Version:  2.0.40 2.2.26 2.4.37 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5

Linux/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/inet_diag.h>
256 #include <linux/init.h>
257 #include <linux/fs.h>
258 #include <linux/skbuff.h>
259 #include <linux/scatterlist.h>
260 #include <linux/splice.h>
261 #include <linux/net.h>
262 #include <linux/socket.h>
263 #include <linux/random.h>
264 #include <linux/bootmem.h>
265 #include <linux/highmem.h>
266 #include <linux/swap.h>
267 #include <linux/cache.h>
268 #include <linux/err.h>
269 #include <linux/crypto.h>
270 #include <linux/time.h>
271 #include <linux/slab.h>
272 
273 #include <net/icmp.h>
274 #include <net/inet_common.h>
275 #include <net/tcp.h>
276 #include <net/xfrm.h>
277 #include <net/ip.h>
278 #include <net/sock.h>
279 
280 #include <asm/uaccess.h>
281 #include <asm/ioctls.h>
282 #include <asm/unaligned.h>
283 #include <net/busy_poll.h>
284 
285 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
286 
287 int sysctl_tcp_min_tso_segs __read_mostly = 2;
288 
289 int sysctl_tcp_autocorking __read_mostly = 1;
290 
291 struct percpu_counter tcp_orphan_count;
292 EXPORT_SYMBOL_GPL(tcp_orphan_count);
293 
294 long sysctl_tcp_mem[3] __read_mostly;
295 int sysctl_tcp_wmem[3] __read_mostly;
296 int sysctl_tcp_rmem[3] __read_mostly;
297 
298 EXPORT_SYMBOL(sysctl_tcp_mem);
299 EXPORT_SYMBOL(sysctl_tcp_rmem);
300 EXPORT_SYMBOL(sysctl_tcp_wmem);
301 
302 atomic_long_t tcp_memory_allocated;     /* Current allocated memory. */
303 EXPORT_SYMBOL(tcp_memory_allocated);
304 
305 /*
306  * Current number of TCP sockets.
307  */
308 struct percpu_counter tcp_sockets_allocated;
309 EXPORT_SYMBOL(tcp_sockets_allocated);
310 
311 /*
312  * TCP splice context
313  */
314 struct tcp_splice_state {
315         struct pipe_inode_info *pipe;
316         size_t len;
317         unsigned int flags;
318 };
319 
320 /*
321  * Pressure flag: try to collapse.
322  * Technical note: it is used by multiple contexts non atomically.
323  * All the __sk_mem_schedule() is of this nature: accounting
324  * is strict, actions are advisory and have some latency.
325  */
326 int tcp_memory_pressure __read_mostly;
327 EXPORT_SYMBOL(tcp_memory_pressure);
328 
329 void tcp_enter_memory_pressure(struct sock *sk)
330 {
331         if (!tcp_memory_pressure) {
332                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
333                 tcp_memory_pressure = 1;
334         }
335 }
336 EXPORT_SYMBOL(tcp_enter_memory_pressure);
337 
338 /* Convert seconds to retransmits based on initial and max timeout */
339 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
340 {
341         u8 res = 0;
342 
343         if (seconds > 0) {
344                 int period = timeout;
345 
346                 res = 1;
347                 while (seconds > period && res < 255) {
348                         res++;
349                         timeout <<= 1;
350                         if (timeout > rto_max)
351                                 timeout = rto_max;
352                         period += timeout;
353                 }
354         }
355         return res;
356 }
357 
358 /* Convert retransmits to seconds based on initial and max timeout */
359 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
360 {
361         int period = 0;
362 
363         if (retrans > 0) {
364                 period = timeout;
365                 while (--retrans) {
366                         timeout <<= 1;
367                         if (timeout > rto_max)
368                                 timeout = rto_max;
369                         period += timeout;
370                 }
371         }
372         return period;
373 }
374 
375 /* Address-family independent initialization for a tcp_sock.
376  *
377  * NOTE: A lot of things set to zero explicitly by call to
378  *       sk_alloc() so need not be done here.
379  */
380 void tcp_init_sock(struct sock *sk)
381 {
382         struct inet_connection_sock *icsk = inet_csk(sk);
383         struct tcp_sock *tp = tcp_sk(sk);
384 
385         __skb_queue_head_init(&tp->out_of_order_queue);
386         tcp_init_xmit_timers(sk);
387         tcp_prequeue_init(tp);
388         INIT_LIST_HEAD(&tp->tsq_node);
389 
390         icsk->icsk_rto = TCP_TIMEOUT_INIT;
391         tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
392         tp->rtt_min[0].rtt = ~0U;
393 
394         /* So many TCP implementations out there (incorrectly) count the
395          * initial SYN frame in their delayed-ACK and congestion control
396          * algorithms that we must have the following bandaid to talk
397          * efficiently to them.  -DaveM
398          */
399         tp->snd_cwnd = TCP_INIT_CWND;
400 
401         /* See draft-stevens-tcpca-spec-01 for discussion of the
402          * initialization of these values.
403          */
404         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
405         tp->snd_cwnd_clamp = ~0;
406         tp->mss_cache = TCP_MSS_DEFAULT;
407         u64_stats_init(&tp->syncp);
408 
409         tp->reordering = sysctl_tcp_reordering;
410         tcp_enable_early_retrans(tp);
411         tcp_assign_congestion_control(sk);
412 
413         tp->tsoffset = 0;
414 
415         sk->sk_state = TCP_CLOSE;
416 
417         sk->sk_write_space = sk_stream_write_space;
418         sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
419 
420         icsk->icsk_sync_mss = tcp_sync_mss;
421 
422         sk->sk_sndbuf = sysctl_tcp_wmem[1];
423         sk->sk_rcvbuf = sysctl_tcp_rmem[1];
424 
425         local_bh_disable();
426         if (mem_cgroup_sockets_enabled)
427                 sock_update_memcg(sk);
428         sk_sockets_allocated_inc(sk);
429         local_bh_enable();
430 }
431 EXPORT_SYMBOL(tcp_init_sock);
432 
433 static void tcp_tx_timestamp(struct sock *sk, struct sk_buff *skb)
434 {
435         if (sk->sk_tsflags) {
436                 struct skb_shared_info *shinfo = skb_shinfo(skb);
437 
438                 sock_tx_timestamp(sk, &shinfo->tx_flags);
439                 if (shinfo->tx_flags & SKBTX_ANY_TSTAMP)
440                         shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
441         }
442 }
443 
444 /*
445  *      Wait for a TCP event.
446  *
447  *      Note that we don't need to lock the socket, as the upper poll layers
448  *      take care of normal races (between the test and the event) and we don't
449  *      go look at any of the socket buffers directly.
450  */
451 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
452 {
453         unsigned int mask;
454         struct sock *sk = sock->sk;
455         const struct tcp_sock *tp = tcp_sk(sk);
456         int state;
457 
458         sock_rps_record_flow(sk);
459 
460         sock_poll_wait(file, sk_sleep(sk), wait);
461 
462         state = sk_state_load(sk);
463         if (state == TCP_LISTEN)
464                 return inet_csk_listen_poll(sk);
465 
466         /* Socket is not locked. We are protected from async events
467          * by poll logic and correct handling of state changes
468          * made by other threads is impossible in any case.
469          */
470 
471         mask = 0;
472 
473         /*
474          * POLLHUP is certainly not done right. But poll() doesn't
475          * have a notion of HUP in just one direction, and for a
476          * socket the read side is more interesting.
477          *
478          * Some poll() documentation says that POLLHUP is incompatible
479          * with the POLLOUT/POLLWR flags, so somebody should check this
480          * all. But careful, it tends to be safer to return too many
481          * bits than too few, and you can easily break real applications
482          * if you don't tell them that something has hung up!
483          *
484          * Check-me.
485          *
486          * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
487          * our fs/select.c). It means that after we received EOF,
488          * poll always returns immediately, making impossible poll() on write()
489          * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
490          * if and only if shutdown has been made in both directions.
491          * Actually, it is interesting to look how Solaris and DUX
492          * solve this dilemma. I would prefer, if POLLHUP were maskable,
493          * then we could set it on SND_SHUTDOWN. BTW examples given
494          * in Stevens' books assume exactly this behaviour, it explains
495          * why POLLHUP is incompatible with POLLOUT.    --ANK
496          *
497          * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
498          * blocking on fresh not-connected or disconnected socket. --ANK
499          */
500         if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
501                 mask |= POLLHUP;
502         if (sk->sk_shutdown & RCV_SHUTDOWN)
503                 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
504 
505         /* Connected or passive Fast Open socket? */
506         if (state != TCP_SYN_SENT &&
507             (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
508                 int target = sock_rcvlowat(sk, 0, INT_MAX);
509 
510                 if (tp->urg_seq == tp->copied_seq &&
511                     !sock_flag(sk, SOCK_URGINLINE) &&
512                     tp->urg_data)
513                         target++;
514 
515                 if (tp->rcv_nxt - tp->copied_seq >= target)
516                         mask |= POLLIN | POLLRDNORM;
517 
518                 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
519                         if (sk_stream_is_writeable(sk)) {
520                                 mask |= POLLOUT | POLLWRNORM;
521                         } else {  /* send SIGIO later */
522                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
523                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
524 
525                                 /* Race breaker. If space is freed after
526                                  * wspace test but before the flags are set,
527                                  * IO signal will be lost. Memory barrier
528                                  * pairs with the input side.
529                                  */
530                                 smp_mb__after_atomic();
531                                 if (sk_stream_is_writeable(sk))
532                                         mask |= POLLOUT | POLLWRNORM;
533                         }
534                 } else
535                         mask |= POLLOUT | POLLWRNORM;
536 
537                 if (tp->urg_data & TCP_URG_VALID)
538                         mask |= POLLPRI;
539         }
540         /* This barrier is coupled with smp_wmb() in tcp_reset() */
541         smp_rmb();
542         if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
543                 mask |= POLLERR;
544 
545         return mask;
546 }
547 EXPORT_SYMBOL(tcp_poll);
548 
549 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
550 {
551         struct tcp_sock *tp = tcp_sk(sk);
552         int answ;
553         bool slow;
554 
555         switch (cmd) {
556         case SIOCINQ:
557                 if (sk->sk_state == TCP_LISTEN)
558                         return -EINVAL;
559 
560                 slow = lock_sock_fast(sk);
561                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
562                         answ = 0;
563                 else if (sock_flag(sk, SOCK_URGINLINE) ||
564                          !tp->urg_data ||
565                          before(tp->urg_seq, tp->copied_seq) ||
566                          !before(tp->urg_seq, tp->rcv_nxt)) {
567 
568                         answ = tp->rcv_nxt - tp->copied_seq;
569 
570                         /* Subtract 1, if FIN was received */
571                         if (answ && sock_flag(sk, SOCK_DONE))
572                                 answ--;
573                 } else
574                         answ = tp->urg_seq - tp->copied_seq;
575                 unlock_sock_fast(sk, slow);
576                 break;
577         case SIOCATMARK:
578                 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
579                 break;
580         case SIOCOUTQ:
581                 if (sk->sk_state == TCP_LISTEN)
582                         return -EINVAL;
583 
584                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
585                         answ = 0;
586                 else
587                         answ = tp->write_seq - tp->snd_una;
588                 break;
589         case SIOCOUTQNSD:
590                 if (sk->sk_state == TCP_LISTEN)
591                         return -EINVAL;
592 
593                 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
594                         answ = 0;
595                 else
596                         answ = tp->write_seq - tp->snd_nxt;
597                 break;
598         default:
599                 return -ENOIOCTLCMD;
600         }
601 
602         return put_user(answ, (int __user *)arg);
603 }
604 EXPORT_SYMBOL(tcp_ioctl);
605 
606 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
607 {
608         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
609         tp->pushed_seq = tp->write_seq;
610 }
611 
612 static inline bool forced_push(const struct tcp_sock *tp)
613 {
614         return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
615 }
616 
617 static void skb_entail(struct sock *sk, struct sk_buff *skb)
618 {
619         struct tcp_sock *tp = tcp_sk(sk);
620         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
621 
622         skb->csum    = 0;
623         tcb->seq     = tcb->end_seq = tp->write_seq;
624         tcb->tcp_flags = TCPHDR_ACK;
625         tcb->sacked  = 0;
626         __skb_header_release(skb);
627         tcp_add_write_queue_tail(sk, skb);
628         sk->sk_wmem_queued += skb->truesize;
629         sk_mem_charge(sk, skb->truesize);
630         if (tp->nonagle & TCP_NAGLE_PUSH)
631                 tp->nonagle &= ~TCP_NAGLE_PUSH;
632 
633         tcp_slow_start_after_idle_check(sk);
634 }
635 
636 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
637 {
638         if (flags & MSG_OOB)
639                 tp->snd_up = tp->write_seq;
640 }
641 
642 /* If a not yet filled skb is pushed, do not send it if
643  * we have data packets in Qdisc or NIC queues :
644  * Because TX completion will happen shortly, it gives a chance
645  * to coalesce future sendmsg() payload into this skb, without
646  * need for a timer, and with no latency trade off.
647  * As packets containing data payload have a bigger truesize
648  * than pure acks (dataless) packets, the last checks prevent
649  * autocorking if we only have an ACK in Qdisc/NIC queues,
650  * or if TX completion was delayed after we processed ACK packet.
651  */
652 static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
653                                 int size_goal)
654 {
655         return skb->len < size_goal &&
656                sysctl_tcp_autocorking &&
657                skb != tcp_write_queue_head(sk) &&
658                atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
659 }
660 
661 static void tcp_push(struct sock *sk, int flags, int mss_now,
662                      int nonagle, int size_goal)
663 {
664         struct tcp_sock *tp = tcp_sk(sk);
665         struct sk_buff *skb;
666 
667         if (!tcp_send_head(sk))
668                 return;
669 
670         skb = tcp_write_queue_tail(sk);
671         if (!(flags & MSG_MORE) || forced_push(tp))
672                 tcp_mark_push(tp, skb);
673 
674         tcp_mark_urg(tp, flags);
675 
676         if (tcp_should_autocork(sk, skb, size_goal)) {
677 
678                 /* avoid atomic op if TSQ_THROTTLED bit is already set */
679                 if (!test_bit(TSQ_THROTTLED, &tp->tsq_flags)) {
680                         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
681                         set_bit(TSQ_THROTTLED, &tp->tsq_flags);
682                 }
683                 /* It is possible TX completion already happened
684                  * before we set TSQ_THROTTLED.
685                  */
686                 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
687                         return;
688         }
689 
690         if (flags & MSG_MORE)
691                 nonagle = TCP_NAGLE_CORK;
692 
693         __tcp_push_pending_frames(sk, mss_now, nonagle);
694 }
695 
696 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
697                                 unsigned int offset, size_t len)
698 {
699         struct tcp_splice_state *tss = rd_desc->arg.data;
700         int ret;
701 
702         ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
703                               min(rd_desc->count, len), tss->flags,
704                               skb_socket_splice);
705         if (ret > 0)
706                 rd_desc->count -= ret;
707         return ret;
708 }
709 
710 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
711 {
712         /* Store TCP splice context information in read_descriptor_t. */
713         read_descriptor_t rd_desc = {
714                 .arg.data = tss,
715                 .count    = tss->len,
716         };
717 
718         return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
719 }
720 
721 /**
722  *  tcp_splice_read - splice data from TCP socket to a pipe
723  * @sock:       socket to splice from
724  * @ppos:       position (not valid)
725  * @pipe:       pipe to splice to
726  * @len:        number of bytes to splice
727  * @flags:      splice modifier flags
728  *
729  * Description:
730  *    Will read pages from given socket and fill them into a pipe.
731  *
732  **/
733 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
734                         struct pipe_inode_info *pipe, size_t len,
735                         unsigned int flags)
736 {
737         struct sock *sk = sock->sk;
738         struct tcp_splice_state tss = {
739                 .pipe = pipe,
740                 .len = len,
741                 .flags = flags,
742         };
743         long timeo;
744         ssize_t spliced;
745         int ret;
746 
747         sock_rps_record_flow(sk);
748         /*
749          * We can't seek on a socket input
750          */
751         if (unlikely(*ppos))
752                 return -ESPIPE;
753 
754         ret = spliced = 0;
755 
756         lock_sock(sk);
757 
758         timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
759         while (tss.len) {
760                 ret = __tcp_splice_read(sk, &tss);
761                 if (ret < 0)
762                         break;
763                 else if (!ret) {
764                         if (spliced)
765                                 break;
766                         if (sock_flag(sk, SOCK_DONE))
767                                 break;
768                         if (sk->sk_err) {
769                                 ret = sock_error(sk);
770                                 break;
771                         }
772                         if (sk->sk_shutdown & RCV_SHUTDOWN)
773                                 break;
774                         if (sk->sk_state == TCP_CLOSE) {
775                                 /*
776                                  * This occurs when user tries to read
777                                  * from never connected socket.
778                                  */
779                                 if (!sock_flag(sk, SOCK_DONE))
780                                         ret = -ENOTCONN;
781                                 break;
782                         }
783                         if (!timeo) {
784                                 ret = -EAGAIN;
785                                 break;
786                         }
787                         sk_wait_data(sk, &timeo, NULL);
788                         if (signal_pending(current)) {
789                                 ret = sock_intr_errno(timeo);
790                                 break;
791                         }
792                         continue;
793                 }
794                 tss.len -= ret;
795                 spliced += ret;
796 
797                 if (!timeo)
798                         break;
799                 release_sock(sk);
800                 lock_sock(sk);
801 
802                 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
803                     (sk->sk_shutdown & RCV_SHUTDOWN) ||
804                     signal_pending(current))
805                         break;
806         }
807 
808         release_sock(sk);
809 
810         if (spliced)
811                 return spliced;
812 
813         return ret;
814 }
815 EXPORT_SYMBOL(tcp_splice_read);
816 
817 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
818                                     bool force_schedule)
819 {
820         struct sk_buff *skb;
821 
822         /* The TCP header must be at least 32-bit aligned.  */
823         size = ALIGN(size, 4);
824 
825         if (unlikely(tcp_under_memory_pressure(sk)))
826                 sk_mem_reclaim_partial(sk);
827 
828         skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
829         if (likely(skb)) {
830                 bool mem_scheduled;
831 
832                 if (force_schedule) {
833                         mem_scheduled = true;
834                         sk_forced_mem_schedule(sk, skb->truesize);
835                 } else {
836                         mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
837                 }
838                 if (likely(mem_scheduled)) {
839                         skb_reserve(skb, sk->sk_prot->max_header);
840                         /*
841                          * Make sure that we have exactly size bytes
842                          * available to the caller, no more, no less.
843                          */
844                         skb->reserved_tailroom = skb->end - skb->tail - size;
845                         return skb;
846                 }
847                 __kfree_skb(skb);
848         } else {
849                 sk->sk_prot->enter_memory_pressure(sk);
850                 sk_stream_moderate_sndbuf(sk);
851         }
852         return NULL;
853 }
854 
855 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
856                                        int large_allowed)
857 {
858         struct tcp_sock *tp = tcp_sk(sk);
859         u32 new_size_goal, size_goal;
860 
861         if (!large_allowed || !sk_can_gso(sk))
862                 return mss_now;
863 
864         /* Note : tcp_tso_autosize() will eventually split this later */
865         new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
866         new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
867 
868         /* We try hard to avoid divides here */
869         size_goal = tp->gso_segs * mss_now;
870         if (unlikely(new_size_goal < size_goal ||
871                      new_size_goal >= size_goal + mss_now)) {
872                 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
873                                      sk->sk_gso_max_segs);
874                 size_goal = tp->gso_segs * mss_now;
875         }
876 
877         return max(size_goal, mss_now);
878 }
879 
880 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
881 {
882         int mss_now;
883 
884         mss_now = tcp_current_mss(sk);
885         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
886 
887         return mss_now;
888 }
889 
890 static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
891                                 size_t size, int flags)
892 {
893         struct tcp_sock *tp = tcp_sk(sk);
894         int mss_now, size_goal;
895         int err;
896         ssize_t copied;
897         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
898 
899         /* Wait for a connection to finish. One exception is TCP Fast Open
900          * (passive side) where data is allowed to be sent before a connection
901          * is fully established.
902          */
903         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
904             !tcp_passive_fastopen(sk)) {
905                 err = sk_stream_wait_connect(sk, &timeo);
906                 if (err != 0)
907                         goto out_err;
908         }
909 
910         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
911 
912         mss_now = tcp_send_mss(sk, &size_goal, flags);
913         copied = 0;
914 
915         err = -EPIPE;
916         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
917                 goto out_err;
918 
919         while (size > 0) {
920                 struct sk_buff *skb = tcp_write_queue_tail(sk);
921                 int copy, i;
922                 bool can_coalesce;
923 
924                 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
925 new_segment:
926                         if (!sk_stream_memory_free(sk))
927                                 goto wait_for_sndbuf;
928 
929                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
930                                                   skb_queue_empty(&sk->sk_write_queue));
931                         if (!skb)
932                                 goto wait_for_memory;
933 
934                         skb_entail(sk, skb);
935                         copy = size_goal;
936                 }
937 
938                 if (copy > size)
939                         copy = size;
940 
941                 i = skb_shinfo(skb)->nr_frags;
942                 can_coalesce = skb_can_coalesce(skb, i, page, offset);
943                 if (!can_coalesce && i >= sysctl_max_skb_frags) {
944                         tcp_mark_push(tp, skb);
945                         goto new_segment;
946                 }
947                 if (!sk_wmem_schedule(sk, copy))
948                         goto wait_for_memory;
949 
950                 if (can_coalesce) {
951                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
952                 } else {
953                         get_page(page);
954                         skb_fill_page_desc(skb, i, page, offset, copy);
955                 }
956                 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
957 
958                 skb->len += copy;
959                 skb->data_len += copy;
960                 skb->truesize += copy;
961                 sk->sk_wmem_queued += copy;
962                 sk_mem_charge(sk, copy);
963                 skb->ip_summed = CHECKSUM_PARTIAL;
964                 tp->write_seq += copy;
965                 TCP_SKB_CB(skb)->end_seq += copy;
966                 tcp_skb_pcount_set(skb, 0);
967 
968                 if (!copied)
969                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
970 
971                 copied += copy;
972                 offset += copy;
973                 size -= copy;
974                 if (!size) {
975                         tcp_tx_timestamp(sk, skb);
976                         goto out;
977                 }
978 
979                 if (skb->len < size_goal || (flags & MSG_OOB))
980                         continue;
981 
982                 if (forced_push(tp)) {
983                         tcp_mark_push(tp, skb);
984                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
985                 } else if (skb == tcp_send_head(sk))
986                         tcp_push_one(sk, mss_now);
987                 continue;
988 
989 wait_for_sndbuf:
990                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
991 wait_for_memory:
992                 tcp_push(sk, flags & ~MSG_MORE, mss_now,
993                          TCP_NAGLE_PUSH, size_goal);
994 
995                 err = sk_stream_wait_memory(sk, &timeo);
996                 if (err != 0)
997                         goto do_error;
998 
999                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1000         }
1001 
1002 out:
1003         if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
1004                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1005         return copied;
1006 
1007 do_error:
1008         if (copied)
1009                 goto out;
1010 out_err:
1011         /* make sure we wake any epoll edge trigger waiter */
1012         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1013                 sk->sk_write_space(sk);
1014         return sk_stream_error(sk, flags, err);
1015 }
1016 
1017 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1018                  size_t size, int flags)
1019 {
1020         ssize_t res;
1021 
1022         if (!(sk->sk_route_caps & NETIF_F_SG) ||
1023             !sk_check_csum_caps(sk))
1024                 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1025                                         flags);
1026 
1027         lock_sock(sk);
1028         res = do_tcp_sendpages(sk, page, offset, size, flags);
1029         release_sock(sk);
1030         return res;
1031 }
1032 EXPORT_SYMBOL(tcp_sendpage);
1033 
1034 static inline int select_size(const struct sock *sk, bool sg)
1035 {
1036         const struct tcp_sock *tp = tcp_sk(sk);
1037         int tmp = tp->mss_cache;
1038 
1039         if (sg) {
1040                 if (sk_can_gso(sk)) {
1041                         /* Small frames wont use a full page:
1042                          * Payload will immediately follow tcp header.
1043                          */
1044                         tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1045                 } else {
1046                         int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1047 
1048                         if (tmp >= pgbreak &&
1049                             tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1050                                 tmp = pgbreak;
1051                 }
1052         }
1053 
1054         return tmp;
1055 }
1056 
1057 void tcp_free_fastopen_req(struct tcp_sock *tp)
1058 {
1059         if (tp->fastopen_req) {
1060                 kfree(tp->fastopen_req);
1061                 tp->fastopen_req = NULL;
1062         }
1063 }
1064 
1065 static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1066                                 int *copied, size_t size)
1067 {
1068         struct tcp_sock *tp = tcp_sk(sk);
1069         int err, flags;
1070 
1071         if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1072                 return -EOPNOTSUPP;
1073         if (tp->fastopen_req)
1074                 return -EALREADY; /* Another Fast Open is in progress */
1075 
1076         tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1077                                    sk->sk_allocation);
1078         if (unlikely(!tp->fastopen_req))
1079                 return -ENOBUFS;
1080         tp->fastopen_req->data = msg;
1081         tp->fastopen_req->size = size;
1082 
1083         flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1084         err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1085                                     msg->msg_namelen, flags);
1086         *copied = tp->fastopen_req->copied;
1087         tcp_free_fastopen_req(tp);
1088         return err;
1089 }
1090 
1091 int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1092 {
1093         struct tcp_sock *tp = tcp_sk(sk);
1094         struct sk_buff *skb;
1095         int flags, err, copied = 0;
1096         int mss_now = 0, size_goal, copied_syn = 0;
1097         bool sg;
1098         long timeo;
1099 
1100         lock_sock(sk);
1101 
1102         flags = msg->msg_flags;
1103         if (flags & MSG_FASTOPEN) {
1104                 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1105                 if (err == -EINPROGRESS && copied_syn > 0)
1106                         goto out;
1107                 else if (err)
1108                         goto out_err;
1109         }
1110 
1111         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1112 
1113         /* Wait for a connection to finish. One exception is TCP Fast Open
1114          * (passive side) where data is allowed to be sent before a connection
1115          * is fully established.
1116          */
1117         if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1118             !tcp_passive_fastopen(sk)) {
1119                 err = sk_stream_wait_connect(sk, &timeo);
1120                 if (err != 0)
1121                         goto do_error;
1122         }
1123 
1124         if (unlikely(tp->repair)) {
1125                 if (tp->repair_queue == TCP_RECV_QUEUE) {
1126                         copied = tcp_send_rcvq(sk, msg, size);
1127                         goto out_nopush;
1128                 }
1129 
1130                 err = -EINVAL;
1131                 if (tp->repair_queue == TCP_NO_QUEUE)
1132                         goto out_err;
1133 
1134                 /* 'common' sending to sendq */
1135         }
1136 
1137         /* This should be in poll */
1138         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1139 
1140         mss_now = tcp_send_mss(sk, &size_goal, flags);
1141 
1142         /* Ok commence sending. */
1143         copied = 0;
1144 
1145         err = -EPIPE;
1146         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1147                 goto out_err;
1148 
1149         sg = !!(sk->sk_route_caps & NETIF_F_SG);
1150 
1151         while (msg_data_left(msg)) {
1152                 int copy = 0;
1153                 int max = size_goal;
1154 
1155                 skb = tcp_write_queue_tail(sk);
1156                 if (tcp_send_head(sk)) {
1157                         if (skb->ip_summed == CHECKSUM_NONE)
1158                                 max = mss_now;
1159                         copy = max - skb->len;
1160                 }
1161 
1162                 if (copy <= 0) {
1163 new_segment:
1164                         /* Allocate new segment. If the interface is SG,
1165                          * allocate skb fitting to single page.
1166                          */
1167                         if (!sk_stream_memory_free(sk))
1168                                 goto wait_for_sndbuf;
1169 
1170                         skb = sk_stream_alloc_skb(sk,
1171                                                   select_size(sk, sg),
1172                                                   sk->sk_allocation,
1173                                                   skb_queue_empty(&sk->sk_write_queue));
1174                         if (!skb)
1175                                 goto wait_for_memory;
1176 
1177                         /*
1178                          * Check whether we can use HW checksum.
1179                          */
1180                         if (sk_check_csum_caps(sk))
1181                                 skb->ip_summed = CHECKSUM_PARTIAL;
1182 
1183                         skb_entail(sk, skb);
1184                         copy = size_goal;
1185                         max = size_goal;
1186 
1187                         /* All packets are restored as if they have
1188                          * already been sent. skb_mstamp isn't set to
1189                          * avoid wrong rtt estimation.
1190                          */
1191                         if (tp->repair)
1192                                 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1193                 }
1194 
1195                 /* Try to append data to the end of skb. */
1196                 if (copy > msg_data_left(msg))
1197                         copy = msg_data_left(msg);
1198 
1199                 /* Where to copy to? */
1200                 if (skb_availroom(skb) > 0) {
1201                         /* We have some space in skb head. Superb! */
1202                         copy = min_t(int, copy, skb_availroom(skb));
1203                         err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1204                         if (err)
1205                                 goto do_fault;
1206                 } else {
1207                         bool merge = true;
1208                         int i = skb_shinfo(skb)->nr_frags;
1209                         struct page_frag *pfrag = sk_page_frag(sk);
1210 
1211                         if (!sk_page_frag_refill(sk, pfrag))
1212                                 goto wait_for_memory;
1213 
1214                         if (!skb_can_coalesce(skb, i, pfrag->page,
1215                                               pfrag->offset)) {
1216                                 if (i == sysctl_max_skb_frags || !sg) {
1217                                         tcp_mark_push(tp, skb);
1218                                         goto new_segment;
1219                                 }
1220                                 merge = false;
1221                         }
1222 
1223                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1224 
1225                         if (!sk_wmem_schedule(sk, copy))
1226                                 goto wait_for_memory;
1227 
1228                         err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1229                                                        pfrag->page,
1230                                                        pfrag->offset,
1231                                                        copy);
1232                         if (err)
1233                                 goto do_error;
1234 
1235                         /* Update the skb. */
1236                         if (merge) {
1237                                 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1238                         } else {
1239                                 skb_fill_page_desc(skb, i, pfrag->page,
1240                                                    pfrag->offset, copy);
1241                                 get_page(pfrag->page);
1242                         }
1243                         pfrag->offset += copy;
1244                 }
1245 
1246                 if (!copied)
1247                         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1248 
1249                 tp->write_seq += copy;
1250                 TCP_SKB_CB(skb)->end_seq += copy;
1251                 tcp_skb_pcount_set(skb, 0);
1252 
1253                 copied += copy;
1254                 if (!msg_data_left(msg)) {
1255                         tcp_tx_timestamp(sk, skb);
1256                         goto out;
1257                 }
1258 
1259                 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1260                         continue;
1261 
1262                 if (forced_push(tp)) {
1263                         tcp_mark_push(tp, skb);
1264                         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1265                 } else if (skb == tcp_send_head(sk))
1266                         tcp_push_one(sk, mss_now);
1267                 continue;
1268 
1269 wait_for_sndbuf:
1270                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1271 wait_for_memory:
1272                 if (copied)
1273                         tcp_push(sk, flags & ~MSG_MORE, mss_now,
1274                                  TCP_NAGLE_PUSH, size_goal);
1275 
1276                 err = sk_stream_wait_memory(sk, &timeo);
1277                 if (err != 0)
1278                         goto do_error;
1279 
1280                 mss_now = tcp_send_mss(sk, &size_goal, flags);
1281         }
1282 
1283 out:
1284         if (copied)
1285                 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1286 out_nopush:
1287         release_sock(sk);
1288         return copied + copied_syn;
1289 
1290 do_fault:
1291         if (!skb->len) {
1292                 tcp_unlink_write_queue(skb, sk);
1293                 /* It is the one place in all of TCP, except connection
1294                  * reset, where we can be unlinking the send_head.
1295                  */
1296                 tcp_check_send_head(sk, skb);
1297                 sk_wmem_free_skb(sk, skb);
1298         }
1299 
1300 do_error:
1301         if (copied + copied_syn)
1302                 goto out;
1303 out_err:
1304         err = sk_stream_error(sk, flags, err);
1305         /* make sure we wake any epoll edge trigger waiter */
1306         if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 && err == -EAGAIN))
1307                 sk->sk_write_space(sk);
1308         release_sock(sk);
1309         return err;
1310 }
1311 EXPORT_SYMBOL(tcp_sendmsg);
1312 
1313 /*
1314  *      Handle reading urgent data. BSD has very simple semantics for
1315  *      this, no blocking and very strange errors 8)
1316  */
1317 
1318 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1319 {
1320         struct tcp_sock *tp = tcp_sk(sk);
1321 
1322         /* No URG data to read. */
1323         if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1324             tp->urg_data == TCP_URG_READ)
1325                 return -EINVAL; /* Yes this is right ! */
1326 
1327         if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1328                 return -ENOTCONN;
1329 
1330         if (tp->urg_data & TCP_URG_VALID) {
1331                 int err = 0;
1332                 char c = tp->urg_data;
1333 
1334                 if (!(flags & MSG_PEEK))
1335                         tp->urg_data = TCP_URG_READ;
1336 
1337                 /* Read urgent data. */
1338                 msg->msg_flags |= MSG_OOB;
1339 
1340                 if (len > 0) {
1341                         if (!(flags & MSG_TRUNC))
1342                                 err = memcpy_to_msg(msg, &c, 1);
1343                         len = 1;
1344                 } else
1345                         msg->msg_flags |= MSG_TRUNC;
1346 
1347                 return err ? -EFAULT : len;
1348         }
1349 
1350         if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1351                 return 0;
1352 
1353         /* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1354          * the available implementations agree in this case:
1355          * this call should never block, independent of the
1356          * blocking state of the socket.
1357          * Mike <pall@rz.uni-karlsruhe.de>
1358          */
1359         return -EAGAIN;
1360 }
1361 
1362 static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1363 {
1364         struct sk_buff *skb;
1365         int copied = 0, err = 0;
1366 
1367         /* XXX -- need to support SO_PEEK_OFF */
1368 
1369         skb_queue_walk(&sk->sk_write_queue, skb) {
1370                 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1371                 if (err)
1372                         break;
1373 
1374                 copied += skb->len;
1375         }
1376 
1377         return err ?: copied;
1378 }
1379 
1380 /* Clean up the receive buffer for full frames taken by the user,
1381  * then send an ACK if necessary.  COPIED is the number of bytes
1382  * tcp_recvmsg has given to the user so far, it speeds up the
1383  * calculation of whether or not we must ACK for the sake of
1384  * a window update.
1385  */
1386 static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1387 {
1388         struct tcp_sock *tp = tcp_sk(sk);
1389         bool time_to_ack = false;
1390 
1391         struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1392 
1393         WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1394              "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1395              tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1396 
1397         if (inet_csk_ack_scheduled(sk)) {
1398                 const struct inet_connection_sock *icsk = inet_csk(sk);
1399                    /* Delayed ACKs frequently hit locked sockets during bulk
1400                     * receive. */
1401                 if (icsk->icsk_ack.blocked ||
1402                     /* Once-per-two-segments ACK was not sent by tcp_input.c */
1403                     tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1404                     /*
1405                      * If this read emptied read buffer, we send ACK, if
1406                      * connection is not bidirectional, user drained
1407                      * receive buffer and there was a small segment
1408                      * in queue.
1409                      */
1410                     (copied > 0 &&
1411                      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1412                       ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1413                        !icsk->icsk_ack.pingpong)) &&
1414                       !atomic_read(&sk->sk_rmem_alloc)))
1415                         time_to_ack = true;
1416         }
1417 
1418         /* We send an ACK if we can now advertise a non-zero window
1419          * which has been raised "significantly".
1420          *
1421          * Even if window raised up to infinity, do not send window open ACK
1422          * in states, where we will not receive more. It is useless.
1423          */
1424         if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1425                 __u32 rcv_window_now = tcp_receive_window(tp);
1426 
1427                 /* Optimize, __tcp_select_window() is not cheap. */
1428                 if (2*rcv_window_now <= tp->window_clamp) {
1429                         __u32 new_window = __tcp_select_window(sk);
1430 
1431                         /* Send ACK now, if this read freed lots of space
1432                          * in our buffer. Certainly, new_window is new window.
1433                          * We can advertise it now, if it is not less than current one.
1434                          * "Lots" means "at least twice" here.
1435                          */
1436                         if (new_window && new_window >= 2 * rcv_window_now)
1437                                 time_to_ack = true;
1438                 }
1439         }
1440         if (time_to_ack)
1441                 tcp_send_ack(sk);
1442 }
1443 
1444 static void tcp_prequeue_process(struct sock *sk)
1445 {
1446         struct sk_buff *skb;
1447         struct tcp_sock *tp = tcp_sk(sk);
1448 
1449         NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1450 
1451         /* RX process wants to run with disabled BHs, though it is not
1452          * necessary */
1453         local_bh_disable();
1454         while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1455                 sk_backlog_rcv(sk, skb);
1456         local_bh_enable();
1457 
1458         /* Clear memory counter. */
1459         tp->ucopy.memory = 0;
1460 }
1461 
1462 static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1463 {
1464         struct sk_buff *skb;
1465         u32 offset;
1466 
1467         while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1468                 offset = seq - TCP_SKB_CB(skb)->seq;
1469                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1470                         offset--;
1471                 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1472                         *off = offset;
1473                         return skb;
1474                 }
1475                 /* This looks weird, but this can happen if TCP collapsing
1476                  * splitted a fat GRO packet, while we released socket lock
1477                  * in skb_splice_bits()
1478                  */
1479                 sk_eat_skb(sk, skb);
1480         }
1481         return NULL;
1482 }
1483 
1484 /*
1485  * This routine provides an alternative to tcp_recvmsg() for routines
1486  * that would like to handle copying from skbuffs directly in 'sendfile'
1487  * fashion.
1488  * Note:
1489  *      - It is assumed that the socket was locked by the caller.
1490  *      - The routine does not block.
1491  *      - At present, there is no support for reading OOB data
1492  *        or for 'peeking' the socket using this routine
1493  *        (although both would be easy to implement).
1494  */
1495 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1496                   sk_read_actor_t recv_actor)
1497 {
1498         struct sk_buff *skb;
1499         struct tcp_sock *tp = tcp_sk(sk);
1500         u32 seq = tp->copied_seq;
1501         u32 offset;
1502         int copied = 0;
1503 
1504         if (sk->sk_state == TCP_LISTEN)
1505                 return -ENOTCONN;
1506         while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1507                 if (offset < skb->len) {
1508                         int used;
1509                         size_t len;
1510 
1511                         len = skb->len - offset;
1512                         /* Stop reading if we hit a patch of urgent data */
1513                         if (tp->urg_data) {
1514                                 u32 urg_offset = tp->urg_seq - seq;
1515                                 if (urg_offset < len)
1516                                         len = urg_offset;
1517                                 if (!len)
1518                                         break;
1519                         }
1520                         used = recv_actor(desc, skb, offset, len);
1521                         if (used <= 0) {
1522                                 if (!copied)
1523                                         copied = used;
1524                                 break;
1525                         } else if (used <= len) {
1526                                 seq += used;
1527                                 copied += used;
1528                                 offset += used;
1529                         }
1530                         /* If recv_actor drops the lock (e.g. TCP splice
1531                          * receive) the skb pointer might be invalid when
1532                          * getting here: tcp_collapse might have deleted it
1533                          * while aggregating skbs from the socket queue.
1534                          */
1535                         skb = tcp_recv_skb(sk, seq - 1, &offset);
1536                         if (!skb)
1537                                 break;
1538                         /* TCP coalescing might have appended data to the skb.
1539                          * Try to splice more frags
1540                          */
1541                         if (offset + 1 != skb->len)
1542                                 continue;
1543                 }
1544                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1545                         sk_eat_skb(sk, skb);
1546                         ++seq;
1547                         break;
1548                 }
1549                 sk_eat_skb(sk, skb);
1550                 if (!desc->count)
1551                         break;
1552                 tp->copied_seq = seq;
1553         }
1554         tp->copied_seq = seq;
1555 
1556         tcp_rcv_space_adjust(sk);
1557 
1558         /* Clean up data we have read: This will do ACK frames. */
1559         if (copied > 0) {
1560                 tcp_recv_skb(sk, seq, &offset);
1561                 tcp_cleanup_rbuf(sk, copied);
1562         }
1563         return copied;
1564 }
1565 EXPORT_SYMBOL(tcp_read_sock);
1566 
1567 /*
1568  *      This routine copies from a sock struct into the user buffer.
1569  *
1570  *      Technical note: in 2.3 we work on _locked_ socket, so that
1571  *      tricks with *seq access order and skb->users are not required.
1572  *      Probably, code can be easily improved even more.
1573  */
1574 
1575 int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1576                 int flags, int *addr_len)
1577 {
1578         struct tcp_sock *tp = tcp_sk(sk);
1579         int copied = 0;
1580         u32 peek_seq;
1581         u32 *seq;
1582         unsigned long used;
1583         int err;
1584         int target;             /* Read at least this many bytes */
1585         long timeo;
1586         struct task_struct *user_recv = NULL;
1587         struct sk_buff *skb, *last;
1588         u32 urg_hole = 0;
1589 
1590         if (unlikely(flags & MSG_ERRQUEUE))
1591                 return inet_recv_error(sk, msg, len, addr_len);
1592 
1593         if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1594             (sk->sk_state == TCP_ESTABLISHED))
1595                 sk_busy_loop(sk, nonblock);
1596 
1597         lock_sock(sk);
1598 
1599         err = -ENOTCONN;
1600         if (sk->sk_state == TCP_LISTEN)
1601                 goto out;
1602 
1603         timeo = sock_rcvtimeo(sk, nonblock);
1604 
1605         /* Urgent data needs to be handled specially. */
1606         if (flags & MSG_OOB)
1607                 goto recv_urg;
1608 
1609         if (unlikely(tp->repair)) {
1610                 err = -EPERM;
1611                 if (!(flags & MSG_PEEK))
1612                         goto out;
1613 
1614                 if (tp->repair_queue == TCP_SEND_QUEUE)
1615                         goto recv_sndq;
1616 
1617                 err = -EINVAL;
1618                 if (tp->repair_queue == TCP_NO_QUEUE)
1619                         goto out;
1620 
1621                 /* 'common' recv queue MSG_PEEK-ing */
1622         }
1623 
1624         seq = &tp->copied_seq;
1625         if (flags & MSG_PEEK) {
1626                 peek_seq = tp->copied_seq;
1627                 seq = &peek_seq;
1628         }
1629 
1630         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1631 
1632         do {
1633                 u32 offset;
1634 
1635                 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1636                 if (tp->urg_data && tp->urg_seq == *seq) {
1637                         if (copied)
1638                                 break;
1639                         if (signal_pending(current)) {
1640                                 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1641                                 break;
1642                         }
1643                 }
1644 
1645                 /* Next get a buffer. */
1646 
1647                 last = skb_peek_tail(&sk->sk_receive_queue);
1648                 skb_queue_walk(&sk->sk_receive_queue, skb) {
1649                         last = skb;
1650                         /* Now that we have two receive queues this
1651                          * shouldn't happen.
1652                          */
1653                         if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1654                                  "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1655                                  *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1656                                  flags))
1657                                 break;
1658 
1659                         offset = *seq - TCP_SKB_CB(skb)->seq;
1660                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
1661                                 offset--;
1662                         if (offset < skb->len)
1663                                 goto found_ok_skb;
1664                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1665                                 goto found_fin_ok;
1666                         WARN(!(flags & MSG_PEEK),
1667                              "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1668                              *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1669                 }
1670 
1671                 /* Well, if we have backlog, try to process it now yet. */
1672 
1673                 if (copied >= target && !sk->sk_backlog.tail)
1674                         break;
1675 
1676                 if (copied) {
1677                         if (sk->sk_err ||
1678                             sk->sk_state == TCP_CLOSE ||
1679                             (sk->sk_shutdown & RCV_SHUTDOWN) ||
1680                             !timeo ||
1681                             signal_pending(current))
1682                                 break;
1683                 } else {
1684                         if (sock_flag(sk, SOCK_DONE))
1685                                 break;
1686 
1687                         if (sk->sk_err) {
1688                                 copied = sock_error(sk);
1689                                 break;
1690                         }
1691 
1692                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1693                                 break;
1694 
1695                         if (sk->sk_state == TCP_CLOSE) {
1696                                 if (!sock_flag(sk, SOCK_DONE)) {
1697                                         /* This occurs when user tries to read
1698                                          * from never connected socket.
1699                                          */
1700                                         copied = -ENOTCONN;
1701                                         break;
1702                                 }
1703                                 break;
1704                         }
1705 
1706                         if (!timeo) {
1707                                 copied = -EAGAIN;
1708                                 break;
1709                         }
1710 
1711                         if (signal_pending(current)) {
1712                                 copied = sock_intr_errno(timeo);
1713                                 break;
1714                         }
1715                 }
1716 
1717                 tcp_cleanup_rbuf(sk, copied);
1718 
1719                 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1720                         /* Install new reader */
1721                         if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1722                                 user_recv = current;
1723                                 tp->ucopy.task = user_recv;
1724                                 tp->ucopy.msg = msg;
1725                         }
1726 
1727                         tp->ucopy.len = len;
1728 
1729                         WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1730                                 !(flags & (MSG_PEEK | MSG_TRUNC)));
1731 
1732                         /* Ugly... If prequeue is not empty, we have to
1733                          * process it before releasing socket, otherwise
1734                          * order will be broken at second iteration.
1735                          * More elegant solution is required!!!
1736                          *
1737                          * Look: we have the following (pseudo)queues:
1738                          *
1739                          * 1. packets in flight
1740                          * 2. backlog
1741                          * 3. prequeue
1742                          * 4. receive_queue
1743                          *
1744                          * Each queue can be processed only if the next ones
1745                          * are empty. At this point we have empty receive_queue.
1746                          * But prequeue _can_ be not empty after 2nd iteration,
1747                          * when we jumped to start of loop because backlog
1748                          * processing added something to receive_queue.
1749                          * We cannot release_sock(), because backlog contains
1750                          * packets arrived _after_ prequeued ones.
1751                          *
1752                          * Shortly, algorithm is clear --- to process all
1753                          * the queues in order. We could make it more directly,
1754                          * requeueing packets from backlog to prequeue, if
1755                          * is not empty. It is more elegant, but eats cycles,
1756                          * unfortunately.
1757                          */
1758                         if (!skb_queue_empty(&tp->ucopy.prequeue))
1759                                 goto do_prequeue;
1760 
1761                         /* __ Set realtime policy in scheduler __ */
1762                 }
1763 
1764                 if (copied >= target) {
1765                         /* Do not sleep, just process backlog. */
1766                         release_sock(sk);
1767                         lock_sock(sk);
1768                 } else {
1769                         sk_wait_data(sk, &timeo, last);
1770                 }
1771 
1772                 if (user_recv) {
1773                         int chunk;
1774 
1775                         /* __ Restore normal policy in scheduler __ */
1776 
1777                         chunk = len - tp->ucopy.len;
1778                         if (chunk != 0) {
1779                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1780                                 len -= chunk;
1781                                 copied += chunk;
1782                         }
1783 
1784                         if (tp->rcv_nxt == tp->copied_seq &&
1785                             !skb_queue_empty(&tp->ucopy.prequeue)) {
1786 do_prequeue:
1787                                 tcp_prequeue_process(sk);
1788 
1789                                 chunk = len - tp->ucopy.len;
1790                                 if (chunk != 0) {
1791                                         NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1792                                         len -= chunk;
1793                                         copied += chunk;
1794                                 }
1795                         }
1796                 }
1797                 if ((flags & MSG_PEEK) &&
1798                     (peek_seq - copied - urg_hole != tp->copied_seq)) {
1799                         net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1800                                             current->comm,
1801                                             task_pid_nr(current));
1802                         peek_seq = tp->copied_seq;
1803                 }
1804                 continue;
1805 
1806         found_ok_skb:
1807                 /* Ok so how much can we use? */
1808                 used = skb->len - offset;
1809                 if (len < used)
1810                         used = len;
1811 
1812                 /* Do we have urgent data here? */
1813                 if (tp->urg_data) {
1814                         u32 urg_offset = tp->urg_seq - *seq;
1815                         if (urg_offset < used) {
1816                                 if (!urg_offset) {
1817                                         if (!sock_flag(sk, SOCK_URGINLINE)) {
1818                                                 ++*seq;
1819                                                 urg_hole++;
1820                                                 offset++;
1821                                                 used--;
1822                                                 if (!used)
1823                                                         goto skip_copy;
1824                                         }
1825                                 } else
1826                                         used = urg_offset;
1827                         }
1828                 }
1829 
1830                 if (!(flags & MSG_TRUNC)) {
1831                         err = skb_copy_datagram_msg(skb, offset, msg, used);
1832                         if (err) {
1833                                 /* Exception. Bailout! */
1834                                 if (!copied)
1835                                         copied = -EFAULT;
1836                                 break;
1837                         }
1838                 }
1839 
1840                 *seq += used;
1841                 copied += used;
1842                 len -= used;
1843 
1844                 tcp_rcv_space_adjust(sk);
1845 
1846 skip_copy:
1847                 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1848                         tp->urg_data = 0;
1849                         tcp_fast_path_check(sk);
1850                 }
1851                 if (used + offset < skb->len)
1852                         continue;
1853 
1854                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1855                         goto found_fin_ok;
1856                 if (!(flags & MSG_PEEK))
1857                         sk_eat_skb(sk, skb);
1858                 continue;
1859 
1860         found_fin_ok:
1861                 /* Process the FIN. */
1862                 ++*seq;
1863                 if (!(flags & MSG_PEEK))
1864                         sk_eat_skb(sk, skb);
1865                 break;
1866         } while (len > 0);
1867 
1868         if (user_recv) {
1869                 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1870                         int chunk;
1871 
1872                         tp->ucopy.len = copied > 0 ? len : 0;
1873 
1874                         tcp_prequeue_process(sk);
1875 
1876                         if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1877                                 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1878                                 len -= chunk;
1879                                 copied += chunk;
1880                         }
1881                 }
1882 
1883                 tp->ucopy.task = NULL;
1884                 tp->ucopy.len = 0;
1885         }
1886 
1887         /* According to UNIX98, msg_name/msg_namelen are ignored
1888          * on connected socket. I was just happy when found this 8) --ANK
1889          */
1890 
1891         /* Clean up data we have read: This will do ACK frames. */
1892         tcp_cleanup_rbuf(sk, copied);
1893 
1894         release_sock(sk);
1895         return copied;
1896 
1897 out:
1898         release_sock(sk);
1899         return err;
1900 
1901 recv_urg:
1902         err = tcp_recv_urg(sk, msg, len, flags);
1903         goto out;
1904 
1905 recv_sndq:
1906         err = tcp_peek_sndq(sk, msg, len);
1907         goto out;
1908 }
1909 EXPORT_SYMBOL(tcp_recvmsg);
1910 
1911 void tcp_set_state(struct sock *sk, int state)
1912 {
1913         int oldstate = sk->sk_state;
1914 
1915         switch (state) {
1916         case TCP_ESTABLISHED:
1917                 if (oldstate != TCP_ESTABLISHED)
1918                         TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1919                 break;
1920 
1921         case TCP_CLOSE:
1922                 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1923                         TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1924 
1925                 sk->sk_prot->unhash(sk);
1926                 if (inet_csk(sk)->icsk_bind_hash &&
1927                     !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1928                         inet_put_port(sk);
1929                 /* fall through */
1930         default:
1931                 if (oldstate == TCP_ESTABLISHED)
1932                         TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1933         }
1934 
1935         /* Change state AFTER socket is unhashed to avoid closed
1936          * socket sitting in hash tables.
1937          */
1938         sk_state_store(sk, state);
1939 
1940 #ifdef STATE_TRACE
1941         SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1942 #endif
1943 }
1944 EXPORT_SYMBOL_GPL(tcp_set_state);
1945 
1946 /*
1947  *      State processing on a close. This implements the state shift for
1948  *      sending our FIN frame. Note that we only send a FIN for some
1949  *      states. A shutdown() may have already sent the FIN, or we may be
1950  *      closed.
1951  */
1952 
1953 static const unsigned char new_state[16] = {
1954   /* current state:        new state:      action:      */
1955   [0 /* (Invalid) */]   = TCP_CLOSE,
1956   [TCP_ESTABLISHED]     = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1957   [TCP_SYN_SENT]        = TCP_CLOSE,
1958   [TCP_SYN_RECV]        = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1959   [TCP_FIN_WAIT1]       = TCP_FIN_WAIT1,
1960   [TCP_FIN_WAIT2]       = TCP_FIN_WAIT2,
1961   [TCP_TIME_WAIT]       = TCP_CLOSE,
1962   [TCP_CLOSE]           = TCP_CLOSE,
1963   [TCP_CLOSE_WAIT]      = TCP_LAST_ACK  | TCP_ACTION_FIN,
1964   [TCP_LAST_ACK]        = TCP_LAST_ACK,
1965   [TCP_LISTEN]          = TCP_CLOSE,
1966   [TCP_CLOSING]         = TCP_CLOSING,
1967   [TCP_NEW_SYN_RECV]    = TCP_CLOSE,    /* should not happen ! */
1968 };
1969 
1970 static int tcp_close_state(struct sock *sk)
1971 {
1972         int next = (int)new_state[sk->sk_state];
1973         int ns = next & TCP_STATE_MASK;
1974 
1975         tcp_set_state(sk, ns);
1976 
1977         return next & TCP_ACTION_FIN;
1978 }
1979 
1980 /*
1981  *      Shutdown the sending side of a connection. Much like close except
1982  *      that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1983  */
1984 
1985 void tcp_shutdown(struct sock *sk, int how)
1986 {
1987         /*      We need to grab some memory, and put together a FIN,
1988          *      and then put it into the queue to be sent.
1989          *              Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1990          */
1991         if (!(how & SEND_SHUTDOWN))
1992                 return;
1993 
1994         /* If we've already sent a FIN, or it's a closed state, skip this. */
1995         if ((1 << sk->sk_state) &
1996             (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1997              TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1998                 /* Clear out any half completed packets.  FIN if needed. */
1999                 if (tcp_close_state(sk))
2000                         tcp_send_fin(sk);
2001         }
2002 }
2003 EXPORT_SYMBOL(tcp_shutdown);
2004 
2005 bool tcp_check_oom(struct sock *sk, int shift)
2006 {
2007         bool too_many_orphans, out_of_socket_memory;
2008 
2009         too_many_orphans = tcp_too_many_orphans(sk, shift);
2010         out_of_socket_memory = tcp_out_of_memory(sk);
2011 
2012         if (too_many_orphans)
2013                 net_info_ratelimited("too many orphaned sockets\n");
2014         if (out_of_socket_memory)
2015                 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2016         return too_many_orphans || out_of_socket_memory;
2017 }
2018 
2019 void tcp_close(struct sock *sk, long timeout)
2020 {
2021         struct sk_buff *skb;
2022         int data_was_unread = 0;
2023         int state;
2024 
2025         lock_sock(sk);
2026         sk->sk_shutdown = SHUTDOWN_MASK;
2027 
2028         if (sk->sk_state == TCP_LISTEN) {
2029                 tcp_set_state(sk, TCP_CLOSE);
2030 
2031                 /* Special case. */
2032                 inet_csk_listen_stop(sk);
2033 
2034                 goto adjudge_to_death;
2035         }
2036 
2037         /*  We need to flush the recv. buffs.  We do this only on the
2038          *  descriptor close, not protocol-sourced closes, because the
2039          *  reader process may not have drained the data yet!
2040          */
2041         while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2042                 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2043 
2044                 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2045                         len--;
2046                 data_was_unread += len;
2047                 __kfree_skb(skb);
2048         }
2049 
2050         sk_mem_reclaim(sk);
2051 
2052         /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2053         if (sk->sk_state == TCP_CLOSE)
2054                 goto adjudge_to_death;
2055 
2056         /* As outlined in RFC 2525, section 2.17, we send a RST here because
2057          * data was lost. To witness the awful effects of the old behavior of
2058          * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2059          * GET in an FTP client, suspend the process, wait for the client to
2060          * advertise a zero window, then kill -9 the FTP client, wheee...
2061          * Note: timeout is always zero in such a case.
2062          */
2063         if (unlikely(tcp_sk(sk)->repair)) {
2064                 sk->sk_prot->disconnect(sk, 0);
2065         } else if (data_was_unread) {
2066                 /* Unread data was tossed, zap the connection. */
2067                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2068                 tcp_set_state(sk, TCP_CLOSE);
2069                 tcp_send_active_reset(sk, sk->sk_allocation);
2070         } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2071                 /* Check zero linger _after_ checking for unread data. */
2072                 sk->sk_prot->disconnect(sk, 0);
2073                 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2074         } else if (tcp_close_state(sk)) {
2075                 /* We FIN if the application ate all the data before
2076                  * zapping the connection.
2077                  */
2078 
2079                 /* RED-PEN. Formally speaking, we have broken TCP state
2080                  * machine. State transitions:
2081                  *
2082                  * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2083                  * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2084                  * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2085                  *
2086                  * are legal only when FIN has been sent (i.e. in window),
2087                  * rather than queued out of window. Purists blame.
2088                  *
2089                  * F.e. "RFC state" is ESTABLISHED,
2090                  * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2091                  *
2092                  * The visible declinations are that sometimes
2093                  * we enter time-wait state, when it is not required really
2094                  * (harmless), do not send active resets, when they are
2095                  * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2096                  * they look as CLOSING or LAST_ACK for Linux)
2097                  * Probably, I missed some more holelets.
2098                  *                                              --ANK
2099                  * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2100                  * in a single packet! (May consider it later but will
2101                  * probably need API support or TCP_CORK SYN-ACK until
2102                  * data is written and socket is closed.)
2103                  */
2104                 tcp_send_fin(sk);
2105         }
2106 
2107         sk_stream_wait_close(sk, timeout);
2108 
2109 adjudge_to_death:
2110         state = sk->sk_state;
2111         sock_hold(sk);
2112         sock_orphan(sk);
2113 
2114         /* It is the last release_sock in its life. It will remove backlog. */
2115         release_sock(sk);
2116 
2117 
2118         /* Now socket is owned by kernel and we acquire BH lock
2119            to finish close. No need to check for user refs.
2120          */
2121         local_bh_disable();
2122         bh_lock_sock(sk);
2123         WARN_ON(sock_owned_by_user(sk));
2124 
2125         percpu_counter_inc(sk->sk_prot->orphan_count);
2126 
2127         /* Have we already been destroyed by a softirq or backlog? */
2128         if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2129                 goto out;
2130 
2131         /*      This is a (useful) BSD violating of the RFC. There is a
2132          *      problem with TCP as specified in that the other end could
2133          *      keep a socket open forever with no application left this end.
2134          *      We use a 1 minute timeout (about the same as BSD) then kill
2135          *      our end. If they send after that then tough - BUT: long enough
2136          *      that we won't make the old 4*rto = almost no time - whoops
2137          *      reset mistake.
2138          *
2139          *      Nope, it was not mistake. It is really desired behaviour
2140          *      f.e. on http servers, when such sockets are useless, but
2141          *      consume significant resources. Let's do it with special
2142          *      linger2 option.                                 --ANK
2143          */
2144 
2145         if (sk->sk_state == TCP_FIN_WAIT2) {
2146                 struct tcp_sock *tp = tcp_sk(sk);
2147                 if (tp->linger2 < 0) {
2148                         tcp_set_state(sk, TCP_CLOSE);
2149                         tcp_send_active_reset(sk, GFP_ATOMIC);
2150                         NET_INC_STATS_BH(sock_net(sk),
2151                                         LINUX_MIB_TCPABORTONLINGER);
2152                 } else {
2153                         const int tmo = tcp_fin_time(sk);
2154 
2155                         if (tmo > TCP_TIMEWAIT_LEN) {
2156                                 inet_csk_reset_keepalive_timer(sk,
2157                                                 tmo - TCP_TIMEWAIT_LEN);
2158                         } else {
2159                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2160                                 goto out;
2161                         }
2162                 }
2163         }
2164         if (sk->sk_state != TCP_CLOSE) {
2165                 sk_mem_reclaim(sk);
2166                 if (tcp_check_oom(sk, 0)) {
2167                         tcp_set_state(sk, TCP_CLOSE);
2168                         tcp_send_active_reset(sk, GFP_ATOMIC);
2169                         NET_INC_STATS_BH(sock_net(sk),
2170                                         LINUX_MIB_TCPABORTONMEMORY);
2171                 }
2172         }
2173 
2174         if (sk->sk_state == TCP_CLOSE) {
2175                 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2176                 /* We could get here with a non-NULL req if the socket is
2177                  * aborted (e.g., closed with unread data) before 3WHS
2178                  * finishes.
2179                  */
2180                 if (req)
2181                         reqsk_fastopen_remove(sk, req, false);
2182                 inet_csk_destroy_sock(sk);
2183         }
2184         /* Otherwise, socket is reprieved until protocol close. */
2185 
2186 out:
2187         bh_unlock_sock(sk);
2188         local_bh_enable();
2189         sock_put(sk);
2190 }
2191 EXPORT_SYMBOL(tcp_close);
2192 
2193 /* These states need RST on ABORT according to RFC793 */
2194 
2195 static inline bool tcp_need_reset(int state)
2196 {
2197         return (1 << state) &
2198                (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2199                 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2200 }
2201 
2202 int tcp_disconnect(struct sock *sk, int flags)
2203 {
2204         struct inet_sock *inet = inet_sk(sk);
2205         struct inet_connection_sock *icsk = inet_csk(sk);
2206         struct tcp_sock *tp = tcp_sk(sk);
2207         int err = 0;
2208         int old_state = sk->sk_state;
2209 
2210         if (old_state != TCP_CLOSE)
2211                 tcp_set_state(sk, TCP_CLOSE);
2212 
2213         /* ABORT function of RFC793 */
2214         if (old_state == TCP_LISTEN) {
2215                 inet_csk_listen_stop(sk);
2216         } else if (unlikely(tp->repair)) {
2217                 sk->sk_err = ECONNABORTED;
2218         } else if (tcp_need_reset(old_state) ||
2219                    (tp->snd_nxt != tp->write_seq &&
2220                     (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2221                 /* The last check adjusts for discrepancy of Linux wrt. RFC
2222                  * states
2223                  */
2224                 tcp_send_active_reset(sk, gfp_any());
2225                 sk->sk_err = ECONNRESET;
2226         } else if (old_state == TCP_SYN_SENT)
2227                 sk->sk_err = ECONNRESET;
2228 
2229         tcp_clear_xmit_timers(sk);
2230         __skb_queue_purge(&sk->sk_receive_queue);
2231         tcp_write_queue_purge(sk);
2232         __skb_queue_purge(&tp->out_of_order_queue);
2233 
2234         inet->inet_dport = 0;
2235 
2236         if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2237                 inet_reset_saddr(sk);
2238 
2239         sk->sk_shutdown = 0;
2240         sock_reset_flag(sk, SOCK_DONE);
2241         tp->srtt_us = 0;
2242         tp->write_seq += tp->max_window + 2;
2243         if (tp->write_seq == 0)
2244                 tp->write_seq = 1;
2245         icsk->icsk_backoff = 0;
2246         tp->snd_cwnd = 2;
2247         icsk->icsk_probes_out = 0;
2248         tp->packets_out = 0;
2249         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2250         tp->snd_cwnd_cnt = 0;
2251         tp->window_clamp = 0;
2252         tcp_set_ca_state(sk, TCP_CA_Open);
2253         tcp_clear_retrans(tp);
2254         inet_csk_delack_init(sk);
2255         tcp_init_send_head(sk);
2256         memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2257         __sk_dst_reset(sk);
2258 
2259         WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2260 
2261         sk->sk_error_report(sk);
2262         return err;
2263 }
2264 EXPORT_SYMBOL(tcp_disconnect);
2265 
2266 static inline bool tcp_can_repair_sock(const struct sock *sk)
2267 {
2268         return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2269                 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_ESTABLISHED));
2270 }
2271 
2272 static int tcp_repair_options_est(struct tcp_sock *tp,
2273                 struct tcp_repair_opt __user *optbuf, unsigned int len)
2274 {
2275         struct tcp_repair_opt opt;
2276 
2277         while (len >= sizeof(opt)) {
2278                 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2279                         return -EFAULT;
2280 
2281                 optbuf++;
2282                 len -= sizeof(opt);
2283 
2284                 switch (opt.opt_code) {
2285                 case TCPOPT_MSS:
2286                         tp->rx_opt.mss_clamp = opt.opt_val;
2287                         break;
2288                 case TCPOPT_WINDOW:
2289                         {
2290                                 u16 snd_wscale = opt.opt_val & 0xFFFF;
2291                                 u16 rcv_wscale = opt.opt_val >> 16;
2292 
2293                                 if (snd_wscale > 14 || rcv_wscale > 14)
2294                                         return -EFBIG;
2295 
2296                                 tp->rx_opt.snd_wscale = snd_wscale;
2297                                 tp->rx_opt.rcv_wscale = rcv_wscale;
2298                                 tp->rx_opt.wscale_ok = 1;
2299                         }
2300                         break;
2301                 case TCPOPT_SACK_PERM:
2302                         if (opt.opt_val != 0)
2303                                 return -EINVAL;
2304 
2305                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2306                         if (sysctl_tcp_fack)
2307                                 tcp_enable_fack(tp);
2308                         break;
2309                 case TCPOPT_TIMESTAMP:
2310                         if (opt.opt_val != 0)
2311                                 return -EINVAL;
2312 
2313                         tp->rx_opt.tstamp_ok = 1;
2314                         break;
2315                 }
2316         }
2317 
2318         return 0;
2319 }
2320 
2321 /*
2322  *      Socket option code for TCP.
2323  */
2324 static int do_tcp_setsockopt(struct sock *sk, int level,
2325                 int optname, char __user *optval, unsigned int optlen)
2326 {
2327         struct tcp_sock *tp = tcp_sk(sk);
2328         struct inet_connection_sock *icsk = inet_csk(sk);
2329         int val;
2330         int err = 0;
2331 
2332         /* These are data/string values, all the others are ints */
2333         switch (optname) {
2334         case TCP_CONGESTION: {
2335                 char name[TCP_CA_NAME_MAX];
2336 
2337                 if (optlen < 1)
2338                         return -EINVAL;
2339 
2340                 val = strncpy_from_user(name, optval,
2341                                         min_t(long, TCP_CA_NAME_MAX-1, optlen));
2342                 if (val < 0)
2343                         return -EFAULT;
2344                 name[val] = 0;
2345 
2346                 lock_sock(sk);
2347                 err = tcp_set_congestion_control(sk, name);
2348                 release_sock(sk);
2349                 return err;
2350         }
2351         default:
2352                 /* fallthru */
2353                 break;
2354         }
2355 
2356         if (optlen < sizeof(int))
2357                 return -EINVAL;
2358 
2359         if (get_user(val, (int __user *)optval))
2360                 return -EFAULT;
2361 
2362         lock_sock(sk);
2363 
2364         switch (optname) {
2365         case TCP_MAXSEG:
2366                 /* Values greater than interface MTU won't take effect. However
2367                  * at the point when this call is done we typically don't yet
2368                  * know which interface is going to be used */
2369                 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2370                         err = -EINVAL;
2371                         break;
2372                 }
2373                 tp->rx_opt.user_mss = val;
2374                 break;
2375 
2376         case TCP_NODELAY:
2377                 if (val) {
2378                         /* TCP_NODELAY is weaker than TCP_CORK, so that
2379                          * this option on corked socket is remembered, but
2380                          * it is not activated until cork is cleared.
2381                          *
2382                          * However, when TCP_NODELAY is set we make
2383                          * an explicit push, which overrides even TCP_CORK
2384                          * for currently queued segments.
2385                          */
2386                         tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2387                         tcp_push_pending_frames(sk);
2388                 } else {
2389                         tp->nonagle &= ~TCP_NAGLE_OFF;
2390                 }
2391                 break;
2392 
2393         case TCP_THIN_LINEAR_TIMEOUTS:
2394                 if (val < 0 || val > 1)
2395                         err = -EINVAL;
2396                 else
2397                         tp->thin_lto = val;
2398                 break;
2399 
2400         case TCP_THIN_DUPACK:
2401                 if (val < 0 || val > 1)
2402                         err = -EINVAL;
2403                 else {
2404                         tp->thin_dupack = val;
2405                         if (tp->thin_dupack)
2406                                 tcp_disable_early_retrans(tp);
2407                 }
2408                 break;
2409 
2410         case TCP_REPAIR:
2411                 if (!tcp_can_repair_sock(sk))
2412                         err = -EPERM;
2413                 else if (val == 1) {
2414                         tp->repair = 1;
2415                         sk->sk_reuse = SK_FORCE_REUSE;
2416                         tp->repair_queue = TCP_NO_QUEUE;
2417                 } else if (val == 0) {
2418                         tp->repair = 0;
2419                         sk->sk_reuse = SK_NO_REUSE;
2420                         tcp_send_window_probe(sk);
2421                 } else
2422                         err = -EINVAL;
2423 
2424                 break;
2425 
2426         case TCP_REPAIR_QUEUE:
2427                 if (!tp->repair)
2428                         err = -EPERM;
2429                 else if (val < TCP_QUEUES_NR)
2430                         tp->repair_queue = val;
2431                 else
2432                         err = -EINVAL;
2433                 break;
2434 
2435         case TCP_QUEUE_SEQ:
2436                 if (sk->sk_state != TCP_CLOSE)
2437                         err = -EPERM;
2438                 else if (tp->repair_queue == TCP_SEND_QUEUE)
2439                         tp->write_seq = val;
2440                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2441                         tp->rcv_nxt = val;
2442                 else
2443                         err = -EINVAL;
2444                 break;
2445 
2446         case TCP_REPAIR_OPTIONS:
2447                 if (!tp->repair)
2448                         err = -EINVAL;
2449                 else if (sk->sk_state == TCP_ESTABLISHED)
2450                         err = tcp_repair_options_est(tp,
2451                                         (struct tcp_repair_opt __user *)optval,
2452                                         optlen);
2453                 else
2454                         err = -EPERM;
2455                 break;
2456 
2457         case TCP_CORK:
2458                 /* When set indicates to always queue non-full frames.
2459                  * Later the user clears this option and we transmit
2460                  * any pending partial frames in the queue.  This is
2461                  * meant to be used alongside sendfile() to get properly
2462                  * filled frames when the user (for example) must write
2463                  * out headers with a write() call first and then use
2464                  * sendfile to send out the data parts.
2465                  *
2466                  * TCP_CORK can be set together with TCP_NODELAY and it is
2467                  * stronger than TCP_NODELAY.
2468                  */
2469                 if (val) {
2470                         tp->nonagle |= TCP_NAGLE_CORK;
2471                 } else {
2472                         tp->nonagle &= ~TCP_NAGLE_CORK;
2473                         if (tp->nonagle&TCP_NAGLE_OFF)
2474                                 tp->nonagle |= TCP_NAGLE_PUSH;
2475                         tcp_push_pending_frames(sk);
2476                 }
2477                 break;
2478 
2479         case TCP_KEEPIDLE:
2480                 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2481                         err = -EINVAL;
2482                 else {
2483                         tp->keepalive_time = val * HZ;
2484                         if (sock_flag(sk, SOCK_KEEPOPEN) &&
2485                             !((1 << sk->sk_state) &
2486                               (TCPF_CLOSE | TCPF_LISTEN))) {
2487                                 u32 elapsed = keepalive_time_elapsed(tp);
2488                                 if (tp->keepalive_time > elapsed)
2489                                         elapsed = tp->keepalive_time - elapsed;
2490                                 else
2491                                         elapsed = 0;
2492                                 inet_csk_reset_keepalive_timer(sk, elapsed);
2493                         }
2494                 }
2495                 break;
2496         case TCP_KEEPINTVL:
2497                 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2498                         err = -EINVAL;
2499                 else
2500                         tp->keepalive_intvl = val * HZ;
2501                 break;
2502         case TCP_KEEPCNT:
2503                 if (val < 1 || val > MAX_TCP_KEEPCNT)
2504                         err = -EINVAL;
2505                 else
2506                         tp->keepalive_probes = val;
2507                 break;
2508         case TCP_SYNCNT:
2509                 if (val < 1 || val > MAX_TCP_SYNCNT)
2510                         err = -EINVAL;
2511                 else
2512                         icsk->icsk_syn_retries = val;
2513                 break;
2514 
2515         case TCP_SAVE_SYN:
2516                 if (val < 0 || val > 1)
2517                         err = -EINVAL;
2518                 else
2519                         tp->save_syn = val;
2520                 break;
2521 
2522         case TCP_LINGER2:
2523                 if (val < 0)
2524                         tp->linger2 = -1;
2525                 else if (val > sysctl_tcp_fin_timeout / HZ)
2526                         tp->linger2 = 0;
2527                 else
2528                         tp->linger2 = val * HZ;
2529                 break;
2530 
2531         case TCP_DEFER_ACCEPT:
2532                 /* Translate value in seconds to number of retransmits */
2533                 icsk->icsk_accept_queue.rskq_defer_accept =
2534                         secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2535                                         TCP_RTO_MAX / HZ);
2536                 break;
2537 
2538         case TCP_WINDOW_CLAMP:
2539                 if (!val) {
2540                         if (sk->sk_state != TCP_CLOSE) {
2541                                 err = -EINVAL;
2542                                 break;
2543                         }
2544                         tp->window_clamp = 0;
2545                 } else
2546                         tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2547                                                 SOCK_MIN_RCVBUF / 2 : val;
2548                 break;
2549 
2550         case TCP_QUICKACK:
2551                 if (!val) {
2552                         icsk->icsk_ack.pingpong = 1;
2553                 } else {
2554                         icsk->icsk_ack.pingpong = 0;
2555                         if ((1 << sk->sk_state) &
2556                             (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2557                             inet_csk_ack_scheduled(sk)) {
2558                                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2559                                 tcp_cleanup_rbuf(sk, 1);
2560                                 if (!(val & 1))
2561                                         icsk->icsk_ack.pingpong = 1;
2562                         }
2563                 }
2564                 break;
2565 
2566 #ifdef CONFIG_TCP_MD5SIG
2567         case TCP_MD5SIG:
2568                 /* Read the IP->Key mappings from userspace */
2569                 err = tp->af_specific->md5_parse(sk, optval, optlen);
2570                 break;
2571 #endif
2572         case TCP_USER_TIMEOUT:
2573                 /* Cap the max time in ms TCP will retry or probe the window
2574                  * before giving up and aborting (ETIMEDOUT) a connection.
2575                  */
2576                 if (val < 0)
2577                         err = -EINVAL;
2578                 else
2579                         icsk->icsk_user_timeout = msecs_to_jiffies(val);
2580                 break;
2581 
2582         case TCP_FASTOPEN:
2583                 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2584                     TCPF_LISTEN))) {
2585                         tcp_fastopen_init_key_once(true);
2586 
2587                         fastopen_queue_tune(sk, val);
2588                 } else {
2589                         err = -EINVAL;
2590                 }
2591                 break;
2592         case TCP_TIMESTAMP:
2593                 if (!tp->repair)
2594                         err = -EPERM;
2595                 else
2596                         tp->tsoffset = val - tcp_time_stamp;
2597                 break;
2598         case TCP_NOTSENT_LOWAT:
2599                 tp->notsent_lowat = val;
2600                 sk->sk_write_space(sk);
2601                 break;
2602         default:
2603                 err = -ENOPROTOOPT;
2604                 break;
2605         }
2606 
2607         release_sock(sk);
2608         return err;
2609 }
2610 
2611 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2612                    unsigned int optlen)
2613 {
2614         const struct inet_connection_sock *icsk = inet_csk(sk);
2615 
2616         if (level != SOL_TCP)
2617                 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2618                                                      optval, optlen);
2619         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2620 }
2621 EXPORT_SYMBOL(tcp_setsockopt);
2622 
2623 #ifdef CONFIG_COMPAT
2624 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2625                           char __user *optval, unsigned int optlen)
2626 {
2627         if (level != SOL_TCP)
2628                 return inet_csk_compat_setsockopt(sk, level, optname,
2629                                                   optval, optlen);
2630         return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2631 }
2632 EXPORT_SYMBOL(compat_tcp_setsockopt);
2633 #endif
2634 
2635 /* Return information about state of tcp endpoint in API format. */
2636 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2637 {
2638         const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2639         const struct inet_connection_sock *icsk = inet_csk(sk);
2640         u32 now = tcp_time_stamp;
2641         unsigned int start;
2642         u64 rate64;
2643         u32 rate;
2644 
2645         memset(info, 0, sizeof(*info));
2646         if (sk->sk_type != SOCK_STREAM)
2647                 return;
2648 
2649         info->tcpi_state = sk_state_load(sk);
2650 
2651         info->tcpi_ca_state = icsk->icsk_ca_state;
2652         info->tcpi_retransmits = icsk->icsk_retransmits;
2653         info->tcpi_probes = icsk->icsk_probes_out;
2654         info->tcpi_backoff = icsk->icsk_backoff;
2655 
2656         if (tp->rx_opt.tstamp_ok)
2657                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2658         if (tcp_is_sack(tp))
2659                 info->tcpi_options |= TCPI_OPT_SACK;
2660         if (tp->rx_opt.wscale_ok) {
2661                 info->tcpi_options |= TCPI_OPT_WSCALE;
2662                 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2663                 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2664         }
2665 
2666         if (tp->ecn_flags & TCP_ECN_OK)
2667                 info->tcpi_options |= TCPI_OPT_ECN;
2668         if (tp->ecn_flags & TCP_ECN_SEEN)
2669                 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2670         if (tp->syn_data_acked)
2671                 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2672 
2673         info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2674         info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2675         info->tcpi_snd_mss = tp->mss_cache;
2676         info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2677 
2678         if (info->tcpi_state == TCP_LISTEN) {
2679                 info->tcpi_unacked = sk->sk_ack_backlog;
2680                 info->tcpi_sacked = sk->sk_max_ack_backlog;
2681         } else {
2682                 info->tcpi_unacked = tp->packets_out;
2683                 info->tcpi_sacked = tp->sacked_out;
2684         }
2685         info->tcpi_lost = tp->lost_out;
2686         info->tcpi_retrans = tp->retrans_out;
2687         info->tcpi_fackets = tp->fackets_out;
2688 
2689         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2690         info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2691         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2692 
2693         info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2694         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2695         info->tcpi_rtt = tp->srtt_us >> 3;
2696         info->tcpi_rttvar = tp->mdev_us >> 2;
2697         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2698         info->tcpi_snd_cwnd = tp->snd_cwnd;
2699         info->tcpi_advmss = tp->advmss;
2700         info->tcpi_reordering = tp->reordering;
2701 
2702         info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2703         info->tcpi_rcv_space = tp->rcvq_space.space;
2704 
2705         info->tcpi_total_retrans = tp->total_retrans;
2706 
2707         rate = READ_ONCE(sk->sk_pacing_rate);
2708         rate64 = rate != ~0U ? rate : ~0ULL;
2709         put_unaligned(rate64, &info->tcpi_pacing_rate);
2710 
2711         rate = READ_ONCE(sk->sk_max_pacing_rate);
2712         rate64 = rate != ~0U ? rate : ~0ULL;
2713         put_unaligned(rate64, &info->tcpi_max_pacing_rate);
2714 
2715         do {
2716                 start = u64_stats_fetch_begin_irq(&tp->syncp);
2717                 put_unaligned(tp->bytes_acked, &info->tcpi_bytes_acked);
2718                 put_unaligned(tp->bytes_received, &info->tcpi_bytes_received);
2719         } while (u64_stats_fetch_retry_irq(&tp->syncp, start));
2720         info->tcpi_segs_out = tp->segs_out;
2721         info->tcpi_segs_in = tp->segs_in;
2722 }
2723 EXPORT_SYMBOL_GPL(tcp_get_info);
2724 
2725 static int do_tcp_getsockopt(struct sock *sk, int level,
2726                 int optname, char __user *optval, int __user *optlen)
2727 {
2728         struct inet_connection_sock *icsk = inet_csk(sk);
2729         struct tcp_sock *tp = tcp_sk(sk);
2730         int val, len;
2731 
2732         if (get_user(len, optlen))
2733                 return -EFAULT;
2734 
2735         len = min_t(unsigned int, len, sizeof(int));
2736 
2737         if (len < 0)
2738                 return -EINVAL;
2739 
2740         switch (optname) {
2741         case TCP_MAXSEG:
2742                 val = tp->mss_cache;
2743                 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2744                         val = tp->rx_opt.user_mss;
2745                 if (tp->repair)
2746                         val = tp->rx_opt.mss_clamp;
2747                 break;
2748         case TCP_NODELAY:
2749                 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2750                 break;
2751         case TCP_CORK:
2752                 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2753                 break;
2754         case TCP_KEEPIDLE:
2755                 val = keepalive_time_when(tp) / HZ;
2756                 break;
2757         case TCP_KEEPINTVL:
2758                 val = keepalive_intvl_when(tp) / HZ;
2759                 break;
2760         case TCP_KEEPCNT:
2761                 val = keepalive_probes(tp);
2762                 break;
2763         case TCP_SYNCNT:
2764                 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2765                 break;
2766         case TCP_LINGER2:
2767                 val = tp->linger2;
2768                 if (val >= 0)
2769                         val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2770                 break;
2771         case TCP_DEFER_ACCEPT:
2772                 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2773                                       TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2774                 break;
2775         case TCP_WINDOW_CLAMP:
2776                 val = tp->window_clamp;
2777                 break;
2778         case TCP_INFO: {
2779                 struct tcp_info info;
2780 
2781                 if (get_user(len, optlen))
2782                         return -EFAULT;
2783 
2784                 tcp_get_info(sk, &info);
2785 
2786                 len = min_t(unsigned int, len, sizeof(info));
2787                 if (put_user(len, optlen))
2788                         return -EFAULT;
2789                 if (copy_to_user(optval, &info, len))
2790                         return -EFAULT;
2791                 return 0;
2792         }
2793         case TCP_CC_INFO: {
2794                 const struct tcp_congestion_ops *ca_ops;
2795                 union tcp_cc_info info;
2796                 size_t sz = 0;
2797                 int attr;
2798 
2799                 if (get_user(len, optlen))
2800                         return -EFAULT;
2801 
2802                 ca_ops = icsk->icsk_ca_ops;
2803                 if (ca_ops && ca_ops->get_info)
2804                         sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2805 
2806                 len = min_t(unsigned int, len, sz);
2807                 if (put_user(len, optlen))
2808                         return -EFAULT;
2809                 if (copy_to_user(optval, &info, len))
2810                         return -EFAULT;
2811                 return 0;
2812         }
2813         case TCP_QUICKACK:
2814                 val = !icsk->icsk_ack.pingpong;
2815                 break;
2816 
2817         case TCP_CONGESTION:
2818                 if (get_user(len, optlen))
2819                         return -EFAULT;
2820                 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2821                 if (put_user(len, optlen))
2822                         return -EFAULT;
2823                 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2824                         return -EFAULT;
2825                 return 0;
2826 
2827         case TCP_THIN_LINEAR_TIMEOUTS:
2828                 val = tp->thin_lto;
2829                 break;
2830         case TCP_THIN_DUPACK:
2831                 val = tp->thin_dupack;
2832                 break;
2833 
2834         case TCP_REPAIR:
2835                 val = tp->repair;
2836                 break;
2837 
2838         case TCP_REPAIR_QUEUE:
2839                 if (tp->repair)
2840                         val = tp->repair_queue;
2841                 else
2842                         return -EINVAL;
2843                 break;
2844 
2845         case TCP_QUEUE_SEQ:
2846                 if (tp->repair_queue == TCP_SEND_QUEUE)
2847                         val = tp->write_seq;
2848                 else if (tp->repair_queue == TCP_RECV_QUEUE)
2849                         val = tp->rcv_nxt;
2850                 else
2851                         return -EINVAL;
2852                 break;
2853 
2854         case TCP_USER_TIMEOUT:
2855                 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2856                 break;
2857 
2858         case TCP_FASTOPEN:
2859                 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
2860                 break;
2861 
2862         case TCP_TIMESTAMP:
2863                 val = tcp_time_stamp + tp->tsoffset;
2864                 break;
2865         case TCP_NOTSENT_LOWAT:
2866                 val = tp->notsent_lowat;
2867                 break;
2868         case TCP_SAVE_SYN:
2869                 val = tp->save_syn;
2870                 break;
2871         case TCP_SAVED_SYN: {
2872                 if (get_user(len, optlen))
2873                         return -EFAULT;
2874 
2875                 lock_sock(sk);
2876                 if (tp->saved_syn) {
2877                         if (len < tp->saved_syn[0]) {
2878                                 if (put_user(tp->saved_syn[0], optlen)) {
2879                                         release_sock(sk);
2880                                         return -EFAULT;
2881                                 }
2882                                 release_sock(sk);
2883                                 return -EINVAL;
2884                         }
2885                         len = tp->saved_syn[0];
2886                         if (put_user(len, optlen)) {
2887                                 release_sock(sk);
2888                                 return -EFAULT;
2889                         }
2890                         if (copy_to_user(optval, tp->saved_syn + 1, len)) {
2891                                 release_sock(sk);
2892                                 return -EFAULT;
2893                         }
2894                         tcp_saved_syn_free(tp);
2895                         release_sock(sk);
2896                 } else {
2897                         release_sock(sk);
2898                         len = 0;
2899                         if (put_user(len, optlen))
2900                                 return -EFAULT;
2901                 }
2902                 return 0;
2903         }
2904         default:
2905                 return -ENOPROTOOPT;
2906         }
2907 
2908         if (put_user(len, optlen))
2909                 return -EFAULT;
2910         if (copy_to_user(optval, &val, len))
2911                 return -EFAULT;
2912         return 0;
2913 }
2914 
2915 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2916                    int __user *optlen)
2917 {
2918         struct inet_connection_sock *icsk = inet_csk(sk);
2919 
2920         if (level != SOL_TCP)
2921                 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2922                                                      optval, optlen);
2923         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2924 }
2925 EXPORT_SYMBOL(tcp_getsockopt);
2926 
2927 #ifdef CONFIG_COMPAT
2928 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2929                           char __user *optval, int __user *optlen)
2930 {
2931         if (level != SOL_TCP)
2932                 return inet_csk_compat_getsockopt(sk, level, optname,
2933                                                   optval, optlen);
2934         return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2935 }
2936 EXPORT_SYMBOL(compat_tcp_getsockopt);
2937 #endif
2938 
2939 #ifdef CONFIG_TCP_MD5SIG
2940 static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
2941 static DEFINE_MUTEX(tcp_md5sig_mutex);
2942 static bool tcp_md5sig_pool_populated = false;
2943 
2944 static void __tcp_alloc_md5sig_pool(void)
2945 {
2946         int cpu;
2947 
2948         for_each_possible_cpu(cpu) {
2949                 if (!per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm) {
2950                         struct crypto_hash *hash;
2951 
2952                         hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2953                         if (IS_ERR(hash))
2954                                 return;
2955                         per_cpu(tcp_md5sig_pool, cpu).md5_desc.tfm = hash;
2956                 }
2957         }
2958         /* before setting tcp_md5sig_pool_populated, we must commit all writes
2959          * to memory. See smp_rmb() in tcp_get_md5sig_pool()
2960          */
2961         smp_wmb();
2962         tcp_md5sig_pool_populated = true;
2963 }
2964 
2965 bool tcp_alloc_md5sig_pool(void)
2966 {
2967         if (unlikely(!tcp_md5sig_pool_populated)) {
2968                 mutex_lock(&tcp_md5sig_mutex);
2969 
2970                 if (!tcp_md5sig_pool_populated)
2971                         __tcp_alloc_md5sig_pool();
2972 
2973                 mutex_unlock(&tcp_md5sig_mutex);
2974         }
2975         return tcp_md5sig_pool_populated;
2976 }
2977 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2978 
2979 
2980 /**
2981  *      tcp_get_md5sig_pool - get md5sig_pool for this user
2982  *
2983  *      We use percpu structure, so if we succeed, we exit with preemption
2984  *      and BH disabled, to make sure another thread or softirq handling
2985  *      wont try to get same context.
2986  */
2987 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2988 {
2989         local_bh_disable();
2990 
2991         if (tcp_md5sig_pool_populated) {
2992                 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
2993                 smp_rmb();
2994                 return this_cpu_ptr(&tcp_md5sig_pool);
2995         }
2996         local_bh_enable();
2997         return NULL;
2998 }
2999 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3000 
3001 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3002                         const struct tcphdr *th)
3003 {
3004         struct scatterlist sg;
3005         struct tcphdr hdr;
3006         int err;
3007 
3008         /* We are not allowed to change tcphdr, make a local copy */
3009         memcpy(&hdr, th, sizeof(hdr));
3010         hdr.check = 0;
3011 
3012         /* options aren't included in the hash */
3013         sg_init_one(&sg, &hdr, sizeof(hdr));
3014         err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3015         return err;
3016 }
3017 EXPORT_SYMBOL(tcp_md5_hash_header);
3018 
3019 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3020                           const struct sk_buff *skb, unsigned int header_len)
3021 {
3022         struct scatterlist sg;
3023         const struct tcphdr *tp = tcp_hdr(skb);
3024         struct hash_desc *desc = &hp->md5_desc;
3025         unsigned int i;
3026         const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3027                                            skb_headlen(skb) - header_len : 0;
3028         const struct skb_shared_info *shi = skb_shinfo(skb);
3029         struct sk_buff *frag_iter;
3030 
3031         sg_init_table(&sg, 1);
3032 
3033         sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3034         if (crypto_hash_update(desc, &sg, head_data_len))
3035                 return 1;
3036 
3037         for (i = 0; i < shi->nr_frags; ++i) {
3038                 const struct skb_frag_struct *f = &shi->frags[i];
3039                 unsigned int offset = f->page_offset;
3040                 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3041 
3042                 sg_set_page(&sg, page, skb_frag_size(f),
3043                             offset_in_page(offset));
3044                 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3045                         return 1;
3046         }
3047 
3048         skb_walk_frags(skb, frag_iter)
3049                 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3050                         return 1;
3051 
3052         return 0;
3053 }
3054 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3055 
3056 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3057 {
3058         struct scatterlist sg;
3059 
3060         sg_init_one(&sg, key->key, key->keylen);
3061         return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3062 }
3063 EXPORT_SYMBOL(tcp_md5_hash_key);
3064 
3065 #endif
3066 
3067 void tcp_done(struct sock *sk)
3068 {
3069         struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3070 
3071         if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3072                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3073 
3074         tcp_set_state(sk, TCP_CLOSE);
3075         tcp_clear_xmit_timers(sk);
3076         if (req)
3077                 reqsk_fastopen_remove(sk, req, false);
3078 
3079         sk->sk_shutdown = SHUTDOWN_MASK;
3080 
3081         if (!sock_flag(sk, SOCK_DEAD))
3082                 sk->sk_state_change(sk);
3083         else
3084                 inet_csk_destroy_sock(sk);
3085 }
3086 EXPORT_SYMBOL_GPL(tcp_done);
3087 
3088 int tcp_abort(struct sock *sk, int err)
3089 {
3090         if (!sk_fullsock(sk)) {
3091                 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3092                         struct request_sock *req = inet_reqsk(sk);
3093 
3094                         local_bh_disable();
3095                         inet_csk_reqsk_queue_drop_and_put(req->rsk_listener,
3096                                                           req);
3097                         local_bh_enable();
3098                         return 0;
3099                 }
3100                 sock_gen_put(sk);
3101                 return -EOPNOTSUPP;
3102         }
3103 
3104         /* Don't race with userspace socket closes such as tcp_close. */
3105         lock_sock(sk);
3106 
3107         if (sk->sk_state == TCP_LISTEN) {
3108                 tcp_set_state(sk, TCP_CLOSE);
3109                 inet_csk_listen_stop(sk);
3110         }
3111 
3112         /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3113         local_bh_disable();
3114         bh_lock_sock(sk);
3115 
3116         if (!sock_flag(sk, SOCK_DEAD)) {
3117                 sk->sk_err = err;
3118                 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3119                 smp_wmb();
3120                 sk->sk_error_report(sk);
3121                 if (tcp_need_reset(sk->sk_state))
3122                         tcp_send_active_reset(sk, GFP_ATOMIC);
3123                 tcp_done(sk);
3124         }
3125 
3126         bh_unlock_sock(sk);
3127         local_bh_enable();
3128         release_sock(sk);
3129         sock_put(sk);
3130         return 0;
3131 }
3132 EXPORT_SYMBOL_GPL(tcp_abort);
3133 
3134 extern struct tcp_congestion_ops tcp_reno;
3135 
3136 static __initdata unsigned long thash_entries;
3137 static int __init set_thash_entries(char *str)
3138 {
3139         ssize_t ret;
3140 
3141         if (!str)
3142                 return 0;
3143 
3144         ret = kstrtoul(str, 0, &thash_entries);
3145         if (ret)
3146                 return 0;
3147 
3148         return 1;
3149 }
3150 __setup("thash_entries=", set_thash_entries);
3151 
3152 static void __init tcp_init_mem(void)
3153 {
3154         unsigned long limit = nr_free_buffer_pages() / 16;
3155 
3156         limit = max(limit, 128UL);
3157         sysctl_tcp_mem[0] = limit / 4 * 3;              /* 4.68 % */
3158         sysctl_tcp_mem[1] = limit;                      /* 6.25 % */
3159         sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;      /* 9.37 % */
3160 }
3161 
3162 void __init tcp_init(void)
3163 {
3164         unsigned long limit;
3165         int max_rshare, max_wshare, cnt;
3166         unsigned int i;
3167 
3168         sock_skb_cb_check_size(sizeof(struct tcp_skb_cb));
3169 
3170         percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3171         percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3172         tcp_hashinfo.bind_bucket_cachep =
3173                 kmem_cache_create("tcp_bind_bucket",
3174                                   sizeof(struct inet_bind_bucket), 0,
3175                                   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3176 
3177         /* Size and allocate the main established and bind bucket
3178          * hash tables.
3179          *
3180          * The methodology is similar to that of the buffer cache.
3181          */
3182         tcp_hashinfo.ehash =
3183                 alloc_large_system_hash("TCP established",
3184                                         sizeof(struct inet_ehash_bucket),
3185                                         thash_entries,
3186                                         17, /* one slot per 128 KB of memory */
3187                                         0,
3188                                         NULL,
3189                                         &tcp_hashinfo.ehash_mask,
3190                                         0,
3191                                         thash_entries ? 0 : 512 * 1024);
3192         for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3193                 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3194 
3195         if (inet_ehash_locks_alloc(&tcp_hashinfo))
3196                 panic("TCP: failed to alloc ehash_locks");
3197         tcp_hashinfo.bhash =
3198                 alloc_large_system_hash("TCP bind",
3199                                         sizeof(struct inet_bind_hashbucket),
3200                                         tcp_hashinfo.ehash_mask + 1,
3201                                         17, /* one slot per 128 KB of memory */
3202                                         0,
3203                                         &tcp_hashinfo.bhash_size,
3204                                         NULL,
3205                                         0,
3206                                         64 * 1024);
3207         tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3208         for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3209                 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3210                 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3211         }
3212 
3213 
3214         cnt = tcp_hashinfo.ehash_mask + 1;
3215 
3216         tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3217         sysctl_tcp_max_orphans = cnt / 2;
3218         sysctl_max_syn_backlog = max(128, cnt / 256);
3219 
3220         tcp_init_mem();
3221         /* Set per-socket limits to no more than 1/128 the pressure threshold */
3222         limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3223         max_wshare = min(4UL*1024*1024, limit);
3224         max_rshare = min(6UL*1024*1024, limit);
3225 
3226         sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3227         sysctl_tcp_wmem[1] = 16*1024;
3228         sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3229 
3230         sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3231         sysctl_tcp_rmem[1] = 87380;
3232         sysctl_tcp_rmem[2] = max(87380, max_rshare);
3233 
3234         pr_info("Hash tables configured (established %u bind %u)\n",
3235                 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3236 
3237         tcp_metrics_init();
3238         BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3239         tcp_tasklet_init();
3240 }
3241 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us