Version:  2.0.40 2.2.26 2.4.37 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19

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

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