net: fix possible wrong checksum generation
[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/inet_common.h>
274 #include <net/tcp.h>
275 #include <net/xfrm.h>
276 #include <net/ip.h>
277 #include <net/netdma.h>
278 #include <net/sock.h>
279
280 #include <asm/uaccess.h>
281 #include <asm/ioctls.h>
282
283 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
284
285 struct percpu_counter tcp_orphan_count;
286 EXPORT_SYMBOL_GPL(tcp_orphan_count);
287
288 int sysctl_tcp_wmem[3] __read_mostly;
289 int sysctl_tcp_rmem[3] __read_mostly;
290
291 EXPORT_SYMBOL(sysctl_tcp_rmem);
292 EXPORT_SYMBOL(sysctl_tcp_wmem);
293
294 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
295 EXPORT_SYMBOL(tcp_memory_allocated);
296
297 /*
298  * Current number of TCP sockets.
299  */
300 struct percpu_counter tcp_sockets_allocated;
301 EXPORT_SYMBOL(tcp_sockets_allocated);
302
303 /*
304  * TCP splice context
305  */
306 struct tcp_splice_state {
307         struct pipe_inode_info *pipe;
308         size_t len;
309         unsigned int flags;
310 };
311
312 /*
313  * Pressure flag: try to collapse.
314  * Technical note: it is used by multiple contexts non atomically.
315  * All the __sk_mem_schedule() is of this nature: accounting
316  * is strict, actions are advisory and have some latency.
317  */
318 int tcp_memory_pressure __read_mostly;
319 EXPORT_SYMBOL(tcp_memory_pressure);
320
321 void tcp_enter_memory_pressure(struct sock *sk)
322 {
323         if (!tcp_memory_pressure) {
324                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
325                 tcp_memory_pressure = 1;
326         }
327 }
328 EXPORT_SYMBOL(tcp_enter_memory_pressure);
329
330 /* Convert seconds to retransmits based on initial and max timeout */
331 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
332 {
333         u8 res = 0;
334
335         if (seconds > 0) {
336                 int period = timeout;
337
338                 res = 1;
339                 while (seconds > period && res < 255) {
340                         res++;
341                         timeout <<= 1;
342                         if (timeout > rto_max)
343                                 timeout = rto_max;
344                         period += timeout;
345                 }
346         }
347         return res;
348 }
349
350 /* Convert retransmits to seconds based on initial and max timeout */
351 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
352 {
353         int period = 0;
354
355         if (retrans > 0) {
356                 period = timeout;
357                 while (--retrans) {
358                         timeout <<= 1;
359                         if (timeout > rto_max)
360                                 timeout = rto_max;
361                         period += timeout;
362                 }
363         }
364         return period;
365 }
366
367 /* Address-family independent initialization for a tcp_sock.
368  *
369  * NOTE: A lot of things set to zero explicitly by call to
370  *       sk_alloc() so need not be done here.
371  */
372 void tcp_init_sock(struct sock *sk)
373 {
374         struct inet_connection_sock *icsk = inet_csk(sk);
375         struct tcp_sock *tp = tcp_sk(sk);
376
377         skb_queue_head_init(&tp->out_of_order_queue);
378         tcp_init_xmit_timers(sk);
379         tcp_prequeue_init(tp);
380         INIT_LIST_HEAD(&tp->tsq_node);
381
382         icsk->icsk_rto = TCP_TIMEOUT_INIT;
383         tp->mdev = TCP_TIMEOUT_INIT;
384
385         /* So many TCP implementations out there (incorrectly) count the
386          * initial SYN frame in their delayed-ACK and congestion control
387          * algorithms that we must have the following bandaid to talk
388          * efficiently to them.  -DaveM
389          */
390         tp->snd_cwnd = TCP_INIT_CWND;
391
392         /* See draft-stevens-tcpca-spec-01 for discussion of the
393          * initialization of these values.
394          */
395         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
396         tp->snd_cwnd_clamp = ~0;
397         tp->mss_cache = TCP_MSS_DEFAULT;
398
399         tp->reordering = sysctl_tcp_reordering;
400         tcp_enable_early_retrans(tp);
401         icsk->icsk_ca_ops = &tcp_init_congestion_ops;
402
403         sk->sk_state = TCP_CLOSE;
404
405         sk->sk_write_space = sk_stream_write_space;
406         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
407
408         icsk->icsk_sync_mss = tcp_sync_mss;
409
410         /* TCP Cookie Transactions */
411         if (sysctl_tcp_cookie_size > 0) {
412                 /* Default, cookies without s_data_payload. */
413                 tp->cookie_values =
414                         kzalloc(sizeof(*tp->cookie_values),
415                                 sk->sk_allocation);
416                 if (tp->cookie_values != NULL)
417                         kref_init(&tp->cookie_values->kref);
418         }
419         /* Presumed zeroed, in order of appearance:
420          *      cookie_in_always, cookie_out_never,
421          *      s_data_constant, s_data_in, s_data_out
422          */
423         sk->sk_sndbuf = sysctl_tcp_wmem[1];
424         sk->sk_rcvbuf = sysctl_tcp_rmem[1];
425
426         local_bh_disable();
427         sock_update_memcg(sk);
428         sk_sockets_allocated_inc(sk);
429         local_bh_enable();
430 }
431 EXPORT_SYMBOL(tcp_init_sock);
432
433 /*
434  *      Wait for a TCP event.
435  *
436  *      Note that we don't need to lock the socket, as the upper poll layers
437  *      take care of normal races (between the test and the event) and we don't
438  *      go look at any of the socket buffers directly.
439  */
440 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
441 {
442         unsigned int mask;
443         struct sock *sk = sock->sk;
444         const struct tcp_sock *tp = tcp_sk(sk);
445
446         sock_poll_wait(file, sk_sleep(sk), wait);
447         if (sk->sk_state == TCP_LISTEN)
448                 return inet_csk_listen_poll(sk);
449
450         /* Socket is not locked. We are protected from async events
451          * by poll logic and correct handling of state changes
452          * made by other threads is impossible in any case.
453          */
454
455         mask = 0;
456
457         /*
458          * POLLHUP is certainly not done right. But poll() doesn't
459          * have a notion of HUP in just one direction, and for a
460          * socket the read side is more interesting.
461          *
462          * Some poll() documentation says that POLLHUP is incompatible
463          * with the POLLOUT/POLLWR flags, so somebody should check this
464          * all. But careful, it tends to be safer to return too many
465          * bits than too few, and you can easily break real applications
466          * if you don't tell them that something has hung up!
467          *
468          * Check-me.
469          *
470          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
471          * our fs/select.c). It means that after we received EOF,
472          * poll always returns immediately, making impossible poll() on write()
473          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
474          * if and only if shutdown has been made in both directions.
475          * Actually, it is interesting to look how Solaris and DUX
476          * solve this dilemma. I would prefer, if POLLHUP were maskable,
477          * then we could set it on SND_SHUTDOWN. BTW examples given
478          * in Stevens' books assume exactly this behaviour, it explains
479          * why POLLHUP is incompatible with POLLOUT.    --ANK
480          *
481          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
482          * blocking on fresh not-connected or disconnected socket. --ANK
483          */
484         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
485                 mask |= POLLHUP;
486         if (sk->sk_shutdown & RCV_SHUTDOWN)
487                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
488
489         /* Connected or passive Fast Open socket? */
490         if (sk->sk_state != TCP_SYN_SENT &&
491             (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
492                 int target = sock_rcvlowat(sk, 0, INT_MAX);
493
494                 if (tp->urg_seq == tp->copied_seq &&
495                     !sock_flag(sk, SOCK_URGINLINE) &&
496                     tp->urg_data)
497                         target++;
498
499                 /* Potential race condition. If read of tp below will
500                  * escape above sk->sk_state, we can be illegally awaken
501                  * in SYN_* states. */
502                 if (tp->rcv_nxt - tp->copied_seq >= target)
503                         mask |= POLLIN | POLLRDNORM;
504
505                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
506                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
507                                 mask |= POLLOUT | POLLWRNORM;
508                         } else {  /* send SIGIO later */
509                                 set_bit(SOCK_ASYNC_NOSPACE,
510                                         &sk->sk_socket->flags);
511                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
512
513                                 /* Race breaker. If space is freed after
514                                  * wspace test but before the flags are set,
515                                  * IO signal will be lost.
516                                  */
517                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
518                                         mask |= POLLOUT | POLLWRNORM;
519                         }
520                 } else
521                         mask |= POLLOUT | POLLWRNORM;
522
523                 if (tp->urg_data & TCP_URG_VALID)
524                         mask |= POLLPRI;
525         }
526         /* This barrier is coupled with smp_wmb() in tcp_reset() */
527         smp_rmb();
528         if (sk->sk_err)
529                 mask |= POLLERR;
530
531         return mask;
532 }
533 EXPORT_SYMBOL(tcp_poll);
534
535 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
536 {
537         struct tcp_sock *tp = tcp_sk(sk);
538         int answ;
539         bool slow;
540
541         switch (cmd) {
542         case SIOCINQ:
543                 if (sk->sk_state == TCP_LISTEN)
544                         return -EINVAL;
545
546                 slow = lock_sock_fast(sk);
547                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
548                         answ = 0;
549                 else if (sock_flag(sk, SOCK_URGINLINE) ||
550                          !tp->urg_data ||
551                          before(tp->urg_seq, tp->copied_seq) ||
552                          !before(tp->urg_seq, tp->rcv_nxt)) {
553
554                         answ = tp->rcv_nxt - tp->copied_seq;
555
556                         /* Subtract 1, if FIN was received */
557                         if (answ && sock_flag(sk, SOCK_DONE))
558                                 answ--;
559                 } else
560                         answ = tp->urg_seq - tp->copied_seq;
561                 unlock_sock_fast(sk, slow);
562                 break;
563         case SIOCATMARK:
564                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
565                 break;
566         case SIOCOUTQ:
567                 if (sk->sk_state == TCP_LISTEN)
568                         return -EINVAL;
569
570                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
571                         answ = 0;
572                 else
573                         answ = tp->write_seq - tp->snd_una;
574                 break;
575         case SIOCOUTQNSD:
576                 if (sk->sk_state == TCP_LISTEN)
577                         return -EINVAL;
578
579                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
580                         answ = 0;
581                 else
582                         answ = tp->write_seq - tp->snd_nxt;
583                 break;
584         default:
585                 return -ENOIOCTLCMD;
586         }
587
588         return put_user(answ, (int __user *)arg);
589 }
590 EXPORT_SYMBOL(tcp_ioctl);
591
592 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
593 {
594         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
595         tp->pushed_seq = tp->write_seq;
596 }
597
598 static inline bool forced_push(const struct tcp_sock *tp)
599 {
600         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
601 }
602
603 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
604 {
605         struct tcp_sock *tp = tcp_sk(sk);
606         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
607
608         skb->csum    = 0;
609         tcb->seq     = tcb->end_seq = tp->write_seq;
610         tcb->tcp_flags = TCPHDR_ACK;
611         tcb->sacked  = 0;
612         skb_header_release(skb);
613         tcp_add_write_queue_tail(sk, skb);
614         sk->sk_wmem_queued += skb->truesize;
615         sk_mem_charge(sk, skb->truesize);
616         if (tp->nonagle & TCP_NAGLE_PUSH)
617                 tp->nonagle &= ~TCP_NAGLE_PUSH;
618 }
619
620 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
621 {
622         if (flags & MSG_OOB)
623                 tp->snd_up = tp->write_seq;
624 }
625
626 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
627                             int nonagle)
628 {
629         if (tcp_send_head(sk)) {
630                 struct tcp_sock *tp = tcp_sk(sk);
631
632                 if (!(flags & MSG_MORE) || forced_push(tp))
633                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
634
635                 tcp_mark_urg(tp, flags);
636                 __tcp_push_pending_frames(sk, mss_now,
637                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
638         }
639 }
640
641 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
642                                 unsigned int offset, size_t len)
643 {
644         struct tcp_splice_state *tss = rd_desc->arg.data;
645         int ret;
646
647         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
648                               tss->flags);
649         if (ret > 0)
650                 rd_desc->count -= ret;
651         return ret;
652 }
653
654 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
655 {
656         /* Store TCP splice context information in read_descriptor_t. */
657         read_descriptor_t rd_desc = {
658                 .arg.data = tss,
659                 .count    = tss->len,
660         };
661
662         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
663 }
664
665 /**
666  *  tcp_splice_read - splice data from TCP socket to a pipe
667  * @sock:       socket to splice from
668  * @ppos:       position (not valid)
669  * @pipe:       pipe to splice to
670  * @len:        number of bytes to splice
671  * @flags:      splice modifier flags
672  *
673  * Description:
674  *    Will read pages from given socket and fill them into a pipe.
675  *
676  **/
677 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
678                         struct pipe_inode_info *pipe, size_t len,
679                         unsigned int flags)
680 {
681         struct sock *sk = sock->sk;
682         struct tcp_splice_state tss = {
683                 .pipe = pipe,
684                 .len = len,
685                 .flags = flags,
686         };
687         long timeo;
688         ssize_t spliced;
689         int ret;
690
691         sock_rps_record_flow(sk);
692         /*
693          * We can't seek on a socket input
694          */
695         if (unlikely(*ppos))
696                 return -ESPIPE;
697
698         ret = spliced = 0;
699
700         lock_sock(sk);
701
702         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
703         while (tss.len) {
704                 ret = __tcp_splice_read(sk, &tss);
705                 if (ret < 0)
706                         break;
707                 else if (!ret) {
708                         if (spliced)
709                                 break;
710                         if (sock_flag(sk, SOCK_DONE))
711                                 break;
712                         if (sk->sk_err) {
713                                 ret = sock_error(sk);
714                                 break;
715                         }
716                         if (sk->sk_shutdown & RCV_SHUTDOWN)
717                                 break;
718                         if (sk->sk_state == TCP_CLOSE) {
719                                 /*
720                                  * This occurs when user tries to read
721                                  * from never connected socket.
722                                  */
723                                 if (!sock_flag(sk, SOCK_DONE))
724                                         ret = -ENOTCONN;
725                                 break;
726                         }
727                         if (!timeo) {
728                                 ret = -EAGAIN;
729                                 break;
730                         }
731                         sk_wait_data(sk, &timeo);
732                         if (signal_pending(current)) {
733                                 ret = sock_intr_errno(timeo);
734                                 break;
735                         }
736                         continue;
737                 }
738                 tss.len -= ret;
739                 spliced += ret;
740
741                 if (!timeo)
742                         break;
743                 release_sock(sk);
744                 lock_sock(sk);
745
746                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
747                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
748                     signal_pending(current))
749                         break;
750         }
751
752         release_sock(sk);
753
754         if (spliced)
755                 return spliced;
756
757         return ret;
758 }
759 EXPORT_SYMBOL(tcp_splice_read);
760
761 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
762 {
763         struct sk_buff *skb;
764
765         /* The TCP header must be at least 32-bit aligned.  */
766         size = ALIGN(size, 4);
767
768         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
769         if (skb) {
770                 if (sk_wmem_schedule(sk, skb->truesize)) {
771                         skb_reserve(skb, sk->sk_prot->max_header);
772                         /*
773                          * Make sure that we have exactly size bytes
774                          * available to the caller, no more, no less.
775                          */
776                         skb->avail_size = size;
777                         return skb;
778                 }
779                 __kfree_skb(skb);
780         } else {
781                 sk->sk_prot->enter_memory_pressure(sk);
782                 sk_stream_moderate_sndbuf(sk);
783         }
784         return NULL;
785 }
786
787 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
788                                        int large_allowed)
789 {
790         struct tcp_sock *tp = tcp_sk(sk);
791         u32 xmit_size_goal, old_size_goal;
792
793         xmit_size_goal = mss_now;
794
795         if (large_allowed && sk_can_gso(sk)) {
796                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
797                                   inet_csk(sk)->icsk_af_ops->net_header_len -
798                                   inet_csk(sk)->icsk_ext_hdr_len -
799                                   tp->tcp_header_len);
800
801                 /* TSQ : try to have two TSO segments in flight */
802                 xmit_size_goal = min_t(u32, xmit_size_goal,
803                                        sysctl_tcp_limit_output_bytes >> 1);
804
805                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
806
807                 /* We try hard to avoid divides here */
808                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
809
810                 if (likely(old_size_goal <= xmit_size_goal &&
811                            old_size_goal + mss_now > xmit_size_goal)) {
812                         xmit_size_goal = old_size_goal;
813                 } else {
814                         tp->xmit_size_goal_segs =
815                                 min_t(u16, xmit_size_goal / mss_now,
816                                       sk->sk_gso_max_segs);
817                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
818                 }
819         }
820
821         return max(xmit_size_goal, mss_now);
822 }
823
824 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
825 {
826         int mss_now;
827
828         mss_now = tcp_current_mss(sk);
829         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
830
831         return mss_now;
832 }
833
834 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
835                                 size_t size, int flags)
836 {
837         struct tcp_sock *tp = tcp_sk(sk);
838         int mss_now, size_goal;
839         int err;
840         ssize_t copied;
841         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
842
843         /* Wait for a connection to finish. One exception is TCP Fast Open
844          * (passive side) where data is allowed to be sent before a connection
845          * is fully established.
846          */
847         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
848             !tcp_passive_fastopen(sk)) {
849                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
850                         goto out_err;
851         }
852
853         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
854
855         mss_now = tcp_send_mss(sk, &size_goal, flags);
856         copied = 0;
857
858         err = -EPIPE;
859         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
860                 goto out_err;
861
862         while (size > 0) {
863                 struct sk_buff *skb = tcp_write_queue_tail(sk);
864                 int copy, i;
865                 bool can_coalesce;
866
867                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
868 new_segment:
869                         if (!sk_stream_memory_free(sk))
870                                 goto wait_for_sndbuf;
871
872                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
873                         if (!skb)
874                                 goto wait_for_memory;
875
876                         skb_entail(sk, skb);
877                         copy = size_goal;
878                 }
879
880                 if (copy > size)
881                         copy = size;
882
883                 i = skb_shinfo(skb)->nr_frags;
884                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
885                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
886                         tcp_mark_push(tp, skb);
887                         goto new_segment;
888                 }
889                 if (!sk_wmem_schedule(sk, copy))
890                         goto wait_for_memory;
891
892                 if (can_coalesce) {
893                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
894                 } else {
895                         get_page(page);
896                         skb_fill_page_desc(skb, i, page, offset, copy);
897                 }
898
899                 skb_shinfo(skb)->gso_type |= SKB_GSO_SHARED_FRAG;
900
901                 skb->len += copy;
902                 skb->data_len += copy;
903                 skb->truesize += copy;
904                 sk->sk_wmem_queued += copy;
905                 sk_mem_charge(sk, copy);
906                 skb->ip_summed = CHECKSUM_PARTIAL;
907                 tp->write_seq += copy;
908                 TCP_SKB_CB(skb)->end_seq += copy;
909                 skb_shinfo(skb)->gso_segs = 0;
910
911                 if (!copied)
912                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
913
914                 copied += copy;
915                 offset += copy;
916                 if (!(size -= copy))
917                         goto out;
918
919                 if (skb->len < size_goal || (flags & MSG_OOB))
920                         continue;
921
922                 if (forced_push(tp)) {
923                         tcp_mark_push(tp, skb);
924                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
925                 } else if (skb == tcp_send_head(sk))
926                         tcp_push_one(sk, mss_now);
927                 continue;
928
929 wait_for_sndbuf:
930                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
931 wait_for_memory:
932                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
933
934                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
935                         goto do_error;
936
937                 mss_now = tcp_send_mss(sk, &size_goal, flags);
938         }
939
940 out:
941         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
942                 tcp_push(sk, flags, mss_now, tp->nonagle);
943         return copied;
944
945 do_error:
946         if (copied)
947                 goto out;
948 out_err:
949         return sk_stream_error(sk, flags, err);
950 }
951
952 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
953                  size_t size, int flags)
954 {
955         ssize_t res;
956
957         if (!(sk->sk_route_caps & NETIF_F_SG) ||
958             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
959                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
960                                         flags);
961
962         lock_sock(sk);
963         res = do_tcp_sendpages(sk, page, offset, size, flags);
964         release_sock(sk);
965         return res;
966 }
967 EXPORT_SYMBOL(tcp_sendpage);
968
969 static inline int select_size(const struct sock *sk, bool sg)
970 {
971         const struct tcp_sock *tp = tcp_sk(sk);
972         int tmp = tp->mss_cache;
973
974         if (sg) {
975                 if (sk_can_gso(sk)) {
976                         /* Small frames wont use a full page:
977                          * Payload will immediately follow tcp header.
978                          */
979                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
980                 } else {
981                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
982
983                         if (tmp >= pgbreak &&
984                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
985                                 tmp = pgbreak;
986                 }
987         }
988
989         return tmp;
990 }
991
992 void tcp_free_fastopen_req(struct tcp_sock *tp)
993 {
994         if (tp->fastopen_req != NULL) {
995                 kfree(tp->fastopen_req);
996                 tp->fastopen_req = NULL;
997         }
998 }
999
1000 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size)
1001 {
1002         struct tcp_sock *tp = tcp_sk(sk);
1003         int err, flags;
1004
1005         if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1006                 return -EOPNOTSUPP;
1007         if (tp->fastopen_req != NULL)
1008                 return -EALREADY; /* Another Fast Open is in progress */
1009
1010         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1011                                    sk->sk_allocation);
1012         if (unlikely(tp->fastopen_req == NULL))
1013                 return -ENOBUFS;
1014         tp->fastopen_req->data = msg;
1015
1016         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1017         err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1018                                     msg->msg_namelen, flags);
1019         *size = tp->fastopen_req->copied;
1020         tcp_free_fastopen_req(tp);
1021         return err;
1022 }
1023
1024 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1025                 size_t size)
1026 {
1027         struct iovec *iov;
1028         struct tcp_sock *tp = tcp_sk(sk);
1029         struct sk_buff *skb;
1030         int iovlen, flags, err, copied = 0;
1031         int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1032         bool sg;
1033         long timeo;
1034
1035         lock_sock(sk);
1036
1037         flags = msg->msg_flags;
1038         if (flags & MSG_FASTOPEN) {
1039                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn);
1040                 if (err == -EINPROGRESS && copied_syn > 0)
1041                         goto out;
1042                 else if (err)
1043                         goto out_err;
1044                 offset = copied_syn;
1045         }
1046
1047         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1048
1049         /* Wait for a connection to finish. One exception is TCP Fast Open
1050          * (passive side) where data is allowed to be sent before a connection
1051          * is fully established.
1052          */
1053         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1054             !tcp_passive_fastopen(sk)) {
1055                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1056                         goto do_error;
1057         }
1058
1059         if (unlikely(tp->repair)) {
1060                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1061                         copied = tcp_send_rcvq(sk, msg, size);
1062                         goto out;
1063                 }
1064
1065                 err = -EINVAL;
1066                 if (tp->repair_queue == TCP_NO_QUEUE)
1067                         goto out_err;
1068
1069                 /* 'common' sending to sendq */
1070         }
1071
1072         /* This should be in poll */
1073         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1074
1075         mss_now = tcp_send_mss(sk, &size_goal, flags);
1076
1077         /* Ok commence sending. */
1078         iovlen = msg->msg_iovlen;
1079         iov = msg->msg_iov;
1080         copied = 0;
1081
1082         err = -EPIPE;
1083         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1084                 goto out_err;
1085
1086         sg = !!(sk->sk_route_caps & NETIF_F_SG);
1087
1088         while (--iovlen >= 0) {
1089                 size_t seglen = iov->iov_len;
1090                 unsigned char __user *from = iov->iov_base;
1091
1092                 iov++;
1093                 if (unlikely(offset > 0)) {  /* Skip bytes copied in SYN */
1094                         if (offset >= seglen) {
1095                                 offset -= seglen;
1096                                 continue;
1097                         }
1098                         seglen -= offset;
1099                         from += offset;
1100                         offset = 0;
1101                 }
1102
1103                 while (seglen > 0) {
1104                         int copy = 0;
1105                         int max = size_goal;
1106
1107                         skb = tcp_write_queue_tail(sk);
1108                         if (tcp_send_head(sk)) {
1109                                 if (skb->ip_summed == CHECKSUM_NONE)
1110                                         max = mss_now;
1111                                 copy = max - skb->len;
1112                         }
1113
1114                         if (copy <= 0) {
1115 new_segment:
1116                                 /* Allocate new segment. If the interface is SG,
1117                                  * allocate skb fitting to single page.
1118                                  */
1119                                 if (!sk_stream_memory_free(sk))
1120                                         goto wait_for_sndbuf;
1121
1122                                 skb = sk_stream_alloc_skb(sk,
1123                                                           select_size(sk, sg),
1124                                                           sk->sk_allocation);
1125                                 if (!skb)
1126                                         goto wait_for_memory;
1127
1128                                 /*
1129                                  * Check whether we can use HW checksum.
1130                                  */
1131                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1132                                         skb->ip_summed = CHECKSUM_PARTIAL;
1133
1134                                 skb_entail(sk, skb);
1135                                 copy = size_goal;
1136                                 max = size_goal;
1137                         }
1138
1139                         /* Try to append data to the end of skb. */
1140                         if (copy > seglen)
1141                                 copy = seglen;
1142
1143                         /* Where to copy to? */
1144                         if (skb_availroom(skb) > 0) {
1145                                 /* We have some space in skb head. Superb! */
1146                                 copy = min_t(int, copy, skb_availroom(skb));
1147                                 err = skb_add_data_nocache(sk, skb, from, copy);
1148                                 if (err)
1149                                         goto do_fault;
1150                         } else {
1151                                 bool merge = true;
1152                                 int i = skb_shinfo(skb)->nr_frags;
1153                                 struct page_frag *pfrag = sk_page_frag(sk);
1154
1155                                 if (!sk_page_frag_refill(sk, pfrag))
1156                                         goto wait_for_memory;
1157
1158                                 if (!skb_can_coalesce(skb, i, pfrag->page,
1159                                                       pfrag->offset)) {
1160                                         if (i == MAX_SKB_FRAGS || !sg) {
1161                                                 tcp_mark_push(tp, skb);
1162                                                 goto new_segment;
1163                                         }
1164                                         merge = false;
1165                                 }
1166
1167                                 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1168
1169                                 if (!sk_wmem_schedule(sk, copy))
1170                                         goto wait_for_memory;
1171
1172                                 err = skb_copy_to_page_nocache(sk, from, skb,
1173                                                                pfrag->page,
1174                                                                pfrag->offset,
1175                                                                copy);
1176                                 if (err)
1177                                         goto do_error;
1178
1179                                 /* Update the skb. */
1180                                 if (merge) {
1181                                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1182                                 } else {
1183                                         skb_fill_page_desc(skb, i, pfrag->page,
1184                                                            pfrag->offset, copy);
1185                                         get_page(pfrag->page);
1186                                 }
1187                                 pfrag->offset += copy;
1188                         }
1189
1190                         if (!copied)
1191                                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1192
1193                         tp->write_seq += copy;
1194                         TCP_SKB_CB(skb)->end_seq += copy;
1195                         skb_shinfo(skb)->gso_segs = 0;
1196
1197                         from += copy;
1198                         copied += copy;
1199                         if ((seglen -= copy) == 0 && iovlen == 0)
1200                                 goto out;
1201
1202                         if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1203                                 continue;
1204
1205                         if (forced_push(tp)) {
1206                                 tcp_mark_push(tp, skb);
1207                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1208                         } else if (skb == tcp_send_head(sk))
1209                                 tcp_push_one(sk, mss_now);
1210                         continue;
1211
1212 wait_for_sndbuf:
1213                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1214 wait_for_memory:
1215                         if (copied)
1216                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1217
1218                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1219                                 goto do_error;
1220
1221                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1222                 }
1223         }
1224
1225 out:
1226         if (copied)
1227                 tcp_push(sk, flags, mss_now, tp->nonagle);
1228         release_sock(sk);
1229         return copied + copied_syn;
1230
1231 do_fault:
1232         if (!skb->len) {
1233                 tcp_unlink_write_queue(skb, sk);
1234                 /* It is the one place in all of TCP, except connection
1235                  * reset, where we can be unlinking the send_head.
1236                  */
1237                 tcp_check_send_head(sk, skb);
1238                 sk_wmem_free_skb(sk, skb);
1239         }
1240
1241 do_error:
1242         if (copied + copied_syn)
1243                 goto out;
1244 out_err:
1245         err = sk_stream_error(sk, flags, err);
1246         release_sock(sk);
1247         return err;
1248 }
1249 EXPORT_SYMBOL(tcp_sendmsg);
1250
1251 /*
1252  *      Handle reading urgent data. BSD has very simple semantics for
1253  *      this, no blocking and very strange errors 8)
1254  */
1255
1256 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1257 {
1258         struct tcp_sock *tp = tcp_sk(sk);
1259
1260         /* No URG data to read. */
1261         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1262             tp->urg_data == TCP_URG_READ)
1263                 return -EINVAL; /* Yes this is right ! */
1264
1265         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1266                 return -ENOTCONN;
1267
1268         if (tp->urg_data & TCP_URG_VALID) {
1269                 int err = 0;
1270                 char c = tp->urg_data;
1271
1272                 if (!(flags & MSG_PEEK))
1273                         tp->urg_data = TCP_URG_READ;
1274
1275                 /* Read urgent data. */
1276                 msg->msg_flags |= MSG_OOB;
1277
1278                 if (len > 0) {
1279                         if (!(flags & MSG_TRUNC))
1280                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1281                         len = 1;
1282                 } else
1283                         msg->msg_flags |= MSG_TRUNC;
1284
1285                 return err ? -EFAULT : len;
1286         }
1287
1288         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1289                 return 0;
1290
1291         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1292          * the available implementations agree in this case:
1293          * this call should never block, independent of the
1294          * blocking state of the socket.
1295          * Mike <pall@rz.uni-karlsruhe.de>
1296          */
1297         return -EAGAIN;
1298 }
1299
1300 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1301 {
1302         struct sk_buff *skb;
1303         int copied = 0, err = 0;
1304
1305         /* XXX -- need to support SO_PEEK_OFF */
1306
1307         skb_queue_walk(&sk->sk_write_queue, skb) {
1308                 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1309                 if (err)
1310                         break;
1311
1312                 copied += skb->len;
1313         }
1314
1315         return err ?: copied;
1316 }
1317
1318 /* Clean up the receive buffer for full frames taken by the user,
1319  * then send an ACK if necessary.  COPIED is the number of bytes
1320  * tcp_recvmsg has given to the user so far, it speeds up the
1321  * calculation of whether or not we must ACK for the sake of
1322  * a window update.
1323  */
1324 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1325 {
1326         struct tcp_sock *tp = tcp_sk(sk);
1327         bool time_to_ack = false;
1328
1329         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1330
1331         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1332              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1333              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1334
1335         if (inet_csk_ack_scheduled(sk)) {
1336                 const struct inet_connection_sock *icsk = inet_csk(sk);
1337                    /* Delayed ACKs frequently hit locked sockets during bulk
1338                     * receive. */
1339                 if (icsk->icsk_ack.blocked ||
1340                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1341                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1342                     /*
1343                      * If this read emptied read buffer, we send ACK, if
1344                      * connection is not bidirectional, user drained
1345                      * receive buffer and there was a small segment
1346                      * in queue.
1347                      */
1348                     (copied > 0 &&
1349                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1350                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1351                        !icsk->icsk_ack.pingpong)) &&
1352                       !atomic_read(&sk->sk_rmem_alloc)))
1353                         time_to_ack = true;
1354         }
1355
1356         /* We send an ACK if we can now advertise a non-zero window
1357          * which has been raised "significantly".
1358          *
1359          * Even if window raised up to infinity, do not send window open ACK
1360          * in states, where we will not receive more. It is useless.
1361          */
1362         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1363                 __u32 rcv_window_now = tcp_receive_window(tp);
1364
1365                 /* Optimize, __tcp_select_window() is not cheap. */
1366                 if (2*rcv_window_now <= tp->window_clamp) {
1367                         __u32 new_window = __tcp_select_window(sk);
1368
1369                         /* Send ACK now, if this read freed lots of space
1370                          * in our buffer. Certainly, new_window is new window.
1371                          * We can advertise it now, if it is not less than current one.
1372                          * "Lots" means "at least twice" here.
1373                          */
1374                         if (new_window && new_window >= 2 * rcv_window_now)
1375                                 time_to_ack = true;
1376                 }
1377         }
1378         if (time_to_ack)
1379                 tcp_send_ack(sk);
1380 }
1381
1382 static void tcp_prequeue_process(struct sock *sk)
1383 {
1384         struct sk_buff *skb;
1385         struct tcp_sock *tp = tcp_sk(sk);
1386
1387         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1388
1389         /* RX process wants to run with disabled BHs, though it is not
1390          * necessary */
1391         local_bh_disable();
1392         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1393                 sk_backlog_rcv(sk, skb);
1394         local_bh_enable();
1395
1396         /* Clear memory counter. */
1397         tp->ucopy.memory = 0;
1398 }
1399
1400 #ifdef CONFIG_NET_DMA
1401 static void tcp_service_net_dma(struct sock *sk, bool wait)
1402 {
1403         dma_cookie_t done, used;
1404         dma_cookie_t last_issued;
1405         struct tcp_sock *tp = tcp_sk(sk);
1406
1407         if (!tp->ucopy.dma_chan)
1408                 return;
1409
1410         last_issued = tp->ucopy.dma_cookie;
1411         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1412
1413         do {
1414                 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1415                                               last_issued, &done,
1416                                               &used) == DMA_SUCCESS) {
1417                         /* Safe to free early-copied skbs now */
1418                         __skb_queue_purge(&sk->sk_async_wait_queue);
1419                         break;
1420                 } else {
1421                         struct sk_buff *skb;
1422                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1423                                (dma_async_is_complete(skb->dma_cookie, done,
1424                                                       used) == DMA_SUCCESS)) {
1425                                 __skb_dequeue(&sk->sk_async_wait_queue);
1426                                 kfree_skb(skb);
1427                         }
1428                 }
1429         } while (wait);
1430 }
1431 #endif
1432
1433 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1434 {
1435         struct sk_buff *skb;
1436         u32 offset;
1437
1438         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1439                 offset = seq - TCP_SKB_CB(skb)->seq;
1440                 if (tcp_hdr(skb)->syn)
1441                         offset--;
1442                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1443                         *off = offset;
1444                         return skb;
1445                 }
1446                 /* This looks weird, but this can happen if TCP collapsing
1447                  * splitted a fat GRO packet, while we released socket lock
1448                  * in skb_splice_bits()
1449                  */
1450                 sk_eat_skb(sk, skb, false);
1451         }
1452         return NULL;
1453 }
1454
1455 /*
1456  * This routine provides an alternative to tcp_recvmsg() for routines
1457  * that would like to handle copying from skbuffs directly in 'sendfile'
1458  * fashion.
1459  * Note:
1460  *      - It is assumed that the socket was locked by the caller.
1461  *      - The routine does not block.
1462  *      - At present, there is no support for reading OOB data
1463  *        or for 'peeking' the socket using this routine
1464  *        (although both would be easy to implement).
1465  */
1466 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1467                   sk_read_actor_t recv_actor)
1468 {
1469         struct sk_buff *skb;
1470         struct tcp_sock *tp = tcp_sk(sk);
1471         u32 seq = tp->copied_seq;
1472         u32 offset;
1473         int copied = 0;
1474
1475         if (sk->sk_state == TCP_LISTEN)
1476                 return -ENOTCONN;
1477         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1478                 if (offset < skb->len) {
1479                         int used;
1480                         size_t len;
1481
1482                         len = skb->len - offset;
1483                         /* Stop reading if we hit a patch of urgent data */
1484                         if (tp->urg_data) {
1485                                 u32 urg_offset = tp->urg_seq - seq;
1486                                 if (urg_offset < len)
1487                                         len = urg_offset;
1488                                 if (!len)
1489                                         break;
1490                         }
1491                         used = recv_actor(desc, skb, offset, len);
1492                         if (used <= 0) {
1493                                 if (!copied)
1494                                         copied = used;
1495                                 break;
1496                         } else if (used <= len) {
1497                                 seq += used;
1498                                 copied += used;
1499                                 offset += used;
1500                         }
1501                         /* If recv_actor drops the lock (e.g. TCP splice
1502                          * receive) the skb pointer might be invalid when
1503                          * getting here: tcp_collapse might have deleted it
1504                          * while aggregating skbs from the socket queue.
1505                          */
1506                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1507                         if (!skb)
1508                                 break;
1509                         /* TCP coalescing might have appended data to the skb.
1510                          * Try to splice more frags
1511                          */
1512                         if (offset + 1 != skb->len)
1513                                 continue;
1514                 }
1515                 if (tcp_hdr(skb)->fin) {
1516                         sk_eat_skb(sk, skb, false);
1517                         ++seq;
1518                         break;
1519                 }
1520                 sk_eat_skb(sk, skb, false);
1521                 if (!desc->count)
1522                         break;
1523                 tp->copied_seq = seq;
1524         }
1525         tp->copied_seq = seq;
1526
1527         tcp_rcv_space_adjust(sk);
1528
1529         /* Clean up data we have read: This will do ACK frames. */
1530         if (copied > 0) {
1531                 tcp_recv_skb(sk, seq, &offset);
1532                 tcp_cleanup_rbuf(sk, copied);
1533         }
1534         return copied;
1535 }
1536 EXPORT_SYMBOL(tcp_read_sock);
1537
1538 /*
1539  *      This routine copies from a sock struct into the user buffer.
1540  *
1541  *      Technical note: in 2.3 we work on _locked_ socket, so that
1542  *      tricks with *seq access order and skb->users are not required.
1543  *      Probably, code can be easily improved even more.
1544  */
1545
1546 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1547                 size_t len, int nonblock, int flags, int *addr_len)
1548 {
1549         struct tcp_sock *tp = tcp_sk(sk);
1550         int copied = 0;
1551         u32 peek_seq;
1552         u32 *seq;
1553         unsigned long used;
1554         int err;
1555         int target;             /* Read at least this many bytes */
1556         long timeo;
1557         struct task_struct *user_recv = NULL;
1558         bool copied_early = false;
1559         struct sk_buff *skb;
1560         u32 urg_hole = 0;
1561
1562         lock_sock(sk);
1563
1564         err = -ENOTCONN;
1565         if (sk->sk_state == TCP_LISTEN)
1566                 goto out;
1567
1568         timeo = sock_rcvtimeo(sk, nonblock);
1569
1570         /* Urgent data needs to be handled specially. */
1571         if (flags & MSG_OOB)
1572                 goto recv_urg;
1573
1574         if (unlikely(tp->repair)) {
1575                 err = -EPERM;
1576                 if (!(flags & MSG_PEEK))
1577                         goto out;
1578
1579                 if (tp->repair_queue == TCP_SEND_QUEUE)
1580                         goto recv_sndq;
1581
1582                 err = -EINVAL;
1583                 if (tp->repair_queue == TCP_NO_QUEUE)
1584                         goto out;
1585
1586                 /* 'common' recv queue MSG_PEEK-ing */
1587         }
1588
1589         seq = &tp->copied_seq;
1590         if (flags & MSG_PEEK) {
1591                 peek_seq = tp->copied_seq;
1592                 seq = &peek_seq;
1593         }
1594
1595         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1596
1597 #ifdef CONFIG_NET_DMA
1598         tp->ucopy.dma_chan = NULL;
1599         preempt_disable();
1600         skb = skb_peek_tail(&sk->sk_receive_queue);
1601         {
1602                 int available = 0;
1603
1604                 if (skb)
1605                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1606                 if ((available < target) &&
1607                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1608                     !sysctl_tcp_low_latency &&
1609                     net_dma_find_channel()) {
1610                         preempt_enable_no_resched();
1611                         tp->ucopy.pinned_list =
1612                                         dma_pin_iovec_pages(msg->msg_iov, len);
1613                 } else {
1614                         preempt_enable_no_resched();
1615                 }
1616         }
1617 #endif
1618
1619         do {
1620                 u32 offset;
1621
1622                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1623                 if (tp->urg_data && tp->urg_seq == *seq) {
1624                         if (copied)
1625                                 break;
1626                         if (signal_pending(current)) {
1627                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1628                                 break;
1629                         }
1630                 }
1631
1632                 /* Next get a buffer. */
1633
1634                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1635                         /* Now that we have two receive queues this
1636                          * shouldn't happen.
1637                          */
1638                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1639                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1640                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1641                                  flags))
1642                                 break;
1643
1644                         offset = *seq - TCP_SKB_CB(skb)->seq;
1645                         if (tcp_hdr(skb)->syn)
1646                                 offset--;
1647                         if (offset < skb->len)
1648                                 goto found_ok_skb;
1649                         if (tcp_hdr(skb)->fin)
1650                                 goto found_fin_ok;
1651                         WARN(!(flags & MSG_PEEK),
1652                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1653                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1654                 }
1655
1656                 /* Well, if we have backlog, try to process it now yet. */
1657
1658                 if (copied >= target && !sk->sk_backlog.tail)
1659                         break;
1660
1661                 if (copied) {
1662                         if (sk->sk_err ||
1663                             sk->sk_state == TCP_CLOSE ||
1664                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1665                             !timeo ||
1666                             signal_pending(current))
1667                                 break;
1668                 } else {
1669                         if (sock_flag(sk, SOCK_DONE))
1670                                 break;
1671
1672                         if (sk->sk_err) {
1673                                 copied = sock_error(sk);
1674                                 break;
1675                         }
1676
1677                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1678                                 break;
1679
1680                         if (sk->sk_state == TCP_CLOSE) {
1681                                 if (!sock_flag(sk, SOCK_DONE)) {
1682                                         /* This occurs when user tries to read
1683                                          * from never connected socket.
1684                                          */
1685                                         copied = -ENOTCONN;
1686                                         break;
1687                                 }
1688                                 break;
1689                         }
1690
1691                         if (!timeo) {
1692                                 copied = -EAGAIN;
1693                                 break;
1694                         }
1695
1696                         if (signal_pending(current)) {
1697                                 copied = sock_intr_errno(timeo);
1698                                 break;
1699                         }
1700                 }
1701
1702                 tcp_cleanup_rbuf(sk, copied);
1703
1704                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1705                         /* Install new reader */
1706                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1707                                 user_recv = current;
1708                                 tp->ucopy.task = user_recv;
1709                                 tp->ucopy.iov = msg->msg_iov;
1710                         }
1711
1712                         tp->ucopy.len = len;
1713
1714                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1715                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1716
1717                         /* Ugly... If prequeue is not empty, we have to
1718                          * process it before releasing socket, otherwise
1719                          * order will be broken at second iteration.
1720                          * More elegant solution is required!!!
1721                          *
1722                          * Look: we have the following (pseudo)queues:
1723                          *
1724                          * 1. packets in flight
1725                          * 2. backlog
1726                          * 3. prequeue
1727                          * 4. receive_queue
1728                          *
1729                          * Each queue can be processed only if the next ones
1730                          * are empty. At this point we have empty receive_queue.
1731                          * But prequeue _can_ be not empty after 2nd iteration,
1732                          * when we jumped to start of loop because backlog
1733                          * processing added something to receive_queue.
1734                          * We cannot release_sock(), because backlog contains
1735                          * packets arrived _after_ prequeued ones.
1736                          *
1737                          * Shortly, algorithm is clear --- to process all
1738                          * the queues in order. We could make it more directly,
1739                          * requeueing packets from backlog to prequeue, if
1740                          * is not empty. It is more elegant, but eats cycles,
1741                          * unfortunately.
1742                          */
1743                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1744                                 goto do_prequeue;
1745
1746                         /* __ Set realtime policy in scheduler __ */
1747                 }
1748
1749 #ifdef CONFIG_NET_DMA
1750                 if (tp->ucopy.dma_chan) {
1751                         if (tp->rcv_wnd == 0 &&
1752                             !skb_queue_empty(&sk->sk_async_wait_queue)) {
1753                                 tcp_service_net_dma(sk, true);
1754                                 tcp_cleanup_rbuf(sk, copied);
1755                         } else
1756                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1757                 }
1758 #endif
1759                 if (copied >= target) {
1760                         /* Do not sleep, just process backlog. */
1761                         release_sock(sk);
1762                         lock_sock(sk);
1763                 } else
1764                         sk_wait_data(sk, &timeo);
1765
1766 #ifdef CONFIG_NET_DMA
1767                 tcp_service_net_dma(sk, false);  /* Don't block */
1768                 tp->ucopy.wakeup = 0;
1769 #endif
1770
1771                 if (user_recv) {
1772                         int chunk;
1773
1774                         /* __ Restore normal policy in scheduler __ */
1775
1776                         if ((chunk = len - tp->ucopy.len) != 0) {
1777                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1778                                 len -= chunk;
1779                                 copied += chunk;
1780                         }
1781
1782                         if (tp->rcv_nxt == tp->copied_seq &&
1783                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1784 do_prequeue:
1785                                 tcp_prequeue_process(sk);
1786
1787                                 if ((chunk = len - tp->ucopy.len) != 0) {
1788                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1789                                         len -= chunk;
1790                                         copied += chunk;
1791                                 }
1792                         }
1793                 }
1794                 if ((flags & MSG_PEEK) &&
1795                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1796                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1797                                             current->comm,
1798                                             task_pid_nr(current));
1799                         peek_seq = tp->copied_seq;
1800                 }
1801                 continue;
1802
1803         found_ok_skb:
1804                 /* Ok so how much can we use? */
1805                 used = skb->len - offset;
1806                 if (len < used)
1807                         used = len;
1808
1809                 /* Do we have urgent data here? */
1810                 if (tp->urg_data) {
1811                         u32 urg_offset = tp->urg_seq - *seq;
1812                         if (urg_offset < used) {
1813                                 if (!urg_offset) {
1814                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1815                                                 ++*seq;
1816                                                 urg_hole++;
1817                                                 offset++;
1818                                                 used--;
1819                                                 if (!used)
1820                                                         goto skip_copy;
1821                                         }
1822                                 } else
1823                                         used = urg_offset;
1824                         }
1825                 }
1826
1827                 if (!(flags & MSG_TRUNC)) {
1828 #ifdef CONFIG_NET_DMA
1829                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1830                                 tp->ucopy.dma_chan = net_dma_find_channel();
1831
1832                         if (tp->ucopy.dma_chan) {
1833                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1834                                         tp->ucopy.dma_chan, skb, offset,
1835                                         msg->msg_iov, used,
1836                                         tp->ucopy.pinned_list);
1837
1838                                 if (tp->ucopy.dma_cookie < 0) {
1839
1840                                         pr_alert("%s: dma_cookie < 0\n",
1841                                                  __func__);
1842
1843                                         /* Exception. Bailout! */
1844                                         if (!copied)
1845                                                 copied = -EFAULT;
1846                                         break;
1847                                 }
1848
1849                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1850
1851                                 if ((offset + used) == skb->len)
1852                                         copied_early = true;
1853
1854                         } else
1855 #endif
1856                         {
1857                                 err = skb_copy_datagram_iovec(skb, offset,
1858                                                 msg->msg_iov, used);
1859                                 if (err) {
1860                                         /* Exception. Bailout! */
1861                                         if (!copied)
1862                                                 copied = -EFAULT;
1863                                         break;
1864                                 }
1865                         }
1866                 }
1867
1868                 *seq += used;
1869                 copied += used;
1870                 len -= used;
1871
1872                 tcp_rcv_space_adjust(sk);
1873
1874 skip_copy:
1875                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1876                         tp->urg_data = 0;
1877                         tcp_fast_path_check(sk);
1878                 }
1879                 if (used + offset < skb->len)
1880                         continue;
1881
1882                 if (tcp_hdr(skb)->fin)
1883                         goto found_fin_ok;
1884                 if (!(flags & MSG_PEEK)) {
1885                         sk_eat_skb(sk, skb, copied_early);
1886                         copied_early = false;
1887                 }
1888                 continue;
1889
1890         found_fin_ok:
1891                 /* Process the FIN. */
1892                 ++*seq;
1893                 if (!(flags & MSG_PEEK)) {
1894                         sk_eat_skb(sk, skb, copied_early);
1895                         copied_early = false;
1896                 }
1897                 break;
1898         } while (len > 0);
1899
1900         if (user_recv) {
1901                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1902                         int chunk;
1903
1904                         tp->ucopy.len = copied > 0 ? len : 0;
1905
1906                         tcp_prequeue_process(sk);
1907
1908                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1909                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1910                                 len -= chunk;
1911                                 copied += chunk;
1912                         }
1913                 }
1914
1915                 tp->ucopy.task = NULL;
1916                 tp->ucopy.len = 0;
1917         }
1918
1919 #ifdef CONFIG_NET_DMA
1920         tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1921         tp->ucopy.dma_chan = NULL;
1922
1923         if (tp->ucopy.pinned_list) {
1924                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1925                 tp->ucopy.pinned_list = NULL;
1926         }
1927 #endif
1928
1929         /* According to UNIX98, msg_name/msg_namelen are ignored
1930          * on connected socket. I was just happy when found this 8) --ANK
1931          */
1932
1933         /* Clean up data we have read: This will do ACK frames. */
1934         tcp_cleanup_rbuf(sk, copied);
1935
1936         release_sock(sk);
1937         return copied;
1938
1939 out:
1940         release_sock(sk);
1941         return err;
1942
1943 recv_urg:
1944         err = tcp_recv_urg(sk, msg, len, flags);
1945         goto out;
1946
1947 recv_sndq:
1948         err = tcp_peek_sndq(sk, msg, len);
1949         goto out;
1950 }
1951 EXPORT_SYMBOL(tcp_recvmsg);
1952
1953 void tcp_set_state(struct sock *sk, int state)
1954 {
1955         int oldstate = sk->sk_state;
1956
1957         switch (state) {
1958         case TCP_ESTABLISHED:
1959                 if (oldstate != TCP_ESTABLISHED)
1960                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1961                 break;
1962
1963         case TCP_CLOSE:
1964                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1965                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1966
1967                 sk->sk_prot->unhash(sk);
1968                 if (inet_csk(sk)->icsk_bind_hash &&
1969                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1970                         inet_put_port(sk);
1971                 /* fall through */
1972         default:
1973                 if (oldstate == TCP_ESTABLISHED)
1974                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1975         }
1976
1977         /* Change state AFTER socket is unhashed to avoid closed
1978          * socket sitting in hash tables.
1979          */
1980         sk->sk_state = state;
1981
1982 #ifdef STATE_TRACE
1983         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1984 #endif
1985 }
1986 EXPORT_SYMBOL_GPL(tcp_set_state);
1987
1988 /*
1989  *      State processing on a close. This implements the state shift for
1990  *      sending our FIN frame. Note that we only send a FIN for some
1991  *      states. A shutdown() may have already sent the FIN, or we may be
1992  *      closed.
1993  */
1994
1995 static const unsigned char new_state[16] = {
1996   /* current state:        new state:      action:      */
1997   /* (Invalid)          */ TCP_CLOSE,
1998   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1999   /* TCP_SYN_SENT       */ TCP_CLOSE,
2000   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2001   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
2002   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
2003   /* TCP_TIME_WAIT      */ TCP_CLOSE,
2004   /* TCP_CLOSE          */ TCP_CLOSE,
2005   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
2006   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
2007   /* TCP_LISTEN         */ TCP_CLOSE,
2008   /* TCP_CLOSING        */ TCP_CLOSING,
2009 };
2010
2011 static int tcp_close_state(struct sock *sk)
2012 {
2013         int next = (int)new_state[sk->sk_state];
2014         int ns = next & TCP_STATE_MASK;
2015
2016         tcp_set_state(sk, ns);
2017
2018         return next & TCP_ACTION_FIN;
2019 }
2020
2021 /*
2022  *      Shutdown the sending side of a connection. Much like close except
2023  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2024  */
2025
2026 void tcp_shutdown(struct sock *sk, int how)
2027 {
2028         /*      We need to grab some memory, and put together a FIN,
2029          *      and then put it into the queue to be sent.
2030          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2031          */
2032         if (!(how & SEND_SHUTDOWN))
2033                 return;
2034
2035         /* If we've already sent a FIN, or it's a closed state, skip this. */
2036         if ((1 << sk->sk_state) &
2037             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2038              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2039                 /* Clear out any half completed packets.  FIN if needed. */
2040                 if (tcp_close_state(sk))
2041                         tcp_send_fin(sk);
2042         }
2043 }
2044 EXPORT_SYMBOL(tcp_shutdown);
2045
2046 bool tcp_check_oom(struct sock *sk, int shift)
2047 {
2048         bool too_many_orphans, out_of_socket_memory;
2049
2050         too_many_orphans = tcp_too_many_orphans(sk, shift);
2051         out_of_socket_memory = tcp_out_of_memory(sk);
2052
2053         if (too_many_orphans)
2054                 net_info_ratelimited("too many orphaned sockets\n");
2055         if (out_of_socket_memory)
2056                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2057         return too_many_orphans || out_of_socket_memory;
2058 }
2059
2060 void tcp_close(struct sock *sk, long timeout)
2061 {
2062         struct sk_buff *skb;
2063         int data_was_unread = 0;
2064         int state;
2065
2066         lock_sock(sk);
2067         sk->sk_shutdown = SHUTDOWN_MASK;
2068
2069         if (sk->sk_state == TCP_LISTEN) {
2070                 tcp_set_state(sk, TCP_CLOSE);
2071
2072                 /* Special case. */
2073                 inet_csk_listen_stop(sk);
2074
2075                 goto adjudge_to_death;
2076         }
2077
2078         /*  We need to flush the recv. buffs.  We do this only on the
2079          *  descriptor close, not protocol-sourced closes, because the
2080          *  reader process may not have drained the data yet!
2081          */
2082         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2083                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2084                           tcp_hdr(skb)->fin;
2085                 data_was_unread += len;
2086                 __kfree_skb(skb);
2087         }
2088
2089         sk_mem_reclaim(sk);
2090
2091         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2092         if (sk->sk_state == TCP_CLOSE)
2093                 goto adjudge_to_death;
2094
2095         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2096          * data was lost. To witness the awful effects of the old behavior of
2097          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2098          * GET in an FTP client, suspend the process, wait for the client to
2099          * advertise a zero window, then kill -9 the FTP client, wheee...
2100          * Note: timeout is always zero in such a case.
2101          */
2102         if (unlikely(tcp_sk(sk)->repair)) {
2103                 sk->sk_prot->disconnect(sk, 0);
2104         } else if (data_was_unread) {
2105                 /* Unread data was tossed, zap the connection. */
2106                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2107                 tcp_set_state(sk, TCP_CLOSE);
2108                 tcp_send_active_reset(sk, sk->sk_allocation);
2109         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2110                 /* Check zero linger _after_ checking for unread data. */
2111                 sk->sk_prot->disconnect(sk, 0);
2112                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2113         } else if (tcp_close_state(sk)) {
2114                 /* We FIN if the application ate all the data before
2115                  * zapping the connection.
2116                  */
2117
2118                 /* RED-PEN. Formally speaking, we have broken TCP state
2119                  * machine. State transitions:
2120                  *
2121                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2122                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2123                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2124                  *
2125                  * are legal only when FIN has been sent (i.e. in window),
2126                  * rather than queued out of window. Purists blame.
2127                  *
2128                  * F.e. "RFC state" is ESTABLISHED,
2129                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2130                  *
2131                  * The visible declinations are that sometimes
2132                  * we enter time-wait state, when it is not required really
2133                  * (harmless), do not send active resets, when they are
2134                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2135                  * they look as CLOSING or LAST_ACK for Linux)
2136                  * Probably, I missed some more holelets.
2137                  *                                              --ANK
2138                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2139                  * in a single packet! (May consider it later but will
2140                  * probably need API support or TCP_CORK SYN-ACK until
2141                  * data is written and socket is closed.)
2142                  */
2143                 tcp_send_fin(sk);
2144         }
2145
2146         sk_stream_wait_close(sk, timeout);
2147
2148 adjudge_to_death:
2149         state = sk->sk_state;
2150         sock_hold(sk);
2151         sock_orphan(sk);
2152
2153         /* It is the last release_sock in its life. It will remove backlog. */
2154         release_sock(sk);
2155
2156
2157         /* Now socket is owned by kernel and we acquire BH lock
2158            to finish close. No need to check for user refs.
2159          */
2160         local_bh_disable();
2161         bh_lock_sock(sk);
2162         WARN_ON(sock_owned_by_user(sk));
2163
2164         percpu_counter_inc(sk->sk_prot->orphan_count);
2165
2166         /* Have we already been destroyed by a softirq or backlog? */
2167         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2168                 goto out;
2169
2170         /*      This is a (useful) BSD violating of the RFC. There is a
2171          *      problem with TCP as specified in that the other end could
2172          *      keep a socket open forever with no application left this end.
2173          *      We use a 3 minute timeout (about the same as BSD) then kill
2174          *      our end. If they send after that then tough - BUT: long enough
2175          *      that we won't make the old 4*rto = almost no time - whoops
2176          *      reset mistake.
2177          *
2178          *      Nope, it was not mistake. It is really desired behaviour
2179          *      f.e. on http servers, when such sockets are useless, but
2180          *      consume significant resources. Let's do it with special
2181          *      linger2 option.                                 --ANK
2182          */
2183
2184         if (sk->sk_state == TCP_FIN_WAIT2) {
2185                 struct tcp_sock *tp = tcp_sk(sk);
2186                 if (tp->linger2 < 0) {
2187                         tcp_set_state(sk, TCP_CLOSE);
2188                         tcp_send_active_reset(sk, GFP_ATOMIC);
2189                         NET_INC_STATS_BH(sock_net(sk),
2190                                         LINUX_MIB_TCPABORTONLINGER);
2191                 } else {
2192                         const int tmo = tcp_fin_time(sk);
2193
2194                         if (tmo > TCP_TIMEWAIT_LEN) {
2195                                 inet_csk_reset_keepalive_timer(sk,
2196                                                 tmo - TCP_TIMEWAIT_LEN);
2197                         } else {
2198                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2199                                 goto out;
2200                         }
2201                 }
2202         }
2203         if (sk->sk_state != TCP_CLOSE) {
2204                 sk_mem_reclaim(sk);
2205                 if (tcp_check_oom(sk, 0)) {
2206                         tcp_set_state(sk, TCP_CLOSE);
2207                         tcp_send_active_reset(sk, GFP_ATOMIC);
2208                         NET_INC_STATS_BH(sock_net(sk),
2209                                         LINUX_MIB_TCPABORTONMEMORY);
2210                 }
2211         }
2212
2213         if (sk->sk_state == TCP_CLOSE) {
2214                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2215                 /* We could get here with a non-NULL req if the socket is
2216                  * aborted (e.g., closed with unread data) before 3WHS
2217                  * finishes.
2218                  */
2219                 if (req != NULL)
2220                         reqsk_fastopen_remove(sk, req, false);
2221                 inet_csk_destroy_sock(sk);
2222         }
2223         /* Otherwise, socket is reprieved until protocol close. */
2224
2225 out:
2226         bh_unlock_sock(sk);
2227         local_bh_enable();
2228         sock_put(sk);
2229 }
2230 EXPORT_SYMBOL(tcp_close);
2231
2232 /* These states need RST on ABORT according to RFC793 */
2233
2234 static inline bool tcp_need_reset(int state)
2235 {
2236         return (1 << state) &
2237                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2238                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2239 }
2240
2241 int tcp_disconnect(struct sock *sk, int flags)
2242 {
2243         struct inet_sock *inet = inet_sk(sk);
2244         struct inet_connection_sock *icsk = inet_csk(sk);
2245         struct tcp_sock *tp = tcp_sk(sk);
2246         int err = 0;
2247         int old_state = sk->sk_state;
2248
2249         if (old_state != TCP_CLOSE)
2250                 tcp_set_state(sk, TCP_CLOSE);
2251
2252         /* ABORT function of RFC793 */
2253         if (old_state == TCP_LISTEN) {
2254                 inet_csk_listen_stop(sk);
2255         } else if (unlikely(tp->repair)) {
2256                 sk->sk_err = ECONNABORTED;
2257         } else if (tcp_need_reset(old_state) ||
2258                    (tp->snd_nxt != tp->write_seq &&
2259                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2260                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2261                  * states
2262                  */
2263                 tcp_send_active_reset(sk, gfp_any());
2264                 sk->sk_err = ECONNRESET;
2265         } else if (old_state == TCP_SYN_SENT)
2266                 sk->sk_err = ECONNRESET;
2267
2268         tcp_clear_xmit_timers(sk);
2269         __skb_queue_purge(&sk->sk_receive_queue);
2270         tcp_write_queue_purge(sk);
2271         __skb_queue_purge(&tp->out_of_order_queue);
2272 #ifdef CONFIG_NET_DMA
2273         __skb_queue_purge(&sk->sk_async_wait_queue);
2274 #endif
2275
2276         inet->inet_dport = 0;
2277
2278         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2279                 inet_reset_saddr(sk);
2280
2281         sk->sk_shutdown = 0;
2282         sock_reset_flag(sk, SOCK_DONE);
2283         tp->srtt = 0;
2284         if ((tp->write_seq += tp->max_window + 2) == 0)
2285                 tp->write_seq = 1;
2286         icsk->icsk_backoff = 0;
2287         tp->snd_cwnd = 2;
2288         icsk->icsk_probes_out = 0;
2289         tp->packets_out = 0;
2290         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2291         tp->snd_cwnd_cnt = 0;
2292         tp->bytes_acked = 0;
2293         tp->window_clamp = 0;
2294         tcp_set_ca_state(sk, TCP_CA_Open);
2295         tcp_clear_retrans(tp);
2296         inet_csk_delack_init(sk);
2297         tcp_init_send_head(sk);
2298         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2299         __sk_dst_reset(sk);
2300
2301         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2302
2303         sk->sk_error_report(sk);
2304         return err;
2305 }
2306 EXPORT_SYMBOL(tcp_disconnect);
2307
2308 void tcp_sock_destruct(struct sock *sk)
2309 {
2310         inet_sock_destruct(sk);
2311
2312         kfree(inet_csk(sk)->icsk_accept_queue.fastopenq);
2313 }
2314
2315 static inline bool tcp_can_repair_sock(const struct sock *sk)
2316 {
2317         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2318                 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2319 }
2320
2321 static int tcp_repair_options_est(struct tcp_sock *tp,
2322                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2323 {
2324         struct tcp_repair_opt opt;
2325
2326         while (len >= sizeof(opt)) {
2327                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2328                         return -EFAULT;
2329
2330                 optbuf++;
2331                 len -= sizeof(opt);
2332
2333                 switch (opt.opt_code) {
2334                 case TCPOPT_MSS:
2335                         tp->rx_opt.mss_clamp = opt.opt_val;
2336                         break;
2337                 case TCPOPT_WINDOW:
2338                         {
2339                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2340                                 u16 rcv_wscale = opt.opt_val >> 16;
2341
2342                                 if (snd_wscale > 14 || rcv_wscale > 14)
2343                                         return -EFBIG;
2344
2345                                 tp->rx_opt.snd_wscale = snd_wscale;
2346                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2347                                 tp->rx_opt.wscale_ok = 1;
2348                         }
2349                         break;
2350                 case TCPOPT_SACK_PERM:
2351                         if (opt.opt_val != 0)
2352                                 return -EINVAL;
2353
2354                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2355                         if (sysctl_tcp_fack)
2356                                 tcp_enable_fack(tp);
2357                         break;
2358                 case TCPOPT_TIMESTAMP:
2359                         if (opt.opt_val != 0)
2360                                 return -EINVAL;
2361
2362                         tp->rx_opt.tstamp_ok = 1;
2363                         break;
2364                 }
2365         }
2366
2367         return 0;
2368 }
2369
2370 /*
2371  *      Socket option code for TCP.
2372  */
2373 static int do_tcp_setsockopt(struct sock *sk, int level,
2374                 int optname, char __user *optval, unsigned int optlen)
2375 {
2376         struct tcp_sock *tp = tcp_sk(sk);
2377         struct inet_connection_sock *icsk = inet_csk(sk);
2378         int val;
2379         int err = 0;
2380
2381         /* These are data/string values, all the others are ints */
2382         switch (optname) {
2383         case TCP_CONGESTION: {
2384                 char name[TCP_CA_NAME_MAX];
2385
2386                 if (optlen < 1)
2387                         return -EINVAL;
2388
2389                 val = strncpy_from_user(name, optval,
2390                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2391                 if (val < 0)
2392                         return -EFAULT;
2393                 name[val] = 0;
2394
2395                 lock_sock(sk);
2396                 err = tcp_set_congestion_control(sk, name);
2397                 release_sock(sk);
2398                 return err;
2399         }
2400         case TCP_COOKIE_TRANSACTIONS: {
2401                 struct tcp_cookie_transactions ctd;
2402                 struct tcp_cookie_values *cvp = NULL;
2403
2404                 if (sizeof(ctd) > optlen)
2405                         return -EINVAL;
2406                 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2407                         return -EFAULT;
2408
2409                 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2410                     ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2411                         return -EINVAL;
2412
2413                 if (ctd.tcpct_cookie_desired == 0) {
2414                         /* default to global value */
2415                 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2416                            ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2417                            ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2418                         return -EINVAL;
2419                 }
2420
2421                 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2422                         /* Supercedes all other values */
2423                         lock_sock(sk);
2424                         if (tp->cookie_values != NULL) {
2425                                 kref_put(&tp->cookie_values->kref,
2426                                          tcp_cookie_values_release);
2427                                 tp->cookie_values = NULL;
2428                         }
2429                         tp->rx_opt.cookie_in_always = 0; /* false */
2430                         tp->rx_opt.cookie_out_never = 1; /* true */
2431                         release_sock(sk);
2432                         return err;
2433                 }
2434
2435                 /* Allocate ancillary memory before locking.
2436                  */
2437                 if (ctd.tcpct_used > 0 ||
2438                     (tp->cookie_values == NULL &&
2439                      (sysctl_tcp_cookie_size > 0 ||
2440                       ctd.tcpct_cookie_desired > 0 ||
2441                       ctd.tcpct_s_data_desired > 0))) {
2442                         cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2443                                       GFP_KERNEL);
2444                         if (cvp == NULL)
2445                                 return -ENOMEM;
2446
2447                         kref_init(&cvp->kref);
2448                 }
2449                 lock_sock(sk);
2450                 tp->rx_opt.cookie_in_always =
2451                         (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2452                 tp->rx_opt.cookie_out_never = 0; /* false */
2453
2454                 if (tp->cookie_values != NULL) {
2455                         if (cvp != NULL) {
2456                                 /* Changed values are recorded by a changed
2457                                  * pointer, ensuring the cookie will differ,
2458                                  * without separately hashing each value later.
2459                                  */
2460                                 kref_put(&tp->cookie_values->kref,
2461                                          tcp_cookie_values_release);
2462                         } else {
2463                                 cvp = tp->cookie_values;
2464                         }
2465                 }
2466
2467                 if (cvp != NULL) {
2468                         cvp->cookie_desired = ctd.tcpct_cookie_desired;
2469
2470                         if (ctd.tcpct_used > 0) {
2471                                 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2472                                        ctd.tcpct_used);
2473                                 cvp->s_data_desired = ctd.tcpct_used;
2474                                 cvp->s_data_constant = 1; /* true */
2475                         } else {
2476                                 /* No constant payload data. */
2477                                 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2478                                 cvp->s_data_constant = 0; /* false */
2479                         }
2480
2481                         tp->cookie_values = cvp;
2482                 }
2483                 release_sock(sk);
2484                 return err;
2485         }
2486         default:
2487                 /* fallthru */
2488                 break;
2489         }
2490
2491         if (optlen < sizeof(int))
2492                 return -EINVAL;
2493
2494         if (get_user(val, (int __user *)optval))
2495                 return -EFAULT;
2496
2497         lock_sock(sk);
2498
2499         switch (optname) {
2500         case TCP_MAXSEG:
2501                 /* Values greater than interface MTU won't take effect. However
2502                  * at the point when this call is done we typically don't yet
2503                  * know which interface is going to be used */
2504                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2505                         err = -EINVAL;
2506                         break;
2507                 }
2508                 tp->rx_opt.user_mss = val;
2509                 break;
2510
2511         case TCP_NODELAY:
2512                 if (val) {
2513                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2514                          * this option on corked socket is remembered, but
2515                          * it is not activated until cork is cleared.
2516                          *
2517                          * However, when TCP_NODELAY is set we make
2518                          * an explicit push, which overrides even TCP_CORK
2519                          * for currently queued segments.
2520                          */
2521                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2522                         tcp_push_pending_frames(sk);
2523                 } else {
2524                         tp->nonagle &= ~TCP_NAGLE_OFF;
2525                 }
2526                 break;
2527
2528         case TCP_THIN_LINEAR_TIMEOUTS:
2529                 if (val < 0 || val > 1)
2530                         err = -EINVAL;
2531                 else
2532                         tp->thin_lto = val;
2533                 break;
2534
2535         case TCP_THIN_DUPACK:
2536                 if (val < 0 || val > 1)
2537                         err = -EINVAL;
2538                 else
2539                         tp->thin_dupack = val;
2540                         if (tp->thin_dupack)
2541                                 tcp_disable_early_retrans(tp);
2542                 break;
2543
2544         case TCP_REPAIR:
2545                 if (!tcp_can_repair_sock(sk))
2546                         err = -EPERM;
2547                 else if (val == 1) {
2548                         tp->repair = 1;
2549                         sk->sk_reuse = SK_FORCE_REUSE;
2550                         tp->repair_queue = TCP_NO_QUEUE;
2551                 } else if (val == 0) {
2552                         tp->repair = 0;
2553                         sk->sk_reuse = SK_NO_REUSE;
2554                         tcp_send_window_probe(sk);
2555                 } else
2556                         err = -EINVAL;
2557
2558                 break;
2559
2560         case TCP_REPAIR_QUEUE:
2561                 if (!tp->repair)
2562                         err = -EPERM;
2563                 else if (val < TCP_QUEUES_NR)
2564                         tp->repair_queue = val;
2565                 else
2566                         err = -EINVAL;
2567                 break;
2568
2569         case TCP_QUEUE_SEQ:
2570                 if (sk->sk_state != TCP_CLOSE)
2571                         err = -EPERM;
2572                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2573                         tp->write_seq = val;
2574                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2575                         tp->rcv_nxt = val;
2576                 else
2577                         err = -EINVAL;
2578                 break;
2579
2580         case TCP_REPAIR_OPTIONS:
2581                 if (!tp->repair)
2582                         err = -EINVAL;
2583                 else if (sk->sk_state == TCP_ESTABLISHED)
2584                         err = tcp_repair_options_est(tp,
2585                                         (struct tcp_repair_opt __user *)optval,
2586                                         optlen);
2587                 else
2588                         err = -EPERM;
2589                 break;
2590
2591         case TCP_CORK:
2592                 /* When set indicates to always queue non-full frames.
2593                  * Later the user clears this option and we transmit
2594                  * any pending partial frames in the queue.  This is
2595                  * meant to be used alongside sendfile() to get properly
2596                  * filled frames when the user (for example) must write
2597                  * out headers with a write() call first and then use
2598                  * sendfile to send out the data parts.
2599                  *
2600                  * TCP_CORK can be set together with TCP_NODELAY and it is
2601                  * stronger than TCP_NODELAY.
2602                  */
2603                 if (val) {
2604                         tp->nonagle |= TCP_NAGLE_CORK;
2605                 } else {
2606                         tp->nonagle &= ~TCP_NAGLE_CORK;
2607                         if (tp->nonagle&TCP_NAGLE_OFF)
2608                                 tp->nonagle |= TCP_NAGLE_PUSH;
2609                         tcp_push_pending_frames(sk);
2610                 }
2611                 break;
2612
2613         case TCP_KEEPIDLE:
2614                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2615                         err = -EINVAL;
2616                 else {
2617                         tp->keepalive_time = val * HZ;
2618                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2619                             !((1 << sk->sk_state) &
2620                               (TCPF_CLOSE | TCPF_LISTEN))) {
2621                                 u32 elapsed = keepalive_time_elapsed(tp);
2622                                 if (tp->keepalive_time > elapsed)
2623                                         elapsed = tp->keepalive_time - elapsed;
2624                                 else
2625                                         elapsed = 0;
2626                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2627                         }
2628                 }
2629                 break;
2630         case TCP_KEEPINTVL:
2631                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2632                         err = -EINVAL;
2633                 else
2634                         tp->keepalive_intvl = val * HZ;
2635                 break;
2636         case TCP_KEEPCNT:
2637                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2638                         err = -EINVAL;
2639                 else
2640                         tp->keepalive_probes = val;
2641                 break;
2642         case TCP_SYNCNT:
2643                 if (val < 1 || val > MAX_TCP_SYNCNT)
2644                         err = -EINVAL;
2645                 else
2646                         icsk->icsk_syn_retries = val;
2647                 break;
2648
2649         case TCP_LINGER2:
2650                 if (val < 0)
2651                         tp->linger2 = -1;
2652                 else if (val > sysctl_tcp_fin_timeout / HZ)
2653                         tp->linger2 = 0;
2654                 else
2655                         tp->linger2 = val * HZ;
2656                 break;
2657
2658         case TCP_DEFER_ACCEPT:
2659                 /* Translate value in seconds to number of retransmits */
2660                 icsk->icsk_accept_queue.rskq_defer_accept =
2661                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2662                                         TCP_RTO_MAX / HZ);
2663                 break;
2664
2665         case TCP_WINDOW_CLAMP:
2666                 if (!val) {
2667                         if (sk->sk_state != TCP_CLOSE) {
2668                                 err = -EINVAL;
2669                                 break;
2670                         }
2671                         tp->window_clamp = 0;
2672                 } else
2673                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2674                                                 SOCK_MIN_RCVBUF / 2 : val;
2675                 break;
2676
2677         case TCP_QUICKACK:
2678                 if (!val) {
2679                         icsk->icsk_ack.pingpong = 1;
2680                 } else {
2681                         icsk->icsk_ack.pingpong = 0;
2682                         if ((1 << sk->sk_state) &
2683                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2684                             inet_csk_ack_scheduled(sk)) {
2685                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2686                                 tcp_cleanup_rbuf(sk, 1);
2687                                 if (!(val & 1))
2688                                         icsk->icsk_ack.pingpong = 1;
2689                         }
2690                 }
2691                 break;
2692
2693 #ifdef CONFIG_TCP_MD5SIG
2694         case TCP_MD5SIG:
2695                 /* Read the IP->Key mappings from userspace */
2696                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2697                 break;
2698 #endif
2699         case TCP_USER_TIMEOUT:
2700                 /* Cap the max timeout in ms TCP will retry/retrans
2701                  * before giving up and aborting (ETIMEDOUT) a connection.
2702                  */
2703                 if (val < 0)
2704                         err = -EINVAL;
2705                 else
2706                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2707                 break;
2708
2709         case TCP_FASTOPEN:
2710                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2711                     TCPF_LISTEN)))
2712                         err = fastopen_init_queue(sk, val);
2713                 else
2714                         err = -EINVAL;
2715                 break;
2716         default:
2717                 err = -ENOPROTOOPT;
2718                 break;
2719         }
2720
2721         release_sock(sk);
2722         return err;
2723 }
2724
2725 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2726                    unsigned int optlen)
2727 {
2728         const struct inet_connection_sock *icsk = inet_csk(sk);
2729
2730         if (level != SOL_TCP)
2731                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2732                                                      optval, optlen);
2733         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2734 }
2735 EXPORT_SYMBOL(tcp_setsockopt);
2736
2737 #ifdef CONFIG_COMPAT
2738 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2739                           char __user *optval, unsigned int optlen)
2740 {
2741         if (level != SOL_TCP)
2742                 return inet_csk_compat_setsockopt(sk, level, optname,
2743                                                   optval, optlen);
2744         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2745 }
2746 EXPORT_SYMBOL(compat_tcp_setsockopt);
2747 #endif
2748
2749 /* Return information about state of tcp endpoint in API format. */
2750 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2751 {
2752         const struct tcp_sock *tp = tcp_sk(sk);
2753         const struct inet_connection_sock *icsk = inet_csk(sk);
2754         u32 now = tcp_time_stamp;
2755
2756         memset(info, 0, sizeof(*info));
2757
2758         info->tcpi_state = sk->sk_state;
2759         info->tcpi_ca_state = icsk->icsk_ca_state;
2760         info->tcpi_retransmits = icsk->icsk_retransmits;
2761         info->tcpi_probes = icsk->icsk_probes_out;
2762         info->tcpi_backoff = icsk->icsk_backoff;
2763
2764         if (tp->rx_opt.tstamp_ok)
2765                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2766         if (tcp_is_sack(tp))
2767                 info->tcpi_options |= TCPI_OPT_SACK;
2768         if (tp->rx_opt.wscale_ok) {
2769                 info->tcpi_options |= TCPI_OPT_WSCALE;
2770                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2771                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2772         }
2773
2774         if (tp->ecn_flags & TCP_ECN_OK)
2775                 info->tcpi_options |= TCPI_OPT_ECN;
2776         if (tp->ecn_flags & TCP_ECN_SEEN)
2777                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2778         if (tp->syn_data_acked)
2779                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2780
2781         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2782         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2783         info->tcpi_snd_mss = tp->mss_cache;
2784         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2785
2786         if (sk->sk_state == TCP_LISTEN) {
2787                 info->tcpi_unacked = sk->sk_ack_backlog;
2788                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2789         } else {
2790                 info->tcpi_unacked = tp->packets_out;
2791                 info->tcpi_sacked = tp->sacked_out;
2792         }
2793         info->tcpi_lost = tp->lost_out;
2794         info->tcpi_retrans = tp->retrans_out;
2795         info->tcpi_fackets = tp->fackets_out;
2796
2797         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2798         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2799         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2800
2801         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2802         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2803         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2804         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2805         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2806         info->tcpi_snd_cwnd = tp->snd_cwnd;
2807         info->tcpi_advmss = tp->advmss;
2808         info->tcpi_reordering = tp->reordering;
2809
2810         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2811         info->tcpi_rcv_space = tp->rcvq_space.space;
2812
2813         info->tcpi_total_retrans = tp->total_retrans;
2814 }
2815 EXPORT_SYMBOL_GPL(tcp_get_info);
2816
2817 static int do_tcp_getsockopt(struct sock *sk, int level,
2818                 int optname, char __user *optval, int __user *optlen)
2819 {
2820         struct inet_connection_sock *icsk = inet_csk(sk);
2821         struct tcp_sock *tp = tcp_sk(sk);
2822         int val, len;
2823
2824         if (get_user(len, optlen))
2825                 return -EFAULT;
2826
2827         len = min_t(unsigned int, len, sizeof(int));
2828
2829         if (len < 0)
2830                 return -EINVAL;
2831
2832         switch (optname) {
2833         case TCP_MAXSEG:
2834                 val = tp->mss_cache;
2835                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2836                         val = tp->rx_opt.user_mss;
2837                 if (tp->repair)
2838                         val = tp->rx_opt.mss_clamp;
2839                 break;
2840         case TCP_NODELAY:
2841                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2842                 break;
2843         case TCP_CORK:
2844                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2845                 break;
2846         case TCP_KEEPIDLE:
2847                 val = keepalive_time_when(tp) / HZ;
2848                 break;
2849         case TCP_KEEPINTVL:
2850                 val = keepalive_intvl_when(tp) / HZ;
2851                 break;
2852         case TCP_KEEPCNT:
2853                 val = keepalive_probes(tp);
2854                 break;
2855         case TCP_SYNCNT:
2856                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2857                 break;
2858         case TCP_LINGER2:
2859                 val = tp->linger2;
2860                 if (val >= 0)
2861                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2862                 break;
2863         case TCP_DEFER_ACCEPT:
2864                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2865                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2866                 break;
2867         case TCP_WINDOW_CLAMP:
2868                 val = tp->window_clamp;
2869                 break;
2870         case TCP_INFO: {
2871                 struct tcp_info info;
2872
2873                 if (get_user(len, optlen))
2874                         return -EFAULT;
2875
2876                 tcp_get_info(sk, &info);
2877
2878                 len = min_t(unsigned int, len, sizeof(info));
2879                 if (put_user(len, optlen))
2880                         return -EFAULT;
2881                 if (copy_to_user(optval, &info, len))
2882                         return -EFAULT;
2883                 return 0;
2884         }
2885         case TCP_QUICKACK:
2886                 val = !icsk->icsk_ack.pingpong;
2887                 break;
2888
2889         case TCP_CONGESTION:
2890                 if (get_user(len, optlen))
2891                         return -EFAULT;
2892                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2893                 if (put_user(len, optlen))
2894                         return -EFAULT;
2895                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2896                         return -EFAULT;
2897                 return 0;
2898
2899         case TCP_COOKIE_TRANSACTIONS: {
2900                 struct tcp_cookie_transactions ctd;
2901                 struct tcp_cookie_values *cvp = tp->cookie_values;
2902
2903                 if (get_user(len, optlen))
2904                         return -EFAULT;
2905                 if (len < sizeof(ctd))
2906                         return -EINVAL;
2907
2908                 memset(&ctd, 0, sizeof(ctd));
2909                 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2910                                    TCP_COOKIE_IN_ALWAYS : 0)
2911                                 | (tp->rx_opt.cookie_out_never ?
2912                                    TCP_COOKIE_OUT_NEVER : 0);
2913
2914                 if (cvp != NULL) {
2915                         ctd.tcpct_flags |= (cvp->s_data_in ?
2916                                             TCP_S_DATA_IN : 0)
2917                                          | (cvp->s_data_out ?
2918                                             TCP_S_DATA_OUT : 0);
2919
2920                         ctd.tcpct_cookie_desired = cvp->cookie_desired;
2921                         ctd.tcpct_s_data_desired = cvp->s_data_desired;
2922
2923                         memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2924                                cvp->cookie_pair_size);
2925                         ctd.tcpct_used = cvp->cookie_pair_size;
2926                 }
2927
2928                 if (put_user(sizeof(ctd), optlen))
2929                         return -EFAULT;
2930                 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2931                         return -EFAULT;
2932                 return 0;
2933         }
2934         case TCP_THIN_LINEAR_TIMEOUTS:
2935                 val = tp->thin_lto;
2936                 break;
2937         case TCP_THIN_DUPACK:
2938                 val = tp->thin_dupack;
2939                 break;
2940
2941         case TCP_REPAIR:
2942                 val = tp->repair;
2943                 break;
2944
2945         case TCP_REPAIR_QUEUE:
2946                 if (tp->repair)
2947                         val = tp->repair_queue;
2948                 else
2949                         return -EINVAL;
2950                 break;
2951
2952         case TCP_QUEUE_SEQ:
2953                 if (tp->repair_queue == TCP_SEND_QUEUE)
2954                         val = tp->write_seq;
2955                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2956                         val = tp->rcv_nxt;
2957                 else
2958                         return -EINVAL;
2959                 break;
2960
2961         case TCP_USER_TIMEOUT:
2962                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2963                 break;
2964         default:
2965                 return -ENOPROTOOPT;
2966         }
2967
2968         if (put_user(len, optlen))
2969                 return -EFAULT;
2970         if (copy_to_user(optval, &val, len))
2971                 return -EFAULT;
2972         return 0;
2973 }
2974
2975 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2976                    int __user *optlen)
2977 {
2978         struct inet_connection_sock *icsk = inet_csk(sk);
2979
2980         if (level != SOL_TCP)
2981                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2982                                                      optval, optlen);
2983         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2984 }
2985 EXPORT_SYMBOL(tcp_getsockopt);
2986
2987 #ifdef CONFIG_COMPAT
2988 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2989                           char __user *optval, int __user *optlen)
2990 {
2991         if (level != SOL_TCP)
2992                 return inet_csk_compat_getsockopt(sk, level, optname,
2993                                                   optval, optlen);
2994         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2995 }
2996 EXPORT_SYMBOL(compat_tcp_getsockopt);
2997 #endif
2998
2999 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
3000         netdev_features_t features)
3001 {
3002         struct sk_buff *segs = ERR_PTR(-EINVAL);
3003         struct tcphdr *th;
3004         unsigned int thlen;
3005         unsigned int seq;
3006         __be32 delta;
3007         unsigned int oldlen;
3008         unsigned int mss;
3009
3010         if (!pskb_may_pull(skb, sizeof(*th)))
3011                 goto out;
3012
3013         th = tcp_hdr(skb);
3014         thlen = th->doff * 4;
3015         if (thlen < sizeof(*th))
3016                 goto out;
3017
3018         if (!pskb_may_pull(skb, thlen))
3019                 goto out;
3020
3021         oldlen = (u16)~skb->len;
3022         __skb_pull(skb, thlen);
3023
3024         mss = skb_shinfo(skb)->gso_size;
3025         if (unlikely(skb->len <= mss))
3026                 goto out;
3027
3028         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
3029                 /* Packet is from an untrusted source, reset gso_segs. */
3030                 int type = skb_shinfo(skb)->gso_type;
3031
3032                 if (unlikely(type &
3033                              ~(SKB_GSO_TCPV4 |
3034                                SKB_GSO_DODGY |
3035                                SKB_GSO_TCP_ECN |
3036                                SKB_GSO_TCPV6 |
3037                                SKB_GSO_SHARED_FRAG |
3038                                0) ||
3039                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
3040                         goto out;
3041
3042                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
3043
3044                 segs = NULL;
3045                 goto out;
3046         }
3047
3048         segs = skb_segment(skb, features);
3049         if (IS_ERR(segs))
3050                 goto out;
3051
3052         delta = htonl(oldlen + (thlen + mss));
3053
3054         skb = segs;
3055         th = tcp_hdr(skb);
3056         seq = ntohl(th->seq);
3057
3058         do {
3059                 th->fin = th->psh = 0;
3060
3061                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3062                                        (__force u32)delta));
3063                 if (skb->ip_summed != CHECKSUM_PARTIAL)
3064                         th->check =
3065                              csum_fold(csum_partial(skb_transport_header(skb),
3066                                                     thlen, skb->csum));
3067
3068                 seq += mss;
3069                 skb = skb->next;
3070                 th = tcp_hdr(skb);
3071
3072                 th->seq = htonl(seq);
3073                 th->cwr = 0;
3074         } while (skb->next);
3075
3076         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
3077                       skb->data_len);
3078         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3079                                 (__force u32)delta));
3080         if (skb->ip_summed != CHECKSUM_PARTIAL)
3081                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
3082                                                    thlen, skb->csum));
3083
3084 out:
3085         return segs;
3086 }
3087 EXPORT_SYMBOL(tcp_tso_segment);
3088
3089 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3090 {
3091         struct sk_buff **pp = NULL;
3092         struct sk_buff *p;
3093         struct tcphdr *th;
3094         struct tcphdr *th2;
3095         unsigned int len;
3096         unsigned int thlen;
3097         __be32 flags;
3098         unsigned int mss = 1;
3099         unsigned int hlen;
3100         unsigned int off;
3101         int flush = 1;
3102         int i;
3103
3104         off = skb_gro_offset(skb);
3105         hlen = off + sizeof(*th);
3106         th = skb_gro_header_fast(skb, off);
3107         if (skb_gro_header_hard(skb, hlen)) {
3108                 th = skb_gro_header_slow(skb, hlen, off);
3109                 if (unlikely(!th))
3110                         goto out;
3111         }
3112
3113         thlen = th->doff * 4;
3114         if (thlen < sizeof(*th))
3115                 goto out;
3116
3117         hlen = off + thlen;
3118         if (skb_gro_header_hard(skb, hlen)) {
3119                 th = skb_gro_header_slow(skb, hlen, off);
3120                 if (unlikely(!th))
3121                         goto out;
3122         }
3123
3124         skb_gro_pull(skb, thlen);
3125
3126         len = skb_gro_len(skb);
3127         flags = tcp_flag_word(th);
3128
3129         for (; (p = *head); head = &p->next) {
3130                 if (!NAPI_GRO_CB(p)->same_flow)
3131                         continue;
3132
3133                 th2 = tcp_hdr(p);
3134
3135                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3136                         NAPI_GRO_CB(p)->same_flow = 0;
3137                         continue;
3138                 }
3139
3140                 goto found;
3141         }
3142
3143         goto out_check_final;
3144
3145 found:
3146         flush = NAPI_GRO_CB(p)->flush;
3147         flush |= (__force int)(flags & TCP_FLAG_CWR);
3148         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3149                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3150         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3151         for (i = sizeof(*th); i < thlen; i += 4)
3152                 flush |= *(u32 *)((u8 *)th + i) ^
3153                          *(u32 *)((u8 *)th2 + i);
3154
3155         mss = skb_shinfo(p)->gso_size;
3156
3157         flush |= (len - 1) >= mss;
3158         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3159
3160         if (flush || skb_gro_receive(head, skb)) {
3161                 mss = 1;
3162                 goto out_check_final;
3163         }
3164
3165         p = *head;
3166         th2 = tcp_hdr(p);
3167         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3168
3169 out_check_final:
3170         flush = len < mss;
3171         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3172                                         TCP_FLAG_RST | TCP_FLAG_SYN |
3173                                         TCP_FLAG_FIN));
3174
3175         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3176                 pp = head;
3177
3178 out:
3179         NAPI_GRO_CB(skb)->flush |= flush;
3180
3181         return pp;
3182 }
3183 EXPORT_SYMBOL(tcp_gro_receive);
3184
3185 int tcp_gro_complete(struct sk_buff *skb)
3186 {
3187         struct tcphdr *th = tcp_hdr(skb);
3188
3189         skb->csum_start = skb_transport_header(skb) - skb->head;
3190         skb->csum_offset = offsetof(struct tcphdr, check);
3191         skb->ip_summed = CHECKSUM_PARTIAL;
3192
3193         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3194
3195         if (th->cwr)
3196                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3197
3198         return 0;
3199 }
3200 EXPORT_SYMBOL(tcp_gro_complete);
3201
3202 #ifdef CONFIG_TCP_MD5SIG
3203 static unsigned long tcp_md5sig_users;
3204 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3205 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3206
3207 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3208 {
3209         int cpu;
3210
3211         for_each_possible_cpu(cpu) {
3212                 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3213
3214                 if (p->md5_desc.tfm)
3215                         crypto_free_hash(p->md5_desc.tfm);
3216         }
3217         free_percpu(pool);
3218 }
3219
3220 void tcp_free_md5sig_pool(void)
3221 {
3222         struct tcp_md5sig_pool __percpu *pool = NULL;
3223
3224         spin_lock_bh(&tcp_md5sig_pool_lock);
3225         if (--tcp_md5sig_users == 0) {
3226                 pool = tcp_md5sig_pool;
3227                 tcp_md5sig_pool = NULL;
3228         }
3229         spin_unlock_bh(&tcp_md5sig_pool_lock);
3230         if (pool)
3231                 __tcp_free_md5sig_pool(pool);
3232 }
3233 EXPORT_SYMBOL(tcp_free_md5sig_pool);
3234
3235 static struct tcp_md5sig_pool __percpu *
3236 __tcp_alloc_md5sig_pool(struct sock *sk)
3237 {
3238         int cpu;
3239         struct tcp_md5sig_pool __percpu *pool;
3240
3241         pool = alloc_percpu(struct tcp_md5sig_pool);
3242         if (!pool)
3243                 return NULL;
3244
3245         for_each_possible_cpu(cpu) {
3246                 struct crypto_hash *hash;
3247
3248                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3249                 if (IS_ERR_OR_NULL(hash))
3250                         goto out_free;
3251
3252                 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3253         }
3254         return pool;
3255 out_free:
3256         __tcp_free_md5sig_pool(pool);
3257         return NULL;
3258 }
3259
3260 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3261 {
3262         struct tcp_md5sig_pool __percpu *pool;
3263         bool alloc = false;
3264
3265 retry:
3266         spin_lock_bh(&tcp_md5sig_pool_lock);
3267         pool = tcp_md5sig_pool;
3268         if (tcp_md5sig_users++ == 0) {
3269                 alloc = true;
3270                 spin_unlock_bh(&tcp_md5sig_pool_lock);
3271         } else if (!pool) {
3272                 tcp_md5sig_users--;
3273                 spin_unlock_bh(&tcp_md5sig_pool_lock);
3274                 cpu_relax();
3275                 goto retry;
3276         } else
3277                 spin_unlock_bh(&tcp_md5sig_pool_lock);
3278
3279         if (alloc) {
3280                 /* we cannot hold spinlock here because this may sleep. */
3281                 struct tcp_md5sig_pool __percpu *p;
3282
3283                 p = __tcp_alloc_md5sig_pool(sk);
3284                 spin_lock_bh(&tcp_md5sig_pool_lock);
3285                 if (!p) {
3286                         tcp_md5sig_users--;
3287                         spin_unlock_bh(&tcp_md5sig_pool_lock);
3288                         return NULL;
3289                 }
3290                 pool = tcp_md5sig_pool;
3291                 if (pool) {
3292                         /* oops, it has already been assigned. */
3293                         spin_unlock_bh(&tcp_md5sig_pool_lock);
3294                         __tcp_free_md5sig_pool(p);
3295                 } else {
3296                         tcp_md5sig_pool = pool = p;
3297                         spin_unlock_bh(&tcp_md5sig_pool_lock);
3298                 }
3299         }
3300         return pool;
3301 }
3302 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3303
3304
3305 /**
3306  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3307  *
3308  *      We use percpu structure, so if we succeed, we exit with preemption
3309  *      and BH disabled, to make sure another thread or softirq handling
3310  *      wont try to get same context.
3311  */
3312 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3313 {
3314         struct tcp_md5sig_pool __percpu *p;
3315
3316         local_bh_disable();
3317
3318         spin_lock(&tcp_md5sig_pool_lock);
3319         p = tcp_md5sig_pool;
3320         if (p)
3321                 tcp_md5sig_users++;
3322         spin_unlock(&tcp_md5sig_pool_lock);
3323
3324         if (p)
3325                 return this_cpu_ptr(p);
3326
3327         local_bh_enable();
3328         return NULL;
3329 }
3330 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3331
3332 void tcp_put_md5sig_pool(void)
3333 {
3334         local_bh_enable();
3335         tcp_free_md5sig_pool();
3336 }
3337 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3338
3339 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3340                         const struct tcphdr *th)
3341 {
3342         struct scatterlist sg;
3343         struct tcphdr hdr;
3344         int err;
3345
3346         /* We are not allowed to change tcphdr, make a local copy */
3347         memcpy(&hdr, th, sizeof(hdr));
3348         hdr.check = 0;
3349
3350         /* options aren't included in the hash */
3351         sg_init_one(&sg, &hdr, sizeof(hdr));
3352         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3353         return err;
3354 }
3355 EXPORT_SYMBOL(tcp_md5_hash_header);
3356
3357 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3358                           const struct sk_buff *skb, unsigned int header_len)
3359 {
3360         struct scatterlist sg;
3361         const struct tcphdr *tp = tcp_hdr(skb);
3362         struct hash_desc *desc = &hp->md5_desc;
3363         unsigned int i;
3364         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3365                                            skb_headlen(skb) - header_len : 0;
3366         const struct skb_shared_info *shi = skb_shinfo(skb);
3367         struct sk_buff *frag_iter;
3368
3369         sg_init_table(&sg, 1);
3370
3371         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3372         if (crypto_hash_update(desc, &sg, head_data_len))
3373                 return 1;
3374
3375         for (i = 0; i < shi->nr_frags; ++i) {
3376                 const struct skb_frag_struct *f = &shi->frags[i];
3377                 struct page *page = skb_frag_page(f);
3378                 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3379                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3380                         return 1;
3381         }
3382
3383         skb_walk_frags(skb, frag_iter)
3384                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3385                         return 1;
3386
3387         return 0;
3388 }
3389 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3390
3391 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3392 {
3393         struct scatterlist sg;
3394
3395         sg_init_one(&sg, key->key, key->keylen);
3396         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3397 }
3398 EXPORT_SYMBOL(tcp_md5_hash_key);
3399
3400 #endif
3401
3402 /* Each Responder maintains up to two secret values concurrently for
3403  * efficient secret rollover.  Each secret value has 4 states:
3404  *
3405  * Generating.  (tcp_secret_generating != tcp_secret_primary)
3406  *    Generates new Responder-Cookies, but not yet used for primary
3407  *    verification.  This is a short-term state, typically lasting only
3408  *    one round trip time (RTT).
3409  *
3410  * Primary.  (tcp_secret_generating == tcp_secret_primary)
3411  *    Used both for generation and primary verification.
3412  *
3413  * Retiring.  (tcp_secret_retiring != tcp_secret_secondary)
3414  *    Used for verification, until the first failure that can be
3415  *    verified by the newer Generating secret.  At that time, this
3416  *    cookie's state is changed to Secondary, and the Generating
3417  *    cookie's state is changed to Primary.  This is a short-term state,
3418  *    typically lasting only one round trip time (RTT).
3419  *
3420  * Secondary.  (tcp_secret_retiring == tcp_secret_secondary)
3421  *    Used for secondary verification, after primary verification
3422  *    failures.  This state lasts no more than twice the Maximum Segment
3423  *    Lifetime (2MSL).  Then, the secret is discarded.
3424  */
3425 struct tcp_cookie_secret {
3426         /* The secret is divided into two parts.  The digest part is the
3427          * equivalent of previously hashing a secret and saving the state,
3428          * and serves as an initialization vector (IV).  The message part
3429          * serves as the trailing secret.
3430          */
3431         u32                             secrets[COOKIE_WORKSPACE_WORDS];
3432         unsigned long                   expires;
3433 };
3434
3435 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3436 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3437 #define TCP_SECRET_LIFE (HZ * 600)
3438
3439 static struct tcp_cookie_secret tcp_secret_one;
3440 static struct tcp_cookie_secret tcp_secret_two;
3441
3442 /* Essentially a circular list, without dynamic allocation. */
3443 static struct tcp_cookie_secret *tcp_secret_generating;
3444 static struct tcp_cookie_secret *tcp_secret_primary;
3445 static struct tcp_cookie_secret *tcp_secret_retiring;
3446 static struct tcp_cookie_secret *tcp_secret_secondary;
3447
3448 static DEFINE_SPINLOCK(tcp_secret_locker);
3449
3450 /* Select a pseudo-random word in the cookie workspace.
3451  */
3452 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3453 {
3454         return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3455 }
3456
3457 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3458  * Called in softirq context.
3459  * Returns: 0 for success.
3460  */
3461 int tcp_cookie_generator(u32 *bakery)
3462 {
3463         unsigned long jiffy = jiffies;
3464
3465         if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3466                 spin_lock_bh(&tcp_secret_locker);
3467                 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3468                         /* refreshed by another */
3469                         memcpy(bakery,
3470                                &tcp_secret_generating->secrets[0],
3471                                COOKIE_WORKSPACE_WORDS);
3472                 } else {
3473                         /* still needs refreshing */
3474                         get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3475
3476                         /* The first time, paranoia assumes that the
3477                          * randomization function isn't as strong.  But,
3478                          * this secret initialization is delayed until
3479                          * the last possible moment (packet arrival).
3480                          * Although that time is observable, it is
3481                          * unpredictably variable.  Mash in the most
3482                          * volatile clock bits available, and expire the
3483                          * secret extra quickly.
3484                          */
3485                         if (unlikely(tcp_secret_primary->expires ==
3486                                      tcp_secret_secondary->expires)) {
3487                                 struct timespec tv;
3488
3489                                 getnstimeofday(&tv);
3490                                 bakery[COOKIE_DIGEST_WORDS+0] ^=
3491                                         (u32)tv.tv_nsec;
3492
3493                                 tcp_secret_secondary->expires = jiffy
3494                                         + TCP_SECRET_1MSL
3495                                         + (0x0f & tcp_cookie_work(bakery, 0));
3496                         } else {
3497                                 tcp_secret_secondary->expires = jiffy
3498                                         + TCP_SECRET_LIFE
3499                                         + (0xff & tcp_cookie_work(bakery, 1));
3500                                 tcp_secret_primary->expires = jiffy
3501                                         + TCP_SECRET_2MSL
3502                                         + (0x1f & tcp_cookie_work(bakery, 2));
3503                         }
3504                         memcpy(&tcp_secret_secondary->secrets[0],
3505                                bakery, COOKIE_WORKSPACE_WORDS);
3506
3507                         rcu_assign_pointer(tcp_secret_generating,
3508                                            tcp_secret_secondary);
3509                         rcu_assign_pointer(tcp_secret_retiring,
3510                                            tcp_secret_primary);
3511                         /*
3512                          * Neither call_rcu() nor synchronize_rcu() needed.
3513                          * Retiring data is not freed.  It is replaced after
3514                          * further (locked) pointer updates, and a quiet time
3515                          * (minimum 1MSL, maximum LIFE - 2MSL).
3516                          */
3517                 }
3518                 spin_unlock_bh(&tcp_secret_locker);
3519         } else {
3520                 rcu_read_lock_bh();
3521                 memcpy(bakery,
3522                        &rcu_dereference(tcp_secret_generating)->secrets[0],
3523                        COOKIE_WORKSPACE_WORDS);
3524                 rcu_read_unlock_bh();
3525         }
3526         return 0;
3527 }
3528 EXPORT_SYMBOL(tcp_cookie_generator);
3529
3530 void tcp_done(struct sock *sk)
3531 {
3532         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3533
3534         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3535                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3536
3537         tcp_set_state(sk, TCP_CLOSE);
3538         tcp_clear_xmit_timers(sk);
3539         if (req != NULL)
3540                 reqsk_fastopen_remove(sk, req, false);
3541
3542         sk->sk_shutdown = SHUTDOWN_MASK;
3543
3544         if (!sock_flag(sk, SOCK_DEAD))
3545                 sk->sk_state_change(sk);
3546         else
3547                 inet_csk_destroy_sock(sk);
3548 }
3549 EXPORT_SYMBOL_GPL(tcp_done);
3550
3551 extern struct tcp_congestion_ops tcp_reno;
3552
3553 static __initdata unsigned long thash_entries;
3554 static int __init set_thash_entries(char *str)
3555 {
3556         ssize_t ret;
3557
3558         if (!str)
3559                 return 0;
3560
3561         ret = kstrtoul(str, 0, &thash_entries);
3562         if (ret)
3563                 return 0;
3564
3565         return 1;
3566 }
3567 __setup("thash_entries=", set_thash_entries);
3568
3569 void tcp_init_mem(struct net *net)
3570 {
3571         unsigned long limit = nr_free_buffer_pages() / 8;
3572         limit = max(limit, 128UL);
3573         net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3574         net->ipv4.sysctl_tcp_mem[1] = limit;
3575         net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3576 }
3577
3578 void __init tcp_init(void)
3579 {
3580         struct sk_buff *skb = NULL;
3581         unsigned long limit;
3582         int max_rshare, max_wshare, cnt;
3583         unsigned int i;
3584         unsigned long jiffy = jiffies;
3585
3586         BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3587
3588         percpu_counter_init(&tcp_sockets_allocated, 0);
3589         percpu_counter_init(&tcp_orphan_count, 0);
3590         tcp_hashinfo.bind_bucket_cachep =
3591                 kmem_cache_create("tcp_bind_bucket",
3592                                   sizeof(struct inet_bind_bucket), 0,
3593                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3594
3595         /* Size and allocate the main established and bind bucket
3596          * hash tables.
3597          *
3598          * The methodology is similar to that of the buffer cache.
3599          */
3600         tcp_hashinfo.ehash =
3601                 alloc_large_system_hash("TCP established",
3602                                         sizeof(struct inet_ehash_bucket),
3603                                         thash_entries,
3604                                         17, /* one slot per 128 KB of memory */
3605                                         0,
3606                                         NULL,
3607                                         &tcp_hashinfo.ehash_mask,
3608                                         0,
3609                                         thash_entries ? 0 : 512 * 1024);
3610         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3611                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3612                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3613         }
3614         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3615                 panic("TCP: failed to alloc ehash_locks");
3616         tcp_hashinfo.bhash =
3617                 alloc_large_system_hash("TCP bind",
3618                                         sizeof(struct inet_bind_hashbucket),
3619                                         tcp_hashinfo.ehash_mask + 1,
3620                                         17, /* one slot per 128 KB of memory */
3621                                         0,
3622                                         &tcp_hashinfo.bhash_size,
3623                                         NULL,
3624                                         0,
3625                                         64 * 1024);
3626         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3627         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3628                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3629                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3630         }
3631
3632
3633         cnt = tcp_hashinfo.ehash_mask + 1;
3634
3635         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3636         sysctl_tcp_max_orphans = cnt / 2;
3637         sysctl_max_syn_backlog = max(128, cnt / 256);
3638
3639         tcp_init_mem(&init_net);
3640         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3641         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3642         max_wshare = min(4UL*1024*1024, limit);
3643         max_rshare = min(6UL*1024*1024, limit);
3644
3645         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3646         sysctl_tcp_wmem[1] = 16*1024;
3647         sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3648
3649         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3650         sysctl_tcp_rmem[1] = 87380;
3651         sysctl_tcp_rmem[2] = max(87380, max_rshare);
3652
3653         pr_info("Hash tables configured (established %u bind %u)\n",
3654                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3655
3656         tcp_metrics_init();
3657
3658         tcp_register_congestion_control(&tcp_reno);
3659
3660         memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3661         memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3662         tcp_secret_one.expires = jiffy; /* past due */
3663         tcp_secret_two.expires = jiffy; /* past due */
3664         tcp_secret_generating = &tcp_secret_one;
3665         tcp_secret_primary = &tcp_secret_one;
3666         tcp_secret_retiring = &tcp_secret_two;
3667         tcp_secret_secondary = &tcp_secret_two;
3668         tcp_tasklet_init();
3669 }