blob: 02fdda68718d0c98d60d1d8de42fc87689e9991c [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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 *
Jesper Juhl02c30a82005-05-05 16:16:16 -070010 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * 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/config.h>
251#include <linux/module.h>
252#include <linux/types.h>
253#include <linux/fcntl.h>
254#include <linux/poll.h>
255#include <linux/init.h>
256#include <linux/smp_lock.h>
257#include <linux/fs.h>
258#include <linux/random.h>
259#include <linux/bootmem.h>
260
261#include <net/icmp.h>
262#include <net/tcp.h>
263#include <net/xfrm.h>
264#include <net/ip.h>
265
266
267#include <asm/uaccess.h>
268#include <asm/ioctls.h>
269
270int sysctl_tcp_fin_timeout = TCP_FIN_TIMEOUT;
271
Eric Dumazetba899662005-08-26 12:05:31 -0700272DEFINE_SNMP_STAT(struct tcp_mib, tcp_statistics) __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274atomic_t tcp_orphan_count = ATOMIC_INIT(0);
275
Arnaldo Carvalho de Melo0a5578c2005-08-09 20:11:41 -0700276EXPORT_SYMBOL_GPL(tcp_orphan_count);
277
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278int sysctl_tcp_mem[3];
279int sysctl_tcp_wmem[3] = { 4 * 1024, 16 * 1024, 128 * 1024 };
280int sysctl_tcp_rmem[3] = { 4 * 1024, 87380, 87380 * 2 };
281
282EXPORT_SYMBOL(sysctl_tcp_mem);
283EXPORT_SYMBOL(sysctl_tcp_rmem);
284EXPORT_SYMBOL(sysctl_tcp_wmem);
285
286atomic_t tcp_memory_allocated; /* Current allocated memory. */
287atomic_t tcp_sockets_allocated; /* Current number of TCP sockets. */
288
289EXPORT_SYMBOL(tcp_memory_allocated);
290EXPORT_SYMBOL(tcp_sockets_allocated);
291
292/*
293 * Pressure flag: try to collapse.
294 * Technical note: it is used by multiple contexts non atomically.
295 * All the sk_stream_mem_schedule() is of this nature: accounting
296 * is strict, actions are advisory and have some latency.
297 */
298int tcp_memory_pressure;
299
300EXPORT_SYMBOL(tcp_memory_pressure);
301
302void tcp_enter_memory_pressure(void)
303{
304 if (!tcp_memory_pressure) {
305 NET_INC_STATS(LINUX_MIB_TCPMEMORYPRESSURES);
306 tcp_memory_pressure = 1;
307 }
308}
309
310EXPORT_SYMBOL(tcp_enter_memory_pressure);
311
312/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313 * Wait for a TCP event.
314 *
315 * Note that we don't need to lock the socket, as the upper poll layers
316 * take care of normal races (between the test and the event) and we don't
317 * go look at any of the socket buffers directly.
318 */
319unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
320{
321 unsigned int mask;
322 struct sock *sk = sock->sk;
323 struct tcp_sock *tp = tcp_sk(sk);
324
325 poll_wait(file, sk->sk_sleep, wait);
326 if (sk->sk_state == TCP_LISTEN)
Arnaldo Carvalho de Melodc40c7b2005-08-23 21:52:58 -0700327 return inet_csk_listen_poll(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328
329 /* Socket is not locked. We are protected from async events
330 by poll logic and correct handling of state changes
331 made by another threads is impossible in any case.
332 */
333
334 mask = 0;
335 if (sk->sk_err)
336 mask = POLLERR;
337
338 /*
339 * POLLHUP is certainly not done right. But poll() doesn't
340 * have a notion of HUP in just one direction, and for a
341 * socket the read side is more interesting.
342 *
343 * Some poll() documentation says that POLLHUP is incompatible
344 * with the POLLOUT/POLLWR flags, so somebody should check this
345 * all. But careful, it tends to be safer to return too many
346 * bits than too few, and you can easily break real applications
347 * if you don't tell them that something has hung up!
348 *
349 * Check-me.
350 *
351 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
352 * our fs/select.c). It means that after we received EOF,
353 * poll always returns immediately, making impossible poll() on write()
354 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
355 * if and only if shutdown has been made in both directions.
356 * Actually, it is interesting to look how Solaris and DUX
357 * solve this dilemma. I would prefer, if PULLHUP were maskable,
358 * then we could set it on SND_SHUTDOWN. BTW examples given
359 * in Stevens' books assume exactly this behaviour, it explains
360 * why PULLHUP is incompatible with POLLOUT. --ANK
361 *
362 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
363 * blocking on fresh not-connected or disconnected socket. --ANK
364 */
365 if (sk->sk_shutdown == SHUTDOWN_MASK || sk->sk_state == TCP_CLOSE)
366 mask |= POLLHUP;
367 if (sk->sk_shutdown & RCV_SHUTDOWN)
368 mask |= POLLIN | POLLRDNORM;
369
370 /* Connected? */
371 if ((1 << sk->sk_state) & ~(TCPF_SYN_SENT | TCPF_SYN_RECV)) {
372 /* Potential race condition. If read of tp below will
373 * escape above sk->sk_state, we can be illegally awaken
374 * in SYN_* states. */
375 if ((tp->rcv_nxt != tp->copied_seq) &&
376 (tp->urg_seq != tp->copied_seq ||
377 tp->rcv_nxt != tp->copied_seq + 1 ||
378 sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data))
379 mask |= POLLIN | POLLRDNORM;
380
381 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
382 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
383 mask |= POLLOUT | POLLWRNORM;
384 } else { /* send SIGIO later */
385 set_bit(SOCK_ASYNC_NOSPACE,
386 &sk->sk_socket->flags);
387 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
388
389 /* Race breaker. If space is freed after
390 * wspace test but before the flags are set,
391 * IO signal will be lost.
392 */
393 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk))
394 mask |= POLLOUT | POLLWRNORM;
395 }
396 }
397
398 if (tp->urg_data & TCP_URG_VALID)
399 mask |= POLLPRI;
400 }
401 return mask;
402}
403
404int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
405{
406 struct tcp_sock *tp = tcp_sk(sk);
407 int answ;
408
409 switch (cmd) {
410 case SIOCINQ:
411 if (sk->sk_state == TCP_LISTEN)
412 return -EINVAL;
413
414 lock_sock(sk);
415 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
416 answ = 0;
417 else if (sock_flag(sk, SOCK_URGINLINE) ||
418 !tp->urg_data ||
419 before(tp->urg_seq, tp->copied_seq) ||
420 !before(tp->urg_seq, tp->rcv_nxt)) {
421 answ = tp->rcv_nxt - tp->copied_seq;
422
423 /* Subtract 1, if FIN is in queue. */
424 if (answ && !skb_queue_empty(&sk->sk_receive_queue))
425 answ -=
426 ((struct sk_buff *)sk->sk_receive_queue.prev)->h.th->fin;
427 } else
428 answ = tp->urg_seq - tp->copied_seq;
429 release_sock(sk);
430 break;
431 case SIOCATMARK:
432 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
433 break;
434 case SIOCOUTQ:
435 if (sk->sk_state == TCP_LISTEN)
436 return -EINVAL;
437
438 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
439 answ = 0;
440 else
441 answ = tp->write_seq - tp->snd_una;
442 break;
443 default:
444 return -ENOIOCTLCMD;
445 };
446
447 return put_user(answ, (int __user *)arg);
448}
449
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
451{
452 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
453 tp->pushed_seq = tp->write_seq;
454}
455
456static inline int forced_push(struct tcp_sock *tp)
457{
458 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
459}
460
461static inline void skb_entail(struct sock *sk, struct tcp_sock *tp,
462 struct sk_buff *skb)
463{
464 skb->csum = 0;
465 TCP_SKB_CB(skb)->seq = tp->write_seq;
466 TCP_SKB_CB(skb)->end_seq = tp->write_seq;
467 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
468 TCP_SKB_CB(skb)->sacked = 0;
469 skb_header_release(skb);
470 __skb_queue_tail(&sk->sk_write_queue, skb);
471 sk_charge_skb(sk, skb);
472 if (!sk->sk_send_head)
473 sk->sk_send_head = skb;
David S. Miller89ebd192005-08-23 10:13:06 -0700474 if (tp->nonagle & TCP_NAGLE_PUSH)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475 tp->nonagle &= ~TCP_NAGLE_PUSH;
476}
477
478static 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
488static inline void tcp_push(struct sock *sk, struct tcp_sock *tp, int flags,
489 int mss_now, int nonagle)
490{
491 if (sk->sk_send_head) {
492 struct sk_buff *skb = sk->sk_write_queue.prev;
493 if (!(flags & MSG_MORE) || forced_push(tp))
494 tcp_mark_push(tp, skb);
495 tcp_mark_urg(tp, flags, skb);
496 __tcp_push_pending_frames(sk, tp, mss_now,
497 (flags & MSG_MORE) ? TCP_NAGLE_CORK : nonagle);
498 }
499}
500
501static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset,
502 size_t psize, int flags)
503{
504 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700505 int mss_now, size_goal;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 int err;
507 ssize_t copied;
508 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
509
510 /* Wait for a connection to finish. */
511 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
512 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
513 goto out_err;
514
515 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
516
517 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700518 size_goal = tp->xmit_size_goal;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 copied = 0;
520
521 err = -EPIPE;
522 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
523 goto do_error;
524
525 while (psize > 0) {
526 struct sk_buff *skb = sk->sk_write_queue.prev;
527 struct page *page = pages[poffset / PAGE_SIZE];
528 int copy, i, can_coalesce;
529 int offset = poffset % PAGE_SIZE;
530 int size = min_t(size_t, psize, PAGE_SIZE - offset);
531
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700532 if (!sk->sk_send_head || (copy = size_goal - skb->len) <= 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533new_segment:
534 if (!sk_stream_memory_free(sk))
535 goto wait_for_sndbuf;
536
537 skb = sk_stream_alloc_pskb(sk, 0, 0,
538 sk->sk_allocation);
539 if (!skb)
540 goto wait_for_memory;
541
542 skb_entail(sk, tp, skb);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700543 copy = size_goal;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544 }
545
546 if (copy > size)
547 copy = size;
548
549 i = skb_shinfo(skb)->nr_frags;
550 can_coalesce = skb_can_coalesce(skb, i, page, offset);
551 if (!can_coalesce && i >= MAX_SKB_FRAGS) {
552 tcp_mark_push(tp, skb);
553 goto new_segment;
554 }
555 if (sk->sk_forward_alloc < copy &&
556 !sk_stream_mem_schedule(sk, copy, 0))
557 goto wait_for_memory;
558
559 if (can_coalesce) {
560 skb_shinfo(skb)->frags[i - 1].size += copy;
561 } else {
562 get_page(page);
563 skb_fill_page_desc(skb, i, page, offset, copy);
564 }
565
566 skb->len += copy;
567 skb->data_len += copy;
568 skb->truesize += copy;
569 sk->sk_wmem_queued += copy;
570 sk->sk_forward_alloc -= copy;
571 skb->ip_summed = CHECKSUM_HW;
572 tp->write_seq += copy;
573 TCP_SKB_CB(skb)->end_seq += copy;
574 skb_shinfo(skb)->tso_segs = 0;
575
576 if (!copied)
577 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_PSH;
578
579 copied += copy;
580 poffset += copy;
581 if (!(psize -= copy))
582 goto out;
583
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700584 if (skb->len < mss_now || (flags & MSG_OOB))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 continue;
586
587 if (forced_push(tp)) {
588 tcp_mark_push(tp, skb);
589 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
590 } else if (skb == sk->sk_send_head)
591 tcp_push_one(sk, mss_now);
592 continue;
593
594wait_for_sndbuf:
595 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
596wait_for_memory:
597 if (copied)
598 tcp_push(sk, tp, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
599
600 if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
601 goto do_error;
602
603 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700604 size_goal = tp->xmit_size_goal;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605 }
606
607out:
608 if (copied)
609 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
610 return copied;
611
612do_error:
613 if (copied)
614 goto out;
615out_err:
616 return sk_stream_error(sk, flags, err);
617}
618
619ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
620 size_t size, int flags)
621{
622 ssize_t res;
623 struct sock *sk = sock->sk;
624
625#define TCP_ZC_CSUM_FLAGS (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM)
626
627 if (!(sk->sk_route_caps & NETIF_F_SG) ||
628 !(sk->sk_route_caps & TCP_ZC_CSUM_FLAGS))
629 return sock_no_sendpage(sock, page, offset, size, flags);
630
631#undef TCP_ZC_CSUM_FLAGS
632
633 lock_sock(sk);
634 TCP_CHECK_TIMER(sk);
635 res = do_tcp_sendpages(sk, &page, offset, size, flags);
636 TCP_CHECK_TIMER(sk);
637 release_sock(sk);
638 return res;
639}
640
641#define TCP_PAGE(sk) (sk->sk_sndmsg_page)
642#define TCP_OFF(sk) (sk->sk_sndmsg_off)
643
644static inline int select_size(struct sock *sk, struct tcp_sock *tp)
645{
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700646 int tmp = tp->mss_cache;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647
David S. Millerb4e26f52005-07-05 15:20:27 -0700648 if (sk->sk_route_caps & NETIF_F_SG) {
649 if (sk->sk_route_caps & NETIF_F_TSO)
650 tmp = 0;
651 else {
652 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
653
654 if (tmp >= pgbreak &&
655 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
656 tmp = pgbreak;
657 }
658 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 return tmp;
661}
662
663int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
664 size_t size)
665{
666 struct iovec *iov;
667 struct tcp_sock *tp = tcp_sk(sk);
668 struct sk_buff *skb;
669 int iovlen, flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700670 int mss_now, size_goal;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 int err, copied;
672 long timeo;
673
674 lock_sock(sk);
675 TCP_CHECK_TIMER(sk);
676
677 flags = msg->msg_flags;
678 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
679
680 /* Wait for a connection to finish. */
681 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
682 if ((err = sk_stream_wait_connect(sk, &timeo)) != 0)
683 goto out_err;
684
685 /* This should be in poll */
686 clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
687
688 mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700689 size_goal = tp->xmit_size_goal;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690
691 /* Ok commence sending. */
692 iovlen = msg->msg_iovlen;
693 iov = msg->msg_iov;
694 copied = 0;
695
696 err = -EPIPE;
697 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
698 goto do_error;
699
700 while (--iovlen >= 0) {
701 int seglen = iov->iov_len;
702 unsigned char __user *from = iov->iov_base;
703
704 iov++;
705
706 while (seglen > 0) {
707 int copy;
708
709 skb = sk->sk_write_queue.prev;
710
711 if (!sk->sk_send_head ||
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700712 (copy = size_goal - skb->len) <= 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713
714new_segment:
715 /* Allocate new segment. If the interface is SG,
716 * allocate skb fitting to single page.
717 */
718 if (!sk_stream_memory_free(sk))
719 goto wait_for_sndbuf;
720
721 skb = sk_stream_alloc_pskb(sk, select_size(sk, tp),
722 0, sk->sk_allocation);
723 if (!skb)
724 goto wait_for_memory;
725
726 /*
727 * Check whether we can use HW checksum.
728 */
729 if (sk->sk_route_caps &
730 (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM |
731 NETIF_F_HW_CSUM))
732 skb->ip_summed = CHECKSUM_HW;
733
734 skb_entail(sk, tp, skb);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700735 copy = size_goal;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 }
737
738 /* Try to append data to the end of skb. */
739 if (copy > seglen)
740 copy = seglen;
741
742 /* Where to copy to? */
743 if (skb_tailroom(skb) > 0) {
744 /* We have some space in skb head. Superb! */
745 if (copy > skb_tailroom(skb))
746 copy = skb_tailroom(skb);
747 if ((err = skb_add_data(skb, from, copy)) != 0)
748 goto do_fault;
749 } else {
750 int merge = 0;
751 int i = skb_shinfo(skb)->nr_frags;
752 struct page *page = TCP_PAGE(sk);
753 int off = TCP_OFF(sk);
754
755 if (skb_can_coalesce(skb, i, page, off) &&
756 off != PAGE_SIZE) {
757 /* We can extend the last page
758 * fragment. */
759 merge = 1;
760 } else if (i == MAX_SKB_FRAGS ||
761 (!i &&
762 !(sk->sk_route_caps & NETIF_F_SG))) {
763 /* Need to add new fragment and cannot
764 * do this because interface is non-SG,
765 * or because all the page slots are
766 * busy. */
767 tcp_mark_push(tp, skb);
768 goto new_segment;
769 } else if (page) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 if (off == PAGE_SIZE) {
771 put_page(page);
772 TCP_PAGE(sk) = page = NULL;
773 }
774 }
775
776 if (!page) {
777 /* Allocate new cache page. */
778 if (!(page = sk_stream_alloc_page(sk)))
779 goto wait_for_memory;
780 off = 0;
781 }
782
783 if (copy > PAGE_SIZE - off)
784 copy = PAGE_SIZE - off;
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)->tso_segs = 0;
824
825 from += copy;
826 copied += copy;
827 if ((seglen -= copy) == 0 && iovlen == 0)
828 goto out;
829
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700830 if (skb->len < mss_now || (flags & MSG_OOB))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 continue;
832
833 if (forced_push(tp)) {
834 tcp_mark_push(tp, skb);
835 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_PUSH);
836 } else if (skb == sk->sk_send_head)
837 tcp_push_one(sk, mss_now);
838 continue;
839
840wait_for_sndbuf:
841 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
842wait_for_memory:
843 if (copied)
844 tcp_push(sk, tp, 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));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700850 size_goal = tp->xmit_size_goal;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 }
852 }
853
854out:
855 if (copied)
856 tcp_push(sk, tp, flags, mss_now, tp->nonagle);
857 TCP_CHECK_TIMER(sk);
858 release_sock(sk);
859 return copied;
860
861do_fault:
862 if (!skb->len) {
863 if (sk->sk_send_head == skb)
864 sk->sk_send_head = NULL;
David S. Miller8728b832005-08-09 19:25:21 -0700865 __skb_unlink(skb, &sk->sk_write_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866 sk_stream_free_skb(sk, skb);
867 }
868
869do_error:
870 if (copied)
871 goto out;
872out_err:
873 err = sk_stream_error(sk, flags, err);
874 TCP_CHECK_TIMER(sk);
875 release_sock(sk);
876 return err;
877}
878
879/*
880 * Handle reading urgent data. BSD has very simple semantics for
881 * this, no blocking and very strange errors 8)
882 */
883
884static int tcp_recv_urg(struct sock *sk, long timeo,
885 struct msghdr *msg, int len, int flags,
886 int *addr_len)
887{
888 struct tcp_sock *tp = tcp_sk(sk);
889
890 /* No URG data to read. */
891 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
892 tp->urg_data == TCP_URG_READ)
893 return -EINVAL; /* Yes this is right ! */
894
895 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
896 return -ENOTCONN;
897
898 if (tp->urg_data & TCP_URG_VALID) {
899 int err = 0;
900 char c = tp->urg_data;
901
902 if (!(flags & MSG_PEEK))
903 tp->urg_data = TCP_URG_READ;
904
905 /* Read urgent data. */
906 msg->msg_flags |= MSG_OOB;
907
908 if (len > 0) {
909 if (!(flags & MSG_TRUNC))
910 err = memcpy_toiovec(msg->msg_iov, &c, 1);
911 len = 1;
912 } else
913 msg->msg_flags |= MSG_TRUNC;
914
915 return err ? -EFAULT : len;
916 }
917
918 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
919 return 0;
920
921 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
922 * the available implementations agree in this case:
923 * this call should never block, independent of the
924 * blocking state of the socket.
925 * Mike <pall@rz.uni-karlsruhe.de>
926 */
927 return -EAGAIN;
928}
929
930/* Clean up the receive buffer for full frames taken by the user,
931 * then send an ACK if necessary. COPIED is the number of bytes
932 * tcp_recvmsg has given to the user so far, it speeds up the
933 * calculation of whether or not we must ACK for the sake of
934 * a window update.
935 */
936static void cleanup_rbuf(struct sock *sk, int copied)
937{
938 struct tcp_sock *tp = tcp_sk(sk);
939 int time_to_ack = 0;
940
941#if TCP_DEBUG
942 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
943
944 BUG_TRAP(!skb || before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq));
945#endif
946
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700947 if (inet_csk_ack_scheduled(sk)) {
948 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 /* Delayed ACKs frequently hit locked sockets during bulk
950 * receive. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700951 if (icsk->icsk_ack.blocked ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952 /* Once-per-two-segments ACK was not sent by tcp_input.c */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700953 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 /*
955 * If this read emptied read buffer, we send ACK, if
956 * connection is not bidirectional, user drained
957 * receive buffer and there was a small segment
958 * in queue.
959 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700960 (copied > 0 && (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
961 !icsk->icsk_ack.pingpong && !atomic_read(&sk->sk_rmem_alloc)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 time_to_ack = 1;
963 }
964
965 /* We send an ACK if we can now advertise a non-zero window
966 * which has been raised "significantly".
967 *
968 * Even if window raised up to infinity, do not send window open ACK
969 * in states, where we will not receive more. It is useless.
970 */
971 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
972 __u32 rcv_window_now = tcp_receive_window(tp);
973
974 /* Optimize, __tcp_select_window() is not cheap. */
975 if (2*rcv_window_now <= tp->window_clamp) {
976 __u32 new_window = __tcp_select_window(sk);
977
978 /* Send ACK now, if this read freed lots of space
979 * in our buffer. Certainly, new_window is new window.
980 * We can advertise it now, if it is not less than current one.
981 * "Lots" means "at least twice" here.
982 */
983 if (new_window && new_window >= 2 * rcv_window_now)
984 time_to_ack = 1;
985 }
986 }
987 if (time_to_ack)
988 tcp_send_ack(sk);
989}
990
991static void tcp_prequeue_process(struct sock *sk)
992{
993 struct sk_buff *skb;
994 struct tcp_sock *tp = tcp_sk(sk);
995
David S. Millerb03efcf2005-07-08 14:57:23 -0700996 NET_INC_STATS_USER(LINUX_MIB_TCPPREQUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997
998 /* RX process wants to run with disabled BHs, though it is not
999 * necessary */
1000 local_bh_disable();
1001 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1002 sk->sk_backlog_rcv(sk, skb);
1003 local_bh_enable();
1004
1005 /* Clear memory counter. */
1006 tp->ucopy.memory = 0;
1007}
1008
1009static inline struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1010{
1011 struct sk_buff *skb;
1012 u32 offset;
1013
1014 skb_queue_walk(&sk->sk_receive_queue, skb) {
1015 offset = seq - TCP_SKB_CB(skb)->seq;
1016 if (skb->h.th->syn)
1017 offset--;
1018 if (offset < skb->len || skb->h.th->fin) {
1019 *off = offset;
1020 return skb;
1021 }
1022 }
1023 return NULL;
1024}
1025
1026/*
1027 * This routine provides an alternative to tcp_recvmsg() for routines
1028 * that would like to handle copying from skbuffs directly in 'sendfile'
1029 * fashion.
1030 * Note:
1031 * - It is assumed that the socket was locked by the caller.
1032 * - The routine does not block.
1033 * - At present, there is no support for reading OOB data
1034 * or for 'peeking' the socket using this routine
1035 * (although both would be easy to implement).
1036 */
1037int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1038 sk_read_actor_t recv_actor)
1039{
1040 struct sk_buff *skb;
1041 struct tcp_sock *tp = tcp_sk(sk);
1042 u32 seq = tp->copied_seq;
1043 u32 offset;
1044 int copied = 0;
1045
1046 if (sk->sk_state == TCP_LISTEN)
1047 return -ENOTCONN;
1048 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1049 if (offset < skb->len) {
1050 size_t used, len;
1051
1052 len = skb->len - offset;
1053 /* Stop reading if we hit a patch of urgent data */
1054 if (tp->urg_data) {
1055 u32 urg_offset = tp->urg_seq - seq;
1056 if (urg_offset < len)
1057 len = urg_offset;
1058 if (!len)
1059 break;
1060 }
1061 used = recv_actor(desc, skb, offset, len);
1062 if (used <= len) {
1063 seq += used;
1064 copied += used;
1065 offset += used;
1066 }
1067 if (offset != skb->len)
1068 break;
1069 }
1070 if (skb->h.th->fin) {
1071 sk_eat_skb(sk, skb);
1072 ++seq;
1073 break;
1074 }
1075 sk_eat_skb(sk, skb);
1076 if (!desc->count)
1077 break;
1078 }
1079 tp->copied_seq = seq;
1080
1081 tcp_rcv_space_adjust(sk);
1082
1083 /* Clean up data we have read: This will do ACK frames. */
1084 if (copied)
1085 cleanup_rbuf(sk, copied);
1086 return copied;
1087}
1088
1089/*
1090 * This routine copies from a sock struct into the user buffer.
1091 *
1092 * Technical note: in 2.3 we work on _locked_ socket, so that
1093 * tricks with *seq access order and skb->users are not required.
1094 * Probably, code can be easily improved even more.
1095 */
1096
1097int tcp_recvmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
1098 size_t len, int nonblock, int flags, int *addr_len)
1099{
1100 struct tcp_sock *tp = tcp_sk(sk);
1101 int copied = 0;
1102 u32 peek_seq;
1103 u32 *seq;
1104 unsigned long used;
1105 int err;
1106 int target; /* Read at least this many bytes */
1107 long timeo;
1108 struct task_struct *user_recv = NULL;
1109
1110 lock_sock(sk);
1111
1112 TCP_CHECK_TIMER(sk);
1113
1114 err = -ENOTCONN;
1115 if (sk->sk_state == TCP_LISTEN)
1116 goto out;
1117
1118 timeo = sock_rcvtimeo(sk, nonblock);
1119
1120 /* Urgent data needs to be handled specially. */
1121 if (flags & MSG_OOB)
1122 goto recv_urg;
1123
1124 seq = &tp->copied_seq;
1125 if (flags & MSG_PEEK) {
1126 peek_seq = tp->copied_seq;
1127 seq = &peek_seq;
1128 }
1129
1130 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1131
1132 do {
1133 struct sk_buff *skb;
1134 u32 offset;
1135
1136 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1137 if (tp->urg_data && tp->urg_seq == *seq) {
1138 if (copied)
1139 break;
1140 if (signal_pending(current)) {
1141 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1142 break;
1143 }
1144 }
1145
1146 /* Next get a buffer. */
1147
1148 skb = skb_peek(&sk->sk_receive_queue);
1149 do {
1150 if (!skb)
1151 break;
1152
1153 /* Now that we have two receive queues this
1154 * shouldn't happen.
1155 */
1156 if (before(*seq, TCP_SKB_CB(skb)->seq)) {
1157 printk(KERN_INFO "recvmsg bug: copied %X "
1158 "seq %X\n", *seq, TCP_SKB_CB(skb)->seq);
1159 break;
1160 }
1161 offset = *seq - TCP_SKB_CB(skb)->seq;
1162 if (skb->h.th->syn)
1163 offset--;
1164 if (offset < skb->len)
1165 goto found_ok_skb;
1166 if (skb->h.th->fin)
1167 goto found_fin_ok;
1168 BUG_TRAP(flags & MSG_PEEK);
1169 skb = skb->next;
1170 } while (skb != (struct sk_buff *)&sk->sk_receive_queue);
1171
1172 /* Well, if we have backlog, try to process it now yet. */
1173
1174 if (copied >= target && !sk->sk_backlog.tail)
1175 break;
1176
1177 if (copied) {
1178 if (sk->sk_err ||
1179 sk->sk_state == TCP_CLOSE ||
1180 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1181 !timeo ||
1182 signal_pending(current) ||
1183 (flags & MSG_PEEK))
1184 break;
1185 } else {
1186 if (sock_flag(sk, SOCK_DONE))
1187 break;
1188
1189 if (sk->sk_err) {
1190 copied = sock_error(sk);
1191 break;
1192 }
1193
1194 if (sk->sk_shutdown & RCV_SHUTDOWN)
1195 break;
1196
1197 if (sk->sk_state == TCP_CLOSE) {
1198 if (!sock_flag(sk, SOCK_DONE)) {
1199 /* This occurs when user tries to read
1200 * from never connected socket.
1201 */
1202 copied = -ENOTCONN;
1203 break;
1204 }
1205 break;
1206 }
1207
1208 if (!timeo) {
1209 copied = -EAGAIN;
1210 break;
1211 }
1212
1213 if (signal_pending(current)) {
1214 copied = sock_intr_errno(timeo);
1215 break;
1216 }
1217 }
1218
1219 cleanup_rbuf(sk, copied);
1220
David S. Miller7df55122005-06-18 23:01:10 -07001221 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222 /* Install new reader */
1223 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1224 user_recv = current;
1225 tp->ucopy.task = user_recv;
1226 tp->ucopy.iov = msg->msg_iov;
1227 }
1228
1229 tp->ucopy.len = len;
1230
1231 BUG_TRAP(tp->copied_seq == tp->rcv_nxt ||
1232 (flags & (MSG_PEEK | MSG_TRUNC)));
1233
1234 /* Ugly... If prequeue is not empty, we have to
1235 * process it before releasing socket, otherwise
1236 * order will be broken at second iteration.
1237 * More elegant solution is required!!!
1238 *
1239 * Look: we have the following (pseudo)queues:
1240 *
1241 * 1. packets in flight
1242 * 2. backlog
1243 * 3. prequeue
1244 * 4. receive_queue
1245 *
1246 * Each queue can be processed only if the next ones
1247 * are empty. At this point we have empty receive_queue.
1248 * But prequeue _can_ be not empty after 2nd iteration,
1249 * when we jumped to start of loop because backlog
1250 * processing added something to receive_queue.
1251 * We cannot release_sock(), because backlog contains
1252 * packets arrived _after_ prequeued ones.
1253 *
1254 * Shortly, algorithm is clear --- to process all
1255 * the queues in order. We could make it more directly,
1256 * requeueing packets from backlog to prequeue, if
1257 * is not empty. It is more elegant, but eats cycles,
1258 * unfortunately.
1259 */
David S. Millerb03efcf2005-07-08 14:57:23 -07001260 if (!skb_queue_empty(&tp->ucopy.prequeue))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 goto do_prequeue;
1262
1263 /* __ Set realtime policy in scheduler __ */
1264 }
1265
1266 if (copied >= target) {
1267 /* Do not sleep, just process backlog. */
1268 release_sock(sk);
1269 lock_sock(sk);
1270 } else
1271 sk_wait_data(sk, &timeo);
1272
1273 if (user_recv) {
1274 int chunk;
1275
1276 /* __ Restore normal policy in scheduler __ */
1277
1278 if ((chunk = len - tp->ucopy.len) != 0) {
1279 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1280 len -= chunk;
1281 copied += chunk;
1282 }
1283
1284 if (tp->rcv_nxt == tp->copied_seq &&
David S. Millerb03efcf2005-07-08 14:57:23 -07001285 !skb_queue_empty(&tp->ucopy.prequeue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286do_prequeue:
1287 tcp_prequeue_process(sk);
1288
1289 if ((chunk = len - tp->ucopy.len) != 0) {
1290 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1291 len -= chunk;
1292 copied += chunk;
1293 }
1294 }
1295 }
1296 if ((flags & MSG_PEEK) && peek_seq != tp->copied_seq) {
1297 if (net_ratelimit())
1298 printk(KERN_DEBUG "TCP(%s:%d): Application bug, race in MSG_PEEK.\n",
1299 current->comm, current->pid);
1300 peek_seq = tp->copied_seq;
1301 }
1302 continue;
1303
1304 found_ok_skb:
1305 /* Ok so how much can we use? */
1306 used = skb->len - offset;
1307 if (len < used)
1308 used = len;
1309
1310 /* Do we have urgent data here? */
1311 if (tp->urg_data) {
1312 u32 urg_offset = tp->urg_seq - *seq;
1313 if (urg_offset < used) {
1314 if (!urg_offset) {
1315 if (!sock_flag(sk, SOCK_URGINLINE)) {
1316 ++*seq;
1317 offset++;
1318 used--;
1319 if (!used)
1320 goto skip_copy;
1321 }
1322 } else
1323 used = urg_offset;
1324 }
1325 }
1326
1327 if (!(flags & MSG_TRUNC)) {
1328 err = skb_copy_datagram_iovec(skb, offset,
1329 msg->msg_iov, used);
1330 if (err) {
1331 /* Exception. Bailout! */
1332 if (!copied)
1333 copied = -EFAULT;
1334 break;
1335 }
1336 }
1337
1338 *seq += used;
1339 copied += used;
1340 len -= used;
1341
1342 tcp_rcv_space_adjust(sk);
1343
1344skip_copy:
1345 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1346 tp->urg_data = 0;
1347 tcp_fast_path_check(sk, tp);
1348 }
1349 if (used + offset < skb->len)
1350 continue;
1351
1352 if (skb->h.th->fin)
1353 goto found_fin_ok;
1354 if (!(flags & MSG_PEEK))
1355 sk_eat_skb(sk, skb);
1356 continue;
1357
1358 found_fin_ok:
1359 /* Process the FIN. */
1360 ++*seq;
1361 if (!(flags & MSG_PEEK))
1362 sk_eat_skb(sk, skb);
1363 break;
1364 } while (len > 0);
1365
1366 if (user_recv) {
David S. Millerb03efcf2005-07-08 14:57:23 -07001367 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 int chunk;
1369
1370 tp->ucopy.len = copied > 0 ? len : 0;
1371
1372 tcp_prequeue_process(sk);
1373
1374 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1375 NET_ADD_STATS_USER(LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1376 len -= chunk;
1377 copied += chunk;
1378 }
1379 }
1380
1381 tp->ucopy.task = NULL;
1382 tp->ucopy.len = 0;
1383 }
1384
1385 /* According to UNIX98, msg_name/msg_namelen are ignored
1386 * on connected socket. I was just happy when found this 8) --ANK
1387 */
1388
1389 /* Clean up data we have read: This will do ACK frames. */
1390 cleanup_rbuf(sk, copied);
1391
1392 TCP_CHECK_TIMER(sk);
1393 release_sock(sk);
1394 return copied;
1395
1396out:
1397 TCP_CHECK_TIMER(sk);
1398 release_sock(sk);
1399 return err;
1400
1401recv_urg:
1402 err = tcp_recv_urg(sk, timeo, msg, len, flags, addr_len);
1403 goto out;
1404}
1405
1406/*
1407 * State processing on a close. This implements the state shift for
1408 * sending our FIN frame. Note that we only send a FIN for some
1409 * states. A shutdown() may have already sent the FIN, or we may be
1410 * closed.
1411 */
1412
1413static unsigned char new_state[16] = {
1414 /* current state: new state: action: */
1415 /* (Invalid) */ TCP_CLOSE,
1416 /* TCP_ESTABLISHED */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1417 /* TCP_SYN_SENT */ TCP_CLOSE,
1418 /* TCP_SYN_RECV */ TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1419 /* TCP_FIN_WAIT1 */ TCP_FIN_WAIT1,
1420 /* TCP_FIN_WAIT2 */ TCP_FIN_WAIT2,
1421 /* TCP_TIME_WAIT */ TCP_CLOSE,
1422 /* TCP_CLOSE */ TCP_CLOSE,
1423 /* TCP_CLOSE_WAIT */ TCP_LAST_ACK | TCP_ACTION_FIN,
1424 /* TCP_LAST_ACK */ TCP_LAST_ACK,
1425 /* TCP_LISTEN */ TCP_CLOSE,
1426 /* TCP_CLOSING */ TCP_CLOSING,
1427};
1428
1429static int tcp_close_state(struct sock *sk)
1430{
1431 int next = (int)new_state[sk->sk_state];
1432 int ns = next & TCP_STATE_MASK;
1433
1434 tcp_set_state(sk, ns);
1435
1436 return next & TCP_ACTION_FIN;
1437}
1438
1439/*
1440 * Shutdown the sending side of a connection. Much like close except
1441 * that we don't receive shut down or set_sock_flag(sk, SOCK_DEAD).
1442 */
1443
1444void tcp_shutdown(struct sock *sk, int how)
1445{
1446 /* We need to grab some memory, and put together a FIN,
1447 * and then put it into the queue to be sent.
1448 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
1449 */
1450 if (!(how & SEND_SHUTDOWN))
1451 return;
1452
1453 /* If we've already sent a FIN, or it's a closed state, skip this. */
1454 if ((1 << sk->sk_state) &
1455 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
1456 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
1457 /* Clear out any half completed packets. FIN if needed. */
1458 if (tcp_close_state(sk))
1459 tcp_send_fin(sk);
1460 }
1461}
1462
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463void tcp_close(struct sock *sk, long timeout)
1464{
1465 struct sk_buff *skb;
1466 int data_was_unread = 0;
1467
1468 lock_sock(sk);
1469 sk->sk_shutdown = SHUTDOWN_MASK;
1470
1471 if (sk->sk_state == TCP_LISTEN) {
1472 tcp_set_state(sk, TCP_CLOSE);
1473
1474 /* Special case. */
Arnaldo Carvalho de Melo0a5578c2005-08-09 20:11:41 -07001475 inet_csk_listen_stop(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476
1477 goto adjudge_to_death;
1478 }
1479
1480 /* We need to flush the recv. buffs. We do this only on the
1481 * descriptor close, not protocol-sourced closes, because the
1482 * reader process may not have drained the data yet!
1483 */
1484 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
1485 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq -
1486 skb->h.th->fin;
1487 data_was_unread += len;
1488 __kfree_skb(skb);
1489 }
1490
1491 sk_stream_mem_reclaim(sk);
1492
1493 /* As outlined in draft-ietf-tcpimpl-prob-03.txt, section
1494 * 3.10, we send a RST here because data was lost. To
1495 * witness the awful effects of the old behavior of always
1496 * doing a FIN, run an older 2.1.x kernel or 2.0.x, start
1497 * a bulk GET in an FTP client, suspend the process, wait
1498 * for the client to advertise a zero window, then kill -9
1499 * the FTP client, wheee... Note: timeout is always zero
1500 * in such a case.
1501 */
1502 if (data_was_unread) {
1503 /* Unread data was tossed, zap the connection. */
1504 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONCLOSE);
1505 tcp_set_state(sk, TCP_CLOSE);
1506 tcp_send_active_reset(sk, GFP_KERNEL);
1507 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
1508 /* Check zero linger _after_ checking for unread data. */
1509 sk->sk_prot->disconnect(sk, 0);
1510 NET_INC_STATS_USER(LINUX_MIB_TCPABORTONDATA);
1511 } else if (tcp_close_state(sk)) {
1512 /* We FIN if the application ate all the data before
1513 * zapping the connection.
1514 */
1515
1516 /* RED-PEN. Formally speaking, we have broken TCP state
1517 * machine. State transitions:
1518 *
1519 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
1520 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
1521 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
1522 *
1523 * are legal only when FIN has been sent (i.e. in window),
1524 * rather than queued out of window. Purists blame.
1525 *
1526 * F.e. "RFC state" is ESTABLISHED,
1527 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
1528 *
1529 * The visible declinations are that sometimes
1530 * we enter time-wait state, when it is not required really
1531 * (harmless), do not send active resets, when they are
1532 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
1533 * they look as CLOSING or LAST_ACK for Linux)
1534 * Probably, I missed some more holelets.
1535 * --ANK
1536 */
1537 tcp_send_fin(sk);
1538 }
1539
1540 sk_stream_wait_close(sk, timeout);
1541
1542adjudge_to_death:
1543 /* It is the last release_sock in its life. It will remove backlog. */
1544 release_sock(sk);
1545
1546
1547 /* Now socket is owned by kernel and we acquire BH lock
1548 to finish close. No need to check for user refs.
1549 */
1550 local_bh_disable();
1551 bh_lock_sock(sk);
1552 BUG_TRAP(!sock_owned_by_user(sk));
1553
1554 sock_hold(sk);
1555 sock_orphan(sk);
1556
1557 /* This is a (useful) BSD violating of the RFC. There is a
1558 * problem with TCP as specified in that the other end could
1559 * keep a socket open forever with no application left this end.
1560 * We use a 3 minute timeout (about the same as BSD) then kill
1561 * our end. If they send after that then tough - BUT: long enough
1562 * that we won't make the old 4*rto = almost no time - whoops
1563 * reset mistake.
1564 *
1565 * Nope, it was not mistake. It is really desired behaviour
1566 * f.e. on http servers, when such sockets are useless, but
1567 * consume significant resources. Let's do it with special
1568 * linger2 option. --ANK
1569 */
1570
1571 if (sk->sk_state == TCP_FIN_WAIT2) {
1572 struct tcp_sock *tp = tcp_sk(sk);
1573 if (tp->linger2 < 0) {
1574 tcp_set_state(sk, TCP_CLOSE);
1575 tcp_send_active_reset(sk, GFP_ATOMIC);
1576 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONLINGER);
1577 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001578 const int tmo = tcp_fin_time(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579
1580 if (tmo > TCP_TIMEWAIT_LEN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001581 inet_csk_reset_keepalive_timer(sk, tcp_fin_time(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582 } else {
Arnaldo Carvalho de Melo0a5578c2005-08-09 20:11:41 -07001583 atomic_inc(sk->sk_prot->orphan_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
1585 goto out;
1586 }
1587 }
1588 }
1589 if (sk->sk_state != TCP_CLOSE) {
1590 sk_stream_mem_reclaim(sk);
Arnaldo Carvalho de Melo0a5578c2005-08-09 20:11:41 -07001591 if (atomic_read(sk->sk_prot->orphan_count) > sysctl_tcp_max_orphans ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592 (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
1593 atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
1594 if (net_ratelimit())
1595 printk(KERN_INFO "TCP: too many of orphaned "
1596 "sockets\n");
1597 tcp_set_state(sk, TCP_CLOSE);
1598 tcp_send_active_reset(sk, GFP_ATOMIC);
1599 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
1600 }
1601 }
Arnaldo Carvalho de Melo0a5578c2005-08-09 20:11:41 -07001602 atomic_inc(sk->sk_prot->orphan_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603
1604 if (sk->sk_state == TCP_CLOSE)
Arnaldo Carvalho de Melo0a5578c2005-08-09 20:11:41 -07001605 inet_csk_destroy_sock(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 /* Otherwise, socket is reprieved until protocol close. */
1607
1608out:
1609 bh_unlock_sock(sk);
1610 local_bh_enable();
1611 sock_put(sk);
1612}
1613
1614/* These states need RST on ABORT according to RFC793 */
1615
1616static inline int tcp_need_reset(int state)
1617{
1618 return (1 << state) &
1619 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
1620 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
1621}
1622
1623int tcp_disconnect(struct sock *sk, int flags)
1624{
1625 struct inet_sock *inet = inet_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001626 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627 struct tcp_sock *tp = tcp_sk(sk);
1628 int err = 0;
1629 int old_state = sk->sk_state;
1630
1631 if (old_state != TCP_CLOSE)
1632 tcp_set_state(sk, TCP_CLOSE);
1633
1634 /* ABORT function of RFC793 */
1635 if (old_state == TCP_LISTEN) {
Arnaldo Carvalho de Melo0a5578c2005-08-09 20:11:41 -07001636 inet_csk_listen_stop(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637 } else if (tcp_need_reset(old_state) ||
1638 (tp->snd_nxt != tp->write_seq &&
1639 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
1640 /* The last check adjusts for discrepance of Linux wrt. RFC
1641 * states
1642 */
1643 tcp_send_active_reset(sk, gfp_any());
1644 sk->sk_err = ECONNRESET;
1645 } else if (old_state == TCP_SYN_SENT)
1646 sk->sk_err = ECONNRESET;
1647
1648 tcp_clear_xmit_timers(sk);
1649 __skb_queue_purge(&sk->sk_receive_queue);
1650 sk_stream_writequeue_purge(sk);
1651 __skb_queue_purge(&tp->out_of_order_queue);
1652
1653 inet->dport = 0;
1654
1655 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
1656 inet_reset_saddr(sk);
1657
1658 sk->sk_shutdown = 0;
1659 sock_reset_flag(sk, SOCK_DONE);
1660 tp->srtt = 0;
1661 if ((tp->write_seq += tp->max_window + 2) == 0)
1662 tp->write_seq = 1;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001663 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664 tp->snd_cwnd = 2;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001665 icsk->icsk_probes_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666 tp->packets_out = 0;
1667 tp->snd_ssthresh = 0x7fffffff;
1668 tp->snd_cwnd_cnt = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001669 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670 tcp_clear_retrans(tp);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001671 inet_csk_delack_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 sk->sk_send_head = NULL;
1673 tp->rx_opt.saw_tstamp = 0;
1674 tcp_sack_reset(&tp->rx_opt);
1675 __sk_dst_reset(sk);
1676
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001677 BUG_TRAP(!inet->num || icsk->icsk_bind_hash);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678
1679 sk->sk_error_report(sk);
1680 return err;
1681}
1682
1683/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 * Socket option code for TCP.
1685 */
1686int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
1687 int optlen)
1688{
1689 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001690 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691 int val;
1692 int err = 0;
1693
1694 if (level != SOL_TCP)
1695 return tp->af_specific->setsockopt(sk, level, optname,
1696 optval, optlen);
1697
Stephen Hemminger5f8ef482005-06-23 20:37:36 -07001698 /* This is a string value all the others are int's */
1699 if (optname == TCP_CONGESTION) {
1700 char name[TCP_CA_NAME_MAX];
1701
1702 if (optlen < 1)
1703 return -EINVAL;
1704
1705 val = strncpy_from_user(name, optval,
1706 min(TCP_CA_NAME_MAX-1, optlen));
1707 if (val < 0)
1708 return -EFAULT;
1709 name[val] = 0;
1710
1711 lock_sock(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001712 err = tcp_set_congestion_control(sk, name);
Stephen Hemminger5f8ef482005-06-23 20:37:36 -07001713 release_sock(sk);
1714 return err;
1715 }
1716
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 if (optlen < sizeof(int))
1718 return -EINVAL;
1719
1720 if (get_user(val, (int __user *)optval))
1721 return -EFAULT;
1722
1723 lock_sock(sk);
1724
1725 switch (optname) {
1726 case TCP_MAXSEG:
1727 /* Values greater than interface MTU won't take effect. However
1728 * at the point when this call is done we typically don't yet
1729 * know which interface is going to be used */
1730 if (val < 8 || val > MAX_TCP_WINDOW) {
1731 err = -EINVAL;
1732 break;
1733 }
1734 tp->rx_opt.user_mss = val;
1735 break;
1736
1737 case TCP_NODELAY:
1738 if (val) {
1739 /* TCP_NODELAY is weaker than TCP_CORK, so that
1740 * this option on corked socket is remembered, but
1741 * it is not activated until cork is cleared.
1742 *
1743 * However, when TCP_NODELAY is set we make
1744 * an explicit push, which overrides even TCP_CORK
1745 * for currently queued segments.
1746 */
1747 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
1748 tcp_push_pending_frames(sk, tp);
1749 } else {
1750 tp->nonagle &= ~TCP_NAGLE_OFF;
1751 }
1752 break;
1753
1754 case TCP_CORK:
1755 /* When set indicates to always queue non-full frames.
1756 * Later the user clears this option and we transmit
1757 * any pending partial frames in the queue. This is
1758 * meant to be used alongside sendfile() to get properly
1759 * filled frames when the user (for example) must write
1760 * out headers with a write() call first and then use
1761 * sendfile to send out the data parts.
1762 *
1763 * TCP_CORK can be set together with TCP_NODELAY and it is
1764 * stronger than TCP_NODELAY.
1765 */
1766 if (val) {
1767 tp->nonagle |= TCP_NAGLE_CORK;
1768 } else {
1769 tp->nonagle &= ~TCP_NAGLE_CORK;
1770 if (tp->nonagle&TCP_NAGLE_OFF)
1771 tp->nonagle |= TCP_NAGLE_PUSH;
1772 tcp_push_pending_frames(sk, tp);
1773 }
1774 break;
1775
1776 case TCP_KEEPIDLE:
1777 if (val < 1 || val > MAX_TCP_KEEPIDLE)
1778 err = -EINVAL;
1779 else {
1780 tp->keepalive_time = val * HZ;
1781 if (sock_flag(sk, SOCK_KEEPOPEN) &&
1782 !((1 << sk->sk_state) &
1783 (TCPF_CLOSE | TCPF_LISTEN))) {
1784 __u32 elapsed = tcp_time_stamp - tp->rcv_tstamp;
1785 if (tp->keepalive_time > elapsed)
1786 elapsed = tp->keepalive_time - elapsed;
1787 else
1788 elapsed = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001789 inet_csk_reset_keepalive_timer(sk, elapsed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 }
1791 }
1792 break;
1793 case TCP_KEEPINTVL:
1794 if (val < 1 || val > MAX_TCP_KEEPINTVL)
1795 err = -EINVAL;
1796 else
1797 tp->keepalive_intvl = val * HZ;
1798 break;
1799 case TCP_KEEPCNT:
1800 if (val < 1 || val > MAX_TCP_KEEPCNT)
1801 err = -EINVAL;
1802 else
1803 tp->keepalive_probes = val;
1804 break;
1805 case TCP_SYNCNT:
1806 if (val < 1 || val > MAX_TCP_SYNCNT)
1807 err = -EINVAL;
1808 else
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001809 icsk->icsk_syn_retries = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 break;
1811
1812 case TCP_LINGER2:
1813 if (val < 0)
1814 tp->linger2 = -1;
1815 else if (val > sysctl_tcp_fin_timeout / HZ)
1816 tp->linger2 = 0;
1817 else
1818 tp->linger2 = val * HZ;
1819 break;
1820
1821 case TCP_DEFER_ACCEPT:
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07001822 icsk->icsk_accept_queue.rskq_defer_accept = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823 if (val > 0) {
1824 /* Translate value in seconds to number of
1825 * retransmits */
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07001826 while (icsk->icsk_accept_queue.rskq_defer_accept < 32 &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 val > ((TCP_TIMEOUT_INIT / HZ) <<
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07001828 icsk->icsk_accept_queue.rskq_defer_accept))
1829 icsk->icsk_accept_queue.rskq_defer_accept++;
1830 icsk->icsk_accept_queue.rskq_defer_accept++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 }
1832 break;
1833
1834 case TCP_WINDOW_CLAMP:
1835 if (!val) {
1836 if (sk->sk_state != TCP_CLOSE) {
1837 err = -EINVAL;
1838 break;
1839 }
1840 tp->window_clamp = 0;
1841 } else
1842 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
1843 SOCK_MIN_RCVBUF / 2 : val;
1844 break;
1845
1846 case TCP_QUICKACK:
1847 if (!val) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001848 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001850 icsk->icsk_ack.pingpong = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851 if ((1 << sk->sk_state) &
1852 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001853 inet_csk_ack_scheduled(sk)) {
1854 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 cleanup_rbuf(sk, 1);
1856 if (!(val & 1))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001857 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858 }
1859 }
1860 break;
1861
1862 default:
1863 err = -ENOPROTOOPT;
1864 break;
1865 };
1866 release_sock(sk);
1867 return err;
1868}
1869
1870/* Return information about state of tcp endpoint in API format. */
1871void tcp_get_info(struct sock *sk, struct tcp_info *info)
1872{
1873 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001874 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 u32 now = tcp_time_stamp;
1876
1877 memset(info, 0, sizeof(*info));
1878
1879 info->tcpi_state = sk->sk_state;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001880 info->tcpi_ca_state = icsk->icsk_ca_state;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001881 info->tcpi_retransmits = icsk->icsk_retransmits;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001882 info->tcpi_probes = icsk->icsk_probes_out;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001883 info->tcpi_backoff = icsk->icsk_backoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884
1885 if (tp->rx_opt.tstamp_ok)
1886 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
1887 if (tp->rx_opt.sack_ok)
1888 info->tcpi_options |= TCPI_OPT_SACK;
1889 if (tp->rx_opt.wscale_ok) {
1890 info->tcpi_options |= TCPI_OPT_WSCALE;
1891 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
1892 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
1893 }
1894
1895 if (tp->ecn_flags&TCP_ECN_OK)
1896 info->tcpi_options |= TCPI_OPT_ECN;
1897
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001898 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
1899 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001900 info->tcpi_snd_mss = tp->mss_cache;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001901 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902
1903 info->tcpi_unacked = tp->packets_out;
1904 info->tcpi_sacked = tp->sacked_out;
1905 info->tcpi_lost = tp->lost_out;
1906 info->tcpi_retrans = tp->retrans_out;
1907 info->tcpi_fackets = tp->fackets_out;
1908
1909 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001910 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
1912
1913 info->tcpi_pmtu = tp->pmtu_cookie;
1914 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
1915 info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
1916 info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>2;
1917 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
1918 info->tcpi_snd_cwnd = tp->snd_cwnd;
1919 info->tcpi_advmss = tp->advmss;
1920 info->tcpi_reordering = tp->reordering;
1921
1922 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
1923 info->tcpi_rcv_space = tp->rcvq_space.space;
1924
1925 info->tcpi_total_retrans = tp->total_retrans;
1926}
1927
1928EXPORT_SYMBOL_GPL(tcp_get_info);
1929
1930int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
1931 int __user *optlen)
1932{
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07001933 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934 struct tcp_sock *tp = tcp_sk(sk);
1935 int val, len;
1936
1937 if (level != SOL_TCP)
1938 return tp->af_specific->getsockopt(sk, level, optname,
1939 optval, optlen);
1940
1941 if (get_user(len, optlen))
1942 return -EFAULT;
1943
1944 len = min_t(unsigned int, len, sizeof(int));
1945
1946 if (len < 0)
1947 return -EINVAL;
1948
1949 switch (optname) {
1950 case TCP_MAXSEG:
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001951 val = tp->mss_cache;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
1953 val = tp->rx_opt.user_mss;
1954 break;
1955 case TCP_NODELAY:
1956 val = !!(tp->nonagle&TCP_NAGLE_OFF);
1957 break;
1958 case TCP_CORK:
1959 val = !!(tp->nonagle&TCP_NAGLE_CORK);
1960 break;
1961 case TCP_KEEPIDLE:
1962 val = (tp->keepalive_time ? : sysctl_tcp_keepalive_time) / HZ;
1963 break;
1964 case TCP_KEEPINTVL:
1965 val = (tp->keepalive_intvl ? : sysctl_tcp_keepalive_intvl) / HZ;
1966 break;
1967 case TCP_KEEPCNT:
1968 val = tp->keepalive_probes ? : sysctl_tcp_keepalive_probes;
1969 break;
1970 case TCP_SYNCNT:
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07001971 val = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001972 break;
1973 case TCP_LINGER2:
1974 val = tp->linger2;
1975 if (val >= 0)
1976 val = (val ? : sysctl_tcp_fin_timeout) / HZ;
1977 break;
1978 case TCP_DEFER_ACCEPT:
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07001979 val = !icsk->icsk_accept_queue.rskq_defer_accept ? 0 :
1980 ((TCP_TIMEOUT_INIT / HZ) << (icsk->icsk_accept_queue.rskq_defer_accept - 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 break;
1982 case TCP_WINDOW_CLAMP:
1983 val = tp->window_clamp;
1984 break;
1985 case TCP_INFO: {
1986 struct tcp_info info;
1987
1988 if (get_user(len, optlen))
1989 return -EFAULT;
1990
1991 tcp_get_info(sk, &info);
1992
1993 len = min_t(unsigned int, len, sizeof(info));
1994 if (put_user(len, optlen))
1995 return -EFAULT;
1996 if (copy_to_user(optval, &info, len))
1997 return -EFAULT;
1998 return 0;
1999 }
2000 case TCP_QUICKACK:
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07002001 val = !icsk->icsk_ack.pingpong;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 break;
Stephen Hemminger5f8ef482005-06-23 20:37:36 -07002003
2004 case TCP_CONGESTION:
2005 if (get_user(len, optlen))
2006 return -EFAULT;
2007 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2008 if (put_user(len, optlen))
2009 return -EFAULT;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002010 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
Stephen Hemminger5f8ef482005-06-23 20:37:36 -07002011 return -EFAULT;
2012 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013 default:
2014 return -ENOPROTOOPT;
2015 };
2016
2017 if (put_user(len, optlen))
2018 return -EFAULT;
2019 if (copy_to_user(optval, &val, len))
2020 return -EFAULT;
2021 return 0;
2022}
2023
2024
2025extern void __skb_cb_too_small_for_tcp(int, int);
Stephen Hemminger5f8ef482005-06-23 20:37:36 -07002026extern struct tcp_congestion_ops tcp_reno;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027
2028static __initdata unsigned long thash_entries;
2029static int __init set_thash_entries(char *str)
2030{
2031 if (!str)
2032 return 0;
2033 thash_entries = simple_strtoul(str, &str, 0);
2034 return 1;
2035}
2036__setup("thash_entries=", set_thash_entries);
2037
2038void __init tcp_init(void)
2039{
2040 struct sk_buff *skb = NULL;
2041 int order, i;
2042
2043 if (sizeof(struct tcp_skb_cb) > sizeof(skb->cb))
2044 __skb_cb_too_small_for_tcp(sizeof(struct tcp_skb_cb),
2045 sizeof(skb->cb));
2046
Arnaldo Carvalho de Melo6e04e022005-08-09 20:07:35 -07002047 tcp_hashinfo.bind_bucket_cachep =
2048 kmem_cache_create("tcp_bind_bucket",
2049 sizeof(struct inet_bind_bucket), 0,
2050 SLAB_HWCACHE_ALIGN, NULL, NULL);
2051 if (!tcp_hashinfo.bind_bucket_cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 panic("tcp_init: Cannot alloc tcp_bind_bucket cache.");
2053
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054 /* Size and allocate the main established and bind bucket
2055 * hash tables.
2056 *
2057 * The methodology is similar to that of the buffer cache.
2058 */
Arnaldo Carvalho de Melo6e04e022005-08-09 20:07:35 -07002059 tcp_hashinfo.ehash =
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060 alloc_large_system_hash("TCP established",
Arnaldo Carvalho de Melo0f7ff922005-08-09 19:59:44 -07002061 sizeof(struct inet_ehash_bucket),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062 thash_entries,
2063 (num_physpages >= 128 * 1024) ?
2064 (25 - PAGE_SHIFT) :
2065 (27 - PAGE_SHIFT),
2066 HASH_HIGHMEM,
Arnaldo Carvalho de Melo6e04e022005-08-09 20:07:35 -07002067 &tcp_hashinfo.ehash_size,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068 NULL,
2069 0);
Arnaldo Carvalho de Melo6e04e022005-08-09 20:07:35 -07002070 tcp_hashinfo.ehash_size = (1 << tcp_hashinfo.ehash_size) >> 1;
2071 for (i = 0; i < (tcp_hashinfo.ehash_size << 1); i++) {
2072 rwlock_init(&tcp_hashinfo.ehash[i].lock);
2073 INIT_HLIST_HEAD(&tcp_hashinfo.ehash[i].chain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074 }
2075
Arnaldo Carvalho de Melo6e04e022005-08-09 20:07:35 -07002076 tcp_hashinfo.bhash =
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 alloc_large_system_hash("TCP bind",
Arnaldo Carvalho de Melo0f7ff922005-08-09 19:59:44 -07002078 sizeof(struct inet_bind_hashbucket),
Arnaldo Carvalho de Melo6e04e022005-08-09 20:07:35 -07002079 tcp_hashinfo.ehash_size,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080 (num_physpages >= 128 * 1024) ?
2081 (25 - PAGE_SHIFT) :
2082 (27 - PAGE_SHIFT),
2083 HASH_HIGHMEM,
Arnaldo Carvalho de Melo6e04e022005-08-09 20:07:35 -07002084 &tcp_hashinfo.bhash_size,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 NULL,
2086 64 * 1024);
Arnaldo Carvalho de Melo6e04e022005-08-09 20:07:35 -07002087 tcp_hashinfo.bhash_size = 1 << tcp_hashinfo.bhash_size;
2088 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
2089 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
2090 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 }
2092
2093 /* Try to be a bit smarter and adjust defaults depending
2094 * on available memory.
2095 */
2096 for (order = 0; ((1 << order) << PAGE_SHIFT) <
Arnaldo Carvalho de Melo6e04e022005-08-09 20:07:35 -07002097 (tcp_hashinfo.bhash_size * sizeof(struct inet_bind_hashbucket));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 order++)
2099 ;
Andi Kleene7626482005-06-13 14:24:52 -07002100 if (order >= 4) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101 sysctl_local_port_range[0] = 32768;
2102 sysctl_local_port_range[1] = 61000;
Arnaldo Carvalho de Melo295ff7e2005-08-09 20:44:40 -07002103 tcp_death_row.sysctl_max_tw_buckets = 180000;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104 sysctl_tcp_max_orphans = 4096 << (order - 4);
2105 sysctl_max_syn_backlog = 1024;
2106 } else if (order < 3) {
2107 sysctl_local_port_range[0] = 1024 * (3 - order);
Arnaldo Carvalho de Melo295ff7e2005-08-09 20:44:40 -07002108 tcp_death_row.sysctl_max_tw_buckets >>= (3 - order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109 sysctl_tcp_max_orphans >>= (3 - order);
2110 sysctl_max_syn_backlog = 128;
2111 }
Arnaldo Carvalho de Melo6e04e022005-08-09 20:07:35 -07002112 tcp_hashinfo.port_rover = sysctl_local_port_range[0] - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113
2114 sysctl_tcp_mem[0] = 768 << order;
2115 sysctl_tcp_mem[1] = 1024 << order;
2116 sysctl_tcp_mem[2] = 1536 << order;
2117
2118 if (order < 3) {
2119 sysctl_tcp_wmem[2] = 64 * 1024;
2120 sysctl_tcp_rmem[0] = PAGE_SIZE;
2121 sysctl_tcp_rmem[1] = 43689;
2122 sysctl_tcp_rmem[2] = 2 * 43689;
2123 }
2124
2125 printk(KERN_INFO "TCP: Hash tables configured "
2126 "(established %d bind %d)\n",
Arnaldo Carvalho de Melo6e04e022005-08-09 20:07:35 -07002127 tcp_hashinfo.ehash_size << 1, tcp_hashinfo.bhash_size);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002128
2129 tcp_register_congestion_control(&tcp_reno);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130}
2131
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132EXPORT_SYMBOL(tcp_close);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133EXPORT_SYMBOL(tcp_disconnect);
2134EXPORT_SYMBOL(tcp_getsockopt);
2135EXPORT_SYMBOL(tcp_ioctl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136EXPORT_SYMBOL(tcp_poll);
2137EXPORT_SYMBOL(tcp_read_sock);
2138EXPORT_SYMBOL(tcp_recvmsg);
2139EXPORT_SYMBOL(tcp_sendmsg);
2140EXPORT_SYMBOL(tcp_sendpage);
2141EXPORT_SYMBOL(tcp_setsockopt);
2142EXPORT_SYMBOL(tcp_shutdown);
2143EXPORT_SYMBOL(tcp_statistics);