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