6afb6d8662b2cebb984bec82657ab4221e95e464
[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
269 #include <net/icmp.h>
270 #include <net/tcp.h>
271 #include <net/xfrm.h>
272 #include <net/ip.h>
273 #include <net/netdma.h>
274 #include <net/sock.h>
275
276 #include <asm/uaccess.h>
277 #include <asm/ioctls.h>
278
279 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
280
281 struct percpu_counter tcp_orphan_count;
282 EXPORT_SYMBOL_GPL(tcp_orphan_count);
283
284 int sysctl_tcp_mem[3] __read_mostly;
285 int sysctl_tcp_wmem[3] __read_mostly;
286 int sysctl_tcp_rmem[3] __read_mostly;
287
288 EXPORT_SYMBOL(sysctl_tcp_mem);
289 EXPORT_SYMBOL(sysctl_tcp_rmem);
290 EXPORT_SYMBOL(sysctl_tcp_wmem);
291
292 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
293 EXPORT_SYMBOL(tcp_memory_allocated);
294
295 /*
296  * Current number of TCP sockets.
297  */
298 struct percpu_counter tcp_sockets_allocated;
299 EXPORT_SYMBOL(tcp_sockets_allocated);
300
301 /*
302  * TCP splice context
303  */
304 struct tcp_splice_state {
305         struct pipe_inode_info *pipe;
306         size_t len;
307         unsigned int flags;
308 };
309
310 /*
311  * Pressure flag: try to collapse.
312  * Technical note: it is used by multiple contexts non atomically.
313  * All the __sk_mem_schedule() is of this nature: accounting
314  * is strict, actions are advisory and have some latency.
315  */
316 int tcp_memory_pressure __read_mostly;
317
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
328 EXPORT_SYMBOL(tcp_enter_memory_pressure);
329
330 /* Convert seconds to retransmits based on initial and max timeout */
331 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
332 {
333         u8 res = 0;
334
335         if (seconds > 0) {
336                 int period = timeout;
337
338                 res = 1;
339                 while (seconds > period && res < 255) {
340                         res++;
341                         timeout <<= 1;
342                         if (timeout > rto_max)
343                                 timeout = rto_max;
344                         period += timeout;
345                 }
346         }
347         return res;
348 }
349
350 /* Convert retransmits to seconds based on initial and max timeout */
351 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
352 {
353         int period = 0;
354
355         if (retrans > 0) {
356                 period = timeout;
357                 while (--retrans) {
358                         timeout <<= 1;
359                         if (timeout > rto_max)
360                                 timeout = rto_max;
361                         period += timeout;
362                 }
363         }
364         return period;
365 }
366
367 /*
368  *      Wait for a TCP event.
369  *
370  *      Note that we don't need to lock the socket, as the upper poll layers
371  *      take care of normal races (between the test and the event) and we don't
372  *      go look at any of the socket buffers directly.
373  */
374 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
375 {
376         unsigned int mask;
377         struct sock *sk = sock->sk;
378         struct tcp_sock *tp = tcp_sk(sk);
379
380         sock_poll_wait(file, sk->sk_sleep, wait);
381         if (sk->sk_state == TCP_LISTEN)
382                 return inet_csk_listen_poll(sk);
383
384         /* Socket is not locked. We are protected from async events
385          * by poll logic and correct handling of state changes
386          * made by other threads is impossible in any case.
387          */
388
389         mask = 0;
390         if (sk->sk_err)
391                 mask = POLLERR;
392
393         /*
394          * POLLHUP is certainly not done right. But poll() doesn't
395          * have a notion of HUP in just one direction, and for a
396          * socket the read side is more interesting.
397          *
398          * Some poll() documentation says that POLLHUP is incompatible
399          * with the POLLOUT/POLLWR flags, so somebody should check this
400          * all. But careful, it tends to be safer to return too many
401          * bits than too few, and you can easily break real applications
402          * if you don't tell them that something has hung up!
403          *
404          * Check-me.
405          *
406          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
407          * our fs/select.c). It means that after we received EOF,
408          * poll always returns immediately, making impossible poll() on write()
409          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
410          * if and only if shutdown has been made in both directions.
411          * Actually, it is interesting to look how Solaris and DUX
412          * solve this dilemma. I would prefer, if POLLHUP were maskable,
413          * then we could set it on SND_SHUTDOWN. BTW examples given
414          * in Stevens' books assume exactly this behaviour, it explains
415          * why POLLHUP is incompatible with POLLOUT.    --ANK
416          *
417          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
418          * blocking on fresh not-connected or disconnected socket. --ANK
419          */
420         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
421                 mask |= POLLHUP;
422         if (sk->sk_shutdown & RCV_SHUTDOWN)
423                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
424
425         /* Connected? */
426         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
427                 int target = sock_rcvlowat(sk, 0, INT_MAX);
428
429                 if (tp->urg_seq == tp->copied_seq &&
430                     !sock_flag(sk, SOCK_URGINLINE) &&
431                     tp->urg_data)
432                         target++;
433
434                 /* Potential race condition. If read of tp below will
435                  * escape above sk->sk_state, we can be illegally awaken
436                  * in SYN_* states. */
437                 if (tp->rcv_nxt - tp->copied_seq >= target)
438                         mask |= POLLIN | POLLRDNORM;
439
440                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
441                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
442                                 mask |= POLLOUT | POLLWRNORM;
443                         } else {  /* send SIGIO later */
444                                 set_bit(SOCK_ASYNC_NOSPACE,
445                                         &sk->sk_socket->flags);
446                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
447
448                                 /* Race breaker. If space is freed after
449                                  * wspace test but before the flags are set,
450                                  * IO signal will be lost.
451                                  */
452                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
453                                         mask |= POLLOUT | POLLWRNORM;
454                         }
455                 }
456
457                 if (tp->urg_data & TCP_URG_VALID)
458                         mask |= POLLPRI;
459         }
460         return mask;
461 }
462
463 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
464 {
465         struct tcp_sock *tp = tcp_sk(sk);
466         int answ;
467
468         switch (cmd) {
469         case SIOCINQ:
470                 if (sk->sk_state == TCP_LISTEN)
471                         return -EINVAL;
472
473                 lock_sock(sk);
474                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
475                         answ = 0;
476                 else if (sock_flag(sk, SOCK_URGINLINE) ||
477                          !tp->urg_data ||
478                          before(tp->urg_seq, tp->copied_seq) ||
479                          !before(tp->urg_seq, tp->rcv_nxt)) {
480                         struct sk_buff *skb;
481
482                         answ = tp->rcv_nxt - tp->copied_seq;
483
484                         /* Subtract 1, if FIN is in queue. */
485                         skb = skb_peek_tail(&sk->sk_receive_queue);
486                         if (answ && skb)
487                                 answ -= tcp_hdr(skb)->fin;
488                 } else
489                         answ = tp->urg_seq - tp->copied_seq;
490                 release_sock(sk);
491                 break;
492         case SIOCATMARK:
493                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
494                 break;
495         case SIOCOUTQ:
496                 if (sk->sk_state == TCP_LISTEN)
497                         return -EINVAL;
498
499                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
500                         answ = 0;
501                 else
502                         answ = tp->write_seq - tp->snd_una;
503                 break;
504         default:
505                 return -ENOIOCTLCMD;
506         }
507
508         return put_user(answ, (int __user *)arg);
509 }
510
511 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
512 {
513         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
514         tp->pushed_seq = tp->write_seq;
515 }
516
517 static inline int forced_push(struct tcp_sock *tp)
518 {
519         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
520 }
521
522 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
523 {
524         struct tcp_sock *tp = tcp_sk(sk);
525         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
526
527         skb->csum    = 0;
528         tcb->seq     = tcb->end_seq = tp->write_seq;
529         tcb->flags   = TCPCB_FLAG_ACK;
530         tcb->sacked  = 0;
531         skb_header_release(skb);
532         tcp_add_write_queue_tail(sk, skb);
533         sk->sk_wmem_queued += skb->truesize;
534         sk_mem_charge(sk, skb->truesize);
535         if (tp->nonagle & TCP_NAGLE_PUSH)
536                 tp->nonagle &= ~TCP_NAGLE_PUSH;
537 }
538
539 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
540 {
541         if (flags & MSG_OOB)
542                 tp->snd_up = tp->write_seq;
543 }
544
545 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
546                             int nonagle)
547 {
548         if (tcp_send_head(sk)) {
549                 struct tcp_sock *tp = tcp_sk(sk);
550
551                 if (!(flags & MSG_MORE) || forced_push(tp))
552                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
553
554                 tcp_mark_urg(tp, flags);
555                 __tcp_push_pending_frames(sk, mss_now,
556                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
557         }
558 }
559
560 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
561                                 unsigned int offset, size_t len)
562 {
563         struct tcp_splice_state *tss = rd_desc->arg.data;
564         int ret;
565
566         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
567                               tss->flags);
568         if (ret > 0)
569                 rd_desc->count -= ret;
570         return ret;
571 }
572
573 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
574 {
575         /* Store TCP splice context information in read_descriptor_t. */
576         read_descriptor_t rd_desc = {
577                 .arg.data = tss,
578                 .count    = tss->len,
579         };
580
581         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
582 }
583
584 /**
585  *  tcp_splice_read - splice data from TCP socket to a pipe
586  * @sock:       socket to splice from
587  * @ppos:       position (not valid)
588  * @pipe:       pipe to splice to
589  * @len:        number of bytes to splice
590  * @flags:      splice modifier flags
591  *
592  * Description:
593  *    Will read pages from given socket and fill them into a pipe.
594  *
595  **/
596 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
597                         struct pipe_inode_info *pipe, size_t len,
598                         unsigned int flags)
599 {
600         struct sock *sk = sock->sk;
601         struct tcp_splice_state tss = {
602                 .pipe = pipe,
603                 .len = len,
604                 .flags = flags,
605         };
606         long timeo;
607         ssize_t spliced;
608         int ret;
609
610         /*
611          * We can't seek on a socket input
612          */
613         if (unlikely(*ppos))
614                 return -ESPIPE;
615
616         ret = spliced = 0;
617
618         lock_sock(sk);
619
620         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
621         while (tss.len) {
622                 ret = __tcp_splice_read(sk, &tss);
623                 if (ret < 0)
624                         break;
625                 else if (!ret) {
626                         if (spliced)
627                                 break;
628                         if (sock_flag(sk, SOCK_DONE))
629                                 break;
630                         if (sk->sk_err) {
631                                 ret = sock_error(sk);
632                                 break;
633                         }
634                         if (sk->sk_shutdown & RCV_SHUTDOWN)
635                                 break;
636                         if (sk->sk_state == TCP_CLOSE) {
637                                 /*
638                                  * This occurs when user tries to read
639                                  * from never connected socket.
640                                  */
641                                 if (!sock_flag(sk, SOCK_DONE))
642                                         ret = -ENOTCONN;
643                                 break;
644                         }
645                         if (!timeo) {
646                                 ret = -EAGAIN;
647                                 break;
648                         }
649                         sk_wait_data(sk, &timeo);
650                         if (signal_pending(current)) {
651                                 ret = sock_intr_errno(timeo);
652                                 break;
653                         }
654                         continue;
655                 }
656                 tss.len -= ret;
657                 spliced += ret;
658
659                 if (!timeo)
660                         break;
661                 release_sock(sk);
662                 lock_sock(sk);
663
664                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
665                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
666                     signal_pending(current))
667                         break;
668         }
669
670         release_sock(sk);
671
672         if (spliced)
673                 return spliced;
674
675         return ret;
676 }
677
678 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
679 {
680         struct sk_buff *skb;
681
682         /* The TCP header must be at least 32-bit aligned.  */
683         size = ALIGN(size, 4);
684
685         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
686         if (skb) {
687                 if (sk_wmem_schedule(sk, skb->truesize)) {
688                         /*
689                          * Make sure that we have exactly size bytes
690                          * available to the caller, no more, no less.
691                          */
692                         skb_reserve(skb, skb_tailroom(skb) - size);
693                         return skb;
694                 }
695                 __kfree_skb(skb);
696         } else {
697                 sk->sk_prot->enter_memory_pressure(sk);
698                 sk_stream_moderate_sndbuf(sk);
699         }
700         return NULL;
701 }
702
703 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
704                                        int large_allowed)
705 {
706         struct tcp_sock *tp = tcp_sk(sk);
707         u32 xmit_size_goal, old_size_goal;
708
709         xmit_size_goal = mss_now;
710
711         if (large_allowed && sk_can_gso(sk)) {
712                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
713                                   inet_csk(sk)->icsk_af_ops->net_header_len -
714                                   inet_csk(sk)->icsk_ext_hdr_len -
715                                   tp->tcp_header_len);
716
717                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
718
719                 /* We try hard to avoid divides here */
720                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
721
722                 if (likely(old_size_goal <= xmit_size_goal &&
723                            old_size_goal + mss_now > xmit_size_goal)) {
724                         xmit_size_goal = old_size_goal;
725                 } else {
726                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
727                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
728                 }
729         }
730
731         return max(xmit_size_goal, mss_now);
732 }
733
734 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
735 {
736         int mss_now;
737
738         mss_now = tcp_current_mss(sk);
739         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
740
741         return mss_now;
742 }
743
744 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
745                          size_t psize, int flags)
746 {
747         struct tcp_sock *tp = tcp_sk(sk);
748         int mss_now, size_goal;
749         int err;
750         ssize_t copied;
751         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
752
753         /* Wait for a connection to finish. */
754         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
755                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
756                         goto out_err;
757
758         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
759
760         mss_now = tcp_send_mss(sk, &size_goal, flags);
761         copied = 0;
762
763         err = -EPIPE;
764         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
765                 goto out_err;
766
767         while (psize > 0) {
768                 struct sk_buff *skb = tcp_write_queue_tail(sk);
769                 struct page *page = pages[poffset / PAGE_SIZE];
770                 int copy, i, can_coalesce;
771                 int offset = poffset % PAGE_SIZE;
772                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
773
774                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
775 new_segment:
776                         if (!sk_stream_memory_free(sk))
777                                 goto wait_for_sndbuf;
778
779                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
780                         if (!skb)
781                                 goto wait_for_memory;
782
783                         skb_entail(sk, skb);
784                         copy = size_goal;
785                 }
786
787                 if (copy > size)
788                         copy = size;
789
790                 i = skb_shinfo(skb)->nr_frags;
791                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
792                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
793                         tcp_mark_push(tp, skb);
794                         goto new_segment;
795                 }
796                 if (!sk_wmem_schedule(sk, copy))
797                         goto wait_for_memory;
798
799                 if (can_coalesce) {
800                         skb_shinfo(skb)->frags[i - 1].size += copy;
801                 } else {
802                         get_page(page);
803                         skb_fill_page_desc(skb, i, page, offset, copy);
804                 }
805
806                 skb->len += copy;
807                 skb->data_len += copy;
808                 skb->truesize += copy;
809                 sk->sk_wmem_queued += copy;
810                 sk_mem_charge(sk, copy);
811                 skb->ip_summed = CHECKSUM_PARTIAL;
812                 tp->write_seq += copy;
813                 TCP_SKB_CB(skb)->end_seq += copy;
814                 skb_shinfo(skb)->gso_segs = 0;
815
816                 if (!copied)
817                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
818
819                 copied += copy;
820                 poffset += copy;
821                 if (!(psize -= copy))
822                         goto out;
823
824                 if (skb->len < size_goal || (flags & MSG_OOB))
825                         continue;
826
827                 if (forced_push(tp)) {
828                         tcp_mark_push(tp, skb);
829                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
830                 } else if (skb == tcp_send_head(sk))
831                         tcp_push_one(sk, mss_now);
832                 continue;
833
834 wait_for_sndbuf:
835                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
836 wait_for_memory:
837                 if (copied)
838                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
839
840                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
841                         goto do_error;
842
843                 mss_now = tcp_send_mss(sk, &size_goal, flags);
844         }
845
846 out:
847         if (copied)
848                 tcp_push(sk, flags, mss_now, tp->nonagle);
849         return copied;
850
851 do_error:
852         if (copied)
853                 goto out;
854 out_err:
855         return sk_stream_error(sk, flags, err);
856 }
857
858 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
859                      size_t size, int flags)
860 {
861         ssize_t res;
862         struct sock *sk = sock->sk;
863
864         if (!(sk->sk_route_caps & NETIF_F_SG) ||
865             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
866                 return sock_no_sendpage(sock, page, offset, size, flags);
867
868         lock_sock(sk);
869         TCP_CHECK_TIMER(sk);
870         res = do_tcp_sendpages(sk, &page, offset, size, flags);
871         TCP_CHECK_TIMER(sk);
872         release_sock(sk);
873         return res;
874 }
875
876 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
877 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
878
879 static inline int select_size(struct sock *sk, int sg)
880 {
881         struct tcp_sock *tp = tcp_sk(sk);
882         int tmp = tp->mss_cache;
883
884         if (sg) {
885                 if (sk_can_gso(sk))
886                         tmp = 0;
887                 else {
888                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
889
890                         if (tmp >= pgbreak &&
891                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
892                                 tmp = pgbreak;
893                 }
894         }
895
896         return tmp;
897 }
898
899 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
900                 size_t size)
901 {
902         struct sock *sk = sock->sk;
903         struct iovec *iov;
904         struct tcp_sock *tp = tcp_sk(sk);
905         struct sk_buff *skb;
906         int iovlen, flags;
907         int mss_now, size_goal;
908         int sg, err, copied;
909         long timeo;
910
911         lock_sock(sk);
912         TCP_CHECK_TIMER(sk);
913
914         flags = msg->msg_flags;
915         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
916
917         /* Wait for a connection to finish. */
918         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
919                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
920                         goto out_err;
921
922         /* This should be in poll */
923         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
924
925         mss_now = tcp_send_mss(sk, &size_goal, flags);
926
927         /* Ok commence sending. */
928         iovlen = msg->msg_iovlen;
929         iov = msg->msg_iov;
930         copied = 0;
931
932         err = -EPIPE;
933         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
934                 goto out_err;
935
936         sg = sk->sk_route_caps & NETIF_F_SG;
937
938         while (--iovlen >= 0) {
939                 int seglen = iov->iov_len;
940                 unsigned char __user *from = iov->iov_base;
941
942                 iov++;
943
944                 while (seglen > 0) {
945                         int copy = 0;
946                         int max = size_goal;
947
948                         skb = tcp_write_queue_tail(sk);
949                         if (tcp_send_head(sk)) {
950                                 if (skb->ip_summed == CHECKSUM_NONE)
951                                         max = mss_now;
952                                 copy = max - skb->len;
953                         }
954
955                         if (copy <= 0) {
956 new_segment:
957                                 /* Allocate new segment. If the interface is SG,
958                                  * allocate skb fitting to single page.
959                                  */
960                                 if (!sk_stream_memory_free(sk))
961                                         goto wait_for_sndbuf;
962
963                                 skb = sk_stream_alloc_skb(sk,
964                                                           select_size(sk, sg),
965                                                           sk->sk_allocation);
966                                 if (!skb)
967                                         goto wait_for_memory;
968
969                                 /*
970                                  * Check whether we can use HW checksum.
971                                  */
972                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
973                                         skb->ip_summed = CHECKSUM_PARTIAL;
974
975                                 skb_entail(sk, skb);
976                                 copy = size_goal;
977                                 max = size_goal;
978                         }
979
980                         /* Try to append data to the end of skb. */
981                         if (copy > seglen)
982                                 copy = seglen;
983
984                         /* Where to copy to? */
985                         if (skb_tailroom(skb) > 0) {
986                                 /* We have some space in skb head. Superb! */
987                                 if (copy > skb_tailroom(skb))
988                                         copy = skb_tailroom(skb);
989                                 if ((err = skb_add_data(skb, from, copy)) != 0)
990                                         goto do_fault;
991                         } else {
992                                 int merge = 0;
993                                 int i = skb_shinfo(skb)->nr_frags;
994                                 struct page *page = TCP_PAGE(sk);
995                                 int off = TCP_OFF(sk);
996
997                                 if (skb_can_coalesce(skb, i, page, off) &&
998                                     off != PAGE_SIZE) {
999                                         /* We can extend the last page
1000                                          * fragment. */
1001                                         merge = 1;
1002                                 } else if (i == MAX_SKB_FRAGS || !sg) {
1003                                         /* Need to add new fragment and cannot
1004                                          * do this because interface is non-SG,
1005                                          * or because all the page slots are
1006                                          * busy. */
1007                                         tcp_mark_push(tp, skb);
1008                                         goto new_segment;
1009                                 } else if (page) {
1010                                         if (off == PAGE_SIZE) {
1011                                                 put_page(page);
1012                                                 TCP_PAGE(sk) = page = NULL;
1013                                                 off = 0;
1014                                         }
1015                                 } else
1016                                         off = 0;
1017
1018                                 if (copy > PAGE_SIZE - off)
1019                                         copy = PAGE_SIZE - off;
1020
1021                                 if (!sk_wmem_schedule(sk, copy))
1022                                         goto wait_for_memory;
1023
1024                                 if (!page) {
1025                                         /* Allocate new cache page. */
1026                                         if (!(page = sk_stream_alloc_page(sk)))
1027                                                 goto wait_for_memory;
1028                                 }
1029
1030                                 /* Time to copy data. We are close to
1031                                  * the end! */
1032                                 err = skb_copy_to_page(sk, from, skb, page,
1033                                                        off, copy);
1034                                 if (err) {
1035                                         /* If this page was new, give it to the
1036                                          * socket so it does not get leaked.
1037                                          */
1038                                         if (!TCP_PAGE(sk)) {
1039                                                 TCP_PAGE(sk) = page;
1040                                                 TCP_OFF(sk) = 0;
1041                                         }
1042                                         goto do_error;
1043                                 }
1044
1045                                 /* Update the skb. */
1046                                 if (merge) {
1047                                         skb_shinfo(skb)->frags[i - 1].size +=
1048                                                                         copy;
1049                                 } else {
1050                                         skb_fill_page_desc(skb, i, page, off, copy);
1051                                         if (TCP_PAGE(sk)) {
1052                                                 get_page(page);
1053                                         } else if (off + copy < PAGE_SIZE) {
1054                                                 get_page(page);
1055                                                 TCP_PAGE(sk) = page;
1056                                         }
1057                                 }
1058
1059                                 TCP_OFF(sk) = off + copy;
1060                         }
1061
1062                         if (!copied)
1063                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
1064
1065                         tp->write_seq += copy;
1066                         TCP_SKB_CB(skb)->end_seq += copy;
1067                         skb_shinfo(skb)->gso_segs = 0;
1068
1069                         from += copy;
1070                         copied += copy;
1071                         if ((seglen -= copy) == 0 && iovlen == 0)
1072                                 goto out;
1073
1074                         if (skb->len < max || (flags & MSG_OOB))
1075                                 continue;
1076
1077                         if (forced_push(tp)) {
1078                                 tcp_mark_push(tp, skb);
1079                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1080                         } else if (skb == tcp_send_head(sk))
1081                                 tcp_push_one(sk, mss_now);
1082                         continue;
1083
1084 wait_for_sndbuf:
1085                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1086 wait_for_memory:
1087                         if (copied)
1088                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1089
1090                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1091                                 goto do_error;
1092
1093                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1094                 }
1095         }
1096
1097 out:
1098         if (copied)
1099                 tcp_push(sk, flags, mss_now, tp->nonagle);
1100         TCP_CHECK_TIMER(sk);
1101         release_sock(sk);
1102         return copied;
1103
1104 do_fault:
1105         if (!skb->len) {
1106                 tcp_unlink_write_queue(skb, sk);
1107                 /* It is the one place in all of TCP, except connection
1108                  * reset, where we can be unlinking the send_head.
1109                  */
1110                 tcp_check_send_head(sk, skb);
1111                 sk_wmem_free_skb(sk, skb);
1112         }
1113
1114 do_error:
1115         if (copied)
1116                 goto out;
1117 out_err:
1118         err = sk_stream_error(sk, flags, err);
1119         TCP_CHECK_TIMER(sk);
1120         release_sock(sk);
1121         return err;
1122 }
1123
1124 /*
1125  *      Handle reading urgent data. BSD has very simple semantics for
1126  *      this, no blocking and very strange errors 8)
1127  */
1128
1129 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1130 {
1131         struct tcp_sock *tp = tcp_sk(sk);
1132
1133         /* No URG data to read. */
1134         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1135             tp->urg_data == TCP_URG_READ)
1136                 return -EINVAL; /* Yes this is right ! */
1137
1138         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1139                 return -ENOTCONN;
1140
1141         if (tp->urg_data & TCP_URG_VALID) {
1142                 int err = 0;
1143                 char c = tp->urg_data;
1144
1145                 if (!(flags & MSG_PEEK))
1146                         tp->urg_data = TCP_URG_READ;
1147
1148                 /* Read urgent data. */
1149                 msg->msg_flags |= MSG_OOB;
1150
1151                 if (len > 0) {
1152                         if (!(flags & MSG_TRUNC))
1153                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1154                         len = 1;
1155                 } else
1156                         msg->msg_flags |= MSG_TRUNC;
1157
1158                 return err ? -EFAULT : len;
1159         }
1160
1161         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1162                 return 0;
1163
1164         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1165          * the available implementations agree in this case:
1166          * this call should never block, independent of the
1167          * blocking state of the socket.
1168          * Mike <pall@rz.uni-karlsruhe.de>
1169          */
1170         return -EAGAIN;
1171 }
1172
1173 /* Clean up the receive buffer for full frames taken by the user,
1174  * then send an ACK if necessary.  COPIED is the number of bytes
1175  * tcp_recvmsg has given to the user so far, it speeds up the
1176  * calculation of whether or not we must ACK for the sake of
1177  * a window update.
1178  */
1179 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1180 {
1181         struct tcp_sock *tp = tcp_sk(sk);
1182         int time_to_ack = 0;
1183
1184 #if TCP_DEBUG
1185         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1186
1187         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1188              KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1189              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1190 #endif
1191
1192         if (inet_csk_ack_scheduled(sk)) {
1193                 const struct inet_connection_sock *icsk = inet_csk(sk);
1194                    /* Delayed ACKs frequently hit locked sockets during bulk
1195                     * receive. */
1196                 if (icsk->icsk_ack.blocked ||
1197                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1198                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1199                     /*
1200                      * If this read emptied read buffer, we send ACK, if
1201                      * connection is not bidirectional, user drained
1202                      * receive buffer and there was a small segment
1203                      * in queue.
1204                      */
1205                     (copied > 0 &&
1206                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1207                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1208                        !icsk->icsk_ack.pingpong)) &&
1209                       !atomic_read(&sk->sk_rmem_alloc)))
1210                         time_to_ack = 1;
1211         }
1212
1213         /* We send an ACK if we can now advertise a non-zero window
1214          * which has been raised "significantly".
1215          *
1216          * Even if window raised up to infinity, do not send window open ACK
1217          * in states, where we will not receive more. It is useless.
1218          */
1219         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1220                 __u32 rcv_window_now = tcp_receive_window(tp);
1221
1222                 /* Optimize, __tcp_select_window() is not cheap. */
1223                 if (2*rcv_window_now <= tp->window_clamp) {
1224                         __u32 new_window = __tcp_select_window(sk);
1225
1226                         /* Send ACK now, if this read freed lots of space
1227                          * in our buffer. Certainly, new_window is new window.
1228                          * We can advertise it now, if it is not less than current one.
1229                          * "Lots" means "at least twice" here.
1230                          */
1231                         if (new_window && new_window >= 2 * rcv_window_now)
1232                                 time_to_ack = 1;
1233                 }
1234         }
1235         if (time_to_ack)
1236                 tcp_send_ack(sk);
1237 }
1238
1239 static void tcp_prequeue_process(struct sock *sk)
1240 {
1241         struct sk_buff *skb;
1242         struct tcp_sock *tp = tcp_sk(sk);
1243
1244         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1245
1246         /* RX process wants to run with disabled BHs, though it is not
1247          * necessary */
1248         local_bh_disable();
1249         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1250                 sk_backlog_rcv(sk, skb);
1251         local_bh_enable();
1252
1253         /* Clear memory counter. */
1254         tp->ucopy.memory = 0;
1255 }
1256
1257 #ifdef CONFIG_NET_DMA
1258 static void tcp_service_net_dma(struct sock *sk, bool wait)
1259 {
1260         dma_cookie_t done, used;
1261         dma_cookie_t last_issued;
1262         struct tcp_sock *tp = tcp_sk(sk);
1263
1264         if (!tp->ucopy.dma_chan)
1265                 return;
1266
1267         last_issued = tp->ucopy.dma_cookie;
1268         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1269
1270         do {
1271                 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1272                                               last_issued, &done,
1273                                               &used) == DMA_SUCCESS) {
1274                         /* Safe to free early-copied skbs now */
1275                         __skb_queue_purge(&sk->sk_async_wait_queue);
1276                         break;
1277                 } else {
1278                         struct sk_buff *skb;
1279                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1280                                (dma_async_is_complete(skb->dma_cookie, done,
1281                                                       used) == DMA_SUCCESS)) {
1282                                 __skb_dequeue(&sk->sk_async_wait_queue);
1283                                 kfree_skb(skb);
1284                         }
1285                 }
1286         } while (wait);
1287 }
1288 #endif
1289
1290 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1291 {
1292         struct sk_buff *skb;
1293         u32 offset;
1294
1295         skb_queue_walk(&sk->sk_receive_queue, skb) {
1296                 offset = seq - TCP_SKB_CB(skb)->seq;
1297                 if (tcp_hdr(skb)->syn)
1298                         offset--;
1299                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1300                         *off = offset;
1301                         return skb;
1302                 }
1303         }
1304         return NULL;
1305 }
1306
1307 /*
1308  * This routine provides an alternative to tcp_recvmsg() for routines
1309  * that would like to handle copying from skbuffs directly in 'sendfile'
1310  * fashion.
1311  * Note:
1312  *      - It is assumed that the socket was locked by the caller.
1313  *      - The routine does not block.
1314  *      - At present, there is no support for reading OOB data
1315  *        or for 'peeking' the socket using this routine
1316  *        (although both would be easy to implement).
1317  */
1318 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1319                   sk_read_actor_t recv_actor)
1320 {
1321         struct sk_buff *skb;
1322         struct tcp_sock *tp = tcp_sk(sk);
1323         u32 seq = tp->copied_seq;
1324         u32 offset;
1325         int copied = 0;
1326
1327         if (sk->sk_state == TCP_LISTEN)
1328                 return -ENOTCONN;
1329         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1330                 if (offset < skb->len) {
1331                         int used;
1332                         size_t len;
1333
1334                         len = skb->len - offset;
1335                         /* Stop reading if we hit a patch of urgent data */
1336                         if (tp->urg_data) {
1337                                 u32 urg_offset = tp->urg_seq - seq;
1338                                 if (urg_offset < len)
1339                                         len = urg_offset;
1340                                 if (!len)
1341                                         break;
1342                         }
1343                         used = recv_actor(desc, skb, offset, len);
1344                         if (used < 0) {
1345                                 if (!copied)
1346                                         copied = used;
1347                                 break;
1348                         } else if (used <= len) {
1349                                 seq += used;
1350                                 copied += used;
1351                                 offset += used;
1352                         }
1353                         /*
1354                          * If recv_actor drops the lock (e.g. TCP splice
1355                          * receive) the skb pointer might be invalid when
1356                          * getting here: tcp_collapse might have deleted it
1357                          * while aggregating skbs from the socket queue.
1358                          */
1359                         skb = tcp_recv_skb(sk, seq-1, &offset);
1360                         if (!skb || (offset+1 != skb->len))
1361                                 break;
1362                 }
1363                 if (tcp_hdr(skb)->fin) {
1364                         sk_eat_skb(sk, skb, 0);
1365                         ++seq;
1366                         break;
1367                 }
1368                 sk_eat_skb(sk, skb, 0);
1369                 if (!desc->count)
1370                         break;
1371         }
1372         tp->copied_seq = seq;
1373
1374         tcp_rcv_space_adjust(sk);
1375
1376         /* Clean up data we have read: This will do ACK frames. */
1377         if (copied > 0)
1378                 tcp_cleanup_rbuf(sk, copied);
1379         return copied;
1380 }
1381
1382 /*
1383  *      This routine copies from a sock struct into the user buffer.
1384  *
1385  *      Technical note: in 2.3 we work on _locked_ socket, so that
1386  *      tricks with *seq access order and skb->users are not required.
1387  *      Probably, code can be easily improved even more.
1388  */
1389
1390 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1391                 size_t len, int nonblock, int flags, int *addr_len)
1392 {
1393         struct tcp_sock *tp = tcp_sk(sk);
1394         int copied = 0;
1395         u32 peek_seq;
1396         u32 *seq;
1397         unsigned long used;
1398         int err;
1399         int target;             /* Read at least this many bytes */
1400         long timeo;
1401         struct task_struct *user_recv = NULL;
1402         int copied_early = 0;
1403         struct sk_buff *skb;
1404         u32 urg_hole = 0;
1405
1406         lock_sock(sk);
1407
1408         TCP_CHECK_TIMER(sk);
1409
1410         err = -ENOTCONN;
1411         if (sk->sk_state == TCP_LISTEN)
1412                 goto out;
1413
1414         timeo = sock_rcvtimeo(sk, nonblock);
1415
1416         /* Urgent data needs to be handled specially. */
1417         if (flags & MSG_OOB)
1418                 goto recv_urg;
1419
1420         seq = &tp->copied_seq;
1421         if (flags & MSG_PEEK) {
1422                 peek_seq = tp->copied_seq;
1423                 seq = &peek_seq;
1424         }
1425
1426         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1427
1428 #ifdef CONFIG_NET_DMA
1429         tp->ucopy.dma_chan = NULL;
1430         preempt_disable();
1431         skb = skb_peek_tail(&sk->sk_receive_queue);
1432         {
1433                 int available = 0;
1434
1435                 if (skb)
1436                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1437                 if ((available < target) &&
1438                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1439                     !sysctl_tcp_low_latency &&
1440                     dma_find_channel(DMA_MEMCPY)) {
1441                         preempt_enable_no_resched();
1442                         tp->ucopy.pinned_list =
1443                                         dma_pin_iovec_pages(msg->msg_iov, len);
1444                 } else {
1445                         preempt_enable_no_resched();
1446                 }
1447         }
1448 #endif
1449
1450         do {
1451                 u32 offset;
1452
1453                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1454                 if (tp->urg_data && tp->urg_seq == *seq) {
1455                         if (copied)
1456                                 break;
1457                         if (signal_pending(current)) {
1458                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1459                                 break;
1460                         }
1461                 }
1462
1463                 /* Next get a buffer. */
1464
1465                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1466                         /* Now that we have two receive queues this
1467                          * shouldn't happen.
1468                          */
1469                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1470                              KERN_INFO "recvmsg bug: copied %X "
1471                                        "seq %X rcvnxt %X fl %X\n", *seq,
1472                                        TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1473                                        flags))
1474                                 break;
1475
1476                         offset = *seq - TCP_SKB_CB(skb)->seq;
1477                         if (tcp_hdr(skb)->syn)
1478                                 offset--;
1479                         if (offset < skb->len)
1480                                 goto found_ok_skb;
1481                         if (tcp_hdr(skb)->fin)
1482                                 goto found_fin_ok;
1483                         WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: "
1484                                         "copied %X seq %X rcvnxt %X fl %X\n",
1485                                         *seq, TCP_SKB_CB(skb)->seq,
1486                                         tp->rcv_nxt, flags);
1487                 }
1488
1489                 /* Well, if we have backlog, try to process it now yet. */
1490
1491                 if (copied >= target && !sk->sk_backlog.tail)
1492                         break;
1493
1494                 if (copied) {
1495                         if (sk->sk_err ||
1496                             sk->sk_state == TCP_CLOSE ||
1497                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1498                             !timeo ||
1499                             signal_pending(current))
1500                                 break;
1501                 } else {
1502                         if (sock_flag(sk, SOCK_DONE))
1503                                 break;
1504
1505                         if (sk->sk_err) {
1506                                 copied = sock_error(sk);
1507                                 break;
1508                         }
1509
1510                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1511                                 break;
1512
1513                         if (sk->sk_state == TCP_CLOSE) {
1514                                 if (!sock_flag(sk, SOCK_DONE)) {
1515                                         /* This occurs when user tries to read
1516                                          * from never connected socket.
1517                                          */
1518                                         copied = -ENOTCONN;
1519                                         break;
1520                                 }
1521                                 break;
1522                         }
1523
1524                         if (!timeo) {
1525                                 copied = -EAGAIN;
1526                                 break;
1527                         }
1528
1529                         if (signal_pending(current)) {
1530                                 copied = sock_intr_errno(timeo);
1531                                 break;
1532                         }
1533                 }
1534
1535                 tcp_cleanup_rbuf(sk, copied);
1536
1537                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1538                         /* Install new reader */
1539                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1540                                 user_recv = current;
1541                                 tp->ucopy.task = user_recv;
1542                                 tp->ucopy.iov = msg->msg_iov;
1543                         }
1544
1545                         tp->ucopy.len = len;
1546
1547                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1548                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1549
1550                         /* Ugly... If prequeue is not empty, we have to
1551                          * process it before releasing socket, otherwise
1552                          * order will be broken at second iteration.
1553                          * More elegant solution is required!!!
1554                          *
1555                          * Look: we have the following (pseudo)queues:
1556                          *
1557                          * 1. packets in flight
1558                          * 2. backlog
1559                          * 3. prequeue
1560                          * 4. receive_queue
1561                          *
1562                          * Each queue can be processed only if the next ones
1563                          * are empty. At this point we have empty receive_queue.
1564                          * But prequeue _can_ be not empty after 2nd iteration,
1565                          * when we jumped to start of loop because backlog
1566                          * processing added something to receive_queue.
1567                          * We cannot release_sock(), because backlog contains
1568                          * packets arrived _after_ prequeued ones.
1569                          *
1570                          * Shortly, algorithm is clear --- to process all
1571                          * the queues in order. We could make it more directly,
1572                          * requeueing packets from backlog to prequeue, if
1573                          * is not empty. It is more elegant, but eats cycles,
1574                          * unfortunately.
1575                          */
1576                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1577                                 goto do_prequeue;
1578
1579                         /* __ Set realtime policy in scheduler __ */
1580                 }
1581
1582 #ifdef CONFIG_NET_DMA
1583                 if (tp->ucopy.dma_chan)
1584                         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1585 #endif
1586                 if (copied >= target) {
1587                         /* Do not sleep, just process backlog. */
1588                         release_sock(sk);
1589                         lock_sock(sk);
1590                 } else
1591                         sk_wait_data(sk, &timeo);
1592
1593 #ifdef CONFIG_NET_DMA
1594                 tcp_service_net_dma(sk, false);  /* Don't block */
1595                 tp->ucopy.wakeup = 0;
1596 #endif
1597
1598                 if (user_recv) {
1599                         int chunk;
1600
1601                         /* __ Restore normal policy in scheduler __ */
1602
1603                         if ((chunk = len - tp->ucopy.len) != 0) {
1604                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1605                                 len -= chunk;
1606                                 copied += chunk;
1607                         }
1608
1609                         if (tp->rcv_nxt == tp->copied_seq &&
1610                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1611 do_prequeue:
1612                                 tcp_prequeue_process(sk);
1613
1614                                 if ((chunk = len - tp->ucopy.len) != 0) {
1615                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1616                                         len -= chunk;
1617                                         copied += chunk;
1618                                 }
1619                         }
1620                 }
1621                 if ((flags & MSG_PEEK) &&
1622                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1623                         if (net_ratelimit())
1624                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1625                                        current->comm, task_pid_nr(current));
1626                         peek_seq = tp->copied_seq;
1627                 }
1628                 continue;
1629
1630         found_ok_skb:
1631                 /* Ok so how much can we use? */
1632                 used = skb->len - offset;
1633                 if (len < used)
1634                         used = len;
1635
1636                 /* Do we have urgent data here? */
1637                 if (tp->urg_data) {
1638                         u32 urg_offset = tp->urg_seq - *seq;
1639                         if (urg_offset < used) {
1640                                 if (!urg_offset) {
1641                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1642                                                 ++*seq;
1643                                                 urg_hole++;
1644                                                 offset++;
1645                                                 used--;
1646                                                 if (!used)
1647                                                         goto skip_copy;
1648                                         }
1649                                 } else
1650                                         used = urg_offset;
1651                         }
1652                 }
1653
1654                 if (!(flags & MSG_TRUNC)) {
1655 #ifdef CONFIG_NET_DMA
1656                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1657                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1658
1659                         if (tp->ucopy.dma_chan) {
1660                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1661                                         tp->ucopy.dma_chan, skb, offset,
1662                                         msg->msg_iov, used,
1663                                         tp->ucopy.pinned_list);
1664
1665                                 if (tp->ucopy.dma_cookie < 0) {
1666
1667                                         printk(KERN_ALERT "dma_cookie < 0\n");
1668
1669                                         /* Exception. Bailout! */
1670                                         if (!copied)
1671                                                 copied = -EFAULT;
1672                                         break;
1673                                 }
1674
1675                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1676
1677                                 if ((offset + used) == skb->len)
1678                                         copied_early = 1;
1679
1680                         } else
1681 #endif
1682                         {
1683                                 err = skb_copy_datagram_iovec(skb, offset,
1684                                                 msg->msg_iov, used);
1685                                 if (err) {
1686                                         /* Exception. Bailout! */
1687                                         if (!copied)
1688                                                 copied = -EFAULT;
1689                                         break;
1690                                 }
1691                         }
1692                 }
1693
1694                 *seq += used;
1695                 copied += used;
1696                 len -= used;
1697
1698                 tcp_rcv_space_adjust(sk);
1699
1700 skip_copy:
1701                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1702                         tp->urg_data = 0;
1703                         tcp_fast_path_check(sk);
1704                 }
1705                 if (used + offset < skb->len)
1706                         continue;
1707
1708                 if (tcp_hdr(skb)->fin)
1709                         goto found_fin_ok;
1710                 if (!(flags & MSG_PEEK)) {
1711                         sk_eat_skb(sk, skb, copied_early);
1712                         copied_early = 0;
1713                 }
1714                 continue;
1715
1716         found_fin_ok:
1717                 /* Process the FIN. */
1718                 ++*seq;
1719                 if (!(flags & MSG_PEEK)) {
1720                         sk_eat_skb(sk, skb, copied_early);
1721                         copied_early = 0;
1722                 }
1723                 break;
1724         } while (len > 0);
1725
1726         if (user_recv) {
1727                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1728                         int chunk;
1729
1730                         tp->ucopy.len = copied > 0 ? len : 0;
1731
1732                         tcp_prequeue_process(sk);
1733
1734                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1735                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1736                                 len -= chunk;
1737                                 copied += chunk;
1738                         }
1739                 }
1740
1741                 tp->ucopy.task = NULL;
1742                 tp->ucopy.len = 0;
1743         }
1744
1745 #ifdef CONFIG_NET_DMA
1746         tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1747         tp->ucopy.dma_chan = NULL;
1748
1749         if (tp->ucopy.pinned_list) {
1750                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1751                 tp->ucopy.pinned_list = NULL;
1752         }
1753 #endif
1754
1755         /* According to UNIX98, msg_name/msg_namelen are ignored
1756          * on connected socket. I was just happy when found this 8) --ANK
1757          */
1758
1759         /* Clean up data we have read: This will do ACK frames. */
1760         tcp_cleanup_rbuf(sk, copied);
1761
1762         TCP_CHECK_TIMER(sk);
1763         release_sock(sk);
1764         return copied;
1765
1766 out:
1767         TCP_CHECK_TIMER(sk);
1768         release_sock(sk);
1769         return err;
1770
1771 recv_urg:
1772         err = tcp_recv_urg(sk, msg, len, flags);
1773         goto out;
1774 }
1775
1776 void tcp_set_state(struct sock *sk, int state)
1777 {
1778         int oldstate = sk->sk_state;
1779
1780         switch (state) {
1781         case TCP_ESTABLISHED:
1782                 if (oldstate != TCP_ESTABLISHED)
1783                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1784                 break;
1785
1786         case TCP_CLOSE:
1787                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1788                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1789
1790                 sk->sk_prot->unhash(sk);
1791                 if (inet_csk(sk)->icsk_bind_hash &&
1792                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1793                         inet_put_port(sk);
1794                 /* fall through */
1795         default:
1796                 if (oldstate == TCP_ESTABLISHED)
1797                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1798         }
1799
1800         /* Change state AFTER socket is unhashed to avoid closed
1801          * socket sitting in hash tables.
1802          */
1803         sk->sk_state = state;
1804
1805 #ifdef STATE_TRACE
1806         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1807 #endif
1808 }
1809 EXPORT_SYMBOL_GPL(tcp_set_state);
1810
1811 /*
1812  *      State processing on a close. This implements the state shift for
1813  *      sending our FIN frame. Note that we only send a FIN for some
1814  *      states. A shutdown() may have already sent the FIN, or we may be
1815  *      closed.
1816  */
1817
1818 static const unsigned char new_state[16] = {
1819   /* current state:        new state:      action:      */
1820   /* (Invalid)          */ TCP_CLOSE,
1821   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1822   /* TCP_SYN_SENT       */ TCP_CLOSE,
1823   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1824   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1825   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1826   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1827   /* TCP_CLOSE          */ TCP_CLOSE,
1828   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1829   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1830   /* TCP_LISTEN         */ TCP_CLOSE,
1831   /* TCP_CLOSING        */ TCP_CLOSING,
1832 };
1833
1834 static int tcp_close_state(struct sock *sk)
1835 {
1836         int next = (int)new_state[sk->sk_state];
1837         int ns = next & TCP_STATE_MASK;
1838
1839         tcp_set_state(sk, ns);
1840
1841         return next & TCP_ACTION_FIN;
1842 }
1843
1844 /*
1845  *      Shutdown the sending side of a connection. Much like close except
1846  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1847  */
1848
1849 void tcp_shutdown(struct sock *sk, int how)
1850 {
1851         /*      We need to grab some memory, and put together a FIN,
1852          *      and then put it into the queue to be sent.
1853          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1854          */
1855         if (!(how & SEND_SHUTDOWN))
1856                 return;
1857
1858         /* If we've already sent a FIN, or it's a closed state, skip this. */
1859         if ((1 << sk->sk_state) &
1860             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1861              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1862                 /* Clear out any half completed packets.  FIN if needed. */
1863                 if (tcp_close_state(sk))
1864                         tcp_send_fin(sk);
1865         }
1866 }
1867
1868 void tcp_close(struct sock *sk, long timeout)
1869 {
1870         struct sk_buff *skb;
1871         int data_was_unread = 0;
1872         int state;
1873
1874         lock_sock(sk);
1875         sk->sk_shutdown = SHUTDOWN_MASK;
1876
1877         if (sk->sk_state == TCP_LISTEN) {
1878                 tcp_set_state(sk, TCP_CLOSE);
1879
1880                 /* Special case. */
1881                 inet_csk_listen_stop(sk);
1882
1883                 goto adjudge_to_death;
1884         }
1885
1886         /*  We need to flush the recv. buffs.  We do this only on the
1887          *  descriptor close, not protocol-sourced closes, because the
1888          *  reader process may not have drained the data yet!
1889          */
1890         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1891                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1892                           tcp_hdr(skb)->fin;
1893                 data_was_unread += len;
1894                 __kfree_skb(skb);
1895         }
1896
1897         sk_mem_reclaim(sk);
1898
1899         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1900          * data was lost. To witness the awful effects of the old behavior of
1901          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1902          * GET in an FTP client, suspend the process, wait for the client to
1903          * advertise a zero window, then kill -9 the FTP client, wheee...
1904          * Note: timeout is always zero in such a case.
1905          */
1906         if (data_was_unread) {
1907                 /* Unread data was tossed, zap the connection. */
1908                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1909                 tcp_set_state(sk, TCP_CLOSE);
1910                 tcp_send_active_reset(sk, sk->sk_allocation);
1911         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1912                 /* Check zero linger _after_ checking for unread data. */
1913                 sk->sk_prot->disconnect(sk, 0);
1914                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1915         } else if (tcp_close_state(sk)) {
1916                 /* We FIN if the application ate all the data before
1917                  * zapping the connection.
1918                  */
1919
1920                 /* RED-PEN. Formally speaking, we have broken TCP state
1921                  * machine. State transitions:
1922                  *
1923                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1924                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1925                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1926                  *
1927                  * are legal only when FIN has been sent (i.e. in window),
1928                  * rather than queued out of window. Purists blame.
1929                  *
1930                  * F.e. "RFC state" is ESTABLISHED,
1931                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1932                  *
1933                  * The visible declinations are that sometimes
1934                  * we enter time-wait state, when it is not required really
1935                  * (harmless), do not send active resets, when they are
1936                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1937                  * they look as CLOSING or LAST_ACK for Linux)
1938                  * Probably, I missed some more holelets.
1939                  *                                              --ANK
1940                  */
1941                 tcp_send_fin(sk);
1942         }
1943
1944         sk_stream_wait_close(sk, timeout);
1945
1946 adjudge_to_death:
1947         state = sk->sk_state;
1948         sock_hold(sk);
1949         sock_orphan(sk);
1950
1951         /* It is the last release_sock in its life. It will remove backlog. */
1952         release_sock(sk);
1953
1954
1955         /* Now socket is owned by kernel and we acquire BH lock
1956            to finish close. No need to check for user refs.
1957          */
1958         local_bh_disable();
1959         bh_lock_sock(sk);
1960         WARN_ON(sock_owned_by_user(sk));
1961
1962         percpu_counter_inc(sk->sk_prot->orphan_count);
1963
1964         /* Have we already been destroyed by a softirq or backlog? */
1965         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1966                 goto out;
1967
1968         /*      This is a (useful) BSD violating of the RFC. There is a
1969          *      problem with TCP as specified in that the other end could
1970          *      keep a socket open forever with no application left this end.
1971          *      We use a 3 minute timeout (about the same as BSD) then kill
1972          *      our end. If they send after that then tough - BUT: long enough
1973          *      that we won't make the old 4*rto = almost no time - whoops
1974          *      reset mistake.
1975          *
1976          *      Nope, it was not mistake. It is really desired behaviour
1977          *      f.e. on http servers, when such sockets are useless, but
1978          *      consume significant resources. Let's do it with special
1979          *      linger2 option.                                 --ANK
1980          */
1981
1982         if (sk->sk_state == TCP_FIN_WAIT2) {
1983                 struct tcp_sock *tp = tcp_sk(sk);
1984                 if (tp->linger2 < 0) {
1985                         tcp_set_state(sk, TCP_CLOSE);
1986                         tcp_send_active_reset(sk, GFP_ATOMIC);
1987                         NET_INC_STATS_BH(sock_net(sk),
1988                                         LINUX_MIB_TCPABORTONLINGER);
1989                 } else {
1990                         const int tmo = tcp_fin_time(sk);
1991
1992                         if (tmo > TCP_TIMEWAIT_LEN) {
1993                                 inet_csk_reset_keepalive_timer(sk,
1994                                                 tmo - TCP_TIMEWAIT_LEN);
1995                         } else {
1996                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1997                                 goto out;
1998                         }
1999                 }
2000         }
2001         if (sk->sk_state != TCP_CLOSE) {
2002                 int orphan_count = percpu_counter_read_positive(
2003                                                 sk->sk_prot->orphan_count);
2004
2005                 sk_mem_reclaim(sk);
2006                 if (tcp_too_many_orphans(sk, orphan_count)) {
2007                         if (net_ratelimit())
2008                                 printk(KERN_INFO "TCP: too many of orphaned "
2009                                        "sockets\n");
2010                         tcp_set_state(sk, TCP_CLOSE);
2011                         tcp_send_active_reset(sk, GFP_ATOMIC);
2012                         NET_INC_STATS_BH(sock_net(sk),
2013                                         LINUX_MIB_TCPABORTONMEMORY);
2014                 }
2015         }
2016
2017         if (sk->sk_state == TCP_CLOSE)
2018                 inet_csk_destroy_sock(sk);
2019         /* Otherwise, socket is reprieved until protocol close. */
2020
2021 out:
2022         bh_unlock_sock(sk);
2023         local_bh_enable();
2024         sock_put(sk);
2025 }
2026
2027 /* These states need RST on ABORT according to RFC793 */
2028
2029 static inline int tcp_need_reset(int state)
2030 {
2031         return (1 << state) &
2032                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2033                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2034 }
2035
2036 int tcp_disconnect(struct sock *sk, int flags)
2037 {
2038         struct inet_sock *inet = inet_sk(sk);
2039         struct inet_connection_sock *icsk = inet_csk(sk);
2040         struct tcp_sock *tp = tcp_sk(sk);
2041         int err = 0;
2042         int old_state = sk->sk_state;
2043
2044         if (old_state != TCP_CLOSE)
2045                 tcp_set_state(sk, TCP_CLOSE);
2046
2047         /* ABORT function of RFC793 */
2048         if (old_state == TCP_LISTEN) {
2049                 inet_csk_listen_stop(sk);
2050         } else if (tcp_need_reset(old_state) ||
2051                    (tp->snd_nxt != tp->write_seq &&
2052                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2053                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2054                  * states
2055                  */
2056                 tcp_send_active_reset(sk, gfp_any());
2057                 sk->sk_err = ECONNRESET;
2058         } else if (old_state == TCP_SYN_SENT)
2059                 sk->sk_err = ECONNRESET;
2060
2061         tcp_clear_xmit_timers(sk);
2062         __skb_queue_purge(&sk->sk_receive_queue);
2063         tcp_write_queue_purge(sk);
2064         __skb_queue_purge(&tp->out_of_order_queue);
2065 #ifdef CONFIG_NET_DMA
2066         __skb_queue_purge(&sk->sk_async_wait_queue);
2067 #endif
2068
2069         inet->inet_dport = 0;
2070
2071         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2072                 inet_reset_saddr(sk);
2073
2074         sk->sk_shutdown = 0;
2075         sock_reset_flag(sk, SOCK_DONE);
2076         tp->srtt = 0;
2077         if ((tp->write_seq += tp->max_window + 2) == 0)
2078                 tp->write_seq = 1;
2079         icsk->icsk_backoff = 0;
2080         tp->snd_cwnd = 2;
2081         icsk->icsk_probes_out = 0;
2082         tp->packets_out = 0;
2083         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2084         tp->snd_cwnd_cnt = 0;
2085         tp->bytes_acked = 0;
2086         tp->window_clamp = 0;
2087         tcp_set_ca_state(sk, TCP_CA_Open);
2088         tcp_clear_retrans(tp);
2089         inet_csk_delack_init(sk);
2090         tcp_init_send_head(sk);
2091         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2092         __sk_dst_reset(sk);
2093
2094         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2095
2096         sk->sk_error_report(sk);
2097         return err;
2098 }
2099
2100 /*
2101  *      Socket option code for TCP.
2102  */
2103 static int do_tcp_setsockopt(struct sock *sk, int level,
2104                 int optname, char __user *optval, unsigned int optlen)
2105 {
2106         struct tcp_sock *tp = tcp_sk(sk);
2107         struct inet_connection_sock *icsk = inet_csk(sk);
2108         int val;
2109         int err = 0;
2110
2111         /* These are data/string values, all the others are ints */
2112         switch (optname) {
2113         case TCP_CONGESTION: {
2114                 char name[TCP_CA_NAME_MAX];
2115
2116                 if (optlen < 1)
2117                         return -EINVAL;
2118
2119                 val = strncpy_from_user(name, optval,
2120                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2121                 if (val < 0)
2122                         return -EFAULT;
2123                 name[val] = 0;
2124
2125                 lock_sock(sk);
2126                 err = tcp_set_congestion_control(sk, name);
2127                 release_sock(sk);
2128                 return err;
2129         }
2130         case TCP_COOKIE_TRANSACTIONS: {
2131                 struct tcp_cookie_transactions ctd;
2132                 struct tcp_cookie_values *cvp = NULL;
2133
2134                 if (sizeof(ctd) > optlen)
2135                         return -EINVAL;
2136                 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2137                         return -EFAULT;
2138
2139                 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2140                     ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2141                         return -EINVAL;
2142
2143                 if (ctd.tcpct_cookie_desired == 0) {
2144                         /* default to global value */
2145                 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2146                            ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2147                            ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2148                         return -EINVAL;
2149                 }
2150
2151                 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2152                         /* Supercedes all other values */
2153                         lock_sock(sk);
2154                         if (tp->cookie_values != NULL) {
2155                                 kref_put(&tp->cookie_values->kref,
2156                                          tcp_cookie_values_release);
2157                                 tp->cookie_values = NULL;
2158                         }
2159                         tp->rx_opt.cookie_in_always = 0; /* false */
2160                         tp->rx_opt.cookie_out_never = 1; /* true */
2161                         release_sock(sk);
2162                         return err;
2163                 }
2164
2165                 /* Allocate ancillary memory before locking.
2166                  */
2167                 if (ctd.tcpct_used > 0 ||
2168                     (tp->cookie_values == NULL &&
2169                      (sysctl_tcp_cookie_size > 0 ||
2170                       ctd.tcpct_cookie_desired > 0 ||
2171                       ctd.tcpct_s_data_desired > 0))) {
2172                         cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2173                                       GFP_KERNEL);
2174                         if (cvp == NULL)
2175                                 return -ENOMEM;
2176                 }
2177                 lock_sock(sk);
2178                 tp->rx_opt.cookie_in_always =
2179                         (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2180                 tp->rx_opt.cookie_out_never = 0; /* false */
2181
2182                 if (tp->cookie_values != NULL) {
2183                         if (cvp != NULL) {
2184                                 /* Changed values are recorded by a changed
2185                                  * pointer, ensuring the cookie will differ,
2186                                  * without separately hashing each value later.
2187                                  */
2188                                 kref_put(&tp->cookie_values->kref,
2189                                          tcp_cookie_values_release);
2190                                 kref_init(&cvp->kref);
2191                                 tp->cookie_values = cvp;
2192                         } else {
2193                                 cvp = tp->cookie_values;
2194                         }
2195                 }
2196                 if (cvp != NULL) {
2197                         cvp->cookie_desired = ctd.tcpct_cookie_desired;
2198
2199                         if (ctd.tcpct_used > 0) {
2200                                 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2201                                        ctd.tcpct_used);
2202                                 cvp->s_data_desired = ctd.tcpct_used;
2203                                 cvp->s_data_constant = 1; /* true */
2204                         } else {
2205                                 /* No constant payload data. */
2206                                 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2207                                 cvp->s_data_constant = 0; /* false */
2208                         }
2209                 }
2210                 release_sock(sk);
2211                 return err;
2212         }
2213         default:
2214                 /* fallthru */
2215                 break;
2216         };
2217
2218         if (optlen < sizeof(int))
2219                 return -EINVAL;
2220
2221         if (get_user(val, (int __user *)optval))
2222                 return -EFAULT;
2223
2224         lock_sock(sk);
2225
2226         switch (optname) {
2227         case TCP_MAXSEG:
2228                 /* Values greater than interface MTU won't take effect. However
2229                  * at the point when this call is done we typically don't yet
2230                  * know which interface is going to be used */
2231                 if (val < 8 || val > MAX_TCP_WINDOW) {
2232                         err = -EINVAL;
2233                         break;
2234                 }
2235                 tp->rx_opt.user_mss = val;
2236                 break;
2237
2238         case TCP_NODELAY:
2239                 if (val) {
2240                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2241                          * this option on corked socket is remembered, but
2242                          * it is not activated until cork is cleared.
2243                          *
2244                          * However, when TCP_NODELAY is set we make
2245                          * an explicit push, which overrides even TCP_CORK
2246                          * for currently queued segments.
2247                          */
2248                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2249                         tcp_push_pending_frames(sk);
2250                 } else {
2251                         tp->nonagle &= ~TCP_NAGLE_OFF;
2252                 }
2253                 break;
2254
2255         case TCP_THIN_LINEAR_TIMEOUTS:
2256                 if (val < 0 || val > 1)
2257                         err = -EINVAL;
2258                 else
2259                         tp->thin_lto = val;
2260                 break;
2261
2262         case TCP_THIN_DUPACK:
2263                 if (val < 0 || val > 1)
2264                         err = -EINVAL;
2265                 else
2266                         tp->thin_dupack = val;
2267                 break;
2268
2269         case TCP_CORK:
2270                 /* When set indicates to always queue non-full frames.
2271                  * Later the user clears this option and we transmit
2272                  * any pending partial frames in the queue.  This is
2273                  * meant to be used alongside sendfile() to get properly
2274                  * filled frames when the user (for example) must write
2275                  * out headers with a write() call first and then use
2276                  * sendfile to send out the data parts.
2277                  *
2278                  * TCP_CORK can be set together with TCP_NODELAY and it is
2279                  * stronger than TCP_NODELAY.
2280                  */
2281                 if (val) {
2282                         tp->nonagle |= TCP_NAGLE_CORK;
2283                 } else {
2284                         tp->nonagle &= ~TCP_NAGLE_CORK;
2285                         if (tp->nonagle&TCP_NAGLE_OFF)
2286                                 tp->nonagle |= TCP_NAGLE_PUSH;
2287                         tcp_push_pending_frames(sk);
2288                 }
2289                 break;
2290
2291         case TCP_KEEPIDLE:
2292                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2293                         err = -EINVAL;
2294                 else {
2295                         tp->keepalive_time = val * HZ;
2296                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2297                             !((1 << sk->sk_state) &
2298                               (TCPF_CLOSE | TCPF_LISTEN))) {
2299                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2300                                 if (tp->keepalive_time > elapsed)
2301                                         elapsed = tp->keepalive_time - elapsed;
2302                                 else
2303                                         elapsed = 0;
2304                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2305                         }
2306                 }
2307                 break;
2308         case TCP_KEEPINTVL:
2309                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2310                         err = -EINVAL;
2311                 else
2312                         tp->keepalive_intvl = val * HZ;
2313                 break;
2314         case TCP_KEEPCNT:
2315                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2316                         err = -EINVAL;
2317                 else
2318                         tp->keepalive_probes = val;
2319                 break;
2320         case TCP_SYNCNT:
2321                 if (val < 1 || val > MAX_TCP_SYNCNT)
2322                         err = -EINVAL;
2323                 else
2324                         icsk->icsk_syn_retries = val;
2325                 break;
2326
2327         case TCP_LINGER2:
2328                 if (val < 0)
2329                         tp->linger2 = -1;
2330                 else if (val > sysctl_tcp_fin_timeout / HZ)
2331                         tp->linger2 = 0;
2332                 else
2333                         tp->linger2 = val * HZ;
2334                 break;
2335
2336         case TCP_DEFER_ACCEPT:
2337                 /* Translate value in seconds to number of retransmits */
2338                 icsk->icsk_accept_queue.rskq_defer_accept =
2339                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2340                                         TCP_RTO_MAX / HZ);
2341                 break;
2342
2343         case TCP_WINDOW_CLAMP:
2344                 if (!val) {
2345                         if (sk->sk_state != TCP_CLOSE) {
2346                                 err = -EINVAL;
2347                                 break;
2348                         }
2349                         tp->window_clamp = 0;
2350                 } else
2351                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2352                                                 SOCK_MIN_RCVBUF / 2 : val;
2353                 break;
2354
2355         case TCP_QUICKACK:
2356                 if (!val) {
2357                         icsk->icsk_ack.pingpong = 1;
2358                 } else {
2359                         icsk->icsk_ack.pingpong = 0;
2360                         if ((1 << sk->sk_state) &
2361                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2362                             inet_csk_ack_scheduled(sk)) {
2363                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2364                                 tcp_cleanup_rbuf(sk, 1);
2365                                 if (!(val & 1))
2366                                         icsk->icsk_ack.pingpong = 1;
2367                         }
2368                 }
2369                 break;
2370
2371 #ifdef CONFIG_TCP_MD5SIG
2372         case TCP_MD5SIG:
2373                 /* Read the IP->Key mappings from userspace */
2374                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2375                 break;
2376 #endif
2377
2378         default:
2379                 err = -ENOPROTOOPT;
2380                 break;
2381         }
2382
2383         release_sock(sk);
2384         return err;
2385 }
2386
2387 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2388                    unsigned int optlen)
2389 {
2390         struct inet_connection_sock *icsk = inet_csk(sk);
2391
2392         if (level != SOL_TCP)
2393                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2394                                                      optval, optlen);
2395         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2396 }
2397
2398 #ifdef CONFIG_COMPAT
2399 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2400                           char __user *optval, unsigned int optlen)
2401 {
2402         if (level != SOL_TCP)
2403                 return inet_csk_compat_setsockopt(sk, level, optname,
2404                                                   optval, optlen);
2405         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2406 }
2407
2408 EXPORT_SYMBOL(compat_tcp_setsockopt);
2409 #endif
2410
2411 /* Return information about state of tcp endpoint in API format. */
2412 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2413 {
2414         struct tcp_sock *tp = tcp_sk(sk);
2415         const struct inet_connection_sock *icsk = inet_csk(sk);
2416         u32 now = tcp_time_stamp;
2417
2418         memset(info, 0, sizeof(*info));
2419
2420         info->tcpi_state = sk->sk_state;
2421         info->tcpi_ca_state = icsk->icsk_ca_state;
2422         info->tcpi_retransmits = icsk->icsk_retransmits;
2423         info->tcpi_probes = icsk->icsk_probes_out;
2424         info->tcpi_backoff = icsk->icsk_backoff;
2425
2426         if (tp->rx_opt.tstamp_ok)
2427                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2428         if (tcp_is_sack(tp))
2429                 info->tcpi_options |= TCPI_OPT_SACK;
2430         if (tp->rx_opt.wscale_ok) {
2431                 info->tcpi_options |= TCPI_OPT_WSCALE;
2432                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2433                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2434         }
2435
2436         if (tp->ecn_flags&TCP_ECN_OK)
2437                 info->tcpi_options |= TCPI_OPT_ECN;
2438
2439         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2440         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2441         info->tcpi_snd_mss = tp->mss_cache;
2442         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2443
2444         if (sk->sk_state == TCP_LISTEN) {
2445                 info->tcpi_unacked = sk->sk_ack_backlog;
2446                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2447         } else {
2448                 info->tcpi_unacked = tp->packets_out;
2449                 info->tcpi_sacked = tp->sacked_out;
2450         }
2451         info->tcpi_lost = tp->lost_out;
2452         info->tcpi_retrans = tp->retrans_out;
2453         info->tcpi_fackets = tp->fackets_out;
2454
2455         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2456         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2457         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2458
2459         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2460         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2461         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2462         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2463         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2464         info->tcpi_snd_cwnd = tp->snd_cwnd;
2465         info->tcpi_advmss = tp->advmss;
2466         info->tcpi_reordering = tp->reordering;
2467
2468         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2469         info->tcpi_rcv_space = tp->rcvq_space.space;
2470
2471         info->tcpi_total_retrans = tp->total_retrans;
2472 }
2473
2474 EXPORT_SYMBOL_GPL(tcp_get_info);
2475
2476 static int do_tcp_getsockopt(struct sock *sk, int level,
2477                 int optname, char __user *optval, int __user *optlen)
2478 {
2479         struct inet_connection_sock *icsk = inet_csk(sk);
2480         struct tcp_sock *tp = tcp_sk(sk);
2481         int val, len;
2482
2483         if (get_user(len, optlen))
2484                 return -EFAULT;
2485
2486         len = min_t(unsigned int, len, sizeof(int));
2487
2488         if (len < 0)
2489                 return -EINVAL;
2490
2491         switch (optname) {
2492         case TCP_MAXSEG:
2493                 val = tp->mss_cache;
2494                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2495                         val = tp->rx_opt.user_mss;
2496                 break;
2497         case TCP_NODELAY:
2498                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2499                 break;
2500         case TCP_CORK:
2501                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2502                 break;
2503         case TCP_KEEPIDLE:
2504                 val = keepalive_time_when(tp) / HZ;
2505                 break;
2506         case TCP_KEEPINTVL:
2507                 val = keepalive_intvl_when(tp) / HZ;
2508                 break;
2509         case TCP_KEEPCNT:
2510                 val = keepalive_probes(tp);
2511                 break;
2512         case TCP_SYNCNT:
2513                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2514                 break;
2515         case TCP_LINGER2:
2516                 val = tp->linger2;
2517                 if (val >= 0)
2518                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2519                 break;
2520         case TCP_DEFER_ACCEPT:
2521                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2522                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2523                 break;
2524         case TCP_WINDOW_CLAMP:
2525                 val = tp->window_clamp;
2526                 break;
2527         case TCP_INFO: {
2528                 struct tcp_info info;
2529
2530                 if (get_user(len, optlen))
2531                         return -EFAULT;
2532
2533                 tcp_get_info(sk, &info);
2534
2535                 len = min_t(unsigned int, len, sizeof(info));
2536                 if (put_user(len, optlen))
2537                         return -EFAULT;
2538                 if (copy_to_user(optval, &info, len))
2539                         return -EFAULT;
2540                 return 0;
2541         }
2542         case TCP_QUICKACK:
2543                 val = !icsk->icsk_ack.pingpong;
2544                 break;
2545
2546         case TCP_CONGESTION:
2547                 if (get_user(len, optlen))
2548                         return -EFAULT;
2549                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2550                 if (put_user(len, optlen))
2551                         return -EFAULT;
2552                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2553                         return -EFAULT;
2554                 return 0;
2555
2556         case TCP_COOKIE_TRANSACTIONS: {
2557                 struct tcp_cookie_transactions ctd;
2558                 struct tcp_cookie_values *cvp = tp->cookie_values;
2559
2560                 if (get_user(len, optlen))
2561                         return -EFAULT;
2562                 if (len < sizeof(ctd))
2563                         return -EINVAL;
2564
2565                 memset(&ctd, 0, sizeof(ctd));
2566                 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2567                                    TCP_COOKIE_IN_ALWAYS : 0)
2568                                 | (tp->rx_opt.cookie_out_never ?
2569                                    TCP_COOKIE_OUT_NEVER : 0);
2570
2571                 if (cvp != NULL) {
2572                         ctd.tcpct_flags |= (cvp->s_data_in ?
2573                                             TCP_S_DATA_IN : 0)
2574                                          | (cvp->s_data_out ?
2575                                             TCP_S_DATA_OUT : 0);
2576
2577                         ctd.tcpct_cookie_desired = cvp->cookie_desired;
2578                         ctd.tcpct_s_data_desired = cvp->s_data_desired;
2579
2580                         memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2581                                cvp->cookie_pair_size);
2582                         ctd.tcpct_used = cvp->cookie_pair_size;
2583                 }
2584
2585                 if (put_user(sizeof(ctd), optlen))
2586                         return -EFAULT;
2587                 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2588                         return -EFAULT;
2589                 return 0;
2590         }
2591         default:
2592                 return -ENOPROTOOPT;
2593         }
2594
2595         if (put_user(len, optlen))
2596                 return -EFAULT;
2597         if (copy_to_user(optval, &val, len))
2598                 return -EFAULT;
2599         return 0;
2600 }
2601
2602 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2603                    int __user *optlen)
2604 {
2605         struct inet_connection_sock *icsk = inet_csk(sk);
2606
2607         if (level != SOL_TCP)
2608                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2609                                                      optval, optlen);
2610         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2611 }
2612
2613 #ifdef CONFIG_COMPAT
2614 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2615                           char __user *optval, int __user *optlen)
2616 {
2617         if (level != SOL_TCP)
2618                 return inet_csk_compat_getsockopt(sk, level, optname,
2619                                                   optval, optlen);
2620         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2621 }
2622
2623 EXPORT_SYMBOL(compat_tcp_getsockopt);
2624 #endif
2625
2626 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2627 {
2628         struct sk_buff *segs = ERR_PTR(-EINVAL);
2629         struct tcphdr *th;
2630         unsigned thlen;
2631         unsigned int seq;
2632         __be32 delta;
2633         unsigned int oldlen;
2634         unsigned int mss;
2635
2636         if (!pskb_may_pull(skb, sizeof(*th)))
2637                 goto out;
2638
2639         th = tcp_hdr(skb);
2640         thlen = th->doff * 4;
2641         if (thlen < sizeof(*th))
2642                 goto out;
2643
2644         if (!pskb_may_pull(skb, thlen))
2645                 goto out;
2646
2647         oldlen = (u16)~skb->len;
2648         __skb_pull(skb, thlen);
2649
2650         mss = skb_shinfo(skb)->gso_size;
2651         if (unlikely(skb->len <= mss))
2652                 goto out;
2653
2654         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2655                 /* Packet is from an untrusted source, reset gso_segs. */
2656                 int type = skb_shinfo(skb)->gso_type;
2657
2658                 if (unlikely(type &
2659                              ~(SKB_GSO_TCPV4 |
2660                                SKB_GSO_DODGY |
2661                                SKB_GSO_TCP_ECN |
2662                                SKB_GSO_TCPV6 |
2663                                0) ||
2664                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2665                         goto out;
2666
2667                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2668
2669                 segs = NULL;
2670                 goto out;
2671         }
2672
2673         segs = skb_segment(skb, features);
2674         if (IS_ERR(segs))
2675                 goto out;
2676
2677         delta = htonl(oldlen + (thlen + mss));
2678
2679         skb = segs;
2680         th = tcp_hdr(skb);
2681         seq = ntohl(th->seq);
2682
2683         do {
2684                 th->fin = th->psh = 0;
2685
2686                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2687                                        (__force u32)delta));
2688                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2689                         th->check =
2690                              csum_fold(csum_partial(skb_transport_header(skb),
2691                                                     thlen, skb->csum));
2692
2693                 seq += mss;
2694                 skb = skb->next;
2695                 th = tcp_hdr(skb);
2696
2697                 th->seq = htonl(seq);
2698                 th->cwr = 0;
2699         } while (skb->next);
2700
2701         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2702                       skb->data_len);
2703         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2704                                 (__force u32)delta));
2705         if (skb->ip_summed != CHECKSUM_PARTIAL)
2706                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2707                                                    thlen, skb->csum));
2708
2709 out:
2710         return segs;
2711 }
2712 EXPORT_SYMBOL(tcp_tso_segment);
2713
2714 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2715 {
2716         struct sk_buff **pp = NULL;
2717         struct sk_buff *p;
2718         struct tcphdr *th;
2719         struct tcphdr *th2;
2720         unsigned int len;
2721         unsigned int thlen;
2722         unsigned int flags;
2723         unsigned int mss = 1;
2724         unsigned int hlen;
2725         unsigned int off;
2726         int flush = 1;
2727         int i;
2728
2729         off = skb_gro_offset(skb);
2730         hlen = off + sizeof(*th);
2731         th = skb_gro_header_fast(skb, off);
2732         if (skb_gro_header_hard(skb, hlen)) {
2733                 th = skb_gro_header_slow(skb, hlen, off);
2734                 if (unlikely(!th))
2735                         goto out;
2736         }
2737
2738         thlen = th->doff * 4;
2739         if (thlen < sizeof(*th))
2740                 goto out;
2741
2742         hlen = off + thlen;
2743         if (skb_gro_header_hard(skb, hlen)) {
2744                 th = skb_gro_header_slow(skb, hlen, off);
2745                 if (unlikely(!th))
2746                         goto out;
2747         }
2748
2749         skb_gro_pull(skb, thlen);
2750
2751         len = skb_gro_len(skb);
2752         flags = tcp_flag_word(th);
2753
2754         for (; (p = *head); head = &p->next) {
2755                 if (!NAPI_GRO_CB(p)->same_flow)
2756                         continue;
2757
2758                 th2 = tcp_hdr(p);
2759
2760                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2761                         NAPI_GRO_CB(p)->same_flow = 0;
2762                         continue;
2763                 }
2764
2765                 goto found;
2766         }
2767
2768         goto out_check_final;
2769
2770 found:
2771         flush = NAPI_GRO_CB(p)->flush;
2772         flush |= flags & TCP_FLAG_CWR;
2773         flush |= (flags ^ tcp_flag_word(th2)) &
2774                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH);
2775         flush |= th->ack_seq ^ th2->ack_seq;
2776         for (i = sizeof(*th); i < thlen; i += 4)
2777                 flush |= *(u32 *)((u8 *)th + i) ^
2778                          *(u32 *)((u8 *)th2 + i);
2779
2780         mss = skb_shinfo(p)->gso_size;
2781
2782         flush |= (len - 1) >= mss;
2783         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2784
2785         if (flush || skb_gro_receive(head, skb)) {
2786                 mss = 1;
2787                 goto out_check_final;
2788         }
2789
2790         p = *head;
2791         th2 = tcp_hdr(p);
2792         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2793
2794 out_check_final:
2795         flush = len < mss;
2796         flush |= flags & (TCP_FLAG_URG | TCP_FLAG_PSH | TCP_FLAG_RST |
2797                           TCP_FLAG_SYN | TCP_FLAG_FIN);
2798
2799         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2800                 pp = head;
2801
2802 out:
2803         NAPI_GRO_CB(skb)->flush |= flush;
2804
2805         return pp;
2806 }
2807 EXPORT_SYMBOL(tcp_gro_receive);
2808
2809 int tcp_gro_complete(struct sk_buff *skb)
2810 {
2811         struct tcphdr *th = tcp_hdr(skb);
2812
2813         skb->csum_start = skb_transport_header(skb) - skb->head;
2814         skb->csum_offset = offsetof(struct tcphdr, check);
2815         skb->ip_summed = CHECKSUM_PARTIAL;
2816
2817         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2818
2819         if (th->cwr)
2820                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2821
2822         return 0;
2823 }
2824 EXPORT_SYMBOL(tcp_gro_complete);
2825
2826 #ifdef CONFIG_TCP_MD5SIG
2827 static unsigned long tcp_md5sig_users;
2828 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool;
2829 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2830
2831 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool)
2832 {
2833         int cpu;
2834         for_each_possible_cpu(cpu) {
2835                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2836                 if (p) {
2837                         if (p->md5_desc.tfm)
2838                                 crypto_free_hash(p->md5_desc.tfm);
2839                         kfree(p);
2840                         p = NULL;
2841                 }
2842         }
2843         free_percpu(pool);
2844 }
2845
2846 void tcp_free_md5sig_pool(void)
2847 {
2848         struct tcp_md5sig_pool * __percpu *pool = NULL;
2849
2850         spin_lock_bh(&tcp_md5sig_pool_lock);
2851         if (--tcp_md5sig_users == 0) {
2852                 pool = tcp_md5sig_pool;
2853                 tcp_md5sig_pool = NULL;
2854         }
2855         spin_unlock_bh(&tcp_md5sig_pool_lock);
2856         if (pool)
2857                 __tcp_free_md5sig_pool(pool);
2858 }
2859
2860 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2861
2862 static struct tcp_md5sig_pool * __percpu *
2863 __tcp_alloc_md5sig_pool(struct sock *sk)
2864 {
2865         int cpu;
2866         struct tcp_md5sig_pool * __percpu *pool;
2867
2868         pool = alloc_percpu(struct tcp_md5sig_pool *);
2869         if (!pool)
2870                 return NULL;
2871
2872         for_each_possible_cpu(cpu) {
2873                 struct tcp_md5sig_pool *p;
2874                 struct crypto_hash *hash;
2875
2876                 p = kzalloc(sizeof(*p), sk->sk_allocation);
2877                 if (!p)
2878                         goto out_free;
2879                 *per_cpu_ptr(pool, cpu) = p;
2880
2881                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2882                 if (!hash || IS_ERR(hash))
2883                         goto out_free;
2884
2885                 p->md5_desc.tfm = hash;
2886         }
2887         return pool;
2888 out_free:
2889         __tcp_free_md5sig_pool(pool);
2890         return NULL;
2891 }
2892
2893 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2894 {
2895         struct tcp_md5sig_pool * __percpu *pool;
2896         int alloc = 0;
2897
2898 retry:
2899         spin_lock_bh(&tcp_md5sig_pool_lock);
2900         pool = tcp_md5sig_pool;
2901         if (tcp_md5sig_users++ == 0) {
2902                 alloc = 1;
2903                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2904         } else if (!pool) {
2905                 tcp_md5sig_users--;
2906                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2907                 cpu_relax();
2908                 goto retry;
2909         } else
2910                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2911
2912         if (alloc) {
2913                 /* we cannot hold spinlock here because this may sleep. */
2914                 struct tcp_md5sig_pool * __percpu *p;
2915
2916                 p = __tcp_alloc_md5sig_pool(sk);
2917                 spin_lock_bh(&tcp_md5sig_pool_lock);
2918                 if (!p) {
2919                         tcp_md5sig_users--;
2920                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2921                         return NULL;
2922                 }
2923                 pool = tcp_md5sig_pool;
2924                 if (pool) {
2925                         /* oops, it has already been assigned. */
2926                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2927                         __tcp_free_md5sig_pool(p);
2928                 } else {
2929                         tcp_md5sig_pool = pool = p;
2930                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2931                 }
2932         }
2933         return pool;
2934 }
2935
2936 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2937
2938 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2939 {
2940         struct tcp_md5sig_pool * __percpu *p;
2941         spin_lock_bh(&tcp_md5sig_pool_lock);
2942         p = tcp_md5sig_pool;
2943         if (p)
2944                 tcp_md5sig_users++;
2945         spin_unlock_bh(&tcp_md5sig_pool_lock);
2946         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2947 }
2948
2949 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2950
2951 void __tcp_put_md5sig_pool(void)
2952 {
2953         tcp_free_md5sig_pool();
2954 }
2955
2956 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2957
2958 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2959                         struct tcphdr *th)
2960 {
2961         struct scatterlist sg;
2962         int err;
2963
2964         __sum16 old_checksum = th->check;
2965         th->check = 0;
2966         /* options aren't included in the hash */
2967         sg_init_one(&sg, th, sizeof(struct tcphdr));
2968         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2969         th->check = old_checksum;
2970         return err;
2971 }
2972
2973 EXPORT_SYMBOL(tcp_md5_hash_header);
2974
2975 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2976                           struct sk_buff *skb, unsigned header_len)
2977 {
2978         struct scatterlist sg;
2979         const struct tcphdr *tp = tcp_hdr(skb);
2980         struct hash_desc *desc = &hp->md5_desc;
2981         unsigned i;
2982         const unsigned head_data_len = skb_headlen(skb) > header_len ?
2983                                        skb_headlen(skb) - header_len : 0;
2984         const struct skb_shared_info *shi = skb_shinfo(skb);
2985
2986         sg_init_table(&sg, 1);
2987
2988         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2989         if (crypto_hash_update(desc, &sg, head_data_len))
2990                 return 1;
2991
2992         for (i = 0; i < shi->nr_frags; ++i) {
2993                 const struct skb_frag_struct *f = &shi->frags[i];
2994                 sg_set_page(&sg, f->page, f->size, f->page_offset);
2995                 if (crypto_hash_update(desc, &sg, f->size))
2996                         return 1;
2997         }
2998
2999         return 0;
3000 }
3001
3002 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3003
3004 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
3005 {
3006         struct scatterlist sg;
3007
3008         sg_init_one(&sg, key->key, key->keylen);
3009         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3010 }
3011
3012 EXPORT_SYMBOL(tcp_md5_hash_key);
3013
3014 #endif
3015
3016 /**
3017  * Each Responder maintains up to two secret values concurrently for
3018  * efficient secret rollover.  Each secret value has 4 states:
3019  *
3020  * Generating.  (tcp_secret_generating != tcp_secret_primary)
3021  *    Generates new Responder-Cookies, but not yet used for primary
3022  *    verification.  This is a short-term state, typically lasting only
3023  *    one round trip time (RTT).
3024  *
3025  * Primary.  (tcp_secret_generating == tcp_secret_primary)
3026  *    Used both for generation and primary verification.
3027  *
3028  * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3029  *    Used for verification, until the first failure that can be
3030  *    verified by the newer Generating secret.  At that time, this
3031  *    cookie's state is changed to Secondary, and the Generating
3032  *    cookie's state is changed to Primary.  This is a short-term state,
3033  *    typically lasting only one round trip time (RTT).
3034  *
3035  * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3036  *    Used for secondary verification, after primary verification
3037  *    failures.  This state lasts no more than twice the Maximum Segment
3038  *    Lifetime (2MSL).  Then, the secret is discarded.
3039  */
3040 struct tcp_cookie_secret {
3041         /* The secret is divided into two parts.  The digest part is the
3042          * equivalent of previously hashing a secret and saving the state,
3043          * and serves as an initialization vector (IV).  The message part
3044          * serves as the trailing secret.
3045          */
3046         u32                             secrets[COOKIE_WORKSPACE_WORDS];
3047         unsigned long                   expires;
3048 };
3049
3050 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3051 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3052 #define TCP_SECRET_LIFE (HZ * 600)
3053
3054 static struct tcp_cookie_secret tcp_secret_one;
3055 static struct tcp_cookie_secret tcp_secret_two;
3056
3057 /* Essentially a circular list, without dynamic allocation. */
3058 static struct tcp_cookie_secret *tcp_secret_generating;
3059 static struct tcp_cookie_secret *tcp_secret_primary;
3060 static struct tcp_cookie_secret *tcp_secret_retiring;
3061 static struct tcp_cookie_secret *tcp_secret_secondary;
3062
3063 static DEFINE_SPINLOCK(tcp_secret_locker);
3064
3065 /* Select a pseudo-random word in the cookie workspace.
3066  */
3067 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3068 {
3069         return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3070 }
3071
3072 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3073  * Called in softirq context.
3074  * Returns: 0 for success.
3075  */
3076 int tcp_cookie_generator(u32 *bakery)
3077 {
3078         unsigned long jiffy = jiffies;
3079
3080         if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3081                 spin_lock_bh(&tcp_secret_locker);
3082                 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3083                         /* refreshed by another */
3084                         memcpy(bakery,
3085                                &tcp_secret_generating->secrets[0],
3086                                COOKIE_WORKSPACE_WORDS);
3087                 } else {
3088                         /* still needs refreshing */
3089                         get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3090
3091                         /* The first time, paranoia assumes that the
3092                          * randomization function isn't as strong.  But,
3093                          * this secret initialization is delayed until
3094                          * the last possible moment (packet arrival).
3095                          * Although that time is observable, it is
3096                          * unpredictably variable.  Mash in the most
3097                          * volatile clock bits available, and expire the
3098                          * secret extra quickly.
3099                          */
3100                         if (unlikely(tcp_secret_primary->expires ==
3101                                      tcp_secret_secondary->expires)) {
3102                                 struct timespec tv;
3103
3104                                 getnstimeofday(&tv);
3105                                 bakery[COOKIE_DIGEST_WORDS+0] ^=
3106                                         (u32)tv.tv_nsec;
3107
3108                                 tcp_secret_secondary->expires = jiffy
3109                                         + TCP_SECRET_1MSL
3110                                         + (0x0f & tcp_cookie_work(bakery, 0));
3111                         } else {
3112                                 tcp_secret_secondary->expires = jiffy
3113                                         + TCP_SECRET_LIFE
3114                                         + (0xff & tcp_cookie_work(bakery, 1));
3115                                 tcp_secret_primary->expires = jiffy
3116                                         + TCP_SECRET_2MSL
3117                                         + (0x1f & tcp_cookie_work(bakery, 2));
3118                         }
3119                         memcpy(&tcp_secret_secondary->secrets[0],
3120                                bakery, COOKIE_WORKSPACE_WORDS);
3121
3122                         rcu_assign_pointer(tcp_secret_generating,
3123                                            tcp_secret_secondary);
3124                         rcu_assign_pointer(tcp_secret_retiring,
3125                                            tcp_secret_primary);
3126                         /*
3127                          * Neither call_rcu() nor synchronize_rcu() needed.
3128                          * Retiring data is not freed.  It is replaced after
3129                          * further (locked) pointer updates, and a quiet time
3130                          * (minimum 1MSL, maximum LIFE - 2MSL).
3131                          */
3132                 }
3133                 spin_unlock_bh(&tcp_secret_locker);
3134         } else {
3135                 rcu_read_lock_bh();
3136                 memcpy(bakery,
3137                        &rcu_dereference(tcp_secret_generating)->secrets[0],
3138                        COOKIE_WORKSPACE_WORDS);
3139                 rcu_read_unlock_bh();
3140         }
3141         return 0;
3142 }
3143 EXPORT_SYMBOL(tcp_cookie_generator);
3144
3145 void tcp_done(struct sock *sk)
3146 {
3147         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3148                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3149
3150         tcp_set_state(sk, TCP_CLOSE);
3151         tcp_clear_xmit_timers(sk);
3152
3153         sk->sk_shutdown = SHUTDOWN_MASK;
3154
3155         if (!sock_flag(sk, SOCK_DEAD))
3156                 sk->sk_state_change(sk);
3157         else
3158                 inet_csk_destroy_sock(sk);
3159 }
3160 EXPORT_SYMBOL_GPL(tcp_done);
3161
3162 extern struct tcp_congestion_ops tcp_reno;
3163
3164 static __initdata unsigned long thash_entries;
3165 static int __init set_thash_entries(char *str)
3166 {
3167         if (!str)
3168                 return 0;
3169         thash_entries = simple_strtoul(str, &str, 0);
3170         return 1;
3171 }
3172 __setup("thash_entries=", set_thash_entries);
3173
3174 void __init tcp_init(void)
3175 {
3176         struct sk_buff *skb = NULL;
3177         unsigned long nr_pages, limit;
3178         int order, i, max_share;
3179         unsigned long jiffy = jiffies;
3180
3181         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3182
3183         percpu_counter_init(&tcp_sockets_allocated, 0);
3184         percpu_counter_init(&tcp_orphan_count, 0);
3185         tcp_hashinfo.bind_bucket_cachep =
3186                 kmem_cache_create("tcp_bind_bucket",
3187                                   sizeof(struct inet_bind_bucket), 0,
3188                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3189
3190         /* Size and allocate the main established and bind bucket
3191          * hash tables.
3192          *
3193          * The methodology is similar to that of the buffer cache.
3194          */
3195         tcp_hashinfo.ehash =
3196                 alloc_large_system_hash("TCP established",
3197                                         sizeof(struct inet_ehash_bucket),
3198                                         thash_entries,
3199                                         (totalram_pages >= 128 * 1024) ?
3200                                         13 : 15,
3201                                         0,
3202                                         NULL,
3203                                         &tcp_hashinfo.ehash_mask,
3204                                         thash_entries ? 0 : 512 * 1024);
3205         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3206                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3207                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3208         }
3209         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3210                 panic("TCP: failed to alloc ehash_locks");
3211         tcp_hashinfo.bhash =
3212                 alloc_large_system_hash("TCP bind",
3213                                         sizeof(struct inet_bind_hashbucket),
3214                                         tcp_hashinfo.ehash_mask + 1,
3215                                         (totalram_pages >= 128 * 1024) ?
3216                                         13 : 15,
3217                                         0,
3218                                         &tcp_hashinfo.bhash_size,
3219                                         NULL,
3220                                         64 * 1024);
3221         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3222         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3223                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3224                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3225         }
3226
3227         /* Try to be a bit smarter and adjust defaults depending
3228          * on available memory.
3229          */
3230         for (order = 0; ((1 << order) << PAGE_SHIFT) <
3231                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
3232                         order++)
3233                 ;
3234         if (order >= 4) {
3235                 tcp_death_row.sysctl_max_tw_buckets = 180000;
3236                 sysctl_tcp_max_orphans = 4096 << (order - 4);
3237                 sysctl_max_syn_backlog = 1024;
3238         } else if (order < 3) {
3239                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
3240                 sysctl_tcp_max_orphans >>= (3 - order);
3241                 sysctl_max_syn_backlog = 128;
3242         }
3243
3244         /* Set the pressure threshold to be a fraction of global memory that
3245          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
3246          * memory, with a floor of 128 pages.
3247          */
3248         nr_pages = totalram_pages - totalhigh_pages;
3249         limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
3250         limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
3251         limit = max(limit, 128UL);
3252         sysctl_tcp_mem[0] = limit / 4 * 3;
3253         sysctl_tcp_mem[1] = limit;
3254         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3255
3256         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3257         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3258         max_share = min(4UL*1024*1024, limit);
3259
3260         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3261         sysctl_tcp_wmem[1] = 16*1024;
3262         sysctl_tcp_wmem[2] = max(64*1024, max_share);
3263
3264         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3265         sysctl_tcp_rmem[1] = 87380;
3266         sysctl_tcp_rmem[2] = max(87380, max_share);
3267
3268         printk(KERN_INFO "TCP: Hash tables configured "
3269                "(established %u bind %u)\n",
3270                tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3271
3272         tcp_register_congestion_control(&tcp_reno);
3273
3274         memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3275         memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3276         tcp_secret_one.expires = jiffy; /* past due */
3277         tcp_secret_two.expires = jiffy; /* past due */
3278         tcp_secret_generating = &tcp_secret_one;
3279         tcp_secret_primary = &tcp_secret_one;
3280         tcp_secret_retiring = &tcp_secret_two;
3281         tcp_secret_secondary = &tcp_secret_two;
3282 }
3283
3284 EXPORT_SYMBOL(tcp_close);
3285 EXPORT_SYMBOL(tcp_disconnect);
3286 EXPORT_SYMBOL(tcp_getsockopt);
3287 EXPORT_SYMBOL(tcp_ioctl);
3288 EXPORT_SYMBOL(tcp_poll);
3289 EXPORT_SYMBOL(tcp_read_sock);
3290 EXPORT_SYMBOL(tcp_recvmsg);
3291 EXPORT_SYMBOL(tcp_sendmsg);
3292 EXPORT_SYMBOL(tcp_splice_read);
3293 EXPORT_SYMBOL(tcp_sendpage);
3294 EXPORT_SYMBOL(tcp_setsockopt);
3295 EXPORT_SYMBOL(tcp_shutdown);