tcp: Fix OOB POLLIN avoidance.
[linux-2.6.git] / net / ipv4 / tcp.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Authors:     Ross Biro
9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
12  *              Florian La Roche, <flla@stud.uni-sb.de>
13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *              Jorge Cwik, <jorge@laser.satlink.net>
19  *
20  * Fixes:
21  *              Alan Cox        :       Numerous verify_area() calls
22  *              Alan Cox        :       Set the ACK bit on a reset
23  *              Alan Cox        :       Stopped it crashing if it closed while
24  *                                      sk->inuse=1 and was trying to connect
25  *                                      (tcp_err()).
26  *              Alan Cox        :       All icmp error handling was broken
27  *                                      pointers passed where wrong and the
28  *                                      socket was looked up backwards. Nobody
29  *                                      tested any icmp error code obviously.
30  *              Alan Cox        :       tcp_err() now handled properly. It
31  *                                      wakes people on errors. poll
32  *                                      behaves and the icmp error race
33  *                                      has gone by moving it into sock.c
34  *              Alan Cox        :       tcp_send_reset() fixed to work for
35  *                                      everything not just packets for
36  *                                      unknown sockets.
37  *              Alan Cox        :       tcp option processing.
38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
39  *                                      syn rule wrong]
40  *              Herp Rosmanith  :       More reset fixes
41  *              Alan Cox        :       No longer acks invalid rst frames.
42  *                                      Acking any kind of RST is right out.
43  *              Alan Cox        :       Sets an ignore me flag on an rst
44  *                                      receive otherwise odd bits of prattle
45  *                                      escape still
46  *              Alan Cox        :       Fixed another acking RST frame bug.
47  *                                      Should stop LAN workplace lockups.
48  *              Alan Cox        :       Some tidyups using the new skb list
49  *                                      facilities
50  *              Alan Cox        :       sk->keepopen now seems to work
51  *              Alan Cox        :       Pulls options out correctly on accepts
52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
54  *                                      bit to skb ops.
55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
56  *                                      nasty.
57  *              Alan Cox        :       Added some better commenting, as the
58  *                                      tcp is hard to follow
59  *              Alan Cox        :       Removed incorrect check for 20 * psh
60  *      Michael O'Reilly        :       ack < copied bug fix.
61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
62  *              Alan Cox        :       FIN with no memory -> CRASH
63  *              Alan Cox        :       Added socket option proto entries.
64  *                                      Also added awareness of them to accept.
65  *              Alan Cox        :       Added TCP options (SOL_TCP)
66  *              Alan Cox        :       Switched wakeup calls to callbacks,
67  *                                      so the kernel can layer network
68  *                                      sockets.
69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
71  *              Alan Cox        :       RST frames sent on unsynchronised
72  *                                      state ack error.
73  *              Alan Cox        :       Put in missing check for SYN bit.
74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
75  *                                      window non shrink trick.
76  *              Alan Cox        :       Added a couple of small NET2E timer
77  *                                      fixes
78  *              Charles Hedrick :       TCP fixes
79  *              Toomas Tamm     :       TCP window fixes
80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
81  *              Charles Hedrick :       Rewrote most of it to actually work
82  *              Linus           :       Rewrote tcp_read() and URG handling
83  *                                      completely
84  *              Gerhard Koerting:       Fixed some missing timer handling
85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
86  *              Gerhard Koerting:       PC/TCP workarounds
87  *              Adam Caldwell   :       Assorted timer/timing errors
88  *              Matthew Dillon  :       Fixed another RST bug
89  *              Alan Cox        :       Move to kernel side addressing changes.
90  *              Alan Cox        :       Beginning work on TCP fastpathing
91  *                                      (not yet usable)
92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
93  *              Alan Cox        :       TCP fast path debugging
94  *              Alan Cox        :       Window clamping
95  *              Michael Riepe   :       Bug in tcp_check()
96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
97  *              Matt Dillon     :       Yet more small nasties remove from the
98  *                                      TCP code (Be very nice to this man if
99  *                                      tcp finally works 100%) 8)
100  *              Alan Cox        :       BSD accept semantics.
101  *              Alan Cox        :       Reset on closedown bug.
102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
103  *              Michael Pall    :       Handle poll() after URG properly in
104  *                                      all cases.
105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
106  *                                      (multi URG PUSH broke rlogin).
107  *              Michael Pall    :       Fix the multi URG PUSH problem in
108  *                                      tcp_readable(), poll() after URG
109  *                                      works now.
110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
111  *                                      BSD api.
112  *              Alan Cox        :       Changed the semantics of sk->socket to
113  *                                      fix a race and a signal problem with
114  *                                      accept() and async I/O.
115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
118  *                                      clients/servers which listen in on
119  *                                      fixed ports.
120  *              Alan Cox        :       Cleaned the above up and shrank it to
121  *                                      a sensible code size.
122  *              Alan Cox        :       Self connect lockup fix.
123  *              Alan Cox        :       No connect to multicast.
124  *              Ross Biro       :       Close unaccepted children on master
125  *                                      socket close.
126  *              Alan Cox        :       Reset tracing code.
127  *              Alan Cox        :       Spurious resets on shutdown.
128  *              Alan Cox        :       Giant 15 minute/60 second timer error
129  *              Alan Cox        :       Small whoops in polling before an
130  *                                      accept.
131  *              Alan Cox        :       Kept the state trace facility since
132  *                                      it's handy for debugging.
133  *              Alan Cox        :       More reset handler fixes.
134  *              Alan Cox        :       Started rewriting the code based on
135  *                                      the RFC's for other useful protocol
136  *                                      references see: Comer, KA9Q NOS, and
137  *                                      for a reference on the difference
138  *                                      between specifications and how BSD
139  *                                      works see the 4.4lite source.
140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
141  *                                      close.
142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
144  *              Alan Cox        :       Reimplemented timers as per the RFC
145  *                                      and using multiple timers for sanity.
146  *              Alan Cox        :       Small bug fixes, and a lot of new
147  *                                      comments.
148  *              Alan Cox        :       Fixed dual reader crash by locking
149  *                                      the buffers (much like datagram.c)
150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
151  *                                      now gets fed up of retrying without
152  *                                      (even a no space) answer.
153  *              Alan Cox        :       Extracted closing code better
154  *              Alan Cox        :       Fixed the closing state machine to
155  *                                      resemble the RFC.
156  *              Alan Cox        :       More 'per spec' fixes.
157  *              Jorge Cwik      :       Even faster checksumming.
158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
159  *                                      only frames. At least one pc tcp stack
160  *                                      generates them.
161  *              Alan Cox        :       Cache last socket.
162  *              Alan Cox        :       Per route irtt.
163  *              Matt Day        :       poll()->select() match BSD precisely on error
164  *              Alan Cox        :       New buffers
165  *              Marc Tamsky     :       Various sk->prot->retransmits and
166  *                                      sk->retransmits misupdating fixed.
167  *                                      Fixed tcp_write_timeout: stuck close,
168  *                                      and TCP syn retries gets used now.
169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
170  *                                      ack if state is TCP_CLOSED.
171  *              Alan Cox        :       Look up device on a retransmit - routes may
172  *                                      change. Doesn't yet cope with MSS shrink right
173  *                                      but it's a start!
174  *              Marc Tamsky     :       Closing in closing fixes.
175  *              Mike Shaver     :       RFC1122 verifications.
176  *              Alan Cox        :       rcv_saddr errors.
177  *              Alan Cox        :       Block double connect().
178  *              Alan Cox        :       Small hooks for enSKIP.
179  *              Alexey Kuznetsov:       Path MTU discovery.
180  *              Alan Cox        :       Support soft errors.
181  *              Alan Cox        :       Fix MTU discovery pathological case
182  *                                      when the remote claims no mtu!
183  *              Marc Tamsky     :       TCP_CLOSE fix.
184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
185  *                                      window but wrong (fixes NT lpd problems)
186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
187  *              Joerg Reuter    :       No modification of locked buffers in
188  *                                      tcp_do_retransmit()
189  *              Eric Schenk     :       Changed receiver side silly window
190  *                                      avoidance algorithm to BSD style
191  *                                      algorithm. This doubles throughput
192  *                                      against machines running Solaris,
193  *                                      and seems to result in general
194  *                                      improvement.
195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
196  *      Willy Konynenberg       :       Transparent proxying support.
197  *      Mike McLagan            :       Routing by source
198  *              Keith Owens     :       Do proper merging with partial SKB's in
199  *                                      tcp_do_sendmsg to avoid burstiness.
200  *              Eric Schenk     :       Fix fast close down bug with
201  *                                      shutdown() followed by close().
202  *              Andi Kleen      :       Make poll agree with SIGIO
203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
204  *                                      lingertime == 0 (RFC 793 ABORT Call)
205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
206  *                                      csum_and_copy_from_user() if possible.
207  *
208  *              This program is free software; you can redistribute it and/or
209  *              modify it under the terms of the GNU General Public License
210  *              as published by the Free Software Foundation; either version
211  *              2 of the License, or(at your option) any later version.
212  *
213  * Description of States:
214  *
215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
216  *
217  *      TCP_SYN_RECV            received a connection request, sent ack,
218  *                              waiting for final ack in three-way handshake.
219  *
220  *      TCP_ESTABLISHED         connection established
221  *
222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
223  *                              transmission of remaining buffered data
224  *
225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
226  *                              to shutdown
227  *
228  *      TCP_CLOSING             both sides have shutdown but we still have
229  *                              data we have to finish sending
230  *
231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
232  *                              closed, can only be entered from FIN_WAIT2
233  *                              or CLOSING.  Required because the other end
234  *                              may not have gotten our last ACK causing it
235  *                              to retransmit the data packet (which we ignore)
236  *
237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
238  *                              us to finish writing our data and to shutdown
239  *                              (we have to close() to move on to LAST_ACK)
240  *
241  *      TCP_LAST_ACK            out side has shutdown after remote has
242  *                              shutdown.  There may still be data in our
243  *                              buffer that we have to finish sending
244  *
245  *      TCP_CLOSE               socket is finished
246  */
247
248 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
267 #include <linux/time.h>
268
269 #include <net/icmp.h>
270 #include <net/tcp.h>
271 #include <net/xfrm.h>
272 #include <net/ip.h>
273 #include <net/netdma.h>
274 #include <net/sock.h>
275
276 #include <asm/uaccess.h>
277 #include <asm/ioctls.h>
278
279 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
280
281 struct percpu_counter tcp_orphan_count;
282 EXPORT_SYMBOL_GPL(tcp_orphan_count);
283
284 int sysctl_tcp_mem[3] __read_mostly;
285 int sysctl_tcp_wmem[3] __read_mostly;
286 int sysctl_tcp_rmem[3] __read_mostly;
287
288 EXPORT_SYMBOL(sysctl_tcp_mem);
289 EXPORT_SYMBOL(sysctl_tcp_rmem);
290 EXPORT_SYMBOL(sysctl_tcp_wmem);
291
292 atomic_t tcp_memory_allocated;  /* Current allocated memory. */
293 EXPORT_SYMBOL(tcp_memory_allocated);
294
295 /*
296  * Current number of TCP sockets.
297  */
298 struct percpu_counter tcp_sockets_allocated;
299 EXPORT_SYMBOL(tcp_sockets_allocated);
300
301 /*
302  * TCP splice context
303  */
304 struct tcp_splice_state {
305         struct pipe_inode_info *pipe;
306         size_t len;
307         unsigned int flags;
308 };
309
310 /*
311  * Pressure flag: try to collapse.
312  * Technical note: it is used by multiple contexts non atomically.
313  * All the __sk_mem_schedule() is of this nature: accounting
314  * is strict, actions are advisory and have some latency.
315  */
316 int tcp_memory_pressure __read_mostly;
317
318 EXPORT_SYMBOL(tcp_memory_pressure);
319
320 void tcp_enter_memory_pressure(struct sock *sk)
321 {
322         if (!tcp_memory_pressure) {
323                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
324                 tcp_memory_pressure = 1;
325         }
326 }
327
328 EXPORT_SYMBOL(tcp_enter_memory_pressure);
329
330 /* Convert seconds to retransmits based on initial and max timeout */
331 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
332 {
333         u8 res = 0;
334
335         if (seconds > 0) {
336                 int period = timeout;
337
338                 res = 1;
339                 while (seconds > period && res < 255) {
340                         res++;
341                         timeout <<= 1;
342                         if (timeout > rto_max)
343                                 timeout = rto_max;
344                         period += timeout;
345                 }
346         }
347         return res;
348 }
349
350 /* Convert retransmits to seconds based on initial and max timeout */
351 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
352 {
353         int period = 0;
354
355         if (retrans > 0) {
356                 period = timeout;
357                 while (--retrans) {
358                         timeout <<= 1;
359                         if (timeout > rto_max)
360                                 timeout = rto_max;
361                         period += timeout;
362                 }
363         }
364         return period;
365 }
366
367 /*
368  *      Wait for a TCP event.
369  *
370  *      Note that we don't need to lock the socket, as the upper poll layers
371  *      take care of normal races (between the test and the event) and we don't
372  *      go look at any of the socket buffers directly.
373  */
374 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
375 {
376         unsigned int mask;
377         struct sock *sk = sock->sk;
378         struct tcp_sock *tp = tcp_sk(sk);
379
380         sock_poll_wait(file, sk->sk_sleep, wait);
381         if (sk->sk_state == TCP_LISTEN)
382                 return inet_csk_listen_poll(sk);
383
384         /* Socket is not locked. We are protected from async events
385          * by poll logic and correct handling of state changes
386          * made by other threads is impossible in any case.
387          */
388
389         mask = 0;
390         if (sk->sk_err)
391                 mask = POLLERR;
392
393         /*
394          * POLLHUP is certainly not done right. But poll() doesn't
395          * have a notion of HUP in just one direction, and for a
396          * socket the read side is more interesting.
397          *
398          * Some poll() documentation says that POLLHUP is incompatible
399          * with the POLLOUT/POLLWR flags, so somebody should check this
400          * all. But careful, it tends to be safer to return too many
401          * bits than too few, and you can easily break real applications
402          * if you don't tell them that something has hung up!
403          *
404          * Check-me.
405          *
406          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
407          * our fs/select.c). It means that after we received EOF,
408          * poll always returns immediately, making impossible poll() on write()
409          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
410          * if and only if shutdown has been made in both directions.
411          * Actually, it is interesting to look how Solaris and DUX
412          * solve this dilemma. I would prefer, if POLLHUP were maskable,
413          * then we could set it on SND_SHUTDOWN. BTW examples given
414          * in Stevens' books assume exactly this behaviour, it explains
415          * why POLLHUP is incompatible with POLLOUT.    --ANK
416          *
417          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
418          * blocking on fresh not-connected or disconnected socket. --ANK
419          */
420         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
421                 mask |= POLLHUP;
422         if (sk->sk_shutdown & RCV_SHUTDOWN)
423                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
424
425         /* Connected? */
426         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
427                 int target = sock_rcvlowat(sk, 0, INT_MAX);
428
429                 if (tp->urg_seq == tp->copied_seq &&
430                     !sock_flag(sk, SOCK_URGINLINE) &&
431                     tp->urg_data)
432                         target++;
433
434                 /* Potential race condition. If read of tp below will
435                  * escape above sk->sk_state, we can be illegally awaken
436                  * in SYN_* states. */
437                 if (tp->rcv_nxt - tp->copied_seq >= target)
438                         mask |= POLLIN | POLLRDNORM;
439
440                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
441                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
442                                 mask |= POLLOUT | POLLWRNORM;
443                         } else {  /* send SIGIO later */
444                                 set_bit(SOCK_ASYNC_NOSPACE,
445                                         &sk->sk_socket->flags);
446                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
447
448                                 /* Race breaker. If space is freed after
449                                  * wspace test but before the flags are set,
450                                  * IO signal will be lost.
451                                  */
452                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
453                                         mask |= POLLOUT | POLLWRNORM;
454                         }
455                 }
456
457                 if (tp->urg_data & TCP_URG_VALID)
458                         mask |= POLLPRI;
459         }
460         return mask;
461 }
462
463 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
464 {
465         struct tcp_sock *tp = tcp_sk(sk);
466         int answ;
467
468         switch (cmd) {
469         case SIOCINQ:
470                 if (sk->sk_state == TCP_LISTEN)
471                         return -EINVAL;
472
473                 lock_sock(sk);
474                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
475                         answ = 0;
476                 else if (sock_flag(sk, SOCK_URGINLINE) ||
477                          !tp->urg_data ||
478                          before(tp->urg_seq, tp->copied_seq) ||
479                          !before(tp->urg_seq, tp->rcv_nxt)) {
480                         struct sk_buff *skb;
481
482                         answ = tp->rcv_nxt - tp->copied_seq;
483
484                         /* Subtract 1, if FIN is in queue. */
485                         skb = skb_peek_tail(&sk->sk_receive_queue);
486                         if (answ && skb)
487                                 answ -= tcp_hdr(skb)->fin;
488                 } else
489                         answ = tp->urg_seq - tp->copied_seq;
490                 release_sock(sk);
491                 break;
492         case SIOCATMARK:
493                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
494                 break;
495         case SIOCOUTQ:
496                 if (sk->sk_state == TCP_LISTEN)
497                         return -EINVAL;
498
499                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
500                         answ = 0;
501                 else
502                         answ = tp->write_seq - tp->snd_una;
503                 break;
504         default:
505                 return -ENOIOCTLCMD;
506         }
507
508         return put_user(answ, (int __user *)arg);
509 }
510
511 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
512 {
513         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
514         tp->pushed_seq = tp->write_seq;
515 }
516
517 static inline int forced_push(struct tcp_sock *tp)
518 {
519         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
520 }
521
522 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
523 {
524         struct tcp_sock *tp = tcp_sk(sk);
525         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
526
527         skb->csum    = 0;
528         tcb->seq     = tcb->end_seq = tp->write_seq;
529         tcb->flags   = TCPCB_FLAG_ACK;
530         tcb->sacked  = 0;
531         skb_header_release(skb);
532         tcp_add_write_queue_tail(sk, skb);
533         sk->sk_wmem_queued += skb->truesize;
534         sk_mem_charge(sk, skb->truesize);
535         if (tp->nonagle & TCP_NAGLE_PUSH)
536                 tp->nonagle &= ~TCP_NAGLE_PUSH;
537 }
538
539 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
540 {
541         if (flags & MSG_OOB)
542                 tp->snd_up = tp->write_seq;
543 }
544
545 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
546                             int nonagle)
547 {
548         if (tcp_send_head(sk)) {
549                 struct tcp_sock *tp = tcp_sk(sk);
550
551                 if (!(flags & MSG_MORE) || forced_push(tp))
552                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
553
554                 tcp_mark_urg(tp, flags);
555                 __tcp_push_pending_frames(sk, mss_now,
556                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
557         }
558 }
559
560 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
561                                 unsigned int offset, size_t len)
562 {
563         struct tcp_splice_state *tss = rd_desc->arg.data;
564         int ret;
565
566         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
567                               tss->flags);
568         if (ret > 0)
569                 rd_desc->count -= ret;
570         return ret;
571 }
572
573 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
574 {
575         /* Store TCP splice context information in read_descriptor_t. */
576         read_descriptor_t rd_desc = {
577                 .arg.data = tss,
578                 .count    = tss->len,
579         };
580
581         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
582 }
583
584 /**
585  *  tcp_splice_read - splice data from TCP socket to a pipe
586  * @sock:       socket to splice from
587  * @ppos:       position (not valid)
588  * @pipe:       pipe to splice to
589  * @len:        number of bytes to splice
590  * @flags:      splice modifier flags
591  *
592  * Description:
593  *    Will read pages from given socket and fill them into a pipe.
594  *
595  **/
596 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
597                         struct pipe_inode_info *pipe, size_t len,
598                         unsigned int flags)
599 {
600         struct sock *sk = sock->sk;
601         struct tcp_splice_state tss = {
602                 .pipe = pipe,
603                 .len = len,
604                 .flags = flags,
605         };
606         long timeo;
607         ssize_t spliced;
608         int ret;
609
610         /*
611          * We can't seek on a socket input
612          */
613         if (unlikely(*ppos))
614                 return -ESPIPE;
615
616         ret = spliced = 0;
617
618         lock_sock(sk);
619
620         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
621         while (tss.len) {
622                 ret = __tcp_splice_read(sk, &tss);
623                 if (ret < 0)
624                         break;
625                 else if (!ret) {
626                         if (spliced)
627                                 break;
628                         if (sock_flag(sk, SOCK_DONE))
629                                 break;
630                         if (sk->sk_err) {
631                                 ret = sock_error(sk);
632                                 break;
633                         }
634                         if (sk->sk_shutdown & RCV_SHUTDOWN)
635                                 break;
636                         if (sk->sk_state == TCP_CLOSE) {
637                                 /*
638                                  * This occurs when user tries to read
639                                  * from never connected socket.
640                                  */
641                                 if (!sock_flag(sk, SOCK_DONE))
642                                         ret = -ENOTCONN;
643                                 break;
644                         }
645                         if (!timeo) {
646                                 ret = -EAGAIN;
647                                 break;
648                         }
649                         sk_wait_data(sk, &timeo);
650                         if (signal_pending(current)) {
651                                 ret = sock_intr_errno(timeo);
652                                 break;
653                         }
654                         continue;
655                 }
656                 tss.len -= ret;
657                 spliced += ret;
658
659                 if (!timeo)
660                         break;
661                 release_sock(sk);
662                 lock_sock(sk);
663
664                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
665                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
666                     signal_pending(current))
667                         break;
668         }
669
670         release_sock(sk);
671
672         if (spliced)
673                 return spliced;
674
675         return ret;
676 }
677
678 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
679 {
680         struct sk_buff *skb;
681
682         /* The TCP header must be at least 32-bit aligned.  */
683         size = ALIGN(size, 4);
684
685         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
686         if (skb) {
687                 if (sk_wmem_schedule(sk, skb->truesize)) {
688                         /*
689                          * Make sure that we have exactly size bytes
690                          * available to the caller, no more, no less.
691                          */
692                         skb_reserve(skb, skb_tailroom(skb) - size);
693                         return skb;
694                 }
695                 __kfree_skb(skb);
696         } else {
697                 sk->sk_prot->enter_memory_pressure(sk);
698                 sk_stream_moderate_sndbuf(sk);
699         }
700         return NULL;
701 }
702
703 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
704                                        int large_allowed)
705 {
706         struct tcp_sock *tp = tcp_sk(sk);
707         u32 xmit_size_goal, old_size_goal;
708
709         xmit_size_goal = mss_now;
710
711         if (large_allowed && sk_can_gso(sk)) {
712                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
713                                   inet_csk(sk)->icsk_af_ops->net_header_len -
714                                   inet_csk(sk)->icsk_ext_hdr_len -
715                                   tp->tcp_header_len);
716
717                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
718
719                 /* We try hard to avoid divides here */
720                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
721
722                 if (likely(old_size_goal <= xmit_size_goal &&
723                            old_size_goal + mss_now > xmit_size_goal)) {
724                         xmit_size_goal = old_size_goal;
725                 } else {
726                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
727                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
728                 }
729         }
730
731         return max(xmit_size_goal, mss_now);
732 }
733
734 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
735 {
736         int mss_now;
737
738         mss_now = tcp_current_mss(sk);
739         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
740
741         return mss_now;
742 }
743
744 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
745                          size_t psize, int flags)
746 {
747         struct tcp_sock *tp = tcp_sk(sk);
748         int mss_now, size_goal;
749         int err;
750         ssize_t copied;
751         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
752
753         /* Wait for a connection to finish. */
754         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
755                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
756                         goto out_err;
757
758         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
759
760         mss_now = tcp_send_mss(sk, &size_goal, flags);
761         copied = 0;
762
763         err = -EPIPE;
764         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
765                 goto out_err;
766
767         while (psize > 0) {
768                 struct sk_buff *skb = tcp_write_queue_tail(sk);
769                 struct page *page = pages[poffset / PAGE_SIZE];
770                 int copy, i, can_coalesce;
771                 int offset = poffset % PAGE_SIZE;
772                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
773
774                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
775 new_segment:
776                         if (!sk_stream_memory_free(sk))
777                                 goto wait_for_sndbuf;
778
779                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
780                         if (!skb)
781                                 goto wait_for_memory;
782
783                         skb_entail(sk, skb);
784                         copy = size_goal;
785                 }
786
787                 if (copy > size)
788                         copy = size;
789
790                 i = skb_shinfo(skb)->nr_frags;
791                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
792                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
793                         tcp_mark_push(tp, skb);
794                         goto new_segment;
795                 }
796                 if (!sk_wmem_schedule(sk, copy))
797                         goto wait_for_memory;
798
799                 if (can_coalesce) {
800                         skb_shinfo(skb)->frags[i - 1].size += copy;
801                 } else {
802                         get_page(page);
803                         skb_fill_page_desc(skb, i, page, offset, copy);
804                 }
805
806                 skb->len += copy;
807                 skb->data_len += copy;
808                 skb->truesize += copy;
809                 sk->sk_wmem_queued += copy;
810                 sk_mem_charge(sk, copy);
811                 skb->ip_summed = CHECKSUM_PARTIAL;
812                 tp->write_seq += copy;
813                 TCP_SKB_CB(skb)->end_seq += copy;
814                 skb_shinfo(skb)->gso_segs = 0;
815
816                 if (!copied)
817                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
818
819                 copied += copy;
820                 poffset += copy;
821                 if (!(psize -= copy))
822                         goto out;
823
824                 if (skb->len < size_goal || (flags & MSG_OOB))
825                         continue;
826
827                 if (forced_push(tp)) {
828                         tcp_mark_push(tp, skb);
829                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
830                 } else if (skb == tcp_send_head(sk))
831                         tcp_push_one(sk, mss_now);
832                 continue;
833
834 wait_for_sndbuf:
835                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
836 wait_for_memory:
837                 if (copied)
838                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
839
840                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
841                         goto do_error;
842
843                 mss_now = tcp_send_mss(sk, &size_goal, flags);
844         }
845
846 out:
847         if (copied)
848                 tcp_push(sk, flags, mss_now, tp->nonagle);
849         return copied;
850
851 do_error:
852         if (copied)
853                 goto out;
854 out_err:
855         return sk_stream_error(sk, flags, err);
856 }
857
858 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
859                      size_t size, int flags)
860 {
861         ssize_t res;
862         struct sock *sk = sock->sk;
863
864         if (!(sk->sk_route_caps & NETIF_F_SG) ||
865             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
866                 return sock_no_sendpage(sock, page, offset, size, flags);
867
868         lock_sock(sk);
869         TCP_CHECK_TIMER(sk);
870         res = do_tcp_sendpages(sk, &page, offset, size, flags);
871         TCP_CHECK_TIMER(sk);
872         release_sock(sk);
873         return res;
874 }
875
876 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
877 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
878
879 static inline int select_size(struct sock *sk, int sg)
880 {
881         struct tcp_sock *tp = tcp_sk(sk);
882         int tmp = tp->mss_cache;
883
884         if (sg) {
885                 if (sk_can_gso(sk))
886                         tmp = 0;
887                 else {
888                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
889
890                         if (tmp >= pgbreak &&
891                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
892                                 tmp = pgbreak;
893                 }
894         }
895
896         return tmp;
897 }
898
899 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
900                 size_t size)
901 {
902         struct sock *sk = sock->sk;
903         struct iovec *iov;
904         struct tcp_sock *tp = tcp_sk(sk);
905         struct sk_buff *skb;
906         int iovlen, flags;
907         int mss_now, size_goal;
908         int sg, err, copied;
909         long timeo;
910
911         lock_sock(sk);
912         TCP_CHECK_TIMER(sk);
913
914         flags = msg->msg_flags;
915         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
916
917         /* Wait for a connection to finish. */
918         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
919                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
920                         goto out_err;
921
922         /* This should be in poll */
923         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
924
925         mss_now = tcp_send_mss(sk, &size_goal, flags);
926
927         /* Ok commence sending. */
928         iovlen = msg->msg_iovlen;
929         iov = msg->msg_iov;
930         copied = 0;
931
932         err = -EPIPE;
933         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
934                 goto out_err;
935
936         sg = sk->sk_route_caps & NETIF_F_SG;
937
938         while (--iovlen >= 0) {
939                 int seglen = iov->iov_len;
940                 unsigned char __user *from = iov->iov_base;
941
942                 iov++;
943
944                 while (seglen > 0) {
945                         int copy = 0;
946                         int max = size_goal;
947
948                         skb = tcp_write_queue_tail(sk);
949                         if (tcp_send_head(sk)) {
950                                 if (skb->ip_summed == CHECKSUM_NONE)
951                                         max = mss_now;
952                                 copy = max - skb->len;
953                         }
954
955                         if (copy <= 0) {
956 new_segment:
957                                 /* Allocate new segment. If the interface is SG,
958                                  * allocate skb fitting to single page.
959                                  */
960                                 if (!sk_stream_memory_free(sk))
961                                         goto wait_for_sndbuf;
962
963                                 skb = sk_stream_alloc_skb(sk,
964                                                           select_size(sk, sg),
965                                                           sk->sk_allocation);
966                                 if (!skb)
967                                         goto wait_for_memory;
968
969                                 /*
970                                  * Check whether we can use HW checksum.
971                                  */
972                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
973                                         skb->ip_summed = CHECKSUM_PARTIAL;
974
975                                 skb_entail(sk, skb);
976                                 copy = size_goal;
977                                 max = size_goal;
978                         }
979
980                         /* Try to append data to the end of skb. */
981                         if (copy > seglen)
982                                 copy = seglen;
983
984                         /* Where to copy to? */
985                         if (skb_tailroom(skb) > 0) {
986                                 /* We have some space in skb head. Superb! */
987                                 if (copy > skb_tailroom(skb))
988                                         copy = skb_tailroom(skb);
989                                 if ((err = skb_add_data(skb, from, copy)) != 0)
990                                         goto do_fault;
991                         } else {
992                                 int merge = 0;
993                                 int i = skb_shinfo(skb)->nr_frags;
994                                 struct page *page = TCP_PAGE(sk);
995                                 int off = TCP_OFF(sk);
996
997                                 if (skb_can_coalesce(skb, i, page, off) &&
998                                     off != PAGE_SIZE) {
999                                         /* We can extend the last page
1000                                          * fragment. */
1001                                         merge = 1;
1002                                 } else if (i == MAX_SKB_FRAGS || !sg) {
1003                                         /* Need to add new fragment and cannot
1004                                          * do this because interface is non-SG,
1005                                          * or because all the page slots are
1006                                          * busy. */
1007                                         tcp_mark_push(tp, skb);
1008                                         goto new_segment;
1009                                 } else if (page) {
1010                                         if (off == PAGE_SIZE) {
1011                                                 put_page(page);
1012                                                 TCP_PAGE(sk) = page = NULL;
1013                                                 off = 0;
1014                                         }
1015                                 } else
1016                                         off = 0;
1017
1018                                 if (copy > PAGE_SIZE - off)
1019                                         copy = PAGE_SIZE - off;
1020
1021                                 if (!sk_wmem_schedule(sk, copy))
1022                                         goto wait_for_memory;
1023
1024                                 if (!page) {
1025                                         /* Allocate new cache page. */
1026                                         if (!(page = sk_stream_alloc_page(sk)))
1027                                                 goto wait_for_memory;
1028                                 }
1029
1030                                 /* Time to copy data. We are close to
1031                                  * the end! */
1032                                 err = skb_copy_to_page(sk, from, skb, page,
1033                                                        off, copy);
1034                                 if (err) {
1035                                         /* If this page was new, give it to the
1036                                          * socket so it does not get leaked.
1037                                          */
1038                                         if (!TCP_PAGE(sk)) {
1039                                                 TCP_PAGE(sk) = page;
1040                                                 TCP_OFF(sk) = 0;
1041                                         }
1042                                         goto do_error;
1043                                 }
1044
1045                                 /* Update the skb. */
1046                                 if (merge) {
1047                                         skb_shinfo(skb)->frags[i - 1].size +=
1048                                                                         copy;
1049                                 } else {
1050                                         skb_fill_page_desc(skb, i, page, off, copy);
1051                                         if (TCP_PAGE(sk)) {
1052                                                 get_page(page);
1053                                         } else if (off + copy < PAGE_SIZE) {
1054                                                 get_page(page);
1055                                                 TCP_PAGE(sk) = page;
1056                                         }
1057                                 }
1058
1059                                 TCP_OFF(sk) = off + copy;
1060                         }
1061
1062                         if (!copied)
1063                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
1064
1065                         tp->write_seq += copy;
1066                         TCP_SKB_CB(skb)->end_seq += copy;
1067                         skb_shinfo(skb)->gso_segs = 0;
1068
1069                         from += copy;
1070                         copied += copy;
1071                         if ((seglen -= copy) == 0 && iovlen == 0)
1072                                 goto out;
1073
1074                         if (skb->len < max || (flags & MSG_OOB))
1075                                 continue;
1076
1077                         if (forced_push(tp)) {
1078                                 tcp_mark_push(tp, skb);
1079                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1080                         } else if (skb == tcp_send_head(sk))
1081                                 tcp_push_one(sk, mss_now);
1082                         continue;
1083
1084 wait_for_sndbuf:
1085                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1086 wait_for_memory:
1087                         if (copied)
1088                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1089
1090                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1091                                 goto do_error;
1092
1093                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1094                 }
1095         }
1096
1097 out:
1098         if (copied)
1099                 tcp_push(sk, flags, mss_now, tp->nonagle);
1100         TCP_CHECK_TIMER(sk);
1101         release_sock(sk);
1102         return copied;
1103
1104 do_fault:
1105         if (!skb->len) {
1106                 tcp_unlink_write_queue(skb, sk);
1107                 /* It is the one place in all of TCP, except connection
1108                  * reset, where we can be unlinking the send_head.
1109                  */
1110                 tcp_check_send_head(sk, skb);
1111                 sk_wmem_free_skb(sk, skb);
1112         }
1113
1114 do_error:
1115         if (copied)
1116                 goto out;
1117 out_err:
1118         err = sk_stream_error(sk, flags, err);
1119         TCP_CHECK_TIMER(sk);
1120         release_sock(sk);
1121         return err;
1122 }
1123
1124 /*
1125  *      Handle reading urgent data. BSD has very simple semantics for
1126  *      this, no blocking and very strange errors 8)
1127  */
1128
1129 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1130 {
1131         struct tcp_sock *tp = tcp_sk(sk);
1132
1133         /* No URG data to read. */
1134         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1135             tp->urg_data == TCP_URG_READ)
1136                 return -EINVAL; /* Yes this is right ! */
1137
1138         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1139                 return -ENOTCONN;
1140
1141         if (tp->urg_data & TCP_URG_VALID) {
1142                 int err = 0;
1143                 char c = tp->urg_data;
1144
1145                 if (!(flags & MSG_PEEK))
1146                         tp->urg_data = TCP_URG_READ;
1147
1148                 /* Read urgent data. */
1149                 msg->msg_flags |= MSG_OOB;
1150
1151                 if (len > 0) {
1152                         if (!(flags & MSG_TRUNC))
1153                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1154                         len = 1;
1155                 } else
1156                         msg->msg_flags |= MSG_TRUNC;
1157
1158                 return err ? -EFAULT : len;
1159         }
1160
1161         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1162                 return 0;
1163
1164         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1165          * the available implementations agree in this case:
1166          * this call should never block, independent of the
1167          * blocking state of the socket.
1168          * Mike <pall@rz.uni-karlsruhe.de>
1169          */
1170         return -EAGAIN;
1171 }
1172
1173 /* Clean up the receive buffer for full frames taken by the user,
1174  * then send an ACK if necessary.  COPIED is the number of bytes
1175  * tcp_recvmsg has given to the user so far, it speeds up the
1176  * calculation of whether or not we must ACK for the sake of
1177  * a window update.
1178  */
1179 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1180 {
1181         struct tcp_sock *tp = tcp_sk(sk);
1182         int time_to_ack = 0;
1183
1184 #if TCP_DEBUG
1185         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1186
1187         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1188              KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1189              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1190 #endif
1191
1192         if (inet_csk_ack_scheduled(sk)) {
1193                 const struct inet_connection_sock *icsk = inet_csk(sk);
1194                    /* Delayed ACKs frequently hit locked sockets during bulk
1195                     * receive. */
1196                 if (icsk->icsk_ack.blocked ||
1197                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1198                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1199                     /*
1200                      * If this read emptied read buffer, we send ACK, if
1201                      * connection is not bidirectional, user drained
1202                      * receive buffer and there was a small segment
1203                      * in queue.
1204                      */
1205                     (copied > 0 &&
1206                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1207                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1208                        !icsk->icsk_ack.pingpong)) &&
1209                       !atomic_read(&sk->sk_rmem_alloc)))
1210                         time_to_ack = 1;
1211         }
1212
1213         /* We send an ACK if we can now advertise a non-zero window
1214          * which has been raised "significantly".
1215          *
1216          * Even if window raised up to infinity, do not send window open ACK
1217          * in states, where we will not receive more. It is useless.
1218          */
1219         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1220                 __u32 rcv_window_now = tcp_receive_window(tp);
1221
1222                 /* Optimize, __tcp_select_window() is not cheap. */
1223                 if (2*rcv_window_now <= tp->window_clamp) {
1224                         __u32 new_window = __tcp_select_window(sk);
1225
1226                         /* Send ACK now, if this read freed lots of space
1227                          * in our buffer. Certainly, new_window is new window.
1228                          * We can advertise it now, if it is not less than current one.
1229                          * "Lots" means "at least twice" here.
1230                          */
1231                         if (new_window && new_window >= 2 * rcv_window_now)
1232                                 time_to_ack = 1;
1233                 }
1234         }
1235         if (time_to_ack)
1236                 tcp_send_ack(sk);
1237 }
1238
1239 static void tcp_prequeue_process(struct sock *sk)
1240 {
1241         struct sk_buff *skb;
1242         struct tcp_sock *tp = tcp_sk(sk);
1243
1244         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1245
1246         /* RX process wants to run with disabled BHs, though it is not
1247          * necessary */
1248         local_bh_disable();
1249         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1250                 sk_backlog_rcv(sk, skb);
1251         local_bh_enable();
1252
1253         /* Clear memory counter. */
1254         tp->ucopy.memory = 0;
1255 }
1256
1257 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1258 {
1259         struct sk_buff *skb;
1260         u32 offset;
1261
1262         skb_queue_walk(&sk->sk_receive_queue, skb) {
1263                 offset = seq - TCP_SKB_CB(skb)->seq;
1264                 if (tcp_hdr(skb)->syn)
1265                         offset--;
1266                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1267                         *off = offset;
1268                         return skb;
1269                 }
1270         }
1271         return NULL;
1272 }
1273
1274 /*
1275  * This routine provides an alternative to tcp_recvmsg() for routines
1276  * that would like to handle copying from skbuffs directly in 'sendfile'
1277  * fashion.
1278  * Note:
1279  *      - It is assumed that the socket was locked by the caller.
1280  *      - The routine does not block.
1281  *      - At present, there is no support for reading OOB data
1282  *        or for 'peeking' the socket using this routine
1283  *        (although both would be easy to implement).
1284  */
1285 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1286                   sk_read_actor_t recv_actor)
1287 {
1288         struct sk_buff *skb;
1289         struct tcp_sock *tp = tcp_sk(sk);
1290         u32 seq = tp->copied_seq;
1291         u32 offset;
1292         int copied = 0;
1293
1294         if (sk->sk_state == TCP_LISTEN)
1295                 return -ENOTCONN;
1296         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1297                 if (offset < skb->len) {
1298                         int used;
1299                         size_t len;
1300
1301                         len = skb->len - offset;
1302                         /* Stop reading if we hit a patch of urgent data */
1303                         if (tp->urg_data) {
1304                                 u32 urg_offset = tp->urg_seq - seq;
1305                                 if (urg_offset < len)
1306                                         len = urg_offset;
1307                                 if (!len)
1308                                         break;
1309                         }
1310                         used = recv_actor(desc, skb, offset, len);
1311                         if (used < 0) {
1312                                 if (!copied)
1313                                         copied = used;
1314                                 break;
1315                         } else if (used <= len) {
1316                                 seq += used;
1317                                 copied += used;
1318                                 offset += used;
1319                         }
1320                         /*
1321                          * If recv_actor drops the lock (e.g. TCP splice
1322                          * receive) the skb pointer might be invalid when
1323                          * getting here: tcp_collapse might have deleted it
1324                          * while aggregating skbs from the socket queue.
1325                          */
1326                         skb = tcp_recv_skb(sk, seq-1, &offset);
1327                         if (!skb || (offset+1 != skb->len))
1328                                 break;
1329                 }
1330                 if (tcp_hdr(skb)->fin) {
1331                         sk_eat_skb(sk, skb, 0);
1332                         ++seq;
1333                         break;
1334                 }
1335                 sk_eat_skb(sk, skb, 0);
1336                 if (!desc->count)
1337                         break;
1338         }
1339         tp->copied_seq = seq;
1340
1341         tcp_rcv_space_adjust(sk);
1342
1343         /* Clean up data we have read: This will do ACK frames. */
1344         if (copied > 0)
1345                 tcp_cleanup_rbuf(sk, copied);
1346         return copied;
1347 }
1348
1349 /*
1350  *      This routine copies from a sock struct into the user buffer.
1351  *
1352  *      Technical note: in 2.3 we work on _locked_ socket, so that
1353  *      tricks with *seq access order and skb->users are not required.
1354  *      Probably, code can be easily improved even more.
1355  */
1356
1357 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1358                 size_t len, int nonblock, int flags, int *addr_len)
1359 {
1360         struct tcp_sock *tp = tcp_sk(sk);
1361         int copied = 0;
1362         u32 peek_seq;
1363         u32 *seq;
1364         unsigned long used;
1365         int err;
1366         int target;             /* Read at least this many bytes */
1367         long timeo;
1368         struct task_struct *user_recv = NULL;
1369         int copied_early = 0;
1370         struct sk_buff *skb;
1371         u32 urg_hole = 0;
1372
1373         lock_sock(sk);
1374
1375         TCP_CHECK_TIMER(sk);
1376
1377         err = -ENOTCONN;
1378         if (sk->sk_state == TCP_LISTEN)
1379                 goto out;
1380
1381         timeo = sock_rcvtimeo(sk, nonblock);
1382
1383         /* Urgent data needs to be handled specially. */
1384         if (flags & MSG_OOB)
1385                 goto recv_urg;
1386
1387         seq = &tp->copied_seq;
1388         if (flags & MSG_PEEK) {
1389                 peek_seq = tp->copied_seq;
1390                 seq = &peek_seq;
1391         }
1392
1393         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1394
1395 #ifdef CONFIG_NET_DMA
1396         tp->ucopy.dma_chan = NULL;
1397         preempt_disable();
1398         skb = skb_peek_tail(&sk->sk_receive_queue);
1399         {
1400                 int available = 0;
1401
1402                 if (skb)
1403                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1404                 if ((available < target) &&
1405                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1406                     !sysctl_tcp_low_latency &&
1407                     dma_find_channel(DMA_MEMCPY)) {
1408                         preempt_enable_no_resched();
1409                         tp->ucopy.pinned_list =
1410                                         dma_pin_iovec_pages(msg->msg_iov, len);
1411                 } else {
1412                         preempt_enable_no_resched();
1413                 }
1414         }
1415 #endif
1416
1417         do {
1418                 u32 offset;
1419
1420                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1421                 if (tp->urg_data && tp->urg_seq == *seq) {
1422                         if (copied)
1423                                 break;
1424                         if (signal_pending(current)) {
1425                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1426                                 break;
1427                         }
1428                 }
1429
1430                 /* Next get a buffer. */
1431
1432                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1433                         /* Now that we have two receive queues this
1434                          * shouldn't happen.
1435                          */
1436                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1437                              KERN_INFO "recvmsg bug: copied %X "
1438                                        "seq %X rcvnxt %X fl %X\n", *seq,
1439                                        TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1440                                        flags))
1441                                 break;
1442
1443                         offset = *seq - TCP_SKB_CB(skb)->seq;
1444                         if (tcp_hdr(skb)->syn)
1445                                 offset--;
1446                         if (offset < skb->len)
1447                                 goto found_ok_skb;
1448                         if (tcp_hdr(skb)->fin)
1449                                 goto found_fin_ok;
1450                         WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: "
1451                                         "copied %X seq %X rcvnxt %X fl %X\n",
1452                                         *seq, TCP_SKB_CB(skb)->seq,
1453                                         tp->rcv_nxt, flags);
1454                 }
1455
1456                 /* Well, if we have backlog, try to process it now yet. */
1457
1458                 if (copied >= target && !sk->sk_backlog.tail)
1459                         break;
1460
1461                 if (copied) {
1462                         if (sk->sk_err ||
1463                             sk->sk_state == TCP_CLOSE ||
1464                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1465                             !timeo ||
1466                             signal_pending(current))
1467                                 break;
1468                 } else {
1469                         if (sock_flag(sk, SOCK_DONE))
1470                                 break;
1471
1472                         if (sk->sk_err) {
1473                                 copied = sock_error(sk);
1474                                 break;
1475                         }
1476
1477                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1478                                 break;
1479
1480                         if (sk->sk_state == TCP_CLOSE) {
1481                                 if (!sock_flag(sk, SOCK_DONE)) {
1482                                         /* This occurs when user tries to read
1483                                          * from never connected socket.
1484                                          */
1485                                         copied = -ENOTCONN;
1486                                         break;
1487                                 }
1488                                 break;
1489                         }
1490
1491                         if (!timeo) {
1492                                 copied = -EAGAIN;
1493                                 break;
1494                         }
1495
1496                         if (signal_pending(current)) {
1497                                 copied = sock_intr_errno(timeo);
1498                                 break;
1499                         }
1500                 }
1501
1502                 tcp_cleanup_rbuf(sk, copied);
1503
1504                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1505                         /* Install new reader */
1506                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1507                                 user_recv = current;
1508                                 tp->ucopy.task = user_recv;
1509                                 tp->ucopy.iov = msg->msg_iov;
1510                         }
1511
1512                         tp->ucopy.len = len;
1513
1514                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1515                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1516
1517                         /* Ugly... If prequeue is not empty, we have to
1518                          * process it before releasing socket, otherwise
1519                          * order will be broken at second iteration.
1520                          * More elegant solution is required!!!
1521                          *
1522                          * Look: we have the following (pseudo)queues:
1523                          *
1524                          * 1. packets in flight
1525                          * 2. backlog
1526                          * 3. prequeue
1527                          * 4. receive_queue
1528                          *
1529                          * Each queue can be processed only if the next ones
1530                          * are empty. At this point we have empty receive_queue.
1531                          * But prequeue _can_ be not empty after 2nd iteration,
1532                          * when we jumped to start of loop because backlog
1533                          * processing added something to receive_queue.
1534                          * We cannot release_sock(), because backlog contains
1535                          * packets arrived _after_ prequeued ones.
1536                          *
1537                          * Shortly, algorithm is clear --- to process all
1538                          * the queues in order. We could make it more directly,
1539                          * requeueing packets from backlog to prequeue, if
1540                          * is not empty. It is more elegant, but eats cycles,
1541                          * unfortunately.
1542                          */
1543                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1544                                 goto do_prequeue;
1545
1546                         /* __ Set realtime policy in scheduler __ */
1547                 }
1548
1549                 if (copied >= target) {
1550                         /* Do not sleep, just process backlog. */
1551                         release_sock(sk);
1552                         lock_sock(sk);
1553                 } else
1554                         sk_wait_data(sk, &timeo);
1555
1556 #ifdef CONFIG_NET_DMA
1557                 tp->ucopy.wakeup = 0;
1558 #endif
1559
1560                 if (user_recv) {
1561                         int chunk;
1562
1563                         /* __ Restore normal policy in scheduler __ */
1564
1565                         if ((chunk = len - tp->ucopy.len) != 0) {
1566                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1567                                 len -= chunk;
1568                                 copied += chunk;
1569                         }
1570
1571                         if (tp->rcv_nxt == tp->copied_seq &&
1572                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1573 do_prequeue:
1574                                 tcp_prequeue_process(sk);
1575
1576                                 if ((chunk = len - tp->ucopy.len) != 0) {
1577                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1578                                         len -= chunk;
1579                                         copied += chunk;
1580                                 }
1581                         }
1582                 }
1583                 if ((flags & MSG_PEEK) &&
1584                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1585                         if (net_ratelimit())
1586                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1587                                        current->comm, task_pid_nr(current));
1588                         peek_seq = tp->copied_seq;
1589                 }
1590                 continue;
1591
1592         found_ok_skb:
1593                 /* Ok so how much can we use? */
1594                 used = skb->len - offset;
1595                 if (len < used)
1596                         used = len;
1597
1598                 /* Do we have urgent data here? */
1599                 if (tp->urg_data) {
1600                         u32 urg_offset = tp->urg_seq - *seq;
1601                         if (urg_offset < used) {
1602                                 if (!urg_offset) {
1603                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1604                                                 ++*seq;
1605                                                 urg_hole++;
1606                                                 offset++;
1607                                                 used--;
1608                                                 if (!used)
1609                                                         goto skip_copy;
1610                                         }
1611                                 } else
1612                                         used = urg_offset;
1613                         }
1614                 }
1615
1616                 if (!(flags & MSG_TRUNC)) {
1617 #ifdef CONFIG_NET_DMA
1618                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1619                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1620
1621                         if (tp->ucopy.dma_chan) {
1622                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1623                                         tp->ucopy.dma_chan, skb, offset,
1624                                         msg->msg_iov, used,
1625                                         tp->ucopy.pinned_list);
1626
1627                                 if (tp->ucopy.dma_cookie < 0) {
1628
1629                                         printk(KERN_ALERT "dma_cookie < 0\n");
1630
1631                                         /* Exception. Bailout! */
1632                                         if (!copied)
1633                                                 copied = -EFAULT;
1634                                         break;
1635                                 }
1636                                 if ((offset + used) == skb->len)
1637                                         copied_early = 1;
1638
1639                         } else
1640 #endif
1641                         {
1642                                 err = skb_copy_datagram_iovec(skb, offset,
1643                                                 msg->msg_iov, used);
1644                                 if (err) {
1645                                         /* Exception. Bailout! */
1646                                         if (!copied)
1647                                                 copied = -EFAULT;
1648                                         break;
1649                                 }
1650                         }
1651                 }
1652
1653                 *seq += used;
1654                 copied += used;
1655                 len -= used;
1656
1657                 tcp_rcv_space_adjust(sk);
1658
1659 skip_copy:
1660                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1661                         tp->urg_data = 0;
1662                         tcp_fast_path_check(sk);
1663                 }
1664                 if (used + offset < skb->len)
1665                         continue;
1666
1667                 if (tcp_hdr(skb)->fin)
1668                         goto found_fin_ok;
1669                 if (!(flags & MSG_PEEK)) {
1670                         sk_eat_skb(sk, skb, copied_early);
1671                         copied_early = 0;
1672                 }
1673                 continue;
1674
1675         found_fin_ok:
1676                 /* Process the FIN. */
1677                 ++*seq;
1678                 if (!(flags & MSG_PEEK)) {
1679                         sk_eat_skb(sk, skb, copied_early);
1680                         copied_early = 0;
1681                 }
1682                 break;
1683         } while (len > 0);
1684
1685         if (user_recv) {
1686                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1687                         int chunk;
1688
1689                         tp->ucopy.len = copied > 0 ? len : 0;
1690
1691                         tcp_prequeue_process(sk);
1692
1693                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1694                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1695                                 len -= chunk;
1696                                 copied += chunk;
1697                         }
1698                 }
1699
1700                 tp->ucopy.task = NULL;
1701                 tp->ucopy.len = 0;
1702         }
1703
1704 #ifdef CONFIG_NET_DMA
1705         if (tp->ucopy.dma_chan) {
1706                 dma_cookie_t done, used;
1707
1708                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1709
1710                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1711                                                  tp->ucopy.dma_cookie, &done,
1712                                                  &used) == DMA_IN_PROGRESS) {
1713                         /* do partial cleanup of sk_async_wait_queue */
1714                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1715                                (dma_async_is_complete(skb->dma_cookie, done,
1716                                                       used) == DMA_SUCCESS)) {
1717                                 __skb_dequeue(&sk->sk_async_wait_queue);
1718                                 kfree_skb(skb);
1719                         }
1720                 }
1721
1722                 /* Safe to free early-copied skbs now */
1723                 __skb_queue_purge(&sk->sk_async_wait_queue);
1724                 tp->ucopy.dma_chan = NULL;
1725         }
1726         if (tp->ucopy.pinned_list) {
1727                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1728                 tp->ucopy.pinned_list = NULL;
1729         }
1730 #endif
1731
1732         /* According to UNIX98, msg_name/msg_namelen are ignored
1733          * on connected socket. I was just happy when found this 8) --ANK
1734          */
1735
1736         /* Clean up data we have read: This will do ACK frames. */
1737         tcp_cleanup_rbuf(sk, copied);
1738
1739         TCP_CHECK_TIMER(sk);
1740         release_sock(sk);
1741         return copied;
1742
1743 out:
1744         TCP_CHECK_TIMER(sk);
1745         release_sock(sk);
1746         return err;
1747
1748 recv_urg:
1749         err = tcp_recv_urg(sk, msg, len, flags);
1750         goto out;
1751 }
1752
1753 void tcp_set_state(struct sock *sk, int state)
1754 {
1755         int oldstate = sk->sk_state;
1756
1757         switch (state) {
1758         case TCP_ESTABLISHED:
1759                 if (oldstate != TCP_ESTABLISHED)
1760                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1761                 break;
1762
1763         case TCP_CLOSE:
1764                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1765                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1766
1767                 sk->sk_prot->unhash(sk);
1768                 if (inet_csk(sk)->icsk_bind_hash &&
1769                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1770                         inet_put_port(sk);
1771                 /* fall through */
1772         default:
1773                 if (oldstate == TCP_ESTABLISHED)
1774                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1775         }
1776
1777         /* Change state AFTER socket is unhashed to avoid closed
1778          * socket sitting in hash tables.
1779          */
1780         sk->sk_state = state;
1781
1782 #ifdef STATE_TRACE
1783         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1784 #endif
1785 }
1786 EXPORT_SYMBOL_GPL(tcp_set_state);
1787
1788 /*
1789  *      State processing on a close. This implements the state shift for
1790  *      sending our FIN frame. Note that we only send a FIN for some
1791  *      states. A shutdown() may have already sent the FIN, or we may be
1792  *      closed.
1793  */
1794
1795 static const unsigned char new_state[16] = {
1796   /* current state:        new state:      action:      */
1797   /* (Invalid)          */ TCP_CLOSE,
1798   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1799   /* TCP_SYN_SENT       */ TCP_CLOSE,
1800   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1801   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1802   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1803   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1804   /* TCP_CLOSE          */ TCP_CLOSE,
1805   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1806   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1807   /* TCP_LISTEN         */ TCP_CLOSE,
1808   /* TCP_CLOSING        */ TCP_CLOSING,
1809 };
1810
1811 static int tcp_close_state(struct sock *sk)
1812 {
1813         int next = (int)new_state[sk->sk_state];
1814         int ns = next & TCP_STATE_MASK;
1815
1816         tcp_set_state(sk, ns);
1817
1818         return next & TCP_ACTION_FIN;
1819 }
1820
1821 /*
1822  *      Shutdown the sending side of a connection. Much like close except
1823  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1824  */
1825
1826 void tcp_shutdown(struct sock *sk, int how)
1827 {
1828         /*      We need to grab some memory, and put together a FIN,
1829          *      and then put it into the queue to be sent.
1830          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1831          */
1832         if (!(how & SEND_SHUTDOWN))
1833                 return;
1834
1835         /* If we've already sent a FIN, or it's a closed state, skip this. */
1836         if ((1 << sk->sk_state) &
1837             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1838              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1839                 /* Clear out any half completed packets.  FIN if needed. */
1840                 if (tcp_close_state(sk))
1841                         tcp_send_fin(sk);
1842         }
1843 }
1844
1845 void tcp_close(struct sock *sk, long timeout)
1846 {
1847         struct sk_buff *skb;
1848         int data_was_unread = 0;
1849         int state;
1850
1851         lock_sock(sk);
1852         sk->sk_shutdown = SHUTDOWN_MASK;
1853
1854         if (sk->sk_state == TCP_LISTEN) {
1855                 tcp_set_state(sk, TCP_CLOSE);
1856
1857                 /* Special case. */
1858                 inet_csk_listen_stop(sk);
1859
1860                 goto adjudge_to_death;
1861         }
1862
1863         /*  We need to flush the recv. buffs.  We do this only on the
1864          *  descriptor close, not protocol-sourced closes, because the
1865          *  reader process may not have drained the data yet!
1866          */
1867         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1868                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1869                           tcp_hdr(skb)->fin;
1870                 data_was_unread += len;
1871                 __kfree_skb(skb);
1872         }
1873
1874         sk_mem_reclaim(sk);
1875
1876         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1877          * data was lost. To witness the awful effects of the old behavior of
1878          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1879          * GET in an FTP client, suspend the process, wait for the client to
1880          * advertise a zero window, then kill -9 the FTP client, wheee...
1881          * Note: timeout is always zero in such a case.
1882          */
1883         if (data_was_unread) {
1884                 /* Unread data was tossed, zap the connection. */
1885                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1886                 tcp_set_state(sk, TCP_CLOSE);
1887                 tcp_send_active_reset(sk, sk->sk_allocation);
1888         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1889                 /* Check zero linger _after_ checking for unread data. */
1890                 sk->sk_prot->disconnect(sk, 0);
1891                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1892         } else if (tcp_close_state(sk)) {
1893                 /* We FIN if the application ate all the data before
1894                  * zapping the connection.
1895                  */
1896
1897                 /* RED-PEN. Formally speaking, we have broken TCP state
1898                  * machine. State transitions:
1899                  *
1900                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1901                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1902                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1903                  *
1904                  * are legal only when FIN has been sent (i.e. in window),
1905                  * rather than queued out of window. Purists blame.
1906                  *
1907                  * F.e. "RFC state" is ESTABLISHED,
1908                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1909                  *
1910                  * The visible declinations are that sometimes
1911                  * we enter time-wait state, when it is not required really
1912                  * (harmless), do not send active resets, when they are
1913                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1914                  * they look as CLOSING or LAST_ACK for Linux)
1915                  * Probably, I missed some more holelets.
1916                  *                                              --ANK
1917                  */
1918                 tcp_send_fin(sk);
1919         }
1920
1921         sk_stream_wait_close(sk, timeout);
1922
1923 adjudge_to_death:
1924         state = sk->sk_state;
1925         sock_hold(sk);
1926         sock_orphan(sk);
1927
1928         /* It is the last release_sock in its life. It will remove backlog. */
1929         release_sock(sk);
1930
1931
1932         /* Now socket is owned by kernel and we acquire BH lock
1933            to finish close. No need to check for user refs.
1934          */
1935         local_bh_disable();
1936         bh_lock_sock(sk);
1937         WARN_ON(sock_owned_by_user(sk));
1938
1939         percpu_counter_inc(sk->sk_prot->orphan_count);
1940
1941         /* Have we already been destroyed by a softirq or backlog? */
1942         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1943                 goto out;
1944
1945         /*      This is a (useful) BSD violating of the RFC. There is a
1946          *      problem with TCP as specified in that the other end could
1947          *      keep a socket open forever with no application left this end.
1948          *      We use a 3 minute timeout (about the same as BSD) then kill
1949          *      our end. If they send after that then tough - BUT: long enough
1950          *      that we won't make the old 4*rto = almost no time - whoops
1951          *      reset mistake.
1952          *
1953          *      Nope, it was not mistake. It is really desired behaviour
1954          *      f.e. on http servers, when such sockets are useless, but
1955          *      consume significant resources. Let's do it with special
1956          *      linger2 option.                                 --ANK
1957          */
1958
1959         if (sk->sk_state == TCP_FIN_WAIT2) {
1960                 struct tcp_sock *tp = tcp_sk(sk);
1961                 if (tp->linger2 < 0) {
1962                         tcp_set_state(sk, TCP_CLOSE);
1963                         tcp_send_active_reset(sk, GFP_ATOMIC);
1964                         NET_INC_STATS_BH(sock_net(sk),
1965                                         LINUX_MIB_TCPABORTONLINGER);
1966                 } else {
1967                         const int tmo = tcp_fin_time(sk);
1968
1969                         if (tmo > TCP_TIMEWAIT_LEN) {
1970                                 inet_csk_reset_keepalive_timer(sk,
1971                                                 tmo - TCP_TIMEWAIT_LEN);
1972                         } else {
1973                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1974                                 goto out;
1975                         }
1976                 }
1977         }
1978         if (sk->sk_state != TCP_CLOSE) {
1979                 int orphan_count = percpu_counter_read_positive(
1980                                                 sk->sk_prot->orphan_count);
1981
1982                 sk_mem_reclaim(sk);
1983                 if (tcp_too_many_orphans(sk, orphan_count)) {
1984                         if (net_ratelimit())
1985                                 printk(KERN_INFO "TCP: too many of orphaned "
1986                                        "sockets\n");
1987                         tcp_set_state(sk, TCP_CLOSE);
1988                         tcp_send_active_reset(sk, GFP_ATOMIC);
1989                         NET_INC_STATS_BH(sock_net(sk),
1990                                         LINUX_MIB_TCPABORTONMEMORY);
1991                 }
1992         }
1993
1994         if (sk->sk_state == TCP_CLOSE)
1995                 inet_csk_destroy_sock(sk);
1996         /* Otherwise, socket is reprieved until protocol close. */
1997
1998 out:
1999         bh_unlock_sock(sk);
2000         local_bh_enable();
2001         sock_put(sk);
2002 }
2003
2004 /* These states need RST on ABORT according to RFC793 */
2005
2006 static inline int tcp_need_reset(int state)
2007 {
2008         return (1 << state) &
2009                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2010                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2011 }
2012
2013 int tcp_disconnect(struct sock *sk, int flags)
2014 {
2015         struct inet_sock *inet = inet_sk(sk);
2016         struct inet_connection_sock *icsk = inet_csk(sk);
2017         struct tcp_sock *tp = tcp_sk(sk);
2018         int err = 0;
2019         int old_state = sk->sk_state;
2020
2021         if (old_state != TCP_CLOSE)
2022                 tcp_set_state(sk, TCP_CLOSE);
2023
2024         /* ABORT function of RFC793 */
2025         if (old_state == TCP_LISTEN) {
2026                 inet_csk_listen_stop(sk);
2027         } else if (tcp_need_reset(old_state) ||
2028                    (tp->snd_nxt != tp->write_seq &&
2029                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2030                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2031                  * states
2032                  */
2033                 tcp_send_active_reset(sk, gfp_any());
2034                 sk->sk_err = ECONNRESET;
2035         } else if (old_state == TCP_SYN_SENT)
2036                 sk->sk_err = ECONNRESET;
2037
2038         tcp_clear_xmit_timers(sk);
2039         __skb_queue_purge(&sk->sk_receive_queue);
2040         tcp_write_queue_purge(sk);
2041         __skb_queue_purge(&tp->out_of_order_queue);
2042 #ifdef CONFIG_NET_DMA
2043         __skb_queue_purge(&sk->sk_async_wait_queue);
2044 #endif
2045
2046         inet->inet_dport = 0;
2047
2048         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2049                 inet_reset_saddr(sk);
2050
2051         sk->sk_shutdown = 0;
2052         sock_reset_flag(sk, SOCK_DONE);
2053         tp->srtt = 0;
2054         if ((tp->write_seq += tp->max_window + 2) == 0)
2055                 tp->write_seq = 1;
2056         icsk->icsk_backoff = 0;
2057         tp->snd_cwnd = 2;
2058         icsk->icsk_probes_out = 0;
2059         tp->packets_out = 0;
2060         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2061         tp->snd_cwnd_cnt = 0;
2062         tp->bytes_acked = 0;
2063         tp->window_clamp = 0;
2064         tcp_set_ca_state(sk, TCP_CA_Open);
2065         tcp_clear_retrans(tp);
2066         inet_csk_delack_init(sk);
2067         tcp_init_send_head(sk);
2068         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2069         __sk_dst_reset(sk);
2070
2071         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2072
2073         sk->sk_error_report(sk);
2074         return err;
2075 }
2076
2077 /*
2078  *      Socket option code for TCP.
2079  */
2080 static int do_tcp_setsockopt(struct sock *sk, int level,
2081                 int optname, char __user *optval, unsigned int optlen)
2082 {
2083         struct tcp_sock *tp = tcp_sk(sk);
2084         struct inet_connection_sock *icsk = inet_csk(sk);
2085         int val;
2086         int err = 0;
2087
2088         /* These are data/string values, all the others are ints */
2089         switch (optname) {
2090         case TCP_CONGESTION: {
2091                 char name[TCP_CA_NAME_MAX];
2092
2093                 if (optlen < 1)
2094                         return -EINVAL;
2095
2096                 val = strncpy_from_user(name, optval,
2097                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2098                 if (val < 0)
2099                         return -EFAULT;
2100                 name[val] = 0;
2101
2102                 lock_sock(sk);
2103                 err = tcp_set_congestion_control(sk, name);
2104                 release_sock(sk);
2105                 return err;
2106         }
2107         case TCP_COOKIE_TRANSACTIONS: {
2108                 struct tcp_cookie_transactions ctd;
2109                 struct tcp_cookie_values *cvp = NULL;
2110
2111                 if (sizeof(ctd) > optlen)
2112                         return -EINVAL;
2113                 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2114                         return -EFAULT;
2115
2116                 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2117                     ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2118                         return -EINVAL;
2119
2120                 if (ctd.tcpct_cookie_desired == 0) {
2121                         /* default to global value */
2122                 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2123                            ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2124                            ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2125                         return -EINVAL;
2126                 }
2127
2128                 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2129                         /* Supercedes all other values */
2130                         lock_sock(sk);
2131                         if (tp->cookie_values != NULL) {
2132                                 kref_put(&tp->cookie_values->kref,
2133                                          tcp_cookie_values_release);
2134                                 tp->cookie_values = NULL;
2135                         }
2136                         tp->rx_opt.cookie_in_always = 0; /* false */
2137                         tp->rx_opt.cookie_out_never = 1; /* true */
2138                         release_sock(sk);
2139                         return err;
2140                 }
2141
2142                 /* Allocate ancillary memory before locking.
2143                  */
2144                 if (ctd.tcpct_used > 0 ||
2145                     (tp->cookie_values == NULL &&
2146                      (sysctl_tcp_cookie_size > 0 ||
2147                       ctd.tcpct_cookie_desired > 0 ||
2148                       ctd.tcpct_s_data_desired > 0))) {
2149                         cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2150                                       GFP_KERNEL);
2151                         if (cvp == NULL)
2152                                 return -ENOMEM;
2153                 }
2154                 lock_sock(sk);
2155                 tp->rx_opt.cookie_in_always =
2156                         (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2157                 tp->rx_opt.cookie_out_never = 0; /* false */
2158
2159                 if (tp->cookie_values != NULL) {
2160                         if (cvp != NULL) {
2161                                 /* Changed values are recorded by a changed
2162                                  * pointer, ensuring the cookie will differ,
2163                                  * without separately hashing each value later.
2164                                  */
2165                                 kref_put(&tp->cookie_values->kref,
2166                                          tcp_cookie_values_release);
2167                                 kref_init(&cvp->kref);
2168                                 tp->cookie_values = cvp;
2169                         } else {
2170                                 cvp = tp->cookie_values;
2171                         }
2172                 }
2173                 if (cvp != NULL) {
2174                         cvp->cookie_desired = ctd.tcpct_cookie_desired;
2175
2176                         if (ctd.tcpct_used > 0) {
2177                                 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2178                                        ctd.tcpct_used);
2179                                 cvp->s_data_desired = ctd.tcpct_used;
2180                                 cvp->s_data_constant = 1; /* true */
2181                         } else {
2182                                 /* No constant payload data. */
2183                                 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2184                                 cvp->s_data_constant = 0; /* false */
2185                         }
2186                 }
2187                 release_sock(sk);
2188                 return err;
2189         }
2190         default:
2191                 /* fallthru */
2192                 break;
2193         };
2194
2195         if (optlen < sizeof(int))
2196                 return -EINVAL;
2197
2198         if (get_user(val, (int __user *)optval))
2199                 return -EFAULT;
2200
2201         lock_sock(sk);
2202
2203         switch (optname) {
2204         case TCP_MAXSEG:
2205                 /* Values greater than interface MTU won't take effect. However
2206                  * at the point when this call is done we typically don't yet
2207                  * know which interface is going to be used */
2208                 if (val < 8 || val > MAX_TCP_WINDOW) {
2209                         err = -EINVAL;
2210                         break;
2211                 }
2212                 tp->rx_opt.user_mss = val;
2213                 break;
2214
2215         case TCP_NODELAY:
2216                 if (val) {
2217                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2218                          * this option on corked socket is remembered, but
2219                          * it is not activated until cork is cleared.
2220                          *
2221                          * However, when TCP_NODELAY is set we make
2222                          * an explicit push, which overrides even TCP_CORK
2223                          * for currently queued segments.
2224                          */
2225                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2226                         tcp_push_pending_frames(sk);
2227                 } else {
2228                         tp->nonagle &= ~TCP_NAGLE_OFF;
2229                 }
2230                 break;
2231
2232         case TCP_THIN_LINEAR_TIMEOUTS:
2233                 if (val < 0 || val > 1)
2234                         err = -EINVAL;
2235                 else
2236                         tp->thin_lto = val;
2237                 break;
2238
2239         case TCP_THIN_DUPACK:
2240                 if (val < 0 || val > 1)
2241                         err = -EINVAL;
2242                 else
2243                         tp->thin_dupack = val;
2244                 break;
2245
2246         case TCP_CORK:
2247                 /* When set indicates to always queue non-full frames.
2248                  * Later the user clears this option and we transmit
2249                  * any pending partial frames in the queue.  This is
2250                  * meant to be used alongside sendfile() to get properly
2251                  * filled frames when the user (for example) must write
2252                  * out headers with a write() call first and then use
2253                  * sendfile to send out the data parts.
2254                  *
2255                  * TCP_CORK can be set together with TCP_NODELAY and it is
2256                  * stronger than TCP_NODELAY.
2257                  */
2258                 if (val) {
2259                         tp->nonagle |= TCP_NAGLE_CORK;
2260                 } else {
2261                         tp->nonagle &= ~TCP_NAGLE_CORK;
2262                         if (tp->nonagle&TCP_NAGLE_OFF)
2263                                 tp->nonagle |= TCP_NAGLE_PUSH;
2264                         tcp_push_pending_frames(sk);
2265                 }
2266                 break;
2267
2268         case TCP_KEEPIDLE:
2269                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2270                         err = -EINVAL;
2271                 else {
2272                         tp->keepalive_time = val * HZ;
2273                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2274                             !((1 << sk->sk_state) &
2275                               (TCPF_CLOSE | TCPF_LISTEN))) {
2276                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2277                                 if (tp->keepalive_time > elapsed)
2278                                         elapsed = tp->keepalive_time - elapsed;
2279                                 else
2280                                         elapsed = 0;
2281                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2282                         }
2283                 }
2284                 break;
2285         case TCP_KEEPINTVL:
2286                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2287                         err = -EINVAL;
2288                 else
2289                         tp->keepalive_intvl = val * HZ;
2290                 break;
2291         case TCP_KEEPCNT:
2292                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2293                         err = -EINVAL;
2294                 else
2295                         tp->keepalive_probes = val;
2296                 break;
2297         case TCP_SYNCNT:
2298                 if (val < 1 || val > MAX_TCP_SYNCNT)
2299                         err = -EINVAL;
2300                 else
2301                         icsk->icsk_syn_retries = val;
2302                 break;
2303
2304         case TCP_LINGER2:
2305                 if (val < 0)
2306                         tp->linger2 = -1;
2307                 else if (val > sysctl_tcp_fin_timeout / HZ)
2308                         tp->linger2 = 0;
2309                 else
2310                         tp->linger2 = val * HZ;
2311                 break;
2312
2313         case TCP_DEFER_ACCEPT:
2314                 /* Translate value in seconds to number of retransmits */
2315                 icsk->icsk_accept_queue.rskq_defer_accept =
2316                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2317                                         TCP_RTO_MAX / HZ);
2318                 break;
2319
2320         case TCP_WINDOW_CLAMP:
2321                 if (!val) {
2322                         if (sk->sk_state != TCP_CLOSE) {
2323                                 err = -EINVAL;
2324                                 break;
2325                         }
2326                         tp->window_clamp = 0;
2327                 } else
2328                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2329                                                 SOCK_MIN_RCVBUF / 2 : val;
2330                 break;
2331
2332         case TCP_QUICKACK:
2333                 if (!val) {
2334                         icsk->icsk_ack.pingpong = 1;
2335                 } else {
2336                         icsk->icsk_ack.pingpong = 0;
2337                         if ((1 << sk->sk_state) &
2338                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2339                             inet_csk_ack_scheduled(sk)) {
2340                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2341                                 tcp_cleanup_rbuf(sk, 1);
2342                                 if (!(val & 1))
2343                                         icsk->icsk_ack.pingpong = 1;
2344                         }
2345                 }
2346                 break;
2347
2348 #ifdef CONFIG_TCP_MD5SIG
2349         case TCP_MD5SIG:
2350                 /* Read the IP->Key mappings from userspace */
2351                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2352                 break;
2353 #endif
2354
2355         default:
2356                 err = -ENOPROTOOPT;
2357                 break;
2358         }
2359
2360         release_sock(sk);
2361         return err;
2362 }
2363
2364 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2365                    unsigned int optlen)
2366 {
2367         struct inet_connection_sock *icsk = inet_csk(sk);
2368
2369         if (level != SOL_TCP)
2370                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2371                                                      optval, optlen);
2372         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2373 }
2374
2375 #ifdef CONFIG_COMPAT
2376 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2377                           char __user *optval, unsigned int optlen)
2378 {
2379         if (level != SOL_TCP)
2380                 return inet_csk_compat_setsockopt(sk, level, optname,
2381                                                   optval, optlen);
2382         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2383 }
2384
2385 EXPORT_SYMBOL(compat_tcp_setsockopt);
2386 #endif
2387
2388 /* Return information about state of tcp endpoint in API format. */
2389 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2390 {
2391         struct tcp_sock *tp = tcp_sk(sk);
2392         const struct inet_connection_sock *icsk = inet_csk(sk);
2393         u32 now = tcp_time_stamp;
2394
2395         memset(info, 0, sizeof(*info));
2396
2397         info->tcpi_state = sk->sk_state;
2398         info->tcpi_ca_state = icsk->icsk_ca_state;
2399         info->tcpi_retransmits = icsk->icsk_retransmits;
2400         info->tcpi_probes = icsk->icsk_probes_out;
2401         info->tcpi_backoff = icsk->icsk_backoff;
2402
2403         if (tp->rx_opt.tstamp_ok)
2404                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2405         if (tcp_is_sack(tp))
2406                 info->tcpi_options |= TCPI_OPT_SACK;
2407         if (tp->rx_opt.wscale_ok) {
2408                 info->tcpi_options |= TCPI_OPT_WSCALE;
2409                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2410                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2411         }
2412
2413         if (tp->ecn_flags&TCP_ECN_OK)
2414                 info->tcpi_options |= TCPI_OPT_ECN;
2415
2416         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2417         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2418         info->tcpi_snd_mss = tp->mss_cache;
2419         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2420
2421         if (sk->sk_state == TCP_LISTEN) {
2422                 info->tcpi_unacked = sk->sk_ack_backlog;
2423                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2424         } else {
2425                 info->tcpi_unacked = tp->packets_out;
2426                 info->tcpi_sacked = tp->sacked_out;
2427         }
2428         info->tcpi_lost = tp->lost_out;
2429         info->tcpi_retrans = tp->retrans_out;
2430         info->tcpi_fackets = tp->fackets_out;
2431
2432         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2433         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2434         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2435
2436         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2437         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2438         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2439         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2440         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2441         info->tcpi_snd_cwnd = tp->snd_cwnd;
2442         info->tcpi_advmss = tp->advmss;
2443         info->tcpi_reordering = tp->reordering;
2444
2445         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2446         info->tcpi_rcv_space = tp->rcvq_space.space;
2447
2448         info->tcpi_total_retrans = tp->total_retrans;
2449 }
2450
2451 EXPORT_SYMBOL_GPL(tcp_get_info);
2452
2453 static int do_tcp_getsockopt(struct sock *sk, int level,
2454                 int optname, char __user *optval, int __user *optlen)
2455 {
2456         struct inet_connection_sock *icsk = inet_csk(sk);
2457         struct tcp_sock *tp = tcp_sk(sk);
2458         int val, len;
2459
2460         if (get_user(len, optlen))
2461                 return -EFAULT;
2462
2463         len = min_t(unsigned int, len, sizeof(int));
2464
2465         if (len < 0)
2466                 return -EINVAL;
2467
2468         switch (optname) {
2469         case TCP_MAXSEG:
2470                 val = tp->mss_cache;
2471                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2472                         val = tp->rx_opt.user_mss;
2473                 break;
2474         case TCP_NODELAY:
2475                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2476                 break;
2477         case TCP_CORK:
2478                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2479                 break;
2480         case TCP_KEEPIDLE:
2481                 val = keepalive_time_when(tp) / HZ;
2482                 break;
2483         case TCP_KEEPINTVL:
2484                 val = keepalive_intvl_when(tp) / HZ;
2485                 break;
2486         case TCP_KEEPCNT:
2487                 val = keepalive_probes(tp);
2488                 break;
2489         case TCP_SYNCNT:
2490                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2491                 break;
2492         case TCP_LINGER2:
2493                 val = tp->linger2;
2494                 if (val >= 0)
2495                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2496                 break;
2497         case TCP_DEFER_ACCEPT:
2498                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2499                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2500                 break;
2501         case TCP_WINDOW_CLAMP:
2502                 val = tp->window_clamp;
2503                 break;
2504         case TCP_INFO: {
2505                 struct tcp_info info;
2506
2507                 if (get_user(len, optlen))
2508                         return -EFAULT;
2509
2510                 tcp_get_info(sk, &info);
2511
2512                 len = min_t(unsigned int, len, sizeof(info));
2513                 if (put_user(len, optlen))
2514                         return -EFAULT;
2515                 if (copy_to_user(optval, &info, len))
2516                         return -EFAULT;
2517                 return 0;
2518         }
2519         case TCP_QUICKACK:
2520                 val = !icsk->icsk_ack.pingpong;
2521                 break;
2522
2523         case TCP_CONGESTION:
2524                 if (get_user(len, optlen))
2525                         return -EFAULT;
2526                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2527                 if (put_user(len, optlen))
2528                         return -EFAULT;
2529                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2530                         return -EFAULT;
2531                 return 0;
2532
2533         case TCP_COOKIE_TRANSACTIONS: {
2534                 struct tcp_cookie_transactions ctd;
2535                 struct tcp_cookie_values *cvp = tp->cookie_values;
2536
2537                 if (get_user(len, optlen))
2538                         return -EFAULT;
2539                 if (len < sizeof(ctd))
2540                         return -EINVAL;
2541
2542                 memset(&ctd, 0, sizeof(ctd));
2543                 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2544                                    TCP_COOKIE_IN_ALWAYS : 0)
2545                                 | (tp->rx_opt.cookie_out_never ?
2546                                    TCP_COOKIE_OUT_NEVER : 0);
2547
2548                 if (cvp != NULL) {
2549                         ctd.tcpct_flags |= (cvp->s_data_in ?
2550                                             TCP_S_DATA_IN : 0)
2551                                          | (cvp->s_data_out ?
2552                                             TCP_S_DATA_OUT : 0);
2553
2554                         ctd.tcpct_cookie_desired = cvp->cookie_desired;
2555                         ctd.tcpct_s_data_desired = cvp->s_data_desired;
2556
2557                         memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2558                                cvp->cookie_pair_size);
2559                         ctd.tcpct_used = cvp->cookie_pair_size;
2560                 }
2561
2562                 if (put_user(sizeof(ctd), optlen))
2563                         return -EFAULT;
2564                 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2565                         return -EFAULT;
2566                 return 0;
2567         }
2568         default:
2569                 return -ENOPROTOOPT;
2570         }
2571
2572         if (put_user(len, optlen))
2573                 return -EFAULT;
2574         if (copy_to_user(optval, &val, len))
2575                 return -EFAULT;
2576         return 0;
2577 }
2578
2579 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2580                    int __user *optlen)
2581 {
2582         struct inet_connection_sock *icsk = inet_csk(sk);
2583
2584         if (level != SOL_TCP)
2585                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2586                                                      optval, optlen);
2587         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2588 }
2589
2590 #ifdef CONFIG_COMPAT
2591 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2592                           char __user *optval, int __user *optlen)
2593 {
2594         if (level != SOL_TCP)
2595                 return inet_csk_compat_getsockopt(sk, level, optname,
2596                                                   optval, optlen);
2597         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2598 }
2599
2600 EXPORT_SYMBOL(compat_tcp_getsockopt);
2601 #endif
2602
2603 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2604 {
2605         struct sk_buff *segs = ERR_PTR(-EINVAL);
2606         struct tcphdr *th;
2607         unsigned thlen;
2608         unsigned int seq;
2609         __be32 delta;
2610         unsigned int oldlen;
2611         unsigned int mss;
2612
2613         if (!pskb_may_pull(skb, sizeof(*th)))
2614                 goto out;
2615
2616         th = tcp_hdr(skb);
2617         thlen = th->doff * 4;
2618         if (thlen < sizeof(*th))
2619                 goto out;
2620
2621         if (!pskb_may_pull(skb, thlen))
2622                 goto out;
2623
2624         oldlen = (u16)~skb->len;
2625         __skb_pull(skb, thlen);
2626
2627         mss = skb_shinfo(skb)->gso_size;
2628         if (unlikely(skb->len <= mss))
2629                 goto out;
2630
2631         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2632                 /* Packet is from an untrusted source, reset gso_segs. */
2633                 int type = skb_shinfo(skb)->gso_type;
2634
2635                 if (unlikely(type &
2636                              ~(SKB_GSO_TCPV4 |
2637                                SKB_GSO_DODGY |
2638                                SKB_GSO_TCP_ECN |
2639                                SKB_GSO_TCPV6 |
2640                                0) ||
2641                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2642                         goto out;
2643
2644                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2645
2646                 segs = NULL;
2647                 goto out;
2648         }
2649
2650         segs = skb_segment(skb, features);
2651         if (IS_ERR(segs))
2652                 goto out;
2653
2654         delta = htonl(oldlen + (thlen + mss));
2655
2656         skb = segs;
2657         th = tcp_hdr(skb);
2658         seq = ntohl(th->seq);
2659
2660         do {
2661                 th->fin = th->psh = 0;
2662
2663                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2664                                        (__force u32)delta));
2665                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2666                         th->check =
2667                              csum_fold(csum_partial(skb_transport_header(skb),
2668                                                     thlen, skb->csum));
2669
2670                 seq += mss;
2671                 skb = skb->next;
2672                 th = tcp_hdr(skb);
2673
2674                 th->seq = htonl(seq);
2675                 th->cwr = 0;
2676         } while (skb->next);
2677
2678         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2679                       skb->data_len);
2680         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2681                                 (__force u32)delta));
2682         if (skb->ip_summed != CHECKSUM_PARTIAL)
2683                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2684                                                    thlen, skb->csum));
2685
2686 out:
2687         return segs;
2688 }
2689 EXPORT_SYMBOL(tcp_tso_segment);
2690
2691 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2692 {
2693         struct sk_buff **pp = NULL;
2694         struct sk_buff *p;
2695         struct tcphdr *th;
2696         struct tcphdr *th2;
2697         unsigned int len;
2698         unsigned int thlen;
2699         unsigned int flags;
2700         unsigned int mss = 1;
2701         unsigned int hlen;
2702         unsigned int off;
2703         int flush = 1;
2704         int i;
2705
2706         off = skb_gro_offset(skb);
2707         hlen = off + sizeof(*th);
2708         th = skb_gro_header_fast(skb, off);
2709         if (skb_gro_header_hard(skb, hlen)) {
2710                 th = skb_gro_header_slow(skb, hlen, off);
2711                 if (unlikely(!th))
2712                         goto out;
2713         }
2714
2715         thlen = th->doff * 4;
2716         if (thlen < sizeof(*th))
2717                 goto out;
2718
2719         hlen = off + thlen;
2720         if (skb_gro_header_hard(skb, hlen)) {
2721                 th = skb_gro_header_slow(skb, hlen, off);
2722                 if (unlikely(!th))
2723                         goto out;
2724         }
2725
2726         skb_gro_pull(skb, thlen);
2727
2728         len = skb_gro_len(skb);
2729         flags = tcp_flag_word(th);
2730
2731         for (; (p = *head); head = &p->next) {
2732                 if (!NAPI_GRO_CB(p)->same_flow)
2733                         continue;
2734
2735                 th2 = tcp_hdr(p);
2736
2737                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2738                         NAPI_GRO_CB(p)->same_flow = 0;
2739                         continue;
2740                 }
2741
2742                 goto found;
2743         }
2744
2745         goto out_check_final;
2746
2747 found:
2748         flush = NAPI_GRO_CB(p)->flush;
2749         flush |= flags & TCP_FLAG_CWR;
2750         flush |= (flags ^ tcp_flag_word(th2)) &
2751                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH);
2752         flush |= th->ack_seq ^ th2->ack_seq;
2753         for (i = sizeof(*th); i < thlen; i += 4)
2754                 flush |= *(u32 *)((u8 *)th + i) ^
2755                          *(u32 *)((u8 *)th2 + i);
2756
2757         mss = skb_shinfo(p)->gso_size;
2758
2759         flush |= (len - 1) >= mss;
2760         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2761
2762         if (flush || skb_gro_receive(head, skb)) {
2763                 mss = 1;
2764                 goto out_check_final;
2765         }
2766
2767         p = *head;
2768         th2 = tcp_hdr(p);
2769         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2770
2771 out_check_final:
2772         flush = len < mss;
2773         flush |= flags & (TCP_FLAG_URG | TCP_FLAG_PSH | TCP_FLAG_RST |
2774                           TCP_FLAG_SYN | TCP_FLAG_FIN);
2775
2776         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2777                 pp = head;
2778
2779 out:
2780         NAPI_GRO_CB(skb)->flush |= flush;
2781
2782         return pp;
2783 }
2784 EXPORT_SYMBOL(tcp_gro_receive);
2785
2786 int tcp_gro_complete(struct sk_buff *skb)
2787 {
2788         struct tcphdr *th = tcp_hdr(skb);
2789
2790         skb->csum_start = skb_transport_header(skb) - skb->head;
2791         skb->csum_offset = offsetof(struct tcphdr, check);
2792         skb->ip_summed = CHECKSUM_PARTIAL;
2793
2794         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2795
2796         if (th->cwr)
2797                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2798
2799         return 0;
2800 }
2801 EXPORT_SYMBOL(tcp_gro_complete);
2802
2803 #ifdef CONFIG_TCP_MD5SIG
2804 static unsigned long tcp_md5sig_users;
2805 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool;
2806 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2807
2808 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool)
2809 {
2810         int cpu;
2811         for_each_possible_cpu(cpu) {
2812                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2813                 if (p) {
2814                         if (p->md5_desc.tfm)
2815                                 crypto_free_hash(p->md5_desc.tfm);
2816                         kfree(p);
2817                         p = NULL;
2818                 }
2819         }
2820         free_percpu(pool);
2821 }
2822
2823 void tcp_free_md5sig_pool(void)
2824 {
2825         struct tcp_md5sig_pool * __percpu *pool = NULL;
2826
2827         spin_lock_bh(&tcp_md5sig_pool_lock);
2828         if (--tcp_md5sig_users == 0) {
2829                 pool = tcp_md5sig_pool;
2830                 tcp_md5sig_pool = NULL;
2831         }
2832         spin_unlock_bh(&tcp_md5sig_pool_lock);
2833         if (pool)
2834                 __tcp_free_md5sig_pool(pool);
2835 }
2836
2837 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2838
2839 static struct tcp_md5sig_pool * __percpu *
2840 __tcp_alloc_md5sig_pool(struct sock *sk)
2841 {
2842         int cpu;
2843         struct tcp_md5sig_pool * __percpu *pool;
2844
2845         pool = alloc_percpu(struct tcp_md5sig_pool *);
2846         if (!pool)
2847                 return NULL;
2848
2849         for_each_possible_cpu(cpu) {
2850                 struct tcp_md5sig_pool *p;
2851                 struct crypto_hash *hash;
2852
2853                 p = kzalloc(sizeof(*p), sk->sk_allocation);
2854                 if (!p)
2855                         goto out_free;
2856                 *per_cpu_ptr(pool, cpu) = p;
2857
2858                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2859                 if (!hash || IS_ERR(hash))
2860                         goto out_free;
2861
2862                 p->md5_desc.tfm = hash;
2863         }
2864         return pool;
2865 out_free:
2866         __tcp_free_md5sig_pool(pool);
2867         return NULL;
2868 }
2869
2870 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2871 {
2872         struct tcp_md5sig_pool * __percpu *pool;
2873         int alloc = 0;
2874
2875 retry:
2876         spin_lock_bh(&tcp_md5sig_pool_lock);
2877         pool = tcp_md5sig_pool;
2878         if (tcp_md5sig_users++ == 0) {
2879                 alloc = 1;
2880                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2881         } else if (!pool) {
2882                 tcp_md5sig_users--;
2883                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2884                 cpu_relax();
2885                 goto retry;
2886         } else
2887                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2888
2889         if (alloc) {
2890                 /* we cannot hold spinlock here because this may sleep. */
2891                 struct tcp_md5sig_pool * __percpu *p;
2892
2893                 p = __tcp_alloc_md5sig_pool(sk);
2894                 spin_lock_bh(&tcp_md5sig_pool_lock);
2895                 if (!p) {
2896                         tcp_md5sig_users--;
2897                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2898                         return NULL;
2899                 }
2900                 pool = tcp_md5sig_pool;
2901                 if (pool) {
2902                         /* oops, it has already been assigned. */
2903                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2904                         __tcp_free_md5sig_pool(p);
2905                 } else {
2906                         tcp_md5sig_pool = pool = p;
2907                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2908                 }
2909         }
2910         return pool;
2911 }
2912
2913 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2914
2915 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2916 {
2917         struct tcp_md5sig_pool * __percpu *p;
2918         spin_lock_bh(&tcp_md5sig_pool_lock);
2919         p = tcp_md5sig_pool;
2920         if (p)
2921                 tcp_md5sig_users++;
2922         spin_unlock_bh(&tcp_md5sig_pool_lock);
2923         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2924 }
2925
2926 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2927
2928 void __tcp_put_md5sig_pool(void)
2929 {
2930         tcp_free_md5sig_pool();
2931 }
2932
2933 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2934
2935 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2936                         struct tcphdr *th)
2937 {
2938         struct scatterlist sg;
2939         int err;
2940
2941         __sum16 old_checksum = th->check;
2942         th->check = 0;
2943         /* options aren't included in the hash */
2944         sg_init_one(&sg, th, sizeof(struct tcphdr));
2945         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2946         th->check = old_checksum;
2947         return err;
2948 }
2949
2950 EXPORT_SYMBOL(tcp_md5_hash_header);
2951
2952 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2953                           struct sk_buff *skb, unsigned header_len)
2954 {
2955         struct scatterlist sg;
2956         const struct tcphdr *tp = tcp_hdr(skb);
2957         struct hash_desc *desc = &hp->md5_desc;
2958         unsigned i;
2959         const unsigned head_data_len = skb_headlen(skb) > header_len ?
2960                                        skb_headlen(skb) - header_len : 0;
2961         const struct skb_shared_info *shi = skb_shinfo(skb);
2962
2963         sg_init_table(&sg, 1);
2964
2965         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2966         if (crypto_hash_update(desc, &sg, head_data_len))
2967                 return 1;
2968
2969         for (i = 0; i < shi->nr_frags; ++i) {
2970                 const struct skb_frag_struct *f = &shi->frags[i];
2971                 sg_set_page(&sg, f->page, f->size, f->page_offset);
2972                 if (crypto_hash_update(desc, &sg, f->size))
2973                         return 1;
2974         }
2975
2976         return 0;
2977 }
2978
2979 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2980
2981 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
2982 {
2983         struct scatterlist sg;
2984
2985         sg_init_one(&sg, key->key, key->keylen);
2986         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2987 }
2988
2989 EXPORT_SYMBOL(tcp_md5_hash_key);
2990
2991 #endif
2992
2993 /**
2994  * Each Responder maintains up to two secret values concurrently for
2995  * efficient secret rollover.  Each secret value has 4 states:
2996  *
2997  * Generating.  (tcp_secret_generating != tcp_secret_primary)
2998  *    Generates new Responder-Cookies, but not yet used for primary
2999  *    verification.  This is a short-term state, typically lasting only
3000  *    one round trip time (RTT).
3001  *
3002  * Primary.  (tcp_secret_generating == tcp_secret_primary)
3003  *    Used both for generation and primary verification.
3004  *
3005  * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3006  *    Used for verification, until the first failure that can be
3007  *    verified by the newer Generating secret.  At that time, this
3008  *    cookie's state is changed to Secondary, and the Generating
3009  *    cookie's state is changed to Primary.  This is a short-term state,
3010  *    typically lasting only one round trip time (RTT).
3011  *
3012  * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3013  *    Used for secondary verification, after primary verification
3014  *    failures.  This state lasts no more than twice the Maximum Segment
3015  *    Lifetime (2MSL).  Then, the secret is discarded.
3016  */
3017 struct tcp_cookie_secret {
3018         /* The secret is divided into two parts.  The digest part is the
3019          * equivalent of previously hashing a secret and saving the state,
3020          * and serves as an initialization vector (IV).  The message part
3021          * serves as the trailing secret.
3022          */
3023         u32                             secrets[COOKIE_WORKSPACE_WORDS];
3024         unsigned long                   expires;
3025 };
3026
3027 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3028 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3029 #define TCP_SECRET_LIFE (HZ * 600)
3030
3031 static struct tcp_cookie_secret tcp_secret_one;
3032 static struct tcp_cookie_secret tcp_secret_two;
3033
3034 /* Essentially a circular list, without dynamic allocation. */
3035 static struct tcp_cookie_secret *tcp_secret_generating;
3036 static struct tcp_cookie_secret *tcp_secret_primary;
3037 static struct tcp_cookie_secret *tcp_secret_retiring;
3038 static struct tcp_cookie_secret *tcp_secret_secondary;
3039
3040 static DEFINE_SPINLOCK(tcp_secret_locker);
3041
3042 /* Select a pseudo-random word in the cookie workspace.
3043  */
3044 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3045 {
3046         return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3047 }
3048
3049 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3050  * Called in softirq context.
3051  * Returns: 0 for success.
3052  */
3053 int tcp_cookie_generator(u32 *bakery)
3054 {
3055         unsigned long jiffy = jiffies;
3056
3057         if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3058                 spin_lock_bh(&tcp_secret_locker);
3059                 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3060                         /* refreshed by another */
3061                         memcpy(bakery,
3062                                &tcp_secret_generating->secrets[0],
3063                                COOKIE_WORKSPACE_WORDS);
3064                 } else {
3065                         /* still needs refreshing */
3066                         get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3067
3068                         /* The first time, paranoia assumes that the
3069                          * randomization function isn't as strong.  But,
3070                          * this secret initialization is delayed until
3071                          * the last possible moment (packet arrival).
3072                          * Although that time is observable, it is
3073                          * unpredictably variable.  Mash in the most
3074                          * volatile clock bits available, and expire the
3075                          * secret extra quickly.
3076                          */
3077                         if (unlikely(tcp_secret_primary->expires ==
3078                                      tcp_secret_secondary->expires)) {
3079                                 struct timespec tv;
3080
3081                                 getnstimeofday(&tv);
3082                                 bakery[COOKIE_DIGEST_WORDS+0] ^=
3083                                         (u32)tv.tv_nsec;
3084
3085                                 tcp_secret_secondary->expires = jiffy
3086                                         + TCP_SECRET_1MSL
3087                                         + (0x0f & tcp_cookie_work(bakery, 0));
3088                         } else {
3089                                 tcp_secret_secondary->expires = jiffy
3090                                         + TCP_SECRET_LIFE
3091                                         + (0xff & tcp_cookie_work(bakery, 1));
3092                                 tcp_secret_primary->expires = jiffy
3093                                         + TCP_SECRET_2MSL
3094                                         + (0x1f & tcp_cookie_work(bakery, 2));
3095                         }
3096                         memcpy(&tcp_secret_secondary->secrets[0],
3097                                bakery, COOKIE_WORKSPACE_WORDS);
3098
3099                         rcu_assign_pointer(tcp_secret_generating,
3100                                            tcp_secret_secondary);
3101                         rcu_assign_pointer(tcp_secret_retiring,
3102                                            tcp_secret_primary);
3103                         /*
3104                          * Neither call_rcu() nor synchronize_rcu() needed.
3105                          * Retiring data is not freed.  It is replaced after
3106                          * further (locked) pointer updates, and a quiet time
3107                          * (minimum 1MSL, maximum LIFE - 2MSL).
3108                          */
3109                 }
3110                 spin_unlock_bh(&tcp_secret_locker);
3111         } else {
3112                 rcu_read_lock_bh();
3113                 memcpy(bakery,
3114                        &rcu_dereference(tcp_secret_generating)->secrets[0],
3115                        COOKIE_WORKSPACE_WORDS);
3116                 rcu_read_unlock_bh();
3117         }
3118         return 0;
3119 }
3120 EXPORT_SYMBOL(tcp_cookie_generator);
3121
3122 void tcp_done(struct sock *sk)
3123 {
3124         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3125                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3126
3127         tcp_set_state(sk, TCP_CLOSE);
3128         tcp_clear_xmit_timers(sk);
3129
3130         sk->sk_shutdown = SHUTDOWN_MASK;
3131
3132         if (!sock_flag(sk, SOCK_DEAD))
3133                 sk->sk_state_change(sk);
3134         else
3135                 inet_csk_destroy_sock(sk);
3136 }
3137 EXPORT_SYMBOL_GPL(tcp_done);
3138
3139 extern struct tcp_congestion_ops tcp_reno;
3140
3141 static __initdata unsigned long thash_entries;
3142 static int __init set_thash_entries(char *str)
3143 {
3144         if (!str)
3145                 return 0;
3146         thash_entries = simple_strtoul(str, &str, 0);
3147         return 1;
3148 }
3149 __setup("thash_entries=", set_thash_entries);
3150
3151 void __init tcp_init(void)
3152 {
3153         struct sk_buff *skb = NULL;
3154         unsigned long nr_pages, limit;
3155         int order, i, max_share;
3156         unsigned long jiffy = jiffies;
3157
3158         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3159
3160         percpu_counter_init(&tcp_sockets_allocated, 0);
3161         percpu_counter_init(&tcp_orphan_count, 0);
3162         tcp_hashinfo.bind_bucket_cachep =
3163                 kmem_cache_create("tcp_bind_bucket",
3164                                   sizeof(struct inet_bind_bucket), 0,
3165                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3166
3167         /* Size and allocate the main established and bind bucket
3168          * hash tables.
3169          *
3170          * The methodology is similar to that of the buffer cache.
3171          */
3172         tcp_hashinfo.ehash =
3173                 alloc_large_system_hash("TCP established",
3174                                         sizeof(struct inet_ehash_bucket),
3175                                         thash_entries,
3176                                         (totalram_pages >= 128 * 1024) ?
3177                                         13 : 15,
3178                                         0,
3179                                         NULL,
3180                                         &tcp_hashinfo.ehash_mask,
3181                                         thash_entries ? 0 : 512 * 1024);
3182         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3183                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3184                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3185         }
3186         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3187                 panic("TCP: failed to alloc ehash_locks");
3188         tcp_hashinfo.bhash =
3189                 alloc_large_system_hash("TCP bind",
3190                                         sizeof(struct inet_bind_hashbucket),
3191                                         tcp_hashinfo.ehash_mask + 1,
3192                                         (totalram_pages >= 128 * 1024) ?
3193                                         13 : 15,
3194                                         0,
3195                                         &tcp_hashinfo.bhash_size,
3196                                         NULL,
3197                                         64 * 1024);
3198         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3199         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3200                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3201                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3202         }
3203
3204         /* Try to be a bit smarter and adjust defaults depending
3205          * on available memory.
3206          */
3207         for (order = 0; ((1 << order) << PAGE_SHIFT) <
3208                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
3209                         order++)
3210                 ;
3211         if (order >= 4) {
3212                 tcp_death_row.sysctl_max_tw_buckets = 180000;
3213                 sysctl_tcp_max_orphans = 4096 << (order - 4);
3214                 sysctl_max_syn_backlog = 1024;
3215         } else if (order < 3) {
3216                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
3217                 sysctl_tcp_max_orphans >>= (3 - order);
3218                 sysctl_max_syn_backlog = 128;
3219         }
3220
3221         /* Set the pressure threshold to be a fraction of global memory that
3222          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
3223          * memory, with a floor of 128 pages.
3224          */
3225         nr_pages = totalram_pages - totalhigh_pages;
3226         limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
3227         limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
3228         limit = max(limit, 128UL);
3229         sysctl_tcp_mem[0] = limit / 4 * 3;
3230         sysctl_tcp_mem[1] = limit;
3231         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3232
3233         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3234         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3235         max_share = min(4UL*1024*1024, limit);
3236
3237         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3238         sysctl_tcp_wmem[1] = 16*1024;
3239         sysctl_tcp_wmem[2] = max(64*1024, max_share);
3240
3241         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3242         sysctl_tcp_rmem[1] = 87380;
3243         sysctl_tcp_rmem[2] = max(87380, max_share);
3244
3245         printk(KERN_INFO "TCP: Hash tables configured "
3246                "(established %u bind %u)\n",
3247                tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3248
3249         tcp_register_congestion_control(&tcp_reno);
3250
3251         memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3252         memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3253         tcp_secret_one.expires = jiffy; /* past due */
3254         tcp_secret_two.expires = jiffy; /* past due */
3255         tcp_secret_generating = &tcp_secret_one;
3256         tcp_secret_primary = &tcp_secret_one;
3257         tcp_secret_retiring = &tcp_secret_two;
3258         tcp_secret_secondary = &tcp_secret_two;
3259 }
3260
3261 EXPORT_SYMBOL(tcp_close);
3262 EXPORT_SYMBOL(tcp_disconnect);
3263 EXPORT_SYMBOL(tcp_getsockopt);
3264 EXPORT_SYMBOL(tcp_ioctl);
3265 EXPORT_SYMBOL(tcp_poll);
3266 EXPORT_SYMBOL(tcp_read_sock);
3267 EXPORT_SYMBOL(tcp_recvmsg);
3268 EXPORT_SYMBOL(tcp_sendmsg);
3269 EXPORT_SYMBOL(tcp_splice_read);
3270 EXPORT_SYMBOL(tcp_sendpage);
3271 EXPORT_SYMBOL(tcp_setsockopt);
3272 EXPORT_SYMBOL(tcp_shutdown);