vfs: fix panic in __d_lookup() with high dentry hashtable counts
[linux-2.6.git] / net / ipv4 / tcp.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Authors:     Ross Biro
9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
12  *              Florian La Roche, <flla@stud.uni-sb.de>
13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *              Jorge Cwik, <jorge@laser.satlink.net>
19  *
20  * Fixes:
21  *              Alan Cox        :       Numerous verify_area() calls
22  *              Alan Cox        :       Set the ACK bit on a reset
23  *              Alan Cox        :       Stopped it crashing if it closed while
24  *                                      sk->inuse=1 and was trying to connect
25  *                                      (tcp_err()).
26  *              Alan Cox        :       All icmp error handling was broken
27  *                                      pointers passed where wrong and the
28  *                                      socket was looked up backwards. Nobody
29  *                                      tested any icmp error code obviously.
30  *              Alan Cox        :       tcp_err() now handled properly. It
31  *                                      wakes people on errors. poll
32  *                                      behaves and the icmp error race
33  *                                      has gone by moving it into sock.c
34  *              Alan Cox        :       tcp_send_reset() fixed to work for
35  *                                      everything not just packets for
36  *                                      unknown sockets.
37  *              Alan Cox        :       tcp option processing.
38  *              Alan Cox        :       Reset tweaked (still not 100%) [Had
39  *                                      syn rule wrong]
40  *              Herp Rosmanith  :       More reset fixes
41  *              Alan Cox        :       No longer acks invalid rst frames.
42  *                                      Acking any kind of RST is right out.
43  *              Alan Cox        :       Sets an ignore me flag on an rst
44  *                                      receive otherwise odd bits of prattle
45  *                                      escape still
46  *              Alan Cox        :       Fixed another acking RST frame bug.
47  *                                      Should stop LAN workplace lockups.
48  *              Alan Cox        :       Some tidyups using the new skb list
49  *                                      facilities
50  *              Alan Cox        :       sk->keepopen now seems to work
51  *              Alan Cox        :       Pulls options out correctly on accepts
52  *              Alan Cox        :       Fixed assorted sk->rqueue->next errors
53  *              Alan Cox        :       PSH doesn't end a TCP read. Switched a
54  *                                      bit to skb ops.
55  *              Alan Cox        :       Tidied tcp_data to avoid a potential
56  *                                      nasty.
57  *              Alan Cox        :       Added some better commenting, as the
58  *                                      tcp is hard to follow
59  *              Alan Cox        :       Removed incorrect check for 20 * psh
60  *      Michael O'Reilly        :       ack < copied bug fix.
61  *      Johannes Stille         :       Misc tcp fixes (not all in yet).
62  *              Alan Cox        :       FIN with no memory -> CRASH
63  *              Alan Cox        :       Added socket option proto entries.
64  *                                      Also added awareness of them to accept.
65  *              Alan Cox        :       Added TCP options (SOL_TCP)
66  *              Alan Cox        :       Switched wakeup calls to callbacks,
67  *                                      so the kernel can layer network
68  *                                      sockets.
69  *              Alan Cox        :       Use ip_tos/ip_ttl settings.
70  *              Alan Cox        :       Handle FIN (more) properly (we hope).
71  *              Alan Cox        :       RST frames sent on unsynchronised
72  *                                      state ack error.
73  *              Alan Cox        :       Put in missing check for SYN bit.
74  *              Alan Cox        :       Added tcp_select_window() aka NET2E
75  *                                      window non shrink trick.
76  *              Alan Cox        :       Added a couple of small NET2E timer
77  *                                      fixes
78  *              Charles Hedrick :       TCP fixes
79  *              Toomas Tamm     :       TCP window fixes
80  *              Alan Cox        :       Small URG fix to rlogin ^C ack fight
81  *              Charles Hedrick :       Rewrote most of it to actually work
82  *              Linus           :       Rewrote tcp_read() and URG handling
83  *                                      completely
84  *              Gerhard Koerting:       Fixed some missing timer handling
85  *              Matthew Dillon  :       Reworked TCP machine states as per RFC
86  *              Gerhard Koerting:       PC/TCP workarounds
87  *              Adam Caldwell   :       Assorted timer/timing errors
88  *              Matthew Dillon  :       Fixed another RST bug
89  *              Alan Cox        :       Move to kernel side addressing changes.
90  *              Alan Cox        :       Beginning work on TCP fastpathing
91  *                                      (not yet usable)
92  *              Arnt Gulbrandsen:       Turbocharged tcp_check() routine.
93  *              Alan Cox        :       TCP fast path debugging
94  *              Alan Cox        :       Window clamping
95  *              Michael Riepe   :       Bug in tcp_check()
96  *              Matt Dillon     :       More TCP improvements and RST bug fixes
97  *              Matt Dillon     :       Yet more small nasties remove from the
98  *                                      TCP code (Be very nice to this man if
99  *                                      tcp finally works 100%) 8)
100  *              Alan Cox        :       BSD accept semantics.
101  *              Alan Cox        :       Reset on closedown bug.
102  *      Peter De Schrijver      :       ENOTCONN check missing in tcp_sendto().
103  *              Michael Pall    :       Handle poll() after URG properly in
104  *                                      all cases.
105  *              Michael Pall    :       Undo the last fix in tcp_read_urg()
106  *                                      (multi URG PUSH broke rlogin).
107  *              Michael Pall    :       Fix the multi URG PUSH problem in
108  *                                      tcp_readable(), poll() after URG
109  *                                      works now.
110  *              Michael Pall    :       recv(...,MSG_OOB) never blocks in the
111  *                                      BSD api.
112  *              Alan Cox        :       Changed the semantics of sk->socket to
113  *                                      fix a race and a signal problem with
114  *                                      accept() and async I/O.
115  *              Alan Cox        :       Relaxed the rules on tcp_sendto().
116  *              Yury Shevchuk   :       Really fixed accept() blocking problem.
117  *              Craig I. Hagan  :       Allow for BSD compatible TIME_WAIT for
118  *                                      clients/servers which listen in on
119  *                                      fixed ports.
120  *              Alan Cox        :       Cleaned the above up and shrank it to
121  *                                      a sensible code size.
122  *              Alan Cox        :       Self connect lockup fix.
123  *              Alan Cox        :       No connect to multicast.
124  *              Ross Biro       :       Close unaccepted children on master
125  *                                      socket close.
126  *              Alan Cox        :       Reset tracing code.
127  *              Alan Cox        :       Spurious resets on shutdown.
128  *              Alan Cox        :       Giant 15 minute/60 second timer error
129  *              Alan Cox        :       Small whoops in polling before an
130  *                                      accept.
131  *              Alan Cox        :       Kept the state trace facility since
132  *                                      it's handy for debugging.
133  *              Alan Cox        :       More reset handler fixes.
134  *              Alan Cox        :       Started rewriting the code based on
135  *                                      the RFC's for other useful protocol
136  *                                      references see: Comer, KA9Q NOS, and
137  *                                      for a reference on the difference
138  *                                      between specifications and how BSD
139  *                                      works see the 4.4lite source.
140  *              A.N.Kuznetsov   :       Don't time wait on completion of tidy
141  *                                      close.
142  *              Linus Torvalds  :       Fin/Shutdown & copied_seq changes.
143  *              Linus Torvalds  :       Fixed BSD port reuse to work first syn
144  *              Alan Cox        :       Reimplemented timers as per the RFC
145  *                                      and using multiple timers for sanity.
146  *              Alan Cox        :       Small bug fixes, and a lot of new
147  *                                      comments.
148  *              Alan Cox        :       Fixed dual reader crash by locking
149  *                                      the buffers (much like datagram.c)
150  *              Alan Cox        :       Fixed stuck sockets in probe. A probe
151  *                                      now gets fed up of retrying without
152  *                                      (even a no space) answer.
153  *              Alan Cox        :       Extracted closing code better
154  *              Alan Cox        :       Fixed the closing state machine to
155  *                                      resemble the RFC.
156  *              Alan Cox        :       More 'per spec' fixes.
157  *              Jorge Cwik      :       Even faster checksumming.
158  *              Alan Cox        :       tcp_data() doesn't ack illegal PSH
159  *                                      only frames. At least one pc tcp stack
160  *                                      generates them.
161  *              Alan Cox        :       Cache last socket.
162  *              Alan Cox        :       Per route irtt.
163  *              Matt Day        :       poll()->select() match BSD precisely on error
164  *              Alan Cox        :       New buffers
165  *              Marc Tamsky     :       Various sk->prot->retransmits and
166  *                                      sk->retransmits misupdating fixed.
167  *                                      Fixed tcp_write_timeout: stuck close,
168  *                                      and TCP syn retries gets used now.
169  *              Mark Yarvis     :       In tcp_read_wakeup(), don't send an
170  *                                      ack if state is TCP_CLOSED.
171  *              Alan Cox        :       Look up device on a retransmit - routes may
172  *                                      change. Doesn't yet cope with MSS shrink right
173  *                                      but it's a start!
174  *              Marc Tamsky     :       Closing in closing fixes.
175  *              Mike Shaver     :       RFC1122 verifications.
176  *              Alan Cox        :       rcv_saddr errors.
177  *              Alan Cox        :       Block double connect().
178  *              Alan Cox        :       Small hooks for enSKIP.
179  *              Alexey Kuznetsov:       Path MTU discovery.
180  *              Alan Cox        :       Support soft errors.
181  *              Alan Cox        :       Fix MTU discovery pathological case
182  *                                      when the remote claims no mtu!
183  *              Marc Tamsky     :       TCP_CLOSE fix.
184  *              Colin (G3TNE)   :       Send a reset on syn ack replies in
185  *                                      window but wrong (fixes NT lpd problems)
186  *              Pedro Roque     :       Better TCP window handling, delayed ack.
187  *              Joerg Reuter    :       No modification of locked buffers in
188  *                                      tcp_do_retransmit()
189  *              Eric Schenk     :       Changed receiver side silly window
190  *                                      avoidance algorithm to BSD style
191  *                                      algorithm. This doubles throughput
192  *                                      against machines running Solaris,
193  *                                      and seems to result in general
194  *                                      improvement.
195  *      Stefan Magdalinski      :       adjusted tcp_readable() to fix FIONREAD
196  *      Willy Konynenberg       :       Transparent proxying support.
197  *      Mike McLagan            :       Routing by source
198  *              Keith Owens     :       Do proper merging with partial SKB's in
199  *                                      tcp_do_sendmsg to avoid burstiness.
200  *              Eric Schenk     :       Fix fast close down bug with
201  *                                      shutdown() followed by close().
202  *              Andi Kleen      :       Make poll agree with SIGIO
203  *      Salvatore Sanfilippo    :       Support SO_LINGER with linger == 1 and
204  *                                      lingertime == 0 (RFC 793 ABORT Call)
205  *      Hirokazu Takahashi      :       Use copy_from_user() instead of
206  *                                      csum_and_copy_from_user() if possible.
207  *
208  *              This program is free software; you can redistribute it and/or
209  *              modify it under the terms of the GNU General Public License
210  *              as published by the Free Software Foundation; either version
211  *              2 of the License, or(at your option) any later version.
212  *
213  * Description of States:
214  *
215  *      TCP_SYN_SENT            sent a connection request, waiting for ack
216  *
217  *      TCP_SYN_RECV            received a connection request, sent ack,
218  *                              waiting for final ack in three-way handshake.
219  *
220  *      TCP_ESTABLISHED         connection established
221  *
222  *      TCP_FIN_WAIT1           our side has shutdown, waiting to complete
223  *                              transmission of remaining buffered data
224  *
225  *      TCP_FIN_WAIT2           all buffered data sent, waiting for remote
226  *                              to shutdown
227  *
228  *      TCP_CLOSING             both sides have shutdown but we still have
229  *                              data we have to finish sending
230  *
231  *      TCP_TIME_WAIT           timeout to catch resent junk before entering
232  *                              closed, can only be entered from FIN_WAIT2
233  *                              or CLOSING.  Required because the other end
234  *                              may not have gotten our last ACK causing it
235  *                              to retransmit the data packet (which we ignore)
236  *
237  *      TCP_CLOSE_WAIT          remote side has shutdown and is waiting for
238  *                              us to finish writing our data and to shutdown
239  *                              (we have to close() to move on to LAST_ACK)
240  *
241  *      TCP_LAST_ACK            out side has shutdown after remote has
242  *                              shutdown.  There may still be data in our
243  *                              buffer that we have to finish sending
244  *
245  *      TCP_CLOSE               socket is finished
246  */
247
248 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
267 #include <linux/time.h>
268 #include <linux/slab.h>
269
270 #include <net/icmp.h>
271 #include <net/tcp.h>
272 #include <net/xfrm.h>
273 #include <net/ip.h>
274 #include <net/netdma.h>
275 #include <net/sock.h>
276
277 #include <asm/uaccess.h>
278 #include <asm/ioctls.h>
279
280 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
281
282 struct percpu_counter tcp_orphan_count;
283 EXPORT_SYMBOL_GPL(tcp_orphan_count);
284
285 int sysctl_tcp_wmem[3] __read_mostly;
286 int sysctl_tcp_rmem[3] __read_mostly;
287
288 EXPORT_SYMBOL(sysctl_tcp_rmem);
289 EXPORT_SYMBOL(sysctl_tcp_wmem);
290
291 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
292 EXPORT_SYMBOL(tcp_memory_allocated);
293
294 /*
295  * Current number of TCP sockets.
296  */
297 struct percpu_counter tcp_sockets_allocated;
298 EXPORT_SYMBOL(tcp_sockets_allocated);
299
300 /*
301  * TCP splice context
302  */
303 struct tcp_splice_state {
304         struct pipe_inode_info *pipe;
305         size_t len;
306         unsigned int flags;
307 };
308
309 /*
310  * Pressure flag: try to collapse.
311  * Technical note: it is used by multiple contexts non atomically.
312  * All the __sk_mem_schedule() is of this nature: accounting
313  * is strict, actions are advisory and have some latency.
314  */
315 int tcp_memory_pressure __read_mostly;
316 EXPORT_SYMBOL(tcp_memory_pressure);
317
318 void tcp_enter_memory_pressure(struct sock *sk)
319 {
320         if (!tcp_memory_pressure) {
321                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
322                 tcp_memory_pressure = 1;
323         }
324 }
325 EXPORT_SYMBOL(tcp_enter_memory_pressure);
326
327 /* Convert seconds to retransmits based on initial and max timeout */
328 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
329 {
330         u8 res = 0;
331
332         if (seconds > 0) {
333                 int period = timeout;
334
335                 res = 1;
336                 while (seconds > period && res < 255) {
337                         res++;
338                         timeout <<= 1;
339                         if (timeout > rto_max)
340                                 timeout = rto_max;
341                         period += timeout;
342                 }
343         }
344         return res;
345 }
346
347 /* Convert retransmits to seconds based on initial and max timeout */
348 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
349 {
350         int period = 0;
351
352         if (retrans > 0) {
353                 period = timeout;
354                 while (--retrans) {
355                         timeout <<= 1;
356                         if (timeout > rto_max)
357                                 timeout = rto_max;
358                         period += timeout;
359                 }
360         }
361         return period;
362 }
363
364 /*
365  *      Wait for a TCP event.
366  *
367  *      Note that we don't need to lock the socket, as the upper poll layers
368  *      take care of normal races (between the test and the event) and we don't
369  *      go look at any of the socket buffers directly.
370  */
371 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
372 {
373         unsigned int mask;
374         struct sock *sk = sock->sk;
375         const struct tcp_sock *tp = tcp_sk(sk);
376
377         sock_poll_wait(file, sk_sleep(sk), wait);
378         if (sk->sk_state == TCP_LISTEN)
379                 return inet_csk_listen_poll(sk);
380
381         /* Socket is not locked. We are protected from async events
382          * by poll logic and correct handling of state changes
383          * made by other threads is impossible in any case.
384          */
385
386         mask = 0;
387
388         /*
389          * POLLHUP is certainly not done right. But poll() doesn't
390          * have a notion of HUP in just one direction, and for a
391          * socket the read side is more interesting.
392          *
393          * Some poll() documentation says that POLLHUP is incompatible
394          * with the POLLOUT/POLLWR flags, so somebody should check this
395          * all. But careful, it tends to be safer to return too many
396          * bits than too few, and you can easily break real applications
397          * if you don't tell them that something has hung up!
398          *
399          * Check-me.
400          *
401          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
402          * our fs/select.c). It means that after we received EOF,
403          * poll always returns immediately, making impossible poll() on write()
404          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
405          * if and only if shutdown has been made in both directions.
406          * Actually, it is interesting to look how Solaris and DUX
407          * solve this dilemma. I would prefer, if POLLHUP were maskable,
408          * then we could set it on SND_SHUTDOWN. BTW examples given
409          * in Stevens' books assume exactly this behaviour, it explains
410          * why POLLHUP is incompatible with POLLOUT.    --ANK
411          *
412          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
413          * blocking on fresh not-connected or disconnected socket. --ANK
414          */
415         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
416                 mask |= POLLHUP;
417         if (sk->sk_shutdown & RCV_SHUTDOWN)
418                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
419
420         /* Connected? */
421         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
422                 int target = sock_rcvlowat(sk, 0, INT_MAX);
423
424                 if (tp->urg_seq == tp->copied_seq &&
425                     !sock_flag(sk, SOCK_URGINLINE) &&
426                     tp->urg_data)
427                         target++;
428
429                 /* Potential race condition. If read of tp below will
430                  * escape above sk->sk_state, we can be illegally awaken
431                  * in SYN_* states. */
432                 if (tp->rcv_nxt - tp->copied_seq >= target)
433                         mask |= POLLIN | POLLRDNORM;
434
435                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
436                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
437                                 mask |= POLLOUT | POLLWRNORM;
438                         } else {  /* send SIGIO later */
439                                 set_bit(SOCK_ASYNC_NOSPACE,
440                                         &sk->sk_socket->flags);
441                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
442
443                                 /* Race breaker. If space is freed after
444                                  * wspace test but before the flags are set,
445                                  * IO signal will be lost.
446                                  */
447                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
448                                         mask |= POLLOUT | POLLWRNORM;
449                         }
450                 } else
451                         mask |= POLLOUT | POLLWRNORM;
452
453                 if (tp->urg_data & TCP_URG_VALID)
454                         mask |= POLLPRI;
455         }
456         /* This barrier is coupled with smp_wmb() in tcp_reset() */
457         smp_rmb();
458         if (sk->sk_err)
459                 mask |= POLLERR;
460
461         return mask;
462 }
463 EXPORT_SYMBOL(tcp_poll);
464
465 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
466 {
467         struct tcp_sock *tp = tcp_sk(sk);
468         int answ;
469
470         switch (cmd) {
471         case SIOCINQ:
472                 if (sk->sk_state == TCP_LISTEN)
473                         return -EINVAL;
474
475                 lock_sock(sk);
476                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
477                         answ = 0;
478                 else if (sock_flag(sk, SOCK_URGINLINE) ||
479                          !tp->urg_data ||
480                          before(tp->urg_seq, tp->copied_seq) ||
481                          !before(tp->urg_seq, tp->rcv_nxt)) {
482                         struct sk_buff *skb;
483
484                         answ = tp->rcv_nxt - tp->copied_seq;
485
486                         /* Subtract 1, if FIN is in queue. */
487                         skb = skb_peek_tail(&sk->sk_receive_queue);
488                         if (answ && skb)
489                                 answ -= tcp_hdr(skb)->fin;
490                 } else
491                         answ = tp->urg_seq - tp->copied_seq;
492                 release_sock(sk);
493                 break;
494         case SIOCATMARK:
495                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
496                 break;
497         case SIOCOUTQ:
498                 if (sk->sk_state == TCP_LISTEN)
499                         return -EINVAL;
500
501                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
502                         answ = 0;
503                 else
504                         answ = tp->write_seq - tp->snd_una;
505                 break;
506         case SIOCOUTQNSD:
507                 if (sk->sk_state == TCP_LISTEN)
508                         return -EINVAL;
509
510                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
511                         answ = 0;
512                 else
513                         answ = tp->write_seq - tp->snd_nxt;
514                 break;
515         default:
516                 return -ENOIOCTLCMD;
517         }
518
519         return put_user(answ, (int __user *)arg);
520 }
521 EXPORT_SYMBOL(tcp_ioctl);
522
523 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
524 {
525         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
526         tp->pushed_seq = tp->write_seq;
527 }
528
529 static inline int forced_push(const struct tcp_sock *tp)
530 {
531         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
532 }
533
534 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
535 {
536         struct tcp_sock *tp = tcp_sk(sk);
537         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
538
539         skb->csum    = 0;
540         tcb->seq     = tcb->end_seq = tp->write_seq;
541         tcb->tcp_flags = TCPHDR_ACK;
542         tcb->sacked  = 0;
543         skb_header_release(skb);
544         tcp_add_write_queue_tail(sk, skb);
545         sk->sk_wmem_queued += skb->truesize;
546         sk_mem_charge(sk, skb->truesize);
547         if (tp->nonagle & TCP_NAGLE_PUSH)
548                 tp->nonagle &= ~TCP_NAGLE_PUSH;
549 }
550
551 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
552 {
553         if (flags & MSG_OOB)
554                 tp->snd_up = tp->write_seq;
555 }
556
557 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
558                             int nonagle)
559 {
560         if (tcp_send_head(sk)) {
561                 struct tcp_sock *tp = tcp_sk(sk);
562
563                 if (!(flags & MSG_MORE) || forced_push(tp))
564                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
565
566                 tcp_mark_urg(tp, flags);
567                 __tcp_push_pending_frames(sk, mss_now,
568                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
569         }
570 }
571
572 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
573                                 unsigned int offset, size_t len)
574 {
575         struct tcp_splice_state *tss = rd_desc->arg.data;
576         int ret;
577
578         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
579                               tss->flags);
580         if (ret > 0)
581                 rd_desc->count -= ret;
582         return ret;
583 }
584
585 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
586 {
587         /* Store TCP splice context information in read_descriptor_t. */
588         read_descriptor_t rd_desc = {
589                 .arg.data = tss,
590                 .count    = tss->len,
591         };
592
593         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
594 }
595
596 /**
597  *  tcp_splice_read - splice data from TCP socket to a pipe
598  * @sock:       socket to splice from
599  * @ppos:       position (not valid)
600  * @pipe:       pipe to splice to
601  * @len:        number of bytes to splice
602  * @flags:      splice modifier flags
603  *
604  * Description:
605  *    Will read pages from given socket and fill them into a pipe.
606  *
607  **/
608 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
609                         struct pipe_inode_info *pipe, size_t len,
610                         unsigned int flags)
611 {
612         struct sock *sk = sock->sk;
613         struct tcp_splice_state tss = {
614                 .pipe = pipe,
615                 .len = len,
616                 .flags = flags,
617         };
618         long timeo;
619         ssize_t spliced;
620         int ret;
621
622         sock_rps_record_flow(sk);
623         /*
624          * We can't seek on a socket input
625          */
626         if (unlikely(*ppos))
627                 return -ESPIPE;
628
629         ret = spliced = 0;
630
631         lock_sock(sk);
632
633         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
634         while (tss.len) {
635                 ret = __tcp_splice_read(sk, &tss);
636                 if (ret < 0)
637                         break;
638                 else if (!ret) {
639                         if (spliced)
640                                 break;
641                         if (sock_flag(sk, SOCK_DONE))
642                                 break;
643                         if (sk->sk_err) {
644                                 ret = sock_error(sk);
645                                 break;
646                         }
647                         if (sk->sk_shutdown & RCV_SHUTDOWN)
648                                 break;
649                         if (sk->sk_state == TCP_CLOSE) {
650                                 /*
651                                  * This occurs when user tries to read
652                                  * from never connected socket.
653                                  */
654                                 if (!sock_flag(sk, SOCK_DONE))
655                                         ret = -ENOTCONN;
656                                 break;
657                         }
658                         if (!timeo) {
659                                 ret = -EAGAIN;
660                                 break;
661                         }
662                         sk_wait_data(sk, &timeo);
663                         if (signal_pending(current)) {
664                                 ret = sock_intr_errno(timeo);
665                                 break;
666                         }
667                         continue;
668                 }
669                 tss.len -= ret;
670                 spliced += ret;
671
672                 if (!timeo)
673                         break;
674                 release_sock(sk);
675                 lock_sock(sk);
676
677                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
678                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
679                     signal_pending(current))
680                         break;
681         }
682
683         release_sock(sk);
684
685         if (spliced)
686                 return spliced;
687
688         return ret;
689 }
690 EXPORT_SYMBOL(tcp_splice_read);
691
692 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
693 {
694         struct sk_buff *skb;
695
696         /* The TCP header must be at least 32-bit aligned.  */
697         size = ALIGN(size, 4);
698
699         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
700         if (skb) {
701                 if (sk_wmem_schedule(sk, skb->truesize)) {
702                         /*
703                          * Make sure that we have exactly size bytes
704                          * available to the caller, no more, no less.
705                          */
706                         skb_reserve(skb, skb_tailroom(skb) - size);
707                         return skb;
708                 }
709                 __kfree_skb(skb);
710         } else {
711                 sk->sk_prot->enter_memory_pressure(sk);
712                 sk_stream_moderate_sndbuf(sk);
713         }
714         return NULL;
715 }
716
717 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
718                                        int large_allowed)
719 {
720         struct tcp_sock *tp = tcp_sk(sk);
721         u32 xmit_size_goal, old_size_goal;
722
723         xmit_size_goal = mss_now;
724
725         if (large_allowed && sk_can_gso(sk)) {
726                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
727                                   inet_csk(sk)->icsk_af_ops->net_header_len -
728                                   inet_csk(sk)->icsk_ext_hdr_len -
729                                   tp->tcp_header_len);
730
731                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
732
733                 /* We try hard to avoid divides here */
734                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
735
736                 if (likely(old_size_goal <= xmit_size_goal &&
737                            old_size_goal + mss_now > xmit_size_goal)) {
738                         xmit_size_goal = old_size_goal;
739                 } else {
740                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
741                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
742                 }
743         }
744
745         return max(xmit_size_goal, mss_now);
746 }
747
748 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
749 {
750         int mss_now;
751
752         mss_now = tcp_current_mss(sk);
753         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
754
755         return mss_now;
756 }
757
758 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
759                          size_t psize, int flags)
760 {
761         struct tcp_sock *tp = tcp_sk(sk);
762         int mss_now, size_goal;
763         int err;
764         ssize_t copied;
765         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
766
767         /* Wait for a connection to finish. */
768         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
769                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
770                         goto out_err;
771
772         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
773
774         mss_now = tcp_send_mss(sk, &size_goal, flags);
775         copied = 0;
776
777         err = -EPIPE;
778         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
779                 goto out_err;
780
781         while (psize > 0) {
782                 struct sk_buff *skb = tcp_write_queue_tail(sk);
783                 struct page *page = pages[poffset / PAGE_SIZE];
784                 int copy, i, can_coalesce;
785                 int offset = poffset % PAGE_SIZE;
786                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
787
788                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
789 new_segment:
790                         if (!sk_stream_memory_free(sk))
791                                 goto wait_for_sndbuf;
792
793                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
794                         if (!skb)
795                                 goto wait_for_memory;
796
797                         skb_entail(sk, skb);
798                         copy = size_goal;
799                 }
800
801                 if (copy > size)
802                         copy = size;
803
804                 i = skb_shinfo(skb)->nr_frags;
805                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
806                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
807                         tcp_mark_push(tp, skb);
808                         goto new_segment;
809                 }
810                 if (!sk_wmem_schedule(sk, copy))
811                         goto wait_for_memory;
812
813                 if (can_coalesce) {
814                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
815                 } else {
816                         get_page(page);
817                         skb_fill_page_desc(skb, i, page, offset, copy);
818                 }
819
820                 skb->len += copy;
821                 skb->data_len += copy;
822                 skb->truesize += copy;
823                 sk->sk_wmem_queued += copy;
824                 sk_mem_charge(sk, copy);
825                 skb->ip_summed = CHECKSUM_PARTIAL;
826                 tp->write_seq += copy;
827                 TCP_SKB_CB(skb)->end_seq += copy;
828                 skb_shinfo(skb)->gso_segs = 0;
829
830                 if (!copied)
831                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
832
833                 copied += copy;
834                 poffset += copy;
835                 if (!(psize -= copy))
836                         goto out;
837
838                 if (skb->len < size_goal || (flags & MSG_OOB))
839                         continue;
840
841                 if (forced_push(tp)) {
842                         tcp_mark_push(tp, skb);
843                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
844                 } else if (skb == tcp_send_head(sk))
845                         tcp_push_one(sk, mss_now);
846                 continue;
847
848 wait_for_sndbuf:
849                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
850 wait_for_memory:
851                 if (copied)
852                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
853
854                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
855                         goto do_error;
856
857                 mss_now = tcp_send_mss(sk, &size_goal, flags);
858         }
859
860 out:
861         if (copied)
862                 tcp_push(sk, flags, mss_now, tp->nonagle);
863         return copied;
864
865 do_error:
866         if (copied)
867                 goto out;
868 out_err:
869         return sk_stream_error(sk, flags, err);
870 }
871
872 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
873                  size_t size, int flags)
874 {
875         ssize_t res;
876
877         if (!(sk->sk_route_caps & NETIF_F_SG) ||
878             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
879                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
880                                         flags);
881
882         lock_sock(sk);
883         res = do_tcp_sendpages(sk, &page, offset, size, flags);
884         release_sock(sk);
885         return res;
886 }
887 EXPORT_SYMBOL(tcp_sendpage);
888
889 static inline int select_size(const struct sock *sk, bool sg)
890 {
891         const struct tcp_sock *tp = tcp_sk(sk);
892         int tmp = tp->mss_cache;
893
894         if (sg) {
895                 if (sk_can_gso(sk)) {
896                         /* Small frames wont use a full page:
897                          * Payload will immediately follow tcp header.
898                          */
899                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
900                 } else {
901                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
902
903                         if (tmp >= pgbreak &&
904                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
905                                 tmp = pgbreak;
906                 }
907         }
908
909         return tmp;
910 }
911
912 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
913                 size_t size)
914 {
915         struct iovec *iov;
916         struct tcp_sock *tp = tcp_sk(sk);
917         struct sk_buff *skb;
918         int iovlen, flags, err, copied;
919         int mss_now, size_goal;
920         bool sg;
921         long timeo;
922
923         lock_sock(sk);
924
925         flags = msg->msg_flags;
926         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
927
928         /* Wait for a connection to finish. */
929         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
930                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
931                         goto out_err;
932
933         /* This should be in poll */
934         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
935
936         mss_now = tcp_send_mss(sk, &size_goal, flags);
937
938         /* Ok commence sending. */
939         iovlen = msg->msg_iovlen;
940         iov = msg->msg_iov;
941         copied = 0;
942
943         err = -EPIPE;
944         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
945                 goto out_err;
946
947         sg = !!(sk->sk_route_caps & NETIF_F_SG);
948
949         while (--iovlen >= 0) {
950                 size_t seglen = iov->iov_len;
951                 unsigned char __user *from = iov->iov_base;
952
953                 iov++;
954
955                 while (seglen > 0) {
956                         int copy = 0;
957                         int max = size_goal;
958
959                         skb = tcp_write_queue_tail(sk);
960                         if (tcp_send_head(sk)) {
961                                 if (skb->ip_summed == CHECKSUM_NONE)
962                                         max = mss_now;
963                                 copy = max - skb->len;
964                         }
965
966                         if (copy <= 0) {
967 new_segment:
968                                 /* Allocate new segment. If the interface is SG,
969                                  * allocate skb fitting to single page.
970                                  */
971                                 if (!sk_stream_memory_free(sk))
972                                         goto wait_for_sndbuf;
973
974                                 skb = sk_stream_alloc_skb(sk,
975                                                           select_size(sk, sg),
976                                                           sk->sk_allocation);
977                                 if (!skb)
978                                         goto wait_for_memory;
979
980                                 /*
981                                  * Check whether we can use HW checksum.
982                                  */
983                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
984                                         skb->ip_summed = CHECKSUM_PARTIAL;
985
986                                 skb_entail(sk, skb);
987                                 copy = size_goal;
988                                 max = size_goal;
989                         }
990
991                         /* Try to append data to the end of skb. */
992                         if (copy > seglen)
993                                 copy = seglen;
994
995                         /* Where to copy to? */
996                         if (skb_tailroom(skb) > 0) {
997                                 /* We have some space in skb head. Superb! */
998                                 if (copy > skb_tailroom(skb))
999                                         copy = skb_tailroom(skb);
1000                                 err = skb_add_data_nocache(sk, skb, from, copy);
1001                                 if (err)
1002                                         goto do_fault;
1003                         } else {
1004                                 int merge = 0;
1005                                 int i = skb_shinfo(skb)->nr_frags;
1006                                 struct page *page = sk->sk_sndmsg_page;
1007                                 int off;
1008
1009                                 if (page && page_count(page) == 1)
1010                                         sk->sk_sndmsg_off = 0;
1011
1012                                 off = sk->sk_sndmsg_off;
1013
1014                                 if (skb_can_coalesce(skb, i, page, off) &&
1015                                     off != PAGE_SIZE) {
1016                                         /* We can extend the last page
1017                                          * fragment. */
1018                                         merge = 1;
1019                                 } else if (i == MAX_SKB_FRAGS || !sg) {
1020                                         /* Need to add new fragment and cannot
1021                                          * do this because interface is non-SG,
1022                                          * or because all the page slots are
1023                                          * busy. */
1024                                         tcp_mark_push(tp, skb);
1025                                         goto new_segment;
1026                                 } else if (page) {
1027                                         if (off == PAGE_SIZE) {
1028                                                 put_page(page);
1029                                                 sk->sk_sndmsg_page = page = NULL;
1030                                                 off = 0;
1031                                         }
1032                                 } else
1033                                         off = 0;
1034
1035                                 if (copy > PAGE_SIZE - off)
1036                                         copy = PAGE_SIZE - off;
1037
1038                                 if (!sk_wmem_schedule(sk, copy))
1039                                         goto wait_for_memory;
1040
1041                                 if (!page) {
1042                                         /* Allocate new cache page. */
1043                                         if (!(page = sk_stream_alloc_page(sk)))
1044                                                 goto wait_for_memory;
1045                                 }
1046
1047                                 /* Time to copy data. We are close to
1048                                  * the end! */
1049                                 err = skb_copy_to_page_nocache(sk, from, skb,
1050                                                                page, off, copy);
1051                                 if (err) {
1052                                         /* If this page was new, give it to the
1053                                          * socket so it does not get leaked.
1054                                          */
1055                                         if (!sk->sk_sndmsg_page) {
1056                                                 sk->sk_sndmsg_page = page;
1057                                                 sk->sk_sndmsg_off = 0;
1058                                         }
1059                                         goto do_error;
1060                                 }
1061
1062                                 /* Update the skb. */
1063                                 if (merge) {
1064                                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1065                                 } else {
1066                                         skb_fill_page_desc(skb, i, page, off, copy);
1067                                         if (sk->sk_sndmsg_page) {
1068                                                 get_page(page);
1069                                         } else if (off + copy < PAGE_SIZE) {
1070                                                 get_page(page);
1071                                                 sk->sk_sndmsg_page = page;
1072                                         }
1073                                 }
1074
1075                                 sk->sk_sndmsg_off = off + copy;
1076                         }
1077
1078                         if (!copied)
1079                                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1080
1081                         tp->write_seq += copy;
1082                         TCP_SKB_CB(skb)->end_seq += copy;
1083                         skb_shinfo(skb)->gso_segs = 0;
1084
1085                         from += copy;
1086                         copied += copy;
1087                         if ((seglen -= copy) == 0 && iovlen == 0)
1088                                 goto out;
1089
1090                         if (skb->len < max || (flags & MSG_OOB))
1091                                 continue;
1092
1093                         if (forced_push(tp)) {
1094                                 tcp_mark_push(tp, skb);
1095                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1096                         } else if (skb == tcp_send_head(sk))
1097                                 tcp_push_one(sk, mss_now);
1098                         continue;
1099
1100 wait_for_sndbuf:
1101                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1102 wait_for_memory:
1103                         if (copied)
1104                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1105
1106                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1107                                 goto do_error;
1108
1109                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1110                 }
1111         }
1112
1113 out:
1114         if (copied)
1115                 tcp_push(sk, flags, mss_now, tp->nonagle);
1116         release_sock(sk);
1117         return copied;
1118
1119 do_fault:
1120         if (!skb->len) {
1121                 tcp_unlink_write_queue(skb, sk);
1122                 /* It is the one place in all of TCP, except connection
1123                  * reset, where we can be unlinking the send_head.
1124                  */
1125                 tcp_check_send_head(sk, skb);
1126                 sk_wmem_free_skb(sk, skb);
1127         }
1128
1129 do_error:
1130         if (copied)
1131                 goto out;
1132 out_err:
1133         err = sk_stream_error(sk, flags, err);
1134         release_sock(sk);
1135         return err;
1136 }
1137 EXPORT_SYMBOL(tcp_sendmsg);
1138
1139 /*
1140  *      Handle reading urgent data. BSD has very simple semantics for
1141  *      this, no blocking and very strange errors 8)
1142  */
1143
1144 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1145 {
1146         struct tcp_sock *tp = tcp_sk(sk);
1147
1148         /* No URG data to read. */
1149         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1150             tp->urg_data == TCP_URG_READ)
1151                 return -EINVAL; /* Yes this is right ! */
1152
1153         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1154                 return -ENOTCONN;
1155
1156         if (tp->urg_data & TCP_URG_VALID) {
1157                 int err = 0;
1158                 char c = tp->urg_data;
1159
1160                 if (!(flags & MSG_PEEK))
1161                         tp->urg_data = TCP_URG_READ;
1162
1163                 /* Read urgent data. */
1164                 msg->msg_flags |= MSG_OOB;
1165
1166                 if (len > 0) {
1167                         if (!(flags & MSG_TRUNC))
1168                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1169                         len = 1;
1170                 } else
1171                         msg->msg_flags |= MSG_TRUNC;
1172
1173                 return err ? -EFAULT : len;
1174         }
1175
1176         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1177                 return 0;
1178
1179         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1180          * the available implementations agree in this case:
1181          * this call should never block, independent of the
1182          * blocking state of the socket.
1183          * Mike <pall@rz.uni-karlsruhe.de>
1184          */
1185         return -EAGAIN;
1186 }
1187
1188 /* Clean up the receive buffer for full frames taken by the user,
1189  * then send an ACK if necessary.  COPIED is the number of bytes
1190  * tcp_recvmsg has given to the user so far, it speeds up the
1191  * calculation of whether or not we must ACK for the sake of
1192  * a window update.
1193  */
1194 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1195 {
1196         struct tcp_sock *tp = tcp_sk(sk);
1197         int time_to_ack = 0;
1198
1199         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1200
1201         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1202              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1203              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1204
1205         if (inet_csk_ack_scheduled(sk)) {
1206                 const struct inet_connection_sock *icsk = inet_csk(sk);
1207                    /* Delayed ACKs frequently hit locked sockets during bulk
1208                     * receive. */
1209                 if (icsk->icsk_ack.blocked ||
1210                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1211                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1212                     /*
1213                      * If this read emptied read buffer, we send ACK, if
1214                      * connection is not bidirectional, user drained
1215                      * receive buffer and there was a small segment
1216                      * in queue.
1217                      */
1218                     (copied > 0 &&
1219                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1220                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1221                        !icsk->icsk_ack.pingpong)) &&
1222                       !atomic_read(&sk->sk_rmem_alloc)))
1223                         time_to_ack = 1;
1224         }
1225
1226         /* We send an ACK if we can now advertise a non-zero window
1227          * which has been raised "significantly".
1228          *
1229          * Even if window raised up to infinity, do not send window open ACK
1230          * in states, where we will not receive more. It is useless.
1231          */
1232         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1233                 __u32 rcv_window_now = tcp_receive_window(tp);
1234
1235                 /* Optimize, __tcp_select_window() is not cheap. */
1236                 if (2*rcv_window_now <= tp->window_clamp) {
1237                         __u32 new_window = __tcp_select_window(sk);
1238
1239                         /* Send ACK now, if this read freed lots of space
1240                          * in our buffer. Certainly, new_window is new window.
1241                          * We can advertise it now, if it is not less than current one.
1242                          * "Lots" means "at least twice" here.
1243                          */
1244                         if (new_window && new_window >= 2 * rcv_window_now)
1245                                 time_to_ack = 1;
1246                 }
1247         }
1248         if (time_to_ack)
1249                 tcp_send_ack(sk);
1250 }
1251
1252 static void tcp_prequeue_process(struct sock *sk)
1253 {
1254         struct sk_buff *skb;
1255         struct tcp_sock *tp = tcp_sk(sk);
1256
1257         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1258
1259         /* RX process wants to run with disabled BHs, though it is not
1260          * necessary */
1261         local_bh_disable();
1262         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1263                 sk_backlog_rcv(sk, skb);
1264         local_bh_enable();
1265
1266         /* Clear memory counter. */
1267         tp->ucopy.memory = 0;
1268 }
1269
1270 #ifdef CONFIG_NET_DMA
1271 static void tcp_service_net_dma(struct sock *sk, bool wait)
1272 {
1273         dma_cookie_t done, used;
1274         dma_cookie_t last_issued;
1275         struct tcp_sock *tp = tcp_sk(sk);
1276
1277         if (!tp->ucopy.dma_chan)
1278                 return;
1279
1280         last_issued = tp->ucopy.dma_cookie;
1281         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1282
1283         do {
1284                 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1285                                               last_issued, &done,
1286                                               &used) == DMA_SUCCESS) {
1287                         /* Safe to free early-copied skbs now */
1288                         __skb_queue_purge(&sk->sk_async_wait_queue);
1289                         break;
1290                 } else {
1291                         struct sk_buff *skb;
1292                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1293                                (dma_async_is_complete(skb->dma_cookie, done,
1294                                                       used) == DMA_SUCCESS)) {
1295                                 __skb_dequeue(&sk->sk_async_wait_queue);
1296                                 kfree_skb(skb);
1297                         }
1298                 }
1299         } while (wait);
1300 }
1301 #endif
1302
1303 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1304 {
1305         struct sk_buff *skb;
1306         u32 offset;
1307
1308         skb_queue_walk(&sk->sk_receive_queue, skb) {
1309                 offset = seq - TCP_SKB_CB(skb)->seq;
1310                 if (tcp_hdr(skb)->syn)
1311                         offset--;
1312                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1313                         *off = offset;
1314                         return skb;
1315                 }
1316         }
1317         return NULL;
1318 }
1319
1320 /*
1321  * This routine provides an alternative to tcp_recvmsg() for routines
1322  * that would like to handle copying from skbuffs directly in 'sendfile'
1323  * fashion.
1324  * Note:
1325  *      - It is assumed that the socket was locked by the caller.
1326  *      - The routine does not block.
1327  *      - At present, there is no support for reading OOB data
1328  *        or for 'peeking' the socket using this routine
1329  *        (although both would be easy to implement).
1330  */
1331 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1332                   sk_read_actor_t recv_actor)
1333 {
1334         struct sk_buff *skb;
1335         struct tcp_sock *tp = tcp_sk(sk);
1336         u32 seq = tp->copied_seq;
1337         u32 offset;
1338         int copied = 0;
1339
1340         if (sk->sk_state == TCP_LISTEN)
1341                 return -ENOTCONN;
1342         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1343                 if (offset < skb->len) {
1344                         int used;
1345                         size_t len;
1346
1347                         len = skb->len - offset;
1348                         /* Stop reading if we hit a patch of urgent data */
1349                         if (tp->urg_data) {
1350                                 u32 urg_offset = tp->urg_seq - seq;
1351                                 if (urg_offset < len)
1352                                         len = urg_offset;
1353                                 if (!len)
1354                                         break;
1355                         }
1356                         used = recv_actor(desc, skb, offset, len);
1357                         if (used < 0) {
1358                                 if (!copied)
1359                                         copied = used;
1360                                 break;
1361                         } else if (used <= len) {
1362                                 seq += used;
1363                                 copied += used;
1364                                 offset += used;
1365                         }
1366                         /*
1367                          * If recv_actor drops the lock (e.g. TCP splice
1368                          * receive) the skb pointer might be invalid when
1369                          * getting here: tcp_collapse might have deleted it
1370                          * while aggregating skbs from the socket queue.
1371                          */
1372                         skb = tcp_recv_skb(sk, seq-1, &offset);
1373                         if (!skb || (offset+1 != skb->len))
1374                                 break;
1375                 }
1376                 if (tcp_hdr(skb)->fin) {
1377                         sk_eat_skb(sk, skb, 0);
1378                         ++seq;
1379                         break;
1380                 }
1381                 sk_eat_skb(sk, skb, 0);
1382                 if (!desc->count)
1383                         break;
1384                 tp->copied_seq = seq;
1385         }
1386         tp->copied_seq = seq;
1387
1388         tcp_rcv_space_adjust(sk);
1389
1390         /* Clean up data we have read: This will do ACK frames. */
1391         if (copied > 0)
1392                 tcp_cleanup_rbuf(sk, copied);
1393         return copied;
1394 }
1395 EXPORT_SYMBOL(tcp_read_sock);
1396
1397 /*
1398  *      This routine copies from a sock struct into the user buffer.
1399  *
1400  *      Technical note: in 2.3 we work on _locked_ socket, so that
1401  *      tricks with *seq access order and skb->users are not required.
1402  *      Probably, code can be easily improved even more.
1403  */
1404
1405 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1406                 size_t len, int nonblock, int flags, int *addr_len)
1407 {
1408         struct tcp_sock *tp = tcp_sk(sk);
1409         int copied = 0;
1410         u32 peek_seq;
1411         u32 *seq;
1412         unsigned long used;
1413         int err;
1414         int target;             /* Read at least this many bytes */
1415         long timeo;
1416         struct task_struct *user_recv = NULL;
1417         int copied_early = 0;
1418         struct sk_buff *skb;
1419         u32 urg_hole = 0;
1420
1421         lock_sock(sk);
1422
1423         err = -ENOTCONN;
1424         if (sk->sk_state == TCP_LISTEN)
1425                 goto out;
1426
1427         timeo = sock_rcvtimeo(sk, nonblock);
1428
1429         /* Urgent data needs to be handled specially. */
1430         if (flags & MSG_OOB)
1431                 goto recv_urg;
1432
1433         seq = &tp->copied_seq;
1434         if (flags & MSG_PEEK) {
1435                 peek_seq = tp->copied_seq;
1436                 seq = &peek_seq;
1437         }
1438
1439         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1440
1441 #ifdef CONFIG_NET_DMA
1442         tp->ucopy.dma_chan = NULL;
1443         preempt_disable();
1444         skb = skb_peek_tail(&sk->sk_receive_queue);
1445         {
1446                 int available = 0;
1447
1448                 if (skb)
1449                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1450                 if ((available < target) &&
1451                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1452                     !sysctl_tcp_low_latency &&
1453                     dma_find_channel(DMA_MEMCPY)) {
1454                         preempt_enable_no_resched();
1455                         tp->ucopy.pinned_list =
1456                                         dma_pin_iovec_pages(msg->msg_iov, len);
1457                 } else {
1458                         preempt_enable_no_resched();
1459                 }
1460         }
1461 #endif
1462
1463         do {
1464                 u32 offset;
1465
1466                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1467                 if (tp->urg_data && tp->urg_seq == *seq) {
1468                         if (copied)
1469                                 break;
1470                         if (signal_pending(current)) {
1471                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1472                                 break;
1473                         }
1474                 }
1475
1476                 /* Next get a buffer. */
1477
1478                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1479                         /* Now that we have two receive queues this
1480                          * shouldn't happen.
1481                          */
1482                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1483                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1484                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1485                                  flags))
1486                                 break;
1487
1488                         offset = *seq - TCP_SKB_CB(skb)->seq;
1489                         if (tcp_hdr(skb)->syn)
1490                                 offset--;
1491                         if (offset < skb->len)
1492                                 goto found_ok_skb;
1493                         if (tcp_hdr(skb)->fin)
1494                                 goto found_fin_ok;
1495                         WARN(!(flags & MSG_PEEK),
1496                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1497                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1498                 }
1499
1500                 /* Well, if we have backlog, try to process it now yet. */
1501
1502                 if (copied >= target && !sk->sk_backlog.tail)
1503                         break;
1504
1505                 if (copied) {
1506                         if (sk->sk_err ||
1507                             sk->sk_state == TCP_CLOSE ||
1508                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1509                             !timeo ||
1510                             signal_pending(current))
1511                                 break;
1512                 } else {
1513                         if (sock_flag(sk, SOCK_DONE))
1514                                 break;
1515
1516                         if (sk->sk_err) {
1517                                 copied = sock_error(sk);
1518                                 break;
1519                         }
1520
1521                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1522                                 break;
1523
1524                         if (sk->sk_state == TCP_CLOSE) {
1525                                 if (!sock_flag(sk, SOCK_DONE)) {
1526                                         /* This occurs when user tries to read
1527                                          * from never connected socket.
1528                                          */
1529                                         copied = -ENOTCONN;
1530                                         break;
1531                                 }
1532                                 break;
1533                         }
1534
1535                         if (!timeo) {
1536                                 copied = -EAGAIN;
1537                                 break;
1538                         }
1539
1540                         if (signal_pending(current)) {
1541                                 copied = sock_intr_errno(timeo);
1542                                 break;
1543                         }
1544                 }
1545
1546                 tcp_cleanup_rbuf(sk, copied);
1547
1548                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1549                         /* Install new reader */
1550                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1551                                 user_recv = current;
1552                                 tp->ucopy.task = user_recv;
1553                                 tp->ucopy.iov = msg->msg_iov;
1554                         }
1555
1556                         tp->ucopy.len = len;
1557
1558                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1559                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1560
1561                         /* Ugly... If prequeue is not empty, we have to
1562                          * process it before releasing socket, otherwise
1563                          * order will be broken at second iteration.
1564                          * More elegant solution is required!!!
1565                          *
1566                          * Look: we have the following (pseudo)queues:
1567                          *
1568                          * 1. packets in flight
1569                          * 2. backlog
1570                          * 3. prequeue
1571                          * 4. receive_queue
1572                          *
1573                          * Each queue can be processed only if the next ones
1574                          * are empty. At this point we have empty receive_queue.
1575                          * But prequeue _can_ be not empty after 2nd iteration,
1576                          * when we jumped to start of loop because backlog
1577                          * processing added something to receive_queue.
1578                          * We cannot release_sock(), because backlog contains
1579                          * packets arrived _after_ prequeued ones.
1580                          *
1581                          * Shortly, algorithm is clear --- to process all
1582                          * the queues in order. We could make it more directly,
1583                          * requeueing packets from backlog to prequeue, if
1584                          * is not empty. It is more elegant, but eats cycles,
1585                          * unfortunately.
1586                          */
1587                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1588                                 goto do_prequeue;
1589
1590                         /* __ Set realtime policy in scheduler __ */
1591                 }
1592
1593 #ifdef CONFIG_NET_DMA
1594                 if (tp->ucopy.dma_chan)
1595                         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1596 #endif
1597                 if (copied >= target) {
1598                         /* Do not sleep, just process backlog. */
1599                         release_sock(sk);
1600                         lock_sock(sk);
1601                 } else
1602                         sk_wait_data(sk, &timeo);
1603
1604 #ifdef CONFIG_NET_DMA
1605                 tcp_service_net_dma(sk, false);  /* Don't block */
1606                 tp->ucopy.wakeup = 0;
1607 #endif
1608
1609                 if (user_recv) {
1610                         int chunk;
1611
1612                         /* __ Restore normal policy in scheduler __ */
1613
1614                         if ((chunk = len - tp->ucopy.len) != 0) {
1615                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1616                                 len -= chunk;
1617                                 copied += chunk;
1618                         }
1619
1620                         if (tp->rcv_nxt == tp->copied_seq &&
1621                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1622 do_prequeue:
1623                                 tcp_prequeue_process(sk);
1624
1625                                 if ((chunk = len - tp->ucopy.len) != 0) {
1626                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1627                                         len -= chunk;
1628                                         copied += chunk;
1629                                 }
1630                         }
1631                 }
1632                 if ((flags & MSG_PEEK) &&
1633                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1634                         if (net_ratelimit())
1635                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1636                                        current->comm, task_pid_nr(current));
1637                         peek_seq = tp->copied_seq;
1638                 }
1639                 continue;
1640
1641         found_ok_skb:
1642                 /* Ok so how much can we use? */
1643                 used = skb->len - offset;
1644                 if (len < used)
1645                         used = len;
1646
1647                 /* Do we have urgent data here? */
1648                 if (tp->urg_data) {
1649                         u32 urg_offset = tp->urg_seq - *seq;
1650                         if (urg_offset < used) {
1651                                 if (!urg_offset) {
1652                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1653                                                 ++*seq;
1654                                                 urg_hole++;
1655                                                 offset++;
1656                                                 used--;
1657                                                 if (!used)
1658                                                         goto skip_copy;
1659                                         }
1660                                 } else
1661                                         used = urg_offset;
1662                         }
1663                 }
1664
1665                 if (!(flags & MSG_TRUNC)) {
1666 #ifdef CONFIG_NET_DMA
1667                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1668                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1669
1670                         if (tp->ucopy.dma_chan) {
1671                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1672                                         tp->ucopy.dma_chan, skb, offset,
1673                                         msg->msg_iov, used,
1674                                         tp->ucopy.pinned_list);
1675
1676                                 if (tp->ucopy.dma_cookie < 0) {
1677
1678                                         printk(KERN_ALERT "dma_cookie < 0\n");
1679
1680                                         /* Exception. Bailout! */
1681                                         if (!copied)
1682                                                 copied = -EFAULT;
1683                                         break;
1684                                 }
1685
1686                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1687
1688                                 if ((offset + used) == skb->len)
1689                                         copied_early = 1;
1690
1691                         } else
1692 #endif
1693                         {
1694                                 err = skb_copy_datagram_iovec(skb, offset,
1695                                                 msg->msg_iov, used);
1696                                 if (err) {
1697                                         /* Exception. Bailout! */
1698                                         if (!copied)
1699                                                 copied = -EFAULT;
1700                                         break;
1701                                 }
1702                         }
1703                 }
1704
1705                 *seq += used;
1706                 copied += used;
1707                 len -= used;
1708
1709                 tcp_rcv_space_adjust(sk);
1710
1711 skip_copy:
1712                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1713                         tp->urg_data = 0;
1714                         tcp_fast_path_check(sk);
1715                 }
1716                 if (used + offset < skb->len)
1717                         continue;
1718
1719                 if (tcp_hdr(skb)->fin)
1720                         goto found_fin_ok;
1721                 if (!(flags & MSG_PEEK)) {
1722                         sk_eat_skb(sk, skb, copied_early);
1723                         copied_early = 0;
1724                 }
1725                 continue;
1726
1727         found_fin_ok:
1728                 /* Process the FIN. */
1729                 ++*seq;
1730                 if (!(flags & MSG_PEEK)) {
1731                         sk_eat_skb(sk, skb, copied_early);
1732                         copied_early = 0;
1733                 }
1734                 break;
1735         } while (len > 0);
1736
1737         if (user_recv) {
1738                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1739                         int chunk;
1740
1741                         tp->ucopy.len = copied > 0 ? len : 0;
1742
1743                         tcp_prequeue_process(sk);
1744
1745                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1746                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1747                                 len -= chunk;
1748                                 copied += chunk;
1749                         }
1750                 }
1751
1752                 tp->ucopy.task = NULL;
1753                 tp->ucopy.len = 0;
1754         }
1755
1756 #ifdef CONFIG_NET_DMA
1757         tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1758         tp->ucopy.dma_chan = NULL;
1759
1760         if (tp->ucopy.pinned_list) {
1761                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1762                 tp->ucopy.pinned_list = NULL;
1763         }
1764 #endif
1765
1766         /* According to UNIX98, msg_name/msg_namelen are ignored
1767          * on connected socket. I was just happy when found this 8) --ANK
1768          */
1769
1770         /* Clean up data we have read: This will do ACK frames. */
1771         tcp_cleanup_rbuf(sk, copied);
1772
1773         release_sock(sk);
1774         return copied;
1775
1776 out:
1777         release_sock(sk);
1778         return err;
1779
1780 recv_urg:
1781         err = tcp_recv_urg(sk, msg, len, flags);
1782         goto out;
1783 }
1784 EXPORT_SYMBOL(tcp_recvmsg);
1785
1786 void tcp_set_state(struct sock *sk, int state)
1787 {
1788         int oldstate = sk->sk_state;
1789
1790         switch (state) {
1791         case TCP_ESTABLISHED:
1792                 if (oldstate != TCP_ESTABLISHED)
1793                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1794                 break;
1795
1796         case TCP_CLOSE:
1797                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1798                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1799
1800                 sk->sk_prot->unhash(sk);
1801                 if (inet_csk(sk)->icsk_bind_hash &&
1802                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1803                         inet_put_port(sk);
1804                 /* fall through */
1805         default:
1806                 if (oldstate == TCP_ESTABLISHED)
1807                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1808         }
1809
1810         /* Change state AFTER socket is unhashed to avoid closed
1811          * socket sitting in hash tables.
1812          */
1813         sk->sk_state = state;
1814
1815 #ifdef STATE_TRACE
1816         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1817 #endif
1818 }
1819 EXPORT_SYMBOL_GPL(tcp_set_state);
1820
1821 /*
1822  *      State processing on a close. This implements the state shift for
1823  *      sending our FIN frame. Note that we only send a FIN for some
1824  *      states. A shutdown() may have already sent the FIN, or we may be
1825  *      closed.
1826  */
1827
1828 static const unsigned char new_state[16] = {
1829   /* current state:        new state:      action:      */
1830   /* (Invalid)          */ TCP_CLOSE,
1831   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1832   /* TCP_SYN_SENT       */ TCP_CLOSE,
1833   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1834   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1835   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1836   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1837   /* TCP_CLOSE          */ TCP_CLOSE,
1838   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1839   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1840   /* TCP_LISTEN         */ TCP_CLOSE,
1841   /* TCP_CLOSING        */ TCP_CLOSING,
1842 };
1843
1844 static int tcp_close_state(struct sock *sk)
1845 {
1846         int next = (int)new_state[sk->sk_state];
1847         int ns = next & TCP_STATE_MASK;
1848
1849         tcp_set_state(sk, ns);
1850
1851         return next & TCP_ACTION_FIN;
1852 }
1853
1854 /*
1855  *      Shutdown the sending side of a connection. Much like close except
1856  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1857  */
1858
1859 void tcp_shutdown(struct sock *sk, int how)
1860 {
1861         /*      We need to grab some memory, and put together a FIN,
1862          *      and then put it into the queue to be sent.
1863          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1864          */
1865         if (!(how & SEND_SHUTDOWN))
1866                 return;
1867
1868         /* If we've already sent a FIN, or it's a closed state, skip this. */
1869         if ((1 << sk->sk_state) &
1870             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1871              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1872                 /* Clear out any half completed packets.  FIN if needed. */
1873                 if (tcp_close_state(sk))
1874                         tcp_send_fin(sk);
1875         }
1876 }
1877 EXPORT_SYMBOL(tcp_shutdown);
1878
1879 bool tcp_check_oom(struct sock *sk, int shift)
1880 {
1881         bool too_many_orphans, out_of_socket_memory;
1882
1883         too_many_orphans = tcp_too_many_orphans(sk, shift);
1884         out_of_socket_memory = tcp_out_of_memory(sk);
1885
1886         if (too_many_orphans && net_ratelimit())
1887                 pr_info("TCP: too many orphaned sockets\n");
1888         if (out_of_socket_memory && net_ratelimit())
1889                 pr_info("TCP: out of memory -- consider tuning tcp_mem\n");
1890         return too_many_orphans || out_of_socket_memory;
1891 }
1892
1893 void tcp_close(struct sock *sk, long timeout)
1894 {
1895         struct sk_buff *skb;
1896         int data_was_unread = 0;
1897         int state;
1898
1899         lock_sock(sk);
1900         sk->sk_shutdown = SHUTDOWN_MASK;
1901
1902         if (sk->sk_state == TCP_LISTEN) {
1903                 tcp_set_state(sk, TCP_CLOSE);
1904
1905                 /* Special case. */
1906                 inet_csk_listen_stop(sk);
1907
1908                 goto adjudge_to_death;
1909         }
1910
1911         /*  We need to flush the recv. buffs.  We do this only on the
1912          *  descriptor close, not protocol-sourced closes, because the
1913          *  reader process may not have drained the data yet!
1914          */
1915         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1916                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1917                           tcp_hdr(skb)->fin;
1918                 data_was_unread += len;
1919                 __kfree_skb(skb);
1920         }
1921
1922         sk_mem_reclaim(sk);
1923
1924         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1925         if (sk->sk_state == TCP_CLOSE)
1926                 goto adjudge_to_death;
1927
1928         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1929          * data was lost. To witness the awful effects of the old behavior of
1930          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1931          * GET in an FTP client, suspend the process, wait for the client to
1932          * advertise a zero window, then kill -9 the FTP client, wheee...
1933          * Note: timeout is always zero in such a case.
1934          */
1935         if (data_was_unread) {
1936                 /* Unread data was tossed, zap the connection. */
1937                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1938                 tcp_set_state(sk, TCP_CLOSE);
1939                 tcp_send_active_reset(sk, sk->sk_allocation);
1940         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1941                 /* Check zero linger _after_ checking for unread data. */
1942                 sk->sk_prot->disconnect(sk, 0);
1943                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1944         } else if (tcp_close_state(sk)) {
1945                 /* We FIN if the application ate all the data before
1946                  * zapping the connection.
1947                  */
1948
1949                 /* RED-PEN. Formally speaking, we have broken TCP state
1950                  * machine. State transitions:
1951                  *
1952                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1953                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1954                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1955                  *
1956                  * are legal only when FIN has been sent (i.e. in window),
1957                  * rather than queued out of window. Purists blame.
1958                  *
1959                  * F.e. "RFC state" is ESTABLISHED,
1960                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1961                  *
1962                  * The visible declinations are that sometimes
1963                  * we enter time-wait state, when it is not required really
1964                  * (harmless), do not send active resets, when they are
1965                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1966                  * they look as CLOSING or LAST_ACK for Linux)
1967                  * Probably, I missed some more holelets.
1968                  *                                              --ANK
1969                  */
1970                 tcp_send_fin(sk);
1971         }
1972
1973         sk_stream_wait_close(sk, timeout);
1974
1975 adjudge_to_death:
1976         state = sk->sk_state;
1977         sock_hold(sk);
1978         sock_orphan(sk);
1979
1980         /* It is the last release_sock in its life. It will remove backlog. */
1981         release_sock(sk);
1982
1983
1984         /* Now socket is owned by kernel and we acquire BH lock
1985            to finish close. No need to check for user refs.
1986          */
1987         local_bh_disable();
1988         bh_lock_sock(sk);
1989         WARN_ON(sock_owned_by_user(sk));
1990
1991         percpu_counter_inc(sk->sk_prot->orphan_count);
1992
1993         /* Have we already been destroyed by a softirq or backlog? */
1994         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1995                 goto out;
1996
1997         /*      This is a (useful) BSD violating of the RFC. There is a
1998          *      problem with TCP as specified in that the other end could
1999          *      keep a socket open forever with no application left this end.
2000          *      We use a 3 minute timeout (about the same as BSD) then kill
2001          *      our end. If they send after that then tough - BUT: long enough
2002          *      that we won't make the old 4*rto = almost no time - whoops
2003          *      reset mistake.
2004          *
2005          *      Nope, it was not mistake. It is really desired behaviour
2006          *      f.e. on http servers, when such sockets are useless, but
2007          *      consume significant resources. Let's do it with special
2008          *      linger2 option.                                 --ANK
2009          */
2010
2011         if (sk->sk_state == TCP_FIN_WAIT2) {
2012                 struct tcp_sock *tp = tcp_sk(sk);
2013                 if (tp->linger2 < 0) {
2014                         tcp_set_state(sk, TCP_CLOSE);
2015                         tcp_send_active_reset(sk, GFP_ATOMIC);
2016                         NET_INC_STATS_BH(sock_net(sk),
2017                                         LINUX_MIB_TCPABORTONLINGER);
2018                 } else {
2019                         const int tmo = tcp_fin_time(sk);
2020
2021                         if (tmo > TCP_TIMEWAIT_LEN) {
2022                                 inet_csk_reset_keepalive_timer(sk,
2023                                                 tmo - TCP_TIMEWAIT_LEN);
2024                         } else {
2025                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2026                                 goto out;
2027                         }
2028                 }
2029         }
2030         if (sk->sk_state != TCP_CLOSE) {
2031                 sk_mem_reclaim(sk);
2032                 if (tcp_check_oom(sk, 0)) {
2033                         tcp_set_state(sk, TCP_CLOSE);
2034                         tcp_send_active_reset(sk, GFP_ATOMIC);
2035                         NET_INC_STATS_BH(sock_net(sk),
2036                                         LINUX_MIB_TCPABORTONMEMORY);
2037                 }
2038         }
2039
2040         if (sk->sk_state == TCP_CLOSE)
2041                 inet_csk_destroy_sock(sk);
2042         /* Otherwise, socket is reprieved until protocol close. */
2043
2044 out:
2045         bh_unlock_sock(sk);
2046         local_bh_enable();
2047         sock_put(sk);
2048 }
2049 EXPORT_SYMBOL(tcp_close);
2050
2051 /* These states need RST on ABORT according to RFC793 */
2052
2053 static inline int tcp_need_reset(int state)
2054 {
2055         return (1 << state) &
2056                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2057                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2058 }
2059
2060 int tcp_disconnect(struct sock *sk, int flags)
2061 {
2062         struct inet_sock *inet = inet_sk(sk);
2063         struct inet_connection_sock *icsk = inet_csk(sk);
2064         struct tcp_sock *tp = tcp_sk(sk);
2065         int err = 0;
2066         int old_state = sk->sk_state;
2067
2068         if (old_state != TCP_CLOSE)
2069                 tcp_set_state(sk, TCP_CLOSE);
2070
2071         /* ABORT function of RFC793 */
2072         if (old_state == TCP_LISTEN) {
2073                 inet_csk_listen_stop(sk);
2074         } else if (tcp_need_reset(old_state) ||
2075                    (tp->snd_nxt != tp->write_seq &&
2076                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2077                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2078                  * states
2079                  */
2080                 tcp_send_active_reset(sk, gfp_any());
2081                 sk->sk_err = ECONNRESET;
2082         } else if (old_state == TCP_SYN_SENT)
2083                 sk->sk_err = ECONNRESET;
2084
2085         tcp_clear_xmit_timers(sk);
2086         __skb_queue_purge(&sk->sk_receive_queue);
2087         tcp_write_queue_purge(sk);
2088         __skb_queue_purge(&tp->out_of_order_queue);
2089 #ifdef CONFIG_NET_DMA
2090         __skb_queue_purge(&sk->sk_async_wait_queue);
2091 #endif
2092
2093         inet->inet_dport = 0;
2094
2095         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2096                 inet_reset_saddr(sk);
2097
2098         sk->sk_shutdown = 0;
2099         sock_reset_flag(sk, SOCK_DONE);
2100         tp->srtt = 0;
2101         if ((tp->write_seq += tp->max_window + 2) == 0)
2102                 tp->write_seq = 1;
2103         icsk->icsk_backoff = 0;
2104         tp->snd_cwnd = 2;
2105         icsk->icsk_probes_out = 0;
2106         tp->packets_out = 0;
2107         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2108         tp->snd_cwnd_cnt = 0;
2109         tp->bytes_acked = 0;
2110         tp->window_clamp = 0;
2111         tcp_set_ca_state(sk, TCP_CA_Open);
2112         tcp_clear_retrans(tp);
2113         inet_csk_delack_init(sk);
2114         tcp_init_send_head(sk);
2115         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2116         __sk_dst_reset(sk);
2117
2118         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2119
2120         sk->sk_error_report(sk);
2121         return err;
2122 }
2123 EXPORT_SYMBOL(tcp_disconnect);
2124
2125 /*
2126  *      Socket option code for TCP.
2127  */
2128 static int do_tcp_setsockopt(struct sock *sk, int level,
2129                 int optname, char __user *optval, unsigned int optlen)
2130 {
2131         struct tcp_sock *tp = tcp_sk(sk);
2132         struct inet_connection_sock *icsk = inet_csk(sk);
2133         int val;
2134         int err = 0;
2135
2136         /* These are data/string values, all the others are ints */
2137         switch (optname) {
2138         case TCP_CONGESTION: {
2139                 char name[TCP_CA_NAME_MAX];
2140
2141                 if (optlen < 1)
2142                         return -EINVAL;
2143
2144                 val = strncpy_from_user(name, optval,
2145                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2146                 if (val < 0)
2147                         return -EFAULT;
2148                 name[val] = 0;
2149
2150                 lock_sock(sk);
2151                 err = tcp_set_congestion_control(sk, name);
2152                 release_sock(sk);
2153                 return err;
2154         }
2155         case TCP_COOKIE_TRANSACTIONS: {
2156                 struct tcp_cookie_transactions ctd;
2157                 struct tcp_cookie_values *cvp = NULL;
2158
2159                 if (sizeof(ctd) > optlen)
2160                         return -EINVAL;
2161                 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2162                         return -EFAULT;
2163
2164                 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2165                     ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2166                         return -EINVAL;
2167
2168                 if (ctd.tcpct_cookie_desired == 0) {
2169                         /* default to global value */
2170                 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2171                            ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2172                            ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2173                         return -EINVAL;
2174                 }
2175
2176                 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2177                         /* Supercedes all other values */
2178                         lock_sock(sk);
2179                         if (tp->cookie_values != NULL) {
2180                                 kref_put(&tp->cookie_values->kref,
2181                                          tcp_cookie_values_release);
2182                                 tp->cookie_values = NULL;
2183                         }
2184                         tp->rx_opt.cookie_in_always = 0; /* false */
2185                         tp->rx_opt.cookie_out_never = 1; /* true */
2186                         release_sock(sk);
2187                         return err;
2188                 }
2189
2190                 /* Allocate ancillary memory before locking.
2191                  */
2192                 if (ctd.tcpct_used > 0 ||
2193                     (tp->cookie_values == NULL &&
2194                      (sysctl_tcp_cookie_size > 0 ||
2195                       ctd.tcpct_cookie_desired > 0 ||
2196                       ctd.tcpct_s_data_desired > 0))) {
2197                         cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2198                                       GFP_KERNEL);
2199                         if (cvp == NULL)
2200                                 return -ENOMEM;
2201
2202                         kref_init(&cvp->kref);
2203                 }
2204                 lock_sock(sk);
2205                 tp->rx_opt.cookie_in_always =
2206                         (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2207                 tp->rx_opt.cookie_out_never = 0; /* false */
2208
2209                 if (tp->cookie_values != NULL) {
2210                         if (cvp != NULL) {
2211                                 /* Changed values are recorded by a changed
2212                                  * pointer, ensuring the cookie will differ,
2213                                  * without separately hashing each value later.
2214                                  */
2215                                 kref_put(&tp->cookie_values->kref,
2216                                          tcp_cookie_values_release);
2217                         } else {
2218                                 cvp = tp->cookie_values;
2219                         }
2220                 }
2221
2222                 if (cvp != NULL) {
2223                         cvp->cookie_desired = ctd.tcpct_cookie_desired;
2224
2225                         if (ctd.tcpct_used > 0) {
2226                                 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2227                                        ctd.tcpct_used);
2228                                 cvp->s_data_desired = ctd.tcpct_used;
2229                                 cvp->s_data_constant = 1; /* true */
2230                         } else {
2231                                 /* No constant payload data. */
2232                                 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2233                                 cvp->s_data_constant = 0; /* false */
2234                         }
2235
2236                         tp->cookie_values = cvp;
2237                 }
2238                 release_sock(sk);
2239                 return err;
2240         }
2241         default:
2242                 /* fallthru */
2243                 break;
2244         }
2245
2246         if (optlen < sizeof(int))
2247                 return -EINVAL;
2248
2249         if (get_user(val, (int __user *)optval))
2250                 return -EFAULT;
2251
2252         lock_sock(sk);
2253
2254         switch (optname) {
2255         case TCP_MAXSEG:
2256                 /* Values greater than interface MTU won't take effect. However
2257                  * at the point when this call is done we typically don't yet
2258                  * know which interface is going to be used */
2259                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2260                         err = -EINVAL;
2261                         break;
2262                 }
2263                 tp->rx_opt.user_mss = val;
2264                 break;
2265
2266         case TCP_NODELAY:
2267                 if (val) {
2268                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2269                          * this option on corked socket is remembered, but
2270                          * it is not activated until cork is cleared.
2271                          *
2272                          * However, when TCP_NODELAY is set we make
2273                          * an explicit push, which overrides even TCP_CORK
2274                          * for currently queued segments.
2275                          */
2276                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2277                         tcp_push_pending_frames(sk);
2278                 } else {
2279                         tp->nonagle &= ~TCP_NAGLE_OFF;
2280                 }
2281                 break;
2282
2283         case TCP_THIN_LINEAR_TIMEOUTS:
2284                 if (val < 0 || val > 1)
2285                         err = -EINVAL;
2286                 else
2287                         tp->thin_lto = val;
2288                 break;
2289
2290         case TCP_THIN_DUPACK:
2291                 if (val < 0 || val > 1)
2292                         err = -EINVAL;
2293                 else
2294                         tp->thin_dupack = val;
2295                 break;
2296
2297         case TCP_CORK:
2298                 /* When set indicates to always queue non-full frames.
2299                  * Later the user clears this option and we transmit
2300                  * any pending partial frames in the queue.  This is
2301                  * meant to be used alongside sendfile() to get properly
2302                  * filled frames when the user (for example) must write
2303                  * out headers with a write() call first and then use
2304                  * sendfile to send out the data parts.
2305                  *
2306                  * TCP_CORK can be set together with TCP_NODELAY and it is
2307                  * stronger than TCP_NODELAY.
2308                  */
2309                 if (val) {
2310                         tp->nonagle |= TCP_NAGLE_CORK;
2311                 } else {
2312                         tp->nonagle &= ~TCP_NAGLE_CORK;
2313                         if (tp->nonagle&TCP_NAGLE_OFF)
2314                                 tp->nonagle |= TCP_NAGLE_PUSH;
2315                         tcp_push_pending_frames(sk);
2316                 }
2317                 break;
2318
2319         case TCP_KEEPIDLE:
2320                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2321                         err = -EINVAL;
2322                 else {
2323                         tp->keepalive_time = val * HZ;
2324                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2325                             !((1 << sk->sk_state) &
2326                               (TCPF_CLOSE | TCPF_LISTEN))) {
2327                                 u32 elapsed = keepalive_time_elapsed(tp);
2328                                 if (tp->keepalive_time > elapsed)
2329                                         elapsed = tp->keepalive_time - elapsed;
2330                                 else
2331                                         elapsed = 0;
2332                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2333                         }
2334                 }
2335                 break;
2336         case TCP_KEEPINTVL:
2337                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2338                         err = -EINVAL;
2339                 else
2340                         tp->keepalive_intvl = val * HZ;
2341                 break;
2342         case TCP_KEEPCNT:
2343                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2344                         err = -EINVAL;
2345                 else
2346                         tp->keepalive_probes = val;
2347                 break;
2348         case TCP_SYNCNT:
2349                 if (val < 1 || val > MAX_TCP_SYNCNT)
2350                         err = -EINVAL;
2351                 else
2352                         icsk->icsk_syn_retries = val;
2353                 break;
2354
2355         case TCP_LINGER2:
2356                 if (val < 0)
2357                         tp->linger2 = -1;
2358                 else if (val > sysctl_tcp_fin_timeout / HZ)
2359                         tp->linger2 = 0;
2360                 else
2361                         tp->linger2 = val * HZ;
2362                 break;
2363
2364         case TCP_DEFER_ACCEPT:
2365                 /* Translate value in seconds to number of retransmits */
2366                 icsk->icsk_accept_queue.rskq_defer_accept =
2367                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2368                                         TCP_RTO_MAX / HZ);
2369                 break;
2370
2371         case TCP_WINDOW_CLAMP:
2372                 if (!val) {
2373                         if (sk->sk_state != TCP_CLOSE) {
2374                                 err = -EINVAL;
2375                                 break;
2376                         }
2377                         tp->window_clamp = 0;
2378                 } else
2379                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2380                                                 SOCK_MIN_RCVBUF / 2 : val;
2381                 break;
2382
2383         case TCP_QUICKACK:
2384                 if (!val) {
2385                         icsk->icsk_ack.pingpong = 1;
2386                 } else {
2387                         icsk->icsk_ack.pingpong = 0;
2388                         if ((1 << sk->sk_state) &
2389                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2390                             inet_csk_ack_scheduled(sk)) {
2391                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2392                                 tcp_cleanup_rbuf(sk, 1);
2393                                 if (!(val & 1))
2394                                         icsk->icsk_ack.pingpong = 1;
2395                         }
2396                 }
2397                 break;
2398
2399 #ifdef CONFIG_TCP_MD5SIG
2400         case TCP_MD5SIG:
2401                 /* Read the IP->Key mappings from userspace */
2402                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2403                 break;
2404 #endif
2405         case TCP_USER_TIMEOUT:
2406                 /* Cap the max timeout in ms TCP will retry/retrans
2407                  * before giving up and aborting (ETIMEDOUT) a connection.
2408                  */
2409                 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2410                 break;
2411         default:
2412                 err = -ENOPROTOOPT;
2413                 break;
2414         }
2415
2416         release_sock(sk);
2417         return err;
2418 }
2419
2420 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2421                    unsigned int optlen)
2422 {
2423         const struct inet_connection_sock *icsk = inet_csk(sk);
2424
2425         if (level != SOL_TCP)
2426                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2427                                                      optval, optlen);
2428         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2429 }
2430 EXPORT_SYMBOL(tcp_setsockopt);
2431
2432 #ifdef CONFIG_COMPAT
2433 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2434                           char __user *optval, unsigned int optlen)
2435 {
2436         if (level != SOL_TCP)
2437                 return inet_csk_compat_setsockopt(sk, level, optname,
2438                                                   optval, optlen);
2439         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2440 }
2441 EXPORT_SYMBOL(compat_tcp_setsockopt);
2442 #endif
2443
2444 /* Return information about state of tcp endpoint in API format. */
2445 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2446 {
2447         const struct tcp_sock *tp = tcp_sk(sk);
2448         const struct inet_connection_sock *icsk = inet_csk(sk);
2449         u32 now = tcp_time_stamp;
2450
2451         memset(info, 0, sizeof(*info));
2452
2453         info->tcpi_state = sk->sk_state;
2454         info->tcpi_ca_state = icsk->icsk_ca_state;
2455         info->tcpi_retransmits = icsk->icsk_retransmits;
2456         info->tcpi_probes = icsk->icsk_probes_out;
2457         info->tcpi_backoff = icsk->icsk_backoff;
2458
2459         if (tp->rx_opt.tstamp_ok)
2460                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2461         if (tcp_is_sack(tp))
2462                 info->tcpi_options |= TCPI_OPT_SACK;
2463         if (tp->rx_opt.wscale_ok) {
2464                 info->tcpi_options |= TCPI_OPT_WSCALE;
2465                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2466                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2467         }
2468
2469         if (tp->ecn_flags & TCP_ECN_OK)
2470                 info->tcpi_options |= TCPI_OPT_ECN;
2471         if (tp->ecn_flags & TCP_ECN_SEEN)
2472                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2473
2474         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2475         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2476         info->tcpi_snd_mss = tp->mss_cache;
2477         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2478
2479         if (sk->sk_state == TCP_LISTEN) {
2480                 info->tcpi_unacked = sk->sk_ack_backlog;
2481                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2482         } else {
2483                 info->tcpi_unacked = tp->packets_out;
2484                 info->tcpi_sacked = tp->sacked_out;
2485         }
2486         info->tcpi_lost = tp->lost_out;
2487         info->tcpi_retrans = tp->retrans_out;
2488         info->tcpi_fackets = tp->fackets_out;
2489
2490         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2491         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2492         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2493
2494         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2495         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2496         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2497         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2498         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2499         info->tcpi_snd_cwnd = tp->snd_cwnd;
2500         info->tcpi_advmss = tp->advmss;
2501         info->tcpi_reordering = tp->reordering;
2502
2503         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2504         info->tcpi_rcv_space = tp->rcvq_space.space;
2505
2506         info->tcpi_total_retrans = tp->total_retrans;
2507 }
2508 EXPORT_SYMBOL_GPL(tcp_get_info);
2509
2510 static int do_tcp_getsockopt(struct sock *sk, int level,
2511                 int optname, char __user *optval, int __user *optlen)
2512 {
2513         struct inet_connection_sock *icsk = inet_csk(sk);
2514         struct tcp_sock *tp = tcp_sk(sk);
2515         int val, len;
2516
2517         if (get_user(len, optlen))
2518                 return -EFAULT;
2519
2520         len = min_t(unsigned int, len, sizeof(int));
2521
2522         if (len < 0)
2523                 return -EINVAL;
2524
2525         switch (optname) {
2526         case TCP_MAXSEG:
2527                 val = tp->mss_cache;
2528                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2529                         val = tp->rx_opt.user_mss;
2530                 break;
2531         case TCP_NODELAY:
2532                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2533                 break;
2534         case TCP_CORK:
2535                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2536                 break;
2537         case TCP_KEEPIDLE:
2538                 val = keepalive_time_when(tp) / HZ;
2539                 break;
2540         case TCP_KEEPINTVL:
2541                 val = keepalive_intvl_when(tp) / HZ;
2542                 break;
2543         case TCP_KEEPCNT:
2544                 val = keepalive_probes(tp);
2545                 break;
2546         case TCP_SYNCNT:
2547                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2548                 break;
2549         case TCP_LINGER2:
2550                 val = tp->linger2;
2551                 if (val >= 0)
2552                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2553                 break;
2554         case TCP_DEFER_ACCEPT:
2555                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2556                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2557                 break;
2558         case TCP_WINDOW_CLAMP:
2559                 val = tp->window_clamp;
2560                 break;
2561         case TCP_INFO: {
2562                 struct tcp_info info;
2563
2564                 if (get_user(len, optlen))
2565                         return -EFAULT;
2566
2567                 tcp_get_info(sk, &info);
2568
2569                 len = min_t(unsigned int, len, sizeof(info));
2570                 if (put_user(len, optlen))
2571                         return -EFAULT;
2572                 if (copy_to_user(optval, &info, len))
2573                         return -EFAULT;
2574                 return 0;
2575         }
2576         case TCP_QUICKACK:
2577                 val = !icsk->icsk_ack.pingpong;
2578                 break;
2579
2580         case TCP_CONGESTION:
2581                 if (get_user(len, optlen))
2582                         return -EFAULT;
2583                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2584                 if (put_user(len, optlen))
2585                         return -EFAULT;
2586                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2587                         return -EFAULT;
2588                 return 0;
2589
2590         case TCP_COOKIE_TRANSACTIONS: {
2591                 struct tcp_cookie_transactions ctd;
2592                 struct tcp_cookie_values *cvp = tp->cookie_values;
2593
2594                 if (get_user(len, optlen))
2595                         return -EFAULT;
2596                 if (len < sizeof(ctd))
2597                         return -EINVAL;
2598
2599                 memset(&ctd, 0, sizeof(ctd));
2600                 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2601                                    TCP_COOKIE_IN_ALWAYS : 0)
2602                                 | (tp->rx_opt.cookie_out_never ?
2603                                    TCP_COOKIE_OUT_NEVER : 0);
2604
2605                 if (cvp != NULL) {
2606                         ctd.tcpct_flags |= (cvp->s_data_in ?
2607                                             TCP_S_DATA_IN : 0)
2608                                          | (cvp->s_data_out ?
2609                                             TCP_S_DATA_OUT : 0);
2610
2611                         ctd.tcpct_cookie_desired = cvp->cookie_desired;
2612                         ctd.tcpct_s_data_desired = cvp->s_data_desired;
2613
2614                         memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2615                                cvp->cookie_pair_size);
2616                         ctd.tcpct_used = cvp->cookie_pair_size;
2617                 }
2618
2619                 if (put_user(sizeof(ctd), optlen))
2620                         return -EFAULT;
2621                 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2622                         return -EFAULT;
2623                 return 0;
2624         }
2625         case TCP_THIN_LINEAR_TIMEOUTS:
2626                 val = tp->thin_lto;
2627                 break;
2628         case TCP_THIN_DUPACK:
2629                 val = tp->thin_dupack;
2630                 break;
2631
2632         case TCP_USER_TIMEOUT:
2633                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2634                 break;
2635         default:
2636                 return -ENOPROTOOPT;
2637         }
2638
2639         if (put_user(len, optlen))
2640                 return -EFAULT;
2641         if (copy_to_user(optval, &val, len))
2642                 return -EFAULT;
2643         return 0;
2644 }
2645
2646 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2647                    int __user *optlen)
2648 {
2649         struct inet_connection_sock *icsk = inet_csk(sk);
2650
2651         if (level != SOL_TCP)
2652                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2653                                                      optval, optlen);
2654         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2655 }
2656 EXPORT_SYMBOL(tcp_getsockopt);
2657
2658 #ifdef CONFIG_COMPAT
2659 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2660                           char __user *optval, int __user *optlen)
2661 {
2662         if (level != SOL_TCP)
2663                 return inet_csk_compat_getsockopt(sk, level, optname,
2664                                                   optval, optlen);
2665         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2666 }
2667 EXPORT_SYMBOL(compat_tcp_getsockopt);
2668 #endif
2669
2670 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2671         netdev_features_t features)
2672 {
2673         struct sk_buff *segs = ERR_PTR(-EINVAL);
2674         struct tcphdr *th;
2675         unsigned thlen;
2676         unsigned int seq;
2677         __be32 delta;
2678         unsigned int oldlen;
2679         unsigned int mss;
2680
2681         if (!pskb_may_pull(skb, sizeof(*th)))
2682                 goto out;
2683
2684         th = tcp_hdr(skb);
2685         thlen = th->doff * 4;
2686         if (thlen < sizeof(*th))
2687                 goto out;
2688
2689         if (!pskb_may_pull(skb, thlen))
2690                 goto out;
2691
2692         oldlen = (u16)~skb->len;
2693         __skb_pull(skb, thlen);
2694
2695         mss = skb_shinfo(skb)->gso_size;
2696         if (unlikely(skb->len <= mss))
2697                 goto out;
2698
2699         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2700                 /* Packet is from an untrusted source, reset gso_segs. */
2701                 int type = skb_shinfo(skb)->gso_type;
2702
2703                 if (unlikely(type &
2704                              ~(SKB_GSO_TCPV4 |
2705                                SKB_GSO_DODGY |
2706                                SKB_GSO_TCP_ECN |
2707                                SKB_GSO_TCPV6 |
2708                                0) ||
2709                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2710                         goto out;
2711
2712                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2713
2714                 segs = NULL;
2715                 goto out;
2716         }
2717
2718         segs = skb_segment(skb, features);
2719         if (IS_ERR(segs))
2720                 goto out;
2721
2722         delta = htonl(oldlen + (thlen + mss));
2723
2724         skb = segs;
2725         th = tcp_hdr(skb);
2726         seq = ntohl(th->seq);
2727
2728         do {
2729                 th->fin = th->psh = 0;
2730
2731                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2732                                        (__force u32)delta));
2733                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2734                         th->check =
2735                              csum_fold(csum_partial(skb_transport_header(skb),
2736                                                     thlen, skb->csum));
2737
2738                 seq += mss;
2739                 skb = skb->next;
2740                 th = tcp_hdr(skb);
2741
2742                 th->seq = htonl(seq);
2743                 th->cwr = 0;
2744         } while (skb->next);
2745
2746         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2747                       skb->data_len);
2748         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2749                                 (__force u32)delta));
2750         if (skb->ip_summed != CHECKSUM_PARTIAL)
2751                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2752                                                    thlen, skb->csum));
2753
2754 out:
2755         return segs;
2756 }
2757 EXPORT_SYMBOL(tcp_tso_segment);
2758
2759 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2760 {
2761         struct sk_buff **pp = NULL;
2762         struct sk_buff *p;
2763         struct tcphdr *th;
2764         struct tcphdr *th2;
2765         unsigned int len;
2766         unsigned int thlen;
2767         __be32 flags;
2768         unsigned int mss = 1;
2769         unsigned int hlen;
2770         unsigned int off;
2771         int flush = 1;
2772         int i;
2773
2774         off = skb_gro_offset(skb);
2775         hlen = off + sizeof(*th);
2776         th = skb_gro_header_fast(skb, off);
2777         if (skb_gro_header_hard(skb, hlen)) {
2778                 th = skb_gro_header_slow(skb, hlen, off);
2779                 if (unlikely(!th))
2780                         goto out;
2781         }
2782
2783         thlen = th->doff * 4;
2784         if (thlen < sizeof(*th))
2785                 goto out;
2786
2787         hlen = off + thlen;
2788         if (skb_gro_header_hard(skb, hlen)) {
2789                 th = skb_gro_header_slow(skb, hlen, off);
2790                 if (unlikely(!th))
2791                         goto out;
2792         }
2793
2794         skb_gro_pull(skb, thlen);
2795
2796         len = skb_gro_len(skb);
2797         flags = tcp_flag_word(th);
2798
2799         for (; (p = *head); head = &p->next) {
2800                 if (!NAPI_GRO_CB(p)->same_flow)
2801                         continue;
2802
2803                 th2 = tcp_hdr(p);
2804
2805                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2806                         NAPI_GRO_CB(p)->same_flow = 0;
2807                         continue;
2808                 }
2809
2810                 goto found;
2811         }
2812
2813         goto out_check_final;
2814
2815 found:
2816         flush = NAPI_GRO_CB(p)->flush;
2817         flush |= (__force int)(flags & TCP_FLAG_CWR);
2818         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2819                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2820         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2821         for (i = sizeof(*th); i < thlen; i += 4)
2822                 flush |= *(u32 *)((u8 *)th + i) ^
2823                          *(u32 *)((u8 *)th2 + i);
2824
2825         mss = skb_shinfo(p)->gso_size;
2826
2827         flush |= (len - 1) >= mss;
2828         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2829
2830         if (flush || skb_gro_receive(head, skb)) {
2831                 mss = 1;
2832                 goto out_check_final;
2833         }
2834
2835         p = *head;
2836         th2 = tcp_hdr(p);
2837         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2838
2839 out_check_final:
2840         flush = len < mss;
2841         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2842                                         TCP_FLAG_RST | TCP_FLAG_SYN |
2843                                         TCP_FLAG_FIN));
2844
2845         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2846                 pp = head;
2847
2848 out:
2849         NAPI_GRO_CB(skb)->flush |= flush;
2850
2851         return pp;
2852 }
2853 EXPORT_SYMBOL(tcp_gro_receive);
2854
2855 int tcp_gro_complete(struct sk_buff *skb)
2856 {
2857         struct tcphdr *th = tcp_hdr(skb);
2858
2859         skb->csum_start = skb_transport_header(skb) - skb->head;
2860         skb->csum_offset = offsetof(struct tcphdr, check);
2861         skb->ip_summed = CHECKSUM_PARTIAL;
2862
2863         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2864
2865         if (th->cwr)
2866                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2867
2868         return 0;
2869 }
2870 EXPORT_SYMBOL(tcp_gro_complete);
2871
2872 #ifdef CONFIG_TCP_MD5SIG
2873 static unsigned long tcp_md5sig_users;
2874 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
2875 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2876
2877 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
2878 {
2879         int cpu;
2880
2881         for_each_possible_cpu(cpu) {
2882                 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
2883
2884                 if (p->md5_desc.tfm)
2885                         crypto_free_hash(p->md5_desc.tfm);
2886         }
2887         free_percpu(pool);
2888 }
2889
2890 void tcp_free_md5sig_pool(void)
2891 {
2892         struct tcp_md5sig_pool __percpu *pool = NULL;
2893
2894         spin_lock_bh(&tcp_md5sig_pool_lock);
2895         if (--tcp_md5sig_users == 0) {
2896                 pool = tcp_md5sig_pool;
2897                 tcp_md5sig_pool = NULL;
2898         }
2899         spin_unlock_bh(&tcp_md5sig_pool_lock);
2900         if (pool)
2901                 __tcp_free_md5sig_pool(pool);
2902 }
2903 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2904
2905 static struct tcp_md5sig_pool __percpu *
2906 __tcp_alloc_md5sig_pool(struct sock *sk)
2907 {
2908         int cpu;
2909         struct tcp_md5sig_pool __percpu *pool;
2910
2911         pool = alloc_percpu(struct tcp_md5sig_pool);
2912         if (!pool)
2913                 return NULL;
2914
2915         for_each_possible_cpu(cpu) {
2916                 struct crypto_hash *hash;
2917
2918                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2919                 if (!hash || IS_ERR(hash))
2920                         goto out_free;
2921
2922                 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
2923         }
2924         return pool;
2925 out_free:
2926         __tcp_free_md5sig_pool(pool);
2927         return NULL;
2928 }
2929
2930 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2931 {
2932         struct tcp_md5sig_pool __percpu *pool;
2933         int alloc = 0;
2934
2935 retry:
2936         spin_lock_bh(&tcp_md5sig_pool_lock);
2937         pool = tcp_md5sig_pool;
2938         if (tcp_md5sig_users++ == 0) {
2939                 alloc = 1;
2940                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2941         } else if (!pool) {
2942                 tcp_md5sig_users--;
2943                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2944                 cpu_relax();
2945                 goto retry;
2946         } else
2947                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2948
2949         if (alloc) {
2950                 /* we cannot hold spinlock here because this may sleep. */
2951                 struct tcp_md5sig_pool __percpu *p;
2952
2953                 p = __tcp_alloc_md5sig_pool(sk);
2954                 spin_lock_bh(&tcp_md5sig_pool_lock);
2955                 if (!p) {
2956                         tcp_md5sig_users--;
2957                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2958                         return NULL;
2959                 }
2960                 pool = tcp_md5sig_pool;
2961                 if (pool) {
2962                         /* oops, it has already been assigned. */
2963                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2964                         __tcp_free_md5sig_pool(p);
2965                 } else {
2966                         tcp_md5sig_pool = pool = p;
2967                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2968                 }
2969         }
2970         return pool;
2971 }
2972 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2973
2974
2975 /**
2976  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2977  *
2978  *      We use percpu structure, so if we succeed, we exit with preemption
2979  *      and BH disabled, to make sure another thread or softirq handling
2980  *      wont try to get same context.
2981  */
2982 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2983 {
2984         struct tcp_md5sig_pool __percpu *p;
2985
2986         local_bh_disable();
2987
2988         spin_lock(&tcp_md5sig_pool_lock);
2989         p = tcp_md5sig_pool;
2990         if (p)
2991                 tcp_md5sig_users++;
2992         spin_unlock(&tcp_md5sig_pool_lock);
2993
2994         if (p)
2995                 return this_cpu_ptr(p);
2996
2997         local_bh_enable();
2998         return NULL;
2999 }
3000 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3001
3002 void tcp_put_md5sig_pool(void)
3003 {
3004         local_bh_enable();
3005         tcp_free_md5sig_pool();
3006 }
3007 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3008
3009 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3010                         const struct tcphdr *th)
3011 {
3012         struct scatterlist sg;
3013         struct tcphdr hdr;
3014         int err;
3015
3016         /* We are not allowed to change tcphdr, make a local copy */
3017         memcpy(&hdr, th, sizeof(hdr));
3018         hdr.check = 0;
3019
3020         /* options aren't included in the hash */
3021         sg_init_one(&sg, &hdr, sizeof(hdr));
3022         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3023         return err;
3024 }
3025 EXPORT_SYMBOL(tcp_md5_hash_header);
3026
3027 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3028                           const struct sk_buff *skb, unsigned int header_len)
3029 {
3030         struct scatterlist sg;
3031         const struct tcphdr *tp = tcp_hdr(skb);
3032         struct hash_desc *desc = &hp->md5_desc;
3033         unsigned i;
3034         const unsigned head_data_len = skb_headlen(skb) > header_len ?
3035                                        skb_headlen(skb) - header_len : 0;
3036         const struct skb_shared_info *shi = skb_shinfo(skb);
3037         struct sk_buff *frag_iter;
3038
3039         sg_init_table(&sg, 1);
3040
3041         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3042         if (crypto_hash_update(desc, &sg, head_data_len))
3043                 return 1;
3044
3045         for (i = 0; i < shi->nr_frags; ++i) {
3046                 const struct skb_frag_struct *f = &shi->frags[i];
3047                 struct page *page = skb_frag_page(f);
3048                 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3049                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3050                         return 1;
3051         }
3052
3053         skb_walk_frags(skb, frag_iter)
3054                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3055                         return 1;
3056
3057         return 0;
3058 }
3059 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3060
3061 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3062 {
3063         struct scatterlist sg;
3064
3065         sg_init_one(&sg, key->key, key->keylen);
3066         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3067 }
3068 EXPORT_SYMBOL(tcp_md5_hash_key);
3069
3070 #endif
3071
3072 /**
3073  * Each Responder maintains up to two secret values concurrently for
3074  * efficient secret rollover.  Each secret value has 4 states:
3075  *
3076  * Generating.  (tcp_secret_generating != tcp_secret_primary)
3077  *    Generates new Responder-Cookies, but not yet used for primary
3078  *    verification.  This is a short-term state, typically lasting only
3079  *    one round trip time (RTT).
3080  *
3081  * Primary.  (tcp_secret_generating == tcp_secret_primary)
3082  *    Used both for generation and primary verification.
3083  *
3084  * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3085  *    Used for verification, until the first failure that can be
3086  *    verified by the newer Generating secret.  At that time, this
3087  *    cookie's state is changed to Secondary, and the Generating
3088  *    cookie's state is changed to Primary.  This is a short-term state,
3089  *    typically lasting only one round trip time (RTT).
3090  *
3091  * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3092  *    Used for secondary verification, after primary verification
3093  *    failures.  This state lasts no more than twice the Maximum Segment
3094  *    Lifetime (2MSL).  Then, the secret is discarded.
3095  */
3096 struct tcp_cookie_secret {
3097         /* The secret is divided into two parts.  The digest part is the
3098          * equivalent of previously hashing a secret and saving the state,
3099          * and serves as an initialization vector (IV).  The message part
3100          * serves as the trailing secret.
3101          */
3102         u32                             secrets[COOKIE_WORKSPACE_WORDS];
3103         unsigned long                   expires;
3104 };
3105
3106 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3107 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3108 #define TCP_SECRET_LIFE (HZ * 600)
3109
3110 static struct tcp_cookie_secret tcp_secret_one;
3111 static struct tcp_cookie_secret tcp_secret_two;
3112
3113 /* Essentially a circular list, without dynamic allocation. */
3114 static struct tcp_cookie_secret *tcp_secret_generating;
3115 static struct tcp_cookie_secret *tcp_secret_primary;
3116 static struct tcp_cookie_secret *tcp_secret_retiring;
3117 static struct tcp_cookie_secret *tcp_secret_secondary;
3118
3119 static DEFINE_SPINLOCK(tcp_secret_locker);
3120
3121 /* Select a pseudo-random word in the cookie workspace.
3122  */
3123 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3124 {
3125         return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3126 }
3127
3128 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3129  * Called in softirq context.
3130  * Returns: 0 for success.
3131  */
3132 int tcp_cookie_generator(u32 *bakery)
3133 {
3134         unsigned long jiffy = jiffies;
3135
3136         if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3137                 spin_lock_bh(&tcp_secret_locker);
3138                 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3139                         /* refreshed by another */
3140                         memcpy(bakery,
3141                                &tcp_secret_generating->secrets[0],
3142                                COOKIE_WORKSPACE_WORDS);
3143                 } else {
3144                         /* still needs refreshing */
3145                         get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3146
3147                         /* The first time, paranoia assumes that the
3148                          * randomization function isn't as strong.  But,
3149                          * this secret initialization is delayed until
3150                          * the last possible moment (packet arrival).
3151                          * Although that time is observable, it is
3152                          * unpredictably variable.  Mash in the most
3153                          * volatile clock bits available, and expire the
3154                          * secret extra quickly.
3155                          */
3156                         if (unlikely(tcp_secret_primary->expires ==
3157                                      tcp_secret_secondary->expires)) {
3158                                 struct timespec tv;
3159
3160                                 getnstimeofday(&tv);
3161                                 bakery[COOKIE_DIGEST_WORDS+0] ^=
3162                                         (u32)tv.tv_nsec;
3163
3164                                 tcp_secret_secondary->expires = jiffy
3165                                         + TCP_SECRET_1MSL
3166                                         + (0x0f & tcp_cookie_work(bakery, 0));
3167                         } else {
3168                                 tcp_secret_secondary->expires = jiffy
3169                                         + TCP_SECRET_LIFE
3170                                         + (0xff & tcp_cookie_work(bakery, 1));
3171                                 tcp_secret_primary->expires = jiffy
3172                                         + TCP_SECRET_2MSL
3173                                         + (0x1f & tcp_cookie_work(bakery, 2));
3174                         }
3175                         memcpy(&tcp_secret_secondary->secrets[0],
3176                                bakery, COOKIE_WORKSPACE_WORDS);
3177
3178                         rcu_assign_pointer(tcp_secret_generating,
3179                                            tcp_secret_secondary);
3180                         rcu_assign_pointer(tcp_secret_retiring,
3181                                            tcp_secret_primary);
3182                         /*
3183                          * Neither call_rcu() nor synchronize_rcu() needed.
3184                          * Retiring data is not freed.  It is replaced after
3185                          * further (locked) pointer updates, and a quiet time
3186                          * (minimum 1MSL, maximum LIFE - 2MSL).
3187                          */
3188                 }
3189                 spin_unlock_bh(&tcp_secret_locker);
3190         } else {
3191                 rcu_read_lock_bh();
3192                 memcpy(bakery,
3193                        &rcu_dereference(tcp_secret_generating)->secrets[0],
3194                        COOKIE_WORKSPACE_WORDS);
3195                 rcu_read_unlock_bh();
3196         }
3197         return 0;
3198 }
3199 EXPORT_SYMBOL(tcp_cookie_generator);
3200
3201 void tcp_done(struct sock *sk)
3202 {
3203         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3204                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3205
3206         tcp_set_state(sk, TCP_CLOSE);
3207         tcp_clear_xmit_timers(sk);
3208
3209         sk->sk_shutdown = SHUTDOWN_MASK;
3210
3211         if (!sock_flag(sk, SOCK_DEAD))
3212                 sk->sk_state_change(sk);
3213         else
3214                 inet_csk_destroy_sock(sk);
3215 }
3216 EXPORT_SYMBOL_GPL(tcp_done);
3217
3218 extern struct tcp_congestion_ops tcp_reno;
3219
3220 static __initdata unsigned long thash_entries;
3221 static int __init set_thash_entries(char *str)
3222 {
3223         if (!str)
3224                 return 0;
3225         thash_entries = simple_strtoul(str, &str, 0);
3226         return 1;
3227 }
3228 __setup("thash_entries=", set_thash_entries);
3229
3230 void tcp_init_mem(struct net *net)
3231 {
3232         unsigned long limit = nr_free_buffer_pages() / 8;
3233         limit = max(limit, 128UL);
3234         net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3235         net->ipv4.sysctl_tcp_mem[1] = limit;
3236         net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3237 }
3238
3239 void __init tcp_init(void)
3240 {
3241         struct sk_buff *skb = NULL;
3242         unsigned long limit;
3243         int max_share, cnt;
3244         unsigned int i;
3245         unsigned long jiffy = jiffies;
3246
3247         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3248
3249         percpu_counter_init(&tcp_sockets_allocated, 0);
3250         percpu_counter_init(&tcp_orphan_count, 0);
3251         tcp_hashinfo.bind_bucket_cachep =
3252                 kmem_cache_create("tcp_bind_bucket",
3253                                   sizeof(struct inet_bind_bucket), 0,
3254                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3255
3256         /* Size and allocate the main established and bind bucket
3257          * hash tables.
3258          *
3259          * The methodology is similar to that of the buffer cache.
3260          */
3261         tcp_hashinfo.ehash =
3262                 alloc_large_system_hash("TCP established",
3263                                         sizeof(struct inet_ehash_bucket),
3264                                         thash_entries,
3265                                         (totalram_pages >= 128 * 1024) ?
3266                                         13 : 15,
3267                                         0,
3268                                         NULL,
3269                                         &tcp_hashinfo.ehash_mask,
3270                                         thash_entries ? 0 : 512 * 1024);
3271         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3272                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3273                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3274         }
3275         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3276                 panic("TCP: failed to alloc ehash_locks");
3277         tcp_hashinfo.bhash =
3278                 alloc_large_system_hash("TCP bind",
3279                                         sizeof(struct inet_bind_hashbucket),
3280                                         tcp_hashinfo.ehash_mask + 1,
3281                                         (totalram_pages >= 128 * 1024) ?
3282                                         13 : 15,
3283                                         0,
3284                                         &tcp_hashinfo.bhash_size,
3285                                         NULL,
3286                                         64 * 1024);
3287         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3288         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3289                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3290                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3291         }
3292
3293
3294         cnt = tcp_hashinfo.ehash_mask + 1;
3295
3296         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3297         sysctl_tcp_max_orphans = cnt / 2;
3298         sysctl_max_syn_backlog = max(128, cnt / 256);
3299
3300         tcp_init_mem(&init_net);
3301         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3302         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 10);
3303         limit = max(limit, 128UL);
3304         max_share = min(4UL*1024*1024, limit);
3305
3306         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3307         sysctl_tcp_wmem[1] = 16*1024;
3308         sysctl_tcp_wmem[2] = max(64*1024, max_share);
3309
3310         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3311         sysctl_tcp_rmem[1] = 87380;
3312         sysctl_tcp_rmem[2] = max(87380, max_share);
3313
3314         printk(KERN_INFO "TCP: Hash tables configured "
3315                "(established %u bind %u)\n",
3316                tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3317
3318         tcp_register_congestion_control(&tcp_reno);
3319
3320         memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3321         memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3322         tcp_secret_one.expires = jiffy; /* past due */
3323         tcp_secret_two.expires = jiffy; /* past due */
3324         tcp_secret_generating = &tcp_secret_one;
3325         tcp_secret_primary = &tcp_secret_one;
3326         tcp_secret_retiring = &tcp_secret_two;
3327         tcp_secret_secondary = &tcp_secret_two;
3328 }