
1270         dma_cookie_t done, used;
1271         dma_cookie_t last_issued;
1272         struct tcp_sock *tp = tcp_sk(sk);
1273
1274         if (!tp->ucopy.dma_chan)
1275                 return;
1276
1277         last_issued = tp->ucopy.dma_cookie;
1278         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1279
1280         do {
1281                 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1282                                               last_issued, &done,
1283                                               &used) == DMA_SUCCESS) {
1284                         /* Safe to free early-copied skbs now */
1285                         __skb_queue_purge(&sk->sk_async_wait_queue);
1286                         break;
1287                 } else {
1288                         struct sk_buff *skb;
1289                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1290                                (dma_async_is_complete(skb->dma_cookie, done,
1291                                                       used) == DMA_SUCCESS)) {
1292                                 __skb_dequeue(&sk->sk_async_wait_queue);
1293                                 kfree_skb(skb);
1294                         }
1295                 }
1296         } while (wait);
1297 }
1298 #endif
1299
1300 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1301 {
1302         struct sk_buff *skb;
1303         u32 offset;
1304
1305         skb_queue_walk(&sk->sk_receive_queue, skb) {
1306                 offset = seq - TCP_SKB_CB(skb)->seq;
1307                 if (tcp_hdr(skb)->syn)
1308                         offset--;
1309                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1310                         *off = offset;
1311                         return skb;
1312                 }
1313         }
1314         return NULL;
1315 }
1316
1317 /*
1318  * This routine provides an alternative to tcp_recvmsg() for routines
1319  * that would like to handle copying from skbuffs directly in 'sendfile'
1320  * fashion.
1321  * Note:
1322  *      - It is assumed that the socket was locked by the caller.
1323  *      - The routine does not block.
1324  *      - At present, there is no support for reading OOB data
1325  *        or for 'peeking' the socket using this routine
1326  *        (although both would be easy to implement).
1327  */
1328 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1329                   sk_read_actor_t recv_actor)
1330 {
1331         struct sk_buff *skb;
1332         struct tcp_sock *tp = tcp_sk(sk);
1333         u32 seq = tp->copied_seq;
1334         u32 offset;
1335         int copied = 0;
1336
1337         if (sk->sk_state == TCP_LISTEN)
1338                 return -ENOTCONN;
1339         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1340                 if (offset < skb->len) {
1341                         int used;
1342                         size_t len;
1343
1344                         len = skb->len - offset;
1345                         /* Stop reading if we hit a patch of urgent data */
1346                         if (tp->urg_data) {
1347                                 u32 urg_offset = tp->urg_seq - seq;
1348                                 if (urg_offset < len)
1349                                         len = urg_offset;
1350                                 if (!len)
1351                                         break;
1352                         }
1353                         used = recv_actor(desc, skb, offset, len);
1354                         if (used < 0) {
1355                                 if (!copied)
1356                                         copied = used;
1357                                 break;
1358                         } else if (used <= len) {
1359                                 seq += used;
1360                                 copied += used;
1361                                 offset += used;
1362                         }
1363                         /*
1364                          * If recv_actor drops the lock (e.g. TCP splice
1365                          * receive) the skb pointer might be invalid when
1366                          * getting here: tcp_collapse might have deleted it
1367                          * while aggregating skbs from the socket queue.
1368                          */
1369                         skb = tcp_recv_skb(sk, seq-1, &offset);
1370                         if (!skb || (offset+1 != skb->len))
1371                                 break;
1372                 }
1373                 if (tcp_hdr(skb)->fin) {
1374                         sk_eat_skb(sk, skb, 0);
1375                         ++seq;
1376                         break;
1377                 }
1378                 sk_eat_skb(sk, skb, 0);
1379                 if (!desc->count)
1380                         break;
1381                 tp->copied_seq = seq;
1382         }
1383         tp->copied_seq = seq;
1384
1385         tcp_rcv_space_adjust(sk);
1386
1387         /* Clean up data we have read: This will do ACK frames. */
1388         if (copied > 0)
1389                 tcp_cleanup_rbuf(sk, copied);
1390         return copied;
1391 }
1392 EXPORT_SYMBOL(tcp_read_sock);
1393
1394 /*
1395  *      This routine copies from a sock struct into the user buffer.
1396  *
1397  *      Technical note: in 2.3 we work on _locked_ socket, so that
1398  *      tricks with *seq access order and skb->users are not required.
1399  *      Probably, code can be easily improved even more.
1400  */
1401
1402 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1403                 size_t len, int nonblock, int flags, int *addr_len)
1404 {
1405         struct tcp_sock *tp = tcp_sk(sk);
1406         int copied = 0;
1407         u32 peek_seq;
1408         u32 *seq;
1409         unsigned long used;
1410         int err;
1411         int target;             /* Read at least this many bytes */
1412         long timeo;
1413         struct task_struct *user_recv = NULL;
1414         int copied_early = 0;
1415         struct sk_buff *skb;
1416         u32 urg_hole = 0;
1417
1418         lock_sock(sk);
1419
1420         err = -ENOTCONN;
1421         if (sk->sk_state == TCP_LISTEN)
1422                 goto out;
1423
1424         timeo = sock_rcvtimeo(sk, nonblock);
1425
1426         /* Urgent data needs to be handled specially. */
1427         if (flags & MSG_OOB)
1428                 goto recv_urg;
1429
1430         seq = &tp->copied_seq;
1431         if (flags & MSG_PEEK) {
1432                 peek_seq = tp->copied_seq;
1433                 seq = &peek_seq;
1434         }
1435
1436         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1437
1438 #ifdef CONFIG_NET_DMA
1439         tp->ucopy.dma_chan = NULL;
1440         preempt_disable();
1441         skb = skb_peek_tail(&sk->sk_receive_queue);
1442         {
1443                 int available = 0;
1444
1445                 if (skb)
1446                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1447                 if ((available < target) &&
1448                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1449                     !sysctl_tcp_low_latency &&
1450                     dma_find_channel(DMA_MEMCPY)) {
1451                         preempt_enable_no_resched();
1452                         tp->ucopy.pinned_list =
1453                                         dma_pin_iovec_pages(msg->msg_iov, len);
1454                 } else {
1455                         preempt_enable_no_resched();
1456                 }
1457         }
1458 #endif
1459
1460         do {
1461                 u32 offset;
1462
1463                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1464                 if (tp->urg_data && tp->urg_seq == *seq) {
1465                         if (copied)
1466                                 break;
1467                         if (signal_pending(current)) {
1468                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1469                                 break;
1470                         }
1471                 }
1472
1473                 /* Next get a buffer. */
1474
1475                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1476                         /* Now that we have two receive queues this
1477                          * shouldn't happen.
1478                          */
1479                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1480                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1481                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1482                                  flags))
1483                                 break;
1484
1485                         offset = *seq - TCP_SKB_CB(skb)->seq;
1486                         if (tcp_hdr(skb)->syn)
1487                                 offset--;
1488                         if (offset < skb->len)
1489                                 goto found_ok_skb;
1490                         if (tcp_hdr(skb)->fin)
1491                                 goto found_fin_ok;
1492                         WARN(!(flags & MSG_PEEK),
1493                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1494                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1495                 }
1496
1497                 /* Well, if we have backlog, try to process it now yet. */
1498
1499                 if (copied >= target && !sk->sk_backlog.tail)
1500                         break;
1501
1502                 if (copied) {
1503                         if (sk->sk_err ||
1504                             sk->sk_state == TCP_CLOSE ||
1505                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1506                             !timeo ||
1507                             signal_pending(current))
1508                                 break;
1509                 } else {
1510                         if (sock_flag(sk, SOCK_DONE))
1511                                 break;
1512
1513                         if (sk->sk_err) {
1514                                 copied = sock_error(sk);
1515                                 break;
1516                         }
1517
1518                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1519                                 break;
1520
1521                         if (sk->sk_state == TCP_CLOSE) {
1522                                 if (!sock_flag(sk, SOCK_DONE)) {
1523                                         /* This occurs when user tries to read
1524                                          * from never connected socket.
1525                                          */
1526                                         copied = -ENOTCONN;
1527                                         break;
1528                                 }
1529                                 break;
1530                         }
1531
1532                         if (!timeo) {
1533                                 copied = -EAGAIN;
1534                                 break;
1535                         }
1536
1537                         if (signal_pending(current)) {
1538                                 copied = sock_intr_errno(timeo);
1539                                 break;
1540                         }
1541                 }
1542
1543                 tcp_cleanup_rbuf(sk, copied);
1544
1545                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1546                         /* Install new reader */
1547                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1548                                 user_recv = current;
1549                                 tp->ucopy.task = user_recv;
1550                                 tp->ucopy.iov = msg->msg_iov;
1551                         }
1552
1553                         tp->ucopy.len = len;
1554
1555                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1556                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1557
1558                         /* Ugly... If prequeue is not empty, we have to
1559                          * process it before releasing socket, otherwise
1560                          * order will be broken at second iteration.
1561                          * More elegant solution is required!!!
1562                          *
1563                          * Look: we have the following (pseudo)queues:
1564                          *
1565                          * 1. packets in flight
1566                          * 2. backlog
1567                          * 3. prequeue
1568                          * 4. receive_queue
1569                          *
1570                          * Each queue can be processed only if the next ones
1571                          * are empty. At this point we have empty receive_queue.
1572                          * But prequeue _can_ be not empty after 2nd iteration,
1573                          * when we jumped to start of loop because backlog
1574                          * processing added something to receive_queue.
1575                          * We cannot release_sock(), because backlog contains
1576                          * packets arrived _after_ prequeued ones.
1577                          *
1578                          * Shortly, algorithm is clear --- to process all
1579                          * the queues in order. We could make it more directly,
1580                          * requeueing packets from backlog to prequeue, if
1581                          * is not empty. It is more elegant, but eats cycles,
1582                          * unfortunately.
1583                          */
1584                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1585                                 goto do_prequeue;
1586
1587                         /* __ Set realtime policy in scheduler __ */
1588                 }
1589
1590 #ifdef CONFIG_NET_DMA
1591                 if (tp->ucopy.dma_chan)
1592                         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1593 #endif
1594                 if (copied >= target) {
1595                         /* Do not sleep, just process backlog. */
1596                         release_sock(sk);
1597                         lock_sock(sk);
1598                 } else
1599                         sk_wait_data(sk, &timeo);
1600
1601 #ifdef CONFIG_NET_DMA
1602                 tcp_service_net_dma(sk, false);  /* Don't block */
1603                 tp->ucopy.wakeup = 0;
1604 #endif
1605
1606                 if (user_recv) {
1607                         int chunk;
1608
1609                         /* __ Restore normal policy in scheduler __ */
1610
1611                         if ((chunk = len - tp->ucopy.len) != 0) {
1612                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1613                                 len -= chunk;
1614                                 copied += chunk;
1615                         }
1616
1617                         if (tp->rcv_nxt == tp->copied_seq &&
1618                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1619 do_prequeue:
1620                                 tcp_prequeue_process(sk);
1621
1622                                 if ((chunk = len - tp->ucopy.len) != 0) {
1623                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1624                                         len -= chunk;
1625                                         copied += chunk;
1626                                 }
1627                         }
1628                 }
1629                 if ((flags & MSG_PEEK) &&
1630                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1631                         if (net_ratelimit())
1632                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1633                                        current->comm, task_pid_nr(current));
1634                         peek_seq = tp->copied_seq;
1635                 }
1636                 continue;
1637
1638         found_ok_skb:
1639                 /* Ok so how much can we use? */
1640                 used = skb->len - offset;
1641                 if (len < used)
1642                         used = len;
1643
1644                 /* Do we have urgent data here? */
1645                 if (tp->urg_data) {
1646                         u32 urg_offset = tp->urg_seq - *seq;
1647                         if (urg_offset < used) {
1648                                 if (!urg_offset) {
1649                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1650                                                 ++*seq;
1651                                                 urg_hole++;
1652                                                 offset++;
1653                                                 used--;
1654                                                 if (!used)
1655                                                         goto skip_copy;
1656                                         }
1657                                 } else
1658                                         used = urg_offset;
1659                         }
1660                 }
1661
1662                 if (!(flags & MSG_TRUNC)) {
1663 #ifdef CONFIG_NET_DMA
1664                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1665                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1666
1667                         if (tp->ucopy.dma_chan) {
1668                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1669                                         tp->ucopy.dma_chan, skb, offset,
1670                                         msg->msg_iov, used,
1671                                         tp->ucopy.pinned_list);
1672
1673                                 if (tp->ucopy.dma_cookie < 0) {
1674
1675                                         printk(KERN_ALERT "dma_cookie < 0\n");
1676
1677                                         /* Exception. Bailout! */
1678                                         if (!copied)
1679                                                 copied = -EFAULT;
1680                                         break;
1681                                 }
1682
1683                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1684
1685                                 if ((offset + used) == skb->len)
1686                                         copied_early = 1;
1687
1688                         } else
1689 #endif
1690                         {
1691                                 err = skb_copy_datagram_iovec(skb, offset,
1692                                                 msg->msg_iov, used);
1693                                 if (err) {
1694                                         /* Exception. Bailout! */
1695                                         if (!copied)
1696                                                 copied = -EFAULT;
1697                                         break;
1698                                 }
1699                         }
1700                 }
1701
1702                 *seq += used;
1703                 copied += used;
1704                 len -= used;
1705
1706                 tcp_rcv_space_adjust(sk);
1707
1708 skip_copy:
1709                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1710                         tp->urg_data = 0;
1711                         tcp_fast_path_check(sk);
1712                 }
1713                 if (used + offset < skb->len)
1714                         continue;
1715
1716                 if (tcp_hdr(skb)->fin)
1717                         goto found_fin_ok;
1718                 if (!(flags & MSG_PEEK)) {
1719                         sk_eat_skb(sk, skb, copied_early);
1720                         copied_early = 0;
1721                 }
1722                 continue;
1723
1724         found_fin_ok:
1725                 /* Process the FIN. */
1726                 ++*seq;
1727                 if (!(flags & MSG_PEEK)) {
1728                         sk_eat_skb(sk, skb, copied_early);
1729                         copied_early = 0;
1730                 }
1731                 break;
1732         } while (len > 0);
1733
1734         if (user_recv) {
1735                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1736                         int chunk;
1737
1738                         tp->ucopy.len = copied > 0 ? len : 0;
1739
1740                         tcp_prequeue_process(sk);
1741
1742                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1743                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1744                                 len -= chunk;
1745                                 copied += chunk;
1746                         }
1747                 }
1748
1749                 tp->ucopy.task = NULL;
1750                 tp->ucopy.len = 0;
1751         }
1752
1753 #ifdef CONFIG_NET_DMA
1754         tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1755         tp->ucopy.dma_chan = NULL;
1756
1757         if (tp->ucopy.pinned_list) {
1758                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1759                 tp->ucopy.pinned_list = NULL;
1760         }
1761 #endif
1762
1763         /* According to UNIX98, msg_name/msg_namelen are ignored
1764          * on connected socket. I was just happy when found this 8) --ANK
1765          */
1766
1767         /* Clean up data we have read: This will do ACK frames. */
1768         tcp_cleanup_rbuf(sk, copied);
1769
1770         release_sock(sk);
1771         return copied;
1772
1773 out:
1774         release_sock(sk);
1775         return err;
1776
1777 recv_urg:
1778         err = tcp_recv_urg(sk, msg, len, flags);
1779         goto out;
1780 }
1781 EXPORT_SYMBOL(tcp_recvmsg);
1782
1783 void tcp_set_state(struct sock *sk, int state)
1784 {
1785         int oldstate = sk->sk_state;
1786
1787         switch (state) {
1788         case TCP_ESTABLISHED:
1789                 if (oldstate != TCP_ESTABLISHED)
1790                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1791                 break;
1792
1793         case TCP_CLOSE:
1794                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1795                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1796
1797                 sk->sk_prot->unhash(sk);
1798                 if (inet_csk(sk)->icsk_bind_hash &&
1799                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1800                         inet_put_port(sk);
1801                 /* fall through */
1802         default:
1803                 if (oldstate == TCP_ESTABLISHED)
1804                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1805         }
1806
1807         /* Change state AFTER socket is unhashed to avoid closed
1808          * socket sitting in hash tables.
1809          */
1810         sk->sk_state = state;
1811
1812 #ifdef STATE_TRACE
1813         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1814 #endif
1815 }
1816 EXPORT_SYMBOL_GPL(tcp_set_state);
1817
1818 /*
1819  *      State processing on a close. This implements the state shift for
1820  *      sending our FIN frame. Note that we only send a FIN for some
1821  *      states. A shutdown() may have already sent the FIN, or we may be
1822  *      closed.
1823  */
1824
1825 static const unsigned char new_state[16] = {
1826   /* current state:        new state:      action:      */
1827   /* (Invalid)          */ TCP_CLOSE,
1828   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1829   /* TCP_SYN_SENT       */ TCP_CLOSE,
1830   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1831   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1832   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1833   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1834   /* TCP_CLOSE          */ TCP_CLOSE,
1835   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1836   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1837   /* TCP_LISTEN         */ TCP_CLOSE,
1838   /* TCP_CLOSING        */ TCP_CLOSING,
1839 };
1840
1841 static int tcp_close_state(struct sock *sk)
1842 {
1843         int next = (int)new_state[sk->sk_state];
1844         int ns = next & TCP_STATE_MASK;
1845
1846         tcp_set_state(sk, ns);
1847
1848         return next & TCP_ACTION_FIN;
1849 }
1850
1851 /*
1852  *      Shutdown the sending side of a connection. Much like close except
1853  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1854  */
1855
1856 void tcp_shutdown(struct sock *sk, int how)
1857 {
1858         /*      We need to grab some memory, and put together a FIN,
1859          *      and then put it into the queue to be sent.
1860          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1861          */
1862         if (!(how & SEND_SHUTDOWN))
1863                 return;
1864
1865         /* If we've already sent a FIN, or it's a closed state, skip this. */
1866         if ((1 << sk->sk_state) &
1867             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1868              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1869                 /* Clear out any half completed packets.  FIN if needed. */
1870                 if (tcp_close_state(sk))
1871                         tcp_send_fin(sk);
1872         }
1873 }
1874 EXPORT_SYMBOL(tcp_shutdown);
1875
1876 void tcp_close(struct sock *sk, long timeout)
1877 {
1878         struct sk_buff *skb;
1879         int data_was_unread = 0;
1880         int state;
1881
1882         lock_sock(sk);
1883         sk->sk_shutdown = SHUTDOWN_MASK;
1884
1885         if (sk->sk_state == TCP_LISTEN) {
1886                 tcp_set_state(sk, TCP_CLOSE);
1887
1888                 /* Special case. */
1889                 inet_csk_listen_stop(sk);
1890
1891                 goto adjudge_to_death;
1892         }
1893
1894         /*  We need to flush the recv. buffs.  We do this only on the
1895          *  descriptor close, not protocol-sourced closes, because the
1896          *  reader process may not have drained the data yet!
1897          */
1898         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1899                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1900                           tcp_hdr(skb)->fin;
1901                 data_was_unread += len;
1902                 __kfree_skb(skb);
1903         }
1904
1905         sk_mem_reclaim(sk);
1906
1907         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1908         if (sk->sk_state == TCP_CLOSE)
1909                 goto adjudge_to_death;
1910
1911         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1912          * data was lost. To witness the awful effects of the old behavior of
1913          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1914          * GET in an FTP client, suspend the process, wait for the client to
1915          * advertise a zero window, then kill -9 the FTP client, wheee...
1916          * Note: timeout is always zero in such a case.
1917          */
1918         if (data_was_unread) {
1919                 /* Unread data was tossed, zap the connection. */
1920                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1921                 tcp_set_state(sk, TCP_CLOSE);
1922                 tcp_send_active_reset(sk, sk->sk_allocation);
1923         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1924                 /* Check zero linger _after_ checking for unread data. */
1925                 sk->sk_prot->disconnect(sk, 0);
1926                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1927         } else if (tcp_close_state(sk)) {
1928                 /* We FIN if the application ate all the data before
1929                  * zapping the connection.
1930                  */
1931
1932                 /* RED-PEN. Formally speaking, we have broken TCP state
1933                  * machine. State transitions:
1934                  *
1935                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1936                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1937                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1938                  *
1939                  * are legal only when FIN has been sent (i.e. in window),
1940                  * rather than queued out of window. Purists blame.
1941                  *
1942                  * F.e. "RFC state" is ESTABLISHED,
1943                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1944                  *
1945                  * The visible declinations are that sometimes
1946                  * we enter time-wait state, when it is not required really
1947                  * (harmless), do not send active resets, when they are
1948                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1949                  * they look as CLOSING or LAST_ACK for Linux)
1950                  * Probably, I missed some more holelets.
1951                  *                                              --ANK
1952                  */
1953                 tcp_send_fin(sk);
1954         }
1955
1956         sk_stream_wait_close(sk, timeout);
1957
1958 adjudge_to_death:
1959         state = sk->sk_state;
1960         sock_hold(sk);
1961         sock_orphan(sk);
1962
1963         /* It is the last release_sock in its life. It will remove backlog. */
1964         release_sock(sk);
1965
1966
1967         /* Now socket is owned by kernel and we acquire BH lock
1968            to finish close. No need to check for user refs.
1969          */
1970         local_bh_disable();
1971         bh_lock_sock(sk);
1972         WARN_ON(sock_owned_by_user(sk));
1973
1974         percpu_counter_inc(sk->sk_prot->orphan_count);
1975
1976         /* Have we already been destroyed by a softirq or backlog? */
1977         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1978                 goto out;
1979
1980         /*      This is a (useful) BSD violating of the RFC. There is a
1981          *      problem with TCP as specified in that the other end could
1982          *      keep a socket open forever with no application left this end.
1983          *      We use a 3 minute timeout (about the same as BSD) then kill
1984          *      our end. If they send after that then tough - BUT: long enough
1985          *      that we won't make the old 4*rto = almost no time - whoops
1986          *      reset mistake.
1987          *
1988          *      Nope, it was not mistake. It is really desired behaviour
1989          *      f.e. on http servers, when such sockets are useless, but
1990          *      consume significant resources. Let's do it with special
1991          *      linger2 option.                                 --ANK
1992          */
1993
1994         if (sk->sk_state == TCP_FIN_WAIT2) {
1995                 struct tcp_sock *tp = tcp_sk(sk);
1996                 if (tp->linger2 < 0) {
1997                         tcp_set_state(sk, TCP_CLOSE);
1998                         tcp_send_active_reset(sk, GFP_ATOMIC);
1999                         NET_INC_STATS_BH(sock_net(sk),
2000                                         LINUX_MIB_TCPABORTONLINGER);
2001                 } else {
2002                         const int tmo = tcp_fin_time(sk);
2003
2004                         if (tmo > TCP_TIMEWAIT_LEN) {
2005                                 inet_csk_reset_keepalive_timer(sk,
2006                                                 tmo - TCP_TIMEWAIT_LEN);
2007                         } else {
2008                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2009                                 goto out;
2010                         }
2011                 }
2012         }
2013         if (sk->sk_state != TCP_CLOSE) {
2014                 sk_mem_reclaim(sk);
2015                 if (tcp_too_many_orphans(sk, 0)) {
2016                         if (net_ratelimit())
2017                                 printk(KERN_INFO "TCP: too many of orphaned "
2018                                        "sockets\n");
2019                         tcp_set_state(sk, TCP_CLOSE);
2020                         tcp_send_active_reset(sk, GFP_ATOMIC);
2021                         NET_INC_STATS_BH(sock_net(sk),
2022                                         LINUX_MIB_TCPABORTONMEMORY);
2023                 }
2024         }
2025
2026         if (sk->sk_state == TCP_CLOSE)
2027                 inet_csk_destroy_sock(sk);
2028         /* Otherwise, socket is reprieved until protocol close. */
2029
2030 out:
2031         bh_unlock_sock(sk);
2032         local_bh_enable();
2033         sock_put(sk);
2034 }
2035 EXPORT_SYMBOL(tcp_close);
2036
2037 /* These states need RST on ABORT according to RFC793 */
2038
2039 static inline int tcp_need_reset(int state)
2040 {
2041         return (1 << state) &
2042                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2043                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2044 }
2045
2046 int tcp_disconnect(struct sock *sk, int flags)
2047 {
2048         struct inet_sock *inet = inet_sk(sk);
2049         struct inet_connection_sock *icsk = inet_csk(sk);
2050         struct tcp_sock *tp = tcp_sk(sk);
2051         int err = 0;
2052         int old_state = sk->sk_state;
2053
2054         if (old_state != TCP_CLOSE)
2055                 tcp_set_state(sk, TCP_CLOSE);
2056
2057         /* ABORT function of RFC793 */
2058         if (old_state == TCP_LISTEN) {
2059                 inet_csk_listen_stop(sk);
2060         } else if (tcp_need_reset(old_state) ||
2061                    (tp->snd_nxt != tp->write_seq &&
2062                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2063                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2064                  * states
2065                  */
2066                 tcp_send_active_reset(sk, gfp_any());
2067                 sk->sk_err = ECONNRESET;
2068         } else if (old_state == TCP_SYN_SENT)
2069                 sk->sk_err = ECONNRESET;
2070
2071         tcp_clear_xmit_timers(sk);
2072         __skb_queue_purge(&sk->sk_receive_queue);
2073         tcp_write_queue_purge(sk);
2074         __skb_queue_purge(&tp->out_of_order_queue);
2075 #ifdef CONFIG_NET_DMA
2076         __skb_queue_purge(&sk->sk_async_wait_queue);
2077 #endif
2078
2079         inet->inet_dport = 0;
2080
2081         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2082                 inet_reset_saddr(sk);
2083
2084         sk->sk_shutdown = 0;
2085         sock_reset_flag(sk, SOCK_DONE);
2086         tp->srtt = 0;
2087         if ((tp->write_seq += tp->max_window + 2) == 0)
2088                 tp->write_seq = 1;
2089         icsk->icsk_backoff = 0;
2090         tp->snd_cwnd = 2;
2091         icsk->icsk_probes_out = 0;
2092         tp->packets_out = 0;
2093         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2094         tp->snd_cwnd_cnt = 0;
2095         tp->bytes_acked = 0;
2096         tp->window_clamp = 0;
2097         tcp_set_ca_state(sk, TCP_CA_Open);
2098         tcp_clear_retrans(tp);
2099         inet_csk_delack_init(sk);
2100         tcp_init_send_head(sk);
2101         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2102         __sk_dst_reset(sk);
2103
2104         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2105
2106         sk->sk_error_report(sk);
2107         return err;
2108 }
2109 EXPORT_SYMBOL(tcp_disconnect);
2110
2111 /*
2112  *      Socket option code for TCP.
2113  */
2114 static int do_tcp_setsockopt(struct sock *sk, int level,
2115                 int optname, char __user *optval, unsigned int optlen)
2116 {
2117         struct tcp_sock *tp = tcp_sk(sk);
2118         struct inet_connection_sock *icsk = inet_csk(sk);
2119         int val;
2120         int err = 0;
2121
2122         /* These are data/string values, all the others are ints */
2123         switch (optname) {
2124         case TCP_CONGESTION: {
2125                 char name[TCP_CA_NAME_MAX];
2126
2127                 if (optlen < 1)
2128                         return -EINVAL;
2129
2130                 val = strncpy_from_user(name, optval,
2131                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2132                 if (val < 0)
2133                         return -EFAULT;
2134                 name[val] = 0;
2135
2136                 lock_sock(sk);
2137                 err = tcp_set_congestion_control(sk, name);
2138                 release_sock(sk);
2139                 return err;
2140         }
2141         case TCP_COOKIE_TRANSACTIONS: {
2142                 struct tcp_cookie_transactions ctd;
2143                 struct tcp_cookie_values *cvp = NULL;
2144
2145                 if (sizeof(ctd) > optlen)
2146                         return -EINVAL;
2147                 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2148                         return -EFAULT;
2149
2150                 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2151                     ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2152                         return -EINVAL;
2153
2154                 if (ctd.tcpct_cookie_desired == 0) {
2155                         /* default to global value */
2156                 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2157                            ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2158                            ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2159                         return -EINVAL;
2160                 }
2161
2162                 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2163                         /* Supercedes all other values */
2164                         lock_sock(sk);
2165                         if (tp->cookie_values != NULL) {
2166                                 kref_put(&tp->cookie_values->kref,
2167                                          tcp_cookie_values_release);
2168                                 tp->cookie_values = NULL;
2169                         }
2170                         tp->rx_opt.cookie_in_always = 0; /* false */
2171                         tp->rx_opt.cookie_out_never = 1; /* true */
2172                         release_sock(sk);
2173                         return err;
2174                 }
2175
2176                 /* Allocate ancillary memory before locking.
2177                  */
2178                 if (ctd.tcpct_used > 0 ||
2179                     (tp->cookie_values == NULL &&
2180                      (sysctl_tcp_cookie_size > 0 ||
2181                       ctd.tcpct_cookie_desired > 0 ||
2182                       ctd.tcpct_s_data_desired > 0))) {
2183                         cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2184                                       GFP_KERNEL);
2185                         if (cvp == NULL)
2186                                 return -ENOMEM;
2187
2188                         kref_init(&cvp->kref);
2189                 }
2190                 lock_sock(sk);
2191                 tp->rx_opt.cookie_in_always =
2192                         (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2193                 tp->rx_opt.cookie_out_never = 0; /* false */
2194
2195                 if (tp->cookie_values != NULL) {
2196                         if (cvp != NULL) {
2197                                 /* Changed values are recorded by a changed
2198                                  * pointer, ensuring the cookie will differ,
2199                                  * without separately hashing each value later.
2200                                  */
2201                                 kref_put(&tp->cookie_values->kref,
2202                                          tcp_cookie_values_release);
2203                         } else {
2204                                 cvp = tp->cookie_values;
2205                         }
2206                 }
2207
2208                 if (cvp != NULL) {
2209                         cvp->cookie_desired = ctd.tcpct_cookie_desired;
2210
2211                         if (ctd.tcpct_used > 0) {
2212                                 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2213                                        ctd.tcpct_used);
2214                                 cvp->s_data_desired = ctd.tcpct_used;
2215                                 cvp->s_data_constant = 1; /* true */
2216                         } else {
2217                                 /* No constant payload data. */
2218                                 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2219                                 cvp->s_data_constant = 0; /* false */
2220                         }
2221
2222                         tp->cookie_values = cvp;
2223                 }
2224                 release_sock(sk);
2225                 return err;
2226         }
2227         default:
2228                 /* fallthru */
2229                 break;
2230         }
2231
2232         if (optlen < sizeof(int))
2233                 return -EINVAL;
2234
2235         if (get_user(val, (int __user *)optval))
2236                 return -EFAULT;
2237
2238         lock_sock(sk);
2239
2240         switch (optname) {
2241         case TCP_MAXSEG:
2242                 /* Values greater than interface MTU won't take effect. However
2243                  * at the point when this call is done we typically don't yet
2244                  * know which interface is going to be used */
2245                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2246                         err = -EINVAL;
2247                         break;
2248                 }
2249                 tp->rx_opt.user_mss = val;
2250                 break;
2251
2252         case TCP_NODELAY:
2253                 if (val) {
2254                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2255                          * this option on corked socket is remembered, but
2256                          * it is not activated until cork is cleared.
2257                          *
2258                          * However, when TCP_NODELAY is set we make
2259                          * an explicit push, which overrides even TCP_CORK
2260                          * for currently queued segments.
2261                          */
2262                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2263                         tcp_push_pending_frames(sk);
2264                 } else {
2265                         tp->nonagle &= ~TCP_NAGLE_OFF;
2266                 }
2267                 break;
2268
2269         case TCP_THIN_LINEAR_TIMEOUTS:
2270                 if (val < 0 || val > 1)
2271                         err = -EINVAL;
2272                 else
2273                         tp->thin_lto = val;
2274                 break;
2275
2276         case TCP_THIN_DUPACK:
2277                 if (val < 0 || val > 1)
2278                         err = -EINVAL;
2279                 else
2280                         tp->thin_dupack = val;
2281                 break;
2282
2283         case TCP_CORK:
2284                 /* When set indicates to always queue non-full frames.
2285                  * Later the user clears this option and we transmit
2286                  * any pending partial frames in the queue.  This is
2287                  * meant to be used alongside sendfile() to get properly
2288                  * filled frames when the user (for example) must write
2289                  * out headers with a write() call first and then use
2290                  * sendfile to send out the data parts.
2291                  *
2292                  * TCP_CORK can be set together with TCP_NODELAY and it is
2293                  * stronger than TCP_NODELAY.
2294                  */
2295                 if (val) {
2296                         tp->nonagle |= TCP_NAGLE_CORK;
2297                 } else {
2298                         tp->nonagle &= ~TCP_NAGLE_CORK;
2299                         if (tp->nonagle&TCP_NAGLE_OFF)
2300                                 tp->nonagle |= TCP_NAGLE_PUSH;
2301                         tcp_push_pending_frames(sk);
2302                 }
2303                 break;
2304
2305         case TCP_KEEPIDLE:
2306                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2307                         err = -EINVAL;
2308                 else {
2309                         tp->keepalive_time = val * HZ;
2310                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2311                             !((1 << sk->sk_state) &
2312                               (TCPF_CLOSE | TCPF_LISTEN))) {
2313                                 u32 elapsed = keepalive_time_elapsed(tp);
2314                                 if (tp->keepalive_time > elapsed)
2315                                         elapsed = tp->keepalive_time - elapsed;
2316                                 else
2317                                         elapsed = 0;
2318                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2319                         }
2320                 }
2321                 break;
2322         case TCP_KEEPINTVL:
2323                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2324                         err = -EINVAL;
2325                 else
2326                         tp->keepalive_intvl = val * HZ;
2327                 break;
2328         case TCP_KEEPCNT:
2329                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2330                         err = -EINVAL;
2331                 else
2332                         tp->keepalive_probes = val;
2333                 break;
2334         case TCP_SYNCNT:
2335                 if (val < 1 || val > MAX_TCP_SYNCNT)
2336                         err = -EINVAL;
2337                 else
2338                         icsk->icsk_syn_retries = val;
2339                 break;
2340
2341         case TCP_LINGER2:
2342                 if (val < 0)
2343                         tp->linger2 = -1;
2344                 else if (val > sysctl_tcp_fin_timeout / HZ)
2345                         tp->linger2 = 0;
2346                 else
2347                         tp->linger2 = val * HZ;
2348                 break;
2349
2350         case TCP_DEFER_ACCEPT:
2351                 /* Translate value in seconds to number of retransmits */
2352                 icsk->icsk_accept_queue.rskq_defer_accept =
2353                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2354                                         TCP_RTO_MAX / HZ);
2355                 break;
2356
2357         case TCP_WINDOW_CLAMP:
2358                 if (!val) {
2359                         if (sk->sk_state != TCP_CLOSE) {
2360                                 err = -EINVAL;
2361                                 break;
2362                         }
2363                         tp->window_clamp = 0;
2364                 } else
2365                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2366                                                 SOCK_MIN_RCVBUF / 2 : val;
2367                 break;
2368
2369         case TCP_QUICKACK:
2370                 if (!val) {
2371                         icsk->icsk_ack.pingpong = 1;
2372                 } else {
2373                         icsk->icsk_ack.pingpong = 0;
2374                         if ((1 << sk->sk_state) &
2375                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2376                             inet_csk_ack_scheduled(sk)) {
2377                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2378                                 tcp_cleanup_rbuf(sk, 1);
2379                                 if (!(val & 1))
2380                                         icsk->icsk_ack.pingpong = 1;
2381                         }
2382                 }
2383                 break;
2384
2385 #ifdef CONFIG_TCP_MD5SIG
2386         case TCP_MD5SIG:
2387                 /* Read the IP->Key mappings from userspace */
2388                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2389                 break;
2390 #endif
2391         case TCP_USER_TIMEOUT:
2392                 /* Cap the max timeout in ms TCP will retry/retrans
2393                  * before giving up and aborting (ETIMEDOUT) a connection.
2394                  */
2395                 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2396                 break;
2397         default:
2398                 err = -ENOPROTOOPT;
2399                 break;
2400         }
2401
2402         release_sock(sk);
2403         return err;
2404 }
2405
2406 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2407                    unsigned int optlen)
2408 {
2409         const struct inet_connection_sock *icsk = inet_csk(sk);
2410
2411         if (level != SOL_TCP)
2412                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2413                                                      optval, optlen);
2414         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2415 }
2416 EXPORT_SYMBOL(tcp_setsockopt);
2417
2418 #ifdef CONFIG_COMPAT
2419 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2420                           char __user *optval, unsigned int optlen)
2421 {
2422         if (level != SOL_TCP)
2423                 return inet_csk_compat_setsockopt(sk, level, optname,
2424                                                   optval, optlen);
2425         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2426 }
2427 EXPORT_SYMBOL(compat_tcp_setsockopt);
2428 #endif
2429
2430 /* Return information about state of tcp endpoint in API format. */
2431 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2432 {
2433         const struct tcp_sock *tp = tcp_sk(sk);
2434         const struct inet_connection_sock *icsk = inet_csk(sk);
2435         u32 now = tcp_time_stamp;
2436
2437         memset(info, 0, sizeof(*info));
2438
2439         info->tcpi_state = sk->sk_state;
2440         info->tcpi_ca_state = icsk->icsk_ca_state;
2441         info->tcpi_retransmits = icsk->icsk_retransmits;
2442         info->tcpi_probes = icsk->icsk_probes_out;
2443         info->tcpi_backoff = icsk->icsk_backoff;
2444
2445         if (tp->rx_opt.tstamp_ok)
2446                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2447         if (tcp_is_sack(tp))
2448                 info->tcpi_options |= TCPI_OPT_SACK;
2449         if (tp->rx_opt.wscale_ok) {
2450                 info->tcpi_options |= TCPI_OPT_WSCALE;
2451                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2452                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2453         }
2454
2455         if (tp->ecn_flags & TCP_ECN_OK)
2456                 info->tcpi_options |= TCPI_OPT_ECN;
2457         if (tp->ecn_flags & TCP_ECN_SEEN)
2458                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2459
2460         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2461         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2462         info->tcpi_snd_mss = tp->mss_cache;
2463         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2464
2465         if (sk->sk_state == TCP_LISTEN) {
2466                 info->tcpi_unacked = sk->sk_ack_backlog;
2467                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2468         } else {
2469                 info->tcpi_unacked = tp->packets_out;
2470                 info->tcpi_sacked = tp->sacked_out;
2471         }
2472         info->tcpi_lost = tp->lost_out;
2473         info->tcpi_retrans = tp->retrans_out;
2474         info->tcpi_fackets = tp->fackets_out;
2475
2476         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2477         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2478         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2479
2480         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2481         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2482         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2483         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2484         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2485         info->tcpi_snd_cwnd = tp->snd_cwnd;
2486         info->tcpi_advmss = tp->advmss;
2487         info->tcpi_reordering = tp->reordering;
2488
2489         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2490         info->tcpi_rcv_space = tp->rcvq_space.space;
2491
2492         info->tcpi_total_retrans = tp->total_retrans;
2493 }
2494 EXPORT_SYMBOL_GPL(tcp_get_info);
2495
2496 static int do_tcp_getsockopt(struct sock *sk, int level,
2497                 int optname, char __user *optval, int __user *optlen)
2498 {
2499         struct inet_connection_sock *icsk = inet_csk(sk);
2500         struct tcp_sock *tp = tcp_sk(sk);
2501         int val, len;
2502
2503         if (get_user(len, optlen))
2504                 return -EFAULT;
2505
2506         len = min_t(unsigned int, len, sizeof(int));
2507
2508         if (len < 0)
2509                 return -EINVAL;
2510
2511         switch (optname) {
2512         case TCP_MAXSEG:
2513                 val = tp->mss_cache;
2514                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2515                         val = tp->rx_opt.user_mss;
2516                 break;
2517         case TCP_NODELAY:
2518                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2519                 break;
2520         case TCP_CORK:
2521                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2522                 break;
2523         case TCP_KEEPIDLE:
2524                 val = keepalive_time_when(tp) / HZ;
2525                 break;
2526         case TCP_KEEPINTVL:
2527                 val = keepalive_intvl_when(tp) / HZ;
2528                 break;
2529         case TCP_KEEPCNT:
2530                 val = keepalive_probes(tp);
2531                 break;
2532         case TCP_SYNCNT:
2533                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2534                 break;
2535         case TCP_LINGER2:
2536                 val = tp->linger2;
2537                 if (val >= 0)
2538                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2539                 break;
2540         case TCP_DEFER_ACCEPT:
2541                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2542                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2543                 break;
2544         case TCP_WINDOW_CLAMP:
2545                 val = tp->window_clamp;
2546                 break;
2547         case TCP_INFO: {
2548                 struct tcp_info info;
2549
2550                 if (get_user(len, optlen))
2551                         return -EFAULT;
2552
2553                 tcp_get_info(sk, &info);
2554
2555                 len = min_t(unsigned int, len, sizeof(info));
2556                 if (put_user(len, optlen))
2557                         return -EFAULT;
2558                 if (copy_to_user(optval, &info, len))
2559                         return -EFAULT;
2560                 return 0;
2561         }
2562         case TCP_QUICKACK:
2563                 val = !icsk->icsk_ack.pingpong;
2564                 break;
2565
2566         case TCP_CONGESTION:
2567                 if (get_user(len, optlen))
2568                         return -EFAULT;
2569                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2570                 if (put_user(len, optlen))
2571                         return -EFAULT;
2572                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2573                         return -EFAULT;
2574                 return 0;
2575
2576         case TCP_COOKIE_TRANSACTIONS: {
2577                 struct tcp_cookie_transactions ctd;
2578                 struct tcp_cookie_values *cvp = tp->cookie_values;
2579
2580                 if (get_user(len, optlen))
2581                         return -EFAULT;
2582                 if (len < sizeof(ctd))
2583                         return -EINVAL;
2584
2585                 memset(&ctd, 0, sizeof(ctd));
2586                 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2587                                    TCP_COOKIE_IN_ALWAYS : 0)
2588                                 | (tp->rx_opt.cookie_out_never ?
2589                                    TCP_COOKIE_OUT_NEVER : 0);
2590
2591                 if (cvp != NULL) {
2592                         ctd.tcpct_flags |= (cvp->s_data_in ?
2593                                             TCP_S_DATA_IN : 0)
2594                                          | (cvp->s_data_out ?
2595                                             TCP_S_DATA_OUT : 0);
2596
2597                         ctd.tcpct_cookie_desired = cvp->cookie_desired;
2598                         ctd.tcpct_s_data_desired = cvp->s_data_desired;
2599
2600                         memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2601                                cvp->cookie_pair_size);
2602                         ctd.tcpct_used = cvp->cookie_pair_size;
2603                 }
2604
2605                 if (put_user(sizeof(ctd), optlen))
2606                         return -EFAULT;
2607                 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2608                         return -EFAULT;
2609                 return 0;
2610         }
2611         case TCP_THIN_LINEAR_TIMEOUTS:
2612                 val = tp->thin_lto;
2613                 break;
2614         case TCP_THIN_DUPACK:
2615                 val = tp->thin_dupack;
2616                 break;
2617
2618         case TCP_USER_TIMEOUT:
2619                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2620                 break;
2621         default:
2622                 return -ENOPROTOOPT;
2623         }
2624
2625         if (put_user(len, optlen))
2626                 return -EFAULT;
2627         if (copy_to_user(optval, &val, len))
2628                 return -EFAULT;
2629         return 0;
2630 }
2631
2632 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2633                    int __user *optlen)
2634 {
2635         struct inet_connection_sock *icsk = inet_csk(sk);
2636
2637         if (level != SOL_TCP)
2638                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2639                                                      optval, optlen);
2640         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2641 }
2642 EXPORT_SYMBOL(tcp_getsockopt);
2643
2644 #ifdef CONFIG_COMPAT
2645 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2646                           char __user *optval, int __user *optlen)
2647 {
2648         if (level != SOL_TCP)
2649                 return inet_csk_compat_getsockopt(sk, level, optname,
2650                                                   optval, optlen);
2651         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2652 }
2653 EXPORT_SYMBOL(compat_tcp_getsockopt);
2654 #endif
2655
2656 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, u32 features)
2657 {
2658         struct sk_buff *segs = ERR_PTR(-EINVAL);
2659         struct tcphdr *th;
2660         unsigned thlen;
2661         unsigned int seq;
2662         __be32 delta;
2663         unsigned int oldlen;
2664         unsigned int mss;
2665
2666         if (!pskb_may_pull(skb, sizeof(*th)))
2667                 goto out;
2668
2669         th = tcp_hdr(skb);
2670         thlen = th->doff * 4;
2671         if (thlen < sizeof(*th))
2672                 goto out;
2673
2674         if (!pskb_may_pull(skb, thlen))
2675                 goto out;
2676
2677         oldlen = (u16)~skb->len;
2678         __skb_pull(skb, thlen);
2679
2680         mss = skb_shinfo(skb)->gso_size;
2681         if (unlikely(skb->len <= mss))
2682                 goto out;
2683
2684         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2685                 /* Packet is from an untrusted source, reset gso_segs. */
2686                 int type = skb_shinfo(skb)->gso_type;
2687
2688                 if (unlikely(type &
2689                              ~(SKB_GSO_TCPV4 |
2690                                SKB_GSO_DODGY |
2691                                SKB_GSO_TCP_ECN |
2692                                SKB_GSO_TCPV6 |
2693                                0) ||
2694                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2695                         goto out;
2696
2697                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2698
2699                 segs = NULL;
2700                 goto out;
2701         }
2702
2703         segs = skb_segment(skb, features);
2704         if (IS_ERR(segs))
2705                 goto out;
2706
2707         delta = htonl(oldlen + (thlen + mss));
2708
2709         skb = segs;
2710         th = tcp_hdr(skb);
2711         seq = ntohl(th->seq);
2712
2713         do {
2714                 th->fin = th->psh = 0;
2715
2716                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2717                                        (__force u32)delta));
2718                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2719                         th->check =
2720                              csum_fold(csum_partial(skb_transport_header(skb),
2721                                                     thlen, skb->csum));
2722
2723                 seq += mss;
2724                 skb = skb->next;
2725                 th = tcp_hdr(skb);
2726
2727                 th->seq = htonl(seq);
2728                 th->cwr = 0;
2729         } while (skb->next);
2730
2731         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2732                       skb->data_len);
2733         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2734                                 (__force u32)delta));
2735         if (skb->ip_summed != CHECKSUM_PARTIAL)
2736                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2737                                                    thlen, skb->csum));
2738
2739 out:
2740         return segs;
2741 }
2742 EXPORT_SYMBOL(tcp_tso_segment);
2743
2744 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2745 {
2746         struct sk_buff **pp = NULL;
2747         struct sk_buff *p;
2748         struct tcphdr *th;
2749         struct tcphdr *th2;
2750         unsigned int len;
2751         unsigned int thlen;
2752         __be32 flags;
2753         unsigned int mss = 1;
2754         unsigned int hlen;
2755         unsigned int off;
2756         int flush = 1;
2757         int i;
2758
2759         off = skb_gro_offset(skb);
2760         hlen = off + sizeof(*th);
2761         th = skb_gro_header_fast(skb, off);
2762         if (skb_gro_header_hard(skb, hlen)) {
2763                 th = skb_gro_header_slow(skb, hlen, off);
2764                 if (unlikely(!th))
2765                         goto out;
2766         }
2767
2768         thlen = th->doff * 4;
2769         if (thlen < sizeof(*th))
2770                 goto out;
2771
2772         hlen = off + thlen;
2773         if (skb_gro_header_hard(skb, hlen)) {
2774                 th = skb_gro_header_slow(skb, hlen, off);
2775                 if (unlikely(!th))
2776                         goto out;
2777         }
2778
2779         skb_gro_pull(skb, thlen);
2780
2781         len = skb_gro_len(skb);
2782         flags = tcp_flag_word(th);
2783
2784         for (; (p = *head); head = &p->next) {
2785                 if (!NAPI_GRO_CB(p)->same_flow)
2786                         continue;
2787
2788                 th2 = tcp_hdr(p);
2789
2790                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2791                         NAPI_GRO_CB(p)->same_flow = 0;
2792                         continue;
2793                 }
2794
2795                 goto found;
2796         }
2797
2798         goto out_check_final;
2799
2800 found:
2801         flush = NAPI_GRO_CB(p)->flush;
2802         flush |= (__force int)(flags & TCP_FLAG_CWR);
2803         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2804                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2805         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2806         for (i = sizeof(*th); i < thlen; i += 4)
2807                 flush |= *(u32 *)((u8 *)th + i) ^
2808                          *(u32 *)((u8 *)th2 + i);
2809
2810         mss = skb_shinfo(p)->gso_size;
2811
2812         flush |= (len - 1) >= mss;
2813         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2814
2815         if (flush || skb_gro_receive(head, skb)) {
2816                 mss = 1;
2817                 goto out_check_final;
2818         }
2819
2820         p = *head;
2821         th2 = tcp_hdr(p);
2822         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2823
2824 out_check_final:
2825         flush = len < mss;
2826         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2827                                         TCP_FLAG_RST | TCP_FLAG_SYN |
2828                                         TCP_FLAG_FIN));
2829
2830         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2831                 pp = head;
2832
2833 out:
2834         NAPI_GRO_CB(skb)->flush |= flush;
2835
2836         return pp;
2837 }
2838 EXPORT_SYMBOL(tcp_gro_receive);
2839
2840 int tcp_gro_complete(struct sk_buff *skb)
2841 {
2842         struct tcphdr *th = tcp_hdr(skb);
2843
2844         skb->csum_start = skb_transport_header(skb) - skb->head;
2845         skb->csum_offset = offsetof(struct tcphdr, check);
2846         skb->ip_summed = CHECKSUM_PARTIAL;
2847
2848         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2849
2850         if (th->cwr)
2851                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2852
2853         return 0;
2854 }
2855 EXPORT_SYMBOL(tcp_gro_complete);
2856
2857 #ifdef CONFIG_TCP_MD5SIG
2858 static unsigned long tcp_md5sig_users;
2859 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
2860 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2861
2862 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
2863 {
2864         int cpu;
2865
2866         for_each_possible_cpu(cpu) {
2867                 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
2868
2869                 if (p->md5_desc.tfm)
2870                         crypto_free_hash(p->md5_desc.tfm);
2871         }
2872         free_percpu(pool);
2873 }
2874
2875 void tcp_free_md5sig_pool(void)
2876 {
2877         struct tcp_md5sig_pool __percpu *pool = NULL;
2878
2879         spin_lock_bh(&tcp_md5sig_pool_lock);
2880         if (--tcp_md5sig_users == 0) {
2881                 pool = tcp_md5sig_pool;
2882                 tcp_md5sig_pool = NULL;
2883         }
2884         spin_unlock_bh(&tcp_md5sig_pool_lock);
2885         if (pool)
2886                 __tcp_free_md5sig_pool(pool);
2887 }
2888 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2889
2890 static struct tcp_md5sig_pool __percpu *
2891 __tcp_alloc_md5sig_pool(struct sock *sk)
2892 {
2893         int cpu;
2894         struct tcp_md5sig_pool __percpu *pool;
2895
2896         pool = alloc_percpu(struct tcp_md5sig_pool);
2897         if (!pool)
2898                 return NULL;
2899
2900         for_each_possible_cpu(cpu) {
2901                 struct crypto_hash *hash;
2902
2903                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2904                 if (!hash || IS_ERR(hash))
2905                         goto out_free;
2906
2907                 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
2908         }
2909         return pool;
2910 out_free:
2911         __tcp_free_md5sig_pool(pool);
2912         return NULL;
2913 }
2914
2915 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2916 {
2917         struct tcp_md5sig_pool __percpu *pool;
2918         int alloc = 0;
2919
2920 retry:
2921         spin_lock_bh(&tcp_md5sig_pool_lock);
2922         pool = tcp_md5sig_pool;
2923         if (tcp_md5sig_users++ == 0) {
2924                 alloc = 1;
2925                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2926         } else if (!pool) {
2927                 tcp_md5sig_users--;
2928                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2929                 cpu_relax();
2930                 goto retry;
2931         } else
2932                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2933
2934         if (alloc) {
2935                 /* we cannot hold spinlock here because this may sleep. */
2936                 struct tcp_md5sig_pool __percpu *p;
2937
2938                 p = __tcp_alloc_md5sig_pool(sk);
2939                 spin_lock_bh(&tcp_md5sig_pool_lock);
2940                 if (!p) {
2941                         tcp_md5sig_users--;
2942                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2943                         return NULL;
2944                 }
2945                 pool = tcp_md5sig_pool;
2946                 if (pool) {
2947                         /* oops, it has already been assigned. */
2948                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2949                         __tcp_free_md5sig_pool(p);
2950                 } else {
2951                         tcp_md5sig_pool = pool = p;
2952                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2953                 }
2954         }
2955         return pool;
2956 }
2957 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2958
2959
2960 /**
2961  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2962  *
2963  *      We use percpu structure, so if we succeed, we exit with preemption
2964  *      and BH disabled, to make sure another thread or softirq handling
2965  *      wont try to get same context.
2966  */
2967 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2968 {
2969         struct tcp_md5sig_pool __percpu *p;
2970
2971         local_bh_disable();
2972
2973         spin_lock(&tcp_md5sig_pool_lock);
2974         p = tcp_md5sig_pool;
2975         if (p)
2976                 tcp_md5sig_users++;
2977         spin_unlock(&tcp_md5sig_pool_lock);
2978
2979         if (p)
2980                 return this_cpu_ptr(p);
2981
2982         local_bh_enable();
2983         return NULL;
2984 }
2985 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2986
2987 void tcp_put_md5sig_pool(void)
2988 {
2989         local_bh_enable();
2990         tcp_free_md5sig_pool();
2991 }
2992 EXPORT_SYMBOL(tcp_put_md5sig_pool);
2993
2994 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2995                         const struct tcphdr *th)
2996 {
2997         struct scatterlist sg;
2998         struct tcphdr hdr;
2999         int err;
3000
3001         /* We are not allowed to change tcphdr, make a local copy */
3002         memcpy(&hdr, th, sizeof(hdr));
3003         hdr.check = 0;
3004
3005         /* options aren't included in the hash */
3006         sg_init_one(&sg, &hdr, sizeof(hdr));
3007         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3008         return err;
3009 }
3010 EXPORT_SYMBOL(tcp_md5_hash_header);
3011
3012 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3013                           const struct sk_buff *skb, unsigned int header_len)
3014 {
3015         struct scatterlist sg;
3016         const struct tcphdr *tp = tcp_hdr(skb);
3017         struct hash_desc *desc = &hp->md5_desc;
3018         unsigned i;
3019         const unsigned head_data_len = skb_headlen(skb) > header_len ?
3020                                        skb_headlen(skb) - header_len : 0;
3021         const struct skb_shared_info *shi = skb_shinfo(skb);
3022         struct sk_buff *frag_iter;
3023
3024         sg_init_table(&sg, 1);
3025
3026         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3027         if (crypto_hash_update(desc, &sg, head_data_len))
3028                 return 1;
3029
3030         for (i = 0; i < shi->nr_frags; ++i) {
3031                 const struct skb_frag_struct *f = &shi->frags[i];
3032                 struct page *page = skb_frag_page(f);
3033                 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3034                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3035                         return 1;
3036         }
3037
3038         skb_walk_frags(skb, frag_iter)
3039                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3040                         return 1;
3041
3042         return 0;
3043 }
3044 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3045
3046 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3047 {
3048         struct scatterlist sg;
3049
3050         sg_init_one(&sg, key->key, key->keylen);
3051         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3052 }
3053 EXPORT_SYMBOL(tcp_md5_hash_key);
3054
3055 #endif
3056
3057 /**
3058  * Each Responder maintains up to two secret values concurrently for
3059  * efficient secret rollover.  Each secret value has 4 states:
3060  *
3061  * Generating.  (tcp_secret_generating != tcp_secret_primary)
3062  *    Generates new Responder-Cookies, but not yet used for primary
3063  *    verification.  This is a short-term state, typically lasting only
3064  *    one round trip time (RTT).
3065  *
3066  * Primary.  (tcp_secret_generating == tcp_secret_primary)
3067  *    Used both for generation and primary verification.
3068  *
3069  * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3070  *    Used for verification, until the first failure that can be
3071  *    verified by the newer Generating secret.  At that time, this
3072  *    cookie's state is changed to Secondary, and the Generating
3073  *    cookie's state is changed to Primary.  This is a short-term state,
3074  *    typically lasting only one round trip time (RTT).
3075  *
3076  * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3077  *    Used for secondary verification, after primary verification
3078  *    failures.  This state lasts no more than twice the Maximum Segment
3079  *    Lifetime (2MSL).  Then, the secret is discarded.
3080  */
3081 struct tcp_cookie_secret {
3082         /* The secret is divided into two parts.  The digest part is the
3083          * equivalent of previously hashing a secret and saving the state,
3084          * and serves as an initialization vector (IV).  The message part
3085          * serves as the trailing secret.
3086          */
3087         u32                             secrets[COOKIE_WORKSPACE_WORDS];
3088         unsigned long                   expires;
3089 };
3090
3091 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3092 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3093 #define TCP_SECRET_LIFE (HZ * 600)
3094
3095 static struct tcp_cookie_secret tcp_secret_one;
3096 static struct tcp_cookie_secret tcp_secret_two;
3097
3098 /* Essentially a circular list, without dynamic allocation. */
3099 static struct tcp_cookie_secret *tcp_secret_generating;
3100 static struct tcp_cookie_secret *tcp_secret_primary;
3101 static struct tcp_cookie_secret *tcp_secret_retiring;
3102 static struct tcp_cookie_secret *tcp_secret_secondary;
3103
3104 static DEFINE_SPINLOCK(tcp_secret_locker);
3105
3106 /* Select a pseudo-random word in the cookie workspace.
3107  */
3108 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3109 {
3110         return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3111 }
3112
3113 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3114  * Called in softirq context.
3115  * Returns: 0 for success.
3116  */
3117 int tcp_cookie_generator(u32 *bakery)
3118 {
3119         unsigned long jiffy = jiffies;
3120
3121         if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3122                 spin_lock_bh(&tcp_secret_locker);
3123                 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3124                         /* refreshed by another */
3125                         memcpy(bakery,
3126                                &tcp_secret_generating->secrets[0],
3127                                COOKIE_WORKSPACE_WORDS);
3128                 } else {
3129                         /* still needs refreshing */
3130                         get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3131
3132                         /* The first time, paranoia assumes that the
3133                          * randomization function isn't as strong.  But,
3134                          * this secret initialization is delayed until
3135                          * the last possible moment (packet arrival).
3136                          * Although that time is observable, it is
3137                          * unpredictably variable.  Mash in the most
3138                          * volatile clock bits available, and expire the
3139                          * secret extra quickly.
3140                          */
3141                         if (unlikely(tcp_secret_primary->expires ==
3142                                      tcp_secret_secondary->expires)) {
3143                                 struct timespec tv;
3144
3145                                 getnstimeofday(&tv);
3146                                 bakery[COOKIE_DIGEST_WORDS+0] ^=
3147                                         (u32)tv.tv_nsec;
3148
3149                                 tcp_secret_secondary->expires = jiffy
3150                                         + TCP_SECRET_1MSL
3151                                         + (0x0f & tcp_cookie_work(bakery, 0));
3152                         } else {
3153                                 tcp_secret_secondary->expires = jiffy
3154                                         + TCP_SECRET_LIFE
3155                                         + (0xff & tcp_cookie_work(bakery, 1));
3156                                 tcp_secret_primary->expires = jiffy
3157                                         + TCP_SECRET_2MSL
3158                                         + (0x1f & tcp_cookie_work(bakery, 2));
3159                         }
3160                         memcpy(&tcp_secret_secondary->secrets[0],
3161                                bakery, COOKIE_WORKSPACE_WORDS);
3162
3163                         rcu_assign_pointer(tcp_secret_generating,
3164                                            tcp_secret_secondary);
3165                         rcu_assign_pointer(tcp_secret_retiring,
3166                                            tcp_secret_primary);
3167                         /*
3168                          * Neither call_rcu() nor synchronize_rcu() needed.
3169                          * Retiring data is not freed.  It is replaced after
3170                          * further (locked) pointer updates, and a quiet time
3171                          * (minimum 1MSL, maximum LIFE - 2MSL).
3172                          */
3173                 }
3174                 spin_unlock_bh(&tcp_secret_locker);
3175         } else {
3176                 rcu_read_lock_bh();
3177                 memcpy(bakery,
3178                        &rcu_dereference(tcp_secret_generating)->secrets[0],
3179                        COOKIE_WORKSPACE_WORDS);
3180                 rcu_read_unlock_bh();
3181         }
3182         return 0;
3183 }
3184 EXPORT_SYMBOL(tcp_cookie_generator);
3185
3186 void tcp_done(struct sock *sk)
3187 {
3188         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3189                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3190
3191         tcp_set_state(sk, TCP_CLOSE);
3192         tcp_clear_xmit_timers(sk);
3193
3194         sk->sk_shutdown = SHUTDOWN_MASK;
3195
3196         if (!sock_flag(sk, SOCK_DEAD))
3197                 sk->sk_state_change(sk);
3198         else
3199                 inet_csk_destroy_sock(sk);
3200 }
3201 EXPORT_SYMBOL_GPL(tcp_done);
3202
3203 extern struct tcp_congestion_ops tcp_reno;
3204
3205 static __initdata unsigned long thash_entries;
3206 static int __init set_thash_entries(char *str)
3207 {
3208         if (!str)
3209                 return 0;
3210         thash_entries = simple_strtoul(str, &str, 0);
3211         return 1;
3212 }
3213 __setup("thash_entries=", set_thash_entries);
3214
3215 void __init tcp_init(void)
3216 {
3217         struct sk_buff *skb = NULL;
3218         unsigned long limit;
3219         int i, max_share, cnt;
3220         unsigned long jiffy = jiffies;
3221
3222         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3223
3224         percpu_counter_init(&tcp_sockets_allocated, 0);
3225         percpu_counter_init(&tcp_orphan_count, 0);
3226         tcp_hashinfo.bind_bucket_cachep =
3227                 kmem_cache_create("tcp_bind_bucket",
3228                                   sizeof(struct inet_bind_bucket), 0,
3229                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3230
3231         /* Size and allocate the main established and bind bucket
3232          * hash tables.
3233          *
3234          * The methodology is similar to that of the buffer cache.
3235          */
3236         tcp_hashinfo.ehash =
3237                 alloc_large_system_hash("TCP established",
3238                                         sizeof(struct inet_ehash_bucket),
3239                                         thash_entries,
3240                                         (totalram_pages >= 128 * 1024) ?
3241                                         13 : 15,
3242                                         0,
3243                                         NULL,
3244                                         &tcp_hashinfo.ehash_mask,
3245                                         thash_entries ? 0 : 512 * 1024);
3246         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3247                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3248                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3249         }
3250         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3251                 panic("TCP: failed to alloc ehash_locks");
3252         tcp_hashinfo.bhash =
3253                 alloc_large_system_hash("TCP bind",
3254                                         sizeof(struct inet_bind_hashbucket),
3255                                         tcp_hashinfo.ehash_mask + 1,
3256                                         (totalram_pages >= 128 * 1024) ?
3257                                         13 : 15,
3258                                         0,
3259                                         &tcp_hashinfo.bhash_size,
3260                                         NULL,
3261                                         64 * 1024);
3262         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3263         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3264                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3265                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3266         }
3267
3268
3269         cnt = tcp_hashinfo.ehash_mask + 1;
3270
3271         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3272         sysctl_tcp_max_orphans = cnt / 2;
3273         sysctl_max_syn_backlog = max(128, cnt / 256);
3274
3275         limit = nr_free_buffer_pages() / 8;
3276         limit = max(limit, 128UL);
3277         sysctl_tcp_mem[0] = limit / 4 * 3;
3278         sysctl_tcp_mem[1] = limit;
3279         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3280
3281         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3282         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3283         max_share = min(4UL*1024*1024, limit);
3284
3285         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3286         sysctl_tcp_wmem[1] = 16*1024;
3287         sysctl_tcp_wmem[2] = max(64*1024, max_share);
3288
3289         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3290         sysctl_tcp_rmem[1] = 87380;
3291         sysctl_tcp_rmem[2] = max(87380, max_share);
3292
3293         printk(KERN_INFO "TCP: Hash tables configured "
3294                "(established %u bind %u)\n",
3295                tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3296
3297         tcp_register_congestion_control(&tcp_reno);
3298
3299         memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3300         memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3301         tcp_secret_one.expires = jiffy; /* past due */
3302         tcp_secret_two.expires = jiffy; /* past due */
3303         tcp_secret_generating = &tcp_secret_one;
3304         tcp_secret_primary = &tcp_secret_one;
3305         tcp_secret_retiring = &tcp_secret_two;
3306         tcp_secret_secondary = &tcp_secret_two;
3307 }