?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
writeback: split writeback_inodes_wb
[linux-2.6/next.git] / net / ipv4 / tcp.c
blob6596b4feeddc7879fc020606f53dabdc8608f3e8
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
319 EXPORT_SYMBOL(tcp_memory_pressure);
320
321 void tcp_enter_memory_pressure(struct sock *sk)
322 {
323 if (!tcp_memory_pressure) {
324 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
325 tcp_memory_pressure = 1;
326 }
327 }
328
329 EXPORT_SYMBOL(tcp_enter_memory_pressure);
330
331 /* Convert seconds to retransmits based on initial and max timeout */
332 static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
333 {
334 u8 res = 0;
335
336 if (seconds > 0) {
337 int period = timeout;
338
339 res = 1;
340 while (seconds > period && res < 255) {
341 res++;
342 timeout <<= 1;
343 if (timeout > rto_max)
344 timeout = rto_max;
345 period += timeout;
346 }
347 }
348 return res;
349 }
350
351 /* Convert retransmits to seconds based on initial and max timeout */
352 static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
353 {
354 int period = 0;
355
356 if (retrans > 0) {
357 period = timeout;
358 while (--retrans) {
359 timeout <<= 1;
360 if (timeout > rto_max)
361 timeout = rto_max;
362 period += timeout;
363 }
364 }
365 return period;
366 }
367
368 /*
369 * Wait for a TCP event.
370 *
371 * Note that we don't need to lock the socket, as the upper poll layers
372 * take care of normal races (between the test and the event) and we don't
373 * go look at any of the socket buffers directly.
374 */
375 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
376 {
377 unsigned int mask;
378 struct sock *sk = sock->sk;
379 struct tcp_sock *tp = tcp_sk(sk);
380
381 sock_poll_wait(file, sk_sleep(sk), wait);
382 if (sk->sk_state == TCP_LISTEN)
383 return inet_csk_listen_poll(sk);
384
385 /* Socket is not locked. We are protected from async events
386 * by poll logic and correct handling of state changes
387 * made by other threads is impossible in any case.
388 */
389
390 mask = 0;
391 if (sk->sk_err)
392 mask = POLLERR;
393
394 /*
395 * POLLHUP is certainly not done right. But poll() doesn't
396 * have a notion of HUP in just one direction, and for a
397 * socket the read side is more interesting.
398 *
399 * Some poll() documentation says that POLLHUP is incompatible
400 * with the POLLOUT/POLLWR flags, so somebody should check this
401 * all. But careful, it tends to be safer to return too many
402 * bits than too few, and you can easily break real applications
403 * if you don't tell them that something has hung up!
404 *
405 * Check-me.
406 *
407 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
408 * our fs/select.c). It means that after we received EOF,
409 * poll always returns immediately, making impossible poll() on write()
410 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
411 * if and only if shutdown has been made in both directions.
412 * Actually, it is interesting to look how Solaris and DUX
413 * solve this dilemma. I would prefer, if POLLHUP were maskable,
414 * then we could set it on SND_SHUTDOWN. BTW examples given
415 * in Stevens' books assume exactly this behaviour, it explains
416 * why POLLHUP is incompatible with POLLOUT. --ANK
417 *
418 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
419 * blocking on fresh not-connected or disconnected socket. --ANK
420 */
421 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
422 mask |= POLLHUP;
423 if (sk->sk_shutdown & RCV_SHUTDOWN)
424 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
425
426 /* Connected? */
427 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
428 int target = sock_rcvlowat(sk, 0, INT_MAX);
429
430 if (tp->urg_seq == tp->copied_seq &&
431 !sock_flag(sk, SOCK_URGINLINE) &&
432 tp->urg_data)
433 target++;
434
435 /* Potential race condition. If read of tp below will
436 * escape above sk->sk_state, we can be illegally awaken
437 * in SYN_* states. */
438 if (tp->rcv_nxt - tp->copied_seq >= target)
439 mask |= POLLIN | POLLRDNORM;
440
441 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
442 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
443 mask |= POLLOUT | POLLWRNORM;
444 } else { /* send SIGIO later */
445 set_bit(SOCK_ASYNC_NOSPACE,
446 &sk->sk_socket->flags);
447 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
448
449 /* Race breaker. If space is freed after
450 * wspace test but before the flags are set,
451 * IO signal will be lost.
452 */
453 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
454 mask |= POLLOUT | POLLWRNORM;
455 }
456 }
457
458 if (tp->urg_data & TCP_URG_VALID)
459 mask |= POLLPRI;
460 }
461 return mask;
462 }
463
464 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
465 {
466 struct tcp_sock *tp = tcp_sk(sk);
467 int answ;
468
469 switch (cmd) {
470 case SIOCINQ:
471 if (sk->sk_state == TCP_LISTEN)
472 return -EINVAL;
473
474 lock_sock(sk);
475 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
476 answ = 0;
477 else if (sock_flag(sk, SOCK_URGINLINE) ||
478 !tp->urg_data ||
479 before(tp->urg_seq, tp->copied_seq) ||
480 !before(tp->urg_seq, tp->rcv_nxt)) {
481 struct sk_buff *skb;
482
483 answ = tp->rcv_nxt - tp->copied_seq;
484
485 /* Subtract 1, if FIN is in queue. */
486 skb = skb_peek_tail(&sk->sk_receive_queue);
487 if (answ && skb)
488 answ -= tcp_hdr(skb)->fin;
489 } else
490 answ = tp->urg_seq - tp->copied_seq;
491 release_sock(sk);
492 break;
493 case SIOCATMARK:
494 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
495 break;
496 case SIOCOUTQ:
497 if (sk->sk_state == TCP_LISTEN)
498 return -EINVAL;
499
500 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
501 answ = 0;
502 else
503 answ = tp->write_seq - tp->snd_una;
504 break;
505 default:
506 return -ENOIOCTLCMD;
507 }
508
509 return put_user(answ, (int __user *)arg);
510 }
511
512 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
513 {
514 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_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 = TCPCB_FLAG_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
679 struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
680 {
681 struct sk_buff *skb;
682
683 /* The TCP header must be at least 32-bit aligned. */
684 size = ALIGN(size, 4);
685
686 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
687 if (skb) {
688 if (sk_wmem_schedule(sk, skb->truesize)) {
689 /*
690 * Make sure that we have exactly size bytes
691 * available to the caller, no more, no less.
692 */
693 skb_reserve(skb, skb_tailroom(skb) - size);
694 return skb;
695 }
696 __kfree_skb(skb);
697 } else {
698 sk->sk_prot->enter_memory_pressure(sk);
699 sk_stream_moderate_sndbuf(sk);
700 }
701 return NULL;
702 }
703
704 static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
705 int large_allowed)
706 {
707 struct tcp_sock *tp = tcp_sk(sk);
708 u32 xmit_size_goal, old_size_goal;
709
710 xmit_size_goal = mss_now;
711
712 if (large_allowed && sk_can_gso(sk)) {
713 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
714 inet_csk(sk)->icsk_af_ops->net_header_len -
715 inet_csk(sk)->icsk_ext_hdr_len -
716 tp->tcp_header_len);
717
718 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
719
720 /* We try hard to avoid divides here */
721 old_size_goal = tp->xmit_size_goal_segs * mss_now;
722
723 if (likely(old_size_goal <= xmit_size_goal &&
724 old_size_goal + mss_now > xmit_size_goal)) {
725 xmit_size_goal = old_size_goal;
726 } else {
727 tp->xmit_size_goal_segs = xmit_size_goal / mss_now;
728 xmit_size_goal = tp->xmit_size_goal_segs * mss_now;
729 }
730 }
731
732 return max(xmit_size_goal, mss_now);
733 }
734
735 static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
736 {
737 int mss_now;
738
739 mss_now = tcp_current_mss(sk);
740 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
741
742 return mss_now;
743 }
744
745 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
746 size_t psize, int flags)
747 {
748 struct tcp_sock *tp = tcp_sk(sk);
749 int mss_now, size_goal;
750 int err;
751 ssize_t copied;
752 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
753
754 /* Wait for a connection to finish. */
755 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
756 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
757 goto out_err;
758
759 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
760
761 mss_now = tcp_send_mss(sk, &size_goal, flags);
762 copied = 0;
763
764 err = -EPIPE;
765 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
766 goto out_err;
767
768 while (psize > 0) {
769 struct sk_buff *skb = tcp_write_queue_tail(sk);
770 struct page *page = pages[poffset / PAGE_SIZE];
771 int copy, i, can_coalesce;
772 int offset = poffset % PAGE_SIZE;
773 int size = min_t(size_t, psize, PAGE_SIZE - offset);
774
775 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
776 new_segment:
777 if (!sk_stream_memory_free(sk))
778 goto wait_for_sndbuf;
779
780 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
781 if (!skb)
782 goto wait_for_memory;
783
784 skb_entail(sk, skb);
785 copy = size_goal;
786 }
787
788 if (copy > size)
789 copy = size;
790
791 i = skb_shinfo(skb)->nr_frags;
792 can_coalesce = skb_can_coalesce(skb, i, page, offset);
793 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
794 tcp_mark_push(tp, skb);
795 goto new_segment;
796 }
797 if (!sk_wmem_schedule(sk, copy))
798 goto wait_for_memory;
799
800 if (can_coalesce) {
801 skb_shinfo(skb)->frags[i - 1].size += copy;
802 } else {
803 get_page(page);
804 skb_fill_page_desc(skb, i, page, offset, copy);
805 }
806
807 skb->len += copy;
808 skb->data_len += copy;
809 skb->truesize += copy;
810 sk->sk_wmem_queued += copy;
811 sk_mem_charge(sk, copy);
812 skb->ip_summed = CHECKSUM_PARTIAL;
813 tp->write_seq += copy;
814 TCP_SKB_CB(skb)->end_seq += copy;
815 skb_shinfo(skb)->gso_segs = 0;
816
817 if (!copied)
818 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
819
820 copied += copy;
821 poffset += copy;
822 if (!(psize -= copy))
823 goto out;
824
825 if (skb->len < size_goal || (flags & MSG_OOB))
826 continue;
827
828 if (forced_push(tp)) {
829 tcp_mark_push(tp, skb);
830 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
831 } else if (skb == tcp_send_head(sk))
832 tcp_push_one(sk, mss_now);
833 continue;
834
835 wait_for_sndbuf:
836 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
837 wait_for_memory:
838 if (copied)
839 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
840
841 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
842 goto do_error;
843
844 mss_now = tcp_send_mss(sk, &size_goal, flags);
845 }
846
847 out:
848 if (copied)
849 tcp_push(sk, flags, mss_now, tp->nonagle);
850 return copied;
851
852 do_error:
853 if (copied)
854 goto out;
855 out_err:
856 return sk_stream_error(sk, flags, err);
857 }
858
859 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
860 size_t size, int flags)
861 {
862 ssize_t res;
863 struct sock *sk = sock->sk;
864
865 if (!(sk->sk_route_caps & NETIF_F_SG) ||
866 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
867 return sock_no_sendpage(sock, page, offset, size, flags);
868
869 lock_sock(sk);
870 TCP_CHECK_TIMER(sk);
871 res = do_tcp_sendpages(sk, &page, offset, size, flags);
872 TCP_CHECK_TIMER(sk);
873 release_sock(sk);
874 return res;
875 }
876
877 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
878 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
879
880 static inline int select_size(struct sock *sk, int sg)
881 {
882 struct tcp_sock *tp = tcp_sk(sk);
883 int tmp = tp->mss_cache;
884
885 if (sg) {
886 if (sk_can_gso(sk))
887 tmp = 0;
888 else {
889 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
890
891 if (tmp >= pgbreak &&
892 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
893 tmp = pgbreak;
894 }
895 }
896
897 return tmp;
898 }
899
900 int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
901 size_t size)
902 {
903 struct sock *sk = sock->sk;
904 struct iovec *iov;
905 struct tcp_sock *tp = tcp_sk(sk);
906 struct sk_buff *skb;
907 int iovlen, flags;
908 int mss_now, size_goal;
909 int sg, err, copied;
910 long timeo;
911
912 lock_sock(sk);
913 TCP_CHECK_TIMER(sk);
914
915 flags = msg->msg_flags;
916 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
917
918 /* Wait for a connection to finish. */
919 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
920 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
921 goto out_err;
922
923 /* This should be in poll */
924 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
925
926 mss_now = tcp_send_mss(sk, &size_goal, flags);
927
928 /* Ok commence sending. */
929 iovlen = msg->msg_iovlen;
930 iov = msg->msg_iov;
931 copied = 0;
932
933 err = -EPIPE;
934 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
935 goto out_err;
936
937 sg = sk->sk_route_caps & NETIF_F_SG;
938
939 while (--iovlen >= 0) {
940 int seglen = iov->iov_len;
941 unsigned char __user *from = iov->iov_base;
942
943 iov++;
944
945 while (seglen > 0) {
946 int copy = 0;
947 int max = size_goal;
948
949 skb = tcp_write_queue_tail(sk);
950 if (tcp_send_head(sk)) {
951 if (skb->ip_summed == CHECKSUM_NONE)
952 max = mss_now;
953 copy = max - skb->len;
954 }
955
956 if (copy <= 0) {
957 new_segment:
958 /* Allocate new segment. If the interface is SG,
959 * allocate skb fitting to single page.
960 */
961 if (!sk_stream_memory_free(sk))
962 goto wait_for_sndbuf;
963
964 skb = sk_stream_alloc_skb(sk,
965 select_size(sk, sg),
966 sk->sk_allocation);
967 if (!skb)
968 goto wait_for_memory;
969
970 /*
971 * Check whether we can use HW checksum.
972 */
973 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
974 skb->ip_summed = CHECKSUM_PARTIAL;
975
976 skb_entail(sk, skb);
977 copy = size_goal;
978 max = size_goal;
979 }
980
981 /* Try to append data to the end of skb. */
982 if (copy > seglen)
983 copy = seglen;
984
985 /* Where to copy to? */
986 if (skb_tailroom(skb) > 0) {
987 /* We have some space in skb head. Superb! */
988 if (copy > skb_tailroom(skb))
989 copy = skb_tailroom(skb);
990 if ((err = skb_add_data(skb, from, copy)) != 0)
991 goto do_fault;
992 } else {
993 int merge = 0;
994 int i = skb_shinfo(skb)->nr_frags;
995 struct page *page = TCP_PAGE(sk);
996 int off = TCP_OFF(sk);
997
998 if (skb_can_coalesce(skb, i, page, off) &&
999 off != PAGE_SIZE) {
1000 /* We can extend the last page
1001 * fragment. */
1002 merge = 1;
1003 } else if (i == MAX_SKB_FRAGS || !sg) {
1004 /* Need to add new fragment and cannot
1005 * do this because interface is non-SG,
1006 * or because all the page slots are
1007 * busy. */
1008 tcp_mark_push(tp, skb);
1009 goto new_segment;
1010 } else if (page) {
1011 if (off == PAGE_SIZE) {
1012 put_page(page);
1013 TCP_PAGE(sk) = page = NULL;
1014 off = 0;
1015 }
1016 } else
1017 off = 0;
1018
1019 if (copy > PAGE_SIZE - off)
1020 copy = PAGE_SIZE - off;
1021
1022 if (!sk_wmem_schedule(sk, copy))
1023 goto wait_for_memory;
1024
1025 if (!page) {
1026 /* Allocate new cache page. */
1027 if (!(page = sk_stream_alloc_page(sk)))
1028 goto wait_for_memory;
1029 }
1030
1031 /* Time to copy data. We are close to
1032 * the end! */
1033 err = skb_copy_to_page(sk, from, skb, page,
1034 off, copy);
1035 if (err) {
1036 /* If this page was new, give it to the
1037 * socket so it does not get leaked.
1038 */
1039 if (!TCP_PAGE(sk)) {
1040 TCP_PAGE(sk) = page;
1041 TCP_OFF(sk) = 0;
1042 }
1043 goto do_error;
1044 }
1045
1046 /* Update the skb. */
1047 if (merge) {
1048 skb_shinfo(skb)->frags[i - 1].size +=
1049 copy;
1050 } else {
1051 skb_fill_page_desc(skb, i, page, off, copy);
1052 if (TCP_PAGE(sk)) {
1053 get_page(page);
1054 } else if (off + copy < PAGE_SIZE) {
1055 get_page(page);
1056 TCP_PAGE(sk) = page;
1057 }
1058 }
1059
1060 TCP_OFF(sk) = off + copy;
1061 }
1062
1063 if (!copied)
1064 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
1065
1066 tp->write_seq += copy;
1067 TCP_SKB_CB(skb)->end_seq += copy;
1068 skb_shinfo(skb)->gso_segs = 0;
1069
1070 from += copy;
1071 copied += copy;
1072 if ((seglen -= copy) == 0 && iovlen == 0)
1073 goto out;
1074
1075 if (skb->len < max || (flags & MSG_OOB))
1076 continue;
1077
1078 if (forced_push(tp)) {
1079 tcp_mark_push(tp, skb);
1080 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1081 } else if (skb == tcp_send_head(sk))
1082 tcp_push_one(sk, mss_now);
1083 continue;
1084
1085 wait_for_sndbuf:
1086 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1087 wait_for_memory:
1088 if (copied)
1089 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
1090
1091 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
1092 goto do_error;
1093
1094 mss_now = tcp_send_mss(sk, &size_goal, flags);
1095 }
1096 }
1097
1098 out:
1099 if (copied)
1100 tcp_push(sk, flags, mss_now, tp->nonagle);
1101 TCP_CHECK_TIMER(sk);
1102 release_sock(sk);
1103 return copied;
1104
1105 do_fault:
1106 if (!skb->len) {
1107 tcp_unlink_write_queue(skb, sk);
1108 /* It is the one place in all of TCP, except connection
1109 * reset, where we can be unlinking the send_head.
1110 */
1111 tcp_check_send_head(sk, skb);
1112 sk_wmem_free_skb(sk, skb);
1113 }
1114
1115 do_error:
1116 if (copied)
1117 goto out;
1118 out_err:
1119 err = sk_stream_error(sk, flags, err);
1120 TCP_CHECK_TIMER(sk);
1121 release_sock(sk);
1122 return err;
1123 }
1124
1125 /*
1126 * Handle reading urgent data. BSD has very simple semantics for
1127 * this, no blocking and very strange errors 8)
1128 */
1129
1130 static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1131 {
1132 struct tcp_sock *tp = tcp_sk(sk);
1133
1134 /* No URG data to read. */
1135 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1136 tp->urg_data == TCP_URG_READ)
1137 return -EINVAL; /* Yes this is right ! */
1138
1139 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1140 return -ENOTCONN;
1141
1142 if (tp->urg_data & TCP_URG_VALID) {
1143 int err = 0;
1144 char c = tp->urg_data;
1145
1146 if (!(flags & MSG_PEEK))
1147 tp->urg_data = TCP_URG_READ;
1148
1149 /* Read urgent data. */
1150 msg->msg_flags |= MSG_OOB;
1151
1152 if (len > 0) {
1153 if (!(flags & MSG_TRUNC))
1154 err = memcpy_toiovec(msg->msg_iov, &c, 1);
1155 len = 1;
1156 } else
1157 msg->msg_flags |= MSG_TRUNC;
1158
1159 return err ? -EFAULT : len;
1160 }
1161
1162 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1163 return 0;
1164
1165 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1166 * the available implementations agree in this case:
1167 * this call should never block, independent of the
1168 * blocking state of the socket.
1169 * Mike <pall@rz.uni-karlsruhe.de>
1170 */
1171 return -EAGAIN;
1172 }
1173
1174 /* Clean up the receive buffer for full frames taken by the user,
1175 * then send an ACK if necessary. COPIED is the number of bytes
1176 * tcp_recvmsg has given to the user so far, it speeds up the
1177 * calculation of whether or not we must ACK for the sake of
1178 * a window update.
1179 */
1180 void tcp_cleanup_rbuf(struct sock *sk, int copied)
1181 {
1182 struct tcp_sock *tp = tcp_sk(sk);
1183 int time_to_ack = 0;
1184
1185 #if TCP_DEBUG
1186 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1187
1188 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1189 KERN_INFO "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1190 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1191 #endif
1192
1193 if (inet_csk_ack_scheduled(sk)) {
1194 const struct inet_connection_sock *icsk = inet_csk(sk);
1195 /* Delayed ACKs frequently hit locked sockets during bulk
1196 * receive. */
1197 if (icsk->icsk_ack.blocked ||
1198 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1199 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1200 /*
1201 * If this read emptied read buffer, we send ACK, if
1202 * connection is not bidirectional, user drained
1203 * receive buffer and there was a small segment
1204 * in queue.
1205 */
1206 (copied > 0 &&
1207 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1208 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1209 !icsk->icsk_ack.pingpong)) &&
1210 !atomic_read(&sk->sk_rmem_alloc)))
1211 time_to_ack = 1;
1212 }
1213
1214 /* We send an ACK if we can now advertise a non-zero window
1215 * which has been raised "significantly".
1216 *
1217 * Even if window raised up to infinity, do not send window open ACK
1218 * in states, where we will not receive more. It is useless.
1219 */
1220 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1221 __u32 rcv_window_now = tcp_receive_window(tp);
1222
1223 /* Optimize, __tcp_select_window() is not cheap. */
1224 if (2*rcv_window_now <= tp->window_clamp) {
1225 __u32 new_window = __tcp_select_window(sk);
1226
1227 /* Send ACK now, if this read freed lots of space
1228 * in our buffer. Certainly, new_window is new window.
1229 * We can advertise it now, if it is not less than current one.
1230 * "Lots" means "at least twice" here.
1231 */
1232 if (new_window && new_window >= 2 * rcv_window_now)
1233 time_to_ack = 1;
1234 }
1235 }
1236 if (time_to_ack)
1237 tcp_send_ack(sk);
1238 }
1239
1240 static void tcp_prequeue_process(struct sock *sk)
1241 {
1242 struct sk_buff *skb;
1243 struct tcp_sock *tp = tcp_sk(sk);
1244
1245 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1246
1247 /* RX process wants to run with disabled BHs, though it is not
1248 * necessary */
1249 local_bh_disable();
1250 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1251 sk_backlog_rcv(sk, skb);
1252 local_bh_enable();
1253
1254 /* Clear memory counter. */
1255 tp->ucopy.memory = 0;
1256 }
1257
1258 #ifdef CONFIG_NET_DMA
1259 static void tcp_service_net_dma(struct sock *sk, bool wait)
1260 {
1261 dma_cookie_t done, used;
1262 dma_cookie_t last_issued;
1263 struct tcp_sock *tp = tcp_sk(sk);
1264
1265 if (!tp->ucopy.dma_chan)
1266 return;
1267
1268 last_issued = tp->ucopy.dma_cookie;
1269 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1270
1271 do {
1272 if (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1273 last_issued, &done,
1274 &used) == DMA_SUCCESS) {
1275 /* Safe to free early-copied skbs now */
1276 __skb_queue_purge(&sk->sk_async_wait_queue);
1277 break;
1278 } else {
1279 struct sk_buff *skb;
1280 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1281 (dma_async_is_complete(skb->dma_cookie, done,
1282 used) == DMA_SUCCESS)) {
1283 __skb_dequeue(&sk->sk_async_wait_queue);
1284 kfree_skb(skb);
1285 }
1286 }
1287 } while (wait);
1288 }
1289 #endif
1290
1291 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1292 {
1293 struct sk_buff *skb;
1294 u32 offset;
1295
1296 skb_queue_walk(&sk->sk_receive_queue, skb) {
1297 offset = seq - TCP_SKB_CB(skb)->seq;
1298 if (tcp_hdr(skb)->syn)
1299 offset--;
1300 if (offset < skb->len || tcp_hdr(skb)->fin) {
1301 *off = offset;
1302 return skb;
1303 }
1304 }
1305 return NULL;
1306 }
1307
1308 /*
1309 * This routine provides an alternative to tcp_recvmsg() for routines
1310 * that would like to handle copying from skbuffs directly in 'sendfile'
1311 * fashion.
1312 * Note:
1313 * - It is assumed that the socket was locked by the caller.
1314 * - The routine does not block.
1315 * - At present, there is no support for reading OOB data
1316 * or for 'peeking' the socket using this routine
1317 * (although both would be easy to implement).
1318 */
1319 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1320 sk_read_actor_t recv_actor)
1321 {
1322 struct sk_buff *skb;
1323 struct tcp_sock *tp = tcp_sk(sk);
1324 u32 seq = tp->copied_seq;
1325 u32 offset;
1326 int copied = 0;
1327
1328 if (sk->sk_state == TCP_LISTEN)
1329 return -ENOTCONN;
1330 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1331 if (offset < skb->len) {
1332 int used;
1333 size_t len;
1334
1335 len = skb->len - offset;
1336 /* Stop reading if we hit a patch of urgent data */
1337 if (tp->urg_data) {
1338 u32 urg_offset = tp->urg_seq - seq;
1339 if (urg_offset < len)
1340 len = urg_offset;
1341 if (!len)
1342 break;
1343 }
1344 used = recv_actor(desc, skb, offset, len);
1345 if (used < 0) {
1346 if (!copied)
1347 copied = used;
1348 break;
1349 } else if (used <= len) {
1350 seq += used;
1351 copied += used;
1352 offset += used;
1353 }
1354 /*
1355 * If recv_actor drops the lock (e.g. TCP splice
1356 * receive) the skb pointer might be invalid when
1357 * getting here: tcp_collapse might have deleted it
1358 * while aggregating skbs from the socket queue.
1359 */
1360 skb = tcp_recv_skb(sk, seq-1, &offset);
1361 if (!skb || (offset+1 != skb->len))
1362 break;
1363 }
1364 if (tcp_hdr(skb)->fin) {
1365 sk_eat_skb(sk, skb, 0);
1366 ++seq;
1367 break;
1368 }
1369 sk_eat_skb(sk, skb, 0);
1370 if (!desc->count)
1371 break;
1372 tp->copied_seq = seq;
1373 }
1374 tp->copied_seq = seq;
1375
1376 tcp_rcv_space_adjust(sk);
1377
1378 /* Clean up data we have read: This will do ACK frames. */
1379 if (copied > 0)
1380 tcp_cleanup_rbuf(sk, copied);
1381 return copied;
1382 }
1383
1384 /*
1385 * This routine copies from a sock struct into the user buffer.
1386 *
1387 * Technical note: in 2.3 we work on _locked_ socket, so that
1388 * tricks with *seq access order and skb->users are not required.
1389 * Probably, code can be easily improved even more.
1390 */
1391
1392 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1393 size_t len, int nonblock, int flags, int *addr_len)
1394 {
1395 struct tcp_sock *tp = tcp_sk(sk);
1396 int copied = 0;
1397 u32 peek_seq;
1398 u32 *seq;
1399 unsigned long used;
1400 int err;
1401 int target; /* Read at least this many bytes */
1402 long timeo;
1403 struct task_struct *user_recv = NULL;
1404 int copied_early = 0;
1405 struct sk_buff *skb;
1406 u32 urg_hole = 0;
1407
1408 lock_sock(sk);
1409
1410 TCP_CHECK_TIMER(sk);
1411
1412 err = -ENOTCONN;
1413 if (sk->sk_state == TCP_LISTEN)
1414 goto out;
1415
1416 timeo = sock_rcvtimeo(sk, nonblock);
1417
1418 /* Urgent data needs to be handled specially. */
1419 if (flags & MSG_OOB)
1420 goto recv_urg;
1421
1422 seq = &tp->copied_seq;
1423 if (flags & MSG_PEEK) {
1424 peek_seq = tp->copied_seq;
1425 seq = &peek_seq;
1426 }
1427
1428 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1429
1430 #ifdef CONFIG_NET_DMA
1431 tp->ucopy.dma_chan = NULL;
1432 preempt_disable();
1433 skb = skb_peek_tail(&sk->sk_receive_queue);
1434 {
1435 int available = 0;
1436
1437 if (skb)
1438 available = TCP_SKB_CB(skb)->seq + skb->len - (*seq);
1439 if ((available < target) &&
1440 (len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1441 !sysctl_tcp_low_latency &&
1442 dma_find_channel(DMA_MEMCPY)) {
1443 preempt_enable_no_resched();
1444 tp->ucopy.pinned_list =
1445 dma_pin_iovec_pages(msg->msg_iov, len);
1446 } else {
1447 preempt_enable_no_resched();
1448 }
1449 }
1450 #endif
1451
1452 do {
1453 u32 offset;
1454
1455 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1456 if (tp->urg_data && tp->urg_seq == *seq) {
1457 if (copied)
1458 break;
1459 if (signal_pending(current)) {
1460 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1461 break;
1462 }
1463 }
1464
1465 /* Next get a buffer. */
1466
1467 skb_queue_walk(&sk->sk_receive_queue, skb) {
1468 /* Now that we have two receive queues this
1469 * shouldn't happen.
1470 */
1471 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1472 KERN_INFO "recvmsg bug: copied %X "
1473 "seq %X rcvnxt %X fl %X\n", *seq,
1474 TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1475 flags))
1476 break;
1477
1478 offset = *seq - TCP_SKB_CB(skb)->seq;
1479 if (tcp_hdr(skb)->syn)
1480 offset--;
1481 if (offset < skb->len)
1482 goto found_ok_skb;
1483 if (tcp_hdr(skb)->fin)
1484 goto found_fin_ok;
1485 WARN(!(flags & MSG_PEEK), KERN_INFO "recvmsg bug 2: "
1486 "copied %X seq %X rcvnxt %X fl %X\n",
1487 *seq, TCP_SKB_CB(skb)->seq,
1488 tp->rcv_nxt, flags);
1489 }
1490
1491 /* Well, if we have backlog, try to process it now yet. */
1492
1493 if (copied >= target && !sk->sk_backlog.tail)
1494 break;
1495
1496 if (copied) {
1497 if (sk->sk_err ||
1498 sk->sk_state == TCP_CLOSE ||
1499 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1500 !timeo ||
1501 signal_pending(current))
1502 break;
1503 } else {
1504 if (sock_flag(sk, SOCK_DONE))
1505 break;
1506
1507 if (sk->sk_err) {
1508 copied = sock_error(sk);
1509 break;
1510 }
1511
1512 if (sk->sk_shutdown & RCV_SHUTDOWN)
1513 break;
1514
1515 if (sk->sk_state == TCP_CLOSE) {
1516 if (!sock_flag(sk, SOCK_DONE)) {
1517 /* This occurs when user tries to read
1518 * from never connected socket.
1519 */
1520 copied = -ENOTCONN;
1521 break;
1522 }
1523 break;
1524 }
1525
1526 if (!timeo) {
1527 copied = -EAGAIN;
1528 break;
1529 }
1530
1531 if (signal_pending(current)) {
1532 copied = sock_intr_errno(timeo);
1533 break;
1534 }
1535 }
1536
1537 tcp_cleanup_rbuf(sk, copied);
1538
1539 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1540 /* Install new reader */
1541 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1542 user_recv = current;
1543 tp->ucopy.task = user_recv;
1544 tp->ucopy.iov = msg->msg_iov;
1545 }
1546
1547 tp->ucopy.len = len;
1548
1549 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1550 !(flags & (MSG_PEEK | MSG_TRUNC)));
1551
1552 /* Ugly... If prequeue is not empty, we have to
1553 * process it before releasing socket, otherwise
1554 * order will be broken at second iteration.
1555 * More elegant solution is required!!!
1556 *
1557 * Look: we have the following (pseudo)queues:
1558 *
1559 * 1. packets in flight
1560 * 2. backlog
1561 * 3. prequeue
1562 * 4. receive_queue
1563 *
1564 * Each queue can be processed only if the next ones
1565 * are empty. At this point we have empty receive_queue.
1566 * But prequeue _can_ be not empty after 2nd iteration,
1567 * when we jumped to start of loop because backlog
1568 * processing added something to receive_queue.
1569 * We cannot release_sock(), because backlog contains
1570 * packets arrived _after_ prequeued ones.
1571 *
1572 * Shortly, algorithm is clear --- to process all
1573 * the queues in order. We could make it more directly,
1574 * requeueing packets from backlog to prequeue, if
1575 * is not empty. It is more elegant, but eats cycles,
1576 * unfortunately.
1577 */
1578 if (!skb_queue_empty(&tp->ucopy.prequeue))
1579 goto do_prequeue;
1580
1581 /* __ Set realtime policy in scheduler __ */
1582 }
1583
1584 #ifdef CONFIG_NET_DMA
1585 if (tp->ucopy.dma_chan)
1586 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1587 #endif
1588 if (copied >= target) {
1589 /* Do not sleep, just process backlog. */
1590 release_sock(sk);
1591 lock_sock(sk);
1592 } else
1593 sk_wait_data(sk, &timeo);
1594
1595 #ifdef CONFIG_NET_DMA
1596 tcp_service_net_dma(sk, false); /* Don't block */
1597 tp->ucopy.wakeup = 0;
1598 #endif
1599
1600 if (user_recv) {
1601 int chunk;
1602
1603 /* __ Restore normal policy in scheduler __ */
1604
1605 if ((chunk = len - tp->ucopy.len) != 0) {
1606 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1607 len -= chunk;
1608 copied += chunk;
1609 }
1610
1611 if (tp->rcv_nxt == tp->copied_seq &&
1612 !skb_queue_empty(&tp->ucopy.prequeue)) {
1613 do_prequeue:
1614 tcp_prequeue_process(sk);
1615
1616 if ((chunk = len - tp->ucopy.len) != 0) {
1617 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1618 len -= chunk;
1619 copied += chunk;
1620 }
1621 }
1622 }
1623 if ((flags & MSG_PEEK) &&
1624 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1625 if (net_ratelimit())
1626 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1627 current->comm, task_pid_nr(current));
1628 peek_seq = tp->copied_seq;
1629 }
1630 continue;
1631
1632 found_ok_skb:
1633 /* Ok so how much can we use? */
1634 used = skb->len - offset;
1635 if (len < used)
1636 used = len;
1637
1638 /* Do we have urgent data here? */
1639 if (tp->urg_data) {
1640 u32 urg_offset = tp->urg_seq - *seq;
1641 if (urg_offset < used) {
1642 if (!urg_offset) {
1643 if (!sock_flag(sk, SOCK_URGINLINE)) {
1644 ++*seq;
1645 urg_hole++;
1646 offset++;
1647 used--;
1648 if (!used)
1649 goto skip_copy;
1650 }
1651 } else
1652 used = urg_offset;
1653 }
1654 }
1655
1656 if (!(flags & MSG_TRUNC)) {
1657 #ifdef CONFIG_NET_DMA
1658 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1659 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
1660
1661 if (tp->ucopy.dma_chan) {
1662 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1663 tp->ucopy.dma_chan, skb, offset,
1664 msg->msg_iov, used,
1665 tp->ucopy.pinned_list);
1666
1667 if (tp->ucopy.dma_cookie < 0) {
1668
1669 printk(KERN_ALERT "dma_cookie < 0\n");
1670
1671 /* Exception. Bailout! */
1672 if (!copied)
1673 copied = -EFAULT;
1674 break;
1675 }
1676
1677 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1678
1679 if ((offset + used) == skb->len)
1680 copied_early = 1;
1681
1682 } else
1683 #endif
1684 {
1685 err = skb_copy_datagram_iovec(skb, offset,
1686 msg->msg_iov, used);
1687 if (err) {
1688 /* Exception. Bailout! */
1689 if (!copied)
1690 copied = -EFAULT;
1691 break;
1692 }
1693 }
1694 }
1695
1696 *seq += used;
1697 copied += used;
1698 len -= used;
1699
1700 tcp_rcv_space_adjust(sk);
1701
1702 skip_copy:
1703 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1704 tp->urg_data = 0;
1705 tcp_fast_path_check(sk);
1706 }
1707 if (used + offset < skb->len)
1708 continue;
1709
1710 if (tcp_hdr(skb)->fin)
1711 goto found_fin_ok;
1712 if (!(flags & MSG_PEEK)) {
1713 sk_eat_skb(sk, skb, copied_early);
1714 copied_early = 0;
1715 }
1716 continue;
1717
1718 found_fin_ok:
1719 /* Process the FIN. */
1720 ++*seq;
1721 if (!(flags & MSG_PEEK)) {
1722 sk_eat_skb(sk, skb, copied_early);
1723 copied_early = 0;
1724 }
1725 break;
1726 } while (len > 0);
1727
1728 if (user_recv) {
1729 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1730 int chunk;
1731
1732 tp->ucopy.len = copied > 0 ? len : 0;
1733
1734 tcp_prequeue_process(sk);
1735
1736 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1737 NET_ADD_STATS_USER(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1738 len -= chunk;
1739 copied += chunk;
1740 }
1741 }
1742
1743 tp->ucopy.task = NULL;
1744 tp->ucopy.len = 0;
1745 }
1746
1747 #ifdef CONFIG_NET_DMA
1748 tcp_service_net_dma(sk, true); /* Wait for queue to drain */
1749 tp->ucopy.dma_chan = NULL;
1750
1751 if (tp->ucopy.pinned_list) {
1752 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1753 tp->ucopy.pinned_list = NULL;
1754 }
1755 #endif
1756
1757 /* According to UNIX98, msg_name/msg_namelen are ignored
1758 * on connected socket. I was just happy when found this 8) --ANK
1759 */
1760
1761 /* Clean up data we have read: This will do ACK frames. */
1762 tcp_cleanup_rbuf(sk, copied);
1763
1764 TCP_CHECK_TIMER(sk);
1765 release_sock(sk);
1766 return copied;
1767
1768 out:
1769 TCP_CHECK_TIMER(sk);
1770 release_sock(sk);
1771 return err;
1772
1773 recv_urg:
1774 err = tcp_recv_urg(sk, msg, len, flags);
1775 goto out;
1776 }
1777
1778 void tcp_set_state(struct sock *sk, int state)
1779 {
1780 int oldstate = sk->sk_state;
1781
1782 switch (state) {
1783 case TCP_ESTABLISHED:
1784 if (oldstate != TCP_ESTABLISHED)
1785 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1786 break;
1787
1788 case TCP_CLOSE:
1789 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1790 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1791
1792 sk->sk_prot->unhash(sk);
1793 if (inet_csk(sk)->icsk_bind_hash &&
1794 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1795 inet_put_port(sk);
1796 /* fall through */
1797 default:
1798 if (oldstate == TCP_ESTABLISHED)
1799 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1800 }
1801
1802 /* Change state AFTER socket is unhashed to avoid closed
1803 * socket sitting in hash tables.
1804 */
1805 sk->sk_state = state;
1806
1807 #ifdef STATE_TRACE
1808 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1809 #endif
1810 }
1811 EXPORT_SYMBOL_GPL(tcp_set_state);
1812
1813 /*
1814 * State processing on a close. This implements the state shift for
1815 * sending our FIN frame. Note that we only send a FIN for some
1816 * states. A shutdown() may have already sent the FIN, or we may be
1817 * closed.
1818 */
1819
1820 static const unsigned char new_state[16] = {
1821 /* current state: new state: action: */
1822 /* (Invalid) */ TCP_CLOSE,
1823 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1824 /* TCP_SYN_SENT */ TCP_CLOSE,
1825 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1826 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1827 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1828 /* TCP_TIME_WAIT */ TCP_CLOSE,
1829 /* TCP_CLOSE */ TCP_CLOSE,
1830 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1831 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1832 /* TCP_LISTEN */ TCP_CLOSE,
1833 /* TCP_CLOSING */ TCP_CLOSING,
1834 };
1835
1836 static int tcp_close_state(struct sock *sk)
1837 {
1838 int next = (int)new_state[sk->sk_state];
1839 int ns = next & TCP_STATE_MASK;
1840
1841 tcp_set_state(sk, ns);
1842
1843 return next & TCP_ACTION_FIN;
1844 }
1845
1846 /*
1847 * Shutdown the sending side of a connection. Much like close except
1848 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
1849 */
1850
1851 void tcp_shutdown(struct sock *sk, int how)
1852 {
1853 /* We need to grab some memory, and put together a FIN,
1854 * and then put it into the queue to be sent.
1855 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1856 */
1857 if (!(how & SEND_SHUTDOWN))
1858 return;
1859
1860 /* If we've already sent a FIN, or it's a closed state, skip this. */
1861 if ((1 << sk->sk_state) &
1862 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1863 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1864 /* Clear out any half completed packets. FIN if needed. */
1865 if (tcp_close_state(sk))
1866 tcp_send_fin(sk);
1867 }
1868 }
1869
1870 void tcp_close(struct sock *sk, long timeout)
1871 {
1872 struct sk_buff *skb;
1873 int data_was_unread = 0;
1874 int state;
1875
1876 lock_sock(sk);
1877 sk->sk_shutdown = SHUTDOWN_MASK;
1878
1879 if (sk->sk_state == TCP_LISTEN) {
1880 tcp_set_state(sk, TCP_CLOSE);
1881
1882 /* Special case. */
1883 inet_csk_listen_stop(sk);
1884
1885 goto adjudge_to_death;
1886 }
1887
1888 /* We need to flush the recv. buffs. We do this only on the
1889 * descriptor close, not protocol-sourced closes, because the
1890 * reader process may not have drained the data yet!
1891 */
1892 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1893 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1894 tcp_hdr(skb)->fin;
1895 data_was_unread += len;
1896 __kfree_skb(skb);
1897 }
1898
1899 sk_mem_reclaim(sk);
1900
1901 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1902 * data was lost. To witness the awful effects of the old behavior of
1903 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1904 * GET in an FTP client, suspend the process, wait for the client to
1905 * advertise a zero window, then kill -9 the FTP client, wheee...
1906 * Note: timeout is always zero in such a case.
1907 */
1908 if (data_was_unread) {
1909 /* Unread data was tossed, zap the connection. */
1910 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
1911 tcp_set_state(sk, TCP_CLOSE);
1912 tcp_send_active_reset(sk, sk->sk_allocation);
1913 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1914 /* Check zero linger _after_ checking for unread data. */
1915 sk->sk_prot->disconnect(sk, 0);
1916 NET_INC_STATS_USER(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
1917 } else if (tcp_close_state(sk)) {
1918 /* We FIN if the application ate all the data before
1919 * zapping the connection.
1920 */
1921
1922 /* RED-PEN. Formally speaking, we have broken TCP state
1923 * machine. State transitions:
1924 *
1925 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1926 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1927 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1928 *
1929 * are legal only when FIN has been sent (i.e. in window),
1930 * rather than queued out of window. Purists blame.
1931 *
1932 * F.e. "RFC state" is ESTABLISHED,
1933 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1934 *
1935 * The visible declinations are that sometimes
1936 * we enter time-wait state, when it is not required really
1937 * (harmless), do not send active resets, when they are
1938 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1939 * they look as CLOSING or LAST_ACK for Linux)
1940 * Probably, I missed some more holelets.
1941 * --ANK
1942 */
1943 tcp_send_fin(sk);
1944 }
1945
1946 sk_stream_wait_close(sk, timeout);
1947
1948 adjudge_to_death:
1949 state = sk->sk_state;
1950 sock_hold(sk);
1951 sock_orphan(sk);
1952
1953 /* It is the last release_sock in its life. It will remove backlog. */
1954 release_sock(sk);
1955
1956
1957 /* Now socket is owned by kernel and we acquire BH lock
1958 to finish close. No need to check for user refs.
1959 */
1960 local_bh_disable();
1961 bh_lock_sock(sk);
1962 WARN_ON(sock_owned_by_user(sk));
1963
1964 percpu_counter_inc(sk->sk_prot->orphan_count);
1965
1966 /* Have we already been destroyed by a softirq or backlog? */
1967 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1968 goto out;
1969
1970 /* This is a (useful) BSD violating of the RFC. There is a
1971 * problem with TCP as specified in that the other end could
1972 * keep a socket open forever with no application left this end.
1973 * We use a 3 minute timeout (about the same as BSD) then kill
1974 * our end. If they send after that then tough - BUT: long enough
1975 * that we won't make the old 4*rto = almost no time - whoops
1976 * reset mistake.
1977 *
1978 * Nope, it was not mistake. It is really desired behaviour
1979 * f.e. on http servers, when such sockets are useless, but
1980 * consume significant resources. Let's do it with special
1981 * linger2 option. --ANK
1982 */
1983
1984 if (sk->sk_state == TCP_FIN_WAIT2) {
1985 struct tcp_sock *tp = tcp_sk(sk);
1986 if (tp->linger2 < 0) {
1987 tcp_set_state(sk, TCP_CLOSE);
1988 tcp_send_active_reset(sk, GFP_ATOMIC);
1989 NET_INC_STATS_BH(sock_net(sk),
1990 LINUX_MIB_TCPABORTONLINGER);
1991 } else {
1992 const int tmo = tcp_fin_time(sk);
1993
1994 if (tmo > TCP_TIMEWAIT_LEN) {
1995 inet_csk_reset_keepalive_timer(sk,
1996 tmo - TCP_TIMEWAIT_LEN);
1997 } else {
1998 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1999 goto out;
2000 }
2001 }
2002 }
2003 if (sk->sk_state != TCP_CLOSE) {
2004 int orphan_count = percpu_counter_read_positive(
2005 sk->sk_prot->orphan_count);
2006
2007 sk_mem_reclaim(sk);
2008 if (tcp_too_many_orphans(sk, orphan_count)) {
2009 if (net_ratelimit())
2010 printk(KERN_INFO "TCP: too many of orphaned "
2011 "sockets\n");
2012 tcp_set_state(sk, TCP_CLOSE);
2013 tcp_send_active_reset(sk, GFP_ATOMIC);
2014 NET_INC_STATS_BH(sock_net(sk),
2015 LINUX_MIB_TCPABORTONMEMORY);
2016 }
2017 }
2018
2019 if (sk->sk_state == TCP_CLOSE)
2020 inet_csk_destroy_sock(sk);
2021 /* Otherwise, socket is reprieved until protocol close. */
2022
2023 out:
2024 bh_unlock_sock(sk);
2025 local_bh_enable();
2026 sock_put(sk);
2027 }
2028
2029 /* These states need RST on ABORT according to RFC793 */
2030
2031 static inline int tcp_need_reset(int state)
2032 {
2033 return (1 << state) &
2034 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2035 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2036 }
2037
2038 int tcp_disconnect(struct sock *sk, int flags)
2039 {
2040 struct inet_sock *inet = inet_sk(sk);
2041 struct inet_connection_sock *icsk = inet_csk(sk);
2042 struct tcp_sock *tp = tcp_sk(sk);
2043 int err = 0;
2044 int old_state = sk->sk_state;
2045
2046 if (old_state != TCP_CLOSE)
2047 tcp_set_state(sk, TCP_CLOSE);
2048
2049 /* ABORT function of RFC793 */
2050 if (old_state == TCP_LISTEN) {
2051 inet_csk_listen_stop(sk);
2052 } else if (tcp_need_reset(old_state) ||
2053 (tp->snd_nxt != tp->write_seq &&
2054 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2055 /* The last check adjusts for discrepancy of Linux wrt. RFC
2056 * states
2057 */
2058 tcp_send_active_reset(sk, gfp_any());
2059 sk->sk_err = ECONNRESET;
2060 } else if (old_state == TCP_SYN_SENT)
2061 sk->sk_err = ECONNRESET;
2062
2063 tcp_clear_xmit_timers(sk);
2064 __skb_queue_purge(&sk->sk_receive_queue);
2065 tcp_write_queue_purge(sk);
2066 __skb_queue_purge(&tp->out_of_order_queue);
2067 #ifdef CONFIG_NET_DMA
2068 __skb_queue_purge(&sk->sk_async_wait_queue);
2069 #endif
2070
2071 inet->inet_dport = 0;
2072
2073 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2074 inet_reset_saddr(sk);
2075
2076 sk->sk_shutdown = 0;
2077 sock_reset_flag(sk, SOCK_DONE);
2078 tp->srtt = 0;
2079 if ((tp->write_seq += tp->max_window + 2) == 0)
2080 tp->write_seq = 1;
2081 icsk->icsk_backoff = 0;
2082 tp->snd_cwnd = 2;
2083 icsk->icsk_probes_out = 0;
2084 tp->packets_out = 0;
2085 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2086 tp->snd_cwnd_cnt = 0;
2087 tp->bytes_acked = 0;
2088 tp->window_clamp = 0;
2089 tcp_set_ca_state(sk, TCP_CA_Open);
2090 tcp_clear_retrans(tp);
2091 inet_csk_delack_init(sk);
2092 tcp_init_send_head(sk);
2093 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2094 __sk_dst_reset(sk);
2095
2096 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2097
2098 sk->sk_error_report(sk);
2099 return err;
2100 }
2101
2102 /*
2103 * Socket option code for TCP.
2104 */
2105 static int do_tcp_setsockopt(struct sock *sk, int level,
2106 int optname, char __user *optval, unsigned int optlen)
2107 {
2108 struct tcp_sock *tp = tcp_sk(sk);
2109 struct inet_connection_sock *icsk = inet_csk(sk);
2110 int val;
2111 int err = 0;
2112
2113 /* These are data/string values, all the others are ints */
2114 switch (optname) {
2115 case TCP_CONGESTION: {
2116 char name[TCP_CA_NAME_MAX];
2117
2118 if (optlen < 1)
2119 return -EINVAL;
2120
2121 val = strncpy_from_user(name, optval,
2122 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2123 if (val < 0)
2124 return -EFAULT;
2125 name[val] = 0;
2126
2127 lock_sock(sk);
2128 err = tcp_set_congestion_control(sk, name);
2129 release_sock(sk);
2130 return err;
2131 }
2132 case TCP_COOKIE_TRANSACTIONS: {
2133 struct tcp_cookie_transactions ctd;
2134 struct tcp_cookie_values *cvp = NULL;
2135
2136 if (sizeof(ctd) > optlen)
2137 return -EINVAL;
2138 if (copy_from_user(&ctd, optval, sizeof(ctd)))
2139 return -EFAULT;
2140
2141 if (ctd.tcpct_used > sizeof(ctd.tcpct_value) ||
2142 ctd.tcpct_s_data_desired > TCP_MSS_DESIRED)
2143 return -EINVAL;
2144
2145 if (ctd.tcpct_cookie_desired == 0) {
2146 /* default to global value */
2147 } else if ((0x1 & ctd.tcpct_cookie_desired) ||
2148 ctd.tcpct_cookie_desired > TCP_COOKIE_MAX ||
2149 ctd.tcpct_cookie_desired < TCP_COOKIE_MIN) {
2150 return -EINVAL;
2151 }
2152
2153 if (TCP_COOKIE_OUT_NEVER & ctd.tcpct_flags) {
2154 /* Supercedes all other values */
2155 lock_sock(sk);
2156 if (tp->cookie_values != NULL) {
2157 kref_put(&tp->cookie_values->kref,
2158 tcp_cookie_values_release);
2159 tp->cookie_values = NULL;
2160 }
2161 tp->rx_opt.cookie_in_always = 0; /* false */
2162 tp->rx_opt.cookie_out_never = 1; /* true */
2163 release_sock(sk);
2164 return err;
2165 }
2166
2167 /* Allocate ancillary memory before locking.
2168 */
2169 if (ctd.tcpct_used > 0 ||
2170 (tp->cookie_values == NULL &&
2171 (sysctl_tcp_cookie_size > 0 ||
2172 ctd.tcpct_cookie_desired > 0 ||
2173 ctd.tcpct_s_data_desired > 0))) {
2174 cvp = kzalloc(sizeof(*cvp) + ctd.tcpct_used,
2175 GFP_KERNEL);
2176 if (cvp == NULL)
2177 return -ENOMEM;
2178 }
2179 lock_sock(sk);
2180 tp->rx_opt.cookie_in_always =
2181 (TCP_COOKIE_IN_ALWAYS & ctd.tcpct_flags);
2182 tp->rx_opt.cookie_out_never = 0; /* false */
2183
2184 if (tp->cookie_values != NULL) {
2185 if (cvp != NULL) {
2186 /* Changed values are recorded by a changed
2187 * pointer, ensuring the cookie will differ,
2188 * without separately hashing each value later.
2189 */
2190 kref_put(&tp->cookie_values->kref,
2191 tcp_cookie_values_release);
2192 kref_init(&cvp->kref);
2193 tp->cookie_values = cvp;
2194 } else {
2195 cvp = tp->cookie_values;
2196 }
2197 }
2198 if (cvp != NULL) {
2199 cvp->cookie_desired = ctd.tcpct_cookie_desired;
2200
2201 if (ctd.tcpct_used > 0) {
2202 memcpy(cvp->s_data_payload, ctd.tcpct_value,
2203 ctd.tcpct_used);
2204 cvp->s_data_desired = ctd.tcpct_used;
2205 cvp->s_data_constant = 1; /* true */
2206 } else {
2207 /* No constant payload data. */
2208 cvp->s_data_desired = ctd.tcpct_s_data_desired;
2209 cvp->s_data_constant = 0; /* false */
2210 }
2211 }
2212 release_sock(sk);
2213 return err;
2214 }
2215 default:
2216 /* fallthru */
2217 break;
2218 }
2219
2220 if (optlen < sizeof(int))
2221 return -EINVAL;
2222
2223 if (get_user(val, (int __user *)optval))
2224 return -EFAULT;
2225
2226 lock_sock(sk);
2227
2228 switch (optname) {
2229 case TCP_MAXSEG:
2230 /* Values greater than interface MTU won't take effect. However
2231 * at the point when this call is done we typically don't yet
2232 * know which interface is going to be used */
2233 if (val < 8 || val > MAX_TCP_WINDOW) {
2234 err = -EINVAL;
2235 break;
2236 }
2237 tp->rx_opt.user_mss = val;
2238 break;
2239
2240 case TCP_NODELAY:
2241 if (val) {
2242 /* TCP_NODELAY is weaker than TCP_CORK, so that
2243 * this option on corked socket is remembered, but
2244 * it is not activated until cork is cleared.
2245 *
2246 * However, when TCP_NODELAY is set we make
2247 * an explicit push, which overrides even TCP_CORK
2248 * for currently queued segments.
2249 */
2250 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2251 tcp_push_pending_frames(sk);
2252 } else {
2253 tp->nonagle &= ~TCP_NAGLE_OFF;
2254 }
2255 break;
2256
2257 case TCP_THIN_LINEAR_TIMEOUTS:
2258 if (val < 0 || val > 1)
2259 err = -EINVAL;
2260 else
2261 tp->thin_lto = val;
2262 break;
2263
2264 case TCP_THIN_DUPACK:
2265 if (val < 0 || val > 1)
2266 err = -EINVAL;
2267 else
2268 tp->thin_dupack = val;
2269 break;
2270
2271 case TCP_CORK:
2272 /* When set indicates to always queue non-full frames.
2273 * Later the user clears this option and we transmit
2274 * any pending partial frames in the queue. This is
2275 * meant to be used alongside sendfile() to get properly
2276 * filled frames when the user (for example) must write
2277 * out headers with a write() call first and then use
2278 * sendfile to send out the data parts.
2279 *
2280 * TCP_CORK can be set together with TCP_NODELAY and it is
2281 * stronger than TCP_NODELAY.
2282 */
2283 if (val) {
2284 tp->nonagle |= TCP_NAGLE_CORK;
2285 } else {
2286 tp->nonagle &= ~TCP_NAGLE_CORK;
2287 if (tp->nonagle&TCP_NAGLE_OFF)
2288 tp->nonagle |= TCP_NAGLE_PUSH;
2289 tcp_push_pending_frames(sk);
2290 }
2291 break;
2292
2293 case TCP_KEEPIDLE:
2294 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2295 err = -EINVAL;
2296 else {
2297 tp->keepalive_time = val * HZ;
2298 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2299 !((1 << sk->sk_state) &
2300 (TCPF_CLOSE | TCPF_LISTEN))) {
2301 u32 elapsed = keepalive_time_elapsed(tp);
2302 if (tp->keepalive_time > elapsed)
2303 elapsed = tp->keepalive_time - elapsed;
2304 else
2305 elapsed = 0;
2306 inet_csk_reset_keepalive_timer(sk, elapsed);
2307 }
2308 }
2309 break;
2310 case TCP_KEEPINTVL:
2311 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2312 err = -EINVAL;
2313 else
2314 tp->keepalive_intvl = val * HZ;
2315 break;
2316 case TCP_KEEPCNT:
2317 if (val < 1 || val > MAX_TCP_KEEPCNT)
2318 err = -EINVAL;
2319 else
2320 tp->keepalive_probes = val;
2321 break;
2322 case TCP_SYNCNT:
2323 if (val < 1 || val > MAX_TCP_SYNCNT)
2324 err = -EINVAL;
2325 else
2326 icsk->icsk_syn_retries = val;
2327 break;
2328
2329 case TCP_LINGER2:
2330 if (val < 0)
2331 tp->linger2 = -1;
2332 else if (val > sysctl_tcp_fin_timeout / HZ)
2333 tp->linger2 = 0;
2334 else
2335 tp->linger2 = val * HZ;
2336 break;
2337
2338 case TCP_DEFER_ACCEPT:
2339 /* Translate value in seconds to number of retransmits */
2340 icsk->icsk_accept_queue.rskq_defer_accept =
2341 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2342 TCP_RTO_MAX / HZ);
2343 break;
2344
2345 case TCP_WINDOW_CLAMP:
2346 if (!val) {
2347 if (sk->sk_state != TCP_CLOSE) {
2348 err = -EINVAL;
2349 break;
2350 }
2351 tp->window_clamp = 0;
2352 } else
2353 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2354 SOCK_MIN_RCVBUF / 2 : val;
2355 break;
2356
2357 case TCP_QUICKACK:
2358 if (!val) {
2359 icsk->icsk_ack.pingpong = 1;
2360 } else {
2361 icsk->icsk_ack.pingpong = 0;
2362 if ((1 << sk->sk_state) &
2363 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2364 inet_csk_ack_scheduled(sk)) {
2365 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2366 tcp_cleanup_rbuf(sk, 1);
2367 if (!(val & 1))
2368 icsk->icsk_ack.pingpong = 1;
2369 }
2370 }
2371 break;
2372
2373 #ifdef CONFIG_TCP_MD5SIG
2374 case TCP_MD5SIG:
2375 /* Read the IP->Key mappings from userspace */
2376 err = tp->af_specific->md5_parse(sk, optval, optlen);
2377 break;
2378 #endif
2379
2380 default:
2381 err = -ENOPROTOOPT;
2382 break;
2383 }
2384
2385 release_sock(sk);
2386 return err;
2387 }
2388
2389 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2390 unsigned int optlen)
2391 {
2392 struct inet_connection_sock *icsk = inet_csk(sk);
2393
2394 if (level != SOL_TCP)
2395 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2396 optval, optlen);
2397 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2398 }
2399
2400 #ifdef CONFIG_COMPAT
2401 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2402 char __user *optval, unsigned int optlen)
2403 {
2404 if (level != SOL_TCP)
2405 return inet_csk_compat_setsockopt(sk, level, optname,
2406 optval, optlen);
2407 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2408 }
2409
2410 EXPORT_SYMBOL(compat_tcp_setsockopt);
2411 #endif
2412
2413 /* Return information about state of tcp endpoint in API format. */
2414 void tcp_get_info(struct sock *sk, struct tcp_info *info)
2415 {
2416 struct tcp_sock *tp = tcp_sk(sk);
2417 const struct inet_connection_sock *icsk = inet_csk(sk);
2418 u32 now = tcp_time_stamp;
2419
2420 memset(info, 0, sizeof(*info));
2421
2422 info->tcpi_state = sk->sk_state;
2423 info->tcpi_ca_state = icsk->icsk_ca_state;
2424 info->tcpi_retransmits = icsk->icsk_retransmits;
2425 info->tcpi_probes = icsk->icsk_probes_out;
2426 info->tcpi_backoff = icsk->icsk_backoff;
2427
2428 if (tp->rx_opt.tstamp_ok)
2429 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2430 if (tcp_is_sack(tp))
2431 info->tcpi_options |= TCPI_OPT_SACK;
2432 if (tp->rx_opt.wscale_ok) {
2433 info->tcpi_options |= TCPI_OPT_WSCALE;
2434 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2435 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2436 }
2437
2438 if (tp->ecn_flags&TCP_ECN_OK)
2439 info->tcpi_options |= TCPI_OPT_ECN;
2440
2441 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2442 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2443 info->tcpi_snd_mss = tp->mss_cache;
2444 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2445
2446 if (sk->sk_state == TCP_LISTEN) {
2447 info->tcpi_unacked = sk->sk_ack_backlog;
2448 info->tcpi_sacked = sk->sk_max_ack_backlog;
2449 } else {
2450 info->tcpi_unacked = tp->packets_out;
2451 info->tcpi_sacked = tp->sacked_out;
2452 }
2453 info->tcpi_lost = tp->lost_out;
2454 info->tcpi_retrans = tp->retrans_out;
2455 info->tcpi_fackets = tp->fackets_out;
2456
2457 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2458 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2459 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2460
2461 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2462 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2463 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2464 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2465 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2466 info->tcpi_snd_cwnd = tp->snd_cwnd;
2467 info->tcpi_advmss = tp->advmss;
2468 info->tcpi_reordering = tp->reordering;
2469
2470 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2471 info->tcpi_rcv_space = tp->rcvq_space.space;
2472
2473 info->tcpi_total_retrans = tp->total_retrans;
2474 }
2475
2476 EXPORT_SYMBOL_GPL(tcp_get_info);
2477
2478 static int do_tcp_getsockopt(struct sock *sk, int level,
2479 int optname, char __user *optval, int __user *optlen)
2480 {
2481 struct inet_connection_sock *icsk = inet_csk(sk);
2482 struct tcp_sock *tp = tcp_sk(sk);
2483 int val, len;
2484
2485 if (get_user(len, optlen))
2486 return -EFAULT;
2487
2488 len = min_t(unsigned int, len, sizeof(int));
2489
2490 if (len < 0)
2491 return -EINVAL;
2492
2493 switch (optname) {
2494 case TCP_MAXSEG:
2495 val = tp->mss_cache;
2496 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2497 val = tp->rx_opt.user_mss;
2498 break;
2499 case TCP_NODELAY:
2500 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2501 break;
2502 case TCP_CORK:
2503 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2504 break;
2505 case TCP_KEEPIDLE:
2506 val = keepalive_time_when(tp) / HZ;
2507 break;
2508 case TCP_KEEPINTVL:
2509 val = keepalive_intvl_when(tp) / HZ;
2510 break;
2511 case TCP_KEEPCNT:
2512 val = keepalive_probes(tp);
2513 break;
2514 case TCP_SYNCNT:
2515 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2516 break;
2517 case TCP_LINGER2:
2518 val = tp->linger2;
2519 if (val >= 0)
2520 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2521 break;
2522 case TCP_DEFER_ACCEPT:
2523 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2524 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2525 break;
2526 case TCP_WINDOW_CLAMP:
2527 val = tp->window_clamp;
2528 break;
2529 case TCP_INFO: {
2530 struct tcp_info info;
2531
2532 if (get_user(len, optlen))
2533 return -EFAULT;
2534
2535 tcp_get_info(sk, &info);
2536
2537 len = min_t(unsigned int, len, sizeof(info));
2538 if (put_user(len, optlen))
2539 return -EFAULT;
2540 if (copy_to_user(optval, &info, len))
2541 return -EFAULT;
2542 return 0;
2543 }
2544 case TCP_QUICKACK:
2545 val = !icsk->icsk_ack.pingpong;
2546 break;
2547
2548 case TCP_CONGESTION:
2549 if (get_user(len, optlen))
2550 return -EFAULT;
2551 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2552 if (put_user(len, optlen))
2553 return -EFAULT;
2554 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2555 return -EFAULT;
2556 return 0;
2557
2558 case TCP_COOKIE_TRANSACTIONS: {
2559 struct tcp_cookie_transactions ctd;
2560 struct tcp_cookie_values *cvp = tp->cookie_values;
2561
2562 if (get_user(len, optlen))
2563 return -EFAULT;
2564 if (len < sizeof(ctd))
2565 return -EINVAL;
2566
2567 memset(&ctd, 0, sizeof(ctd));
2568 ctd.tcpct_flags = (tp->rx_opt.cookie_in_always ?
2569 TCP_COOKIE_IN_ALWAYS : 0)
2570 | (tp->rx_opt.cookie_out_never ?
2571 TCP_COOKIE_OUT_NEVER : 0);
2572
2573 if (cvp != NULL) {
2574 ctd.tcpct_flags |= (cvp->s_data_in ?
2575 TCP_S_DATA_IN : 0)
2576 | (cvp->s_data_out ?
2577 TCP_S_DATA_OUT : 0);
2578
2579 ctd.tcpct_cookie_desired = cvp->cookie_desired;
2580 ctd.tcpct_s_data_desired = cvp->s_data_desired;
2581
2582 memcpy(&ctd.tcpct_value[0], &cvp->cookie_pair[0],
2583 cvp->cookie_pair_size);
2584 ctd.tcpct_used = cvp->cookie_pair_size;
2585 }
2586
2587 if (put_user(sizeof(ctd), optlen))
2588 return -EFAULT;
2589 if (copy_to_user(optval, &ctd, sizeof(ctd)))
2590 return -EFAULT;
2591 return 0;
2592 }
2593 default:
2594 return -ENOPROTOOPT;
2595 }
2596
2597 if (put_user(len, optlen))
2598 return -EFAULT;
2599 if (copy_to_user(optval, &val, len))
2600 return -EFAULT;
2601 return 0;
2602 }
2603
2604 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2605 int __user *optlen)
2606 {
2607 struct inet_connection_sock *icsk = inet_csk(sk);
2608
2609 if (level != SOL_TCP)
2610 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2611 optval, optlen);
2612 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2613 }
2614
2615 #ifdef CONFIG_COMPAT
2616 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2617 char __user *optval, int __user *optlen)
2618 {
2619 if (level != SOL_TCP)
2620 return inet_csk_compat_getsockopt(sk, level, optname,
2621 optval, optlen);
2622 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2623 }
2624
2625 EXPORT_SYMBOL(compat_tcp_getsockopt);
2626 #endif
2627
2628 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2629 {
2630 struct sk_buff *segs = ERR_PTR(-EINVAL);
2631 struct tcphdr *th;
2632 unsigned thlen;
2633 unsigned int seq;
2634 __be32 delta;
2635 unsigned int oldlen;
2636 unsigned int mss;
2637
2638 if (!pskb_may_pull(skb, sizeof(*th)))
2639 goto out;
2640
2641 th = tcp_hdr(skb);
2642 thlen = th->doff * 4;
2643 if (thlen < sizeof(*th))
2644 goto out;
2645
2646 if (!pskb_may_pull(skb, thlen))
2647 goto out;
2648
2649 oldlen = (u16)~skb->len;
2650 __skb_pull(skb, thlen);
2651
2652 mss = skb_shinfo(skb)->gso_size;
2653 if (unlikely(skb->len <= mss))
2654 goto out;
2655
2656 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2657 /* Packet is from an untrusted source, reset gso_segs. */
2658 int type = skb_shinfo(skb)->gso_type;
2659
2660 if (unlikely(type &
2661 ~(SKB_GSO_TCPV4 |
2662 SKB_GSO_DODGY |
2663 SKB_GSO_TCP_ECN |
2664 SKB_GSO_TCPV6 |
2665 0) ||
2666 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2667 goto out;
2668
2669 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
2670
2671 segs = NULL;
2672 goto out;
2673 }
2674
2675 segs = skb_segment(skb, features);
2676 if (IS_ERR(segs))
2677 goto out;
2678
2679 delta = htonl(oldlen + (thlen + mss));
2680
2681 skb = segs;
2682 th = tcp_hdr(skb);
2683 seq = ntohl(th->seq);
2684
2685 do {
2686 th->fin = th->psh = 0;
2687
2688 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2689 (__force u32)delta));
2690 if (skb->ip_summed != CHECKSUM_PARTIAL)
2691 th->check =
2692 csum_fold(csum_partial(skb_transport_header(skb),
2693 thlen, skb->csum));
2694
2695 seq += mss;
2696 skb = skb->next;
2697 th = tcp_hdr(skb);
2698
2699 th->seq = htonl(seq);
2700 th->cwr = 0;
2701 } while (skb->next);
2702
2703 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2704 skb->data_len);
2705 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2706 (__force u32)delta));
2707 if (skb->ip_summed != CHECKSUM_PARTIAL)
2708 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2709 thlen, skb->csum));
2710
2711 out:
2712 return segs;
2713 }
2714 EXPORT_SYMBOL(tcp_tso_segment);
2715
2716 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
2717 {
2718 struct sk_buff **pp = NULL;
2719 struct sk_buff *p;
2720 struct tcphdr *th;
2721 struct tcphdr *th2;
2722 unsigned int len;
2723 unsigned int thlen;
2724 __be32 flags;
2725 unsigned int mss = 1;
2726 unsigned int hlen;
2727 unsigned int off;
2728 int flush = 1;
2729 int i;
2730
2731 off = skb_gro_offset(skb);
2732 hlen = off + sizeof(*th);
2733 th = skb_gro_header_fast(skb, off);
2734 if (skb_gro_header_hard(skb, hlen)) {
2735 th = skb_gro_header_slow(skb, hlen, off);
2736 if (unlikely(!th))
2737 goto out;
2738 }
2739
2740 thlen = th->doff * 4;
2741 if (thlen < sizeof(*th))
2742 goto out;
2743
2744 hlen = off + thlen;
2745 if (skb_gro_header_hard(skb, hlen)) {
2746 th = skb_gro_header_slow(skb, hlen, off);
2747 if (unlikely(!th))
2748 goto out;
2749 }
2750
2751 skb_gro_pull(skb, thlen);
2752
2753 len = skb_gro_len(skb);
2754 flags = tcp_flag_word(th);
2755
2756 for (; (p = *head); head = &p->next) {
2757 if (!NAPI_GRO_CB(p)->same_flow)
2758 continue;
2759
2760 th2 = tcp_hdr(p);
2761
2762 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
2763 NAPI_GRO_CB(p)->same_flow = 0;
2764 continue;
2765 }
2766
2767 goto found;
2768 }
2769
2770 goto out_check_final;
2771
2772 found:
2773 flush = NAPI_GRO_CB(p)->flush;
2774 flush |= (__force int)(flags & TCP_FLAG_CWR);
2775 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
2776 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
2777 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
2778 for (i = sizeof(*th); i < thlen; i += 4)
2779 flush |= *(u32 *)((u8 *)th + i) ^
2780 *(u32 *)((u8 *)th2 + i);
2781
2782 mss = skb_shinfo(p)->gso_size;
2783
2784 flush |= (len - 1) >= mss;
2785 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
2786
2787 if (flush || skb_gro_receive(head, skb)) {
2788 mss = 1;
2789 goto out_check_final;
2790 }
2791
2792 p = *head;
2793 th2 = tcp_hdr(p);
2794 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
2795
2796 out_check_final:
2797 flush = len < mss;
2798 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
2799 TCP_FLAG_RST | TCP_FLAG_SYN |
2800 TCP_FLAG_FIN));
2801
2802 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
2803 pp = head;
2804
2805 out:
2806 NAPI_GRO_CB(skb)->flush |= flush;
2807
2808 return pp;
2809 }
2810 EXPORT_SYMBOL(tcp_gro_receive);
2811
2812 int tcp_gro_complete(struct sk_buff *skb)
2813 {
2814 struct tcphdr *th = tcp_hdr(skb);
2815
2816 skb->csum_start = skb_transport_header(skb) - skb->head;
2817 skb->csum_offset = offsetof(struct tcphdr, check);
2818 skb->ip_summed = CHECKSUM_PARTIAL;
2819
2820 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
2821
2822 if (th->cwr)
2823 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
2824
2825 return 0;
2826 }
2827 EXPORT_SYMBOL(tcp_gro_complete);
2828
2829 #ifdef CONFIG_TCP_MD5SIG
2830 static unsigned long tcp_md5sig_users;
2831 static struct tcp_md5sig_pool * __percpu *tcp_md5sig_pool;
2832 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2833
2834 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool * __percpu *pool)
2835 {
2836 int cpu;
2837 for_each_possible_cpu(cpu) {
2838 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2839 if (p) {
2840 if (p->md5_desc.tfm)
2841 crypto_free_hash(p->md5_desc.tfm);
2842 kfree(p);
2843 }
2844 }
2845 free_percpu(pool);
2846 }
2847
2848 void tcp_free_md5sig_pool(void)
2849 {
2850 struct tcp_md5sig_pool * __percpu *pool = NULL;
2851
2852 spin_lock_bh(&tcp_md5sig_pool_lock);
2853 if (--tcp_md5sig_users == 0) {
2854 pool = tcp_md5sig_pool;
2855 tcp_md5sig_pool = NULL;
2856 }
2857 spin_unlock_bh(&tcp_md5sig_pool_lock);
2858 if (pool)
2859 __tcp_free_md5sig_pool(pool);
2860 }
2861
2862 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2863
2864 static struct tcp_md5sig_pool * __percpu *
2865 __tcp_alloc_md5sig_pool(struct sock *sk)
2866 {
2867 int cpu;
2868 struct tcp_md5sig_pool * __percpu *pool;
2869
2870 pool = alloc_percpu(struct tcp_md5sig_pool *);
2871 if (!pool)
2872 return NULL;
2873
2874 for_each_possible_cpu(cpu) {
2875 struct tcp_md5sig_pool *p;
2876 struct crypto_hash *hash;
2877
2878 p = kzalloc(sizeof(*p), sk->sk_allocation);
2879 if (!p)
2880 goto out_free;
2881 *per_cpu_ptr(pool, cpu) = p;
2882
2883 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2884 if (!hash || IS_ERR(hash))
2885 goto out_free;
2886
2887 p->md5_desc.tfm = hash;
2888 }
2889 return pool;
2890 out_free:
2891 __tcp_free_md5sig_pool(pool);
2892 return NULL;
2893 }
2894
2895 struct tcp_md5sig_pool * __percpu *tcp_alloc_md5sig_pool(struct sock *sk)
2896 {
2897 struct tcp_md5sig_pool * __percpu *pool;
2898 int alloc = 0;
2899
2900 retry:
2901 spin_lock_bh(&tcp_md5sig_pool_lock);
2902 pool = tcp_md5sig_pool;
2903 if (tcp_md5sig_users++ == 0) {
2904 alloc = 1;
2905 spin_unlock_bh(&tcp_md5sig_pool_lock);
2906 } else if (!pool) {
2907 tcp_md5sig_users--;
2908 spin_unlock_bh(&tcp_md5sig_pool_lock);
2909 cpu_relax();
2910 goto retry;
2911 } else
2912 spin_unlock_bh(&tcp_md5sig_pool_lock);
2913
2914 if (alloc) {
2915 /* we cannot hold spinlock here because this may sleep. */
2916 struct tcp_md5sig_pool * __percpu *p;
2917
2918 p = __tcp_alloc_md5sig_pool(sk);
2919 spin_lock_bh(&tcp_md5sig_pool_lock);
2920 if (!p) {
2921 tcp_md5sig_users--;
2922 spin_unlock_bh(&tcp_md5sig_pool_lock);
2923 return NULL;
2924 }
2925 pool = tcp_md5sig_pool;
2926 if (pool) {
2927 /* oops, it has already been assigned. */
2928 spin_unlock_bh(&tcp_md5sig_pool_lock);
2929 __tcp_free_md5sig_pool(p);
2930 } else {
2931 tcp_md5sig_pool = pool = p;
2932 spin_unlock_bh(&tcp_md5sig_pool_lock);
2933 }
2934 }
2935 return pool;
2936 }
2937
2938 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2939
2940
2941 /**
2942 * tcp_get_md5sig_pool - get md5sig_pool for this user
2943 *
2944 * We use percpu structure, so if we succeed, we exit with preemption
2945 * and BH disabled, to make sure another thread or softirq handling
2946 * wont try to get same context.
2947 */
2948 struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
2949 {
2950 struct tcp_md5sig_pool * __percpu *p;
2951
2952 local_bh_disable();
2953
2954 spin_lock(&tcp_md5sig_pool_lock);
2955 p = tcp_md5sig_pool;
2956 if (p)
2957 tcp_md5sig_users++;
2958 spin_unlock(&tcp_md5sig_pool_lock);
2959
2960 if (p)
2961 return *per_cpu_ptr(p, smp_processor_id());
2962
2963 local_bh_enable();
2964 return NULL;
2965 }
2966 EXPORT_SYMBOL(tcp_get_md5sig_pool);
2967
2968 void tcp_put_md5sig_pool(void)
2969 {
2970 local_bh_enable();
2971 tcp_free_md5sig_pool();
2972 }
2973 EXPORT_SYMBOL(tcp_put_md5sig_pool);
2974
2975 int tcp_md5_hash_header(struct tcp_md5sig_pool *hp,
2976 struct tcphdr *th)
2977 {
2978 struct scatterlist sg;
2979 int err;
2980
2981 __sum16 old_checksum = th->check;
2982 th->check = 0;
2983 /* options aren't included in the hash */
2984 sg_init_one(&sg, th, sizeof(struct tcphdr));
2985 err = crypto_hash_update(&hp->md5_desc, &sg, sizeof(struct tcphdr));
2986 th->check = old_checksum;
2987 return err;
2988 }
2989
2990 EXPORT_SYMBOL(tcp_md5_hash_header);
2991
2992 int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
2993 struct sk_buff *skb, unsigned header_len)
2994 {
2995 struct scatterlist sg;
2996 const struct tcphdr *tp = tcp_hdr(skb);
2997 struct hash_desc *desc = &hp->md5_desc;
2998 unsigned i;
2999 const unsigned head_data_len = skb_headlen(skb) > header_len ?
3000 skb_headlen(skb) - header_len : 0;
3001 const struct skb_shared_info *shi = skb_shinfo(skb);
3002
3003 sg_init_table(&sg, 1);
3004
3005 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3006 if (crypto_hash_update(desc, &sg, head_data_len))
3007 return 1;
3008
3009 for (i = 0; i < shi->nr_frags; ++i) {
3010 const struct skb_frag_struct *f = &shi->frags[i];
3011 sg_set_page(&sg, f->page, f->size, f->page_offset);
3012 if (crypto_hash_update(desc, &sg, f->size))
3013 return 1;
3014 }
3015
3016 return 0;
3017 }
3018
3019 EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3020
3021 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, struct tcp_md5sig_key *key)
3022 {
3023 struct scatterlist sg;
3024
3025 sg_init_one(&sg, key->key, key->keylen);
3026 return crypto_hash_update(&hp->md5_desc, &sg, key->keylen);
3027 }
3028
3029 EXPORT_SYMBOL(tcp_md5_hash_key);
3030
3031 #endif
3032
3033 /**
3034 * Each Responder maintains up to two secret values concurrently for
3035 * efficient secret rollover. Each secret value has 4 states:
3036 *
3037 * Generating. (tcp_secret_generating != tcp_secret_primary)
3038 * Generates new Responder-Cookies, but not yet used for primary
3039 * verification. This is a short-term state, typically lasting only
3040 * one round trip time (RTT).
3041 *
3042 * Primary. (tcp_secret_generating == tcp_secret_primary)
3043 * Used both for generation and primary verification.
3044 *
3045 * Retiring. (tcp_secret_retiring != tcp_secret_secondary)
3046 * Used for verification, until the first failure that can be
3047 * verified by the newer Generating secret. At that time, this
3048 * cookie's state is changed to Secondary, and the Generating
3049 * cookie's state is changed to Primary. This is a short-term state,
3050 * typically lasting only one round trip time (RTT).
3051 *
3052 * Secondary. (tcp_secret_retiring == tcp_secret_secondary)
3053 * Used for secondary verification, after primary verification
3054 * failures. This state lasts no more than twice the Maximum Segment
3055 * Lifetime (2MSL). Then, the secret is discarded.
3056 */
3057 struct tcp_cookie_secret {
3058 /* The secret is divided into two parts. The digest part is the
3059 * equivalent of previously hashing a secret and saving the state,
3060 * and serves as an initialization vector (IV). The message part
3061 * serves as the trailing secret.
3062 */
3063 u32 secrets[COOKIE_WORKSPACE_WORDS];
3064 unsigned long expires;
3065 };
3066
3067 #define TCP_SECRET_1MSL (HZ * TCP_PAWS_MSL)
3068 #define TCP_SECRET_2MSL (HZ * TCP_PAWS_MSL * 2)
3069 #define TCP_SECRET_LIFE (HZ * 600)
3070
3071 static struct tcp_cookie_secret tcp_secret_one;
3072 static struct tcp_cookie_secret tcp_secret_two;
3073
3074 /* Essentially a circular list, without dynamic allocation. */
3075 static struct tcp_cookie_secret *tcp_secret_generating;
3076 static struct tcp_cookie_secret *tcp_secret_primary;
3077 static struct tcp_cookie_secret *tcp_secret_retiring;
3078 static struct tcp_cookie_secret *tcp_secret_secondary;
3079
3080 static DEFINE_SPINLOCK(tcp_secret_locker);
3081
3082 /* Select a pseudo-random word in the cookie workspace.
3083 */
3084 static inline u32 tcp_cookie_work(const u32 *ws, const int n)
3085 {
3086 return ws[COOKIE_DIGEST_WORDS + ((COOKIE_MESSAGE_WORDS-1) & ws[n])];
3087 }
3088
3089 /* Fill bakery[COOKIE_WORKSPACE_WORDS] with generator, updating as needed.
3090 * Called in softirq context.
3091 * Returns: 0 for success.
3092 */
3093 int tcp_cookie_generator(u32 *bakery)
3094 {
3095 unsigned long jiffy = jiffies;
3096
3097 if (unlikely(time_after_eq(jiffy, tcp_secret_generating->expires))) {
3098 spin_lock_bh(&tcp_secret_locker);
3099 if (!time_after_eq(jiffy, tcp_secret_generating->expires)) {
3100 /* refreshed by another */
3101 memcpy(bakery,
3102 &tcp_secret_generating->secrets[0],
3103 COOKIE_WORKSPACE_WORDS);
3104 } else {
3105 /* still needs refreshing */
3106 get_random_bytes(bakery, COOKIE_WORKSPACE_WORDS);
3107
3108 /* The first time, paranoia assumes that the
3109 * randomization function isn't as strong. But,
3110 * this secret initialization is delayed until
3111 * the last possible moment (packet arrival).
3112 * Although that time is observable, it is
3113 * unpredictably variable. Mash in the most
3114 * volatile clock bits available, and expire the
3115 * secret extra quickly.
3116 */
3117 if (unlikely(tcp_secret_primary->expires ==
3118 tcp_secret_secondary->expires)) {
3119 struct timespec tv;
3120
3121 getnstimeofday(&tv);
3122 bakery[COOKIE_DIGEST_WORDS+0] ^=
3123 (u32)tv.tv_nsec;
3124
3125 tcp_secret_secondary->expires = jiffy
3126 + TCP_SECRET_1MSL
3127 + (0x0f & tcp_cookie_work(bakery, 0));
3128 } else {
3129 tcp_secret_secondary->expires = jiffy
3130 + TCP_SECRET_LIFE
3131 + (0xff & tcp_cookie_work(bakery, 1));
3132 tcp_secret_primary->expires = jiffy
3133 + TCP_SECRET_2MSL
3134 + (0x1f & tcp_cookie_work(bakery, 2));
3135 }
3136 memcpy(&tcp_secret_secondary->secrets[0],
3137 bakery, COOKIE_WORKSPACE_WORDS);
3138
3139 rcu_assign_pointer(tcp_secret_generating,
3140 tcp_secret_secondary);
3141 rcu_assign_pointer(tcp_secret_retiring,
3142 tcp_secret_primary);
3143 /*
3144 * Neither call_rcu() nor synchronize_rcu() needed.
3145 * Retiring data is not freed. It is replaced after
3146 * further (locked) pointer updates, and a quiet time
3147 * (minimum 1MSL, maximum LIFE - 2MSL).
3148 */
3149 }
3150 spin_unlock_bh(&tcp_secret_locker);
3151 } else {
3152 rcu_read_lock_bh();
3153 memcpy(bakery,
3154 &rcu_dereference(tcp_secret_generating)->secrets[0],
3155 COOKIE_WORKSPACE_WORDS);
3156 rcu_read_unlock_bh();
3157 }
3158 return 0;
3159 }
3160 EXPORT_SYMBOL(tcp_cookie_generator);
3161
3162 void tcp_done(struct sock *sk)
3163 {
3164 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3165 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3166
3167 tcp_set_state(sk, TCP_CLOSE);
3168 tcp_clear_xmit_timers(sk);
3169
3170 sk->sk_shutdown = SHUTDOWN_MASK;
3171
3172 if (!sock_flag(sk, SOCK_DEAD))
3173 sk->sk_state_change(sk);
3174 else
3175 inet_csk_destroy_sock(sk);
3176 }
3177 EXPORT_SYMBOL_GPL(tcp_done);
3178
3179 extern struct tcp_congestion_ops tcp_reno;
3180
3181 static __initdata unsigned long thash_entries;
3182 static int __init set_thash_entries(char *str)
3183 {
3184 if (!str)
3185 return 0;
3186 thash_entries = simple_strtoul(str, &str, 0);
3187 return 1;
3188 }
3189 __setup("thash_entries=", set_thash_entries);
3190
3191 void __init tcp_init(void)
3192 {
3193 struct sk_buff *skb = NULL;
3194 unsigned long nr_pages, limit;
3195 int order, i, max_share;
3196 unsigned long jiffy = jiffies;
3197
3198 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
3199
3200 percpu_counter_init(&tcp_sockets_allocated, 0);
3201 percpu_counter_init(&tcp_orphan_count, 0);
3202 tcp_hashinfo.bind_bucket_cachep =
3203 kmem_cache_create("tcp_bind_bucket",
3204 sizeof(struct inet_bind_bucket), 0,
3205 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3206
3207 /* Size and allocate the main established and bind bucket
3208 * hash tables.
3209 *
3210 * The methodology is similar to that of the buffer cache.
3211 */
3212 tcp_hashinfo.ehash =
3213 alloc_large_system_hash("TCP established",
3214 sizeof(struct inet_ehash_bucket),
3215 thash_entries,
3216 (totalram_pages >= 128 * 1024) ?
3217 13 : 15,
3218 0,
3219 NULL,
3220 &tcp_hashinfo.ehash_mask,
3221 thash_entries ? 0 : 512 * 1024);
3222 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++) {
3223 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3224 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].twchain, i);
3225 }
3226 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3227 panic("TCP: failed to alloc ehash_locks");
3228 tcp_hashinfo.bhash =
3229 alloc_large_system_hash("TCP bind",
3230 sizeof(struct inet_bind_hashbucket),
3231 tcp_hashinfo.ehash_mask + 1,
3232 (totalram_pages >= 128 * 1024) ?
3233 13 : 15,
3234 0,
3235 &tcp_hashinfo.bhash_size,
3236 NULL,
3237 64 * 1024);
3238 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
3239 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3240 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3241 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3242 }
3243
3244 /* Try to be a bit smarter and adjust defaults depending
3245 * on available memory.
3246 */
3247 for (order = 0; ((1 << order) << PAGE_SHIFT) <
3248 (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
3249 order++)
3250 ;
3251 if (order >= 4) {
3252 tcp_death_row.sysctl_max_tw_buckets = 180000;
3253 sysctl_tcp_max_orphans = 4096 << (order - 4);
3254 sysctl_max_syn_backlog = 1024;
3255 } else if (order < 3) {
3256 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
3257 sysctl_tcp_max_orphans >>= (3 - order);
3258 sysctl_max_syn_backlog = 128;
3259 }
3260
3261 /* Set the pressure threshold to be a fraction of global memory that
3262 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
3263 * memory, with a floor of 128 pages.
3264 */
3265 nr_pages = totalram_pages - totalhigh_pages;
3266 limit = min(nr_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
3267 limit = (limit * (nr_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
3268 limit = max(limit, 128UL);
3269 sysctl_tcp_mem[0] = limit / 4 * 3;
3270 sysctl_tcp_mem[1] = limit;
3271 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
3272
3273 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3274 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
3275 max_share = min(4UL*1024*1024, limit);
3276
3277 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3278 sysctl_tcp_wmem[1] = 16*1024;
3279 sysctl_tcp_wmem[2] = max(64*1024, max_share);
3280
3281 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3282 sysctl_tcp_rmem[1] = 87380;
3283 sysctl_tcp_rmem[2] = max(87380, max_share);
3284
3285 printk(KERN_INFO "TCP: Hash tables configured "
3286 "(established %u bind %u)\n",
3287 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3288
3289 tcp_register_congestion_control(&tcp_reno);
3290
3291 memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
3292 memset(&tcp_secret_two.secrets[0], 0, sizeof(tcp_secret_two.secrets));
3293 tcp_secret_one.expires = jiffy; /* past due */
3294 tcp_secret_two.expires = jiffy; /* past due */
3295 tcp_secret_generating = &tcp_secret_one;
3296 tcp_secret_primary = &tcp_secret_one;
3297 tcp_secret_retiring = &tcp_secret_two;
3298 tcp_secret_secondary = &tcp_secret_two;
3299 }
3300
3301 EXPORT_SYMBOL(tcp_close);
3302 EXPORT_SYMBOL(tcp_disconnect);
3303 EXPORT_SYMBOL(tcp_getsockopt);
3304 EXPORT_SYMBOL(tcp_ioctl);
3305 EXPORT_SYMBOL(tcp_poll);
3306 EXPORT_SYMBOL(tcp_read_sock);
3307 EXPORT_SYMBOL(tcp_recvmsg);
3308 EXPORT_SYMBOL(tcp_sendmsg);
3309 EXPORT_SYMBOL(tcp_splice_read);
3310 EXPORT_SYMBOL(tcp_sendpage);
3311 EXPORT_SYMBOL(tcp_setsockopt);
3312 EXPORT_SYMBOL(tcp_shutdown);
linux-next