?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael...
[linux-2.6.git] / net / ipv4 / tcp.c
blobb0a26bb25e2e3b341b2aad2e5f911fd99f6833fe
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 #include <linux/kernel.h>
249 #include <linux/module.h>
250 #include <linux/types.h>
251 #include <linux/fcntl.h>
252 #include <linux/poll.h>
253 #include <linux/init.h>
254 #include <linux/fs.h>
255 #include <linux/skbuff.h>
256 #include <linux/scatterlist.h>
257 #include <linux/splice.h>
258 #include <linux/net.h>
259 #include <linux/socket.h>
260 #include <linux/random.h>
261 #include <linux/bootmem.h>
262 #include <linux/highmem.h>
263 #include <linux/swap.h>
264 #include <linux/cache.h>
265 #include <linux/err.h>
266 #include <linux/crypto.h>
267 #include <linux/time.h>
268
269 #include <net/icmp.h>
270 #include <net/tcp.h>
271 #include <net/xfrm.h>
272 #include <net/ip.h>
273 #include <net/netdma.h>
274 #include <net/sock.h>
275
276 #include <asm/uaccess.h>
277 #include <asm/ioctls.h>
278
279 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
280
281 struct percpu_counter tcp_orphan_count;
282 EXPORT_SYMBOL_GPL(tcp_orphan_count);
283
284 int sysctl_tcp_mem[3] __read_mostly;
285 int sysctl_tcp_wmem[3] __read_mostly;
286 int sysctl_tcp_rmem[3] __read_mostly;
287
288 EXPORT_SYMBOL(sysctl_tcp_mem);
289 EXPORT_SYMBOL(sysctl_tcp_rmem);
290 EXPORT_SYMBOL(sysctl_tcp_wmem);
291
292 atomic_t tcp_memory_allocated; /* Current allocated memory. */
293 EXPORT_SYMBOL(tcp_memory_allocated);
294
295 /*
296 * Current number of TCP sockets.
297 */
298 struct percpu_counter tcp_sockets_allocated;
299 EXPORT_SYMBOL(tcp_sockets_allocated);
300
301 /*
302 * TCP splice context
303 */
304 struct tcp_splice_state {
305 struct pipe_inode_info *pipe;
306 size_t len;
307 unsigned int flags;
308 };
309
310 /*
311 * Pressure flag: try to collapse.
312 * Technical note: it is used by multiple contexts non atomically.
313 * All the __sk_mem_schedule() is of this nature: accounting
314 * is strict, actions are advisory and have some latency.
315 */
316 int tcp_memory_pressure __read_mostly;
317
318 EXPORT_SYMBOL(tcp_memory_pressure);
319
320 void tcp_enter_memory_pressure(struct sock *sk)
321 {
322 if (!tcp_memory_pressure) {
323 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
324 tcp_memory_pressure = 1;
325 }
326 }
327
328 EXPORT_SYMBOL(tcp_enter_memory_pressure);
329
330 /* Convert seconds to retransmits based on initial and max timeout */
331 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
332 {
333 u8 res = 0;
334
335 if (seconds > 0) {
336 int period = timeout;
337
338 res = 1;
339 while (seconds > period && res < 255) {
340 res++;
341 timeout <<= 1;
342 if (timeout > rto_max)
343 timeout = rto_max;
344 period += timeout;
345 }
346 }
347 return res;
348 }
349
350 /* Convert retransmits to seconds based on initial and max timeout */
351 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
352 {
353 int period = 0;
354
355 if (retrans > 0) {
356 period = timeout;
357 while (--retrans) {
358 timeout <<= 1;
359 if (timeout > rto_max)
360 timeout = rto_max;
361 period += timeout;
362 }
363 }
364 return period;
365 }
366
367 /*
368 * Wait for a TCP event.
369 *
370 * Note that we don't need to lock the socket, as the upper poll layers
371 * take care of normal races (between the test and the event) and we don't
372 * go look at any of the socket buffers directly.
373 */
374 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
375 {
376 unsigned int mask;
377 struct sock *sk = sock->sk;
378 struct tcp_sock *tp = tcp_sk(sk);
379
380 sock_poll_wait(file, sk->sk_sleep, wait);
381 if (sk->sk_state == TCP_LISTEN)
382 return inet_csk_listen_poll(sk);
383
384 /* Socket is not locked. We are protected from async events
385 * by poll logic and correct handling of state changes
386 * made by other threads is impossible in any case.
387 */
388
389 mask = 0;
390 if (sk->sk_err)
391 mask = POLLERR;
392
393 /*
394 * POLLHUP is certainly not done right. But poll() doesn't
395 * have a notion of HUP in just one direction, and for a
396 * socket the read side is more interesting.
397 *
398 * Some poll() documentation says that POLLHUP is incompatible
399 * with the POLLOUT/POLLWR flags, so somebody should check this
400 * all. But careful, it tends to be safer to return too many
401 * bits than too few, and you can easily break real applications
402 * if you don't tell them that something has hung up!
403 *
404 * Check-me.
405 *
406 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
407 * our fs/select.c). It means that after we received EOF,
408 * poll always returns immediately, making impossible poll() on write()
409 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
410 * if and only if shutdown has been made in both directions.
411 * Actually, it is interesting to look how Solaris and DUX
412 * solve this dilemma. I would prefer, if POLLHUP were maskable,
413 * then we could set it on SND_SHUTDOWN. BTW examples given
414 * in Stevens' books assume exactly this behaviour, it explains
415 * why POLLHUP is incompatible with POLLOUT. --ANK
416 *
417 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
418 * blocking on fresh not-connected or disconnected socket. --ANK
419 */
420 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
421 mask |= POLLHUP;
422 if (sk->sk_shutdown & RCV_SHUTDOWN)
423 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
424
425 /* Connected? */
426 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
427 int target = sock_rcvlowat(sk, 0, INT_MAX);
428
429 if (tp->urg_seq == tp->copied_seq &&
430 !sock_flag(sk, SOCK_URGINLINE) &&
431 tp->urg_data)
432 target--;
433
434 /* Potential race condition. If read of tp below will
435 * escape above sk->sk_state, we can be illegally awaken
436 * in SYN_* states. */
437 if (tp->rcv_nxt - tp->copied_seq >= target)
438 mask |= POLLIN | POLLRDNORM;
439
440 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
441 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
442 mask |= POLLOUT | POLLWRNORM;
443 } else { /* send SIGIO later */
444 set_bit(SOCK_ASYNC_NOSPACE,
445 &sk->sk_socket->flags);
446 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
447
448 /* Race breaker. If space is freed after
449 * wspace test but before the flags are set,
450 * IO signal will be lost.
451 */
452 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
453 mask |= POLLOUT | POLLWRNORM;
454 }
455 }
456
457 if (tp->urg_data & TCP_URG_VALID)
458 mask |= POLLPRI;
459 }
460 return mask;
461 }
462
463 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
464 {
465 struct tcp_sock *tp = tcp_sk(sk);
466 int answ;
467
468 switch (cmd) {
469 case SIOCINQ:
470 if (sk->sk_state == TCP_LISTEN)
471 return -EINVAL;
472
473 lock_sock(sk);
474 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
475 answ = 0;
476 else if (sock_flag(sk, SOCK_URGINLINE) ||
477 !tp->urg_data ||
478 before(tp->urg_seq, tp->copied_seq) ||
479 !before(tp->urg_seq, tp->rcv_nxt)) {
480 struct sk_buff *skb;
481
482 answ = tp->rcv_nxt - tp->copied_seq;
483
484 /* Subtract 1, if FIN is in queue. */
485 skb = skb_peek_tail(&sk->sk_receive_queue);
486 if (answ && skb)
487 answ -= tcp_hdr(skb)->fin;
488 } else
489 answ = tp->urg_seq - tp->copied_seq;
490 release_sock(sk);
491 break;
492 case SIOCATMARK:
493 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
494 break;
495 case SIOCOUTQ:
496 if (sk->sk_state == TCP_LISTEN)
497 return -EINVAL;
498
499 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
500 answ = 0;
501 else
502 answ = tp->write_seq - tp->snd_una;
503 break;
504 default:
505 return -ENOIOCTLCMD;
506 }
507
508 return put_user(answ, (int __user *)arg);
509 }
510
511 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
512 {
513 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
514 tp->pushed_seq = tp->write_seq;
515 }
516
517 static inline int forced_push(struct tcp_sock *tp)
518 {
519 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
520 }
521
522 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
523 {
524 struct tcp_sock *tp = tcp_sk(sk);
525 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
526
527 skb->csum = 0;
528 tcb->seq = tcb->end_seq = tp->write_seq;
529 tcb->flags = TCPCB_FLAG_ACK;
530 tcb->sacked = 0;
531 skb_header_release(skb);
532 tcp_add_write_queue_tail(sk, skb);
533 sk->sk_wmem_queued += skb->truesize;
534 sk_mem_charge(sk, skb->truesize);
535 if (tp->nonagle & TCP_NAGLE_PUSH)
536 tp->nonagle &= ~TCP_NAGLE_PUSH;
537 }
538
539 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
540 struct sk_buff *skb)
541 {
542 if (flags & MSG_OOB)
543 tp->snd_up = tp->write_seq;
544 }
545
546 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
547 int nonagle)
548 {
549 struct tcp_sock *tp = tcp_sk(sk);
550
551 if (tcp_send_head(sk)) {
552 struct sk_buff *skb = tcp_write_queue_tail(sk);
553 if (!(flags & MSG_MORE) || forced_push(tp))
554 tcp_mark_push(tp, skb);
555 tcp_mark_urg(tp, flags, skb);
556 __tcp_push_pending_frames(sk, mss_now,
557 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
558 }
559 }
560
561 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
562 unsigned int offset, size_t len)
563 {
564 struct tcp_splice_state *tss = rd_desc->arg.data;
565 int ret;
566
567 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
568 tss->flags);
569 if (ret > 0)
570 rd_desc->count -= ret;
571 return ret;
572 }
573
574 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
575 {
576 /* Store TCP splice context information in read_descriptor_t. */
577 read_descriptor_t rd_desc = {
578 .arg.data = tss,
579 .count = tss->len,
580 };
581
582 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
583 }
584
585 /**
586 * tcp_splice_read - splice data from TCP socket to a pipe
587 * @sock: socket to splice from
588 * @ppos: position (not valid)
589 * @pipe: pipe to splice to
590 * @len: number of bytes to splice
591 * @flags: splice modifier flags
592 *
593 * Description:
594 * Will read pages from given socket and fill them into a pipe.
595 *
596 **/
597 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
598 struct pipe_inode_info *pipe, size_t len,
599 unsigned int flags)
600 {
601 struct sock *sk = sock->sk;
602 struct tcp_splice_state tss = {
603 .pipe = pipe,
604 .len = len,
605 .flags = flags,
606 };
607 long timeo;
608 ssize_t spliced;
609 int ret;
610
611 /*
612 * We can't seek on a socket input
613 */
614 if (unlikely(*ppos))
615 return -ESPIPE;
616
617 ret = spliced = 0;
618
619 lock_sock(sk);
620
621 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
622 while (tss.len) {
623 ret = __tcp_splice_read(sk, &tss);
624 if (ret < 0)
625 break;
626 else if (!ret) {
627 if (spliced)
628 break;
629 if (sock_flag(sk, SOCK_DONE))
630 break;
631 if (sk->sk_err) {
632 ret = sock_error(sk);
633 break;
634 }
635 if (sk->sk_shutdown & RCV_SHUTDOWN)
636 break;
637 if (sk->sk_state == TCP_CLOSE) {
638 /*
639 * This occurs when user tries to read
640 * from never connected socket.
641 */
642 if (!sock_flag(sk, SOCK_DONE))
643 ret = -ENOTCONN;
644 break;
645 }
646 if (!timeo) {
647 ret = -EAGAIN;
648 break;
649 }
650 sk_wait_data(sk, &timeo);
651 if (signal_pending(current)) {
652 ret = sock_intr_errno(timeo);
653 break;
654 }
655 continue;
656 }
657 tss.len -= ret;
658 spliced += ret;
659
660 if (!timeo)
661 break;
662 release_sock(sk);
663 lock_sock(sk);
664
665 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
666 (sk->sk_shutdown & RCV_SHUTDOWN) ||
667 signal_pending(current))
668 break;
669 }
670
671 release_sock(sk);
672
673 if (spliced)
674 return spliced;
675
676 return ret;
677 }
678
679 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
680 {
681 struct sk_buff *skb;
682
683 /* The TCP header must be at least 32-bit aligned. */
684 size = ALIGN(size, 4);
685
686 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
687 if (skb) {
688 if (sk_wmem_schedule(sk, skb->truesize)) {
689 /*
690 * Make sure that we have exactly size bytes
691 * available to the caller, no more, no less.
692 */
693 skb_reserve(skb, skb_tailroom(skb) - size);
694 return skb;
695 }
696 __kfree_skb(skb);
697 } else {
698 sk->sk_prot->enter_memory_pressure(sk);
699 sk_stream_moderate_sndbuf(sk);
700 }
701 return NULL;
702 }
703
704 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
705 int large_allowed)
706 {
707 struct tcp_sock *tp = tcp_sk(sk);
708 u32 xmit_size_goal, old_size_goal;
709
710 xmit_size_goal = mss_now;
711
712 if (large_allowed && sk_can_gso(sk)) {
713 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
714 inet_csk(sk)->icsk_af_ops->net_header_len -
715 inet_csk(sk)->icsk_ext_hdr_len -
716 tp->tcp_header_len);
717
718 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
719
720 /* We try hard to avoid divides here */
721 old_size_goal = tp->xmit_size_goal_segs * mss_now;
722
723 if (likely(old_size_goal <= xmit_size_goal &&
724 old_size_goal + mss_now > xmit_size_goal)) {
725 xmit_size_goal = old_size_goal;
726 } else {
727 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
728 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
729 }
730 }
731
732 return max(xmit_size_goal, mss_now);
733 }
734
735 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
736 {
737 int mss_now;
738
739 mss_now = tcp_current_mss(sk);
740 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
741
742 return mss_now;
743 }
744
745 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
746 size_t psize, int flags)
747 {
748 struct tcp_sock *tp = tcp_sk(sk);
749 int mss_now, size_goal;
750 int err;
751 ssize_t copied;
752 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
753
754 /* Wait for a connection to finish. */
755 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
756 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
757 goto out_err;
758
759 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
760
761 mss_now = tcp_send_mss(sk, &size_goal, flags);
762 copied = 0;
763
764 err = -EPIPE;
765 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
766 goto out_err;
767
768 while (psize > 0) {
769 struct sk_buff *skb = tcp_write_queue_tail(sk);
770 struct page *page = pages[poffset / PAGE_SIZE];
771 int copy, i, can_coalesce;
772 int offset = poffset % PAGE_SIZE;
773 int size = min_t(size_t, psize, PAGE_SIZE - offset);
774
775 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
776 new_segment:
777 if (!sk_stream_memory_free(sk))
778 goto wait_for_sndbuf;
779
780 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
781 if (!skb)
782 goto wait_for_memory;
783
784 skb_entail(sk, skb);
785 copy = size_goal;
786 }
787
788 if (copy > size)
789 copy = size;
790
791 i = skb_shinfo(skb)->nr_frags;
792 can_coalesce = skb_can_coalesce(skb, i, page, offset);
793 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
794 tcp_mark_push(tp, skb);
795 goto new_segment;
796 }
797 if (!sk_wmem_schedule(sk, copy))
798 goto wait_for_memory;
799
800 if (can_coalesce) {
801 skb_shinfo(skb)->frags[i - 1].size += copy;
802 } else {
803 get_page(page);
804 skb_fill_page_desc(skb, i, page, offset, copy);
805 }
806
807 skb->len += copy;
808 skb->data_len += copy;
809 skb->truesize += copy;
810 sk->sk_wmem_queued += copy;
811 sk_mem_charge(sk, copy);
812 skb->ip_summed = CHECKSUM_PARTIAL;
813 tp->write_seq += copy;
814 TCP_SKB_CB(skb)->end_seq += copy;
815 skb_shinfo(skb)->gso_segs = 0;
816
817 if (!copied)
818 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
819
820 copied += copy;
821 poffset += copy;
822 if (!(psize -= copy))
823 goto out;
824
825 if (skb->len < size_goal || (flags & MSG_OOB))
826 continue;
827
828 if (forced_push(tp)) {
829 tcp_mark_push(tp, skb);
830 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
831 } else if (skb == tcp_send_head(sk))
832 tcp_push_one(sk, mss_now);
833 continue;
834
835 wait_for_sndbuf:
836 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
837 wait_for_memory:
838 if (copied)
839 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
840
841 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
842 goto do_error;
843
844 mss_now = tcp_send_mss(sk, &size_goal, flags);
845 }
846
847 out:
848 if (copied)
849 tcp_push(sk, flags, mss_now, tp->nonagle);
850 return copied;
851
852 do_error:
853 if (copied)
854 goto out;
855 out_err:
856 return sk_stream_error(sk, flags, err);
857 }
858
859 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
860 size_t size, int flags)
861 {
862 ssize_t res;
863 struct sock *sk = sock->sk;
864
865 if (!(sk->sk_route_caps & NETIF_F_SG) ||
866 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
867 return sock_no_sendpage(sock, page, offset, size, flags);
868
869 lock_sock(sk);
870 TCP_CHECK_TIMER(sk);
871 res = do_tcp_sendpages(sk, &page, offset, size, flags);
872 TCP_CHECK_TIMER(sk);
873 release_sock(sk);
874 return res;
875 }
876
877 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
878 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
879
880 static inline int select_size(struct sock *sk)
881 {
882 struct tcp_sock *tp = tcp_sk(sk);
883 int tmp = tp->mss_cache;
884
885 if (sk->sk_route_caps & NETIF_F_SG) {
886 if (sk_can_gso(sk))
887 tmp = 0;
888 else {
889 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
890
891 if (tmp >= pgbreak &&
892 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
893 tmp = pgbreak;
894 }
895 }
896
897 return tmp;
898 }
899
900 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
901 size_t size)
902 {
903 struct sock *sk = sock->sk;
904 struct iovec *iov;
905 struct tcp_sock *tp = tcp_sk(sk);
906 struct sk_buff *skb;
907 int iovlen, flags;
908 int mss_now, size_goal;
909 int err, copied;
910 long timeo;
911
912 lock_sock(sk);
913 TCP_CHECK_TIMER(sk);
914
915 flags = msg->msg_flags;
916 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
917
918 /* Wait for a connection to finish. */
919 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
920 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
921 goto out_err;
922
923 /* This should be in poll */
924 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
925
926 mss_now = tcp_send_mss(sk, &size_goal, flags);
927
928 /* Ok commence sending. */
929 iovlen = msg->msg_iovlen;
930 iov = msg->msg_iov;
931 copied = 0;
932
933 err = -EPIPE;
934 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
935 goto out_err;
936
937 while (--iovlen >= 0) {
938 int seglen = iov->iov_len;
939 unsigned char __user *from = iov->iov_base;
940
941 iov++;
942
943 while (seglen > 0) {
944 int copy = 0;
945 int max = size_goal;
946
947 skb = tcp_write_queue_tail(sk);
948 if (tcp_send_head(sk)) {
949 if (skb->ip_summed == CHECKSUM_NONE)
950 max = mss_now;
951 copy = max - skb->len;
952 }
953
954 if (copy <= 0) {
955 new_segment:
956 /* Allocate new segment. If the interface is SG,
957 * allocate skb fitting to single page.
958 */
959 if (!sk_stream_memory_free(sk))
960 goto wait_for_sndbuf;
961
962 skb = sk_stream_alloc_skb(sk, select_size(sk),
963 sk->sk_allocation);
964 if (!skb)
965 goto wait_for_memory;
966
967 /*
968 * Check whether we can use HW checksum.
969 */
970 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
971 skb->ip_summed = CHECKSUM_PARTIAL;
972
973 skb_entail(sk, skb);
974 copy = size_goal;
975 max = size_goal;
976 }
977
978 /* Try to append data to the end of skb. */
979 if (copy > seglen)
980 copy = seglen;
981
982 /* Where to copy to? */
983 if (skb_tailroom(skb) > 0) {
984 /* We have some space in skb head. Superb! */
985 if (copy > skb_tailroom(skb))
986 copy = skb_tailroom(skb);
987 if ((err = skb_add_data(skb, from, copy)) != 0)
988 goto do_fault;
989 } else {
990 int merge = 0;
991 int i = skb_shinfo(skb)->nr_frags;
992 struct page *page = TCP_PAGE(sk);
993 int off = TCP_OFF(sk);
994
995 if (skb_can_coalesce(skb, i, page, off) &&
996 off != PAGE_SIZE) {
997 /* We can extend the last page
998 * fragment. */
999 merge = 1;
1000 } else if (i == MAX_SKB_FRAGS ||
1001 (!i &&
1002 !(sk->sk_route_caps & NETIF_F_SG))) {
1003 /* Need to add new fragment and cannot
1004 * do this because interface is non-SG,
1005 * or because all the page slots are
1006 * busy. */
1007 tcp_mark_push(tp, skb);
1008 goto new_segment;
1009 } else if (page) {
1010 if (off == PAGE_SIZE) {
1011 put_page(page);
1012 TCP_PAGE(sk) = page = NULL;
1013 off = 0;
1014 }
1015 } else
1016 off = 0;
1017
1018 if (copy > PAGE_SIZE - off)
1019 copy = PAGE_SIZE - off;
1020
1021 if (!sk_wmem_schedule(sk, copy))
1022 goto wait_for_memory;
1023
1024 if (!page) {
1025 /* Allocate new cache page. */
1026 if (!(page = sk_stream_alloc_page(sk)))
1027 goto wait_for_memory;
1028 }
1029
1030 /* Time to copy data. We are close to
1031 * the end! */
1032 err = skb_copy_to_page(sk, from, skb, page,
1033 off, copy);
1034 if (err) {
1035 /* If this page was new, give it to the
1036 * socket so it does not get leaked.
1037 */
1038 if (!TCP_PAGE(sk)) {
1039 TCP_PAGE(sk) = page;
1040 TCP_OFF(sk) = 0;
1041 }
1042 goto do_error;
1043 }
1044
1045 /* Update the skb. */
1046 if (merge) {
1047 skb_shinfo(skb)->frags[i - 1].size +=
1048 copy;
1049 } else {
1050 skb_fill_page_desc(skb, i, page, off, copy);
1051 if (TCP_PAGE(sk)) {
1052 get_page(page);
1053 } else if (off + copy < PAGE_SIZE) {
1054 get_page(page);
1055 TCP_PAGE(sk) = page;
1056 }
1057 }
1058
1059 TCP_OFF(sk) = off + copy;
1060 }
1061
1062 if (!copied)
1063 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
1064
1065 tp->write_seq += copy;
1066 TCP_SKB_CB(skb)->end_seq += copy;
1067 skb_shinfo(skb)->gso_segs = 0;
1068
1069 from += copy;
1070 copied += copy;
1071 if ((seglen -= copy) == 0 && iovlen == 0)
1072 goto out;
1073
1074 if (skb->len < max || (flags & MSG_OOB))
1075 continue;
1076
1077 if (forced_push(tp)) {
1078 tcp_mark_push(tp, skb);
1079 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1080 } else if (skb == tcp_send_head(sk))
1081 tcp_push_one(sk, mss_now);
1082 continue;
1083
1084 wait_for_sndbuf:
1085 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1086 wait_for_memory:
1087 if (copied)
1088 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1089
1090 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1091 goto do_error;
1092
1093 mss_now = tcp_send_mss(sk, &size_goal, flags);
1094 }
1095 }
1096
1097 out:
1098 if (copied)
1099 tcp_push(sk, flags, mss_now, tp->nonagle);
1100 TCP_CHECK_TIMER(sk);
1101 release_sock(sk);
1102 return copied;
1103
1104 do_fault:
1105 if (!skb->len) {
1106 tcp_unlink_write_queue(skb, sk);
1107 /* It is the one place in all of TCP, except connection
1108 * reset, where we can be unlinking the send_head.
1109 */
1110 tcp_check_send_head(sk, skb);
1111 sk_wmem_free_skb(sk, skb);
1112 }
1113
1114 do_error:
1115 if (copied)
1116 goto out;
1117 out_err:
1118 err = sk_stream_error(sk, flags, err);
1119 TCP_CHECK_TIMER(sk);
1120 release_sock(sk);
1121 return err;
1122 }
1123
1124 /*
1125 * Handle reading urgent data. BSD has very simple semantics for
1126 * this, no blocking and very strange errors 8)
1127 */
1128
1129 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1130 {
1131 struct tcp_sock *tp = tcp_sk(sk);
1132
1133 /* No URG data to read. */
1134 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1135 tp->urg_data == TCP_URG_READ)
1136 return -EINVAL; /* Yes this is right ! */
1137
1138 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1139 return -ENOTCONN;
1140
1141 if (tp->urg_data & TCP_URG_VALID) {
1142 int err = 0;
1143 char c = tp->urg_data;
1144
1145 if (!(flags & MSG_PEEK))
1146 tp->urg_data = TCP_URG_READ;
1147
1148 /* Read urgent data. */
1149 msg->msg_flags |= MSG_OOB;
1150
1151 if (len > 0) {
1152 if (!(flags & MSG_TRUNC))
1153 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1154 len = 1;
1155 } else
1156 msg->msg_flags |= MSG_TRUNC;
1157
1158 return err ? -EFAULT : len;
1159 }
1160
1161 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1162 return 0;
1163
1164 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1165 * the available implementations agree in this case:
1166 * this call should never block, independent of the
1167 * blocking state of the socket.
1168 * Mike <pall@rz.uni-karlsruhe.de>
1169 */
1170 return -EAGAIN;
1171 }
1172
1173 /* Clean up the receive buffer for full frames taken by the user,
1174 * then send an ACK if necessary. COPIED is the number of bytes
1175 * tcp_recvmsg has given to the user so far, it speeds up the
1176 * calculation of whether or not we must ACK for the sake of
1177 * a window update.
1178 */
1179 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1180 {
1181 struct tcp_sock *tp = tcp_sk(sk);
1182 int time_to_ack = 0;
1183
1184 #if TCP_DEBUG
1185 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1186
1187 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1188 KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1189 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1190 #endif
1191
1192 if (inet_csk_ack_scheduled(sk)) {
1193 const struct inet_connection_sock *icsk = inet_csk(sk);
1194 /* Delayed ACKs frequently hit locked sockets during bulk
1195 * receive. */
1196 if (icsk->icsk_ack.blocked ||
1197 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1198 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1199 /*
1200 * If this read emptied read buffer, we send ACK, if
1201 * connection is not bidirectional, user drained
1202 * receive buffer and there was a small segment
1203 * in queue.
1204 */
1205 (copied > 0 &&
1206 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1207 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1208 !icsk->icsk_ack.pingpong)) &&
1209 !atomic_read(&sk->sk_rmem_alloc)))
1210 time_to_ack = 1;
1211 }
1212
1213 /* We send an ACK if we can now advertise a non-zero window
1214 * which has been raised "significantly".
1215 *
1216 * Even if window raised up to infinity, do not send window open ACK
1217 * in states, where we will not receive more. It is useless.
1218 */
1219 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1220 __u32 rcv_window_now = tcp_receive_window(tp);
1221
1222 /* Optimize, __tcp_select_window() is not cheap. */
1223 if (2*rcv_window_now <= tp->window_clamp) {
1224 __u32 new_window = __tcp_select_window(sk);
1225
1226 /* Send ACK now, if this read freed lots of space
1227 * in our buffer. Certainly, new_window is new window.
1228 * We can advertise it now, if it is not less than current one.
1229 * "Lots" means "at least twice" here.
1230 */
1231 if (new_window && new_window >= 2 * rcv_window_now)
1232 time_to_ack = 1;
1233 }
1234 }
1235 if (time_to_ack)
1236 tcp_send_ack(sk);
1237 }
1238
1239 static void tcp_prequeue_process(struct sock *sk)
1240 {
1241 struct sk_buff *skb;
1242 struct tcp_sock *tp = tcp_sk(sk);
1243
1244 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1245
1246 /* RX process wants to run with disabled BHs, though it is not
1247 * necessary */
1248 local_bh_disable();
1249 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1250 sk_backlog_rcv(sk, skb);
1251 local_bh_enable();
1252
1253 /* Clear memory counter. */
1254 tp->ucopy.memory = 0;
1255 }
1256
1257 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1258 {
1259 struct sk_buff *skb;
1260 u32 offset;
1261
1262 skb_queue_walk(&sk->sk_receive_queue, skb) {
1263 offset = seq - TCP_SKB_CB(skb)->seq;
1264 if (tcp_hdr(skb)->syn)
1265 offset--;
1266 if (offset < skb->len || tcp_hdr(skb)->fin) {
1267 *off = offset;
1268 return skb;
1269 }
1270 }
1271 return NULL;
1272 }
1273
1274 /*
1275 * This routine provides an alternative to tcp_recvmsg() for routines
1276 * that would like to handle copying from skbuffs directly in 'sendfile'
1277 * fashion.
1278 * Note:
1279 * - It is assumed that the socket was locked by the caller.
1280 * - The routine does not block.
1281 * - At present, there is no support for reading OOB data
1282 * or for 'peeking' the socket using this routine
1283 * (although both would be easy to implement).
1284 */
1285 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1286 sk_read_actor_t recv_actor)
1287 {
1288 struct sk_buff *skb;
1289 struct tcp_sock *tp = tcp_sk(sk);
1290 u32 seq = tp->copied_seq;
1291 u32 offset;
1292 int copied = 0;
1293
1294 if (sk->sk_state == TCP_LISTEN)
1295 return -ENOTCONN;
1296 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1297 if (offset < skb->len) {
1298 int used;
1299 size_t len;
1300
1301 len = skb->len - offset;
1302 /* Stop reading if we hit a patch of urgent data */
1303 if (tp->urg_data) {
1304 u32 urg_offset = tp->urg_seq - seq;
1305 if (urg_offset < len)
1306 len = urg_offset;
1307 if (!len)
1308 break;
1309 }
1310 used = recv_actor(desc, skb, offset, len);
1311 if (used < 0) {
1312 if (!copied)
1313 copied = used;
1314 break;
1315 } else if (used <= len) {
1316 seq += used;
1317 copied += used;
1318 offset += used;
1319 }
1320 /*
1321 * If recv_actor drops the lock (e.g. TCP splice
1322 * receive) the skb pointer might be invalid when
1323 * getting here: tcp_collapse might have deleted it
1324 * while aggregating skbs from the socket queue.
1325 */
1326 skb = tcp_recv_skb(sk, seq-1, &offset);
1327 if (!skb || (offset+1 != skb->len))
1328 break;
1329 }
1330 if (tcp_hdr(skb)->fin) {
1331 sk_eat_skb(sk, skb, 0);
1332 ++seq;
1333 break;
1334 }
1335 sk_eat_skb(sk, skb, 0);
1336 if (!desc->count)
1337 break;
1338 }
1339 tp->copied_seq = seq;
1340
1341 tcp_rcv_space_adjust(sk);
1342
1343 /* Clean up data we have read: This will do ACK frames. */
1344 if (copied > 0)
1345 tcp_cleanup_rbuf(sk, copied);
1346 return copied;
1347 }
1348
1349 /*
1350 * This routine copies from a sock struct into the user buffer.
1351 *
1352 * Technical note: in 2.3 we work on _locked_ socket, so that
1353 * tricks with *seq access order and skb->users are not required.
1354 * Probably, code can be easily improved even more.
1355 */
1356
1357 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1358 size_t len, int nonblock, int flags, int *addr_len)
1359 {
1360 struct tcp_sock *tp = tcp_sk(sk);
1361 int copied = 0;
1362 u32 peek_seq;
1363 u32 *seq;
1364 unsigned long used;
1365 int err;
1366 int target; /* Read at least this many bytes */
1367 long timeo;
1368 struct task_struct *user_recv = NULL;
1369 int copied_early = 0;
1370 struct sk_buff *skb;
1371 u32 urg_hole = 0;
1372
1373 lock_sock(sk);
1374
1375 TCP_CHECK_TIMER(sk);
1376
1377 err = -ENOTCONN;
1378 if (sk->sk_state == TCP_LISTEN)
1379 goto out;
1380
1381 timeo = sock_rcvtimeo(sk, nonblock);
1382
1383 /* Urgent data needs to be handled specially. */
1384 if (flags & MSG_OOB)
1385 goto recv_urg;
1386
1387 seq = &tp->copied_seq;
1388 if (flags & MSG_PEEK) {
1389 peek_seq = tp->copied_seq;
1390 seq = &peek_seq;
1391 }
1392
1393 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1394
1395 #ifdef CONFIG_NET_DMA
1396 tp->ucopy.dma_chan = NULL;
1397 preempt_disable();
1398 skb = skb_peek_tail(&sk->sk_receive_queue);
1399 {
1400 int available = 0;
1401
1402 if (skb)
1403 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1404 if ((available < target) &&
1405 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1406 !sysctl_tcp_low_latency &&
1407 dma_find_channel(DMA_MEMCPY)) {
1408 preempt_enable_no_resched();
1409 tp->ucopy.pinned_list =
1410 dma_pin_iovec_pages(msg->msg_iov, len);
1411 } else {
1412 preempt_enable_no_resched();
1413 }
1414 }
1415 #endif
1416
1417 do {
1418 u32 offset;
1419
1420 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1421 if (tp->urg_data && tp->urg_seq == *seq) {
1422 if (copied)
1423 break;
1424 if (signal_pending(current)) {
1425 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1426 break;
1427 }
1428 }
1429
1430 /* Next get a buffer. */
1431
1432 skb_queue_walk(&sk->sk_receive_queue, skb) {
1433 /* Now that we have two receive queues this
1434 * shouldn't happen.
1435 */
1436 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1437 KERN_INFO "recvmsg bug: copied %X "
1438 "seq %X rcvnxt %X fl %X\n", *seq,
1439 TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1440 flags))
1441 break;
1442
1443 offset = *seq - TCP_SKB_CB(skb)->seq;
1444 if (tcp_hdr(skb)->syn)
1445 offset--;
1446 if (offset < skb->len)
1447 goto found_ok_skb;
1448 if (tcp_hdr(skb)->fin)
1449 goto found_fin_ok;
1450 WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: "
1451 "copied %X seq %X rcvnxt %X fl %X\n",
1452 *seq, TCP_SKB_CB(skb)->seq,
1453 tp->rcv_nxt, flags);
1454 }
1455
1456 /* Well, if we have backlog, try to process it now yet. */
1457
1458 if (copied >= target && !sk->sk_backlog.tail)
1459 break;
1460
1461 if (copied) {
1462 if (sk->sk_err ||
1463 sk->sk_state == TCP_CLOSE ||
1464 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1465 !timeo ||
1466 signal_pending(current))
1467 break;
1468 } else {
1469 if (sock_flag(sk, SOCK_DONE))
1470 break;
1471
1472 if (sk->sk_err) {
1473 copied = sock_error(sk);
1474 break;
1475 }
1476
1477 if (sk->sk_shutdown & RCV_SHUTDOWN)
1478 break;
1479
1480 if (sk->sk_state == TCP_CLOSE) {
1481 if (!sock_flag(sk, SOCK_DONE)) {
1482 /* This occurs when user tries to read
1483 * from never connected socket.
1484 */
1485 copied = -ENOTCONN;
1486 break;
1487 }
1488 break;
1489 }
1490
1491 if (!timeo) {
1492 copied = -EAGAIN;
1493 break;
1494 }
1495
1496 if (signal_pending(current)) {
1497 copied = sock_intr_errno(timeo);
1498 break;
1499 }
1500 }
1501
1502 tcp_cleanup_rbuf(sk, copied);
1503
1504 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1505 /* Install new reader */
1506 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1507 user_recv = current;
1508 tp->ucopy.task = user_recv;
1509 tp->ucopy.iov = msg->msg_iov;
1510 }
1511
1512 tp->ucopy.len = len;
1513
1514 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1515 !(flags & (MSG_PEEK | MSG_TRUNC)));
1516
1517 /* Ugly... If prequeue is not empty, we have to
1518 * process it before releasing socket, otherwise
1519 * order will be broken at second iteration.
1520 * More elegant solution is required!!!
1521 *
1522 * Look: we have the following (pseudo)queues:
1523 *
1524 * 1. packets in flight
1525 * 2. backlog
1526 * 3. prequeue
1527 * 4. receive_queue
1528 *
1529 * Each queue can be processed only if the next ones
1530 * are empty. At this point we have empty receive_queue.
1531 * But prequeue _can_ be not empty after 2nd iteration,
1532 * when we jumped to start of loop because backlog
1533 * processing added something to receive_queue.
1534 * We cannot release_sock(), because backlog contains
1535 * packets arrived _after_ prequeued ones.
1536 *
1537 * Shortly, algorithm is clear --- to process all
1538 * the queues in order. We could make it more directly,
1539 * requeueing packets from backlog to prequeue, if
1540 * is not empty. It is more elegant, but eats cycles,
1541 * unfortunately.
1542 */
1543 if (!skb_queue_empty(&tp->ucopy.prequeue))
1544 goto do_prequeue;
1545
1546 /* __ Set realtime policy in scheduler __ */
1547 }
1548
1549 if (copied >= target) {
1550 /* Do not sleep, just process backlog. */
1551 release_sock(sk);
1552 lock_sock(sk);
1553 } else
1554 sk_wait_data(sk, &timeo);
1555
1556 #ifdef CONFIG_NET_DMA
1557 tp->ucopy.wakeup = 0;
1558 #endif
1559
1560 if (user_recv) {
1561 int chunk;
1562
1563 /* __ Restore normal policy in scheduler __ */
1564
1565 if ((chunk = len - tp->ucopy.len) != 0) {
1566 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1567 len -= chunk;
1568 copied += chunk;
1569 }
1570
1571 if (tp->rcv_nxt == tp->copied_seq &&
1572 !skb_queue_empty(&tp->ucopy.prequeue)) {
1573 do_prequeue:
1574 tcp_prequeue_process(sk);
1575
1576 if ((chunk = len - tp->ucopy.len) != 0) {
1577 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1578 len -= chunk;
1579 copied += chunk;
1580 }
1581 }
1582 }
1583 if ((flags & MSG_PEEK) &&
1584 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1585 if (net_ratelimit())
1586 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1587 current->comm, task_pid_nr(current));
1588 peek_seq = tp->copied_seq;
1589 }
1590 continue;
1591
1592 found_ok_skb:
1593 /* Ok so how much can we use? */
1594 used = skb->len - offset;
1595 if (len < used)
1596 used = len;
1597
1598 /* Do we have urgent data here? */
1599 if (tp->urg_data) {
1600 u32 urg_offset = tp->urg_seq - *seq;
1601 if (urg_offset < used) {
1602 if (!urg_offset) {
1603 if (!sock_flag(sk, SOCK_URGINLINE)) {
1604 ++*seq;
1605 urg_hole++;
1606 offset++;
1607 used--;
1608 if (!used)
1609 goto skip_copy;
1610 }
1611 } else
1612 used = urg_offset;
1613 }
1614 }
1615
1616 if (!(flags & MSG_TRUNC)) {
1617 #ifdef CONFIG_NET_DMA
1618 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1619 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1620
1621 if (tp->ucopy.dma_chan) {
1622 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1623 tp->ucopy.dma_chan, skb, offset,
1624 msg->msg_iov, used,
1625 tp->ucopy.pinned_list);
1626
1627 if (tp->ucopy.dma_cookie < 0) {
1628
1629 printk(KERN_ALERT "dma_cookie < 0\n");
1630
1631 /* Exception. Bailout! */
1632 if (!copied)
1633 copied = -EFAULT;
1634 break;
1635 }
1636 if ((offset + used) == skb->len)
1637 copied_early = 1;
1638
1639 } else
1640 #endif
1641 {
1642 err = skb_copy_datagram_iovec(skb, offset,
1643 msg->msg_iov, used);
1644 if (err) {
1645 /* Exception. Bailout! */
1646 if (!copied)
1647 copied = -EFAULT;
1648 break;
1649 }
1650 }
1651 }
1652
1653 *seq += used;
1654 copied += used;
1655 len -= used;
1656
1657 tcp_rcv_space_adjust(sk);
1658
1659 skip_copy:
1660 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1661 tp->urg_data = 0;
1662 tcp_fast_path_check(sk);
1663 }
1664 if (used + offset < skb->len)
1665 continue;
1666
1667 if (tcp_hdr(skb)->fin)
1668 goto found_fin_ok;
1669 if (!(flags & MSG_PEEK)) {
1670 sk_eat_skb(sk, skb, copied_early);
1671 copied_early = 0;
1672 }
1673 continue;
1674
1675 found_fin_ok:
1676 /* Process the FIN. */
1677 ++*seq;
1678 if (!(flags & MSG_PEEK)) {
1679 sk_eat_skb(sk, skb, copied_early);
1680 copied_early = 0;
1681 }
1682 break;
1683 } while (len > 0);
1684
1685 if (user_recv) {
1686 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1687 int chunk;
1688
1689 tp->ucopy.len = copied > 0 ? len : 0;
1690
1691 tcp_prequeue_process(sk);
1692
1693 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1694 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1695 len -= chunk;
1696 copied += chunk;
1697 }
1698 }
1699
1700 tp->ucopy.task = NULL;
1701 tp->ucopy.len = 0;
1702 }
1703
1704 #ifdef CONFIG_NET_DMA
1705 if (tp->ucopy.dma_chan) {
1706 dma_cookie_t done, used;
1707
1708 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1709
1710 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1711 tp->ucopy.dma_cookie, &done,
1712 &used) == DMA_IN_PROGRESS) {
1713 /* do partial cleanup of sk_async_wait_queue */
1714 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1715 (dma_async_is_complete(skb->dma_cookie, done,
1716 used) == DMA_SUCCESS)) {
1717 __skb_dequeue(&sk->sk_async_wait_queue);
1718 kfree_skb(skb);
1719 }
1720 }
1721
1722 /* Safe to free early-copied skbs now */
1723 __skb_queue_purge(&sk->sk_async_wait_queue);
1724 tp->ucopy.dma_chan = NULL;
1725 }
1726 if (tp->ucopy.pinned_list) {
1727 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1728 tp->ucopy.pinned_list = NULL;
1729 }
1730 #endif
1731
1732 /* According to UNIX98, msg_name/msg_namelen are ignored
1733 * on connected socket. I was just happy when found this 8) --ANK
1734 */
1735
1736 /* Clean up data we have read: This will do ACK frames. */
1737 tcp_cleanup_rbuf(sk, copied);
1738
1739 TCP_CHECK_TIMER(sk);
1740 release_sock(sk);
1741 return copied;
1742
1743 out:
1744 TCP_CHECK_TIMER(sk);
1745 release_sock(sk);
1746 return err;
1747
1748 recv_urg:
1749 err = tcp_recv_urg(sk, msg, len, flags);
1750 goto out;
1751 }
1752
1753 void tcp_set_state(struct sock *sk, int state)
1754 {
1755 int oldstate = sk->sk_state;
1756
1757 switch (state) {
1758 case TCP_ESTABLISHED:
1759 if (oldstate != TCP_ESTABLISHED)
1760 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1761 break;
1762
1763 case TCP_CLOSE:
1764 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1765 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1766
1767 sk->sk_prot->unhash(sk);
1768 if (inet_csk(sk)->icsk_bind_hash &&
1769 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1770 inet_put_port(sk);
1771 /* fall through */
1772 default:
1773 if (oldstate == TCP_ESTABLISHED)
1774 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1775 }
1776
1777 /* Change state AFTER socket is unhashed to avoid closed
1778 * socket sitting in hash tables.
1779 */
1780 sk->sk_state = state;
1781
1782 #ifdef STATE_TRACE
1783 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1784 #endif
1785 }
1786 EXPORT_SYMBOL_GPL(tcp_set_state);
1787
1788 /*
1789 * State processing on a close. This implements the state shift for
1790 * sending our FIN frame. Note that we only send a FIN for some
1791 * states. A shutdown() may have already sent the FIN, or we may be
1792 * closed.
1793 */
1794
1795 static const unsigned char new_state[16] = {
1796 /* current state: new state: action: */
1797 /* (Invalid) */ TCP_CLOSE,
1798 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1799 /* TCP_SYN_SENT */ TCP_CLOSE,
1800 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1801 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1802 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1803 /* TCP_TIME_WAIT */ TCP_CLOSE,
1804 /* TCP_CLOSE */ TCP_CLOSE,
1805 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1806 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1807 /* TCP_LISTEN */ TCP_CLOSE,
1808 /* TCP_CLOSING */ TCP_CLOSING,
1809 };
1810
1811 static int tcp_close_state(struct sock *sk)
1812 {
1813 int next = (int)new_state[sk->sk_state];
1814 int ns = next & TCP_STATE_MASK;
1815
1816 tcp_set_state(sk, ns);
1817
1818 return next & TCP_ACTION_FIN;
1819 }
1820
1821 /*
1822 * Shutdown the sending side of a connection. Much like close except
1823 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1824 */
1825
1826 void tcp_shutdown(struct sock *sk, int how)
1827 {
1828 /* We need to grab some memory, and put together a FIN,
1829 * and then put it into the queue to be sent.
1830 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1831 */
1832 if (!(how & SEND_SHUTDOWN))
1833 return;
1834
1835 /* If we've already sent a FIN, or it's a closed state, skip this. */
1836 if ((1 << sk->sk_state) &
1837 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1838 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1839 /* Clear out any half completed packets. FIN if needed. */
1840 if (tcp_close_state(sk))
1841 tcp_send_fin(sk);
1842 }
1843 }
1844
1845 void tcp_close(struct sock *sk, long timeout)
1846 {
1847 struct sk_buff *skb;
1848 int data_was_unread = 0;
1849 int state;
1850
1851 lock_sock(sk);
1852 sk->sk_shutdown = SHUTDOWN_MASK;
1853
1854 if (sk->sk_state == TCP_LISTEN) {
1855 tcp_set_state(sk, TCP_CLOSE);
1856
1857 /* Special case. */
1858 inet_csk_listen_stop(sk);
1859
1860 goto adjudge_to_death;
1861 }
1862
1863 /* We need to flush the recv. buffs. We do this only on the
1864 * descriptor close, not protocol-sourced closes, because the
1865 * reader process may not have drained the data yet!
1866 */
1867 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1868 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1869 tcp_hdr(skb)->fin;
1870 data_was_unread += len;
1871 __kfree_skb(skb);
1872 }
1873
1874 sk_mem_reclaim(sk);
1875
1876 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1877 * data was lost. To witness the awful effects of the old behavior of
1878 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1879 * GET in an FTP client, suspend the process, wait for the client to
1880 * advertise a zero window, then kill -9 the FTP client, wheee...
1881 * Note: timeout is always zero in such a case.
1882 */
1883 if (data_was_unread) {
1884 /* Unread data was tossed, zap the connection. */
1885 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1886 tcp_set_state(sk, TCP_CLOSE);
1887 tcp_send_active_reset(sk, sk->sk_allocation);
1888 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1889 /* Check zero linger _after_ checking for unread data. */
1890 sk->sk_prot->disconnect(sk, 0);
1891 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1892 } else if (tcp_close_state(sk)) {
1893 /* We FIN if the application ate all the data before
1894 * zapping the connection.
1895 */
1896
1897 /* RED-PEN. Formally speaking, we have broken TCP state
1898 * machine. State transitions:
1899 *
1900 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1901 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1902 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1903 *
1904 * are legal only when FIN has been sent (i.e. in window),
1905 * rather than queued out of window. Purists blame.
1906 *
1907 * F.e. "RFC state" is ESTABLISHED,
1908 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1909 *
1910 * The visible declinations are that sometimes
1911 * we enter time-wait state, when it is not required really
1912 * (harmless), do not send active resets, when they are
1913 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1914 * they look as CLOSING or LAST_ACK for Linux)
1915 * Probably, I missed some more holelets.
1916 * --ANK
1917 */
1918 tcp_send_fin(sk);
1919 }
1920
1921 sk_stream_wait_close(sk, timeout);
1922
1923 adjudge_to_death:
1924 state = sk->sk_state;
1925 sock_hold(sk);
1926 sock_orphan(sk);
1927
1928 /* It is the last release_sock in its life. It will remove backlog. */
1929 release_sock(sk);
1930
1931
1932 /* Now socket is owned by kernel and we acquire BH lock
1933 to finish close. No need to check for user refs.
1934 */
1935 local_bh_disable();
1936 bh_lock_sock(sk);
1937 WARN_ON(sock_owned_by_user(sk));
1938
1939 percpu_counter_inc(sk->sk_prot->orphan_count);
1940
1941 /* Have we already been destroyed by a softirq or backlog? */
1942 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1943 goto out;
1944
1945 /* This is a (useful) BSD violating of the RFC. There is a
1946 * problem with TCP as specified in that the other end could
1947 * keep a socket open forever with no application left this end.
1948 * We use a 3 minute timeout (about the same as BSD) then kill
1949 * our end. If they send after that then tough - BUT: long enough
1950 * that we won't make the old 4*rto = almost no time - whoops
1951 * reset mistake.
1952 *
1953 * Nope, it was not mistake. It is really desired behaviour
1954 * f.e. on http servers, when such sockets are useless, but
1955 * consume significant resources. Let's do it with special
1956 * linger2 option. --ANK
1957 */
1958
1959 if (sk->sk_state == TCP_FIN_WAIT2) {
1960 struct tcp_sock *tp = tcp_sk(sk);
1961 if (tp->linger2 < 0) {
1962 tcp_set_state(sk, TCP_CLOSE);
1963 tcp_send_active_reset(sk, GFP_ATOMIC);
1964 NET_INC_STATS_BH(sock_net(sk),
1965 LINUX_MIB_TCPABORTONLINGER);
1966 } else {
1967 const int tmo = tcp_fin_time(sk);
1968
1969 if (tmo > TCP_TIMEWAIT_LEN) {
1970 inet_csk_reset_keepalive_timer(sk,
1971 tmo - TCP_TIMEWAIT_LEN);
1972 } else {
1973 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1974 goto out;
1975 }
1976 }
1977 }
1978 if (sk->sk_state != TCP_CLOSE) {
1979 int orphan_count = percpu_counter_read_positive(
1980 sk->sk_prot->orphan_count);
1981
1982 sk_mem_reclaim(sk);
1983 if (tcp_too_many_orphans(sk, orphan_count)) {
1984 if (net_ratelimit())
1985 printk(KERN_INFO "TCP: too many of orphaned "
1986 "sockets\n");
1987 tcp_set_state(sk, TCP_CLOSE);
1988 tcp_send_active_reset(sk, GFP_ATOMIC);
1989 NET_INC_STATS_BH(sock_net(sk),
1990 LINUX_MIB_TCPABORTONMEMORY);
1991 }
1992 }
1993
1994 if (sk->sk_state == TCP_CLOSE)
1995 inet_csk_destroy_sock(sk);
1996 /* Otherwise, socket is reprieved until protocol close. */
1997
1998 out:
1999 bh_unlock_sock(sk);
2000 local_bh_enable();
2001 sock_put(sk);
2002 }
2003
2004 /* These states need RST on ABORT according to RFC793 */
2005
2006 static inline int tcp_need_reset(int state)
2007 {
2008 return (1 << state) &
2009 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2010 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2011 }
2012
2013 int tcp_disconnect(struct sock *sk, int flags)
2014 {
2015 struct inet_sock *inet = inet_sk(sk);
2016 struct inet_connection_sock *icsk = inet_csk(sk);
2017 struct tcp_sock *tp = tcp_sk(sk);
2018 int err = 0;
2019 int old_state = sk->sk_state;
2020
2021 if (old_state != TCP_CLOSE)
2022 tcp_set_state(sk, TCP_CLOSE);
2023
2024 /* ABORT function of RFC793 */
2025 if (old_state == TCP_LISTEN) {
2026 inet_csk_listen_stop(sk);
2027 } else if (tcp_need_reset(old_state) ||
2028 (tp->snd_nxt != tp->write_seq &&
2029 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2030 /* The last check adjusts for discrepancy of Linux wrt. RFC
2031 * states
2032 */
2033 tcp_send_active_reset(sk, gfp_any());
2034 sk->sk_err = ECONNRESET;
2035 } else if (old_state == TCP_SYN_SENT)
2036 sk->sk_err = ECONNRESET;
2037
2038 tcp_clear_xmit_timers(sk);
2039 __skb_queue_purge(&sk->sk_receive_queue);
2040 tcp_write_queue_purge(sk);
2041 __skb_queue_purge(&tp->out_of_order_queue);
2042 #ifdef CONFIG_NET_DMA
2043 __skb_queue_purge(&sk->sk_async_wait_queue);
2044 #endif
2045
2046 inet->inet_dport = 0;
2047
2048 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2049 inet_reset_saddr(sk);
2050
2051 sk->sk_shutdown = 0;
2052 sock_reset_flag(sk, SOCK_DONE);
2053 tp->srtt = 0;
2054 if ((tp->write_seq += tp->max_window + 2) == 0)
2055 tp->write_seq = 1;
2056 icsk->icsk_backoff = 0;
2057 tp->snd_cwnd = 2;
2058 icsk->icsk_probes_out = 0;
2059 tp->packets_out = 0;
2060 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2061 tp->snd_cwnd_cnt = 0;
2062 tp->bytes_acked = 0;
2063 tp->window_clamp = 0;
2064 tcp_set_ca_state(sk, TCP_CA_Open);
2065 tcp_clear_retrans(tp);
2066 inet_csk_delack_init(sk);
2067 tcp_init_send_head(sk);
2068 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2069 __sk_dst_reset(sk);
2070
2071 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2072
2073 sk->sk_error_report(sk);
2074 return err;
2075 }
2076
2077 /*
2078 * Socket option code for TCP.
2079 */
2080 static int do_tcp_setsockopt(struct sock *sk, int level,
2081 int optname, char __user *optval, unsigned int optlen)
2082 {
2083 struct tcp_sock *tp = tcp_sk(sk);
2084 struct inet_connection_sock *icsk = inet_csk(sk);
2085 int val;
2086 int err = 0;
2087
2088 /* These are data/string values, all the others are ints */
2089 switch (optname) {
2090 case TCP_CONGESTION: {
2091 char name[TCP_CA_NAME_MAX];
2092
2093 if (optlen < 1)
2094 return -EINVAL;
2095
2096 val = strncpy_from_user(name, optval,
2097 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2098 if (val < 0)
2099 return -EFAULT;
2100 name[val] = 0;
2101
2102 lock_sock(sk);
2103 err = tcp_set_congestion_control(sk, name);
2104 release_sock(sk);
2105 return err;
2106 }
2107 case TCP_COOKIE_TRANSACTIONS: {
2108 struct tcp_cookie_transactions ctd;
2109 struct tcp_cookie_values *cvp = NULL;
2110
2111 if (sizeof(ctd) > optlen)
2112 return -EINVAL;
2113 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2114 return -EFAULT;
2115
2116 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2117 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2118 return -EINVAL;
2119
2120 if (ctd.tcpct_cookie_desired == 0) {
2121 /* default to global value */
2122 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2123 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2124 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2125 return -EINVAL;
2126 }
2127
2128 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2129 /* Supercedes all other values */
2130 lock_sock(sk);
2131 if (tp->cookie_values != NULL) {
2132 kref_put(&tp->cookie_values->kref,
2133 tcp_cookie_values_release);
2134 tp->cookie_values = NULL;
2135 }
2136 tp->rx_opt.cookie_in_always = 0; /* false */
2137 tp->rx_opt.cookie_out_never = 1; /* true */
2138 release_sock(sk);
2139 return err;
2140 }
2141
2142 /* Allocate ancillary memory before locking.
2143 */
2144 if (ctd.tcpct_used > 0 ||
2145 (tp->cookie_values == NULL &&
2146 (sysctl_tcp_cookie_size > 0 ||
2147 ctd.tcpct_cookie_desired > 0 ||
2148 ctd.tcpct_s_data_desired > 0))) {
2149 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2150 GFP_KERNEL);
2151 if (cvp == NULL)
2152 return -ENOMEM;
2153 }
2154 lock_sock(sk);
2155 tp->rx_opt.cookie_in_always =
2156 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2157 tp->rx_opt.cookie_out_never = 0; /* false */
2158
2159 if (tp->cookie_values != NULL) {
2160 if (cvp != NULL) {
2161 /* Changed values are recorded by a changed
2162 * pointer, ensuring the cookie will differ,
2163 * without separately hashing each value later.
2164 */
2165 kref_put(&tp->cookie_values->kref,
2166 tcp_cookie_values_release);
2167 kref_init(&cvp->kref);
2168 tp->cookie_values = cvp;
2169 } else {
2170 cvp = tp->cookie_values;
2171 }
2172 }
2173 if (cvp != NULL) {
2174 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2175
2176 if (ctd.tcpct_used > 0) {
2177 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2178 ctd.tcpct_used);
2179 cvp->s_data_desired = ctd.tcpct_used;
2180 cvp->s_data_constant = 1; /* true */
2181 } else {
2182 /* No constant payload data. */
2183 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2184 cvp->s_data_constant = 0; /* false */
2185 }
2186 }
2187 release_sock(sk);
2188 return err;
2189 }
2190 default:
2191 /* fallthru */
2192 break;
2193 };
2194
2195 if (optlen < sizeof(int))
2196 return -EINVAL;
2197
2198 if (get_user(val, (int __user *)optval))
2199 return -EFAULT;
2200
2201 lock_sock(sk);
2202
2203 switch (optname) {
2204 case TCP_MAXSEG:
2205 /* Values greater than interface MTU won't take effect. However
2206 * at the point when this call is done we typically don't yet
2207 * know which interface is going to be used */
2208 if (val < 8 || val > MAX_TCP_WINDOW) {
2209 err = -EINVAL;
2210 break;
2211 }
2212 tp->rx_opt.user_mss = val;
2213 break;
2214
2215 case TCP_NODELAY:
2216 if (val) {
2217 /* TCP_NODELAY is weaker than TCP_CORK, so that
2218 * this option on corked socket is remembered, but
2219 * it is not activated until cork is cleared.
2220 *
2221 * However, when TCP_NODELAY is set we make
2222 * an explicit push, which overrides even TCP_CORK
2223 * for currently queued segments.
2224 */
2225 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2226 tcp_push_pending_frames(sk);
2227 } else {
2228 tp->nonagle &= ~TCP_NAGLE_OFF;
2229 }
2230 break;
2231
2232 case TCP_CORK:
2233 /* When set indicates to always queue non-full frames.
2234 * Later the user clears this option and we transmit
2235 * any pending partial frames in the queue. This is
2236 * meant to be used alongside sendfile() to get properly
2237 * filled frames when the user (for example) must write
2238 * out headers with a write() call first and then use
2239 * sendfile to send out the data parts.
2240 *
2241 * TCP_CORK can be set together with TCP_NODELAY and it is
2242 * stronger than TCP_NODELAY.
2243 */
2244 if (val) {
2245 tp->nonagle |= TCP_NAGLE_CORK;
2246 } else {
2247 tp->nonagle &= ~TCP_NAGLE_CORK;
2248 if (tp->nonagle&TCP_NAGLE_OFF)
2249 tp->nonagle |= TCP_NAGLE_PUSH;
2250 tcp_push_pending_frames(sk);
2251 }
2252 break;
2253
2254 case TCP_KEEPIDLE:
2255 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2256 err = -EINVAL;
2257 else {
2258 tp->keepalive_time = val * HZ;
2259 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2260 !((1 << sk->sk_state) &
2261 (TCPF_CLOSE | TCPF_LISTEN))) {
2262 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
2263 if (tp->keepalive_time > elapsed)
2264 elapsed = tp->keepalive_time - elapsed;
2265 else
2266 elapsed = 0;
2267 inet_csk_reset_keepalive_timer(sk, elapsed);
2268 }
2269 }
2270 break;
2271 case TCP_KEEPINTVL:
2272 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2273 err = -EINVAL;
2274 else
2275 tp->keepalive_intvl = val * HZ;
2276 break;
2277 case TCP_KEEPCNT:
2278 if (val < 1 || val > MAX_TCP_KEEPCNT)
2279 err = -EINVAL;
2280 else
2281 tp->keepalive_probes = val;
2282 break;
2283 case TCP_SYNCNT:
2284 if (val < 1 || val > MAX_TCP_SYNCNT)
2285 err = -EINVAL;
2286 else
2287 icsk->icsk_syn_retries = val;
2288 break;
2289
2290 case TCP_LINGER2:
2291 if (val < 0)
2292 tp->linger2 = -1;
2293 else if (val > sysctl_tcp_fin_timeout / HZ)
2294 tp->linger2 = 0;
2295 else
2296 tp->linger2 = val * HZ;
2297 break;
2298
2299 case TCP_DEFER_ACCEPT:
2300 /* Translate value in seconds to number of retransmits */
2301 icsk->icsk_accept_queue.rskq_defer_accept =
2302 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2303 TCP_RTO_MAX / HZ);
2304 break;
2305
2306 case TCP_WINDOW_CLAMP:
2307 if (!val) {
2308 if (sk->sk_state != TCP_CLOSE) {
2309 err = -EINVAL;
2310 break;
2311 }
2312 tp->window_clamp = 0;
2313 } else
2314 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2315 SOCK_MIN_RCVBUF / 2 : val;
2316 break;
2317
2318 case TCP_QUICKACK:
2319 if (!val) {
2320 icsk->icsk_ack.pingpong = 1;
2321 } else {
2322 icsk->icsk_ack.pingpong = 0;
2323 if ((1 << sk->sk_state) &
2324 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2325 inet_csk_ack_scheduled(sk)) {
2326 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2327 tcp_cleanup_rbuf(sk, 1);
2328 if (!(val & 1))
2329 icsk->icsk_ack.pingpong = 1;
2330 }
2331 }
2332 break;
2333
2334 #ifdef CONFIG_TCP_MD5SIG
2335 case TCP_MD5SIG:
2336 /* Read the IP->Key mappings from userspace */
2337 err = tp->af_specific->md5_parse(sk, optval, optlen);
2338 break;
2339 #endif
2340
2341 default:
2342 err = -ENOPROTOOPT;
2343 break;
2344 }
2345
2346 release_sock(sk);
2347 return err;
2348 }
2349
2350 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2351 unsigned int optlen)
2352 {
2353 struct inet_connection_sock *icsk = inet_csk(sk);
2354
2355 if (level != SOL_TCP)
2356 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2357 optval, optlen);
2358 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2359 }
2360
2361 #ifdef CONFIG_COMPAT
2362 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2363 char __user *optval, unsigned int optlen)
2364 {
2365 if (level != SOL_TCP)
2366 return inet_csk_compat_setsockopt(sk, level, optname,
2367 optval, optlen);
2368 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2369 }
2370
2371 EXPORT_SYMBOL(compat_tcp_setsockopt);
2372 #endif
2373
2374 /* Return information about state of tcp endpoint in API format. */
2375 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2376 {
2377 struct tcp_sock *tp = tcp_sk(sk);
2378 const struct inet_connection_sock *icsk = inet_csk(sk);
2379 u32 now = tcp_time_stamp;
2380
2381 memset(info, 0, sizeof(*info));
2382
2383 info->tcpi_state = sk->sk_state;
2384 info->tcpi_ca_state = icsk->icsk_ca_state;
2385 info->tcpi_retransmits = icsk->icsk_retransmits;
2386 info->tcpi_probes = icsk->icsk_probes_out;
2387 info->tcpi_backoff = icsk->icsk_backoff;
2388
2389 if (tp->rx_opt.tstamp_ok)
2390 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2391 if (tcp_is_sack(tp))
2392 info->tcpi_options |= TCPI_OPT_SACK;
2393 if (tp->rx_opt.wscale_ok) {
2394 info->tcpi_options |= TCPI_OPT_WSCALE;
2395 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2396 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2397 }
2398
2399 if (tp->ecn_flags&TCP_ECN_OK)
2400 info->tcpi_options |= TCPI_OPT_ECN;
2401
2402 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2403 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2404 info->tcpi_snd_mss = tp->mss_cache;
2405 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2406
2407 if (sk->sk_state == TCP_LISTEN) {
2408 info->tcpi_unacked = sk->sk_ack_backlog;
2409 info->tcpi_sacked = sk->sk_max_ack_backlog;
2410 } else {
2411 info->tcpi_unacked = tp->packets_out;
2412 info->tcpi_sacked = tp->sacked_out;
2413 }
2414 info->tcpi_lost = tp->lost_out;
2415 info->tcpi_retrans = tp->retrans_out;
2416 info->tcpi_fackets = tp->fackets_out;
2417
2418 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2419 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2420 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2421
2422 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2423 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2424 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2425 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2426 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2427 info->tcpi_snd_cwnd = tp->snd_cwnd;
2428 info->tcpi_advmss = tp->advmss;
2429 info->tcpi_reordering = tp->reordering;
2430
2431 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2432 info->tcpi_rcv_space = tp->rcvq_space.space;
2433
2434 info->tcpi_total_retrans = tp->total_retrans;
2435 }
2436
2437 EXPORT_SYMBOL_GPL(tcp_get_info);
2438
2439 static int do_tcp_getsockopt(struct sock *sk, int level,
2440 int optname, char __user *optval, int __user *optlen)
2441 {
2442 struct inet_connection_sock *icsk = inet_csk(sk);
2443 struct tcp_sock *tp = tcp_sk(sk);
2444 int val, len;
2445
2446 if (get_user(len, optlen))
2447 return -EFAULT;
2448
2449 len = min_t(unsigned int, len, sizeof(int));
2450
2451 if (len < 0)
2452 return -EINVAL;
2453
2454 switch (optname) {
2455 case TCP_MAXSEG:
2456 val = tp->mss_cache;
2457 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2458 val = tp->rx_opt.user_mss;
2459 break;
2460 case TCP_NODELAY:
2461 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2462 break;
2463 case TCP_CORK:
2464 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2465 break;
2466 case TCP_KEEPIDLE:
2467 val = keepalive_time_when(tp) / HZ;
2468 break;
2469 case TCP_KEEPINTVL:
2470 val = keepalive_intvl_when(tp) / HZ;
2471 break;
2472 case TCP_KEEPCNT:
2473 val = keepalive_probes(tp);
2474 break;
2475 case TCP_SYNCNT:
2476 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2477 break;
2478 case TCP_LINGER2:
2479 val = tp->linger2;
2480 if (val >= 0)
2481 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2482 break;
2483 case TCP_DEFER_ACCEPT:
2484 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2485 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2486 break;
2487 case TCP_WINDOW_CLAMP:
2488 val = tp->window_clamp;
2489 break;
2490 case TCP_INFO: {
2491 struct tcp_info info;
2492
2493 if (get_user(len, optlen))
2494 return -EFAULT;
2495
2496 tcp_get_info(sk, &info);
2497
2498 len = min_t(unsigned int, len, sizeof(info));
2499 if (put_user(len, optlen))
2500 return -EFAULT;
2501 if (copy_to_user(optval, &info, len))
2502 return -EFAULT;
2503 return 0;
2504 }
2505 case TCP_QUICKACK:
2506 val = !icsk->icsk_ack.pingpong;
2507 break;
2508
2509 case TCP_CONGESTION:
2510 if (get_user(len, optlen))
2511 return -EFAULT;
2512 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2513 if (put_user(len, optlen))
2514 return -EFAULT;
2515 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2516 return -EFAULT;
2517 return 0;
2518
2519 case TCP_COOKIE_TRANSACTIONS: {
2520 struct tcp_cookie_transactions ctd;
2521 struct tcp_cookie_values *cvp = tp->cookie_values;
2522
2523 if (get_user(len, optlen))
2524 return -EFAULT;
2525 if (len < sizeof(ctd))
2526 return -EINVAL;
2527
2528 memset(&ctd, 0, sizeof(ctd));
2529 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2530 TCP_COOKIE_IN_ALWAYS : 0)
2531 | (tp->rx_opt.cookie_out_never ?
2532 TCP_COOKIE_OUT_NEVER : 0);
2533
2534 if (cvp != NULL) {
2535 ctd.tcpct_flags |= (cvp->s_data_in ?
2536 TCP_S_DATA_IN : 0)
2537 | (cvp->s_data_out ?
2538 TCP_S_DATA_OUT : 0);
2539
2540 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2541 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2542
2543 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2544 cvp->cookie_pair_size);
2545 ctd.tcpct_used = cvp->cookie_pair_size;
2546 }
2547
2548 if (put_user(sizeof(ctd), optlen))
2549 return -EFAULT;
2550 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2551 return -EFAULT;
2552 return 0;
2553 }
2554 default:
2555 return -ENOPROTOOPT;
2556 }
2557
2558 if (put_user(len, optlen))
2559 return -EFAULT;
2560 if (copy_to_user(optval, &val, len))
2561 return -EFAULT;
2562 return 0;
2563 }
2564
2565 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2566 int __user *optlen)
2567 {
2568 struct inet_connection_sock *icsk = inet_csk(sk);
2569
2570 if (level != SOL_TCP)
2571 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2572 optval, optlen);
2573 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2574 }
2575
2576 #ifdef CONFIG_COMPAT
2577 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2578 char __user *optval, int __user *optlen)
2579 {
2580 if (level != SOL_TCP)
2581 return inet_csk_compat_getsockopt(sk, level, optname,
2582 optval, optlen);
2583 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2584 }
2585
2586 EXPORT_SYMBOL(compat_tcp_getsockopt);
2587 #endif
2588
2589 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2590 {
2591 struct sk_buff *segs = ERR_PTR(-EINVAL);
2592 struct tcphdr *th;
2593 unsigned thlen;
2594 unsigned int seq;
2595 __be32 delta;
2596 unsigned int oldlen;
2597 unsigned int mss;
2598
2599 if (!pskb_may_pull(skb, sizeof(*th)))
2600 goto out;
2601
2602 th = tcp_hdr(skb);
2603 thlen = th->doff * 4;
2604 if (thlen < sizeof(*th))
2605 goto out;
2606
2607 if (!pskb_may_pull(skb, thlen))
2608 goto out;
2609
2610 oldlen = (u16)~skb->len;
2611 __skb_pull(skb, thlen);
2612
2613 mss = skb_shinfo(skb)->gso_size;
2614 if (unlikely(skb->len <= mss))
2615 goto out;
2616
2617 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2618 /* Packet is from an untrusted source, reset gso_segs. */
2619 int type = skb_shinfo(skb)->gso_type;
2620
2621 if (unlikely(type &
2622 ~(SKB_GSO_TCPV4 |
2623 SKB_GSO_DODGY |
2624 SKB_GSO_TCP_ECN |
2625 SKB_GSO_TCPV6 |
2626 0) ||
2627 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2628 goto out;
2629
2630 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2631
2632 segs = NULL;
2633 goto out;
2634 }
2635
2636 segs = skb_segment(skb, features);
2637 if (IS_ERR(segs))
2638 goto out;
2639
2640 delta = htonl(oldlen + (thlen + mss));
2641
2642 skb = segs;
2643 th = tcp_hdr(skb);
2644 seq = ntohl(th->seq);
2645
2646 do {
2647 th->fin = th->psh = 0;
2648
2649 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2650 (__force u32)delta));
2651 if (skb->ip_summed != CHECKSUM_PARTIAL)
2652 th->check =
2653 csum_fold(csum_partial(skb_transport_header(skb),
2654 thlen, skb->csum));
2655
2656 seq += mss;
2657 skb = skb->next;
2658 th = tcp_hdr(skb);
2659
2660 th->seq = htonl(seq);
2661 th->cwr = 0;
2662 } while (skb->next);
2663
2664 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2665 skb->data_len);
2666 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2667 (__force u32)delta));
2668 if (skb->ip_summed != CHECKSUM_PARTIAL)
2669 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2670 thlen, skb->csum));
2671
2672 out:
2673 return segs;
2674 }
2675 EXPORT_SYMBOL(tcp_tso_segment);
2676
2677 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2678 {
2679 struct sk_buff **pp = NULL;
2680 struct sk_buff *p;
2681 struct tcphdr *th;
2682 struct tcphdr *th2;
2683 unsigned int len;
2684 unsigned int thlen;
2685 unsigned int flags;
2686 unsigned int mss = 1;
2687 unsigned int hlen;
2688 unsigned int off;
2689 int flush = 1;
2690 int i;
2691
2692 off = skb_gro_offset(skb);
2693 hlen = off + sizeof(*th);
2694 th = skb_gro_header_fast(skb, off);
2695 if (skb_gro_header_hard(skb, hlen)) {
2696 th = skb_gro_header_slow(skb, hlen, off);
2697 if (unlikely(!th))
2698 goto out;
2699 }
2700
2701 thlen = th->doff * 4;
2702 if (thlen < sizeof(*th))
2703 goto out;
2704
2705 hlen = off + thlen;
2706 if (skb_gro_header_hard(skb, hlen)) {
2707 th = skb_gro_header_slow(skb, hlen, off);
2708 if (unlikely(!th))
2709 goto out;
2710 }
2711
2712 skb_gro_pull(skb, thlen);
2713
2714 len = skb_gro_len(skb);
2715 flags = tcp_flag_word(th);
2716
2717 for (; (p = *head); head = &p->next) {
2718 if (!NAPI_GRO_CB(p)->same_flow)
2719 continue;
2720
2721 th2 = tcp_hdr(p);
2722
2723 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2724 NAPI_GRO_CB(p)->same_flow = 0;
2725 continue;
2726 }
2727
2728 goto found;
2729 }
2730
2731 goto out_check_final;
2732
2733 found:
2734 flush = NAPI_GRO_CB(p)->flush;
2735 flush |= flags & TCP_FLAG_CWR;
2736 flush |= (flags ^ tcp_flag_word(th2)) &
2737 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH);
2738 flush |= th->ack_seq ^ th2->ack_seq;
2739 for (i = sizeof(*th); i < thlen; i += 4)
2740 flush |= *(u32 *)((u8 *)th + i) ^
2741 *(u32 *)((u8 *)th2 + i);
2742
2743 mss = skb_shinfo(p)->gso_size;
2744
2745 flush |= (len - 1) >= mss;
2746 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2747
2748 if (flush || skb_gro_receive(head, skb)) {
2749 mss = 1;
2750 goto out_check_final;
2751 }
2752
2753 p = *head;
2754 th2 = tcp_hdr(p);
2755 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2756
2757 out_check_final:
2758 flush = len < mss;
2759 flush |= flags & (TCP_FLAG_URG | TCP_FLAG_PSH | TCP_FLAG_RST |
2760 TCP_FLAG_SYN | TCP_FLAG_FIN);
2761
2762 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2763 pp = head;
2764
2765 out:
2766 NAPI_GRO_CB(skb)->flush |= flush;
2767
2768 return pp;
2769 }
2770 EXPORT_SYMBOL(tcp_gro_receive);
2771
2772 int tcp_gro_complete(struct sk_buff *skb)
2773 {
2774 struct tcphdr *th = tcp_hdr(skb);
2775
2776 skb->csum_start = skb_transport_header(skb) - skb->head;
2777 skb->csum_offset = offsetof(struct tcphdr, check);
2778 skb->ip_summed = CHECKSUM_PARTIAL;
2779
2780 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2781
2782 if (th->cwr)
2783 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2784
2785 return 0;
2786 }
2787 EXPORT_SYMBOL(tcp_gro_complete);
2788
2789 #ifdef CONFIG_TCP_MD5SIG
2790 static unsigned long tcp_md5sig_users;
2791 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2792 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2793
2794 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2795 {
2796 int cpu;
2797 for_each_possible_cpu(cpu) {
2798 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2799 if (p) {
2800 if (p->md5_desc.tfm)
2801 crypto_free_hash(p->md5_desc.tfm);
2802 kfree(p);
2803 p = NULL;
2804 }
2805 }
2806 free_percpu(pool);
2807 }
2808
2809 void tcp_free_md5sig_pool(void)
2810 {
2811 struct tcp_md5sig_pool **pool = NULL;
2812
2813 spin_lock_bh(&tcp_md5sig_pool_lock);
2814 if (--tcp_md5sig_users == 0) {
2815 pool = tcp_md5sig_pool;
2816 tcp_md5sig_pool = NULL;
2817 }
2818 spin_unlock_bh(&tcp_md5sig_pool_lock);
2819 if (pool)
2820 __tcp_free_md5sig_pool(pool);
2821 }
2822
2823 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2824
2825 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(struct sock *sk)
2826 {
2827 int cpu;
2828 struct tcp_md5sig_pool **pool;
2829
2830 pool = alloc_percpu(struct tcp_md5sig_pool *);
2831 if (!pool)
2832 return NULL;
2833
2834 for_each_possible_cpu(cpu) {
2835 struct tcp_md5sig_pool *p;
2836 struct crypto_hash *hash;
2837
2838 p = kzalloc(sizeof(*p), sk->sk_allocation);
2839 if (!p)
2840 goto out_free;
2841 *per_cpu_ptr(pool, cpu) = p;
2842
2843 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2844 if (!hash || IS_ERR(hash))
2845 goto out_free;
2846
2847 p->md5_desc.tfm = hash;
2848 }
2849 return pool;
2850 out_free:
2851 __tcp_free_md5sig_pool(pool);
2852 return NULL;
2853 }
2854
2855 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(struct sock *sk)
2856 {
2857 struct tcp_md5sig_pool **pool;
2858 int alloc = 0;
2859
2860 retry:
2861 spin_lock_bh(&tcp_md5sig_pool_lock);
2862 pool = tcp_md5sig_pool;
2863 if (tcp_md5sig_users++ == 0) {
2864 alloc = 1;
2865 spin_unlock_bh(&tcp_md5sig_pool_lock);
2866 } else if (!pool) {
2867 tcp_md5sig_users--;
2868 spin_unlock_bh(&tcp_md5sig_pool_lock);
2869 cpu_relax();
2870 goto retry;
2871 } else
2872 spin_unlock_bh(&tcp_md5sig_pool_lock);
2873
2874 if (alloc) {
2875 /* we cannot hold spinlock here because this may sleep. */
2876 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool(sk);
2877 spin_lock_bh(&tcp_md5sig_pool_lock);
2878 if (!p) {
2879 tcp_md5sig_users--;
2880 spin_unlock_bh(&tcp_md5sig_pool_lock);
2881 return NULL;
2882 }
2883 pool = tcp_md5sig_pool;
2884 if (pool) {
2885 /* oops, it has already been assigned. */
2886 spin_unlock_bh(&tcp_md5sig_pool_lock);
2887 __tcp_free_md5sig_pool(p);
2888 } else {
2889 tcp_md5sig_pool = pool = p;
2890 spin_unlock_bh(&tcp_md5sig_pool_lock);
2891 }
2892 }
2893 return pool;
2894 }
2895
2896 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2897
2898 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2899 {
2900 struct tcp_md5sig_pool **p;
2901 spin_lock_bh(&tcp_md5sig_pool_lock);
2902 p = tcp_md5sig_pool;
2903 if (p)
2904 tcp_md5sig_users++;
2905 spin_unlock_bh(&tcp_md5sig_pool_lock);
2906 return (p ? *per_cpu_ptr(p, cpu) : NULL);
2907 }
2908
2909 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2910
2911 void __tcp_put_md5sig_pool(void)
2912 {
2913 tcp_free_md5sig_pool();
2914 }
2915
2916 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2917
2918 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2919 struct tcphdr *th)
2920 {
2921 struct scatterlist sg;
2922 int err;
2923
2924 __sum16 old_checksum = th->check;
2925 th->check = 0;
2926 /* options aren't included in the hash */
2927 sg_init_one(&sg, th, sizeof(struct tcphdr));
2928 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2929 th->check = old_checksum;
2930 return err;
2931 }
2932
2933 EXPORT_SYMBOL(tcp_md5_hash_header);
2934
2935 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2936 struct sk_buff *skb, unsigned header_len)
2937 {
2938 struct scatterlist sg;
2939 const struct tcphdr *tp = tcp_hdr(skb);
2940 struct hash_desc *desc = &hp->md5_desc;
2941 unsigned i;
2942 const unsigned head_data_len = skb_headlen(skb) > header_len ?
2943 skb_headlen(skb) - header_len : 0;
2944 const struct skb_shared_info *shi = skb_shinfo(skb);
2945
2946 sg_init_table(&sg, 1);
2947
2948 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
2949 if (crypto_hash_update(desc, &sg, head_data_len))
2950 return 1;
2951
2952 for (i = 0; i < shi->nr_frags; ++i) {
2953 const struct skb_frag_struct *f = &shi->frags[i];
2954 sg_set_page(&sg, f->page, f->size, f->page_offset);
2955 if (crypto_hash_update(desc, &sg, f->size))
2956 return 1;
2957 }
2958
2959 return 0;
2960 }
2961
2962 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
2963
2964 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
2965 {
2966 struct scatterlist sg;
2967
2968 sg_init_one(&sg, key->key, key->keylen);
2969 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
2970 }
2971
2972 EXPORT_SYMBOL(tcp_md5_hash_key);
2973
2974 #endif
2975
2976 /**
2977 * Each Responder maintains up to two secret values concurrently for
2978 * efficient secret rollover. Each secret value has 4 states:
2979 *
2980 * Generating. (tcp_secret_generating != tcp_secret_primary)
2981 * Generates new Responder-Cookies, but not yet used for primary
2982 * verification. This is a short-term state, typically lasting only
2983 * one round trip time (RTT).
2984 *
2985 * Primary. (tcp_secret_generating == tcp_secret_primary)
2986 * Used both for generation and primary verification.
2987 *
2988 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
2989 * Used for verification, until the first failure that can be
2990 * verified by the newer Generating secret. At that time, this
2991 * cookie's state is changed to Secondary, and the Generating
2992 * cookie's state is changed to Primary. This is a short-term state,
2993 * typically lasting only one round trip time (RTT).
2994 *
2995 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
2996 * Used for secondary verification, after primary verification
2997 * failures. This state lasts no more than twice the Maximum Segment
2998 * Lifetime (2MSL). Then, the secret is discarded.
2999 */
3000 struct tcp_cookie_secret {
3001 /* The secret is divided into two parts. The digest part is the
3002 * equivalent of previously hashing a secret and saving the state,
3003 * and serves as an initialization vector (IV). The message part
3004 * serves as the trailing secret.
3005 */
3006 u32 secrets[COOKIE_WORKSPACE_WORDS];
3007 unsigned long expires;
3008 };
3009
3010 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3011 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3012 #define TCP_SECRET_LIFE (HZ * 600)
3013
3014 static struct tcp_cookie_secret tcp_secret_one;
3015 static struct tcp_cookie_secret tcp_secret_two;
3016
3017 /* Essentially a circular list, without dynamic allocation. */
3018 static struct tcp_cookie_secret *tcp_secret_generating;
3019 static struct tcp_cookie_secret *tcp_secret_primary;
3020 static struct tcp_cookie_secret *tcp_secret_retiring;
3021 static struct tcp_cookie_secret *tcp_secret_secondary;
3022
3023 static DEFINE_SPINLOCK(tcp_secret_locker);
3024
3025 /* Select a pseudo-random word in the cookie workspace.
3026 */
3027 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3028 {
3029 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3030 }
3031
3032 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3033 * Called in softirq context.
3034 * Returns: 0 for success.
3035 */
3036 int tcp_cookie_generator(u32 *bakery)
3037 {
3038 unsigned long jiffy = jiffies;
3039
3040 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3041 spin_lock_bh(&tcp_secret_locker);
3042 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3043 /* refreshed by another */
3044 memcpy(bakery,
3045 &tcp_secret_generating->secrets[0],
3046 COOKIE_WORKSPACE_WORDS);
3047 } else {
3048 /* still needs refreshing */
3049 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3050
3051 /* The first time, paranoia assumes that the
3052 * randomization function isn't as strong. But,
3053 * this secret initialization is delayed until
3054 * the last possible moment (packet arrival).
3055 * Although that time is observable, it is
3056 * unpredictably variable. Mash in the most
3057 * volatile clock bits available, and expire the
3058 * secret extra quickly.
3059 */
3060 if (unlikely(tcp_secret_primary->expires ==
3061 tcp_secret_secondary->expires)) {
3062 struct timespec tv;
3063
3064 getnstimeofday(&tv);
3065 bakery[COOKIE_DIGEST_WORDS+0] ^=
3066 (u32)tv.tv_nsec;
3067
3068 tcp_secret_secondary->expires = jiffy
3069 + TCP_SECRET_1MSL
3070 + (0x0f & tcp_cookie_work(bakery, 0));
3071 } else {
3072 tcp_secret_secondary->expires = jiffy
3073 + TCP_SECRET_LIFE
3074 + (0xff & tcp_cookie_work(bakery, 1));
3075 tcp_secret_primary->expires = jiffy
3076 + TCP_SECRET_2MSL
3077 + (0x1f & tcp_cookie_work(bakery, 2));
3078 }
3079 memcpy(&tcp_secret_secondary->secrets[0],
3080 bakery, COOKIE_WORKSPACE_WORDS);
3081
3082 rcu_assign_pointer(tcp_secret_generating,
3083 tcp_secret_secondary);
3084 rcu_assign_pointer(tcp_secret_retiring,
3085 tcp_secret_primary);
3086 /*
3087 * Neither call_rcu() nor synchronize_rcu() needed.
3088 * Retiring data is not freed. It is replaced after
3089 * further (locked) pointer updates, and a quiet time
3090 * (minimum 1MSL, maximum LIFE - 2MSL).
3091 */
3092 }
3093 spin_unlock_bh(&tcp_secret_locker);
3094 } else {
3095 rcu_read_lock_bh();
3096 memcpy(bakery,
3097 &rcu_dereference(tcp_secret_generating)->secrets[0],
3098 COOKIE_WORKSPACE_WORDS);
3099 rcu_read_unlock_bh();
3100 }
3101 return 0;
3102 }
3103 EXPORT_SYMBOL(tcp_cookie_generator);
3104
3105 void tcp_done(struct sock *sk)
3106 {
3107 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3108 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3109
3110 tcp_set_state(sk, TCP_CLOSE);
3111 tcp_clear_xmit_timers(sk);
3112
3113 sk->sk_shutdown = SHUTDOWN_MASK;
3114
3115 if (!sock_flag(sk, SOCK_DEAD))
3116 sk->sk_state_change(sk);
3117 else
3118 inet_csk_destroy_sock(sk);
3119 }
3120 EXPORT_SYMBOL_GPL(tcp_done);
3121
3122 extern struct tcp_congestion_ops tcp_reno;
3123
3124 static __initdata unsigned long thash_entries;
3125 static int __init set_thash_entries(char *str)
3126 {
3127 if (!str)
3128 return 0;
3129 thash_entries = simple_strtoul(str, &str, 0);
3130 return 1;
3131 }
3132 __setup("thash_entries=", set_thash_entries);
3133
3134 void __init tcp_init(void)
3135 {
3136 struct sk_buff *skb = NULL;
3137 unsigned long nr_pages, limit;
3138 int order, i, max_share;
3139 unsigned long jiffy = jiffies;
3140
3141 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3142
3143 percpu_counter_init(&tcp_sockets_allocated, 0);
3144 percpu_counter_init(&tcp_orphan_count, 0);
3145 tcp_hashinfo.bind_bucket_cachep =
3146 kmem_cache_create("tcp_bind_bucket",
3147 sizeof(struct inet_bind_bucket), 0,
3148 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3149
3150 /* Size and allocate the main established and bind bucket
3151 * hash tables.
3152 *
3153 * The methodology is similar to that of the buffer cache.
3154 */
3155 tcp_hashinfo.ehash =
3156 alloc_large_system_hash("TCP established",
3157 sizeof(struct inet_ehash_bucket),
3158 thash_entries,
3159 (totalram_pages >= 128 * 1024) ?
3160 13 : 15,
3161 0,
3162 NULL,
3163 &tcp_hashinfo.ehash_mask,
3164 thash_entries ? 0 : 512 * 1024);
3165 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3166 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3167 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3168 }
3169 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3170 panic("TCP: failed to alloc ehash_locks");
3171 tcp_hashinfo.bhash =
3172 alloc_large_system_hash("TCP bind",
3173 sizeof(struct inet_bind_hashbucket),
3174 tcp_hashinfo.ehash_mask + 1,
3175 (totalram_pages >= 128 * 1024) ?
3176 13 : 15,
3177 0,
3178 &tcp_hashinfo.bhash_size,
3179 NULL,
3180 64 * 1024);
3181 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3182 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3183 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3184 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3185 }
3186
3187 /* Try to be a bit smarter and adjust defaults depending
3188 * on available memory.
3189 */
3190 for (order = 0; ((1 << order) << PAGE_SHIFT) <
3191 (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
3192 order++)
3193 ;
3194 if (order >= 4) {
3195 tcp_death_row.sysctl_max_tw_buckets = 180000;
3196 sysctl_tcp_max_orphans = 4096 << (order - 4);
3197 sysctl_max_syn_backlog = 1024;
3198 } else if (order < 3) {
3199 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
3200 sysctl_tcp_max_orphans >>= (3 - order);
3201 sysctl_max_syn_backlog = 128;
3202 }
3203
3204 /* Set the pressure threshold to be a fraction of global memory that
3205 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
3206 * memory, with a floor of 128 pages.
3207 */
3208 nr_pages = totalram_pages - totalhigh_pages;
3209 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
3210 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
3211 limit = max(limit, 128UL);
3212 sysctl_tcp_mem[0] = limit / 4 * 3;
3213 sysctl_tcp_mem[1] = limit;
3214 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3215
3216 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3217 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3218 max_share = min(4UL*1024*1024, limit);
3219
3220 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3221 sysctl_tcp_wmem[1] = 16*1024;
3222 sysctl_tcp_wmem[2] = max(64*1024, max_share);
3223
3224 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3225 sysctl_tcp_rmem[1] = 87380;
3226 sysctl_tcp_rmem[2] = max(87380, max_share);
3227
3228 printk(KERN_INFO "TCP: Hash tables configured "
3229 "(established %u bind %u)\n",
3230 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3231
3232 tcp_register_congestion_control(&tcp_reno);
3233
3234 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3235 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3236 tcp_secret_one.expires = jiffy; /* past due */
3237 tcp_secret_two.expires = jiffy; /* past due */
3238 tcp_secret_generating = &tcp_secret_one;
3239 tcp_secret_primary = &tcp_secret_one;
3240 tcp_secret_retiring = &tcp_secret_two;
3241 tcp_secret_secondary = &tcp_secret_two;
3242 }
3243
3244 EXPORT_SYMBOL(tcp_close);
3245 EXPORT_SYMBOL(tcp_disconnect);
3246 EXPORT_SYMBOL(tcp_getsockopt);
3247 EXPORT_SYMBOL(tcp_ioctl);
3248 EXPORT_SYMBOL(tcp_poll);
3249 EXPORT_SYMBOL(tcp_read_sock);
3250 EXPORT_SYMBOL(tcp_recvmsg);
3251 EXPORT_SYMBOL(tcp_sendmsg);
3252 EXPORT_SYMBOL(tcp_splice_read);
3253 EXPORT_SYMBOL(tcp_sendpage);
3254 EXPORT_SYMBOL(tcp_setsockopt);
3255 EXPORT_SYMBOL(tcp_shutdown);
Linus' kernel tree