tcp: Remove unnecessary window comparisons for GRO
[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
268 #include <net/icmp.h>
269 #include <net/tcp.h>
270 #include <net/xfrm.h>
271 #include <net/ip.h>
272 #include <net/netdma.h>
273 #include <net/sock.h>
274
275 #include <asm/uaccess.h>
276 #include <asm/ioctls.h>
277
278 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
279
280 struct percpu_counter tcp_orphan_count;
281 EXPORT_SYMBOL_GPL(tcp_orphan_count);
282
283 int sysctl_tcp_mem[3] __read_mostly;
284 int sysctl_tcp_wmem[3] __read_mostly;
285 int sysctl_tcp_rmem[3] __read_mostly;
286
287 EXPORT_SYMBOL(sysctl_tcp_mem);
288 EXPORT_SYMBOL(sysctl_tcp_rmem);
289 EXPORT_SYMBOL(sysctl_tcp_wmem);
290
291 atomic_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
317 EXPORT_SYMBOL(tcp_memory_pressure);
318
319 void tcp_enter_memory_pressure(struct sock *sk)
320 {
321         if (!tcp_memory_pressure) {
322                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
323                 tcp_memory_pressure = 1;
324         }
325 }
326
327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
328
329 /*
330  *      Wait for a TCP event.
331  *
332  *      Note that we don't need to lock the socket, as the upper poll layers
333  *      take care of normal races (between the test and the event) and we don't
334  *      go look at any of the socket buffers directly.
335  */
336 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
337 {
338         unsigned int mask;
339         struct sock *sk = sock->sk;
340         struct tcp_sock *tp = tcp_sk(sk);
341
342         poll_wait(file, sk->sk_sleep, wait);
343         if (sk->sk_state == TCP_LISTEN)
344                 return inet_csk_listen_poll(sk);
345
346         /* Socket is not locked. We are protected from async events
347          * by poll logic and correct handling of state changes
348          * made by other threads is impossible in any case.
349          */
350
351         mask = 0;
352         if (sk->sk_err)
353                 mask = POLLERR;
354
355         /*
356          * POLLHUP is certainly not done right. But poll() doesn't
357          * have a notion of HUP in just one direction, and for a
358          * socket the read side is more interesting.
359          *
360          * Some poll() documentation says that POLLHUP is incompatible
361          * with the POLLOUT/POLLWR flags, so somebody should check this
362          * all. But careful, it tends to be safer to return too many
363          * bits than too few, and you can easily break real applications
364          * if you don't tell them that something has hung up!
365          *
366          * Check-me.
367          *
368          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
369          * our fs/select.c). It means that after we received EOF,
370          * poll always returns immediately, making impossible poll() on write()
371          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
372          * if and only if shutdown has been made in both directions.
373          * Actually, it is interesting to look how Solaris and DUX
374          * solve this dilemma. I would prefer, if POLLHUP were maskable,
375          * then we could set it on SND_SHUTDOWN. BTW examples given
376          * in Stevens' books assume exactly this behaviour, it explains
377          * why POLLHUP is incompatible with POLLOUT.    --ANK
378          *
379          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
380          * blocking on fresh not-connected or disconnected socket. --ANK
381          */
382         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
383                 mask |= POLLHUP;
384         if (sk->sk_shutdown & RCV_SHUTDOWN)
385                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
386
387         /* Connected? */
388         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
389                 int target = sock_rcvlowat(sk, 0, INT_MAX);
390
391                 if (tp->urg_seq == tp->copied_seq &&
392                     !sock_flag(sk, SOCK_URGINLINE) &&
393                     tp->urg_data)
394                         target--;
395
396                 /* Potential race condition. If read of tp below will
397                  * escape above sk->sk_state, we can be illegally awaken
398                  * in SYN_* states. */
399                 if (tp->rcv_nxt - tp->copied_seq >= target)
400                         mask |= POLLIN | POLLRDNORM;
401
402                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
403                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
404                                 mask |= POLLOUT | POLLWRNORM;
405                         } else {  /* send SIGIO later */
406                                 set_bit(SOCK_ASYNC_NOSPACE,
407                                         &sk->sk_socket->flags);
408                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
409
410                                 /* Race breaker. If space is freed after
411                                  * wspace test but before the flags are set,
412                                  * IO signal will be lost.
413                                  */
414                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
415                                         mask |= POLLOUT | POLLWRNORM;
416                         }
417                 }
418
419                 if (tp->urg_data & TCP_URG_VALID)
420                         mask |= POLLPRI;
421         }
422         return mask;
423 }
424
425 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
426 {
427         struct tcp_sock *tp = tcp_sk(sk);
428         int answ;
429
430         switch (cmd) {
431         case SIOCINQ:
432                 if (sk->sk_state == TCP_LISTEN)
433                         return -EINVAL;
434
435                 lock_sock(sk);
436                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
437                         answ = 0;
438                 else if (sock_flag(sk, SOCK_URGINLINE) ||
439                          !tp->urg_data ||
440                          before(tp->urg_seq, tp->copied_seq) ||
441                          !before(tp->urg_seq, tp->rcv_nxt)) {
442                         answ = tp->rcv_nxt - tp->copied_seq;
443
444                         /* Subtract 1, if FIN is in queue. */
445                         if (answ && !skb_queue_empty(&sk->sk_receive_queue))
446                                 answ -=
447                        tcp_hdr((struct sk_buff *)sk->sk_receive_queue.prev)->fin;
448                 } else
449                         answ = tp->urg_seq - tp->copied_seq;
450                 release_sock(sk);
451                 break;
452         case SIOCATMARK:
453                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
454                 break;
455         case SIOCOUTQ:
456                 if (sk->sk_state == TCP_LISTEN)
457                         return -EINVAL;
458
459                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
460                         answ = 0;
461                 else
462                         answ = tp->write_seq - tp->snd_una;
463                 break;
464         default:
465                 return -ENOIOCTLCMD;
466         }
467
468         return put_user(answ, (int __user *)arg);
469 }
470
471 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
472 {
473         TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
474         tp->pushed_seq = tp->write_seq;
475 }
476
477 static inline int forced_push(struct tcp_sock *tp)
478 {
479         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
480 }
481
482 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
483 {
484         struct tcp_sock *tp = tcp_sk(sk);
485         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
486
487         skb->csum    = 0;
488         tcb->seq     = tcb->end_seq = tp->write_seq;
489         tcb->flags   = TCPCB_FLAG_ACK;
490         tcb->sacked  = 0;
491         skb_header_release(skb);
492         tcp_add_write_queue_tail(sk, skb);
493         sk->sk_wmem_queued += skb->truesize;
494         sk_mem_charge(sk, skb->truesize);
495         if (tp->nonagle & TCP_NAGLE_PUSH)
496                 tp->nonagle &= ~TCP_NAGLE_PUSH;
497 }
498
499 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
500                                 struct sk_buff *skb)
501 {
502         if (flags & MSG_OOB)
503                 tp->snd_up = tp->write_seq;
504 }
505
506 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
507                             int nonagle)
508 {
509         struct tcp_sock *tp = tcp_sk(sk);
510
511         if (tcp_send_head(sk)) {
512                 struct sk_buff *skb = tcp_write_queue_tail(sk);
513                 if (!(flags & MSG_MORE) || forced_push(tp))
514                         tcp_mark_push(tp, skb);
515                 tcp_mark_urg(tp, flags, skb);
516                 __tcp_push_pending_frames(sk, mss_now,
517                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
518         }
519 }
520
521 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
522                                 unsigned int offset, size_t len)
523 {
524         struct tcp_splice_state *tss = rd_desc->arg.data;
525         int ret;
526
527         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
528                               tss->flags);
529         if (ret > 0)
530                 rd_desc->count -= ret;
531         return ret;
532 }
533
534 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
535 {
536         /* Store TCP splice context information in read_descriptor_t. */
537         read_descriptor_t rd_desc = {
538                 .arg.data = tss,
539                 .count    = tss->len,
540         };
541
542         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
543 }
544
545 /**
546  *  tcp_splice_read - splice data from TCP socket to a pipe
547  * @sock:       socket to splice from
548  * @ppos:       position (not valid)
549  * @pipe:       pipe to splice to
550  * @len:        number of bytes to splice
551  * @flags:      splice modifier flags
552  *
553  * Description:
554  *    Will read pages from given socket and fill them into a pipe.
555  *
556  **/
557 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
558                         struct pipe_inode_info *pipe, size_t len,
559                         unsigned int flags)
560 {
561         struct sock *sk = sock->sk;
562         struct tcp_splice_state tss = {
563                 .pipe = pipe,
564                 .len = len,
565                 .flags = flags,
566         };
567         long timeo;
568         ssize_t spliced;
569         int ret;
570
571         /*
572          * We can't seek on a socket input
573          */
574         if (unlikely(*ppos))
575                 return -ESPIPE;
576
577         ret = spliced = 0;
578
579         lock_sock(sk);
580
581         timeo = sock_rcvtimeo(sk, flags & SPLICE_F_NONBLOCK);
582         while (tss.len) {
583                 ret = __tcp_splice_read(sk, &tss);
584                 if (ret < 0)
585                         break;
586                 else if (!ret) {
587                         if (spliced)
588                                 break;
589                         if (sock_flag(sk, SOCK_DONE))
590                                 break;
591                         if (sk->sk_err) {
592                                 ret = sock_error(sk);
593                                 break;
594                         }
595                         if (sk->sk_shutdown & RCV_SHUTDOWN)
596                                 break;
597                         if (sk->sk_state == TCP_CLOSE) {
598                                 /*
599                                  * This occurs when user tries to read
600                                  * from never connected socket.
601                                  */
602                                 if (!sock_flag(sk, SOCK_DONE))
603                                         ret = -ENOTCONN;
604                                 break;
605                         }
606                         if (!timeo) {
607                                 ret = -EAGAIN;
608                                 break;
609                         }
610                         sk_wait_data(sk, &timeo);
611                         if (signal_pending(current)) {
612                                 ret = sock_intr_errno(timeo);
613                                 break;
614                         }
615                         continue;
616                 }
617                 tss.len -= ret;
618                 spliced += ret;
619
620                 if (!timeo)
621                         break;
622                 release_sock(sk);
623                 lock_sock(sk);
624
625                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
626                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
627                     signal_pending(current))
628                         break;
629         }
630
631         release_sock(sk);
632
633         if (spliced)
634                 return spliced;
635
636         return ret;
637 }
638
639 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
640 {
641         struct sk_buff *skb;
642
643         /* The TCP header must be at least 32-bit aligned.  */
644         size = ALIGN(size, 4);
645
646         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
647         if (skb) {
648                 if (sk_wmem_schedule(sk, skb->truesize)) {
649                         /*
650                          * Make sure that we have exactly size bytes
651                          * available to the caller, no more, no less.
652                          */
653                         skb_reserve(skb, skb_tailroom(skb) - size);
654                         return skb;
655                 }
656                 __kfree_skb(skb);
657         } else {
658                 sk->sk_prot->enter_memory_pressure(sk);
659                 sk_stream_moderate_sndbuf(sk);
660         }
661         return NULL;
662 }
663
664 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
665                                        int large_allowed)
666 {
667         struct tcp_sock *tp = tcp_sk(sk);
668         u32 xmit_size_goal, old_size_goal;
669
670         xmit_size_goal = mss_now;
671
672         if (large_allowed && sk_can_gso(sk)) {
673                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
674                                   inet_csk(sk)->icsk_af_ops->net_header_len -
675                                   inet_csk(sk)->icsk_ext_hdr_len -
676                                   tp->tcp_header_len);
677
678                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
679
680                 /* We try hard to avoid divides here */
681                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
682
683                 if (likely(old_size_goal <= xmit_size_goal &&
684                            old_size_goal + mss_now > xmit_size_goal)) {
685                         xmit_size_goal = old_size_goal;
686                 } else {
687                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
688                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
689                 }
690         }
691
692         return max(xmit_size_goal, mss_now);
693 }
694
695 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
696 {
697         int mss_now;
698
699         mss_now = tcp_current_mss(sk);
700         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
701
702         return mss_now;
703 }
704
705 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
706                          size_t psize, int flags)
707 {
708         struct tcp_sock *tp = tcp_sk(sk);
709         int mss_now, size_goal;
710         int err;
711         ssize_t copied;
712         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
713
714         /* Wait for a connection to finish. */
715         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
716                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
717                         goto out_err;
718
719         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
720
721         mss_now = tcp_send_mss(sk, &size_goal, flags);
722         copied = 0;
723
724         err = -EPIPE;
725         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
726                 goto out_err;
727
728         while (psize > 0) {
729                 struct sk_buff *skb = tcp_write_queue_tail(sk);
730                 struct page *page = pages[poffset / PAGE_SIZE];
731                 int copy, i, can_coalesce;
732                 int offset = poffset % PAGE_SIZE;
733                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
734
735                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
736 new_segment:
737                         if (!sk_stream_memory_free(sk))
738                                 goto wait_for_sndbuf;
739
740                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
741                         if (!skb)
742                                 goto wait_for_memory;
743
744                         skb_entail(sk, skb);
745                         copy = size_goal;
746                 }
747
748                 if (copy > size)
749                         copy = size;
750
751                 i = skb_shinfo(skb)->nr_frags;
752                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
753                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
754                         tcp_mark_push(tp, skb);
755                         goto new_segment;
756                 }
757                 if (!sk_wmem_schedule(sk, copy))
758                         goto wait_for_memory;
759
760                 if (can_coalesce) {
761                         skb_shinfo(skb)->frags[i - 1].size += copy;
762                 } else {
763                         get_page(page);
764                         skb_fill_page_desc(skb, i, page, offset, copy);
765                 }
766
767                 skb->len += copy;
768                 skb->data_len += copy;
769                 skb->truesize += copy;
770                 sk->sk_wmem_queued += copy;
771                 sk_mem_charge(sk, copy);
772                 skb->ip_summed = CHECKSUM_PARTIAL;
773                 tp->write_seq += copy;
774                 TCP_SKB_CB(skb)->end_seq += copy;
775                 skb_shinfo(skb)->gso_segs = 0;
776
777                 if (!copied)
778                         TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
779
780                 copied += copy;
781                 poffset += copy;
782                 if (!(psize -= copy))
783                         goto out;
784
785                 if (skb->len < size_goal || (flags & MSG_OOB))
786                         continue;
787
788                 if (forced_push(tp)) {
789                         tcp_mark_push(tp, skb);
790                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
791                 } else if (skb == tcp_send_head(sk))
792                         tcp_push_one(sk, mss_now);
793                 continue;
794
795 wait_for_sndbuf:
796                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
797 wait_for_memory:
798                 if (copied)
799                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
800
801                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
802                         goto do_error;
803
804                 mss_now = tcp_send_mss(sk, &size_goal, flags);
805         }
806
807 out:
808         if (copied)
809                 tcp_push(sk, flags, mss_now, tp->nonagle);
810         return copied;
811
812 do_error:
813         if (copied)
814                 goto out;
815 out_err:
816         return sk_stream_error(sk, flags, err);
817 }
818
819 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
820                      size_t size, int flags)
821 {
822         ssize_t res;
823         struct sock *sk = sock->sk;
824
825         if (!(sk->sk_route_caps & NETIF_F_SG) ||
826             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
827                 return sock_no_sendpage(sock, page, offset, size, flags);
828
829         lock_sock(sk);
830         TCP_CHECK_TIMER(sk);
831         res = do_tcp_sendpages(sk, &page, offset, size, flags);
832         TCP_CHECK_TIMER(sk);
833         release_sock(sk);
834         return res;
835 }
836
837 #define TCP_PAGE(sk)    (sk->sk_sndmsg_page)
838 #define TCP_OFF(sk)     (sk->sk_sndmsg_off)
839
840 static inline int select_size(struct sock *sk)
841 {
842         struct tcp_sock *tp = tcp_sk(sk);
843         int tmp = tp->mss_cache;
844
845         if (sk->sk_route_caps & NETIF_F_SG) {
846                 if (sk_can_gso(sk))
847                         tmp = 0;
848                 else {
849                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
850
851                         if (tmp >= pgbreak &&
852                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
853                                 tmp = pgbreak;
854                 }
855         }
856
857         return tmp;
858 }
859
860 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
861                 size_t size)
862 {
863         struct sock *sk = sock->sk;
864         struct iovec *iov;
865         struct tcp_sock *tp = tcp_sk(sk);
866         struct sk_buff *skb;
867         int iovlen, flags;
868         int mss_now, size_goal;
869         int err, copied;
870         long timeo;
871
872         lock_sock(sk);
873         TCP_CHECK_TIMER(sk);
874
875         flags = msg->msg_flags;
876         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
877
878         /* Wait for a connection to finish. */
879         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
880                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
881                         goto out_err;
882
883         /* This should be in poll */
884         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
885
886         mss_now = tcp_send_mss(sk, &size_goal, flags);
887
888         /* Ok commence sending. */
889         iovlen = msg->msg_iovlen;
890         iov = msg->msg_iov;
891         copied = 0;
892
893         err = -EPIPE;
894         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
895                 goto out_err;
896
897         while (--iovlen >= 0) {
898                 int seglen = iov->iov_len;
899                 unsigned char __user *from = iov->iov_base;
900
901                 iov++;
902
903                 while (seglen > 0) {
904                         int copy;
905
906                         skb = tcp_write_queue_tail(sk);
907
908                         if (!tcp_send_head(sk) ||
909                             (copy = size_goal - skb->len) <= 0) {
910
911 new_segment:
912                                 /* Allocate new segment. If the interface is SG,
913                                  * allocate skb fitting to single page.
914                                  */
915                                 if (!sk_stream_memory_free(sk))
916                                         goto wait_for_sndbuf;
917
918                                 skb = sk_stream_alloc_skb(sk, select_size(sk),
919                                                 sk->sk_allocation);
920                                 if (!skb)
921                                         goto wait_for_memory;
922
923                                 /*
924                                  * Check whether we can use HW checksum.
925                                  */
926                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
927                                         skb->ip_summed = CHECKSUM_PARTIAL;
928
929                                 skb_entail(sk, skb);
930                                 copy = size_goal;
931                         }
932
933                         /* Try to append data to the end of skb. */
934                         if (copy > seglen)
935                                 copy = seglen;
936
937                         /* Where to copy to? */
938                         if (skb_tailroom(skb) > 0) {
939                                 /* We have some space in skb head. Superb! */
940                                 if (copy > skb_tailroom(skb))
941                                         copy = skb_tailroom(skb);
942                                 if ((err = skb_add_data(skb, from, copy)) != 0)
943                                         goto do_fault;
944                         } else {
945                                 int merge = 0;
946                                 int i = skb_shinfo(skb)->nr_frags;
947                                 struct page *page = TCP_PAGE(sk);
948                                 int off = TCP_OFF(sk);
949
950                                 if (skb_can_coalesce(skb, i, page, off) &&
951                                     off != PAGE_SIZE) {
952                                         /* We can extend the last page
953                                          * fragment. */
954                                         merge = 1;
955                                 } else if (i == MAX_SKB_FRAGS ||
956                                            (!i &&
957                                            !(sk->sk_route_caps & NETIF_F_SG))) {
958                                         /* Need to add new fragment and cannot
959                                          * do this because interface is non-SG,
960                                          * or because all the page slots are
961                                          * busy. */
962                                         tcp_mark_push(tp, skb);
963                                         goto new_segment;
964                                 } else if (page) {
965                                         if (off == PAGE_SIZE) {
966                                                 put_page(page);
967                                                 TCP_PAGE(sk) = page = NULL;
968                                                 off = 0;
969                                         }
970                                 } else
971                                         off = 0;
972
973                                 if (copy > PAGE_SIZE - off)
974                                         copy = PAGE_SIZE - off;
975
976                                 if (!sk_wmem_schedule(sk, copy))
977                                         goto wait_for_memory;
978
979                                 if (!page) {
980                                         /* Allocate new cache page. */
981                                         if (!(page = sk_stream_alloc_page(sk)))
982                                                 goto wait_for_memory;
983                                 }
984
985                                 /* Time to copy data. We are close to
986                                  * the end! */
987                                 err = skb_copy_to_page(sk, from, skb, page,
988                                                        off, copy);
989                                 if (err) {
990                                         /* If this page was new, give it to the
991                                          * socket so it does not get leaked.
992                                          */
993                                         if (!TCP_PAGE(sk)) {
994                                                 TCP_PAGE(sk) = page;
995                                                 TCP_OFF(sk) = 0;
996                                         }
997                                         goto do_error;
998                                 }
999
1000                                 /* Update the skb. */
1001                                 if (merge) {
1002                                         skb_shinfo(skb)->frags[i - 1].size +=
1003                                                                         copy;
1004                                 } else {
1005                                         skb_fill_page_desc(skb, i, page, off, copy);
1006                                         if (TCP_PAGE(sk)) {
1007                                                 get_page(page);
1008                                         } else if (off + copy < PAGE_SIZE) {
1009                                                 get_page(page);
1010                                                 TCP_PAGE(sk) = page;
1011                                         }
1012                                 }
1013
1014                                 TCP_OFF(sk) = off + copy;
1015                         }
1016
1017                         if (!copied)
1018                                 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
1019
1020                         tp->write_seq += copy;
1021                         TCP_SKB_CB(skb)->end_seq += copy;
1022                         skb_shinfo(skb)->gso_segs = 0;
1023
1024                         from += copy;
1025                         copied += copy;
1026                         if ((seglen -= copy) == 0 && iovlen == 0)
1027                                 goto out;
1028
1029                         if (skb->len < size_goal || (flags & MSG_OOB))
1030                                 continue;
1031
1032                         if (forced_push(tp)) {
1033                                 tcp_mark_push(tp, skb);
1034                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1035                         } else if (skb == tcp_send_head(sk))
1036                                 tcp_push_one(sk, mss_now);
1037                         continue;
1038
1039 wait_for_sndbuf:
1040                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1041 wait_for_memory:
1042                         if (copied)
1043                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1044
1045                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1046                                 goto do_error;
1047
1048                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1049                 }
1050         }
1051
1052 out:
1053         if (copied)
1054                 tcp_push(sk, flags, mss_now, tp->nonagle);
1055         TCP_CHECK_TIMER(sk);
1056         release_sock(sk);
1057         return copied;
1058
1059 do_fault:
1060         if (!skb->len) {
1061                 tcp_unlink_write_queue(skb, sk);
1062                 /* It is the one place in all of TCP, except connection
1063                  * reset, where we can be unlinking the send_head.
1064                  */
1065                 tcp_check_send_head(sk, skb);
1066                 sk_wmem_free_skb(sk, skb);
1067         }
1068
1069 do_error:
1070         if (copied)
1071                 goto out;
1072 out_err:
1073         err = sk_stream_error(sk, flags, err);
1074         TCP_CHECK_TIMER(sk);
1075         release_sock(sk);
1076         return err;
1077 }
1078
1079 /*
1080  *      Handle reading urgent data. BSD has very simple semantics for
1081  *      this, no blocking and very strange errors 8)
1082  */
1083
1084 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1085 {
1086         struct tcp_sock *tp = tcp_sk(sk);
1087
1088         /* No URG data to read. */
1089         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1090             tp->urg_data == TCP_URG_READ)
1091                 return -EINVAL; /* Yes this is right ! */
1092
1093         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1094                 return -ENOTCONN;
1095
1096         if (tp->urg_data & TCP_URG_VALID) {
1097                 int err = 0;
1098                 char c = tp->urg_data;
1099
1100                 if (!(flags & MSG_PEEK))
1101                         tp->urg_data = TCP_URG_READ;
1102
1103                 /* Read urgent data. */
1104                 msg->msg_flags |= MSG_OOB;
1105
1106                 if (len > 0) {
1107                         if (!(flags & MSG_TRUNC))
1108                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1109                         len = 1;
1110                 } else
1111                         msg->msg_flags |= MSG_TRUNC;
1112
1113                 return err ? -EFAULT : len;
1114         }
1115
1116         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1117                 return 0;
1118
1119         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1120          * the available implementations agree in this case:
1121          * this call should never block, independent of the
1122          * blocking state of the socket.
1123          * Mike <pall@rz.uni-karlsruhe.de>
1124          */
1125         return -EAGAIN;
1126 }
1127
1128 /* Clean up the receive buffer for full frames taken by the user,
1129  * then send an ACK if necessary.  COPIED is the number of bytes
1130  * tcp_recvmsg has given to the user so far, it speeds up the
1131  * calculation of whether or not we must ACK for the sake of
1132  * a window update.
1133  */
1134 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1135 {
1136         struct tcp_sock *tp = tcp_sk(sk);
1137         int time_to_ack = 0;
1138
1139 #if TCP_DEBUG
1140         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1141
1142         WARN_ON(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
1143 #endif
1144
1145         if (inet_csk_ack_scheduled(sk)) {
1146                 const struct inet_connection_sock *icsk = inet_csk(sk);
1147                    /* Delayed ACKs frequently hit locked sockets during bulk
1148                     * receive. */
1149                 if (icsk->icsk_ack.blocked ||
1150                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1151                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1152                     /*
1153                      * If this read emptied read buffer, we send ACK, if
1154                      * connection is not bidirectional, user drained
1155                      * receive buffer and there was a small segment
1156                      * in queue.
1157                      */
1158                     (copied > 0 &&
1159                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1160                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1161                        !icsk->icsk_ack.pingpong)) &&
1162                       !atomic_read(&sk->sk_rmem_alloc)))
1163                         time_to_ack = 1;
1164         }
1165
1166         /* We send an ACK if we can now advertise a non-zero window
1167          * which has been raised "significantly".
1168          *
1169          * Even if window raised up to infinity, do not send window open ACK
1170          * in states, where we will not receive more. It is useless.
1171          */
1172         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1173                 __u32 rcv_window_now = tcp_receive_window(tp);
1174
1175                 /* Optimize, __tcp_select_window() is not cheap. */
1176                 if (2*rcv_window_now <= tp->window_clamp) {
1177                         __u32 new_window = __tcp_select_window(sk);
1178
1179                         /* Send ACK now, if this read freed lots of space
1180                          * in our buffer. Certainly, new_window is new window.
1181                          * We can advertise it now, if it is not less than current one.
1182                          * "Lots" means "at least twice" here.
1183                          */
1184                         if (new_window && new_window >= 2 * rcv_window_now)
1185                                 time_to_ack = 1;
1186                 }
1187         }
1188         if (time_to_ack)
1189                 tcp_send_ack(sk);
1190 }
1191
1192 static void tcp_prequeue_process(struct sock *sk)
1193 {
1194         struct sk_buff *skb;
1195         struct tcp_sock *tp = tcp_sk(sk);
1196
1197         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1198
1199         /* RX process wants to run with disabled BHs, though it is not
1200          * necessary */
1201         local_bh_disable();
1202         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1203                 sk_backlog_rcv(sk, skb);
1204         local_bh_enable();
1205
1206         /* Clear memory counter. */
1207         tp->ucopy.memory = 0;
1208 }
1209
1210 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1211 {
1212         struct sk_buff *skb;
1213         u32 offset;
1214
1215         skb_queue_walk(&sk->sk_receive_queue, skb) {
1216                 offset = seq - TCP_SKB_CB(skb)->seq;
1217                 if (tcp_hdr(skb)->syn)
1218                         offset--;
1219                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1220                         *off = offset;
1221                         return skb;
1222                 }
1223         }
1224         return NULL;
1225 }
1226
1227 /*
1228  * This routine provides an alternative to tcp_recvmsg() for routines
1229  * that would like to handle copying from skbuffs directly in 'sendfile'
1230  * fashion.
1231  * Note:
1232  *      - It is assumed that the socket was locked by the caller.
1233  *      - The routine does not block.
1234  *      - At present, there is no support for reading OOB data
1235  *        or for 'peeking' the socket using this routine
1236  *        (although both would be easy to implement).
1237  */
1238 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1239                   sk_read_actor_t recv_actor)
1240 {
1241         struct sk_buff *skb;
1242         struct tcp_sock *tp = tcp_sk(sk);
1243         u32 seq = tp->copied_seq;
1244         u32 offset;
1245         int copied = 0;
1246
1247         if (sk->sk_state == TCP_LISTEN)
1248                 return -ENOTCONN;
1249         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1250                 if (offset < skb->len) {
1251                         int used;
1252                         size_t len;
1253
1254                         len = skb->len - offset;
1255                         /* Stop reading if we hit a patch of urgent data */
1256                         if (tp->urg_data) {
1257                                 u32 urg_offset = tp->urg_seq - seq;
1258                                 if (urg_offset < len)
1259                                         len = urg_offset;
1260                                 if (!len)
1261                                         break;
1262                         }
1263                         used = recv_actor(desc, skb, offset, len);
1264                         if (used < 0) {
1265                                 if (!copied)
1266                                         copied = used;
1267                                 break;
1268                         } else if (used <= len) {
1269                                 seq += used;
1270                                 copied += used;
1271                                 offset += used;
1272                         }
1273                         /*
1274                          * If recv_actor drops the lock (e.g. TCP splice
1275                          * receive) the skb pointer might be invalid when
1276                          * getting here: tcp_collapse might have deleted it
1277                          * while aggregating skbs from the socket queue.
1278                          */
1279                         skb = tcp_recv_skb(sk, seq-1, &offset);
1280                         if (!skb || (offset+1 != skb->len))
1281                                 break;
1282                 }
1283                 if (tcp_hdr(skb)->fin) {
1284                         sk_eat_skb(sk, skb, 0);
1285                         ++seq;
1286                         break;
1287                 }
1288                 sk_eat_skb(sk, skb, 0);
1289                 if (!desc->count)
1290                         break;
1291         }
1292         tp->copied_seq = seq;
1293
1294         tcp_rcv_space_adjust(sk);
1295
1296         /* Clean up data we have read: This will do ACK frames. */
1297         if (copied > 0)
1298                 tcp_cleanup_rbuf(sk, copied);
1299         return copied;
1300 }
1301
1302 /*
1303  *      This routine copies from a sock struct into the user buffer.
1304  *
1305  *      Technical note: in 2.3 we work on _locked_ socket, so that
1306  *      tricks with *seq access order and skb->users are not required.
1307  *      Probably, code can be easily improved even more.
1308  */
1309
1310 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1311                 size_t len, int nonblock, int flags, int *addr_len)
1312 {
1313         struct tcp_sock *tp = tcp_sk(sk);
1314         int copied = 0;
1315         u32 peek_seq;
1316         u32 *seq;
1317         unsigned long used;
1318         int err;
1319         int target;             /* Read at least this many bytes */
1320         long timeo;
1321         struct task_struct *user_recv = NULL;
1322         int copied_early = 0;
1323         struct sk_buff *skb;
1324         u32 urg_hole = 0;
1325
1326         lock_sock(sk);
1327
1328         TCP_CHECK_TIMER(sk);
1329
1330         err = -ENOTCONN;
1331         if (sk->sk_state == TCP_LISTEN)
1332                 goto out;
1333
1334         timeo = sock_rcvtimeo(sk, nonblock);
1335
1336         /* Urgent data needs to be handled specially. */
1337         if (flags & MSG_OOB)
1338                 goto recv_urg;
1339
1340         seq = &tp->copied_seq;
1341         if (flags & MSG_PEEK) {
1342                 peek_seq = tp->copied_seq;
1343                 seq = &peek_seq;
1344         }
1345
1346         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1347
1348 #ifdef CONFIG_NET_DMA
1349         tp->ucopy.dma_chan = NULL;
1350         preempt_disable();
1351         skb = skb_peek_tail(&sk->sk_receive_queue);
1352         {
1353                 int available = 0;
1354
1355                 if (skb)
1356                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1357                 if ((available < target) &&
1358                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1359                     !sysctl_tcp_low_latency &&
1360                     dma_find_channel(DMA_MEMCPY)) {
1361                         preempt_enable_no_resched();
1362                         tp->ucopy.pinned_list =
1363                                         dma_pin_iovec_pages(msg->msg_iov, len);
1364                 } else {
1365                         preempt_enable_no_resched();
1366                 }
1367         }
1368 #endif
1369
1370         do {
1371                 u32 offset;
1372
1373                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1374                 if (tp->urg_data && tp->urg_seq == *seq) {
1375                         if (copied)
1376                                 break;
1377                         if (signal_pending(current)) {
1378                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1379                                 break;
1380                         }
1381                 }
1382
1383                 /* Next get a buffer. */
1384
1385                 skb = skb_peek(&sk->sk_receive_queue);
1386                 do {
1387                         if (!skb)
1388                                 break;
1389
1390                         /* Now that we have two receive queues this
1391                          * shouldn't happen.
1392                          */
1393                         if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1394                                 printk(KERN_INFO "recvmsg bug: copied %X "
1395                                        "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1396                                 break;
1397                         }
1398                         offset = *seq - TCP_SKB_CB(skb)->seq;
1399                         if (tcp_hdr(skb)->syn)
1400                                 offset--;
1401                         if (offset < skb->len)
1402                                 goto found_ok_skb;
1403                         if (tcp_hdr(skb)->fin)
1404                                 goto found_fin_ok;
1405                         WARN_ON(!(flags & MSG_PEEK));
1406                         skb = skb->next;
1407                 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1408
1409                 /* Well, if we have backlog, try to process it now yet. */
1410
1411                 if (copied >= target && !sk->sk_backlog.tail)
1412                         break;
1413
1414                 if (copied) {
1415                         if (sk->sk_err ||
1416                             sk->sk_state == TCP_CLOSE ||
1417                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1418                             !timeo ||
1419                             signal_pending(current))
1420                                 break;
1421                 } else {
1422                         if (sock_flag(sk, SOCK_DONE))
1423                                 break;
1424
1425                         if (sk->sk_err) {
1426                                 copied = sock_error(sk);
1427                                 break;
1428                         }
1429
1430                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1431                                 break;
1432
1433                         if (sk->sk_state == TCP_CLOSE) {
1434                                 if (!sock_flag(sk, SOCK_DONE)) {
1435                                         /* This occurs when user tries to read
1436                                          * from never connected socket.
1437                                          */
1438                                         copied = -ENOTCONN;
1439                                         break;
1440                                 }
1441                                 break;
1442                         }
1443
1444                         if (!timeo) {
1445                                 copied = -EAGAIN;
1446                                 break;
1447                         }
1448
1449                         if (signal_pending(current)) {
1450                                 copied = sock_intr_errno(timeo);
1451                                 break;
1452                         }
1453                 }
1454
1455                 tcp_cleanup_rbuf(sk, copied);
1456
1457                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1458                         /* Install new reader */
1459                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1460                                 user_recv = current;
1461                                 tp->ucopy.task = user_recv;
1462                                 tp->ucopy.iov = msg->msg_iov;
1463                         }
1464
1465                         tp->ucopy.len = len;
1466
1467                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1468                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1469
1470                         /* Ugly... If prequeue is not empty, we have to
1471                          * process it before releasing socket, otherwise
1472                          * order will be broken at second iteration.
1473                          * More elegant solution is required!!!
1474                          *
1475                          * Look: we have the following (pseudo)queues:
1476                          *
1477                          * 1. packets in flight
1478                          * 2. backlog
1479                          * 3. prequeue
1480                          * 4. receive_queue
1481                          *
1482                          * Each queue can be processed only if the next ones
1483                          * are empty. At this point we have empty receive_queue.
1484                          * But prequeue _can_ be not empty after 2nd iteration,
1485                          * when we jumped to start of loop because backlog
1486                          * processing added something to receive_queue.
1487                          * We cannot release_sock(), because backlog contains
1488                          * packets arrived _after_ prequeued ones.
1489                          *
1490                          * Shortly, algorithm is clear --- to process all
1491                          * the queues in order. We could make it more directly,
1492                          * requeueing packets from backlog to prequeue, if
1493                          * is not empty. It is more elegant, but eats cycles,
1494                          * unfortunately.
1495                          */
1496                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1497                                 goto do_prequeue;
1498
1499                         /* __ Set realtime policy in scheduler __ */
1500                 }
1501
1502                 if (copied >= target) {
1503                         /* Do not sleep, just process backlog. */
1504                         release_sock(sk);
1505                         lock_sock(sk);
1506                 } else
1507                         sk_wait_data(sk, &timeo);
1508
1509 #ifdef CONFIG_NET_DMA
1510                 tp->ucopy.wakeup = 0;
1511 #endif
1512
1513                 if (user_recv) {
1514                         int chunk;
1515
1516                         /* __ Restore normal policy in scheduler __ */
1517
1518                         if ((chunk = len - tp->ucopy.len) != 0) {
1519                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1520                                 len -= chunk;
1521                                 copied += chunk;
1522                         }
1523
1524                         if (tp->rcv_nxt == tp->copied_seq &&
1525                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1526 do_prequeue:
1527                                 tcp_prequeue_process(sk);
1528
1529                                 if ((chunk = len - tp->ucopy.len) != 0) {
1530                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1531                                         len -= chunk;
1532                                         copied += chunk;
1533                                 }
1534                         }
1535                 }
1536                 if ((flags & MSG_PEEK) &&
1537                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1538                         if (net_ratelimit())
1539                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1540                                        current->comm, task_pid_nr(current));
1541                         peek_seq = tp->copied_seq;
1542                 }
1543                 continue;
1544
1545         found_ok_skb:
1546                 /* Ok so how much can we use? */
1547                 used = skb->len - offset;
1548                 if (len < used)
1549                         used = len;
1550
1551                 /* Do we have urgent data here? */
1552                 if (tp->urg_data) {
1553                         u32 urg_offset = tp->urg_seq - *seq;
1554                         if (urg_offset < used) {
1555                                 if (!urg_offset) {
1556                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1557                                                 ++*seq;
1558                                                 urg_hole++;
1559                                                 offset++;
1560                                                 used--;
1561                                                 if (!used)
1562                                                         goto skip_copy;
1563                                         }
1564                                 } else
1565                                         used = urg_offset;
1566                         }
1567                 }
1568
1569                 if (!(flags & MSG_TRUNC)) {
1570 #ifdef CONFIG_NET_DMA
1571                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1572                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1573
1574                         if (tp->ucopy.dma_chan) {
1575                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1576                                         tp->ucopy.dma_chan, skb, offset,
1577                                         msg->msg_iov, used,
1578                                         tp->ucopy.pinned_list);
1579
1580                                 if (tp->ucopy.dma_cookie < 0) {
1581
1582                                         printk(KERN_ALERT "dma_cookie < 0\n");
1583
1584                                         /* Exception. Bailout! */
1585                                         if (!copied)
1586                                                 copied = -EFAULT;
1587                                         break;
1588                                 }
1589                                 if ((offset + used) == skb->len)
1590                                         copied_early = 1;
1591
1592                         } else
1593 #endif
1594                         {
1595                                 err = skb_copy_datagram_iovec(skb, offset,
1596                                                 msg->msg_iov, used);
1597                                 if (err) {
1598                                         /* Exception. Bailout! */
1599                                         if (!copied)
1600                                                 copied = -EFAULT;
1601                                         break;
1602                                 }
1603                         }
1604                 }
1605
1606                 *seq += used;
1607                 copied += used;
1608                 len -= used;
1609
1610                 tcp_rcv_space_adjust(sk);
1611
1612 skip_copy:
1613                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1614                         tp->urg_data = 0;
1615                         tcp_fast_path_check(sk);
1616                 }
1617                 if (used + offset < skb->len)
1618                         continue;
1619
1620                 if (tcp_hdr(skb)->fin)
1621                         goto found_fin_ok;
1622                 if (!(flags & MSG_PEEK)) {
1623                         sk_eat_skb(sk, skb, copied_early);
1624                         copied_early = 0;
1625                 }
1626                 continue;
1627
1628         found_fin_ok:
1629                 /* Process the FIN. */
1630                 ++*seq;
1631                 if (!(flags & MSG_PEEK)) {
1632                         sk_eat_skb(sk, skb, copied_early);
1633                         copied_early = 0;
1634                 }
1635                 break;
1636         } while (len > 0);
1637
1638         if (user_recv) {
1639                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1640                         int chunk;
1641
1642                         tp->ucopy.len = copied > 0 ? len : 0;
1643
1644                         tcp_prequeue_process(sk);
1645
1646                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1647                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1648                                 len -= chunk;
1649                                 copied += chunk;
1650                         }
1651                 }
1652
1653                 tp->ucopy.task = NULL;
1654                 tp->ucopy.len = 0;
1655         }
1656
1657 #ifdef CONFIG_NET_DMA
1658         if (tp->ucopy.dma_chan) {
1659                 dma_cookie_t done, used;
1660
1661                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1662
1663                 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1664                                                  tp->ucopy.dma_cookie, &done,
1665                                                  &used) == DMA_IN_PROGRESS) {
1666                         /* do partial cleanup of sk_async_wait_queue */
1667                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1668                                (dma_async_is_complete(skb->dma_cookie, done,
1669                                                       used) == DMA_SUCCESS)) {
1670                                 __skb_dequeue(&sk->sk_async_wait_queue);
1671                                 kfree_skb(skb);
1672                         }
1673                 }
1674
1675                 /* Safe to free early-copied skbs now */
1676                 __skb_queue_purge(&sk->sk_async_wait_queue);
1677                 tp->ucopy.dma_chan = NULL;
1678         }
1679         if (tp->ucopy.pinned_list) {
1680                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1681                 tp->ucopy.pinned_list = NULL;
1682         }
1683 #endif
1684
1685         /* According to UNIX98, msg_name/msg_namelen are ignored
1686          * on connected socket. I was just happy when found this 8) --ANK
1687          */
1688
1689         /* Clean up data we have read: This will do ACK frames. */
1690         tcp_cleanup_rbuf(sk, copied);
1691
1692         TCP_CHECK_TIMER(sk);
1693         release_sock(sk);
1694         return copied;
1695
1696 out:
1697         TCP_CHECK_TIMER(sk);
1698         release_sock(sk);
1699         return err;
1700
1701 recv_urg:
1702         err = tcp_recv_urg(sk, msg, len, flags);
1703         goto out;
1704 }
1705
1706 void tcp_set_state(struct sock *sk, int state)
1707 {
1708         int oldstate = sk->sk_state;
1709
1710         switch (state) {
1711         case TCP_ESTABLISHED:
1712                 if (oldstate != TCP_ESTABLISHED)
1713                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1714                 break;
1715
1716         case TCP_CLOSE:
1717                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1718                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1719
1720                 sk->sk_prot->unhash(sk);
1721                 if (inet_csk(sk)->icsk_bind_hash &&
1722                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1723                         inet_put_port(sk);
1724                 /* fall through */
1725         default:
1726                 if (oldstate == TCP_ESTABLISHED)
1727                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1728         }
1729
1730         /* Change state AFTER socket is unhashed to avoid closed
1731          * socket sitting in hash tables.
1732          */
1733         sk->sk_state = state;
1734
1735 #ifdef STATE_TRACE
1736         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1737 #endif
1738 }
1739 EXPORT_SYMBOL_GPL(tcp_set_state);
1740
1741 /*
1742  *      State processing on a close. This implements the state shift for
1743  *      sending our FIN frame. Note that we only send a FIN for some
1744  *      states. A shutdown() may have already sent the FIN, or we may be
1745  *      closed.
1746  */
1747
1748 static const unsigned char new_state[16] = {
1749   /* current state:        new state:      action:      */
1750   /* (Invalid)          */ TCP_CLOSE,
1751   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1752   /* TCP_SYN_SENT       */ TCP_CLOSE,
1753   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1754   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1755   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1756   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1757   /* TCP_CLOSE          */ TCP_CLOSE,
1758   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1759   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1760   /* TCP_LISTEN         */ TCP_CLOSE,
1761   /* TCP_CLOSING        */ TCP_CLOSING,
1762 };
1763
1764 static int tcp_close_state(struct sock *sk)
1765 {
1766         int next = (int)new_state[sk->sk_state];
1767         int ns = next & TCP_STATE_MASK;
1768
1769         tcp_set_state(sk, ns);
1770
1771         return next & TCP_ACTION_FIN;
1772 }
1773
1774 /*
1775  *      Shutdown the sending side of a connection. Much like close except
1776  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1777  */
1778
1779 void tcp_shutdown(struct sock *sk, int how)
1780 {
1781         /*      We need to grab some memory, and put together a FIN,
1782          *      and then put it into the queue to be sent.
1783          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1784          */
1785         if (!(how & SEND_SHUTDOWN))
1786                 return;
1787
1788         /* If we've already sent a FIN, or it's a closed state, skip this. */
1789         if ((1 << sk->sk_state) &
1790             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1791              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1792                 /* Clear out any half completed packets.  FIN if needed. */
1793                 if (tcp_close_state(sk))
1794                         tcp_send_fin(sk);
1795         }
1796 }
1797
1798 void tcp_close(struct sock *sk, long timeout)
1799 {
1800         struct sk_buff *skb;
1801         int data_was_unread = 0;
1802         int state;
1803
1804         lock_sock(sk);
1805         sk->sk_shutdown = SHUTDOWN_MASK;
1806
1807         if (sk->sk_state == TCP_LISTEN) {
1808                 tcp_set_state(sk, TCP_CLOSE);
1809
1810                 /* Special case. */
1811                 inet_csk_listen_stop(sk);
1812
1813                 goto adjudge_to_death;
1814         }
1815
1816         /*  We need to flush the recv. buffs.  We do this only on the
1817          *  descriptor close, not protocol-sourced closes, because the
1818          *  reader process may not have drained the data yet!
1819          */
1820         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1821                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1822                           tcp_hdr(skb)->fin;
1823                 data_was_unread += len;
1824                 __kfree_skb(skb);
1825         }
1826
1827         sk_mem_reclaim(sk);
1828
1829         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1830          * data was lost. To witness the awful effects of the old behavior of
1831          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1832          * GET in an FTP client, suspend the process, wait for the client to
1833          * advertise a zero window, then kill -9 the FTP client, wheee...
1834          * Note: timeout is always zero in such a case.
1835          */
1836         if (data_was_unread) {
1837                 /* Unread data was tossed, zap the connection. */
1838                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1839                 tcp_set_state(sk, TCP_CLOSE);
1840                 tcp_send_active_reset(sk, GFP_KERNEL);
1841         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1842                 /* Check zero linger _after_ checking for unread data. */
1843                 sk->sk_prot->disconnect(sk, 0);
1844                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1845         } else if (tcp_close_state(sk)) {
1846                 /* We FIN if the application ate all the data before
1847                  * zapping the connection.
1848                  */
1849
1850                 /* RED-PEN. Formally speaking, we have broken TCP state
1851                  * machine. State transitions:
1852                  *
1853                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1854                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1855                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1856                  *
1857                  * are legal only when FIN has been sent (i.e. in window),
1858                  * rather than queued out of window. Purists blame.
1859                  *
1860                  * F.e. "RFC state" is ESTABLISHED,
1861                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1862                  *
1863                  * The visible declinations are that sometimes
1864                  * we enter time-wait state, when it is not required really
1865                  * (harmless), do not send active resets, when they are
1866                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1867                  * they look as CLOSING or LAST_ACK for Linux)
1868                  * Probably, I missed some more holelets.
1869                  *                                              --ANK
1870                  */
1871                 tcp_send_fin(sk);
1872         }
1873
1874         sk_stream_wait_close(sk, timeout);
1875
1876 adjudge_to_death:
1877         state = sk->sk_state;
1878         sock_hold(sk);
1879         sock_orphan(sk);
1880
1881         /* It is the last release_sock in its life. It will remove backlog. */
1882         release_sock(sk);
1883
1884
1885         /* Now socket is owned by kernel and we acquire BH lock
1886            to finish close. No need to check for user refs.
1887          */
1888         local_bh_disable();
1889         bh_lock_sock(sk);
1890         WARN_ON(sock_owned_by_user(sk));
1891
1892         percpu_counter_inc(sk->sk_prot->orphan_count);
1893
1894         /* Have we already been destroyed by a softirq or backlog? */
1895         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1896                 goto out;
1897
1898         /*      This is a (useful) BSD violating of the RFC. There is a
1899          *      problem with TCP as specified in that the other end could
1900          *      keep a socket open forever with no application left this end.
1901          *      We use a 3 minute timeout (about the same as BSD) then kill
1902          *      our end. If they send after that then tough - BUT: long enough
1903          *      that we won't make the old 4*rto = almost no time - whoops
1904          *      reset mistake.
1905          *
1906          *      Nope, it was not mistake. It is really desired behaviour
1907          *      f.e. on http servers, when such sockets are useless, but
1908          *      consume significant resources. Let's do it with special
1909          *      linger2 option.                                 --ANK
1910          */
1911
1912         if (sk->sk_state == TCP_FIN_WAIT2) {
1913                 struct tcp_sock *tp = tcp_sk(sk);
1914                 if (tp->linger2 < 0) {
1915                         tcp_set_state(sk, TCP_CLOSE);
1916                         tcp_send_active_reset(sk, GFP_ATOMIC);
1917                         NET_INC_STATS_BH(sock_net(sk),
1918                                         LINUX_MIB_TCPABORTONLINGER);
1919                 } else {
1920                         const int tmo = tcp_fin_time(sk);
1921
1922                         if (tmo > TCP_TIMEWAIT_LEN) {
1923                                 inet_csk_reset_keepalive_timer(sk,
1924                                                 tmo - TCP_TIMEWAIT_LEN);
1925                         } else {
1926                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1927                                 goto out;
1928                         }
1929                 }
1930         }
1931         if (sk->sk_state != TCP_CLOSE) {
1932                 int orphan_count = percpu_counter_read_positive(
1933                                                 sk->sk_prot->orphan_count);
1934
1935                 sk_mem_reclaim(sk);
1936                 if (tcp_too_many_orphans(sk, orphan_count)) {
1937                         if (net_ratelimit())
1938                                 printk(KERN_INFO "TCP: too many of orphaned "
1939                                        "sockets\n");
1940                         tcp_set_state(sk, TCP_CLOSE);
1941                         tcp_send_active_reset(sk, GFP_ATOMIC);
1942                         NET_INC_STATS_BH(sock_net(sk),
1943                                         LINUX_MIB_TCPABORTONMEMORY);
1944                 }
1945         }
1946
1947         if (sk->sk_state == TCP_CLOSE)
1948                 inet_csk_destroy_sock(sk);
1949         /* Otherwise, socket is reprieved until protocol close. */
1950
1951 out:
1952         bh_unlock_sock(sk);
1953         local_bh_enable();
1954         sock_put(sk);
1955 }
1956
1957 /* These states need RST on ABORT according to RFC793 */
1958
1959 static inline int tcp_need_reset(int state)
1960 {
1961         return (1 << state) &
1962                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1963                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1964 }
1965
1966 int tcp_disconnect(struct sock *sk, int flags)
1967 {
1968         struct inet_sock *inet = inet_sk(sk);
1969         struct inet_connection_sock *icsk = inet_csk(sk);
1970         struct tcp_sock *tp = tcp_sk(sk);
1971         int err = 0;
1972         int old_state = sk->sk_state;
1973
1974         if (old_state != TCP_CLOSE)
1975                 tcp_set_state(sk, TCP_CLOSE);
1976
1977         /* ABORT function of RFC793 */
1978         if (old_state == TCP_LISTEN) {
1979                 inet_csk_listen_stop(sk);
1980         } else if (tcp_need_reset(old_state) ||
1981                    (tp->snd_nxt != tp->write_seq &&
1982                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1983                 /* The last check adjusts for discrepancy of Linux wrt. RFC
1984                  * states
1985                  */
1986                 tcp_send_active_reset(sk, gfp_any());
1987                 sk->sk_err = ECONNRESET;
1988         } else if (old_state == TCP_SYN_SENT)
1989                 sk->sk_err = ECONNRESET;
1990
1991         tcp_clear_xmit_timers(sk);
1992         __skb_queue_purge(&sk->sk_receive_queue);
1993         tcp_write_queue_purge(sk);
1994         __skb_queue_purge(&tp->out_of_order_queue);
1995 #ifdef CONFIG_NET_DMA
1996         __skb_queue_purge(&sk->sk_async_wait_queue);
1997 #endif
1998
1999         inet->dport = 0;
2000
2001         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2002                 inet_reset_saddr(sk);
2003
2004         sk->sk_shutdown = 0;
2005         sock_reset_flag(sk, SOCK_DONE);
2006         tp->srtt = 0;
2007         if ((tp->write_seq += tp->max_window + 2) == 0)
2008                 tp->write_seq = 1;
2009         icsk->icsk_backoff = 0;
2010         tp->snd_cwnd = 2;
2011         icsk->icsk_probes_out = 0;
2012         tp->packets_out = 0;
2013         tp->snd_ssthresh = 0x7fffffff;
2014         tp->snd_cwnd_cnt = 0;
2015         tp->bytes_acked = 0;
2016         tcp_set_ca_state(sk, TCP_CA_Open);
2017         tcp_clear_retrans(tp);
2018         inet_csk_delack_init(sk);
2019         tcp_init_send_head(sk);
2020         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2021         __sk_dst_reset(sk);
2022
2023         WARN_ON(inet->num && !icsk->icsk_bind_hash);
2024
2025         sk->sk_error_report(sk);
2026         return err;
2027 }
2028
2029 /*
2030  *      Socket option code for TCP.
2031  */
2032 static int do_tcp_setsockopt(struct sock *sk, int level,
2033                 int optname, char __user *optval, int optlen)
2034 {
2035         struct tcp_sock *tp = tcp_sk(sk);
2036         struct inet_connection_sock *icsk = inet_csk(sk);
2037         int val;
2038         int err = 0;
2039
2040         /* This is a string value all the others are int's */
2041         if (optname == TCP_CONGESTION) {
2042                 char name[TCP_CA_NAME_MAX];
2043
2044                 if (optlen < 1)
2045                         return -EINVAL;
2046
2047                 val = strncpy_from_user(name, optval,
2048                                         min(TCP_CA_NAME_MAX-1, optlen));
2049                 if (val < 0)
2050                         return -EFAULT;
2051                 name[val] = 0;
2052
2053                 lock_sock(sk);
2054                 err = tcp_set_congestion_control(sk, name);
2055                 release_sock(sk);
2056                 return err;
2057         }
2058
2059         if (optlen < sizeof(int))
2060                 return -EINVAL;
2061
2062         if (get_user(val, (int __user *)optval))
2063                 return -EFAULT;
2064
2065         lock_sock(sk);
2066
2067         switch (optname) {
2068         case TCP_MAXSEG:
2069                 /* Values greater than interface MTU won't take effect. However
2070                  * at the point when this call is done we typically don't yet
2071                  * know which interface is going to be used */
2072                 if (val < 8 || val > MAX_TCP_WINDOW) {
2073                         err = -EINVAL;
2074                         break;
2075                 }
2076                 tp->rx_opt.user_mss = val;
2077                 break;
2078
2079         case TCP_NODELAY:
2080                 if (val) {
2081                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2082                          * this option on corked socket is remembered, but
2083                          * it is not activated until cork is cleared.
2084                          *
2085                          * However, when TCP_NODELAY is set we make
2086                          * an explicit push, which overrides even TCP_CORK
2087                          * for currently queued segments.
2088                          */
2089                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2090                         tcp_push_pending_frames(sk);
2091                 } else {
2092                         tp->nonagle &= ~TCP_NAGLE_OFF;
2093                 }
2094                 break;
2095
2096         case TCP_CORK:
2097                 /* When set indicates to always queue non-full frames.
2098                  * Later the user clears this option and we transmit
2099                  * any pending partial frames in the queue.  This is
2100                  * meant to be used alongside sendfile() to get properly
2101                  * filled frames when the user (for example) must write
2102                  * out headers with a write() call first and then use
2103                  * sendfile to send out the data parts.
2104                  *
2105                  * TCP_CORK can be set together with TCP_NODELAY and it is
2106                  * stronger than TCP_NODELAY.
2107                  */
2108                 if (val) {
2109                         tp->nonagle |= TCP_NAGLE_CORK;
2110                 } else {
2111                         tp->nonagle &= ~TCP_NAGLE_CORK;
2112                         if (tp->nonagle&TCP_NAGLE_OFF)
2113                                 tp->nonagle |= TCP_NAGLE_PUSH;
2114                         tcp_push_pending_frames(sk);
2115                 }
2116                 break;
2117
2118         case TCP_KEEPIDLE:
2119                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2120                         err = -EINVAL;
2121                 else {
2122                         tp->keepalive_time = val * HZ;
2123                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2124                             !((1 << sk->sk_state) &
2125                               (TCPF_CLOSE | TCPF_LISTEN))) {
2126                                 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2127                                 if (tp->keepalive_time > elapsed)
2128                                         elapsed = tp->keepalive_time - elapsed;
2129                                 else
2130                                         elapsed = 0;
2131                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2132                         }
2133                 }
2134                 break;
2135         case TCP_KEEPINTVL:
2136                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2137                         err = -EINVAL;
2138                 else
2139                         tp->keepalive_intvl = val * HZ;
2140                 break;
2141         case TCP_KEEPCNT:
2142                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2143                         err = -EINVAL;
2144                 else
2145                         tp->keepalive_probes = val;
2146                 break;
2147         case TCP_SYNCNT:
2148                 if (val < 1 || val > MAX_TCP_SYNCNT)
2149                         err = -EINVAL;
2150                 else
2151                         icsk->icsk_syn_retries = val;
2152                 break;
2153
2154         case TCP_LINGER2:
2155                 if (val < 0)
2156                         tp->linger2 = -1;
2157                 else if (val > sysctl_tcp_fin_timeout / HZ)
2158                         tp->linger2 = 0;
2159                 else
2160                         tp->linger2 = val * HZ;
2161                 break;
2162
2163         case TCP_DEFER_ACCEPT:
2164                 icsk->icsk_accept_queue.rskq_defer_accept = 0;
2165                 if (val > 0) {
2166                         /* Translate value in seconds to number of
2167                          * retransmits */
2168                         while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
2169                                val > ((TCP_TIMEOUT_INIT / HZ) <<
2170                                        icsk->icsk_accept_queue.rskq_defer_accept))
2171                                 icsk->icsk_accept_queue.rskq_defer_accept++;
2172                         icsk->icsk_accept_queue.rskq_defer_accept++;
2173                 }
2174                 break;
2175
2176         case TCP_WINDOW_CLAMP:
2177                 if (!val) {
2178                         if (sk->sk_state != TCP_CLOSE) {
2179                                 err = -EINVAL;
2180                                 break;
2181                         }
2182                         tp->window_clamp = 0;
2183                 } else
2184                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2185                                                 SOCK_MIN_RCVBUF / 2 : val;
2186                 break;
2187
2188         case TCP_QUICKACK:
2189                 if (!val) {
2190                         icsk->icsk_ack.pingpong = 1;
2191                 } else {
2192                         icsk->icsk_ack.pingpong = 0;
2193                         if ((1 << sk->sk_state) &
2194                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2195                             inet_csk_ack_scheduled(sk)) {
2196                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2197                                 tcp_cleanup_rbuf(sk, 1);
2198                                 if (!(val & 1))
2199                                         icsk->icsk_ack.pingpong = 1;
2200                         }
2201                 }
2202                 break;
2203
2204 #ifdef CONFIG_TCP_MD5SIG
2205         case TCP_MD5SIG:
2206                 /* Read the IP->Key mappings from userspace */
2207                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2208                 break;
2209 #endif
2210
2211         default:
2212                 err = -ENOPROTOOPT;
2213                 break;
2214         }
2215
2216         release_sock(sk);
2217         return err;
2218 }
2219
2220 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2221                    int optlen)
2222 {
2223         struct inet_connection_sock *icsk = inet_csk(sk);
2224
2225         if (level != SOL_TCP)
2226                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2227                                                      optval, optlen);
2228         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2229 }
2230
2231 #ifdef CONFIG_COMPAT
2232 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2233                           char __user *optval, int optlen)
2234 {
2235         if (level != SOL_TCP)
2236                 return inet_csk_compat_setsockopt(sk, level, optname,
2237                                                   optval, optlen);
2238         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2239 }
2240
2241 EXPORT_SYMBOL(compat_tcp_setsockopt);
2242 #endif
2243
2244 /* Return information about state of tcp endpoint in API format. */
2245 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2246 {
2247         struct tcp_sock *tp = tcp_sk(sk);
2248         const struct inet_connection_sock *icsk = inet_csk(sk);
2249         u32 now = tcp_time_stamp;
2250
2251         memset(info, 0, sizeof(*info));
2252
2253         info->tcpi_state = sk->sk_state;
2254         info->tcpi_ca_state = icsk->icsk_ca_state;
2255         info->tcpi_retransmits = icsk->icsk_retransmits;
2256         info->tcpi_probes = icsk->icsk_probes_out;
2257         info->tcpi_backoff = icsk->icsk_backoff;
2258
2259         if (tp->rx_opt.tstamp_ok)
2260                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2261         if (tcp_is_sack(tp))
2262                 info->tcpi_options |= TCPI_OPT_SACK;
2263         if (tp->rx_opt.wscale_ok) {
2264                 info->tcpi_options |= TCPI_OPT_WSCALE;
2265                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2266                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2267         }
2268
2269         if (tp->ecn_flags&TCP_ECN_OK)
2270                 info->tcpi_options |= TCPI_OPT_ECN;
2271
2272         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2273         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2274         info->tcpi_snd_mss = tp->mss_cache;
2275         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2276
2277         if (sk->sk_state == TCP_LISTEN) {
2278                 info->tcpi_unacked = sk->sk_ack_backlog;
2279                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2280         } else {
2281                 info->tcpi_unacked = tp->packets_out;
2282                 info->tcpi_sacked = tp->sacked_out;
2283         }
2284         info->tcpi_lost = tp->lost_out;
2285         info->tcpi_retrans = tp->retrans_out;
2286         info->tcpi_fackets = tp->fackets_out;
2287
2288         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2289         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2290         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2291
2292         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2293         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2294         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2295         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2296         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2297         info->tcpi_snd_cwnd = tp->snd_cwnd;
2298         info->tcpi_advmss = tp->advmss;
2299         info->tcpi_reordering = tp->reordering;
2300
2301         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2302         info->tcpi_rcv_space = tp->rcvq_space.space;
2303
2304         info->tcpi_total_retrans = tp->total_retrans;
2305 }
2306
2307 EXPORT_SYMBOL_GPL(tcp_get_info);
2308
2309 static int do_tcp_getsockopt(struct sock *sk, int level,
2310                 int optname, char __user *optval, int __user *optlen)
2311 {
2312         struct inet_connection_sock *icsk = inet_csk(sk);
2313         struct tcp_sock *tp = tcp_sk(sk);
2314         int val, len;
2315
2316         if (get_user(len, optlen))
2317                 return -EFAULT;
2318
2319         len = min_t(unsigned int, len, sizeof(int));
2320
2321         if (len < 0)
2322                 return -EINVAL;
2323
2324         switch (optname) {
2325         case TCP_MAXSEG:
2326                 val = tp->mss_cache;
2327                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2328                         val = tp->rx_opt.user_mss;
2329                 break;
2330         case TCP_NODELAY:
2331                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2332                 break;
2333         case TCP_CORK:
2334                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2335                 break;
2336         case TCP_KEEPIDLE:
2337                 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2338                 break;
2339         case TCP_KEEPINTVL:
2340                 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2341                 break;
2342         case TCP_KEEPCNT:
2343                 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2344                 break;
2345         case TCP_SYNCNT:
2346                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2347                 break;
2348         case TCP_LINGER2:
2349                 val = tp->linger2;
2350                 if (val >= 0)
2351                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2352                 break;
2353         case TCP_DEFER_ACCEPT:
2354                 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2355                         ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2356                 break;
2357         case TCP_WINDOW_CLAMP:
2358                 val = tp->window_clamp;
2359                 break;
2360         case TCP_INFO: {
2361                 struct tcp_info info;
2362
2363                 if (get_user(len, optlen))
2364                         return -EFAULT;
2365
2366                 tcp_get_info(sk, &info);
2367
2368                 len = min_t(unsigned int, len, sizeof(info));
2369                 if (put_user(len, optlen))
2370                         return -EFAULT;
2371                 if (copy_to_user(optval, &info, len))
2372                         return -EFAULT;
2373                 return 0;
2374         }
2375         case TCP_QUICKACK:
2376                 val = !icsk->icsk_ack.pingpong;
2377                 break;
2378
2379         case TCP_CONGESTION:
2380                 if (get_user(len, optlen))
2381                         return -EFAULT;
2382                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2383                 if (put_user(len, optlen))
2384                         return -EFAULT;
2385                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2386                         return -EFAULT;
2387                 return 0;
2388         default:
2389                 return -ENOPROTOOPT;
2390         }
2391
2392         if (put_user(len, optlen))
2393                 return -EFAULT;
2394         if (copy_to_user(optval, &val, len))
2395                 return -EFAULT;
2396         return 0;
2397 }
2398
2399 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2400                    int __user *optlen)
2401 {
2402         struct inet_connection_sock *icsk = inet_csk(sk);
2403
2404         if (level != SOL_TCP)
2405                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2406                                                      optval, optlen);
2407         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2408 }
2409
2410 #ifdef CONFIG_COMPAT
2411 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2412                           char __user *optval, int __user *optlen)
2413 {
2414         if (level != SOL_TCP)
2415                 return inet_csk_compat_getsockopt(sk, level, optname,
2416                                                   optval, optlen);
2417         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2418 }
2419
2420 EXPORT_SYMBOL(compat_tcp_getsockopt);
2421 #endif
2422
2423 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2424 {
2425         struct sk_buff *segs = ERR_PTR(-EINVAL);
2426         struct tcphdr *th;
2427         unsigned thlen;
2428         unsigned int seq;
2429         __be32 delta;
2430         unsigned int oldlen;
2431         unsigned int mss;
2432
2433         if (!pskb_may_pull(skb, sizeof(*th)))
2434                 goto out;
2435
2436         th = tcp_hdr(skb);
2437         thlen = th->doff * 4;
2438         if (thlen < sizeof(*th))
2439                 goto out;
2440
2441         if (!pskb_may_pull(skb, thlen))
2442                 goto out;
2443
2444         oldlen = (u16)~skb->len;
2445         __skb_pull(skb, thlen);
2446
2447         mss = skb_shinfo(skb)->gso_size;
2448         if (unlikely(skb->len <= mss))
2449                 goto out;
2450
2451         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2452                 /* Packet is from an untrusted source, reset gso_segs. */
2453                 int type = skb_shinfo(skb)->gso_type;
2454
2455                 if (unlikely(type &
2456                              ~(SKB_GSO_TCPV4 |
2457                                SKB_GSO_DODGY |
2458                                SKB_GSO_TCP_ECN |
2459                                SKB_GSO_TCPV6 |
2460                                0) ||
2461                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2462                         goto out;
2463
2464                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2465
2466                 segs = NULL;
2467                 goto out;
2468         }
2469
2470         segs = skb_segment(skb, features);
2471         if (IS_ERR(segs))
2472                 goto out;
2473
2474         delta = htonl(oldlen + (thlen + mss));
2475
2476         skb = segs;
2477         th = tcp_hdr(skb);
2478         seq = ntohl(th->seq);
2479
2480         do {
2481                 th->fin = th->psh = 0;
2482
2483                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2484                                        (__force u32)delta));
2485                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2486                         th->check =
2487                              csum_fold(csum_partial(skb_transport_header(skb),
2488                                                     thlen, skb->csum));
2489
2490                 seq += mss;
2491                 skb = skb->next;
2492                 th = tcp_hdr(skb);
2493
2494                 th->seq = htonl(seq);
2495                 th->cwr = 0;
2496         } while (skb->next);
2497
2498         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2499                       skb->data_len);
2500         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2501                                 (__force u32)delta));
2502         if (skb->ip_summed != CHECKSUM_PARTIAL)
2503                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2504                                                    thlen, skb->csum));
2505
2506 out:
2507         return segs;
2508 }
2509 EXPORT_SYMBOL(tcp_tso_segment);
2510
2511 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2512 {
2513         struct sk_buff **pp = NULL;
2514         struct sk_buff *p;
2515         struct tcphdr *th;
2516         struct tcphdr *th2;
2517         unsigned int len;
2518         unsigned int thlen;
2519         unsigned int flags;
2520         unsigned int mss = 1;
2521         int flush = 1;
2522         int i;
2523
2524         th = skb_gro_header(skb, sizeof(*th));
2525         if (unlikely(!th))
2526                 goto out;
2527
2528         thlen = th->doff * 4;
2529         if (thlen < sizeof(*th))
2530                 goto out;
2531
2532         th = skb_gro_header(skb, thlen);
2533         if (unlikely(!th))
2534                 goto out;
2535
2536         skb_gro_pull(skb, thlen);
2537
2538         len = skb_gro_len(skb);
2539         flags = tcp_flag_word(th);
2540
2541         for (; (p = *head); head = &p->next) {
2542                 if (!NAPI_GRO_CB(p)->same_flow)
2543                         continue;
2544
2545                 th2 = tcp_hdr(p);
2546
2547                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2548                         NAPI_GRO_CB(p)->same_flow = 0;
2549                         continue;
2550                 }
2551
2552                 goto found;
2553         }
2554
2555         goto out_check_final;
2556
2557 found:
2558         flush = NAPI_GRO_CB(p)->flush;
2559         flush |= flags & TCP_FLAG_CWR;
2560         flush |= (flags ^ tcp_flag_word(th2)) &
2561                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH);
2562         flush |= th->ack_seq ^ th2->ack_seq;
2563         for (i = sizeof(*th); !flush && i < thlen; i += 4)
2564                 flush |= *(u32 *)((u8 *)th + i) ^
2565                          *(u32 *)((u8 *)th2 + i);
2566
2567         mss = skb_shinfo(p)->gso_size;
2568
2569         flush |= (len > mss) | !len;
2570         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2571
2572         if (flush || skb_gro_receive(head, skb)) {
2573                 mss = 1;
2574                 goto out_check_final;
2575         }
2576
2577         p = *head;
2578         th2 = tcp_hdr(p);
2579         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2580
2581 out_check_final:
2582         flush = len < mss;
2583         flush |= flags & (TCP_FLAG_URG | TCP_FLAG_PSH | TCP_FLAG_RST |
2584                           TCP_FLAG_SYN | TCP_FLAG_FIN);
2585
2586         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2587                 pp = head;
2588
2589 out:
2590         NAPI_GRO_CB(skb)->flush |= flush;
2591
2592         return pp;
2593 }
2594 EXPORT_SYMBOL(tcp_gro_receive);
2595
2596 int tcp_gro_complete(struct sk_buff *skb)
2597 {
2598         struct tcphdr *th = tcp_hdr(skb);
2599
2600         skb->csum_start = skb_transport_header(skb) - skb->head;
2601         skb->csum_offset = offsetof(struct tcphdr, check);
2602         skb->ip_summed = CHECKSUM_PARTIAL;
2603
2604         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2605
2606         if (th->cwr)
2607                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2608
2609         return 0;
2610 }
2611 EXPORT_SYMBOL(tcp_gro_complete);
2612
2613 #ifdef CONFIG_TCP_MD5SIG
2614 static unsigned long tcp_md5sig_users;
2615 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2616 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2617
2618 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2619 {
2620         int cpu;
2621         for_each_possible_cpu(cpu) {
2622                 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2623                 if (p) {
2624                         if (p->md5_desc.tfm)
2625                                 crypto_free_hash(p->md5_desc.tfm);
2626                         kfree(p);
2627                         p = NULL;
2628                 }
2629         }
2630         free_percpu(pool);
2631 }
2632
2633 void tcp_free_md5sig_pool(void)
2634 {
2635         struct tcp_md5sig_pool **pool = NULL;
2636
2637         spin_lock_bh(&tcp_md5sig_pool_lock);
2638         if (--tcp_md5sig_users == 0) {
2639                 pool = tcp_md5sig_pool;
2640                 tcp_md5sig_pool = NULL;
2641         }
2642         spin_unlock_bh(&tcp_md5sig_pool_lock);
2643         if (pool)
2644                 __tcp_free_md5sig_pool(pool);
2645 }
2646
2647 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2648
2649 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2650 {
2651         int cpu;
2652         struct tcp_md5sig_pool **pool;
2653
2654         pool = alloc_percpu(struct tcp_md5sig_pool *);
2655         if (!pool)
2656                 return NULL;
2657
2658         for_each_possible_cpu(cpu) {
2659                 struct tcp_md5sig_pool *p;
2660                 struct crypto_hash *hash;
2661
2662                 p = kzalloc(sizeof(*p), GFP_KERNEL);
2663                 if (!p)
2664                         goto out_free;
2665                 *per_cpu_ptr(pool, cpu) = p;
2666
2667                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2668                 if (!hash || IS_ERR(hash))
2669                         goto out_free;
2670
2671                 p->md5_desc.tfm = hash;
2672         }
2673         return pool;
2674 out_free:
2675         __tcp_free_md5sig_pool(pool);
2676         return NULL;
2677 }
2678
2679 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2680 {
2681         struct tcp_md5sig_pool **pool;
2682         int alloc = 0;
2683
2684 retry:
2685         spin_lock_bh(&tcp_md5sig_pool_lock);
2686         pool = tcp_md5sig_pool;
2687         if (tcp_md5sig_users++ == 0) {
2688                 alloc = 1;
2689                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2690         } else if (!pool) {
2691                 tcp_md5sig_users--;
2692                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2693                 cpu_relax();
2694                 goto retry;
2695         } else
2696                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2697
2698         if (alloc) {
2699                 /* we cannot hold spinlock here because this may sleep. */
2700                 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2701                 spin_lock_bh(&tcp_md5sig_pool_lock);
2702                 if (!p) {
2703                         tcp_md5sig_users--;
2704                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2705                         return NULL;
2706                 }
2707                 pool = tcp_md5sig_pool;
2708                 if (pool) {
2709                         /* oops, it has already been assigned. */
2710                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2711                         __tcp_free_md5sig_pool(p);
2712                 } else {
2713                         tcp_md5sig_pool = pool = p;
2714                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2715                 }
2716         }
2717         return pool;
2718 }
2719
2720 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2721
2722 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2723 {
2724         struct tcp_md5sig_pool **p;
2725         spin_lock_bh(&tcp_md5sig_pool_lock);
2726         p = tcp_md5sig_pool;
2727         if (p)
2728                 tcp_md5sig_users++;
2729         spin_unlock_bh(&tcp_md5sig_pool_lock);
2730         return (p ? *per_cpu_ptr(p, cpu) : NULL);
2731 }
2732
2733 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2734
2735 void __tcp_put_md5sig_pool(void)
2736 {
2737         tcp_free_md5sig_pool();
2738 }
2739
2740 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2741
2742 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2743                         struct tcphdr *th)
2744 {
2745         struct scatterlist sg;
2746         int err;
2747
2748         __sum16 old_checksum = th->check;
2749         th->check = 0;
2750         /* options aren't included in the hash */
2751         sg_init_one(&sg, th, sizeof(struct tcphdr));
2752         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2753         th->check = old_checksum;
2754         return err;
2755 }
2756
2757 EXPORT_SYMBOL(tcp_md5_hash_header);
2758
2759 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2760                           struct sk_buff *skb, unsigned header_len)
2761 {
2762         struct scatterlist sg;
2763         const struct tcphdr *tp = tcp_hdr(skb);
2764         struct hash_desc *desc = &hp->md5_desc;
2765         unsigned i;
2766         const unsigned head_data_len = skb_headlen(skb) > header_len ?
2767                                        skb_headlen(skb) - header_len : 0;
2768         const struct skb_shared_info *shi = skb_shinfo(skb);
2769
2770         sg_init_table(&sg, 1);
2771
2772         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2773         if (crypto_hash_update(desc, &sg, head_data_len))
2774                 return 1;
2775
2776         for (i = 0; i < shi->nr_frags; ++i) {
2777                 const struct skb_frag_struct *f = &shi->frags[i];
2778                 sg_set_page(&sg, f->page, f->size, f->page_offset);
2779                 if (crypto_hash_update(desc, &sg, f->size))
2780                         return 1;
2781         }
2782
2783         return 0;
2784 }
2785
2786 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2787
2788 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
2789 {
2790         struct scatterlist sg;
2791
2792         sg_init_one(&sg, key->key, key->keylen);
2793         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2794 }
2795
2796 EXPORT_SYMBOL(tcp_md5_hash_key);
2797
2798 #endif
2799
2800 void tcp_done(struct sock *sk)
2801 {
2802         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2803                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
2804
2805         tcp_set_state(sk, TCP_CLOSE);
2806         tcp_clear_xmit_timers(sk);
2807
2808         sk->sk_shutdown = SHUTDOWN_MASK;
2809
2810         if (!sock_flag(sk, SOCK_DEAD))
2811                 sk->sk_state_change(sk);
2812         else
2813                 inet_csk_destroy_sock(sk);
2814 }
2815 EXPORT_SYMBOL_GPL(tcp_done);
2816
2817 extern struct tcp_congestion_ops tcp_reno;
2818
2819 static __initdata unsigned long thash_entries;
2820 static int __init set_thash_entries(char *str)
2821 {
2822         if (!str)
2823                 return 0;
2824         thash_entries = simple_strtoul(str, &str, 0);
2825         return 1;
2826 }
2827 __setup("thash_entries=", set_thash_entries);
2828
2829 void __init tcp_init(void)
2830 {
2831         struct sk_buff *skb = NULL;
2832         unsigned long nr_pages, limit;
2833         int order, i, max_share;
2834
2835         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
2836
2837         percpu_counter_init(&tcp_sockets_allocated, 0);
2838         percpu_counter_init(&tcp_orphan_count, 0);
2839         tcp_hashinfo.bind_bucket_cachep =
2840                 kmem_cache_create("tcp_bind_bucket",
2841                                   sizeof(struct inet_bind_bucket), 0,
2842                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
2843
2844         /* Size and allocate the main established and bind bucket
2845          * hash tables.
2846          *
2847          * The methodology is similar to that of the buffer cache.
2848          */
2849         tcp_hashinfo.ehash =
2850                 alloc_large_system_hash("TCP established",
2851                                         sizeof(struct inet_ehash_bucket),
2852                                         thash_entries,
2853                                         (num_physpages >= 128 * 1024) ?
2854                                         13 : 15,
2855                                         0,
2856                                         &tcp_hashinfo.ehash_size,
2857                                         NULL,
2858                                         thash_entries ? 0 : 512 * 1024);
2859         tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2860         for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2861                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
2862                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
2863         }
2864         if (inet_ehash_locks_alloc(&tcp_hashinfo))
2865                 panic("TCP: failed to alloc ehash_locks");
2866         tcp_hashinfo.bhash =
2867                 alloc_large_system_hash("TCP bind",
2868                                         sizeof(struct inet_bind_hashbucket),
2869                                         tcp_hashinfo.ehash_size,
2870                                         (num_physpages >= 128 * 1024) ?
2871                                         13 : 15,
2872                                         0,
2873                                         &tcp_hashinfo.bhash_size,
2874                                         NULL,
2875                                         64 * 1024);
2876         tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2877         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2878                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2879                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2880         }
2881
2882         /* Try to be a bit smarter and adjust defaults depending
2883          * on available memory.
2884          */
2885         for (order = 0; ((1 << order) << PAGE_SHIFT) <
2886                         (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2887                         order++)
2888                 ;
2889         if (order >= 4) {
2890                 tcp_death_row.sysctl_max_tw_buckets = 180000;
2891                 sysctl_tcp_max_orphans = 4096 << (order - 4);
2892                 sysctl_max_syn_backlog = 1024;
2893         } else if (order < 3) {
2894                 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2895                 sysctl_tcp_max_orphans >>= (3 - order);
2896                 sysctl_max_syn_backlog = 128;
2897         }
2898
2899         /* Set the pressure threshold to be a fraction of global memory that
2900          * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2901          * memory, with a floor of 128 pages.
2902          */
2903         nr_pages = totalram_pages - totalhigh_pages;
2904         limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2905         limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2906         limit = max(limit, 128UL);
2907         sysctl_tcp_mem[0] = limit / 4 * 3;
2908         sysctl_tcp_mem[1] = limit;
2909         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2910
2911         /* Set per-socket limits to no more than 1/128 the pressure threshold */
2912         limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2913         max_share = min(4UL*1024*1024, limit);
2914
2915         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
2916         sysctl_tcp_wmem[1] = 16*1024;
2917         sysctl_tcp_wmem[2] = max(64*1024, max_share);
2918
2919         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
2920         sysctl_tcp_rmem[1] = 87380;
2921         sysctl_tcp_rmem[2] = max(87380, max_share);
2922
2923         printk(KERN_INFO "TCP: Hash tables configured "
2924                "(established %d bind %d)\n",
2925                tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2926
2927         tcp_register_congestion_control(&tcp_reno);
2928 }
2929
2930 EXPORT_SYMBOL(tcp_close);
2931 EXPORT_SYMBOL(tcp_disconnect);
2932 EXPORT_SYMBOL(tcp_getsockopt);
2933 EXPORT_SYMBOL(tcp_ioctl);
2934 EXPORT_SYMBOL(tcp_poll);
2935 EXPORT_SYMBOL(tcp_read_sock);
2936 EXPORT_SYMBOL(tcp_recvmsg);
2937 EXPORT_SYMBOL(tcp_sendmsg);
2938 EXPORT_SYMBOL(tcp_splice_read);
2939 EXPORT_SYMBOL(tcp_sendpage);
2940 EXPORT_SYMBOL(tcp_setsockopt);
2941 EXPORT_SYMBOL(tcp_shutdown);