?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
[NETFILTER]: nf_conntrack: round up hashsize to next multiple of PAGE_SIZE
[linux-2.6/x86.git] / net / ipv4 / tcp.c
blob450f44bb2c8e6591b90e846a056dee2a04f8d40a
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 * Version: $Id: tcp.c,v 1.216 2002/02/01 22:01:04 davem Exp $
9 *
10 * Authors: Ross Biro
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
21 *
22 * Fixes:
23 * Alan Cox : Numerous verify_area() calls
24 * Alan Cox : Set the ACK bit on a reset
25 * Alan Cox : Stopped it crashing if it closed while
26 * sk->inuse=1 and was trying to connect
27 * (tcp_err()).
28 * Alan Cox : All icmp error handling was broken
29 * pointers passed where wrong and the
30 * socket was looked up backwards. Nobody
31 * tested any icmp error code obviously.
32 * Alan Cox : tcp_err() now handled properly. It
33 * wakes people on errors. poll
34 * behaves and the icmp error race
35 * has gone by moving it into sock.c
36 * Alan Cox : tcp_send_reset() fixed to work for
37 * everything not just packets for
38 * unknown sockets.
39 * Alan Cox : tcp option processing.
40 * Alan Cox : Reset tweaked (still not 100%) [Had
41 * syn rule wrong]
42 * Herp Rosmanith : More reset fixes
43 * Alan Cox : No longer acks invalid rst frames.
44 * Acking any kind of RST is right out.
45 * Alan Cox : Sets an ignore me flag on an rst
46 * receive otherwise odd bits of prattle
47 * escape still
48 * Alan Cox : Fixed another acking RST frame bug.
49 * Should stop LAN workplace lockups.
50 * Alan Cox : Some tidyups using the new skb list
51 * facilities
52 * Alan Cox : sk->keepopen now seems to work
53 * Alan Cox : Pulls options out correctly on accepts
54 * Alan Cox : Fixed assorted sk->rqueue->next errors
55 * Alan Cox : PSH doesn't end a TCP read. Switched a
56 * bit to skb ops.
57 * Alan Cox : Tidied tcp_data to avoid a potential
58 * nasty.
59 * Alan Cox : Added some better commenting, as the
60 * tcp is hard to follow
61 * Alan Cox : Removed incorrect check for 20 * psh
62 * Michael O'Reilly : ack < copied bug fix.
63 * Johannes Stille : Misc tcp fixes (not all in yet).
64 * Alan Cox : FIN with no memory -> CRASH
65 * Alan Cox : Added socket option proto entries.
66 * Also added awareness of them to accept.
67 * Alan Cox : Added TCP options (SOL_TCP)
68 * Alan Cox : Switched wakeup calls to callbacks,
69 * so the kernel can layer network
70 * sockets.
71 * Alan Cox : Use ip_tos/ip_ttl settings.
72 * Alan Cox : Handle FIN (more) properly (we hope).
73 * Alan Cox : RST frames sent on unsynchronised
74 * state ack error.
75 * Alan Cox : Put in missing check for SYN bit.
76 * Alan Cox : Added tcp_select_window() aka NET2E
77 * window non shrink trick.
78 * Alan Cox : Added a couple of small NET2E timer
79 * fixes
80 * Charles Hedrick : TCP fixes
81 * Toomas Tamm : TCP window fixes
82 * Alan Cox : Small URG fix to rlogin ^C ack fight
83 * Charles Hedrick : Rewrote most of it to actually work
84 * Linus : Rewrote tcp_read() and URG handling
85 * completely
86 * Gerhard Koerting: Fixed some missing timer handling
87 * Matthew Dillon : Reworked TCP machine states as per RFC
88 * Gerhard Koerting: PC/TCP workarounds
89 * Adam Caldwell : Assorted timer/timing errors
90 * Matthew Dillon : Fixed another RST bug
91 * Alan Cox : Move to kernel side addressing changes.
92 * Alan Cox : Beginning work on TCP fastpathing
93 * (not yet usable)
94 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
95 * Alan Cox : TCP fast path debugging
96 * Alan Cox : Window clamping
97 * Michael Riepe : Bug in tcp_check()
98 * Matt Dillon : More TCP improvements and RST bug fixes
99 * Matt Dillon : Yet more small nasties remove from the
100 * TCP code (Be very nice to this man if
101 * tcp finally works 100%) 8)
102 * Alan Cox : BSD accept semantics.
103 * Alan Cox : Reset on closedown bug.
104 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
105 * Michael Pall : Handle poll() after URG properly in
106 * all cases.
107 * Michael Pall : Undo the last fix in tcp_read_urg()
108 * (multi URG PUSH broke rlogin).
109 * Michael Pall : Fix the multi URG PUSH problem in
110 * tcp_readable(), poll() after URG
111 * works now.
112 * Michael Pall : recv(...,MSG_OOB) never blocks in the
113 * BSD api.
114 * Alan Cox : Changed the semantics of sk->socket to
115 * fix a race and a signal problem with
116 * accept() and async I/O.
117 * Alan Cox : Relaxed the rules on tcp_sendto().
118 * Yury Shevchuk : Really fixed accept() blocking problem.
119 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
120 * clients/servers which listen in on
121 * fixed ports.
122 * Alan Cox : Cleaned the above up and shrank it to
123 * a sensible code size.
124 * Alan Cox : Self connect lockup fix.
125 * Alan Cox : No connect to multicast.
126 * Ross Biro : Close unaccepted children on master
127 * socket close.
128 * Alan Cox : Reset tracing code.
129 * Alan Cox : Spurious resets on shutdown.
130 * Alan Cox : Giant 15 minute/60 second timer error
131 * Alan Cox : Small whoops in polling before an
132 * accept.
133 * Alan Cox : Kept the state trace facility since
134 * it's handy for debugging.
135 * Alan Cox : More reset handler fixes.
136 * Alan Cox : Started rewriting the code based on
137 * the RFC's for other useful protocol
138 * references see: Comer, KA9Q NOS, and
139 * for a reference on the difference
140 * between specifications and how BSD
141 * works see the 4.4lite source.
142 * A.N.Kuznetsov : Don't time wait on completion of tidy
143 * close.
144 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
145 * Linus Torvalds : Fixed BSD port reuse to work first syn
146 * Alan Cox : Reimplemented timers as per the RFC
147 * and using multiple timers for sanity.
148 * Alan Cox : Small bug fixes, and a lot of new
149 * comments.
150 * Alan Cox : Fixed dual reader crash by locking
151 * the buffers (much like datagram.c)
152 * Alan Cox : Fixed stuck sockets in probe. A probe
153 * now gets fed up of retrying without
154 * (even a no space) answer.
155 * Alan Cox : Extracted closing code better
156 * Alan Cox : Fixed the closing state machine to
157 * resemble the RFC.
158 * Alan Cox : More 'per spec' fixes.
159 * Jorge Cwik : Even faster checksumming.
160 * Alan Cox : tcp_data() doesn't ack illegal PSH
161 * only frames. At least one pc tcp stack
162 * generates them.
163 * Alan Cox : Cache last socket.
164 * Alan Cox : Per route irtt.
165 * Matt Day : poll()->select() match BSD precisely on error
166 * Alan Cox : New buffers
167 * Marc Tamsky : Various sk->prot->retransmits and
168 * sk->retransmits misupdating fixed.
169 * Fixed tcp_write_timeout: stuck close,
170 * and TCP syn retries gets used now.
171 * Mark Yarvis : In tcp_read_wakeup(), don't send an
172 * ack if state is TCP_CLOSED.
173 * Alan Cox : Look up device on a retransmit - routes may
174 * change. Doesn't yet cope with MSS shrink right
175 * but it's a start!
176 * Marc Tamsky : Closing in closing fixes.
177 * Mike Shaver : RFC1122 verifications.
178 * Alan Cox : rcv_saddr errors.
179 * Alan Cox : Block double connect().
180 * Alan Cox : Small hooks for enSKIP.
181 * Alexey Kuznetsov: Path MTU discovery.
182 * Alan Cox : Support soft errors.
183 * Alan Cox : Fix MTU discovery pathological case
184 * when the remote claims no mtu!
185 * Marc Tamsky : TCP_CLOSE fix.
186 * Colin (G3TNE) : Send a reset on syn ack replies in
187 * window but wrong (fixes NT lpd problems)
188 * Pedro Roque : Better TCP window handling, delayed ack.
189 * Joerg Reuter : No modification of locked buffers in
190 * tcp_do_retransmit()
191 * Eric Schenk : Changed receiver side silly window
192 * avoidance algorithm to BSD style
193 * algorithm. This doubles throughput
194 * against machines running Solaris,
195 * and seems to result in general
196 * improvement.
197 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
198 * Willy Konynenberg : Transparent proxying support.
199 * Mike McLagan : Routing by source
200 * Keith Owens : Do proper merging with partial SKB's in
201 * tcp_do_sendmsg to avoid burstiness.
202 * Eric Schenk : Fix fast close down bug with
203 * shutdown() followed by close().
204 * Andi Kleen : Make poll agree with SIGIO
205 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
206 * lingertime == 0 (RFC 793 ABORT Call)
207 * Hirokazu Takahashi : Use copy_from_user() instead of
208 * csum_and_copy_from_user() if possible.
209 *
210 * This program is free software; you can redistribute it and/or
211 * modify it under the terms of the GNU General Public License
212 * as published by the Free Software Foundation; either version
213 * 2 of the License, or(at your option) any later version.
214 *
215 * Description of States:
216 *
217 * TCP_SYN_SENT sent a connection request, waiting for ack
218 *
219 * TCP_SYN_RECV received a connection request, sent ack,
220 * waiting for final ack in three-way handshake.
221 *
222 * TCP_ESTABLISHED connection established
223 *
224 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
225 * transmission of remaining buffered data
226 *
227 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
228 * to shutdown
229 *
230 * TCP_CLOSING both sides have shutdown but we still have
231 * data we have to finish sending
232 *
233 * TCP_TIME_WAIT timeout to catch resent junk before entering
234 * closed, can only be entered from FIN_WAIT2
235 * or CLOSING. Required because the other end
236 * may not have gotten our last ACK causing it
237 * to retransmit the data packet (which we ignore)
238 *
239 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
240 * us to finish writing our data and to shutdown
241 * (we have to close() to move on to LAST_ACK)
242 *
243 * TCP_LAST_ACK out side has shutdown after remote has
244 * shutdown. There may still be data in our
245 * buffer that we have to finish sending
246 *
247 * TCP_CLOSE socket is finished
248 */
249
250 #include <linux/module.h>
251 #include <linux/types.h>
252 #include <linux/fcntl.h>
253 #include <linux/poll.h>
254 #include <linux/init.h>
255 #include <linux/fs.h>
256 #include <linux/random.h>
257 #include <linux/bootmem.h>
258 #include <linux/cache.h>
259 #include <linux/err.h>
260 #include <linux/crypto.h>
261
262 #include <net/icmp.h>
263 #include <net/tcp.h>
264 #include <net/xfrm.h>
265 #include <net/ip.h>
266 #include <net/netdma.h>
267
268 #include <asm/uaccess.h>
269 #include <asm/ioctls.h>
270
271 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
272
273 DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
274
275 atomic_t tcp_orphan_count = ATOMIC_INIT(0);
276
277 EXPORT_SYMBOL_GPL(tcp_orphan_count);
278
279 int sysctl_tcp_mem[3] __read_mostly;
280 int sysctl_tcp_wmem[3] __read_mostly;
281 int sysctl_tcp_rmem[3] __read_mostly;
282
283 EXPORT_SYMBOL(sysctl_tcp_mem);
284 EXPORT_SYMBOL(sysctl_tcp_rmem);
285 EXPORT_SYMBOL(sysctl_tcp_wmem);
286
287 atomic_t tcp_memory_allocated; /* Current allocated memory. */
288 atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
289
290 EXPORT_SYMBOL(tcp_memory_allocated);
291 EXPORT_SYMBOL(tcp_sockets_allocated);
292
293 /*
294 * Pressure flag: try to collapse.
295 * Technical note: it is used by multiple contexts non atomically.
296 * All the sk_stream_mem_schedule() is of this nature: accounting
297 * is strict, actions are advisory and have some latency.
298 */
299 int tcp_memory_pressure __read_mostly;
300
301 EXPORT_SYMBOL(tcp_memory_pressure);
302
303 void tcp_enter_memory_pressure(void)
304 {
305 if (!tcp_memory_pressure) {
306 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
307 tcp_memory_pressure = 1;
308 }
309 }
310
311 EXPORT_SYMBOL(tcp_enter_memory_pressure);
312
313 /*
314 * Wait for a TCP event.
315 *
316 * Note that we don't need to lock the socket, as the upper poll layers
317 * take care of normal races (between the test and the event) and we don't
318 * go look at any of the socket buffers directly.
319 */
320 unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
321 {
322 unsigned int mask;
323 struct sock *sk = sock->sk;
324 struct tcp_sock *tp = tcp_sk(sk);
325
326 poll_wait(file, sk->sk_sleep, wait);
327 if (sk->sk_state == TCP_LISTEN)
328 return inet_csk_listen_poll(sk);
329
330 /* Socket is not locked. We are protected from async events
331 by poll logic and correct handling of state changes
332 made by another threads is impossible in any case.
333 */
334
335 mask = 0;
336 if (sk->sk_err)
337 mask = POLLERR;
338
339 /*
340 * POLLHUP is certainly not done right. But poll() doesn't
341 * have a notion of HUP in just one direction, and for a
342 * socket the read side is more interesting.
343 *
344 * Some poll() documentation says that POLLHUP is incompatible
345 * with the POLLOUT/POLLWR flags, so somebody should check this
346 * all. But careful, it tends to be safer to return too many
347 * bits than too few, and you can easily break real applications
348 * if you don't tell them that something has hung up!
349 *
350 * Check-me.
351 *
352 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
353 * our fs/select.c). It means that after we received EOF,
354 * poll always returns immediately, making impossible poll() on write()
355 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
356 * if and only if shutdown has been made in both directions.
357 * Actually, it is interesting to look how Solaris and DUX
358 * solve this dilemma. I would prefer, if PULLHUP were maskable,
359 * then we could set it on SND_SHUTDOWN. BTW examples given
360 * in Stevens' books assume exactly this behaviour, it explains
361 * why PULLHUP is incompatible with POLLOUT. --ANK
362 *
363 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
364 * blocking on fresh not-connected or disconnected socket. --ANK
365 */
366 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
367 mask |= POLLHUP;
368 if (sk->sk_shutdown & RCV_SHUTDOWN)
369 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
370
371 /* Connected? */
372 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
373 /* Potential race condition. If read of tp below will
374 * escape above sk->sk_state, we can be illegally awaken
375 * in SYN_* states. */
376 if ((tp->rcv_nxt != tp->copied_seq) &&
377 (tp->urg_seq != tp->copied_seq ||
378 tp->rcv_nxt != tp->copied_seq + 1 ||
379 sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
380 mask |= POLLIN | POLLRDNORM;
381
382 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
383 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
384 mask |= POLLOUT | POLLWRNORM;
385 } else { /* send SIGIO later */
386 set_bit(SOCK_ASYNC_NOSPACE,
387 &sk->sk_socket->flags);
388 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
389
390 /* Race breaker. If space is freed after
391 * wspace test but before the flags are set,
392 * IO signal will be lost.
393 */
394 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
395 mask |= POLLOUT | POLLWRNORM;
396 }
397 }
398
399 if (tp->urg_data & TCP_URG_VALID)
400 mask |= POLLPRI;
401 }
402 return mask;
403 }
404
405 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
406 {
407 struct tcp_sock *tp = tcp_sk(sk);
408 int answ;
409
410 switch (cmd) {
411 case SIOCINQ:
412 if (sk->sk_state == TCP_LISTEN)
413 return -EINVAL;
414
415 lock_sock(sk);
416 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
417 answ = 0;
418 else if (sock_flag(sk, SOCK_URGINLINE) ||
419 !tp->urg_data ||
420 before(tp->urg_seq, tp->copied_seq) ||
421 !before(tp->urg_seq, tp->rcv_nxt)) {
422 answ = tp->rcv_nxt - tp->copied_seq;
423
424 /* Subtract 1, if FIN is in queue. */
425 if (answ && !skb_queue_empty(&sk->sk_receive_queue))
426 answ -=
427 tcp_hdr((struct sk_buff *)sk->sk_receive_queue.prev)->fin;
428 } else
429 answ = tp->urg_seq - tp->copied_seq;
430 release_sock(sk);
431 break;
432 case SIOCATMARK:
433 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
434 break;
435 case SIOCOUTQ:
436 if (sk->sk_state == TCP_LISTEN)
437 return -EINVAL;
438
439 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
440 answ = 0;
441 else
442 answ = tp->write_seq - tp->snd_una;
443 break;
444 default:
445 return -ENOIOCTLCMD;
446 }
447
448 return put_user(answ, (int __user *)arg);
449 }
450
451 static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
452 {
453 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
454 tp->pushed_seq = tp->write_seq;
455 }
456
457 static inline int forced_push(struct tcp_sock *tp)
458 {
459 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
460 }
461
462 static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
463 {
464 struct tcp_sock *tp = tcp_sk(sk);
465 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
466
467 skb->csum = 0;
468 tcb->seq = tcb->end_seq = tp->write_seq;
469 tcb->flags = TCPCB_FLAG_ACK;
470 tcb->sacked = 0;
471 skb_header_release(skb);
472 tcp_add_write_queue_tail(sk, skb);
473 sk_charge_skb(sk, skb);
474 if (tp->nonagle & TCP_NAGLE_PUSH)
475 tp->nonagle &= ~TCP_NAGLE_PUSH;
476 }
477
478 static inline void tcp_mark_urg(struct tcp_sock *tp, int flags,
479 struct sk_buff *skb)
480 {
481 if (flags & MSG_OOB) {
482 tp->urg_mode = 1;
483 tp->snd_up = tp->write_seq;
484 TCP_SKB_CB(skb)->sacked |= TCPCB_URG;
485 }
486 }
487
488 static inline void tcp_push(struct sock *sk, int flags, int mss_now,
489 int nonagle)
490 {
491 struct tcp_sock *tp = tcp_sk(sk);
492
493 if (tcp_send_head(sk)) {
494 struct sk_buff *skb = tcp_write_queue_tail(sk);
495 if (!(flags & MSG_MORE) || forced_push(tp))
496 tcp_mark_push(tp, skb);
497 tcp_mark_urg(tp, flags, skb);
498 __tcp_push_pending_frames(sk, mss_now,
499 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
500 }
501 }
502
503 static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
504 size_t psize, int flags)
505 {
506 struct tcp_sock *tp = tcp_sk(sk);
507 int mss_now, size_goal;
508 int err;
509 ssize_t copied;
510 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
511
512 /* Wait for a connection to finish. */
513 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
514 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
515 goto out_err;
516
517 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
518
519 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
520 size_goal = tp->xmit_size_goal;
521 copied = 0;
522
523 err = -EPIPE;
524 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
525 goto do_error;
526
527 while (psize > 0) {
528 struct sk_buff *skb = tcp_write_queue_tail(sk);
529 struct page *page = pages[poffset / PAGE_SIZE];
530 int copy, i, can_coalesce;
531 int offset = poffset % PAGE_SIZE;
532 int size = min_t(size_t, psize, PAGE_SIZE - offset);
533
534 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0) {
535 new_segment:
536 if (!sk_stream_memory_free(sk))
537 goto wait_for_sndbuf;
538
539 skb = sk_stream_alloc_pskb(sk, 0, 0,
540 sk->sk_allocation);
541 if (!skb)
542 goto wait_for_memory;
543
544 skb_entail(sk, skb);
545 copy = size_goal;
546 }
547
548 if (copy > size)
549 copy = size;
550
551 i = skb_shinfo(skb)->nr_frags;
552 can_coalesce = skb_can_coalesce(skb, i, page, offset);
553 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
554 tcp_mark_push(tp, skb);
555 goto new_segment;
556 }
557 if (!sk_stream_wmem_schedule(sk, copy))
558 goto wait_for_memory;
559
560 if (can_coalesce) {
561 skb_shinfo(skb)->frags[i - 1].size += copy;
562 } else {
563 get_page(page);
564 skb_fill_page_desc(skb, i, page, offset, copy);
565 }
566
567 skb->len += copy;
568 skb->data_len += copy;
569 skb->truesize += copy;
570 sk->sk_wmem_queued += copy;
571 sk->sk_forward_alloc -= copy;
572 skb->ip_summed = CHECKSUM_PARTIAL;
573 tp->write_seq += copy;
574 TCP_SKB_CB(skb)->end_seq += copy;
575 skb_shinfo(skb)->gso_segs = 0;
576
577 if (!copied)
578 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
579
580 copied += copy;
581 poffset += copy;
582 if (!(psize -= copy))
583 goto out;
584
585 if (skb->len < mss_now || (flags & MSG_OOB))
586 continue;
587
588 if (forced_push(tp)) {
589 tcp_mark_push(tp, skb);
590 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
591 } else if (skb == tcp_send_head(sk))
592 tcp_push_one(sk, mss_now);
593 continue;
594
595 wait_for_sndbuf:
596 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
597 wait_for_memory:
598 if (copied)
599 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
600
601 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
602 goto do_error;
603
604 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
605 size_goal = tp->xmit_size_goal;
606 }
607
608 out:
609 if (copied)
610 tcp_push(sk, flags, mss_now, tp->nonagle);
611 return copied;
612
613 do_error:
614 if (copied)
615 goto out;
616 out_err:
617 return sk_stream_error(sk, flags, err);
618 }
619
620 ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
621 size_t size, int flags)
622 {
623 ssize_t res;
624 struct sock *sk = sock->sk;
625
626 if (!(sk->sk_route_caps & NETIF_F_SG) ||
627 !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
628 return sock_no_sendpage(sock, page, offset, size, flags);
629
630 lock_sock(sk);
631 TCP_CHECK_TIMER(sk);
632 res = do_tcp_sendpages(sk, &page, offset, size, flags);
633 TCP_CHECK_TIMER(sk);
634 release_sock(sk);
635 return res;
636 }
637
638 #define TCP_PAGE(sk) (sk->sk_sndmsg_page)
639 #define TCP_OFF(sk) (sk->sk_sndmsg_off)
640
641 static inline int select_size(struct sock *sk)
642 {
643 struct tcp_sock *tp = tcp_sk(sk);
644 int tmp = tp->mss_cache;
645
646 if (sk->sk_route_caps & NETIF_F_SG) {
647 if (sk_can_gso(sk))
648 tmp = 0;
649 else {
650 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
651
652 if (tmp >= pgbreak &&
653 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
654 tmp = pgbreak;
655 }
656 }
657
658 return tmp;
659 }
660
661 int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
662 size_t size)
663 {
664 struct iovec *iov;
665 struct tcp_sock *tp = tcp_sk(sk);
666 struct sk_buff *skb;
667 int iovlen, flags;
668 int mss_now, size_goal;
669 int err, copied;
670 long timeo;
671
672 lock_sock(sk);
673 TCP_CHECK_TIMER(sk);
674
675 flags = msg->msg_flags;
676 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
677
678 /* Wait for a connection to finish. */
679 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
680 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
681 goto out_err;
682
683 /* This should be in poll */
684 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
685
686 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
687 size_goal = tp->xmit_size_goal;
688
689 /* Ok commence sending. */
690 iovlen = msg->msg_iovlen;
691 iov = msg->msg_iov;
692 copied = 0;
693
694 err = -EPIPE;
695 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
696 goto do_error;
697
698 while (--iovlen >= 0) {
699 int seglen = iov->iov_len;
700 unsigned char __user *from = iov->iov_base;
701
702 iov++;
703
704 while (seglen > 0) {
705 int copy;
706
707 skb = tcp_write_queue_tail(sk);
708
709 if (!tcp_send_head(sk) ||
710 (copy = size_goal - skb->len) <= 0) {
711
712 new_segment:
713 /* Allocate new segment. If the interface is SG,
714 * allocate skb fitting to single page.
715 */
716 if (!sk_stream_memory_free(sk))
717 goto wait_for_sndbuf;
718
719 skb = sk_stream_alloc_pskb(sk, select_size(sk),
720 0, sk->sk_allocation);
721 if (!skb)
722 goto wait_for_memory;
723
724 /*
725 * Check whether we can use HW checksum.
726 */
727 if (sk->sk_route_caps & NETIF_F_ALL_CSUM)
728 skb->ip_summed = CHECKSUM_PARTIAL;
729
730 skb_entail(sk, skb);
731 copy = size_goal;
732 }
733
734 /* Try to append data to the end of skb. */
735 if (copy > seglen)
736 copy = seglen;
737
738 /* Where to copy to? */
739 if (skb_tailroom(skb) > 0) {
740 /* We have some space in skb head. Superb! */
741 if (copy > skb_tailroom(skb))
742 copy = skb_tailroom(skb);
743 if ((err = skb_add_data(skb, from, copy)) != 0)
744 goto do_fault;
745 } else {
746 int merge = 0;
747 int i = skb_shinfo(skb)->nr_frags;
748 struct page *page = TCP_PAGE(sk);
749 int off = TCP_OFF(sk);
750
751 if (skb_can_coalesce(skb, i, page, off) &&
752 off != PAGE_SIZE) {
753 /* We can extend the last page
754 * fragment. */
755 merge = 1;
756 } else if (i == MAX_SKB_FRAGS ||
757 (!i &&
758 !(sk->sk_route_caps & NETIF_F_SG))) {
759 /* Need to add new fragment and cannot
760 * do this because interface is non-SG,
761 * or because all the page slots are
762 * busy. */
763 tcp_mark_push(tp, skb);
764 goto new_segment;
765 } else if (page) {
766 if (off == PAGE_SIZE) {
767 put_page(page);
768 TCP_PAGE(sk) = page = NULL;
769 off = 0;
770 }
771 } else
772 off = 0;
773
774 if (copy > PAGE_SIZE - off)
775 copy = PAGE_SIZE - off;
776
777 if (!sk_stream_wmem_schedule(sk, copy))
778 goto wait_for_memory;
779
780 if (!page) {
781 /* Allocate new cache page. */
782 if (!(page = sk_stream_alloc_page(sk)))
783 goto wait_for_memory;
784 }
785
786 /* Time to copy data. We are close to
787 * the end! */
788 err = skb_copy_to_page(sk, from, skb, page,
789 off, copy);
790 if (err) {
791 /* If this page was new, give it to the
792 * socket so it does not get leaked.
793 */
794 if (!TCP_PAGE(sk)) {
795 TCP_PAGE(sk) = page;
796 TCP_OFF(sk) = 0;
797 }
798 goto do_error;
799 }
800
801 /* Update the skb. */
802 if (merge) {
803 skb_shinfo(skb)->frags[i - 1].size +=
804 copy;
805 } else {
806 skb_fill_page_desc(skb, i, page, off, copy);
807 if (TCP_PAGE(sk)) {
808 get_page(page);
809 } else if (off + copy < PAGE_SIZE) {
810 get_page(page);
811 TCP_PAGE(sk) = page;
812 }
813 }
814
815 TCP_OFF(sk) = off + copy;
816 }
817
818 if (!copied)
819 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
820
821 tp->write_seq += copy;
822 TCP_SKB_CB(skb)->end_seq += copy;
823 skb_shinfo(skb)->gso_segs = 0;
824
825 from += copy;
826 copied += copy;
827 if ((seglen -= copy) == 0 && iovlen == 0)
828 goto out;
829
830 if (skb->len < mss_now || (flags & MSG_OOB))
831 continue;
832
833 if (forced_push(tp)) {
834 tcp_mark_push(tp, skb);
835 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
836 } else if (skb == tcp_send_head(sk))
837 tcp_push_one(sk, mss_now);
838 continue;
839
840 wait_for_sndbuf:
841 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
842 wait_for_memory:
843 if (copied)
844 tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
845
846 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
847 goto do_error;
848
849 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
850 size_goal = tp->xmit_size_goal;
851 }
852 }
853
854 out:
855 if (copied)
856 tcp_push(sk, flags, mss_now, tp->nonagle);
857 TCP_CHECK_TIMER(sk);
858 release_sock(sk);
859 return copied;
860
861 do_fault:
862 if (!skb->len) {
863 tcp_unlink_write_queue(skb, sk);
864 /* It is the one place in all of TCP, except connection
865 * reset, where we can be unlinking the send_head.
866 */
867 tcp_check_send_head(sk, skb);
868 sk_stream_free_skb(sk, skb);
869 }
870
871 do_error:
872 if (copied)
873 goto out;
874 out_err:
875 err = sk_stream_error(sk, flags, err);
876 TCP_CHECK_TIMER(sk);
877 release_sock(sk);
878 return err;
879 }
880
881 /*
882 * Handle reading urgent data. BSD has very simple semantics for
883 * this, no blocking and very strange errors 8)
884 */
885
886 static int tcp_recv_urg(struct sock *sk, long timeo,
887 struct msghdr *msg, int len, int flags,
888 int *addr_len)
889 {
890 struct tcp_sock *tp = tcp_sk(sk);
891
892 /* No URG data to read. */
893 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
894 tp->urg_data == TCP_URG_READ)
895 return -EINVAL; /* Yes this is right ! */
896
897 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
898 return -ENOTCONN;
899
900 if (tp->urg_data & TCP_URG_VALID) {
901 int err = 0;
902 char c = tp->urg_data;
903
904 if (!(flags & MSG_PEEK))
905 tp->urg_data = TCP_URG_READ;
906
907 /* Read urgent data. */
908 msg->msg_flags |= MSG_OOB;
909
910 if (len > 0) {
911 if (!(flags & MSG_TRUNC))
912 err = memcpy_toiovec(msg->msg_iov, &c, 1);
913 len = 1;
914 } else
915 msg->msg_flags |= MSG_TRUNC;
916
917 return err ? -EFAULT : len;
918 }
919
920 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
921 return 0;
922
923 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
924 * the available implementations agree in this case:
925 * this call should never block, independent of the
926 * blocking state of the socket.
927 * Mike <pall@rz.uni-karlsruhe.de>
928 */
929 return -EAGAIN;
930 }
931
932 /* Clean up the receive buffer for full frames taken by the user,
933 * then send an ACK if necessary. COPIED is the number of bytes
934 * tcp_recvmsg has given to the user so far, it speeds up the
935 * calculation of whether or not we must ACK for the sake of
936 * a window update.
937 */
938 void tcp_cleanup_rbuf(struct sock *sk, int copied)
939 {
940 struct tcp_sock *tp = tcp_sk(sk);
941 int time_to_ack = 0;
942
943 #if TCP_DEBUG
944 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
945
946 BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
947 #endif
948
949 if (inet_csk_ack_scheduled(sk)) {
950 const struct inet_connection_sock *icsk = inet_csk(sk);
951 /* Delayed ACKs frequently hit locked sockets during bulk
952 * receive. */
953 if (icsk->icsk_ack.blocked ||
954 /* Once-per-two-segments ACK was not sent by tcp_input.c */
955 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
956 /*
957 * If this read emptied read buffer, we send ACK, if
958 * connection is not bidirectional, user drained
959 * receive buffer and there was a small segment
960 * in queue.
961 */
962 (copied > 0 &&
963 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
964 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
965 !icsk->icsk_ack.pingpong)) &&
966 !atomic_read(&sk->sk_rmem_alloc)))
967 time_to_ack = 1;
968 }
969
970 /* We send an ACK if we can now advertise a non-zero window
971 * which has been raised "significantly".
972 *
973 * Even if window raised up to infinity, do not send window open ACK
974 * in states, where we will not receive more. It is useless.
975 */
976 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
977 __u32 rcv_window_now = tcp_receive_window(tp);
978
979 /* Optimize, __tcp_select_window() is not cheap. */
980 if (2*rcv_window_now <= tp->window_clamp) {
981 __u32 new_window = __tcp_select_window(sk);
982
983 /* Send ACK now, if this read freed lots of space
984 * in our buffer. Certainly, new_window is new window.
985 * We can advertise it now, if it is not less than current one.
986 * "Lots" means "at least twice" here.
987 */
988 if (new_window && new_window >= 2 * rcv_window_now)
989 time_to_ack = 1;
990 }
991 }
992 if (time_to_ack)
993 tcp_send_ack(sk);
994 }
995
996 static void tcp_prequeue_process(struct sock *sk)
997 {
998 struct sk_buff *skb;
999 struct tcp_sock *tp = tcp_sk(sk);
1000
1001 NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
1002
1003 /* RX process wants to run with disabled BHs, though it is not
1004 * necessary */
1005 local_bh_disable();
1006 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1007 sk->sk_backlog_rcv(sk, skb);
1008 local_bh_enable();
1009
1010 /* Clear memory counter. */
1011 tp->ucopy.memory = 0;
1012 }
1013
1014 static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1015 {
1016 struct sk_buff *skb;
1017 u32 offset;
1018
1019 skb_queue_walk(&sk->sk_receive_queue, skb) {
1020 offset = seq - TCP_SKB_CB(skb)->seq;
1021 if (tcp_hdr(skb)->syn)
1022 offset--;
1023 if (offset < skb->len || tcp_hdr(skb)->fin) {
1024 *off = offset;
1025 return skb;
1026 }
1027 }
1028 return NULL;
1029 }
1030
1031 /*
1032 * This routine provides an alternative to tcp_recvmsg() for routines
1033 * that would like to handle copying from skbuffs directly in 'sendfile'
1034 * fashion.
1035 * Note:
1036 * - It is assumed that the socket was locked by the caller.
1037 * - The routine does not block.
1038 * - At present, there is no support for reading OOB data
1039 * or for 'peeking' the socket using this routine
1040 * (although both would be easy to implement).
1041 */
1042 int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1043 sk_read_actor_t recv_actor)
1044 {
1045 struct sk_buff *skb;
1046 struct tcp_sock *tp = tcp_sk(sk);
1047 u32 seq = tp->copied_seq;
1048 u32 offset;
1049 int copied = 0;
1050
1051 if (sk->sk_state == TCP_LISTEN)
1052 return -ENOTCONN;
1053 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1054 if (offset < skb->len) {
1055 size_t used, len;
1056
1057 len = skb->len - offset;
1058 /* Stop reading if we hit a patch of urgent data */
1059 if (tp->urg_data) {
1060 u32 urg_offset = tp->urg_seq - seq;
1061 if (urg_offset < len)
1062 len = urg_offset;
1063 if (!len)
1064 break;
1065 }
1066 used = recv_actor(desc, skb, offset, len);
1067 if (used < 0) {
1068 if (!copied)
1069 copied = used;
1070 break;
1071 } else if (used <= len) {
1072 seq += used;
1073 copied += used;
1074 offset += used;
1075 }
1076 if (offset != skb->len)
1077 break;
1078 }
1079 if (tcp_hdr(skb)->fin) {
1080 sk_eat_skb(sk, skb, 0);
1081 ++seq;
1082 break;
1083 }
1084 sk_eat_skb(sk, skb, 0);
1085 if (!desc->count)
1086 break;
1087 }
1088 tp->copied_seq = seq;
1089
1090 tcp_rcv_space_adjust(sk);
1091
1092 /* Clean up data we have read: This will do ACK frames. */
1093 if (copied > 0)
1094 tcp_cleanup_rbuf(sk, copied);
1095 return copied;
1096 }
1097
1098 /*
1099 * This routine copies from a sock struct into the user buffer.
1100 *
1101 * Technical note: in 2.3 we work on _locked_ socket, so that
1102 * tricks with *seq access order and skb->users are not required.
1103 * Probably, code can be easily improved even more.
1104 */
1105
1106 int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1107 size_t len, int nonblock, int flags, int *addr_len)
1108 {
1109 struct tcp_sock *tp = tcp_sk(sk);
1110 int copied = 0;
1111 u32 peek_seq;
1112 u32 *seq;
1113 unsigned long used;
1114 int err;
1115 int target; /* Read at least this many bytes */
1116 long timeo;
1117 struct task_struct *user_recv = NULL;
1118 int copied_early = 0;
1119
1120 lock_sock(sk);
1121
1122 TCP_CHECK_TIMER(sk);
1123
1124 err = -ENOTCONN;
1125 if (sk->sk_state == TCP_LISTEN)
1126 goto out;
1127
1128 timeo = sock_rcvtimeo(sk, nonblock);
1129
1130 /* Urgent data needs to be handled specially. */
1131 if (flags & MSG_OOB)
1132 goto recv_urg;
1133
1134 seq = &tp->copied_seq;
1135 if (flags & MSG_PEEK) {
1136 peek_seq = tp->copied_seq;
1137 seq = &peek_seq;
1138 }
1139
1140 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1141
1142 #ifdef CONFIG_NET_DMA
1143 tp->ucopy.dma_chan = NULL;
1144 preempt_disable();
1145 if ((len > sysctl_tcp_dma_copybreak) && !(flags & MSG_PEEK) &&
1146 !sysctl_tcp_low_latency && __get_cpu_var(softnet_data).net_dma) {
1147 preempt_enable_no_resched();
1148 tp->ucopy.pinned_list = dma_pin_iovec_pages(msg->msg_iov, len);
1149 } else
1150 preempt_enable_no_resched();
1151 #endif
1152
1153 do {
1154 struct sk_buff *skb;
1155 u32 offset;
1156
1157 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1158 if (tp->urg_data && tp->urg_seq == *seq) {
1159 if (copied)
1160 break;
1161 if (signal_pending(current)) {
1162 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1163 break;
1164 }
1165 }
1166
1167 /* Next get a buffer. */
1168
1169 skb = skb_peek(&sk->sk_receive_queue);
1170 do {
1171 if (!skb)
1172 break;
1173
1174 /* Now that we have two receive queues this
1175 * shouldn't happen.
1176 */
1177 if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1178 printk(KERN_INFO "recvmsg bug: copied %X "
1179 "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1180 break;
1181 }
1182 offset = *seq - TCP_SKB_CB(skb)->seq;
1183 if (tcp_hdr(skb)->syn)
1184 offset--;
1185 if (offset < skb->len)
1186 goto found_ok_skb;
1187 if (tcp_hdr(skb)->fin)
1188 goto found_fin_ok;
1189 BUG_TRAP(flags & MSG_PEEK);
1190 skb = skb->next;
1191 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1192
1193 /* Well, if we have backlog, try to process it now yet. */
1194
1195 if (copied >= target && !sk->sk_backlog.tail)
1196 break;
1197
1198 if (copied) {
1199 if (sk->sk_err ||
1200 sk->sk_state == TCP_CLOSE ||
1201 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1202 !timeo ||
1203 signal_pending(current) ||
1204 (flags & MSG_PEEK))
1205 break;
1206 } else {
1207 if (sock_flag(sk, SOCK_DONE))
1208 break;
1209
1210 if (sk->sk_err) {
1211 copied = sock_error(sk);
1212 break;
1213 }
1214
1215 if (sk->sk_shutdown & RCV_SHUTDOWN)
1216 break;
1217
1218 if (sk->sk_state == TCP_CLOSE) {
1219 if (!sock_flag(sk, SOCK_DONE)) {
1220 /* This occurs when user tries to read
1221 * from never connected socket.
1222 */
1223 copied = -ENOTCONN;
1224 break;
1225 }
1226 break;
1227 }
1228
1229 if (!timeo) {
1230 copied = -EAGAIN;
1231 break;
1232 }
1233
1234 if (signal_pending(current)) {
1235 copied = sock_intr_errno(timeo);
1236 break;
1237 }
1238 }
1239
1240 tcp_cleanup_rbuf(sk, copied);
1241
1242 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1243 /* Install new reader */
1244 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1245 user_recv = current;
1246 tp->ucopy.task = user_recv;
1247 tp->ucopy.iov = msg->msg_iov;
1248 }
1249
1250 tp->ucopy.len = len;
1251
1252 BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1253 (flags & (MSG_PEEK | MSG_TRUNC)));
1254
1255 /* Ugly... If prequeue is not empty, we have to
1256 * process it before releasing socket, otherwise
1257 * order will be broken at second iteration.
1258 * More elegant solution is required!!!
1259 *
1260 * Look: we have the following (pseudo)queues:
1261 *
1262 * 1. packets in flight
1263 * 2. backlog
1264 * 3. prequeue
1265 * 4. receive_queue
1266 *
1267 * Each queue can be processed only if the next ones
1268 * are empty. At this point we have empty receive_queue.
1269 * But prequeue _can_ be not empty after 2nd iteration,
1270 * when we jumped to start of loop because backlog
1271 * processing added something to receive_queue.
1272 * We cannot release_sock(), because backlog contains
1273 * packets arrived _after_ prequeued ones.
1274 *
1275 * Shortly, algorithm is clear --- to process all
1276 * the queues in order. We could make it more directly,
1277 * requeueing packets from backlog to prequeue, if
1278 * is not empty. It is more elegant, but eats cycles,
1279 * unfortunately.
1280 */
1281 if (!skb_queue_empty(&tp->ucopy.prequeue))
1282 goto do_prequeue;
1283
1284 /* __ Set realtime policy in scheduler __ */
1285 }
1286
1287 if (copied >= target) {
1288 /* Do not sleep, just process backlog. */
1289 release_sock(sk);
1290 lock_sock(sk);
1291 } else
1292 sk_wait_data(sk, &timeo);
1293
1294 #ifdef CONFIG_NET_DMA
1295 tp->ucopy.wakeup = 0;
1296 #endif
1297
1298 if (user_recv) {
1299 int chunk;
1300
1301 /* __ Restore normal policy in scheduler __ */
1302
1303 if ((chunk = len - tp->ucopy.len) != 0) {
1304 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1305 len -= chunk;
1306 copied += chunk;
1307 }
1308
1309 if (tp->rcv_nxt == tp->copied_seq &&
1310 !skb_queue_empty(&tp->ucopy.prequeue)) {
1311 do_prequeue:
1312 tcp_prequeue_process(sk);
1313
1314 if ((chunk = len - tp->ucopy.len) != 0) {
1315 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1316 len -= chunk;
1317 copied += chunk;
1318 }
1319 }
1320 }
1321 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1322 if (net_ratelimit())
1323 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1324 current->comm, current->pid);
1325 peek_seq = tp->copied_seq;
1326 }
1327 continue;
1328
1329 found_ok_skb:
1330 /* Ok so how much can we use? */
1331 used = skb->len - offset;
1332 if (len < used)
1333 used = len;
1334
1335 /* Do we have urgent data here? */
1336 if (tp->urg_data) {
1337 u32 urg_offset = tp->urg_seq - *seq;
1338 if (urg_offset < used) {
1339 if (!urg_offset) {
1340 if (!sock_flag(sk, SOCK_URGINLINE)) {
1341 ++*seq;
1342 offset++;
1343 used--;
1344 if (!used)
1345 goto skip_copy;
1346 }
1347 } else
1348 used = urg_offset;
1349 }
1350 }
1351
1352 if (!(flags & MSG_TRUNC)) {
1353 #ifdef CONFIG_NET_DMA
1354 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
1355 tp->ucopy.dma_chan = get_softnet_dma();
1356
1357 if (tp->ucopy.dma_chan) {
1358 tp->ucopy.dma_cookie = dma_skb_copy_datagram_iovec(
1359 tp->ucopy.dma_chan, skb, offset,
1360 msg->msg_iov, used,
1361 tp->ucopy.pinned_list);
1362
1363 if (tp->ucopy.dma_cookie < 0) {
1364
1365 printk(KERN_ALERT "dma_cookie < 0\n");
1366
1367 /* Exception. Bailout! */
1368 if (!copied)
1369 copied = -EFAULT;
1370 break;
1371 }
1372 if ((offset + used) == skb->len)
1373 copied_early = 1;
1374
1375 } else
1376 #endif
1377 {
1378 err = skb_copy_datagram_iovec(skb, offset,
1379 msg->msg_iov, used);
1380 if (err) {
1381 /* Exception. Bailout! */
1382 if (!copied)
1383 copied = -EFAULT;
1384 break;
1385 }
1386 }
1387 }
1388
1389 *seq += used;
1390 copied += used;
1391 len -= used;
1392
1393 tcp_rcv_space_adjust(sk);
1394
1395 skip_copy:
1396 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1397 tp->urg_data = 0;
1398 tcp_fast_path_check(sk);
1399 }
1400 if (used + offset < skb->len)
1401 continue;
1402
1403 if (tcp_hdr(skb)->fin)
1404 goto found_fin_ok;
1405 if (!(flags & MSG_PEEK)) {
1406 sk_eat_skb(sk, skb, copied_early);
1407 copied_early = 0;
1408 }
1409 continue;
1410
1411 found_fin_ok:
1412 /* Process the FIN. */
1413 ++*seq;
1414 if (!(flags & MSG_PEEK)) {
1415 sk_eat_skb(sk, skb, copied_early);
1416 copied_early = 0;
1417 }
1418 break;
1419 } while (len > 0);
1420
1421 if (user_recv) {
1422 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1423 int chunk;
1424
1425 tp->ucopy.len = copied > 0 ? len : 0;
1426
1427 tcp_prequeue_process(sk);
1428
1429 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1430 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1431 len -= chunk;
1432 copied += chunk;
1433 }
1434 }
1435
1436 tp->ucopy.task = NULL;
1437 tp->ucopy.len = 0;
1438 }
1439
1440 #ifdef CONFIG_NET_DMA
1441 if (tp->ucopy.dma_chan) {
1442 struct sk_buff *skb;
1443 dma_cookie_t done, used;
1444
1445 dma_async_memcpy_issue_pending(tp->ucopy.dma_chan);
1446
1447 while (dma_async_memcpy_complete(tp->ucopy.dma_chan,
1448 tp->ucopy.dma_cookie, &done,
1449 &used) == DMA_IN_PROGRESS) {
1450 /* do partial cleanup of sk_async_wait_queue */
1451 while ((skb = skb_peek(&sk->sk_async_wait_queue)) &&
1452 (dma_async_is_complete(skb->dma_cookie, done,
1453 used) == DMA_SUCCESS)) {
1454 __skb_dequeue(&sk->sk_async_wait_queue);
1455 kfree_skb(skb);
1456 }
1457 }
1458
1459 /* Safe to free early-copied skbs now */
1460 __skb_queue_purge(&sk->sk_async_wait_queue);
1461 dma_chan_put(tp->ucopy.dma_chan);
1462 tp->ucopy.dma_chan = NULL;
1463 }
1464 if (tp->ucopy.pinned_list) {
1465 dma_unpin_iovec_pages(tp->ucopy.pinned_list);
1466 tp->ucopy.pinned_list = NULL;
1467 }
1468 #endif
1469
1470 /* According to UNIX98, msg_name/msg_namelen are ignored
1471 * on connected socket. I was just happy when found this 8) --ANK
1472 */
1473
1474 /* Clean up data we have read: This will do ACK frames. */
1475 tcp_cleanup_rbuf(sk, copied);
1476
1477 TCP_CHECK_TIMER(sk);
1478 release_sock(sk);
1479 return copied;
1480
1481 out:
1482 TCP_CHECK_TIMER(sk);
1483 release_sock(sk);
1484 return err;
1485
1486 recv_urg:
1487 err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1488 goto out;
1489 }
1490
1491 /*
1492 * State processing on a close. This implements the state shift for
1493 * sending our FIN frame. Note that we only send a FIN for some
1494 * states. A shutdown() may have already sent the FIN, or we may be
1495 * closed.
1496 */
1497
1498 static const unsigned char new_state[16] = {
1499 /* current state: new state: action: */
1500 /* (Invalid) */ TCP_CLOSE,
1501 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1502 /* TCP_SYN_SENT */ TCP_CLOSE,
1503 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1504 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1505 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1506 /* TCP_TIME_WAIT */ TCP_CLOSE,
1507 /* TCP_CLOSE */ TCP_CLOSE,
1508 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1509 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1510 /* TCP_LISTEN */ TCP_CLOSE,
1511 /* TCP_CLOSING */ TCP_CLOSING,
1512 };
1513
1514 static int tcp_close_state(struct sock *sk)
1515 {
1516 int next = (int)new_state[sk->sk_state];
1517 int ns = next & TCP_STATE_MASK;
1518
1519 tcp_set_state(sk, ns);
1520
1521 return next & TCP_ACTION_FIN;
1522 }
1523
1524 /*
1525 * Shutdown the sending side of a connection. Much like close except
1526 * that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1527 */
1528
1529 void tcp_shutdown(struct sock *sk, int how)
1530 {
1531 /* We need to grab some memory, and put together a FIN,
1532 * and then put it into the queue to be sent.
1533 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1534 */
1535 if (!(how & SEND_SHUTDOWN))
1536 return;
1537
1538 /* If we've already sent a FIN, or it's a closed state, skip this. */
1539 if ((1 << sk->sk_state) &
1540 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1541 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1542 /* Clear out any half completed packets. FIN if needed. */
1543 if (tcp_close_state(sk))
1544 tcp_send_fin(sk);
1545 }
1546 }
1547
1548 void tcp_close(struct sock *sk, long timeout)
1549 {
1550 struct sk_buff *skb;
1551 int data_was_unread = 0;
1552 int state;
1553
1554 lock_sock(sk);
1555 sk->sk_shutdown = SHUTDOWN_MASK;
1556
1557 if (sk->sk_state == TCP_LISTEN) {
1558 tcp_set_state(sk, TCP_CLOSE);
1559
1560 /* Special case. */
1561 inet_csk_listen_stop(sk);
1562
1563 goto adjudge_to_death;
1564 }
1565
1566 /* We need to flush the recv. buffs. We do this only on the
1567 * descriptor close, not protocol-sourced closes, because the
1568 * reader process may not have drained the data yet!
1569 */
1570 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1571 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1572 tcp_hdr(skb)->fin;
1573 data_was_unread += len;
1574 __kfree_skb(skb);
1575 }
1576
1577 sk_stream_mem_reclaim(sk);
1578
1579 /* As outlined in RFC 2525, section 2.17, we send a RST here because
1580 * data was lost. To witness the awful effects of the old behavior of
1581 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
1582 * GET in an FTP client, suspend the process, wait for the client to
1583 * advertise a zero window, then kill -9 the FTP client, wheee...
1584 * Note: timeout is always zero in such a case.
1585 */
1586 if (data_was_unread) {
1587 /* Unread data was tossed, zap the connection. */
1588 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1589 tcp_set_state(sk, TCP_CLOSE);
1590 tcp_send_active_reset(sk, GFP_KERNEL);
1591 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1592 /* Check zero linger _after_ checking for unread data. */
1593 sk->sk_prot->disconnect(sk, 0);
1594 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1595 } else if (tcp_close_state(sk)) {
1596 /* We FIN if the application ate all the data before
1597 * zapping the connection.
1598 */
1599
1600 /* RED-PEN. Formally speaking, we have broken TCP state
1601 * machine. State transitions:
1602 *
1603 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1604 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1605 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1606 *
1607 * are legal only when FIN has been sent (i.e. in window),
1608 * rather than queued out of window. Purists blame.
1609 *
1610 * F.e. "RFC state" is ESTABLISHED,
1611 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1612 *
1613 * The visible declinations are that sometimes
1614 * we enter time-wait state, when it is not required really
1615 * (harmless), do not send active resets, when they are
1616 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1617 * they look as CLOSING or LAST_ACK for Linux)
1618 * Probably, I missed some more holelets.
1619 * --ANK
1620 */
1621 tcp_send_fin(sk);
1622 }
1623
1624 sk_stream_wait_close(sk, timeout);
1625
1626 adjudge_to_death:
1627 state = sk->sk_state;
1628 sock_hold(sk);
1629 sock_orphan(sk);
1630 atomic_inc(sk->sk_prot->orphan_count);
1631
1632 /* It is the last release_sock in its life. It will remove backlog. */
1633 release_sock(sk);
1634
1635
1636 /* Now socket is owned by kernel and we acquire BH lock
1637 to finish close. No need to check for user refs.
1638 */
1639 local_bh_disable();
1640 bh_lock_sock(sk);
1641 BUG_TRAP(!sock_owned_by_user(sk));
1642
1643 /* Have we already been destroyed by a softirq or backlog? */
1644 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
1645 goto out;
1646
1647 /* This is a (useful) BSD violating of the RFC. There is a
1648 * problem with TCP as specified in that the other end could
1649 * keep a socket open forever with no application left this end.
1650 * We use a 3 minute timeout (about the same as BSD) then kill
1651 * our end. If they send after that then tough - BUT: long enough
1652 * that we won't make the old 4*rto = almost no time - whoops
1653 * reset mistake.
1654 *
1655 * Nope, it was not mistake. It is really desired behaviour
1656 * f.e. on http servers, when such sockets are useless, but
1657 * consume significant resources. Let's do it with special
1658 * linger2 option. --ANK
1659 */
1660
1661 if (sk->sk_state == TCP_FIN_WAIT2) {
1662 struct tcp_sock *tp = tcp_sk(sk);
1663 if (tp->linger2 < 0) {
1664 tcp_set_state(sk, TCP_CLOSE);
1665 tcp_send_active_reset(sk, GFP_ATOMIC);
1666 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1667 } else {
1668 const int tmo = tcp_fin_time(sk);
1669
1670 if (tmo > TCP_TIMEWAIT_LEN) {
1671 inet_csk_reset_keepalive_timer(sk,
1672 tmo - TCP_TIMEWAIT_LEN);
1673 } else {
1674 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1675 goto out;
1676 }
1677 }
1678 }
1679 if (sk->sk_state != TCP_CLOSE) {
1680 sk_stream_mem_reclaim(sk);
1681 if (tcp_too_many_orphans(sk,
1682 atomic_read(sk->sk_prot->orphan_count))) {
1683 if (net_ratelimit())
1684 printk(KERN_INFO "TCP: too many of orphaned "
1685 "sockets\n");
1686 tcp_set_state(sk, TCP_CLOSE);
1687 tcp_send_active_reset(sk, GFP_ATOMIC);
1688 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1689 }
1690 }
1691
1692 if (sk->sk_state == TCP_CLOSE)
1693 inet_csk_destroy_sock(sk);
1694 /* Otherwise, socket is reprieved until protocol close. */
1695
1696 out:
1697 bh_unlock_sock(sk);
1698 local_bh_enable();
1699 sock_put(sk);
1700 }
1701
1702 /* These states need RST on ABORT according to RFC793 */
1703
1704 static inline int tcp_need_reset(int state)
1705 {
1706 return (1 << state) &
1707 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1708 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1709 }
1710
1711 int tcp_disconnect(struct sock *sk, int flags)
1712 {
1713 struct inet_sock *inet = inet_sk(sk);
1714 struct inet_connection_sock *icsk = inet_csk(sk);
1715 struct tcp_sock *tp = tcp_sk(sk);
1716 int err = 0;
1717 int old_state = sk->sk_state;
1718
1719 if (old_state != TCP_CLOSE)
1720 tcp_set_state(sk, TCP_CLOSE);
1721
1722 /* ABORT function of RFC793 */
1723 if (old_state == TCP_LISTEN) {
1724 inet_csk_listen_stop(sk);
1725 } else if (tcp_need_reset(old_state) ||
1726 (tp->snd_nxt != tp->write_seq &&
1727 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1728 /* The last check adjusts for discrepancy of Linux wrt. RFC
1729 * states
1730 */
1731 tcp_send_active_reset(sk, gfp_any());
1732 sk->sk_err = ECONNRESET;
1733 } else if (old_state == TCP_SYN_SENT)
1734 sk->sk_err = ECONNRESET;
1735
1736 tcp_clear_xmit_timers(sk);
1737 __skb_queue_purge(&sk->sk_receive_queue);
1738 tcp_write_queue_purge(sk);
1739 __skb_queue_purge(&tp->out_of_order_queue);
1740 #ifdef CONFIG_NET_DMA
1741 __skb_queue_purge(&sk->sk_async_wait_queue);
1742 #endif
1743
1744 inet->dport = 0;
1745
1746 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1747 inet_reset_saddr(sk);
1748
1749 sk->sk_shutdown = 0;
1750 sock_reset_flag(sk, SOCK_DONE);
1751 tp->srtt = 0;
1752 if ((tp->write_seq += tp->max_window + 2) == 0)
1753 tp->write_seq = 1;
1754 icsk->icsk_backoff = 0;
1755 tp->snd_cwnd = 2;
1756 icsk->icsk_probes_out = 0;
1757 tp->packets_out = 0;
1758 tp->snd_ssthresh = 0x7fffffff;
1759 tp->snd_cwnd_cnt = 0;
1760 tp->bytes_acked = 0;
1761 tcp_set_ca_state(sk, TCP_CA_Open);
1762 tcp_clear_retrans(tp);
1763 inet_csk_delack_init(sk);
1764 tcp_init_send_head(sk);
1765 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
1766 __sk_dst_reset(sk);
1767
1768 BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
1769
1770 sk->sk_error_report(sk);
1771 return err;
1772 }
1773
1774 /*
1775 * Socket option code for TCP.
1776 */
1777 static int do_tcp_setsockopt(struct sock *sk, int level,
1778 int optname, char __user *optval, int optlen)
1779 {
1780 struct tcp_sock *tp = tcp_sk(sk);
1781 struct inet_connection_sock *icsk = inet_csk(sk);
1782 int val;
1783 int err = 0;
1784
1785 /* This is a string value all the others are int's */
1786 if (optname == TCP_CONGESTION) {
1787 char name[TCP_CA_NAME_MAX];
1788
1789 if (optlen < 1)
1790 return -EINVAL;
1791
1792 val = strncpy_from_user(name, optval,
1793 min(TCP_CA_NAME_MAX-1, optlen));
1794 if (val < 0)
1795 return -EFAULT;
1796 name[val] = 0;
1797
1798 lock_sock(sk);
1799 err = tcp_set_congestion_control(sk, name);
1800 release_sock(sk);
1801 return err;
1802 }
1803
1804 if (optlen < sizeof(int))
1805 return -EINVAL;
1806
1807 if (get_user(val, (int __user *)optval))
1808 return -EFAULT;
1809
1810 lock_sock(sk);
1811
1812 switch (optname) {
1813 case TCP_MAXSEG:
1814 /* Values greater than interface MTU won't take effect. However
1815 * at the point when this call is done we typically don't yet
1816 * know which interface is going to be used */
1817 if (val < 8 || val > MAX_TCP_WINDOW) {
1818 err = -EINVAL;
1819 break;
1820 }
1821 tp->rx_opt.user_mss = val;
1822 break;
1823
1824 case TCP_NODELAY:
1825 if (val) {
1826 /* TCP_NODELAY is weaker than TCP_CORK, so that
1827 * this option on corked socket is remembered, but
1828 * it is not activated until cork is cleared.
1829 *
1830 * However, when TCP_NODELAY is set we make
1831 * an explicit push, which overrides even TCP_CORK
1832 * for currently queued segments.
1833 */
1834 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1835 tcp_push_pending_frames(sk);
1836 } else {
1837 tp->nonagle &= ~TCP_NAGLE_OFF;
1838 }
1839 break;
1840
1841 case TCP_CORK:
1842 /* When set indicates to always queue non-full frames.
1843 * Later the user clears this option and we transmit
1844 * any pending partial frames in the queue. This is
1845 * meant to be used alongside sendfile() to get properly
1846 * filled frames when the user (for example) must write
1847 * out headers with a write() call first and then use
1848 * sendfile to send out the data parts.
1849 *
1850 * TCP_CORK can be set together with TCP_NODELAY and it is
1851 * stronger than TCP_NODELAY.
1852 */
1853 if (val) {
1854 tp->nonagle |= TCP_NAGLE_CORK;
1855 } else {
1856 tp->nonagle &= ~TCP_NAGLE_CORK;
1857 if (tp->nonagle&TCP_NAGLE_OFF)
1858 tp->nonagle |= TCP_NAGLE_PUSH;
1859 tcp_push_pending_frames(sk);
1860 }
1861 break;
1862
1863 case TCP_KEEPIDLE:
1864 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1865 err = -EINVAL;
1866 else {
1867 tp->keepalive_time = val * HZ;
1868 if (sock_flag(sk, SOCK_KEEPOPEN) &&
1869 !((1 << sk->sk_state) &
1870 (TCPF_CLOSE | TCPF_LISTEN))) {
1871 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1872 if (tp->keepalive_time > elapsed)
1873 elapsed = tp->keepalive_time - elapsed;
1874 else
1875 elapsed = 0;
1876 inet_csk_reset_keepalive_timer(sk, elapsed);
1877 }
1878 }
1879 break;
1880 case TCP_KEEPINTVL:
1881 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1882 err = -EINVAL;
1883 else
1884 tp->keepalive_intvl = val * HZ;
1885 break;
1886 case TCP_KEEPCNT:
1887 if (val < 1 || val > MAX_TCP_KEEPCNT)
1888 err = -EINVAL;
1889 else
1890 tp->keepalive_probes = val;
1891 break;
1892 case TCP_SYNCNT:
1893 if (val < 1 || val > MAX_TCP_SYNCNT)
1894 err = -EINVAL;
1895 else
1896 icsk->icsk_syn_retries = val;
1897 break;
1898
1899 case TCP_LINGER2:
1900 if (val < 0)
1901 tp->linger2 = -1;
1902 else if (val > sysctl_tcp_fin_timeout / HZ)
1903 tp->linger2 = 0;
1904 else
1905 tp->linger2 = val * HZ;
1906 break;
1907
1908 case TCP_DEFER_ACCEPT:
1909 icsk->icsk_accept_queue.rskq_defer_accept = 0;
1910 if (val > 0) {
1911 /* Translate value in seconds to number of
1912 * retransmits */
1913 while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
1914 val > ((TCP_TIMEOUT_INIT / HZ) <<
1915 icsk->icsk_accept_queue.rskq_defer_accept))
1916 icsk->icsk_accept_queue.rskq_defer_accept++;
1917 icsk->icsk_accept_queue.rskq_defer_accept++;
1918 }
1919 break;
1920
1921 case TCP_WINDOW_CLAMP:
1922 if (!val) {
1923 if (sk->sk_state != TCP_CLOSE) {
1924 err = -EINVAL;
1925 break;
1926 }
1927 tp->window_clamp = 0;
1928 } else
1929 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1930 SOCK_MIN_RCVBUF / 2 : val;
1931 break;
1932
1933 case TCP_QUICKACK:
1934 if (!val) {
1935 icsk->icsk_ack.pingpong = 1;
1936 } else {
1937 icsk->icsk_ack.pingpong = 0;
1938 if ((1 << sk->sk_state) &
1939 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
1940 inet_csk_ack_scheduled(sk)) {
1941 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
1942 tcp_cleanup_rbuf(sk, 1);
1943 if (!(val & 1))
1944 icsk->icsk_ack.pingpong = 1;
1945 }
1946 }
1947 break;
1948
1949 #ifdef CONFIG_TCP_MD5SIG
1950 case TCP_MD5SIG:
1951 /* Read the IP->Key mappings from userspace */
1952 err = tp->af_specific->md5_parse(sk, optval, optlen);
1953 break;
1954 #endif
1955
1956 default:
1957 err = -ENOPROTOOPT;
1958 break;
1959 }
1960
1961 release_sock(sk);
1962 return err;
1963 }
1964
1965 int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1966 int optlen)
1967 {
1968 struct inet_connection_sock *icsk = inet_csk(sk);
1969
1970 if (level != SOL_TCP)
1971 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1972 optval, optlen);
1973 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1974 }
1975
1976 #ifdef CONFIG_COMPAT
1977 int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
1978 char __user *optval, int optlen)
1979 {
1980 if (level != SOL_TCP)
1981 return inet_csk_compat_setsockopt(sk, level, optname,
1982 optval, optlen);
1983 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
1984 }
1985
1986 EXPORT_SYMBOL(compat_tcp_setsockopt);
1987 #endif
1988
1989 /* Return information about state of tcp endpoint in API format. */
1990 void tcp_get_info(struct sock *sk, struct tcp_info *info)
1991 {
1992 struct tcp_sock *tp = tcp_sk(sk);
1993 const struct inet_connection_sock *icsk = inet_csk(sk);
1994 u32 now = tcp_time_stamp;
1995
1996 memset(info, 0, sizeof(*info));
1997
1998 info->tcpi_state = sk->sk_state;
1999 info->tcpi_ca_state = icsk->icsk_ca_state;
2000 info->tcpi_retransmits = icsk->icsk_retransmits;
2001 info->tcpi_probes = icsk->icsk_probes_out;
2002 info->tcpi_backoff = icsk->icsk_backoff;
2003
2004 if (tp->rx_opt.tstamp_ok)
2005 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2006 if (tp->rx_opt.sack_ok)
2007 info->tcpi_options |= TCPI_OPT_SACK;
2008 if (tp->rx_opt.wscale_ok) {
2009 info->tcpi_options |= TCPI_OPT_WSCALE;
2010 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2011 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2012 }
2013
2014 if (tp->ecn_flags&TCP_ECN_OK)
2015 info->tcpi_options |= TCPI_OPT_ECN;
2016
2017 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2018 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2019 info->tcpi_snd_mss = tp->mss_cache;
2020 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2021
2022 info->tcpi_unacked = tp->packets_out;
2023 info->tcpi_sacked = tp->sacked_out;
2024 info->tcpi_lost = tp->lost_out;
2025 info->tcpi_retrans = tp->retrans_out;
2026 info->tcpi_fackets = tp->fackets_out;
2027
2028 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2029 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2030 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2031
2032 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2033 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2034 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
2035 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
2036 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2037 info->tcpi_snd_cwnd = tp->snd_cwnd;
2038 info->tcpi_advmss = tp->advmss;
2039 info->tcpi_reordering = tp->reordering;
2040
2041 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2042 info->tcpi_rcv_space = tp->rcvq_space.space;
2043
2044 info->tcpi_total_retrans = tp->total_retrans;
2045 }
2046
2047 EXPORT_SYMBOL_GPL(tcp_get_info);
2048
2049 static int do_tcp_getsockopt(struct sock *sk, int level,
2050 int optname, char __user *optval, int __user *optlen)
2051 {
2052 struct inet_connection_sock *icsk = inet_csk(sk);
2053 struct tcp_sock *tp = tcp_sk(sk);
2054 int val, len;
2055
2056 if (get_user(len, optlen))
2057 return -EFAULT;
2058
2059 len = min_t(unsigned int, len, sizeof(int));
2060
2061 if (len < 0)
2062 return -EINVAL;
2063
2064 switch (optname) {
2065 case TCP_MAXSEG:
2066 val = tp->mss_cache;
2067 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2068 val = tp->rx_opt.user_mss;
2069 break;
2070 case TCP_NODELAY:
2071 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2072 break;
2073 case TCP_CORK:
2074 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2075 break;
2076 case TCP_KEEPIDLE:
2077 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
2078 break;
2079 case TCP_KEEPINTVL:
2080 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
2081 break;
2082 case TCP_KEEPCNT:
2083 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
2084 break;
2085 case TCP_SYNCNT:
2086 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
2087 break;
2088 case TCP_LINGER2:
2089 val = tp->linger2;
2090 if (val >= 0)
2091 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
2092 break;
2093 case TCP_DEFER_ACCEPT:
2094 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
2095 ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
2096 break;
2097 case TCP_WINDOW_CLAMP:
2098 val = tp->window_clamp;
2099 break;
2100 case TCP_INFO: {
2101 struct tcp_info info;
2102
2103 if (get_user(len, optlen))
2104 return -EFAULT;
2105
2106 tcp_get_info(sk, &info);
2107
2108 len = min_t(unsigned int, len, sizeof(info));
2109 if (put_user(len, optlen))
2110 return -EFAULT;
2111 if (copy_to_user(optval, &info, len))
2112 return -EFAULT;
2113 return 0;
2114 }
2115 case TCP_QUICKACK:
2116 val = !icsk->icsk_ack.pingpong;
2117 break;
2118
2119 case TCP_CONGESTION:
2120 if (get_user(len, optlen))
2121 return -EFAULT;
2122 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2123 if (put_user(len, optlen))
2124 return -EFAULT;
2125 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2126 return -EFAULT;
2127 return 0;
2128 default:
2129 return -ENOPROTOOPT;
2130 }
2131
2132 if (put_user(len, optlen))
2133 return -EFAULT;
2134 if (copy_to_user(optval, &val, len))
2135 return -EFAULT;
2136 return 0;
2137 }
2138
2139 int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
2140 int __user *optlen)
2141 {
2142 struct inet_connection_sock *icsk = inet_csk(sk);
2143
2144 if (level != SOL_TCP)
2145 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
2146 optval, optlen);
2147 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2148 }
2149
2150 #ifdef CONFIG_COMPAT
2151 int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
2152 char __user *optval, int __user *optlen)
2153 {
2154 if (level != SOL_TCP)
2155 return inet_csk_compat_getsockopt(sk, level, optname,
2156 optval, optlen);
2157 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
2158 }
2159
2160 EXPORT_SYMBOL(compat_tcp_getsockopt);
2161 #endif
2162
2163 struct sk_buff *tcp_tso_segment(struct sk_buff *skb, int features)
2164 {
2165 struct sk_buff *segs = ERR_PTR(-EINVAL);
2166 struct tcphdr *th;
2167 unsigned thlen;
2168 unsigned int seq;
2169 __be32 delta;
2170 unsigned int oldlen;
2171 unsigned int len;
2172
2173 if (!pskb_may_pull(skb, sizeof(*th)))
2174 goto out;
2175
2176 th = tcp_hdr(skb);
2177 thlen = th->doff * 4;
2178 if (thlen < sizeof(*th))
2179 goto out;
2180
2181 if (!pskb_may_pull(skb, thlen))
2182 goto out;
2183
2184 oldlen = (u16)~skb->len;
2185 __skb_pull(skb, thlen);
2186
2187 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
2188 /* Packet is from an untrusted source, reset gso_segs. */
2189 int type = skb_shinfo(skb)->gso_type;
2190 int mss;
2191
2192 if (unlikely(type &
2193 ~(SKB_GSO_TCPV4 |
2194 SKB_GSO_DODGY |
2195 SKB_GSO_TCP_ECN |
2196 SKB_GSO_TCPV6 |
2197 0) ||
2198 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
2199 goto out;
2200
2201 mss = skb_shinfo(skb)->gso_size;
2202 skb_shinfo(skb)->gso_segs = (skb->len + mss - 1) / mss;
2203
2204 segs = NULL;
2205 goto out;
2206 }
2207
2208 segs = skb_segment(skb, features);
2209 if (IS_ERR(segs))
2210 goto out;
2211
2212 len = skb_shinfo(skb)->gso_size;
2213 delta = htonl(oldlen + (thlen + len));
2214
2215 skb = segs;
2216 th = tcp_hdr(skb);
2217 seq = ntohl(th->seq);
2218
2219 do {
2220 th->fin = th->psh = 0;
2221
2222 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2223 (__force u32)delta));
2224 if (skb->ip_summed != CHECKSUM_PARTIAL)
2225 th->check =
2226 csum_fold(csum_partial(skb_transport_header(skb),
2227 thlen, skb->csum));
2228
2229 seq += len;
2230 skb = skb->next;
2231 th = tcp_hdr(skb);
2232
2233 th->seq = htonl(seq);
2234 th->cwr = 0;
2235 } while (skb->next);
2236
2237 delta = htonl(oldlen + (skb->tail - skb->transport_header) +
2238 skb->data_len);
2239 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
2240 (__force u32)delta));
2241 if (skb->ip_summed != CHECKSUM_PARTIAL)
2242 th->check = csum_fold(csum_partial(skb_transport_header(skb),
2243 thlen, skb->csum));
2244
2245 out:
2246 return segs;
2247 }
2248 EXPORT_SYMBOL(tcp_tso_segment);
2249
2250 #ifdef CONFIG_TCP_MD5SIG
2251 static unsigned long tcp_md5sig_users;
2252 static struct tcp_md5sig_pool **tcp_md5sig_pool;
2253 static DEFINE_SPINLOCK(tcp_md5sig_pool_lock);
2254
2255 static void __tcp_free_md5sig_pool(struct tcp_md5sig_pool **pool)
2256 {
2257 int cpu;
2258 for_each_possible_cpu(cpu) {
2259 struct tcp_md5sig_pool *p = *per_cpu_ptr(pool, cpu);
2260 if (p) {
2261 if (p->md5_desc.tfm)
2262 crypto_free_hash(p->md5_desc.tfm);
2263 kfree(p);
2264 p = NULL;
2265 }
2266 }
2267 free_percpu(pool);
2268 }
2269
2270 void tcp_free_md5sig_pool(void)
2271 {
2272 struct tcp_md5sig_pool **pool = NULL;
2273
2274 spin_lock_bh(&tcp_md5sig_pool_lock);
2275 if (--tcp_md5sig_users == 0) {
2276 pool = tcp_md5sig_pool;
2277 tcp_md5sig_pool = NULL;
2278 }
2279 spin_unlock_bh(&tcp_md5sig_pool_lock);
2280 if (pool)
2281 __tcp_free_md5sig_pool(pool);
2282 }
2283
2284 EXPORT_SYMBOL(tcp_free_md5sig_pool);
2285
2286 static struct tcp_md5sig_pool **__tcp_alloc_md5sig_pool(void)
2287 {
2288 int cpu;
2289 struct tcp_md5sig_pool **pool;
2290
2291 pool = alloc_percpu(struct tcp_md5sig_pool *);
2292 if (!pool)
2293 return NULL;
2294
2295 for_each_possible_cpu(cpu) {
2296 struct tcp_md5sig_pool *p;
2297 struct crypto_hash *hash;
2298
2299 p = kzalloc(sizeof(*p), GFP_KERNEL);
2300 if (!p)
2301 goto out_free;
2302 *per_cpu_ptr(pool, cpu) = p;
2303
2304 hash = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC);
2305 if (!hash || IS_ERR(hash))
2306 goto out_free;
2307
2308 p->md5_desc.tfm = hash;
2309 }
2310 return pool;
2311 out_free:
2312 __tcp_free_md5sig_pool(pool);
2313 return NULL;
2314 }
2315
2316 struct tcp_md5sig_pool **tcp_alloc_md5sig_pool(void)
2317 {
2318 struct tcp_md5sig_pool **pool;
2319 int alloc = 0;
2320
2321 retry:
2322 spin_lock_bh(&tcp_md5sig_pool_lock);
2323 pool = tcp_md5sig_pool;
2324 if (tcp_md5sig_users++ == 0) {
2325 alloc = 1;
2326 spin_unlock_bh(&tcp_md5sig_pool_lock);
2327 } else if (!pool) {
2328 tcp_md5sig_users--;
2329 spin_unlock_bh(&tcp_md5sig_pool_lock);
2330 cpu_relax();
2331 goto retry;
2332 } else
2333 spin_unlock_bh(&tcp_md5sig_pool_lock);
2334
2335 if (alloc) {
2336 /* we cannot hold spinlock here because this may sleep. */
2337 struct tcp_md5sig_pool **p = __tcp_alloc_md5sig_pool();
2338 spin_lock_bh(&tcp_md5sig_pool_lock);
2339 if (!p) {
2340 tcp_md5sig_users--;
2341 spin_unlock_bh(&tcp_md5sig_pool_lock);
2342 return NULL;
2343 }
2344 pool = tcp_md5sig_pool;
2345 if (pool) {
2346 /* oops, it has already been assigned. */
2347 spin_unlock_bh(&tcp_md5sig_pool_lock);
2348 __tcp_free_md5sig_pool(p);
2349 } else {
2350 tcp_md5sig_pool = pool = p;
2351 spin_unlock_bh(&tcp_md5sig_pool_lock);
2352 }
2353 }
2354 return pool;
2355 }
2356
2357 EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
2358
2359 struct tcp_md5sig_pool *__tcp_get_md5sig_pool(int cpu)
2360 {
2361 struct tcp_md5sig_pool **p;
2362 spin_lock_bh(&tcp_md5sig_pool_lock);
2363 p = tcp_md5sig_pool;
2364 if (p)
2365 tcp_md5sig_users++;
2366 spin_unlock_bh(&tcp_md5sig_pool_lock);
2367 return (p ? *per_cpu_ptr(p, cpu) : NULL);
2368 }
2369
2370 EXPORT_SYMBOL(__tcp_get_md5sig_pool);
2371
2372 void __tcp_put_md5sig_pool(void)
2373 {
2374 tcp_free_md5sig_pool();
2375 }
2376
2377 EXPORT_SYMBOL(__tcp_put_md5sig_pool);
2378 #endif
2379
2380 void tcp_done(struct sock *sk)
2381 {
2382 if(sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
2383 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
2384
2385 tcp_set_state(sk, TCP_CLOSE);
2386 tcp_clear_xmit_timers(sk);
2387
2388 sk->sk_shutdown = SHUTDOWN_MASK;
2389
2390 if (!sock_flag(sk, SOCK_DEAD))
2391 sk->sk_state_change(sk);
2392 else
2393 inet_csk_destroy_sock(sk);
2394 }
2395 EXPORT_SYMBOL_GPL(tcp_done);
2396
2397 extern void __skb_cb_too_small_for_tcp(int, int);
2398 extern struct tcp_congestion_ops tcp_reno;
2399
2400 static __initdata unsigned long thash_entries;
2401 static int __init set_thash_entries(char *str)
2402 {
2403 if (!str)
2404 return 0;
2405 thash_entries = simple_strtoul(str, &str, 0);
2406 return 1;
2407 }
2408 __setup("thash_entries=", set_thash_entries);
2409
2410 void __init tcp_init(void)
2411 {
2412 struct sk_buff *skb = NULL;
2413 unsigned long limit;
2414 int order, i, max_share;
2415
2416 if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2417 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2418 sizeof(skb->cb));
2419
2420 tcp_hashinfo.bind_bucket_cachep =
2421 kmem_cache_create("tcp_bind_bucket",
2422 sizeof(struct inet_bind_bucket), 0,
2423 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
2424
2425 /* Size and allocate the main established and bind bucket
2426 * hash tables.
2427 *
2428 * The methodology is similar to that of the buffer cache.
2429 */
2430 tcp_hashinfo.ehash =
2431 alloc_large_system_hash("TCP established",
2432 sizeof(struct inet_ehash_bucket),
2433 thash_entries,
2434 (num_physpages >= 128 * 1024) ?
2435 13 : 15,
2436 0,
2437 &tcp_hashinfo.ehash_size,
2438 NULL,
2439 0);
2440 tcp_hashinfo.ehash_size = 1 << tcp_hashinfo.ehash_size;
2441 for (i = 0; i < tcp_hashinfo.ehash_size; i++) {
2442 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2443 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
2444 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].twchain);
2445 }
2446
2447 tcp_hashinfo.bhash =
2448 alloc_large_system_hash("TCP bind",
2449 sizeof(struct inet_bind_hashbucket),
2450 tcp_hashinfo.ehash_size,
2451 (num_physpages >= 128 * 1024) ?
2452 13 : 15,
2453 0,
2454 &tcp_hashinfo.bhash_size,
2455 NULL,
2456 64 * 1024);
2457 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2458 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2459 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2460 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
2461 }
2462
2463 /* Try to be a bit smarter and adjust defaults depending
2464 * on available memory.
2465 */
2466 for (order = 0; ((1 << order) << PAGE_SHIFT) <
2467 (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
2468 order++)
2469 ;
2470 if (order >= 4) {
2471 tcp_death_row.sysctl_max_tw_buckets = 180000;
2472 sysctl_tcp_max_orphans = 4096 << (order - 4);
2473 sysctl_max_syn_backlog = 1024;
2474 } else if (order < 3) {
2475 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
2476 sysctl_tcp_max_orphans >>= (3 - order);
2477 sysctl_max_syn_backlog = 128;
2478 }
2479
2480 /* Set the pressure threshold to be a fraction of global memory that
2481 * is up to 1/2 at 256 MB, decreasing toward zero with the amount of
2482 * memory, with a floor of 128 pages.
2483 */
2484 limit = min(nr_all_pages, 1UL<<(28-PAGE_SHIFT)) >> (20-PAGE_SHIFT);
2485 limit = (limit * (nr_all_pages >> (20-PAGE_SHIFT))) >> (PAGE_SHIFT-11);
2486 limit = max(limit, 128UL);
2487 sysctl_tcp_mem[0] = limit / 4 * 3;
2488 sysctl_tcp_mem[1] = limit;
2489 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
2490
2491 /* Set per-socket limits to no more than 1/128 the pressure threshold */
2492 limit = ((unsigned long)sysctl_tcp_mem[1]) << (PAGE_SHIFT - 7);
2493 max_share = min(4UL*1024*1024, limit);
2494
2495 sysctl_tcp_wmem[0] = SK_STREAM_MEM_QUANTUM;
2496 sysctl_tcp_wmem[1] = 16*1024;
2497 sysctl_tcp_wmem[2] = max(64*1024, max_share);
2498
2499 sysctl_tcp_rmem[0] = SK_STREAM_MEM_QUANTUM;
2500 sysctl_tcp_rmem[1] = 87380;
2501 sysctl_tcp_rmem[2] = max(87380, max_share);
2502
2503 printk(KERN_INFO "TCP: Hash tables configured "
2504 "(established %d bind %d)\n",
2505 tcp_hashinfo.ehash_size, tcp_hashinfo.bhash_size);
2506
2507 tcp_register_congestion_control(&tcp_reno);
2508 }
2509
2510 EXPORT_SYMBOL(tcp_close);
2511 EXPORT_SYMBOL(tcp_disconnect);
2512 EXPORT_SYMBOL(tcp_getsockopt);
2513 EXPORT_SYMBOL(tcp_ioctl);
2514 EXPORT_SYMBOL(tcp_poll);
2515 EXPORT_SYMBOL(tcp_read_sock);
2516 EXPORT_SYMBOL(tcp_recvmsg);
2517 EXPORT_SYMBOL(tcp_sendmsg);
2518 EXPORT_SYMBOL(tcp_sendpage);
2519 EXPORT_SYMBOL(tcp_setsockopt);
2520 EXPORT_SYMBOL(tcp_shutdown);
2521 EXPORT_SYMBOL(tcp_statistics);
x86 "subsystem" repository