tcp: avoid order-1 allocations on wifi and tx path
[linux-3.10.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/tcp.h>
274 #include <net/xfrm.h>
275 #include <net/ip.h>
276 #include <net/netdma.h>
277 #include <net/sock.h>
278
279 #include <asm/uaccess.h>
280 #include <asm/ioctls.h>
281
282 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
283
284 struct percpu_counter tcp_orphan_count;
285 EXPORT_SYMBOL_GPL(tcp_orphan_count);
286
287 int sysctl_tcp_wmem[3] __read_mostly;
288 int sysctl_tcp_rmem[3] __read_mostly;
289
290 EXPORT_SYMBOL(sysctl_tcp_rmem);
291 EXPORT_SYMBOL(sysctl_tcp_wmem);
292
293 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
294 EXPORT_SYMBOL(tcp_memory_allocated);
295
296 /*
297  * Current number of TCP sockets.
298  */
299 struct percpu_counter tcp_sockets_allocated;
300 EXPORT_SYMBOL(tcp_sockets_allocated);
301
302 /*
303  * TCP splice context
304  */
305 struct tcp_splice_state {
306         struct pipe_inode_info *pipe;
307         size_t len;
308         unsigned int flags;
309 };
310
311 /*
312  * Pressure flag: try to collapse.
313  * Technical note: it is used by multiple contexts non atomically.
314  * All the __sk_mem_schedule() is of this nature: accounting
315  * is strict, actions are advisory and have some latency.
316  */
317 int tcp_memory_pressure __read_mostly;
318 EXPORT_SYMBOL(tcp_memory_pressure);
319
320 void tcp_enter_memory_pressure(struct sock *sk)
321 {
322         if (!tcp_memory_pressure) {
323                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
324                 tcp_memory_pressure = 1;
325         }
326 }
327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
328
329 /* Convert seconds to retransmits based on initial and max timeout */
330 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
331 {
332         u8 res = 0;
333
334         if (seconds > 0) {
335                 int period = timeout;
336
337                 res = 1;
338                 while (seconds > period && res < 255) {
339                         res++;
340                         timeout <<= 1;
341                         if (timeout > rto_max)
342                                 timeout = rto_max;
343                         period += timeout;
344                 }
345         }
346         return res;
347 }
348
349 /* Convert retransmits to seconds based on initial and max timeout */
350 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
351 {
352         int period = 0;
353
354         if (retrans > 0) {
355                 period = timeout;
356                 while (--retrans) {
357                         timeout <<= 1;
358                         if (timeout > rto_max)
359                                 timeout = rto_max;
360                         period += timeout;
361                 }
362         }
363         return period;
364 }
365
366 /*
367  *      Wait for a TCP event.
368  *
369  *      Note that we don't need to lock the socket, as the upper poll layers
370  *      take care of normal races (between the test and the event) and we don't
371  *      go look at any of the socket buffers directly.
372  */
373 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
374 {
375         unsigned int mask;
376         struct sock *sk = sock->sk;
377         const struct tcp_sock *tp = tcp_sk(sk);
378
379         sock_poll_wait(file, sk_sleep(sk), wait);
380         if (sk->sk_state == TCP_LISTEN)
381                 return inet_csk_listen_poll(sk);
382
383         /* Socket is not locked. We are protected from async events
384          * by poll logic and correct handling of state changes
385          * made by other threads is impossible in any case.
386          */
387
388         mask = 0;
389
390         /*
391          * POLLHUP is certainly not done right. But poll() doesn't
392          * have a notion of HUP in just one direction, and for a
393          * socket the read side is more interesting.
394          *
395          * Some poll() documentation says that POLLHUP is incompatible
396          * with the POLLOUT/POLLWR flags, so somebody should check this
397          * all. But careful, it tends to be safer to return too many
398          * bits than too few, and you can easily break real applications
399          * if you don't tell them that something has hung up!
400          *
401          * Check-me.
402          *
403          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
404          * our fs/select.c). It means that after we received EOF,
405          * poll always returns immediately, making impossible poll() on write()
406          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
407          * if and only if shutdown has been made in both directions.
408          * Actually, it is interesting to look how Solaris and DUX
409          * solve this dilemma. I would prefer, if POLLHUP were maskable,
410          * then we could set it on SND_SHUTDOWN. BTW examples given
411          * in Stevens' books assume exactly this behaviour, it explains
412          * why POLLHUP is incompatible with POLLOUT.    --ANK
413          *
414          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
415          * blocking on fresh not-connected or disconnected socket. --ANK
416          */
417         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
418                 mask |= POLLHUP;
419         if (sk->sk_shutdown & RCV_SHUTDOWN)
420                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
421
422         /* Connected? */
423         if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
424                 int target = sock_rcvlowat(sk, 0, INT_MAX);
425
426                 if (tp->urg_seq == tp->copied_seq &&
427                     !sock_flag(sk, SOCK_URGINLINE) &&
428                     tp->urg_data)
429                         target++;
430
431                 /* Potential race condition. If read of tp below will
432                  * escape above sk->sk_state, we can be illegally awaken
433                  * in SYN_* states. */
434                 if (tp->rcv_nxt - tp->copied_seq >= target)
435                         mask |= POLLIN | POLLRDNORM;
436
437                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
438                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
439                                 mask |= POLLOUT | POLLWRNORM;
440                         } else {  /* send SIGIO later */
441                                 set_bit(SOCK_ASYNC_NOSPACE,
442                                         &sk->sk_socket->flags);
443                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
444
445                                 /* Race breaker. If space is freed after
446                                  * wspace test but before the flags are set,
447                                  * IO signal will be lost.
448                                  */
449                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
450                                         mask |= POLLOUT | POLLWRNORM;
451                         }
452                 } else
453                         mask |= POLLOUT | POLLWRNORM;
454
455                 if (tp->urg_data & TCP_URG_VALID)
456                         mask |= POLLPRI;
457         }
458         /* This barrier is coupled with smp_wmb() in tcp_reset() */
459         smp_rmb();
460         if (sk->sk_err)
461                 mask |= POLLERR;
462
463         return mask;
464 }
465 EXPORT_SYMBOL(tcp_poll);
466
467 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
468 {
469         struct tcp_sock *tp = tcp_sk(sk);
470         int answ;
471
472         switch (cmd) {
473         case SIOCINQ:
474                 if (sk->sk_state == TCP_LISTEN)
475                         return -EINVAL;
476
477                 lock_sock(sk);
478                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
479                         answ = 0;
480                 else if (sock_flag(sk, SOCK_URGINLINE) ||
481                          !tp->urg_data ||
482                          before(tp->urg_seq, tp->copied_seq) ||
483                          !before(tp->urg_seq, tp->rcv_nxt)) {
484                         struct sk_buff *skb;
485
486                         answ = tp->rcv_nxt - tp->copied_seq;
487
488                         /* Subtract 1, if FIN is in queue. */
489                         skb = skb_peek_tail(&sk->sk_receive_queue);
490                         if (answ && skb)
491                                 answ -= tcp_hdr(skb)->fin;
492                 } else
493                         answ = tp->urg_seq - tp->copied_seq;
494                 release_sock(sk);
495                 break;
496         case SIOCATMARK:
497                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
498                 break;
499         case SIOCOUTQ:
500                 if (sk->sk_state == TCP_LISTEN)
501                         return -EINVAL;
502
503                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
504                         answ = 0;
505                 else
506                         answ = tp->write_seq - tp->snd_una;
507                 break;
508         case SIOCOUTQNSD:
509                 if (sk->sk_state == TCP_LISTEN)
510                         return -EINVAL;
511
512                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
513                         answ = 0;
514                 else
515                         answ = tp->write_seq - tp->snd_nxt;
516                 break;
517         default:
518                 return -ENOIOCTLCMD;
519         }
520
521         return put_user(answ, (int __user *)arg);
522 }
523 EXPORT_SYMBOL(tcp_ioctl);
524
525 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
526 {
527         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
528         tp->pushed_seq = tp->write_seq;
529 }
530
531 static inline int forced_push(const struct tcp_sock *tp)
532 {
533         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
534 }
535
536 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
537 {
538         struct tcp_sock *tp = tcp_sk(sk);
539         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
540
541         skb->csum    = 0;
542         tcb->seq     = tcb->end_seq = tp->write_seq;
543         tcb->tcp_flags = TCPHDR_ACK;
544         tcb->sacked  = 0;
545         skb_header_release(skb);
546         tcp_add_write_queue_tail(sk, skb);
547         sk->sk_wmem_queued += skb->truesize;
548         sk_mem_charge(sk, skb->truesize);
549         if (tp->nonagle & TCP_NAGLE_PUSH)
550                 tp->nonagle &= ~TCP_NAGLE_PUSH;
551 }
552
553 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
554 {
555         if (flags & MSG_OOB)
556                 tp->snd_up = tp->write_seq;
557 }
558
559 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
560                             int nonagle)
561 {
562         if (tcp_send_head(sk)) {
563                 struct tcp_sock *tp = tcp_sk(sk);
564
565                 if (!(flags & MSG_MORE) || forced_push(tp))
566                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
567
568                 tcp_mark_urg(tp, flags);
569                 __tcp_push_pending_frames(sk, mss_now,
570                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
571         }
572 }
573
574 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
575                                 unsigned int offset, size_t len)
576 {
577         struct tcp_splice_state *tss = rd_desc->arg.data;
578         int ret;
579
580         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
581                               tss->flags);
582         if (ret > 0)
583                 rd_desc->count -= ret;
584         return ret;
585 }
586
587 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
588 {
589         /* Store TCP splice context information in read_descriptor_t. */
590         read_descriptor_t rd_desc = {
591                 .arg.data = tss,
592                 .count    = tss->len,
593         };
594
595         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
596 }
597
598 /**
599  *  tcp_splice_read - splice data from TCP socket to a pipe
600  * @sock:       socket to splice from
601  * @ppos:       position (not valid)
602  * @pipe:       pipe to splice to
603  * @len:        number of bytes to splice
604  * @flags:      splice modifier flags
605  *
606  * Description:
607  *    Will read pages from given socket and fill them into a pipe.
608  *
609  **/
610 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
611                         struct pipe_inode_info *pipe, size_t len,
612                         unsigned int flags)
613 {
614         struct sock *sk = sock->sk;
615         struct tcp_splice_state tss = {
616                 .pipe = pipe,
617                 .len = len,
618                 .flags = flags,
619         };
620         long timeo;
621         ssize_t spliced;
622         int ret;
623
624         sock_rps_record_flow(sk);
625         /*
626          * We can't seek on a socket input
627          */
628         if (unlikely(*ppos))
629                 return -ESPIPE;
630
631         ret = spliced = 0;
632
633         lock_sock(sk);
634
635         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
636         while (tss.len) {
637                 ret = __tcp_splice_read(sk, &tss);
638                 if (ret < 0)
639                         break;
640                 else if (!ret) {
641                         if (spliced)
642                                 break;
643                         if (sock_flag(sk, SOCK_DONE))
644                                 break;
645                         if (sk->sk_err) {
646                                 ret = sock_error(sk);
647                                 break;
648                         }
649                         if (sk->sk_shutdown & RCV_SHUTDOWN)
650                                 break;
651                         if (sk->sk_state == TCP_CLOSE) {
652                                 /*
653                                  * This occurs when user tries to read
654                                  * from never connected socket.
655                                  */
656                                 if (!sock_flag(sk, SOCK_DONE))
657                                         ret = -ENOTCONN;
658                                 break;
659                         }
660                         if (!timeo) {
661                                 ret = -EAGAIN;
662                                 break;
663                         }
664                         sk_wait_data(sk, &timeo);
665                         if (signal_pending(current)) {
666                                 ret = sock_intr_errno(timeo);
667                                 break;
668                         }
669                         continue;
670                 }
671                 tss.len -= ret;
672                 spliced += ret;
673
674                 if (!timeo)
675                         break;
676                 release_sock(sk);
677                 lock_sock(sk);
678
679                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
680                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
681                     signal_pending(current))
682                         break;
683         }
684
685         release_sock(sk);
686
687         if (spliced)
688                 return spliced;
689
690         return ret;
691 }
692 EXPORT_SYMBOL(tcp_splice_read);
693
694 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
695 {
696         struct sk_buff *skb;
697
698         /* The TCP header must be at least 32-bit aligned.  */
699         size = ALIGN(size, 4);
700
701         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
702         if (skb) {
703                 if (sk_wmem_schedule(sk, skb->truesize)) {
704                         skb_reserve(skb, sk->sk_prot->max_header);
705                         /*
706                          * Make sure that we have exactly size bytes
707                          * available to the caller, no more, no less.
708                          */
709                         skb->avail_size = size;
710                         return skb;
711                 }
712                 __kfree_skb(skb);
713         } else {
714                 sk->sk_prot->enter_memory_pressure(sk);
715                 sk_stream_moderate_sndbuf(sk);
716         }
717         return NULL;
718 }
719
720 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
721                                        int large_allowed)
722 {
723         struct tcp_sock *tp = tcp_sk(sk);
724         u32 xmit_size_goal, old_size_goal;
725
726         xmit_size_goal = mss_now;
727
728         if (large_allowed && sk_can_gso(sk)) {
729                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
730                                   inet_csk(sk)->icsk_af_ops->net_header_len -
731                                   inet_csk(sk)->icsk_ext_hdr_len -
732                                   tp->tcp_header_len);
733
734                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
735
736                 /* We try hard to avoid divides here */
737                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
738
739                 if (likely(old_size_goal <= xmit_size_goal &&
740                            old_size_goal + mss_now > xmit_size_goal)) {
741                         xmit_size_goal = old_size_goal;
742                 } else {
743                         tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
744                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
745                 }
746         }
747
748         return max(xmit_size_goal, mss_now);
749 }
750
751 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
752 {
753         int mss_now;
754
755         mss_now = tcp_current_mss(sk);
756         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
757
758         return mss_now;
759 }
760
761 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
762                          size_t psize, int flags)
763 {
764         struct tcp_sock *tp = tcp_sk(sk);
765         int mss_now, size_goal;
766         int err;
767         ssize_t copied;
768         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
769
770         /* Wait for a connection to finish. */
771         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
772                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
773                         goto out_err;
774
775         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
776
777         mss_now = tcp_send_mss(sk, &size_goal, flags);
778         copied = 0;
779
780         err = -EPIPE;
781         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
782                 goto out_err;
783
784         while (psize > 0) {
785                 struct sk_buff *skb = tcp_write_queue_tail(sk);
786                 struct page *page = pages[poffset / PAGE_SIZE];
787                 int copy, i, can_coalesce;
788                 int offset = poffset % PAGE_SIZE;
789                 int size = min_t(size_t, psize, PAGE_SIZE - offset);
790
791                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
792 new_segment:
793                         if (!sk_stream_memory_free(sk))
794                                 goto wait_for_sndbuf;
795
796                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
797                         if (!skb)
798                                 goto wait_for_memory;
799
800                         skb_entail(sk, skb);
801                         copy = size_goal;
802                 }
803
804                 if (copy > size)
805                         copy = size;
806
807                 i = skb_shinfo(skb)->nr_frags;
808                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
809                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
810                         tcp_mark_push(tp, skb);
811                         goto new_segment;
812                 }
813                 if (!sk_wmem_schedule(sk, copy))
814                         goto wait_for_memory;
815
816                 if (can_coalesce) {
817                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
818                 } else {
819                         get_page(page);
820                         skb_fill_page_desc(skb, i, page, offset, copy);
821                 }
822
823                 skb->len += copy;
824                 skb->data_len += copy;
825                 skb->truesize += copy;
826                 sk->sk_wmem_queued += copy;
827                 sk_mem_charge(sk, copy);
828                 skb->ip_summed = CHECKSUM_PARTIAL;
829                 tp->write_seq += copy;
830                 TCP_SKB_CB(skb)->end_seq += copy;
831                 skb_shinfo(skb)->gso_segs = 0;
832
833                 if (!copied)
834                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
835
836                 copied += copy;
837                 poffset += copy;
838                 if (!(psize -= copy))
839                         goto out;
840
841                 if (skb->len < size_goal || (flags & MSG_OOB))
842                         continue;
843
844                 if (forced_push(tp)) {
845                         tcp_mark_push(tp, skb);
846                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
847                 } else if (skb == tcp_send_head(sk))
848                         tcp_push_one(sk, mss_now);
849                 continue;
850
851 wait_for_sndbuf:
852                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
853 wait_for_memory:
854                 if (copied)
855                         tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
856
857                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
858                         goto do_error;
859
860                 mss_now = tcp_send_mss(sk, &size_goal, flags);
861         }
862
863 out:
864         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
865                 tcp_push(sk, flags, mss_now, tp->nonagle);
866         return copied;
867
868 do_error:
869         if (copied)
870                 goto out;
871 out_err:
872         return sk_stream_error(sk, flags, err);
873 }
874
875 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
876                  size_t size, int flags)
877 {
878         ssize_t res;
879
880         if (!(sk->sk_route_caps & NETIF_F_SG) ||
881             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
882                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
883                                         flags);
884
885         lock_sock(sk);
886         res = do_tcp_sendpages(sk, &page, offset, size, flags);
887         release_sock(sk);
888         return res;
889 }
890 EXPORT_SYMBOL(tcp_sendpage);
891
892 static inline int select_size(const struct sock *sk, bool sg)
893 {
894         const struct tcp_sock *tp = tcp_sk(sk);
895         int tmp = tp->mss_cache;
896
897         if (sg) {
898                 if (sk_can_gso(sk)) {
899                         /* Small frames wont use a full page:
900                          * Payload will immediately follow tcp header.
901                          */
902                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
903                 } else {
904                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
905
906                         if (tmp >= pgbreak &&
907                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
908                                 tmp = pgbreak;
909                 }
910         }
911
912         return tmp;
913 }
914
915 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
916                 size_t size)
917 {
918         struct iovec *iov;
919         struct tcp_sock *tp = tcp_sk(sk);
920         struct sk_buff *skb;
921         int iovlen, flags, err, copied;
922         int mss_now, size_goal;
923         bool sg;
924         long timeo;
925
926         lock_sock(sk);
927
928         flags = msg->msg_flags;
929         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
930
931         /* Wait for a connection to finish. */
932         if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
933                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
934                         goto out_err;
935
936         /* This should be in poll */
937         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
938
939         mss_now = tcp_send_mss(sk, &size_goal, flags);
940
941         /* Ok commence sending. */
942         iovlen = msg->msg_iovlen;
943         iov = msg->msg_iov;
944         copied = 0;
945
946         err = -EPIPE;
947         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
948                 goto out_err;
949
950         sg = !!(sk->sk_route_caps & NETIF_F_SG);
951
952         while (--iovlen >= 0) {
953                 size_t seglen = iov->iov_len;
954                 unsigned char __user *from = iov->iov_base;
955
956                 iov++;
957
958                 while (seglen > 0) {
959                         int copy = 0;
960                         int max = size_goal;
961
962                         skb = tcp_write_queue_tail(sk);
963                         if (tcp_send_head(sk)) {
964                                 if (skb->ip_summed == CHECKSUM_NONE)
965                                         max = mss_now;
966                                 copy = max - skb->len;
967                         }
968
969                         if (copy <= 0) {
970 new_segment:
971                                 /* Allocate new segment. If the interface is SG,
972                                  * allocate skb fitting to single page.
973                                  */
974                                 if (!sk_stream_memory_free(sk))
975                                         goto wait_for_sndbuf;
976
977                                 skb = sk_stream_alloc_skb(sk,
978                                                           select_size(sk, sg),
979                                                           sk->sk_allocation);
980                                 if (!skb)
981                                         goto wait_for_memory;
982
983                                 /*
984                                  * Check whether we can use HW checksum.
985                                  */
986                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
987                                         skb->ip_summed = CHECKSUM_PARTIAL;
988
989                                 skb_entail(sk, skb);
990                                 copy = size_goal;
991                                 max = size_goal;
992                         }
993
994                         /* Try to append data to the end of skb. */
995                         if (copy > seglen)
996                                 copy = seglen;
997
998                         /* Where to copy to? */
999                         if (skb_availroom(skb) > 0) {
1000                                 /* We have some space in skb head. Superb! */
1001                                 copy = min_t(int, copy, skb_availroom(skb));
1002                                 err = skb_add_data_nocache(sk, skb, from, copy);
1003                                 if (err)
1004                                         goto do_fault;
1005                         } else {
1006                                 int merge = 0;
1007                                 int i = skb_shinfo(skb)->nr_frags;
1008                                 struct page *page = sk->sk_sndmsg_page;
1009                                 int off;
1010
1011                                 if (page && page_count(page) == 1)
1012                                         sk->sk_sndmsg_off = 0;
1013
1014                                 off = sk->sk_sndmsg_off;
1015
1016                                 if (skb_can_coalesce(skb, i, page, off) &&
1017                                     off != PAGE_SIZE) {
1018                                         /* We can extend the last page
1019                                          * fragment. */
1020                                         merge = 1;
1021                                 } else if (i == MAX_SKB_FRAGS || !sg) {
1022                                         /* Need to add new fragment and cannot
1023                                          * do this because interface is non-SG,
1024                                          * or because all the page slots are
1025                                          * busy. */
1026                                         tcp_mark_push(tp, skb);
1027                                         goto new_segment;
1028                                 } else if (page) {
1029                                         if (off == PAGE_SIZE) {
1030                                                 put_page(page);
1031                                                 sk->sk_sndmsg_page = page = NULL;
1032                                                 off = 0;
1033                                         }
1034                                 } else
1035                                         off = 0;
1036
1037                                 if (copy > PAGE_SIZE - off)
1038                                         copy = PAGE_SIZE - off;
1039
1040                                 if (!sk_wmem_schedule(sk, copy))
1041                                         goto wait_for_memory;
1042
1043                                 if (!page) {
1044                                         /* Allocate new cache page. */
1045                                         if (!(page = sk_stream_alloc_page(sk)))
1046                                                 goto wait_for_memory;
1047                                 }
1048
1049                                 /* Time to copy data. We are close to
1050                                  * the end! */
1051                                 err = skb_copy_to_page_nocache(sk, from, skb,
1052                                                                page, off, copy);
1053                                 if (err) {
1054                                         /* If this page was new, give it to the
1055                                          * socket so it does not get leaked.
1056                                          */
1057                                         if (!sk->sk_sndmsg_page) {
1058                                                 sk->sk_sndmsg_page = page;
1059                                                 sk->sk_sndmsg_off = 0;
1060                                         }
1061                                         goto do_error;
1062                                 }
1063
1064                                 /* Update the skb. */
1065                                 if (merge) {
1066                                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1067                                 } else {
1068                                         skb_fill_page_desc(skb, i, page, off, copy);
1069                                         if (sk->sk_sndmsg_page) {
1070                                                 get_page(page);
1071                                         } else if (off + copy < PAGE_SIZE) {
1072                                                 get_page(page);
1073                                                 sk->sk_sndmsg_page = page;
1074                                         }
1075                                 }
1076
1077                                 sk->sk_sndmsg_off = off + copy;
1078                         }
1079
1080                         if (!copied)
1081                                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1082
1083                         tp->write_seq += copy;
1084                         TCP_SKB_CB(skb)->end_seq += copy;
1085                         skb_shinfo(skb)->gso_segs = 0;
1086
1087                         from += copy;
1088                         copied += copy;
1089                         if ((seglen -= copy) == 0 && iovlen == 0)
1090                                 goto out;
1091
1092                         if (skb->len < max || (flags & MSG_OOB))
1093                                 continue;
1094
1095                         if (forced_push(tp)) {
1096                                 tcp_mark_push(tp, skb);
1097                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1098                         } else if (skb == tcp_send_head(sk))
1099                                 tcp_push_one(sk, mss_now);
1100                         continue;
1101
1102 wait_for_sndbuf:
1103                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1104 wait_for_memory:
1105                         if (copied)
1106                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1107
1108                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1109                                 goto do_error;
1110
1111                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1112                 }
1113         }
1114
1115 out:
1116         if (copied)
1117                 tcp_push(sk, flags, mss_now, tp->nonagle);
1118         release_sock(sk);
1119         return copied;
1120
1121 do_fault:
1122         if (!skb->len) {
1123                 tcp_unlink_write_queue(skb, sk);
1124                 /* It is the one place in all of TCP, except connection
1125                  * reset, where we can be unlinking the send_head.
1126                  */
1127                 tcp_check_send_head(sk, skb);
1128                 sk_wmem_free_skb(sk, skb);
1129         }
1130
1131 do_error:
1132         if (copied)
1133                 goto out;
1134 out_err:
1135         err = sk_stream_error(sk, flags, err);
1136         release_sock(sk);
1137         return err;
1138 }
1139 EXPORT_SYMBOL(tcp_sendmsg);
1140
1141 /*
1142  *      Handle reading urgent data. BSD has very simple semantics for
1143  *      this, no blocking and very strange errors 8)
1144  */
1145
1146 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1147 {
1148         struct tcp_sock *tp = tcp_sk(sk);
1149
1150         /* No URG data to read. */
1151         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1152             tp->urg_data == TCP_URG_READ)
1153                 return -EINVAL; /* Yes this is right ! */
1154
1155         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1156                 return -ENOTCONN;
1157
1158         if (tp->urg_data & TCP_URG_VALID) {
1159                 int err = 0;
1160                 char c = tp->urg_data;
1161
1162                 if (!(flags & MSG_PEEK))
1163                         tp->urg_data = TCP_URG_READ;
1164
1165                 /* Read urgent data. */
1166                 msg->msg_flags |= MSG_OOB;
1167
1168                 if (len > 0) {
1169                         if (!(flags & MSG_TRUNC))
1170                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1171                         len = 1;
1172                 } else
1173                         msg->msg_flags |= MSG_TRUNC;
1174
1175                 return err ? -EFAULT : len;
1176         }
1177
1178         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1179                 return 0;
1180
1181         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1182          * the available implementations agree in this case:
1183          * this call should never block, independent of the
1184          * blocking state of the socket.
1185          * Mike <pall@rz.uni-karlsruhe.de>
1186          */
1187         return -EAGAIN;
1188 }
1189
1190 /* Clean up the receive buffer for full frames taken by the user,
1191  * then send an ACK if necessary.  COPIED is the number of bytes
1192  * tcp_recvmsg has given to the user so far, it speeds up the
1193  * calculation of whether or not we must ACK for the sake of
1194  * a window update.
1195  */
1196 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1197 {
1198         struct tcp_sock *tp = tcp_sk(sk);
1199         int time_to_ack = 0;
1200
1201         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1202
1203         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1204              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1205              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1206
1207         if (inet_csk_ack_scheduled(sk)) {
1208                 const struct inet_connection_sock *icsk = inet_csk(sk);
1209                    /* Delayed ACKs frequently hit locked sockets during bulk
1210                     * receive. */
1211                 if (icsk->icsk_ack.blocked ||
1212                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1213                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1214                     /*
1215                      * If this read emptied read buffer, we send ACK, if
1216                      * connection is not bidirectional, user drained
1217                      * receive buffer and there was a small segment
1218                      * in queue.
1219                      */
1220                     (copied > 0 &&
1221                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1222                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1223                        !icsk->icsk_ack.pingpong)) &&
1224                       !atomic_read(&sk->sk_rmem_alloc)))
1225                         time_to_ack = 1;
1226         }
1227
1228         /* We send an ACK if we can now advertise a non-zero window
1229          * which has been raised "significantly".
1230          *
1231          * Even if window raised up to infinity, do not send window open ACK
1232          * in states, where we will not receive more. It is useless.
1233          */
1234         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1235                 __u32 rcv_window_now = tcp_receive_window(tp);
1236
1237                 /* Optimize, __tcp_select_window() is not cheap. */
1238                 if (2*rcv_window_now <= tp->window_clamp) {
1239                         __u32 new_window = __tcp_select_window(sk);
1240
1241                         /* Send ACK now, if this read freed lots of space
1242                          * in our buffer. Certainly, new_window is new window.
1243                          * We can advertise it now, if it is not less than current one.
1244                          * "Lots" means "at least twice" here.
1245                          */
1246                         if (new_window && new_window >= 2 * rcv_window_now)
1247                                 time_to_ack = 1;
1248                 }
1249         }
1250         if (time_to_ack)
1251                 tcp_send_ack(sk);
1252 }
1253
1254 static void tcp_prequeue_process(struct sock *sk)
1255 {
1256         struct sk_buff *skb;
1257         struct tcp_sock *tp = tcp_sk(sk);
1258
1259         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1260
1261         /* RX process wants to run with disabled BHs, though it is not
1262          * necessary */
1263         local_bh_disable();
1264         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1265                 sk_backlog_rcv(sk, skb);
1266         local_bh_enable();
1267
1268         /* Clear memory counter. */
1269         tp->ucopy.memory = 0;
1270 }
1271
1272 #ifdef CONFIG_NET_DMA
1273 static void tcp_service_net_dma(struct sock *sk, bool wait)
1274 {
1275         dma_cookie_t done, used;
1276         dma_cookie_t last_issued;
1277         struct tcp_sock *tp = tcp_sk(sk);
1278
1279         if (!tp->ucopy.dma_chan)
1280                 return;
1281
1282         last_issued = tp->ucopy.dma_cookie;
1283         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1284
1285         do {
1286                 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1287                                               last_issued, &done,
1288                                               &used) == DMA_SUCCESS) {
1289                         /* Safe to free early-copied skbs now */
1290                         __skb_queue_purge(&sk->sk_async_wait_queue);
1291                         break;
1292                 } else {
1293                         struct sk_buff *skb;
1294                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1295                                (dma_async_is_complete(skb->dma_cookie, done,
1296                                                       used) == DMA_SUCCESS)) {
1297                                 __skb_dequeue(&sk->sk_async_wait_queue);
1298                                 kfree_skb(skb);
1299                         }
1300                 }
1301         } while (wait);
1302 }
1303 #endif
1304
1305 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1306 {
1307         struct sk_buff *skb;
1308         u32 offset;
1309
1310         skb_queue_walk(&sk->sk_receive_queue, skb) {
1311                 offset = seq - TCP_SKB_CB(skb)->seq;
1312                 if (tcp_hdr(skb)->syn)
1313                         offset--;
1314                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1315                         *off = offset;
1316                         return skb;
1317                 }
1318         }
1319         return NULL;
1320 }
1321
1322 /*
1323  * This routine provides an alternative to tcp_recvmsg() for routines
1324  * that would like to handle copying from skbuffs directly in 'sendfile'
1325  * fashion.
1326  * Note:
1327  *      - It is assumed that the socket was locked by the caller.
1328  *      - The routine does not block.
1329  *      - At present, there is no support for reading OOB data
1330  *        or for 'peeking' the socket using this routine
1331  *        (although both would be easy to implement).
1332  */
1333 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1334                   sk_read_actor_t recv_actor)
1335 {
1336         struct sk_buff *skb;
1337         struct tcp_sock *tp = tcp_sk(sk);
1338         u32 seq = tp->copied_seq;
1339         u32 offset;
1340         int copied = 0;
1341
1342         if (sk->sk_state == TCP_LISTEN)
1343                 return -ENOTCONN;
1344         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1345                 if (offset < skb->len) {
1346                         int used;
1347                         size_t len;
1348
1349                         len = skb->len - offset;
1350                         /* Stop reading if we hit a patch of urgent data */
1351                         if (tp->urg_data) {
1352                                 u32 urg_offset = tp->urg_seq - seq;
1353                                 if (urg_offset < len)
1354                                         len = urg_offset;
1355                                 if (!len)
1356                                         break;
1357                         }
1358                         used = recv_actor(desc, skb, offset, len);
1359                         if (used < 0) {
1360                                 if (!copied)
1361                                         copied = used;
1362                                 break;
1363                         } else if (used <= len) {
1364                                 seq += used;
1365                                 copied += used;
1366                                 offset += used;
1367                         }
1368                         /*
1369                          * If recv_actor drops the lock (e.g. TCP splice
1370                          * receive) the skb pointer might be invalid when
1371                          * getting here: tcp_collapse might have deleted it
1372                          * while aggregating skbs from the socket queue.
1373                          */
1374                         skb = tcp_recv_skb(sk, seq-1, &offset);
1375                         if (!skb || (offset+1 != skb->len))
1376                                 break;
1377                 }
1378                 if (tcp_hdr(skb)->fin) {
1379                         sk_eat_skb(sk, skb, 0);
1380                         ++seq;
1381                         break;
1382                 }
1383                 sk_eat_skb(sk, skb, 0);
1384                 if (!desc->count)
1385                         break;
1386                 tp->copied_seq = seq;
1387         }
1388         tp->copied_seq = seq;
1389
1390         tcp_rcv_space_adjust(sk);
1391
1392         /* Clean up data we have read: This will do ACK frames. */
1393         if (copied > 0)
1394                 tcp_cleanup_rbuf(sk, copied);
1395         return copied;
1396 }
1397 EXPORT_SYMBOL(tcp_read_sock);
1398
1399 /*
1400  *      This routine copies from a sock struct into the user buffer.
1401  *
1402  *      Technical note: in 2.3 we work on _locked_ socket, so that
1403  *      tricks with *seq access order and skb->users are not required.
1404  *      Probably, code can be easily improved even more.
1405  */
1406
1407 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1408                 size_t len, int nonblock, int flags, int *addr_len)
1409 {
1410         struct tcp_sock *tp = tcp_sk(sk);
1411         int copied = 0;
1412         u32 peek_seq;
1413         u32 *seq;
1414         unsigned long used;
1415         int err;
1416         int target;             /* Read at least this many bytes */
1417         long timeo;
1418         struct task_struct *user_recv = NULL;
1419         int copied_early = 0;
1420         struct sk_buff *skb;
1421         u32 urg_hole = 0;
1422
1423         lock_sock(sk);
1424
1425         err = -ENOTCONN;
1426         if (sk->sk_state == TCP_LISTEN)
1427                 goto out;
1428
1429         timeo = sock_rcvtimeo(sk, nonblock);
1430
1431         /* Urgent data needs to be handled specially. */
1432         if (flags & MSG_OOB)
1433                 goto recv_urg;
1434
1435         seq = &tp->copied_seq;
1436         if (flags & MSG_PEEK) {
1437                 peek_seq = tp->copied_seq;
1438                 seq = &peek_seq;
1439         }
1440
1441         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1442
1443 #ifdef CONFIG_NET_DMA
1444         tp->ucopy.dma_chan = NULL;
1445         preempt_disable();
1446         skb = skb_peek_tail(&sk->sk_receive_queue);
1447         {
1448                 int available = 0;
1449
1450                 if (skb)
1451                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1452                 if ((available < target) &&
1453                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1454                     !sysctl_tcp_low_latency &&
1455                     dma_find_channel(DMA_MEMCPY)) {
1456                         preempt_enable_no_resched();
1457                         tp->ucopy.pinned_list =
1458                                         dma_pin_iovec_pages(msg->msg_iov, len);
1459                 } else {
1460                         preempt_enable_no_resched();
1461                 }
1462         }
1463 #endif
1464
1465         do {
1466                 u32 offset;
1467
1468                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1469                 if (tp->urg_data && tp->urg_seq == *seq) {
1470                         if (copied)
1471                                 break;
1472                         if (signal_pending(current)) {
1473                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1474                                 break;
1475                         }
1476                 }
1477
1478                 /* Next get a buffer. */
1479
1480                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1481                         /* Now that we have two receive queues this
1482                          * shouldn't happen.
1483                          */
1484                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1485                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1486                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1487                                  flags))
1488                                 break;
1489
1490                         offset = *seq - TCP_SKB_CB(skb)->seq;
1491                         if (tcp_hdr(skb)->syn)
1492                                 offset--;
1493                         if (offset < skb->len)
1494                                 goto found_ok_skb;
1495                         if (tcp_hdr(skb)->fin)
1496                                 goto found_fin_ok;
1497                         WARN(!(flags & MSG_PEEK),
1498                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1499                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1500                 }
1501
1502                 /* Well, if we have backlog, try to process it now yet. */
1503
1504                 if (copied >= target && !sk->sk_backlog.tail)
1505                         break;
1506
1507                 if (copied) {
1508                         if (sk->sk_err ||
1509                             sk->sk_state == TCP_CLOSE ||
1510                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1511                             !timeo ||
1512                             signal_pending(current))
1513                                 break;
1514                 } else {
1515                         if (sock_flag(sk, SOCK_DONE))
1516                                 break;
1517
1518                         if (sk->sk_err) {
1519                                 copied = sock_error(sk);
1520                                 break;
1521                         }
1522
1523                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1524                                 break;
1525
1526                         if (sk->sk_state == TCP_CLOSE) {
1527                                 if (!sock_flag(sk, SOCK_DONE)) {
1528                                         /* This occurs when user tries to read
1529                                          * from never connected socket.
1530                                          */
1531                                         copied = -ENOTCONN;
1532                                         break;
1533                                 }
1534                                 break;
1535                         }
1536
1537                         if (!timeo) {
1538                                 copied = -EAGAIN;
1539                                 break;
1540                         }
1541
1542                         if (signal_pending(current)) {
1543                                 copied = sock_intr_errno(timeo);
1544                                 break;
1545                         }
1546                 }
1547
1548                 tcp_cleanup_rbuf(sk, copied);
1549
1550                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1551                         /* Install new reader */
1552                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1553                                 user_recv = current;
1554                                 tp->ucopy.task = user_recv;
1555                                 tp->ucopy.iov = msg->msg_iov;
1556                         }
1557
1558                         tp->ucopy.len = len;
1559
1560                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1561                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1562
1563                         /* Ugly... If prequeue is not empty, we have to
1564                          * process it before releasing socket, otherwise
1565                          * order will be broken at second iteration.
1566                          * More elegant solution is required!!!
1567                          *
1568                          * Look: we have the following (pseudo)queues:
1569                          *
1570                          * 1. packets in flight
1571                          * 2. backlog
1572                          * 3. prequeue
1573                          * 4. receive_queue
1574                          *
1575                          * Each queue can be processed only if the next ones
1576                          * are empty. At this point we have empty receive_queue.
1577                          * But prequeue _can_ be not empty after 2nd iteration,
1578                          * when we jumped to start of loop because backlog
1579                          * processing added something to receive_queue.
1580                          * We cannot release_sock(), because backlog contains
1581                          * packets arrived _after_ prequeued ones.
1582                          *
1583                          * Shortly, algorithm is clear --- to process all
1584                          * the queues in order. We could make it more directly,
1585                          * requeueing packets from backlog to prequeue, if
1586                          * is not empty. It is more elegant, but eats cycles,
1587                          * unfortunately.
1588                          */
1589                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1590                                 goto do_prequeue;
1591
1592                         /* __ Set realtime policy in scheduler __ */
1593                 }
1594
1595 #ifdef CONFIG_NET_DMA
1596                 if (tp->ucopy.dma_chan)
1597                         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1598 #endif
1599                 if (copied >= target) {
1600                         /* Do not sleep, just process backlog. */
1601                         release_sock(sk);
1602                         lock_sock(sk);
1603                 } else
1604                         sk_wait_data(sk, &timeo);
1605
1606 #ifdef CONFIG_NET_DMA
1607                 tcp_service_net_dma(sk, false);  /* Don't block */
1608                 tp->ucopy.wakeup = 0;
1609 #endif
1610
1611                 if (user_recv) {
1612                         int chunk;
1613
1614                         /* __ Restore normal policy in scheduler __ */
1615
1616                         if ((chunk = len - tp->ucopy.len) != 0) {
1617                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1618                                 len -= chunk;
1619                                 copied += chunk;
1620                         }
1621
1622                         if (tp->rcv_nxt == tp->copied_seq &&
1623                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1624 do_prequeue:
1625                                 tcp_prequeue_process(sk);
1626
1627                                 if ((chunk = len - tp->ucopy.len) != 0) {
1628                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1629                                         len -= chunk;
1630                                         copied += chunk;
1631                                 }
1632                         }
1633                 }
1634                 if ((flags & MSG_PEEK) &&
1635                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1636                         if (net_ratelimit())
1637                                 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1638                                        current->comm, task_pid_nr(current));
1639                         peek_seq = tp->copied_seq;
1640                 }
1641                 continue;
1642
1643         found_ok_skb:
1644                 /* Ok so how much can we use? */
1645                 used = skb->len - offset;
1646                 if (len < used)
1647                         used = len;
1648
1649                 /* Do we have urgent data here? */
1650                 if (tp->urg_data) {
1651                         u32 urg_offset = tp->urg_seq - *seq;
1652                         if (urg_offset < used) {
1653                                 if (!urg_offset) {
1654                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1655                                                 ++*seq;
1656                                                 urg_hole++;
1657                                                 offset++;
1658                                                 used--;
1659                                                 if (!used)
1660                                                         goto skip_copy;
1661                                         }
1662                                 } else
1663                                         used = urg_offset;
1664                         }
1665                 }
1666
1667                 if (!(flags & MSG_TRUNC)) {
1668 #ifdef CONFIG_NET_DMA
1669                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1670                                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1671
1672                         if (tp->ucopy.dma_chan) {
1673                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1674                                         tp->ucopy.dma_chan, skb, offset,
1675                                         msg->msg_iov, used,
1676                                         tp->ucopy.pinned_list);
1677
1678                                 if (tp->ucopy.dma_cookie < 0) {
1679
1680                                         pr_alert("%s: dma_cookie < 0\n",
1681                                                  __func__);
1682
1683                                         /* Exception. Bailout! */
1684                                         if (!copied)
1685                                                 copied = -EFAULT;
1686                                         break;
1687                                 }
1688
1689                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1690
1691                                 if ((offset + used) == skb->len)
1692                                         copied_early = 1;
1693
1694                         } else
1695 #endif
1696                         {
1697                                 err = skb_copy_datagram_iovec(skb, offset,
1698                                                 msg->msg_iov, used);
1699                                 if (err) {
1700                                         /* Exception. Bailout! */
1701                                         if (!copied)
1702                                                 copied = -EFAULT;
1703                                         break;
1704                                 }
1705                         }
1706                 }
1707
1708                 *seq += used;
1709                 copied += used;
1710                 len -= used;
1711
1712                 tcp_rcv_space_adjust(sk);
1713
1714 skip_copy:
1715                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1716                         tp->urg_data = 0;
1717                         tcp_fast_path_check(sk);
1718                 }
1719                 if (used + offset < skb->len)
1720                         continue;
1721
1722                 if (tcp_hdr(skb)->fin)
1723                         goto found_fin_ok;
1724                 if (!(flags & MSG_PEEK)) {
1725                         sk_eat_skb(sk, skb, copied_early);
1726                         copied_early = 0;
1727                 }
1728                 continue;
1729
1730         found_fin_ok:
1731                 /* Process the FIN. */
1732                 ++*seq;
1733                 if (!(flags & MSG_PEEK)) {
1734                         sk_eat_skb(sk, skb, copied_early);
1735                         copied_early = 0;
1736                 }
1737                 break;
1738         } while (len > 0);
1739
1740         if (user_recv) {
1741                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1742                         int chunk;
1743
1744                         tp->ucopy.len = copied > 0 ? len : 0;
1745
1746                         tcp_prequeue_process(sk);
1747
1748                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1749                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1750                                 len -= chunk;
1751                                 copied += chunk;
1752                         }
1753                 }
1754
1755                 tp->ucopy.task = NULL;
1756                 tp->ucopy.len = 0;
1757         }
1758
1759 #ifdef CONFIG_NET_DMA
1760         tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1761         tp->ucopy.dma_chan = NULL;
1762
1763         if (tp->ucopy.pinned_list) {
1764                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1765                 tp->ucopy.pinned_list = NULL;
1766         }
1767 #endif
1768
1769         /* According to UNIX98, msg_name/msg_namelen are ignored
1770          * on connected socket. I was just happy when found this 8) --ANK
1771          */
1772
1773         /* Clean up data we have read: This will do ACK frames. */
1774         tcp_cleanup_rbuf(sk, copied);
1775
1776         release_sock(sk);
1777         return copied;
1778
1779 out:
1780         release_sock(sk);
1781         return err;
1782
1783 recv_urg:
1784         err = tcp_recv_urg(sk, msg, len, flags);
1785         goto out;
1786 }
1787 EXPORT_SYMBOL(tcp_recvmsg);
1788
1789 void tcp_set_state(struct sock *sk, int state)
1790 {
1791         int oldstate = sk->sk_state;
1792
1793         switch (state) {
1794         case TCP_ESTABLISHED:
1795                 if (oldstate != TCP_ESTABLISHED)
1796                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1797                 break;
1798
1799         case TCP_CLOSE:
1800                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1801                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1802
1803                 sk->sk_prot->unhash(sk);
1804                 if (inet_csk(sk)->icsk_bind_hash &&
1805                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1806                         inet_put_port(sk);
1807                 /* fall through */
1808         default:
1809                 if (oldstate == TCP_ESTABLISHED)
1810                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1811         }
1812
1813         /* Change state AFTER socket is unhashed to avoid closed
1814          * socket sitting in hash tables.
1815          */
1816         sk->sk_state = state;
1817
1818 #ifdef STATE_TRACE
1819         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1820 #endif
1821 }
1822 EXPORT_SYMBOL_GPL(tcp_set_state);
1823
1824 /*
1825  *      State processing on a close. This implements the state shift for
1826  *      sending our FIN frame. Note that we only send a FIN for some
1827  *      states. A shutdown() may have already sent the FIN, or we may be
1828  *      closed.
1829  */
1830
1831 static const unsigned char new_state[16] = {
1832   /* current state:        new state:      action:      */
1833   /* (Invalid)          */ TCP_CLOSE,
1834   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1835   /* TCP_SYN_SENT       */ TCP_CLOSE,
1836   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1837   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
1838   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
1839   /* TCP_TIME_WAIT      */ TCP_CLOSE,
1840   /* TCP_CLOSE          */ TCP_CLOSE,
1841   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
1842   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
1843   /* TCP_LISTEN         */ TCP_CLOSE,
1844   /* TCP_CLOSING        */ TCP_CLOSING,
1845 };
1846
1847 static int tcp_close_state(struct sock *sk)
1848 {
1849         int next = (int)new_state[sk->sk_state];
1850         int ns = next & TCP_STATE_MASK;
1851
1852         tcp_set_state(sk, ns);
1853
1854         return next & TCP_ACTION_FIN;
1855 }
1856
1857 /*
1858  *      Shutdown the sending side of a connection. Much like close except
1859  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1860  */
1861
1862 void tcp_shutdown(struct sock *sk, int how)
1863 {
1864         /*      We need to grab some memory, and put together a FIN,
1865          *      and then put it into the queue to be sent.
1866          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1867          */
1868         if (!(how & SEND_SHUTDOWN))
1869                 return;
1870
1871         /* If we've already sent a FIN, or it's a closed state, skip this. */
1872         if ((1 << sk->sk_state) &
1873             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1874              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1875                 /* Clear out any half completed packets.  FIN if needed. */
1876                 if (tcp_close_state(sk))
1877                         tcp_send_fin(sk);
1878         }
1879 }
1880 EXPORT_SYMBOL(tcp_shutdown);
1881
1882 bool tcp_check_oom(struct sock *sk, int shift)
1883 {
1884         bool too_many_orphans, out_of_socket_memory;
1885
1886         too_many_orphans = tcp_too_many_orphans(sk, shift);
1887         out_of_socket_memory = tcp_out_of_memory(sk);
1888
1889         if (too_many_orphans && net_ratelimit())
1890                 pr_info("too many orphaned sockets\n");
1891         if (out_of_socket_memory && net_ratelimit())
1892                 pr_info("out of memory -- consider tuning tcp_mem\n");
1893         return too_many_orphans || out_of_socket_memory;
1894 }
1895
1896 void tcp_close(struct sock *sk, long timeout)
1897 {
1898         struct sk_buff *skb;
1899         int data_was_unread = 0;
1900         int state;
1901
1902         lock_sock(sk);
1903         sk->sk_shutdown = SHUTDOWN_MASK;
1904
1905         if (sk->sk_state == TCP_LISTEN) {
1906                 tcp_set_state(sk, TCP_CLOSE);
1907
1908                 /* Special case. */
1909                 inet_csk_listen_stop(sk);
1910
1911                 goto adjudge_to_death;
1912         }
1913
1914         /*  We need to flush the recv. buffs.  We do this only on the
1915          *  descriptor close, not protocol-sourced closes, because the
1916          *  reader process may not have drained the data yet!
1917          */
1918         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1919                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1920                           tcp_hdr(skb)->fin;
1921                 data_was_unread += len;
1922                 __kfree_skb(skb);
1923         }
1924
1925         sk_mem_reclaim(sk);
1926
1927         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1928         if (sk->sk_state == TCP_CLOSE)
1929                 goto adjudge_to_death;
1930
1931         /* As outlined in RFC 2525, section 2.17, we send a RST here because
1932          * data was lost. To witness the awful effects of the old behavior of
1933          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1934          * GET in an FTP client, suspend the process, wait for the client to
1935          * advertise a zero window, then kill -9 the FTP client, wheee...
1936          * Note: timeout is always zero in such a case.
1937          */
1938         if (data_was_unread) {
1939                 /* Unread data was tossed, zap the connection. */
1940                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1941                 tcp_set_state(sk, TCP_CLOSE);
1942                 tcp_send_active_reset(sk, sk->sk_allocation);
1943         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1944                 /* Check zero linger _after_ checking for unread data. */
1945                 sk->sk_prot->disconnect(sk, 0);
1946                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1947         } else if (tcp_close_state(sk)) {
1948                 /* We FIN if the application ate all the data before
1949                  * zapping the connection.
1950                  */
1951
1952                 /* RED-PEN. Formally speaking, we have broken TCP state
1953                  * machine. State transitions:
1954                  *
1955                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1956                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1957                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1958                  *
1959                  * are legal only when FIN has been sent (i.e. in window),
1960                  * rather than queued out of window. Purists blame.
1961                  *
1962                  * F.e. "RFC state" is ESTABLISHED,
1963                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1964                  *
1965                  * The visible declinations are that sometimes
1966                  * we enter time-wait state, when it is not required really
1967                  * (harmless), do not send active resets, when they are
1968                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1969                  * they look as CLOSING or LAST_ACK for Linux)
1970                  * Probably, I missed some more holelets.
1971                  *                                              --ANK
1972                  */
1973                 tcp_send_fin(sk);
1974         }
1975
1976         sk_stream_wait_close(sk, timeout);
1977
1978 adjudge_to_death:
1979         state = sk->sk_state;
1980         sock_hold(sk);
1981         sock_orphan(sk);
1982
1983         /* It is the last release_sock in its life. It will remove backlog. */
1984         release_sock(sk);
1985
1986
1987         /* Now socket is owned by kernel and we acquire BH lock
1988            to finish close. No need to check for user refs.
1989          */
1990         local_bh_disable();
1991         bh_lock_sock(sk);
1992         WARN_ON(sock_owned_by_user(sk));
1993
1994         percpu_counter_inc(sk->sk_prot->orphan_count);
1995
1996         /* Have we already been destroyed by a softirq or backlog? */
1997         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1998                 goto out;
1999
2000         /*      This is a (useful) BSD violating of the RFC. There is a
2001          *      problem with TCP as specified in that the other end could
2002          *      keep a socket open forever with no application left this end.
2003          *      We use a 3 minute timeout (about the same as BSD) then kill
2004          *      our end. If they send after that then tough - BUT: long enough
2005          *      that we won't make the old 4*rto = almost no time - whoops
2006          *      reset mistake.
2007          *
2008          *      Nope, it was not mistake. It is really desired behaviour
2009          *      f.e. on http servers, when such sockets are useless, but
2010          *      consume significant resources. Let's do it with special
2011          *      linger2 option.                                 --ANK
2012          */
2013
2014         if (sk->sk_state == TCP_FIN_WAIT2) {
2015                 struct tcp_sock *tp = tcp_sk(sk);
2016                 if (tp->linger2 < 0) {
2017                         tcp_set_state(sk, TCP_CLOSE);
2018                         tcp_send_active_reset(sk, GFP_ATOMIC);
2019                         NET_INC_STATS_BH(sock_net(sk),
2020                                         LINUX_MIB_TCPABORTONLINGER);
2021                 } else {
2022                         const int tmo = tcp_fin_time(sk);
2023
2024                         if (tmo > TCP_TIMEWAIT_LEN) {
2025                                 inet_csk_reset_keepalive_timer(sk,
2026                                                 tmo - TCP_TIMEWAIT_LEN);
2027                         } else {
2028                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2029                                 goto out;
2030                         }
2031                 }
2032         }
2033         if (sk->sk_state != TCP_CLOSE) {
2034                 sk_mem_reclaim(sk);
2035                 if (tcp_check_oom(sk, 0)) {
2036                         tcp_set_state(sk, TCP_CLOSE);
2037                         tcp_send_active_reset(sk, GFP_ATOMIC);
2038                         NET_INC_STATS_BH(sock_net(sk),
2039                                         LINUX_MIB_TCPABORTONMEMORY);
2040                 }
2041         }
2042
2043         if (sk->sk_state == TCP_CLOSE)
2044                 inet_csk_destroy_sock(sk);
2045         /* Otherwise, socket is reprieved until protocol close. */
2046
2047 out:
2048         bh_unlock_sock(sk);
2049         local_bh_enable();
2050         sock_put(sk);
2051 }
2052 EXPORT_SYMBOL(tcp_close);
2053
2054 /* These states need RST on ABORT according to RFC793 */
2055
2056 static inline int tcp_need_reset(int state)
2057 {
2058         return (1 << state) &
2059                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2060                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2061 }
2062
2063 int tcp_disconnect(struct sock *sk, int flags)
2064 {
2065         struct inet_sock *inet = inet_sk(sk);
2066         struct inet_connection_sock *icsk = inet_csk(sk);
2067         struct tcp_sock *tp = tcp_sk(sk);
2068         int err = 0;
2069         int old_state = sk->sk_state;
2070
2071         if (old_state != TCP_CLOSE)
2072                 tcp_set_state(sk, TCP_CLOSE);
2073
2074         /* ABORT function of RFC793 */
2075         if (old_state == TCP_LISTEN) {
2076                 inet_csk_listen_stop(sk);
2077         } else if (tcp_need_reset(old_state) ||
2078                    (tp->snd_nxt != tp->write_seq &&
2079                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2080                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2081                  * states
2082                  */
2083                 tcp_send_active_reset(sk, gfp_any());
2084                 sk->sk_err = ECONNRESET;
2085         } else if (old_state == TCP_SYN_SENT)
2086                 sk->sk_err = ECONNRESET;
2087
2088         tcp_clear_xmit_timers(sk);
2089         __skb_queue_purge(&sk->sk_receive_queue);
2090         tcp_write_queue_purge(sk);
2091         __skb_queue_purge(&tp->out_of_order_queue);
2092 #ifdef CONFIG_NET_DMA
2093         __skb_queue_purge(&sk->sk_async_wait_queue);
2094 #endif
2095
2096         inet->inet_dport = 0;
2097
2098         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2099                 inet_reset_saddr(sk);
2100
2101         sk->sk_shutdown = 0;
2102         sock_reset_flag(sk, SOCK_DONE);
2103         tp->srtt = 0;
2104         if ((tp->write_seq += tp->max_window + 2) == 0)
2105                 tp->write_seq = 1;
2106         icsk->icsk_backoff = 0;
2107         tp->snd_cwnd = 2;
2108         icsk->icsk_probes_out = 0;
2109         tp->packets_out = 0;
2110         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2111         tp->snd_cwnd_cnt = 0;
2112         tp->bytes_acked = 0;
2113         tp->window_clamp = 0;
2114         tcp_set_ca_state(sk, TCP_CA_Open);
2115         tcp_clear_retrans(tp);
2116         inet_csk_delack_init(sk);
2117         tcp_init_send_head(sk);
2118         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2119         __sk_dst_reset(sk);
2120
2121         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2122
2123         sk->sk_error_report(sk);
2124         return err;
2125 }
2126 EXPORT_SYMBOL(tcp_disconnect);
2127
2128 /*
2129  *      Socket option code for TCP.
2130  */
2131 static int do_tcp_setsockopt(struct sock *sk, int level,
2132                 int optname, char __user *optval, unsigned int optlen)
2133 {
2134         struct tcp_sock *tp = tcp_sk(sk);
2135         struct inet_connection_sock *icsk = inet_csk(sk);
2136         int val;
2137         int err = 0;
2138
2139         /* These are data/string values, all the others are ints */
2140         switch (optname) {
2141         case TCP_CONGESTION: {
2142                 char name[TCP_CA_NAME_MAX];
2143
2144                 if (optlen < 1)
2145                         return -EINVAL;
2146
2147                 val = strncpy_from_user(name, optval,
2148                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2149                 if (val < 0)
2150                         return -EFAULT;
2151                 name[val] = 0;
2152
2153                 lock_sock(sk);
2154                 err = tcp_set_congestion_control(sk, name);
2155                 release_sock(sk);
2156                 return err;
2157         }
2158         case TCP_COOKIE_TRANSACTIONS: {
2159                 struct tcp_cookie_transactions ctd;
2160                 struct tcp_cookie_values *cvp = NULL;
2161
2162                 if (sizeof(ctd) > optlen)
2163                         return -EINVAL;
2164                 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2165                         return -EFAULT;
2166
2167                 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2168                     ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2169                         return -EINVAL;
2170
2171                 if (ctd.tcpct_cookie_desired == 0) {
2172                         /* default to global value */
2173                 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2174                            ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2175                            ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2176                         return -EINVAL;
2177                 }
2178
2179                 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2180                         /* Supercedes all other values */
2181                         lock_sock(sk);
2182                         if (tp->cookie_values != NULL) {
2183                                 kref_put(&tp->cookie_values->kref,
2184                                          tcp_cookie_values_release);
2185                                 tp->cookie_values = NULL;
2186                         }
2187                         tp->rx_opt.cookie_in_always = 0; /* false */
2188                         tp->rx_opt.cookie_out_never = 1; /* true */
2189                         release_sock(sk);
2190                         return err;
2191                 }
2192
2193                 /* Allocate ancillary memory before locking.
2194                  */
2195                 if (ctd.tcpct_used > 0 ||
2196                     (tp->cookie_values == NULL &&
2197                      (sysctl_tcp_cookie_size > 0 ||
2198                       ctd.tcpct_cookie_desired > 0 ||
2199                       ctd.tcpct_s_data_desired > 0))) {
2200                         cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2201                                       GFP_KERNEL);
2202                         if (cvp == NULL)
2203                                 return -ENOMEM;
2204
2205                         kref_init(&cvp->kref);
2206                 }
2207                 lock_sock(sk);
2208                 tp->rx_opt.cookie_in_always =
2209                         (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2210                 tp->rx_opt.cookie_out_never = 0; /* false */
2211
2212                 if (tp->cookie_values != NULL) {
2213                         if (cvp != NULL) {
2214                                 /* Changed values are recorded by a changed
2215                                  * pointer, ensuring the cookie will differ,
2216                                  * without separately hashing each value later.
2217                                  */
2218                                 kref_put(&tp->cookie_values->kref,
2219                                          tcp_cookie_values_release);
2220                         } else {
2221                                 cvp = tp->cookie_values;
2222                         }
2223                 }
2224
2225                 if (cvp != NULL) {
2226                         cvp->cookie_desired = ctd.tcpct_cookie_desired;
2227
2228                         if (ctd.tcpct_used > 0) {
2229                                 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2230                                        ctd.tcpct_used);
2231                                 cvp->s_data_desired = ctd.tcpct_used;
2232                                 cvp->s_data_constant = 1; /* true */
2233                         } else {
2234                                 /* No constant payload data. */
2235                                 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2236                                 cvp->s_data_constant = 0; /* false */
2237                         }
2238
2239                         tp->cookie_values = cvp;
2240                 }
2241                 release_sock(sk);
2242                 return err;
2243         }
2244         default:
2245                 /* fallthru */
2246                 break;
2247         }
2248
2249         if (optlen < sizeof(int))
2250                 return -EINVAL;
2251
2252         if (get_user(val, (int __user *)optval))
2253                 return -EFAULT;
2254
2255         lock_sock(sk);
2256
2257         switch (optname) {
2258         case TCP_MAXSEG:
2259                 /* Values greater than interface MTU won't take effect. However
2260                  * at the point when this call is done we typically don't yet
2261                  * know which interface is going to be used */
2262                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2263                         err = -EINVAL;
2264                         break;
2265                 }
2266                 tp->rx_opt.user_mss = val;
2267                 break;
2268
2269         case TCP_NODELAY:
2270                 if (val) {
2271                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2272                          * this option on corked socket is remembered, but
2273                          * it is not activated until cork is cleared.
2274                          *
2275                          * However, when TCP_NODELAY is set we make
2276                          * an explicit push, which overrides even TCP_CORK
2277                          * for currently queued segments.
2278                          */
2279                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2280                         tcp_push_pending_frames(sk);
2281                 } else {
2282                         tp->nonagle &= ~TCP_NAGLE_OFF;
2283                 }
2284                 break;
2285
2286         case TCP_THIN_LINEAR_TIMEOUTS:
2287                 if (val < 0 || val > 1)
2288                         err = -EINVAL;
2289                 else
2290                         tp->thin_lto = val;
2291                 break;
2292
2293         case TCP_THIN_DUPACK:
2294                 if (val < 0 || val > 1)
2295                         err = -EINVAL;
2296                 else
2297                         tp->thin_dupack = val;
2298                 break;
2299
2300         case TCP_CORK:
2301                 /* When set indicates to always queue non-full frames.
2302                  * Later the user clears this option and we transmit
2303                  * any pending partial frames in the queue.  This is
2304                  * meant to be used alongside sendfile() to get properly
2305                  * filled frames when the user (for example) must write
2306                  * out headers with a write() call first and then use
2307                  * sendfile to send out the data parts.
2308                  *
2309                  * TCP_CORK can be set together with TCP_NODELAY and it is
2310                  * stronger than TCP_NODELAY.
2311                  */
2312                 if (val) {
2313                         tp->nonagle |= TCP_NAGLE_CORK;
2314                 } else {
2315                         tp->nonagle &= ~TCP_NAGLE_CORK;
2316                         if (tp->nonagle&TCP_NAGLE_OFF)
2317                                 tp->nonagle |= TCP_NAGLE_PUSH;
2318                         tcp_push_pending_frames(sk);
2319                 }
2320                 break;
2321
2322         case TCP_KEEPIDLE:
2323                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2324                         err = -EINVAL;
2325                 else {
2326                         tp->keepalive_time = val * HZ;
2327                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2328                             !((1 << sk->sk_state) &
2329                               (TCPF_CLOSE | TCPF_LISTEN))) {
2330                                 u32 elapsed = keepalive_time_elapsed(tp);
2331                                 if (tp->keepalive_time > elapsed)
2332                                         elapsed = tp->keepalive_time - elapsed;
2333                                 else
2334                                         elapsed = 0;
2335                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2336                         }
2337                 }
2338                 break;
2339         case TCP_KEEPINTVL:
2340                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2341                         err = -EINVAL;
2342                 else
2343                         tp->keepalive_intvl = val * HZ;
2344                 break;
2345         case TCP_KEEPCNT:
2346                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2347                         err = -EINVAL;
2348                 else
2349                         tp->keepalive_probes = val;
2350                 break;
2351         case TCP_SYNCNT:
2352                 if (val < 1 || val > MAX_TCP_SYNCNT)
2353                         err = -EINVAL;
2354                 else
2355                         icsk->icsk_syn_retries = val;
2356                 break;
2357
2358         case TCP_LINGER2:
2359                 if (val < 0)
2360                         tp->linger2 = -1;
2361                 else if (val > sysctl_tcp_fin_timeout / HZ)
2362                         tp->linger2 = 0;
2363                 else
2364                         tp->linger2 = val * HZ;
2365                 break;
2366
2367         case TCP_DEFER_ACCEPT:
2368                 /* Translate value in seconds to number of retransmits */
2369                 icsk->icsk_accept_queue.rskq_defer_accept =
2370                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2371                                         TCP_RTO_MAX / HZ);
2372                 break;
2373
2374         case TCP_WINDOW_CLAMP:
2375                 if (!val) {
2376                         if (sk->sk_state != TCP_CLOSE) {
2377                                 err = -EINVAL;
2378                                 break;
2379                         }
2380                         tp->window_clamp = 0;
2381                 } else
2382                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2383                                                 SOCK_MIN_RCVBUF / 2 : val;
2384                 break;
2385
2386         case TCP_QUICKACK:
2387                 if (!val) {
2388                         icsk->icsk_ack.pingpong = 1;
2389                 } else {
2390                         icsk->icsk_ack.pingpong = 0;
2391                         if ((1 << sk->sk_state) &
2392                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2393                             inet_csk_ack_scheduled(sk)) {
2394                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2395                                 tcp_cleanup_rbuf(sk, 1);
2396                                 if (!(val & 1))
2397                                         icsk->icsk_ack.pingpong = 1;
2398                         }
2399                 }
2400                 break;
2401
2402 #ifdef CONFIG_TCP_MD5SIG
2403         case TCP_MD5SIG:
2404                 /* Read the IP->Key mappings from userspace */
2405                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2406                 break;
2407 #endif
2408         case TCP_USER_TIMEOUT:
2409                 /* Cap the max timeout in ms TCP will retry/retrans
2410                  * before giving up and aborting (ETIMEDOUT) a connection.
2411                  */
2412                 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2413                 break;
2414         default:
2415                 err = -ENOPROTOOPT;
2416                 break;
2417         }
2418
2419         release_sock(sk);
2420         return err;
2421 }
2422
2423 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2424                    unsigned int optlen)
2425 {
2426         const struct inet_connection_sock *icsk = inet_csk(sk);
2427
2428         if (level != SOL_TCP)
2429                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2430                                                      optval, optlen);
2431         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2432 }
2433 EXPORT_SYMBOL(tcp_setsockopt);
2434
2435 #ifdef CONFIG_COMPAT
2436 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2437                           char __user *optval, unsigned int optlen)
2438 {
2439         if (level != SOL_TCP)
2440                 return inet_csk_compat_setsockopt(sk, level, optname,
2441                                                   optval, optlen);
2442         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2443 }
2444 EXPORT_SYMBOL(compat_tcp_setsockopt);
2445 #endif
2446
2447 /* Return information about state of tcp endpoint in API format. */
2448 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2449 {
2450         const struct tcp_sock *tp = tcp_sk(sk);
2451         const struct inet_connection_sock *icsk = inet_csk(sk);
2452         u32 now = tcp_time_stamp;
2453
2454         memset(info, 0, sizeof(*info));
2455
2456         info->tcpi_state = sk->sk_state;
2457         info->tcpi_ca_state = icsk->icsk_ca_state;
2458         info->tcpi_retransmits = icsk->icsk_retransmits;
2459         info->tcpi_probes = icsk->icsk_probes_out;
2460         info->tcpi_backoff = icsk->icsk_backoff;
2461
2462         if (tp->rx_opt.tstamp_ok)
2463                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2464         if (tcp_is_sack(tp))
2465                 info->tcpi_options |= TCPI_OPT_SACK;
2466         if (tp->rx_opt.wscale_ok) {
2467                 info->tcpi_options |= TCPI_OPT_WSCALE;
2468                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2469                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2470         }
2471
2472         if (tp->ecn_flags & TCP_ECN_OK)
2473                 info->tcpi_options |= TCPI_OPT_ECN;
2474         if (tp->ecn_flags & TCP_ECN_SEEN)
2475                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2476
2477         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2478         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2479         info->tcpi_snd_mss = tp->mss_cache;
2480         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2481
2482         if (sk->sk_state == TCP_LISTEN) {
2483                 info->tcpi_unacked = sk->sk_ack_backlog;
2484                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2485         } else {
2486                 info->tcpi_unacked = tp->packets_out;
2487                 info->tcpi_sacked = tp->sacked_out;
2488         }
2489         info->tcpi_lost = tp->lost_out;
2490         info->tcpi_retrans = tp->retrans_out;
2491         info->tcpi_fackets = tp->fackets_out;
2492
2493         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2494         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2495         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2496
2497         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2498         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2499         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2500         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2501         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2502         info->tcpi_snd_cwnd = tp->snd_cwnd;
2503         info->tcpi_advmss = tp->advmss;
2504         info->tcpi_reordering = tp->reordering;
2505
2506         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2507         info->tcpi_rcv_space = tp->rcvq_space.space;
2508
2509         info->tcpi_total_retrans = tp->total_retrans;
2510 }
2511 EXPORT_SYMBOL_GPL(tcp_get_info);
2512
2513 static int do_tcp_getsockopt(struct sock *sk, int level,
2514                 int optname, char __user *optval, int __user *optlen)
2515 {
2516         struct inet_connection_sock *icsk = inet_csk(sk);
2517         struct tcp_sock *tp = tcp_sk(sk);
2518         int val, len;
2519
2520         if (get_user(len, optlen))
2521                 return -EFAULT;
2522
2523         len = min_t(unsigned int, len, sizeof(int));
2524
2525         if (len < 0)
2526                 return -EINVAL;
2527
2528         switch (optname) {
2529         case TCP_MAXSEG:
2530                 val = tp->mss_cache;
2531                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2532                         val = tp->rx_opt.user_mss;
2533                 break;
2534         case TCP_NODELAY:
2535                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2536                 break;
2537         case TCP_CORK:
2538                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2539                 break;
2540         case TCP_KEEPIDLE:
2541                 val = keepalive_time_when(tp) / HZ;
2542                 break;
2543         case TCP_KEEPINTVL:
2544                 val = keepalive_intvl_when(tp) / HZ;
2545                 break;
2546         case TCP_KEEPCNT:
2547                 val = keepalive_probes(tp);
2548                 break;
2549         case TCP_SYNCNT:
2550                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2551                 break;
2552         case TCP_LINGER2:
2553                 val = tp->linger2;
2554                 if (val >= 0)
2555                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2556                 break;
2557         case TCP_DEFER_ACCEPT:
2558                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2559                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2560                 break;
2561         case TCP_WINDOW_CLAMP:
2562                 val = tp->window_clamp;
2563                 break;
2564         case TCP_INFO: {
2565                 struct tcp_info info;
2566
2567                 if (get_user(len, optlen))
2568                         return -EFAULT;
2569
2570                 tcp_get_info(sk, &info);
2571
2572                 len = min_t(unsigned int, len, sizeof(info));
2573                 if (put_user(len, optlen))
2574                         return -EFAULT;
2575                 if (copy_to_user(optval, &info, len))
2576                         return -EFAULT;
2577                 return 0;
2578         }
2579         case TCP_QUICKACK:
2580                 val = !icsk->icsk_ack.pingpong;
2581                 break;
2582
2583         case TCP_CONGESTION:
2584                 if (get_user(len, optlen))
2585                         return -EFAULT;
2586                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2587                 if (put_user(len, optlen))
2588                         return -EFAULT;
2589                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2590                         return -EFAULT;
2591                 return 0;
2592
2593         case TCP_COOKIE_TRANSACTIONS: {
2594                 struct tcp_cookie_transactions ctd;
2595                 struct tcp_cookie_values *cvp = tp->cookie_values;
2596
2597                 if (get_user(len, optlen))
2598                         return -EFAULT;
2599                 if (len < sizeof(ctd))
2600                         return -EINVAL;
2601
2602                 memset(&ctd, 0, sizeof(ctd));
2603                 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2604                                    TCP_COOKIE_IN_ALWAYS : 0)
2605                                 | (tp->rx_opt.cookie_out_never ?
2606                                    TCP_COOKIE_OUT_NEVER : 0);
2607
2608                 if (cvp != NULL) {
2609                         ctd.tcpct_flags |= (cvp->s_data_in ?
2610                                             TCP_S_DATA_IN : 0)
2611                                          | (cvp->s_data_out ?
2612                                             TCP_S_DATA_OUT : 0);
2613
2614                         ctd.tcpct_cookie_desired = cvp->cookie_desired;
2615                         ctd.tcpct_s_data_desired = cvp->s_data_desired;
2616
2617                         memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2618                                cvp->cookie_pair_size);
2619                         ctd.tcpct_used = cvp->cookie_pair_size;
2620                 }
2621
2622                 if (put_user(sizeof(ctd), optlen))
2623                         return -EFAULT;
2624                 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2625                         return -EFAULT;
2626                 return 0;
2627         }
2628         case TCP_THIN_LINEAR_TIMEOUTS:
2629                 val = tp->thin_lto;
2630                 break;
2631         case TCP_THIN_DUPACK:
2632                 val = tp->thin_dupack;
2633                 break;
2634
2635         case TCP_USER_TIMEOUT:
2636                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2637                 break;
2638         default:
2639                 return -ENOPROTOOPT;
2640         }
2641
2642         if (put_user(len, optlen))
2643                 return -EFAULT;
2644         if (copy_to_user(optval, &val, len))
2645                 return -EFAULT;
2646         return 0;
2647 }
2648
2649 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2650                    int __user *optlen)
2651 {
2652         struct inet_connection_sock *icsk = inet_csk(sk);
2653
2654         if (level != SOL_TCP)
2655                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2656                                                      optval, optlen);
2657         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2658 }
2659 EXPORT_SYMBOL(tcp_getsockopt);
2660
2661 #ifdef CONFIG_COMPAT
2662 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2663                           char __user *optval, int __user *optlen)
2664 {
2665         if (level != SOL_TCP)
2666                 return inet_csk_compat_getsockopt(sk, level, optname,
2667                                                   optval, optlen);
2668         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2669 }
2670 EXPORT_SYMBOL(compat_tcp_getsockopt);
2671 #endif
2672
2673 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2674         netdev_features_t features)
2675 {
2676         struct sk_buff *segs = ERR_PTR(-EINVAL);
2677         struct tcphdr *th;
2678         unsigned thlen;
2679         unsigned int seq;
2680         __be32 delta;
2681         unsigned int oldlen;
2682         unsigned int mss;
2683
2684         if (!pskb_may_pull(skb, sizeof(*th)))
2685                 goto out;
2686
2687         th = tcp_hdr(skb);
2688         thlen = th->doff * 4;
2689         if (thlen < sizeof(*th))
2690                 goto out;
2691
2692         if (!pskb_may_pull(skb, thlen))
2693                 goto out;
2694
2695         oldlen = (u16)~skb->len;
2696         __skb_pull(skb, thlen);
2697
2698         mss = skb_shinfo(skb)->gso_size;
2699         if (unlikely(skb->len <= mss))
2700                 goto out;
2701
2702         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2703                 /* Packet is from an untrusted source, reset gso_segs. */
2704                 int type = skb_shinfo(skb)->gso_type;
2705
2706                 if (unlikely(type &
2707                              ~(SKB_GSO_TCPV4 |
2708                                SKB_GSO_DODGY |
2709                                SKB_GSO_TCP_ECN |
2710                                SKB_GSO_TCPV6 |
2711                                0) ||
2712                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2713                         goto out;
2714
2715                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2716
2717                 segs = NULL;
2718                 goto out;
2719         }
2720
2721         segs = skb_segment(skb, features);
2722         if (IS_ERR(segs))
2723                 goto out;
2724
2725         delta = htonl(oldlen + (thlen + mss));
2726
2727         skb = segs;
2728         th = tcp_hdr(skb);
2729         seq = ntohl(th->seq);
2730
2731         do {
2732                 th->fin = th->psh = 0;
2733
2734                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2735                                        (__force u32)delta));
2736                 if (skb->ip_summed != CHECKSUM_PARTIAL)
2737                         th->check =
2738                              csum_fold(csum_partial(skb_transport_header(skb),
2739                                                     thlen, skb->csum));
2740
2741                 seq += mss;
2742                 skb = skb->next;
2743                 th = tcp_hdr(skb);
2744
2745                 th->seq = htonl(seq);
2746                 th->cwr = 0;
2747         } while (skb->next);
2748
2749         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2750                       skb->data_len);
2751         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2752                                 (__force u32)delta));
2753         if (skb->ip_summed != CHECKSUM_PARTIAL)
2754                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2755                                                    thlen, skb->csum));
2756
2757 out:
2758         return segs;
2759 }
2760 EXPORT_SYMBOL(tcp_tso_segment);
2761
2762 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2763 {
2764         struct sk_buff **pp = NULL;
2765         struct sk_buff *p;
2766         struct tcphdr *th;
2767         struct tcphdr *th2;
2768         unsigned int len;
2769         unsigned int thlen;
2770         __be32 flags;
2771         unsigned int mss = 1;
2772         unsigned int hlen;
2773         unsigned int off;
2774         int flush = 1;
2775         int i;
2776
2777         off = skb_gro_offset(skb);
2778         hlen = off + sizeof(*th);
2779         th = skb_gro_header_fast(skb, off);
2780         if (skb_gro_header_hard(skb, hlen)) {
2781                 th = skb_gro_header_slow(skb, hlen, off);
2782                 if (unlikely(!th))
2783                         goto out;
2784         }
2785
2786         thlen = th->doff * 4;
2787         if (thlen < sizeof(*th))
2788                 goto out;
2789
2790         hlen = off + thlen;
2791         if (skb_gro_header_hard(skb, hlen)) {
2792                 th = skb_gro_header_slow(skb, hlen, off);
2793                 if (unlikely(!th))
2794                         goto out;
2795         }
2796
2797         skb_gro_pull(skb, thlen);
2798
2799         len = skb_gro_len(skb);
2800         flags = tcp_flag_word(th);
2801
2802         for (; (p = *head); head = &p->next) {
2803                 if (!NAPI_GRO_CB(p)->same_flow)
2804                         continue;
2805
2806                 th2 = tcp_hdr(p);
2807
2808                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2809                         NAPI_GRO_CB(p)->same_flow = 0;
2810                         continue;
2811                 }
2812
2813                 goto found;
2814         }
2815
2816         goto out_check_final;
2817
2818 found:
2819         flush = NAPI_GRO_CB(p)->flush;
2820         flush |= (__force int)(flags & TCP_FLAG_CWR);
2821         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2822                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2823         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2824         for (i = sizeof(*th); i < thlen; i += 4)
2825                 flush |= *(u32 *)((u8 *)th + i) ^
2826                          *(u32 *)((u8 *)th2 + i);
2827
2828         mss = skb_shinfo(p)->gso_size;
2829
2830         flush |= (len - 1) >= mss;
2831         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2832
2833         if (flush || skb_gro_receive(head, skb)) {
2834                 mss = 1;
2835                 goto out_check_final;
2836         }
2837
2838         p = *head;
2839         th2 = tcp_hdr(p);
2840         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2841
2842 out_check_final:
2843         flush = len < mss;
2844         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2845                                         TCP_FLAG_RST | TCP_FLAG_SYN |
2846                                         TCP_FLAG_FIN));
2847
2848         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2849                 pp = head;
2850
2851 out:
2852         NAPI_GRO_CB(skb)->flush |= flush;
2853
2854         return pp;
2855 }
2856 EXPORT_SYMBOL(tcp_gro_receive);
2857
2858 int tcp_gro_complete(struct sk_buff *skb)
2859 {
2860         struct tcphdr *th = tcp_hdr(skb);
2861
2862         skb->csum_start = skb_transport_header(skb) - skb->head;
2863         skb->csum_offset = offsetof(struct tcphdr, check);
2864         skb->ip_summed = CHECKSUM_PARTIAL;
2865
2866         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2867
2868         if (th->cwr)
2869                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2870
2871         return 0;
2872 }
2873 EXPORT_SYMBOL(tcp_gro_complete);
2874
2875 #ifdef CONFIG_TCP_MD5SIG
2876 static unsigned long tcp_md5sig_users;
2877 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
2878 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2879
2880 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
2881 {
2882         int cpu;
2883
2884         for_each_possible_cpu(cpu) {
2885                 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
2886
2887                 if (p->md5_desc.tfm)
2888                         crypto_free_hash(p->md5_desc.tfm);
2889         }
2890         free_percpu(pool);
2891 }
2892
2893 void tcp_free_md5sig_pool(void)
2894 {
2895         struct tcp_md5sig_pool __percpu *pool = NULL;
2896
2897         spin_lock_bh(&tcp_md5sig_pool_lock);
2898         if (--tcp_md5sig_users == 0) {
2899                 pool = tcp_md5sig_pool;
2900                 tcp_md5sig_pool = NULL;
2901         }
2902         spin_unlock_bh(&tcp_md5sig_pool_lock);
2903         if (pool)
2904                 __tcp_free_md5sig_pool(pool);
2905 }
2906 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2907
2908 static struct tcp_md5sig_pool __percpu *
2909 __tcp_alloc_md5sig_pool(struct sock *sk)
2910 {
2911         int cpu;
2912         struct tcp_md5sig_pool __percpu *pool;
2913
2914         pool = alloc_percpu(struct tcp_md5sig_pool);
2915         if (!pool)
2916                 return NULL;
2917
2918         for_each_possible_cpu(cpu) {
2919                 struct crypto_hash *hash;
2920
2921                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2922                 if (!hash || IS_ERR(hash))
2923                         goto out_free;
2924
2925                 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
2926         }
2927         return pool;
2928 out_free:
2929         __tcp_free_md5sig_pool(pool);
2930         return NULL;
2931 }
2932
2933 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2934 {
2935         struct tcp_md5sig_pool __percpu *pool;
2936         int alloc = 0;
2937
2938 retry:
2939         spin_lock_bh(&tcp_md5sig_pool_lock);
2940         pool = tcp_md5sig_pool;
2941         if (tcp_md5sig_users++ == 0) {
2942                 alloc = 1;
2943                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2944         } else if (!pool) {
2945                 tcp_md5sig_users--;
2946                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2947                 cpu_relax();
2948                 goto retry;
2949         } else
2950                 spin_unlock_bh(&tcp_md5sig_pool_lock);
2951
2952         if (alloc) {
2953                 /* we cannot hold spinlock here because this may sleep. */
2954                 struct tcp_md5sig_pool __percpu *p;
2955
2956                 p = __tcp_alloc_md5sig_pool(sk);
2957                 spin_lock_bh(&tcp_md5sig_pool_lock);
2958                 if (!p) {
2959                         tcp_md5sig_users--;
2960                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2961                         return NULL;
2962                 }
2963                 pool = tcp_md5sig_pool;
2964                 if (pool) {
2965                         /* oops, it has already been assigned. */
2966                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2967                         __tcp_free_md5sig_pool(p);
2968                 } else {
2969                         tcp_md5sig_pool = pool = p;
2970                         spin_unlock_bh(&tcp_md5sig_pool_lock);
2971                 }
2972         }
2973         return pool;
2974 }
2975 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2976
2977
2978 /**
2979  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2980  *
2981  *      We use percpu structure, so if we succeed, we exit with preemption
2982  *      and BH disabled, to make sure another thread or softirq handling
2983  *      wont try to get same context.
2984  */
2985 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2986 {
2987         struct tcp_md5sig_pool __percpu *p;
2988
2989         local_bh_disable();
2990
2991         spin_lock(&tcp_md5sig_pool_lock);
2992         p = tcp_md5sig_pool;
2993         if (p)
2994                 tcp_md5sig_users++;
2995         spin_unlock(&tcp_md5sig_pool_lock);
2996
2997         if (p)
2998                 return this_cpu_ptr(p);
2999
3000         local_bh_enable();
3001         return NULL;
3002 }
3003 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3004
3005 void tcp_put_md5sig_pool(void)
3006 {
3007         local_bh_enable();
3008         tcp_free_md5sig_pool();
3009 }
3010 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3011
3012 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3013                         const struct tcphdr *th)
3014 {
3015         struct scatterlist sg;
3016         struct tcphdr hdr;
3017         int err;
3018
3019         /* We are not allowed to change tcphdr, make a local copy */
3020         memcpy(&hdr, th, sizeof(hdr));
3021         hdr.check = 0;
3022
3023         /* options aren't included in the hash */
3024         sg_init_one(&sg, &hdr, sizeof(hdr));
3025         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3026         return err;
3027 }
3028 EXPORT_SYMBOL(tcp_md5_hash_header);
3029
3030 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3031                           const struct sk_buff *skb, unsigned int header_len)
3032 {
3033         struct scatterlist sg;
3034         const struct tcphdr *tp = tcp_hdr(skb);
3035         struct hash_desc *desc = &hp->md5_desc;
3036         unsigned i;
3037         const unsigned head_data_len = skb_headlen(skb) > header_len ?
3038                                        skb_headlen(skb) - header_len : 0;
3039         const struct skb_shared_info *shi = skb_shinfo(skb);
3040         struct sk_buff *frag_iter;
3041
3042         sg_init_table(&sg, 1);
3043
3044         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3045         if (crypto_hash_update(desc, &sg, head_data_len))
3046                 return 1;
3047
3048         for (i = 0; i < shi->nr_frags; ++i) {
3049                 const struct skb_frag_struct *f = &shi->frags[i];
3050                 struct page *page = skb_frag_page(f);
3051                 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3052                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3053                         return 1;
3054         }
3055
3056         skb_walk_frags(skb, frag_iter)
3057                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3058                         return 1;
3059
3060         return 0;
3061 }
3062 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3063
3064 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3065 {
3066         struct scatterlist sg;
3067
3068         sg_init_one(&sg, key->key, key->keylen);
3069         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3070 }
3071 EXPORT_SYMBOL(tcp_md5_hash_key);
3072
3073 #endif
3074
3075 /**
3076  * Each Responder maintains up to two secret values concurrently for
3077  * efficient secret rollover.  Each secret value has 4 states:
3078  *
3079  * Generating.  (tcp_secret_generating != tcp_secret_primary)
3080  *    Generates new Responder-Cookies, but not yet used for primary
3081  *    verification.  This is a short-term state, typically lasting only
3082  *    one round trip time (RTT).
3083  *
3084  * Primary.  (tcp_secret_generating == tcp_secret_primary)
3085  *    Used both for generation and primary verification.
3086  *
3087  * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3088  *    Used for verification, until the first failure that can be
3089  *    verified by the newer Generating secret.  At that time, this
3090  *    cookie's state is changed to Secondary, and the Generating
3091  *    cookie's state is changed to Primary.  This is a short-term state,
3092  *    typically lasting only one round trip time (RTT).
3093  *
3094  * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3095  *    Used for secondary verification, after primary verification
3096  *    failures.  This state lasts no more than twice the Maximum Segment
3097  *    Lifetime (2MSL).  Then, the secret is discarded.
3098  */
3099 struct tcp_cookie_secret {
3100         /* The secret is divided into two parts.  The digest part is the
3101          * equivalent of previously hashing a secret and saving the state,
3102          * and serves as an initialization vector (IV).  The message part
3103          * serves as the trailing secret.
3104          */
3105         u32                             secrets[COOKIE_WORKSPACE_WORDS];
3106         unsigned long                   expires;
3107 };
3108
3109 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3110 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3111 #define TCP_SECRET_LIFE (HZ * 600)
3112
3113 static struct tcp_cookie_secret tcp_secret_one;
3114 static struct tcp_cookie_secret tcp_secret_two;
3115
3116 /* Essentially a circular list, without dynamic allocation. */
3117 static struct tcp_cookie_secret *tcp_secret_generating;
3118 static struct tcp_cookie_secret *tcp_secret_primary;
3119 static struct tcp_cookie_secret *tcp_secret_retiring;
3120 static struct tcp_cookie_secret *tcp_secret_secondary;
3121
3122 static DEFINE_SPINLOCK(tcp_secret_locker);
3123
3124 /* Select a pseudo-random word in the cookie workspace.
3125  */
3126 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3127 {
3128         return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3129 }
3130
3131 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3132  * Called in softirq context.
3133  * Returns: 0 for success.
3134  */
3135 int tcp_cookie_generator(u32 *bakery)
3136 {
3137         unsigned long jiffy = jiffies;
3138
3139         if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3140                 spin_lock_bh(&tcp_secret_locker);
3141                 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3142                         /* refreshed by another */
3143                         memcpy(bakery,
3144                                &tcp_secret_generating->secrets[0],
3145                                COOKIE_WORKSPACE_WORDS);
3146                 } else {
3147                         /* still needs refreshing */
3148                         get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3149
3150                         /* The first time, paranoia assumes that the
3151                          * randomization function isn't as strong.  But,
3152                          * this secret initialization is delayed until
3153                          * the last possible moment (packet arrival).
3154                          * Although that time is observable, it is
3155                          * unpredictably variable.  Mash in the most
3156                          * volatile clock bits available, and expire the
3157                          * secret extra quickly.
3158                          */
3159                         if (unlikely(tcp_secret_primary->expires ==
3160                                      tcp_secret_secondary->expires)) {
3161                                 struct timespec tv;
3162
3163                                 getnstimeofday(&tv);
3164                                 bakery[COOKIE_DIGEST_WORDS+0] ^=
3165                                         (u32)tv.tv_nsec;
3166
3167                                 tcp_secret_secondary->expires = jiffy
3168                                         + TCP_SECRET_1MSL
3169                                         + (0x0f & tcp_cookie_work(bakery, 0));
3170                         } else {
3171                                 tcp_secret_secondary->expires = jiffy
3172                                         + TCP_SECRET_LIFE
3173                                         + (0xff & tcp_cookie_work(bakery, 1));
3174                                 tcp_secret_primary->expires = jiffy
3175                                         + TCP_SECRET_2MSL
3176                                         + (0x1f & tcp_cookie_work(bakery, 2));
3177                         }
3178                         memcpy(&tcp_secret_secondary->secrets[0],
3179                                bakery, COOKIE_WORKSPACE_WORDS);
3180
3181                         rcu_assign_pointer(tcp_secret_generating,
3182                                            tcp_secret_secondary);
3183                         rcu_assign_pointer(tcp_secret_retiring,
3184                                            tcp_secret_primary);
3185                         /*
3186                          * Neither call_rcu() nor synchronize_rcu() needed.
3187                          * Retiring data is not freed.  It is replaced after
3188                          * further (locked) pointer updates, and a quiet time
3189                          * (minimum 1MSL, maximum LIFE - 2MSL).
3190                          */
3191                 }
3192                 spin_unlock_bh(&tcp_secret_locker);
3193         } else {
3194                 rcu_read_lock_bh();
3195                 memcpy(bakery,
3196                        &rcu_dereference(tcp_secret_generating)->secrets[0],
3197                        COOKIE_WORKSPACE_WORDS);
3198                 rcu_read_unlock_bh();
3199         }
3200         return 0;
3201 }
3202 EXPORT_SYMBOL(tcp_cookie_generator);
3203
3204 void tcp_done(struct sock *sk)
3205 {
3206         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3207                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3208
3209         tcp_set_state(sk, TCP_CLOSE);
3210         tcp_clear_xmit_timers(sk);
3211
3212         sk->sk_shutdown = SHUTDOWN_MASK;
3213
3214         if (!sock_flag(sk, SOCK_DEAD))
3215                 sk->sk_state_change(sk);
3216         else
3217                 inet_csk_destroy_sock(sk);
3218 }
3219 EXPORT_SYMBOL_GPL(tcp_done);
3220
3221 extern struct tcp_congestion_ops tcp_reno;
3222
3223 static __initdata unsigned long thash_entries;
3224 static int __init set_thash_entries(char *str)
3225 {
3226         if (!str)
3227                 return 0;
3228         thash_entries = simple_strtoul(str, &str, 0);
3229         return 1;
3230 }
3231 __setup("thash_entries=", set_thash_entries);
3232
3233 void tcp_init_mem(struct net *net)
3234 {
3235         unsigned long limit = nr_free_buffer_pages() / 8;
3236         limit = max(limit, 128UL);
3237         net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3238         net->ipv4.sysctl_tcp_mem[1] = limit;
3239         net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3240 }
3241
3242 void __init tcp_init(void)
3243 {
3244         struct sk_buff *skb = NULL;
3245         unsigned long limit;
3246         int max_share, cnt;
3247         unsigned int i;
3248         unsigned long jiffy = jiffies;
3249
3250         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3251
3252         percpu_counter_init(&tcp_sockets_allocated, 0);
3253         percpu_counter_init(&tcp_orphan_count, 0);
3254         tcp_hashinfo.bind_bucket_cachep =
3255                 kmem_cache_create("tcp_bind_bucket",
3256                                   sizeof(struct inet_bind_bucket), 0,
3257                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3258
3259         /* Size and allocate the main established and bind bucket
3260          * hash tables.
3261          *
3262          * The methodology is similar to that of the buffer cache.
3263          */
3264         tcp_hashinfo.ehash =
3265                 alloc_large_system_hash("TCP established",
3266                                         sizeof(struct inet_ehash_bucket),
3267                                         thash_entries,
3268                                         (totalram_pages >= 128 * 1024) ?
3269                                         13 : 15,
3270                                         0,
3271                                         NULL,
3272                                         &tcp_hashinfo.ehash_mask,
3273                                         thash_entries ? 0 : 512 * 1024);
3274         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3275                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3276                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3277         }
3278         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3279                 panic("TCP: failed to alloc ehash_locks");
3280         tcp_hashinfo.bhash =
3281                 alloc_large_system_hash("TCP bind",
3282                                         sizeof(struct inet_bind_hashbucket),
3283                                         tcp_hashinfo.ehash_mask + 1,
3284                                         (totalram_pages >= 128 * 1024) ?
3285                                         13 : 15,
3286                                         0,
3287                                         &tcp_hashinfo.bhash_size,
3288                                         NULL,
3289                                         64 * 1024);
3290         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3291         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3292                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3293                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3294         }
3295
3296
3297         cnt = tcp_hashinfo.ehash_mask + 1;
3298
3299         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3300         sysctl_tcp_max_orphans = cnt / 2;
3301         sysctl_max_syn_backlog = max(128, cnt / 256);
3302
3303         tcp_init_mem(&init_net);
3304         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3305         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3306         max_share = min(4UL*1024*1024, limit);
3307
3308         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3309         sysctl_tcp_wmem[1] = 16*1024;
3310         sysctl_tcp_wmem[2] = max(64*1024, max_share);
3311
3312         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3313         sysctl_tcp_rmem[1] = 87380;
3314         sysctl_tcp_rmem[2] = max(87380, max_share);
3315
3316         pr_info("Hash tables configured (established %u bind %u)\n",
3317                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3318
3319         tcp_register_congestion_control(&tcp_reno);
3320
3321         memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3322         memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3323         tcp_secret_one.expires = jiffy; /* past due */
3324         tcp_secret_two.expires = jiffy; /* past due */
3325         tcp_secret_generating = &tcp_secret_one;
3326         tcp_secret_primary = &tcp_secret_one;
3327         tcp_secret_retiring = &tcp_secret_two;
3328         tcp_secret_secondary = &tcp_secret_two;
3329 }