![[first]](../icons/first.png){kind=icon}
![[previous]](../icons/n_left.png){kind=icon}
![[next]](../icons/n_right.png){kind=icon}
![[last]](../icons/last.png){kind=icon}
![[top]](../icons/n_top.png){kind=icon}
![[bottom]](../icons/bottom.png){kind=icon}
![[index]](../icons/index.png){kind=icon}
![[help]](../icons/help.png){kind=icon}
*/
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: @(#)tcp.c 1.0.16 05/25/93
9 *
10 * Authors: Ross Biro, <bir7@leland.Stanford.Edu>
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. select
34 * behaves and the icmp error race
35 * has gone by moving it into sock.c
36 * Alan Cox : tcp_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 select() 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(), select() 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 selecting 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 : 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 stat 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 its a start!
176 * Marc Tamsky : Closing in closing fixes.
177 * Mike Shaver : RFC1122 verifications.
178 * Alan Cox : rcv_saddr errors.
179 *
180 *
181 * To Fix:
182 * Fast path the code. Two things here - fix the window calculation
183 * so it doesn't iterate over the queue, also spot packets with no funny
184 * options arriving in order and process directly.
185 *
186 * Implement RFC 1191 [Path MTU discovery]
187 * Look at the effect of implementing RFC 1337 suggestions and their impact.
188 * Rewrite output state machine to use a single queue and do low window
189 * situations as per the spec (RFC 1122)
190 * Speed up input assembly algorithm.
191 * RFC1323 - PAWS and window scaling. PAWS is required for IPv6 so we
192 * could do with it working on IPv4
193 * User settable/learned rtt/max window/mtu
194 * Cope with MTU/device switches when retransmitting in tcp.
195 * Fix the window handling to use PR's new code.
196 *
197 * Change the fundamental structure to a single send queue maintained
198 * by TCP (removing the bogus ip stuff [thus fixing mtu drops on
199 * active routes too]). Cut the queue off in tcp_retransmit/
200 * tcp_transmit.
201 * Change the receive queue to assemble as it goes. This lets us
202 * dispose of most of tcp_sequence, half of tcp_ack and chunks of
203 * tcp_data/tcp_read as well as the window shrink crud.
204 * Separate out duplicated code - tcp_alloc_skb, tcp_build_ack
205 * tcp_queue_skb seem obvious routines to extract.
206 *
207 * This program is free software; you can redistribute it and/or
208 * modify it under the terms of the GNU General Public License
209 * as published by the Free Software Foundation; either version
210 * 2 of the License, or(at your option) any later version.
211 *
212 * Description of States:
213 *
214 * TCP_SYN_SENT sent a connection request, waiting for ack
215 *
216 * TCP_SYN_RECV received a connection request, sent ack,
217 * waiting for final ack in three-way handshake.
218 *
219 * TCP_ESTABLISHED connection established
220 *
221 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
222 * transmission of remaining buffered data
223 *
224 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
225 * to shutdown
226 *
227 * TCP_CLOSING both sides have shutdown but we still have
228 * data we have to finish sending
229 *
230 * TCP_TIME_WAIT timeout to catch resent junk before entering
231 * closed, can only be entered from FIN_WAIT2
232 * or CLOSING. Required because the other end
233 * may not have gotten our last ACK causing it
234 * to retransmit the data packet (which we ignore)
235 *
236 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
237 * us to finish writing our data and to shutdown
238 * (we have to close() to move on to LAST_ACK)
239 *
240 * TCP_LAST_ACK out side has shutdown after remote has
241 * shutdown. There may still be data in our
242 * buffer that we have to finish sending
243 *
244 * TCP_CLOSE socket is finished
245 */
246
247 /*
248 * RFC1122 status:
249 * NOTE: I'm not going to be doing comments in the code for this one except
250 * for violations and the like. tcp.c is just too big... If I say something
251 * "does?" or "doesn't?", it means I'm not sure, and will have to hash it out
252 * with Alan. -- MS 950903
253 *
254 * Use of PSH (4.2.2.2)
255 * MAY aggregate data sent without the PSH flag. (does)
256 * MAY queue data recieved without the PSH flag. (does)
257 * SHOULD collapse successive PSH flags when it packetizes data. (doesn't)
258 * MAY implement PSH on send calls. (doesn't, thus:)
259 * MUST NOT buffer data indefinitely (doesn't [1 second])
260 * MUST set PSH on last segment (does)
261 * MAY pass received PSH to application layer (doesn't)
262 * SHOULD send maximum-sized segment whenever possible. (almost always does)
263 *
264 * Window Size (4.2.2.3, 4.2.2.16)
265 * MUST treat window size as an unsigned number (does)
266 * SHOULD treat window size as a 32-bit number (does not)
267 * MUST NOT shrink window once it is offered (does not normally)
268 *
269 * Urgent Pointer (4.2.2.4)
270 * **MUST point urgent pointer to last byte of urgent data (not right
271 * after). (doesn't, to be like BSD)
272 * MUST inform application layer asynchronously of incoming urgent
273 * data. (does)
274 * MUST provide application with means of determining the amount of
275 * urgent data pending. (does)
276 * **MUST support urgent data sequence of arbitrary length. (doesn't, but
277 * it's sort of tricky to fix, as urg_ptr is a 16-bit quantity)
278 * [Follows BSD 1 byte of urgent data]
279 *
280 * TCP Options (4.2.2.5)
281 * MUST be able to recieve TCP options in any segment. (does)
282 * MUST ignore unsupported options (does)
283 *
284 * Maximum Segment Size Option (4.2.2.6)
285 * MUST implement both sending and receiving MSS. (does)
286 * SHOULD send an MSS with every SYN where recieve MSS != 536 (MAY send
287 * it always). (does, even when MSS == 536, which is legal)
288 * MUST assume MSS == 536 if no MSS received at connection setup (does)
289 * MUST calculate "effective send MSS" correctly:
290 * min(physical_MTU, remote_MSS+20) - sizeof(tcphdr) - sizeof(ipopts)
291 * (does - but allows operator override)
292 *
293 * TCP Checksum (4.2.2.7)
294 * MUST generate and check TCP checksum. (does)
295 *
296 * Initial Sequence Number Selection (4.2.2.8)
297 * MUST use the RFC 793 clock selection mechanism. (doesn't, but it's
298 * OK: RFC 793 specifies a 250KHz clock, while we use 1MHz, which is
299 * necessary for 10Mbps networks - and harder than BSD to spoof!)
300 *
301 * Simultaneous Open Attempts (4.2.2.10)
302 * MUST support simultaneous open attempts (does)
303 *
304 * Recovery from Old Duplicate SYN (4.2.2.11)
305 * MUST keep track of active vs. passive open (does)
306 *
307 * RST segment (4.2.2.12)
308 * SHOULD allow an RST segment to contain data (does, but doesn't do
309 * anything with it, which is standard)
310 *
311 * Closing a Connection (4.2.2.13)
312 * MUST inform application of whether connectin was closed by RST or
313 * normal close. (does)
314 * MAY allow "half-duplex" close (treat connection as closed for the
315 * local app, even before handshake is done). (does)
316 * MUST linger in TIME_WAIT for 2 * MSL (does)
317 *
318 * Retransmission Timeout (4.2.2.15)
319 * MUST implement Jacobson's slow start and congestion avoidance
320 * stuff. (does)
321 *
322 * Probing Zero Windows (4.2.2.17)
323 * MUST support probing of zero windows. (does)
324 * MAY keep offered window closed indefinitely. (does)
325 * MUST allow remote window to stay closed indefinitely. (does)
326 *
327 * Passive Open Calls (4.2.2.18)
328 * MUST NOT let new passive open affect other connections. (doesn't)
329 * MUST support passive opens (LISTENs) concurrently. (does)
330 *
331 * Time to Live (4.2.2.19)
332 * MUST make TCP TTL configurable. (does - IP_TTL option)
333 *
334 * Event Processing (4.2.2.20)
335 * SHOULD queue out-of-order segments. (does)
336 * MUST aggregate ACK segments whenever possible. (does but badly)
337 *
338 * Retransmission Timeout Calculation (4.2.3.1)
339 * MUST implement Karn's algorithm and Jacobson's algorithm for RTO
340 * calculation. (does, or at least explains them in the comments 8*b)
341 * SHOULD initialize RTO to 0 and RTT to 3. (does)
342 *
343 * When to Send an ACK Segment (4.2.3.2)
344 * SHOULD implement delayed ACK. (does not)
345 * MUST keep ACK delay < 0.5 sec. (N/A)
346 *
347 * When to Send a Window Update (4.2.3.3)
348 * MUST implement receiver-side SWS. (does)
349 *
350 * When to Send Data (4.2.3.4)
351 * MUST implement sender-side SWS. (does - imperfectly)
352 * SHOULD implement Nagle algorithm. (does)
353 *
354 * TCP Connection Failures (4.2.3.5)
355 * MUST handle excessive retransmissions "properly" (see the RFC). (does)
356 * SHOULD inform application layer of soft errors. (doesn't)
357 *
358 * TCP Keep-Alives (4.2.3.6)
359 * MAY provide keep-alives. (does)
360 * MUST make keep-alives configurable on a per-connection basis. (does)
361 * MUST default to no keep-alives. (does)
362 * **MUST make keep-alive interval configurable. (doesn't)
363 * **MUST make default keep-alive interval > 2 hours. (doesn't)
364 * MUST NOT interpret failure to ACK keep-alive packet as dead
365 * connection. (doesn't)
366 * SHOULD send keep-alive with no data. (does)
367 *
368 * TCP Multihoming (4.2.3.7)
369 * MUST get source address from IP layer before sending first
370 * SYN. (does)
371 * MUST use same local address for all segments of a connection. (does)
372 *
373 * IP Options (4.2.3.8)
374 * (I don't think the IP layer sees the IP options, yet.)
375 * MUST ignore unsupported IP options. (does, I guess 8*b)
376 * MAY support Time Stamp and Record Route. (doesn't)
377 * **MUST allow application to specify a source route. (doesn't?)
378 * **MUST allow receieved Source Route option to set route for all future
379 * segments on this connection. (doesn't, not that I think it's a
380 * huge problem)
381 *
382 * ICMP messages (4.2.3.9)
383 * MUST act on ICMP errors. (does)
384 * MUST slow transmission upon receipt of a Source Quench. (does)
385 * MUST NOT abort connection upon receipt of soft Destination
386 * Unreachables (0, 1, 5), Time Exceededs and Parameter
387 * Problems. (doesn't)
388 * SHOULD report soft Destination Unreachables etc. to the
389 * application. (doesn't)
390 * SHOULD abort connection upon receipt of hard Destination Unreachable
391 * messages (2, 3, 4). (does)
392 *
393 * Remote Address Validation (4.2.3.10)
394 * MUST reject as an error OPEN for invalid remote IP address. (does)
395 * MUST ignore SYN with invalid source address. (does)
396 * MUST silently discard incoming SYN for broadcast/multicast
397 * address. (does)
398 *
399 * Asynchronous Reports (4.2.4.1)
400 * **MUST provide mechanism for reporting soft errors to application
401 * layer. (doesn't)
402 *
403 * Type of Service (4.2.4.2)
404 * MUST allow application layer to set Type of Service. (does IP_TOS)
405 *
406 * (Whew. -- MS 950903)
407 **/
408
409 #include <linux/types.h>
410 #include <linux/sched.h>
411 #include <linux/mm.h>
412 #include <linux/time.h>
413 #include <linux/string.h>
414 #include <linux/config.h>
415 #include <linux/socket.h>
416 #include <linux/sockios.h>
417 #include <linux/termios.h>
418 #include <linux/in.h>
419 #include <linux/fcntl.h>
420 #include <linux/inet.h>
421 #include <linux/netdevice.h>
422 #include <net/snmp.h>
423 #include <net/ip.h>
424 #include <net/protocol.h>
425 #include <net/icmp.h>
426 #include <net/tcp.h>
427 #include <net/arp.h>
428 #include <linux/skbuff.h>
429 #include <net/sock.h>
430 #include <net/route.h>
431 #include <linux/errno.h>
432 #include <linux/timer.h>
433 #include <asm/system.h>
434 #include <asm/segment.h>
435 #include <linux/mm.h>
436 #include <net/checksum.h>
437
438 /*
439 * The MSL timer is the 'normal' timer.
440 */
441
442 #define reset_msl_timer(x,y,z) reset_timer(x,y,z)
443
444 #define SEQ_TICK 3
445 unsigned long seq_offset;
446 struct tcp_mib tcp_statistics;
447
448 /*
449 * Cached last hit socket
450 */
451
452 volatile unsigned long th_cache_saddr,th_cache_daddr;
453 volatile unsigned short th_cache_dport, th_cache_sport;
454 volatile struct sock *th_cache_sk;
455
456 void tcp_cache_zap(void)
/* ![[next]](../icons/right.png)
![[first]](../icons/n_first.png)
![[top]](../icons/top.png)
*/
457 {
458 unsigned long flags;
459 save_flags(flags);
460 cli();
461 th_cache_saddr=0;
462 th_cache_daddr=0;
463 th_cache_dport=0;
464 th_cache_sport=0;
465 th_cache_sk=NULL;
466 restore_flags(flags);
467 }
468
469 static void tcp_close(struct sock *sk, int timeout);
470
471
472 /*
473 * The less said about this the better, but it works and will do for 1.2
474 */
475
476 static struct wait_queue *master_select_wakeup;
477
478 static __inline__ int min(unsigned int a, unsigned int b)
/* ![[previous]](../icons/left.png)
*/
479 {
480 if (a < b)
481 return(a);
482 return(b);
483 }
484
485 #undef STATE_TRACE
486
487 #ifdef STATE_TRACE
488 static char *statename[]={
489 "Unused","Established","Syn Sent","Syn Recv",
490 "Fin Wait 1","Fin Wait 2","Time Wait", "Close",
491 "Close Wait","Last ACK","Listen","Closing"
492 };
493 #endif
494
495 static __inline__ void tcp_set_state(struct sock *sk, int state)
/* */
496 {
497 if(sk->state==TCP_ESTABLISHED)
498 tcp_statistics.TcpCurrEstab--;
499 #ifdef STATE_TRACE
500 if(sk->debug)
501 printk("TCP sk=%p, State %s -> %s\n",sk, statename[sk->state],statename[state]);
502 #endif
503 /* This is a hack but it doesn't occur often and it's going to
504 be a real to fix nicely */
505
506 if(state==TCP_ESTABLISHED && sk->state==TCP_SYN_RECV)
507 {
508 wake_up_interruptible(&master_select_wakeup);
509 }
510 sk->state=state;
511 if(state==TCP_ESTABLISHED)
512 tcp_statistics.TcpCurrEstab++;
513 if(sk->state==TCP_CLOSE)
514 tcp_cache_zap();
515 }
516
517 /*
518 * This routine picks a TCP windows for a socket based on
519 * the following constraints
520 *
521 * 1. The window can never be shrunk once it is offered (RFC 793)
522 * 2. We limit memory per socket
523 *
524 * For now we use NET2E3's heuristic of offering half the memory
525 * we have handy. All is not as bad as this seems however because
526 * of two things. Firstly we will bin packets even within the window
527 * in order to get the data we are waiting for into the memory limit.
528 * Secondly we bin common duplicate forms at receive time
529 * Better heuristics welcome
530 */
531
532 int tcp_select_window(struct sock *sk)
/* */
533 {
534 int new_window = sock_rspace(sk);
535
536 if(sk->window_clamp)
537 new_window=min(sk->window_clamp,new_window);
538 /*
539 * Two things are going on here. First, we don't ever offer a
540 * window less than min(sk->mss, MAX_WINDOW/2). This is the
541 * receiver side of SWS as specified in RFC1122.
542 * Second, we always give them at least the window they
543 * had before, in order to avoid retracting window. This
544 * is technically allowed, but RFC1122 advises against it and
545 * in practice it causes trouble.
546 *
547 * Fixme: This doesn't correctly handle the case where
548 * new_window > sk->window but not by enough to allow for the
549 * shift in sequence space.
550 */
551 if (new_window < min(sk->mss, MAX_WINDOW/2) || new_window < sk->window)
552 return(sk->window);
553 return(new_window);
554 }
555
556 /*
557 * Find someone to 'accept'. Must be called with
558 * sk->inuse=1 or cli()
559 */
560
561 static struct sk_buff *tcp_find_established(struct sock *s)
/* */
562 {
563 struct sk_buff *p=skb_peek(&s->receive_queue);
564 if(p==NULL)
565 return NULL;
566 do
567 {
568 if(p->sk->state == TCP_ESTABLISHED || p->sk->state >= TCP_FIN_WAIT1)
569 return p;
570 p=p->next;
571 }
572 while(p!=(struct sk_buff *)&s->receive_queue);
573 return NULL;
574 }
575
576 /*
577 * Remove a completed connection and return it. This is used by
578 * tcp_accept() to get connections from the queue.
579 */
580
581 static struct sk_buff *tcp_dequeue_established(struct sock *s)
/* */
582 {
583 struct sk_buff *skb;
584 unsigned long flags;
585 save_flags(flags);
586 cli();
587 skb=tcp_find_established(s);
588 if(skb!=NULL)
589 skb_unlink(skb); /* Take it off the queue */
590 restore_flags(flags);
591 return skb;
592 }
593
594 /*
595 * This routine closes sockets which have been at least partially
596 * opened, but not yet accepted. Currently it is only called by
597 * tcp_close, and timeout mirrors the value there.
598 */
599
600 static void tcp_close_pending (struct sock *sk)
/* */
601 {
602 struct sk_buff *skb;
603
604 while ((skb = skb_dequeue(&sk->receive_queue)) != NULL)
605 {
606 skb->sk->dead=1;
607 tcp_close(skb->sk, 0);
608 kfree_skb(skb, FREE_READ);
609 }
610 return;
611 }
612
613 /*
614 * Enter the time wait state.
615 */
616
617 static void tcp_time_wait(struct sock *sk)
/* */
618 {
619 tcp_set_state(sk,TCP_TIME_WAIT);
620 sk->shutdown = SHUTDOWN_MASK;
621 if (!sk->dead)
622 sk->state_change(sk);
623 reset_msl_timer(sk, TIME_CLOSE, TCP_TIMEWAIT_LEN);
624 }
625
626 /*
627 * A socket has timed out on its send queue and wants to do a
628 * little retransmitting. Currently this means TCP.
629 */
630
631 void tcp_do_retransmit(struct sock *sk, int all)
/* */
632 {
633 struct sk_buff * skb;
634 struct proto *prot;
635 struct device *dev;
636 int ct=0;
637 struct rtable *rt;
638
639 prot = sk->prot;
640 skb = sk->send_head;
641
642 while (skb != NULL)
643 {
644 struct tcphdr *th;
645 struct iphdr *iph;
646 int size;
647
648 dev = skb->dev;
649 IS_SKB(skb);
650 skb->when = jiffies;
651
652 /*
653 * Discard the surplus MAC header
654 */
655
656 skb_pull(skb,((unsigned char *)skb->ip_hdr)-skb->data);
657
658 /*
659 * In general it's OK just to use the old packet. However we
660 * need to use the current ack and window fields. Urg and
661 * urg_ptr could possibly stand to be updated as well, but we
662 * don't keep the necessary data. That shouldn't be a problem,
663 * if the other end is doing the right thing. Since we're
664 * changing the packet, we have to issue a new IP identifier.
665 */
666
667 iph = (struct iphdr *)skb->data;
668 th = (struct tcphdr *)(((char *)iph) + (iph->ihl << 2));
669 size = ntohs(iph->tot_len) - (iph->ihl<<2);
670
671 /*
672 * Note: We ought to check for window limits here but
673 * currently this is done (less efficiently) elsewhere.
674 */
675
676 iph->id = htons(ip_id_count++);
677 ip_send_check(iph);
678
679 /*
680 * Put a MAC header back on (may cause ARPing)
681 */
682
683 if(skb->localroute)
684 rt=ip_rt_local(iph->daddr,NULL,NULL);
685 else
686 rt=ip_rt_route(iph->daddr,NULL,NULL);
687
688 if(rt==NULL) /* Deep poo */
689 {
690 if(skb->sk)
691 {
692 skb->sk->err=ENETUNREACH;
693 skb->sk->error_report(skb->sk);
694 }
695 }
696 else
697 {
698 dev=rt->rt_dev;
699 skb->raddr=rt->rt_gateway;
700 if(skb->raddr==0)
701 skb->raddr=iph->daddr;
702 skb->dev=dev;
703 skb->arp=1;
704 if(dev->hard_header)
705 {
706 if(dev->hard_header(skb, dev, ETH_P_IP, NULL, NULL, skb->len)<0)
707 skb->arp=0;
708 }
709
710 /*
711 * This is not the right way to handle this. We have to
712 * issue an up to date window and ack report with this
713 * retransmit to keep the odd buggy tcp that relies on
714 * the fact BSD does this happy.
715 * We don't however need to recalculate the entire
716 * checksum, so someone wanting a small problem to play
717 * with might like to implement RFC1141/RFC1624 and speed
718 * this up by avoiding a full checksum.
719 */
720
721 th->ack_seq = ntohl(sk->acked_seq);
722 th->window = ntohs(tcp_select_window(sk));
723 tcp_send_check(th, sk->saddr, sk->daddr, size, sk);
724
725 /*
726 * If the interface is (still) up and running, kick it.
727 */
728
729 if (dev->flags & IFF_UP)
730 {
731 /*
732 * If the packet is still being sent by the device/protocol
733 * below then don't retransmit. This is both needed, and good -
734 * especially with connected mode AX.25 where it stops resends
735 * occurring of an as yet unsent anyway frame!
736 * We still add up the counts as the round trip time wants
737 * adjusting.
738 */
739 if (sk && !skb_device_locked(skb))
740 {
741 /* Remove it from any existing driver queue first! */
742 skb_unlink(skb);
743 /* Now queue it */
744 ip_statistics.IpOutRequests++;
745 dev_queue_xmit(skb, dev, sk->priority);
746 }
747 }
748 }
749
750 /*
751 * Count retransmissions
752 */
753
754 ct++;
755 sk->prot->retransmits ++;
756 tcp_statistics.TcpRetransSegs++;
757
758
759 /*
760 * Only one retransmit requested.
761 */
762
763 if (!all)
764 break;
765
766 /*
767 * This should cut it off before we send too many packets.
768 */
769
770 if (ct >= sk->cong_window)
771 break;
772 skb = skb->link3;
773 }
774 }
775
776 /*
777 * Reset the retransmission timer
778 */
779
780 static void reset_xmit_timer(struct sock *sk, int why, unsigned long when)
/* */
781 {
782 del_timer(&sk->retransmit_timer);
783 sk->ip_xmit_timeout = why;
784 if((int)when < 0)
785 {
786 when=3;
787 printk("Error: Negative timer in xmit_timer\n");
788 }
789 sk->retransmit_timer.expires=jiffies+when;
790 add_timer(&sk->retransmit_timer);
791 }
792
793 /*
794 * This is the normal code called for timeouts. It does the retransmission
795 * and then does backoff. tcp_do_retransmit is separated out because
796 * tcp_ack needs to send stuff from the retransmit queue without
797 * initiating a backoff.
798 */
799
800
801 void tcp_retransmit_time(struct sock *sk, int all)
/* */
802 {
803 tcp_do_retransmit(sk, all);
804
805 /*
806 * Increase the timeout each time we retransmit. Note that
807 * we do not increase the rtt estimate. rto is initialized
808 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests
809 * that doubling rto each time is the least we can get away with.
810 * In KA9Q, Karn uses this for the first few times, and then
811 * goes to quadratic. netBSD doubles, but only goes up to *64,
812 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is
813 * defined in the protocol as the maximum possible RTT. I guess
814 * we'll have to use something other than TCP to talk to the
815 * University of Mars.
816 *
817 * PAWS allows us longer timeouts and large windows, so once
818 * implemented ftp to mars will work nicely. We will have to fix
819 * the 120 second clamps though!
820 */
821
822 sk->retransmits++;
823 sk->prot->retransmits++;
824 sk->backoff++;
825 sk->rto = min(sk->rto << 1, 120*HZ);
826 reset_xmit_timer(sk, TIME_WRITE, sk->rto);
827 }
828
829
830 /*
831 * A timer event has trigger a tcp retransmit timeout. The
832 * socket xmit queue is ready and set up to send. Because
833 * the ack receive code keeps the queue straight we do
834 * nothing clever here.
835 */
836
837 static void tcp_retransmit(struct sock *sk, int all)
/* */
838 {
839 if (all)
840 {
841 tcp_retransmit_time(sk, all);
842 return;
843 }
844
845 sk->ssthresh = sk->cong_window >> 1; /* remember window where we lost */
846 /* sk->ssthresh in theory can be zero. I guess that's OK */
847 sk->cong_count = 0;
848
849 sk->cong_window = 1;
850
851 /* Do the actual retransmit. */
852 tcp_retransmit_time(sk, all);
853 }
854
855 /*
856 * A write timeout has occurred. Process the after effects.
857 */
858
859 static int tcp_write_timeout(struct sock *sk)
/* */
860 {
861 /*
862 * Look for a 'soft' timeout.
863 */
864 if ((sk->state == TCP_ESTABLISHED && sk->retransmits && !(sk->retransmits & 7))
865 || (sk->state != TCP_ESTABLISHED && sk->retransmits > TCP_RETR1))
866 {
867 /*
868 * Attempt to recover if arp has changed (unlikely!) or
869 * a route has shifted (not supported prior to 1.3).
870 */
871 arp_destroy (sk->daddr, 0);
872 /*ip_route_check (sk->daddr);*/
873 }
874
875 /*
876 * Have we tried to SYN too many times (repent repent 8))
877 */
878
879 if(sk->retransmits > TCP_SYN_RETRIES && sk->state==TCP_SYN_SENT)
880 {
881 sk->err=ETIMEDOUT;
882 sk->error_report(sk);
883 del_timer(&sk->retransmit_timer);
884 tcp_statistics.TcpAttemptFails++; /* Is this right ??? - FIXME - */
885 tcp_set_state(sk,TCP_CLOSE);
886 /* Don't FIN, we got nothing back */
887 release_sock(sk);
888 return 0;
889 }
890 /*
891 * Has it gone just too far ?
892 */
893 if (sk->retransmits > TCP_RETR2)
894 {
895 sk->err = ETIMEDOUT;
896 sk->error_report(sk);
897 del_timer(&sk->retransmit_timer);
898 /*
899 * Time wait the socket
900 */
901 if (sk->state == TCP_FIN_WAIT1 || sk->state == TCP_FIN_WAIT2 || sk->state == TCP_CLOSING )
902 {
903 tcp_set_state(sk,TCP_TIME_WAIT);
904 reset_msl_timer (sk, TIME_CLOSE, TCP_TIMEWAIT_LEN);
905 }
906 else
907 {
908 /*
909 * Clean up time.
910 */
911 tcp_set_state(sk, TCP_CLOSE);
912 release_sock(sk);
913 return 0;
914 }
915 }
916 return 1;
917 }
918
919 /*
920 * The TCP retransmit timer. This lacks a few small details.
921 *
922 * 1. An initial rtt timeout on the probe0 should cause what we can
923 * of the first write queue buffer to be split and sent.
924 * 2. On a 'major timeout' as defined by RFC1122 we shouldn't report
925 * ETIMEDOUT if we know an additional 'soft' error caused this.
926 * tcp_err should save a 'soft error' for us.
927 */
928
929 static void retransmit_timer(unsigned long data)
/* */
930 {
931 struct sock *sk = (struct sock*)data;
932 int why = sk->ip_xmit_timeout;
933
934 /*
935 * only process if socket is not in use
936 */
937
938 cli();
939 if (sk->inuse || in_bh)
940 {
941 /* Try again in 1 second */
942 sk->retransmit_timer.expires = jiffies+HZ;
943 add_timer(&sk->retransmit_timer);
944 sti();
945 return;
946 }
947
948 sk->inuse = 1;
949 sti();
950
951 /* Always see if we need to send an ack. */
952
953 if (sk->ack_backlog && !sk->zapped)
954 {
955 sk->prot->read_wakeup (sk);
956 if (! sk->dead)
957 sk->data_ready(sk,0);
958 }
959
960 /* Now we need to figure out why the socket was on the timer. */
961
962 switch (why)
963 {
964 /* Window probing */
965 case TIME_PROBE0:
966 tcp_send_probe0(sk);
967 tcp_write_timeout(sk);
968 break;
969 /* Retransmitting */
970 case TIME_WRITE:
971 /* It could be we got here because we needed to send an ack.
972 * So we need to check for that.
973 */
974 {
975 struct sk_buff *skb;
976 unsigned long flags;
977
978 save_flags(flags);
979 cli();
980 skb = sk->send_head;
981 if (!skb)
982 {
983 restore_flags(flags);
984 }
985 else
986 {
987 /*
988 * Kicked by a delayed ack. Reset timer
989 * correctly now
990 */
991 if (jiffies < skb->when + sk->rto)
992 {
993 reset_xmit_timer (sk, TIME_WRITE, skb->when + sk->rto - jiffies);
994 restore_flags(flags);
995 break;
996 }
997 restore_flags(flags);
998 /*
999 * Retransmission
1000 */
1001 sk->retransmits++;
1002 sk->prot->retransmits++;
1003 sk->prot->retransmit (sk, 0);
1004 tcp_write_timeout(sk);
1005 }
1006 break;
1007 }
1008 /* Sending Keepalives */
1009 case TIME_KEEPOPEN:
1010 /*
1011 * this reset_timer() call is a hack, this is not
1012 * how KEEPOPEN is supposed to work.
1013 */
1014 reset_xmit_timer (sk, TIME_KEEPOPEN, TCP_TIMEOUT_LEN);
1015
1016 /* Send something to keep the connection open. */
1017 if (sk->prot->write_wakeup)
1018 sk->prot->write_wakeup (sk);
1019 sk->retransmits++;
1020 sk->prot->retransmits++;
1021 tcp_write_timeout(sk);
1022 break;
1023 default:
1024 printk ("rexmit_timer: timer expired - reason unknown\n");
1025 break;
1026 }
1027 release_sock(sk);
1028 }
1029
1030 /*
1031 * This routine is called by the ICMP module when it gets some
1032 * sort of error condition. If err < 0 then the socket should
1033 * be closed and the error returned to the user. If err > 0
1034 * it's just the icmp type << 8 | icmp code. After adjustment
1035 * header points to the first 8 bytes of the tcp header. We need
1036 * to find the appropriate port.
1037 */
1038
1039 void tcp_err(int type, int code, unsigned char *header, __u32 daddr,
/* */
1040 __u32 saddr, struct inet_protocol *protocol)
1041 {
1042 struct tcphdr *th;
1043 struct sock *sk;
1044 struct iphdr *iph=(struct iphdr *)header;
1045
1046 header+=4*iph->ihl;
1047
1048
1049 th =(struct tcphdr *)header;
1050 sk = get_sock(&tcp_prot, th->source, daddr, th->dest, saddr);
1051
1052 if (sk == NULL)
1053 return;
1054
1055 if (type == ICMP_SOURCE_QUENCH)
1056 {
1057 /*
1058 * FIXME:
1059 * For now we will just trigger a linear backoff.
1060 * The slow start code should cause a real backoff here.
1061 */
1062 if (sk->cong_window > 4)
1063 sk->cong_window--;
1064 return;
1065 }
1066
1067 if (type == ICMP_PARAMETERPROB)
1068 {
1069 sk->err=EPROTO;
1070 sk->error_report(sk);
1071 }
1072
1073 /*
1074 * If we've already connected we will keep trying
1075 * until we time out, or the user gives up.
1076 */
1077
1078 if (code < 13 && (icmp_err_convert[code].fatal || sk->state == TCP_SYN_SENT || sk->state == TCP_SYN_RECV))
1079 {
1080 sk->err = icmp_err_convert[code].errno;
1081 if (sk->state == TCP_SYN_SENT || sk->state == TCP_SYN_RECV)
1082 {
1083 tcp_statistics.TcpAttemptFails++;
1084 tcp_set_state(sk,TCP_CLOSE);
1085 sk->error_report(sk); /* Wake people up to see the error (see connect in sock.c) */
1086 }
1087 }
1088 return;
1089 }
1090
1091
1092 /*
1093 * Walk down the receive queue counting readable data until we hit the end or we find a gap
1094 * in the received data queue (ie a frame missing that needs sending to us). Not
1095 * sorting using two queues as data arrives makes life so much harder.
1096 */
1097
1098 static int tcp_readable(struct sock *sk)
/* */
1099 {
1100 unsigned long counted;
1101 unsigned long amount;
1102 struct sk_buff *skb;
1103 int sum;
1104 unsigned long flags;
1105
1106 if(sk && sk->debug)
1107 printk("tcp_readable: %p - ",sk);
1108
1109 save_flags(flags);
1110 cli();
1111 if (sk == NULL || (skb = skb_peek(&sk->receive_queue)) == NULL)
1112 {
1113 restore_flags(flags);
1114 if(sk && sk->debug)
1115 printk("empty\n");
1116 return(0);
1117 }
1118
1119 counted = sk->copied_seq; /* Where we are at the moment */
1120 amount = 0;
1121
1122 /*
1123 * Do until a push or until we are out of data.
1124 */
1125
1126 do
1127 {
1128 if (before(counted, skb->h.th->seq)) /* Found a hole so stops here */
1129 break;
1130 sum = skb->len -(counted - skb->h.th->seq); /* Length - header but start from where we are up to (avoid overlaps) */
1131 if (skb->h.th->syn)
1132 sum++;
1133 if (sum > 0)
1134 { /* Add it up, move on */
1135 amount += sum;
1136 if (skb->h.th->syn)
1137 amount--;
1138 counted += sum;
1139 }
1140 /*
1141 * Don't count urg data ... but do it in the right place!
1142 * Consider: "old_data (ptr is here) URG PUSH data"
1143 * The old code would stop at the first push because
1144 * it counted the urg (amount==1) and then does amount--
1145 * *after* the loop. This means tcp_readable() always
1146 * returned zero if any URG PUSH was in the queue, even
1147 * though there was normal data available. If we subtract
1148 * the urg data right here, we even get it to work for more
1149 * than one URG PUSH skb without normal data.
1150 * This means that select() finally works now with urg data
1151 * in the queue. Note that rlogin was never affected
1152 * because it doesn't use select(); it uses two processes
1153 * and a blocking read(). And the queue scan in tcp_read()
1154 * was correct. Mike <pall@rz.uni-karlsruhe.de>
1155 */
1156 if (skb->h.th->urg)
1157 amount--; /* don't count urg data */
1158 if (amount && skb->h.th->psh) break;
1159 skb = skb->next;
1160 }
1161 while(skb != (struct sk_buff *)&sk->receive_queue);
1162
1163 restore_flags(flags);
1164 if(sk->debug)
1165 printk("got %lu bytes.\n",amount);
1166 return(amount);
1167 }
1168
1169 /*
1170 * LISTEN is a special case for select..
1171 */
1172 static int tcp_listen_select(struct sock *sk, int sel_type, select_table *wait)
/* */
1173 {
1174 if (sel_type == SEL_IN) {
1175 int retval;
1176
1177 sk->inuse = 1;
1178 retval = (tcp_find_established(sk) != NULL);
1179 release_sock(sk);
1180 if (!retval)
1181 select_wait(&master_select_wakeup,wait);
1182 return retval;
1183 }
1184 return 0;
1185 }
1186
1187
1188 /*
1189 * Wait for a TCP event.
1190 *
1191 * Note that we don't need to set "sk->inuse", as the upper select layers
1192 * take care of normal races (between the test and the event) and we don't
1193 * go look at any of the socket buffers directly.
1194 */
1195 static int tcp_select(struct sock *sk, int sel_type, select_table *wait)
/* */
1196 {
1197 if (sk->state == TCP_LISTEN)
1198 return tcp_listen_select(sk, sel_type, wait);
1199
1200 switch(sel_type) {
1201 case SEL_IN:
1202 if (sk->err)
1203 return 1;
1204 if (sk->state == TCP_SYN_SENT || sk->state == TCP_SYN_RECV)
1205 break;
1206
1207 if (sk->shutdown & RCV_SHUTDOWN)
1208 return 1;
1209
1210 if (sk->acked_seq == sk->copied_seq)
1211 break;
1212
1213 if (sk->urg_seq != sk->copied_seq ||
1214 sk->acked_seq != sk->copied_seq+1 ||
1215 sk->urginline || !sk->urg_data)
1216 return 1;
1217 break;
1218
1219 case SEL_OUT:
1220 if (sk->err)
1221 return 1;
1222 if (sk->shutdown & SEND_SHUTDOWN)
1223 return 0;
1224 if (sk->state == TCP_SYN_SENT || sk->state == TCP_SYN_RECV)
1225 break;
1226 /*
1227 * This is now right thanks to a small fix
1228 * by Matt Dillon.
1229 */
1230
1231 if (sock_wspace(sk) < sk->mtu+128+sk->prot->max_header)
1232 break;
1233 return 1;
1234
1235 case SEL_EX:
1236 if (sk->urg_data)
1237 return 1;
1238 break;
1239 }
1240 select_wait(sk->sleep, wait);
1241 return 0;
1242 }
1243
1244 int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
/* */
1245 {
1246 int err;
1247 switch(cmd)
1248 {
1249
1250 case TIOCINQ:
1251 #ifdef FIXME /* FIXME: */
1252 case FIONREAD:
1253 #endif
1254 {
1255 unsigned long amount;
1256
1257 if (sk->state == TCP_LISTEN)
1258 return(-EINVAL);
1259
1260 sk->inuse = 1;
1261 amount = tcp_readable(sk);
1262 release_sock(sk);
1263 err=verify_area(VERIFY_WRITE,(void *)arg, sizeof(int));
1264 if(err)
1265 return err;
1266 put_user(amount, (int *)arg);
1267 return(0);
1268 }
1269 case SIOCATMARK:
1270 {
1271 int answ = sk->urg_data && sk->urg_seq == sk->copied_seq;
1272
1273 err = verify_area(VERIFY_WRITE,(void *) arg, sizeof(int));
1274 if (err)
1275 return err;
1276 put_user(answ,(int *) arg);
1277 return(0);
1278 }
1279 case TIOCOUTQ:
1280 {
1281 unsigned long amount;
1282
1283 if (sk->state == TCP_LISTEN) return(-EINVAL);
1284 amount = sock_wspace(sk);
1285 err=verify_area(VERIFY_WRITE,(void *)arg, sizeof(int));
1286 if(err)
1287 return err;
1288 put_user(amount, (int *)arg);
1289 return(0);
1290 }
1291 default:
1292 return(-EINVAL);
1293 }
1294 }
1295
1296
1297 /*
1298 * This routine computes a TCP checksum.
1299 *
1300 * Modified January 1995 from a go-faster DOS routine by
1301 * Jorge Cwik <jorge@laser.satlink.net>
1302 */
1303
1304 unsigned short tcp_check(struct tcphdr *th, int len,
/* */
1305 unsigned long saddr, unsigned long daddr, unsigned long base)
1306 {
1307 return csum_tcpudp_magic(saddr,daddr,len,IPPROTO_TCP,base);
1308 }
1309
1310
1311
1312 void tcp_send_check(struct tcphdr *th, unsigned long saddr,
/* */
1313 unsigned long daddr, int len, struct sock *sk)
1314 {
1315 th->check = 0;
1316 th->check = tcp_check(th, len, saddr, daddr,
1317 csum_partial((char *)th,len,0));
1318 return;
1319 }
1320
1321 /*
1322 * This is the main buffer sending routine. We queue the buffer
1323 * having checked it is sane seeming.
1324 */
1325
1326 static void tcp_send_skb(struct sock *sk, struct sk_buff *skb)
/* */
1327 {
1328 int size;
1329 struct tcphdr * th = skb->h.th;
1330
1331 /*
1332 * length of packet (not counting length of pre-tcp headers)
1333 */
1334
1335 size = skb->len - ((unsigned char *) th - skb->data);
1336
1337 /*
1338 * Sanity check it..
1339 */
1340
1341 if (size < sizeof(struct tcphdr) || size > skb->len)
1342 {
1343 printk("tcp_send_skb: bad skb (skb = %p, data = %p, th = %p, len = %lu)\n",
1344 skb, skb->data, th, skb->len);
1345 kfree_skb(skb, FREE_WRITE);
1346 return;
1347 }
1348
1349 /*
1350 * If we have queued a header size packet.. (these crash a few
1351 * tcp stacks if ack is not set)
1352 */
1353
1354 if (size == sizeof(struct tcphdr))
1355 {
1356 /* If it's got a syn or fin it's notionally included in the size..*/
1357 if(!th->syn && !th->fin)
1358 {
1359 printk("tcp_send_skb: attempt to queue a bogon.\n");
1360 kfree_skb(skb,FREE_WRITE);
1361 return;
1362 }
1363 }
1364
1365 /*
1366 * Actual processing.
1367 */
1368
1369 tcp_statistics.TcpOutSegs++;
1370 skb->h.seq = ntohl(th->seq) + size - 4*th->doff;
1371
1372 /*
1373 * We must queue if
1374 *
1375 * a) The right edge of this frame exceeds the window
1376 * b) We are retransmitting (Nagle's rule)
1377 * c) We have too many packets 'in flight'
1378 */
1379
1380 if (after(skb->h.seq, sk->window_seq) ||
1381 (sk->retransmits && sk->ip_xmit_timeout == TIME_WRITE) ||
1382 sk->packets_out >= sk->cong_window)
1383 {
1384 /* checksum will be supplied by tcp_write_xmit. So
1385 * we shouldn't need to set it at all. I'm being paranoid */
1386 th->check = 0;
1387 if (skb->next != NULL)
1388 {
1389 printk("tcp_send_partial: next != NULL\n");
1390 skb_unlink(skb);
1391 }
1392 skb_queue_tail(&sk->write_queue, skb);
1393
1394 /*
1395 * If we don't fit we have to start the zero window
1396 * probes. This is broken - we really need to do a partial
1397 * send _first_ (This is what causes the Cisco and PC/TCP
1398 * grief).
1399 */
1400
1401 if (before(sk->window_seq, sk->write_queue.next->h.seq) &&
1402 sk->send_head == NULL && sk->ack_backlog == 0)
1403 reset_xmit_timer(sk, TIME_PROBE0, sk->rto);
1404 }
1405 else
1406 {
1407 /*
1408 * This is going straight out
1409 */
1410
1411 th->ack_seq = ntohl(sk->acked_seq);
1412 th->window = ntohs(tcp_select_window(sk));
1413
1414 tcp_send_check(th, sk->saddr, sk->daddr, size, sk);
1415
1416 sk->sent_seq = sk->write_seq;
1417
1418 /*
1419 * This is mad. The tcp retransmit queue is put together
1420 * by the ip layer. This causes half the problems with
1421 * unroutable FIN's and other things.
1422 */
1423
1424 sk->prot->queue_xmit(sk, skb->dev, skb, 0);
1425
1426 /*
1427 * Set for next retransmit based on expected ACK time.
1428 * FIXME: We set this every time which means our
1429 * retransmits are really about a window behind.
1430 */
1431
1432 reset_xmit_timer(sk, TIME_WRITE, sk->rto);
1433 }
1434 }
1435
1436 /*
1437 * Locking problems lead us to a messy situation where we can have
1438 * multiple partially complete buffers queued up. This is really bad
1439 * as we don't want to be sending partial buffers. Fix this with
1440 * a semaphore or similar to lock tcp_write per socket.
1441 *
1442 * These routines are pretty self descriptive.
1443 */
1444
1445 struct sk_buff * tcp_dequeue_partial(struct sock * sk)
/* */
1446 {
1447 struct sk_buff * skb;
1448 unsigned long flags;
1449
1450 save_flags(flags);
1451 cli();
1452 skb = sk->partial;
1453 if (skb) {
1454 sk->partial = NULL;
1455 del_timer(&sk->partial_timer);
1456 }
1457 restore_flags(flags);
1458 return skb;
1459 }
1460
1461 /*
1462 * Empty the partial queue
1463 */
1464
1465 static void tcp_send_partial(struct sock *sk)
/* */
1466 {
1467 struct sk_buff *skb;
1468
1469 if (sk == NULL)
1470 return;
1471 while ((skb = tcp_dequeue_partial(sk)) != NULL)
1472 tcp_send_skb(sk, skb);
1473 }
1474
1475 /*
1476 * Queue a partial frame
1477 */
1478
1479 void tcp_enqueue_partial(struct sk_buff * skb, struct sock * sk)
/* */
1480 {
1481 struct sk_buff * tmp;
1482 unsigned long flags;
1483
1484 save_flags(flags);
1485 cli();
1486 tmp = sk->partial;
1487 if (tmp)
1488 del_timer(&sk->partial_timer);
1489 sk->partial = skb;
1490 init_timer(&sk->partial_timer);
1491 /*
1492 * Wait up to 1 second for the buffer to fill.
1493 */
1494 sk->partial_timer.expires = jiffies+HZ;
1495 sk->partial_timer.function = (void (*)(unsigned long)) tcp_send_partial;
1496 sk->partial_timer.data = (unsigned long) sk;
1497 add_timer(&sk->partial_timer);
1498 restore_flags(flags);
1499 if (tmp)
1500 tcp_send_skb(sk, tmp);
1501 }
1502
1503
1504 /*
1505 * This routine sends an ack and also updates the window.
1506 */
1507
1508 static void tcp_send_ack(u32 sequence, u32 ack,
/* */
1509 struct sock *sk,
1510 struct tcphdr *th, unsigned long daddr)
1511 {
1512 struct sk_buff *buff;
1513 struct tcphdr *t1;
1514 struct device *dev = NULL;
1515 int tmp;
1516
1517 if(sk->zapped)
1518 return; /* We have been reset, we may not send again */
1519
1520 /*
1521 * We need to grab some memory, and put together an ack,
1522 * and then put it into the queue to be sent.
1523 */
1524
1525 buff = sock_wmalloc(sk, MAX_ACK_SIZE, 1, GFP_ATOMIC);
1526 if (buff == NULL)
1527 {
1528 /*
1529 * Force it to send an ack. We don't have to do this
1530 * (ACK is unreliable) but it's much better use of
1531 * bandwidth on slow links to send a spare ack than
1532 * resend packets.
1533 */
1534
1535 sk->ack_backlog++;
1536 if (sk->ip_xmit_timeout != TIME_WRITE && tcp_connected(sk->state))
1537 {
1538 reset_xmit_timer(sk, TIME_WRITE, HZ);
1539 }
1540 return;
1541 }
1542
1543 /*
1544 * Assemble a suitable TCP frame
1545 */
1546
1547 buff->sk = sk;
1548 buff->localroute = sk->localroute;
1549
1550 /*
1551 * Put in the IP header and routing stuff.
1552 */
1553
1554 tmp = sk->prot->build_header(buff, sk->saddr, daddr, &dev,
1555 IPPROTO_TCP, sk->opt, MAX_ACK_SIZE,sk->ip_tos,sk->ip_ttl);
1556 if (tmp < 0)
1557 {
1558 buff->free = 1;
1559 sock_wfree(sk, buff);
1560 return;
1561 }
1562 t1 =(struct tcphdr *)skb_put(buff,sizeof(struct tcphdr));
1563
1564 memcpy(t1, th, sizeof(*t1));
1565
1566 /*
1567 * Swap the send and the receive.
1568 */
1569
1570 t1->dest = th->source;
1571 t1->source = th->dest;
1572 t1->seq = ntohl(sequence);
1573 t1->ack = 1;
1574 sk->window = tcp_select_window(sk);
1575 t1->window = ntohs(sk->window);
1576 t1->res1 = 0;
1577 t1->res2 = 0;
1578 t1->rst = 0;
1579 t1->urg = 0;
1580 t1->syn = 0;
1581 t1->psh = 0;
1582 t1->fin = 0;
1583
1584 /*
1585 * If we have nothing queued for transmit and the transmit timer
1586 * is on we are just doing an ACK timeout and need to switch
1587 * to a keepalive.
1588 */
1589
1590 if (ack == sk->acked_seq)
1591 {
1592 sk->ack_backlog = 0;
1593 sk->bytes_rcv = 0;
1594 sk->ack_timed = 0;
1595 if (sk->send_head == NULL && skb_peek(&sk->write_queue) == NULL
1596 && sk->ip_xmit_timeout == TIME_WRITE)
1597 {
1598 if(sk->keepopen) {
1599 reset_xmit_timer(sk,TIME_KEEPOPEN,TCP_TIMEOUT_LEN);
1600 } else {
1601 delete_timer(sk);
1602 }
1603 }
1604 }
1605
1606 /*
1607 * Fill in the packet and send it
1608 */
1609
1610 t1->ack_seq = ntohl(ack);
1611 t1->doff = sizeof(*t1)/4;
1612 tcp_send_check(t1, sk->saddr, daddr, sizeof(*t1), sk);
1613 if (sk->debug)
1614 printk("\rtcp_ack: seq %x ack %x\n", sequence, ack);
1615 tcp_statistics.TcpOutSegs++;
1616 sk->prot->queue_xmit(sk, dev, buff, 1);
1617 }
1618
1619
1620 /*
1621 * This routine builds a generic TCP header.
1622 */
1623
1624 extern __inline int tcp_build_header(struct tcphdr *th, struct sock *sk, int push)
/* */
1625 {
1626
1627 memcpy(th,(void *) &(sk->dummy_th), sizeof(*th));
1628 th->seq = htonl(sk->write_seq);
1629 th->psh =(push == 0) ? 1 : 0;
1630 th->doff = sizeof(*th)/4;
1631 th->ack = 1;
1632 th->fin = 0;
1633 sk->ack_backlog = 0;
1634 sk->bytes_rcv = 0;
1635 sk->ack_timed = 0;
1636 th->ack_seq = htonl(sk->acked_seq);
1637 sk->window = tcp_select_window(sk);
1638 th->window = htons(sk->window);
1639
1640 return(sizeof(*th));
1641 }
1642
1643 /*
1644 * This routine copies from a user buffer into a socket,
1645 * and starts the transmit system.
1646 */
1647
1648 static int tcp_sendmsg(struct sock *sk, struct msghdr *msg,
/* */
1649 int len, int nonblock, int flags)
1650 {
1651 int copied = 0;
1652 int copy;
1653 int tmp;
1654 int seglen;
1655 int iovct=0;
1656 struct sk_buff *skb;
1657 struct sk_buff *send_tmp;
1658 struct proto *prot;
1659 struct device *dev = NULL;
1660 unsigned char *from;
1661
1662 /*
1663 * Do sanity checking for sendmsg/sendto/send
1664 */
1665
1666 if (flags & ~(MSG_OOB|MSG_DONTROUTE))
1667 return -EINVAL;
1668 if (msg->msg_name)
1669 {
1670 struct sockaddr_in *addr=(struct sockaddr_in *)msg->msg_name;
1671 if(sk->state == TCP_CLOSE)
1672 return -ENOTCONN;
1673 if (msg->msg_namelen < sizeof(*addr))
1674 return -EINVAL;
1675 if (addr->sin_family && addr->sin_family != AF_INET)
1676 return -EINVAL;
1677 if (addr->sin_port != sk->dummy_th.dest)
1678 return -EISCONN;
1679 if (addr->sin_addr.s_addr != sk->daddr)
1680 return -EISCONN;
1681 }
1682
1683 /*
1684 * Ok commence sending
1685 */
1686
1687 while(iovct<msg->msg_iovlen)
1688 {
1689 seglen=msg->msg_iov[iovct].iov_len;
1690 from=msg->msg_iov[iovct++].iov_base;
1691 sk->inuse=1;
1692 prot = sk->prot;
1693 while(seglen > 0)
1694 {
1695 if (sk->err)
1696 { /* Stop on an error */
1697 release_sock(sk);
1698 if (copied)
1699 return(copied);
1700 tmp = -sk->err;
1701 sk->err = 0;
1702 return(tmp);
1703 }
1704
1705 /*
1706 * First thing we do is make sure that we are established.
1707 */
1708
1709 if (sk->shutdown & SEND_SHUTDOWN)
1710 {
1711 release_sock(sk);
1712 sk->err = EPIPE;
1713 if (copied)
1714 return(copied);
1715 sk->err = 0;
1716 return(-EPIPE);
1717 }
1718
1719 /*
1720 * Wait for a connection to finish.
1721 */
1722
1723 while(sk->state != TCP_ESTABLISHED && sk->state != TCP_CLOSE_WAIT)
1724 {
1725 if (sk->err)
1726 {
1727 release_sock(sk);
1728 if (copied)
1729 return(copied);
1730 tmp = -sk->err;
1731 sk->err = 0;
1732 return(tmp);
1733 }
1734
1735 if (sk->state != TCP_SYN_SENT && sk->state != TCP_SYN_RECV)
1736 {
1737 release_sock(sk);
1738 if (copied)
1739 return(copied);
1740
1741 if (sk->err)
1742 {
1743 tmp = -sk->err;
1744 sk->err = 0;
1745 return(tmp);
1746 }
1747
1748 if (sk->keepopen)
1749 {
1750 send_sig(SIGPIPE, current, 0);
1751 }
1752 return(-EPIPE);
1753 }
1754
1755 if (nonblock || copied)
1756 {
1757 release_sock(sk);
1758 if (copied)
1759 return(copied);
1760 return(-EAGAIN);
1761 }
1762
1763 release_sock(sk);
1764 cli();
1765
1766 if (sk->state != TCP_ESTABLISHED &&
1767 sk->state != TCP_CLOSE_WAIT && sk->err == 0)
1768 {
1769 interruptible_sleep_on(sk->sleep);
1770 if (current->signal & ~current->blocked)
1771 {
1772 sti();
1773 if (copied)
1774 return(copied);
1775 return(-ERESTARTSYS);
1776 }
1777 }
1778 sk->inuse = 1;
1779 sti();
1780 }
1781
1782 /*
1783 * The following code can result in copy <= if sk->mss is ever
1784 * decreased. It shouldn't be. sk->mss is min(sk->mtu, sk->max_window).
1785 * sk->mtu is constant once SYN processing is finished. I.e. we
1786 * had better not get here until we've seen his SYN and at least one
1787 * valid ack. (The SYN sets sk->mtu and the ack sets sk->max_window.)
1788 * But ESTABLISHED should guarantee that. sk->max_window is by definition
1789 * non-decreasing. Note that any ioctl to set user_mss must be done
1790 * before the exchange of SYN's. If the initial ack from the other
1791 * end has a window of 0, max_window and thus mss will both be 0.
1792 */
1793
1794 /*
1795 * Now we need to check if we have a half built packet.
1796 */
1797
1798 if ((skb = tcp_dequeue_partial(sk)) != NULL)
1799 {
1800 int hdrlen;
1801
1802 /* IP header + TCP header */
1803 hdrlen = ((unsigned long)skb->h.th - (unsigned long)skb->data)
1804 + sizeof(struct tcphdr);
1805
1806 /* Add more stuff to the end of skb->len */
1807 if (!(flags & MSG_OOB))
1808 {
1809 copy = min(sk->mss - (skb->len - hdrlen), len);
1810 /* FIXME: this is really a bug. */
1811 if (copy <= 0)
1812 {
1813 printk("TCP: **bug**: \"copy\" <= 0!!\n");
1814 copy = 0;
1815 }
1816 memcpy_fromfs(skb_put(skb,copy), from, copy);
1817 from += copy;
1818 copied += copy;
1819 len -= copy;
1820 sk->write_seq += copy;
1821 seglen -= copy;
1822 }
1823 if ((skb->len - hdrlen) >= sk->mss ||
1824 (flags & MSG_OOB) || !sk->packets_out)
1825 tcp_send_skb(sk, skb);
1826 else
1827 tcp_enqueue_partial(skb, sk);
1828 continue;
1829 }
1830
1831 /*
1832 * We also need to worry about the window.
1833 * If window < 1/2 the maximum window we've seen from this
1834 * host, don't use it. This is sender side
1835 * silly window prevention, as specified in RFC1122.
1836 * (Note that this is different than earlier versions of
1837 * SWS prevention, e.g. RFC813.). What we actually do is
1838 * use the whole MSS. Since the results in the right
1839 * edge of the packet being outside the window, it will
1840 * be queued for later rather than sent.
1841 */
1842
1843 copy = sk->window_seq - sk->write_seq;
1844 if (copy <= 0 || copy < (sk->max_window >> 1) || copy > sk->mss)
1845 copy = sk->mss;
1846 if (copy > len)
1847 copy = len;
1848
1849 /*
1850 * We should really check the window here also.
1851 */
1852
1853 send_tmp = NULL;
1854 if (copy < sk->mss && !(flags & MSG_OOB))
1855 {
1856 /*
1857 * We will release the socket in case we sleep here.
1858 */
1859 release_sock(sk);
1860 /*
1861 * NB: following must be mtu, because mss can be increased.
1862 * mss is always <= mtu
1863 */
1864 skb = sock_wmalloc(sk, sk->mtu + 128 + prot->max_header + 15, 0, GFP_KERNEL);
1865 sk->inuse = 1;
1866 send_tmp = skb;
1867 }
1868 else
1869 {
1870 /*
1871 * We will release the socket in case we sleep here.
1872 */
1873 release_sock(sk);
1874 skb = sock_wmalloc(sk, copy + prot->max_header + 15 , 0, GFP_KERNEL);
1875 sk->inuse = 1;
1876 }
1877
1878 /*
1879 * If we didn't get any memory, we need to sleep.
1880 */
1881
1882 if (skb == NULL)
1883 {
1884 sk->socket->flags |= SO_NOSPACE;
1885 if (nonblock)
1886 {
1887 release_sock(sk);
1888 if (copied)
1889 return(copied);
1890 return(-EAGAIN);
1891 }
1892
1893 /*
1894 * FIXME: here is another race condition.
1895 */
1896
1897 tmp = sk->wmem_alloc;
1898 release_sock(sk);
1899 cli();
1900 /*
1901 * Again we will try to avoid it.
1902 */
1903 if (tmp <= sk->wmem_alloc &&
1904 (sk->state == TCP_ESTABLISHED||sk->state == TCP_CLOSE_WAIT)
1905 && sk->err == 0)
1906 {
1907 sk->socket->flags &= ~SO_NOSPACE;
1908 interruptible_sleep_on(sk->sleep);
1909 if (current->signal & ~current->blocked)
1910 {
1911 sti();
1912 if (copied)
1913 return(copied);
1914 return(-ERESTARTSYS);
1915 }
1916 }
1917 sk->inuse = 1;
1918 sti();
1919 continue;
1920 }
1921
1922 skb->sk = sk;
1923 skb->free = 0;
1924 skb->localroute = sk->localroute|(flags&MSG_DONTROUTE);
1925
1926 /*
1927 * FIXME: we need to optimize this.
1928 * Perhaps some hints here would be good.
1929 */
1930
1931 tmp = prot->build_header(skb, sk->saddr, sk->daddr, &dev,
1932 IPPROTO_TCP, sk->opt, skb->truesize,sk->ip_tos,sk->ip_ttl);
1933 if (tmp < 0 )
1934 {
1935 sock_wfree(sk, skb);
1936 release_sock(sk);
1937 if (copied)
1938 return(copied);
1939 return(tmp);
1940 }
1941 skb->dev = dev;
1942 skb->h.th =(struct tcphdr *)skb_put(skb,sizeof(struct tcphdr));
1943 tmp = tcp_build_header(skb->h.th, sk, len-copy);
1944 if (tmp < 0)
1945 {
1946 sock_wfree(sk, skb);
1947 release_sock(sk);
1948 if (copied)
1949 return(copied);
1950 return(tmp);
1951 }
1952
1953 if (flags & MSG_OOB)
1954 {
1955 skb->h.th->urg = 1;
1956 skb->h.th->urg_ptr = ntohs(copy);
1957 }
1958
1959 memcpy_fromfs(skb_put(skb,copy), from, copy);
1960
1961 from += copy;
1962 copied += copy;
1963 len -= copy;
1964 seglen -= copy;
1965 skb->free = 0;
1966 sk->write_seq += copy;
1967
1968 if (send_tmp != NULL && sk->packets_out)
1969 {
1970 tcp_enqueue_partial(send_tmp, sk);
1971 continue;
1972 }
1973 tcp_send_skb(sk, skb);
1974 }
1975 }
1976 sk->err = 0;
1977
1978 /*
1979 * Nagle's rule. Turn Nagle off with TCP_NODELAY for highly
1980 * interactive fast network servers. It's meant to be on and
1981 * it really improves the throughput though not the echo time
1982 * on my slow slip link - Alan
1983 */
1984
1985 /*
1986 * Avoid possible race on send_tmp - c/o Johannes Stille
1987 */
1988
1989 if(sk->partial && ((!sk->packets_out)
1990 /* If not nagling we can send on the before case too.. */
1991 || (sk->nonagle && before(sk->write_seq , sk->window_seq))
1992 ))
1993 tcp_send_partial(sk);
1994
1995 release_sock(sk);
1996 return(copied);
1997 }
1998
1999 static int tcp_sendto(struct sock *sk, const unsigned char *ubuf, int size, int noblock, unsigned flags,
/* */
2000 struct sockaddr_in *sin, int addr_len)
2001 {
2002 struct iovec iov;
2003 struct msghdr msg;
2004
2005 iov.iov_base = (void *)ubuf;
2006 iov.iov_len = size;
2007
2008 msg.msg_name = (void *)sin;
2009 msg.msg_namelen = addr_len;
2010 msg.msg_accrights = NULL;
2011 msg.msg_iov = &iov;
2012 msg.msg_iovlen = 1;
2013
2014 return tcp_sendmsg(sk, &msg, size, noblock, flags);
2015 }
2016
2017 static int tcp_write(struct sock *sk, const unsigned char *ubuf, int size, int noblock, unsigned flags)
/* */
2018 {
2019 return tcp_sendto(sk,ubuf,size,noblock,flags,NULL,0);
2020 }
2021
2022
2023 /*
2024 * Send an ack if one is backlogged at this point. Ought to merge
2025 * this with tcp_send_ack().
2026 */
2027
2028 static void tcp_read_wakeup(struct sock *sk)
/* */
2029 {
2030 int tmp;
2031 struct device *dev = NULL;
2032 struct tcphdr *t1;
2033 struct sk_buff *buff;
2034
2035 if (!sk->ack_backlog)
2036 return;
2037
2038 /*
2039 * If we're closed, don't send an ack, or we'll get a RST
2040 * from the closed destination.
2041 */
2042 if ((sk->state == TCP_CLOSE) || (sk->state == TCP_TIME_WAIT))
2043 return;
2044
2045 /*
2046 * FIXME: we need to put code here to prevent this routine from
2047 * being called. Being called once in a while is ok, so only check
2048 * if this is the second time in a row.
2049 */
2050
2051 /*
2052 * We need to grab some memory, and put together an ack,
2053 * and then put it into the queue to be sent.
2054 */
2055
2056 buff = sock_wmalloc(sk,MAX_ACK_SIZE,1, GFP_ATOMIC);
2057 if (buff == NULL)
2058 {
2059 /* Try again real soon. */
2060 reset_xmit_timer(sk, TIME_WRITE, HZ);
2061 return;
2062 }
2063
2064 buff->sk = sk;
2065 buff->localroute = sk->localroute;
2066
2067 /*
2068 * Put in the IP header and routing stuff.
2069 */
2070
2071 tmp = sk->prot->build_header(buff, sk->saddr, sk->daddr, &dev,
2072 IPPROTO_TCP, sk->opt, MAX_ACK_SIZE,sk->ip_tos,sk->ip_ttl);
2073 if (tmp < 0)
2074 {
2075 buff->free = 1;
2076 sock_wfree(sk, buff);
2077 return;
2078 }
2079
2080 t1 =(struct tcphdr *)skb_put(buff,sizeof(struct tcphdr));
2081
2082 memcpy(t1,(void *) &sk->dummy_th, sizeof(*t1));
2083 t1->seq = htonl(sk->sent_seq);
2084 t1->ack = 1;
2085 t1->res1 = 0;
2086 t1->res2 = 0;
2087 t1->rst = 0;
2088 t1->urg = 0;
2089 t1->syn = 0;
2090 t1->psh = 0;
2091 sk->ack_backlog = 0;
2092 sk->bytes_rcv = 0;
2093 sk->window = tcp_select_window(sk);
2094 t1->window = ntohs(sk->window);
2095 t1->ack_seq = ntohl(sk->acked_seq);
2096 t1->doff = sizeof(*t1)/4;
2097 tcp_send_check(t1, sk->saddr, sk->daddr, sizeof(*t1), sk);
2098 sk->prot->queue_xmit(sk, dev, buff, 1);
2099 tcp_statistics.TcpOutSegs++;
2100 }
2101
2102
2103 /*
2104 * FIXME:
2105 * This routine frees used buffers.
2106 * It should consider sending an ACK to let the
2107 * other end know we now have a bigger window.
2108 */
2109
2110 static void cleanup_rbuf(struct sock *sk)
/* */
2111 {
2112 unsigned long flags;
2113 unsigned long left;
2114 struct sk_buff *skb;
2115 unsigned long rspace;
2116
2117 if(sk->debug)
2118 printk("cleaning rbuf for sk=%p\n", sk);
2119
2120 save_flags(flags);
2121 cli();
2122
2123 left = sock_rspace(sk);
2124
2125 /*
2126 * We have to loop through all the buffer headers,
2127 * and try to free up all the space we can.
2128 */
2129
2130 while((skb=skb_peek(&sk->receive_queue)) != NULL)
2131 {
2132 if (!skb->used || skb->users)
2133 break;
2134 skb_unlink(skb);
2135 skb->sk = sk;
2136 kfree_skb(skb, FREE_READ);
2137 }
2138
2139 restore_flags(flags);
2140
2141 /*
2142 * FIXME:
2143 * At this point we should send an ack if the difference
2144 * in the window, and the amount of space is bigger than
2145 * TCP_WINDOW_DIFF.
2146 */
2147
2148 if(sk->debug)
2149 printk("sk->rspace = %lu, was %lu\n", sock_rspace(sk),
2150 left);
2151 if ((rspace=sock_rspace(sk)) != left)
2152 {
2153 /*
2154 * This area has caused the most trouble. The current strategy
2155 * is to simply do nothing if the other end has room to send at
2156 * least 3 full packets, because the ack from those will auto-
2157 * matically update the window. If the other end doesn't think
2158 * we have much space left, but we have room for at least 1 more
2159 * complete packet than it thinks we do, we will send an ack
2160 * immediately. Otherwise we will wait up to .5 seconds in case
2161 * the user reads some more.
2162 */
2163 sk->ack_backlog++;
2164 /*
2165 * It's unclear whether to use sk->mtu or sk->mss here. They differ only
2166 * if the other end is offering a window smaller than the agreed on MSS
2167 * (called sk->mtu here). In theory there's no connection between send
2168 * and receive, and so no reason to think that they're going to send
2169 * small packets. For the moment I'm using the hack of reducing the mss
2170 * only on the send side, so I'm putting mtu here.
2171 */
2172
2173 if (rspace > (sk->window - sk->bytes_rcv + sk->mtu))
2174 {
2175 /* Send an ack right now. */
2176 tcp_read_wakeup(sk);
2177 }
2178 else
2179 {
2180 /* Force it to send an ack soon. */
2181 int was_active = del_timer(&sk->retransmit_timer);
2182 if (!was_active || jiffies+TCP_ACK_TIME < sk->timer.expires)
2183 {
2184 reset_xmit_timer(sk, TIME_WRITE, TCP_ACK_TIME);
2185 }
2186 else
2187 add_timer(&sk->retransmit_timer);
2188 }
2189 }
2190 }
2191
2192
2193 /*
2194 * Handle reading urgent data. BSD has very simple semantics for
2195 * this, no blocking and very strange errors 8)
2196 */
2197
2198 static int tcp_recv_urg(struct sock * sk, int nonblock,
/* */
2199 struct msghdr *msg, int len, int flags, int *addr_len)
2200 {
2201 /*
2202 * No URG data to read
2203 */
2204 if (sk->urginline || !sk->urg_data || sk->urg_data == URG_READ)
2205 return -EINVAL; /* Yes this is right ! */
2206
2207 if (sk->err)
2208 {
2209 int tmp = -sk->err;
2210 sk->err = 0;
2211 return tmp;
2212 }
2213
2214 if (sk->state == TCP_CLOSE || sk->done)
2215 {
2216 if (!sk->done)
2217 {
2218 sk->done = 1;
2219 return 0;
2220 }
2221 return -ENOTCONN;
2222 }
2223
2224 if (sk->shutdown & RCV_SHUTDOWN)
2225 {
2226 sk->done = 1;
2227 return 0;
2228 }
2229 sk->inuse = 1;
2230 if (sk->urg_data & URG_VALID)
2231 {
2232 char c = sk->urg_data;
2233 if (!(flags & MSG_PEEK))
2234 sk->urg_data = URG_READ;
2235 memcpy_toiovec(msg->msg_iov, &c, 1);
2236 if(msg->msg_name)
2237 {
2238 struct sockaddr_in *sin=(struct sockaddr_in *)msg->msg_name;
2239 sin->sin_family=AF_INET;
2240 sin->sin_addr.s_addr=sk->daddr;
2241 sin->sin_port=sk->dummy_th.dest;
2242 }
2243 if(addr_len)
2244 *addr_len=sizeof(struct sockaddr_in);
2245 release_sock(sk);
2246 return 1;
2247 }
2248 release_sock(sk);
2249
2250 /*
2251 * Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
2252 * the available implementations agree in this case:
2253 * this call should never block, independent of the
2254 * blocking state of the socket.
2255 * Mike <pall@rz.uni-karlsruhe.de>
2256 */
2257 return -EAGAIN;
2258 }
2259
2260
2261 /*
2262 * This routine copies from a sock struct into the user buffer.
2263 */
2264
2265 static int tcp_recvmsg(struct sock *sk, struct msghdr *msg,
/* */
2266 int len, int nonblock, int flags, int *addr_len)
2267 {
2268 struct wait_queue wait = { current, NULL };
2269 int copied = 0;
2270 u32 peek_seq;
2271 volatile u32 *seq; /* So gcc doesn't overoptimise */
2272 unsigned long used;
2273
2274 /*
2275 * This error should be checked.
2276 */
2277
2278 if (sk->state == TCP_LISTEN)
2279 return -ENOTCONN;
2280
2281 /*
2282 * Urgent data needs to be handled specially.
2283 */
2284
2285 if (flags & MSG_OOB)
2286 return tcp_recv_urg(sk, nonblock, msg, len, flags, addr_len);
2287
2288 /*
2289 * Copying sequence to update. This is volatile to handle
2290 * the multi-reader case neatly (memcpy_to/fromfs might be
2291 * inline and thus not flush cached variables otherwise).
2292 */
2293
2294 peek_seq = sk->copied_seq;
2295 seq = &sk->copied_seq;
2296 if (flags & MSG_PEEK)
2297 seq = &peek_seq;
2298
2299 add_wait_queue(sk->sleep, &wait);
2300 sk->inuse = 1;
2301 while (len > 0)
2302 {
2303 struct sk_buff * skb;
2304 u32 offset;
2305
2306 /*
2307 * Are we at urgent data? Stop if we have read anything.
2308 */
2309
2310 if (copied && sk->urg_data && sk->urg_seq == *seq)
2311 break;
2312
2313 /*
2314 * Next get a buffer.
2315 */
2316
2317 current->state = TASK_INTERRUPTIBLE;
2318
2319 skb = skb_peek(&sk->receive_queue);
2320 do
2321 {
2322 if (!skb)
2323 break;
2324 if (before(*seq, skb->h.th->seq))
2325 break;
2326 offset = *seq - skb->h.th->seq;
2327 if (skb->h.th->syn)
2328 offset--;
2329 if (offset < skb->len)
2330 goto found_ok_skb;
2331 if (skb->h.th->fin)
2332 goto found_fin_ok;
2333 if (!(flags & MSG_PEEK))
2334 skb->used = 1;
2335 skb = skb->next;
2336 }
2337 while (skb != (struct sk_buff *)&sk->receive_queue);
2338
2339 if (copied)
2340 break;
2341
2342 if (sk->err)
2343 {
2344 copied = -sk->err;
2345 sk->err = 0;
2346 break;
2347 }
2348
2349 if (sk->state == TCP_CLOSE)
2350 {
2351 if (!sk->done)
2352 {
2353 sk->done = 1;
2354 break;
2355 }
2356 copied = -ENOTCONN;
2357 break;
2358 }
2359
2360 if (sk->shutdown & RCV_SHUTDOWN)
2361 {
2362 sk->done = 1;
2363 break;
2364 }
2365
2366 if (nonblock)
2367 {
2368 copied = -EAGAIN;
2369 break;
2370 }
2371
2372 cleanup_rbuf(sk);
2373 release_sock(sk);
2374 sk->socket->flags |= SO_WAITDATA;
2375 schedule();
2376 sk->socket->flags &= ~SO_WAITDATA;
2377 sk->inuse = 1;
2378
2379 if (current->signal & ~current->blocked)
2380 {
2381 copied = -ERESTARTSYS;
2382 break;
2383 }
2384 continue;
2385
2386 found_ok_skb:
2387 /*
2388 * Lock the buffer. We can be fairly relaxed as
2389 * an interrupt will never steal a buffer we are
2390 * using unless I've missed something serious in
2391 * tcp_data.
2392 */
2393
2394 skb->users++;
2395
2396 /*
2397 * Ok so how much can we use ?
2398 */
2399
2400 used = skb->len - offset;
2401 if (len < used)
2402 used = len;
2403 /*
2404 * Do we have urgent data here?
2405 */
2406
2407 if (sk->urg_data)
2408 {
2409 u32 urg_offset = sk->urg_seq - *seq;
2410 if (urg_offset < used)
2411 {
2412 if (!urg_offset)
2413 {
2414 if (!sk->urginline)
2415 {
2416 ++*seq;
2417 offset++;
2418 used--;
2419 }
2420 }
2421 else
2422 used = urg_offset;
2423 }
2424 }
2425
2426 /*
2427 * Copy it - We _MUST_ update *seq first so that we
2428 * don't ever double read when we have dual readers
2429 */
2430
2431 *seq += used;
2432
2433 /*
2434 * This memcpy_tofs can sleep. If it sleeps and we
2435 * do a second read it relies on the skb->users to avoid
2436 * a crash when cleanup_rbuf() gets called.
2437 */
2438
2439 memcpy_toiovec(msg->msg_iov,((unsigned char *)skb->h.th) +
2440 skb->h.th->doff*4 + offset, used);
2441 copied += used;
2442 len -= used;
2443
2444 /*
2445 * We now will not sleep again until we are finished
2446 * with skb. Sorry if you are doing the SMP port
2447 * but you'll just have to fix it neatly ;)
2448 */
2449
2450 skb->users --;
2451
2452 if (after(sk->copied_seq,sk->urg_seq))
2453 sk->urg_data = 0;
2454 if (used + offset < skb->len)
2455 continue;
2456
2457 /*
2458 * Process the FIN.
2459 */
2460
2461 if (skb->h.th->fin)
2462 goto found_fin_ok;
2463 if (flags & MSG_PEEK)
2464 continue;
2465 skb->used = 1;
2466 continue;
2467
2468 found_fin_ok:
2469 ++*seq;
2470 if (flags & MSG_PEEK)
2471 break;
2472
2473 /*
2474 * All is done
2475 */
2476
2477 skb->used = 1;
2478 sk->shutdown |= RCV_SHUTDOWN;
2479 break;
2480
2481 }
2482
2483 if(copied>0 && msg->msg_name)
2484 {
2485 struct sockaddr_in *sin=(struct sockaddr_in *)msg->msg_name;
2486 sin->sin_family=AF_INET;
2487 sin->sin_addr.s_addr=sk->daddr;
2488 sin->sin_port=sk->dummy_th.dest;
2489 }
2490 if(addr_len)
2491 *addr_len=sizeof(struct sockaddr_in);
2492
2493 remove_wait_queue(sk->sleep, &wait);
2494 current->state = TASK_RUNNING;
2495
2496 /* Clean up data we have read: This will do ACK frames */
2497 cleanup_rbuf(sk);
2498 release_sock(sk);
2499 return copied;
2500 }
2501
2502
2503 static int tcp_recvfrom(struct sock *sk, unsigned char *ubuf, int size, int noblock, unsigned flags,
/* */
2504 struct sockaddr_in *sa, int *addr_len)
2505 {
2506 struct iovec iov;
2507 struct msghdr msg;
2508
2509 iov.iov_base = (void *)ubuf;
2510 iov.iov_len = size;
2511
2512 msg.msg_name = (void *)sa;
2513 msg.msg_namelen = 0;
2514 if (addr_len)
2515 msg.msg_namelen = *addr_len;
2516 msg.msg_accrights = NULL;
2517 msg.msg_iov = &iov;
2518 msg.msg_iovlen = 1;
2519
2520 return tcp_recvmsg(sk, &msg, size, noblock, flags, addr_len);
2521 }
2522
2523 int tcp_read(struct sock *sk, unsigned char *buff, int len, int noblock,
/* */
2524 unsigned flags)
2525 {
2526 return(tcp_recvfrom(sk, buff, len, noblock, flags, NULL, NULL));
2527 }
2528
2529
2530 /*
2531 * State processing on a close. This implements the state shift for
2532 * sending our FIN frame. Note that we only send a FIN for some
2533 * states. A shutdown() may have already sent the FIN, or we may be
2534 * closed.
2535 */
2536
2537 static int tcp_close_state(struct sock *sk, int dead)
/* */
2538 {
2539 int ns=TCP_CLOSE;
2540 int send_fin=0;
2541 switch(sk->state)
2542 {
2543 case TCP_SYN_SENT: /* No SYN back, no FIN needed */
2544 break;
2545 case TCP_SYN_RECV:
2546 case TCP_ESTABLISHED: /* Closedown begin */
2547 ns=TCP_FIN_WAIT1;
2548 send_fin=1;
2549 break;
2550 case TCP_FIN_WAIT1: /* Already closing, or FIN sent: no change */
2551 case TCP_FIN_WAIT2:
2552 case TCP_CLOSING:
2553 ns=sk->state;
2554 break;
2555 case TCP_CLOSE:
2556 case TCP_LISTEN:
2557 break;
2558 case TCP_CLOSE_WAIT: /* They have FIN'd us. We send our FIN and
2559 wait only for the ACK */
2560 ns=TCP_LAST_ACK;
2561 send_fin=1;
2562 }
2563
2564 tcp_set_state(sk,ns);
2565
2566 /*
2567 * This is a (useful) BSD violating of the RFC. There is a
2568 * problem with TCP as specified in that the other end could
2569 * keep a socket open forever with no application left this end.
2570 * We use a 3 minute timeout (about the same as BSD) then kill
2571 * our end. If they send after that then tough - BUT: long enough
2572 * that we won't make the old 4*rto = almost no time - whoops
2573 * reset mistake.
2574 */
2575 if(dead && ns==TCP_FIN_WAIT2)
2576 {
2577 int timer_active=del_timer(&sk->timer);
2578 if(timer_active)
2579 add_timer(&sk->timer);
2580 else
2581 reset_msl_timer(sk, TIME_CLOSE, TCP_FIN_TIMEOUT);
2582 }
2583
2584 return send_fin;
2585 }
2586
2587 /*
2588 * Send a fin.
2589 */
2590
2591 static void tcp_send_fin(struct sock *sk)
/* */
2592 {
2593 struct proto *prot =(struct proto *)sk->prot;
2594 struct tcphdr *th =(struct tcphdr *)&sk->dummy_th;
2595 struct tcphdr *t1;
2596 struct sk_buff *buff;
2597 struct device *dev=NULL;
2598 int tmp;
2599
2600 release_sock(sk); /* in case the malloc sleeps. */
2601
2602 buff = sock_wmalloc(sk, MAX_RESET_SIZE,1 , GFP_KERNEL);
2603 sk->inuse = 1;
2604
2605 if (buff == NULL)
2606 {
2607 /* This is a disaster if it occurs */
2608 printk("tcp_send_fin: Impossible malloc failure");
2609 return;
2610 }
2611
2612 /*
2613 * Administrivia
2614 */
2615
2616 buff->sk = sk;
2617 buff->localroute = sk->localroute;
2618
2619 /*
2620 * Put in the IP header and routing stuff.
2621 */
2622
2623 tmp = prot->build_header(buff,sk->saddr, sk->daddr, &dev,
2624 IPPROTO_TCP, sk->opt,
2625 sizeof(struct tcphdr),sk->ip_tos,sk->ip_ttl);
2626 if (tmp < 0)
2627 {
2628 int t;
2629 /*
2630 * Finish anyway, treat this as a send that got lost.
2631 * (Not good).
2632 */
2633
2634 buff->free = 1;
2635 sock_wfree(sk,buff);
2636 sk->write_seq++;
2637 t=del_timer(&sk->timer);
2638 if(t)
2639 add_timer(&sk->timer);
2640 else
2641 reset_msl_timer(sk, TIME_CLOSE, TCP_TIMEWAIT_LEN);
2642 return;
2643 }
2644
2645 /*
2646 * We ought to check if the end of the queue is a buffer and
2647 * if so simply add the fin to that buffer, not send it ahead.
2648 */
2649
2650 t1 =(struct tcphdr *)skb_put(buff,sizeof(struct tcphdr));
2651 buff->dev = dev;
2652 memcpy(t1, th, sizeof(*t1));
2653 t1->seq = ntohl(sk->write_seq);
2654 sk->write_seq++;
2655 buff->h.seq = sk->write_seq;
2656 t1->ack = 1;
2657 t1->ack_seq = ntohl(sk->acked_seq);
2658 t1->window = ntohs(sk->window=tcp_select_window(sk));
2659 t1->fin = 1;
2660 t1->rst = 0;
2661 t1->doff = sizeof(*t1)/4;
2662 tcp_send_check(t1, sk->saddr, sk->daddr, sizeof(*t1), sk);
2663
2664 /*
2665 * If there is data in the write queue, the fin must be appended to
2666 * the write queue.
2667 */
2668
2669 if (skb_peek(&sk->write_queue) != NULL)
2670 {
2671 buff->free = 0;
2672 if (buff->next != NULL)
2673 {
2674 printk("tcp_send_fin: next != NULL\n");
2675 skb_unlink(buff);
2676 }
2677 skb_queue_tail(&sk->write_queue, buff);
2678 }
2679 else
2680 {
2681 sk->sent_seq = sk->write_seq;
2682 sk->prot->queue_xmit(sk, dev, buff, 0);
2683 reset_xmit_timer(sk, TIME_WRITE, sk->rto);
2684 }
2685 }
2686
2687 /*
2688 * Shutdown the sending side of a connection. Much like close except
2689 * that we don't receive shut down or set sk->dead=1.
2690 */
2691
2692 void tcp_shutdown(struct sock *sk, int how)
/* */
2693 {
2694 /*
2695 * We need to grab some memory, and put together a FIN,
2696 * and then put it into the queue to be sent.
2697 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2698 */
2699
2700 if (!(how & SEND_SHUTDOWN))
2701 return;
2702
2703 /*
2704 * If we've already sent a FIN, or it's a closed state
2705 */
2706
2707 if (sk->state == TCP_FIN_WAIT1 ||
2708 sk->state == TCP_FIN_WAIT2 ||
2709 sk->state == TCP_CLOSING ||
2710 sk->state == TCP_LAST_ACK ||
2711 sk->state == TCP_TIME_WAIT ||
2712 sk->state == TCP_CLOSE ||
2713 sk->state == TCP_LISTEN
2714 )
2715 {
2716 return;
2717 }
2718 sk->inuse = 1;
2719
2720 /*
2721 * flag that the sender has shutdown
2722 */
2723
2724 sk->shutdown |= SEND_SHUTDOWN;
2725
2726 /*
2727 * Clear out any half completed packets.
2728 */
2729
2730 if (sk->partial)
2731 tcp_send_partial(sk);
2732
2733 /*
2734 * FIN if needed
2735 */
2736
2737 if(tcp_close_state(sk,0))
2738 tcp_send_fin(sk);
2739
2740 release_sock(sk);
2741 }
2742
2743 /*
2744 * This routine will send an RST to the other tcp.
2745 */
2746
2747 static void tcp_reset(unsigned long saddr, unsigned long daddr, struct tcphdr *th,
/* */
2748 struct proto *prot, struct options *opt, struct device *dev, int tos, int ttl)
2749 {
2750 struct sk_buff *buff;
2751 struct tcphdr *t1;
2752 int tmp;
2753 struct device *ndev=NULL;
2754
2755 /*
2756 * Cannot reset a reset (Think about it).
2757 */
2758
2759 if(th->rst)
2760 return;
2761
2762 /*
2763 * We need to grab some memory, and put together an RST,
2764 * and then put it into the queue to be sent.
2765 */
2766
2767 buff = sock_wmalloc(NULL, MAX_RESET_SIZE, 1, GFP_ATOMIC);
2768 if (buff == NULL)
2769 return;
2770
2771 buff->sk = NULL;
2772 buff->dev = dev;
2773 buff->localroute = 0;
2774
2775 /*
2776 * Put in the IP header and routing stuff.
2777 */
2778
2779 tmp = prot->build_header(buff, saddr, daddr, &ndev, IPPROTO_TCP, opt,
2780 sizeof(struct tcphdr),tos,ttl);
2781 if (tmp < 0)
2782 {
2783 buff->free = 1;
2784 sock_wfree(NULL, buff);
2785 return;
2786 }
2787
2788 t1 =(struct tcphdr *)skb_put(buff,sizeof(struct tcphdr));
2789 memcpy(t1, th, sizeof(*t1));
2790
2791 /*
2792 * Swap the send and the receive.
2793 */
2794
2795 t1->dest = th->source;
2796 t1->source = th->dest;
2797 t1->rst = 1;
2798 t1->window = 0;
2799
2800 if(th->ack)
2801 {
2802 t1->ack = 0;
2803 t1->seq = th->ack_seq;
2804 t1->ack_seq = 0;
2805 }
2806 else
2807 {
2808 t1->ack = 1;
2809 if(!th->syn)
2810 t1->ack_seq=htonl(th->seq);
2811 else
2812 t1->ack_seq=htonl(th->seq+1);
2813 t1->seq=0;
2814 }
2815
2816 t1->syn = 0;
2817 t1->urg = 0;
2818 t1->fin = 0;
2819 t1->psh = 0;
2820 t1->doff = sizeof(*t1)/4;
2821 tcp_send_check(t1, saddr, daddr, sizeof(*t1), NULL);
2822 prot->queue_xmit(NULL, ndev, buff, 1);
2823 tcp_statistics.TcpOutSegs++;
2824 }
2825
2826
2827 /*
2828 * Look for tcp options. Parses everything but only knows about MSS.
2829 * This routine is always called with the packet containing the SYN.
2830 * However it may also be called with the ack to the SYN. So you
2831 * can't assume this is always the SYN. It's always called after
2832 * we have set up sk->mtu to our own MTU.
2833 *
2834 * We need at minimum to add PAWS support here. Possibly large windows
2835 * as Linux gets deployed on 100Mb/sec networks.
2836 */
2837
2838 static void tcp_options(struct sock *sk, struct tcphdr *th)
/* */
2839 {
2840 unsigned char *ptr;
2841 int length=(th->doff*4)-sizeof(struct tcphdr);
2842 int mss_seen = 0;
2843
2844 ptr = (unsigned char *)(th + 1);
2845
2846 while(length>0)
2847 {
2848 int opcode=*ptr++;
2849 int opsize=*ptr++;
2850 switch(opcode)
2851 {
2852 case TCPOPT_EOL:
2853 return;
2854 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
2855 length--;
2856 ptr--; /* the opsize=*ptr++ above was a mistake */
2857 continue;
2858
2859 default:
2860 if(opsize<=2) /* Avoid silly options looping forever */
2861 return;
2862 switch(opcode)
2863 {
2864 case TCPOPT_MSS:
2865 if(opsize==4 && th->syn)
2866 {
2867 sk->mtu=min(sk->mtu,ntohs(*(unsigned short *)ptr));
2868 mss_seen = 1;
2869 }
2870 break;
2871 /* Add other options here as people feel the urge to implement stuff like large windows */
2872 }
2873 ptr+=opsize-2;
2874 length-=opsize;
2875 }
2876 }
2877 if (th->syn)
2878 {
2879 if (! mss_seen)
2880 sk->mtu=min(sk->mtu, 536); /* default MSS if none sent */
2881 }
2882 #ifdef CONFIG_INET_PCTCP
2883 sk->mss = min(sk->max_window >> 1, sk->mtu);
2884 #else
2885 sk->mss = min(sk->max_window, sk->mtu);
2886 #endif
2887 }
2888
2889 static inline unsigned long default_mask(unsigned long dst)
/* */
2890 {
2891 dst = ntohl(dst);
2892 if (IN_CLASSA(dst))
2893 return htonl(IN_CLASSA_NET);
2894 if (IN_CLASSB(dst))
2895 return htonl(IN_CLASSB_NET);
2896 return htonl(IN_CLASSC_NET);
2897 }
2898
2899 /*
2900 * Default sequence number picking algorithm.
2901 * As close as possible to RFC 793, which
2902 * suggests using a 250kHz clock.
2903 * Further reading shows this assumes 2MB/s networks.
2904 * For 10MB/s ethernet, a 1MHz clock is appropriate.
2905 * That's funny, Linux has one built in! Use it!
2906 */
2907
2908 extern inline u32 tcp_init_seq(void)
/* */
2909 {
2910 struct timeval tv;
2911 do_gettimeofday(&tv);
2912 return tv.tv_usec+tv.tv_sec*1000000;
2913 }
2914
2915 /*
2916 * This routine handles a connection request.
2917 * It should make sure we haven't already responded.
2918 * Because of the way BSD works, we have to send a syn/ack now.
2919 * This also means it will be harder to close a socket which is
2920 * listening.
2921 */
2922
2923 static void tcp_conn_request(struct sock *sk, struct sk_buff *skb,
/* */
2924 unsigned long daddr, unsigned long saddr,
2925 struct options *opt, struct device *dev, u32 seq)
2926 {
2927 struct sk_buff *buff;
2928 struct tcphdr *t1;
2929 unsigned char *ptr;
2930 struct sock *newsk;
2931 struct tcphdr *th;
2932 struct device *ndev=NULL;
2933 int tmp;
2934 struct rtable *rt;
2935
2936 th = skb->h.th;
2937
2938 /* If the socket is dead, don't accept the connection. */
2939 if (!sk->dead)
2940 {
2941 sk->data_ready(sk,0);
2942 }
2943 else
2944 {
2945 if(sk->debug)
2946 printk("Reset on %p: Connect on dead socket.\n",sk);
2947 tcp_reset(daddr, saddr, th, sk->prot, opt, dev, sk->ip_tos,sk->ip_ttl);
2948 tcp_statistics.TcpAttemptFails++;
2949 kfree_skb(skb, FREE_READ);
2950 return;
2951 }
2952
2953 /*
2954 * Make sure we can accept more. This will prevent a
2955 * flurry of syns from eating up all our memory.
2956 */
2957
2958 if (sk->ack_backlog >= sk->max_ack_backlog)
2959 {
2960 tcp_statistics.TcpAttemptFails++;
2961 kfree_skb(skb, FREE_READ);
2962 return;
2963 }
2964
2965 /*
2966 * We need to build a new sock struct.
2967 * It is sort of bad to have a socket without an inode attached
2968 * to it, but the wake_up's will just wake up the listening socket,
2969 * and if the listening socket is destroyed before this is taken
2970 * off of the queue, this will take care of it.
2971 */
2972
2973 newsk = (struct sock *) kmalloc(sizeof(struct sock), GFP_ATOMIC);
2974 if (newsk == NULL)
2975 {
2976 /* just ignore the syn. It will get retransmitted. */
2977 tcp_statistics.TcpAttemptFails++;
2978 kfree_skb(skb, FREE_READ);
2979 return;
2980 }
2981
2982 memcpy(newsk, sk, sizeof(*newsk));
2983 newsk->opt = NULL;
2984 if (opt && opt->optlen) {
2985 sk->opt = (struct options*)kmalloc(sizeof(struct options)+opt->optlen, GFP_ATOMIC);
2986 if (!sk->opt) {
2987 kfree_s(newsk, sizeof(struct sock));
2988 tcp_statistics.TcpAttemptFails++;
2989 kfree_skb(skb, FREE_READ);
2990 return;
2991 }
2992 if (ip_options_echo(sk->opt, opt, daddr, saddr, skb)) {
2993 kfree_s(sk->opt, sizeof(struct options)+opt->optlen);
2994 kfree_s(newsk, sizeof(struct sock));
2995 tcp_statistics.TcpAttemptFails++;
2996 kfree_skb(skb, FREE_READ);
2997 return;
2998 }
2999 }
3000 skb_queue_head_init(&newsk->write_queue);
3001 skb_queue_head_init(&newsk->receive_queue);
3002 newsk->send_head = NULL;
3003 newsk->send_tail = NULL;
3004 skb_queue_head_init(&newsk->back_log);
3005 newsk->rtt = 0; /*TCP_CONNECT_TIME<<3*/
3006 newsk->rto = TCP_TIMEOUT_INIT;
3007 newsk->mdev = 0;
3008 newsk->max_window = 0;
3009 newsk->cong_window = 1;
3010 newsk->cong_count = 0;
3011 newsk->ssthresh = 0;
3012 newsk->backoff = 0;
3013 newsk->blog = 0;
3014 newsk->intr = 0;
3015 newsk->proc = 0;
3016 newsk->done = 0;
3017 newsk->partial = NULL;
3018 newsk->pair = NULL;
3019 newsk->wmem_alloc = 0;
3020 newsk->rmem_alloc = 0;
3021 newsk->localroute = sk->localroute;
3022
3023 newsk->max_unacked = MAX_WINDOW - TCP_WINDOW_DIFF;
3024
3025 newsk->err = 0;
3026 newsk->shutdown = 0;
3027 newsk->ack_backlog = 0;
3028 newsk->acked_seq = skb->h.th->seq+1;
3029 newsk->copied_seq = skb->h.th->seq+1;
3030 newsk->fin_seq = skb->h.th->seq;
3031 newsk->state = TCP_SYN_RECV;
3032 newsk->timeout = 0;
3033 newsk->ip_xmit_timeout = 0;
3034 newsk->write_seq = seq;
3035 newsk->window_seq = newsk->write_seq;
3036 newsk->rcv_ack_seq = newsk->write_seq;
3037 newsk->urg_data = 0;
3038 newsk->retransmits = 0;
3039 newsk->linger=0;
3040 newsk->destroy = 0;
3041 init_timer(&newsk->timer);
3042 newsk->timer.data = (unsigned long)newsk;
3043 newsk->timer.function = &net_timer;
3044 init_timer(&newsk->retransmit_timer);
3045 newsk->retransmit_timer.data = (unsigned long)newsk;
3046 newsk->retransmit_timer.function=&retransmit_timer;
3047 newsk->dummy_th.source = skb->h.th->dest;
3048 newsk->dummy_th.dest = skb->h.th->source;
3049
3050 /*
3051 * Swap these two, they are from our point of view.
3052 */
3053
3054 newsk->daddr = saddr;
3055 newsk->saddr = daddr;
3056 newsk->rcv_saddr = daddr;
3057
3058 put_sock(newsk->num,newsk);
3059 newsk->dummy_th.res1 = 0;
3060 newsk->dummy_th.doff = 6;
3061 newsk->dummy_th.fin = 0;
3062 newsk->dummy_th.syn = 0;
3063 newsk->dummy_th.rst = 0;
3064 newsk->dummy_th.psh = 0;
3065 newsk->dummy_th.ack = 0;
3066 newsk->dummy_th.urg = 0;
3067 newsk->dummy_th.res2 = 0;
3068 newsk->acked_seq = skb->h.th->seq + 1;
3069 newsk->copied_seq = skb->h.th->seq + 1;
3070 newsk->socket = NULL;
3071
3072 /*
3073 * Grab the ttl and tos values and use them
3074 */
3075
3076 newsk->ip_ttl=sk->ip_ttl;
3077 newsk->ip_tos=skb->ip_hdr->tos;
3078
3079 /*
3080 * Use 512 or whatever user asked for
3081 */
3082
3083 /*
3084 * Note use of sk->user_mss, since user has no direct access to newsk
3085 */
3086
3087 rt=ip_rt_route(saddr, NULL,NULL);
3088
3089 if(rt!=NULL && (rt->rt_flags&RTF_WINDOW))
3090 newsk->window_clamp = rt->rt_window;
3091 else
3092 newsk->window_clamp = 0;
3093
3094 if (sk->user_mss)
3095 newsk->mtu = sk->user_mss;
3096 else if(rt!=NULL && (rt->rt_flags&RTF_MSS))
3097 newsk->mtu = rt->rt_mss - sizeof(struct iphdr) - sizeof(struct tcphdr);
3098 else
3099 {
3100 #ifdef CONFIG_INET_SNARL /* Sub Nets Are Local */
3101 if ((saddr ^ daddr) & default_mask(saddr))
3102 #else
3103 if ((saddr ^ daddr) & dev->pa_mask)
3104 #endif
3105 newsk->mtu = 576 - sizeof(struct iphdr) - sizeof(struct tcphdr);
3106 else
3107 newsk->mtu = MAX_WINDOW;
3108 }
3109
3110 /*
3111 * But not bigger than device MTU
3112 */
3113
3114 newsk->mtu = min(newsk->mtu, dev->mtu - sizeof(struct iphdr) - sizeof(struct tcphdr));
3115
3116 /*
3117 * This will min with what arrived in the packet
3118 */
3119
3120 tcp_options(newsk,skb->h.th);
3121
3122 tcp_cache_zap();
3123
3124 buff = sock_wmalloc(newsk, MAX_SYN_SIZE, 1, GFP_ATOMIC);
3125 if (buff == NULL)
3126 {
3127 sk->err = ENOMEM;
3128 newsk->dead = 1;
3129 newsk->state = TCP_CLOSE;
3130 /* And this will destroy it */
3131 release_sock(newsk);
3132 kfree_skb(skb, FREE_READ);
3133 tcp_statistics.TcpAttemptFails++;
3134 return;
3135 }
3136
3137 buff->sk = newsk;
3138 buff->localroute = newsk->localroute;
3139
3140 /*
3141 * Put in the IP header and routing stuff.
3142 */
3143
3144 tmp = sk->prot->build_header(buff, newsk->saddr, newsk->daddr, &ndev,
3145 IPPROTO_TCP, NULL, MAX_SYN_SIZE,sk->ip_tos,sk->ip_ttl);
3146
3147 /*
3148 * Something went wrong.
3149 */
3150
3151 if (tmp < 0)
3152 {
3153 sk->err = tmp;
3154 buff->free = 1;
3155 kfree_skb(buff,FREE_WRITE);
3156 newsk->dead = 1;
3157 newsk->state = TCP_CLOSE;
3158 release_sock(newsk);
3159 skb->sk = sk;
3160 kfree_skb(skb, FREE_READ);
3161 tcp_statistics.TcpAttemptFails++;
3162 return;
3163 }
3164
3165 t1 =(struct tcphdr *)skb_put(buff,sizeof(struct tcphdr));
3166
3167 memcpy(t1, skb->h.th, sizeof(*t1));
3168 buff->h.seq = newsk->write_seq;
3169 /*
3170 * Swap the send and the receive.
3171 */
3172 t1->dest = skb->h.th->source;
3173 t1->source = newsk->dummy_th.source;
3174 t1->seq = ntohl(newsk->write_seq++);
3175 t1->ack = 1;
3176 newsk->window = tcp_select_window(newsk);
3177 newsk->sent_seq = newsk->write_seq;
3178 t1->window = ntohs(newsk->window);
3179 t1->res1 = 0;
3180 t1->res2 = 0;
3181 t1->rst = 0;
3182 t1->urg = 0;
3183 t1->psh = 0;
3184 t1->syn = 1;
3185 t1->ack_seq = ntohl(skb->h.th->seq+1);
3186 t1->doff = sizeof(*t1)/4+1;
3187 ptr = skb_put(buff,4);
3188 ptr[0] = 2;
3189 ptr[1] = 4;
3190 ptr[2] = ((newsk->mtu) >> 8) & 0xff;
3191 ptr[3] =(newsk->mtu) & 0xff;
3192
3193 tcp_send_check(t1, daddr, saddr, sizeof(*t1)+4, newsk);
3194 newsk->prot->queue_xmit(newsk, ndev, buff, 0);
3195 reset_xmit_timer(newsk, TIME_WRITE , TCP_TIMEOUT_INIT);
3196 skb->sk = newsk;
3197
3198 /*
3199 * Charge the sock_buff to newsk.
3200 */
3201
3202 sk->rmem_alloc -= skb->truesize;
3203 newsk->rmem_alloc += skb->truesize;
3204
3205 skb_queue_tail(&sk->receive_queue,skb);
3206 sk->ack_backlog++;
3207 release_sock(newsk);
3208 tcp_statistics.TcpOutSegs++;
3209 }
3210
3211
3212 static void tcp_close(struct sock *sk, int timeout)
/* */
3213 {
3214 /*
3215 * We need to grab some memory, and put together a FIN,
3216 * and then put it into the queue to be sent.
3217 */
3218
3219 sk->inuse = 1;
3220
3221 if(th_cache_sk==sk)
3222 tcp_cache_zap();
3223 if(sk->state == TCP_LISTEN)
3224 {
3225 /* Special case */
3226 tcp_set_state(sk, TCP_CLOSE);
3227 tcp_close_pending(sk);
3228 release_sock(sk);
3229 return;
3230 }
3231
3232 sk->keepopen = 1;
3233 sk->shutdown = SHUTDOWN_MASK;
3234
3235 if (!sk->dead)
3236 sk->state_change(sk);
3237
3238 if (timeout == 0)
3239 {
3240 struct sk_buff *skb;
3241
3242 /*
3243 * We need to flush the recv. buffs. We do this only on the
3244 * descriptor close, not protocol-sourced closes, because the
3245 * reader process may not have drained the data yet!
3246 */
3247
3248 while((skb=skb_dequeue(&sk->receive_queue))!=NULL)
3249 kfree_skb(skb, FREE_READ);
3250 /*
3251 * Get rid off any half-completed packets.
3252 */
3253
3254 if (sk->partial)
3255 tcp_send_partial(sk);
3256 }
3257
3258
3259 /*
3260 * Timeout is not the same thing - however the code likes
3261 * to send both the same way (sigh).
3262 */
3263
3264 if(timeout)
3265 {
3266 tcp_set_state(sk, TCP_CLOSE); /* Dead */
3267 }
3268 else
3269 {
3270 if(tcp_close_state(sk,1)==1)
3271 {
3272 tcp_send_fin(sk);
3273 }
3274 }
3275 release_sock(sk);
3276 }
3277
3278
3279 /*
3280 * This routine takes stuff off of the write queue,
3281 * and puts it in the xmit queue. This happens as incoming acks
3282 * open up the remote window for us.
3283 */
3284
3285 static void tcp_write_xmit(struct sock *sk)
/* */
3286 {
3287 struct sk_buff *skb;
3288
3289 /*
3290 * The bytes will have to remain here. In time closedown will
3291 * empty the write queue and all will be happy
3292 */
3293
3294 if(sk->zapped)
3295 return;
3296
3297 /*
3298 * Anything on the transmit queue that fits the window can
3299 * be added providing we are not
3300 *
3301 * a) retransmitting (Nagle's rule)
3302 * b) exceeding our congestion window.
3303 */
3304
3305 while((skb = skb_peek(&sk->write_queue)) != NULL &&
3306 before(skb->h.seq, sk->window_seq + 1) &&
3307 (sk->retransmits == 0 ||
3308 sk->ip_xmit_timeout != TIME_WRITE ||
3309 before(skb->h.seq, sk->rcv_ack_seq + 1))
3310 && sk->packets_out < sk->cong_window)
3311 {
3312 IS_SKB(skb);
3313 skb_unlink(skb);
3314
3315 /*
3316 * See if we really need to send the packet.
3317 */
3318
3319 if (before(skb->h.seq, sk->rcv_ack_seq +1))
3320 {
3321 /*
3322 * This is acked data. We can discard it. This
3323 * cannot currently occur.
3324 */
3325
3326 sk->retransmits = 0;
3327 kfree_skb(skb, FREE_WRITE);
3328 if (!sk->dead)
3329 sk->write_space(sk);
3330 }
3331 else
3332 {
3333 struct tcphdr *th;
3334 struct iphdr *iph;
3335 int size;
3336 /*
3337 * put in the ack seq and window at this point rather than earlier,
3338 * in order to keep them monotonic. We really want to avoid taking
3339 * back window allocations. That's legal, but RFC1122 says it's frowned on.
3340 * Ack and window will in general have changed since this packet was put
3341 * on the write queue.
3342 */
3343 iph = skb->ip_hdr;
3344 th = (struct tcphdr *)(((char *)iph) +(iph->ihl << 2));
3345 size = skb->len - (((unsigned char *) th) - skb->data);
3346
3347 th->ack_seq = ntohl(sk->acked_seq);
3348 th->window = ntohs(tcp_select_window(sk));
3349
3350 tcp_send_check(th, sk->saddr, sk->daddr, size, sk);
3351
3352 sk->sent_seq = skb->h.seq;
3353
3354 /*
3355 * IP manages our queue for some crazy reason
3356 */
3357
3358 sk->prot->queue_xmit(sk, skb->dev, skb, skb->free);
3359
3360 /*
3361 * Again we slide the timer wrongly
3362 */
3363
3364 reset_xmit_timer(sk, TIME_WRITE, sk->rto);
3365 }
3366 }
3367 }
3368
3369
3370 /*
3371 * This routine deals with incoming acks, but not outgoing ones.
3372 */
3373
3374 extern __inline__ int tcp_ack(struct sock *sk, struct tcphdr *th, unsigned long saddr, int len)
/* */
3375 {
3376 u32 ack;
3377 int flag = 0;
3378
3379 /*
3380 * 1 - there was data in packet as well as ack or new data is sent or
3381 * in shutdown state
3382 * 2 - data from retransmit queue was acked and removed
3383 * 4 - window shrunk or data from retransmit queue was acked and removed
3384 */
3385
3386 if(sk->zapped)
3387 return(1); /* Dead, cant ack any more so why bother */
3388
3389 /*
3390 * Have we discovered a larger window
3391 */
3392
3393 ack = ntohl(th->ack_seq);
3394
3395 if (ntohs(th->window) > sk->max_window)
3396 {
3397 sk->max_window = ntohs(th->window);
3398 #ifdef CONFIG_INET_PCTCP
3399 /* Hack because we don't send partial packets to non SWS
3400 handling hosts */
3401 sk->mss = min(sk->max_window>>1, sk->mtu);
3402 #else
3403 sk->mss = min(sk->max_window, sk->mtu);
3404 #endif
3405 }
3406
3407 /*
3408 * We have dropped back to keepalive timeouts. Thus we have
3409 * no retransmits pending.
3410 */
3411
3412 if (sk->retransmits && sk->ip_xmit_timeout == TIME_KEEPOPEN)
3413 sk->retransmits = 0;
3414
3415 /*
3416 * If the ack is newer than sent or older than previous acks
3417 * then we can probably ignore it.
3418 */
3419
3420 if (after(ack, sk->sent_seq) || before(ack, sk->rcv_ack_seq))
3421 {
3422 if(sk->debug)
3423 printk("Ack ignored %u %u\n",ack,sk->sent_seq);
3424
3425 /*
3426 * Keepalive processing.
3427 */
3428
3429 if (after(ack, sk->sent_seq))
3430 {
3431 return(0);
3432 }
3433
3434 /*
3435 * Restart the keepalive timer.
3436 */
3437
3438 if (sk->keepopen)
3439 {
3440 if(sk->ip_xmit_timeout==TIME_KEEPOPEN)
3441 reset_xmit_timer(sk, TIME_KEEPOPEN, TCP_TIMEOUT_LEN);
3442 }
3443 return(1);
3444 }
3445
3446 /*
3447 * If there is data set flag 1
3448 */
3449
3450 if (len != th->doff*4)
3451 flag |= 1;
3452
3453 /*
3454 * See if our window has been shrunk.
3455 */
3456
3457 if (after(sk->window_seq, ack+ntohs(th->window)))
3458 {
3459 /*
3460 * We may need to move packets from the send queue
3461 * to the write queue, if the window has been shrunk on us.
3462 * The RFC says you are not allowed to shrink your window
3463 * like this, but if the other end does, you must be able
3464 * to deal with it.
3465 */
3466 struct sk_buff *skb;
3467 struct sk_buff *skb2;
3468 struct sk_buff *wskb = NULL;
3469
3470 skb2 = sk->send_head;
3471 sk->send_head = NULL;
3472 sk->send_tail = NULL;
3473
3474 /*
3475 * This is an artifact of a flawed concept. We want one
3476 * queue and a smarter send routine when we send all.
3477 */
3478
3479 flag |= 4; /* Window changed */
3480
3481 sk->window_seq = ack + ntohs(th->window);
3482 cli();
3483 while (skb2 != NULL)
3484 {
3485 skb = skb2;
3486 skb2 = skb->link3;
3487 skb->link3 = NULL;
3488 if (after(skb->h.seq, sk->window_seq))
3489 {
3490 if (sk->packets_out > 0)
3491 sk->packets_out--;
3492 /* We may need to remove this from the dev send list. */
3493 if (skb->next != NULL)
3494 {
3495 skb_unlink(skb);
3496 }
3497 /* Now add it to the write_queue. */
3498 if (wskb == NULL)
3499 skb_queue_head(&sk->write_queue,skb);
3500 else
3501 skb_append(wskb,skb);
3502 wskb = skb;
3503 }
3504 else
3505 {
3506 if (sk->send_head == NULL)
3507 {
3508 sk->send_head = skb;
3509 sk->send_tail = skb;
3510 }
3511 else
3512 {
3513 sk->send_tail->link3 = skb;
3514 sk->send_tail = skb;
3515 }
3516 skb->link3 = NULL;
3517 }
3518 }
3519 sti();
3520 }
3521
3522 /*
3523 * Pipe has emptied
3524 */
3525
3526 if (sk->send_tail == NULL || sk->send_head == NULL)
3527 {
3528 sk->send_head = NULL;
3529 sk->send_tail = NULL;
3530 sk->packets_out= 0;
3531 }
3532
3533 /*
3534 * Update the right hand window edge of the host
3535 */
3536
3537 sk->window_seq = ack + ntohs(th->window);
3538
3539 /*
3540 * We don't want too many packets out there.
3541 */
3542
3543 if (sk->ip_xmit_timeout == TIME_WRITE &&
3544 sk->cong_window < 2048 && after(ack, sk->rcv_ack_seq))
3545 {
3546 /*
3547 * This is Jacobson's slow start and congestion avoidance.
3548 * SIGCOMM '88, p. 328. Because we keep cong_window in integral
3549 * mss's, we can't do cwnd += 1 / cwnd. Instead, maintain a
3550 * counter and increment it once every cwnd times. It's possible
3551 * that this should be done only if sk->retransmits == 0. I'm
3552 * interpreting "new data is acked" as including data that has
3553 * been retransmitted but is just now being acked.
3554 */
3555 if (sk->cong_window < sk->ssthresh)
3556 /*
3557 * In "safe" area, increase
3558 */
3559 sk->cong_window++;
3560 else
3561 {
3562 /*
3563 * In dangerous area, increase slowly. In theory this is
3564 * sk->cong_window += 1 / sk->cong_window
3565 */
3566 if (sk->cong_count >= sk->cong_window)
3567 {
3568 sk->cong_window++;
3569 sk->cong_count = 0;
3570 }
3571 else
3572 sk->cong_count++;
3573 }
3574 }
3575
3576 /*
3577 * Remember the highest ack received.
3578 */
3579
3580 sk->rcv_ack_seq = ack;
3581
3582 /*
3583 * If this ack opens up a zero window, clear backoff. It was
3584 * being used to time the probes, and is probably far higher than
3585 * it needs to be for normal retransmission.
3586 */
3587
3588 if (sk->ip_xmit_timeout == TIME_PROBE0)
3589 {
3590 sk->retransmits = 0; /* Our probe was answered */
3591
3592 /*
3593 * Was it a usable window open ?
3594 */
3595
3596 if (skb_peek(&sk->write_queue) != NULL && /* should always be non-null */
3597 ! before (sk->window_seq, sk->write_queue.next->h.seq))
3598 {
3599 sk->backoff = 0;
3600
3601 /*
3602 * Recompute rto from rtt. this eliminates any backoff.
3603 */
3604
3605 sk->rto = ((sk->rtt >> 2) + sk->mdev) >> 1;
3606 if (sk->rto > 120*HZ)
3607 sk->rto = 120*HZ;
3608 if (sk->rto < 20) /* Was 1*HZ, then 1 - turns out we must allow about
3609 .2 of a second because of BSD delayed acks - on a 100Mb/sec link
3610 .2 of a second is going to need huge windows (SIGH) */
3611 sk->rto = 20;
3612 }
3613 }
3614
3615 /*
3616 * See if we can take anything off of the retransmit queue.
3617 */
3618
3619 while(sk->send_head != NULL)
3620 {
3621 /* Check for a bug. */
3622 if (sk->send_head->link3 &&
3623 after(sk->send_head->h.seq, sk->send_head->link3->h.seq))
3624 printk("INET: tcp.c: *** bug send_list out of order.\n");
3625
3626 /*
3627 * If our packet is before the ack sequence we can
3628 * discard it as it's confirmed to have arrived the other end.
3629 */
3630
3631 if (before(sk->send_head->h.seq, ack+1))
3632 {
3633 struct sk_buff *oskb;
3634 if (sk->retransmits)
3635 {
3636 /*
3637 * We were retransmitting. don't count this in RTT est
3638 */
3639 flag |= 2;
3640
3641 /*
3642 * even though we've gotten an ack, we're still
3643 * retransmitting as long as we're sending from
3644 * the retransmit queue. Keeping retransmits non-zero
3645 * prevents us from getting new data interspersed with
3646 * retransmissions.
3647 */
3648
3649 if (sk->send_head->link3) /* Any more queued retransmits? */
3650 sk->retransmits = 1;
3651 else
3652 sk->retransmits = 0;
3653 }
3654 /*
3655 * Note that we only reset backoff and rto in the
3656 * rtt recomputation code. And that doesn't happen
3657 * if there were retransmissions in effect. So the
3658 * first new packet after the retransmissions is
3659 * sent with the backoff still in effect. Not until
3660 * we get an ack from a non-retransmitted packet do
3661 * we reset the backoff and rto. This allows us to deal
3662 * with a situation where the network delay has increased
3663 * suddenly. I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3664 */
3665
3666 /*
3667 * We have one less packet out there.
3668 */
3669
3670 if (sk->packets_out > 0)
3671 sk->packets_out --;
3672 /*
3673 * Wake up the process, it can probably write more.
3674 */
3675 if (!sk->dead)
3676 sk->write_space(sk);
3677 oskb = sk->send_head;
3678
3679 if (!(flag&2)) /* Not retransmitting */
3680 {
3681 long m;
3682
3683 /*
3684 * The following amusing code comes from Jacobson's
3685 * article in SIGCOMM '88. Note that rtt and mdev
3686 * are scaled versions of rtt and mean deviation.
3687 * This is designed to be as fast as possible
3688 * m stands for "measurement".
3689 */
3690
3691 m = jiffies - oskb->when; /* RTT */
3692 if(m<=0)
3693 m=1; /* IS THIS RIGHT FOR <0 ??? */
3694 m -= (sk->rtt >> 3); /* m is now error in rtt est */
3695 sk->rtt += m; /* rtt = 7/8 rtt + 1/8 new */
3696 if (m < 0)
3697 m = -m; /* m is now abs(error) */
3698 m -= (sk->mdev >> 2); /* similar update on mdev */
3699 sk->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
3700
3701 /*
3702 * Now update timeout. Note that this removes any backoff.
3703 */
3704
3705 sk->rto = ((sk->rtt >> 2) + sk->mdev) >> 1;
3706 if (sk->rto > 120*HZ)
3707 sk->rto = 120*HZ;
3708 if (sk->rto < 20) /* Was 1*HZ - keep .2 as minimum cos of the BSD delayed acks */
3709 sk->rto = 20;
3710 sk->backoff = 0;
3711 }
3712 flag |= (2|4); /* 2 is really more like 'don't adjust the rtt
3713 In this case as we just set it up */
3714 cli();
3715 oskb = sk->send_head;
3716 IS_SKB(oskb);
3717 sk->send_head = oskb->link3;
3718 if (sk->send_head == NULL)
3719 {
3720 sk->send_tail = NULL;
3721 }
3722
3723 /*
3724 * We may need to remove this from the dev send list.
3725 */
3726
3727 if (oskb->next)
3728 skb_unlink(oskb);
3729 sti();
3730 kfree_skb(oskb, FREE_WRITE); /* write. */
3731 if (!sk->dead)
3732 sk->write_space(sk);
3733 }
3734 else
3735 {
3736 break;
3737 }
3738 }
3739
3740 /*
3741 * XXX someone ought to look at this too.. at the moment, if skb_peek()
3742 * returns non-NULL, we complete ignore the timer stuff in the else
3743 * clause. We ought to organize the code so that else clause can
3744 * (should) be executed regardless, possibly moving the PROBE timer
3745 * reset over. The skb_peek() thing should only move stuff to the
3746 * write queue, NOT also manage the timer functions.
3747 */
3748
3749 /*
3750 * Maybe we can take some stuff off of the write queue,
3751 * and put it onto the xmit queue.
3752 */
3753 if (skb_peek(&sk->write_queue) != NULL)
3754 {
3755 if (after (sk->window_seq+1, sk->write_queue.next->h.seq) &&
3756 (sk->retransmits == 0 ||
3757 sk->ip_xmit_timeout != TIME_WRITE ||
3758 before(sk->write_queue.next->h.seq, sk->rcv_ack_seq + 1))
3759 && sk->packets_out < sk->cong_window)
3760 {
3761 /*
3762 * Add more data to the send queue.
3763 */
3764 flag |= 1;
3765 tcp_write_xmit(sk);
3766 }
3767 else if (before(sk->window_seq, sk->write_queue.next->h.seq) &&
3768 sk->send_head == NULL &&
3769 sk->ack_backlog == 0 &&
3770 sk->state != TCP_TIME_WAIT)
3771 {
3772 /*
3773 * Data to queue but no room.
3774 */
3775 reset_xmit_timer(sk, TIME_PROBE0, sk->rto);
3776 }
3777 }
3778 else
3779 {
3780 /*
3781 * from TIME_WAIT we stay in TIME_WAIT as long as we rx packets
3782 * from TCP_CLOSE we don't do anything
3783 *
3784 * from anything else, if there is write data (or fin) pending,
3785 * we use a TIME_WRITE timeout, else if keepalive we reset to
3786 * a KEEPALIVE timeout, else we delete the timer.
3787 *
3788 * We do not set flag for nominal write data, otherwise we may
3789 * force a state where we start to write itsy bitsy tidbits
3790 * of data.
3791 */
3792
3793 switch(sk->state) {
3794 case TCP_TIME_WAIT:
3795 /*
3796 * keep us in TIME_WAIT until we stop getting packets,
3797 * reset the timeout.
3798 */
3799 reset_msl_timer(sk, TIME_CLOSE, TCP_TIMEWAIT_LEN);
3800 break;
3801 case TCP_CLOSE:
3802 /*
3803 * don't touch the timer.
3804 */
3805 break;
3806 default:
3807 /*
3808 * Must check send_head, write_queue, and ack_backlog
3809 * to determine which timeout to use.
3810 */
3811 if (sk->send_head || skb_peek(&sk->write_queue) != NULL || sk->ack_backlog) {
3812 reset_xmit_timer(sk, TIME_WRITE, sk->rto);
3813 } else if (sk->keepopen) {
3814 reset_xmit_timer(sk, TIME_KEEPOPEN, TCP_TIMEOUT_LEN);
3815 } else {
3816 del_timer(&sk->retransmit_timer);
3817 sk->ip_xmit_timeout = 0;
3818 }
3819 break;
3820 }
3821 }
3822
3823 /*
3824 * We have nothing queued but space to send. Send any partial
3825 * packets immediately (end of Nagle rule application).
3826 */
3827
3828 if (sk->packets_out == 0 && sk->partial != NULL &&
3829 skb_peek(&sk->write_queue) == NULL && sk->send_head == NULL)
3830 {
3831 flag |= 1;
3832 tcp_send_partial(sk);
3833 }
3834
3835 /*
3836 * In the LAST_ACK case, the other end FIN'd us. We then FIN'd them, and
3837 * we are now waiting for an acknowledge to our FIN. The other end is
3838 * already in TIME_WAIT.
3839 *
3840 * Move to TCP_CLOSE on success.
3841 */
3842
3843 if (sk->state == TCP_LAST_ACK)
3844 {
3845 if (!sk->dead)
3846 sk->state_change(sk);
3847 if(sk->debug)
3848 printk("rcv_ack_seq: %X==%X, acked_seq: %X==%X\n",
3849 sk->rcv_ack_seq,sk->write_seq,sk->acked_seq,sk->fin_seq);
3850 if (sk->rcv_ack_seq == sk->write_seq /*&& sk->acked_seq == sk->fin_seq*/)
3851 {
3852 flag |= 1;
3853 tcp_set_state(sk,TCP_CLOSE);
3854 sk->shutdown = SHUTDOWN_MASK;
3855 }
3856 }
3857
3858 /*
3859 * Incoming ACK to a FIN we sent in the case of our initiating the close.
3860 *
3861 * Move to FIN_WAIT2 to await a FIN from the other end. Set
3862 * SEND_SHUTDOWN but not RCV_SHUTDOWN as data can still be coming in.
3863 */
3864
3865 if (sk->state == TCP_FIN_WAIT1)
3866 {
3867
3868 if (!sk->dead)
3869 sk->state_change(sk);
3870 if (sk->rcv_ack_seq == sk->write_seq)
3871 {
3872 flag |= 1;
3873 sk->shutdown |= SEND_SHUTDOWN;
3874 tcp_set_state(sk, TCP_FIN_WAIT2);
3875 }
3876 }
3877
3878 /*
3879 * Incoming ACK to a FIN we sent in the case of a simultaneous close.
3880 *
3881 * Move to TIME_WAIT
3882 */
3883
3884 if (sk->state == TCP_CLOSING)
3885 {
3886
3887 if (!sk->dead)
3888 sk->state_change(sk);
3889 if (sk->rcv_ack_seq == sk->write_seq)
3890 {
3891 flag |= 1;
3892 tcp_time_wait(sk);
3893 }
3894 }
3895
3896 /*
3897 * Final ack of a three way shake
3898 */
3899
3900 if(sk->state==TCP_SYN_RECV)
3901 {
3902 tcp_set_state(sk, TCP_ESTABLISHED);
3903 tcp_options(sk,th);
3904 sk->dummy_th.dest=th->source;
3905 sk->copied_seq = sk->acked_seq;
3906 if(!sk->dead)
3907 sk->state_change(sk);
3908 if(sk->max_window==0)
3909 {
3910 sk->max_window=32; /* Sanity check */
3911 sk->mss=min(sk->max_window,sk->mtu);
3912 }
3913 }
3914
3915 /*
3916 * I make no guarantees about the first clause in the following
3917 * test, i.e. "(!flag) || (flag&4)". I'm not entirely sure under
3918 * what conditions "!flag" would be true. However I think the rest
3919 * of the conditions would prevent that from causing any
3920 * unnecessary retransmission.
3921 * Clearly if the first packet has expired it should be
3922 * retransmitted. The other alternative, "flag&2 && retransmits", is
3923 * harder to explain: You have to look carefully at how and when the
3924 * timer is set and with what timeout. The most recent transmission always
3925 * sets the timer. So in general if the most recent thing has timed
3926 * out, everything before it has as well. So we want to go ahead and
3927 * retransmit some more. If we didn't explicitly test for this
3928 * condition with "flag&2 && retransmits", chances are "when + rto < jiffies"
3929 * would not be true. If you look at the pattern of timing, you can
3930 * show that rto is increased fast enough that the next packet would
3931 * almost never be retransmitted immediately. Then you'd end up
3932 * waiting for a timeout to send each packet on the retransmission
3933 * queue. With my implementation of the Karn sampling algorithm,
3934 * the timeout would double each time. The net result is that it would
3935 * take a hideous amount of time to recover from a single dropped packet.
3936 * It's possible that there should also be a test for TIME_WRITE, but
3937 * I think as long as "send_head != NULL" and "retransmit" is on, we've
3938 * got to be in real retransmission mode.
3939 * Note that tcp_do_retransmit is called with all==1. Setting cong_window
3940 * back to 1 at the timeout will cause us to send 1, then 2, etc. packets.
3941 * As long as no further losses occur, this seems reasonable.
3942 */
3943
3944 if (((!flag) || (flag&4)) && sk->send_head != NULL &&
3945 (((flag&2) && sk->retransmits) ||
3946 (sk->send_head->when + sk->rto < jiffies)))
3947 {
3948 if(sk->send_head->when + sk->rto < jiffies)
3949 tcp_retransmit(sk,0);
3950 else
3951 {
3952 tcp_do_retransmit(sk, 1);
3953 reset_xmit_timer(sk, TIME_WRITE, sk->rto);
3954 }
3955 }
3956
3957 return(1);
3958 }
3959
3960
3961 /*
3962 * Process the FIN bit. This now behaves as it is supposed to work
3963 * and the FIN takes effect when it is validly part of sequence
3964 * space. Not before when we get holes.
3965 *
3966 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3967 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3968 * TIME-WAIT)
3969 *
3970 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3971 * close and we go into CLOSING (and later onto TIME-WAIT)
3972 *
3973 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3974 *
3975 */
3976
3977 static int tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
/* */
3978 {
3979 sk->fin_seq = th->seq + skb->len + th->syn + th->fin;
3980
3981 if (!sk->dead)
3982 {
3983 sk->state_change(sk);
3984 sock_wake_async(sk->socket, 1);
3985 }
3986
3987 switch(sk->state)
3988 {
3989 case TCP_SYN_RECV:
3990 case TCP_SYN_SENT:
3991 case TCP_ESTABLISHED:
3992 /*
3993 * move to CLOSE_WAIT, tcp_data() already handled
3994 * sending the ack.
3995 */
3996 tcp_set_state(sk,TCP_CLOSE_WAIT);
3997 if (th->rst)
3998 sk->shutdown = SHUTDOWN_MASK;
3999 break;
4000
4001 case TCP_CLOSE_WAIT:
4002 case TCP_CLOSING:
4003 /*
4004 * received a retransmission of the FIN, do
4005 * nothing.
4006 */
4007 break;
4008 case TCP_TIME_WAIT:
4009 /*
4010 * received a retransmission of the FIN,
4011 * restart the TIME_WAIT timer.
4012 */
4013 reset_msl_timer(sk, TIME_CLOSE, TCP_TIMEWAIT_LEN);
4014 return(0);
4015 case TCP_FIN_WAIT1:
4016 /*
4017 * This case occurs when a simultaneous close
4018 * happens, we must ack the received FIN and
4019 * enter the CLOSING state.
4020 *
4021 * This causes a WRITE timeout, which will either
4022 * move on to TIME_WAIT when we timeout, or resend
4023 * the FIN properly (maybe we get rid of that annoying
4024 * FIN lost hang). The TIME_WRITE code is already correct
4025 * for handling this timeout.
4026 */
4027
4028 if(sk->ip_xmit_timeout != TIME_WRITE)
4029 reset_xmit_timer(sk, TIME_WRITE, sk->rto);
4030 tcp_set_state(sk,TCP_CLOSING);
4031 break;
4032 case TCP_FIN_WAIT2:
4033 /*
4034 * received a FIN -- send ACK and enter TIME_WAIT
4035 */
4036 reset_msl_timer(sk, TIME_CLOSE, TCP_TIMEWAIT_LEN);
4037 sk->shutdown|=SHUTDOWN_MASK;
4038 tcp_set_state(sk,TCP_TIME_WAIT);
4039 break;
4040 case TCP_CLOSE:
4041 /*
4042 * already in CLOSE
4043 */
4044 break;
4045 default:
4046 tcp_set_state(sk,TCP_LAST_ACK);
4047
4048 /* Start the timers. */
4049 reset_msl_timer(sk, TIME_CLOSE, TCP_TIMEWAIT_LEN);
4050 return(0);
4051 }
4052
4053 return(0);
4054 }
4055
4056
4057
4058 /*
4059 * This routine handles the data. If there is room in the buffer,
4060 * it will be have already been moved into it. If there is no
4061 * room, then we will just have to discard the packet.
4062 */
4063
4064 extern __inline__ int tcp_data(struct sk_buff *skb, struct sock *sk,
/* */
4065 unsigned long saddr, unsigned short len)
4066 {
4067 struct sk_buff *skb1, *skb2;
4068 struct tcphdr *th;
4069 int dup_dumped=0;
4070 u32 new_seq, shut_seq;
4071
4072 th = skb->h.th;
4073 skb_pull(skb,th->doff*4);
4074 skb_trim(skb,len-(th->doff*4));
4075
4076 /*
4077 * The bytes in the receive read/assembly queue has increased. Needed for the
4078 * low memory discard algorithm
4079 */
4080
4081 sk->bytes_rcv += skb->len;
4082
4083 if (skb->len == 0 && !th->fin)
4084 {
4085 /*
4086 * Don't want to keep passing ack's back and forth.
4087 * (someone sent us dataless, boring frame)
4088 */
4089 if (!th->ack)
4090 tcp_send_ack(sk->sent_seq, sk->acked_seq,sk, th, saddr);
4091 kfree_skb(skb, FREE_READ);
4092 return(0);
4093 }
4094
4095 /*
4096 * We no longer have anyone receiving data on this connection.
4097 */
4098
4099 #ifndef TCP_DONT_RST_SHUTDOWN
4100
4101 if(sk->shutdown & RCV_SHUTDOWN)
4102 {
4103 /*
4104 * FIXME: BSD has some magic to avoid sending resets to
4105 * broken 4.2 BSD keepalives. Much to my surprise a few non
4106 * BSD stacks still have broken keepalives so we want to
4107 * cope with it.
4108 */
4109
4110 if(skb->len) /* We don't care if it's just an ack or
4111 a keepalive/window probe */
4112 {
4113 new_seq= th->seq + skb->len + th->syn; /* Right edge of _data_ part of frame */
4114
4115 /* Do this the way 4.4BSD treats it. Not what I'd
4116 regard as the meaning of the spec but it's what BSD
4117 does and clearly they know everything 8) */
4118
4119 /*
4120 * This is valid because of two things
4121 *
4122 * a) The way tcp_data behaves at the bottom.
4123 * b) A fin takes effect when read not when received.
4124 */
4125
4126 shut_seq=sk->acked_seq+1; /* Last byte */
4127
4128 if(after(new_seq,shut_seq))
4129 {
4130 if(sk->debug)
4131 printk("Data arrived on %p after close [Data right edge %X, Socket shut on %X] %d\n",
4132 sk, new_seq, shut_seq, sk->blog);
4133 if(sk->dead)
4134 {
4135 sk->acked_seq = new_seq + th->fin;
4136 tcp_reset(sk->saddr, sk->daddr, skb->h.th,
4137 sk->prot, NULL, skb->dev, sk->ip_tos, sk->ip_ttl);
4138 tcp_statistics.TcpEstabResets++;
4139 tcp_set_state(sk,TCP_CLOSE);
4140 sk->err = EPIPE;
4141 sk->shutdown = SHUTDOWN_MASK;
4142 kfree_skb(skb, FREE_READ);
4143 return 0;
4144 }
4145 }
4146 }
4147 }
4148
4149 #endif
4150
4151 /*
4152 * Now we have to walk the chain, and figure out where this one
4153 * goes into it. This is set up so that the last packet we received
4154 * will be the first one we look at, that way if everything comes
4155 * in order, there will be no performance loss, and if they come
4156 * out of order we will be able to fit things in nicely.
4157 *
4158 * [AC: This is wrong. We should assume in order first and then walk
4159 * forwards from the first hole based upon real traffic patterns.]
4160 *
4161 */
4162
4163 if (skb_peek(&sk->receive_queue) == NULL) /* Empty queue is easy case */
4164 {
4165 skb_queue_head(&sk->receive_queue,skb);
4166 skb1= NULL;
4167 }
4168 else
4169 {
4170 for(skb1=sk->receive_queue.prev; ; skb1 = skb1->prev)
4171 {
4172 if(sk->debug)
4173 {
4174 printk("skb1=%p :", skb1);
4175 printk("skb1->h.th->seq = %d: ", skb1->h.th->seq);
4176 printk("skb->h.th->seq = %d\n",skb->h.th->seq);
4177 printk("copied_seq = %d acked_seq = %d\n", sk->copied_seq,
4178 sk->acked_seq);
4179 }
4180
4181 /*
4182 * Optimisation: Duplicate frame or extension of previous frame from
4183 * same sequence point (lost ack case).
4184 * The frame contains duplicate data or replaces a previous frame
4185 * discard the previous frame (safe as sk->inuse is set) and put
4186 * the new one in its place.
4187 */
4188
4189 if (th->seq==skb1->h.th->seq && skb->len>= skb1->len)
4190 {
4191 skb_append(skb1,skb);
4192 skb_unlink(skb1);
4193 kfree_skb(skb1,FREE_READ);
4194 dup_dumped=1;
4195 skb1=NULL;
4196 break;
4197 }
4198
4199 /*
4200 * Found where it fits
4201 */
4202
4203 if (after(th->seq+1, skb1->h.th->seq))
4204 {
4205 skb_append(skb1,skb);
4206 break;
4207 }
4208
4209 /*
4210 * See if we've hit the start. If so insert.
4211 */
4212 if (skb1 == skb_peek(&sk->receive_queue))
4213 {
4214 skb_queue_head(&sk->receive_queue, skb);
4215 break;
4216 }
4217 }
4218 }
4219
4220 /*
4221 * Figure out what the ack value for this frame is
4222 */
4223
4224 th->ack_seq = th->seq + skb->len;
4225 if (th->syn)
4226 th->ack_seq++;
4227 if (th->fin)
4228 th->ack_seq++;
4229
4230 if (before(sk->acked_seq, sk->copied_seq))
4231 {
4232 printk("*** tcp.c:tcp_data bug acked < copied\n");
4233 sk->acked_seq = sk->copied_seq;
4234 }
4235
4236 /*
4237 * Now figure out if we can ack anything. This is very messy because we really want two
4238 * receive queues, a completed and an assembly queue. We also want only one transmit
4239 * queue.
4240 */
4241
4242 if ((!dup_dumped && (skb1 == NULL || skb1->acked)) || before(th->seq, sk->acked_seq+1))
4243 {
4244 if (before(th->seq, sk->acked_seq+1))
4245 {
4246 int newwindow;
4247
4248 if (after(th->ack_seq, sk->acked_seq))
4249 {
4250 newwindow = sk->window-(th->ack_seq - sk->acked_seq);
4251 if (newwindow < 0)
4252 newwindow = 0;
4253 sk->window = newwindow;
4254 sk->acked_seq = th->ack_seq;
4255 }
4256 skb->acked = 1;
4257
4258 /*
4259 * When we ack the fin, we do the FIN
4260 * processing.
4261 */
4262
4263 if (skb->h.th->fin)
4264 {
4265 tcp_fin(skb,sk,skb->h.th);
4266 }
4267
4268 for(skb2 = skb->next;
4269 skb2 != (struct sk_buff *)&sk->receive_queue;
4270 skb2 = skb2->next)
4271 {
4272 if (before(skb2->h.th->seq, sk->acked_seq+1))
4273 {
4274 if (after(skb2->h.th->ack_seq, sk->acked_seq))
4275 {
4276 newwindow = sk->window -
4277 (skb2->h.th->ack_seq - sk->acked_seq);
4278 if (newwindow < 0)
4279 newwindow = 0;
4280 sk->window = newwindow;
4281 sk->acked_seq = skb2->h.th->ack_seq;
4282 }
4283 skb2->acked = 1;
4284 /*
4285 * When we ack the fin, we do
4286 * the fin handling.
4287 */
4288 if (skb2->h.th->fin)
4289 {
4290 tcp_fin(skb,sk,skb->h.th);
4291 }
4292
4293 /*
4294 * Force an immediate ack.
4295 */
4296
4297 sk->ack_backlog = sk->max_ack_backlog;
4298 }
4299 else
4300 {
4301 break;
4302 }
4303 }
4304
4305 /*
4306 * This also takes care of updating the window.
4307 * This if statement needs to be simplified.
4308 */
4309 if (!sk->delay_acks ||
4310 sk->ack_backlog >= sk->max_ack_backlog ||
4311 sk->bytes_rcv > sk->max_unacked || th->fin) {
4312 /* tcp_send_ack(sk->sent_seq, sk->acked_seq,sk,th, saddr); */
4313 }
4314 else
4315 {
4316 sk->ack_backlog++;
4317 if(sk->debug)
4318 printk("Ack queued.\n");
4319 reset_xmit_timer(sk, TIME_WRITE, TCP_ACK_TIME);
4320 }
4321 }
4322 }
4323
4324 /*
4325 * If we've missed a packet, send an ack.
4326 * Also start a timer to send another.
4327 */
4328
4329 if (!skb->acked)
4330 {
4331
4332 /*
4333 * This is important. If we don't have much room left,
4334 * we need to throw out a few packets so we have a good
4335 * window. Note that mtu is used, not mss, because mss is really
4336 * for the send side. He could be sending us stuff as large as mtu.
4337 */
4338
4339 while (sock_rspace(sk) < sk->mtu)
4340 {
4341 skb1 = skb_peek(&sk->receive_queue);
4342 if (skb1 == NULL)
4343 {
4344 printk("INET: tcp.c:tcp_data memory leak detected.\n");
4345 break;
4346 }
4347
4348 /*
4349 * Don't throw out something that has been acked.
4350 */
4351
4352 if (skb1->acked)
4353 {
4354 break;
4355 }
4356
4357 skb_unlink(skb1);
4358 kfree_skb(skb1, FREE_READ);
4359 }
4360 tcp_send_ack(sk->sent_seq, sk->acked_seq, sk, th, saddr);
4361 sk->ack_backlog++;
4362 reset_xmit_timer(sk, TIME_WRITE, TCP_ACK_TIME);
4363 }
4364 else
4365 {
4366 tcp_send_ack(sk->sent_seq, sk->acked_seq, sk, th, saddr);
4367 }
4368
4369 /*
4370 * Now tell the user we may have some data.
4371 */
4372
4373 if (!sk->dead)
4374 {
4375 if(sk->debug)
4376 printk("Data wakeup.\n");
4377 sk->data_ready(sk,0);
4378 }
4379 return(0);
4380 }
4381
4382
4383 /*
4384 * This routine is only called when we have urgent data
4385 * signalled. Its the 'slow' part of tcp_urg. It could be
4386 * moved inline now as tcp_urg is only called from one
4387 * place. We handle URGent data wrong. We have to - as
4388 * BSD still doesn't use the correction from RFC961.
4389 */
4390
4391 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
/* */
4392 {
4393 u32 ptr = ntohs(th->urg_ptr);
4394
4395 if (ptr)
4396 ptr--;
4397 ptr += th->seq;
4398
4399 /* ignore urgent data that we've already seen and read */
4400 if (after(sk->copied_seq, ptr))
4401 return;
4402
4403 /* do we already have a newer (or duplicate) urgent pointer? */
4404 if (sk->urg_data && !after(ptr, sk->urg_seq))
4405 return;
4406
4407 /* tell the world about our new urgent pointer */
4408 if (sk->proc != 0) {
4409 if (sk->proc > 0) {
4410 kill_proc(sk->proc, SIGURG, 1);
4411 } else {
4412 kill_pg(-sk->proc, SIGURG, 1);
4413 }
4414 }
4415 sk->urg_data = URG_NOTYET;
4416 sk->urg_seq = ptr;
4417 }
4418
4419 /*
4420 * This is the 'fast' part of urgent handling.
4421 */
4422
4423 extern __inline__ int tcp_urg(struct sock *sk, struct tcphdr *th,
/* */
4424 unsigned long saddr, unsigned long len)
4425 {
4426 u32 ptr;
4427
4428 /*
4429 * Check if we get a new urgent pointer - normally not
4430 */
4431
4432 if (th->urg)
4433 tcp_check_urg(sk,th);
4434
4435 /*
4436 * Do we wait for any urgent data? - normally not
4437 */
4438
4439 if (sk->urg_data != URG_NOTYET)
4440 return 0;
4441
4442 /*
4443 * Is the urgent pointer pointing into this packet?
4444 */
4445
4446 ptr = sk->urg_seq - th->seq + th->doff*4;
4447 if (ptr >= len)
4448 return 0;
4449
4450 /*
4451 * Ok, got the correct packet, update info
4452 */
4453
4454 sk->urg_data = URG_VALID | *(ptr + (unsigned char *) th);
4455 if (!sk->dead)
4456 sk->data_ready(sk,0);
4457 return 0;
4458 }
4459
4460 /*
4461 * This will accept the next outstanding connection.
4462 */
4463
4464 static struct sock *tcp_accept(struct sock *sk, int flags)
/* */
4465 {
4466 struct sock *newsk;
4467 struct sk_buff *skb;
4468
4469 /*
4470 * We need to make sure that this socket is listening,
4471 * and that it has something pending.
4472 */
4473
4474 if (sk->state != TCP_LISTEN)
4475 {
4476 sk->err = EINVAL;
4477 return(NULL);
4478 }
4479
4480 /* Avoid the race. */
4481 cli();
4482 sk->inuse = 1;
4483
4484 while((skb = tcp_dequeue_established(sk)) == NULL)
4485 {
4486 if (flags & O_NONBLOCK)
4487 {
4488 sti();
4489 release_sock(sk);
4490 sk->err = EAGAIN;
4491 return(NULL);
4492 }
4493
4494 release_sock(sk);
4495 interruptible_sleep_on(sk->sleep);
4496 if (current->signal & ~current->blocked)
4497 {
4498 sti();
4499 sk->err = ERESTARTSYS;
4500 return(NULL);
4501 }
4502 sk->inuse = 1;
4503 }
4504 sti();
4505
4506 /*
4507 * Now all we need to do is return skb->sk.
4508 */
4509
4510 newsk = skb->sk;
4511
4512 kfree_skb(skb, FREE_READ);
4513 sk->ack_backlog--;
4514 release_sock(sk);
4515 return(newsk);
4516 }
4517
4518
4519 /*
4520 * This will initiate an outgoing connection.
4521 */
4522
4523 static int tcp_connect(struct sock *sk, struct sockaddr_in *usin, int addr_len)
/* */
4524 {
4525 struct sk_buff *buff;
4526 struct device *dev=NULL;
4527 unsigned char *ptr;
4528 int tmp;
4529 int atype;
4530 struct tcphdr *t1;
4531 struct rtable *rt;
4532
4533 if (sk->state != TCP_CLOSE)
4534 {
4535 return(-EISCONN);
4536 }
4537
4538 if (addr_len < 8)
4539 return(-EINVAL);
4540
4541 if (usin->sin_family && usin->sin_family != AF_INET)
4542 return(-EAFNOSUPPORT);
4543
4544 /*
4545 * connect() to INADDR_ANY means loopback (BSD'ism).
4546 */
4547
4548 if(usin->sin_addr.s_addr==INADDR_ANY)
4549 usin->sin_addr.s_addr=ip_my_addr();
4550
4551 /*
4552 * Don't want a TCP connection going to a broadcast address
4553 */
4554
4555 if ((atype=ip_chk_addr(usin->sin_addr.s_addr)) == IS_BROADCAST || atype==IS_MULTICAST)
4556 return -ENETUNREACH;
4557
4558 sk->inuse = 1;
4559 sk->daddr = usin->sin_addr.s_addr;
4560 sk->write_seq = tcp_init_seq();
4561 sk->window_seq = sk->write_seq;
4562 sk->rcv_ack_seq = sk->write_seq -1;
4563 sk->err = 0;
4564 sk->dummy_th.dest = usin->sin_port;
4565 release_sock(sk);
4566
4567 buff = sock_wmalloc(sk,MAX_SYN_SIZE,0, GFP_KERNEL);
4568 if (buff == NULL)
4569 {
4570 return(-ENOMEM);
4571 }
4572 sk->inuse = 1;
4573 buff->sk = sk;
4574 buff->free = 0;
4575 buff->localroute = sk->localroute;
4576
4577
4578 /*
4579 * Put in the IP header and routing stuff.
4580 */
4581
4582 if (sk->localroute)
4583 rt=ip_rt_local(sk->daddr, NULL, sk->saddr ? NULL : &sk->saddr);
4584 else
4585 rt=ip_rt_route(sk->daddr, NULL, sk->saddr ? NULL : &sk->saddr);
4586
4587 /*
4588 * When we connect we enforce receive requirements too.
4589 */
4590
4591 sk->rcv_saddr=sk->saddr;
4592
4593 /*
4594 * We need to build the routing stuff from the things saved in skb.
4595 */
4596
4597 tmp = sk->prot->build_header(buff, sk->saddr, sk->daddr, &dev,
4598 IPPROTO_TCP, NULL, MAX_SYN_SIZE,sk->ip_tos,sk->ip_ttl);
4599 if (tmp < 0)
4600 {
4601 sock_wfree(sk, buff);
4602 release_sock(sk);
4603 return(-ENETUNREACH);
4604 }
4605
4606 t1 = (struct tcphdr *) skb_put(buff,sizeof(struct tcphdr));
4607
4608 memcpy(t1,(void *)&(sk->dummy_th), sizeof(*t1));
4609 t1->seq = ntohl(sk->write_seq++);
4610 sk->sent_seq = sk->write_seq;
4611 buff->h.seq = sk->write_seq;
4612 t1->ack = 0;
4613 t1->window = 2;
4614 t1->res1=0;
4615 t1->res2=0;
4616 t1->rst = 0;
4617 t1->urg = 0;
4618 t1->psh = 0;
4619 t1->syn = 1;
4620 t1->urg_ptr = 0;
4621 t1->doff = 6;
4622 /* use 512 or whatever user asked for */
4623
4624 if(rt!=NULL && (rt->rt_flags&RTF_WINDOW))
4625 sk->window_clamp=rt->rt_window;
4626 else
4627 sk->window_clamp=0;
4628
4629 if (sk->user_mss)
4630 sk->mtu = sk->user_mss;
4631 else if(rt!=NULL && (rt->rt_flags&RTF_MSS))
4632 sk->mtu = rt->rt_mss;
4633 else
4634 {
4635 #ifdef CONFIG_INET_SNARL
4636 if ((sk->saddr ^ sk->daddr) & default_mask(sk->saddr))
4637 #else
4638 if ((sk->saddr ^ sk->daddr) & dev->pa_mask)
4639 #endif
4640 sk->mtu = 576 - sizeof(struct iphdr) - sizeof(struct tcphdr);
4641 else
4642 sk->mtu = MAX_WINDOW;
4643 }
4644 /*
4645 * but not bigger than device MTU
4646 */
4647
4648 if(sk->mtu <32)
4649 sk->mtu = 32; /* Sanity limit */
4650
4651 sk->mtu = min(sk->mtu, dev->mtu - sizeof(struct iphdr) - sizeof(struct tcphdr));
4652
4653 /*
4654 * Put in the TCP options to say MTU.
4655 */
4656
4657 ptr = skb_put(buff,4);
4658 ptr[0] = 2;
4659 ptr[1] = 4;
4660 ptr[2] = (sk->mtu) >> 8;
4661 ptr[3] = (sk->mtu) & 0xff;
4662 tcp_send_check(t1, sk->saddr, sk->daddr,
4663 sizeof(struct tcphdr) + 4, sk);
4664
4665 /*
4666 * This must go first otherwise a really quick response will get reset.
4667 */
4668
4669 tcp_cache_zap();
4670 tcp_set_state(sk,TCP_SYN_SENT);
4671 if(rt&&rt->rt_flags&RTF_IRTT)
4672 sk->rto = rt->rt_irtt;
4673 else
4674 sk->rto = TCP_TIMEOUT_INIT;
4675 sk->retransmit_timer.function=&retransmit_timer;
4676 sk->retransmit_timer.data = (unsigned long)sk;
4677 reset_xmit_timer(sk, TIME_WRITE, sk->rto); /* Timer for repeating the SYN until an answer */
4678 sk->retransmits = 0; /* Now works the right way instead of a hacked initial setting */
4679
4680 sk->prot->queue_xmit(sk, dev, buff, 0);
4681 reset_xmit_timer(sk, TIME_WRITE, sk->rto);
4682 tcp_statistics.TcpActiveOpens++;
4683 tcp_statistics.TcpOutSegs++;
4684
4685 release_sock(sk);
4686 return(0);
4687 }
4688
4689
4690 /* This functions checks to see if the tcp header is actually acceptable. */
4691 extern __inline__ int tcp_sequence(struct sock *sk, struct tcphdr *th, short len,
/* */
4692 struct options *opt, unsigned long saddr, struct device *dev)
4693 {
4694 u32 next_seq;
4695
4696 next_seq = len - 4*th->doff;
4697 if (th->fin)
4698 next_seq++;
4699 /* if we have a zero window, we can't have any data in the packet.. */
4700 if (next_seq && !sk->window)
4701 goto ignore_it;
4702 next_seq += th->seq;
4703
4704 /*
4705 * This isn't quite right. sk->acked_seq could be more recent
4706 * than sk->window. This is however close enough. We will accept
4707 * slightly more packets than we should, but it should not cause
4708 * problems unless someone is trying to forge packets.
4709 */
4710
4711 /* have we already seen all of this packet? */
4712 if (!after(next_seq+1, sk->acked_seq))
4713 goto ignore_it;
4714 /* or does it start beyond the window? */
4715 if (!before(th->seq, sk->acked_seq + sk->window + 1))
4716 goto ignore_it;
4717
4718 /* ok, at least part of this packet would seem interesting.. */
4719 return 1;
4720
4721 ignore_it:
4722 if (th->rst)
4723 return 0;
4724
4725 /*
4726 * Send a reset if we get something not ours and we are
4727 * unsynchronized. Note: We don't do anything to our end. We
4728 * are just killing the bogus remote connection then we will
4729 * connect again and it will work (with luck).
4730 */
4731
4732 if (sk->state==TCP_SYN_SENT || sk->state==TCP_SYN_RECV)
4733 {
4734 tcp_reset(sk->saddr,sk->daddr,th,sk->prot,NULL,dev, sk->ip_tos,sk->ip_ttl);
4735 return 1;
4736 }
4737
4738 /* Try to resync things. */
4739 tcp_send_ack(sk->sent_seq, sk->acked_seq, sk, th, saddr);
4740 return 0;
4741 }
4742
4743 /*
4744 * When we get a reset we do this.
4745 */
4746
4747 static int tcp_std_reset(struct sock *sk, struct sk_buff *skb)
/* */
4748 {
4749 sk->zapped = 1;
4750 sk->err = ECONNRESET;
4751 if (sk->state == TCP_SYN_SENT)
4752 sk->err = ECONNREFUSED;
4753 if (sk->state == TCP_CLOSE_WAIT)
4754 sk->err = EPIPE;
4755 #ifdef TCP_DO_RFC1337
4756 /*
4757 * Time wait assassination protection [RFC1337]
4758 */
4759 if(sk->state!=TCP_TIME_WAIT)
4760 {
4761 tcp_set_state(sk,TCP_CLOSE);
4762 sk->shutdown = SHUTDOWN_MASK;
4763 }
4764 #else
4765 tcp_set_state(sk,TCP_CLOSE);
4766 sk->shutdown = SHUTDOWN_MASK;
4767 #endif
4768 if (!sk->dead)
4769 sk->state_change(sk);
4770 kfree_skb(skb, FREE_READ);
4771 release_sock(sk);
4772 return(0);
4773 }
4774
4775 /*
4776 * A TCP packet has arrived.
4777 * skb->h.raw is the TCP header.
4778 */
4779
4780 int tcp_rcv(struct sk_buff *skb, struct device *dev, struct options *opt,
/* */
4781 __u32 daddr, unsigned short len,
4782 __u32 saddr, int redo, struct inet_protocol * protocol)
4783 {
4784 struct tcphdr *th;
4785 struct sock *sk;
4786 int syn_ok=0;
4787
4788 tcp_statistics.TcpInSegs++;
4789 if(skb->pkt_type!=PACKET_HOST)
4790 {
4791 kfree_skb(skb,FREE_READ);
4792 return(0);
4793 }
4794
4795 th = skb->h.th;
4796
4797 /*
4798 * Find the socket, using the last hit cache if applicable.
4799 */
4800
4801 if(saddr==th_cache_saddr && daddr==th_cache_daddr && th->dest==th_cache_dport && th->source==th_cache_sport)
4802 {
4803 sk=(struct sock *)th_cache_sk;
4804 /*
4805 * We think this is causing the bug so
4806 */
4807 if(sk!=get_sock(&tcp_prot,th->dest, saddr, th->source, daddr))
4808 printk("Cache mismatch on TCP.\n");
4809 }
4810 else
4811 {
4812 sk = get_sock(&tcp_prot, th->dest, saddr, th->source, daddr);
4813 th_cache_saddr=saddr;
4814 th_cache_daddr=daddr;
4815 th_cache_dport=th->dest;
4816 th_cache_sport=th->source;
4817 th_cache_sk=sk;
4818 }
4819
4820 /*
4821 * If this socket has got a reset it's to all intents and purposes
4822 * really dead. Count closed sockets as dead.
4823 *
4824 * Note: BSD appears to have a bug here. A 'closed' TCP in BSD
4825 * simply drops data. This seems incorrect as a 'closed' TCP doesn't
4826 * exist so should cause resets as if the port was unreachable.
4827 */
4828
4829 if (sk!=NULL && (sk->zapped || sk->state==TCP_CLOSE))
4830 sk=NULL;
4831
4832 if (!redo)
4833 {
4834 /*
4835 * Pull up the IP header.
4836 */
4837 skb_pull(skb, skb->h.raw-skb->data);
4838 /*
4839 * Try to use the device checksum if provided.
4840 */
4841 if (
4842 (skb->ip_summed && tcp_check(th, len, saddr, daddr, skb->csum ))||
4843 (!skb->ip_summed && tcp_check(th, len, saddr, daddr, csum_partial((char *)th, len, 0)))
4844 )
4845 {
4846 skb->sk = NULL;
4847 kfree_skb(skb,FREE_READ);
4848 /*
4849 * We don't release the socket because it was
4850 * never marked in use.
4851 */
4852 return(0);
4853 }
4854 th->seq = ntohl(th->seq);
4855
4856 /* See if we know about the socket. */
4857 if (sk == NULL)
4858 {
4859 /*
4860 * No such TCB. If th->rst is 0 send a reset (checked in tcp_reset)
4861 */
4862 tcp_reset(daddr, saddr, th, &tcp_prot, opt,dev,skb->ip_hdr->tos,255);
4863 skb->sk = NULL;
4864 /*
4865 * Discard frame
4866 */
4867 kfree_skb(skb, FREE_READ);
4868 return(0);
4869 }
4870
4871 /* skb->len = len;*/
4872 skb->acked = 0;
4873 skb->used = 0;
4874 skb->free = 0;
4875 skb->saddr = daddr;
4876 skb->daddr = saddr;
4877
4878 /* We may need to add it to the backlog here. */
4879 cli();
4880 if (sk->inuse)
4881 {
4882 skb_queue_tail(&sk->back_log, skb);
4883 sti();
4884 return(0);
4885 }
4886 sk->inuse = 1;
4887 sti();
4888 }
4889 else
4890 {
4891 if (sk==NULL)
4892 {
4893 tcp_reset(daddr, saddr, th, &tcp_prot, opt,dev,skb->ip_hdr->tos,255);
4894 skb->sk = NULL;
4895 kfree_skb(skb, FREE_READ);
4896 return(0);
4897 }
4898 }
4899
4900
4901 if (!sk->prot)
4902 {
4903 printk("IMPOSSIBLE 3\n");
4904 return(0);
4905 }
4906
4907
4908 /*
4909 * Charge the memory to the socket.
4910 */
4911
4912 if (sk->rmem_alloc + skb->truesize >= sk->rcvbuf)
4913 {
4914 kfree_skb(skb, FREE_READ);
4915 release_sock(sk);
4916 return(0);
4917 }
4918
4919 skb->sk=sk;
4920 sk->rmem_alloc += skb->truesize;
4921
4922 /*
4923 * This basically follows the flow suggested by RFC793, with the corrections in RFC1122. We
4924 * don't implement precedence and we process URG incorrectly (deliberately so) for BSD bug
4925 * compatibility. We also set up variables more thoroughly [Karn notes in the
4926 * KA9Q code the RFC793 incoming segment rules don't initialise the variables for all paths].
4927 */
4928
4929 if(sk->state!=TCP_ESTABLISHED) /* Skip this lot for normal flow */
4930 {
4931
4932 /*
4933 * Now deal with unusual cases.
4934 */
4935
4936 if(sk->state==TCP_LISTEN)
4937 {
4938 if(th->ack) /* These use the socket TOS.. might want to be the received TOS */
4939 tcp_reset(daddr,saddr,th,sk->prot,opt,dev,sk->ip_tos, sk->ip_ttl);
4940
4941 /*
4942 * We don't care for RST, and non SYN are absorbed (old segments)
4943 * Broadcast/multicast SYN isn't allowed. Note - bug if you change the
4944 * netmask on a running connection it can go broadcast. Even Sun's have
4945 * this problem so I'm ignoring it
4946 */
4947
4948 if(th->rst || !th->syn || th->ack || ip_chk_addr(daddr)!=IS_MYADDR)
4949 {
4950 kfree_skb(skb, FREE_READ);
4951 release_sock(sk);
4952 return 0;
4953 }
4954
4955 /*
4956 * Guess we need to make a new socket up
4957 */
4958
4959 tcp_conn_request(sk, skb, daddr, saddr, opt, dev, tcp_init_seq());
4960
4961 /*
4962 * Now we have several options: In theory there is nothing else
4963 * in the frame. KA9Q has an option to send data with the syn,
4964 * BSD accepts data with the syn up to the [to be] advertised window
4965 * and Solaris 2.1 gives you a protocol error. For now we just ignore
4966 * it, that fits the spec precisely and avoids incompatibilities. It
4967 * would be nice in future to drop through and process the data.
4968 */
4969
4970 release_sock(sk);
4971 return 0;
4972 }
4973
4974 /* retransmitted SYN? */
4975 if (sk->state == TCP_SYN_RECV && th->syn && th->seq+1 == sk->acked_seq)
4976 {
4977 kfree_skb(skb, FREE_READ);
4978 release_sock(sk);
4979 return 0;
4980 }
4981
4982 /*
4983 * SYN sent means we have to look for a suitable ack and either reset
4984 * for bad matches or go to connected
4985 */
4986
4987 if(sk->state==TCP_SYN_SENT)
4988 {
4989 /* Crossed SYN or previous junk segment */
4990 if(th->ack)
4991 {
4992 /* We got an ack, but it's not a good ack */
4993 if(!tcp_ack(sk,th,saddr,len))
4994 {
4995 /* Reset the ack - its an ack from a
4996 different connection [ th->rst is checked in tcp_reset()] */
4997 tcp_statistics.TcpAttemptFails++;
4998 tcp_reset(daddr, saddr, th,
4999 sk->prot, opt,dev,sk->ip_tos,sk->ip_ttl);
5000 kfree_skb(skb, FREE_READ);
5001 release_sock(sk);
5002 return(0);
5003 }
5004 if(th->rst)
5005 return tcp_std_reset(sk,skb);
5006 if(!th->syn)
5007 {
5008 /* A valid ack from a different connection
5009 start. Shouldn't happen but cover it */
5010 kfree_skb(skb, FREE_READ);
5011 release_sock(sk);
5012 return 0;
5013 }
5014 /*
5015 * Ok.. it's good. Set up sequence numbers and
5016 * move to established.
5017 */
5018 syn_ok=1; /* Don't reset this connection for the syn */
5019 sk->acked_seq=th->seq+1;
5020 sk->fin_seq=th->seq;
5021 tcp_send_ack(sk->sent_seq,sk->acked_seq,sk,th,sk->daddr);
5022 tcp_set_state(sk, TCP_ESTABLISHED);
5023 tcp_options(sk,th);
5024 sk->dummy_th.dest=th->source;
5025 sk->copied_seq = sk->acked_seq;
5026 if(!sk->dead)
5027 {
5028 sk->state_change(sk);
5029 sock_wake_async(sk->socket, 0);
5030 }
5031 if(sk->max_window==0)
5032 {
5033 sk->max_window = 32;
5034 sk->mss = min(sk->max_window, sk->mtu);
5035 }
5036 }
5037 else
5038 {
5039 /* See if SYN's cross. Drop if boring */
5040 if(th->syn && !th->rst)
5041 {
5042 /* Crossed SYN's are fine - but talking to
5043 yourself is right out... */
5044 if(sk->saddr==saddr && sk->daddr==daddr &&
5045 sk->dummy_th.source==th->source &&
5046 sk->dummy_th.dest==th->dest)
5047 {
5048 tcp_statistics.TcpAttemptFails++;
5049 return tcp_std_reset(sk,skb);
5050 }
5051 tcp_set_state(sk,TCP_SYN_RECV);
5052
5053 /*
5054 * FIXME:
5055 * Must send SYN|ACK here
5056 */
5057 }
5058 /* Discard junk segment */
5059 kfree_skb(skb, FREE_READ);
5060 release_sock(sk);
5061 return 0;
5062 }
5063 /*
5064 * SYN_RECV with data maybe.. drop through
5065 */
5066 goto rfc_step6;
5067 }
5068
5069 /*
5070 * BSD has a funny hack with TIME_WAIT and fast reuse of a port. There is
5071 * a more complex suggestion for fixing these reuse issues in RFC1644
5072 * but not yet ready for general use. Also see RFC1379.
5073 */
5074
5075 #define BSD_TIME_WAIT
5076 #ifdef BSD_TIME_WAIT
5077 if (sk->state == TCP_TIME_WAIT && th->syn && sk->dead &&
5078 after(th->seq, sk->acked_seq) && !th->rst)
5079 {
5080 u32 seq = sk->write_seq;
5081 if(sk->debug)
5082 printk("Doing a BSD time wait\n");
5083 tcp_statistics.TcpEstabResets++;
5084 sk->rmem_alloc -= skb->truesize;
5085 skb->sk = NULL;
5086 sk->err=ECONNRESET;
5087 tcp_set_state(sk, TCP_CLOSE);
5088 sk->shutdown = SHUTDOWN_MASK;
5089 release_sock(sk);
5090 sk=get_sock(&tcp_prot, th->dest, saddr, th->source, daddr);
5091 if (sk && sk->state==TCP_LISTEN)
5092 {
5093 sk->inuse=1;
5094 skb->sk = sk;
5095 sk->rmem_alloc += skb->truesize;
5096 tcp_conn_request(sk, skb, daddr, saddr,opt, dev,seq+128000);
5097 release_sock(sk);
5098 return 0;
5099 }
5100 kfree_skb(skb, FREE_READ);
5101 return 0;
5102 }
5103 #endif
5104 }
5105
5106 /*
5107 * We are now in normal data flow (see the step list in the RFC)
5108 * Note most of these are inline now. I'll inline the lot when
5109 * I have time to test it hard and look at what gcc outputs
5110 */
5111
5112 if(!tcp_sequence(sk,th,len,opt,saddr,dev))
5113 {
5114 kfree_skb(skb, FREE_READ);
5115 release_sock(sk);
5116 return 0;
5117 }
5118
5119 if(th->rst)
5120 return tcp_std_reset(sk,skb);
5121
5122 /*
5123 * !syn_ok is effectively the state test in RFC793.
5124 */
5125
5126 if(th->syn && !syn_ok)
5127 {
5128 tcp_reset(daddr,saddr,th, &tcp_prot, opt, dev, skb->ip_hdr->tos, 255);
5129 return tcp_std_reset(sk,skb);
5130 }
5131
5132 /*
5133 * Process the ACK
5134 */
5135
5136
5137 if(th->ack && !tcp_ack(sk,th,saddr,len))
5138 {
5139 /*
5140 * Our three way handshake failed.
5141 */
5142
5143 if(sk->state==TCP_SYN_RECV)
5144 {
5145 tcp_reset(daddr, saddr, th,sk->prot, opt, dev,sk->ip_tos,sk->ip_ttl);
5146 }
5147 kfree_skb(skb, FREE_READ);
5148 release_sock(sk);
5149 return 0;
5150 }
5151
5152 rfc_step6: /* I'll clean this up later */
5153
5154 /*
5155 * Process urgent data
5156 */
5157
5158 if(tcp_urg(sk, th, saddr, len))
5159 {
5160 kfree_skb(skb, FREE_READ);
5161 release_sock(sk);
5162 return 0;
5163 }
5164
5165
5166 /*
5167 * Process the encapsulated data
5168 */
5169
5170 if(tcp_data(skb,sk, saddr, len))
5171 {
5172 kfree_skb(skb, FREE_READ);
5173 release_sock(sk);
5174 return 0;
5175 }
5176
5177 /*
5178 * And done
5179 */
5180
5181 release_sock(sk);
5182 return 0;
5183 }
5184
5185 /*
5186 * This routine sends a packet with an out of date sequence
5187 * number. It assumes the other end will try to ack it.
5188 */
5189
5190 static void tcp_write_wakeup(struct sock *sk)
/* */
5191 {
5192 struct sk_buff *buff,*skb;
5193 struct tcphdr *t1;
5194 struct device *dev=NULL;
5195 int tmp;
5196
5197 if (sk->zapped)
5198 return; /* After a valid reset we can send no more */
5199
5200 /*
5201 * Write data can still be transmitted/retransmitted in the
5202 * following states. If any other state is encountered, return.
5203 * [listen/close will never occur here anyway]
5204 */
5205
5206 if (sk->state != TCP_ESTABLISHED &&
5207 sk->state != TCP_CLOSE_WAIT &&
5208 sk->state != TCP_FIN_WAIT1 &&
5209 sk->state != TCP_LAST_ACK &&
5210 sk->state != TCP_CLOSING
5211 )
5212 {
5213 return;
5214 }
5215 if ( before(sk->sent_seq, sk->window_seq) &&
5216 (skb=skb_peek(&sk->write_queue)))
5217 {
5218 /*
5219 * We are probing the opening of a window
5220 * but the window size is != 0
5221 * must have been a result SWS advoidance ( sender )
5222 */
5223
5224 struct iphdr *iph;
5225 struct tcphdr *th;
5226 struct tcphdr *nth;
5227 unsigned long win_size;
5228 #if 0
5229 unsigned long ow_size;
5230 #endif
5231 void * tcp_data_start;
5232
5233 /*
5234 * How many bytes can we send ?
5235 */
5236
5237 win_size = sk->window_seq - sk->sent_seq;
5238
5239 /*
5240 * Recover the buffer pointers
5241 */
5242
5243 iph = (struct iphdr *)skb->ip_hdr;
5244 th = (struct tcphdr *)(((char *)iph) +(iph->ihl << 2));
5245
5246 /*
5247 * Grab the data for a temporary frame
5248 */
5249
5250 buff = sock_wmalloc(sk, win_size + th->doff * 4 +
5251 (iph->ihl << 2) +
5252 sk->prot->max_header + 15,
5253 1, GFP_ATOMIC);
5254 if ( buff == NULL )
5255 return;
5256
5257 /*
5258 * If we strip the packet on the write queue we must
5259 * be ready to retransmit this one
5260 */
5261
5262 buff->free = /*0*/1;
5263
5264 buff->sk = sk;
5265 buff->localroute = sk->localroute;
5266
5267 /*
5268 * Put headers on the new packet
5269 */
5270
5271 tmp = sk->prot->build_header(buff, sk->saddr, sk->daddr, &dev,
5272 IPPROTO_TCP, sk->opt, buff->truesize,
5273 sk->ip_tos,sk->ip_ttl);
5274 if (tmp < 0)
5275 {
5276 sock_wfree(sk, buff);
5277 return;
5278 }
5279
5280 /*
5281 * Move the TCP header over
5282 */
5283
5284 buff->dev = dev;
5285
5286 nth = (struct tcphdr *) skb_put(buff,th->doff*4);
5287
5288 memcpy(nth, th, th->doff * 4);
5289
5290 /*
5291 * Correct the new header
5292 */
5293
5294 nth->ack = 1;
5295 nth->ack_seq = ntohl(sk->acked_seq);
5296 nth->window = ntohs(tcp_select_window(sk));
5297 nth->check = 0;
5298
5299 /*
5300 * Find the first data byte.
5301 */
5302
5303 tcp_data_start = skb->data + skb->dev->hard_header_len +
5304 (iph->ihl << 2) + th->doff * 4;
5305
5306 /*
5307 * Add it to our new buffer
5308 */
5309 memcpy(skb_put(buff,win_size), tcp_data_start, win_size);
5310
5311 /*
5312 * Remember our right edge sequence number.
5313 */
5314
5315 buff->h.seq = sk->sent_seq + win_size;
5316 sk->sent_seq = buff->h.seq; /* Hack */
5317 #if 0
5318
5319 /*
5320 * now: shrink the queue head segment
5321 */
5322
5323 th->check = 0;
5324 ow_size = skb->len - win_size -
5325 ((unsigned long) (tcp_data_start - (void *) skb->data));
5326
5327 memmove(tcp_data_start, tcp_data_start + win_size, ow_size);
5328 skb_trim(skb,skb->len-win_size);
5329 sk->sent_seq += win_size;
5330 th->seq = htonl(sk->sent_seq);
5331 if (th->urg)
5332 {
5333 unsigned short urg_ptr;
5334
5335 urg_ptr = ntohs(th->urg_ptr);
5336 if (urg_ptr <= win_size)
5337 th->urg = 0;
5338 else
5339 {
5340 urg_ptr -= win_size;
5341 th->urg_ptr = htons(urg_ptr);
5342 nth->urg_ptr = htons(win_size);
5343 }
5344 }
5345 #else
5346 if(th->urg && ntohs(th->urg_ptr) < win_size)
5347 nth->urg = 0;
5348 #endif
5349
5350 /*
5351 * Checksum the split buffer
5352 */
5353
5354 tcp_send_check(nth, sk->saddr, sk->daddr,
5355 nth->doff * 4 + win_size , sk);
5356 }
5357 else
5358 {
5359 buff = sock_wmalloc(sk,MAX_ACK_SIZE,1, GFP_ATOMIC);
5360 if (buff == NULL)
5361 return;
5362
5363 buff->free = 1;
5364 buff->sk = sk;
5365 buff->localroute = sk->localroute;
5366
5367 /*
5368 * Put in the IP header and routing stuff.
5369 */
5370
5371 tmp = sk->prot->build_header(buff, sk->saddr, sk->daddr, &dev,
5372 IPPROTO_TCP, sk->opt, MAX_ACK_SIZE,sk->ip_tos,sk->ip_ttl);
5373 if (tmp < 0)
5374 {
5375 sock_wfree(sk, buff);
5376 return;
5377 }
5378
5379 t1 = (struct tcphdr *)skb_put(buff,sizeof(struct tcphdr));
5380 memcpy(t1,(void *) &sk->dummy_th, sizeof(*t1));
5381
5382 /*
5383 * Use a previous sequence.
5384 * This should cause the other end to send an ack.
5385 */
5386
5387 t1->seq = htonl(sk->sent_seq-1);
5388 t1->ack = 1;
5389 t1->res1= 0;
5390 t1->res2= 0;
5391 t1->rst = 0;
5392 t1->urg = 0;
5393 t1->psh = 0;
5394 t1->fin = 0; /* We are sending a 'previous' sequence, and 0 bytes of data - thus no FIN bit */
5395 t1->syn = 0;
5396 t1->ack_seq = ntohl(sk->acked_seq);
5397 t1->window = ntohs(tcp_select_window(sk));
5398 t1->doff = sizeof(*t1)/4;
5399 tcp_send_check(t1, sk->saddr, sk->daddr, sizeof(*t1), sk);
5400
5401 }
5402
5403 /*
5404 * Send it.
5405 */
5406
5407 sk->prot->queue_xmit(sk, dev, buff, 1);
5408 tcp_statistics.TcpOutSegs++;
5409 }
5410
5411 /*
5412 * A window probe timeout has occurred.
5413 */
5414
5415 void tcp_send_probe0(struct sock *sk)
/* */
5416 {
5417 if (sk->zapped)
5418 return; /* After a valid reset we can send no more */
5419
5420 tcp_write_wakeup(sk);
5421
5422 sk->backoff++;
5423 sk->rto = min(sk->rto << 1, 120*HZ);
5424 reset_xmit_timer (sk, TIME_PROBE0, sk->rto);
5425 sk->retransmits++;
5426 sk->prot->retransmits ++;
5427 }
5428
5429 /*
5430 * Socket option code for TCP.
5431 */
5432
5433 int tcp_setsockopt(struct sock *sk, int level, int optname, char *optval, int optlen)
/* */
5434 {
5435 int val,err;
5436
5437 if(level!=SOL_TCP)
5438 return ip_setsockopt(sk,level,optname,optval,optlen);
5439
5440 if (optval == NULL)
5441 return(-EINVAL);
5442
5443 err=verify_area(VERIFY_READ, optval, sizeof(int));
5444 if(err)
5445 return err;
5446
5447 val = get_user((int *)optval);
5448
5449 switch(optname)
5450 {
5451 case TCP_MAXSEG:
5452 /*
5453 * values greater than interface MTU won't take effect. however at
5454 * the point when this call is done we typically don't yet know
5455 * which interface is going to be used
5456 */
5457 if(val<1||val>MAX_WINDOW)
5458 return -EINVAL;
5459 sk->user_mss=val;
5460 return 0;
5461 case TCP_NODELAY:
5462 sk->nonagle=(val==0)?0:1;
5463 return 0;
5464 default:
5465 return(-ENOPROTOOPT);
5466 }
5467 }
5468
5469 int tcp_getsockopt(struct sock *sk, int level, int optname, char *optval, int *optlen)
/* ![[last]](../icons/n_last.png)
*/
5470 {
5471 int val,err;
5472
5473 if(level!=SOL_TCP)
5474 return ip_getsockopt(sk,level,optname,optval,optlen);
5475
5476 switch(optname)
5477 {
5478 case TCP_MAXSEG:
5479 val=sk->user_mss;
5480 break;
5481 case TCP_NODELAY:
5482 val=sk->nonagle;
5483 break;
5484 default:
5485 return(-ENOPROTOOPT);
5486 }
5487 err=verify_area(VERIFY_WRITE, optlen, sizeof(int));
5488 if(err)
5489 return err;
5490 put_user(sizeof(int),(int *) optlen);
5491
5492 err=verify_area(VERIFY_WRITE, optval, sizeof(int));
5493 if(err)
5494 return err;
5495 put_user(val,(int *)optval);
5496
5497 return(0);
5498 }
5499
5500
5501 struct proto tcp_prot = {
5502 tcp_close,
5503 tcp_read,
5504 tcp_write,
5505 tcp_sendto,
5506 tcp_recvfrom,
5507 ip_build_header,
5508 tcp_connect,
5509 tcp_accept,
5510 ip_queue_xmit,
5511 tcp_retransmit,
5512 tcp_write_wakeup,
5513 tcp_read_wakeup,
5514 tcp_rcv,
5515 tcp_select,
5516 tcp_ioctl,
5517 NULL,
5518 tcp_shutdown,
5519 tcp_setsockopt,
5520 tcp_getsockopt,
5521 tcp_sendmsg,
5522 tcp_recvmsg,
5523 128,
5524 0,
5525 "TCP",
5526 0, 0,
5527 {NULL,}
5528 };
![[bottom]](../icons/n_bottom.png){kind=icon}
*/