projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/linville/wirel...
[deliverable/linux.git] / net / ipv4 / tcp.c
1 /*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 *
20 * Fixes:
21 * Alan Cox : Numerous verify_area() calls
22 * Alan Cox : Set the ACK bit on a reset
23 * Alan Cox : Stopped it crashing if it closed while
24 * sk->inuse=1 and was trying to connect
25 * (tcp_err()).
26 * Alan Cox : All icmp error handling was broken
27 * pointers passed where wrong and the
28 * socket was looked up backwards. Nobody
29 * tested any icmp error code obviously.
30 * Alan Cox : tcp_err() now handled properly. It
31 * wakes people on errors. poll
32 * behaves and the icmp error race
33 * has gone by moving it into sock.c
34 * Alan Cox : tcp_send_reset() fixed to work for
35 * everything not just packets for
36 * unknown sockets.
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
39 * syn rule wrong]
40 * Herp Rosmanith : More reset fixes
41 * Alan Cox : No longer acks invalid rst frames.
42 * Acking any kind of RST is right out.
43 * Alan Cox : Sets an ignore me flag on an rst
44 * receive otherwise odd bits of prattle
45 * escape still
46 * Alan Cox : Fixed another acking RST frame bug.
47 * Should stop LAN workplace lockups.
48 * Alan Cox : Some tidyups using the new skb list
49 * facilities
50 * Alan Cox : sk->keepopen now seems to work
51 * Alan Cox : Pulls options out correctly on accepts
52 * Alan Cox : Fixed assorted sk->rqueue->next errors
53 * Alan Cox : PSH doesn't end a TCP read. Switched a
54 * bit to skb ops.
55 * Alan Cox : Tidied tcp_data to avoid a potential
56 * nasty.
57 * Alan Cox : Added some better commenting, as the
58 * tcp is hard to follow
59 * Alan Cox : Removed incorrect check for 20 * psh
60 * Michael O'Reilly : ack < copied bug fix.
61 * Johannes Stille : Misc tcp fixes (not all in yet).
62 * Alan Cox : FIN with no memory -> CRASH
63 * Alan Cox : Added socket option proto entries.
64 * Also added awareness of them to accept.
65 * Alan Cox : Added TCP options (SOL_TCP)
66 * Alan Cox : Switched wakeup calls to callbacks,
67 * so the kernel can layer network
68 * sockets.
69 * Alan Cox : Use ip_tos/ip_ttl settings.
70 * Alan Cox : Handle FIN (more) properly (we hope).
71 * Alan Cox : RST frames sent on unsynchronised
72 * state ack error.
73 * Alan Cox : Put in missing check for SYN bit.
74 * Alan Cox : Added tcp_select_window() aka NET2E
75 * window non shrink trick.
76 * Alan Cox : Added a couple of small NET2E timer
77 * fixes
78 * Charles Hedrick : TCP fixes
79 * Toomas Tamm : TCP window fixes
80 * Alan Cox : Small URG fix to rlogin ^C ack fight
81 * Charles Hedrick : Rewrote most of it to actually work
82 * Linus : Rewrote tcp_read() and URG handling
83 * completely
84 * Gerhard Koerting: Fixed some missing timer handling
85 * Matthew Dillon : Reworked TCP machine states as per RFC
86 * Gerhard Koerting: PC/TCP workarounds
87 * Adam Caldwell : Assorted timer/timing errors
88 * Matthew Dillon : Fixed another RST bug
89 * Alan Cox : Move to kernel side addressing changes.
90 * Alan Cox : Beginning work on TCP fastpathing
91 * (not yet usable)
92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
93 * Alan Cox : TCP fast path debugging
94 * Alan Cox : Window clamping
95 * Michael Riepe : Bug in tcp_check()
96 * Matt Dillon : More TCP improvements and RST bug fixes
97 * Matt Dillon : Yet more small nasties remove from the
98 * TCP code (Be very nice to this man if
99 * tcp finally works 100%) 8)
100 * Alan Cox : BSD accept semantics.
101 * Alan Cox : Reset on closedown bug.
102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
103 * Michael Pall : Handle poll() after URG properly in
104 * all cases.
105 * Michael Pall : Undo the last fix in tcp_read_urg()
106 * (multi URG PUSH broke rlogin).
107 * Michael Pall : Fix the multi URG PUSH problem in
108 * tcp_readable(), poll() after URG
109 * works now.
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
111 * BSD api.
112 * Alan Cox : Changed the semantics of sk->socket to
113 * fix a race and a signal problem with
114 * accept() and async I/O.
115 * Alan Cox : Relaxed the rules on tcp_sendto().
116 * Yury Shevchuk : Really fixed accept() blocking problem.
117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
118 * clients/servers which listen in on
119 * fixed ports.
120 * Alan Cox : Cleaned the above up and shrank it to
121 * a sensible code size.
122 * Alan Cox : Self connect lockup fix.
123 * Alan Cox : No connect to multicast.
124 * Ross Biro : Close unaccepted children on master
125 * socket close.
126 * Alan Cox : Reset tracing code.
127 * Alan Cox : Spurious resets on shutdown.
128 * Alan Cox : Giant 15 minute/60 second timer error
129 * Alan Cox : Small whoops in polling before an
130 * accept.
131 * Alan Cox : Kept the state trace facility since
132 * it's handy for debugging.
133 * Alan Cox : More reset handler fixes.
134 * Alan Cox : Started rewriting the code based on
135 * the RFC's for other useful protocol
136 * references see: Comer, KA9Q NOS, and
137 * for a reference on the difference
138 * between specifications and how BSD
139 * works see the 4.4lite source.
140 * A.N.Kuznetsov : Don't time wait on completion of tidy
141 * close.
142 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
143 * Linus Torvalds : Fixed BSD port reuse to work first syn
144 * Alan Cox : Reimplemented timers as per the RFC
145 * and using multiple timers for sanity.
146 * Alan Cox : Small bug fixes, and a lot of new
147 * comments.
148 * Alan Cox : Fixed dual reader crash by locking
149 * the buffers (much like datagram.c)
150 * Alan Cox : Fixed stuck sockets in probe. A probe
151 * now gets fed up of retrying without
152 * (even a no space) answer.
153 * Alan Cox : Extracted closing code better
154 * Alan Cox : Fixed the closing state machine to
155 * resemble the RFC.
156 * Alan Cox : More 'per spec' fixes.
157 * Jorge Cwik : Even faster checksumming.
158 * Alan Cox : tcp_data() doesn't ack illegal PSH
159 * only frames. At least one pc tcp stack
160 * generates them.
161 * Alan Cox : Cache last socket.
162 * Alan Cox : Per route irtt.
163 * Matt Day : poll()->select() match BSD precisely on error
164 * Alan Cox : New buffers
165 * Marc Tamsky : Various sk->prot->retransmits and
166 * sk->retransmits misupdating fixed.
167 * Fixed tcp_write_timeout: stuck close,
168 * and TCP syn retries gets used now.
169 * Mark Yarvis : In tcp_read_wakeup(), don't send an
170 * ack if state is TCP_CLOSED.
171 * Alan Cox : Look up device on a retransmit - routes may
172 * change. Doesn't yet cope with MSS shrink right
173 * but it's a start!
174 * Marc Tamsky : Closing in closing fixes.
175 * Mike Shaver : RFC1122 verifications.
176 * Alan Cox : rcv_saddr errors.
177 * Alan Cox : Block double connect().
178 * Alan Cox : Small hooks for enSKIP.
179 * Alexey Kuznetsov: Path MTU discovery.
180 * Alan Cox : Support soft errors.
181 * Alan Cox : Fix MTU discovery pathological case
182 * when the remote claims no mtu!
183 * Marc Tamsky : TCP_CLOSE fix.
184 * Colin (G3TNE) : Send a reset on syn ack replies in
185 * window but wrong (fixes NT lpd problems)
186 * Pedro Roque : Better TCP window handling, delayed ack.
187 * Joerg Reuter : No modification of locked buffers in
188 * tcp_do_retransmit()
189 * Eric Schenk : Changed receiver side silly window
190 * avoidance algorithm to BSD style
191 * algorithm. This doubles throughput
192 * against machines running Solaris,
193 * and seems to result in general
194 * improvement.
195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
196 * Willy Konynenberg : Transparent proxying support.
197 * Mike McLagan : Routing by source
198 * Keith Owens : Do proper merging with partial SKB's in
199 * tcp_do_sendmsg to avoid burstiness.
200 * Eric Schenk : Fix fast close down bug with
201 * shutdown() followed by close().
202 * Andi Kleen : Make poll agree with SIGIO
203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
204 * lingertime == 0 (RFC 793 ABORT Call)
205 * Hirokazu Takahashi : Use copy_from_user() instead of
206 * csum_and_copy_from_user() if possible.
207 *
208 * This program is free software; you can redistribute it and/or
209 * modify it under the terms of the GNU General Public License
210 * as published by the Free Software Foundation; either version
211 * 2 of the License, or(at your option) any later version.
212 *
213 * Description of States:
214 *
215 * TCP_SYN_SENT sent a connection request, waiting for ack
216 *
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
219 *
220 * TCP_ESTABLISHED connection established
221 *
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
224 *
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
226 * to shutdown
227 *
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
230 *
231 * TCP_TIME_WAIT timeout to catch resent junk before entering
232 * closed, can only be entered from FIN_WAIT2
233 * or CLOSING. Required because the other end
234 * may not have gotten our last ACK causing it
235 * to retransmit the data packet (which we ignore)
236 *
237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
238 * us to finish writing our data and to shutdown
239 * (we have to close() to move on to LAST_ACK)
240 *
241 * TCP_LAST_ACK out side has shutdown after remote has
242 * shutdown. There may still be data in our
243 * buffer that we have to finish sending
244 *
245 * TCP_CLOSE socket is finished
246 */
247
248 #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_mem[3] __read_mostly;
286 int sysctl_tcp_wmem[3] __read_mostly;
287 int sysctl_tcp_rmem[3] __read_mostly;
288
289 EXPORT_SYMBOL(sysctl_tcp_mem);
290 EXPORT_SYMBOL(sysctl_tcp_rmem);
291 EXPORT_SYMBOL(sysctl_tcp_wmem);
292
293 atomic_t tcp_memory_allocated; /* Current allocated memory. */
294 EXPORT_SYMBOL(tcp_memory_allocated);
295
296 /*
297 * Current number of TCP sockets.
298 */
299 struct percpu_counter tcp_sockets_allocated;
300 EXPORT_SYMBOL(tcp_sockets_allocated);
301
302 /*
303 * TCP splice context
304 */
305 struct tcp_splice_state {
306 struct pipe_inode_info *pipe;
307 size_t len;
308 unsigned int flags;
309 };
310
311 /*
312 * Pressure flag: try to collapse.
313 * Technical note: it is used by multiple contexts non atomically.
314 * All the __sk_mem_schedule() is of this nature: accounting
315 * is strict, actions are advisory and have some latency.
316 */
317 int tcp_memory_pressure __read_mostly;
318 EXPORT_SYMBOL(tcp_memory_pressure);
319
320 void tcp_enter_memory_pressure(struct sock *sk)
321 {
322 if (!tcp_memory_pressure) {
323 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
324 tcp_memory_pressure = 1;
325 }
326 }
327 EXPORT_SYMBOL(tcp_enter_memory_pressure);
328
329 /* Convert seconds to retransmits based on initial and max timeout */
330 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
331 {
332 u8 res = 0;
333
334 if (seconds > 0) {
335 int period = timeout;
336
337 res = 1;
338 while (seconds > period && res < 255) {
339 res++;
340 timeout <<= 1;
341 if (timeout > rto_max)
342 timeout = rto_max;
343 period += timeout;
344 }
345 }
346 return res;
347 }
348
349 /* Convert retransmits to seconds based on initial and max timeout */
350 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
351 {
352 int period = 0;
353
354 if (retrans > 0) {
355 period = timeout;
356 while (--retrans) {
357 timeout <<= 1;
358 if (timeout > rto_max)
359 timeout = rto_max;
360 period += timeout;
361 }
362 }
363 return period;
364 }
365
366 /*
367 * Wait for a TCP event.
368 *
369 * Note that we don't need to lock the socket, as the upper poll layers
370 * take care of normal races (between the test and the event) and we don't
371 * go look at any of the socket buffers directly.
372 */
373 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
374 {
375 unsigned int mask;
376 struct sock *sk = sock->sk;
377 struct tcp_sock *tp = tcp_sk(sk);
378
379 sock_poll_wait(file, sk_sleep(sk), wait);
380 if (sk->sk_state == TCP_LISTEN)
381 return inet_csk_listen_poll(sk);
382
383 /* Socket is not locked. We are protected from async events
384 * by poll logic and correct handling of state changes
385 * made by other threads is impossible in any case.
386 */
387
388 mask = 0;
389 if (sk->sk_err)
390 mask = POLLERR;
391
392 /*
393 * POLLHUP is certainly not done right. But poll() doesn't
394 * have a notion of HUP in just one direction, and for a
395 * socket the read side is more interesting.
396 *
397 * Some poll() documentation says that POLLHUP is incompatible
398 * with the POLLOUT/POLLWR flags, so somebody should check this
399 * all. But careful, it tends to be safer to return too many
400 * bits than too few, and you can easily break real applications
401 * if you don't tell them that something has hung up!
402 *
403 * Check-me.
404 *
405 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
406 * our fs/select.c). It means that after we received EOF,
407 * poll always returns immediately, making impossible poll() on write()
408 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
409 * if and only if shutdown has been made in both directions.
410 * Actually, it is interesting to look how Solaris and DUX
411 * solve this dilemma. I would prefer, if POLLHUP were maskable,
412 * then we could set it on SND_SHUTDOWN. BTW examples given
413 * in Stevens' books assume exactly this behaviour, it explains
414 * why POLLHUP is incompatible with POLLOUT. --ANK
415 *
416 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
417 * blocking on fresh not-connected or disconnected socket. --ANK
418 */
419 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
420 mask |= POLLHUP;
421 if (sk->sk_shutdown & RCV_SHUTDOWN)
422 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
423
424 /* Connected? */
425 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
426 int target = sock_rcvlowat(sk, 0, INT_MAX);
427
428 if (tp->urg_seq == tp->copied_seq &&
429 !sock_flag(sk, SOCK_URGINLINE) &&
430 tp->urg_data)
431 target++;
432
433 /* Potential race condition. If read of tp below will
434 * escape above sk->sk_state, we can be illegally awaken
435 * in SYN_* states. */
436 if (tp->rcv_nxt - tp->copied_seq >= target)
437 mask |= POLLIN | POLLRDNORM;
438
439 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
440 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
441 mask |= POLLOUT | POLLWRNORM;
442 } else { /* send SIGIO later */
443 set_bit(SOCK_ASYNC_NOSPACE,
444 &sk->sk_socket->flags);
445 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
446
447 /* Race breaker. If space is freed after
448 * wspace test but before the flags are set,
449 * IO signal will be lost.
450 */
451 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
452 mask |= POLLOUT | POLLWRNORM;
453 }
454 }
455
456 if (tp->urg_data & TCP_URG_VALID)
457 mask |= POLLPRI;
458 }
459 return mask;
460 }
461 EXPORT_SYMBOL(tcp_poll);
462
463 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
464 {
465 struct tcp_sock *tp = tcp_sk(sk);
466 int answ;
467
468 switch (cmd) {
469 case SIOCINQ:
470 if (sk->sk_state == TCP_LISTEN)
471 return -EINVAL;
472
473 lock_sock(sk);
474 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
475 answ = 0;
476 else if (sock_flag(sk, SOCK_URGINLINE) ||
477 !tp->urg_data ||
478 before(tp->urg_seq, tp->copied_seq) ||
479 !before(tp->urg_seq, tp->rcv_nxt)) {
480 struct sk_buff *skb;
481
482 answ = tp->rcv_nxt - tp->copied_seq;
483
484 /* Subtract 1, if FIN is in queue. */
485 skb = skb_peek_tail(&sk->sk_receive_queue);
486 if (answ && skb)
487 answ -= tcp_hdr(skb)->fin;
488 } else
489 answ = tp->urg_seq - tp->copied_seq;
490 release_sock(sk);
491 break;
492 case SIOCATMARK:
493 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
494 break;
495 case SIOCOUTQ:
496 if (sk->sk_state == TCP_LISTEN)
497 return -EINVAL;
498
499 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
500 answ = 0;
501 else
502 answ = tp->write_seq - tp->snd_una;
503 break;
504 default:
505 return -ENOIOCTLCMD;
506 }
507
508 return put_user(answ, (int __user *)arg);
509 }
510 EXPORT_SYMBOL(tcp_ioctl);
511
512 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
513 {
514 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
515 tp->pushed_seq = tp->write_seq;
516 }
517
518 static inline int forced_push(struct tcp_sock *tp)
519 {
520 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
521 }
522
523 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
524 {
525 struct tcp_sock *tp = tcp_sk(sk);
526 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
527
528 skb->csum = 0;
529 tcb->seq = tcb->end_seq = tp->write_seq;
530 tcb->flags = TCPHDR_ACK;
531 tcb->sacked = 0;
532 skb_header_release(skb);
533 tcp_add_write_queue_tail(sk, skb);
534 sk->sk_wmem_queued += skb->truesize;
535 sk_mem_charge(sk, skb->truesize);
536 if (tp->nonagle & TCP_NAGLE_PUSH)
537 tp->nonagle &= ~TCP_NAGLE_PUSH;
538 }
539
540 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
541 {
542 if (flags & MSG_OOB)
543 tp->snd_up = tp->write_seq;
544 }
545
546 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
547 int nonagle)
548 {
549 if (tcp_send_head(sk)) {
550 struct tcp_sock *tp = tcp_sk(sk);
551
552 if (!(flags & MSG_MORE) || forced_push(tp))
553 tcp_mark_push(tp, tcp_write_queue_tail(sk));
554
555 tcp_mark_urg(tp, flags);
556 __tcp_push_pending_frames(sk, mss_now,
557 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
558 }
559 }
560
561 static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
562 unsigned int offset, size_t len)
563 {
564 struct tcp_splice_state *tss = rd_desc->arg.data;
565 int ret;
566
567 ret = skb_splice_bits(skb, offset, tss->pipe, min(rd_desc->count, len),
568 tss->flags);
569 if (ret > 0)
570 rd_desc->count -= ret;
571 return ret;
572 }
573
574 static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
575 {
576 /* Store TCP splice context information in read_descriptor_t. */
577 read_descriptor_t rd_desc = {
578 .arg.data = tss,
579 .count = tss->len,
580 };
581
582 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
583 }
584
585 /**
586 * tcp_splice_read - splice data from TCP socket to a pipe
587 * @sock: socket to splice from
588 * @ppos: position (not valid)
589 * @pipe: pipe to splice to
590 * @len: number of bytes to splice
591 * @flags: splice modifier flags
592 *
593 * Description:
594 * Will read pages from given socket and fill them into a pipe.
595 *
596 **/
597 ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
598 struct pipe_inode_info *pipe, size_t len,
599 unsigned int flags)
600 {
601 struct sock *sk = sock->sk;
602 struct tcp_splice_state tss = {
603 .pipe = pipe,
604 .len = len,
605 .flags = flags,
606 };
607 long timeo;
608 ssize_t spliced;
609 int ret;
610
611 /*
612 * We can't seek on a socket input
613 */
614 if (unlikely(*ppos))
615 return -ESPIPE;
616
617 ret = spliced = 0;
618
619 lock_sock(sk);
620
621 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
622 while (tss.len) {
623 ret = __tcp_splice_read(sk, &tss);
624 if (ret < 0)
625 break;
626 else if (!ret) {
627 if (spliced)
628 break;
629 if (sock_flag(sk, SOCK_DONE))
630 break;
631 if (sk->sk_err) {
632 ret = sock_error(sk);
633 break;
634 }
635 if (sk->sk_shutdown & RCV_SHUTDOWN)
636 break;
637 if (sk->sk_state == TCP_CLOSE) {
638 /*
639 * This occurs when user tries to read
640 * from never connected socket.
641 */
642 if (!sock_flag(sk, SOCK_DONE))
643 ret = -ENOTCONN;
644 break;
645 }
646 if (!timeo) {
647 ret = -EAGAIN;
648 break;
649 }
650 sk_wait_data(sk, &timeo);
651 if (signal_pending(current)) {
652 ret = sock_intr_errno(timeo);
653 break;
654 }
655 continue;
656 }
657 tss.len -= ret;
658 spliced += ret;
659
660 if (!timeo)
661 break;
662 release_sock(sk);
663 lock_sock(sk);
664
665 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
666 (sk->sk_shutdown & RCV_SHUTDOWN) ||
667 signal_pending(current))
668 break;
669 }
670
671 release_sock(sk);
672
673 if (spliced)
674 return spliced;
675
676 return ret;
677 }
678 EXPORT_SYMBOL(tcp_splice_read);
679
680 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
681 {
682 struct sk_buff *skb;
683
684 /* The TCP header must be at least 32-bit aligned. */
685 size = ALIGN(size, 4);
686
687 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
688 if (skb) {
689 if (sk_wmem_schedule(sk, skb->truesize)) {
690 /*
691 * Make sure that we have exactly size bytes
692 * available to the caller, no more, no less.
693 */
694 skb_reserve(skb, skb_tailroom(skb) - size);
695 return skb;
696 }
697 __kfree_skb(skb);
698 } else {
699 sk->sk_prot->enter_memory_pressure(sk);
700 sk_stream_moderate_sndbuf(sk);
701 }
702 return NULL;
703 }
704
705 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
706 int large_allowed)
707 {
708 struct tcp_sock *tp = tcp_sk(sk);
709 u32 xmit_size_goal, old_size_goal;
710
711 xmit_size_goal = mss_now;
712
713 if (large_allowed && sk_can_gso(sk)) {
714 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
715 inet_csk(sk)->icsk_af_ops->net_header_len -
716 inet_csk(sk)->icsk_ext_hdr_len -
717 tp->tcp_header_len);
718
719 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
720
721 /* We try hard to avoid divides here */
722 old_size_goal = tp->xmit_size_goal_segs * mss_now;
723
724 if (likely(old_size_goal <= xmit_size_goal &&
725 old_size_goal + mss_now > xmit_size_goal)) {
726 xmit_size_goal = old_size_goal;
727 } else {
728 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
729 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
730 }
731 }
732
733 return max(xmit_size_goal, mss_now);
734 }
735
736 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
737 {
738 int mss_now;
739
740 mss_now = tcp_current_mss(sk);
741 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
742
743 return mss_now;
744 }
745
746 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
747 size_t psize, int flags)
748 {
749 struct tcp_sock *tp = tcp_sk(sk);
750 int mss_now, size_goal;
751 int err;
752 ssize_t copied;
753 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
754
755 /* Wait for a connection to finish. */
756 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
757 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
758 goto out_err;
759
760 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
761
762 mss_now = tcp_send_mss(sk, &size_goal, flags);
763 copied = 0;
764
765 err = -EPIPE;
766 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
767 goto out_err;
768
769 while (psize > 0) {
770 struct sk_buff *skb = tcp_write_queue_tail(sk);
771 struct page *page = pages[poffset / PAGE_SIZE];
772 int copy, i, can_coalesce;
773 int offset = poffset % PAGE_SIZE;
774 int size = min_t(size_t, psize, PAGE_SIZE - offset);
775
776 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
777 new_segment:
778 if (!sk_stream_memory_free(sk))
779 goto wait_for_sndbuf;
780
781 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
782 if (!skb)
783 goto wait_for_memory;
784
785 skb_entail(sk, skb);
786 copy = size_goal;
787 }
788
789 if (copy > size)
790 copy = size;
791
792 i = skb_shinfo(skb)->nr_frags;
793 can_coalesce = skb_can_coalesce(skb, i, page, offset);
794 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
795 tcp_mark_push(tp, skb);
796 goto new_segment;
797 }
798 if (!sk_wmem_schedule(sk, copy))
799 goto wait_for_memory;
800
801 if (can_coalesce) {
802 skb_shinfo(skb)->frags[i - 1].size += copy;
803 } else {
804 get_page(page);
805 skb_fill_page_desc(skb, i, page, offset, copy);
806 }
807
808 skb->len += copy;
809 skb->data_len += copy;
810 skb->truesize += copy;
811 sk->sk_wmem_queued += copy;
812 sk_mem_charge(sk, copy);
813 skb->ip_summed = CHECKSUM_PARTIAL;
814 tp->write_seq += copy;
815 TCP_SKB_CB(skb)->end_seq += copy;
816 skb_shinfo(skb)->gso_segs = 0;
817
818 if (!copied)
819 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
820
821 copied += copy;
822 poffset += copy;
823 if (!(psize -= copy))
824 goto out;
825
826 if (skb->len < size_goal || (flags & MSG_OOB))
827 continue;
828
829 if (forced_push(tp)) {
830 tcp_mark_push(tp, skb);
831 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
832 } else if (skb == tcp_send_head(sk))
833 tcp_push_one(sk, mss_now);
834 continue;
835
836 wait_for_sndbuf:
837 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
838 wait_for_memory:
839 if (copied)
840 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
841
842 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
843 goto do_error;
844
845 mss_now = tcp_send_mss(sk, &size_goal, flags);
846 }
847
848 out:
849 if (copied)
850 tcp_push(sk, flags, mss_now, tp->nonagle);
851 return copied;
852
853 do_error:
854 if (copied)
855 goto out;
856 out_err:
857 return sk_stream_error(sk, flags, err);
858 }
859
860 int tcp_sendpage(struct sock *sk, struct page *page, int offset,
861 size_t size, int flags)
862 {
863 ssize_t res;
864
865 if (!(sk->sk_route_caps & NETIF_F_SG) ||
866 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
867 return sock_no_sendpage(sk->sk_socket, page, offset, size,
868 flags);
869
870 lock_sock(sk);
871 TCP_CHECK_TIMER(sk);
872 res = do_tcp_sendpages(sk, &page, offset, size, flags);
873 TCP_CHECK_TIMER(sk);
874 release_sock(sk);
875 return res;
876 }
877 EXPORT_SYMBOL(tcp_sendpage);
878
879 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
880 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
881
882 static inline int select_size(struct sock *sk, int sg)
883 {
884 struct tcp_sock *tp = tcp_sk(sk);
885 int tmp = tp->mss_cache;
886
887 if (sg) {
888 if (sk_can_gso(sk))
889 tmp = 0;
890 else {
891 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
892
893 if (tmp >= pgbreak &&
894 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
895 tmp = pgbreak;
896 }
897 }
898
899 return tmp;
900 }
901
902 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
903 size_t size)
904 {
905 struct iovec *iov;
906 struct tcp_sock *tp = tcp_sk(sk);
907 struct sk_buff *skb;
908 int iovlen, flags;
909 int mss_now, size_goal;
910 int sg, err, copied;
911 long timeo;
912
913 lock_sock(sk);
914 TCP_CHECK_TIMER(sk);
915
916 flags = msg->msg_flags;
917 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
918
919 /* Wait for a connection to finish. */
920 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
921 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
922 goto out_err;
923
924 /* This should be in poll */
925 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
926
927 mss_now = tcp_send_mss(sk, &size_goal, flags);
928
929 /* Ok commence sending. */
930 iovlen = msg->msg_iovlen;
931 iov = msg->msg_iov;
932 copied = 0;
933
934 err = -EPIPE;
935 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
936 goto out_err;
937
938 sg = sk->sk_route_caps & NETIF_F_SG;
939
940 while (--iovlen >= 0) {
941 int seglen = iov->iov_len;
942 unsigned char __user *from = iov->iov_base;
943
944 iov++;
945
946 while (seglen > 0) {
947 int copy = 0;
948 int max = size_goal;
949
950 skb = tcp_write_queue_tail(sk);
951 if (tcp_send_head(sk)) {
952 if (skb->ip_summed == CHECKSUM_NONE)
953 max = mss_now;
954 copy = max - skb->len;
955 }
956
957 if (copy <= 0) {
958 new_segment:
959 /* Allocate new segment. If the interface is SG,
960 * allocate skb fitting to single page.
961 */
962 if (!sk_stream_memory_free(sk))
963 goto wait_for_sndbuf;
964
965 skb = sk_stream_alloc_skb(sk,
966 select_size(sk, sg),
967 sk->sk_allocation);
968 if (!skb)
969 goto wait_for_memory;
970
971 /*
972 * Check whether we can use HW checksum.
973 */
974 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
975 skb->ip_summed = CHECKSUM_PARTIAL;
976
977 skb_entail(sk, skb);
978 copy = size_goal;
979 max = size_goal;
980 }
981
982 /* Try to append data to the end of skb. */
983 if (copy > seglen)
984 copy = seglen;
985
986 /* Where to copy to? */
987 if (skb_tailroom(skb) > 0) {
988 /* We have some space in skb head. Superb! */
989 if (copy > skb_tailroom(skb))
990 copy = skb_tailroom(skb);
991 if ((err = skb_add_data(skb, from, copy)) != 0)
992 goto do_fault;
993 } else {
994 int merge = 0;
995 int i = skb_shinfo(skb)->nr_frags;
996 struct page *page = TCP_PAGE(sk);
997 int off = TCP_OFF(sk);
998
999 if (skb_can_coalesce(skb, i, page, off) &&
1000 off != PAGE_SIZE) {
1001 /* We can extend the last page
1002 * fragment. */
1003 merge = 1;
1004 } else if (i == MAX_SKB_FRAGS || !sg) {
1005 /* Need to add new fragment and cannot
1006 * do this because interface is non-SG,
1007 * or because all the page slots are
1008 * busy. */
1009 tcp_mark_push(tp, skb);
1010 goto new_segment;
1011 } else if (page) {
1012 if (off == PAGE_SIZE) {
1013 put_page(page);
1014 TCP_PAGE(sk) = page = NULL;
1015 off = 0;
1016 }
1017 } else
1018 off = 0;
1019
1020 if (copy > PAGE_SIZE - off)
1021 copy = PAGE_SIZE - off;
1022
1023 if (!sk_wmem_schedule(sk, copy))
1024 goto wait_for_memory;
1025
1026 if (!page) {
1027 /* Allocate new cache page. */
1028 if (!(page = sk_stream_alloc_page(sk)))
1029 goto wait_for_memory;
1030 }
1031
1032 /* Time to copy data. We are close to
1033 * the end! */
1034 err = skb_copy_to_page(sk, from, skb, page,
1035 off, copy);
1036 if (err) {
1037 /* If this page was new, give it to the
1038 * socket so it does not get leaked.
1039 */
1040 if (!TCP_PAGE(sk)) {
1041 TCP_PAGE(sk) = page;
1042 TCP_OFF(sk) = 0;
1043 }
1044 goto do_error;
1045 }
1046
1047 /* Update the skb. */
1048 if (merge) {
1049 skb_shinfo(skb)->frags[i - 1].size +=
1050 copy;
1051 } else {
1052 skb_fill_page_desc(skb, i, page, off, copy);
1053 if (TCP_PAGE(sk)) {
1054 get_page(page);
1055 } else if (off + copy < PAGE_SIZE) {
1056 get_page(page);
1057 TCP_PAGE(sk) = page;
1058 }
1059 }
1060
1061 TCP_OFF(sk) = off + copy;
1062 }
1063
1064 if (!copied)
1065 TCP_SKB_CB(skb)->flags &= ~TCPHDR_PSH;
1066
1067 tp->write_seq += copy;
1068 TCP_SKB_CB(skb)->end_seq += copy;
1069 skb_shinfo(skb)->gso_segs = 0;
1070
1071 from += copy;
1072 copied += copy;
1073 if ((seglen -= copy) == 0 && iovlen == 0)
1074 goto out;
1075
1076 if (skb->len < max || (flags & MSG_OOB))
1077 continue;
1078
1079 if (forced_push(tp)) {
1080 tcp_mark_push(tp, skb);
1081 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1082 } else if (skb == tcp_send_head(sk))
1083 tcp_push_one(sk, mss_now);
1084 continue;
1085
1086 wait_for_sndbuf:
1087 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1088 wait_for_memory:
1089 if (copied)
1090 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1091
1092 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1093 goto do_error;
1094
1095 mss_now = tcp_send_mss(sk, &size_goal, flags);
1096 }
1097 }
1098
1099 out:
1100 if (copied)
1101 tcp_push(sk, flags, mss_now, tp->nonagle);
1102 TCP_CHECK_TIMER(sk);
1103 release_sock(sk);
1104 return copied;
1105
1106 do_fault:
1107 if (!skb->len) {
1108 tcp_unlink_write_queue(skb, sk);
1109 /* It is the one place in all of TCP, except connection
1110 * reset, where we can be unlinking the send_head.
1111 */
1112 tcp_check_send_head(sk, skb);
1113 sk_wmem_free_skb(sk, skb);
1114 }
1115
1116 do_error:
1117 if (copied)
1118 goto out;
1119 out_err:
1120 err = sk_stream_error(sk, flags, err);
1121 TCP_CHECK_TIMER(sk);
1122 release_sock(sk);
1123 return err;
1124 }
1125 EXPORT_SYMBOL(tcp_sendmsg);
1126
1127 /*
1128 * Handle reading urgent data. BSD has very simple semantics for
1129 * this, no blocking and very strange errors 8)
1130 */
1131
1132 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1133 {
1134 struct tcp_sock *tp = tcp_sk(sk);
1135
1136 /* No URG data to read. */
1137 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1138 tp->urg_data == TCP_URG_READ)
1139 return -EINVAL; /* Yes this is right ! */
1140
1141 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1142 return -ENOTCONN;
1143
1144 if (tp->urg_data & TCP_URG_VALID) {
1145 int err = 0;
1146 char c = tp->urg_data;
1147
1148 if (!(flags & MSG_PEEK))
1149 tp->urg_data = TCP_URG_READ;
1150
1151 /* Read urgent data. */
1152 msg->msg_flags |= MSG_OOB;
1153
1154 if (len > 0) {
1155 if (!(flags & MSG_TRUNC))
1156 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1157 len = 1;
1158 } else
1159 msg->msg_flags |= MSG_TRUNC;
1160
1161 return err ? -EFAULT : len;
1162 }
1163
1164 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1165 return 0;
1166
1167 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1168 * the available implementations agree in this case:
1169 * this call should never block, independent of the
1170 * blocking state of the socket.
1171 * Mike <pall@rz.uni-karlsruhe.de>
1172 */
1173 return -EAGAIN;
1174 }
1175
1176 /* Clean up the receive buffer for full frames taken by the user,
1177 * then send an ACK if necessary. COPIED is the number of bytes
1178 * tcp_recvmsg has given to the user so far, it speeds up the
1179 * calculation of whether or not we must ACK for the sake of
1180 * a window update.
1181 */
1182 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1183 {
1184 struct tcp_sock *tp = tcp_sk(sk);
1185 int time_to_ack = 0;
1186
1187 #if TCP_DEBUG
1188 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1189
1190 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1191 KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1192 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1193 #endif
1194
1195 if (inet_csk_ack_scheduled(sk)) {
1196 const struct inet_connection_sock *icsk = inet_csk(sk);
1197 /* Delayed ACKs frequently hit locked sockets during bulk
1198 * receive. */
1199 if (icsk->icsk_ack.blocked ||
1200 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1201 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1202 /*
1203 * If this read emptied read buffer, we send ACK, if
1204 * connection is not bidirectional, user drained
1205 * receive buffer and there was a small segment
1206 * in queue.
1207 */
1208 (copied > 0 &&
1209 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1210 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1211 !icsk->icsk_ack.pingpong)) &&
1212 !atomic_read(&sk->sk_rmem_alloc)))
1213 time_to_ack = 1;
1214 }
1215
1216 /* We send an ACK if we can now advertise a non-zero window
1217 * which has been raised "significantly".
1218 *
1219 * Even if window raised up to infinity, do not send window open ACK
1220 * in states, where we will not receive more. It is useless.
1221 */
1222 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1223 __u32 rcv_window_now = tcp_receive_window(tp);
1224
1225 /* Optimize, __tcp_select_window() is not cheap. */
1226 if (2*rcv_window_now <= tp->window_clamp) {
1227 __u32 new_window = __tcp_select_window(sk);
1228
1229 /* Send ACK now, if this read freed lots of space
1230 * in our buffer. Certainly, new_window is new window.
1231 * We can advertise it now, if it is not less than current one.
1232 * "Lots" means "at least twice" here.
1233 */
1234 if (new_window && new_window >= 2 * rcv_window_now)
1235 time_to_ack = 1;
1236 }
1237 }
1238 if (time_to_ack)
1239 tcp_send_ack(sk);
1240 }
1241
1242 static void tcp_prequeue_process(struct sock *sk)
1243 {
1244 struct sk_buff *skb;
1245 struct tcp_sock *tp = tcp_sk(sk);
1246
1247 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1248
1249 /* RX process wants to run with disabled BHs, though it is not
1250 * necessary */
1251 local_bh_disable();
1252 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1253 sk_backlog_rcv(sk, skb);
1254 local_bh_enable();
1255
1256 /* Clear memory counter. */
1257 tp->ucopy.memory = 0;
1258 }
1259
1260 #ifdef CONFIG_NET_DMA
1261 static void tcp_service_net_dma(struct sock *sk, bool wait)
1262 {
1263 dma_cookie_t done, used;
1264 dma_cookie_t last_issued;
1265 struct tcp_sock *tp = tcp_sk(sk);
1266
1267 if (!tp->ucopy.dma_chan)
1268 return;
1269
1270 last_issued = tp->ucopy.dma_cookie;
1271 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1272
1273 do {
1274 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1275 last_issued, &done,
1276 &used) == DMA_SUCCESS) {
1277 /* Safe to free early-copied skbs now */
1278 __skb_queue_purge(&sk->sk_async_wait_queue);
1279 break;
1280 } else {
1281 struct sk_buff *skb;
1282 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1283 (dma_async_is_complete(skb->dma_cookie, done,
1284 used) == DMA_SUCCESS)) {
1285 __skb_dequeue(&sk->sk_async_wait_queue);
1286 kfree_skb(skb);
1287 }
1288 }
1289 } while (wait);
1290 }
1291 #endif
1292
1293 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1294 {
1295 struct sk_buff *skb;
1296 u32 offset;
1297
1298 skb_queue_walk(&sk->sk_receive_queue, skb) {
1299 offset = seq - TCP_SKB_CB(skb)->seq;
1300 if (tcp_hdr(skb)->syn)
1301 offset--;
1302 if (offset < skb->len || tcp_hdr(skb)->fin) {
1303 *off = offset;
1304 return skb;
1305 }
1306 }
1307 return NULL;
1308 }
1309
1310 /*
1311 * This routine provides an alternative to tcp_recvmsg() for routines
1312 * that would like to handle copying from skbuffs directly in 'sendfile'
1313 * fashion.
1314 * Note:
1315 * - It is assumed that the socket was locked by the caller.
1316 * - The routine does not block.
1317 * - At present, there is no support for reading OOB data
1318 * or for 'peeking' the socket using this routine
1319 * (although both would be easy to implement).
1320 */
1321 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1322 sk_read_actor_t recv_actor)
1323 {
1324 struct sk_buff *skb;
1325 struct tcp_sock *tp = tcp_sk(sk);
1326 u32 seq = tp->copied_seq;
1327 u32 offset;
1328 int copied = 0;
1329
1330 if (sk->sk_state == TCP_LISTEN)
1331 return -ENOTCONN;
1332 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1333 if (offset < skb->len) {
1334 int used;
1335 size_t len;
1336
1337 len = skb->len - offset;
1338 /* Stop reading if we hit a patch of urgent data */
1339 if (tp->urg_data) {
1340 u32 urg_offset = tp->urg_seq - seq;
1341 if (urg_offset < len)
1342 len = urg_offset;
1343 if (!len)
1344 break;
1345 }
1346 used = recv_actor(desc, skb, offset, len);
1347 if (used < 0) {
1348 if (!copied)
1349 copied = used;
1350 break;
1351 } else if (used <= len) {
1352 seq += used;
1353 copied += used;
1354 offset += used;
1355 }
1356 /*
1357 * If recv_actor drops the lock (e.g. TCP splice
1358 * receive) the skb pointer might be invalid when
1359 * getting here: tcp_collapse might have deleted it
1360 * while aggregating skbs from the socket queue.
1361 */
1362 skb = tcp_recv_skb(sk, seq-1, &offset);
1363 if (!skb || (offset+1 != skb->len))
1364 break;
1365 }
1366 if (tcp_hdr(skb)->fin) {
1367 sk_eat_skb(sk, skb, 0);
1368 ++seq;
1369 break;
1370 }
1371 sk_eat_skb(sk, skb, 0);
1372 if (!desc->count)
1373 break;
1374 tp->copied_seq = seq;
1375 }
1376 tp->copied_seq = seq;
1377
1378 tcp_rcv_space_adjust(sk);
1379
1380 /* Clean up data we have read: This will do ACK frames. */
1381 if (copied > 0)
1382 tcp_cleanup_rbuf(sk, copied);
1383 return copied;
1384 }
1385 EXPORT_SYMBOL(tcp_read_sock);
1386
1387 /*
1388 * This routine copies from a sock struct into the user buffer.
1389 *
1390 * Technical note: in 2.3 we work on _locked_ socket, so that
1391 * tricks with *seq access order and skb->users are not required.
1392 * Probably, code can be easily improved even more.
1393 */
1394
1395 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1396 size_t len, int nonblock, int flags, int *addr_len)
1397 {
1398 struct tcp_sock *tp = tcp_sk(sk);
1399 int copied = 0;
1400 u32 peek_seq;
1401 u32 *seq;
1402 unsigned long used;
1403 int err;
1404 int target; /* Read at least this many bytes */
1405 long timeo;
1406 struct task_struct *user_recv = NULL;
1407 int copied_early = 0;
1408 struct sk_buff *skb;
1409 u32 urg_hole = 0;
1410
1411 lock_sock(sk);
1412
1413 TCP_CHECK_TIMER(sk);
1414
1415 err = -ENOTCONN;
1416 if (sk->sk_state == TCP_LISTEN)
1417 goto out;
1418
1419 timeo = sock_rcvtimeo(sk, nonblock);
1420
1421 /* Urgent data needs to be handled specially. */
1422 if (flags & MSG_OOB)
1423 goto recv_urg;
1424
1425 seq = &tp->copied_seq;
1426 if (flags & MSG_PEEK) {
1427 peek_seq = tp->copied_seq;
1428 seq = &peek_seq;
1429 }
1430
1431 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1432
1433 #ifdef CONFIG_NET_DMA
1434 tp->ucopy.dma_chan = NULL;
1435 preempt_disable();
1436 skb = skb_peek_tail(&sk->sk_receive_queue);
1437 {
1438 int available = 0;
1439
1440 if (skb)
1441 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1442 if ((available < target) &&
1443 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1444 !sysctl_tcp_low_latency &&
1445 dma_find_channel(DMA_MEMCPY)) {
1446 preempt_enable_no_resched();
1447 tp->ucopy.pinned_list =
1448 dma_pin_iovec_pages(msg->msg_iov, len);
1449 } else {
1450 preempt_enable_no_resched();
1451 }
1452 }
1453 #endif
1454
1455 do {
1456 u32 offset;
1457
1458 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1459 if (tp->urg_data && tp->urg_seq == *seq) {
1460 if (copied)
1461 break;
1462 if (signal_pending(current)) {
1463 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1464 break;
1465 }
1466 }
1467
1468 /* Next get a buffer. */
1469
1470 skb_queue_walk(&sk->sk_receive_queue, skb) {
1471 /* Now that we have two receive queues this
1472 * shouldn't happen.
1473 */
1474 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1475 KERN_INFO "recvmsg bug: copied %X "
1476 "seq %X rcvnxt %X fl %X\n", *seq,
1477 TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1478 flags))
1479 break;
1480
1481 offset = *seq - TCP_SKB_CB(skb)->seq;
1482 if (tcp_hdr(skb)->syn)
1483 offset--;
1484 if (offset < skb->len)
1485 goto found_ok_skb;
1486 if (tcp_hdr(skb)->fin)
1487 goto found_fin_ok;
1488 WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: "
1489 "copied %X seq %X rcvnxt %X fl %X\n",
1490 *seq, TCP_SKB_CB(skb)->seq,
1491 tp->rcv_nxt, flags);
1492 }
1493
1494 /* Well, if we have backlog, try to process it now yet. */
1495
1496 if (copied >= target && !sk->sk_backlog.tail)
1497 break;
1498
1499 if (copied) {
1500 if (sk->sk_err ||
1501 sk->sk_state == TCP_CLOSE ||
1502 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1503 !timeo ||
1504 signal_pending(current))
1505 break;
1506 } else {
1507 if (sock_flag(sk, SOCK_DONE))
1508 break;
1509
1510 if (sk->sk_err) {
1511 copied = sock_error(sk);
1512 break;
1513 }
1514
1515 if (sk->sk_shutdown & RCV_SHUTDOWN)
1516 break;
1517
1518 if (sk->sk_state == TCP_CLOSE) {
1519 if (!sock_flag(sk, SOCK_DONE)) {
1520 /* This occurs when user tries to read
1521 * from never connected socket.
1522 */
1523 copied = -ENOTCONN;
1524 break;
1525 }
1526 break;
1527 }
1528
1529 if (!timeo) {
1530 copied = -EAGAIN;
1531 break;
1532 }
1533
1534 if (signal_pending(current)) {
1535 copied = sock_intr_errno(timeo);
1536 break;
1537 }
1538 }
1539
1540 tcp_cleanup_rbuf(sk, copied);
1541
1542 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1543 /* Install new reader */
1544 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1545 user_recv = current;
1546 tp->ucopy.task = user_recv;
1547 tp->ucopy.iov = msg->msg_iov;
1548 }
1549
1550 tp->ucopy.len = len;
1551
1552 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1553 !(flags & (MSG_PEEK | MSG_TRUNC)));
1554
1555 /* Ugly... If prequeue is not empty, we have to
1556 * process it before releasing socket, otherwise
1557 * order will be broken at second iteration.
1558 * More elegant solution is required!!!
1559 *
1560 * Look: we have the following (pseudo)queues:
1561 *
1562 * 1. packets in flight
1563 * 2. backlog
1564 * 3. prequeue
1565 * 4. receive_queue
1566 *
1567 * Each queue can be processed only if the next ones
1568 * are empty. At this point we have empty receive_queue.
1569 * But prequeue _can_ be not empty after 2nd iteration,
1570 * when we jumped to start of loop because backlog
1571 * processing added something to receive_queue.
1572 * We cannot release_sock(), because backlog contains
1573 * packets arrived _after_ prequeued ones.
1574 *
1575 * Shortly, algorithm is clear --- to process all
1576 * the queues in order. We could make it more directly,
1577 * requeueing packets from backlog to prequeue, if
1578 * is not empty. It is more elegant, but eats cycles,
1579 * unfortunately.
1580 */
1581 if (!skb_queue_empty(&tp->ucopy.prequeue))
1582 goto do_prequeue;
1583
1584 /* __ Set realtime policy in scheduler __ */
1585 }
1586
1587 #ifdef CONFIG_NET_DMA
1588 if (tp->ucopy.dma_chan)
1589 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1590 #endif
1591 if (copied >= target) {
1592 /* Do not sleep, just process backlog. */
1593 release_sock(sk);
1594 lock_sock(sk);
1595 } else
1596 sk_wait_data(sk, &timeo);
1597
1598 #ifdef CONFIG_NET_DMA
1599 tcp_service_net_dma(sk, false); /* Don't block */
1600 tp->ucopy.wakeup = 0;
1601 #endif
1602
1603 if (user_recv) {
1604 int chunk;
1605
1606 /* __ Restore normal policy in scheduler __ */
1607
1608 if ((chunk = len - tp->ucopy.len) != 0) {
1609 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1610 len -= chunk;
1611 copied += chunk;
1612 }
1613
1614 if (tp->rcv_nxt == tp->copied_seq &&
1615 !skb_queue_empty(&tp->ucopy.prequeue)) {
1616 do_prequeue:
1617 tcp_prequeue_process(sk);
1618
1619 if ((chunk = len - tp->ucopy.len) != 0) {
1620 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1621 len -= chunk;
1622 copied += chunk;
1623 }
1624 }
1625 }
1626 if ((flags & MSG_PEEK) &&
1627 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1628 if (net_ratelimit())
1629 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1630 current->comm, task_pid_nr(current));
1631 peek_seq = tp->copied_seq;
1632 }
1633 continue;
1634
1635 found_ok_skb:
1636 /* Ok so how much can we use? */
1637 used = skb->len - offset;
1638 if (len < used)
1639 used = len;
1640
1641 /* Do we have urgent data here? */
1642 if (tp->urg_data) {
1643 u32 urg_offset = tp->urg_seq - *seq;
1644 if (urg_offset < used) {
1645 if (!urg_offset) {
1646 if (!sock_flag(sk, SOCK_URGINLINE)) {
1647 ++*seq;
1648 urg_hole++;
1649 offset++;
1650 used--;
1651 if (!used)
1652 goto skip_copy;
1653 }
1654 } else
1655 used = urg_offset;
1656 }
1657 }
1658
1659 if (!(flags & MSG_TRUNC)) {
1660 #ifdef CONFIG_NET_DMA
1661 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1662 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1663
1664 if (tp->ucopy.dma_chan) {
1665 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1666 tp->ucopy.dma_chan, skb, offset,
1667 msg->msg_iov, used,
1668 tp->ucopy.pinned_list);
1669
1670 if (tp->ucopy.dma_cookie < 0) {
1671
1672 printk(KERN_ALERT "dma_cookie < 0\n");
1673
1674 /* Exception. Bailout! */
1675 if (!copied)
1676 copied = -EFAULT;
1677 break;
1678 }
1679
1680 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1681
1682 if ((offset + used) == skb->len)
1683 copied_early = 1;
1684
1685 } else
1686 #endif
1687 {
1688 err = skb_copy_datagram_iovec(skb, offset,
1689 msg->msg_iov, used);
1690 if (err) {
1691 /* Exception. Bailout! */
1692 if (!copied)
1693 copied = -EFAULT;
1694 break;
1695 }
1696 }
1697 }
1698
1699 *seq += used;
1700 copied += used;
1701 len -= used;
1702
1703 tcp_rcv_space_adjust(sk);
1704
1705 skip_copy:
1706 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1707 tp->urg_data = 0;
1708 tcp_fast_path_check(sk);
1709 }
1710 if (used + offset < skb->len)
1711 continue;
1712
1713 if (tcp_hdr(skb)->fin)
1714 goto found_fin_ok;
1715 if (!(flags & MSG_PEEK)) {
1716 sk_eat_skb(sk, skb, copied_early);
1717 copied_early = 0;
1718 }
1719 continue;
1720
1721 found_fin_ok:
1722 /* Process the FIN. */
1723 ++*seq;
1724 if (!(flags & MSG_PEEK)) {
1725 sk_eat_skb(sk, skb, copied_early);
1726 copied_early = 0;
1727 }
1728 break;
1729 } while (len > 0);
1730
1731 if (user_recv) {
1732 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1733 int chunk;
1734
1735 tp->ucopy.len = copied > 0 ? len : 0;
1736
1737 tcp_prequeue_process(sk);
1738
1739 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1740 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1741 len -= chunk;
1742 copied += chunk;
1743 }
1744 }
1745
1746 tp->ucopy.task = NULL;
1747 tp->ucopy.len = 0;
1748 }
1749
1750 #ifdef CONFIG_NET_DMA
1751 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1752 tp->ucopy.dma_chan = NULL;
1753
1754 if (tp->ucopy.pinned_list) {
1755 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1756 tp->ucopy.pinned_list = NULL;
1757 }
1758 #endif
1759
1760 /* According to UNIX98, msg_name/msg_namelen are ignored
1761 * on connected socket. I was just happy when found this 8) --ANK
1762 */
1763
1764 /* Clean up data we have read: This will do ACK frames. */
1765 tcp_cleanup_rbuf(sk, copied);
1766
1767 TCP_CHECK_TIMER(sk);
1768 release_sock(sk);
1769 return copied;
1770
1771 out:
1772 TCP_CHECK_TIMER(sk);
1773 release_sock(sk);
1774 return err;
1775
1776 recv_urg:
1777 err = tcp_recv_urg(sk, msg, len, flags);
1778 goto out;
1779 }
1780 EXPORT_SYMBOL(tcp_recvmsg);
1781
1782 void tcp_set_state(struct sock *sk, int state)
1783 {
1784 int oldstate = sk->sk_state;
1785
1786 switch (state) {
1787 case TCP_ESTABLISHED:
1788 if (oldstate != TCP_ESTABLISHED)
1789 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1790 break;
1791
1792 case TCP_CLOSE:
1793 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1794 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1795
1796 sk->sk_prot->unhash(sk);
1797 if (inet_csk(sk)->icsk_bind_hash &&
1798 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1799 inet_put_port(sk);
1800 /* fall through */
1801 default:
1802 if (oldstate == TCP_ESTABLISHED)
1803 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1804 }
1805
1806 /* Change state AFTER socket is unhashed to avoid closed
1807 * socket sitting in hash tables.
1808 */
1809 sk->sk_state = state;
1810
1811 #ifdef STATE_TRACE
1812 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1813 #endif
1814 }
1815 EXPORT_SYMBOL_GPL(tcp_set_state);
1816
1817 /*
1818 * State processing on a close. This implements the state shift for
1819 * sending our FIN frame. Note that we only send a FIN for some
1820 * states. A shutdown() may have already sent the FIN, or we may be
1821 * closed.
1822 */
1823
1824 static const unsigned char new_state[16] = {
1825 /* current state: new state: action: */
1826 /* (Invalid) */ TCP_CLOSE,
1827 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1828 /* TCP_SYN_SENT */ TCP_CLOSE,
1829 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1830 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1831 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1832 /* TCP_TIME_WAIT */ TCP_CLOSE,
1833 /* TCP_CLOSE */ TCP_CLOSE,
1834 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1835 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1836 /* TCP_LISTEN */ TCP_CLOSE,
1837 /* TCP_CLOSING */ TCP_CLOSING,
1838 };
1839
1840 static int tcp_close_state(struct sock *sk)
1841 {
1842 int next = (int)new_state[sk->sk_state];
1843 int ns = next & TCP_STATE_MASK;
1844
1845 tcp_set_state(sk, ns);
1846
1847 return next & TCP_ACTION_FIN;
1848 }
1849
1850 /*
1851 * Shutdown the sending side of a connection. Much like close except
1852 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1853 */
1854
1855 void tcp_shutdown(struct sock *sk, int how)
1856 {
1857 /* We need to grab some memory, and put together a FIN,
1858 * and then put it into the queue to be sent.
1859 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1860 */
1861 if (!(how & SEND_SHUTDOWN))
1862 return;
1863
1864 /* If we've already sent a FIN, or it's a closed state, skip this. */
1865 if ((1 << sk->sk_state) &
1866 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1867 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1868 /* Clear out any half completed packets. FIN if needed. */
1869 if (tcp_close_state(sk))
1870 tcp_send_fin(sk);
1871 }
1872 }
1873 EXPORT_SYMBOL(tcp_shutdown);
1874
1875 void tcp_close(struct sock *sk, long timeout)
1876 {
1877 struct sk_buff *skb;
1878 int data_was_unread = 0;
1879 int state;
1880
1881 lock_sock(sk);
1882 sk->sk_shutdown = SHUTDOWN_MASK;
1883
1884 if (sk->sk_state == TCP_LISTEN) {
1885 tcp_set_state(sk, TCP_CLOSE);
1886
1887 /* Special case. */
1888 inet_csk_listen_stop(sk);
1889
1890 goto adjudge_to_death;
1891 }
1892
1893 /* We need to flush the recv. buffs. We do this only on the
1894 * descriptor close, not protocol-sourced closes, because the
1895 * reader process may not have drained the data yet!
1896 */
1897 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1898 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1899 tcp_hdr(skb)->fin;
1900 data_was_unread += len;
1901 __kfree_skb(skb);
1902 }
1903
1904 sk_mem_reclaim(sk);
1905
1906 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
1907 if (sk->sk_state == TCP_CLOSE)
1908 goto adjudge_to_death;
1909
1910 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1911 * data was lost. To witness the awful effects of the old behavior of
1912 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1913 * GET in an FTP client, suspend the process, wait for the client to
1914 * advertise a zero window, then kill -9 the FTP client, wheee...
1915 * Note: timeout is always zero in such a case.
1916 */
1917 if (data_was_unread) {
1918 /* Unread data was tossed, zap the connection. */
1919 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1920 tcp_set_state(sk, TCP_CLOSE);
1921 tcp_send_active_reset(sk, sk->sk_allocation);
1922 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1923 /* Check zero linger _after_ checking for unread data. */
1924 sk->sk_prot->disconnect(sk, 0);
1925 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1926 } else if (tcp_close_state(sk)) {
1927 /* We FIN if the application ate all the data before
1928 * zapping the connection.
1929 */
1930
1931 /* RED-PEN. Formally speaking, we have broken TCP state
1932 * machine. State transitions:
1933 *
1934 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1935 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1936 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1937 *
1938 * are legal only when FIN has been sent (i.e. in window),
1939 * rather than queued out of window. Purists blame.
1940 *
1941 * F.e. "RFC state" is ESTABLISHED,
1942 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1943 *
1944 * The visible declinations are that sometimes
1945 * we enter time-wait state, when it is not required really
1946 * (harmless), do not send active resets, when they are
1947 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1948 * they look as CLOSING or LAST_ACK for Linux)
1949 * Probably, I missed some more holelets.
1950 * --ANK
1951 */
1952 tcp_send_fin(sk);
1953 }
1954
1955 sk_stream_wait_close(sk, timeout);
1956
1957 adjudge_to_death:
1958 state = sk->sk_state;
1959 sock_hold(sk);
1960 sock_orphan(sk);
1961
1962 /* It is the last release_sock in its life. It will remove backlog. */
1963 release_sock(sk);
1964
1965
1966 /* Now socket is owned by kernel and we acquire BH lock
1967 to finish close. No need to check for user refs.
1968 */
1969 local_bh_disable();
1970 bh_lock_sock(sk);
1971 WARN_ON(sock_owned_by_user(sk));
1972
1973 percpu_counter_inc(sk->sk_prot->orphan_count);
1974
1975 /* Have we already been destroyed by a softirq or backlog? */
1976 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1977 goto out;
1978
1979 /* This is a (useful) BSD violating of the RFC. There is a
1980 * problem with TCP as specified in that the other end could
1981 * keep a socket open forever with no application left this end.
1982 * We use a 3 minute timeout (about the same as BSD) then kill
1983 * our end. If they send after that then tough - BUT: long enough
1984 * that we won't make the old 4*rto = almost no time - whoops
1985 * reset mistake.
1986 *
1987 * Nope, it was not mistake. It is really desired behaviour
1988 * f.e. on http servers, when such sockets are useless, but
1989 * consume significant resources. Let's do it with special
1990 * linger2 option. --ANK
1991 */
1992
1993 if (sk->sk_state == TCP_FIN_WAIT2) {
1994 struct tcp_sock *tp = tcp_sk(sk);
1995 if (tp->linger2 < 0) {
1996 tcp_set_state(sk, TCP_CLOSE);
1997 tcp_send_active_reset(sk, GFP_ATOMIC);
1998 NET_INC_STATS_BH(sock_net(sk),
1999 LINUX_MIB_TCPABORTONLINGER);
2000 } else {
2001 const int tmo = tcp_fin_time(sk);
2002
2003 if (tmo > TCP_TIMEWAIT_LEN) {
2004 inet_csk_reset_keepalive_timer(sk,
2005 tmo - TCP_TIMEWAIT_LEN);
2006 } else {
2007 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2008 goto out;
2009 }
2010 }
2011 }
2012 if (sk->sk_state != TCP_CLOSE) {
2013 int orphan_count = percpu_counter_read_positive(
2014 sk->sk_prot->orphan_count);
2015
2016 sk_mem_reclaim(sk);
2017 if (tcp_too_many_orphans(sk, orphan_count)) {
2018 if (net_ratelimit())
2019 printk(KERN_INFO "TCP: too many of orphaned "
2020 "sockets\n");
2021 tcp_set_state(sk, TCP_CLOSE);
2022 tcp_send_active_reset(sk, GFP_ATOMIC);
2023 NET_INC_STATS_BH(sock_net(sk),
2024 LINUX_MIB_TCPABORTONMEMORY);
2025 }
2026 }
2027
2028 if (sk->sk_state == TCP_CLOSE)
2029 inet_csk_destroy_sock(sk);
2030 /* Otherwise, socket is reprieved until protocol close. */
2031
2032 out:
2033 bh_unlock_sock(sk);
2034 local_bh_enable();
2035 sock_put(sk);
2036 }
2037 EXPORT_SYMBOL(tcp_close);
2038
2039 /* These states need RST on ABORT according to RFC793 */
2040
2041 static inline int tcp_need_reset(int state)
2042 {
2043 return (1 << state) &
2044 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2045 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2046 }
2047
2048 int tcp_disconnect(struct sock *sk, int flags)
2049 {
2050 struct inet_sock *inet = inet_sk(sk);
2051 struct inet_connection_sock *icsk = inet_csk(sk);
2052 struct tcp_sock *tp = tcp_sk(sk);
2053 int err = 0;
2054 int old_state = sk->sk_state;
2055
2056 if (old_state != TCP_CLOSE)
2057 tcp_set_state(sk, TCP_CLOSE);
2058
2059 /* ABORT function of RFC793 */
2060 if (old_state == TCP_LISTEN) {
2061 inet_csk_listen_stop(sk);
2062 } else if (tcp_need_reset(old_state) ||
2063 (tp->snd_nxt != tp->write_seq &&
2064 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2065 /* The last check adjusts for discrepancy of Linux wrt. RFC
2066 * states
2067 */
2068 tcp_send_active_reset(sk, gfp_any());
2069 sk->sk_err = ECONNRESET;
2070 } else if (old_state == TCP_SYN_SENT)
2071 sk->sk_err = ECONNRESET;
2072
2073 tcp_clear_xmit_timers(sk);
2074 __skb_queue_purge(&sk->sk_receive_queue);
2075 tcp_write_queue_purge(sk);
2076 __skb_queue_purge(&tp->out_of_order_queue);
2077 #ifdef CONFIG_NET_DMA
2078 __skb_queue_purge(&sk->sk_async_wait_queue);
2079 #endif
2080
2081 inet->inet_dport = 0;
2082
2083 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2084 inet_reset_saddr(sk);
2085
2086 sk->sk_shutdown = 0;
2087 sock_reset_flag(sk, SOCK_DONE);
2088 tp->srtt = 0;
2089 if ((tp->write_seq += tp->max_window + 2) == 0)
2090 tp->write_seq = 1;
2091 icsk->icsk_backoff = 0;
2092 tp->snd_cwnd = 2;
2093 icsk->icsk_probes_out = 0;
2094 tp->packets_out = 0;
2095 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2096 tp->snd_cwnd_cnt = 0;
2097 tp->bytes_acked = 0;
2098 tp->window_clamp = 0;
2099 tcp_set_ca_state(sk, TCP_CA_Open);
2100 tcp_clear_retrans(tp);
2101 inet_csk_delack_init(sk);
2102 tcp_init_send_head(sk);
2103 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2104 __sk_dst_reset(sk);
2105
2106 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2107
2108 sk->sk_error_report(sk);
2109 return err;
2110 }
2111 EXPORT_SYMBOL(tcp_disconnect);
2112
2113 /*
2114 * Socket option code for TCP.
2115 */
2116 static int do_tcp_setsockopt(struct sock *sk, int level,
2117 int optname, char __user *optval, unsigned int optlen)
2118 {
2119 struct tcp_sock *tp = tcp_sk(sk);
2120 struct inet_connection_sock *icsk = inet_csk(sk);
2121 int val;
2122 int err = 0;
2123
2124 /* These are data/string values, all the others are ints */
2125 switch (optname) {
2126 case TCP_CONGESTION: {
2127 char name[TCP_CA_NAME_MAX];
2128
2129 if (optlen < 1)
2130 return -EINVAL;
2131
2132 val = strncpy_from_user(name, optval,
2133 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2134 if (val < 0)
2135 return -EFAULT;
2136 name[val] = 0;
2137
2138 lock_sock(sk);
2139 err = tcp_set_congestion_control(sk, name);
2140 release_sock(sk);
2141 return err;
2142 }
2143 case TCP_COOKIE_TRANSACTIONS: {
2144 struct tcp_cookie_transactions ctd;
2145 struct tcp_cookie_values *cvp = NULL;
2146
2147 if (sizeof(ctd) > optlen)
2148 return -EINVAL;
2149 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2150 return -EFAULT;
2151
2152 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2153 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2154 return -EINVAL;
2155
2156 if (ctd.tcpct_cookie_desired == 0) {
2157 /* default to global value */
2158 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2159 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2160 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2161 return -EINVAL;
2162 }
2163
2164 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2165 /* Supercedes all other values */
2166 lock_sock(sk);
2167 if (tp->cookie_values != NULL) {
2168 kref_put(&tp->cookie_values->kref,
2169 tcp_cookie_values_release);
2170 tp->cookie_values = NULL;
2171 }
2172 tp->rx_opt.cookie_in_always = 0; /* false */
2173 tp->rx_opt.cookie_out_never = 1; /* true */
2174 release_sock(sk);
2175 return err;
2176 }
2177
2178 /* Allocate ancillary memory before locking.
2179 */
2180 if (ctd.tcpct_used > 0 ||
2181 (tp->cookie_values == NULL &&
2182 (sysctl_tcp_cookie_size > 0 ||
2183 ctd.tcpct_cookie_desired > 0 ||
2184 ctd.tcpct_s_data_desired > 0))) {
2185 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2186 GFP_KERNEL);
2187 if (cvp == NULL)
2188 return -ENOMEM;
2189 }
2190 lock_sock(sk);
2191 tp->rx_opt.cookie_in_always =
2192 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2193 tp->rx_opt.cookie_out_never = 0; /* false */
2194
2195 if (tp->cookie_values != NULL) {
2196 if (cvp != NULL) {
2197 /* Changed values are recorded by a changed
2198 * pointer, ensuring the cookie will differ,
2199 * without separately hashing each value later.
2200 */
2201 kref_put(&tp->cookie_values->kref,
2202 tcp_cookie_values_release);
2203 kref_init(&cvp->kref);
2204 tp->cookie_values = cvp;
2205 } else {
2206 cvp = tp->cookie_values;
2207 }
2208 }
2209 if (cvp != NULL) {
2210 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2211
2212 if (ctd.tcpct_used > 0) {
2213 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2214 ctd.tcpct_used);
2215 cvp->s_data_desired = ctd.tcpct_used;
2216 cvp->s_data_constant = 1; /* true */
2217 } else {
2218 /* No constant payload data. */
2219 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2220 cvp->s_data_constant = 0; /* false */
2221 }
2222 }
2223 release_sock(sk);
2224 return err;
2225 }
2226 default:
2227 /* fallthru */
2228 break;
2229 }
2230
2231 if (optlen < sizeof(int))
2232 return -EINVAL;
2233
2234 if (get_user(val, (int __user *)optval))
2235 return -EFAULT;
2236
2237 lock_sock(sk);
2238
2239 switch (optname) {
2240 case TCP_MAXSEG:
2241 /* Values greater than interface MTU won't take effect. However
2242 * at the point when this call is done we typically don't yet
2243 * know which interface is going to be used */
2244 if (val < 8 || val > MAX_TCP_WINDOW) {
2245 err = -EINVAL;
2246 break;
2247 }
2248 tp->rx_opt.user_mss = val;
2249 break;
2250
2251 case TCP_NODELAY:
2252 if (val) {
2253 /* TCP_NODELAY is weaker than TCP_CORK, so that
2254 * this option on corked socket is remembered, but
2255 * it is not activated until cork is cleared.
2256 *
2257 * However, when TCP_NODELAY is set we make
2258 * an explicit push, which overrides even TCP_CORK
2259 * for currently queued segments.
2260 */
2261 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2262 tcp_push_pending_frames(sk);
2263 } else {
2264 tp->nonagle &= ~TCP_NAGLE_OFF;
2265 }
2266 break;
2267
2268 case TCP_THIN_LINEAR_TIMEOUTS:
2269 if (val < 0 || val > 1)
2270 err = -EINVAL;
2271 else
2272 tp->thin_lto = val;
2273 break;
2274
2275 case TCP_THIN_DUPACK:
2276 if (val < 0 || val > 1)
2277 err = -EINVAL;
2278 else
2279 tp->thin_dupack = val;
2280 break;
2281
2282 case TCP_CORK:
2283 /* When set indicates to always queue non-full frames.
2284 * Later the user clears this option and we transmit
2285 * any pending partial frames in the queue. This is
2286 * meant to be used alongside sendfile() to get properly
2287 * filled frames when the user (for example) must write
2288 * out headers with a write() call first and then use
2289 * sendfile to send out the data parts.
2290 *
2291 * TCP_CORK can be set together with TCP_NODELAY and it is
2292 * stronger than TCP_NODELAY.
2293 */
2294 if (val) {
2295 tp->nonagle |= TCP_NAGLE_CORK;
2296 } else {
2297 tp->nonagle &= ~TCP_NAGLE_CORK;
2298 if (tp->nonagle&TCP_NAGLE_OFF)
2299 tp->nonagle |= TCP_NAGLE_PUSH;
2300 tcp_push_pending_frames(sk);
2301 }
2302 break;
2303
2304 case TCP_KEEPIDLE:
2305 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2306 err = -EINVAL;
2307 else {
2308 tp->keepalive_time = val * HZ;
2309 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2310 !((1 << sk->sk_state) &
2311 (TCPF_CLOSE | TCPF_LISTEN))) {
2312 u32 elapsed = keepalive_time_elapsed(tp);
2313 if (tp->keepalive_time > elapsed)
2314 elapsed = tp->keepalive_time - elapsed;
2315 else
2316 elapsed = 0;
2317 inet_csk_reset_keepalive_timer(sk, elapsed);
2318 }
2319 }
2320 break;
2321 case TCP_KEEPINTVL:
2322 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2323 err = -EINVAL;
2324 else
2325 tp->keepalive_intvl = val * HZ;
2326 break;
2327 case TCP_KEEPCNT:
2328 if (val < 1 || val > MAX_TCP_KEEPCNT)
2329 err = -EINVAL;
2330 else
2331 tp->keepalive_probes = val;
2332 break;
2333 case TCP_SYNCNT:
2334 if (val < 1 || val > MAX_TCP_SYNCNT)
2335 err = -EINVAL;
2336 else
2337 icsk->icsk_syn_retries = val;
2338 break;
2339
2340 case TCP_LINGER2:
2341 if (val < 0)
2342 tp->linger2 = -1;
2343 else if (val > sysctl_tcp_fin_timeout / HZ)
2344 tp->linger2 = 0;
2345 else
2346 tp->linger2 = val * HZ;
2347 break;
2348
2349 case TCP_DEFER_ACCEPT:
2350 /* Translate value in seconds to number of retransmits */
2351 icsk->icsk_accept_queue.rskq_defer_accept =
2352 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2353 TCP_RTO_MAX / HZ);
2354 break;
2355
2356 case TCP_WINDOW_CLAMP:
2357 if (!val) {
2358 if (sk->sk_state != TCP_CLOSE) {
2359 err = -EINVAL;
2360 break;
2361 }
2362 tp->window_clamp = 0;
2363 } else
2364 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2365 SOCK_MIN_RCVBUF / 2 : val;
2366 break;
2367
2368 case TCP_QUICKACK:
2369 if (!val) {
2370 icsk->icsk_ack.pingpong = 1;
2371 } else {
2372 icsk->icsk_ack.pingpong = 0;
2373 if ((1 << sk->sk_state) &
2374 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2375 inet_csk_ack_scheduled(sk)) {
2376 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2377 tcp_cleanup_rbuf(sk, 1);
2378 if (!(val & 1))
2379 icsk->icsk_ack.pingpong = 1;
2380 }
2381 }
2382 break;
2383
2384 #ifdef CONFIG_TCP_MD5SIG
2385 case TCP_MD5SIG:
2386 /* Read the IP->Key mappings from userspace */
2387 err = tp->af_specific->md5_parse(sk, optval, optlen);
2388 break;
2389 #endif
2390
2391 default:
2392 err = -ENOPROTOOPT;
2393 break;
2394 }
2395
2396 release_sock(sk);
2397 return err;
2398 }
2399
2400 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2401 unsigned int optlen)
2402 {
2403 struct inet_connection_sock *icsk = inet_csk(sk);
2404
2405 if (level != SOL_TCP)
2406 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2407 optval, optlen);
2408 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2409 }
2410 EXPORT_SYMBOL(tcp_setsockopt);
2411
2412 #ifdef CONFIG_COMPAT
2413 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2414 char __user *optval, unsigned int optlen)
2415 {
2416 if (level != SOL_TCP)
2417 return inet_csk_compat_setsockopt(sk, level, optname,
2418 optval, optlen);
2419 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2420 }
2421 EXPORT_SYMBOL(compat_tcp_setsockopt);
2422 #endif
2423
2424 /* Return information about state of tcp endpoint in API format. */
2425 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2426 {
2427 struct tcp_sock *tp = tcp_sk(sk);
2428 const struct inet_connection_sock *icsk = inet_csk(sk);
2429 u32 now = tcp_time_stamp;
2430
2431 memset(info, 0, sizeof(*info));
2432
2433 info->tcpi_state = sk->sk_state;
2434 info->tcpi_ca_state = icsk->icsk_ca_state;
2435 info->tcpi_retransmits = icsk->icsk_retransmits;
2436 info->tcpi_probes = icsk->icsk_probes_out;
2437 info->tcpi_backoff = icsk->icsk_backoff;
2438
2439 if (tp->rx_opt.tstamp_ok)
2440 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2441 if (tcp_is_sack(tp))
2442 info->tcpi_options |= TCPI_OPT_SACK;
2443 if (tp->rx_opt.wscale_ok) {
2444 info->tcpi_options |= TCPI_OPT_WSCALE;
2445 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2446 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2447 }
2448
2449 if (tp->ecn_flags&TCP_ECN_OK)
2450 info->tcpi_options |= TCPI_OPT_ECN;
2451
2452 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2453 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2454 info->tcpi_snd_mss = tp->mss_cache;
2455 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2456
2457 if (sk->sk_state == TCP_LISTEN) {
2458 info->tcpi_unacked = sk->sk_ack_backlog;
2459 info->tcpi_sacked = sk->sk_max_ack_backlog;
2460 } else {
2461 info->tcpi_unacked = tp->packets_out;
2462 info->tcpi_sacked = tp->sacked_out;
2463 }
2464 info->tcpi_lost = tp->lost_out;
2465 info->tcpi_retrans = tp->retrans_out;
2466 info->tcpi_fackets = tp->fackets_out;
2467
2468 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2469 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2470 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2471
2472 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2473 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2474 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2475 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2476 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2477 info->tcpi_snd_cwnd = tp->snd_cwnd;
2478 info->tcpi_advmss = tp->advmss;
2479 info->tcpi_reordering = tp->reordering;
2480
2481 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2482 info->tcpi_rcv_space = tp->rcvq_space.space;
2483
2484 info->tcpi_total_retrans = tp->total_retrans;
2485 }
2486 EXPORT_SYMBOL_GPL(tcp_get_info);
2487
2488 static int do_tcp_getsockopt(struct sock *sk, int level,
2489 int optname, char __user *optval, int __user *optlen)
2490 {
2491 struct inet_connection_sock *icsk = inet_csk(sk);
2492 struct tcp_sock *tp = tcp_sk(sk);
2493 int val, len;
2494
2495 if (get_user(len, optlen))
2496 return -EFAULT;
2497
2498 len = min_t(unsigned int, len, sizeof(int));
2499
2500 if (len < 0)
2501 return -EINVAL;
2502
2503 switch (optname) {
2504 case TCP_MAXSEG:
2505 val = tp->mss_cache;
2506 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2507 val = tp->rx_opt.user_mss;
2508 break;
2509 case TCP_NODELAY:
2510 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2511 break;
2512 case TCP_CORK:
2513 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2514 break;
2515 case TCP_KEEPIDLE:
2516 val = keepalive_time_when(tp) / HZ;
2517 break;
2518 case TCP_KEEPINTVL:
2519 val = keepalive_intvl_when(tp) / HZ;
2520 break;
2521 case TCP_KEEPCNT:
2522 val = keepalive_probes(tp);
2523 break;
2524 case TCP_SYNCNT:
2525 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2526 break;
2527 case TCP_LINGER2:
2528 val = tp->linger2;
2529 if (val >= 0)
2530 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2531 break;
2532 case TCP_DEFER_ACCEPT:
2533 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2534 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2535 break;
2536 case TCP_WINDOW_CLAMP:
2537 val = tp->window_clamp;
2538 break;
2539 case TCP_INFO: {
2540 struct tcp_info info;
2541
2542 if (get_user(len, optlen))
2543 return -EFAULT;
2544
2545 tcp_get_info(sk, &info);
2546
2547 len = min_t(unsigned int, len, sizeof(info));
2548 if (put_user(len, optlen))
2549 return -EFAULT;
2550 if (copy_to_user(optval, &info, len))
2551 return -EFAULT;
2552 return 0;
2553 }
2554 case TCP_QUICKACK:
2555 val = !icsk->icsk_ack.pingpong;
2556 break;
2557
2558 case TCP_CONGESTION:
2559 if (get_user(len, optlen))
2560 return -EFAULT;
2561 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2562 if (put_user(len, optlen))
2563 return -EFAULT;
2564 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2565 return -EFAULT;
2566 return 0;
2567
2568 case TCP_COOKIE_TRANSACTIONS: {
2569 struct tcp_cookie_transactions ctd;
2570 struct tcp_cookie_values *cvp = tp->cookie_values;
2571
2572 if (get_user(len, optlen))
2573 return -EFAULT;
2574 if (len < sizeof(ctd))
2575 return -EINVAL;
2576
2577 memset(&ctd, 0, sizeof(ctd));
2578 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2579 TCP_COOKIE_IN_ALWAYS : 0)
2580 | (tp->rx_opt.cookie_out_never ?
2581 TCP_COOKIE_OUT_NEVER : 0);
2582
2583 if (cvp != NULL) {
2584 ctd.tcpct_flags |= (cvp->s_data_in ?
2585 TCP_S_DATA_IN : 0)
2586 | (cvp->s_data_out ?
2587 TCP_S_DATA_OUT : 0);
2588
2589 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2590 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2591
2592 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2593 cvp->cookie_pair_size);
2594 ctd.tcpct_used = cvp->cookie_pair_size;
2595 }
2596
2597 if (put_user(sizeof(ctd), optlen))
2598 return -EFAULT;
2599 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2600 return -EFAULT;
2601 return 0;
2602 }
2603 default:
2604 return -ENOPROTOOPT;
2605 }
2606
2607 if (put_user(len, optlen))
2608 return -EFAULT;
2609 if (copy_to_user(optval, &val, len))
2610 return -EFAULT;
2611 return 0;
2612 }
2613
2614 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2615 int __user *optlen)
2616 {
2617 struct inet_connection_sock *icsk = inet_csk(sk);
2618
2619 if (level != SOL_TCP)
2620 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2621 optval, optlen);
2622 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2623 }
2624 EXPORT_SYMBOL(tcp_getsockopt);
2625
2626 #ifdef CONFIG_COMPAT
2627 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2628 char __user *optval, int __user *optlen)
2629 {
2630 if (level != SOL_TCP)
2631 return inet_csk_compat_getsockopt(sk, level, optname,
2632 optval, optlen);
2633 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2634 }
2635 EXPORT_SYMBOL(compat_tcp_getsockopt);
2636 #endif
2637
2638 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2639 {
2640 struct sk_buff *segs = ERR_PTR(-EINVAL);
2641 struct tcphdr *th;
2642 unsigned thlen;
2643 unsigned int seq;
2644 __be32 delta;
2645 unsigned int oldlen;
2646 unsigned int mss;
2647
2648 if (!pskb_may_pull(skb, sizeof(*th)))
2649 goto out;
2650
2651 th = tcp_hdr(skb);
2652 thlen = th->doff * 4;
2653 if (thlen < sizeof(*th))
2654 goto out;
2655
2656 if (!pskb_may_pull(skb, thlen))
2657 goto out;
2658
2659 oldlen = (u16)~skb->len;
2660 __skb_pull(skb, thlen);
2661
2662 mss = skb_shinfo(skb)->gso_size;
2663 if (unlikely(skb->len <= mss))
2664 goto out;
2665
2666 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2667 /* Packet is from an untrusted source, reset gso_segs. */
2668 int type = skb_shinfo(skb)->gso_type;
2669
2670 if (unlikely(type &
2671 ~(SKB_GSO_TCPV4 |
2672 SKB_GSO_DODGY |
2673 SKB_GSO_TCP_ECN |
2674 SKB_GSO_TCPV6 |
2675 0) ||
2676 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2677 goto out;
2678
2679 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2680
2681 segs = NULL;
2682 goto out;
2683 }
2684
2685 segs = skb_segment(skb, features);
2686 if (IS_ERR(segs))
2687 goto out;
2688
2689 delta = htonl(oldlen + (thlen + mss));
2690
2691 skb = segs;
2692 th = tcp_hdr(skb);
2693 seq = ntohl(th->seq);
2694
2695 do {
2696 th->fin = th->psh = 0;
2697
2698 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2699 (__force u32)delta));
2700 if (skb->ip_summed != CHECKSUM_PARTIAL)
2701 th->check =
2702 csum_fold(csum_partial(skb_transport_header(skb),
2703 thlen, skb->csum));
2704
2705 seq += mss;
2706 skb = skb->next;
2707 th = tcp_hdr(skb);
2708
2709 th->seq = htonl(seq);
2710 th->cwr = 0;
2711 } while (skb->next);
2712
2713 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2714 skb->data_len);
2715 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2716 (__force u32)delta));
2717 if (skb->ip_summed != CHECKSUM_PARTIAL)
2718 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2719 thlen, skb->csum));
2720
2721 out:
2722 return segs;
2723 }
2724 EXPORT_SYMBOL(tcp_tso_segment);
2725
2726 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2727 {
2728 struct sk_buff **pp = NULL;
2729 struct sk_buff *p;
2730 struct tcphdr *th;
2731 struct tcphdr *th2;
2732 unsigned int len;
2733 unsigned int thlen;
2734 __be32 flags;
2735 unsigned int mss = 1;
2736 unsigned int hlen;
2737 unsigned int off;
2738 int flush = 1;
2739 int i;
2740
2741 off = skb_gro_offset(skb);
2742 hlen = off + sizeof(*th);
2743 th = skb_gro_header_fast(skb, off);
2744 if (skb_gro_header_hard(skb, hlen)) {
2745 th = skb_gro_header_slow(skb, hlen, off);
2746 if (unlikely(!th))
2747 goto out;
2748 }
2749
2750 thlen = th->doff * 4;
2751 if (thlen < sizeof(*th))
2752 goto out;
2753
2754 hlen = off + thlen;
2755 if (skb_gro_header_hard(skb, hlen)) {
2756 th = skb_gro_header_slow(skb, hlen, off);
2757 if (unlikely(!th))
2758 goto out;
2759 }
2760
2761 skb_gro_pull(skb, thlen);
2762
2763 len = skb_gro_len(skb);
2764 flags = tcp_flag_word(th);
2765
2766 for (; (p = *head); head = &p->next) {
2767 if (!NAPI_GRO_CB(p)->same_flow)
2768 continue;
2769
2770 th2 = tcp_hdr(p);
2771
2772 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2773 NAPI_GRO_CB(p)->same_flow = 0;
2774 continue;
2775 }
2776
2777 goto found;
2778 }
2779
2780 goto out_check_final;
2781
2782 found:
2783 flush = NAPI_GRO_CB(p)->flush;
2784 flush |= (__force int)(flags & TCP_FLAG_CWR);
2785 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2786 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2787 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2788 for (i = sizeof(*th); i < thlen; i += 4)
2789 flush |= *(u32 *)((u8 *)th + i) ^
2790 *(u32 *)((u8 *)th2 + i);
2791
2792 mss = skb_shinfo(p)->gso_size;
2793
2794 flush |= (len - 1) >= mss;
2795 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2796
2797 if (flush || skb_gro_receive(head, skb)) {
2798 mss = 1;
2799 goto out_check_final;
2800 }
2801
2802 p = *head;
2803 th2 = tcp_hdr(p);
2804 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2805
2806 out_check_final:
2807 flush = len < mss;
2808 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2809 TCP_FLAG_RST | TCP_FLAG_SYN |
2810 TCP_FLAG_FIN));
2811
2812 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2813 pp = head;
2814
2815 out:
2816 NAPI_GRO_CB(skb)->flush |= flush;
2817
2818 return pp;
2819 }
2820 EXPORT_SYMBOL(tcp_gro_receive);
2821
2822 int tcp_gro_complete(struct sk_buff *skb)
2823 {
2824 struct tcphdr *th = tcp_hdr(skb);
2825
2826 skb->csum_start = skb_transport_header(skb) - skb->head;
2827 skb->csum_offset = offsetof(struct tcphdr, check);
2828 skb->ip_summed = CHECKSUM_PARTIAL;
2829
2830 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2831
2832 if (th->cwr)
2833 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2834
2835 return 0;
2836 }
2837 EXPORT_SYMBOL(tcp_gro_complete);
2838
2839 #ifdef CONFIG_TCP_MD5SIG
2840 static unsigned long tcp_md5sig_users;
2841 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool;
2842 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2843
2844 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool)
2845 {
2846 int cpu;
2847 for_each_possible_cpu(cpu) {
2848 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2849 if (p) {
2850 if (p->md5_desc.tfm)
2851 crypto_free_hash(p->md5_desc.tfm);
2852 kfree(p);
2853 }
2854 }
2855 free_percpu(pool);
2856 }
2857
2858 void tcp_free_md5sig_pool(void)
2859 {
2860 struct tcp_md5sig_pool * __percpu *pool = NULL;
2861
2862 spin_lock_bh(&tcp_md5sig_pool_lock);
2863 if (--tcp_md5sig_users == 0) {
2864 pool = tcp_md5sig_pool;
2865 tcp_md5sig_pool = NULL;
2866 }
2867 spin_unlock_bh(&tcp_md5sig_pool_lock);
2868 if (pool)
2869 __tcp_free_md5sig_pool(pool);
2870 }
2871 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2872
2873 static struct tcp_md5sig_pool * __percpu *
2874 __tcp_alloc_md5sig_pool(struct sock *sk)
2875 {
2876 int cpu;
2877 struct tcp_md5sig_pool * __percpu *pool;
2878
2879 pool = alloc_percpu(struct tcp_md5sig_pool *);
2880 if (!pool)
2881 return NULL;
2882
2883 for_each_possible_cpu(cpu) {
2884 struct tcp_md5sig_pool *p;
2885 struct crypto_hash *hash;
2886
2887 p = kzalloc(sizeof(*p), sk->sk_allocation);
2888 if (!p)
2889 goto out_free;
2890 *per_cpu_ptr(pool, cpu) = p;
2891
2892 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2893 if (!hash || IS_ERR(hash))
2894 goto out_free;
2895
2896 p->md5_desc.tfm = hash;
2897 }
2898 return pool;
2899 out_free:
2900 __tcp_free_md5sig_pool(pool);
2901 return NULL;
2902 }
2903
2904 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2905 {
2906 struct tcp_md5sig_pool * __percpu *pool;
2907 int alloc = 0;
2908
2909 retry:
2910 spin_lock_bh(&tcp_md5sig_pool_lock);
2911 pool = tcp_md5sig_pool;
2912 if (tcp_md5sig_users++ == 0) {
2913 alloc = 1;
2914 spin_unlock_bh(&tcp_md5sig_pool_lock);
2915 } else if (!pool) {
2916 tcp_md5sig_users--;
2917 spin_unlock_bh(&tcp_md5sig_pool_lock);
2918 cpu_relax();
2919 goto retry;
2920 } else
2921 spin_unlock_bh(&tcp_md5sig_pool_lock);
2922
2923 if (alloc) {
2924 /* we cannot hold spinlock here because this may sleep. */
2925 struct tcp_md5sig_pool * __percpu *p;
2926
2927 p = __tcp_alloc_md5sig_pool(sk);
2928 spin_lock_bh(&tcp_md5sig_pool_lock);
2929 if (!p) {
2930 tcp_md5sig_users--;
2931 spin_unlock_bh(&tcp_md5sig_pool_lock);
2932 return NULL;
2933 }
2934 pool = tcp_md5sig_pool;
2935 if (pool) {
2936 /* oops, it has already been assigned. */
2937 spin_unlock_bh(&tcp_md5sig_pool_lock);
2938 __tcp_free_md5sig_pool(p);
2939 } else {
2940 tcp_md5sig_pool = pool = p;
2941 spin_unlock_bh(&tcp_md5sig_pool_lock);
2942 }
2943 }
2944 return pool;
2945 }
2946 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2947
2948
2949 /**
2950 * tcp_get_md5sig_pool - get md5sig_pool for this user
2951 *
2952 * We use percpu structure, so if we succeed, we exit with preemption
2953 * and BH disabled, to make sure another thread or softirq handling
2954 * wont try to get same context.
2955 */
2956 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2957 {
2958 struct tcp_md5sig_pool * __percpu *p;
2959
2960 local_bh_disable();
2961
2962 spin_lock(&tcp_md5sig_pool_lock);
2963 p = tcp_md5sig_pool;
2964 if (p)
2965 tcp_md5sig_users++;
2966 spin_unlock(&tcp_md5sig_pool_lock);
2967
2968 if (p)
2969 return *this_cpu_ptr(p);
2970
2971 local_bh_enable();
2972 return NULL;
2973 }
2974 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2975
2976 void tcp_put_md5sig_pool(void)
2977 {
2978 local_bh_enable();
2979 tcp_free_md5sig_pool();
2980 }
2981 EXPORT_SYMBOL(tcp_put_md5sig_pool);
2982
2983 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2984 struct tcphdr *th)
2985 {
2986 struct scatterlist sg;
2987 int err;
2988
2989 __sum16 old_checksum = th->check;
2990 th->check = 0;
2991 /* options aren't included in the hash */
2992 sg_init_one(&sg, th, sizeof(struct tcphdr));
2993 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2994 th->check = old_checksum;
2995 return err;
2996 }
2997 EXPORT_SYMBOL(tcp_md5_hash_header);
2998
2999 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3000 struct sk_buff *skb, unsigned header_len)
3001 {
3002 struct scatterlist sg;
3003 const struct tcphdr *tp = tcp_hdr(skb);
3004 struct hash_desc *desc = &hp->md5_desc;
3005 unsigned i;
3006 const unsigned head_data_len = skb_headlen(skb) > header_len ?
3007 skb_headlen(skb) - header_len : 0;
3008 const struct skb_shared_info *shi = skb_shinfo(skb);
3009 struct sk_buff *frag_iter;
3010
3011 sg_init_table(&sg, 1);
3012
3013 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3014 if (crypto_hash_update(desc, &sg, head_data_len))
3015 return 1;
3016
3017 for (i = 0; i < shi->nr_frags; ++i) {
3018 const struct skb_frag_struct *f = &shi->frags[i];
3019 sg_set_page(&sg, f->page, f->size, f->page_offset);
3020 if (crypto_hash_update(desc, &sg, f->size))
3021 return 1;
3022 }
3023
3024 skb_walk_frags(skb, frag_iter)
3025 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3026 return 1;
3027
3028 return 0;
3029 }
3030 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3031
3032 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
3033 {
3034 struct scatterlist sg;
3035
3036 sg_init_one(&sg, key->key, key->keylen);
3037 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3038 }
3039 EXPORT_SYMBOL(tcp_md5_hash_key);
3040
3041 #endif
3042
3043 /**
3044 * Each Responder maintains up to two secret values concurrently for
3045 * efficient secret rollover. Each secret value has 4 states:
3046 *
3047 * Generating. (tcp_secret_generating != tcp_secret_primary)
3048 * Generates new Responder-Cookies, but not yet used for primary
3049 * verification. This is a short-term state, typically lasting only
3050 * one round trip time (RTT).
3051 *
3052 * Primary. (tcp_secret_generating == tcp_secret_primary)
3053 * Used both for generation and primary verification.
3054 *
3055 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3056 * Used for verification, until the first failure that can be
3057 * verified by the newer Generating secret. At that time, this
3058 * cookie's state is changed to Secondary, and the Generating
3059 * cookie's state is changed to Primary. This is a short-term state,
3060 * typically lasting only one round trip time (RTT).
3061 *
3062 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3063 * Used for secondary verification, after primary verification
3064 * failures. This state lasts no more than twice the Maximum Segment
3065 * Lifetime (2MSL). Then, the secret is discarded.
3066 */
3067 struct tcp_cookie_secret {
3068 /* The secret is divided into two parts. The digest part is the
3069 * equivalent of previously hashing a secret and saving the state,
3070 * and serves as an initialization vector (IV). The message part
3071 * serves as the trailing secret.
3072 */
3073 u32 secrets[COOKIE_WORKSPACE_WORDS];
3074 unsigned long expires;
3075 };
3076
3077 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3078 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3079 #define TCP_SECRET_LIFE (HZ * 600)
3080
3081 static struct tcp_cookie_secret tcp_secret_one;
3082 static struct tcp_cookie_secret tcp_secret_two;
3083
3084 /* Essentially a circular list, without dynamic allocation. */
3085 static struct tcp_cookie_secret *tcp_secret_generating;
3086 static struct tcp_cookie_secret *tcp_secret_primary;
3087 static struct tcp_cookie_secret *tcp_secret_retiring;
3088 static struct tcp_cookie_secret *tcp_secret_secondary;
3089
3090 static DEFINE_SPINLOCK(tcp_secret_locker);
3091
3092 /* Select a pseudo-random word in the cookie workspace.
3093 */
3094 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3095 {
3096 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3097 }
3098
3099 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3100 * Called in softirq context.
3101 * Returns: 0 for success.
3102 */
3103 int tcp_cookie_generator(u32 *bakery)
3104 {
3105 unsigned long jiffy = jiffies;
3106
3107 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3108 spin_lock_bh(&tcp_secret_locker);
3109 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3110 /* refreshed by another */
3111 memcpy(bakery,
3112 &tcp_secret_generating->secrets[0],
3113 COOKIE_WORKSPACE_WORDS);
3114 } else {
3115 /* still needs refreshing */
3116 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3117
3118 /* The first time, paranoia assumes that the
3119 * randomization function isn't as strong. But,
3120 * this secret initialization is delayed until
3121 * the last possible moment (packet arrival).
3122 * Although that time is observable, it is
3123 * unpredictably variable. Mash in the most
3124 * volatile clock bits available, and expire the
3125 * secret extra quickly.
3126 */
3127 if (unlikely(tcp_secret_primary->expires ==
3128 tcp_secret_secondary->expires)) {
3129 struct timespec tv;
3130
3131 getnstimeofday(&tv);
3132 bakery[COOKIE_DIGEST_WORDS+0] ^=
3133 (u32)tv.tv_nsec;
3134
3135 tcp_secret_secondary->expires = jiffy
3136 + TCP_SECRET_1MSL
3137 + (0x0f & tcp_cookie_work(bakery, 0));
3138 } else {
3139 tcp_secret_secondary->expires = jiffy
3140 + TCP_SECRET_LIFE
3141 + (0xff & tcp_cookie_work(bakery, 1));
3142 tcp_secret_primary->expires = jiffy
3143 + TCP_SECRET_2MSL
3144 + (0x1f & tcp_cookie_work(bakery, 2));
3145 }
3146 memcpy(&tcp_secret_secondary->secrets[0],
3147 bakery, COOKIE_WORKSPACE_WORDS);
3148
3149 rcu_assign_pointer(tcp_secret_generating,
3150 tcp_secret_secondary);
3151 rcu_assign_pointer(tcp_secret_retiring,
3152 tcp_secret_primary);
3153 /*
3154 * Neither call_rcu() nor synchronize_rcu() needed.
3155 * Retiring data is not freed. It is replaced after
3156 * further (locked) pointer updates, and a quiet time
3157 * (minimum 1MSL, maximum LIFE - 2MSL).
3158 */
3159 }
3160 spin_unlock_bh(&tcp_secret_locker);
3161 } else {
3162 rcu_read_lock_bh();
3163 memcpy(bakery,
3164 &rcu_dereference(tcp_secret_generating)->secrets[0],
3165 COOKIE_WORKSPACE_WORDS);
3166 rcu_read_unlock_bh();
3167 }
3168 return 0;
3169 }
3170 EXPORT_SYMBOL(tcp_cookie_generator);
3171
3172 void tcp_done(struct sock *sk)
3173 {
3174 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3175 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3176
3177 tcp_set_state(sk, TCP_CLOSE);
3178 tcp_clear_xmit_timers(sk);
3179
3180 sk->sk_shutdown = SHUTDOWN_MASK;
3181
3182 if (!sock_flag(sk, SOCK_DEAD))
3183 sk->sk_state_change(sk);
3184 else
3185 inet_csk_destroy_sock(sk);
3186 }
3187 EXPORT_SYMBOL_GPL(tcp_done);
3188
3189 extern struct tcp_congestion_ops tcp_reno;
3190
3191 static __initdata unsigned long thash_entries;
3192 static int __init set_thash_entries(char *str)
3193 {
3194 if (!str)
3195 return 0;
3196 thash_entries = simple_strtoul(str, &str, 0);
3197 return 1;
3198 }
3199 __setup("thash_entries=", set_thash_entries);
3200
3201 void __init tcp_init(void)
3202 {
3203 struct sk_buff *skb = NULL;
3204 unsigned long nr_pages, limit;
3205 int order, i, max_share;
3206 unsigned long jiffy = jiffies;
3207
3208 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3209
3210 percpu_counter_init(&tcp_sockets_allocated, 0);
3211 percpu_counter_init(&tcp_orphan_count, 0);
3212 tcp_hashinfo.bind_bucket_cachep =
3213 kmem_cache_create("tcp_bind_bucket",
3214 sizeof(struct inet_bind_bucket), 0,
3215 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3216
3217 /* Size and allocate the main established and bind bucket
3218 * hash tables.
3219 *
3220 * The methodology is similar to that of the buffer cache.
3221 */
3222 tcp_hashinfo.ehash =
3223 alloc_large_system_hash("TCP established",
3224 sizeof(struct inet_ehash_bucket),
3225 thash_entries,
3226 (totalram_pages >= 128 * 1024) ?
3227 13 : 15,
3228 0,
3229 NULL,
3230 &tcp_hashinfo.ehash_mask,
3231 thash_entries ? 0 : 512 * 1024);
3232 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3233 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3234 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3235 }
3236 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3237 panic("TCP: failed to alloc ehash_locks");
3238 tcp_hashinfo.bhash =
3239 alloc_large_system_hash("TCP bind",
3240 sizeof(struct inet_bind_hashbucket),
3241 tcp_hashinfo.ehash_mask + 1,
3242 (totalram_pages >= 128 * 1024) ?
3243 13 : 15,
3244 0,
3245 &tcp_hashinfo.bhash_size,
3246 NULL,
3247 64 * 1024);
3248 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3249 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3250 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3251 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3252 }
3253
3254 /* Try to be a bit smarter and adjust defaults depending
3255 * on available memory.
3256 */
3257 for (order = 0; ((1 << order) << PAGE_SHIFT) <
3258 (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
3259 order++)
3260 ;
3261 if (order >= 4) {
3262 tcp_death_row.sysctl_max_tw_buckets = 180000;
3263 sysctl_tcp_max_orphans = 4096 << (order - 4);
3264 sysctl_max_syn_backlog = 1024;
3265 } else if (order < 3) {
3266 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
3267 sysctl_tcp_max_orphans >>= (3 - order);
3268 sysctl_max_syn_backlog = 128;
3269 }
3270
3271 /* Set the pressure threshold to be a fraction of global memory that
3272 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
3273 * memory, with a floor of 128 pages.
3274 */
3275 nr_pages = totalram_pages - totalhigh_pages;
3276 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
3277 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
3278 limit = max(limit, 128UL);
3279 sysctl_tcp_mem[0] = limit / 4 * 3;
3280 sysctl_tcp_mem[1] = limit;
3281 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3282
3283 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3284 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3285 max_share = min(4UL*1024*1024, limit);
3286
3287 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3288 sysctl_tcp_wmem[1] = 16*1024;
3289 sysctl_tcp_wmem[2] = max(64*1024, max_share);
3290
3291 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3292 sysctl_tcp_rmem[1] = 87380;
3293 sysctl_tcp_rmem[2] = max(87380, max_share);
3294
3295 printk(KERN_INFO "TCP: Hash tables configured "
3296 "(established %u bind %u)\n",
3297 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3298
3299 tcp_register_congestion_control(&tcp_reno);
3300
3301 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3302 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3303 tcp_secret_one.expires = jiffy; /* past due */
3304 tcp_secret_two.expires = jiffy; /* past due */
3305 tcp_secret_generating = &tcp_secret_one;
3306 tcp_secret_primary = &tcp_secret_one;
3307 tcp_secret_retiring = &tcp_secret_two;
3308 tcp_secret_secondary = &tcp_secret_two;
3309 }
Linux kernel - Deliverables
This page took 0.098259 seconds and 6 git commands to generate.