tcp: fix splice() and tcp collapsing interaction
[opensuse:kernel.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 #define pr_fmt(fmt) "TCP: " fmt
249
250 #include <linux/kernel.h>
251 #include <linux/module.h>
252 #include <linux/types.h>
253 #include <linux/fcntl.h>
254 #include <linux/poll.h>
255 #include <linux/init.h>
256 #include <linux/fs.h>
257 #include <linux/skbuff.h>
258 #include <linux/scatterlist.h>
259 #include <linux/splice.h>
260 #include <linux/net.h>
261 #include <linux/socket.h>
262 #include <linux/random.h>
263 #include <linux/bootmem.h>
264 #include <linux/highmem.h>
265 #include <linux/swap.h>
266 #include <linux/cache.h>
267 #include <linux/err.h>
268 #include <linux/crypto.h>
269 #include <linux/time.h>
270 #include <linux/slab.h>
271
272 #include <net/icmp.h>
273 #include <net/inet_common.h>
274 #include <net/tcp.h>
275 #include <net/xfrm.h>
276 #include <net/ip.h>
277 #include <net/netdma.h>
278 #include <net/sock.h>
279
280 #include <asm/uaccess.h>
281 #include <asm/ioctls.h>
282
283 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
284
285 struct percpu_counter tcp_orphan_count;
286 EXPORT_SYMBOL_GPL(tcp_orphan_count);
287
288 int sysctl_tcp_wmem[3] __read_mostly;
289 int sysctl_tcp_rmem[3] __read_mostly;
290
291 EXPORT_SYMBOL(sysctl_tcp_rmem);
292 EXPORT_SYMBOL(sysctl_tcp_wmem);
293
294 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
295 EXPORT_SYMBOL(tcp_memory_allocated);
296
297 /*
298  * Current number of TCP sockets.
299  */
300 struct percpu_counter tcp_sockets_allocated;
301 EXPORT_SYMBOL(tcp_sockets_allocated);
302
303 /*
304  * TCP splice context
305  */
306 struct tcp_splice_state {
307         struct pipe_inode_info *pipe;
308         size_t len;
309         unsigned int flags;
310 };
311
312 /*
313  * Pressure flag: try to collapse.
314  * Technical note: it is used by multiple contexts non atomically.
315  * All the __sk_mem_schedule() is of this nature: accounting
316  * is strict, actions are advisory and have some latency.
317  */
318 int tcp_memory_pressure __read_mostly;
319 EXPORT_SYMBOL(tcp_memory_pressure);
320
321 void tcp_enter_memory_pressure(struct sock *sk)
322 {
323         if (!tcp_memory_pressure) {
324                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
325                 tcp_memory_pressure = 1;
326         }
327 }
328 EXPORT_SYMBOL(tcp_enter_memory_pressure);
329
330 /* Convert seconds to retransmits based on initial and max timeout */
331 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
332 {
333         u8 res = 0;
334
335         if (seconds > 0) {
336                 int period = timeout;
337
338                 res = 1;
339                 while (seconds > period && res < 255) {
340                         res++;
341                         timeout <<= 1;
342                         if (timeout > rto_max)
343                                 timeout = rto_max;
344                         period += timeout;
345                 }
346         }
347         return res;
348 }
349
350 /* Convert retransmits to seconds based on initial and max timeout */
351 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
352 {
353         int period = 0;
354
355         if (retrans > 0) {
356                 period = timeout;
357                 while (--retrans) {
358                         timeout <<= 1;
359                         if (timeout > rto_max)
360                                 timeout = rto_max;
361                         period += timeout;
362                 }
363         }
364         return period;
365 }
366
367 /* Address-family independent initialization for a tcp_sock.
368  *
369  * NOTE: A lot of things set to zero explicitly by call to
370  *       sk_alloc() so need not be done here.
371  */
372 void tcp_init_sock(struct sock *sk)
373 {
374         struct inet_connection_sock *icsk = inet_csk(sk);
375         struct tcp_sock *tp = tcp_sk(sk);
376
377         skb_queue_head_init(&tp->out_of_order_queue);
378         tcp_init_xmit_timers(sk);
379         tcp_prequeue_init(tp);
380         INIT_LIST_HEAD(&tp->tsq_node);
381
382         icsk->icsk_rto = TCP_TIMEOUT_INIT;
383         tp->mdev = TCP_TIMEOUT_INIT;
384
385         /* So many TCP implementations out there (incorrectly) count the
386          * initial SYN frame in their delayed-ACK and congestion control
387          * algorithms that we must have the following bandaid to talk
388          * efficiently to them.  -DaveM
389          */
390         tp->snd_cwnd = TCP_INIT_CWND;
391
392         /* See draft-stevens-tcpca-spec-01 for discussion of the
393          * initialization of these values.
394          */
395         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
396         tp->snd_cwnd_clamp = ~0;
397         tp->mss_cache = TCP_MSS_DEFAULT;
398
399         tp->reordering = sysctl_tcp_reordering;
400         tcp_enable_early_retrans(tp);
401         icsk->icsk_ca_ops = &tcp_init_congestion_ops;
402
403         sk->sk_state = TCP_CLOSE;
404
405         sk->sk_write_space = sk_stream_write_space;
406         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
407
408         icsk->icsk_sync_mss = tcp_sync_mss;
409
410         /* TCP Cookie Transactions */
411         if (sysctl_tcp_cookie_size > 0) {
412                 /* Default, cookies without s_data_payload. */
413                 tp->cookie_values =
414                         kzalloc(sizeof(*tp->cookie_values),
415                                 sk->sk_allocation);
416                 if (tp->cookie_values != NULL)
417                         kref_init(&tp->cookie_values->kref);
418         }
419         /* Presumed zeroed, in order of appearance:
420          *      cookie_in_always, cookie_out_never,
421          *      s_data_constant, s_data_in, s_data_out
422          */
423         sk->sk_sndbuf = sysctl_tcp_wmem[1];
424         sk->sk_rcvbuf = sysctl_tcp_rmem[1];
425
426         local_bh_disable();
427         sock_update_memcg(sk);
428         sk_sockets_allocated_inc(sk);
429         local_bh_enable();
430 }
431 EXPORT_SYMBOL(tcp_init_sock);
432
433 /*
434  *      Wait for a TCP event.
435  *
436  *      Note that we don't need to lock the socket, as the upper poll layers
437  *      take care of normal races (between the test and the event) and we don't
438  *      go look at any of the socket buffers directly.
439  */
440 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
441 {
442         unsigned int mask;
443         struct sock *sk = sock->sk;
444         const struct tcp_sock *tp = tcp_sk(sk);
445
446         sock_poll_wait(file, sk_sleep(sk), wait);
447         if (sk->sk_state == TCP_LISTEN)
448                 return inet_csk_listen_poll(sk);
449
450         /* Socket is not locked. We are protected from async events
451          * by poll logic and correct handling of state changes
452          * made by other threads is impossible in any case.
453          */
454
455         mask = 0;
456
457         /*
458          * POLLHUP is certainly not done right. But poll() doesn't
459          * have a notion of HUP in just one direction, and for a
460          * socket the read side is more interesting.
461          *
462          * Some poll() documentation says that POLLHUP is incompatible
463          * with the POLLOUT/POLLWR flags, so somebody should check this
464          * all. But careful, it tends to be safer to return too many
465          * bits than too few, and you can easily break real applications
466          * if you don't tell them that something has hung up!
467          *
468          * Check-me.
469          *
470          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
471          * our fs/select.c). It means that after we received EOF,
472          * poll always returns immediately, making impossible poll() on write()
473          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
474          * if and only if shutdown has been made in both directions.
475          * Actually, it is interesting to look how Solaris and DUX
476          * solve this dilemma. I would prefer, if POLLHUP were maskable,
477          * then we could set it on SND_SHUTDOWN. BTW examples given
478          * in Stevens' books assume exactly this behaviour, it explains
479          * why POLLHUP is incompatible with POLLOUT.    --ANK
480          *
481          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
482          * blocking on fresh not-connected or disconnected socket. --ANK
483          */
484         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
485                 mask |= POLLHUP;
486         if (sk->sk_shutdown & RCV_SHUTDOWN)
487                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
488
489         /* Connected or passive Fast Open socket? */
490         if (sk->sk_state != TCP_SYN_SENT &&
491             (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
492                 int target = sock_rcvlowat(sk, 0, INT_MAX);
493
494                 if (tp->urg_seq == tp->copied_seq &&
495                     !sock_flag(sk, SOCK_URGINLINE) &&
496                     tp->urg_data)
497                         target++;
498
499                 /* Potential race condition. If read of tp below will
500                  * escape above sk->sk_state, we can be illegally awaken
501                  * in SYN_* states. */
502                 if (tp->rcv_nxt - tp->copied_seq >= target)
503                         mask |= POLLIN | POLLRDNORM;
504
505                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
506                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
507                                 mask |= POLLOUT | POLLWRNORM;
508                         } else {  /* send SIGIO later */
509                                 set_bit(SOCK_ASYNC_NOSPACE,
510                                         &sk->sk_socket->flags);
511                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
512
513                                 /* Race breaker. If space is freed after
514                                  * wspace test but before the flags are set,
515                                  * IO signal will be lost.
516                                  */
517                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
518                                         mask |= POLLOUT | POLLWRNORM;
519                         }
520                 } else
521                         mask |= POLLOUT | POLLWRNORM;
522
523                 if (tp->urg_data & TCP_URG_VALID)
524                         mask |= POLLPRI;
525         }
526         /* This barrier is coupled with smp_wmb() in tcp_reset() */
527         smp_rmb();
528         if (sk->sk_err)
529                 mask |= POLLERR;
530
531         return mask;
532 }
533 EXPORT_SYMBOL(tcp_poll);
534
535 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
536 {
537         struct tcp_sock *tp = tcp_sk(sk);
538         int answ;
539
540         switch (cmd) {
541         case SIOCINQ:
542                 if (sk->sk_state == TCP_LISTEN)
543                         return -EINVAL;
544
545                 lock_sock(sk);
546                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
547                         answ = 0;
548                 else if (sock_flag(sk, SOCK_URGINLINE) ||
549                          !tp->urg_data ||
550                          before(tp->urg_seq, tp->copied_seq) ||
551                          !before(tp->urg_seq, tp->rcv_nxt)) {
552
553                         answ = tp->rcv_nxt - tp->copied_seq;
554
555                         /* Subtract 1, if FIN was received */
556                         if (answ && sock_flag(sk, SOCK_DONE))
557                                 answ--;
558                 } else
559                         answ = tp->urg_seq - tp->copied_seq;
560                 release_sock(sk);
561                 break;
562         case SIOCATMARK:
563                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
564                 break;
565         case SIOCOUTQ:
566                 if (sk->sk_state == TCP_LISTEN)
567                         return -EINVAL;
568
569                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
570                         answ = 0;
571                 else
572                         answ = tp->write_seq - tp->snd_una;
573                 break;
574         case SIOCOUTQNSD:
575                 if (sk->sk_state == TCP_LISTEN)
576                         return -EINVAL;
577
578                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
579                         answ = 0;
580                 else
581                         answ = tp->write_seq - tp->snd_nxt;
582                 break;
583         default:
584                 return -ENOIOCTLCMD;
585         }
586
587         return put_user(answ, (int __user *)arg);
588 }
589 EXPORT_SYMBOL(tcp_ioctl);
590
591 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
592 {
593         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
594         tp->pushed_seq = tp->write_seq;
595 }
596
597 static inline bool forced_push(const struct tcp_sock *tp)
598 {
599         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
600 }
601
602 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
603 {
604         struct tcp_sock *tp = tcp_sk(sk);
605         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
606
607         skb->csum    = 0;
608         tcb->seq     = tcb->end_seq = tp->write_seq;
609         tcb->tcp_flags = TCPHDR_ACK;
610         tcb->sacked  = 0;
611         skb_header_release(skb);
612         tcp_add_write_queue_tail(sk, skb);
613         sk->sk_wmem_queued += skb->truesize;
614         sk_mem_charge(sk, skb->truesize);
615         if (tp->nonagle & TCP_NAGLE_PUSH)
616                 tp->nonagle &= ~TCP_NAGLE_PUSH;
617 }
618
619 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
620 {
621         if (flags & MSG_OOB)
622                 tp->snd_up = tp->write_seq;
623 }
624
625 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
626                             int nonagle)
627 {
628         if (tcp_send_head(sk)) {
629                 struct tcp_sock *tp = tcp_sk(sk);
630
631                 if (!(flags & MSG_MORE) || forced_push(tp))
632                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
633
634                 tcp_mark_urg(tp, flags);
635                 __tcp_push_pending_frames(sk, mss_now,
636                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
637         }
638 }
639
640 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
641                                 unsigned int offset, size_t len)
642 {
643         struct tcp_splice_state *tss = rd_desc->arg.data;
644         int ret;
645
646         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
647                               tss->flags);
648         if (ret > 0)
649                 rd_desc->count -= ret;
650         return ret;
651 }
652
653 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
654 {
655         /* Store TCP splice context information in read_descriptor_t. */
656         read_descriptor_t rd_desc = {
657                 .arg.data = tss,
658                 .count    = tss->len,
659         };
660
661         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
662 }
663
664 /**
665  *  tcp_splice_read - splice data from TCP socket to a pipe
666  * @sock:       socket to splice from
667  * @ppos:       position (not valid)
668  * @pipe:       pipe to splice to
669  * @len:        number of bytes to splice
670  * @flags:      splice modifier flags
671  *
672  * Description:
673  *    Will read pages from given socket and fill them into a pipe.
674  *
675  **/
676 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
677                         struct pipe_inode_info *pipe, size_t len,
678                         unsigned int flags)
679 {
680         struct sock *sk = sock->sk;
681         struct tcp_splice_state tss = {
682                 .pipe = pipe,
683                 .len = len,
684                 .flags = flags,
685         };
686         long timeo;
687         ssize_t spliced;
688         int ret;
689
690         sock_rps_record_flow(sk);
691         /*
692          * We can't seek on a socket input
693          */
694         if (unlikely(*ppos))
695                 return -ESPIPE;
696
697         ret = spliced = 0;
698
699         lock_sock(sk);
700
701         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
702         while (tss.len) {
703                 ret = __tcp_splice_read(sk, &tss);
704                 if (ret < 0)
705                         break;
706                 else if (!ret) {
707                         if (spliced)
708                                 break;
709                         if (sock_flag(sk, SOCK_DONE))
710                                 break;
711                         if (sk->sk_err) {
712                                 ret = sock_error(sk);
713                                 break;
714                         }
715                         if (sk->sk_shutdown & RCV_SHUTDOWN)
716                                 break;
717                         if (sk->sk_state == TCP_CLOSE) {
718                                 /*
719                                  * This occurs when user tries to read
720                                  * from never connected socket.
721                                  */
722                                 if (!sock_flag(sk, SOCK_DONE))
723                                         ret = -ENOTCONN;
724                                 break;
725                         }
726                         if (!timeo) {
727                                 ret = -EAGAIN;
728                                 break;
729                         }
730                         sk_wait_data(sk, &timeo);
731                         if (signal_pending(current)) {
732                                 ret = sock_intr_errno(timeo);
733                                 break;
734                         }
735                         continue;
736                 }
737                 tss.len -= ret;
738                 spliced += ret;
739
740                 if (!timeo)
741                         break;
742                 release_sock(sk);
743                 lock_sock(sk);
744
745                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
746                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
747                     signal_pending(current))
748                         break;
749         }
750
751         release_sock(sk);
752
753         if (spliced)
754                 return spliced;
755
756         return ret;
757 }
758 EXPORT_SYMBOL(tcp_splice_read);
759
760 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
761 {
762         struct sk_buff *skb;
763
764         /* The TCP header must be at least 32-bit aligned.  */
765         size = ALIGN(size, 4);
766
767         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
768         if (skb) {
769                 if (sk_wmem_schedule(sk, skb->truesize)) {
770                         skb_reserve(skb, sk->sk_prot->max_header);
771                         /*
772                          * Make sure that we have exactly size bytes
773                          * available to the caller, no more, no less.
774                          */
775                         skb->avail_size = size;
776                         return skb;
777                 }
778                 __kfree_skb(skb);
779         } else {
780                 sk->sk_prot->enter_memory_pressure(sk);
781                 sk_stream_moderate_sndbuf(sk);
782         }
783         return NULL;
784 }
785
786 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
787                                        int large_allowed)
788 {
789         struct tcp_sock *tp = tcp_sk(sk);
790         u32 xmit_size_goal, old_size_goal;
791
792         xmit_size_goal = mss_now;
793
794         if (large_allowed && sk_can_gso(sk)) {
795                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
796                                   inet_csk(sk)->icsk_af_ops->net_header_len -
797                                   inet_csk(sk)->icsk_ext_hdr_len -
798                                   tp->tcp_header_len);
799
800                 /* TSQ : try to have two TSO segments in flight */
801                 xmit_size_goal = min_t(u32, xmit_size_goal,
802                                        sysctl_tcp_limit_output_bytes >> 1);
803
804                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
805
806                 /* We try hard to avoid divides here */
807                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
808
809                 if (likely(old_size_goal <= xmit_size_goal &&
810                            old_size_goal + mss_now > xmit_size_goal)) {
811                         xmit_size_goal = old_size_goal;
812                 } else {
813                         tp->xmit_size_goal_segs =
814                                 min_t(u16, xmit_size_goal / mss_now,
815                                       sk->sk_gso_max_segs);
816                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
817                 }
818         }
819
820         return max(xmit_size_goal, mss_now);
821 }
822
823 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
824 {
825         int mss_now;
826
827         mss_now = tcp_current_mss(sk);
828         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
829
830         return mss_now;
831 }
832
833 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
834                                 size_t size, int flags)
835 {
836         struct tcp_sock *tp = tcp_sk(sk);
837         int mss_now, size_goal;
838         int err;
839         ssize_t copied;
840         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
841
842         /* Wait for a connection to finish. One exception is TCP Fast Open
843          * (passive side) where data is allowed to be sent before a connection
844          * is fully established.
845          */
846         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
847             !tcp_passive_fastopen(sk)) {
848                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
849                         goto out_err;
850         }
851
852         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
853
854         mss_now = tcp_send_mss(sk, &size_goal, flags);
855         copied = 0;
856
857         err = -EPIPE;
858         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
859                 goto out_err;
860
861         while (size > 0) {
862                 struct sk_buff *skb = tcp_write_queue_tail(sk);
863                 int copy, i;
864                 bool can_coalesce;
865
866                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
867 new_segment:
868                         if (!sk_stream_memory_free(sk))
869                                 goto wait_for_sndbuf;
870
871                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
872                         if (!skb)
873                                 goto wait_for_memory;
874
875                         skb_entail(sk, skb);
876                         copy = size_goal;
877                 }
878
879                 if (copy > size)
880                         copy = size;
881
882                 i = skb_shinfo(skb)->nr_frags;
883                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
884                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
885                         tcp_mark_push(tp, skb);
886                         goto new_segment;
887                 }
888                 if (!sk_wmem_schedule(sk, copy))
889                         goto wait_for_memory;
890
891                 if (can_coalesce) {
892                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
893                 } else {
894                         get_page(page);
895                         skb_fill_page_desc(skb, i, page, offset, copy);
896                 }
897
898                 skb->len += copy;
899                 skb->data_len += copy;
900                 skb->truesize += copy;
901                 sk->sk_wmem_queued += copy;
902                 sk_mem_charge(sk, copy);
903                 skb->ip_summed = CHECKSUM_PARTIAL;
904                 tp->write_seq += copy;
905                 TCP_SKB_CB(skb)->end_seq += copy;
906                 skb_shinfo(skb)->gso_segs = 0;
907
908                 if (!copied)
909                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
910
911                 copied += copy;
912                 offset += copy;
913                 if (!(size -= copy))
914                         goto out;
915
916                 if (skb->len < size_goal || (flags & MSG_OOB))
917                         continue;
918
919                 if (forced_push(tp)) {
920                         tcp_mark_push(tp, skb);
921                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
922                 } else if (skb == tcp_send_head(sk))
923                         tcp_push_one(sk, mss_now);
924                 continue;
925
926 wait_for_sndbuf:
927                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
928 wait_for_memory:
929                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
930
931                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
932                         goto do_error;
933
934                 mss_now = tcp_send_mss(sk, &size_goal, flags);
935         }
936
937 out:
938         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
939                 tcp_push(sk, flags, mss_now, tp->nonagle);
940         return copied;
941
942 do_error:
943         if (copied)
944                 goto out;
945 out_err:
946         return sk_stream_error(sk, flags, err);
947 }
948
949 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
950                  size_t size, int flags)
951 {
952         ssize_t res;
953
954         if (!(sk->sk_route_caps & NETIF_F_SG) ||
955             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
956                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
957                                         flags);
958
959         lock_sock(sk);
960         res = do_tcp_sendpages(sk, page, offset, size, flags);
961         release_sock(sk);
962         return res;
963 }
964 EXPORT_SYMBOL(tcp_sendpage);
965
966 static inline int select_size(const struct sock *sk, bool sg)
967 {
968         const struct tcp_sock *tp = tcp_sk(sk);
969         int tmp = tp->mss_cache;
970
971         if (sg) {
972                 if (sk_can_gso(sk)) {
973                         /* Small frames wont use a full page:
974                          * Payload will immediately follow tcp header.
975                          */
976                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
977                 } else {
978                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
979
980                         if (tmp >= pgbreak &&
981                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
982                                 tmp = pgbreak;
983                 }
984         }
985
986         return tmp;
987 }
988
989 void tcp_free_fastopen_req(struct tcp_sock *tp)
990 {
991         if (tp->fastopen_req != NULL) {
992                 kfree(tp->fastopen_req);
993                 tp->fastopen_req = NULL;
994         }
995 }
996
997 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size)
998 {
999         struct tcp_sock *tp = tcp_sk(sk);
1000         int err, flags;
1001
1002         if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1003                 return -EOPNOTSUPP;
1004         if (tp->fastopen_req != NULL)
1005                 return -EALREADY; /* Another Fast Open is in progress */
1006
1007         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1008                                    sk->sk_allocation);
1009         if (unlikely(tp->fastopen_req == NULL))
1010                 return -ENOBUFS;
1011         tp->fastopen_req->data = msg;
1012
1013         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1014         err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1015                                     msg->msg_namelen, flags);
1016         *size = tp->fastopen_req->copied;
1017         tcp_free_fastopen_req(tp);
1018         return err;
1019 }
1020
1021 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1022                 size_t size)
1023 {
1024         struct iovec *iov;
1025         struct tcp_sock *tp = tcp_sk(sk);
1026         struct sk_buff *skb;
1027         int iovlen, flags, err, copied = 0;
1028         int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1029         bool sg;
1030         long timeo;
1031
1032         lock_sock(sk);
1033
1034         flags = msg->msg_flags;
1035         if (flags & MSG_FASTOPEN) {
1036                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn);
1037                 if (err == -EINPROGRESS && copied_syn > 0)
1038                         goto out;
1039                 else if (err)
1040                         goto out_err;
1041                 offset = copied_syn;
1042         }
1043
1044         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1045
1046         /* Wait for a connection to finish. One exception is TCP Fast Open
1047          * (passive side) where data is allowed to be sent before a connection
1048          * is fully established.
1049          */
1050         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1051             !tcp_passive_fastopen(sk)) {
1052                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1053                         goto do_error;
1054         }
1055
1056         if (unlikely(tp->repair)) {
1057                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1058                         copied = tcp_send_rcvq(sk, msg, size);
1059                         goto out;
1060                 }
1061
1062                 err = -EINVAL;
1063                 if (tp->repair_queue == TCP_NO_QUEUE)
1064                         goto out_err;
1065
1066                 /* 'common' sending to sendq */
1067         }
1068
1069         /* This should be in poll */
1070         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1071
1072         mss_now = tcp_send_mss(sk, &size_goal, flags);
1073
1074         /* Ok commence sending. */
1075         iovlen = msg->msg_iovlen;
1076         iov = msg->msg_iov;
1077         copied = 0;
1078
1079         err = -EPIPE;
1080         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1081                 goto out_err;
1082
1083         sg = !!(sk->sk_route_caps & NETIF_F_SG);
1084
1085         while (--iovlen >= 0) {
1086                 size_t seglen = iov->iov_len;
1087                 unsigned char __user *from = iov->iov_base;
1088
1089                 iov++;
1090                 if (unlikely(offset > 0)) {  /* Skip bytes copied in SYN */
1091                         if (offset >= seglen) {
1092                                 offset -= seglen;
1093                                 continue;
1094                         }
1095                         seglen -= offset;
1096                         from += offset;
1097                         offset = 0;
1098                 }
1099
1100                 while (seglen > 0) {
1101                         int copy = 0;
1102                         int max = size_goal;
1103
1104                         skb = tcp_write_queue_tail(sk);
1105                         if (tcp_send_head(sk)) {
1106                                 if (skb->ip_summed == CHECKSUM_NONE)
1107                                         max = mss_now;
1108                                 copy = max - skb->len;
1109                         }
1110
1111                         if (copy <= 0) {
1112 new_segment:
1113                                 /* Allocate new segment. If the interface is SG,
1114                                  * allocate skb fitting to single page.
1115                                  */
1116                                 if (!sk_stream_memory_free(sk))
1117                                         goto wait_for_sndbuf;
1118
1119                                 skb = sk_stream_alloc_skb(sk,
1120                                                           select_size(sk, sg),
1121                                                           sk->sk_allocation);
1122                                 if (!skb)
1123                                         goto wait_for_memory;
1124
1125                                 /*
1126                                  * Check whether we can use HW checksum.
1127                                  */
1128                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1129                                         skb->ip_summed = CHECKSUM_PARTIAL;
1130
1131                                 skb_entail(sk, skb);
1132                                 copy = size_goal;
1133                                 max = size_goal;
1134                         }
1135
1136                         /* Try to append data to the end of skb. */
1137                         if (copy > seglen)
1138                                 copy = seglen;
1139
1140                         /* Where to copy to? */
1141                         if (skb_availroom(skb) > 0) {
1142                                 /* We have some space in skb head. Superb! */
1143                                 copy = min_t(int, copy, skb_availroom(skb));
1144                                 err = skb_add_data_nocache(sk, skb, from, copy);
1145                                 if (err)
1146                                         goto do_fault;
1147                         } else {
1148                                 bool merge = true;
1149                                 int i = skb_shinfo(skb)->nr_frags;
1150                                 struct page_frag *pfrag = sk_page_frag(sk);
1151
1152                                 if (!sk_page_frag_refill(sk, pfrag))
1153                                         goto wait_for_memory;
1154
1155                                 if (!skb_can_coalesce(skb, i, pfrag->page,
1156                                                       pfrag->offset)) {
1157                                         if (i == MAX_SKB_FRAGS || !sg) {
1158                                                 tcp_mark_push(tp, skb);
1159                                                 goto new_segment;
1160                                         }
1161                                         merge = false;
1162                                 }
1163
1164                                 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1165
1166                                 if (!sk_wmem_schedule(sk, copy))
1167                                         goto wait_for_memory;
1168
1169                                 err = skb_copy_to_page_nocache(sk, from, skb,
1170                                                                pfrag->page,
1171                                                                pfrag->offset,
1172                                                                copy);
1173                                 if (err)
1174                                         goto do_error;
1175
1176                                 /* Update the skb. */
1177                                 if (merge) {
1178                                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1179                                 } else {
1180                                         skb_fill_page_desc(skb, i, pfrag->page,
1181                                                            pfrag->offset, copy);
1182                                         get_page(pfrag->page);
1183                                 }
1184                                 pfrag->offset += copy;
1185                         }
1186
1187                         if (!copied)
1188                                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1189
1190                         tp->write_seq += copy;
1191                         TCP_SKB_CB(skb)->end_seq += copy;
1192                         skb_shinfo(skb)->gso_segs = 0;
1193
1194                         from += copy;
1195                         copied += copy;
1196                         if ((seglen -= copy) == 0 && iovlen == 0)
1197                                 goto out;
1198
1199                         if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1200                                 continue;
1201
1202                         if (forced_push(tp)) {
1203                                 tcp_mark_push(tp, skb);
1204                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1205                         } else if (skb == tcp_send_head(sk))
1206                                 tcp_push_one(sk, mss_now);
1207                         continue;
1208
1209 wait_for_sndbuf:
1210                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1211 wait_for_memory:
1212                         if (copied)
1213                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1214
1215                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1216                                 goto do_error;
1217
1218                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1219                 }
1220         }
1221
1222 out:
1223         if (copied)
1224                 tcp_push(sk, flags, mss_now, tp->nonagle);
1225         release_sock(sk);
1226         return copied + copied_syn;
1227
1228 do_fault:
1229         if (!skb->len) {
1230                 tcp_unlink_write_queue(skb, sk);
1231                 /* It is the one place in all of TCP, except connection
1232                  * reset, where we can be unlinking the send_head.
1233                  */
1234                 tcp_check_send_head(sk, skb);
1235                 sk_wmem_free_skb(sk, skb);
1236         }
1237
1238 do_error:
1239         if (copied + copied_syn)
1240                 goto out;
1241 out_err:
1242         err = sk_stream_error(sk, flags, err);
1243         release_sock(sk);
1244         return err;
1245 }
1246 EXPORT_SYMBOL(tcp_sendmsg);
1247
1248 /*
1249  *      Handle reading urgent data. BSD has very simple semantics for
1250  *      this, no blocking and very strange errors 8)
1251  */
1252
1253 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1254 {
1255         struct tcp_sock *tp = tcp_sk(sk);
1256
1257         /* No URG data to read. */
1258         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1259             tp->urg_data == TCP_URG_READ)
1260                 return -EINVAL; /* Yes this is right ! */
1261
1262         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1263                 return -ENOTCONN;
1264
1265         if (tp->urg_data & TCP_URG_VALID) {
1266                 int err = 0;
1267                 char c = tp->urg_data;
1268
1269                 if (!(flags & MSG_PEEK))
1270                         tp->urg_data = TCP_URG_READ;
1271
1272                 /* Read urgent data. */
1273                 msg->msg_flags |= MSG_OOB;
1274
1275                 if (len > 0) {
1276                         if (!(flags & MSG_TRUNC))
1277                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1278                         len = 1;
1279                 } else
1280                         msg->msg_flags |= MSG_TRUNC;
1281
1282                 return err ? -EFAULT : len;
1283         }
1284
1285         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1286                 return 0;
1287
1288         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1289          * the available implementations agree in this case:
1290          * this call should never block, independent of the
1291          * blocking state of the socket.
1292          * Mike <pall@rz.uni-karlsruhe.de>
1293          */
1294         return -EAGAIN;
1295 }
1296
1297 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1298 {
1299         struct sk_buff *skb;
1300         int copied = 0, err = 0;
1301
1302         /* XXX -- need to support SO_PEEK_OFF */
1303
1304         skb_queue_walk(&sk->sk_write_queue, skb) {
1305                 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1306                 if (err)
1307                         break;
1308
1309                 copied += skb->len;
1310         }
1311
1312         return err ?: copied;
1313 }
1314
1315 /* Clean up the receive buffer for full frames taken by the user,
1316  * then send an ACK if necessary.  COPIED is the number of bytes
1317  * tcp_recvmsg has given to the user so far, it speeds up the
1318  * calculation of whether or not we must ACK for the sake of
1319  * a window update.
1320  */
1321 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1322 {
1323         struct tcp_sock *tp = tcp_sk(sk);
1324         bool time_to_ack = false;
1325
1326         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1327
1328         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1329              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1330              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1331
1332         if (inet_csk_ack_scheduled(sk)) {
1333                 const struct inet_connection_sock *icsk = inet_csk(sk);
1334                    /* Delayed ACKs frequently hit locked sockets during bulk
1335                     * receive. */
1336                 if (icsk->icsk_ack.blocked ||
1337                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1338                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1339                     /*
1340                      * If this read emptied read buffer, we send ACK, if
1341                      * connection is not bidirectional, user drained
1342                      * receive buffer and there was a small segment
1343                      * in queue.
1344                      */
1345                     (copied > 0 &&
1346                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1347                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1348                        !icsk->icsk_ack.pingpong)) &&
1349                       !atomic_read(&sk->sk_rmem_alloc)))
1350                         time_to_ack = true;
1351         }
1352
1353         /* We send an ACK if we can now advertise a non-zero window
1354          * which has been raised "significantly".
1355          *
1356          * Even if window raised up to infinity, do not send window open ACK
1357          * in states, where we will not receive more. It is useless.
1358          */
1359         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1360                 __u32 rcv_window_now = tcp_receive_window(tp);
1361
1362                 /* Optimize, __tcp_select_window() is not cheap. */
1363                 if (2*rcv_window_now <= tp->window_clamp) {
1364                         __u32 new_window = __tcp_select_window(sk);
1365
1366                         /* Send ACK now, if this read freed lots of space
1367                          * in our buffer. Certainly, new_window is new window.
1368                          * We can advertise it now, if it is not less than current one.
1369                          * "Lots" means "at least twice" here.
1370                          */
1371                         if (new_window && new_window >= 2 * rcv_window_now)
1372                                 time_to_ack = true;
1373                 }
1374         }
1375         if (time_to_ack)
1376                 tcp_send_ack(sk);
1377 }
1378
1379 static void tcp_prequeue_process(struct sock *sk)
1380 {
1381         struct sk_buff *skb;
1382         struct tcp_sock *tp = tcp_sk(sk);
1383
1384         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1385
1386         /* RX process wants to run with disabled BHs, though it is not
1387          * necessary */
1388         local_bh_disable();
1389         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1390                 sk_backlog_rcv(sk, skb);
1391         local_bh_enable();
1392
1393         /* Clear memory counter. */
1394         tp->ucopy.memory = 0;
1395 }
1396
1397 #ifdef CONFIG_NET_DMA
1398 static void tcp_service_net_dma(struct sock *sk, bool wait)
1399 {
1400         dma_cookie_t done, used;
1401         dma_cookie_t last_issued;
1402         struct tcp_sock *tp = tcp_sk(sk);
1403
1404         if (!tp->ucopy.dma_chan)
1405                 return;
1406
1407         last_issued = tp->ucopy.dma_cookie;
1408         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1409
1410         do {
1411                 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1412                                               last_issued, &done,
1413                                               &used) == DMA_SUCCESS) {
1414                         /* Safe to free early-copied skbs now */
1415                         __skb_queue_purge(&sk->sk_async_wait_queue);
1416                         break;
1417                 } else {
1418                         struct sk_buff *skb;
1419                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1420                                (dma_async_is_complete(skb->dma_cookie, done,
1421                                                       used) == DMA_SUCCESS)) {
1422                                 __skb_dequeue(&sk->sk_async_wait_queue);
1423                                 kfree_skb(skb);
1424                         }
1425                 }
1426         } while (wait);
1427 }
1428 #endif
1429
1430 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1431 {
1432         struct sk_buff *skb;
1433         u32 offset;
1434
1435         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1436                 offset = seq - TCP_SKB_CB(skb)->seq;
1437                 if (tcp_hdr(skb)->syn)
1438                         offset--;
1439                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1440                         *off = offset;
1441                         return skb;
1442                 }
1443                 /* This looks weird, but this can happen if TCP collapsing
1444                  * splitted a fat GRO packet, while we released socket lock
1445                  * in skb_splice_bits()
1446                  */
1447                 sk_eat_skb(sk, skb, false);
1448         }
1449         return NULL;
1450 }
1451
1452 /*
1453  * This routine provides an alternative to tcp_recvmsg() for routines
1454  * that would like to handle copying from skbuffs directly in 'sendfile'
1455  * fashion.
1456  * Note:
1457  *      - It is assumed that the socket was locked by the caller.
1458  *      - The routine does not block.
1459  *      - At present, there is no support for reading OOB data
1460  *        or for 'peeking' the socket using this routine
1461  *        (although both would be easy to implement).
1462  */
1463 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1464                   sk_read_actor_t recv_actor)
1465 {
1466         struct sk_buff *skb;
1467         struct tcp_sock *tp = tcp_sk(sk);
1468         u32 seq = tp->copied_seq;
1469         u32 offset;
1470         int copied = 0;
1471
1472         if (sk->sk_state == TCP_LISTEN)
1473                 return -ENOTCONN;
1474         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1475                 if (offset < skb->len) {
1476                         int used;
1477                         size_t len;
1478
1479                         len = skb->len - offset;
1480                         /* Stop reading if we hit a patch of urgent data */
1481                         if (tp->urg_data) {
1482                                 u32 urg_offset = tp->urg_seq - seq;
1483                                 if (urg_offset < len)
1484                                         len = urg_offset;
1485                                 if (!len)
1486                                         break;
1487                         }
1488                         used = recv_actor(desc, skb, offset, len);
1489                         if (used <= 0) {
1490                                 if (!copied)
1491                                         copied = used;
1492                                 break;
1493                         } else if (used <= len) {
1494                                 seq += used;
1495                                 copied += used;
1496                                 offset += used;
1497                         }
1498                         /* If recv_actor drops the lock (e.g. TCP splice
1499                          * receive) the skb pointer might be invalid when
1500                          * getting here: tcp_collapse might have deleted it
1501                          * while aggregating skbs from the socket queue.
1502                          */
1503                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1504                         if (!skb)
1505                                 break;
1506                         /* TCP coalescing might have appended data to the skb.
1507                          * Try to splice more frags
1508                          */
1509                         if (offset + 1 != skb->len)
1510                                 continue;
1511                 }
1512                 if (tcp_hdr(skb)->fin) {
1513                         sk_eat_skb(sk, skb, false);
1514                         ++seq;
1515                         break;
1516                 }
1517                 sk_eat_skb(sk, skb, false);
1518                 if (!desc->count)
1519                         break;
1520                 tp->copied_seq = seq;
1521         }
1522         tp->copied_seq = seq;
1523
1524         tcp_rcv_space_adjust(sk);
1525
1526         /* Clean up data we have read: This will do ACK frames. */
1527         if (copied > 0) {
1528                 tcp_recv_skb(sk, seq, &offset);
1529                 tcp_cleanup_rbuf(sk, copied);
1530         }
1531         return copied;
1532 }
1533 EXPORT_SYMBOL(tcp_read_sock);
1534
1535 /*
1536  *      This routine copies from a sock struct into the user buffer.
1537  *
1538  *      Technical note: in 2.3 we work on _locked_ socket, so that
1539  *      tricks with *seq access order and skb->users are not required.
1540  *      Probably, code can be easily improved even more.
1541  */
1542
1543 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1544                 size_t len, int nonblock, int flags, int *addr_len)
1545 {
1546         struct tcp_sock *tp = tcp_sk(sk);
1547         int copied = 0;
1548         u32 peek_seq;
1549         u32 *seq;
1550         unsigned long used;
1551         int err;
1552         int target;             /* Read at least this many bytes */
1553         long timeo;
1554         struct task_struct *user_recv = NULL;
1555         bool copied_early = false;
1556         struct sk_buff *skb;
1557         u32 urg_hole = 0;
1558
1559         lock_sock(sk);
1560
1561         err = -ENOTCONN;
1562         if (sk->sk_state == TCP_LISTEN)
1563                 goto out;
1564
1565         timeo = sock_rcvtimeo(sk, nonblock);
1566
1567         /* Urgent data needs to be handled specially. */
1568         if (flags & MSG_OOB)
1569                 goto recv_urg;
1570
1571         if (unlikely(tp->repair)) {
1572                 err = -EPERM;
1573                 if (!(flags & MSG_PEEK))
1574                         goto out;
1575
1576                 if (tp->repair_queue == TCP_SEND_QUEUE)
1577                         goto recv_sndq;
1578
1579                 err = -EINVAL;
1580                 if (tp->repair_queue == TCP_NO_QUEUE)
1581                         goto out;
1582
1583                 /* 'common' recv queue MSG_PEEK-ing */
1584         }
1585
1586         seq = &tp->copied_seq;
1587         if (flags & MSG_PEEK) {
1588                 peek_seq = tp->copied_seq;
1589                 seq = &peek_seq;
1590         }
1591
1592         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1593
1594 #ifdef CONFIG_NET_DMA
1595         tp->ucopy.dma_chan = NULL;
1596         preempt_disable();
1597         skb = skb_peek_tail(&sk->sk_receive_queue);
1598         {
1599                 int available = 0;
1600
1601                 if (skb)
1602                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1603                 if ((available < target) &&
1604                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1605                     !sysctl_tcp_low_latency &&
1606                     net_dma_find_channel()) {
1607                         preempt_enable_no_resched();
1608                         tp->ucopy.pinned_list =
1609                                         dma_pin_iovec_pages(msg->msg_iov, len);
1610                 } else {
1611                         preempt_enable_no_resched();
1612                 }
1613         }
1614 #endif
1615
1616         do {
1617                 u32 offset;
1618
1619                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1620                 if (tp->urg_data && tp->urg_seq == *seq) {
1621                         if (copied)
1622                                 break;
1623                         if (signal_pending(current)) {
1624                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1625                                 break;
1626                         }
1627                 }
1628
1629                 /* Next get a buffer. */
1630
1631                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1632                         /* Now that we have two receive queues this
1633                          * shouldn't happen.
1634                          */
1635                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1636                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1637                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1638                                  flags))
1639                                 break;
1640
1641                         offset = *seq - TCP_SKB_CB(skb)->seq;
1642                         if (tcp_hdr(skb)->syn)
1643                                 offset--;
1644                         if (offset < skb->len)
1645                                 goto found_ok_skb;
1646                         if (tcp_hdr(skb)->fin)
1647                                 goto found_fin_ok;
1648                         WARN(!(flags & MSG_PEEK),
1649                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1650                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1651                 }
1652
1653                 /* Well, if we have backlog, try to process it now yet. */
1654
1655                 if (copied >= target && !sk->sk_backlog.tail)
1656                         break;
1657
1658                 if (copied) {
1659                         if (sk->sk_err ||
1660                             sk->sk_state == TCP_CLOSE ||
1661                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1662                             !timeo ||
1663                             signal_pending(current))
1664                                 break;
1665                 } else {
1666                         if (sock_flag(sk, SOCK_DONE))
1667                                 break;
1668
1669                         if (sk->sk_err) {
1670                                 copied = sock_error(sk);
1671                                 break;
1672                         }
1673
1674                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1675                                 break;
1676
1677                         if (sk->sk_state == TCP_CLOSE) {
1678                                 if (!sock_flag(sk, SOCK_DONE)) {
1679                                         /* This occurs when user tries to read
1680                                          * from never connected socket.
1681                                          */
1682                                         copied = -ENOTCONN;
1683                                         break;
1684                                 }
1685                                 break;
1686                         }
1687
1688                         if (!timeo) {
1689                                 copied = -EAGAIN;
1690                                 break;
1691                         }
1692
1693                         if (signal_pending(current)) {
1694                                 copied = sock_intr_errno(timeo);
1695                                 break;
1696                         }
1697                 }
1698
1699                 tcp_cleanup_rbuf(sk, copied);
1700
1701                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1702                         /* Install new reader */
1703                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1704                                 user_recv = current;
1705                                 tp->ucopy.task = user_recv;
1706                                 tp->ucopy.iov = msg->msg_iov;
1707                         }
1708
1709                         tp->ucopy.len = len;
1710
1711                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1712                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1713
1714                         /* Ugly... If prequeue is not empty, we have to
1715                          * process it before releasing socket, otherwise
1716                          * order will be broken at second iteration.
1717                          * More elegant solution is required!!!
1718                          *
1719                          * Look: we have the following (pseudo)queues:
1720                          *
1721                          * 1. packets in flight
1722                          * 2. backlog
1723                          * 3. prequeue
1724                          * 4. receive_queue
1725                          *
1726                          * Each queue can be processed only if the next ones
1727                          * are empty. At this point we have empty receive_queue.
1728                          * But prequeue _can_ be not empty after 2nd iteration,
1729                          * when we jumped to start of loop because backlog
1730                          * processing added something to receive_queue.
1731                          * We cannot release_sock(), because backlog contains
1732                          * packets arrived _after_ prequeued ones.
1733                          *
1734                          * Shortly, algorithm is clear --- to process all
1735                          * the queues in order. We could make it more directly,
1736                          * requeueing packets from backlog to prequeue, if
1737                          * is not empty. It is more elegant, but eats cycles,
1738                          * unfortunately.
1739                          */
1740                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1741                                 goto do_prequeue;
1742
1743                         /* __ Set realtime policy in scheduler __ */
1744                 }
1745
1746 #ifdef CONFIG_NET_DMA
1747                 if (tp->ucopy.dma_chan) {
1748                         if (tp->rcv_wnd == 0 &&
1749                             !skb_queue_empty(&sk->sk_async_wait_queue)) {
1750                                 tcp_service_net_dma(sk, true);
1751                                 tcp_cleanup_rbuf(sk, copied);
1752                         } else
1753                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1754                 }
1755 #endif
1756                 if (copied >= target) {
1757                         /* Do not sleep, just process backlog. */
1758                         release_sock(sk);
1759                         lock_sock(sk);
1760                 } else
1761                         sk_wait_data(sk, &timeo);
1762
1763 #ifdef CONFIG_NET_DMA
1764                 tcp_service_net_dma(sk, false);  /* Don't block */
1765                 tp->ucopy.wakeup = 0;
1766 #endif
1767
1768                 if (user_recv) {
1769                         int chunk;
1770
1771                         /* __ Restore normal policy in scheduler __ */
1772
1773                         if ((chunk = len - tp->ucopy.len) != 0) {
1774                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1775                                 len -= chunk;
1776                                 copied += chunk;
1777                         }
1778
1779                         if (tp->rcv_nxt == tp->copied_seq &&
1780                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1781 do_prequeue:
1782                                 tcp_prequeue_process(sk);
1783
1784                                 if ((chunk = len - tp->ucopy.len) != 0) {
1785                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1786                                         len -= chunk;
1787                                         copied += chunk;
1788                                 }
1789                         }
1790                 }
1791                 if ((flags & MSG_PEEK) &&
1792                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1793                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1794                                             current->comm,
1795                                             task_pid_nr(current));
1796                         peek_seq = tp->copied_seq;
1797                 }
1798                 continue;
1799
1800         found_ok_skb:
1801                 /* Ok so how much can we use? */
1802                 used = skb->len - offset;
1803                 if (len < used)
1804                         used = len;
1805
1806                 /* Do we have urgent data here? */
1807                 if (tp->urg_data) {
1808                         u32 urg_offset = tp->urg_seq - *seq;
1809                         if (urg_offset < used) {
1810                                 if (!urg_offset) {
1811                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1812                                                 ++*seq;
1813                                                 urg_hole++;
1814                                                 offset++;
1815                                                 used--;
1816                                                 if (!used)
1817                                                         goto skip_copy;
1818                                         }
1819                                 } else
1820                                         used = urg_offset;
1821                         }
1822                 }
1823
1824                 if (!(flags & MSG_TRUNC)) {
1825 #ifdef CONFIG_NET_DMA
1826                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1827                                 tp->ucopy.dma_chan = net_dma_find_channel();
1828
1829                         if (tp->ucopy.dma_chan) {
1830                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1831                                         tp->ucopy.dma_chan, skb, offset,
1832                                         msg->msg_iov, used,
1833                                         tp->ucopy.pinned_list);
1834
1835                                 if (tp->ucopy.dma_cookie < 0) {
1836
1837                                         pr_alert("%s: dma_cookie < 0\n",
1838                                                  __func__);
1839
1840                                         /* Exception. Bailout! */
1841                                         if (!copied)
1842                                                 copied = -EFAULT;
1843                                         break;
1844                                 }
1845
1846                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1847
1848                                 if ((offset + used) == skb->len)
1849                                         copied_early = true;
1850
1851                         } else
1852 #endif
1853                         {
1854                                 err = skb_copy_datagram_iovec(skb, offset,
1855                                                 msg->msg_iov, used);
1856                                 if (err) {
1857                                         /* Exception. Bailout! */
1858                                         if (!copied)
1859                                                 copied = -EFAULT;
1860                                         break;
1861                                 }
1862                         }
1863                 }
1864
1865                 *seq += used;
1866                 copied += used;
1867                 len -= used;
1868
1869                 tcp_rcv_space_adjust(sk);
1870
1871 skip_copy:
1872                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1873                         tp->urg_data = 0;
1874                         tcp_fast_path_check(sk);
1875                 }
1876                 if (used + offset < skb->len)
1877                         continue;
1878
1879                 if (tcp_hdr(skb)->fin)
1880                         goto found_fin_ok;
1881                 if (!(flags & MSG_PEEK)) {
1882                         sk_eat_skb(sk, skb, copied_early);
1883                         copied_early = false;
1884                 }
1885                 continue;
1886
1887         found_fin_ok:
1888                 /* Process the FIN. */
1889                 ++*seq;
1890                 if (!(flags & MSG_PEEK)) {
1891                         sk_eat_skb(sk, skb, copied_early);
1892                         copied_early = false;
1893                 }
1894                 break;
1895         } while (len > 0);
1896
1897         if (user_recv) {
1898                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1899                         int chunk;
1900
1901                         tp->ucopy.len = copied > 0 ? len : 0;
1902
1903                         tcp_prequeue_process(sk);
1904
1905                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1906                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1907                                 len -= chunk;
1908                                 copied += chunk;
1909                         }
1910                 }
1911
1912                 tp->ucopy.task = NULL;
1913                 tp->ucopy.len = 0;
1914         }
1915
1916 #ifdef CONFIG_NET_DMA
1917         tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1918         tp->ucopy.dma_chan = NULL;
1919
1920         if (tp->ucopy.pinned_list) {
1921                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1922                 tp->ucopy.pinned_list = NULL;
1923         }
1924 #endif
1925
1926         /* According to UNIX98, msg_name/msg_namelen are ignored
1927          * on connected socket. I was just happy when found this 8) --ANK
1928          */
1929
1930         /* Clean up data we have read: This will do ACK frames. */
1931         tcp_cleanup_rbuf(sk, copied);
1932
1933         release_sock(sk);
1934         return copied;
1935
1936 out:
1937         release_sock(sk);
1938         return err;
1939
1940 recv_urg:
1941         err = tcp_recv_urg(sk, msg, len, flags);
1942         goto out;
1943
1944 recv_sndq:
1945         err = tcp_peek_sndq(sk, msg, len);
1946         goto out;
1947 }
1948 EXPORT_SYMBOL(tcp_recvmsg);
1949
1950 void tcp_set_state(struct sock *sk, int state)
1951 {
1952         int oldstate = sk->sk_state;
1953
1954         switch (state) {
1955         case TCP_ESTABLISHED:
1956                 if (oldstate != TCP_ESTABLISHED)
1957                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1958                 break;
1959
1960         case TCP_CLOSE:
1961                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1962                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1963
1964                 sk->sk_prot->unhash(sk);
1965                 if (inet_csk(sk)->icsk_bind_hash &&
1966                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1967                         inet_put_port(sk);
1968                 /* fall through */
1969         default:
1970                 if (oldstate == TCP_ESTABLISHED)
1971                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1972         }
1973
1974         /* Change state AFTER socket is unhashed to avoid closed
1975          * socket sitting in hash tables.
1976          */
1977         sk->sk_state = state;
1978
1979 #ifdef STATE_TRACE
1980         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1981 #endif
1982 }
1983 EXPORT_SYMBOL_GPL(tcp_set_state);
1984
1985 /*
1986  *      State processing on a close. This implements the state shift for
1987  *      sending our FIN frame. Note that we only send a FIN for some
1988  *      states. A shutdown() may have already sent the FIN, or we may be
1989  *      closed.
1990  */
1991
1992 static const unsigned char new_state[16] = {
1993   /* current state:        new state:      action:      */
1994   /* (Invalid)          */ TCP_CLOSE,
1995   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1996   /* TCP_SYN_SENT       */ TCP_CLOSE,
1997   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1998   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1999   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
2000   /* TCP_TIME_WAIT      */ TCP_CLOSE,
2001   /* TCP_CLOSE          */ TCP_CLOSE,
2002   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
2003   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
2004   /* TCP_LISTEN         */ TCP_CLOSE,
2005   /* TCP_CLOSING        */ TCP_CLOSING,
2006 };
2007
2008 static int tcp_close_state(struct sock *sk)
2009 {
2010         int next = (int)new_state[sk->sk_state];
2011         int ns = next & TCP_STATE_MASK;
2012
2013         tcp_set_state(sk, ns);
2014
2015         return next & TCP_ACTION_FIN;
2016 }
2017
2018 /*
2019  *      Shutdown the sending side of a connection. Much like close except
2020  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2021  */
2022
2023 void tcp_shutdown(struct sock *sk, int how)
2024 {
2025         /*      We need to grab some memory, and put together a FIN,
2026          *      and then put it into the queue to be sent.
2027          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2028          */
2029         if (!(how & SEND_SHUTDOWN))
2030                 return;
2031
2032         /* If we've already sent a FIN, or it's a closed state, skip this. */
2033         if ((1 << sk->sk_state) &
2034             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2035              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2036                 /* Clear out any half completed packets.  FIN if needed. */
2037                 if (tcp_close_state(sk))
2038                         tcp_send_fin(sk);
2039         }
2040 }
2041 EXPORT_SYMBOL(tcp_shutdown);
2042
2043 bool tcp_check_oom(struct sock *sk, int shift)
2044 {
2045         bool too_many_orphans, out_of_socket_memory;
2046
2047         too_many_orphans = tcp_too_many_orphans(sk, shift);
2048         out_of_socket_memory = tcp_out_of_memory(sk);
2049
2050         if (too_many_orphans)
2051                 net_info_ratelimited("too many orphaned sockets\n");
2052         if (out_of_socket_memory)
2053                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2054         return too_many_orphans || out_of_socket_memory;
2055 }
2056
2057 void tcp_close(struct sock *sk, long timeout)
2058 {
2059         struct sk_buff *skb;
2060         int data_was_unread = 0;
2061         int state;
2062
2063         lock_sock(sk);
2064         sk->sk_shutdown = SHUTDOWN_MASK;
2065
2066         if (sk->sk_state == TCP_LISTEN) {
2067                 tcp_set_state(sk, TCP_CLOSE);
2068
2069                 /* Special case. */
2070                 inet_csk_listen_stop(sk);
2071
2072                 goto adjudge_to_death;
2073         }
2074
2075         /*  We need to flush the recv. buffs.  We do this only on the
2076          *  descriptor close, not protocol-sourced closes, because the
2077          *  reader process may not have drained the data yet!
2078          */
2079         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2080                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2081                           tcp_hdr(skb)->fin;
2082                 data_was_unread += len;
2083                 __kfree_skb(skb);
2084         }
2085
2086         sk_mem_reclaim(sk);
2087
2088         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2089         if (sk->sk_state == TCP_CLOSE)
2090                 goto adjudge_to_death;
2091
2092         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2093          * data was lost. To witness the awful effects of the old behavior of
2094          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2095          * GET in an FTP client, suspend the process, wait for the client to
2096          * advertise a zero window, then kill -9 the FTP client, wheee...
2097          * Note: timeout is always zero in such a case.
2098          */
2099         if (unlikely(tcp_sk(sk)->repair)) {
2100                 sk->sk_prot->disconnect(sk, 0);
2101         } else if (data_was_unread) {
2102                 /* Unread data was tossed, zap the connection. */
2103                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2104                 tcp_set_state(sk, TCP_CLOSE);
2105                 tcp_send_active_reset(sk, sk->sk_allocation);
2106         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2107                 /* Check zero linger _after_ checking for unread data. */
2108                 sk->sk_prot->disconnect(sk, 0);
2109                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2110         } else if (tcp_close_state(sk)) {
2111                 /* We FIN if the application ate all the data before
2112                  * zapping the connection.
2113                  */
2114
2115                 /* RED-PEN. Formally speaking, we have broken TCP state
2116                  * machine. State transitions:
2117                  *
2118                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2119                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2120                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2121                  *
2122                  * are legal only when FIN has been sent (i.e. in window),
2123                  * rather than queued out of window. Purists blame.
2124                  *
2125                  * F.e. "RFC state" is ESTABLISHED,
2126                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2127                  *
2128                  * The visible declinations are that sometimes
2129                  * we enter time-wait state, when it is not required really
2130                  * (harmless), do not send active resets, when they are
2131                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2132                  * they look as CLOSING or LAST_ACK for Linux)
2133                  * Probably, I missed some more holelets.
2134                  *                                              --ANK
2135                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2136                  * in a single packet! (May consider it later but will
2137                  * probably need API support or TCP_CORK SYN-ACK until
2138                  * data is written and socket is closed.)
2139                  */
2140                 tcp_send_fin(sk);
2141         }
2142
2143         sk_stream_wait_close(sk, timeout);
2144
2145 adjudge_to_death:
2146         state = sk->sk_state;
2147         sock_hold(sk);
2148         sock_orphan(sk);
2149
2150         /* It is the last release_sock in its life. It will remove backlog. */
2151         release_sock(sk);
2152
2153
2154         /* Now socket is owned by kernel and we acquire BH lock
2155            to finish close. No need to check for user refs.
2156          */
2157         local_bh_disable();
2158         bh_lock_sock(sk);
2159         WARN_ON(sock_owned_by_user(sk));
2160
2161         percpu_counter_inc(sk->sk_prot->orphan_count);
2162
2163         /* Have we already been destroyed by a softirq or backlog? */
2164         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2165                 goto out;
2166
2167         /*      This is a (useful) BSD violating of the RFC. There is a
2168          *      problem with TCP as specified in that the other end could
2169          *      keep a socket open forever with no application left this end.
2170          *      We use a 3 minute timeout (about the same as BSD) then kill
2171          *      our end. If they send after that then tough - BUT: long enough
2172          *      that we won't make the old 4*rto = almost no time - whoops
2173          *      reset mistake.
2174          *
2175          *      Nope, it was not mistake. It is really desired behaviour
2176          *      f.e. on http servers, when such sockets are useless, but
2177          *      consume significant resources. Let's do it with special
2178          *      linger2 option.                                 --ANK
2179          */
2180
2181         if (sk->sk_state == TCP_FIN_WAIT2) {
2182                 struct tcp_sock *tp = tcp_sk(sk);
2183                 if (tp->linger2 < 0) {
2184                         tcp_set_state(sk, TCP_CLOSE);
2185                         tcp_send_active_reset(sk, GFP_ATOMIC);
2186                         NET_INC_STATS_BH(sock_net(sk),
2187                                         LINUX_MIB_TCPABORTONLINGER);
2188                 } else {
2189                         const int tmo = tcp_fin_time(sk);
2190
2191                         if (tmo > TCP_TIMEWAIT_LEN) {
2192                                 inet_csk_reset_keepalive_timer(sk,
2193                                                 tmo - TCP_TIMEWAIT_LEN);
2194                         } else {
2195                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2196                                 goto out;
2197                         }
2198                 }
2199         }
2200         if (sk->sk_state != TCP_CLOSE) {
2201                 sk_mem_reclaim(sk);
2202                 if (tcp_check_oom(sk, 0)) {
2203                         tcp_set_state(sk, TCP_CLOSE);
2204                         tcp_send_active_reset(sk, GFP_ATOMIC);
2205                         NET_INC_STATS_BH(sock_net(sk),
2206                                         LINUX_MIB_TCPABORTONMEMORY);
2207                 }
2208         }
2209
2210         if (sk->sk_state == TCP_CLOSE) {
2211                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2212                 /* We could get here with a non-NULL req if the socket is
2213                  * aborted (e.g., closed with unread data) before 3WHS
2214                  * finishes.
2215                  */
2216                 if (req != NULL)
2217                         reqsk_fastopen_remove(sk, req, false);
2218                 inet_csk_destroy_sock(sk);
2219         }
2220         /* Otherwise, socket is reprieved until protocol close. */
2221
2222 out:
2223         bh_unlock_sock(sk);
2224         local_bh_enable();
2225         sock_put(sk);
2226 }
2227 EXPORT_SYMBOL(tcp_close);
2228
2229 /* These states need RST on ABORT according to RFC793 */
2230
2231 static inline bool tcp_need_reset(int state)
2232 {
2233         return (1 << state) &
2234                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2235                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2236 }
2237
2238 int tcp_disconnect(struct sock *sk, int flags)
2239 {
2240         struct inet_sock *inet = inet_sk(sk);
2241         struct inet_connection_sock *icsk = inet_csk(sk);
2242         struct tcp_sock *tp = tcp_sk(sk);
2243         int err = 0;
2244         int old_state = sk->sk_state;
2245
2246         if (old_state != TCP_CLOSE)
2247                 tcp_set_state(sk, TCP_CLOSE);
2248
2249         /* ABORT function of RFC793 */
2250         if (old_state == TCP_LISTEN) {
2251                 inet_csk_listen_stop(sk);
2252         } else if (unlikely(tp->repair)) {
2253                 sk->sk_err = ECONNABORTED;
2254         } else if (tcp_need_reset(old_state) ||
2255                    (tp->snd_nxt != tp->write_seq &&
2256                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2257                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2258                  * states
2259                  */
2260                 tcp_send_active_reset(sk, gfp_any());
2261                 sk->sk_err = ECONNRESET;
2262         } else if (old_state == TCP_SYN_SENT)
2263                 sk->sk_err = ECONNRESET;
2264
2265         tcp_clear_xmit_timers(sk);
2266         __skb_queue_purge(&sk->sk_receive_queue);
2267         tcp_write_queue_purge(sk);
2268         __skb_queue_purge(&tp->out_of_order_queue);
2269 #ifdef CONFIG_NET_DMA
2270         __skb_queue_purge(&sk->sk_async_wait_queue);
2271 #endif
2272
2273         inet->inet_dport = 0;
2274
2275         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2276                 inet_reset_saddr(sk);
2277
2278         sk->sk_shutdown = 0;
2279         sock_reset_flag(sk, SOCK_DONE);
2280         tp->srtt = 0;
2281         if ((tp->write_seq += tp->max_window + 2) == 0)
2282                 tp->write_seq = 1;
2283         icsk->icsk_backoff = 0;
2284         tp->snd_cwnd = 2;
2285         icsk->icsk_probes_out = 0;
2286         tp->packets_out = 0;
2287         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2288         tp->snd_cwnd_cnt = 0;
2289         tp->bytes_acked = 0;
2290         tp->window_clamp = 0;
2291         tcp_set_ca_state(sk, TCP_CA_Open);
2292         tcp_clear_retrans(tp);
2293         inet_csk_delack_init(sk);
2294         tcp_init_send_head(sk);
2295         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2296         __sk_dst_reset(sk);
2297
2298         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2299
2300         sk->sk_error_report(sk);
2301         return err;
2302 }
2303 EXPORT_SYMBOL(tcp_disconnect);
2304
2305 void tcp_sock_destruct(struct sock *sk)
2306 {
2307         inet_sock_destruct(sk);
2308
2309         kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2310 }
2311
2312 static inline bool tcp_can_repair_sock(const struct sock *sk)
2313 {
2314         return capable(CAP_NET_ADMIN) &&
2315                 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2316 }
2317
2318 static int tcp_repair_options_est(struct tcp_sock *tp,
2319                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2320 {
2321         struct tcp_repair_opt opt;
2322
2323         while (len >= sizeof(opt)) {
2324                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2325                         return -EFAULT;
2326
2327                 optbuf++;
2328                 len -= sizeof(opt);
2329
2330                 switch (opt.opt_code) {
2331                 case TCPOPT_MSS:
2332                         tp->rx_opt.mss_clamp = opt.opt_val;
2333                         break;
2334                 case TCPOPT_WINDOW:
2335                         {
2336                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2337                                 u16 rcv_wscale = opt.opt_val >> 16;
2338
2339                                 if (snd_wscale > 14 || rcv_wscale > 14)
2340                                         return -EFBIG;
2341
2342                                 tp->rx_opt.snd_wscale = snd_wscale;
2343                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2344                                 tp->rx_opt.wscale_ok = 1;
2345                         }
2346                         break;
2347                 case TCPOPT_SACK_PERM:
2348                         if (opt.opt_val != 0)
2349                                 return -EINVAL;
2350
2351                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2352                         if (sysctl_tcp_fack)
2353                                 tcp_enable_fack(tp);
2354                         break;
2355                 case TCPOPT_TIMESTAMP:
2356                         if (opt.opt_val != 0)
2357                                 return -EINVAL;
2358
2359                         tp->rx_opt.tstamp_ok = 1;
2360                         break;
2361                 }
2362         }
2363
2364         return 0;
2365 }
2366
2367 /*
2368  *      Socket option code for TCP.
2369  */
2370 static int do_tcp_setsockopt(struct sock *sk, int level,
2371                 int optname, char __user *optval, unsigned int optlen)
2372 {
2373         struct tcp_sock *tp = tcp_sk(sk);
2374         struct inet_connection_sock *icsk = inet_csk(sk);
2375         int val;
2376         int err = 0;
2377
2378         /* These are data/string values, all the others are ints */
2379         switch (optname) {
2380         case TCP_CONGESTION: {
2381                 char name[TCP_CA_NAME_MAX];
2382
2383                 if (optlen < 1)
2384                         return -EINVAL;
2385
2386                 val = strncpy_from_user(name, optval,
2387                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2388                 if (val < 0)
2389                         return -EFAULT;
2390                 name[val] = 0;
2391
2392                 lock_sock(sk);
2393                 err = tcp_set_congestion_control(sk, name);
2394                 release_sock(sk);
2395                 return err;
2396         }
2397         case TCP_COOKIE_TRANSACTIONS: {
2398                 struct tcp_cookie_transactions ctd;
2399                 struct tcp_cookie_values *cvp = NULL;
2400
2401                 if (sizeof(ctd) > optlen)
2402                         return -EINVAL;
2403                 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2404                         return -EFAULT;
2405
2406                 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2407                     ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2408                         return -EINVAL;
2409
2410                 if (ctd.tcpct_cookie_desired == 0) {
2411                         /* default to global value */
2412                 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2413                            ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2414                            ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2415                         return -EINVAL;
2416                 }
2417
2418                 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2419                         /* Supercedes all other values */
2420                         lock_sock(sk);
2421                         if (tp->cookie_values != NULL) {
2422                                 kref_put(&tp->cookie_values->kref,
2423                                          tcp_cookie_values_release);
2424                                 tp->cookie_values = NULL;
2425                         }
2426                         tp->rx_opt.cookie_in_always = 0; /* false */
2427                         tp->rx_opt.cookie_out_never = 1; /* true */
2428                         release_sock(sk);
2429                         return err;
2430                 }
2431
2432                 /* Allocate ancillary memory before locking.
2433                  */
2434                 if (ctd.tcpct_used > 0 ||
2435                     (tp->cookie_values == NULL &&
2436                      (sysctl_tcp_cookie_size > 0 ||
2437                       ctd.tcpct_cookie_desired > 0 ||
2438                       ctd.tcpct_s_data_desired > 0))) {
2439                         cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2440                                       GFP_KERNEL);
2441                         if (cvp == NULL)
2442                                 return -ENOMEM;
2443
2444                         kref_init(&cvp->kref);
2445                 }
2446                 lock_sock(sk);
2447                 tp->rx_opt.cookie_in_always =
2448                         (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2449                 tp->rx_opt.cookie_out_never = 0; /* false */
2450
2451                 if (tp->cookie_values != NULL) {
2452                         if (cvp != NULL) {
2453                                 /* Changed values are recorded by a changed
2454                                  * pointer, ensuring the cookie will differ,
2455                                  * without separately hashing each value later.
2456                                  */
2457                                 kref_put(&tp->cookie_values->kref,
2458                                          tcp_cookie_values_release);
2459                         } else {
2460                                 cvp = tp->cookie_values;
2461                         }
2462                 }
2463
2464                 if (cvp != NULL) {
2465                         cvp->cookie_desired = ctd.tcpct_cookie_desired;
2466
2467                         if (ctd.tcpct_used > 0) {
2468                                 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2469                                        ctd.tcpct_used);
2470                                 cvp->s_data_desired = ctd.tcpct_used;
2471                                 cvp->s_data_constant = 1; /* true */
2472                         } else {
2473                                 /* No constant payload data. */
2474                                 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2475                                 cvp->s_data_constant = 0; /* false */
2476                         }
2477
2478                         tp->cookie_values = cvp;
2479                 }
2480                 release_sock(sk);
2481                 return err;
2482         }
2483         default:
2484                 /* fallthru */
2485                 break;
2486         }
2487
2488         if (optlen < sizeof(int))
2489                 return -EINVAL;
2490
2491         if (get_user(val, (int __user *)optval))
2492                 return -EFAULT;
2493
2494         lock_sock(sk);
2495
2496         switch (optname) {
2497         case TCP_MAXSEG:
2498                 /* Values greater than interface MTU won't take effect. However
2499                  * at the point when this call is done we typically don't yet
2500                  * know which interface is going to be used */
2501                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2502                         err = -EINVAL;
2503                         break;
2504                 }
2505                 tp->rx_opt.user_mss = val;
2506                 break;
2507
2508         case TCP_NODELAY:
2509                 if (val) {
2510                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2511                          * this option on corked socket is remembered, but
2512                          * it is not activated until cork is cleared.
2513                          *
2514                          * However, when TCP_NODELAY is set we make
2515                          * an explicit push, which overrides even TCP_CORK
2516                          * for currently queued segments.
2517                          */
2518                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2519                         tcp_push_pending_frames(sk);
2520                 } else {
2521                         tp->nonagle &= ~TCP_NAGLE_OFF;
2522                 }
2523                 break;
2524
2525         case TCP_THIN_LINEAR_TIMEOUTS:
2526                 if (val < 0 || val > 1)
2527                         err = -EINVAL;
2528                 else
2529                         tp->thin_lto = val;
2530                 break;
2531
2532         case TCP_THIN_DUPACK:
2533                 if (val < 0 || val > 1)
2534                         err = -EINVAL;
2535                 else
2536                         tp->thin_dupack = val;
2537                         if (tp->thin_dupack)
2538                                 tcp_disable_early_retrans(tp);
2539                 break;
2540
2541         case TCP_REPAIR:
2542                 if (!tcp_can_repair_sock(sk))
2543                         err = -EPERM;
2544                 else if (val == 1) {
2545                         tp->repair = 1;
2546                         sk->sk_reuse = SK_FORCE_REUSE;
2547                         tp->repair_queue = TCP_NO_QUEUE;
2548                 } else if (val == 0) {
2549                         tp->repair = 0;
2550                         sk->sk_reuse = SK_NO_REUSE;
2551                         tcp_send_window_probe(sk);
2552                 } else
2553                         err = -EINVAL;
2554
2555                 break;
2556
2557         case TCP_REPAIR_QUEUE:
2558                 if (!tp->repair)
2559                         err = -EPERM;
2560                 else if (val < TCP_QUEUES_NR)
2561                         tp->repair_queue = val;
2562                 else
2563                         err = -EINVAL;
2564                 break;
2565
2566         case TCP_QUEUE_SEQ:
2567                 if (sk->sk_state != TCP_CLOSE)
2568                         err = -EPERM;
2569                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2570                         tp->write_seq = val;
2571                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2572                         tp->rcv_nxt = val;
2573                 else
2574                         err = -EINVAL;
2575                 break;
2576
2577         case TCP_REPAIR_OPTIONS:
2578                 if (!tp->repair)
2579                         err = -EINVAL;
2580                 else if (sk->sk_state == TCP_ESTABLISHED)
2581                         err = tcp_repair_options_est(tp,
2582                                         (struct tcp_repair_opt __user *)optval,
2583                                         optlen);
2584                 else
2585                         err = -EPERM;
2586                 break;
2587
2588         case TCP_CORK:
2589                 /* When set indicates to always queue non-full frames.
2590                  * Later the user clears this option and we transmit
2591                  * any pending partial frames in the queue.  This is
2592                  * meant to be used alongside sendfile() to get properly
2593                  * filled frames when the user (for example) must write
2594                  * out headers with a write() call first and then use
2595                  * sendfile to send out the data parts.
2596                  *
2597                  * TCP_CORK can be set together with TCP_NODELAY and it is
2598                  * stronger than TCP_NODELAY.
2599                  */
2600                 if (val) {
2601                         tp->nonagle |= TCP_NAGLE_CORK;
2602                 } else {
2603                         tp->nonagle &= ~TCP_NAGLE_CORK;
2604                         if (tp->nonagle&TCP_NAGLE_OFF)
2605                                 tp->nonagle |= TCP_NAGLE_PUSH;
2606                         tcp_push_pending_frames(sk);
2607                 }
2608                 break;
2609
2610         case TCP_KEEPIDLE:
2611                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2612                         err = -EINVAL;
2613                 else {
2614                         tp->keepalive_time = val * HZ;
2615                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2616                             !((1 << sk->sk_state) &
2617                               (TCPF_CLOSE | TCPF_LISTEN))) {
2618                                 u32 elapsed = keepalive_time_elapsed(tp);
2619                 &