]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - net/ipv4/tcp.c
net: Fix possible wrong checksum generation.
[~shefty/rdma-dev.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         tp->tsoffset = 0;
404
405         sk->sk_state = TCP_CLOSE;
406
407         sk->sk_write_space = sk_stream_write_space;
408         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
409
410         icsk->icsk_sync_mss = tcp_sync_mss;
411
412         /* TCP Cookie Transactions */
413         if (sysctl_tcp_cookie_size > 0) {
414                 /* Default, cookies without s_data_payload. */
415                 tp->cookie_values =
416                         kzalloc(sizeof(*tp->cookie_values),
417                                 sk->sk_allocation);
418                 if (tp->cookie_values != NULL)
419                         kref_init(&tp->cookie_values->kref);
420         }
421         /* Presumed zeroed, in order of appearance:
422          *      cookie_in_always, cookie_out_never,
423          *      s_data_constant, s_data_in, s_data_out
424          */
425         sk->sk_sndbuf = sysctl_tcp_wmem[1];
426         sk->sk_rcvbuf = sysctl_tcp_rmem[1];
427
428         local_bh_disable();
429         sock_update_memcg(sk);
430         sk_sockets_allocated_inc(sk);
431         local_bh_enable();
432 }
433 EXPORT_SYMBOL(tcp_init_sock);
434
435 /*
436  *      Wait for a TCP event.
437  *
438  *      Note that we don't need to lock the socket, as the upper poll layers
439  *      take care of normal races (between the test and the event) and we don't
440  *      go look at any of the socket buffers directly.
441  */
442 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
443 {
444         unsigned int mask;
445         struct sock *sk = sock->sk;
446         const struct tcp_sock *tp = tcp_sk(sk);
447
448         sock_poll_wait(file, sk_sleep(sk), wait);
449         if (sk->sk_state == TCP_LISTEN)
450                 return inet_csk_listen_poll(sk);
451
452         /* Socket is not locked. We are protected from async events
453          * by poll logic and correct handling of state changes
454          * made by other threads is impossible in any case.
455          */
456
457         mask = 0;
458
459         /*
460          * POLLHUP is certainly not done right. But poll() doesn't
461          * have a notion of HUP in just one direction, and for a
462          * socket the read side is more interesting.
463          *
464          * Some poll() documentation says that POLLHUP is incompatible
465          * with the POLLOUT/POLLWR flags, so somebody should check this
466          * all. But careful, it tends to be safer to return too many
467          * bits than too few, and you can easily break real applications
468          * if you don't tell them that something has hung up!
469          *
470          * Check-me.
471          *
472          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
473          * our fs/select.c). It means that after we received EOF,
474          * poll always returns immediately, making impossible poll() on write()
475          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
476          * if and only if shutdown has been made in both directions.
477          * Actually, it is interesting to look how Solaris and DUX
478          * solve this dilemma. I would prefer, if POLLHUP were maskable,
479          * then we could set it on SND_SHUTDOWN. BTW examples given
480          * in Stevens' books assume exactly this behaviour, it explains
481          * why POLLHUP is incompatible with POLLOUT.    --ANK
482          *
483          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
484          * blocking on fresh not-connected or disconnected socket. --ANK
485          */
486         if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
487                 mask |= POLLHUP;
488         if (sk->sk_shutdown & RCV_SHUTDOWN)
489                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
490
491         /* Connected or passive Fast Open socket? */
492         if (sk->sk_state != TCP_SYN_SENT &&
493             (sk->sk_state != TCP_SYN_RECV || tp->fastopen_rsk != NULL)) {
494                 int target = sock_rcvlowat(sk, 0, INT_MAX);
495
496                 if (tp->urg_seq == tp->copied_seq &&
497                     !sock_flag(sk, SOCK_URGINLINE) &&
498                     tp->urg_data)
499                         target++;
500
501                 /* Potential race condition. If read of tp below will
502                  * escape above sk->sk_state, we can be illegally awaken
503                  * in SYN_* states. */
504                 if (tp->rcv_nxt - tp->copied_seq >= target)
505                         mask |= POLLIN | POLLRDNORM;
506
507                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
508                         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
509                                 mask |= POLLOUT | POLLWRNORM;
510                         } else {  /* send SIGIO later */
511                                 set_bit(SOCK_ASYNC_NOSPACE,
512                                         &sk->sk_socket->flags);
513                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
514
515                                 /* Race breaker. If space is freed after
516                                  * wspace test but before the flags are set,
517                                  * IO signal will be lost.
518                                  */
519                                 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
520                                         mask |= POLLOUT | POLLWRNORM;
521                         }
522                 } else
523                         mask |= POLLOUT | POLLWRNORM;
524
525                 if (tp->urg_data & TCP_URG_VALID)
526                         mask |= POLLPRI;
527         }
528         /* This barrier is coupled with smp_wmb() in tcp_reset() */
529         smp_rmb();
530         if (sk->sk_err)
531                 mask |= POLLERR;
532
533         return mask;
534 }
535 EXPORT_SYMBOL(tcp_poll);
536
537 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
538 {
539         struct tcp_sock *tp = tcp_sk(sk);
540         int answ;
541         bool slow;
542
543         switch (cmd) {
544         case SIOCINQ:
545                 if (sk->sk_state == TCP_LISTEN)
546                         return -EINVAL;
547
548                 slow = lock_sock_fast(sk);
549                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
550                         answ = 0;
551                 else if (sock_flag(sk, SOCK_URGINLINE) ||
552                          !tp->urg_data ||
553                          before(tp->urg_seq, tp->copied_seq) ||
554                          !before(tp->urg_seq, tp->rcv_nxt)) {
555
556                         answ = tp->rcv_nxt - tp->copied_seq;
557
558                         /* Subtract 1, if FIN was received */
559                         if (answ && sock_flag(sk, SOCK_DONE))
560                                 answ--;
561                 } else
562                         answ = tp->urg_seq - tp->copied_seq;
563                 unlock_sock_fast(sk, slow);
564                 break;
565         case SIOCATMARK:
566                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
567                 break;
568         case SIOCOUTQ:
569                 if (sk->sk_state == TCP_LISTEN)
570                         return -EINVAL;
571
572                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
573                         answ = 0;
574                 else
575                         answ = tp->write_seq - tp->snd_una;
576                 break;
577         case SIOCOUTQNSD:
578                 if (sk->sk_state == TCP_LISTEN)
579                         return -EINVAL;
580
581                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
582                         answ = 0;
583                 else
584                         answ = tp->write_seq - tp->snd_nxt;
585                 break;
586         default:
587                 return -ENOIOCTLCMD;
588         }
589
590         return put_user(answ, (int __user *)arg);
591 }
592 EXPORT_SYMBOL(tcp_ioctl);
593
594 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
595 {
596         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
597         tp->pushed_seq = tp->write_seq;
598 }
599
600 static inline bool forced_push(const struct tcp_sock *tp)
601 {
602         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
603 }
604
605 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
606 {
607         struct tcp_sock *tp = tcp_sk(sk);
608         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
609
610         skb->csum    = 0;
611         tcb->seq     = tcb->end_seq = tp->write_seq;
612         tcb->tcp_flags = TCPHDR_ACK;
613         tcb->sacked  = 0;
614         skb_header_release(skb);
615         tcp_add_write_queue_tail(sk, skb);
616         sk->sk_wmem_queued += skb->truesize;
617         sk_mem_charge(sk, skb->truesize);
618         if (tp->nonagle & TCP_NAGLE_PUSH)
619                 tp->nonagle &= ~TCP_NAGLE_PUSH;
620 }
621
622 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
623 {
624         if (flags & MSG_OOB)
625                 tp->snd_up = tp->write_seq;
626 }
627
628 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
629                             int nonagle)
630 {
631         if (tcp_send_head(sk)) {
632                 struct tcp_sock *tp = tcp_sk(sk);
633
634                 if (!(flags & MSG_MORE) || forced_push(tp))
635                         tcp_mark_push(tp, tcp_write_queue_tail(sk));
636
637                 tcp_mark_urg(tp, flags);
638                 __tcp_push_pending_frames(sk, mss_now,
639                                           (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
640         }
641 }
642
643 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
644                                 unsigned int offset, size_t len)
645 {
646         struct tcp_splice_state *tss = rd_desc->arg.data;
647         int ret;
648
649         ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
650                               tss->flags);
651         if (ret > 0)
652                 rd_desc->count -= ret;
653         return ret;
654 }
655
656 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
657 {
658         /* Store TCP splice context information in read_descriptor_t. */
659         read_descriptor_t rd_desc = {
660                 .arg.data = tss,
661                 .count    = tss->len,
662         };
663
664         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
665 }
666
667 /**
668  *  tcp_splice_read - splice data from TCP socket to a pipe
669  * @sock:       socket to splice from
670  * @ppos:       position (not valid)
671  * @pipe:       pipe to splice to
672  * @len:        number of bytes to splice
673  * @flags:      splice modifier flags
674  *
675  * Description:
676  *    Will read pages from given socket and fill them into a pipe.
677  *
678  **/
679 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
680                         struct pipe_inode_info *pipe, size_t len,
681                         unsigned int flags)
682 {
683         struct sock *sk = sock->sk;
684         struct tcp_splice_state tss = {
685                 .pipe = pipe,
686                 .len = len,
687                 .flags = flags,
688         };
689         long timeo;
690         ssize_t spliced;
691         int ret;
692
693         sock_rps_record_flow(sk);
694         /*
695          * We can't seek on a socket input
696          */
697         if (unlikely(*ppos))
698                 return -ESPIPE;
699
700         ret = spliced = 0;
701
702         lock_sock(sk);
703
704         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
705         while (tss.len) {
706                 ret = __tcp_splice_read(sk, &tss);
707                 if (ret < 0)
708                         break;
709                 else if (!ret) {
710                         if (spliced)
711                                 break;
712                         if (sock_flag(sk, SOCK_DONE))
713                                 break;
714                         if (sk->sk_err) {
715                                 ret = sock_error(sk);
716                                 break;
717                         }
718                         if (sk->sk_shutdown & RCV_SHUTDOWN)
719                                 break;
720                         if (sk->sk_state == TCP_CLOSE) {
721                                 /*
722                                  * This occurs when user tries to read
723                                  * from never connected socket.
724                                  */
725                                 if (!sock_flag(sk, SOCK_DONE))
726                                         ret = -ENOTCONN;
727                                 break;
728                         }
729                         if (!timeo) {
730                                 ret = -EAGAIN;
731                                 break;
732                         }
733                         sk_wait_data(sk, &timeo);
734                         if (signal_pending(current)) {
735                                 ret = sock_intr_errno(timeo);
736                                 break;
737                         }
738                         continue;
739                 }
740                 tss.len -= ret;
741                 spliced += ret;
742
743                 if (!timeo)
744                         break;
745                 release_sock(sk);
746                 lock_sock(sk);
747
748                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
749                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
750                     signal_pending(current))
751                         break;
752         }
753
754         release_sock(sk);
755
756         if (spliced)
757                 return spliced;
758
759         return ret;
760 }
761 EXPORT_SYMBOL(tcp_splice_read);
762
763 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
764 {
765         struct sk_buff *skb;
766
767         /* The TCP header must be at least 32-bit aligned.  */
768         size = ALIGN(size, 4);
769
770         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
771         if (skb) {
772                 if (sk_wmem_schedule(sk, skb->truesize)) {
773                         skb_reserve(skb, sk->sk_prot->max_header);
774                         /*
775                          * Make sure that we have exactly size bytes
776                          * available to the caller, no more, no less.
777                          */
778                         skb->avail_size = size;
779                         return skb;
780                 }
781                 __kfree_skb(skb);
782         } else {
783                 sk->sk_prot->enter_memory_pressure(sk);
784                 sk_stream_moderate_sndbuf(sk);
785         }
786         return NULL;
787 }
788
789 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
790                                        int large_allowed)
791 {
792         struct tcp_sock *tp = tcp_sk(sk);
793         u32 xmit_size_goal, old_size_goal;
794
795         xmit_size_goal = mss_now;
796
797         if (large_allowed && sk_can_gso(sk)) {
798                 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
799                                   inet_csk(sk)->icsk_af_ops->net_header_len -
800                                   inet_csk(sk)->icsk_ext_hdr_len -
801                                   tp->tcp_header_len);
802
803                 /* TSQ : try to have two TSO segments in flight */
804                 xmit_size_goal = min_t(u32, xmit_size_goal,
805                                        sysctl_tcp_limit_output_bytes >> 1);
806
807                 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
808
809                 /* We try hard to avoid divides here */
810                 old_size_goal = tp->xmit_size_goal_segs * mss_now;
811
812                 if (likely(old_size_goal <= xmit_size_goal &&
813                            old_size_goal + mss_now > xmit_size_goal)) {
814                         xmit_size_goal = old_size_goal;
815                 } else {
816                         tp->xmit_size_goal_segs =
817                                 min_t(u16, xmit_size_goal / mss_now,
818                                       sk->sk_gso_max_segs);
819                         xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
820                 }
821         }
822
823         return max(xmit_size_goal, mss_now);
824 }
825
826 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
827 {
828         int mss_now;
829
830         mss_now = tcp_current_mss(sk);
831         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
832
833         return mss_now;
834 }
835
836 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
837                                 size_t size, int flags)
838 {
839         struct tcp_sock *tp = tcp_sk(sk);
840         int mss_now, size_goal;
841         int err;
842         ssize_t copied;
843         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
844
845         /* Wait for a connection to finish. One exception is TCP Fast Open
846          * (passive side) where data is allowed to be sent before a connection
847          * is fully established.
848          */
849         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
850             !tcp_passive_fastopen(sk)) {
851                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
852                         goto out_err;
853         }
854
855         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
856
857         mss_now = tcp_send_mss(sk, &size_goal, flags);
858         copied = 0;
859
860         err = -EPIPE;
861         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
862                 goto out_err;
863
864         while (size > 0) {
865                 struct sk_buff *skb = tcp_write_queue_tail(sk);
866                 int copy, i;
867                 bool can_coalesce;
868
869                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
870 new_segment:
871                         if (!sk_stream_memory_free(sk))
872                                 goto wait_for_sndbuf;
873
874                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
875                         if (!skb)
876                                 goto wait_for_memory;
877
878                         skb_entail(sk, skb);
879                         copy = size_goal;
880                 }
881
882                 if (copy > size)
883                         copy = size;
884
885                 i = skb_shinfo(skb)->nr_frags;
886                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
887                 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
888                         tcp_mark_push(tp, skb);
889                         goto new_segment;
890                 }
891                 if (!sk_wmem_schedule(sk, copy))
892                         goto wait_for_memory;
893
894                 if (can_coalesce) {
895                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
896                 } else {
897                         get_page(page);
898                         skb_fill_page_desc(skb, i, page, offset, copy);
899                 }
900                 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
901
902                 skb->len += copy;
903                 skb->data_len += copy;
904                 skb->truesize += copy;
905                 sk->sk_wmem_queued += copy;
906                 sk_mem_charge(sk, copy);
907                 skb->ip_summed = CHECKSUM_PARTIAL;
908                 tp->write_seq += copy;
909                 TCP_SKB_CB(skb)->end_seq += copy;
910                 skb_shinfo(skb)->gso_segs = 0;
911
912                 if (!copied)
913                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
914
915                 copied += copy;
916                 offset += copy;
917                 if (!(size -= copy))
918                         goto out;
919
920                 if (skb->len < size_goal || (flags & MSG_OOB))
921                         continue;
922
923                 if (forced_push(tp)) {
924                         tcp_mark_push(tp, skb);
925                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
926                 } else if (skb == tcp_send_head(sk))
927                         tcp_push_one(sk, mss_now);
928                 continue;
929
930 wait_for_sndbuf:
931                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
932 wait_for_memory:
933                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
934
935                 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
936                         goto do_error;
937
938                 mss_now = tcp_send_mss(sk, &size_goal, flags);
939         }
940
941 out:
942         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
943                 tcp_push(sk, flags, mss_now, tp->nonagle);
944         return copied;
945
946 do_error:
947         if (copied)
948                 goto out;
949 out_err:
950         return sk_stream_error(sk, flags, err);
951 }
952
953 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
954                  size_t size, int flags)
955 {
956         ssize_t res;
957
958         if (!(sk->sk_route_caps & NETIF_F_SG) ||
959             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
960                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
961                                         flags);
962
963         lock_sock(sk);
964         res = do_tcp_sendpages(sk, page, offset, size, flags);
965         release_sock(sk);
966         return res;
967 }
968 EXPORT_SYMBOL(tcp_sendpage);
969
970 static inline int select_size(const struct sock *sk, bool sg)
971 {
972         const struct tcp_sock *tp = tcp_sk(sk);
973         int tmp = tp->mss_cache;
974
975         if (sg) {
976                 if (sk_can_gso(sk)) {
977                         /* Small frames wont use a full page:
978                          * Payload will immediately follow tcp header.
979                          */
980                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
981                 } else {
982                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
983
984                         if (tmp >= pgbreak &&
985                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
986                                 tmp = pgbreak;
987                 }
988         }
989
990         return tmp;
991 }
992
993 void tcp_free_fastopen_req(struct tcp_sock *tp)
994 {
995         if (tp->fastopen_req != NULL) {
996                 kfree(tp->fastopen_req);
997                 tp->fastopen_req = NULL;
998         }
999 }
1000
1001 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *size)
1002 {
1003         struct tcp_sock *tp = tcp_sk(sk);
1004         int err, flags;
1005
1006         if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1007                 return -EOPNOTSUPP;
1008         if (tp->fastopen_req != NULL)
1009                 return -EALREADY; /* Another Fast Open is in progress */
1010
1011         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1012                                    sk->sk_allocation);
1013         if (unlikely(tp->fastopen_req == NULL))
1014                 return -ENOBUFS;
1015         tp->fastopen_req->data = msg;
1016
1017         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1018         err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1019                                     msg->msg_namelen, flags);
1020         *size = tp->fastopen_req->copied;
1021         tcp_free_fastopen_req(tp);
1022         return err;
1023 }
1024
1025 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1026                 size_t size)
1027 {
1028         struct iovec *iov;
1029         struct tcp_sock *tp = tcp_sk(sk);
1030         struct sk_buff *skb;
1031         int iovlen, flags, err, copied = 0;
1032         int mss_now = 0, size_goal, copied_syn = 0, offset = 0;
1033         bool sg;
1034         long timeo;
1035
1036         lock_sock(sk);
1037
1038         flags = msg->msg_flags;
1039         if (flags & MSG_FASTOPEN) {
1040                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn);
1041                 if (err == -EINPROGRESS && copied_syn > 0)
1042                         goto out;
1043                 else if (err)
1044                         goto out_err;
1045                 offset = copied_syn;
1046         }
1047
1048         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1049
1050         /* Wait for a connection to finish. One exception is TCP Fast Open
1051          * (passive side) where data is allowed to be sent before a connection
1052          * is fully established.
1053          */
1054         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1055             !tcp_passive_fastopen(sk)) {
1056                 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
1057                         goto do_error;
1058         }
1059
1060         if (unlikely(tp->repair)) {
1061                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1062                         copied = tcp_send_rcvq(sk, msg, size);
1063                         goto out;
1064                 }
1065
1066                 err = -EINVAL;
1067                 if (tp->repair_queue == TCP_NO_QUEUE)
1068                         goto out_err;
1069
1070                 /* 'common' sending to sendq */
1071         }
1072
1073         /* This should be in poll */
1074         clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
1075
1076         mss_now = tcp_send_mss(sk, &size_goal, flags);
1077
1078         /* Ok commence sending. */
1079         iovlen = msg->msg_iovlen;
1080         iov = msg->msg_iov;
1081         copied = 0;
1082
1083         err = -EPIPE;
1084         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1085                 goto out_err;
1086
1087         sg = !!(sk->sk_route_caps & NETIF_F_SG);
1088
1089         while (--iovlen >= 0) {
1090                 size_t seglen = iov->iov_len;
1091                 unsigned char __user *from = iov->iov_base;
1092
1093                 iov++;
1094                 if (unlikely(offset > 0)) {  /* Skip bytes copied in SYN */
1095                         if (offset >= seglen) {
1096                                 offset -= seglen;
1097                                 continue;
1098                         }
1099                         seglen -= offset;
1100                         from += offset;
1101                         offset = 0;
1102                 }
1103
1104                 while (seglen > 0) {
1105                         int copy = 0;
1106                         int max = size_goal;
1107
1108                         skb = tcp_write_queue_tail(sk);
1109                         if (tcp_send_head(sk)) {
1110                                 if (skb->ip_summed == CHECKSUM_NONE)
1111                                         max = mss_now;
1112                                 copy = max - skb->len;
1113                         }
1114
1115                         if (copy <= 0) {
1116 new_segment:
1117                                 /* Allocate new segment. If the interface is SG,
1118                                  * allocate skb fitting to single page.
1119                                  */
1120                                 if (!sk_stream_memory_free(sk))
1121                                         goto wait_for_sndbuf;
1122
1123                                 skb = sk_stream_alloc_skb(sk,
1124                                                           select_size(sk, sg),
1125                                                           sk->sk_allocation);
1126                                 if (!skb)
1127                                         goto wait_for_memory;
1128
1129                                 /*
1130                                  * Check whether we can use HW checksum.
1131                                  */
1132                                 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
1133                                         skb->ip_summed = CHECKSUM_PARTIAL;
1134
1135                                 skb_entail(sk, skb);
1136                                 copy = size_goal;
1137                                 max = size_goal;
1138                         }
1139
1140                         /* Try to append data to the end of skb. */
1141                         if (copy > seglen)
1142                                 copy = seglen;
1143
1144                         /* Where to copy to? */
1145                         if (skb_availroom(skb) > 0) {
1146                                 /* We have some space in skb head. Superb! */
1147                                 copy = min_t(int, copy, skb_availroom(skb));
1148                                 err = skb_add_data_nocache(sk, skb, from, copy);
1149                                 if (err)
1150                                         goto do_fault;
1151                         } else {
1152                                 bool merge = true;
1153                                 int i = skb_shinfo(skb)->nr_frags;
1154                                 struct page_frag *pfrag = sk_page_frag(sk);
1155
1156                                 if (!sk_page_frag_refill(sk, pfrag))
1157                                         goto wait_for_memory;
1158
1159                                 if (!skb_can_coalesce(skb, i, pfrag->page,
1160                                                       pfrag->offset)) {
1161                                         if (i == MAX_SKB_FRAGS || !sg) {
1162                                                 tcp_mark_push(tp, skb);
1163                                                 goto new_segment;
1164                                         }
1165                                         merge = false;
1166                                 }
1167
1168                                 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1169
1170                                 if (!sk_wmem_schedule(sk, copy))
1171                                         goto wait_for_memory;
1172
1173                                 err = skb_copy_to_page_nocache(sk, from, skb,
1174                                                                pfrag->page,
1175                                                                pfrag->offset,
1176                                                                copy);
1177                                 if (err)
1178                                         goto do_error;
1179
1180                                 /* Update the skb. */
1181                                 if (merge) {
1182                                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1183                                 } else {
1184                                         skb_fill_page_desc(skb, i, pfrag->page,
1185                                                            pfrag->offset, copy);
1186                                         get_page(pfrag->page);
1187                                 }
1188                                 pfrag->offset += copy;
1189                         }
1190
1191                         if (!copied)
1192                                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1193
1194                         tp->write_seq += copy;
1195                         TCP_SKB_CB(skb)->end_seq += copy;
1196                         skb_shinfo(skb)->gso_segs = 0;
1197
1198                         from += copy;
1199                         copied += copy;
1200                         if ((seglen -= copy) == 0 && iovlen == 0)
1201                                 goto out;
1202
1203                         if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1204                                 continue;
1205
1206                         if (forced_push(tp)) {
1207                                 tcp_mark_push(tp, skb);
1208                                 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1209                         } else if (skb == tcp_send_head(sk))
1210                                 tcp_push_one(sk, mss_now);
1211                         continue;
1212
1213 wait_for_sndbuf:
1214                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1215 wait_for_memory:
1216                         if (copied)
1217                                 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1218
1219                         if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1220                                 goto do_error;
1221
1222                         mss_now = tcp_send_mss(sk, &size_goal, flags);
1223                 }
1224         }
1225
1226 out:
1227         if (copied)
1228                 tcp_push(sk, flags, mss_now, tp->nonagle);
1229         release_sock(sk);
1230         return copied + copied_syn;
1231
1232 do_fault:
1233         if (!skb->len) {
1234                 tcp_unlink_write_queue(skb, sk);
1235                 /* It is the one place in all of TCP, except connection
1236                  * reset, where we can be unlinking the send_head.
1237                  */
1238                 tcp_check_send_head(sk, skb);
1239                 sk_wmem_free_skb(sk, skb);
1240         }
1241
1242 do_error:
1243         if (copied + copied_syn)
1244                 goto out;
1245 out_err:
1246         err = sk_stream_error(sk, flags, err);
1247         release_sock(sk);
1248         return err;
1249 }
1250 EXPORT_SYMBOL(tcp_sendmsg);
1251
1252 /*
1253  *      Handle reading urgent data. BSD has very simple semantics for
1254  *      this, no blocking and very strange errors 8)
1255  */
1256
1257 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1258 {
1259         struct tcp_sock *tp = tcp_sk(sk);
1260
1261         /* No URG data to read. */
1262         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1263             tp->urg_data == TCP_URG_READ)
1264                 return -EINVAL; /* Yes this is right ! */
1265
1266         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1267                 return -ENOTCONN;
1268
1269         if (tp->urg_data & TCP_URG_VALID) {
1270                 int err = 0;
1271                 char c = tp->urg_data;
1272
1273                 if (!(flags & MSG_PEEK))
1274                         tp->urg_data = TCP_URG_READ;
1275
1276                 /* Read urgent data. */
1277                 msg->msg_flags |= MSG_OOB;
1278
1279                 if (len > 0) {
1280                         if (!(flags & MSG_TRUNC))
1281                                 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1282                         len = 1;
1283                 } else
1284                         msg->msg_flags |= MSG_TRUNC;
1285
1286                 return err ? -EFAULT : len;
1287         }
1288
1289         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1290                 return 0;
1291
1292         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1293          * the available implementations agree in this case:
1294          * this call should never block, independent of the
1295          * blocking state of the socket.
1296          * Mike <pall@rz.uni-karlsruhe.de>
1297          */
1298         return -EAGAIN;
1299 }
1300
1301 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1302 {
1303         struct sk_buff *skb;
1304         int copied = 0, err = 0;
1305
1306         /* XXX -- need to support SO_PEEK_OFF */
1307
1308         skb_queue_walk(&sk->sk_write_queue, skb) {
1309                 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, skb->len);
1310                 if (err)
1311                         break;
1312
1313                 copied += skb->len;
1314         }
1315
1316         return err ?: copied;
1317 }
1318
1319 /* Clean up the receive buffer for full frames taken by the user,
1320  * then send an ACK if necessary.  COPIED is the number of bytes
1321  * tcp_recvmsg has given to the user so far, it speeds up the
1322  * calculation of whether or not we must ACK for the sake of
1323  * a window update.
1324  */
1325 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1326 {
1327         struct tcp_sock *tp = tcp_sk(sk);
1328         bool time_to_ack = false;
1329
1330         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1331
1332         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1333              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1334              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1335
1336         if (inet_csk_ack_scheduled(sk)) {
1337                 const struct inet_connection_sock *icsk = inet_csk(sk);
1338                    /* Delayed ACKs frequently hit locked sockets during bulk
1339                     * receive. */
1340                 if (icsk->icsk_ack.blocked ||
1341                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1342                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1343                     /*
1344                      * If this read emptied read buffer, we send ACK, if
1345                      * connection is not bidirectional, user drained
1346                      * receive buffer and there was a small segment
1347                      * in queue.
1348                      */
1349                     (copied > 0 &&
1350                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1351                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1352                        !icsk->icsk_ack.pingpong)) &&
1353                       !atomic_read(&sk->sk_rmem_alloc)))
1354                         time_to_ack = true;
1355         }
1356
1357         /* We send an ACK if we can now advertise a non-zero window
1358          * which has been raised "significantly".
1359          *
1360          * Even if window raised up to infinity, do not send window open ACK
1361          * in states, where we will not receive more. It is useless.
1362          */
1363         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1364                 __u32 rcv_window_now = tcp_receive_window(tp);
1365
1366                 /* Optimize, __tcp_select_window() is not cheap. */
1367                 if (2*rcv_window_now <= tp->window_clamp) {
1368                         __u32 new_window = __tcp_select_window(sk);
1369
1370                         /* Send ACK now, if this read freed lots of space
1371                          * in our buffer. Certainly, new_window is new window.
1372                          * We can advertise it now, if it is not less than current one.
1373                          * "Lots" means "at least twice" here.
1374                          */
1375                         if (new_window && new_window >= 2 * rcv_window_now)
1376                                 time_to_ack = true;
1377                 }
1378         }
1379         if (time_to_ack)
1380                 tcp_send_ack(sk);
1381 }
1382
1383 static void tcp_prequeue_process(struct sock *sk)
1384 {
1385         struct sk_buff *skb;
1386         struct tcp_sock *tp = tcp_sk(sk);
1387
1388         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1389
1390         /* RX process wants to run with disabled BHs, though it is not
1391          * necessary */
1392         local_bh_disable();
1393         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1394                 sk_backlog_rcv(sk, skb);
1395         local_bh_enable();
1396
1397         /* Clear memory counter. */
1398         tp->ucopy.memory = 0;
1399 }
1400
1401 #ifdef CONFIG_NET_DMA
1402 static void tcp_service_net_dma(struct sock *sk, bool wait)
1403 {
1404         dma_cookie_t done, used;
1405         dma_cookie_t last_issued;
1406         struct tcp_sock *tp = tcp_sk(sk);
1407
1408         if (!tp->ucopy.dma_chan)
1409                 return;
1410
1411         last_issued = tp->ucopy.dma_cookie;
1412         dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1413
1414         do {
1415                 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1416                                               last_issued, &done,
1417                                               &used) == DMA_SUCCESS) {
1418                         /* Safe to free early-copied skbs now */
1419                         __skb_queue_purge(&sk->sk_async_wait_queue);
1420                         break;
1421                 } else {
1422                         struct sk_buff *skb;
1423                         while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1424                                (dma_async_is_complete(skb->dma_cookie, done,
1425                                                       used) == DMA_SUCCESS)) {
1426                                 __skb_dequeue(&sk->sk_async_wait_queue);
1427                                 kfree_skb(skb);
1428                         }
1429                 }
1430         } while (wait);
1431 }
1432 #endif
1433
1434 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1435 {
1436         struct sk_buff *skb;
1437         u32 offset;
1438
1439         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1440                 offset = seq - TCP_SKB_CB(skb)->seq;
1441                 if (tcp_hdr(skb)->syn)
1442                         offset--;
1443                 if (offset < skb->len || tcp_hdr(skb)->fin) {
1444                         *off = offset;
1445                         return skb;
1446                 }
1447                 /* This looks weird, but this can happen if TCP collapsing
1448                  * splitted a fat GRO packet, while we released socket lock
1449                  * in skb_splice_bits()
1450                  */
1451                 sk_eat_skb(sk, skb, false);
1452         }
1453         return NULL;
1454 }
1455
1456 /*
1457  * This routine provides an alternative to tcp_recvmsg() for routines
1458  * that would like to handle copying from skbuffs directly in 'sendfile'
1459  * fashion.
1460  * Note:
1461  *      - It is assumed that the socket was locked by the caller.
1462  *      - The routine does not block.
1463  *      - At present, there is no support for reading OOB data
1464  *        or for 'peeking' the socket using this routine
1465  *        (although both would be easy to implement).
1466  */
1467 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1468                   sk_read_actor_t recv_actor)
1469 {
1470         struct sk_buff *skb;
1471         struct tcp_sock *tp = tcp_sk(sk);
1472         u32 seq = tp->copied_seq;
1473         u32 offset;
1474         int copied = 0;
1475
1476         if (sk->sk_state == TCP_LISTEN)
1477                 return -ENOTCONN;
1478         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1479                 if (offset < skb->len) {
1480                         int used;
1481                         size_t len;
1482
1483                         len = skb->len - offset;
1484                         /* Stop reading if we hit a patch of urgent data */
1485                         if (tp->urg_data) {
1486                                 u32 urg_offset = tp->urg_seq - seq;
1487                                 if (urg_offset < len)
1488                                         len = urg_offset;
1489                                 if (!len)
1490                                         break;
1491                         }
1492                         used = recv_actor(desc, skb, offset, len);
1493                         if (used <= 0) {
1494                                 if (!copied)
1495                                         copied = used;
1496                                 break;
1497                         } else if (used <= len) {
1498                                 seq += used;
1499                                 copied += used;
1500                                 offset += used;
1501                         }
1502                         /* If recv_actor drops the lock (e.g. TCP splice
1503                          * receive) the skb pointer might be invalid when
1504                          * getting here: tcp_collapse might have deleted it
1505                          * while aggregating skbs from the socket queue.
1506                          */
1507                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1508                         if (!skb)
1509                                 break;
1510                         /* TCP coalescing might have appended data to the skb.
1511                          * Try to splice more frags
1512                          */
1513                         if (offset + 1 != skb->len)
1514                                 continue;
1515                 }
1516                 if (tcp_hdr(skb)->fin) {
1517                         sk_eat_skb(sk, skb, false);
1518                         ++seq;
1519                         break;
1520                 }
1521                 sk_eat_skb(sk, skb, false);
1522                 if (!desc->count)
1523                         break;
1524                 tp->copied_seq = seq;
1525         }
1526         tp->copied_seq = seq;
1527
1528         tcp_rcv_space_adjust(sk);
1529
1530         /* Clean up data we have read: This will do ACK frames. */
1531         if (copied > 0) {
1532                 tcp_recv_skb(sk, seq, &offset);
1533                 tcp_cleanup_rbuf(sk, copied);
1534         }
1535         return copied;
1536 }
1537 EXPORT_SYMBOL(tcp_read_sock);
1538
1539 /*
1540  *      This routine copies from a sock struct into the user buffer.
1541  *
1542  *      Technical note: in 2.3 we work on _locked_ socket, so that
1543  *      tricks with *seq access order and skb->users are not required.
1544  *      Probably, code can be easily improved even more.
1545  */
1546
1547 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1548                 size_t len, int nonblock, int flags, int *addr_len)
1549 {
1550         struct tcp_sock *tp = tcp_sk(sk);
1551         int copied = 0;
1552         u32 peek_seq;
1553         u32 *seq;
1554         unsigned long used;
1555         int err;
1556         int target;             /* Read at least this many bytes */
1557         long timeo;
1558         struct task_struct *user_recv = NULL;
1559         bool copied_early = false;
1560         struct sk_buff *skb;
1561         u32 urg_hole = 0;
1562
1563         lock_sock(sk);
1564
1565         err = -ENOTCONN;
1566         if (sk->sk_state == TCP_LISTEN)
1567                 goto out;
1568
1569         timeo = sock_rcvtimeo(sk, nonblock);
1570
1571         /* Urgent data needs to be handled specially. */
1572         if (flags & MSG_OOB)
1573                 goto recv_urg;
1574
1575         if (unlikely(tp->repair)) {
1576                 err = -EPERM;
1577                 if (!(flags & MSG_PEEK))
1578                         goto out;
1579
1580                 if (tp->repair_queue == TCP_SEND_QUEUE)
1581                         goto recv_sndq;
1582
1583                 err = -EINVAL;
1584                 if (tp->repair_queue == TCP_NO_QUEUE)
1585                         goto out;
1586
1587                 /* 'common' recv queue MSG_PEEK-ing */
1588         }
1589
1590         seq = &tp->copied_seq;
1591         if (flags & MSG_PEEK) {
1592                 peek_seq = tp->copied_seq;
1593                 seq = &peek_seq;
1594         }
1595
1596         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1597
1598 #ifdef CONFIG_NET_DMA
1599         tp->ucopy.dma_chan = NULL;
1600         preempt_disable();
1601         skb = skb_peek_tail(&sk->sk_receive_queue);
1602         {
1603                 int available = 0;
1604
1605                 if (skb)
1606                         available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1607                 if ((available < target) &&
1608                     (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1609                     !sysctl_tcp_low_latency &&
1610                     net_dma_find_channel()) {
1611                         preempt_enable_no_resched();
1612                         tp->ucopy.pinned_list =
1613                                         dma_pin_iovec_pages(msg->msg_iov, len);
1614                 } else {
1615                         preempt_enable_no_resched();
1616                 }
1617         }
1618 #endif
1619
1620         do {
1621                 u32 offset;
1622
1623                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1624                 if (tp->urg_data && tp->urg_seq == *seq) {
1625                         if (copied)
1626                                 break;
1627                         if (signal_pending(current)) {
1628                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1629                                 break;
1630                         }
1631                 }
1632
1633                 /* Next get a buffer. */
1634
1635                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1636                         /* Now that we have two receive queues this
1637                          * shouldn't happen.
1638                          */
1639                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1640                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1641                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1642                                  flags))
1643                                 break;
1644
1645                         offset = *seq - TCP_SKB_CB(skb)->seq;
1646                         if (tcp_hdr(skb)->syn)
1647                                 offset--;
1648                         if (offset < skb->len)
1649                                 goto found_ok_skb;
1650                         if (tcp_hdr(skb)->fin)
1651                                 goto found_fin_ok;
1652                         WARN(!(flags & MSG_PEEK),
1653                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1654                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1655                 }
1656
1657                 /* Well, if we have backlog, try to process it now yet. */
1658
1659                 if (copied >= target && !sk->sk_backlog.tail)
1660                         break;
1661
1662                 if (copied) {
1663                         if (sk->sk_err ||
1664                             sk->sk_state == TCP_CLOSE ||
1665                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1666                             !timeo ||
1667                             signal_pending(current))
1668                                 break;
1669                 } else {
1670                         if (sock_flag(sk, SOCK_DONE))
1671                                 break;
1672
1673                         if (sk->sk_err) {
1674                                 copied = sock_error(sk);
1675                                 break;
1676                         }
1677
1678                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1679                                 break;
1680
1681                         if (sk->sk_state == TCP_CLOSE) {
1682                                 if (!sock_flag(sk, SOCK_DONE)) {
1683                                         /* This occurs when user tries to read
1684                                          * from never connected socket.
1685                                          */
1686                                         copied = -ENOTCONN;
1687                                         break;
1688                                 }
1689                                 break;
1690                         }
1691
1692                         if (!timeo) {
1693                                 copied = -EAGAIN;
1694                                 break;
1695                         }
1696
1697                         if (signal_pending(current)) {
1698                                 copied = sock_intr_errno(timeo);
1699                                 break;
1700                         }
1701                 }
1702
1703                 tcp_cleanup_rbuf(sk, copied);
1704
1705                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1706                         /* Install new reader */
1707                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1708                                 user_recv = current;
1709                                 tp->ucopy.task = user_recv;
1710                                 tp->ucopy.iov = msg->msg_iov;
1711                         }
1712
1713                         tp->ucopy.len = len;
1714
1715                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1716                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1717
1718                         /* Ugly... If prequeue is not empty, we have to
1719                          * process it before releasing socket, otherwise
1720                          * order will be broken at second iteration.
1721                          * More elegant solution is required!!!
1722                          *
1723                          * Look: we have the following (pseudo)queues:
1724                          *
1725                          * 1. packets in flight
1726                          * 2. backlog
1727                          * 3. prequeue
1728                          * 4. receive_queue
1729                          *
1730                          * Each queue can be processed only if the next ones
1731                          * are empty. At this point we have empty receive_queue.
1732                          * But prequeue _can_ be not empty after 2nd iteration,
1733                          * when we jumped to start of loop because backlog
1734                          * processing added something to receive_queue.
1735                          * We cannot release_sock(), because backlog contains
1736                          * packets arrived _after_ prequeued ones.
1737                          *
1738                          * Shortly, algorithm is clear --- to process all
1739                          * the queues in order. We could make it more directly,
1740                          * requeueing packets from backlog to prequeue, if
1741                          * is not empty. It is more elegant, but eats cycles,
1742                          * unfortunately.
1743                          */
1744                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1745                                 goto do_prequeue;
1746
1747                         /* __ Set realtime policy in scheduler __ */
1748                 }
1749
1750 #ifdef CONFIG_NET_DMA
1751                 if (tp->ucopy.dma_chan) {
1752                         if (tp->rcv_wnd == 0 &&
1753                             !skb_queue_empty(&sk->sk_async_wait_queue)) {
1754                                 tcp_service_net_dma(sk, true);
1755                                 tcp_cleanup_rbuf(sk, copied);
1756                         } else
1757                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1758                 }
1759 #endif
1760                 if (copied >= target) {
1761                         /* Do not sleep, just process backlog. */
1762                         release_sock(sk);
1763                         lock_sock(sk);
1764                 } else
1765                         sk_wait_data(sk, &timeo);
1766
1767 #ifdef CONFIG_NET_DMA
1768                 tcp_service_net_dma(sk, false);  /* Don't block */
1769                 tp->ucopy.wakeup = 0;
1770 #endif
1771
1772                 if (user_recv) {
1773                         int chunk;
1774
1775                         /* __ Restore normal policy in scheduler __ */
1776
1777                         if ((chunk = len - tp->ucopy.len) != 0) {
1778                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1779                                 len -= chunk;
1780                                 copied += chunk;
1781                         }
1782
1783                         if (tp->rcv_nxt == tp->copied_seq &&
1784                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1785 do_prequeue:
1786                                 tcp_prequeue_process(sk);
1787
1788                                 if ((chunk = len - tp->ucopy.len) != 0) {
1789                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1790                                         len -= chunk;
1791                                         copied += chunk;
1792                                 }
1793                         }
1794                 }
1795                 if ((flags & MSG_PEEK) &&
1796                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1797                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1798                                             current->comm,
1799                                             task_pid_nr(current));
1800                         peek_seq = tp->copied_seq;
1801                 }
1802                 continue;
1803
1804         found_ok_skb:
1805                 /* Ok so how much can we use? */
1806                 used = skb->len - offset;
1807                 if (len < used)
1808                         used = len;
1809
1810                 /* Do we have urgent data here? */
1811                 if (tp->urg_data) {
1812                         u32 urg_offset = tp->urg_seq - *seq;
1813                         if (urg_offset < used) {
1814                                 if (!urg_offset) {
1815                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1816                                                 ++*seq;
1817                                                 urg_hole++;
1818                                                 offset++;
1819                                                 used--;
1820                                                 if (!used)
1821                                                         goto skip_copy;
1822                                         }
1823                                 } else
1824                                         used = urg_offset;
1825                         }
1826                 }
1827
1828                 if (!(flags & MSG_TRUNC)) {
1829 #ifdef CONFIG_NET_DMA
1830                         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1831                                 tp->ucopy.dma_chan = net_dma_find_channel();
1832
1833                         if (tp->ucopy.dma_chan) {
1834                                 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1835                                         tp->ucopy.dma_chan, skb, offset,
1836                                         msg->msg_iov, used,
1837                                         tp->ucopy.pinned_list);
1838
1839                                 if (tp->ucopy.dma_cookie < 0) {
1840
1841                                         pr_alert("%s: dma_cookie < 0\n",
1842                                                  __func__);
1843
1844                                         /* Exception. Bailout! */
1845                                         if (!copied)
1846                                                 copied = -EFAULT;
1847                                         break;
1848                                 }
1849
1850                                 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1851
1852                                 if ((offset + used) == skb->len)
1853                                         copied_early = true;
1854
1855                         } else
1856 #endif
1857                         {
1858                                 err = skb_copy_datagram_iovec(skb, offset,
1859                                                 msg->msg_iov, used);
1860                                 if (err) {
1861                                         /* Exception. Bailout! */
1862                                         if (!copied)
1863                                                 copied = -EFAULT;
1864                                         break;
1865                                 }
1866                         }
1867                 }
1868
1869                 *seq += used;
1870                 copied += used;
1871                 len -= used;
1872
1873                 tcp_rcv_space_adjust(sk);
1874
1875 skip_copy:
1876                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1877                         tp->urg_data = 0;
1878                         tcp_fast_path_check(sk);
1879                 }
1880                 if (used + offset < skb->len)
1881                         continue;
1882
1883                 if (tcp_hdr(skb)->fin)
1884                         goto found_fin_ok;
1885                 if (!(flags & MSG_PEEK)) {
1886                         sk_eat_skb(sk, skb, copied_early);
1887                         copied_early = false;
1888                 }
1889                 continue;
1890
1891         found_fin_ok:
1892                 /* Process the FIN. */
1893                 ++*seq;
1894                 if (!(flags & MSG_PEEK)) {
1895                         sk_eat_skb(sk, skb, copied_early);
1896                         copied_early = false;
1897                 }
1898                 break;
1899         } while (len > 0);
1900
1901         if (user_recv) {
1902                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1903                         int chunk;
1904
1905                         tp->ucopy.len = copied > 0 ? len : 0;
1906
1907                         tcp_prequeue_process(sk);
1908
1909                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1910                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1911                                 len -= chunk;
1912                                 copied += chunk;
1913                         }
1914                 }
1915
1916                 tp->ucopy.task = NULL;
1917                 tp->ucopy.len = 0;
1918         }
1919
1920 #ifdef CONFIG_NET_DMA
1921         tcp_service_net_dma(sk, true);  /* Wait for queue to drain */
1922         tp->ucopy.dma_chan = NULL;
1923
1924         if (tp->ucopy.pinned_list) {
1925                 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1926                 tp->ucopy.pinned_list = NULL;
1927         }
1928 #endif
1929
1930         /* According to UNIX98, msg_name/msg_namelen are ignored
1931          * on connected socket. I was just happy when found this 8) --ANK
1932          */
1933
1934         /* Clean up data we have read: This will do ACK frames. */
1935         tcp_cleanup_rbuf(sk, copied);
1936
1937         release_sock(sk);
1938         return copied;
1939
1940 out:
1941         release_sock(sk);
1942         return err;
1943
1944 recv_urg:
1945         err = tcp_recv_urg(sk, msg, len, flags);
1946         goto out;
1947
1948 recv_sndq:
1949         err = tcp_peek_sndq(sk, msg, len);
1950         goto out;
1951 }
1952 EXPORT_SYMBOL(tcp_recvmsg);
1953
1954 void tcp_set_state(struct sock *sk, int state)
1955 {
1956         int oldstate = sk->sk_state;
1957
1958         switch (state) {
1959         case TCP_ESTABLISHED:
1960                 if (oldstate != TCP_ESTABLISHED)
1961                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1962                 break;
1963
1964         case TCP_CLOSE:
1965                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1966                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1967
1968                 sk->sk_prot->unhash(sk);
1969                 if (inet_csk(sk)->icsk_bind_hash &&
1970                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1971                         inet_put_port(sk);
1972                 /* fall through */
1973         default:
1974                 if (oldstate == TCP_ESTABLISHED)
1975                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1976         }
1977
1978         /* Change state AFTER socket is unhashed to avoid closed
1979          * socket sitting in hash tables.
1980          */
1981         sk->sk_state = state;
1982
1983 #ifdef STATE_TRACE
1984         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1985 #endif
1986 }
1987 EXPORT_SYMBOL_GPL(tcp_set_state);
1988
1989 /*
1990  *      State processing on a close. This implements the state shift for
1991  *      sending our FIN frame. Note that we only send a FIN for some
1992  *      states. A shutdown() may have already sent the FIN, or we may be
1993  *      closed.
1994  */
1995
1996 static const unsigned char new_state[16] = {
1997   /* current state:        new state:      action:      */
1998   /* (Invalid)          */ TCP_CLOSE,
1999   /* TCP_ESTABLISHED    */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2000   /* TCP_SYN_SENT       */ TCP_CLOSE,
2001   /* TCP_SYN_RECV       */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2002   /* TCP_FIN_WAIT1      */ TCP_FIN_WAIT1,
2003   /* TCP_FIN_WAIT2      */ TCP_FIN_WAIT2,
2004   /* TCP_TIME_WAIT      */ TCP_CLOSE,
2005   /* TCP_CLOSE          */ TCP_CLOSE,
2006   /* TCP_CLOSE_WAIT     */ TCP_LAST_ACK  | TCP_ACTION_FIN,
2007   /* TCP_LAST_ACK       */ TCP_LAST_ACK,
2008   /* TCP_LISTEN         */ TCP_CLOSE,
2009   /* TCP_CLOSING        */ TCP_CLOSING,
2010 };
2011
2012 static int tcp_close_state(struct sock *sk)
2013 {
2014         int next = (int)new_state[sk->sk_state];
2015         int ns = next & TCP_STATE_MASK;
2016
2017         tcp_set_state(sk, ns);
2018
2019         return next & TCP_ACTION_FIN;
2020 }
2021
2022 /*
2023  *      Shutdown the sending side of a connection. Much like close except
2024  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2025  */
2026
2027 void tcp_shutdown(struct sock *sk, int how)
2028 {
2029         /*      We need to grab some memory, and put together a FIN,
2030          *      and then put it into the queue to be sent.
2031          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2032          */
2033         if (!(how & SEND_SHUTDOWN))
2034                 return;
2035
2036         /* If we've already sent a FIN, or it's a closed state, skip this. */
2037         if ((1 << sk->sk_state) &
2038             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2039              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2040                 /* Clear out any half completed packets.  FIN if needed. */
2041                 if (tcp_close_state(sk))
2042                         tcp_send_fin(sk);
2043         }
2044 }
2045 EXPORT_SYMBOL(tcp_shutdown);
2046
2047 bool tcp_check_oom(struct sock *sk, int shift)
2048 {
2049         bool too_many_orphans, out_of_socket_memory;
2050
2051         too_many_orphans = tcp_too_many_orphans(sk, shift);
2052         out_of_socket_memory = tcp_out_of_memory(sk);
2053
2054         if (too_many_orphans)
2055                 net_info_ratelimited("too many orphaned sockets\n");
2056         if (out_of_socket_memory)
2057                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2058         return too_many_orphans || out_of_socket_memory;
2059 }
2060
2061 void tcp_close(struct sock *sk, long timeout)
2062 {
2063         struct sk_buff *skb;
2064         int data_was_unread = 0;
2065         int state;
2066
2067         lock_sock(sk);
2068         sk->sk_shutdown = SHUTDOWN_MASK;
2069
2070         if (sk->sk_state == TCP_LISTEN) {
2071                 tcp_set_state(sk, TCP_CLOSE);
2072
2073                 /* Special case. */
2074                 inet_csk_listen_stop(sk);
2075
2076                 goto adjudge_to_death;
2077         }
2078
2079         /*  We need to flush the recv. buffs.  We do this only on the
2080          *  descriptor close, not protocol-sourced closes, because the
2081          *  reader process may not have drained the data yet!
2082          */
2083         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2084                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
2085                           tcp_hdr(skb)->fin;
2086                 data_was_unread += len;
2087                 __kfree_skb(skb);
2088         }
2089
2090         sk_mem_reclaim(sk);
2091
2092         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2093         if (sk->sk_state == TCP_CLOSE)
2094                 goto adjudge_to_death;
2095
2096         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2097          * data was lost. To witness the awful effects of the old behavior of
2098          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2099          * GET in an FTP client, suspend the process, wait for the client to
2100          * advertise a zero window, then kill -9 the FTP client, wheee...
2101          * Note: timeout is always zero in such a case.
2102          */
2103         if (unlikely(tcp_sk(sk)->repair)) {
2104                 sk->sk_prot->disconnect(sk, 0);
2105         } else if (data_was_unread) {
2106                 /* Unread data was tossed, zap the connection. */
2107                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2108                 tcp_set_state(sk, TCP_CLOSE);
2109                 tcp_send_active_reset(sk, sk->sk_allocation);
2110         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2111                 /* Check zero linger _after_ checking for unread data. */
2112                 sk->sk_prot->disconnect(sk, 0);
2113                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2114         } else if (tcp_close_state(sk)) {
2115                 /* We FIN if the application ate all the data before
2116                  * zapping the connection.
2117                  */
2118
2119                 /* RED-PEN. Formally speaking, we have broken TCP state
2120                  * machine. State transitions:
2121                  *
2122                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2123                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2124                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2125                  *
2126                  * are legal only when FIN has been sent (i.e. in window),
2127                  * rather than queued out of window. Purists blame.
2128                  *
2129                  * F.e. "RFC state" is ESTABLISHED,
2130                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2131                  *
2132                  * The visible declinations are that sometimes
2133                  * we enter time-wait state, when it is not required really
2134                  * (harmless), do not send active resets, when they are
2135                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2136                  * they look as CLOSING or LAST_ACK for Linux)
2137                  * Probably, I missed some more holelets.
2138                  *                                              --ANK
2139                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2140                  * in a single packet! (May consider it later but will
2141                  * probably need API support or TCP_CORK SYN-ACK until
2142                  * data is written and socket is closed.)
2143                  */
2144                 tcp_send_fin(sk);
2145         }
2146
2147         sk_stream_wait_close(sk, timeout);
2148
2149 adjudge_to_death:
2150         state = sk->sk_state;
2151         sock_hold(sk);
2152         sock_orphan(sk);
2153
2154         /* It is the last release_sock in its life. It will remove backlog. */
2155         release_sock(sk);
2156
2157
2158         /* Now socket is owned by kernel and we acquire BH lock
2159            to finish close. No need to check for user refs.
2160          */
2161         local_bh_disable();
2162         bh_lock_sock(sk);
2163         WARN_ON(sock_owned_by_user(sk));
2164
2165         percpu_counter_inc(sk->sk_prot->orphan_count);
2166
2167         /* Have we already been destroyed by a softirq or backlog? */
2168         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2169                 goto out;
2170
2171         /*      This is a (useful) BSD violating of the RFC. There is a
2172          *      problem with TCP as specified in that the other end could
2173          *      keep a socket open forever with no application left this end.
2174          *      We use a 3 minute timeout (about the same as BSD) then kill
2175          *      our end. If they send after that then tough - BUT: long enough
2176          *      that we won't make the old 4*rto = almost no time - whoops
2177          *      reset mistake.
2178          *
2179          *      Nope, it was not mistake. It is really desired behaviour
2180          *      f.e. on http servers, when such sockets are useless, but
2181          *      consume significant resources. Let's do it with special
2182          *      linger2 option.                                 --ANK
2183          */
2184
2185         if (sk->sk_state == TCP_FIN_WAIT2) {
2186                 struct tcp_sock *tp = tcp_sk(sk);
2187                 if (tp->linger2 < 0) {
2188                         tcp_set_state(sk, TCP_CLOSE);
2189                         tcp_send_active_reset(sk, GFP_ATOMIC);
2190                         NET_INC_STATS_BH(sock_net(sk),
2191                                         LINUX_MIB_TCPABORTONLINGER);
2192                 } else {
2193                         const int tmo = tcp_fin_time(sk);
2194
2195                         if (tmo > TCP_TIMEWAIT_LEN) {
2196                                 inet_csk_reset_keepalive_timer(sk,
2197                                                 tmo - TCP_TIMEWAIT_LEN);
2198                         } else {
2199                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2200                                 goto out;
2201                         }
2202                 }
2203         }
2204         if (sk->sk_state != TCP_CLOSE) {
2205                 sk_mem_reclaim(sk);
2206                 if (tcp_check_oom(sk, 0)) {
2207                         tcp_set_state(sk, TCP_CLOSE);
2208                         tcp_send_active_reset(sk, GFP_ATOMIC);
2209                         NET_INC_STATS_BH(sock_net(sk),
2210                                         LINUX_MIB_TCPABORTONMEMORY);
2211                 }
2212         }
2213
2214         if (sk->sk_state == TCP_CLOSE) {
2215                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2216                 /* We could get here with a non-NULL req if the socket is
2217                  * aborted (e.g., closed with unread data) before 3WHS
2218                  * finishes.
2219                  */
2220                 if (req != NULL)
2221                         reqsk_fastopen_remove(sk, req, false);
2222                 inet_csk_destroy_sock(sk);
2223         }
2224         /* Otherwise, socket is reprieved until protocol close. */
2225
2226 out:
2227         bh_unlock_sock(sk);
2228         local_bh_enable();
2229         sock_put(sk);
2230 }
2231 EXPORT_SYMBOL(tcp_close);
2232
2233 /* These states need RST on ABORT according to RFC793 */
2234
2235 static inline bool tcp_need_reset(int state)
2236 {
2237         return (1 << state) &
2238                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2239                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2240 }
2241
2242 int tcp_disconnect(struct sock *sk, int flags)
2243 {
2244         struct inet_sock *inet = inet_sk(sk);
2245         struct inet_connection_sock *icsk = inet_csk(sk);
2246         struct tcp_sock *tp = tcp_sk(sk);
2247         int err = 0;
2248         int old_state = sk->sk_state;
2249
2250         if (old_state != TCP_CLOSE)
2251                 tcp_set_state(sk, TCP_CLOSE);
2252
2253         /* ABORT function of RFC793 */
2254         if (old_state == TCP_LISTEN) {
2255                 inet_csk_listen_stop(sk);
2256         } else if (unlikely(tp->repair)) {
2257                 sk->sk_err = ECONNABORTED;
2258         } else if (tcp_need_reset(old_state) ||
2259                    (tp->snd_nxt != tp->write_seq &&
2260                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2261                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2262                  * states
2263                  */
2264                 tcp_send_active_reset(sk, gfp_any());
2265                 sk->sk_err = ECONNRESET;
2266         } else if (old_state == TCP_SYN_SENT)
2267                 sk->sk_err = ECONNRESET;
2268
2269         tcp_clear_xmit_timers(sk);
2270         __skb_queue_purge(&sk->sk_receive_queue);
2271         tcp_write_queue_purge(sk);
2272         __skb_queue_purge(&tp->out_of_order_queue);
2273 #ifdef CONFIG_NET_DMA
2274         __skb_queue_purge(&sk->sk_async_wait_queue);
2275 #endif
2276
2277         inet->inet_dport = 0;
2278
2279         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2280                 inet_reset_saddr(sk);
2281
2282         sk->sk_shutdown = 0;
2283         sock_reset_flag(sk, SOCK_DONE);
2284         tp->srtt = 0;
2285         if ((tp->write_seq += tp->max_window + 2) == 0)
2286                 tp->write_seq = 1;
2287         icsk->icsk_backoff = 0;
2288         tp->snd_cwnd = 2;
2289         icsk->icsk_probes_out = 0;
2290         tp->packets_out = 0;
2291         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2292         tp->snd_cwnd_cnt = 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         case TCP_TIMESTAMP:
2717                 if (!tp->repair)
2718                         err = -EPERM;
2719                 else
2720                         tp->tsoffset = val - tcp_time_stamp;
2721                 break;
2722         default:
2723                 err = -ENOPROTOOPT;
2724                 break;
2725         }
2726
2727         release_sock(sk);
2728         return err;
2729 }
2730
2731 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2732                    unsigned int optlen)
2733 {
2734         const struct inet_connection_sock *icsk = inet_csk(sk);
2735
2736         if (level != SOL_TCP)
2737                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2738                                                      optval, optlen);
2739         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2740 }
2741 EXPORT_SYMBOL(tcp_setsockopt);
2742
2743 #ifdef CONFIG_COMPAT
2744 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2745                           char __user *optval, unsigned int optlen)
2746 {
2747         if (level != SOL_TCP)
2748                 return inet_csk_compat_setsockopt(sk, level, optname,
2749                                                   optval, optlen);
2750         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2751 }
2752 EXPORT_SYMBOL(compat_tcp_setsockopt);
2753 #endif
2754
2755 /* Return information about state of tcp endpoint in API format. */
2756 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2757 {
2758         const struct tcp_sock *tp = tcp_sk(sk);
2759         const struct inet_connection_sock *icsk = inet_csk(sk);
2760         u32 now = tcp_time_stamp;
2761
2762         memset(info, 0, sizeof(*info));
2763
2764         info->tcpi_state = sk->sk_state;
2765         info->tcpi_ca_state = icsk->icsk_ca_state;
2766         info->tcpi_retransmits = icsk->icsk_retransmits;
2767         info->tcpi_probes = icsk->icsk_probes_out;
2768         info->tcpi_backoff = icsk->icsk_backoff;
2769
2770         if (tp->rx_opt.tstamp_ok)
2771                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2772         if (tcp_is_sack(tp))
2773                 info->tcpi_options |= TCPI_OPT_SACK;
2774         if (tp->rx_opt.wscale_ok) {
2775                 info->tcpi_options |= TCPI_OPT_WSCALE;
2776                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2777                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2778         }
2779
2780         if (tp->ecn_flags & TCP_ECN_OK)
2781                 info->tcpi_options |= TCPI_OPT_ECN;
2782         if (tp->ecn_flags & TCP_ECN_SEEN)
2783                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2784         if (tp->syn_data_acked)
2785                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2786
2787         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2788         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2789         info->tcpi_snd_mss = tp->mss_cache;
2790         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2791
2792         if (sk->sk_state == TCP_LISTEN) {
2793                 info->tcpi_unacked = sk->sk_ack_backlog;
2794                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2795         } else {
2796                 info->tcpi_unacked = tp->packets_out;
2797                 info->tcpi_sacked = tp->sacked_out;
2798         }
2799         info->tcpi_lost = tp->lost_out;
2800         info->tcpi_retrans = tp->retrans_out;
2801         info->tcpi_fackets = tp->fackets_out;
2802
2803         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2804         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2805         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2806
2807         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2808         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2809         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2810         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2811         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2812         info->tcpi_snd_cwnd = tp->snd_cwnd;
2813         info->tcpi_advmss = tp->advmss;
2814         info->tcpi_reordering = tp->reordering;
2815
2816         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2817         info->tcpi_rcv_space = tp->rcvq_space.space;
2818
2819         info->tcpi_total_retrans = tp->total_retrans;
2820 }
2821 EXPORT_SYMBOL_GPL(tcp_get_info);
2822
2823 static int do_tcp_getsockopt(struct sock *sk, int level,
2824                 int optname, char __user *optval, int __user *optlen)
2825 {
2826         struct inet_connection_sock *icsk = inet_csk(sk);
2827         struct tcp_sock *tp = tcp_sk(sk);
2828         int val, len;
2829
2830         if (get_user(len, optlen))
2831                 return -EFAULT;
2832
2833         len = min_t(unsigned int, len, sizeof(int));
2834
2835         if (len < 0)
2836                 return -EINVAL;
2837
2838         switch (optname) {
2839         case TCP_MAXSEG:
2840                 val = tp->mss_cache;
2841                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2842                         val = tp->rx_opt.user_mss;
2843                 if (tp->repair)
2844                         val = tp->rx_opt.mss_clamp;
2845                 break;
2846         case TCP_NODELAY:
2847                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2848                 break;
2849         case TCP_CORK:
2850                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2851                 break;
2852         case TCP_KEEPIDLE:
2853                 val = keepalive_time_when(tp) / HZ;
2854                 break;
2855         case TCP_KEEPINTVL:
2856                 val = keepalive_intvl_when(tp) / HZ;
2857                 break;
2858         case TCP_KEEPCNT:
2859                 val = keepalive_probes(tp);
2860                 break;
2861         case TCP_SYNCNT:
2862                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2863                 break;
2864         case TCP_LINGER2:
2865                 val = tp->linger2;
2866                 if (val >= 0)
2867                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2868                 break;
2869         case TCP_DEFER_ACCEPT:
2870                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2871                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2872                 break;
2873         case TCP_WINDOW_CLAMP:
2874                 val = tp->window_clamp;
2875                 break;
2876         case TCP_INFO: {
2877                 struct tcp_info info;
2878
2879                 if (get_user(len, optlen))
2880                         return -EFAULT;
2881
2882                 tcp_get_info(sk, &info);
2883
2884                 len = min_t(unsigned int, len, sizeof(info));
2885                 if (put_user(len, optlen))
2886                         return -EFAULT;
2887                 if (copy_to_user(optval, &info, len))
2888                         return -EFAULT;
2889                 return 0;
2890         }
2891         case TCP_QUICKACK:
2892                 val = !icsk->icsk_ack.pingpong;
2893                 break;
2894
2895         case TCP_CONGESTION:
2896                 if (get_user(len, optlen))
2897                         return -EFAULT;
2898                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2899                 if (put_user(len, optlen))
2900                         return -EFAULT;
2901                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2902                         return -EFAULT;
2903                 return 0;
2904
2905         case TCP_COOKIE_TRANSACTIONS: {
2906                 struct tcp_cookie_transactions ctd;
2907                 struct tcp_cookie_values *cvp = tp->cookie_values;
2908
2909                 if (get_user(len, optlen))
2910                         return -EFAULT;
2911                 if (len < sizeof(ctd))
2912                         return -EINVAL;
2913
2914                 memset(&ctd, 0, sizeof(ctd));
2915                 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2916                                    TCP_COOKIE_IN_ALWAYS : 0)
2917                                 | (tp->rx_opt.cookie_out_never ?
2918                                    TCP_COOKIE_OUT_NEVER : 0);
2919
2920                 if (cvp != NULL) {
2921                         ctd.tcpct_flags |= (cvp->s_data_in ?
2922                                             TCP_S_DATA_IN : 0)
2923                                          | (cvp->s_data_out ?
2924                                             TCP_S_DATA_OUT : 0);
2925
2926                         ctd.tcpct_cookie_desired = cvp->cookie_desired;
2927                         ctd.tcpct_s_data_desired = cvp->s_data_desired;
2928
2929                         memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2930                                cvp->cookie_pair_size);
2931                         ctd.tcpct_used = cvp->cookie_pair_size;
2932                 }
2933
2934                 if (put_user(sizeof(ctd), optlen))
2935                         return -EFAULT;
2936                 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2937                         return -EFAULT;
2938                 return 0;
2939         }
2940         case TCP_THIN_LINEAR_TIMEOUTS:
2941                 val = tp->thin_lto;
2942                 break;
2943         case TCP_THIN_DUPACK:
2944                 val = tp->thin_dupack;
2945                 break;
2946
2947         case TCP_REPAIR:
2948                 val = tp->repair;
2949                 break;
2950
2951         case TCP_REPAIR_QUEUE:
2952                 if (tp->repair)
2953                         val = tp->repair_queue;
2954                 else
2955                         return -EINVAL;
2956                 break;
2957
2958         case TCP_QUEUE_SEQ:
2959                 if (tp->repair_queue == TCP_SEND_QUEUE)
2960                         val = tp->write_seq;
2961                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2962                         val = tp->rcv_nxt;
2963                 else
2964                         return -EINVAL;
2965                 break;
2966
2967         case TCP_USER_TIMEOUT:
2968                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2969                 break;
2970         case TCP_TIMESTAMP:
2971                 val = tcp_time_stamp + tp->tsoffset;
2972                 break;
2973         default:
2974                 return -ENOPROTOOPT;
2975         }
2976
2977         if (put_user(len, optlen))
2978                 return -EFAULT;
2979         if (copy_to_user(optval, &val, len))
2980                 return -EFAULT;
2981         return 0;
2982 }
2983
2984 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2985                    int __user *optlen)
2986 {
2987         struct inet_connection_sock *icsk = inet_csk(sk);
2988
2989         if (level != SOL_TCP)
2990                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2991                                                      optval, optlen);
2992         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2993 }
2994 EXPORT_SYMBOL(tcp_getsockopt);
2995
2996 #ifdef CONFIG_COMPAT
2997 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2998                           char __user *optval, int __user *optlen)
2999 {
3000         if (level != SOL_TCP)
3001                 return inet_csk_compat_getsockopt(sk, level, optname,
3002                                                   optval, optlen);
3003         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3004 }
3005 EXPORT_SYMBOL(compat_tcp_getsockopt);
3006 #endif
3007
3008 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
3009         netdev_features_t features)
3010 {
3011         struct sk_buff *segs = ERR_PTR(-EINVAL);
3012         struct tcphdr *th;
3013         unsigned int thlen;
3014         unsigned int seq;
3015         __be32 delta;
3016         unsigned int oldlen;
3017         unsigned int mss;
3018
3019         if (!pskb_may_pull(skb, sizeof(*th)))
3020                 goto out;
3021
3022         th = tcp_hdr(skb);
3023         thlen = th->doff * 4;
3024         if (thlen < sizeof(*th))
3025                 goto out;
3026
3027         if (!pskb_may_pull(skb, thlen))
3028                 goto out;
3029
3030         oldlen = (u16)~skb->len;
3031         __skb_pull(skb, thlen);
3032
3033         mss = skb_shinfo(skb)->gso_size;
3034         if (unlikely(skb->len <= mss))
3035                 goto out;
3036
3037         if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
3038                 /* Packet is from an untrusted source, reset gso_segs. */
3039                 int type = skb_shinfo(skb)->gso_type;
3040
3041                 if (unlikely(type &
3042                              ~(SKB_GSO_TCPV4 |
3043                                SKB_GSO_DODGY |
3044                                SKB_GSO_TCP_ECN |
3045                                SKB_GSO_TCPV6 |
3046                                0) ||
3047                              !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
3048                         goto out;
3049
3050                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
3051
3052                 segs = NULL;
3053                 goto out;
3054         }
3055
3056         segs = skb_segment(skb, features);
3057         if (IS_ERR(segs))
3058                 goto out;
3059
3060         delta = htonl(oldlen + (thlen + mss));
3061
3062         skb = segs;
3063         th = tcp_hdr(skb);
3064         seq = ntohl(th->seq);
3065
3066         do {
3067                 th->fin = th->psh = 0;
3068
3069                 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3070                                        (__force u32)delta));
3071                 if (skb->ip_summed != CHECKSUM_PARTIAL)
3072                         th->check =
3073                              csum_fold(csum_partial(skb_transport_header(skb),
3074                                                     thlen, skb->csum));
3075
3076                 seq += mss;
3077                 skb = skb->next;
3078                 th = tcp_hdr(skb);
3079
3080                 th->seq = htonl(seq);
3081                 th->cwr = 0;
3082         } while (skb->next);
3083
3084         delta = htonl(oldlen + (skb->tail - skb->transport_header) +
3085                       skb->data_len);
3086         th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
3087                                 (__force u32)delta));
3088         if (skb->ip_summed != CHECKSUM_PARTIAL)
3089                 th->check = csum_fold(csum_partial(skb_transport_header(skb),
3090                                                    thlen, skb->csum));
3091
3092 out:
3093         return segs;
3094 }
3095 EXPORT_SYMBOL(tcp_tso_segment);
3096
3097 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
3098 {
3099         struct sk_buff **pp = NULL;
3100         struct sk_buff *p;
3101         struct tcphdr *th;
3102         struct tcphdr *th2;
3103         unsigned int len;
3104         unsigned int thlen;
3105         __be32 flags;
3106         unsigned int mss = 1;
3107         unsigned int hlen;
3108         unsigned int off;
3109         int flush = 1;
3110         int i;
3111
3112         off = skb_gro_offset(skb);
3113         hlen = off + sizeof(*th);
3114         th = skb_gro_header_fast(skb, off);
3115         if (skb_gro_header_hard(skb, hlen)) {
3116                 th = skb_gro_header_slow(skb, hlen, off);
3117                 if (unlikely(!th))
3118                         goto out;
3119         }
3120
3121         thlen = th->doff * 4;
3122         if (thlen < sizeof(*th))
3123                 goto out;
3124
3125         hlen = off + thlen;
3126         if (skb_gro_header_hard(skb, hlen)) {
3127                 th = skb_gro_header_slow(skb, hlen, off);
3128                 if (unlikely(!th))
3129                         goto out;
3130         }
3131
3132         skb_gro_pull(skb, thlen);
3133
3134         len = skb_gro_len(skb);
3135         flags = tcp_flag_word(th);
3136
3137         for (; (p = *head); head = &p->next) {
3138                 if (!NAPI_GRO_CB(p)->same_flow)
3139                         continue;
3140
3141                 th2 = tcp_hdr(p);
3142
3143                 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
3144                         NAPI_GRO_CB(p)->same_flow = 0;
3145                         continue;
3146                 }
3147
3148                 goto found;
3149         }
3150
3151         goto out_check_final;
3152
3153 found:
3154         flush = NAPI_GRO_CB(p)->flush;
3155         flush |= (__force int)(flags & TCP_FLAG_CWR);
3156         flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
3157                   ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
3158         flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
3159         for (i = sizeof(*th); i < thlen; i += 4)
3160                 flush |= *(u32 *)((u8 *)th + i) ^
3161                          *(u32 *)((u8 *)th2 + i);
3162
3163         mss = skb_shinfo(p)->gso_size;
3164
3165         flush |= (len - 1) >= mss;
3166         flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
3167
3168         if (flush || skb_gro_receive(head, skb)) {
3169                 mss = 1;
3170                 goto out_check_final;
3171         }
3172
3173         p = *head;
3174         th2 = tcp_hdr(p);
3175         tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
3176
3177 out_check_final:
3178         flush = len < mss;
3179         flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
3180                                         TCP_FLAG_RST | TCP_FLAG_SYN |
3181                                         TCP_FLAG_FIN));
3182
3183         if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
3184                 pp = head;
3185
3186 out:
3187         NAPI_GRO_CB(skb)->flush |= flush;
3188
3189         return pp;
3190 }
3191 EXPORT_SYMBOL(tcp_gro_receive);
3192
3193 int tcp_gro_complete(struct sk_buff *skb)
3194 {
3195         struct tcphdr *th = tcp_hdr(skb);
3196
3197         skb->csum_start = skb_transport_header(skb) - skb->head;
3198         skb->csum_offset = offsetof(struct tcphdr, check);
3199         skb->ip_summed = CHECKSUM_PARTIAL;
3200
3201         skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
3202
3203         if (th->cwr)
3204                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
3205
3206         return 0;
3207 }
3208 EXPORT_SYMBOL(tcp_gro_complete);
3209
3210 #ifdef CONFIG_TCP_MD5SIG
3211 static unsigned long tcp_md5sig_users;
3212 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
3213 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
3214
3215 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
3216 {
3217         int cpu;
3218
3219         for_each_possible_cpu(cpu) {
3220                 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
3221
3222                 if (p->md5_desc.tfm)
3223                         crypto_free_hash(p->md5_desc.tfm);
3224         }
3225         free_percpu(pool);
3226 }
3227
3228 void tcp_free_md5sig_pool(void)
3229 {
3230         struct tcp_md5sig_pool __percpu *pool = NULL;
3231
3232         spin_lock_bh(&tcp_md5sig_pool_lock);
3233         if (--tcp_md5sig_users == 0) {
3234                 pool = tcp_md5sig_pool;
3235                 tcp_md5sig_pool = NULL;
3236         }
3237         spin_unlock_bh(&tcp_md5sig_pool_lock);
3238         if (pool)
3239                 __tcp_free_md5sig_pool(pool);
3240 }
3241 EXPORT_SYMBOL(tcp_free_md5sig_pool);
3242
3243 static struct tcp_md5sig_pool __percpu *
3244 __tcp_alloc_md5sig_pool(struct sock *sk)
3245 {
3246         int cpu;
3247         struct tcp_md5sig_pool __percpu *pool;
3248
3249         pool = alloc_percpu(struct tcp_md5sig_pool);
3250         if (!pool)
3251                 return NULL;
3252
3253         for_each_possible_cpu(cpu) {
3254                 struct crypto_hash *hash;
3255
3256                 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
3257                 if (IS_ERR_OR_NULL(hash))
3258                         goto out_free;
3259
3260                 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
3261         }
3262         return pool;
3263 out_free:
3264         __tcp_free_md5sig_pool(pool);
3265         return NULL;
3266 }
3267
3268 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
3269 {
3270         struct tcp_md5sig_pool __percpu *pool;
3271         bool alloc = false;
3272
3273 retry:
3274         spin_lock_bh(&tcp_md5sig_pool_lock);
3275         pool = tcp_md5sig_pool;
3276         if (tcp_md5sig_users++ == 0) {
3277                 alloc = true;
3278                 spin_unlock_bh(&tcp_md5sig_pool_lock);
3279         } else if (!pool) {
3280                 tcp_md5sig_users--;
3281                 spin_unlock_bh(&tcp_md5sig_pool_lock);
3282                 cpu_relax();
3283                 goto retry;
3284         } else
3285                 spin_unlock_bh(&tcp_md5sig_pool_lock);
3286
3287         if (alloc) {
3288                 /* we cannot hold spinlock here because this may sleep. */
3289                 struct tcp_md5sig_pool __percpu *p;
3290
3291                 p = __tcp_alloc_md5sig_pool(sk);
3292                 spin_lock_bh(&tcp_md5sig_pool_lock);
3293                 if (!p) {
3294                         tcp_md5sig_users--;
3295                         spin_unlock_bh(&tcp_md5sig_pool_lock);
3296                         return NULL;
3297                 }
3298                 pool = tcp_md5sig_pool;
3299                 if (pool) {
3300                         /* oops, it has already been assigned. */
3301                         spin_unlock_bh(&tcp_md5sig_pool_lock);
3302                         __tcp_free_md5sig_pool(p);
3303                 } else {
3304                         tcp_md5sig_pool = pool = p;
3305                         spin_unlock_bh(&tcp_md5sig_pool_lock);
3306                 }
3307         }
3308         return pool;
3309 }
3310 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3311
3312
3313 /**
3314  *      tcp_get_md5sig_pool - get md5sig_pool for this user
3315  *
3316  *      We use percpu structure, so if we succeed, we exit with preemption
3317  *      and BH disabled, to make sure another thread or softirq handling
3318  *      wont try to get same context.
3319  */
3320 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3321 {
3322         struct tcp_md5sig_pool __percpu *p;
3323
3324         local_bh_disable();
3325
3326         spin_lock(&tcp_md5sig_pool_lock);
3327         p = tcp_md5sig_pool;
3328         if (p)
3329                 tcp_md5sig_users++;
3330         spin_unlock(&tcp_md5sig_pool_lock);
3331
3332         if (p)
3333                 return this_cpu_ptr(p);
3334
3335         local_bh_enable();
3336         return NULL;
3337 }
3338 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3339
3340 void tcp_put_md5sig_pool(void)
3341 {
3342         local_bh_enable();
3343         tcp_free_md5sig_pool();
3344 }
3345 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3346
3347 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3348                         const struct tcphdr *th)
3349 {
3350         struct scatterlist sg;
3351         struct tcphdr hdr;
3352         int err;
3353
3354         /* We are not allowed to change tcphdr, make a local copy */
3355         memcpy(&hdr, th, sizeof(hdr));
3356         hdr.check = 0;
3357