misc: uidstat: Adding uid stat driver to collect network statistics.
[linux-2.6.git] / net / ipv4 / tcp.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Authors:     Ross Biro
9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
12  *              Florian La Roche, <flla@stud.uni-sb.de>
13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *              Jorge Cwik, <jorge@laser.satlink.net>
19  *
20  * Fixes:
21  *              Alan Cox        :       Numerous verify_area() calls
22  *              Alan Cox        :       Set the ACK bit on a reset
23  *              Alan Cox        :       Stopped it crashing if it closed while
24  *                                      sk->inuse=1 and was trying to connect
25  *                                      (tcp_err()).
26  *              Alan Cox        :       All icmp error handling was broken
27  *                                      pointers passed where wrong and the
28  *                                      socket was looked up backwards. Nobody
29  *                                      tested any icmp error code obviously.
30  *              Alan Cox        :       tcp_err() now handled properly. It
31  *                                      wakes people on errors. poll
32  *                                      behaves and the icmp error race
33  *                                      has gone by moving it into sock.c
34  *              Alan Cox        :       tcp_send_reset() fixed to work for
35  *                                      everything not just packets for
36  *                                      unknown sockets.
37  *              Alan Cox        :       tcp option processing.
38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
39  *                                      syn rule wrong]
40  *              Herp Rosmanith  :       More reset fixes
41  *              Alan Cox        :       No longer acks invalid rst frames.
42  *                                      Acking any kind of RST is right out.
43  *              Alan Cox        :       Sets an ignore me flag on an rst
44  *                                      receive otherwise odd bits of prattle
45  *                                      escape still
46  *              Alan Cox        :       Fixed another acking RST frame bug.
47  *                                      Should stop LAN workplace lockups.
48  *              Alan Cox        :       Some tidyups using the new skb list
49  *                                      facilities
50  *              Alan Cox        :       sk->keepopen now seems to work
51  *              Alan Cox        :       Pulls options out correctly on accepts
52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
54  *                                      bit to skb ops.
55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
56  *                                      nasty.
57  *              Alan Cox        :       Added some better commenting, as the
58  *                                      tcp is hard to follow
59  *              Alan Cox        :       Removed incorrect check for 20 * psh
60  *      Michael O'Reilly        :       ack < copied bug fix.
61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
62  *              Alan Cox        :       FIN with no memory -> CRASH
63  *              Alan Cox        :       Added socket option proto entries.
64  *                                      Also added awareness of them to accept.
65  *              Alan Cox        :       Added TCP options (SOL_TCP)
66  *              Alan Cox        :       Switched wakeup calls to callbacks,
67  *                                      so the kernel can layer network
68  *                                      sockets.
69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
71  *              Alan Cox        :       RST frames sent on unsynchronised
72  *                                      state ack error.
73  *              Alan Cox        :       Put in missing check for SYN bit.
74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
75  *                                      window non shrink trick.
76  *              Alan Cox        :       Added a couple of small NET2E timer
77  *                                      fixes
78  *              Charles Hedrick :       TCP fixes
79  *              Toomas Tamm     :       TCP window fixes
80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
81  *              Charles Hedrick :       Rewrote most of it to actually work
82  *              Linus           :       Rewrote tcp_read() and URG handling
83  *                                      completely
84  *              Gerhard Koerting:       Fixed some missing timer handling
85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
86  *              Gerhard Koerting:       PC/TCP workarounds
87  *              Adam Caldwell   :       Assorted timer/timing errors
88  *              Matthew Dillon  :       Fixed another RST bug
89  *              Alan Cox        :       Move to kernel side addressing changes.
90  *              Alan Cox        :       Beginning work on TCP fastpathing
91  *                                      (not yet usable)
92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
93  *              Alan Cox        :       TCP fast path debugging
94  *              Alan Cox        :       Window clamping
95  *              Michael Riepe   :       Bug in tcp_check()
96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
97  *              Matt Dillon     :       Yet more small nasties remove from the
98  *                                      TCP code (Be very nice to this man if
99  *                                      tcp finally works 100%) 8)
100  *              Alan Cox        :       BSD accept semantics.
101  *              Alan Cox        :       Reset on closedown bug.
102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
103  *              Michael Pall    :       Handle poll() after URG properly in
104  *                                      all cases.
105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
106  *                                      (multi URG PUSH broke rlogin).
107  *              Michael Pall    :       Fix the multi URG PUSH problem in
108  *                                      tcp_readable(), poll() after URG
109  *                                      works now.
110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
111  *                                      BSD api.
112  *              Alan Cox        :       Changed the semantics of sk->socket to
113  *                                      fix a race and a signal problem with
114  *                                      accept() and async I/O.
115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
118  *                                      clients/servers which listen in on
119  *                                      fixed ports.
120  *              Alan Cox        :       Cleaned the above up and shrank it to
121  *                                      a sensible code size.
122  *              Alan Cox        :       Self connect lockup fix.
123  *              Alan Cox        :       No connect to multicast.
124  *              Ross Biro       :       Close unaccepted children on master
125  *                                      socket close.
126  *              Alan Cox        :       Reset tracing code.
127  *              Alan Cox        :       Spurious resets on shutdown.
128  *              Alan Cox        :       Giant 15 minute/60 second timer error
129  *              Alan Cox        :       Small whoops in polling before an
130  *                                      accept.
131  *              Alan Cox        :       Kept the state trace facility since
132  *                                      it's handy for debugging.
133  *              Alan Cox        :       More reset handler fixes.
134  *              Alan Cox        :       Started rewriting the code based on
135  *                                      the RFC's for other useful protocol
136  *                                      references see: Comer, KA9Q NOS, and
137  *                                      for a reference on the difference
138  *                                      between specifications and how BSD
139  *                                      works see the 4.4lite source.
140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
141  *                                      close.
142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
144  *              Alan Cox        :       Reimplemented timers as per the RFC
145  *                                      and using multiple timers for sanity.
146  *              Alan Cox        :       Small bug fixes, and a lot of new
147  *                                      comments.
148  *              Alan Cox        :       Fixed dual reader crash by locking
149  *                                      the buffers (much like datagram.c)
150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
151  *                                      now gets fed up of retrying without
152  *                                      (even a no space) answer.
153  *              Alan Cox        :       Extracted closing code better
154  *              Alan Cox        :       Fixed the closing state machine to
155  *                                      resemble the RFC.
156  *              Alan Cox        :       More 'per spec' fixes.
157  *              Jorge Cwik      :       Even faster checksumming.
158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
159  *                                      only frames. At least one pc tcp stack
160  *                                      generates them.
161  *              Alan Cox        :       Cache last socket.
162  *              Alan Cox        :       Per route irtt.
163  *              Matt Day        :       poll()->select() match BSD precisely on error
164  *              Alan Cox        :       New buffers
165  *              Marc Tamsky     :       Various sk->prot->retransmits and
166  *                                      sk->retransmits misupdating fixed.
167  *                                      Fixed tcp_write_timeout: stuck close,
168  *                                      and TCP syn retries gets used now.
169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
170  *                                      ack if state is TCP_CLOSED.
171  *              Alan Cox        :       Look up device on a retransmit - routes may
172  *                                      change. Doesn't yet cope with MSS shrink right
173  *                                      but it's a start!
174  *              Marc Tamsky     :       Closing in closing fixes.
175  *              Mike Shaver     :       RFC1122 verifications.
176  *              Alan Cox        :       rcv_saddr errors.
177  *              Alan Cox        :       Block double connect().
178  *              Alan Cox        :       Small hooks for enSKIP.
179  *              Alexey Kuznetsov:       Path MTU discovery.
180  *              Alan Cox        :       Support soft errors.
181  *              Alan Cox        :       Fix MTU discovery pathological case
182  *                                      when the remote claims no mtu!
183  *              Marc Tamsky     :       TCP_CLOSE fix.
184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
185  *                                      window but wrong (fixes NT lpd problems)
186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
187  *              Joerg Reuter    :       No modification of locked buffers in
188  *                                      tcp_do_retransmit()
189  *              Eric Schenk     :       Changed receiver side silly window
190  *                                      avoidance algorithm to BSD style
191  *                                      algorithm. This doubles throughput
192  *                                      against machines running Solaris,
193  *                                      and seems to result in general
194  *                                      improvement.
195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
196  *      Willy Konynenberg       :       Transparent proxying support.
197  *      Mike McLagan            :       Routing by source
198  *              Keith Owens     :       Do proper merging with partial SKB's in
199  *                                      tcp_do_sendmsg to avoid burstiness.
200  *              Eric Schenk     :       Fix fast close down bug with
201  *                                      shutdown() followed by close().
202  *              Andi Kleen      :       Make poll agree with SIGIO
203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
204  *                                      lingertime == 0 (RFC 793 ABORT Call)
205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
206  *                                      csum_and_copy_from_user() if possible.
207  *
208  *              This program is free software; you can redistribute it and/or
209  *              modify it under the terms of the GNU General Public License
210  *              as published by the Free Software Foundation; either version
211  *              2 of the License, or(at your option) any later version.
212  *
213  * Description of States:
214  *
215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
216  *
217  *      TCP_SYN_RECV            received a connection request, sent ack,
218  *                              waiting for final ack in three-way handshake.
219  *
220  *      TCP_ESTABLISHED         connection established
221  *
222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
223  *                              transmission of remaining buffered data
224  *
225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
226  *                              to shutdown
227  *
228  *      TCP_CLOSING             both sides have shutdown but we still have
229  *                              data we have to finish sending
230  *
231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
232  *                              closed, can only be entered from FIN_WAIT2
233  *                              or CLOSING.  Required because the other end
234  *                              may not have gotten our last ACK causing it
235  *                              to retransmit the data packet (which we ignore)
236  *
237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
238  *                              us to finish writing our data and to shutdown
239  *                              (we have to close() to move on to LAST_ACK)
240  *
241  *      TCP_LAST_ACK            out side has shutdown after remote has
242  *                              shutdown.  There may still be data in our
243  *                              buffer that we have to finish sending
244  *
245  *      TCP_CLOSE               socket is finished
246  */
247
248 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
267 #include <linux/time.h>
268 #include <linux/slab.h>
269 #include <linux/uid_stat.h>
270
271 #include <net/icmp.h>
272 #include <net/tcp.h>
273 #include <net/xfrm.h>
274 #include <net/ip.h>
275 #include <net/netdma.h>
276 #include <net/sock.h>
277
278 #include <asm/uaccess.h>
279 #include <asm/ioctls.h>
280
281 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
282
283 struct percpu_counter tcp_orphan_count;
284 EXPORT_SYMBOL_GPL(tcp_orphan_count);
285
286 long sysctl_tcp_mem[3] __read_mostly;
287 int sysctl_tcp_wmem[3] __read_mostly;
288 int sysctl_tcp_rmem[3] __read_mostly;
289
290 EXPORT_SYMBOL(sysctl_tcp_mem);
291 EXPORT_SYMBOL(sysctl_tcp_rmem);
292 EXPORT_SYMBOL(sysctl_tcp_wmem);
293
294 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
295 EXPORT_SYMBOL(tcp_memory_allocated);
296
297 /*
298  * Current number of TCP sockets.
299  */
300 struct percpu_counter tcp_sockets_allocated;
301 EXPORT_SYMBOL(tcp_sockets_allocated);
302
303 /*
304  * TCP splice context
305  */
306 struct tcp_splice_state {
307         struct pipe_inode_info *pipe;
308         size_t len;
309         unsigned int flags;
310 };
311
312 /*
313  * Pressure flag: try to collapse.
314  * Technical note: it is used by multiple contexts non atomically.
315  * All the __sk_mem_schedule() is of this nature: accounting
316  * is strict, actions are advisory and have some latency.
317  */
318 int tcp_memory_pressure __read_mostly;
319 EXPORT_SYMBOL(tcp_memory_pressure);
320
321 void tcp_enter_memory_pressure(struct sock *sk)
322 {
323         if (!tcp_memory_pressure) {
324                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
325                 tcp_memory_pressure = 1;
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_sleep(sk), 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
391         /*
392          * POLLHUP is certainly not done right. But poll() doesn't
393          * have a notion of HUP in just one direction, and for a
394          * socket the read side is more interesting.
395          *
396          * Some poll() documentation says that POLLHUP is incompatible
397          * with the POLLOUT/POLLWR flags, so somebody should check this
398          * all. But careful, it tends to be safer to return too many
399          * bits than too few, and you can easily break real applications
400          * if you don't tell them that something has hung up!
401          *
402          * Check-me.
403          *
404          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
405          * our fs/select.c). It means that after we received EOF,
406          * poll always returns immediately, making impossible poll() on write()
407          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
408          * if and only if shutdown has been made in both directions.
409          * Actually, it is interesting to look how Solaris and DUX
410          * solve this dilemma. I would prefer, if POLLHUP were maskable,
411          * then we could set it on SND_SHUTDOWN. BTW examples given
412          * in Stevens' books assume exactly this behaviour, it explains
413          * why POLLHUP is incompatible with POLLOUT.    --ANK
414          *
415          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
416          * blocking on fresh not-connected or disconnected socket. --ANK
417          */
418         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
419                 mask |= POLLHUP;
420         if (sk->sk_shutdown & RCV_SHUTDOWN)
421                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
422
423         /* Connected? */
424         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
425                 int target = sock_rcvlowat(sk, 0, INT_MAX);
426
427                 if (tp->urg_seq == tp->copied_seq &&
428                     !sock_flag(sk, SOCK_URGINLINE) &&
429                     tp->urg_data)
430                         target++;
431
432                 /* Potential race condition. If read of tp below will
433                  * escape above sk->sk_state, we can be illegally awaken
434                  * in SYN_* states. */
435                 if (tp->rcv_nxt - tp->copied_seq >= target)
436                         mask |= POLLIN | POLLRDNORM;
437
438                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
439                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
440                                 mask |= POLLOUT | POLLWRNORM;
441                         } else {  /* send SIGIO later */
442                                 set_bit(SOCK_ASYNC_NOSPACE,
443                                         &sk->sk_socket->flags);
444                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
445
446                                 /* Race breaker. If space is freed after
447                                  * wspace test but before the flags are set,
448                                  * IO signal will be lost.
449                                  */
450                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
451                                         mask |= POLLOUT | POLLWRNORM;
452                         }
453                 } else
454                         mask |= POLLOUT | POLLWRNORM;
455
456                 if (tp->urg_data & TCP_URG_VALID)
457                         mask |= POLLPRI;
458         }
459         /* This barrier is coupled with smp_wmb() in tcp_reset() */
460         smp_rmb();
461         if (sk->sk_err)
462                 mask |= POLLERR;
463
464         return mask;
465 }
466 EXPORT_SYMBOL(tcp_poll);
467
468 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
469 {
470         struct tcp_sock *tp = tcp_sk(sk);
471         int answ;
472
473         switch (cmd) {
474         case SIOCINQ:
475                 if (sk->sk_state == TCP_LISTEN)
476                         return -EINVAL;
477
478                 lock_sock(sk);
479                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
480                         answ = 0;
481                 else if (sock_flag(sk, SOCK_URGINLINE) ||
482                          !tp->urg_data ||
483                          before(tp->urg_seq, tp->copied_seq) ||
484                          !before(tp->urg_seq, tp->rcv_nxt)) {
485                         struct sk_buff *skb;
486
487                         answ = tp->rcv_nxt - tp->copied_seq;
488
489                         /* Subtract 1, if FIN is in queue. */
490                         skb = skb_peek_tail(&sk->sk_receive_queue);
491                         if (answ && skb)
492                                 answ -= tcp_hdr(skb)->fin;
493                 } else
494                         answ = tp->urg_seq - tp->copied_seq;
495                 release_sock(sk);
496                 break;
497         case SIOCATMARK:
498                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
499                 break;
500         case SIOCOUTQ:
501                 if (sk->sk_state == TCP_LISTEN)
502                         return -EINVAL;
503
504                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
505                         answ = 0;
506                 else
507                         answ = tp->write_seq - tp->snd_una;
508                 break;
509         case SIOCOUTQNSD:
510                 if (sk->sk_state == TCP_LISTEN)
511                         return -EINVAL;
512
513                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
514                         answ = 0;
515                 else
516                         answ = tp->write_seq - tp->snd_nxt;
517                 break;
518         default:
519                 return -ENOIOCTLCMD;
520         }
521
522         return put_user(answ, (int __user *)arg);
523 }
524 EXPORT_SYMBOL(tcp_ioctl);
525
526 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
527 {
528         TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
529         tp->pushed_seq = tp->write_seq;
530 }
531
532 static inline int forced_push(struct tcp_sock *tp)
533 {
534         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
535 }
536
537 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
538 {
539         struct tcp_sock *tp = tcp_sk(sk);
540         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
541
542         skb->csum    = 0;
543         tcb->seq     = tcb->end_seq = tp->write_seq;
544         tcb->flags   = TCPHDR_ACK;
545         tcb->sacked  = 0;
546         skb_header_release(skb);
547         tcp_add_write_queue_tail(sk, skb);
548         sk->sk_wmem_queued += skb->truesize;
549         sk_mem_charge(sk, skb->truesize);
550         if (tp->nonagle & TCP_NAGLE_PUSH)
551                 tp->nonagle &= ~TCP_NAGLE_PUSH;
552 }
553
554 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
555 {
556         if (flags & MSG_OOB)
557                 tp->snd_up = tp->write_seq;
558 }
559
560 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
561                             int nonagle)
562 {
563         if (tcp_send_head(sk)) {
564                 struct tcp_sock *tp = tcp_sk(sk);
565
566                 if (!(flags & MSG_MORE) || forced_push(tp))
567                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
568
569                 tcp_mark_urg(tp, flags);
570                 __tcp_push_pending_frames(sk, mss_now,
571                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
572         }
573 }
574
575 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
576                                 unsigned int offset, size_t len)
577 {
578         struct tcp_splice_state *tss = rd_desc->arg.data;
579         int ret;
580
581         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
582                               tss->flags);
583         if (ret > 0)
584                 rd_desc->count -= ret;
585         return ret;
586 }
587
588 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
589 {
590         /* Store TCP splice context information in read_descriptor_t. */
591         read_descriptor_t rd_desc = {
592                 .arg.data = tss,
593                 .count    = tss->len,
594         };
595
596         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
597 }
598
599 /**
600  *  tcp_splice_read - splice data from TCP socket to a pipe
601  * @sock:       socket to splice from
602  * @ppos:       position (not valid)
603  * @pipe:       pipe to splice to
604  * @len:        number of bytes to splice
605  * @flags:      splice modifier flags
606  *
607  * Description:
608  *    Will read pages from given socket and fill them into a pipe.
609  *
610  **/
611 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
612                         struct pipe_inode_info *pipe, size_t len,
613                         unsigned int flags)
614 {
615         struct sock *sk = sock->sk;
616         struct tcp_splice_state tss = {
617                 .pipe = pipe,
618                 .len = len,
619                 .flags = flags,
620         };
621         long timeo;
622         ssize_t spliced;
623         int ret;
624
625         sock_rps_record_flow(sk);
626         /*
627          * We can't seek on a socket input
628          */
629         if (unlikely(*ppos))
630                 return -ESPIPE;
631
632         ret = spliced = 0;
633
634         lock_sock(sk);
635
636         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
637         while (tss.len) {
638                 ret = __tcp_splice_read(sk, &tss);
639                 if (ret < 0)
640                         break;
641                 else if (!ret) {
642                         if (spliced)
643                                 break;
644                         if (sock_flag(sk, SOCK_DONE))
645                                 break;
646                         if (sk->sk_err) {
647                                 ret = sock_error(sk);
648                                 break;
649                         }
650                         if (sk->sk_shutdown & RCV_SHUTDOWN)
651                                 break;
652                         if (sk->sk_state == TCP_CLOSE) {
653                                 /*
654                                  * This occurs when user tries to read
655                                  * from never connected socket.
656                                  */
657                                 if (!sock_flag(sk, SOCK_DONE))
658                                         ret = -ENOTCONN;
659                                 break;
660                         }
661                         if (!timeo) {
662                                 ret = -EAGAIN;
663                                 break;
664                         }
665                         sk_wait_data(sk, &timeo);
666                         if (signal_pending(current)) {
667                                 ret = sock_intr_errno(timeo);
668                                 break;
669                         }
670                         continue;
671                 }
672                 tss.len -= ret;
673                 spliced += ret;
674
675                 if (!timeo)
676                         break;
677                 release_sock(sk);
678                 lock_sock(sk);
679
680                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
681                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
682                     signal_pending(current))
683                         break;
684         }
685
686         release_sock(sk);
687
688         if (spliced)
689                 return spliced;
690
691         return ret;
692 }
693 EXPORT_SYMBOL(tcp_splice_read);
694
695 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
696 {
697         struct sk_buff *skb;
698
699         /* The TCP header must be at least 32-bit aligned.  */
700         size = ALIGN(size, 4);
701
702         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
703         if (skb) {
704                 if (sk_wmem_schedule(sk, skb->truesize)) {
705                         /*
706                          * Make sure that we have exactly size bytes
707                          * available to the caller, no more, no less.
708                          */
709                         skb_reserve(skb, skb_tailroom(skb) - size);
710                         return skb;
711                 }
712                 __kfree_skb(skb);
713         } else {
714                 sk->sk_prot->enter_memory_pressure(sk);
715                 sk_stream_moderate_sndbuf(sk);
716         }
717         return NULL;
718 }
719
720 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
721                                        int large_allowed)
722 {
723         struct tcp_sock *tp = tcp_sk(sk);
724         u32 xmit_size_goal, old_size_goal;
725
726         xmit_size_goal = mss_now;
727
728         if (large_allowed && sk_can_gso(sk)) {
729                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
730                                   inet_csk(sk)->icsk_af_ops->net_header_len -
731                                   inet_csk(sk)->icsk_ext_hdr_len -
732                                   tp->tcp_header_len);
733
734                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
735
736                 /* We try hard to avoid divides here */
737                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
738
739                 if (likely(old_size_goal <= xmit_size_goal &&
740                            old_size_goal + mss_now > xmit_size_goal)) {
741                         xmit_size_goal = old_size_goal;
742                 } else {
743                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
744                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
745                 }
746         }
747
748         return max(xmit_size_goal, mss_now);
749 }
750
751 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
752 {
753         int mss_now;
754
755         mss_now = tcp_current_mss(sk);
756         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
757
758         return mss_now;
759 }
760
761 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
762                          size_t psize, int flags)
763 {
764         struct tcp_sock *tp = tcp_sk(sk);
765         int mss_now, size_goal;
766         int err;
767         ssize_t copied;
768         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
769
770         /* Wait for a connection to finish. */
771         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
772                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
773                         goto out_err;
774
775         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
776
777         mss_now = tcp_send_mss(sk, &size_goal, flags);
778         copied = 0;
779
780         err = -EPIPE;
781         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
782                 goto out_err;
783
784         while (psize > 0) {
785                 struct sk_buff *skb = tcp_write_queue_tail(sk);
786                 struct page *page = pages[poffset / PAGE_SIZE];
787                 int copy, i, can_coalesce;
788                 int offset = poffset % PAGE_SIZE;
789                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
790
791                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
792 new_segment:
793                         if (!sk_stream_memory_free(sk))
794                                 goto wait_for_sndbuf;
795
796                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
797                         if (!skb)
798                                 goto wait_for_memory;
799
800                         skb_entail(sk, skb);
801                         copy = size_goal;
802                 }
803
804                 if (copy > size)
805                         copy = size;
806
807                 i = skb_shinfo(skb)->nr_frags;
808                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
809                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
810                         tcp_mark_push(tp, skb);
811                         goto new_segment;
812                 }
813                 if (!sk_wmem_schedule(sk, copy))
814                         goto wait_for_memory;
815
816                 if (can_coalesce) {
817                         skb_shinfo(skb)->frags[i - 1].size += copy;
818                 } else {
819                         get_page(page);
820                         skb_fill_page_desc(skb, i, page, offset, copy);
821                 }
822
823                 skb->len += copy;
824                 skb->data_len += copy;
825                 skb->truesize += copy;
826                 sk->sk_wmem_queued += copy;
827                 sk_mem_charge(sk, copy);
828                 skb->ip_summed = CHECKSUM_PARTIAL;
829                 tp->write_seq += copy;
830                 TCP_SKB_CB(skb)->end_seq += copy;
831                 skb_shinfo(skb)->gso_segs = 0;
832
833                 if (!copied)
834                         TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
835
836                 copied += copy;
837                 poffset += copy;
838                 if (!(psize -= copy))
839                         goto out;
840
841                 if (skb->len < size_goal || (flags & MSG_OOB))
842                         continue;
843
844                 if (forced_push(tp)) {
845                         tcp_mark_push(tp, skb);
846                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
847                 } else if (skb == tcp_send_head(sk))
848                         tcp_push_one(sk, mss_now);
849                 continue;
850
851 wait_for_sndbuf:
852                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
853 wait_for_memory:
854                 if (copied)
855                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
856
857                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
858                         goto do_error;
859
860                 mss_now = tcp_send_mss(sk, &size_goal, flags);
861         }
862
863 out:
864         if (copied)
865                 tcp_push(sk, flags, mss_now, tp->nonagle);
866         return copied;
867
868 do_error:
869         if (copied)
870                 goto out;
871 out_err:
872         return sk_stream_error(sk, flags, err);
873 }
874
875 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
876                  size_t size, int flags)
877 {
878         ssize_t res;
879
880         if (!(sk->sk_route_caps & NETIF_F_SG) ||
881             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
882                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
883                                         flags);
884
885         lock_sock(sk);
886         res = do_tcp_sendpages(sk, &page, offset, size, flags);
887         release_sock(sk);
888         return res;
889 }
890 EXPORT_SYMBOL(tcp_sendpage);
891
892 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
893 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
894
895 static inline int select_size(struct sock *sk, int sg)
896 {
897         struct tcp_sock *tp = tcp_sk(sk);
898         int tmp = tp->mss_cache;
899
900         if (sg) {
901                 if (sk_can_gso(sk))
902                         tmp = 0;
903                 else {
904                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
905
906                         if (tmp >= pgbreak &&
907                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
908                                 tmp = pgbreak;
909                 }
910         }
911
912         return tmp;
913 }
914
915 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
916                 size_t size)
917 {
918         struct iovec *iov;
919         struct tcp_sock *tp = tcp_sk(sk);
920         struct sk_buff *skb;
921         int iovlen, flags;
922         int mss_now, size_goal;
923         int sg, err, copied;
924         long timeo;
925
926         lock_sock(sk);
927
928         flags = msg->msg_flags;
929         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
930
931         /* Wait for a connection to finish. */
932         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
933                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
934                         goto out_err;
935
936         /* This should be in poll */
937         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
938
939         mss_now = tcp_send_mss(sk, &size_goal, flags);
940
941         /* Ok commence sending. */
942         iovlen = msg->msg_iovlen;
943         iov = msg->msg_iov;
944         copied = 0;
945
946         err = -EPIPE;
947         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
948                 goto out_err;
949
950         sg = sk->sk_route_caps & NETIF_F_SG;
951
952         while (--iovlen >= 0) {
953                 size_t seglen = iov->iov_len;
954                 unsigned char __user *from = iov->iov_base;
955
956                 iov++;
957
958                 while (seglen > 0) {
959                         int copy = 0;
960                         int max = size_goal;
961
962                         skb = tcp_write_queue_tail(sk);
963                         if (tcp_send_head(sk)) {
964                                 if (skb->ip_summed == CHECKSUM_NONE)
965                                         max = mss_now;
966                                 copy = max - skb->len;
967                         }
968
969                         if (copy <= 0) {
970 new_segment:
971                                 /* Allocate new segment. If the interface is SG,
972                                  * allocate skb fitting to single page.
973                                  */
974                                 if (!sk_stream_memory_free(sk))
975                                         goto wait_for_sndbuf;
976
977                                 skb = sk_stream_alloc_skb(sk,
978                                                           select_size(sk, sg),
979                                                           sk->sk_allocation);
980                                 if (!skb)
981                                         goto wait_for_memory;
982
983                                 /*
984                                  * Check whether we can use HW checksum.
985                                  */
986                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
987                                         skb->ip_summed = CHECKSUM_PARTIAL;
988
989                                 skb_entail(sk, skb);
990                                 copy = size_goal;
991                                 max = size_goal;
992                         }
993
994                         /* Try to append data to the end of skb. */
995                         if (copy > seglen)
996                                 copy = seglen;
997
998                         /* Where to copy to? */
999                         if (skb_tailroom(skb) > 0) {
1000                                 /* We have some space in skb head. Superb! */
1001                                 if (copy > skb_tailroom(skb))
1002                                         copy = skb_tailroom(skb);
1003                                 err = skb_add_data_nocache(sk, skb, from, copy);
1004                                 if (err)
1005                                         goto do_fault;
1006                         } else {
1007                                 int merge = 0;
1008                                 int i = skb_shinfo(skb)->nr_frags;
1009                                 struct page *page = TCP_PAGE(sk);
1010                                 int off = TCP_OFF(sk);
1011
1012                                 if (skb_can_coalesce(skb, i, page, off) &&
1013                                     off != PAGE_SIZE) {
1014                                         /* We can extend the last page
1015                                          * fragment. */
1016                                         merge = 1;
1017                                 } else if (i == MAX_SKB_FRAGS || !sg) {
1018                                         /* Need to add new fragment and cannot
1019                                          * do this because interface is non-SG,
1020                                          * or because all the page slots are
1021                                          * busy. */
1022                                         tcp_mark_push(tp, skb);
1023                                         goto new_segment;
1024                                 } else if (page) {
1025                                         if (off == PAGE_SIZE) {
1026                                                 put_page(page);
1027                                                 TCP_PAGE(sk) = page = NULL;
1028                                                 off = 0;
1029                                         }
1030                                 } else
1031                                         off = 0;
1032
1033                                 if (copy > PAGE_SIZE - off)
1034                                         copy = PAGE_SIZE - off;
1035
1036                                 if (!sk_wmem_schedule(sk, copy))
1037                                         goto wait_for_memory;
1038
1039                                 if (!page) {
1040                                         /* Allocate new cache page. */
1041                                         if (!(page = sk_stream_alloc_page(sk)))
1042                                                 goto wait_for_memory;
1043                                 }
1044
1045                                 /* Time to copy data. We are close to
1046                                  * the end! */
1047                                 err = skb_copy_to_page_nocache(sk, from, skb,
1048                                                                page, off, copy);
1049                                 if (err) {
1050                                         /* If this page was new, give it to the
1051                                          * socket so it does not get leaked.
1052                                          */
1053                                         if (!TCP_PAGE(sk)) {
1054                                                 TCP_PAGE(sk) = page;
1055                                                 TCP_OFF(sk) = 0;
1056                                         }
1057                                         goto do_error;
1058                                 }
1059
1060                                 /* Update the skb. */
1061                                 if (merge) {
1062                                         skb_shinfo(skb)->frags[i - 1].size +=
1063                                                                         copy;
1064                                 } else {
1065                                         skb_fill_page_desc(skb, i, page, off, copy);
1066                                         if (TCP_PAGE(sk)) {
1067                                                 get_page(page);
1068                                         } else if (off + copy < PAGE_SIZE) {
1069                                                 get_page(page);
1070                                                 TCP_PAGE(sk) = page;
1071                                         }
1072                                 }
1073
1074                                 TCP_OFF(sk) = off + copy;
1075                         }
1076
1077                         if (!copied)
1078                                 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
1079
1080                         tp->write_seq += copy;
1081                         TCP_SKB_CB(skb)->end_seq += copy;
1082                         skb_shinfo(skb)->gso_segs = 0;
1083
1084                         from += copy;
1085                         copied += copy;
1086                         if ((seglen -= copy) == 0 && iovlen == 0)
1087                                 goto out;
1088
1089                         if (skb->len < max || (flags & MSG_OOB))
1090                                 continue;
1091
1092                         if (forced_push(tp)) {
1093                                 tcp_mark_push(tp, skb);
1094                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1095                         } else if (skb == tcp_send_head(sk))
1096                                 tcp_push_one(sk, mss_now);
1097                         continue;
1098
1099 wait_for_sndbuf:
1100                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1101 wait_for_memory:
1102                         if (copied)
1103                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1104
1105                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1106                                 goto do_error;
1107
1108                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1109                 }
1110         }
1111
1112 out:
1113         if (copied)
1114                 tcp_push(sk, flags, mss_now, tp->nonagle);
1115         release_sock(sk);
1116
1117         if (copied > 0)
1118                 uid_stat_tcp_snd(current_uid(), copied);
1119         return copied;
1120
1121 do_fault:
1122         if (!skb->len) {
1123                 tcp_unlink_write_queue(skb, sk);
1124                 /* It is the one place in all of TCP, except connection
1125                  * reset, where we can be unlinking the send_head.
1126                  */
1127                 tcp_check_send_head(sk, skb);
1128                 sk_wmem_free_skb(sk, skb);
1129         }
1130
1131 do_error:
1132         if (copied)
1133                 goto out;
1134 out_err:
1135         err = sk_stream_error(sk, flags, err);
1136         release_sock(sk);
1137         return err;
1138 }
1139 EXPORT_SYMBOL(tcp_sendmsg);
1140
1141 /*
1142  *      Handle reading urgent data. BSD has very simple semantics for
1143  *      this, no blocking and very strange errors 8)
1144  */
1145
1146 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1147 {
1148         struct tcp_sock *tp = tcp_sk(sk);
1149
1150         /* No URG data to read. */
1151         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1152             tp->urg_data == TCP_URG_READ)
1153                 return -EINVAL; /* Yes this is right ! */
1154
1155         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1156                 return -ENOTCONN;
1157
1158         if (tp->urg_data & TCP_URG_VALID) {
1159                 int err = 0;
1160                 char c = tp->urg_data;
1161
1162                 if (!(flags & MSG_PEEK))
1163                         tp->urg_data = TCP_URG_READ;
1164
1165                 /* Read urgent data. */
1166                 msg->msg_flags |= MSG_OOB;
1167
1168                 if (len > 0) {
1169                         if (!(flags & MSG_TRUNC))
1170                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1171                         len = 1;
1172                 } else
1173                         msg->msg_flags |= MSG_TRUNC;
1174
1175                 return err ? -EFAULT : len;
1176         }
1177
1178         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1179                 return 0;
1180
1181         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1182          * the available implementations agree in this case:
1183          * this call should never block, independent of the
1184          * blocking state of the socket.
1185          * Mike <pall@rz.uni-karlsruhe.de>
1186          */
1187         return -EAGAIN;
1188 }
1189
1190 /* Clean up the receive buffer for full frames taken by the user,
1191  * then send an ACK if necessary.  COPIED is the number of bytes
1192  * tcp_recvmsg has given to the user so far, it speeds up the
1193  * calculation of whether or not we must ACK for the sake of
1194  * a window update.
1195  */
1196 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1197 {
1198         struct tcp_sock *tp = tcp_sk(sk);
1199         int time_to_ack = 0;
1200
1201 #if TCP_DEBUG
1202         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1203
1204         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1205              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1206              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1207 #endif
1208
1209         if (inet_csk_ack_scheduled(sk)) {
1210                 const struct inet_connection_sock *icsk = inet_csk(sk);
1211                    /* Delayed ACKs frequently hit locked sockets during bulk
1212                     * receive. */
1213                 if (icsk->icsk_ack.blocked ||
1214                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1215                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1216                     /*
1217                      * If this read emptied read buffer, we send ACK, if
1218                      * connection is not bidirectional, user drained
1219                      * receive buffer and there was a small segment
1220                      * in queue.
1221                      */
1222                     (copied > 0 &&
1223                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1224                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1225                        !icsk->icsk_ack.pingpong)) &&
1226                       !atomic_read(&sk->sk_rmem_alloc)))
1227                         time_to_ack = 1;
1228         }
1229
1230         /* We send an ACK if we can now advertise a non-zero window
1231          * which has been raised "significantly".
1232          *
1233          * Even if window raised up to infinity, do not send window open ACK
1234          * in states, where we will not receive more. It is useless.
1235          */
1236         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1237                 __u32 rcv_window_now = tcp_receive_window(tp);
1238
1239                 /* Optimize, __tcp_select_window() is not cheap. */
1240                 if (2*rcv_window_now <= tp->window_clamp) {
1241                         __u32 new_window = __tcp_select_window(sk);
1242
1243                         /* Send ACK now, if this read freed lots of space
1244                          * in our buffer. Certainly, new_window is new window.
1245                          * We can advertise it now, if it is not less than current one.
1246                          * "Lots" means "at least twice" here.
1247                          */
1248                         if (new_window && new_window >= 2 * rcv_window_now)
1249                                 time_to_ack = 1;
1250                 }
1251         }
1252         if (time_to_ack)
1253                 tcp_send_ack(sk);
1254 }
1255
1256 static void tcp_prequeue_process(struct sock *sk)
1257 {
1258         struct sk_buff *skb;
1259         struct tcp_sock *tp = tcp_sk(sk);
1260
1261         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1262
1263         /* RX process wants to run with disabled BHs, though it is not
1264          * necessary */
1265         local_bh_disable();
1266         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1267                 sk_backlog_rcv(sk, skb);
1268         local_bh_enable();
1269
1270         /* Clear memory counter. */
1271         tp->ucopy.memory = 0;
1272 }
1273
1274 #ifdef CONFIG_NET_DMA
1275 static void tcp_service_net_dma(struct sock *sk, bool wait)
1276 {
1277         dma_cookie_t done, used;
1278         dma_cookie_t last_issued;
1279         struct tcp_sock *tp = tcp_sk(sk);
1280
1281         if (!tp->ucopy.dma_chan)
1282                 return;
1283
1284         last_issued = tp->ucopy.dma_cookie;
1285         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1286
1287         do {
1288                 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1289                                               last_issued, &done,
1290                                               &used) == DMA_SUCCESS) {
1291                         /* Safe to free early-copied skbs now */
1292                         __skb_queue_purge(&sk->sk_async_wait_queue);
1293                         break;
1294                 } else {
1295                         struct sk_buff *skb;
1296                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1297                                (dma_async_is_complete(skb->dma_cookie, done,
1298                                                       used) == DMA_SUCCESS)) {
1299                                 __skb_dequeue(&sk->sk_async_wait_queue);
1300                                 kfree_skb(skb);
1301                         }
1302                 }
1303         } while (wait);
1304 }
1305 #endif
1306
1307 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1308 {
1309         struct sk_buff *skb;
1310         u32 offset;
1311
1312         skb_queue_walk(&sk->sk_receive_queue, skb) {
1313                 offset = seq - TCP_SKB_CB(skb)->seq;
1314                 if (tcp_hdr(skb)->syn)
1315                         offset--;
1316                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1317                         *off = offset;
1318                         return skb;
1319                 }
1320         }
1321         return NULL;
1322 }
1323
1324 /*
1325  * This routine provides an alternative to tcp_recvmsg() for routines
1326  * that would like to handle copying from skbuffs directly in 'sendfile'
1327  * fashion.
1328  * Note:
1329  *      - It is assumed that the socket was locked by the caller.
1330  *      - The routine does not block.
1331  *      - At present, there is no support for reading OOB data
1332  *        or for 'peeking' the socket using this routine
1333  *        (although both would be easy to implement).
1334  */
1335 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1336                   sk_read_actor_t recv_actor)
1337 {
1338         struct sk_buff *skb;
1339         struct tcp_sock *tp = tcp_sk(sk);
1340         u32 seq = tp->copied_seq;
1341         u32 offset;
1342         int copied = 0;
1343
1344         if (sk->sk_state == TCP_LISTEN)
1345                 return -ENOTCONN;
1346         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1347                 if (offset < skb->len) {
1348                         int used;
1349                         size_t len;
1350
1351                         len = skb->len - offset;
1352                         /* Stop reading if we hit a patch of urgent data */
1353                         if (tp->urg_data) {
1354                                 u32 urg_offset = tp->urg_seq - seq;
1355                                 if (urg_offset < len)
1356                                         len = urg_offset;
1357                                 if (!len)
1358                                         break;
1359                         }
1360                         used = recv_actor(desc, skb, offset, len);
1361                         if (used < 0) {
1362                                 if (!copied)
1363                                         copied = used;
1364                                 break;
1365                         } else if (used <= len) {
1366                                 seq += used;
1367                                 copied += used;
1368                                 offset += used;
1369                         }
1370                         /*
1371                          * If recv_actor drops the lock (e.g. TCP splice
1372                          * receive) the skb pointer might be invalid when
1373                          * getting here: tcp_collapse might have deleted it
1374                          * while aggregating skbs from the socket queue.
1375                          */
1376                         skb = tcp_recv_skb(sk, seq-1, &offset);
1377                         if (!skb || (offset+1 != skb->len))
1378                                 break;
1379                 }
1380                 if (tcp_hdr(skb)->fin) {
1381                         sk_eat_skb(sk, skb, 0);
1382                         ++seq;
1383                         break;
1384                 }
1385                 sk_eat_skb(sk, skb, 0);
1386                 if (!desc->count)
1387                         break;
1388                 tp->copied_seq = seq;
1389         }
1390         tp->copied_seq = seq;
1391
1392         tcp_rcv_space_adjust(sk);
1393
1394         /* Clean up data we have read: This will do ACK frames. */
1395         if (copied > 0) {
1396                 tcp_cleanup_rbuf(sk, copied);
1397                 uid_stat_tcp_rcv(current_uid(), copied);
1398         }
1399
1400         return copied;
1401 }
1402 EXPORT_SYMBOL(tcp_read_sock);
1403
1404 /*
1405  *      This routine copies from a sock struct into the user buffer.
1406  *
1407  *      Technical note: in 2.3 we work on _locked_ socket, so that
1408  *      tricks with *seq access order and skb->users are not required.
1409  *      Probably, code can be easily improved even more.
1410  */
1411
1412 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1413                 size_t len, int nonblock, int flags, int *addr_len)
1414 {
1415         struct tcp_sock *tp = tcp_sk(sk);
1416         int copied = 0;
1417         u32 peek_seq;
1418         u32 *seq;
1419         unsigned long used;
1420         int err;
1421         int target;             /* Read at least this many bytes */
1422         long timeo;
1423         struct task_struct *user_recv = NULL;
1424         int copied_early = 0;
1425         struct sk_buff *skb;
1426         u32 urg_hole = 0;
1427
1428         lock_sock(sk);
1429
1430         err = -ENOTCONN;
1431         if (sk->sk_state == TCP_LISTEN)
1432                 goto out;
1433
1434         timeo = sock_rcvtimeo(sk, nonblock);
1435
1436         /* Urgent data needs to be handled specially. */
1437         if (flags & MSG_OOB)
1438                 goto recv_urg;
1439
1440         seq = &tp->copied_seq;
1441         if (flags & MSG_PEEK) {
1442                 peek_seq = tp->copied_seq;
1443                 seq = &peek_seq;
1444         }
1445
1446         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1447
1448 #ifdef CONFIG_NET_DMA
1449         tp->ucopy.dma_chan = NULL;
1450         preempt_disable();
1451         skb = skb_peek_tail(&sk->sk_receive_queue);
1452         {
1453                 int available = 0;
1454
1455                 if (skb)
1456                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1457                 if ((available < target) &&
1458                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1459                     !sysctl_tcp_low_latency &&
1460                     dma_find_channel(DMA_MEMCPY)) {
1461                         preempt_enable_no_resched();
1462                         tp->ucopy.pinned_list =
1463                                         dma_pin_iovec_pages(msg->msg_iov, len);
1464                 } else {
1465                         preempt_enable_no_resched();
1466                 }
1467         }
1468 #endif
1469
1470         do {
1471                 u32 offset;
1472
1473                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1474                 if (tp->urg_data && tp->urg_seq == *seq) {
1475                         if (copied)
1476                                 break;
1477                         if (signal_pending(current)) {
1478                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1479                                 break;
1480                         }
1481                 }
1482
1483                 /* Next get a buffer. */
1484
1485                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1486                         /* Now that we have two receive queues this
1487                          * shouldn't happen.
1488                          */
1489                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1490                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1491                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1492                                  flags))
1493                                 break;
1494
1495                         offset = *seq - TCP_SKB_CB(skb)->seq;
1496                         if (tcp_hdr(skb)->syn)
1497                                 offset--;
1498                         if (offset < skb->len)
1499                                 goto found_ok_skb;
1500                         if (tcp_hdr(skb)->fin)
1501                                 goto found_fin_ok;
1502                         WARN(!(flags & MSG_PEEK),
1503                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1504                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1505                 }
1506
1507                 /* Well, if we have backlog, try to process it now yet. */
1508
1509                 if (copied >= target && !sk->sk_backlog.tail)
1510                         break;
1511
1512                 if (copied) {
1513                         if (sk->sk_err ||
1514                             sk->sk_state == TCP_CLOSE ||
1515                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1516                             !timeo ||
1517                             signal_pending(current))
1518                                 break;
1519                 } else {
1520                         if (sock_flag(sk, SOCK_DONE))
1521                                 break;
1522
1523                         if (sk->sk_err) {
1524                                 copied = sock_error(sk);
1525                                 break;
1526                         }
1527
1528                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1529                                 break;
1530
1531                         if (sk->sk_state == TCP_CLOSE) {
1532                                 if (!sock_flag(sk, SOCK_DONE)) {
1533                                         /* This occurs when user tries to read
1534                                          * from never connected socket.
1535                                          */
1536                                         copied = -ENOTCONN;
1537                                         break;
1538                                 }
1539                                 break;
1540                         }
1541
1542                         if (!timeo) {
1543                                 copied = -EAGAIN;
1544                                 break;
1545                         }
1546
1547                         if (signal_pending(current)) {
1548                                 copied = sock_intr_errno(timeo);
1549                                 break;
1550                         }
1551                 }
1552
1553                 tcp_cleanup_rbuf(sk, copied);
1554
1555                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1556                         /* Install new reader */
1557                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1558                                 user_recv = current;
1559                                 tp->ucopy.task = user_recv;
1560                                 tp->ucopy.iov = msg->msg_iov;
1561                         }
1562
1563                         tp->ucopy.len = len;
1564
1565                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1566                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1567
1568                         /* Ugly... If prequeue is not empty, we have to
1569                          * process it before releasing socket, otherwise
1570                          * order will be broken at second iteration.
1571                          * More elegant solution is required!!!
1572                          *
1573                          * Look: we have the following (pseudo)queues:
1574                          *
1575                          * 1. packets in flight
1576                          * 2. backlog
1577                          * 3. prequeue
1578                          * 4. receive_queue
1579                          *
1580                          * Each queue can be processed only if the next ones
1581                          * are empty. At this point we have empty receive_queue.
1582                          * But prequeue _can_ be not empty after 2nd iteration,
1583                          * when we jumped to start of loop because backlog
1584                          * processing added something to receive_queue.
1585                          * We cannot release_sock(), because backlog contains
1586                          * packets arrived _after_ prequeued ones.
1587                          *
1588                          * Shortly, algorithm is clear --- to process all
1589                          * the queues in order. We could make it more directly,
1590                          * requeueing packets from backlog to prequeue, if
1591                          * is not empty. It is more elegant, but eats cycles,
1592                          * unfortunately.
1593                          */
1594                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1595                                 goto do_prequeue;
1596
1597                         /* __ Set realtime policy in scheduler __ */
1598                 }
1599
1600 #ifdef CONFIG_NET_DMA
1601                 if (tp->ucopy.dma_chan)
1602                         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1603 #endif
1604                 if (copied >= target) {
1605                         /* Do not sleep, just process backlog. */
1606                         release_sock(sk);
1607                         lock_sock(sk);
1608                 } else
1609                         sk_wait_data(sk, &timeo);
1610
1611 #ifdef CONFIG_NET_DMA
1612                 tcp_service_net_dma(sk, false);  /* Don't block */
1613                 tp->ucopy.wakeup = 0;
1614 #endif
1615
1616                 if (user_recv) {
1617                         int chunk;
1618
1619                         /* __ Restore normal policy in scheduler __ */
1620
1621                         if ((chunk = len - tp->ucopy.len) != 0) {
1622                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1623                                 len -= chunk;
1624                                 copied += chunk;
1625                         }
1626
1627                         if (tp->rcv_nxt == tp->copied_seq &&
1628                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1629 do_prequeue:
1630                                 tcp_prequeue_process(sk);
1631
1632                                 if ((chunk = len - tp->ucopy.len) != 0) {
1633                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1634                                         len -= chunk;
1635                                         copied += chunk;
1636                                 }
1637                         }
1638                 }
1639                 if ((flags & MSG_PEEK) &&
1640                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1641                         if (net_ratelimit())
1642                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1643                                        current->comm, task_pid_nr(current));
1644                         peek_seq = tp->copied_seq;
1645                 }
1646                 continue;
1647
1648         found_ok_skb:
1649                 /* Ok so how much can we use? */
1650                 used = skb->len - offset;
1651                 if (len < used)
1652                         used = len;
1653
1654                 /* Do we have urgent data here? */
1655                 if (tp->urg_data) {
1656                         u32 urg_offset = tp->urg_seq - *seq;
1657                         if (urg_offset < used) {
1658                                 if (!urg_offset) {
1659                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1660                                                 ++*seq;
1661                                                 urg_hole++;
1662                                                 offset++;
1663                                                 used--;
1664                                                 if (!used)
1665                                                         goto skip_copy;
1666                                         }
1667                                 } else
1668                                         used = urg_offset;
1669                         }
1670                 }
1671
1672                 if (!(flags & MSG_TRUNC)) {
1673 #ifdef CONFIG_NET_DMA
1674                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1675                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1676
1677                         if (tp->ucopy.dma_chan) {
1678                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1679                                         tp->ucopy.dma_chan, skb, offset,
1680                                         msg->msg_iov, used,
1681                                         tp->ucopy.pinned_list);
1682
1683                                 if (tp->ucopy.dma_cookie < 0) {
1684
1685                                         printk(KERN_ALERT "dma_cookie < 0\n");
1686
1687                                         /* Exception. Bailout! */
1688                                         if (!copied)
1689                                                 copied = -EFAULT;
1690                                         break;
1691                                 }
1692
1693                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1694
1695                                 if ((offset + used) == skb->len)
1696                                         copied_early = 1;
1697
1698                         } else
1699 #endif
1700                         {
1701                                 err = skb_copy_datagram_iovec(skb, offset,
1702                                                 msg->msg_iov, used);
1703                                 if (err) {
1704                                         /* Exception. Bailout! */
1705                                         if (!copied)
1706                                                 copied = -EFAULT;
1707                                         break;
1708                                 }
1709                         }
1710                 }
1711
1712                 *seq += used;
1713                 copied += used;
1714                 len -= used;
1715
1716                 tcp_rcv_space_adjust(sk);
1717
1718 skip_copy:
1719                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1720                         tp->urg_data = 0;
1721                         tcp_fast_path_check(sk);
1722                 }
1723                 if (used + offset < skb->len)
1724                         continue;
1725
1726                 if (tcp_hdr(skb)->fin)
1727                         goto found_fin_ok;
1728                 if (!(flags & MSG_PEEK)) {
1729                         sk_eat_skb(sk, skb, copied_early);
1730                         copied_early = 0;
1731                 }
1732                 continue;
1733
1734         found_fin_ok:
1735                 /* Process the FIN. */
1736                 ++*seq;
1737                 if (!(flags & MSG_PEEK)) {
1738                         sk_eat_skb(sk, skb, copied_early);
1739                         copied_early = 0;
1740                 }
1741                 break;
1742         } while (len > 0);
1743
1744         if (user_recv) {
1745                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1746                         int chunk;
1747
1748                         tp->ucopy.len = copied > 0 ? len : 0;
1749
1750                         tcp_prequeue_process(sk);
1751
1752                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1753                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1754                                 len -= chunk;
1755                                 copied += chunk;
1756                         }
1757                 }
1758
1759                 tp->ucopy.task = NULL;
1760                 tp->ucopy.len = 0;
1761         }
1762
1763 #ifdef CONFIG_NET_DMA
1764         tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1765         tp->ucopy.dma_chan = NULL;
1766
1767         if (tp->ucopy.pinned_list) {
1768                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1769                 tp->ucopy.pinned_list = NULL;
1770         }
1771 #endif
1772
1773         /* According to UNIX98, msg_name/msg_namelen are ignored
1774          * on connected socket. I was just happy when found this 8) --ANK
1775          */
1776
1777         /* Clean up data we have read: This will do ACK frames. */
1778         tcp_cleanup_rbuf(sk, copied);
1779
1780         release_sock(sk);
1781
1782         if (copied > 0)
1783                 uid_stat_tcp_rcv(current_uid(), copied);
1784         return copied;
1785
1786 out:
1787         release_sock(sk);
1788         return err;
1789
1790 recv_urg:
1791         err = tcp_recv_urg(sk, msg, len, flags);
1792         if (err > 0)
1793                 uid_stat_tcp_rcv(current_uid(), err);
1794         goto out;
1795 }
1796 EXPORT_SYMBOL(tcp_recvmsg);
1797
1798 void tcp_set_state(struct sock *sk, int state)
1799 {
1800         int oldstate = sk->sk_state;
1801
1802         switch (state) {
1803         case TCP_ESTABLISHED:
1804                 if (oldstate != TCP_ESTABLISHED)
1805                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1806                 break;
1807
1808         case TCP_CLOSE:
1809                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1810                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1811
1812                 sk->sk_prot->unhash(sk);
1813                 if (inet_csk(sk)->icsk_bind_hash &&
1814                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1815                         inet_put_port(sk);
1816                 /* fall through */
1817         default:
1818                 if (oldstate == TCP_ESTABLISHED)
1819                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1820         }
1821
1822         /* Change state AFTER socket is unhashed to avoid closed
1823          * socket sitting in hash tables.
1824          */
1825         sk->sk_state = state;
1826
1827 #ifdef STATE_TRACE
1828         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1829 #endif
1830 }
1831 EXPORT_SYMBOL_GPL(tcp_set_state);
1832
1833 /*
1834  *      State processing on a close. This implements the state shift for
1835  *      sending our FIN frame. Note that we only send a FIN for some
1836  *      states. A shutdown() may have already sent the FIN, or we may be
1837  *      closed.
1838  */
1839
1840 static const unsigned char new_state[16] = {
1841   /* current state:        new state:      action:      */
1842   /* (Invalid)          */ TCP_CLOSE,
1843   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1844   /* TCP_SYN_SENT       */ TCP_CLOSE,
1845   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1846   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1847   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1848   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1849   /* TCP_CLOSE          */ TCP_CLOSE,
1850   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1851   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1852   /* TCP_LISTEN         */ TCP_CLOSE,
1853   /* TCP_CLOSING        */ TCP_CLOSING,
1854 };
1855
1856 static int tcp_close_state(struct sock *sk)
1857 {
1858         int next = (int)new_state[sk->sk_state];
1859         int ns = next & TCP_STATE_MASK;
1860
1861         tcp_set_state(sk, ns);
1862
1863         return next & TCP_ACTION_FIN;
1864 }
1865
1866 /*
1867  *      Shutdown the sending side of a connection. Much like close except
1868  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1869  */
1870
1871 void tcp_shutdown(struct sock *sk, int how)
1872 {
1873         /*      We need to grab some memory, and put together a FIN,
1874          *      and then put it into the queue to be sent.
1875          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1876          */
1877         if (!(how & SEND_SHUTDOWN))
1878                 return;
1879
1880         /* If we've already sent a FIN, or it's a closed state, skip this. */
1881         if ((1 << sk->sk_state) &
1882             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1883              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1884                 /* Clear out any half completed packets.  FIN if needed. */
1885                 if (tcp_close_state(sk))
1886                         tcp_send_fin(sk);
1887         }
1888 }
1889 EXPORT_SYMBOL(tcp_shutdown);
1890
1891 void tcp_close(struct sock *sk, long timeout)
1892 {
1893         struct sk_buff *skb;
1894         int data_was_unread = 0;
1895         int state;
1896
1897         lock_sock(sk);
1898         sk->sk_shutdown = SHUTDOWN_MASK;
1899
1900         if (sk->sk_state == TCP_LISTEN) {
1901                 tcp_set_state(sk, TCP_CLOSE);
1902
1903                 /* Special case. */
1904                 inet_csk_listen_stop(sk);
1905
1906                 goto adjudge_to_death;
1907         }
1908
1909         /*  We need to flush the recv. buffs.  We do this only on the
1910          *  descriptor close, not protocol-sourced closes, because the
1911          *  reader process may not have drained the data yet!
1912          */
1913         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1914                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1915                           tcp_hdr(skb)->fin;
1916                 data_was_unread += len;
1917                 __kfree_skb(skb);
1918         }
1919
1920         sk_mem_reclaim(sk);
1921
1922         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1923         if (sk->sk_state == TCP_CLOSE)
1924                 goto adjudge_to_death;
1925
1926         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1927          * data was lost. To witness the awful effects of the old behavior of
1928          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1929          * GET in an FTP client, suspend the process, wait for the client to
1930          * advertise a zero window, then kill -9 the FTP client, wheee...
1931          * Note: timeout is always zero in such a case.
1932          */
1933         if (data_was_unread) {
1934                 /* Unread data was tossed, zap the connection. */
1935                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1936                 tcp_set_state(sk, TCP_CLOSE);
1937                 tcp_send_active_reset(sk, sk->sk_allocation);
1938         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1939                 /* Check zero linger _after_ checking for unread data. */
1940                 sk->sk_prot->disconnect(sk, 0);
1941                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1942         } else if (tcp_close_state(sk)) {
1943                 /* We FIN if the application ate all the data before
1944                  * zapping the connection.
1945                  */
1946
1947                 /* RED-PEN. Formally speaking, we have broken TCP state
1948                  * machine. State transitions:
1949                  *
1950                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1951                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1952                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1953                  *
1954                  * are legal only when FIN has been sent (i.e. in window),
1955                  * rather than queued out of window. Purists blame.
1956                  *
1957                  * F.e. "RFC state" is ESTABLISHED,
1958                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1959                  *
1960                  * The visible declinations are that sometimes
1961                  * we enter time-wait state, when it is not required really
1962                  * (harmless), do not send active resets, when they are
1963                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1964                  * they look as CLOSING or LAST_ACK for Linux)
1965                  * Probably, I missed some more holelets.
1966                  *                                              --ANK
1967                  */
1968                 tcp_send_fin(sk);
1969         }
1970
1971         sk_stream_wait_close(sk, timeout);
1972
1973 adjudge_to_death:
1974         state = sk->sk_state;
1975         sock_hold(sk);
1976         sock_orphan(sk);
1977
1978         /* It is the last release_sock in its life. It will remove backlog. */
1979         release_sock(sk);
1980
1981
1982         /* Now socket is owned by kernel and we acquire BH lock
1983            to finish close. No need to check for user refs.
1984          */
1985         local_bh_disable();
1986         bh_lock_sock(sk);
1987         WARN_ON(sock_owned_by_user(sk));
1988
1989         percpu_counter_inc(sk->sk_prot->orphan_count);
1990
1991         /* Have we already been destroyed by a softirq or backlog? */
1992         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1993                 goto out;
1994
1995         /*      This is a (useful) BSD violating of the RFC. There is a
1996          *      problem with TCP as specified in that the other end could
1997          *      keep a socket open forever with no application left this end.
1998          *      We use a 3 minute timeout (about the same as BSD) then kill
1999          *      our end. If they send after that then tough - BUT: long enough
2000          *      that we won't make the old 4*rto = almost no time - whoops
2001          *      reset mistake.
2002          *
2003          *      Nope, it was not mistake. It is really desired behaviour
2004          *      f.e. on http servers, when such sockets are useless, but
2005          *      consume significant resources. Let's do it with special
2006          *      linger2 option.                                 --ANK
2007          */
2008
2009         if (sk->sk_state == TCP_FIN_WAIT2) {
2010                 struct tcp_sock *tp = tcp_sk(sk);
2011                 if (tp->linger2 < 0) {
2012                         tcp_set_state(sk, TCP_CLOSE);
2013                         tcp_send_active_reset(sk, GFP_ATOMIC);
2014                         NET_INC_STATS_BH(sock_net(sk),
2015                                         LINUX_MIB_TCPABORTONLINGER);
2016                 } else {
2017                         const int tmo = tcp_fin_time(sk);
2018
2019                         if (tmo > TCP_TIMEWAIT_LEN) {
2020                                 inet_csk_reset_keepalive_timer(sk,
2021                                                 tmo - TCP_TIMEWAIT_LEN);
2022                         } else {
2023                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2024                                 goto out;
2025                         }
2026                 }
2027         }
2028         if (sk->sk_state != TCP_CLOSE) {
2029                 sk_mem_reclaim(sk);
2030                 if (tcp_too_many_orphans(sk, 0)) {
2031                         if (net_ratelimit())
2032                                 printk(KERN_INFO "TCP: too many of orphaned "
2033                                        "sockets\n");
2034                         tcp_set_state(sk, TCP_CLOSE);
2035                         tcp_send_active_reset(sk, GFP_ATOMIC);
2036                         NET_INC_STATS_BH(sock_net(sk),
2037                                         LINUX_MIB_TCPABORTONMEMORY);
2038                 }
2039         }
2040
2041         if (sk->sk_state == TCP_CLOSE)
2042                 inet_csk_destroy_sock(sk);
2043         /* Otherwise, socket is reprieved until protocol close. */
2044
2045 out:
2046         bh_unlock_sock(sk);
2047         local_bh_enable();
2048         sock_put(sk);
2049 }
2050 EXPORT_SYMBOL(tcp_close);
2051
2052 /* These states need RST on ABORT according to RFC793 */
2053
2054 static inline int tcp_need_reset(int state)
2055 {
2056         return (1 << state) &
2057                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2058                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2059 }
2060
2061 int tcp_disconnect(struct sock *sk, int flags)
2062 {
2063         struct inet_sock *inet = inet_sk(sk);
2064         struct inet_connection_sock *icsk = inet_csk(sk);
2065         struct tcp_sock *tp = tcp_sk(sk);
2066         int err = 0;
2067         int old_state = sk->sk_state;
2068
2069         if (old_state != TCP_CLOSE)
2070                 tcp_set_state(sk, TCP_CLOSE);
2071
2072         /* ABORT function of RFC793 */
2073         if (old_state == TCP_LISTEN) {
2074                 inet_csk_listen_stop(sk);
2075         } else if (tcp_need_reset(old_state) ||
2076                    (tp->snd_nxt != tp->write_seq &&
2077                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2078                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2079                  * states
2080                  */
2081                 tcp_send_active_reset(sk, gfp_any());
2082                 sk->sk_err = ECONNRESET;
2083         } else if (old_state == TCP_SYN_SENT)
2084                 sk->sk_err = ECONNRESET;
2085
2086         tcp_clear_xmit_timers(sk);
2087         __skb_queue_purge(&sk->sk_receive_queue);
2088         tcp_write_queue_purge(sk);
2089         __skb_queue_purge(&tp->out_of_order_queue);
2090 #ifdef CONFIG_NET_DMA
2091         __skb_queue_purge(&sk->sk_async_wait_queue);
2092 #endif
2093
2094         inet->inet_dport = 0;
2095
2096         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2097                 inet_reset_saddr(sk);
2098
2099         sk->sk_shutdown = 0;
2100         sock_reset_flag(sk, SOCK_DONE);
2101         tp->srtt = 0;
2102         if ((tp->write_seq += tp->max_window + 2) == 0)
2103                 tp->write_seq = 1;
2104         icsk->icsk_backoff = 0;
2105         tp->snd_cwnd = 2;
2106         icsk->icsk_probes_out = 0;
2107         tp->packets_out = 0;
2108         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2109         tp->snd_cwnd_cnt = 0;
2110         tp->bytes_acked = 0;
2111         tp->window_clamp = 0;
2112         tcp_set_ca_state(sk, TCP_CA_Open);
2113         tcp_clear_retrans(tp);
2114         inet_csk_delack_init(sk);
2115         tcp_init_send_head(sk);
2116         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2117         __sk_dst_reset(sk);
2118
2119         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2120
2121         sk->sk_error_report(sk);
2122         return err;
2123 }
2124 EXPORT_SYMBOL(tcp_disconnect);
2125
2126 /*
2127  *      Socket option code for TCP.
2128  */
2129 static int do_tcp_setsockopt(struct sock *sk, int level,
2130                 int optname, char __user *optval, unsigned int optlen)
2131 {
2132         struct tcp_sock *tp = tcp_sk(sk);
2133         struct inet_connection_sock *icsk = inet_csk(sk);
2134         int val;
2135         int err = 0;
2136
2137         /* These are data/string values, all the others are ints */
2138         switch (optname) {
2139         case TCP_CONGESTION: {
2140                 char name[TCP_CA_NAME_MAX];
2141
2142                 if (optlen < 1)
2143                         return -EINVAL;
2144
2145                 val = strncpy_from_user(name, optval,
2146                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2147                 if (val < 0)
2148                         return -EFAULT;
2149                 name[val] = 0;
2150
2151                 lock_sock(sk);
2152                 err = tcp_set_congestion_control(sk, name);
2153                 release_sock(sk);
2154                 return err;
2155         }
2156         case TCP_COOKIE_TRANSACTIONS: {
2157                 struct tcp_cookie_transactions ctd;
2158                 struct tcp_cookie_values *cvp = NULL;
2159
2160                 if (sizeof(ctd) > optlen)
2161                         return -EINVAL;
2162                 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2163                         return -EFAULT;
2164
2165                 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2166                     ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2167                         return -EINVAL;
2168
2169                 if (ctd.tcpct_cookie_desired == 0) {
2170                         /* default to global value */
2171                 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2172                            ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2173                            ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2174                         return -EINVAL;
2175                 }
2176
2177                 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2178                         /* Supercedes all other values */
2179                         lock_sock(sk);
2180                         if (tp->cookie_values != NULL) {
2181                                 kref_put(&tp->cookie_values->kref,
2182                                          tcp_cookie_values_release);
2183                                 tp->cookie_values = NULL;
2184                         }
2185                         tp->rx_opt.cookie_in_always = 0; /* false */
2186                         tp->rx_opt.cookie_out_never = 1; /* true */
2187                         release_sock(sk);
2188                         return err;
2189                 }
2190
2191                 /* Allocate ancillary memory before locking.
2192                  */
2193                 if (ctd.tcpct_used > 0 ||
2194                     (tp->cookie_values == NULL &&
2195                      (sysctl_tcp_cookie_size > 0 ||
2196                       ctd.tcpct_cookie_desired > 0 ||
2197                       ctd.tcpct_s_data_desired > 0))) {
2198                         cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2199                                       GFP_KERNEL);
2200                         if (cvp == NULL)
2201                                 return -ENOMEM;
2202
2203                         kref_init(&cvp->kref);
2204                 }
2205                 lock_sock(sk);
2206                 tp->rx_opt.cookie_in_always =
2207                         (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2208                 tp->rx_opt.cookie_out_never = 0; /* false */
2209
2210                 if (tp->cookie_values != NULL) {
2211                         if (cvp != NULL) {
2212                                 /* Changed values are recorded by a changed
2213                                  * pointer, ensuring the cookie will differ,
2214                                  * without separately hashing each value later.
2215                                  */
2216                                 kref_put(&tp->cookie_values->kref,
2217                                          tcp_cookie_values_release);
2218                         } else {
2219                                 cvp = tp->cookie_values;
2220                         }
2221                 }
2222
2223                 if (cvp != NULL) {
2224                         cvp->cookie_desired = ctd.tcpct_cookie_desired;
2225
2226                         if (ctd.tcpct_used > 0) {
2227                                 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2228                                        ctd.tcpct_used);
2229                                 cvp->s_data_desired = ctd.tcpct_used;
2230                                 cvp->s_data_constant = 1; /* true */
2231                         } else {
2232                                 /* No constant payload data. */
2233                                 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2234                                 cvp->s_data_constant = 0; /* false */
2235                         }
2236
2237                         tp->cookie_values = cvp;
2238                 }
2239                 release_sock(sk);
2240                 return err;
2241         }
2242         default:
2243                 /* fallthru */
2244                 break;
2245         }
2246
2247         if (optlen < sizeof(int))
2248                 return -EINVAL;
2249
2250         if (get_user(val, (int __user *)optval))
2251                 return -EFAULT;
2252
2253         lock_sock(sk);
2254
2255         switch (optname) {
2256         case TCP_MAXSEG:
2257                 /* Values greater than interface MTU won't take effect. However
2258                  * at the point when this call is done we typically don't yet
2259                  * know which interface is going to be used */
2260                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2261                         err = -EINVAL;
2262                         break;
2263                 }
2264                 tp->rx_opt.user_mss = val;
2265                 break;
2266
2267         case TCP_NODELAY:
2268                 if (val) {
2269                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2270                          * this option on corked socket is remembered, but
2271                          * it is not activated until cork is cleared.
2272                          *
2273                          * However, when TCP_NODELAY is set we make
2274                          * an explicit push, which overrides even TCP_CORK
2275                          * for currently queued segments.
2276                          */
2277                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2278                         tcp_push_pending_frames(sk);
2279                 } else {
2280                         tp->nonagle &= ~TCP_NAGLE_OFF;
2281                 }
2282                 break;
2283
2284         case TCP_THIN_LINEAR_TIMEOUTS:
2285                 if (val < 0 || val > 1)
2286                         err = -EINVAL;
2287                 else
2288                         tp->thin_lto = val;
2289                 break;
2290
2291         case TCP_THIN_DUPACK:
2292                 if (val < 0 || val > 1)
2293                         err = -EINVAL;
2294                 else
2295                         tp->thin_dupack = val;
2296                 break;
2297
2298         case TCP_CORK:
2299                 /* When set indicates to always queue non-full frames.
2300                  * Later the user clears this option and we transmit
2301                  * any pending partial frames in the queue.  This is
2302                  * meant to be used alongside sendfile() to get properly
2303                  * filled frames when the user (for example) must write
2304                  * out headers with a write() call first and then use
2305                  * sendfile to send out the data parts.
2306                  *
2307                  * TCP_CORK can be set together with TCP_NODELAY and it is
2308                  * stronger than TCP_NODELAY.
2309                  */
2310                 if (val) {
2311                         tp->nonagle |= TCP_NAGLE_CORK;
2312                 } else {
2313                         tp->nonagle &= ~TCP_NAGLE_CORK;
2314                         if (tp->nonagle&TCP_NAGLE_OFF)
2315                                 tp->nonagle |= TCP_NAGLE_PUSH;
2316                         tcp_push_pending_frames(sk);
2317                 }
2318                 break;
2319
2320         case TCP_KEEPIDLE:
2321                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2322                         err = -EINVAL;
2323                 else {
2324                         tp->keepalive_time = val * HZ;
2325                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2326                             !((1 << sk->sk_state) &
2327                               (TCPF_CLOSE | TCPF_LISTEN))) {
2328                                 u32 elapsed = keepalive_time_elapsed(tp);
2329                                 if (tp->keepalive_time > elapsed)
2330                                         elapsed = tp->keepalive_time - elapsed;
2331                                 else
2332                                         elapsed = 0;
2333                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2334                         }
2335                 }
2336                 break;
2337         case TCP_KEEPINTVL:
2338                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2339                         err = -EINVAL;
2340                 else
2341                         tp->keepalive_intvl = val * HZ;
2342                 break;
2343         case TCP_KEEPCNT:
2344                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2345                         err = -EINVAL;
2346                 else
2347                         tp->keepalive_probes = val;
2348                 break;
2349         case TCP_SYNCNT:
2350                 if (val < 1 || val > MAX_TCP_SYNCNT)
2351                         err = -EINVAL;
2352                 else
2353                         icsk->icsk_syn_retries = val;
2354                 break;
2355
2356         case TCP_LINGER2:
2357                 if (val < 0)
2358                         tp->linger2 = -1;
2359                 else if (val > sysctl_tcp_fin_timeout / HZ)
2360                         tp->linger2 = 0;
2361                 else
2362                         tp->linger2 = val * HZ;
2363                 break;
2364
2365         case TCP_DEFER_ACCEPT:
2366                 /* Translate value in seconds to number of retransmits */
2367                 icsk->icsk_accept_queue.rskq_defer_accept =
2368                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2369                                         TCP_RTO_MAX / HZ);
2370                 break;
2371
2372         case TCP_WINDOW_CLAMP:
2373                 if (!val) {
2374                         if (sk->sk_state != TCP_CLOSE) {
2375                                 err = -EINVAL;
2376                                 break;
2377                         }
2378                         tp->window_clamp = 0;
2379                 } else
2380                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2381                                                 SOCK_MIN_RCVBUF / 2 : val;
2382                 break;
2383
2384         case TCP_QUICKACK:
2385                 if (!val) {
2386                         icsk->icsk_ack.pingpong = 1;
2387                 } else {
2388                         icsk->icsk_ack.pingpong = 0;
2389                         if ((1 << sk->sk_state) &
2390                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2391                             inet_csk_ack_scheduled(sk)) {
2392                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2393                                 tcp_cleanup_rbuf(sk, 1);
2394                                 if (!(val & 1))
2395                                         icsk->icsk_ack.pingpong = 1;
2396                         }
2397                 }
2398                 break;
2399
2400 #ifdef CONFIG_TCP_MD5SIG
2401         case TCP_MD5SIG:
2402                 /* Read the IP->Key mappings from userspace */
2403                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2404                 break;
2405 #endif
2406         case TCP_USER_TIMEOUT:
2407                 /* Cap the max timeout in ms TCP will retry/retrans
2408                  * before giving up and aborting (ETIMEDOUT) a connection.
2409                  */
2410                 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2411                 break;
2412         default:
2413                 err = -ENOPROTOOPT;
2414                 break;
2415         }
2416
2417         release_sock(sk);
2418         return err;
2419 }
2420
2421 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2422                    unsigned int optlen)
2423 {
2424         struct inet_connection_sock *icsk = inet_csk(sk);
2425
2426         if (level != SOL_TCP)
2427                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2428                                                      optval, optlen);
2429         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2430 }
2431 EXPORT_SYMBOL(tcp_setsockopt);
2432
2433 #ifdef CONFIG_COMPAT
2434 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2435                           char __user *optval, unsigned int optlen)
2436 {
2437         if (level != SOL_TCP)
2438                 return inet_csk_compat_setsockopt(sk, level, optname,
2439                                                   optval, optlen);
2440         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2441 }
2442 EXPORT_SYMBOL(compat_tcp_setsockopt);
2443 #endif
2444
2445 /* Return information about state of tcp endpoint in API format. */
2446 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2447 {
2448         struct tcp_sock *tp = tcp_sk(sk);
2449         const struct inet_connection_sock *icsk = inet_csk(sk);
2450         u32 now = tcp_time_stamp;
2451
2452         memset(info, 0, sizeof(*info));
2453
2454         info->tcpi_state = sk->sk_state;
2455         info->tcpi_ca_state = icsk->icsk_ca_state;
2456         info->tcpi_retransmits = icsk->icsk_retransmits;
2457         info->tcpi_probes = icsk->icsk_probes_out;
2458         info->tcpi_backoff = icsk->icsk_backoff;
2459
2460         if (tp->rx_opt.tstamp_ok)
2461                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2462         if (tcp_is_sack(tp))
2463                 info->tcpi_options |= TCPI_OPT_SACK;
2464         if (tp->rx_opt.wscale_ok) {
2465                 info->tcpi_options |= TCPI_OPT_WSCALE;
2466                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2467                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2468         }
2469
2470         if (tp->ecn_flags&TCP_ECN_OK)
2471                 info->tcpi_options |= TCPI_OPT_ECN;
2472
2473         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2474         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2475         info->tcpi_snd_mss = tp->mss_cache;
2476         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2477
2478         if (sk->sk_state == TCP_LISTEN) {
2479                 info->tcpi_unacked = sk->sk_ack_backlog;
2480                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2481         } else {
2482                 info->tcpi_unacked = tp->packets_out;
2483                 info->tcpi_sacked = tp->sacked_out;
2484         }
2485         info->tcpi_lost = tp->lost_out;
2486         info->tcpi_retrans = tp->retrans_out;
2487         info->tcpi_fackets = tp->fackets_out;
2488
2489         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2490         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2491         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2492
2493         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2494         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2495         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2496         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2497         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2498         info->tcpi_snd_cwnd = tp->snd_cwnd;
2499         info->tcpi_advmss = tp->advmss;
2500         info->tcpi_reordering = tp->reordering;
2501
2502         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2503         info->tcpi_rcv_space = tp->rcvq_space.space;
2504
2505         info->tcpi_total_retrans = tp->total_retrans;
2506 }
2507 EXPORT_SYMBOL_GPL(tcp_get_info);
2508
2509 static int do_tcp_getsockopt(struct sock *sk, int level,
2510                 int optname, char __user *optval, int __user *optlen)
2511 {
2512         struct inet_connection_sock *icsk = inet_csk(sk);
2513         struct tcp_sock *tp = tcp_sk(sk);
2514         int val, len;
2515
2516         if (get_user(len, optlen))
2517                 return -EFAULT;
2518
2519         len = min_t(unsigned int, len, sizeof(int));
2520
2521         if (len < 0)
2522                 return -EINVAL;
2523
2524         switch (optname) {
2525         case TCP_MAXSEG:
2526                 val = tp->mss_cache;
2527                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2528                         val = tp->rx_opt.user_mss;
2529                 break;
2530         case TCP_NODELAY:
2531                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2532                 break;
2533         case TCP_CORK:
2534                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2535                 break;
2536         case TCP_KEEPIDLE:
2537                 val = keepalive_time_when(tp) / HZ;
2538                 break;
2539         case TCP_KEEPINTVL:
2540                 val = keepalive_intvl_when(tp) / HZ;
2541                 break;
2542         case TCP_KEEPCNT:
2543                 val = keepalive_probes(tp);
2544                 break;
2545         case TCP_SYNCNT:
2546                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2547                 break;
2548         case TCP_LINGER2:
2549                 val = tp->linger2;
2550                 if (val >= 0)
2551                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2552                 break;
2553         case TCP_DEFER_ACCEPT:
2554                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2555                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2556                 break;
2557         case TCP_WINDOW_CLAMP:
2558                 val = tp->window_clamp;
2559                 break;
2560         case TCP_INFO: {
2561                 struct tcp_info info;
2562
2563                 if (get_user(len, optlen))
2564                         return -EFAULT;
2565
2566                 tcp_get_info(sk, &info);
2567
2568                 len = min_t(unsigned int, len, sizeof(info));
2569                 if (put_user(len, optlen))
2570                         return -EFAULT;
2571                 if (copy_to_user(optval, &info, len))
2572                         return -EFAULT;
2573                 return 0;
2574         }
2575         case TCP_QUICKACK:
2576                 val = !icsk->icsk_ack.pingpong;
2577                 break;
2578
2579         case TCP_CONGESTION:
2580                 if (get_user(len, optlen))
2581                         return -EFAULT;
2582                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2583                 if (put_user(len, optlen))
2584                         return -EFAULT;
2585                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2586                         return -EFAULT;
2587                 return 0;
2588
2589         case TCP_COOKIE_TRANSACTIONS: {
2590                 struct tcp_cookie_transactions ctd;
2591                 struct tcp_cookie_values *cvp = tp->cookie_values;
2592
2593                 if (get_user(len, optlen))
2594                         return -EFAULT;
2595                 if (len < sizeof(ctd))
2596                         return -EINVAL;
2597
2598                 memset(&ctd, 0, sizeof(ctd));
2599                 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2600                                    TCP_COOKIE_IN_ALWAYS : 0)
2601                                 | (tp->rx_opt.cookie_out_never ?
2602                                    TCP_COOKIE_OUT_NEVER : 0);
2603
2604                 if (cvp != NULL) {
2605                         ctd.tcpct_flags |= (cvp->s_data_in ?
2606                                             TCP_S_DATA_IN : 0)
2607                                          | (cvp->s_data_out ?
2608                                             TCP_S_DATA_OUT : 0);
2609
2610                         ctd.tcpct_cookie_desired = cvp->cookie_desired;
2611                         ctd.tcpct_s_data_desired = cvp->s_data_desired;
2612
2613                         memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2614                                cvp->cookie_pair_size);
2615                         ctd.tcpct_used = cvp->cookie_pair_size;
2616                 }
2617
2618                 if (put_user(sizeof(ctd), optlen))
2619                         return -EFAULT;
2620                 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2621                         return -EFAULT;
2622                 return 0;
2623         }
2624         case TCP_THIN_LINEAR_TIMEOUTS:
2625                 val = tp->thin_lto;
2626                 break;
2627         case TCP_THIN_DUPACK:
2628                 val = tp->thin_dupack;
2629                 break;
2630
2631         case TCP_USER_TIMEOUT:
2632                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2633                 break;
2634         default:
2635                 return -ENOPROTOOPT;
2636         }
2637
2638         if (put_user(len, optlen))
2639                 return -EFAULT;
2640         if (copy_to_user(optval, &val, len))
2641                 return -EFAULT;
2642         return 0;
2643 }
2644
2645 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2646                    int __user *optlen)
2647 {
2648         struct inet_connection_sock *icsk = inet_csk(sk);
2649
2650         if (level != SOL_TCP)
2651                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2652                                                      optval, optlen);
2653         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2654 }
2655 EXPORT_SYMBOL(tcp_getsockopt);
2656
2657 #ifdef CONFIG_COMPAT
2658 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2659                           char __user *optval, int __user *optlen)
2660 {
2661         if (level != SOL_TCP)
2662                 return inet_csk_compat_getsockopt(sk, level, optname,
2663                                                   optval, optlen);
2664         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2665 }
2666 EXPORT_SYMBOL(compat_tcp_getsockopt);
2667 #endif
2668
2669 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, u32 features)
2670 {
2671         struct sk_buff *segs = ERR_PTR(-EINVAL);
2672         struct tcphdr *th;
2673         unsigned thlen;
2674         unsigned int seq;
2675         __be32 delta;
2676         unsigned int oldlen;
2677         unsigned int mss;
2678
2679         if (!pskb_may_pull(skb, sizeof(*th)))
2680                 goto out;
2681
2682         th = tcp_hdr(skb);
2683         thlen = th->doff * 4;
2684         if (thlen < sizeof(*th))
2685                 goto out;
2686
2687         if (!pskb_may_pull(skb, thlen))
2688                 goto out;
2689
2690         oldlen = (u16)~skb->len;
2691         __skb_pull(skb, thlen);
2692
2693         mss = skb_shinfo(skb)->gso_size;
2694         if (unlikely(skb->len <= mss))
2695                 goto out;
2696
2697         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2698                 /* Packet is from an untrusted source, reset gso_segs. */
2699                 int type = skb_shinfo(skb)->gso_type;
2700
2701                 if (unlikely(type &
2702                              ~(SKB_GSO_TCPV4 |
2703                                SKB_GSO_DODGY |
2704                                SKB_GSO_TCP_ECN |
2705                                SKB_GSO_TCPV6 |
2706                                0) ||
2707                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2708                         goto out;
2709
2710                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2711
2712                 segs = NULL;
2713                 goto out;
2714         }
2715
2716         segs = skb_segment(skb, features);
2717         if (IS_ERR(segs))
2718                 goto out;
2719
2720         delta = htonl(oldlen + (thlen + mss));
2721
2722         skb = segs;
2723         th = tcp_hdr(skb);
2724         seq = ntohl(th->seq);
2725
2726         do {
2727                 th->fin = th->psh = 0;
2728
2729                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2730                                        (__force u32)delta));
2731                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2732                         th->check =
2733                              csum_fold(csum_partial(skb_transport_header(skb),
2734                                                     thlen, skb->csum));
2735
2736                 seq += mss;
2737                 skb = skb->next;
2738                 th = tcp_hdr(skb);
2739
2740                 th->seq = htonl(seq);
2741                 th->cwr = 0;
2742         } while (skb->next);
2743
2744         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2745                       skb->data_len);
2746         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2747                                 (__force u32)delta));
2748         if (skb->ip_summed != CHECKSUM_PARTIAL)
2749                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2750                                                    thlen, skb->csum));
2751
2752 out:
2753         return segs;
2754 }
2755 EXPORT_SYMBOL(tcp_tso_segment);
2756
2757 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2758 {
2759         struct sk_buff **pp = NULL;
2760         struct sk_buff *p;
2761         struct tcphdr *th;
2762         struct tcphdr *th2;
2763         unsigned int len;
2764         unsigned int thlen;
2765         __be32 flags;
2766         unsigned int mss = 1;
2767         unsigned int hlen;
2768         unsigned int off;
2769         int flush = 1;
2770         int i;
2771
2772         off = skb_gro_offset(skb);
2773         hlen = off + sizeof(*th);
2774         th = skb_gro_header_fast(skb, off);
2775         if (skb_gro_header_hard(skb, hlen)) {
2776                 th = skb_gro_header_slow(skb, hlen, off);
2777                 if (unlikely(!th))
2778                         goto out;
2779         }
2780
2781         thlen = th->doff * 4;
2782         if (thlen < sizeof(*th))
2783                 goto out;
2784
2785         hlen = off + thlen;
2786         if (skb_gro_header_hard(skb, hlen)) {
2787                 th = skb_gro_header_slow(skb, hlen, off);
2788                 if (unlikely(!th))
2789                         goto out;
2790         }
2791
2792         skb_gro_pull(skb, thlen);
2793
2794         len = skb_gro_len(skb);
2795         flags = tcp_flag_word(th);
2796
2797         for (; (p = *head); head = &p->next) {
2798                 if (!NAPI_GRO_CB(p)->same_flow)
2799                         continue;
2800
2801                 th2 = tcp_hdr(p);
2802
2803                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2804                         NAPI_GRO_CB(p)->same_flow = 0;
2805                         continue;
2806                 }
2807
2808                 goto found;
2809         }
2810
2811         goto out_check_final;
2812
2813 found:
2814         flush = NAPI_GRO_CB(p)->flush;
2815         flush |= (__force int)(flags & TCP_FLAG_CWR);
2816         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2817                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2818         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2819         for (i = sizeof(*th); i < thlen; i += 4)
2820                 flush |= *(u32 *)((u8 *)th + i) ^
2821                          *(u32 *)((u8 *)th2 + i);
2822
2823         mss = skb_shinfo(p)->gso_size;
2824
2825         flush |= (len - 1) >= mss;
2826         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2827
2828         if (flush || skb_gro_receive(head, skb)) {
2829                 mss = 1;
2830                 goto out_check_final;
2831         }
2832
2833         p = *head;
2834         th2 = tcp_hdr(p);
2835         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2836
2837 out_check_final:
2838         flush = len < mss;
2839         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2840                                         TCP_FLAG_RST | TCP_FLAG_SYN |
2841                                         TCP_FLAG_FIN));
2842
2843         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2844                 pp = head;
2845
2846 out:
2847         NAPI_GRO_CB(skb)->flush |= flush;
2848
2849         return pp;
2850 }
2851 EXPORT_SYMBOL(tcp_gro_receive);
2852
2853 int tcp_gro_complete(struct sk_buff *skb)
2854 {
2855         struct tcphdr *th = tcp_hdr(skb);
2856
2857         skb->csum_start = skb_transport_header(skb) - skb->head;
2858         skb->csum_offset = offsetof(struct tcphdr, check);
2859         skb->ip_summed = CHECKSUM_PARTIAL;
2860
2861         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2862
2863         if (th->cwr)
2864                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2865
2866         return 0;
2867 }
2868 EXPORT_SYMBOL(tcp_gro_complete);
2869
2870 #ifdef CONFIG_TCP_MD5SIG
2871 static unsigned long tcp_md5sig_users;
2872 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool;
2873 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2874
2875 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool)
2876 {
2877         int cpu;
2878         for_each_possible_cpu(cpu) {
2879                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2880                 if (p) {
2881                         if (p->md5_desc.tfm)
2882                                 crypto_free_hash(p->md5_desc.tfm);
2883                         kfree(p);
2884                 }
2885         }
2886         free_percpu(pool);
2887 }
2888
2889 void tcp_free_md5sig_pool(void)
2890 {
2891         struct tcp_md5sig_pool * __percpu *pool = NULL;
2892
2893         spin_lock_bh(&tcp_md5sig_pool_lock);
2894         if (--tcp_md5sig_users == 0) {
2895                 pool = tcp_md5sig_pool;
2896                 tcp_md5sig_pool = NULL;
2897         }
2898         spin_unlock_bh(&tcp_md5sig_pool_lock);
2899         if (pool)
2900                 __tcp_free_md5sig_pool(pool);
2901 }
2902 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2903
2904 static struct tcp_md5sig_pool * __percpu *
2905 __tcp_alloc_md5sig_pool(struct sock *sk)
2906 {
2907         int cpu;
2908         struct tcp_md5sig_pool * __percpu *pool;
2909
2910         pool = alloc_percpu(struct tcp_md5sig_pool *);
2911         if (!pool)
2912                 return NULL;
2913
2914         for_each_possible_cpu(cpu) {
2915                 struct tcp_md5sig_pool *p;
2916                 struct crypto_hash *hash;
2917
2918                 p = kzalloc(sizeof(*p), sk->sk_allocation);
2919                 if (!p)
2920                         goto out_free;
2921                 *per_cpu_ptr(pool, cpu) = p;
2922
2923                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2924                 if (!hash || IS_ERR(hash))
2925                         goto out_free;
2926
2927                 p->md5_desc.tfm = hash;
2928         }
2929         return pool;
2930 out_free:
2931         __tcp_free_md5sig_pool(pool);
2932         return NULL;
2933 }
2934
2935 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2936 {
2937         struct tcp_md5sig_pool * __percpu *pool;
2938         int alloc = 0;
2939
2940 retry:
2941         spin_lock_bh(&tcp_md5sig_pool_lock);
2942         pool = tcp_md5sig_pool;
2943         if (tcp_md5sig_users++ == 0) {
2944                 alloc = 1;
2945                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2946         } else if (!pool) {
2947                 tcp_md5sig_users--;
2948                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2949                 cpu_relax();
2950                 goto retry;
2951         } else
2952                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2953
2954         if (alloc) {
2955                 /* we cannot hold spinlock here because this may sleep. */
2956                 struct tcp_md5sig_pool * __percpu *p;
2957
2958                 p = __tcp_alloc_md5sig_pool(sk);
2959                 spin_lock_bh(&tcp_md5sig_pool_lock);
2960                 if (!p) {
2961                         tcp_md5sig_users--;
2962                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2963                         return NULL;
2964                 }
2965                 pool = tcp_md5sig_pool;
2966                 if (pool) {
2967                         /* oops, it has already been assigned. */
2968                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2969                         __tcp_free_md5sig_pool(p);
2970                 } else {
2971                         tcp_md5sig_pool = pool = p;
2972                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2973                 }
2974         }
2975         return pool;
2976 }
2977 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2978
2979
2980 /**
2981  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2982  *
2983  *      We use percpu structure, so if we succeed, we exit with preemption
2984  *      and BH disabled, to make sure another thread or softirq handling
2985  *      wont try to get same context.
2986  */
2987 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2988 {
2989         struct tcp_md5sig_pool * __percpu *p;
2990
2991         local_bh_disable();
2992
2993         spin_lock(&tcp_md5sig_pool_lock);
2994         p = tcp_md5sig_pool;
2995         if (p)
2996                 tcp_md5sig_users++;
2997         spin_unlock(&tcp_md5sig_pool_lock);
2998
2999         if (p)
3000                 return *this_cpu_ptr(p);
3001
3002         local_bh_enable();
3003         return NULL;
3004 }
3005 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3006
3007 void tcp_put_md5sig_pool(void)
3008 {
3009         local_bh_enable();
3010         tcp_free_md5sig_pool();
3011 }
3012 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3013
3014 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3015                         struct tcphdr *th)
3016 {
3017         struct scatterlist sg;
3018         int err;
3019
3020         __sum16 old_checksum = th->check;
3021         th->check = 0;
3022         /* options aren't included in the hash */
3023         sg_init_one(&sg, th, sizeof(struct tcphdr));
3024         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
3025         th->check = old_checksum;
3026         return err;
3027 }
3028 EXPORT_SYMBOL(tcp_md5_hash_header);
3029
3030 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3031                           struct sk_buff *skb, unsigned header_len)
3032 {
3033         struct scatterlist sg;
3034         const struct tcphdr *tp = tcp_hdr(skb);
3035         struct hash_desc *desc = &hp->md5_desc;
3036         unsigned i;
3037         const unsigned head_data_len = skb_headlen(skb) > header_len ?
3038                                        skb_headlen(skb) - header_len : 0;
3039         const struct skb_shared_info *shi = skb_shinfo(skb);
3040         struct sk_buff *frag_iter;
3041
3042         sg_init_table(&sg, 1);
3043
3044         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3045         if (crypto_hash_update(desc, &sg, head_data_len))
3046                 return 1;
3047
3048         for (i = 0; i < shi->nr_frags; ++i) {
3049                 const struct skb_frag_struct *f = &shi->frags[i];
3050                 sg_set_page(&sg, f->page, f->size, f->page_offset);
3051                 if (crypto_hash_update(desc, &sg, f->size))
3052                         return 1;
3053         }
3054
3055         skb_walk_frags(skb, frag_iter)
3056                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3057                         return 1;
3058
3059         return 0;
3060 }
3061 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3062
3063 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
3064 {
3065         struct scatterlist sg;
3066
3067         sg_init_one(&sg, key->key, key->keylen);
3068         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3069 }
3070 EXPORT_SYMBOL(tcp_md5_hash_key);
3071
3072 #endif
3073
3074 /**
3075  * Each Responder maintains up to two secret values concurrently for
3076  * efficient secret rollover.  Each secret value has 4 states:
3077  *
3078  * Generating.  (tcp_secret_generating != tcp_secret_primary)
3079  *    Generates new Responder-Cookies, but not yet used for primary
3080  *    verification.  This is a short-term state, typically lasting only
3081  *    one round trip time (RTT).
3082  *
3083  * Primary.  (tcp_secret_generating == tcp_secret_primary)
3084  *    Used both for generation and primary verification.
3085  *
3086  * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3087  *    Used for verification, until the first failure that can be
3088  *    verified by the newer Generating secret.  At that time, this
3089  *    cookie's state is changed to Secondary, and the Generating
3090  *    cookie's state is changed to Primary.  This is a short-term state,
3091  *    typically lasting only one round trip time (RTT).
3092  *
3093  * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3094  *    Used for secondary verification, after primary verification
3095  *    failures.  This state lasts no more than twice the Maximum Segment
3096  *    Lifetime (2MSL).  Then, the secret is discarded.
3097  */
3098 struct tcp_cookie_secret {
3099         /* The secret is divided into two parts.  The digest part is the
3100          * equivalent of previously hashing a secret and saving the state,
3101          * and serves as an initialization vector (IV).  The message part
3102          * serves as the trailing secret.
3103          */
3104         u32                             secrets[COOKIE_WORKSPACE_WORDS];
3105         unsigned long                   expires;
3106 };
3107
3108 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3109 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3110 #define TCP_SECRET_LIFE (HZ * 600)
3111
3112 static struct tcp_cookie_secret tcp_secret_one;
3113 static struct tcp_cookie_secret tcp_secret_two;
3114
3115 /* Essentially a circular list, without dynamic allocation. */
3116 static struct tcp_cookie_secret *tcp_secret_generating;
3117 static struct tcp_cookie_secret *tcp_secret_primary;
3118 static struct tcp_cookie_secret *tcp_secret_retiring;
3119 static struct tcp_cookie_secret *tcp_secret_secondary;
3120
3121 static DEFINE_SPINLOCK(tcp_secret_locker);
3122
3123 /* Select a pseudo-random word in the cookie workspace.
3124  */
3125 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3126 {
3127         return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3128 }
3129
3130 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3131  * Called in softirq context.
3132  * Returns: 0 for success.
3133  */
3134 int tcp_cookie_generator(u32 *bakery)
3135 {
3136         unsigned long jiffy = jiffies;
3137
3138         if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3139                 spin_lock_bh(&tcp_secret_locker);
3140                 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3141                         /* refreshed by another */
3142                         memcpy(bakery,
3143                                &tcp_secret_generating->secrets[0],
3144                                COOKIE_WORKSPACE_WORDS);
3145                 } else {
3146                         /* still needs refreshing */
3147                         get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3148
3149                         /* The first time, paranoia assumes that the
3150                          * randomization function isn't as strong.  But,
3151                          * this secret initialization is delayed until
3152                          * the last possible moment (packet arrival).
3153                          * Although that time is observable, it is
3154                          * unpredictably variable.  Mash in the most
3155                          * volatile clock bits available, and expire the
3156                          * secret extra quickly.
3157                          */
3158                         if (unlikely(tcp_secret_primary->expires ==
3159                                      tcp_secret_secondary->expires)) {
3160                                 struct timespec tv;
3161
3162                                 getnstimeofday(&tv);
3163                                 bakery[COOKIE_DIGEST_WORDS+0] ^=
3164                                         (u32)tv.tv_nsec;
3165
3166                                 tcp_secret_secondary->expires = jiffy
3167                                         + TCP_SECRET_1MSL
3168                                         + (0x0f & tcp_cookie_work(bakery, 0));
3169                         } else {
3170                                 tcp_secret_secondary->expires = jiffy
3171                                         + TCP_SECRET_LIFE
3172                                         + (0xff & tcp_cookie_work(bakery, 1));
3173                                 tcp_secret_primary->expires = jiffy
3174                                         + TCP_SECRET_2MSL
3175                                         + (0x1f & tcp_cookie_work(bakery, 2));
3176                         }
3177                         memcpy(&tcp_secret_secondary->secrets[0],
3178                                bakery, COOKIE_WORKSPACE_WORDS);
3179
3180                         rcu_assign_pointer(tcp_secret_generating,
3181                                            tcp_secret_secondary);
3182                         rcu_assign_pointer(tcp_secret_retiring,
3183                                            tcp_secret_primary);
3184                         /*
3185                          * Neither call_rcu() nor synchronize_rcu() needed.
3186                          * Retiring data is not freed.  It is replaced after
3187                          * further (locked) pointer updates, and a quiet time
3188                          * (minimum 1MSL, maximum LIFE - 2MSL).
3189                          */
3190                 }
3191                 spin_unlock_bh(&tcp_secret_locker);
3192         } else {
3193                 rcu_read_lock_bh();
3194                 memcpy(bakery,
3195                        &rcu_dereference(tcp_secret_generating)->secrets[0],
3196                        COOKIE_WORKSPACE_WORDS);
3197                 rcu_read_unlock_bh();
3198         }
3199         return 0;
3200 }
3201 EXPORT_SYMBOL(tcp_cookie_generator);
3202
3203 void tcp_done(struct sock *sk)
3204 {
3205         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3206                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3207
3208         tcp_set_state(sk, TCP_CLOSE);
3209         tcp_clear_xmit_timers(sk);
3210
3211         sk->sk_shutdown = SHUTDOWN_MASK;
3212
3213         if (!sock_flag(sk, SOCK_DEAD))
3214                 sk->sk_state_change(sk);
3215         else
3216                 inet_csk_destroy_sock(sk);
3217 }
3218 EXPORT_SYMBOL_GPL(tcp_done);
3219
3220 extern struct tcp_congestion_ops tcp_reno;
3221
3222 static __initdata unsigned long thash_entries;
3223 static int __init set_thash_entries(char *str)
3224 {
3225         if (!str)
3226                 return 0;
3227         thash_entries = simple_strtoul(str, &str, 0);
3228         return 1;
3229 }
3230 __setup("thash_entries=", set_thash_entries);
3231
3232 void __init tcp_init(void)
3233 {
3234         struct sk_buff *skb = NULL;
3235         unsigned long limit;
3236         int i, max_share, cnt;
3237         unsigned long jiffy = jiffies;
3238
3239         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3240
3241         percpu_counter_init(&tcp_sockets_allocated, 0);
3242         percpu_counter_init(&tcp_orphan_count, 0);
3243         tcp_hashinfo.bind_bucket_cachep =
3244                 kmem_cache_create("tcp_bind_bucket",
3245                                   sizeof(struct inet_bind_bucket), 0,
3246                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3247
3248         /* Size and allocate the main established and bind bucket
3249          * hash tables.
3250          *
3251          * The methodology is similar to that of the buffer cache.
3252          */
3253         tcp_hashinfo.ehash =
3254                 alloc_large_system_hash("TCP established",
3255                                         sizeof(struct inet_ehash_bucket),
3256                                         thash_entries,
3257                                         (totalram_pages >= 128 * 1024) ?
3258                                         13 : 15,
3259                                         0,
3260                                         NULL,
3261                                         &tcp_hashinfo.ehash_mask,
3262                                         thash_entries ? 0 : 512 * 1024);
3263         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3264                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3265                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3266         }
3267         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3268                 panic("TCP: failed to alloc ehash_locks");
3269         tcp_hashinfo.bhash =
3270                 alloc_large_system_hash("TCP bind",
3271                                         sizeof(struct inet_bind_hashbucket),
3272                                         tcp_hashinfo.ehash_mask + 1,
3273                                         (totalram_pages >= 128 * 1024) ?
3274                                         13 : 15,
3275                                         0,
3276                                         &tcp_hashinfo.bhash_size,
3277                                         NULL,
3278                                         64 * 1024);
3279         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3280         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3281                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3282                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3283         }
3284
3285
3286         cnt = tcp_hashinfo.ehash_mask + 1;
3287
3288         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3289         sysctl_tcp_max_orphans = cnt / 2;
3290         sysctl_max_syn_backlog = max(128, cnt / 256);
3291
3292         limit = nr_free_buffer_pages() / 8;
3293         limit = max(limit, 128UL);
3294         sysctl_tcp_mem[0] = limit / 4 * 3;
3295         sysctl_tcp_mem[1] = limit;
3296         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3297
3298         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3299         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3300         max_share = min(4UL*1024*1024, limit);
3301
3302         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3303         sysctl_tcp_wmem[1] = 16*1024;
3304         sysctl_tcp_wmem[2] = max(64*1024, max_share);
3305
3306         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3307         sysctl_tcp_rmem[1] = 87380;
3308         sysctl_tcp_rmem[2] = max(87380, max_share);
3309
3310         printk(KERN_INFO "TCP: Hash tables configured "
3311                "(established %u bind %u)\n",
3312                tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3313
3314         tcp_register_congestion_control(&tcp_reno);
3315
3316         memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3317         memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3318         tcp_secret_one.expires = jiffy; /* past due */
3319         tcp_secret_two.expires = jiffy; /* past due */
3320         tcp_secret_generating = &tcp_secret_one;
3321         tcp_secret_primary = &tcp_secret_one;
3322         tcp_secret_retiring = &tcp_secret_two;
3323         tcp_secret_secondary = &tcp_secret_two;
3324 }