?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
powerpc: Fix broken cpu_idle_wait() implementation
[linux/fpc-iii.git] / net / ipv4 / tcp.c
blobbf29e7c442a35a5a64ac36e24e46f7d8f12f8642
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 #include <linux/slab.h>
269
270 #include <net/icmp.h>
271 #include <net/tcp.h>
272 #include <net/xfrm.h>
273 #include <net/ip.h>
274 #include <net/netdma.h>
275 #include <net/sock.h>
276
277 #include <asm/uaccess.h>
278 #include <asm/ioctls.h>
279
280 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
281
282 struct percpu_counter tcp_orphan_count;
283 EXPORT_SYMBOL_GPL(tcp_orphan_count);
284
285 int sysctl_tcp_wmem[3] __read_mostly;
286 int sysctl_tcp_rmem[3] __read_mostly;
287
288 EXPORT_SYMBOL(sysctl_tcp_rmem);
289 EXPORT_SYMBOL(sysctl_tcp_wmem);
290
291 atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
292 EXPORT_SYMBOL(tcp_memory_allocated);
293
294 /*
295 * Current number of TCP sockets.
296 */
297 struct percpu_counter tcp_sockets_allocated;
298 EXPORT_SYMBOL(tcp_sockets_allocated);
299
300 /*
301 * TCP splice context
302 */
303 struct tcp_splice_state {
304 struct pipe_inode_info *pipe;
305 size_t len;
306 unsigned int flags;
307 };
308
309 /*
310 * Pressure flag: try to collapse.
311 * Technical note: it is used by multiple contexts non atomically.
312 * All the __sk_mem_schedule() is of this nature: accounting
313 * is strict, actions are advisory and have some latency.
314 */
315 int tcp_memory_pressure __read_mostly;
316 EXPORT_SYMBOL(tcp_memory_pressure);
317
318 void tcp_enter_memory_pressure(struct sock *sk)
319 {
320 if (!tcp_memory_pressure) {
321 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
322 tcp_memory_pressure = 1;
323 }
324 }
325 EXPORT_SYMBOL(tcp_enter_memory_pressure);
326
327 /* Convert seconds to retransmits based on initial and max timeout */
328 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
329 {
330 u8 res = 0;
331
332 if (seconds > 0) {
333 int period = timeout;
334
335 res = 1;
336 while (seconds > period && res < 255) {
337 res++;
338 timeout <<= 1;
339 if (timeout > rto_max)
340 timeout = rto_max;
341 period += timeout;
342 }
343 }
344 return res;
345 }
346
347 /* Convert retransmits to seconds based on initial and max timeout */
348 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
349 {
350 int period = 0;
351
352 if (retrans > 0) {
353 period = timeout;
354 while (--retrans) {
355 timeout <<= 1;
356 if (timeout > rto_max)
357 timeout = rto_max;
358 period += timeout;
359 }
360 }
361 return period;
362 }
363
364 /*
365 * Wait for a TCP event.
366 *
367 * Note that we don't need to lock the socket, as the upper poll layers
368 * take care of normal races (between the test and the event) and we don't
369 * go look at any of the socket buffers directly.
370 */
371 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
372 {
373 unsigned int mask;
374 struct sock *sk = sock->sk;
375 const struct tcp_sock *tp = tcp_sk(sk);
376
377 sock_poll_wait(file, sk_sleep(sk), wait);
378 if (sk->sk_state == TCP_LISTEN)
379 return inet_csk_listen_poll(sk);
380
381 /* Socket is not locked. We are protected from async events
382 * by poll logic and correct handling of state changes
383 * made by other threads is impossible in any case.
384 */
385
386 mask = 0;
387
388 /*
389 * POLLHUP is certainly not done right. But poll() doesn't
390 * have a notion of HUP in just one direction, and for a
391 * socket the read side is more interesting.
392 *
393 * Some poll() documentation says that POLLHUP is incompatible
394 * with the POLLOUT/POLLWR flags, so somebody should check this
395 * all. But careful, it tends to be safer to return too many
396 * bits than too few, and you can easily break real applications
397 * if you don't tell them that something has hung up!
398 *
399 * Check-me.
400 *
401 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
402 * our fs/select.c). It means that after we received EOF,
403 * poll always returns immediately, making impossible poll() on write()
404 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
405 * if and only if shutdown has been made in both directions.
406 * Actually, it is interesting to look how Solaris and DUX
407 * solve this dilemma. I would prefer, if POLLHUP were maskable,
408 * then we could set it on SND_SHUTDOWN. BTW examples given
409 * in Stevens' books assume exactly this behaviour, it explains
410 * why POLLHUP is incompatible with POLLOUT. --ANK
411 *
412 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
413 * blocking on fresh not-connected or disconnected socket. --ANK
414 */
415 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
416 mask |= POLLHUP;
417 if (sk->sk_shutdown & RCV_SHUTDOWN)
418 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
419
420 /* Connected? */
421 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
422 int target = sock_rcvlowat(sk, 0, INT_MAX);
423
424 if (tp->urg_seq == tp->copied_seq &&
425 !sock_flag(sk, SOCK_URGINLINE) &&
426 tp->urg_data)
427 target++;
428
429 /* Potential race condition. If read of tp below will
430 * escape above sk->sk_state, we can be illegally awaken
431 * in SYN_* states. */
432 if (tp->rcv_nxt - tp->copied_seq >= target)
433 mask |= POLLIN | POLLRDNORM;
434
435 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
436 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
437 mask |= POLLOUT | POLLWRNORM;
438 } else { /* send SIGIO later */
439 set_bit(SOCK_ASYNC_NOSPACE,
440 &sk->sk_socket->flags);
441 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
442
443 /* Race breaker. If space is freed after
444 * wspace test but before the flags are set,
445 * IO signal will be lost.
446 */
447 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
448 mask |= POLLOUT | POLLWRNORM;
449 }
450 } else
451 mask |= POLLOUT | POLLWRNORM;
452
453 if (tp->urg_data & TCP_URG_VALID)
454 mask |= POLLPRI;
455 }
456 /* This barrier is coupled with smp_wmb() in tcp_reset() */
457 smp_rmb();
458 if (sk->sk_err)
459 mask |= POLLERR;
460
461 return mask;
462 }
463 EXPORT_SYMBOL(tcp_poll);
464
465 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
466 {
467 struct tcp_sock *tp = tcp_sk(sk);
468 int answ;
469
470 switch (cmd) {
471 case SIOCINQ:
472 if (sk->sk_state == TCP_LISTEN)
473 return -EINVAL;
474
475 lock_sock(sk);
476 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
477 answ = 0;
478 else if (sock_flag(sk, SOCK_URGINLINE) ||
479 !tp->urg_data ||
480 before(tp->urg_seq, tp->copied_seq) ||
481 !before(tp->urg_seq, tp->rcv_nxt)) {
482 struct sk_buff *skb;
483
484 answ = tp->rcv_nxt - tp->copied_seq;
485
486 /* Subtract 1, if FIN is in queue. */
487 skb = skb_peek_tail(&sk->sk_receive_queue);
488 if (answ && skb)
489 answ -= tcp_hdr(skb)->fin;
490 } else
491 answ = tp->urg_seq - tp->copied_seq;
492 release_sock(sk);
493 break;
494 case SIOCATMARK:
495 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
496 break;
497 case SIOCOUTQ:
498 if (sk->sk_state == TCP_LISTEN)
499 return -EINVAL;
500
501 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
502 answ = 0;
503 else
504 answ = tp->write_seq - tp->snd_una;
505 break;
506 case SIOCOUTQNSD:
507 if (sk->sk_state == TCP_LISTEN)
508 return -EINVAL;
509
510 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
511 answ = 0;
512 else
513 answ = tp->write_seq - tp->snd_nxt;
514 break;
515 default:
516 return -ENOIOCTLCMD;
517 }
518
519 return put_user(answ, (int __user *)arg);
520 }
521 EXPORT_SYMBOL(tcp_ioctl);
522
523 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
524 {
525 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
526 tp->pushed_seq = tp->write_seq;
527 }
528
529 static inline int forced_push(const struct tcp_sock *tp)
530 {
531 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
532 }
533
534 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
535 {
536 struct tcp_sock *tp = tcp_sk(sk);
537 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
538
539 skb->csum = 0;
540 tcb->seq = tcb->end_seq = tp->write_seq;
541 tcb->tcp_flags = TCPHDR_ACK;
542 tcb->sacked = 0;
543 skb_header_release(skb);
544 tcp_add_write_queue_tail(sk, skb);
545 sk->sk_wmem_queued += skb->truesize;
546 sk_mem_charge(sk, skb->truesize);
547 if (tp->nonagle & TCP_NAGLE_PUSH)
548 tp->nonagle &= ~TCP_NAGLE_PUSH;
549 }
550
551 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
552 {
553 if (flags & MSG_OOB)
554 tp->snd_up = tp->write_seq;
555 }
556
557 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
558 int nonagle)
559 {
560 if (tcp_send_head(sk)) {
561 struct tcp_sock *tp = tcp_sk(sk);
562
563 if (!(flags & MSG_MORE) || forced_push(tp))
564 tcp_mark_push(tp, tcp_write_queue_tail(sk));
565
566 tcp_mark_urg(tp, flags);
567 __tcp_push_pending_frames(sk, mss_now,
568 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
569 }
570 }
571
572 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
573 unsigned int offset, size_t len)
574 {
575 struct tcp_splice_state *tss = rd_desc->arg.data;
576 int ret;
577
578 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
579 tss->flags);
580 if (ret > 0)
581 rd_desc->count -= ret;
582 return ret;
583 }
584
585 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
586 {
587 /* Store TCP splice context information in read_descriptor_t. */
588 read_descriptor_t rd_desc = {
589 .arg.data = tss,
590 .count = tss->len,
591 };
592
593 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
594 }
595
596 /**
597 * tcp_splice_read - splice data from TCP socket to a pipe
598 * @sock: socket to splice from
599 * @ppos: position (not valid)
600 * @pipe: pipe to splice to
601 * @len: number of bytes to splice
602 * @flags: splice modifier flags
603 *
604 * Description:
605 * Will read pages from given socket and fill them into a pipe.
606 *
607 **/
608 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
609 struct pipe_inode_info *pipe, size_t len,
610 unsigned int flags)
611 {
612 struct sock *sk = sock->sk;
613 struct tcp_splice_state tss = {
614 .pipe = pipe,
615 .len = len,
616 .flags = flags,
617 };
618 long timeo;
619 ssize_t spliced;
620 int ret;
621
622 sock_rps_record_flow(sk);
623 /*
624 * We can't seek on a socket input
625 */
626 if (unlikely(*ppos))
627 return -ESPIPE;
628
629 ret = spliced = 0;
630
631 lock_sock(sk);
632
633 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
634 while (tss.len) {
635 ret = __tcp_splice_read(sk, &tss);
636 if (ret < 0)
637 break;
638 else if (!ret) {
639 if (spliced)
640 break;
641 if (sock_flag(sk, SOCK_DONE))
642 break;
643 if (sk->sk_err) {
644 ret = sock_error(sk);
645 break;
646 }
647 if (sk->sk_shutdown & RCV_SHUTDOWN)
648 break;
649 if (sk->sk_state == TCP_CLOSE) {
650 /*
651 * This occurs when user tries to read
652 * from never connected socket.
653 */
654 if (!sock_flag(sk, SOCK_DONE))
655 ret = -ENOTCONN;
656 break;
657 }
658 if (!timeo) {
659 ret = -EAGAIN;
660 break;
661 }
662 sk_wait_data(sk, &timeo);
663 if (signal_pending(current)) {
664 ret = sock_intr_errno(timeo);
665 break;
666 }
667 continue;
668 }
669 tss.len -= ret;
670 spliced += ret;
671
672 if (!timeo)
673 break;
674 release_sock(sk);
675 lock_sock(sk);
676
677 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
678 (sk->sk_shutdown & RCV_SHUTDOWN) ||
679 signal_pending(current))
680 break;
681 }
682
683 release_sock(sk);
684
685 if (spliced)
686 return spliced;
687
688 return ret;
689 }
690 EXPORT_SYMBOL(tcp_splice_read);
691
692 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
693 {
694 struct sk_buff *skb;
695
696 /* The TCP header must be at least 32-bit aligned. */
697 size = ALIGN(size, 4);
698
699 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
700 if (skb) {
701 if (sk_wmem_schedule(sk, skb->truesize)) {
702 skb_reserve(skb, sk->sk_prot->max_header);
703 /*
704 * Make sure that we have exactly size bytes
705 * available to the caller, no more, no less.
706 */
707 skb->avail_size = size;
708 return skb;
709 }
710 __kfree_skb(skb);
711 } else {
712 sk->sk_prot->enter_memory_pressure(sk);
713 sk_stream_moderate_sndbuf(sk);
714 }
715 return NULL;
716 }
717
718 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
719 int large_allowed)
720 {
721 struct tcp_sock *tp = tcp_sk(sk);
722 u32 xmit_size_goal, old_size_goal;
723
724 xmit_size_goal = mss_now;
725
726 if (large_allowed && sk_can_gso(sk)) {
727 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
728 inet_csk(sk)->icsk_af_ops->net_header_len -
729 inet_csk(sk)->icsk_ext_hdr_len -
730 tp->tcp_header_len);
731
732 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
733
734 /* We try hard to avoid divides here */
735 old_size_goal = tp->xmit_size_goal_segs * mss_now;
736
737 if (likely(old_size_goal <= xmit_size_goal &&
738 old_size_goal + mss_now > xmit_size_goal)) {
739 xmit_size_goal = old_size_goal;
740 } else {
741 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
742 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
743 }
744 }
745
746 return max(xmit_size_goal, mss_now);
747 }
748
749 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
750 {
751 int mss_now;
752
753 mss_now = tcp_current_mss(sk);
754 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
755
756 return mss_now;
757 }
758
759 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
760 size_t psize, int flags)
761 {
762 struct tcp_sock *tp = tcp_sk(sk);
763 int mss_now, size_goal;
764 int err;
765 ssize_t copied;
766 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
767
768 /* Wait for a connection to finish. */
769 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
770 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
771 goto out_err;
772
773 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
774
775 mss_now = tcp_send_mss(sk, &size_goal, flags);
776 copied = 0;
777
778 err = -EPIPE;
779 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
780 goto out_err;
781
782 while (psize > 0) {
783 struct sk_buff *skb = tcp_write_queue_tail(sk);
784 struct page *page = pages[poffset / PAGE_SIZE];
785 int copy, i, can_coalesce;
786 int offset = poffset % PAGE_SIZE;
787 int size = min_t(size_t, psize, PAGE_SIZE - offset);
788
789 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
790 new_segment:
791 if (!sk_stream_memory_free(sk))
792 goto wait_for_sndbuf;
793
794 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
795 if (!skb)
796 goto wait_for_memory;
797
798 skb_entail(sk, skb);
799 copy = size_goal;
800 }
801
802 if (copy > size)
803 copy = size;
804
805 i = skb_shinfo(skb)->nr_frags;
806 can_coalesce = skb_can_coalesce(skb, i, page, offset);
807 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
808 tcp_mark_push(tp, skb);
809 goto new_segment;
810 }
811 if (!sk_wmem_schedule(sk, copy))
812 goto wait_for_memory;
813
814 if (can_coalesce) {
815 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
816 } else {
817 get_page(page);
818 skb_fill_page_desc(skb, i, page, offset, copy);
819 }
820
821 skb->len += copy;
822 skb->data_len += copy;
823 skb->truesize += copy;
824 sk->sk_wmem_queued += copy;
825 sk_mem_charge(sk, copy);
826 skb->ip_summed = CHECKSUM_PARTIAL;
827 tp->write_seq += copy;
828 TCP_SKB_CB(skb)->end_seq += copy;
829 skb_shinfo(skb)->gso_segs = 0;
830
831 if (!copied)
832 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
833
834 copied += copy;
835 poffset += copy;
836 if (!(psize -= copy))
837 goto out;
838
839 if (skb->len < size_goal || (flags & MSG_OOB))
840 continue;
841
842 if (forced_push(tp)) {
843 tcp_mark_push(tp, skb);
844 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
845 } else if (skb == tcp_send_head(sk))
846 tcp_push_one(sk, mss_now);
847 continue;
848
849 wait_for_sndbuf:
850 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
851 wait_for_memory:
852 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
853
854 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
855 goto do_error;
856
857 mss_now = tcp_send_mss(sk, &size_goal, flags);
858 }
859
860 out:
861 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
862 tcp_push(sk, flags, mss_now, tp->nonagle);
863 return copied;
864
865 do_error:
866 if (copied)
867 goto out;
868 out_err:
869 return sk_stream_error(sk, flags, err);
870 }
871
872 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
873 size_t size, int flags)
874 {
875 ssize_t res;
876
877 if (!(sk->sk_route_caps & NETIF_F_SG) ||
878 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
879 return sock_no_sendpage(sk->sk_socket, page, offset, size,
880 flags);
881
882 lock_sock(sk);
883 res = do_tcp_sendpages(sk, &page, offset, size, flags);
884 release_sock(sk);
885 return res;
886 }
887 EXPORT_SYMBOL(tcp_sendpage);
888
889 static inline int select_size(const struct sock *sk, bool sg)
890 {
891 const struct tcp_sock *tp = tcp_sk(sk);
892 int tmp = tp->mss_cache;
893
894 if (sg) {
895 if (sk_can_gso(sk)) {
896 /* Small frames wont use a full page:
897 * Payload will immediately follow tcp header.
898 */
899 tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
900 } else {
901 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
902
903 if (tmp >= pgbreak &&
904 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
905 tmp = pgbreak;
906 }
907 }
908
909 return tmp;
910 }
911
912 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
913 size_t size)
914 {
915 struct iovec *iov;
916 struct tcp_sock *tp = tcp_sk(sk);
917 struct sk_buff *skb;
918 int iovlen, flags, err, copied;
919 int mss_now, size_goal;
920 bool sg;
921 long timeo;
922
923 lock_sock(sk);
924
925 flags = msg->msg_flags;
926 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
927
928 /* Wait for a connection to finish. */
929 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
930 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
931 goto out_err;
932
933 /* This should be in poll */
934 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
935
936 mss_now = tcp_send_mss(sk, &size_goal, flags);
937
938 /* Ok commence sending. */
939 iovlen = msg->msg_iovlen;
940 iov = msg->msg_iov;
941 copied = 0;
942
943 err = -EPIPE;
944 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
945 goto out_err;
946
947 sg = !!(sk->sk_route_caps & NETIF_F_SG);
948
949 while (--iovlen >= 0) {
950 size_t seglen = iov->iov_len;
951 unsigned char __user *from = iov->iov_base;
952
953 iov++;
954
955 while (seglen > 0) {
956 int copy = 0;
957 int max = size_goal;
958
959 skb = tcp_write_queue_tail(sk);
960 if (tcp_send_head(sk)) {
961 if (skb->ip_summed == CHECKSUM_NONE)
962 max = mss_now;
963 copy = max - skb->len;
964 }
965
966 if (copy <= 0) {
967 new_segment:
968 /* Allocate new segment. If the interface is SG,
969 * allocate skb fitting to single page.
970 */
971 if (!sk_stream_memory_free(sk))
972 goto wait_for_sndbuf;
973
974 skb = sk_stream_alloc_skb(sk,
975 select_size(sk, sg),
976 sk->sk_allocation);
977 if (!skb)
978 goto wait_for_memory;
979
980 /*
981 * Check whether we can use HW checksum.
982 */
983 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
984 skb->ip_summed = CHECKSUM_PARTIAL;
985
986 skb_entail(sk, skb);
987 copy = size_goal;
988 max = size_goal;
989 }
990
991 /* Try to append data to the end of skb. */
992 if (copy > seglen)
993 copy = seglen;
994
995 /* Where to copy to? */
996 if (skb_availroom(skb) > 0) {
997 /* We have some space in skb head. Superb! */
998 copy = min_t(int, copy, skb_availroom(skb));
999 err = skb_add_data_nocache(sk, skb, from, copy);
1000 if (err)
1001 goto do_fault;
1002 } else {
1003 int merge = 0;
1004 int i = skb_shinfo(skb)->nr_frags;
1005 struct page *page = sk->sk_sndmsg_page;
1006 int off;
1007
1008 if (page && page_count(page) == 1)
1009 sk->sk_sndmsg_off = 0;
1010
1011 off = sk->sk_sndmsg_off;
1012
1013 if (skb_can_coalesce(skb, i, page, off) &&
1014 off != PAGE_SIZE) {
1015 /* We can extend the last page
1016 * fragment. */
1017 merge = 1;
1018 } else if (i == MAX_SKB_FRAGS || !sg) {
1019 /* Need to add new fragment and cannot
1020 * do this because interface is non-SG,
1021 * or because all the page slots are
1022 * busy. */
1023 tcp_mark_push(tp, skb);
1024 goto new_segment;
1025 } else if (page) {
1026 if (off == PAGE_SIZE) {
1027 put_page(page);
1028 sk->sk_sndmsg_page = page = NULL;
1029 off = 0;
1030 }
1031 } else
1032 off = 0;
1033
1034 if (copy > PAGE_SIZE - off)
1035 copy = PAGE_SIZE - off;
1036
1037 if (!sk_wmem_schedule(sk, copy))
1038 goto wait_for_memory;
1039
1040 if (!page) {
1041 /* Allocate new cache page. */
1042 if (!(page = sk_stream_alloc_page(sk)))
1043 goto wait_for_memory;
1044 }
1045
1046 /* Time to copy data. We are close to
1047 * the end! */
1048 err = skb_copy_to_page_nocache(sk, from, skb,
1049 page, off, copy);
1050 if (err) {
1051 /* If this page was new, give it to the
1052 * socket so it does not get leaked.
1053 */
1054 if (!sk->sk_sndmsg_page) {
1055 sk->sk_sndmsg_page = page;
1056 sk->sk_sndmsg_off = 0;
1057 }
1058 goto do_error;
1059 }
1060
1061 /* Update the skb. */
1062 if (merge) {
1063 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1064 } else {
1065 skb_fill_page_desc(skb, i, page, off, copy);
1066 if (sk->sk_sndmsg_page) {
1067 get_page(page);
1068 } else if (off + copy < PAGE_SIZE) {
1069 get_page(page);
1070 sk->sk_sndmsg_page = page;
1071 }
1072 }
1073
1074 sk->sk_sndmsg_off = off + copy;
1075 }
1076
1077 if (!copied)
1078 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1079
1080 tp->write_seq += copy;
1081 TCP_SKB_CB(skb)->end_seq += copy;
1082 skb_shinfo(skb)->gso_segs = 0;
1083
1084 from += copy;
1085 copied += copy;
1086 if ((seglen -= copy) == 0 && iovlen == 0)
1087 goto out;
1088
1089 if (skb->len < max || (flags & MSG_OOB))
1090 continue;
1091
1092 if (forced_push(tp)) {
1093 tcp_mark_push(tp, skb);
1094 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1095 } else if (skb == tcp_send_head(sk))
1096 tcp_push_one(sk, mss_now);
1097 continue;
1098
1099 wait_for_sndbuf:
1100 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1101 wait_for_memory:
1102 if (copied)
1103 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1104
1105 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1106 goto do_error;
1107
1108 mss_now = tcp_send_mss(sk, &size_goal, flags);
1109 }
1110 }
1111
1112 out:
1113 if (copied)
1114 tcp_push(sk, flags, mss_now, tp->nonagle);
1115 release_sock(sk);
1116 return copied;
1117
1118 do_fault:
1119 if (!skb->len) {
1120 tcp_unlink_write_queue(skb, sk);
1121 /* It is the one place in all of TCP, except connection
1122 * reset, where we can be unlinking the send_head.
1123 */
1124 tcp_check_send_head(sk, skb);
1125 sk_wmem_free_skb(sk, skb);
1126 }
1127
1128 do_error:
1129 if (copied)
1130 goto out;
1131 out_err:
1132 err = sk_stream_error(sk, flags, err);
1133 release_sock(sk);
1134 return err;
1135 }
1136 EXPORT_SYMBOL(tcp_sendmsg);
1137
1138 /*
1139 * Handle reading urgent data. BSD has very simple semantics for
1140 * this, no blocking and very strange errors 8)
1141 */
1142
1143 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1144 {
1145 struct tcp_sock *tp = tcp_sk(sk);
1146
1147 /* No URG data to read. */
1148 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1149 tp->urg_data == TCP_URG_READ)
1150 return -EINVAL; /* Yes this is right ! */
1151
1152 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1153 return -ENOTCONN;
1154
1155 if (tp->urg_data & TCP_URG_VALID) {
1156 int err = 0;
1157 char c = tp->urg_data;
1158
1159 if (!(flags & MSG_PEEK))
1160 tp->urg_data = TCP_URG_READ;
1161
1162 /* Read urgent data. */
1163 msg->msg_flags |= MSG_OOB;
1164
1165 if (len > 0) {
1166 if (!(flags & MSG_TRUNC))
1167 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1168 len = 1;
1169 } else
1170 msg->msg_flags |= MSG_TRUNC;
1171
1172 return err ? -EFAULT : len;
1173 }
1174
1175 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1176 return 0;
1177
1178 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1179 * the available implementations agree in this case:
1180 * this call should never block, independent of the
1181 * blocking state of the socket.
1182 * Mike <pall@rz.uni-karlsruhe.de>
1183 */
1184 return -EAGAIN;
1185 }
1186
1187 /* Clean up the receive buffer for full frames taken by the user,
1188 * then send an ACK if necessary. COPIED is the number of bytes
1189 * tcp_recvmsg has given to the user so far, it speeds up the
1190 * calculation of whether or not we must ACK for the sake of
1191 * a window update.
1192 */
1193 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1194 {
1195 struct tcp_sock *tp = tcp_sk(sk);
1196 int time_to_ack = 0;
1197
1198 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1199
1200 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1201 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1202 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1203
1204 if (inet_csk_ack_scheduled(sk)) {
1205 const struct inet_connection_sock *icsk = inet_csk(sk);
1206 /* Delayed ACKs frequently hit locked sockets during bulk
1207 * receive. */
1208 if (icsk->icsk_ack.blocked ||
1209 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1210 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1211 /*
1212 * If this read emptied read buffer, we send ACK, if
1213 * connection is not bidirectional, user drained
1214 * receive buffer and there was a small segment
1215 * in queue.
1216 */
1217 (copied > 0 &&
1218 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1219 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1220 !icsk->icsk_ack.pingpong)) &&
1221 !atomic_read(&sk->sk_rmem_alloc)))
1222 time_to_ack = 1;
1223 }
1224
1225 /* We send an ACK if we can now advertise a non-zero window
1226 * which has been raised "significantly".
1227 *
1228 * Even if window raised up to infinity, do not send window open ACK
1229 * in states, where we will not receive more. It is useless.
1230 */
1231 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1232 __u32 rcv_window_now = tcp_receive_window(tp);
1233
1234 /* Optimize, __tcp_select_window() is not cheap. */
1235 if (2*rcv_window_now <= tp->window_clamp) {
1236 __u32 new_window = __tcp_select_window(sk);
1237
1238 /* Send ACK now, if this read freed lots of space
1239 * in our buffer. Certainly, new_window is new window.
1240 * We can advertise it now, if it is not less than current one.
1241 * "Lots" means "at least twice" here.
1242 */
1243 if (new_window && new_window >= 2 * rcv_window_now)
1244 time_to_ack = 1;
1245 }
1246 }
1247 if (time_to_ack)
1248 tcp_send_ack(sk);
1249 }
1250
1251 static void tcp_prequeue_process(struct sock *sk)
1252 {
1253 struct sk_buff *skb;
1254 struct tcp_sock *tp = tcp_sk(sk);
1255
1256 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1257
1258 /* RX process wants to run with disabled BHs, though it is not
1259 * necessary */
1260 local_bh_disable();
1261 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1262 sk_backlog_rcv(sk, skb);
1263 local_bh_enable();
1264
1265 /* Clear memory counter. */
1266 tp->ucopy.memory = 0;
1267 }
1268
1269 #ifdef CONFIG_NET_DMA
1270 static void tcp_service_net_dma(struct sock *sk, bool wait)
1271 {
1272 dma_cookie_t done, used;
1273 dma_cookie_t last_issued;
1274 struct tcp_sock *tp = tcp_sk(sk);
1275
1276 if (!tp->ucopy.dma_chan)
1277 return;
1278
1279 last_issued = tp->ucopy.dma_cookie;
1280 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1281
1282 do {
1283 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1284 last_issued, &done,
1285 &used) == DMA_SUCCESS) {
1286 /* Safe to free early-copied skbs now */
1287 __skb_queue_purge(&sk->sk_async_wait_queue);
1288 break;
1289 } else {
1290 struct sk_buff *skb;
1291 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1292 (dma_async_is_complete(skb->dma_cookie, done,
1293 used) == DMA_SUCCESS)) {
1294 __skb_dequeue(&sk->sk_async_wait_queue);
1295 kfree_skb(skb);
1296 }
1297 }
1298 } while (wait);
1299 }
1300 #endif
1301
1302 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1303 {
1304 struct sk_buff *skb;
1305 u32 offset;
1306
1307 skb_queue_walk(&sk->sk_receive_queue, skb) {
1308 offset = seq - TCP_SKB_CB(skb)->seq;
1309 if (tcp_hdr(skb)->syn)
1310 offset--;
1311 if (offset < skb->len || tcp_hdr(skb)->fin) {
1312 *off = offset;
1313 return skb;
1314 }
1315 }
1316 return NULL;
1317 }
1318
1319 /*
1320 * This routine provides an alternative to tcp_recvmsg() for routines
1321 * that would like to handle copying from skbuffs directly in 'sendfile'
1322 * fashion.
1323 * Note:
1324 * - It is assumed that the socket was locked by the caller.
1325 * - The routine does not block.
1326 * - At present, there is no support for reading OOB data
1327 * or for 'peeking' the socket using this routine
1328 * (although both would be easy to implement).
1329 */
1330 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1331 sk_read_actor_t recv_actor)
1332 {
1333 struct sk_buff *skb;
1334 struct tcp_sock *tp = tcp_sk(sk);
1335 u32 seq = tp->copied_seq;
1336 u32 offset;
1337 int copied = 0;
1338
1339 if (sk->sk_state == TCP_LISTEN)
1340 return -ENOTCONN;
1341 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1342 if (offset < skb->len) {
1343 int used;
1344 size_t len;
1345
1346 len = skb->len - offset;
1347 /* Stop reading if we hit a patch of urgent data */
1348 if (tp->urg_data) {
1349 u32 urg_offset = tp->urg_seq - seq;
1350 if (urg_offset < len)
1351 len = urg_offset;
1352 if (!len)
1353 break;
1354 }
1355 used = recv_actor(desc, skb, offset, len);
1356 if (used < 0) {
1357 if (!copied)
1358 copied = used;
1359 break;
1360 } else if (used <= len) {
1361 seq += used;
1362 copied += used;
1363 offset += used;
1364 }
1365 /*
1366 * If recv_actor drops the lock (e.g. TCP splice
1367 * receive) the skb pointer might be invalid when
1368 * getting here: tcp_collapse might have deleted it
1369 * while aggregating skbs from the socket queue.
1370 */
1371 skb = tcp_recv_skb(sk, seq-1, &offset);
1372 if (!skb || (offset+1 != skb->len))
1373 break;
1374 }
1375 if (tcp_hdr(skb)->fin) {
1376 sk_eat_skb(sk, skb, 0);
1377 ++seq;
1378 break;
1379 }
1380 sk_eat_skb(sk, skb, 0);
1381 if (!desc->count)
1382 break;
1383 tp->copied_seq = seq;
1384 }
1385 tp->copied_seq = seq;
1386
1387 tcp_rcv_space_adjust(sk);
1388
1389 /* Clean up data we have read: This will do ACK frames. */
1390 if (copied > 0)
1391 tcp_cleanup_rbuf(sk, copied);
1392 return copied;
1393 }
1394 EXPORT_SYMBOL(tcp_read_sock);
1395
1396 /*
1397 * This routine copies from a sock struct into the user buffer.
1398 *
1399 * Technical note: in 2.3 we work on _locked_ socket, so that
1400 * tricks with *seq access order and skb->users are not required.
1401 * Probably, code can be easily improved even more.
1402 */
1403
1404 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1405 size_t len, int nonblock, int flags, int *addr_len)
1406 {
1407 struct tcp_sock *tp = tcp_sk(sk);
1408 int copied = 0;
1409 u32 peek_seq;
1410 u32 *seq;
1411 unsigned long used;
1412 int err;
1413 int target; /* Read at least this many bytes */
1414 long timeo;
1415 struct task_struct *user_recv = NULL;
1416 int copied_early = 0;
1417 struct sk_buff *skb;
1418 u32 urg_hole = 0;
1419
1420 lock_sock(sk);
1421
1422 err = -ENOTCONN;
1423 if (sk->sk_state == TCP_LISTEN)
1424 goto out;
1425
1426 timeo = sock_rcvtimeo(sk, nonblock);
1427
1428 /* Urgent data needs to be handled specially. */
1429 if (flags & MSG_OOB)
1430 goto recv_urg;
1431
1432 seq = &tp->copied_seq;
1433 if (flags & MSG_PEEK) {
1434 peek_seq = tp->copied_seq;
1435 seq = &peek_seq;
1436 }
1437
1438 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1439
1440 #ifdef CONFIG_NET_DMA
1441 tp->ucopy.dma_chan = NULL;
1442 preempt_disable();
1443 skb = skb_peek_tail(&sk->sk_receive_queue);
1444 {
1445 int available = 0;
1446
1447 if (skb)
1448 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1449 if ((available < target) &&
1450 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1451 !sysctl_tcp_low_latency &&
1452 dma_find_channel(DMA_MEMCPY)) {
1453 preempt_enable_no_resched();
1454 tp->ucopy.pinned_list =
1455 dma_pin_iovec_pages(msg->msg_iov, len);
1456 } else {
1457 preempt_enable_no_resched();
1458 }
1459 }
1460 #endif
1461
1462 do {
1463 u32 offset;
1464
1465 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1466 if (tp->urg_data && tp->urg_seq == *seq) {
1467 if (copied)
1468 break;
1469 if (signal_pending(current)) {
1470 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1471 break;
1472 }
1473 }
1474
1475 /* Next get a buffer. */
1476
1477 skb_queue_walk(&sk->sk_receive_queue, skb) {
1478 /* Now that we have two receive queues this
1479 * shouldn't happen.
1480 */
1481 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1482 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1483 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1484 flags))
1485 break;
1486
1487 offset = *seq - TCP_SKB_CB(skb)->seq;
1488 if (tcp_hdr(skb)->syn)
1489 offset--;
1490 if (offset < skb->len)
1491 goto found_ok_skb;
1492 if (tcp_hdr(skb)->fin)
1493 goto found_fin_ok;
1494 WARN(!(flags & MSG_PEEK),
1495 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1496 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1497 }
1498
1499 /* Well, if we have backlog, try to process it now yet. */
1500
1501 if (copied >= target && !sk->sk_backlog.tail)
1502 break;
1503
1504 if (copied) {
1505 if (sk->sk_err ||
1506 sk->sk_state == TCP_CLOSE ||
1507 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1508 !timeo ||
1509 signal_pending(current))
1510 break;
1511 } else {
1512 if (sock_flag(sk, SOCK_DONE))
1513 break;
1514
1515 if (sk->sk_err) {
1516 copied = sock_error(sk);
1517 break;
1518 }
1519
1520 if (sk->sk_shutdown & RCV_SHUTDOWN)
1521 break;
1522
1523 if (sk->sk_state == TCP_CLOSE) {
1524 if (!sock_flag(sk, SOCK_DONE)) {
1525 /* This occurs when user tries to read
1526 * from never connected socket.
1527 */
1528 copied = -ENOTCONN;
1529 break;
1530 }
1531 break;
1532 }
1533
1534 if (!timeo) {
1535 copied = -EAGAIN;
1536 break;
1537 }
1538
1539 if (signal_pending(current)) {
1540 copied = sock_intr_errno(timeo);
1541 break;
1542 }
1543 }
1544
1545 tcp_cleanup_rbuf(sk, copied);
1546
1547 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1548 /* Install new reader */
1549 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1550 user_recv = current;
1551 tp->ucopy.task = user_recv;
1552 tp->ucopy.iov = msg->msg_iov;
1553 }
1554
1555 tp->ucopy.len = len;
1556
1557 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1558 !(flags & (MSG_PEEK | MSG_TRUNC)));
1559
1560 /* Ugly... If prequeue is not empty, we have to
1561 * process it before releasing socket, otherwise
1562 * order will be broken at second iteration.
1563 * More elegant solution is required!!!
1564 *
1565 * Look: we have the following (pseudo)queues:
1566 *
1567 * 1. packets in flight
1568 * 2. backlog
1569 * 3. prequeue
1570 * 4. receive_queue
1571 *
1572 * Each queue can be processed only if the next ones
1573 * are empty. At this point we have empty receive_queue.
1574 * But prequeue _can_ be not empty after 2nd iteration,
1575 * when we jumped to start of loop because backlog
1576 * processing added something to receive_queue.
1577 * We cannot release_sock(), because backlog contains
1578 * packets arrived _after_ prequeued ones.
1579 *
1580 * Shortly, algorithm is clear --- to process all
1581 * the queues in order. We could make it more directly,
1582 * requeueing packets from backlog to prequeue, if
1583 * is not empty. It is more elegant, but eats cycles,
1584 * unfortunately.
1585 */
1586 if (!skb_queue_empty(&tp->ucopy.prequeue))
1587 goto do_prequeue;
1588
1589 /* __ Set realtime policy in scheduler __ */
1590 }
1591
1592 #ifdef CONFIG_NET_DMA
1593 if (tp->ucopy.dma_chan)
1594 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1595 #endif
1596 if (copied >= target) {
1597 /* Do not sleep, just process backlog. */
1598 release_sock(sk);
1599 lock_sock(sk);
1600 } else
1601 sk_wait_data(sk, &timeo);
1602
1603 #ifdef CONFIG_NET_DMA
1604 tcp_service_net_dma(sk, false); /* Don't block */
1605 tp->ucopy.wakeup = 0;
1606 #endif
1607
1608 if (user_recv) {
1609 int chunk;
1610
1611 /* __ Restore normal policy in scheduler __ */
1612
1613 if ((chunk = len - tp->ucopy.len) != 0) {
1614 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1615 len -= chunk;
1616 copied += chunk;
1617 }
1618
1619 if (tp->rcv_nxt == tp->copied_seq &&
1620 !skb_queue_empty(&tp->ucopy.prequeue)) {
1621 do_prequeue:
1622 tcp_prequeue_process(sk);
1623
1624 if ((chunk = len - tp->ucopy.len) != 0) {
1625 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1626 len -= chunk;
1627 copied += chunk;
1628 }
1629 }
1630 }
1631 if ((flags & MSG_PEEK) &&
1632 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1633 if (net_ratelimit())
1634 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1635 current->comm, task_pid_nr(current));
1636 peek_seq = tp->copied_seq;
1637 }
1638 continue;
1639
1640 found_ok_skb:
1641 /* Ok so how much can we use? */
1642 used = skb->len - offset;
1643 if (len < used)
1644 used = len;
1645
1646 /* Do we have urgent data here? */
1647 if (tp->urg_data) {
1648 u32 urg_offset = tp->urg_seq - *seq;
1649 if (urg_offset < used) {
1650 if (!urg_offset) {
1651 if (!sock_flag(sk, SOCK_URGINLINE)) {
1652 ++*seq;
1653 urg_hole++;
1654 offset++;
1655 used--;
1656 if (!used)
1657 goto skip_copy;
1658 }
1659 } else
1660 used = urg_offset;
1661 }
1662 }
1663
1664 if (!(flags & MSG_TRUNC)) {
1665 #ifdef CONFIG_NET_DMA
1666 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1667 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1668
1669 if (tp->ucopy.dma_chan) {
1670 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1671 tp->ucopy.dma_chan, skb, offset,
1672 msg->msg_iov, used,
1673 tp->ucopy.pinned_list);
1674
1675 if (tp->ucopy.dma_cookie < 0) {
1676
1677 printk(KERN_ALERT "dma_cookie < 0\n");
1678
1679 /* Exception. Bailout! */
1680 if (!copied)
1681 copied = -EFAULT;
1682 break;
1683 }
1684
1685 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1686
1687 if ((offset + used) == skb->len)
1688 copied_early = 1;
1689
1690 } else
1691 #endif
1692 {
1693 err = skb_copy_datagram_iovec(skb, offset,
1694 msg->msg_iov, used);
1695 if (err) {
1696 /* Exception. Bailout! */
1697 if (!copied)
1698 copied = -EFAULT;
1699 break;
1700 }
1701 }
1702 }
1703
1704 *seq += used;
1705 copied += used;
1706 len -= used;
1707
1708 tcp_rcv_space_adjust(sk);
1709
1710 skip_copy:
1711 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1712 tp->urg_data = 0;
1713 tcp_fast_path_check(sk);
1714 }
1715 if (used + offset < skb->len)
1716 continue;
1717
1718 if (tcp_hdr(skb)->fin)
1719 goto found_fin_ok;
1720 if (!(flags & MSG_PEEK)) {
1721 sk_eat_skb(sk, skb, copied_early);
1722 copied_early = 0;
1723 }
1724 continue;
1725
1726 found_fin_ok:
1727 /* Process the FIN. */
1728 ++*seq;
1729 if (!(flags & MSG_PEEK)) {
1730 sk_eat_skb(sk, skb, copied_early);
1731 copied_early = 0;
1732 }
1733 break;
1734 } while (len > 0);
1735
1736 if (user_recv) {
1737 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1738 int chunk;
1739
1740 tp->ucopy.len = copied > 0 ? len : 0;
1741
1742 tcp_prequeue_process(sk);
1743
1744 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1745 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1746 len -= chunk;
1747 copied += chunk;
1748 }
1749 }
1750
1751 tp->ucopy.task = NULL;
1752 tp->ucopy.len = 0;
1753 }
1754
1755 #ifdef CONFIG_NET_DMA
1756 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1757 tp->ucopy.dma_chan = NULL;
1758
1759 if (tp->ucopy.pinned_list) {
1760 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1761 tp->ucopy.pinned_list = NULL;
1762 }
1763 #endif
1764
1765 /* According to UNIX98, msg_name/msg_namelen are ignored
1766 * on connected socket. I was just happy when found this 8) --ANK
1767 */
1768
1769 /* Clean up data we have read: This will do ACK frames. */
1770 tcp_cleanup_rbuf(sk, copied);
1771
1772 release_sock(sk);
1773 return copied;
1774
1775 out:
1776 release_sock(sk);
1777 return err;
1778
1779 recv_urg:
1780 err = tcp_recv_urg(sk, msg, len, flags);
1781 goto out;
1782 }
1783 EXPORT_SYMBOL(tcp_recvmsg);
1784
1785 void tcp_set_state(struct sock *sk, int state)
1786 {
1787 int oldstate = sk->sk_state;
1788
1789 switch (state) {
1790 case TCP_ESTABLISHED:
1791 if (oldstate != TCP_ESTABLISHED)
1792 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1793 break;
1794
1795 case TCP_CLOSE:
1796 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1797 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1798
1799 sk->sk_prot->unhash(sk);
1800 if (inet_csk(sk)->icsk_bind_hash &&
1801 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1802 inet_put_port(sk);
1803 /* fall through */
1804 default:
1805 if (oldstate == TCP_ESTABLISHED)
1806 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1807 }
1808
1809 /* Change state AFTER socket is unhashed to avoid closed
1810 * socket sitting in hash tables.
1811 */
1812 sk->sk_state = state;
1813
1814 #ifdef STATE_TRACE
1815 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1816 #endif
1817 }
1818 EXPORT_SYMBOL_GPL(tcp_set_state);
1819
1820 /*
1821 * State processing on a close. This implements the state shift for
1822 * sending our FIN frame. Note that we only send a FIN for some
1823 * states. A shutdown() may have already sent the FIN, or we may be
1824 * closed.
1825 */
1826
1827 static const unsigned char new_state[16] = {
1828 /* current state: new state: action: */
1829 /* (Invalid) */ TCP_CLOSE,
1830 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1831 /* TCP_SYN_SENT */ TCP_CLOSE,
1832 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1833 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1834 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1835 /* TCP_TIME_WAIT */ TCP_CLOSE,
1836 /* TCP_CLOSE */ TCP_CLOSE,
1837 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1838 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1839 /* TCP_LISTEN */ TCP_CLOSE,
1840 /* TCP_CLOSING */ TCP_CLOSING,
1841 };
1842
1843 static int tcp_close_state(struct sock *sk)
1844 {
1845 int next = (int)new_state[sk->sk_state];
1846 int ns = next & TCP_STATE_MASK;
1847
1848 tcp_set_state(sk, ns);
1849
1850 return next & TCP_ACTION_FIN;
1851 }
1852
1853 /*
1854 * Shutdown the sending side of a connection. Much like close except
1855 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1856 */
1857
1858 void tcp_shutdown(struct sock *sk, int how)
1859 {
1860 /* We need to grab some memory, and put together a FIN,
1861 * and then put it into the queue to be sent.
1862 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1863 */
1864 if (!(how & SEND_SHUTDOWN))
1865 return;
1866
1867 /* If we've already sent a FIN, or it's a closed state, skip this. */
1868 if ((1 << sk->sk_state) &
1869 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1870 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1871 /* Clear out any half completed packets. FIN if needed. */
1872 if (tcp_close_state(sk))
1873 tcp_send_fin(sk);
1874 }
1875 }
1876 EXPORT_SYMBOL(tcp_shutdown);
1877
1878 bool tcp_check_oom(struct sock *sk, int shift)
1879 {
1880 bool too_many_orphans, out_of_socket_memory;
1881
1882 too_many_orphans = tcp_too_many_orphans(sk, shift);
1883 out_of_socket_memory = tcp_out_of_memory(sk);
1884
1885 if (too_many_orphans && net_ratelimit())
1886 pr_info("TCP: too many orphaned sockets\n");
1887 if (out_of_socket_memory && net_ratelimit())
1888 pr_info("TCP: out of memory -- consider tuning tcp_mem\n");
1889 return too_many_orphans || out_of_socket_memory;
1890 }
1891
1892 void tcp_close(struct sock *sk, long timeout)
1893 {
1894 struct sk_buff *skb;
1895 int data_was_unread = 0;
1896 int state;
1897
1898 lock_sock(sk);
1899 sk->sk_shutdown = SHUTDOWN_MASK;
1900
1901 if (sk->sk_state == TCP_LISTEN) {
1902 tcp_set_state(sk, TCP_CLOSE);
1903
1904 /* Special case. */
1905 inet_csk_listen_stop(sk);
1906
1907 goto adjudge_to_death;
1908 }
1909
1910 /* We need to flush the recv. buffs. We do this only on the
1911 * descriptor close, not protocol-sourced closes, because the
1912 * reader process may not have drained the data yet!
1913 */
1914 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1915 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1916 tcp_hdr(skb)->fin;
1917 data_was_unread += len;
1918 __kfree_skb(skb);
1919 }
1920
1921 sk_mem_reclaim(sk);
1922
1923 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1924 if (sk->sk_state == TCP_CLOSE)
1925 goto adjudge_to_death;
1926
1927 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1928 * data was lost. To witness the awful effects of the old behavior of
1929 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1930 * GET in an FTP client, suspend the process, wait for the client to
1931 * advertise a zero window, then kill -9 the FTP client, wheee...
1932 * Note: timeout is always zero in such a case.
1933 */
1934 if (data_was_unread) {
1935 /* Unread data was tossed, zap the connection. */
1936 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1937 tcp_set_state(sk, TCP_CLOSE);
1938 tcp_send_active_reset(sk, sk->sk_allocation);
1939 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1940 /* Check zero linger _after_ checking for unread data. */
1941 sk->sk_prot->disconnect(sk, 0);
1942 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1943 } else if (tcp_close_state(sk)) {
1944 /* We FIN if the application ate all the data before
1945 * zapping the connection.
1946 */
1947
1948 /* RED-PEN. Formally speaking, we have broken TCP state
1949 * machine. State transitions:
1950 *
1951 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1952 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1953 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1954 *
1955 * are legal only when FIN has been sent (i.e. in window),
1956 * rather than queued out of window. Purists blame.
1957 *
1958 * F.e. "RFC state" is ESTABLISHED,
1959 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1960 *
1961 * The visible declinations are that sometimes
1962 * we enter time-wait state, when it is not required really
1963 * (harmless), do not send active resets, when they are
1964 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1965 * they look as CLOSING or LAST_ACK for Linux)
1966 * Probably, I missed some more holelets.
1967 * --ANK
1968 */
1969 tcp_send_fin(sk);
1970 }
1971
1972 sk_stream_wait_close(sk, timeout);
1973
1974 adjudge_to_death:
1975 state = sk->sk_state;
1976 sock_hold(sk);
1977 sock_orphan(sk);
1978
1979 /* It is the last release_sock in its life. It will remove backlog. */
1980 release_sock(sk);
1981
1982
1983 /* Now socket is owned by kernel and we acquire BH lock
1984 to finish close. No need to check for user refs.
1985 */
1986 local_bh_disable();
1987 bh_lock_sock(sk);
1988 WARN_ON(sock_owned_by_user(sk));
1989
1990 percpu_counter_inc(sk->sk_prot->orphan_count);
1991
1992 /* Have we already been destroyed by a softirq or backlog? */
1993 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1994 goto out;
1995
1996 /* This is a (useful) BSD violating of the RFC. There is a
1997 * problem with TCP as specified in that the other end could
1998 * keep a socket open forever with no application left this end.
1999 * We use a 3 minute timeout (about the same as BSD) then kill
2000 * our end. If they send after that then tough - BUT: long enough
2001 * that we won't make the old 4*rto = almost no time - whoops
2002 * reset mistake.
2003 *
2004 * Nope, it was not mistake. It is really desired behaviour
2005 * f.e. on http servers, when such sockets are useless, but
2006 * consume significant resources. Let's do it with special
2007 * linger2 option. --ANK
2008 */
2009
2010 if (sk->sk_state == TCP_FIN_WAIT2) {
2011 struct tcp_sock *tp = tcp_sk(sk);
2012 if (tp->linger2 < 0) {
2013 tcp_set_state(sk, TCP_CLOSE);
2014 tcp_send_active_reset(sk, GFP_ATOMIC);
2015 NET_INC_STATS_BH(sock_net(sk),
2016 LINUX_MIB_TCPABORTONLINGER);
2017 } else {
2018 const int tmo = tcp_fin_time(sk);
2019
2020 if (tmo > TCP_TIMEWAIT_LEN) {
2021 inet_csk_reset_keepalive_timer(sk,
2022 tmo - TCP_TIMEWAIT_LEN);
2023 } else {
2024 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2025 goto out;
2026 }
2027 }
2028 }
2029 if (sk->sk_state != TCP_CLOSE) {
2030 sk_mem_reclaim(sk);
2031 if (tcp_check_oom(sk, 0)) {
2032 tcp_set_state(sk, TCP_CLOSE);
2033 tcp_send_active_reset(sk, GFP_ATOMIC);
2034 NET_INC_STATS_BH(sock_net(sk),
2035 LINUX_MIB_TCPABORTONMEMORY);
2036 }
2037 }
2038
2039 if (sk->sk_state == TCP_CLOSE)
2040 inet_csk_destroy_sock(sk);
2041 /* Otherwise, socket is reprieved until protocol close. */
2042
2043 out:
2044 bh_unlock_sock(sk);
2045 local_bh_enable();
2046 sock_put(sk);
2047 }
2048 EXPORT_SYMBOL(tcp_close);
2049
2050 /* These states need RST on ABORT according to RFC793 */
2051
2052 static inline int tcp_need_reset(int state)
2053 {
2054 return (1 << state) &
2055 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2056 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2057 }
2058
2059 int tcp_disconnect(struct sock *sk, int flags)
2060 {
2061 struct inet_sock *inet = inet_sk(sk);
2062 struct inet_connection_sock *icsk = inet_csk(sk);
2063 struct tcp_sock *tp = tcp_sk(sk);
2064 int err = 0;
2065 int old_state = sk->sk_state;
2066
2067 if (old_state != TCP_CLOSE)
2068 tcp_set_state(sk, TCP_CLOSE);
2069
2070 /* ABORT function of RFC793 */
2071 if (old_state == TCP_LISTEN) {
2072 inet_csk_listen_stop(sk);
2073 } else if (tcp_need_reset(old_state) ||
2074 (tp->snd_nxt != tp->write_seq &&
2075 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2076 /* The last check adjusts for discrepancy of Linux wrt. RFC
2077 * states
2078 */
2079 tcp_send_active_reset(sk, gfp_any());
2080 sk->sk_err = ECONNRESET;
2081 } else if (old_state == TCP_SYN_SENT)
2082 sk->sk_err = ECONNRESET;
2083
2084 tcp_clear_xmit_timers(sk);
2085 __skb_queue_purge(&sk->sk_receive_queue);
2086 tcp_write_queue_purge(sk);
2087 __skb_queue_purge(&tp->out_of_order_queue);
2088 #ifdef CONFIG_NET_DMA
2089 __skb_queue_purge(&sk->sk_async_wait_queue);
2090 #endif
2091
2092 inet->inet_dport = 0;
2093
2094 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2095 inet_reset_saddr(sk);
2096
2097 sk->sk_shutdown = 0;
2098 sock_reset_flag(sk, SOCK_DONE);
2099 tp->srtt = 0;
2100 if ((tp->write_seq += tp->max_window + 2) == 0)
2101 tp->write_seq = 1;
2102 icsk->icsk_backoff = 0;
2103 tp->snd_cwnd = 2;
2104 icsk->icsk_probes_out = 0;
2105 tp->packets_out = 0;
2106 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2107 tp->snd_cwnd_cnt = 0;
2108 tp->bytes_acked = 0;
2109 tp->window_clamp = 0;
2110 tcp_set_ca_state(sk, TCP_CA_Open);
2111 tcp_clear_retrans(tp);
2112 inet_csk_delack_init(sk);
2113 tcp_init_send_head(sk);
2114 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2115 __sk_dst_reset(sk);
2116
2117 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2118
2119 sk->sk_error_report(sk);
2120 return err;
2121 }
2122 EXPORT_SYMBOL(tcp_disconnect);
2123
2124 /*
2125 * Socket option code for TCP.
2126 */
2127 static int do_tcp_setsockopt(struct sock *sk, int level,
2128 int optname, char __user *optval, unsigned int optlen)
2129 {
2130 struct tcp_sock *tp = tcp_sk(sk);
2131 struct inet_connection_sock *icsk = inet_csk(sk);
2132 int val;
2133 int err = 0;
2134
2135 /* These are data/string values, all the others are ints */
2136 switch (optname) {
2137 case TCP_CONGESTION: {
2138 char name[TCP_CA_NAME_MAX];
2139
2140 if (optlen < 1)
2141 return -EINVAL;
2142
2143 val = strncpy_from_user(name, optval,
2144 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2145 if (val < 0)
2146 return -EFAULT;
2147 name[val] = 0;
2148
2149 lock_sock(sk);
2150 err = tcp_set_congestion_control(sk, name);
2151 release_sock(sk);
2152 return err;
2153 }
2154 case TCP_COOKIE_TRANSACTIONS: {
2155 struct tcp_cookie_transactions ctd;
2156 struct tcp_cookie_values *cvp = NULL;
2157
2158 if (sizeof(ctd) > optlen)
2159 return -EINVAL;
2160 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2161 return -EFAULT;
2162
2163 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2164 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2165 return -EINVAL;
2166
2167 if (ctd.tcpct_cookie_desired == 0) {
2168 /* default to global value */
2169 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2170 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2171 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2172 return -EINVAL;
2173 }
2174
2175 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2176 /* Supercedes all other values */
2177 lock_sock(sk);
2178 if (tp->cookie_values != NULL) {
2179 kref_put(&tp->cookie_values->kref,
2180 tcp_cookie_values_release);
2181 tp->cookie_values = NULL;
2182 }
2183 tp->rx_opt.cookie_in_always = 0; /* false */
2184 tp->rx_opt.cookie_out_never = 1; /* true */
2185 release_sock(sk);
2186 return err;
2187 }
2188
2189 /* Allocate ancillary memory before locking.
2190 */
2191 if (ctd.tcpct_used > 0 ||
2192 (tp->cookie_values == NULL &&
2193 (sysctl_tcp_cookie_size > 0 ||
2194 ctd.tcpct_cookie_desired > 0 ||
2195 ctd.tcpct_s_data_desired > 0))) {
2196 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2197 GFP_KERNEL);
2198 if (cvp == NULL)
2199 return -ENOMEM;
2200
2201 kref_init(&cvp->kref);
2202 }
2203 lock_sock(sk);
2204 tp->rx_opt.cookie_in_always =
2205 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2206 tp->rx_opt.cookie_out_never = 0; /* false */
2207
2208 if (tp->cookie_values != NULL) {
2209 if (cvp != NULL) {
2210 /* Changed values are recorded by a changed
2211 * pointer, ensuring the cookie will differ,
2212 * without separately hashing each value later.
2213 */
2214 kref_put(&tp->cookie_values->kref,
2215 tcp_cookie_values_release);
2216 } else {
2217 cvp = tp->cookie_values;
2218 }
2219 }
2220
2221 if (cvp != NULL) {
2222 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2223
2224 if (ctd.tcpct_used > 0) {
2225 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2226 ctd.tcpct_used);
2227 cvp->s_data_desired = ctd.tcpct_used;
2228 cvp->s_data_constant = 1; /* true */
2229 } else {
2230 /* No constant payload data. */
2231 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2232 cvp->s_data_constant = 0; /* false */
2233 }
2234
2235 tp->cookie_values = cvp;
2236 }
2237 release_sock(sk);
2238 return err;
2239 }
2240 default:
2241 /* fallthru */
2242 break;
2243 }
2244
2245 if (optlen < sizeof(int))
2246 return -EINVAL;
2247
2248 if (get_user(val, (int __user *)optval))
2249 return -EFAULT;
2250
2251 lock_sock(sk);
2252
2253 switch (optname) {
2254 case TCP_MAXSEG:
2255 /* Values greater than interface MTU won't take effect. However
2256 * at the point when this call is done we typically don't yet
2257 * know which interface is going to be used */
2258 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2259 err = -EINVAL;
2260 break;
2261 }
2262 tp->rx_opt.user_mss = val;
2263 break;
2264
2265 case TCP_NODELAY:
2266 if (val) {
2267 /* TCP_NODELAY is weaker than TCP_CORK, so that
2268 * this option on corked socket is remembered, but
2269 * it is not activated until cork is cleared.
2270 *
2271 * However, when TCP_NODELAY is set we make
2272 * an explicit push, which overrides even TCP_CORK
2273 * for currently queued segments.
2274 */
2275 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2276 tcp_push_pending_frames(sk);
2277 } else {
2278 tp->nonagle &= ~TCP_NAGLE_OFF;
2279 }
2280 break;
2281
2282 case TCP_THIN_LINEAR_TIMEOUTS:
2283 if (val < 0 || val > 1)
2284 err = -EINVAL;
2285 else
2286 tp->thin_lto = val;
2287 break;
2288
2289 case TCP_THIN_DUPACK:
2290 if (val < 0 || val > 1)
2291 err = -EINVAL;
2292 else
2293 tp->thin_dupack = val;
2294 break;
2295
2296 case TCP_CORK:
2297 /* When set indicates to always queue non-full frames.
2298 * Later the user clears this option and we transmit
2299 * any pending partial frames in the queue. This is
2300 * meant to be used alongside sendfile() to get properly
2301 * filled frames when the user (for example) must write
2302 * out headers with a write() call first and then use
2303 * sendfile to send out the data parts.
2304 *
2305 * TCP_CORK can be set together with TCP_NODELAY and it is
2306 * stronger than TCP_NODELAY.
2307 */
2308 if (val) {
2309 tp->nonagle |= TCP_NAGLE_CORK;
2310 } else {
2311 tp->nonagle &= ~TCP_NAGLE_CORK;
2312 if (tp->nonagle&TCP_NAGLE_OFF)
2313 tp->nonagle |= TCP_NAGLE_PUSH;
2314 tcp_push_pending_frames(sk);
2315 }
2316 break;
2317
2318 case TCP_KEEPIDLE:
2319 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2320 err = -EINVAL;
2321 else {
2322 tp->keepalive_time = val * HZ;
2323 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2324 !((1 << sk->sk_state) &
2325 (TCPF_CLOSE | TCPF_LISTEN))) {
2326 u32 elapsed = keepalive_time_elapsed(tp);
2327 if (tp->keepalive_time > elapsed)
2328 elapsed = tp->keepalive_time - elapsed;
2329 else
2330 elapsed = 0;
2331 inet_csk_reset_keepalive_timer(sk, elapsed);
2332 }
2333 }
2334 break;
2335 case TCP_KEEPINTVL:
2336 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2337 err = -EINVAL;
2338 else
2339 tp->keepalive_intvl = val * HZ;
2340 break;
2341 case TCP_KEEPCNT:
2342 if (val < 1 || val > MAX_TCP_KEEPCNT)
2343 err = -EINVAL;
2344 else
2345 tp->keepalive_probes = val;
2346 break;
2347 case TCP_SYNCNT:
2348 if (val < 1 || val > MAX_TCP_SYNCNT)
2349 err = -EINVAL;
2350 else
2351 icsk->icsk_syn_retries = val;
2352 break;
2353
2354 case TCP_LINGER2:
2355 if (val < 0)
2356 tp->linger2 = -1;
2357 else if (val > sysctl_tcp_fin_timeout / HZ)
2358 tp->linger2 = 0;
2359 else
2360 tp->linger2 = val * HZ;
2361 break;
2362
2363 case TCP_DEFER_ACCEPT:
2364 /* Translate value in seconds to number of retransmits */
2365 icsk->icsk_accept_queue.rskq_defer_accept =
2366 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2367 TCP_RTO_MAX / HZ);
2368 break;
2369
2370 case TCP_WINDOW_CLAMP:
2371 if (!val) {
2372 if (sk->sk_state != TCP_CLOSE) {
2373 err = -EINVAL;
2374 break;
2375 }
2376 tp->window_clamp = 0;
2377 } else
2378 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2379 SOCK_MIN_RCVBUF / 2 : val;
2380 break;
2381
2382 case TCP_QUICKACK:
2383 if (!val) {
2384 icsk->icsk_ack.pingpong = 1;
2385 } else {
2386 icsk->icsk_ack.pingpong = 0;
2387 if ((1 << sk->sk_state) &
2388 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2389 inet_csk_ack_scheduled(sk)) {
2390 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2391 tcp_cleanup_rbuf(sk, 1);
2392 if (!(val & 1))
2393 icsk->icsk_ack.pingpong = 1;
2394 }
2395 }
2396 break;
2397
2398 #ifdef CONFIG_TCP_MD5SIG
2399 case TCP_MD5SIG:
2400 /* Read the IP->Key mappings from userspace */
2401 err = tp->af_specific->md5_parse(sk, optval, optlen);
2402 break;
2403 #endif
2404 case TCP_USER_TIMEOUT:
2405 /* Cap the max timeout in ms TCP will retry/retrans
2406 * before giving up and aborting (ETIMEDOUT) a connection.
2407 */
2408 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2409 break;
2410 default:
2411 err = -ENOPROTOOPT;
2412 break;
2413 }
2414
2415 release_sock(sk);
2416 return err;
2417 }
2418
2419 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2420 unsigned int optlen)
2421 {
2422 const struct inet_connection_sock *icsk = inet_csk(sk);
2423
2424 if (level != SOL_TCP)
2425 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2426 optval, optlen);
2427 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2428 }
2429 EXPORT_SYMBOL(tcp_setsockopt);
2430
2431 #ifdef CONFIG_COMPAT
2432 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2433 char __user *optval, unsigned int optlen)
2434 {
2435 if (level != SOL_TCP)
2436 return inet_csk_compat_setsockopt(sk, level, optname,
2437 optval, optlen);
2438 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2439 }
2440 EXPORT_SYMBOL(compat_tcp_setsockopt);
2441 #endif
2442
2443 /* Return information about state of tcp endpoint in API format. */
2444 void tcp_get_info(const struct sock *sk, struct tcp_info *info)
2445 {
2446 const struct tcp_sock *tp = tcp_sk(sk);
2447 const struct inet_connection_sock *icsk = inet_csk(sk);
2448 u32 now = tcp_time_stamp;
2449
2450 memset(info, 0, sizeof(*info));
2451
2452 info->tcpi_state = sk->sk_state;
2453 info->tcpi_ca_state = icsk->icsk_ca_state;
2454 info->tcpi_retransmits = icsk->icsk_retransmits;
2455 info->tcpi_probes = icsk->icsk_probes_out;
2456 info->tcpi_backoff = icsk->icsk_backoff;
2457
2458 if (tp->rx_opt.tstamp_ok)
2459 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2460 if (tcp_is_sack(tp))
2461 info->tcpi_options |= TCPI_OPT_SACK;
2462 if (tp->rx_opt.wscale_ok) {
2463 info->tcpi_options |= TCPI_OPT_WSCALE;
2464 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2465 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2466 }
2467
2468 if (tp->ecn_flags & TCP_ECN_OK)
2469 info->tcpi_options |= TCPI_OPT_ECN;
2470 if (tp->ecn_flags & TCP_ECN_SEEN)
2471 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2472
2473 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2474 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2475 info->tcpi_snd_mss = tp->mss_cache;
2476 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2477
2478 if (sk->sk_state == TCP_LISTEN) {
2479 info->tcpi_unacked = sk->sk_ack_backlog;
2480 info->tcpi_sacked = sk->sk_max_ack_backlog;
2481 } else {
2482 info->tcpi_unacked = tp->packets_out;
2483 info->tcpi_sacked = tp->sacked_out;
2484 }
2485 info->tcpi_lost = tp->lost_out;
2486 info->tcpi_retrans = tp->retrans_out;
2487 info->tcpi_fackets = tp->fackets_out;
2488
2489 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2490 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2491 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2492
2493 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2494 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2495 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2496 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2497 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2498 info->tcpi_snd_cwnd = tp->snd_cwnd;
2499 info->tcpi_advmss = tp->advmss;
2500 info->tcpi_reordering = tp->reordering;
2501
2502 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2503 info->tcpi_rcv_space = tp->rcvq_space.space;
2504
2505 info->tcpi_total_retrans = tp->total_retrans;
2506 }
2507 EXPORT_SYMBOL_GPL(tcp_get_info);
2508
2509 static int do_tcp_getsockopt(struct sock *sk, int level,
2510 int optname, char __user *optval, int __user *optlen)
2511 {
2512 struct inet_connection_sock *icsk = inet_csk(sk);
2513 struct tcp_sock *tp = tcp_sk(sk);
2514 int val, len;
2515
2516 if (get_user(len, optlen))
2517 return -EFAULT;
2518
2519 len = min_t(unsigned int, len, sizeof(int));
2520
2521 if (len < 0)
2522 return -EINVAL;
2523
2524 switch (optname) {
2525 case TCP_MAXSEG:
2526 val = tp->mss_cache;
2527 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2528 val = tp->rx_opt.user_mss;
2529 break;
2530 case TCP_NODELAY:
2531 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2532 break;
2533 case TCP_CORK:
2534 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2535 break;
2536 case TCP_KEEPIDLE:
2537 val = keepalive_time_when(tp) / HZ;
2538 break;
2539 case TCP_KEEPINTVL:
2540 val = keepalive_intvl_when(tp) / HZ;
2541 break;
2542 case TCP_KEEPCNT:
2543 val = keepalive_probes(tp);
2544 break;
2545 case TCP_SYNCNT:
2546 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2547 break;
2548 case TCP_LINGER2:
2549 val = tp->linger2;
2550 if (val >= 0)
2551 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2552 break;
2553 case TCP_DEFER_ACCEPT:
2554 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2555 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2556 break;
2557 case TCP_WINDOW_CLAMP:
2558 val = tp->window_clamp;
2559 break;
2560 case TCP_INFO: {
2561 struct tcp_info info;
2562
2563 if (get_user(len, optlen))
2564 return -EFAULT;
2565
2566 tcp_get_info(sk, &info);
2567
2568 len = min_t(unsigned int, len, sizeof(info));
2569 if (put_user(len, optlen))
2570 return -EFAULT;
2571 if (copy_to_user(optval, &info, len))
2572 return -EFAULT;
2573 return 0;
2574 }
2575 case TCP_QUICKACK:
2576 val = !icsk->icsk_ack.pingpong;
2577 break;
2578
2579 case TCP_CONGESTION:
2580 if (get_user(len, optlen))
2581 return -EFAULT;
2582 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2583 if (put_user(len, optlen))
2584 return -EFAULT;
2585 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2586 return -EFAULT;
2587 return 0;
2588
2589 case TCP_COOKIE_TRANSACTIONS: {
2590 struct tcp_cookie_transactions ctd;
2591 struct tcp_cookie_values *cvp = tp->cookie_values;
2592
2593 if (get_user(len, optlen))
2594 return -EFAULT;
2595 if (len < sizeof(ctd))
2596 return -EINVAL;
2597
2598 memset(&ctd, 0, sizeof(ctd));
2599 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2600 TCP_COOKIE_IN_ALWAYS : 0)
2601 | (tp->rx_opt.cookie_out_never ?
2602 TCP_COOKIE_OUT_NEVER : 0);
2603
2604 if (cvp != NULL) {
2605 ctd.tcpct_flags |= (cvp->s_data_in ?
2606 TCP_S_DATA_IN : 0)
2607 | (cvp->s_data_out ?
2608 TCP_S_DATA_OUT : 0);
2609
2610 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2611 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2612
2613 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2614 cvp->cookie_pair_size);
2615 ctd.tcpct_used = cvp->cookie_pair_size;
2616 }
2617
2618 if (put_user(sizeof(ctd), optlen))
2619 return -EFAULT;
2620 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2621 return -EFAULT;
2622 return 0;
2623 }
2624 case TCP_THIN_LINEAR_TIMEOUTS:
2625 val = tp->thin_lto;
2626 break;
2627 case TCP_THIN_DUPACK:
2628 val = tp->thin_dupack;
2629 break;
2630
2631 case TCP_USER_TIMEOUT:
2632 val = jiffies_to_msecs(icsk->icsk_user_timeout);
2633 break;
2634 default:
2635 return -ENOPROTOOPT;
2636 }
2637
2638 if (put_user(len, optlen))
2639 return -EFAULT;
2640 if (copy_to_user(optval, &val, len))
2641 return -EFAULT;
2642 return 0;
2643 }
2644
2645 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2646 int __user *optlen)
2647 {
2648 struct inet_connection_sock *icsk = inet_csk(sk);
2649
2650 if (level != SOL_TCP)
2651 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2652 optval, optlen);
2653 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2654 }
2655 EXPORT_SYMBOL(tcp_getsockopt);
2656
2657 #ifdef CONFIG_COMPAT
2658 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2659 char __user *optval, int __user *optlen)
2660 {
2661 if (level != SOL_TCP)
2662 return inet_csk_compat_getsockopt(sk, level, optname,
2663 optval, optlen);
2664 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2665 }
2666 EXPORT_SYMBOL(compat_tcp_getsockopt);
2667 #endif
2668
2669 struct sk_buff *tcp_tso_segment(struct sk_buff *skb,
2670 netdev_features_t features)
2671 {
2672 struct sk_buff *segs = ERR_PTR(-EINVAL);
2673 struct tcphdr *th;
2674 unsigned thlen;
2675 unsigned int seq;
2676 __be32 delta;
2677 unsigned int oldlen;
2678 unsigned int mss;
2679
2680 if (!pskb_may_pull(skb, sizeof(*th)))
2681 goto out;
2682
2683 th = tcp_hdr(skb);
2684 thlen = th->doff * 4;
2685 if (thlen < sizeof(*th))
2686 goto out;
2687
2688 if (!pskb_may_pull(skb, thlen))
2689 goto out;
2690
2691 oldlen = (u16)~skb->len;
2692 __skb_pull(skb, thlen);
2693
2694 mss = skb_shinfo(skb)->gso_size;
2695 if (unlikely(skb->len <= mss))
2696 goto out;
2697
2698 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2699 /* Packet is from an untrusted source, reset gso_segs. */
2700 int type = skb_shinfo(skb)->gso_type;
2701
2702 if (unlikely(type &
2703 ~(SKB_GSO_TCPV4 |
2704 SKB_GSO_DODGY |
2705 SKB_GSO_TCP_ECN |
2706 SKB_GSO_TCPV6 |
2707 0) ||
2708 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2709 goto out;
2710
2711 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2712
2713 segs = NULL;
2714 goto out;
2715 }
2716
2717 segs = skb_segment(skb, features);
2718 if (IS_ERR(segs))
2719 goto out;
2720
2721 delta = htonl(oldlen + (thlen + mss));
2722
2723 skb = segs;
2724 th = tcp_hdr(skb);
2725 seq = ntohl(th->seq);
2726
2727 do {
2728 th->fin = th->psh = 0;
2729
2730 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2731 (__force u32)delta));
2732 if (skb->ip_summed != CHECKSUM_PARTIAL)
2733 th->check =
2734 csum_fold(csum_partial(skb_transport_header(skb),
2735 thlen, skb->csum));
2736
2737 seq += mss;
2738 skb = skb->next;
2739 th = tcp_hdr(skb);
2740
2741 th->seq = htonl(seq);
2742 th->cwr = 0;
2743 } while (skb->next);
2744
2745 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2746 skb->data_len);
2747 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2748 (__force u32)delta));
2749 if (skb->ip_summed != CHECKSUM_PARTIAL)
2750 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2751 thlen, skb->csum));
2752
2753 out:
2754 return segs;
2755 }
2756 EXPORT_SYMBOL(tcp_tso_segment);
2757
2758 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2759 {
2760 struct sk_buff **pp = NULL;
2761 struct sk_buff *p;
2762 struct tcphdr *th;
2763 struct tcphdr *th2;
2764 unsigned int len;
2765 unsigned int thlen;
2766 __be32 flags;
2767 unsigned int mss = 1;
2768 unsigned int hlen;
2769 unsigned int off;
2770 int flush = 1;
2771 int i;
2772
2773 off = skb_gro_offset(skb);
2774 hlen = off + sizeof(*th);
2775 th = skb_gro_header_fast(skb, off);
2776 if (skb_gro_header_hard(skb, hlen)) {
2777 th = skb_gro_header_slow(skb, hlen, off);
2778 if (unlikely(!th))
2779 goto out;
2780 }
2781
2782 thlen = th->doff * 4;
2783 if (thlen < sizeof(*th))
2784 goto out;
2785
2786 hlen = off + thlen;
2787 if (skb_gro_header_hard(skb, hlen)) {
2788 th = skb_gro_header_slow(skb, hlen, off);
2789 if (unlikely(!th))
2790 goto out;
2791 }
2792
2793 skb_gro_pull(skb, thlen);
2794
2795 len = skb_gro_len(skb);
2796 flags = tcp_flag_word(th);
2797
2798 for (; (p = *head); head = &p->next) {
2799 if (!NAPI_GRO_CB(p)->same_flow)
2800 continue;
2801
2802 th2 = tcp_hdr(p);
2803
2804 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2805 NAPI_GRO_CB(p)->same_flow = 0;
2806 continue;
2807 }
2808
2809 goto found;
2810 }
2811
2812 goto out_check_final;
2813
2814 found:
2815 flush = NAPI_GRO_CB(p)->flush;
2816 flush |= (__force int)(flags & TCP_FLAG_CWR);
2817 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2818 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2819 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2820 for (i = sizeof(*th); i < thlen; i += 4)
2821 flush |= *(u32 *)((u8 *)th + i) ^
2822 *(u32 *)((u8 *)th2 + i);
2823
2824 mss = skb_shinfo(p)->gso_size;
2825
2826 flush |= (len - 1) >= mss;
2827 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2828
2829 if (flush || skb_gro_receive(head, skb)) {
2830 mss = 1;
2831 goto out_check_final;
2832 }
2833
2834 p = *head;
2835 th2 = tcp_hdr(p);
2836 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2837
2838 out_check_final:
2839 flush = len < mss;
2840 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2841 TCP_FLAG_RST | TCP_FLAG_SYN |
2842 TCP_FLAG_FIN));
2843
2844 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2845 pp = head;
2846
2847 out:
2848 NAPI_GRO_CB(skb)->flush |= flush;
2849
2850 return pp;
2851 }
2852 EXPORT_SYMBOL(tcp_gro_receive);
2853
2854 int tcp_gro_complete(struct sk_buff *skb)
2855 {
2856 struct tcphdr *th = tcp_hdr(skb);
2857
2858 skb->csum_start = skb_transport_header(skb) - skb->head;
2859 skb->csum_offset = offsetof(struct tcphdr, check);
2860 skb->ip_summed = CHECKSUM_PARTIAL;
2861
2862 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2863
2864 if (th->cwr)
2865 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2866
2867 return 0;
2868 }
2869 EXPORT_SYMBOL(tcp_gro_complete);
2870
2871 #ifdef CONFIG_TCP_MD5SIG
2872 static unsigned long tcp_md5sig_users;
2873 static struct tcp_md5sig_pool __percpu *tcp_md5sig_pool;
2874 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2875
2876 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool __percpu *pool)
2877 {
2878 int cpu;
2879
2880 for_each_possible_cpu(cpu) {
2881 struct tcp_md5sig_pool *p = per_cpu_ptr(pool, cpu);
2882
2883 if (p->md5_desc.tfm)
2884 crypto_free_hash(p->md5_desc.tfm);
2885 }
2886 free_percpu(pool);
2887 }
2888
2889 void tcp_free_md5sig_pool(void)
2890 {
2891 struct tcp_md5sig_pool __percpu *pool = NULL;
2892
2893 spin_lock_bh(&tcp_md5sig_pool_lock);
2894 if (--tcp_md5sig_users == 0) {
2895 pool = tcp_md5sig_pool;
2896 tcp_md5sig_pool = NULL;
2897 }
2898 spin_unlock_bh(&tcp_md5sig_pool_lock);
2899 if (pool)
2900 __tcp_free_md5sig_pool(pool);
2901 }
2902 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2903
2904 static struct tcp_md5sig_pool __percpu *
2905 __tcp_alloc_md5sig_pool(struct sock *sk)
2906 {
2907 int cpu;
2908 struct tcp_md5sig_pool __percpu *pool;
2909
2910 pool = alloc_percpu(struct tcp_md5sig_pool);
2911 if (!pool)
2912 return NULL;
2913
2914 for_each_possible_cpu(cpu) {
2915 struct crypto_hash *hash;
2916
2917 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2918 if (!hash || IS_ERR(hash))
2919 goto out_free;
2920
2921 per_cpu_ptr(pool, cpu)->md5_desc.tfm = hash;
2922 }
2923 return pool;
2924 out_free:
2925 __tcp_free_md5sig_pool(pool);
2926 return NULL;
2927 }
2928
2929 struct tcp_md5sig_pool __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2930 {
2931 struct tcp_md5sig_pool __percpu *pool;
2932 int alloc = 0;
2933
2934 retry:
2935 spin_lock_bh(&tcp_md5sig_pool_lock);
2936 pool = tcp_md5sig_pool;
2937 if (tcp_md5sig_users++ == 0) {
2938 alloc = 1;
2939 spin_unlock_bh(&tcp_md5sig_pool_lock);
2940 } else if (!pool) {
2941 tcp_md5sig_users--;
2942 spin_unlock_bh(&tcp_md5sig_pool_lock);
2943 cpu_relax();
2944 goto retry;
2945 } else
2946 spin_unlock_bh(&tcp_md5sig_pool_lock);
2947
2948 if (alloc) {
2949 /* we cannot hold spinlock here because this may sleep. */
2950 struct tcp_md5sig_pool __percpu *p;
2951
2952 p = __tcp_alloc_md5sig_pool(sk);
2953 spin_lock_bh(&tcp_md5sig_pool_lock);
2954 if (!p) {
2955 tcp_md5sig_users--;
2956 spin_unlock_bh(&tcp_md5sig_pool_lock);
2957 return NULL;
2958 }
2959 pool = tcp_md5sig_pool;
2960 if (pool) {
2961 /* oops, it has already been assigned. */
2962 spin_unlock_bh(&tcp_md5sig_pool_lock);
2963 __tcp_free_md5sig_pool(p);
2964 } else {
2965 tcp_md5sig_pool = pool = p;
2966 spin_unlock_bh(&tcp_md5sig_pool_lock);
2967 }
2968 }
2969 return pool;
2970 }
2971 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2972
2973
2974 /**
2975 * tcp_get_md5sig_pool - get md5sig_pool for this user
2976 *
2977 * We use percpu structure, so if we succeed, we exit with preemption
2978 * and BH disabled, to make sure another thread or softirq handling
2979 * wont try to get same context.
2980 */
2981 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2982 {
2983 struct tcp_md5sig_pool __percpu *p;
2984
2985 local_bh_disable();
2986
2987 spin_lock(&tcp_md5sig_pool_lock);
2988 p = tcp_md5sig_pool;
2989 if (p)
2990 tcp_md5sig_users++;
2991 spin_unlock(&tcp_md5sig_pool_lock);
2992
2993 if (p)
2994 return this_cpu_ptr(p);
2995
2996 local_bh_enable();
2997 return NULL;
2998 }
2999 EXPORT_SYMBOL(tcp_get_md5sig_pool);
3000
3001 void tcp_put_md5sig_pool(void)
3002 {
3003 local_bh_enable();
3004 tcp_free_md5sig_pool();
3005 }
3006 EXPORT_SYMBOL(tcp_put_md5sig_pool);
3007
3008 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
3009 const struct tcphdr *th)
3010 {
3011 struct scatterlist sg;
3012 struct tcphdr hdr;
3013 int err;
3014
3015 /* We are not allowed to change tcphdr, make a local copy */
3016 memcpy(&hdr, th, sizeof(hdr));
3017 hdr.check = 0;
3018
3019 /* options aren't included in the hash */
3020 sg_init_one(&sg, &hdr, sizeof(hdr));
3021 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(hdr));
3022 return err;
3023 }
3024 EXPORT_SYMBOL(tcp_md5_hash_header);
3025
3026 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3027 const struct sk_buff *skb, unsigned int header_len)
3028 {
3029 struct scatterlist sg;
3030 const struct tcphdr *tp = tcp_hdr(skb);
3031 struct hash_desc *desc = &hp->md5_desc;
3032 unsigned i;
3033 const unsigned head_data_len = skb_headlen(skb) > header_len ?
3034 skb_headlen(skb) - header_len : 0;
3035 const struct skb_shared_info *shi = skb_shinfo(skb);
3036 struct sk_buff *frag_iter;
3037
3038 sg_init_table(&sg, 1);
3039
3040 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3041 if (crypto_hash_update(desc, &sg, head_data_len))
3042 return 1;
3043
3044 for (i = 0; i < shi->nr_frags; ++i) {
3045 const struct skb_frag_struct *f = &shi->frags[i];
3046 struct page *page = skb_frag_page(f);
3047 sg_set_page(&sg, page, skb_frag_size(f), f->page_offset);
3048 if (crypto_hash_update(desc, &sg, skb_frag_size(f)))
3049 return 1;
3050 }
3051
3052 skb_walk_frags(skb, frag_iter)
3053 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3054 return 1;
3055
3056 return 0;
3057 }
3058 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3059
3060 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3061 {
3062 struct scatterlist sg;
3063
3064 sg_init_one(&sg, key->key, key->keylen);
3065 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3066 }
3067 EXPORT_SYMBOL(tcp_md5_hash_key);
3068
3069 #endif
3070
3071 /**
3072 * Each Responder maintains up to two secret values concurrently for
3073 * efficient secret rollover. Each secret value has 4 states:
3074 *
3075 * Generating. (tcp_secret_generating != tcp_secret_primary)
3076 * Generates new Responder-Cookies, but not yet used for primary
3077 * verification. This is a short-term state, typically lasting only
3078 * one round trip time (RTT).
3079 *
3080 * Primary. (tcp_secret_generating == tcp_secret_primary)
3081 * Used both for generation and primary verification.
3082 *
3083 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3084 * Used for verification, until the first failure that can be
3085 * verified by the newer Generating secret. At that time, this
3086 * cookie's state is changed to Secondary, and the Generating
3087 * cookie's state is changed to Primary. This is a short-term state,
3088 * typically lasting only one round trip time (RTT).
3089 *
3090 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3091 * Used for secondary verification, after primary verification
3092 * failures. This state lasts no more than twice the Maximum Segment
3093 * Lifetime (2MSL). Then, the secret is discarded.
3094 */
3095 struct tcp_cookie_secret {
3096 /* The secret is divided into two parts. The digest part is the
3097 * equivalent of previously hashing a secret and saving the state,
3098 * and serves as an initialization vector (IV). The message part
3099 * serves as the trailing secret.
3100 */
3101 u32 secrets[COOKIE_WORKSPACE_WORDS];
3102 unsigned long expires;
3103 };
3104
3105 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3106 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3107 #define TCP_SECRET_LIFE (HZ * 600)
3108
3109 static struct tcp_cookie_secret tcp_secret_one;
3110 static struct tcp_cookie_secret tcp_secret_two;
3111
3112 /* Essentially a circular list, without dynamic allocation. */
3113 static struct tcp_cookie_secret *tcp_secret_generating;
3114 static struct tcp_cookie_secret *tcp_secret_primary;
3115 static struct tcp_cookie_secret *tcp_secret_retiring;
3116 static struct tcp_cookie_secret *tcp_secret_secondary;
3117
3118 static DEFINE_SPINLOCK(tcp_secret_locker);
3119
3120 /* Select a pseudo-random word in the cookie workspace.
3121 */
3122 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3123 {
3124 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3125 }
3126
3127 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3128 * Called in softirq context.
3129 * Returns: 0 for success.
3130 */
3131 int tcp_cookie_generator(u32 *bakery)
3132 {
3133 unsigned long jiffy = jiffies;
3134
3135 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3136 spin_lock_bh(&tcp_secret_locker);
3137 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3138 /* refreshed by another */
3139 memcpy(bakery,
3140 &tcp_secret_generating->secrets[0],
3141 COOKIE_WORKSPACE_WORDS);
3142 } else {
3143 /* still needs refreshing */
3144 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3145
3146 /* The first time, paranoia assumes that the
3147 * randomization function isn't as strong. But,
3148 * this secret initialization is delayed until
3149 * the last possible moment (packet arrival).
3150 * Although that time is observable, it is
3151 * unpredictably variable. Mash in the most
3152 * volatile clock bits available, and expire the
3153 * secret extra quickly.
3154 */
3155 if (unlikely(tcp_secret_primary->expires ==
3156 tcp_secret_secondary->expires)) {
3157 struct timespec tv;
3158
3159 getnstimeofday(&tv);
3160 bakery[COOKIE_DIGEST_WORDS+0] ^=
3161 (u32)tv.tv_nsec;
3162
3163 tcp_secret_secondary->expires = jiffy
3164 + TCP_SECRET_1MSL
3165 + (0x0f & tcp_cookie_work(bakery, 0));
3166 } else {
3167 tcp_secret_secondary->expires = jiffy
3168 + TCP_SECRET_LIFE
3169 + (0xff & tcp_cookie_work(bakery, 1));
3170 tcp_secret_primary->expires = jiffy
3171 + TCP_SECRET_2MSL
3172 + (0x1f & tcp_cookie_work(bakery, 2));
3173 }
3174 memcpy(&tcp_secret_secondary->secrets[0],
3175 bakery, COOKIE_WORKSPACE_WORDS);
3176
3177 rcu_assign_pointer(tcp_secret_generating,
3178 tcp_secret_secondary);
3179 rcu_assign_pointer(tcp_secret_retiring,
3180 tcp_secret_primary);
3181 /*
3182 * Neither call_rcu() nor synchronize_rcu() needed.
3183 * Retiring data is not freed. It is replaced after
3184 * further (locked) pointer updates, and a quiet time
3185 * (minimum 1MSL, maximum LIFE - 2MSL).
3186 */
3187 }
3188 spin_unlock_bh(&tcp_secret_locker);
3189 } else {
3190 rcu_read_lock_bh();
3191 memcpy(bakery,
3192 &rcu_dereference(tcp_secret_generating)->secrets[0],
3193 COOKIE_WORKSPACE_WORDS);
3194 rcu_read_unlock_bh();
3195 }
3196 return 0;
3197 }
3198 EXPORT_SYMBOL(tcp_cookie_generator);
3199
3200 void tcp_done(struct sock *sk)
3201 {
3202 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3203 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3204
3205 tcp_set_state(sk, TCP_CLOSE);
3206 tcp_clear_xmit_timers(sk);
3207
3208 sk->sk_shutdown = SHUTDOWN_MASK;
3209
3210 if (!sock_flag(sk, SOCK_DEAD))
3211 sk->sk_state_change(sk);
3212 else
3213 inet_csk_destroy_sock(sk);
3214 }
3215 EXPORT_SYMBOL_GPL(tcp_done);
3216
3217 extern struct tcp_congestion_ops tcp_reno;
3218
3219 static __initdata unsigned long thash_entries;
3220 static int __init set_thash_entries(char *str)
3221 {
3222 if (!str)
3223 return 0;
3224 thash_entries = simple_strtoul(str, &str, 0);
3225 return 1;
3226 }
3227 __setup("thash_entries=", set_thash_entries);
3228
3229 void tcp_init_mem(struct net *net)
3230 {
3231 unsigned long limit = nr_free_buffer_pages() / 8;
3232 limit = max(limit, 128UL);
3233 net->ipv4.sysctl_tcp_mem[0] = limit / 4 * 3;
3234 net->ipv4.sysctl_tcp_mem[1] = limit;
3235 net->ipv4.sysctl_tcp_mem[2] = net->ipv4.sysctl_tcp_mem[0] * 2;
3236 }
3237
3238 void __init tcp_init(void)
3239 {
3240 struct sk_buff *skb = NULL;
3241 unsigned long limit;
3242 int max_rshare, max_wshare, cnt;
3243 unsigned int i;
3244 unsigned long jiffy = jiffies;
3245
3246 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3247
3248 percpu_counter_init(&tcp_sockets_allocated, 0);
3249 percpu_counter_init(&tcp_orphan_count, 0);
3250 tcp_hashinfo.bind_bucket_cachep =
3251 kmem_cache_create("tcp_bind_bucket",
3252 sizeof(struct inet_bind_bucket), 0,
3253 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3254
3255 /* Size and allocate the main established and bind bucket
3256 * hash tables.
3257 *
3258 * The methodology is similar to that of the buffer cache.
3259 */
3260 tcp_hashinfo.ehash =
3261 alloc_large_system_hash("TCP established",
3262 sizeof(struct inet_ehash_bucket),
3263 thash_entries,
3264 (totalram_pages >= 128 * 1024) ?
3265 13 : 15,
3266 0,
3267 NULL,
3268 &tcp_hashinfo.ehash_mask,
3269 thash_entries ? 0 : 512 * 1024);
3270 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3271 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3272 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3273 }
3274 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3275 panic("TCP: failed to alloc ehash_locks");
3276 tcp_hashinfo.bhash =
3277 alloc_large_system_hash("TCP bind",
3278 sizeof(struct inet_bind_hashbucket),
3279 tcp_hashinfo.ehash_mask + 1,
3280 (totalram_pages >= 128 * 1024) ?
3281 13 : 15,
3282 0,
3283 &tcp_hashinfo.bhash_size,
3284 NULL,
3285 64 * 1024);
3286 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3287 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3288 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3289 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3290 }
3291
3292
3293 cnt = tcp_hashinfo.ehash_mask + 1;
3294
3295 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3296 sysctl_tcp_max_orphans = cnt / 2;
3297 sysctl_max_syn_backlog = max(128, cnt / 256);
3298
3299 tcp_init_mem(&init_net);
3300 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3301 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3302 max_wshare = min(4UL*1024*1024, limit);
3303 max_rshare = min(6UL*1024*1024, limit);
3304
3305 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3306 sysctl_tcp_wmem[1] = 16*1024;
3307 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3308
3309 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3310 sysctl_tcp_rmem[1] = 87380;
3311 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3312
3313 printk(KERN_INFO "TCP: Hash tables configured "
3314 "(established %u bind %u)\n",
3315 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3316
3317 tcp_register_congestion_control(&tcp_reno);
3318
3319 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3320 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3321 tcp_secret_one.expires = jiffy; /* past due */
3322 tcp_secret_two.expires = jiffy; /* past due */
3323 tcp_secret_generating = &tcp_secret_one;
3324 tcp_secret_primary = &tcp_secret_one;
3325 tcp_secret_retiring = &tcp_secret_two;
3326 tcp_secret_secondary = &tcp_secret_two;
3327 }
Linux with added FPC-III support