![[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_input.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 * Pedro Roque : Double ACK bug
24 */
25
26 #include <linux/config.h>
27 #include <net/tcp.h>
28
29 /*
30 * Policy code extracted so its now seperate
31 */
32
33 /*
34 * Called each time to estimate the delayed ack timeout. This is
35 * how it should be done so a fast link isnt impacted by ack delay.
36 */
37
38 extern __inline__ void tcp_delack_estimator(struct sock *sk)
/* ![[next]](../icons/right.png)
![[first]](../icons/n_first.png)
![[top]](../icons/top.png)
*/
39 {
40 /*
41 * Delayed ACK time estimator.
42 */
43
44 if (sk->lrcvtime == 0)
45 {
46 sk->lrcvtime = jiffies;
47 sk->ato = HZ/3;
48 }
49 else
50 {
51 int m;
52
53 m = jiffies - sk->lrcvtime;
54
55 sk->lrcvtime = jiffies;
56
57 if (m <= 0)
58 m = 1;
59
60 if (m > (sk->rtt >> 3))
61 {
62 sk->ato = sk->rtt >> 3;
63 /*
64 * printk(KERN_DEBUG "ato: rtt %lu\n", sk->ato);
65 */
66 }
67 else
68 {
69 sk->ato = (sk->ato >> 1) + m;
70 /*
71 * printk(KERN_DEBUG "ato: m %lu\n", sk->ato);
72 */
73 }
74 }
75 }
76
77 /*
78 * Called on frames that were known _not_ to have been
79 * retransmitted [see Karn/Partridge Proceedings SIGCOMM 87].
80 * The algorithm is from the SIGCOMM 88 piece by Van Jacobson.
81 */
82
83 extern __inline__ void tcp_rtt_estimator(struct sock *sk, struct sk_buff *oskb)
/* ![[previous]](../icons/left.png)
*/
84 {
85 long m;
86 /*
87 * The following amusing code comes from Jacobson's
88 * article in SIGCOMM '88. Note that rtt and mdev
89 * are scaled versions of rtt and mean deviation.
90 * This is designed to be as fast as possible
91 * m stands for "measurement".
92 */
93
94 m = jiffies - oskb->when; /* RTT */
95 if(m<=0)
96 m=1; /* IS THIS RIGHT FOR <0 ??? */
97 m -= (sk->rtt >> 3); /* m is now error in rtt est */
98 sk->rtt += m; /* rtt = 7/8 rtt + 1/8 new */
99 if (m < 0)
100 m = -m; /* m is now abs(error) */
101 m -= (sk->mdev >> 2); /* similar update on mdev */
102 sk->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
103
104 /*
105 * Now update timeout. Note that this removes any backoff.
106 */
107
108 sk->rto = ((sk->rtt >> 2) + sk->mdev) >> 1;
109 if (sk->rto > 120*HZ)
110 sk->rto = 120*HZ;
111 if (sk->rto < HZ/5) /* Was 1*HZ - keep .2 as minimum cos of the BSD delayed acks */
112 sk->rto = HZ/5;
113 sk->backoff = 0;
114 }
115
116 /*
117 * Cached last hit socket
118 */
119
120 static volatile unsigned long th_cache_saddr, th_cache_daddr;
121 static volatile unsigned short th_cache_dport, th_cache_sport;
122 static volatile struct sock *th_cache_sk;
123
124 void tcp_cache_zap(void)
/* */
125 {
126 th_cache_sk=NULL;
127 }
128
129 /*
130 * Find the socket, using the last hit cache if applicable. The cache is not quite
131 * right...
132 */
133
134 static inline struct sock * get_tcp_sock(u32 saddr, u16 sport, u32 daddr, u16 dport)
/* */
135 {
136 struct sock * sk;
137
138 sk = (struct sock *) th_cache_sk;
139 if (!sk || saddr != th_cache_saddr || daddr != th_cache_daddr ||
140 sport != th_cache_sport || dport != th_cache_dport) {
141 sk = get_sock(&tcp_prot, dport, saddr, sport, daddr);
142 if (sk) {
143 th_cache_saddr=saddr;
144 th_cache_daddr=daddr;
145 th_cache_dport=dport;
146 th_cache_sport=sport;
147 th_cache_sk=sk;
148 }
149 }
150 return sk;
151 }
152
153 /*
154 * React to a out-of-window TCP sequence number in an incoming packet
155 */
156
157 static void bad_tcp_sequence(struct sock *sk, struct tcphdr *th, short len,
/* */
158 struct options *opt, unsigned long saddr, struct device *dev)
159 {
160 if (th->rst)
161 return;
162
163 /*
164 * Send a reset if we get something not ours and we are
165 * unsynchronized. Note: We don't do anything to our end. We
166 * are just killing the bogus remote connection then we will
167 * connect again and it will work (with luck).
168 */
169
170 if (sk->state==TCP_SYN_SENT || sk->state==TCP_SYN_RECV)
171 {
172 tcp_send_reset(sk->saddr,sk->daddr,th,sk->prot,NULL,dev, sk->ip_tos,sk->ip_ttl);
173 return;
174 }
175
176 /*
177 * 4.3reno machines look for these kind of acks so they can do fast
178 * recovery. Three identical 'old' acks lets it know that one frame has
179 * been lost and should be resent. Because this is before the whole window
180 * of data has timed out it can take one lost frame per window without
181 * stalling. [See Jacobson RFC1323, Stevens TCP/IP illus vol2]
182 *
183 * We also should be spotting triple bad sequences.
184 */
185 tcp_send_ack(sk->sent_seq, sk->acked_seq, sk, th, saddr);
186 return;
187 }
188
189 /*
190 * This functions checks to see if the tcp header is actually acceptable.
191 */
192
193 extern __inline__ int tcp_sequence(struct sock *sk, u32 seq, u32 end_seq)
/* */
194 {
195 u32 end_window = sk->acked_seq + sk->window;
196 return /* if start is at end of window, end must be too (zero window) */
197 (seq == end_window && seq == end_seq) ||
198 /* if start is before end of window, check for interest */
199 (before(seq, end_window) && !before(end_seq, sk->acked_seq));
200 }
201
202 /*
203 * When we get a reset we do this. This probably is a tcp_output routine
204 * really.
205 */
206
207 static int tcp_reset(struct sock *sk, struct sk_buff *skb)
/* */
208 {
209 sk->zapped = 1;
210 /*
211 * We want the right error as BSD sees it (and indeed as we do).
212 */
213 sk->err = ECONNRESET;
214 if (sk->state == TCP_SYN_SENT)
215 sk->err = ECONNREFUSED;
216 if (sk->state == TCP_CLOSE_WAIT)
217 sk->err = EPIPE;
218 #ifdef CONFIG_TCP_RFC1337
219 /*
220 * Time wait assassination protection [RFC1337]
221 *
222 * This is a good idea, but causes more sockets to take time to close.
223 *
224 * Ian Heavens has since shown this is an inadequate fix for the protocol
225 * bug in question.
226 */
227 if(sk->state!=TCP_TIME_WAIT)
228 {
229 tcp_set_state(sk,TCP_CLOSE);
230 sk->shutdown = SHUTDOWN_MASK;
231 }
232 #else
233 tcp_set_state(sk,TCP_CLOSE);
234 sk->shutdown = SHUTDOWN_MASK;
235 #endif
236 if (!sk->dead)
237 sk->state_change(sk);
238 kfree_skb(skb, FREE_READ);
239 return(0);
240 }
241
242
243 /*
244 * Look for tcp options. Parses everything but only knows about MSS.
245 * This routine is always called with the packet containing the SYN.
246 * However it may also be called with the ack to the SYN. So you
247 * can't assume this is always the SYN. It's always called after
248 * we have set up sk->mtu to our own MTU.
249 *
250 * We need at minimum to add PAWS support here. Possibly large windows
251 * as Linux gets deployed on 100Mb/sec networks.
252 */
253
254 static void tcp_options(struct sock *sk, struct tcphdr *th)
/* */
255 {
256 unsigned char *ptr;
257 int length=(th->doff*4)-sizeof(struct tcphdr);
258 int mss_seen = 0;
259
260 ptr = (unsigned char *)(th + 1);
261
262 while(length>0)
263 {
264 int opcode=*ptr++;
265 int opsize=*ptr++;
266 switch(opcode)
267 {
268 case TCPOPT_EOL:
269 return;
270 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
271 length--;
272 ptr--; /* the opsize=*ptr++ above was a mistake */
273 continue;
274
275 default:
276 if(opsize<=2) /* Avoid silly options looping forever */
277 return;
278 switch(opcode)
279 {
280 case TCPOPT_MSS:
281 if(opsize==4 && th->syn)
282 {
283 sk->mtu=min(sk->mtu,ntohs(*(unsigned short *)ptr));
284 mss_seen = 1;
285 }
286 break;
287 /* Add other options here as people feel the urge to implement stuff like large windows */
288 }
289 ptr+=opsize-2;
290 length-=opsize;
291 }
292 }
293 if (th->syn)
294 {
295 if (! mss_seen)
296 sk->mtu=min(sk->mtu, 536); /* default MSS if none sent */
297 }
298 #ifdef CONFIG_INET_PCTCP
299 sk->mss = min(sk->max_window >> 1, sk->mtu);
300 #else
301 sk->mss = min(sk->max_window, sk->mtu);
302 sk->max_unacked = 2 * sk->mss;
303 #endif
304 }
305
306
307 /*
308 * This routine handles a connection request.
309 * It should make sure we haven't already responded.
310 * Because of the way BSD works, we have to send a syn/ack now.
311 * This also means it will be harder to close a socket which is
312 * listening.
313 */
314
315 static void tcp_conn_request(struct sock *sk, struct sk_buff *skb,
/* */
316 u32 daddr, u32 saddr, struct options *opt, struct device *dev, u32 seq)
317 {
318 struct sock *newsk;
319 struct tcphdr *th;
320 struct rtable *rt;
321
322 th = skb->h.th;
323
324 /* If the socket is dead, don't accept the connection. */
325 if (!sk->dead)
326 {
327 sk->data_ready(sk,0);
328 }
329 else
330 {
331 if(sk->debug)
332 printk("Reset on %p: Connect on dead socket.\n",sk);
333 tcp_send_reset(daddr, saddr, th, sk->prot, opt, dev, sk->ip_tos,sk->ip_ttl);
334 tcp_statistics.TcpAttemptFails++;
335 kfree_skb(skb, FREE_READ);
336 return;
337 }
338
339 /*
340 * Make sure we can accept more. This will prevent a
341 * flurry of syns from eating up all our memory.
342 *
343 * BSD does some funnies here and allows 3/2 times the
344 * set backlog as a fudge factor. Thats just too gross.
345 */
346
347 if (sk->ack_backlog >= sk->max_ack_backlog)
348 {
349 tcp_statistics.TcpAttemptFails++;
350 kfree_skb(skb, FREE_READ);
351 return;
352 }
353
354 /*
355 * We need to build a new sock struct.
356 * It is sort of bad to have a socket without an inode attached
357 * to it, but the wake_up's will just wake up the listening socket,
358 * and if the listening socket is destroyed before this is taken
359 * off of the queue, this will take care of it.
360 */
361
362 newsk = (struct sock *) kmalloc(sizeof(struct sock), GFP_ATOMIC);
363 if (newsk == NULL)
364 {
365 /* just ignore the syn. It will get retransmitted. */
366 tcp_statistics.TcpAttemptFails++;
367 kfree_skb(skb, FREE_READ);
368 return;
369 }
370
371 memcpy(newsk, sk, sizeof(*newsk));
372 newsk->opt = NULL;
373 newsk->ip_route_cache = NULL;
374 if (opt && opt->optlen)
375 {
376 sk->opt = (struct options*)kmalloc(sizeof(struct options)+opt->optlen, GFP_ATOMIC);
377 if (!sk->opt)
378 {
379 kfree_s(newsk, sizeof(struct sock));
380 tcp_statistics.TcpAttemptFails++;
381 kfree_skb(skb, FREE_READ);
382 return;
383 }
384 if (ip_options_echo(sk->opt, opt, daddr, saddr, skb))
385 {
386 kfree_s(sk->opt, sizeof(struct options)+opt->optlen);
387 kfree_s(newsk, sizeof(struct sock));
388 tcp_statistics.TcpAttemptFails++;
389 kfree_skb(skb, FREE_READ);
390 return;
391 }
392 }
393 skb_queue_head_init(&newsk->write_queue);
394 skb_queue_head_init(&newsk->receive_queue);
395 newsk->send_head = NULL;
396 newsk->send_tail = NULL;
397 skb_queue_head_init(&newsk->back_log);
398 newsk->rtt = 0; /*TCP_CONNECT_TIME<<3*/
399 newsk->ato = HZ/3;
400 newsk->rto = TCP_TIMEOUT_INIT;
401 newsk->mdev = 0;
402 newsk->max_window = 0;
403 newsk->cong_window = 1;
404 newsk->cong_count = 0;
405 newsk->ssthresh = 0;
406 newsk->backoff = 0;
407 newsk->blog = 0;
408 newsk->intr = 0;
409 newsk->proc = 0;
410 newsk->done = 0;
411 newsk->partial = NULL;
412 newsk->pair = NULL;
413 newsk->wmem_alloc = 0;
414 newsk->rmem_alloc = 0;
415 newsk->localroute = sk->localroute;
416
417 newsk->max_unacked = MAX_WINDOW - TCP_WINDOW_DIFF;
418
419 newsk->err = 0;
420 newsk->shutdown = 0;
421 newsk->ack_backlog = 0;
422 newsk->acked_seq = skb->seq+1;
423 newsk->lastwin_seq = skb->seq+1;
424 newsk->delay_acks = 1;
425 newsk->copied_seq = skb->seq+1;
426 newsk->fin_seq = skb->seq;
427 newsk->state = TCP_SYN_RECV;
428 newsk->timeout = 0;
429 newsk->ip_xmit_timeout = 0;
430 newsk->write_seq = seq;
431 newsk->window_seq = newsk->write_seq;
432 newsk->rcv_ack_seq = newsk->write_seq;
433 newsk->urg_data = 0;
434 newsk->retransmits = 0;
435 newsk->linger=0;
436 newsk->destroy = 0;
437 init_timer(&newsk->timer);
438 newsk->timer.data = (unsigned long)newsk;
439 newsk->timer.function = &net_timer;
440 init_timer(&newsk->retransmit_timer);
441 newsk->retransmit_timer.data = (unsigned long)newsk;
442 newsk->retransmit_timer.function=&tcp_retransmit_timer;
443 newsk->dummy_th.source = skb->h.th->dest;
444 newsk->dummy_th.dest = skb->h.th->source;
445
446 /*
447 * Swap these two, they are from our point of view.
448 */
449
450 newsk->daddr = saddr;
451 newsk->saddr = daddr;
452 newsk->rcv_saddr = daddr;
453
454 put_sock(newsk->num,newsk);
455 newsk->acked_seq = skb->seq + 1;
456 newsk->copied_seq = skb->seq + 1;
457 newsk->socket = NULL;
458
459 /*
460 * Grab the ttl and tos values and use them
461 */
462
463 newsk->ip_ttl=sk->ip_ttl;
464 newsk->ip_tos=skb->ip_hdr->tos;
465
466 /*
467 * Use 512 or whatever user asked for
468 */
469
470 /*
471 * Note use of sk->user_mss, since user has no direct access to newsk
472 */
473
474 rt = ip_rt_route(newsk->opt && newsk->opt->srr ? newsk->opt->faddr : saddr, 0);
475 newsk->ip_route_cache = rt;
476
477 if(rt!=NULL && (rt->rt_flags&RTF_WINDOW))
478 newsk->window_clamp = rt->rt_window;
479 else
480 newsk->window_clamp = 0;
481
482 if (sk->user_mss)
483 newsk->mtu = sk->user_mss;
484 else if (rt)
485 newsk->mtu = rt->rt_mtu - sizeof(struct iphdr) - sizeof(struct tcphdr);
486 else
487 newsk->mtu = 576 - sizeof(struct iphdr) - sizeof(struct tcphdr);
488
489 /*
490 * But not bigger than device MTU
491 */
492
493 newsk->mtu = min(newsk->mtu, dev->mtu - sizeof(struct iphdr) - sizeof(struct tcphdr));
494
495 #ifdef CONFIG_SKIP
496
497 /*
498 * SKIP devices set their MTU to 65535. This is so they can take packets
499 * unfragmented to security process then fragment. They could lie to the
500 * TCP layer about a suitable MTU, but its easier to let skip sort it out
501 * simply because the final package we want unfragmented is going to be
502 *
503 * [IPHDR][IPSP][Security data][Modified TCP data][Security data]
504 */
505
506 if(skip_pick_mtu!=NULL) /* If SKIP is loaded.. */
507 sk->mtu=skip_pick_mtu(sk->mtu,dev);
508 #endif
509 /*
510 * This will min with what arrived in the packet
511 */
512
513 tcp_options(newsk,skb->h.th);
514
515 tcp_cache_zap();
516 tcp_send_synack(newsk, sk, skb);
517 }
518
519
520 /*
521 * Handle a TCP window that shrunk on us. It shouldn't happen,
522 * but..
523 *
524 * We may need to move packets from the send queue
525 * to the write queue, if the window has been shrunk on us.
526 * The RFC says you are not allowed to shrink your window
527 * like this, but if the other end does, you must be able
528 * to deal with it.
529 */
530 void tcp_window_shrunk(struct sock * sk, u32 window_seq)
/* */
531 {
532 struct sk_buff *skb;
533 struct sk_buff *skb2;
534 struct sk_buff *wskb = NULL;
535
536 skb2 = sk->send_head;
537 sk->send_head = NULL;
538 sk->send_tail = NULL;
539
540 /*
541 * This is an artifact of a flawed concept. We want one
542 * queue and a smarter send routine when we send all.
543 */
544 cli();
545 while (skb2 != NULL)
546 {
547 skb = skb2;
548 skb2 = skb->link3;
549 skb->link3 = NULL;
550 if (after(skb->end_seq, window_seq))
551 {
552 if (sk->packets_out > 0)
553 sk->packets_out--;
554 /* We may need to remove this from the dev send list. */
555 if (skb->next != NULL)
556 {
557 skb_unlink(skb);
558 }
559 /* Now add it to the write_queue. */
560 if (wskb == NULL)
561 skb_queue_head(&sk->write_queue,skb);
562 else
563 skb_append(wskb,skb);
564 wskb = skb;
565 }
566 else
567 {
568 if (sk->send_head == NULL)
569 {
570 sk->send_head = skb;
571 sk->send_tail = skb;
572 }
573 else
574 {
575 sk->send_tail->link3 = skb;
576 sk->send_tail = skb;
577 }
578 skb->link3 = NULL;
579 }
580 }
581 sti();
582 }
583
584
585 /*
586 * This routine deals with incoming acks, but not outgoing ones.
587 *
588 * This routine is totally _WRONG_. The list structuring is wrong,
589 * the algorithm is wrong, the code is wrong.
590 */
591
592 static int tcp_ack(struct sock *sk, struct tcphdr *th, u32 ack, int len)
/* */
593 {
594 int flag = 0;
595 u32 window_seq;
596
597 /*
598 * 1 - there was data in packet as well as ack or new data is sent or
599 * in shutdown state
600 * 2 - data from retransmit queue was acked and removed
601 * 4 - window shrunk or data from retransmit queue was acked and removed
602 */
603
604 if(sk->zapped)
605 return(1); /* Dead, cant ack any more so why bother */
606
607 /*
608 * We have dropped back to keepalive timeouts. Thus we have
609 * no retransmits pending.
610 */
611
612 if (sk->ip_xmit_timeout == TIME_KEEPOPEN)
613 sk->retransmits = 0;
614
615 /*
616 * If the ack is newer than sent or older than previous acks
617 * then we can probably ignore it.
618 */
619
620 if (after(ack, sk->sent_seq) || before(ack, sk->rcv_ack_seq))
621 goto uninteresting_ack;
622
623 /*
624 * If there is data set flag 1
625 */
626
627 if (len != th->doff*4)
628 flag |= 1;
629
630 /*
631 * Have we discovered a larger window
632 */
633 window_seq = ntohs(th->window);
634 if (window_seq > sk->max_window)
635 {
636 sk->max_window = window_seq;
637 #ifdef CONFIG_INET_PCTCP
638 /* Hack because we don't send partial packets to non SWS
639 handling hosts */
640 sk->mss = min(window_seq>>1, sk->mtu);
641 #else
642 sk->mss = min(window_seq, sk->mtu);
643 #endif
644 }
645 window_seq += ack;
646
647 /*
648 * See if our window has been shrunk.
649 */
650 if (after(sk->window_seq, window_seq)) {
651 flag |= 4;
652 tcp_window_shrunk(sk, window_seq);
653 }
654
655 /*
656 * Update the right hand window edge of the host
657 */
658 sk->window_seq = window_seq;
659
660 /*
661 * Pipe has emptied
662 */
663 if (sk->send_tail == NULL || sk->send_head == NULL)
664 {
665 sk->send_head = NULL;
666 sk->send_tail = NULL;
667 sk->packets_out= 0;
668 }
669
670 /*
671 * We don't want too many packets out there.
672 */
673
674 if (sk->ip_xmit_timeout == TIME_WRITE &&
675 sk->cong_window < 2048 && after(ack, sk->rcv_ack_seq))
676 {
677
678 /*
679 * This is Jacobson's slow start and congestion avoidance.
680 * SIGCOMM '88, p. 328. Because we keep cong_window in integral
681 * mss's, we can't do cwnd += 1 / cwnd. Instead, maintain a
682 * counter and increment it once every cwnd times. It's possible
683 * that this should be done only if sk->retransmits == 0. I'm
684 * interpreting "new data is acked" as including data that has
685 * been retransmitted but is just now being acked.
686 */
687 if (sk->cong_window < sk->ssthresh)
688 /*
689 * In "safe" area, increase
690 */
691 sk->cong_window++;
692 else
693 {
694 /*
695 * In dangerous area, increase slowly. In theory this is
696 * sk->cong_window += 1 / sk->cong_window
697 */
698 if (sk->cong_count >= sk->cong_window)
699 {
700 sk->cong_window++;
701 sk->cong_count = 0;
702 }
703 else
704 sk->cong_count++;
705 }
706 }
707
708 /*
709 * Remember the highest ack received.
710 */
711
712 sk->rcv_ack_seq = ack;
713
714 /*
715 * We passed data and got it acked, remove any soft error
716 * log. Something worked...
717 */
718
719 sk->err_soft = 0;
720
721 /*
722 * If this ack opens up a zero window, clear backoff. It was
723 * being used to time the probes, and is probably far higher than
724 * it needs to be for normal retransmission.
725 */
726
727 if (sk->ip_xmit_timeout == TIME_PROBE0)
728 {
729 sk->retransmits = 0; /* Our probe was answered */
730
731 /*
732 * Was it a usable window open ?
733 */
734
735 if (skb_peek(&sk->write_queue) != NULL && /* should always be non-null */
736 ! before (sk->window_seq, sk->write_queue.next->end_seq))
737 {
738 sk->backoff = 0;
739
740 /*
741 * Recompute rto from rtt. this eliminates any backoff.
742 */
743
744 sk->rto = ((sk->rtt >> 2) + sk->mdev) >> 1;
745 if (sk->rto > 120*HZ)
746 sk->rto = 120*HZ;
747 if (sk->rto < HZ/5) /* Was 1*HZ, then 1 - turns out we must allow about
748 .2 of a second because of BSD delayed acks - on a 100Mb/sec link
749 .2 of a second is going to need huge windows (SIGH) */
750 sk->rto = HZ/5;
751 }
752 }
753
754 /*
755 * See if we can take anything off of the retransmit queue.
756 */
757
758 while(sk->send_head != NULL)
759 {
760 /* Check for a bug. */
761 if (sk->send_head->link3 &&
762 after(sk->send_head->end_seq, sk->send_head->link3->end_seq))
763 printk("INET: tcp.c: *** bug send_list out of order.\n");
764
765 /*
766 * If our packet is before the ack sequence we can
767 * discard it as it's confirmed to have arrived the other end.
768 */
769
770 if (before(sk->send_head->end_seq, ack+1))
771 {
772 struct sk_buff *oskb;
773 if (sk->retransmits)
774 {
775 /*
776 * We were retransmitting. don't count this in RTT est
777 */
778 flag |= 2;
779
780 /*
781 * even though we've gotten an ack, we're still
782 * retransmitting as long as we're sending from
783 * the retransmit queue. Keeping retransmits non-zero
784 * prevents us from getting new data interspersed with
785 * retransmissions.
786 */
787
788 if (sk->send_head->link3) /* Any more queued retransmits? */
789 sk->retransmits = 1;
790 else
791 sk->retransmits = 0;
792 }
793 /*
794 * Note that we only reset backoff and rto in the
795 * rtt recomputation code. And that doesn't happen
796 * if there were retransmissions in effect. So the
797 * first new packet after the retransmissions is
798 * sent with the backoff still in effect. Not until
799 * we get an ack from a non-retransmitted packet do
800 * we reset the backoff and rto. This allows us to deal
801 * with a situation where the network delay has increased
802 * suddenly. I.e. Karn's algorithm. (SIGCOMM '87, p5.)
803 */
804
805 /*
806 * We have one less packet out there.
807 */
808
809 if (sk->packets_out > 0)
810 sk->packets_out --;
811
812 oskb = sk->send_head;
813
814 if (!(flag&2)) /* Not retransmitting */
815 tcp_rtt_estimator(sk,oskb);
816 flag |= (2|4); /* 2 is really more like 'don't adjust the rtt
817 In this case as we just set it up */
818 cli();
819 oskb = sk->send_head;
820 IS_SKB(oskb);
821 sk->send_head = oskb->link3;
822 if (sk->send_head == NULL)
823 {
824 sk->send_tail = NULL;
825 }
826
827 /*
828 * We may need to remove this from the dev send list.
829 */
830
831 if (oskb->next)
832 skb_unlink(oskb);
833 sti();
834 kfree_skb(oskb, FREE_WRITE); /* write. */
835 if (!sk->dead)
836 sk->write_space(sk);
837 }
838 else
839 {
840 break;
841 }
842 }
843
844 /*
845 * XXX someone ought to look at this too.. at the moment, if skb_peek()
846 * returns non-NULL, we complete ignore the timer stuff in the else
847 * clause. We ought to organize the code so that else clause can
848 * (should) be executed regardless, possibly moving the PROBE timer
849 * reset over. The skb_peek() thing should only move stuff to the
850 * write queue, NOT also manage the timer functions.
851 */
852
853 /*
854 * Maybe we can take some stuff off of the write queue,
855 * and put it onto the xmit queue.
856 */
857 if (skb_peek(&sk->write_queue) != NULL)
858 {
859 if (after (sk->window_seq+1, sk->write_queue.next->end_seq) &&
860 (sk->retransmits == 0 ||
861 sk->ip_xmit_timeout != TIME_WRITE ||
862 before(sk->write_queue.next->end_seq, sk->rcv_ack_seq + 1))
863 && sk->packets_out < sk->cong_window)
864 {
865 /*
866 * Add more data to the send queue.
867 */
868 flag |= 1;
869 tcp_write_xmit(sk);
870 }
871 else if (before(sk->window_seq, sk->write_queue.next->end_seq) &&
872 sk->send_head == NULL &&
873 sk->ack_backlog == 0 &&
874 sk->state != TCP_TIME_WAIT)
875 {
876 /*
877 * Data to queue but no room.
878 */
879 tcp_reset_xmit_timer(sk, TIME_PROBE0, sk->rto);
880 }
881 }
882 else
883 {
884 /*
885 * from TIME_WAIT we stay in TIME_WAIT as long as we rx packets
886 * from TCP_CLOSE we don't do anything
887 *
888 * from anything else, if there is write data (or fin) pending,
889 * we use a TIME_WRITE timeout, else if keepalive we reset to
890 * a KEEPALIVE timeout, else we delete the timer.
891 *
892 * We do not set flag for nominal write data, otherwise we may
893 * force a state where we start to write itsy bitsy tidbits
894 * of data.
895 */
896
897 switch(sk->state) {
898 case TCP_TIME_WAIT:
899 /*
900 * keep us in TIME_WAIT until we stop getting packets,
901 * reset the timeout.
902 */
903 tcp_reset_msl_timer(sk, TIME_CLOSE, TCP_TIMEWAIT_LEN);
904 break;
905 case TCP_CLOSE:
906 /*
907 * don't touch the timer.
908 */
909 break;
910 default:
911 /*
912 * Must check send_head, write_queue, and ack_backlog
913 * to determine which timeout to use.
914 */
915 if (sk->send_head || skb_peek(&sk->write_queue) != NULL || sk->ack_backlog) {
916 tcp_reset_xmit_timer(sk, TIME_WRITE, sk->rto);
917 } else if (sk->keepopen) {
918 tcp_reset_xmit_timer(sk, TIME_KEEPOPEN, TCP_TIMEOUT_LEN);
919 } else {
920 del_timer(&sk->retransmit_timer);
921 sk->ip_xmit_timeout = 0;
922 }
923 break;
924 }
925 }
926
927 /*
928 * We have nothing queued but space to send. Send any partial
929 * packets immediately (end of Nagle rule application).
930 */
931
932 if (sk->packets_out == 0 && sk->partial != NULL &&
933 skb_peek(&sk->write_queue) == NULL && sk->send_head == NULL)
934 {
935 flag |= 1;
936 tcp_send_partial(sk);
937 }
938
939 /*
940 * In the LAST_ACK case, the other end FIN'd us. We then FIN'd them, and
941 * we are now waiting for an acknowledge to our FIN. The other end is
942 * already in TIME_WAIT.
943 *
944 * Move to TCP_CLOSE on success.
945 */
946
947 if (sk->state == TCP_LAST_ACK)
948 {
949 if (!sk->dead)
950 sk->state_change(sk);
951 if(sk->debug)
952 printk("rcv_ack_seq: %X==%X, acked_seq: %X==%X\n",
953 sk->rcv_ack_seq,sk->write_seq,sk->acked_seq,sk->fin_seq);
954 if (sk->rcv_ack_seq == sk->write_seq /*&& sk->acked_seq == sk->fin_seq*/)
955 {
956 flag |= 1;
957 sk->shutdown = SHUTDOWN_MASK;
958 tcp_set_state(sk,TCP_CLOSE);
959 return 1;
960 }
961 }
962
963 /*
964 * Incoming ACK to a FIN we sent in the case of our initiating the close.
965 *
966 * Move to FIN_WAIT2 to await a FIN from the other end. Set
967 * SEND_SHUTDOWN but not RCV_SHUTDOWN as data can still be coming in.
968 */
969
970 if (sk->state == TCP_FIN_WAIT1)
971 {
972
973 if (!sk->dead)
974 sk->state_change(sk);
975 if (sk->rcv_ack_seq == sk->write_seq)
976 {
977 flag |= 1;
978 sk->shutdown |= SEND_SHUTDOWN;
979 tcp_set_state(sk, TCP_FIN_WAIT2);
980 }
981 }
982
983 /*
984 * Incoming ACK to a FIN we sent in the case of a simultaneous close.
985 *
986 * Move to TIME_WAIT
987 */
988
989 if (sk->state == TCP_CLOSING)
990 {
991
992 if (!sk->dead)
993 sk->state_change(sk);
994 if (sk->rcv_ack_seq == sk->write_seq)
995 {
996 flag |= 1;
997 tcp_time_wait(sk);
998 }
999 }
1000
1001 /*
1002 * Final ack of a three way shake
1003 */
1004
1005 if(sk->state==TCP_SYN_RECV)
1006 {
1007 tcp_set_state(sk, TCP_ESTABLISHED);
1008 tcp_options(sk,th);
1009 sk->dummy_th.dest=th->source;
1010 sk->copied_seq = sk->acked_seq;
1011 if(!sk->dead)
1012 sk->state_change(sk);
1013 if(sk->max_window==0)
1014 {
1015 sk->max_window=32; /* Sanity check */
1016 sk->mss=min(sk->max_window,sk->mtu);
1017 }
1018 }
1019
1020 /*
1021 * I make no guarantees about the first clause in the following
1022 * test, i.e. "(!flag) || (flag&4)". I'm not entirely sure under
1023 * what conditions "!flag" would be true. However I think the rest
1024 * of the conditions would prevent that from causing any
1025 * unnecessary retransmission.
1026 * Clearly if the first packet has expired it should be
1027 * retransmitted. The other alternative, "flag&2 && retransmits", is
1028 * harder to explain: You have to look carefully at how and when the
1029 * timer is set and with what timeout. The most recent transmission always
1030 * sets the timer. So in general if the most recent thing has timed
1031 * out, everything before it has as well. So we want to go ahead and
1032 * retransmit some more. If we didn't explicitly test for this
1033 * condition with "flag&2 && retransmits", chances are "when + rto < jiffies"
1034 * would not be true. If you look at the pattern of timing, you can
1035 * show that rto is increased fast enough that the next packet would
1036 * almost never be retransmitted immediately. Then you'd end up
1037 * waiting for a timeout to send each packet on the retransmission
1038 * queue. With my implementation of the Karn sampling algorithm,
1039 * the timeout would double each time. The net result is that it would
1040 * take a hideous amount of time to recover from a single dropped packet.
1041 * It's possible that there should also be a test for TIME_WRITE, but
1042 * I think as long as "send_head != NULL" and "retransmit" is on, we've
1043 * got to be in real retransmission mode.
1044 * Note that tcp_do_retransmit is called with all==1. Setting cong_window
1045 * back to 1 at the timeout will cause us to send 1, then 2, etc. packets.
1046 * As long as no further losses occur, this seems reasonable.
1047 */
1048
1049 if (((!flag) || (flag&4)) && sk->send_head != NULL &&
1050 (((flag&2) && sk->retransmits) ||
1051 (sk->send_head->when + sk->rto < jiffies)))
1052 {
1053 if(sk->send_head->when + sk->rto < jiffies)
1054 tcp_retransmit(sk,0);
1055 else
1056 {
1057 tcp_do_retransmit(sk, 1);
1058 tcp_reset_xmit_timer(sk, TIME_WRITE, sk->rto);
1059 }
1060 }
1061
1062 return 1;
1063
1064 uninteresting_ack:
1065 if(sk->debug)
1066 printk("Ack ignored %u %u\n",ack,sk->sent_seq);
1067
1068 /*
1069 * Keepalive processing.
1070 */
1071
1072 if (after(ack, sk->sent_seq))
1073 {
1074 return 0;
1075 }
1076
1077 /*
1078 * Restart the keepalive timer.
1079 */
1080
1081 if (sk->keepopen)
1082 {
1083 if(sk->ip_xmit_timeout==TIME_KEEPOPEN)
1084 tcp_reset_xmit_timer(sk, TIME_KEEPOPEN, TCP_TIMEOUT_LEN);
1085 }
1086 return 1;
1087 }
1088
1089
1090 /*
1091 * Process the FIN bit. This now behaves as it is supposed to work
1092 * and the FIN takes effect when it is validly part of sequence
1093 * space. Not before when we get holes.
1094 *
1095 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
1096 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
1097 * TIME-WAIT)
1098 *
1099 * If we are in FINWAIT-1, a received FIN indicates simultaneous
1100 * close and we go into CLOSING (and later onto TIME-WAIT)
1101 *
1102 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
1103 *
1104 */
1105
1106 static int tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
/* */
1107 {
1108 sk->fin_seq = skb->end_seq;
1109
1110 if (!sk->dead)
1111 {
1112 sk->state_change(sk);
1113 sock_wake_async(sk->socket, 1);
1114 }
1115
1116 switch(sk->state)
1117 {
1118 case TCP_SYN_RECV:
1119 case TCP_SYN_SENT:
1120 case TCP_ESTABLISHED:
1121 /*
1122 * move to CLOSE_WAIT, tcp_data() already handled
1123 * sending the ack.
1124 */
1125 tcp_set_state(sk,TCP_CLOSE_WAIT);
1126 if (th->rst)
1127 sk->shutdown = SHUTDOWN_MASK;
1128 break;
1129
1130 case TCP_CLOSE_WAIT:
1131 case TCP_CLOSING:
1132 /*
1133 * received a retransmission of the FIN, do
1134 * nothing.
1135 */
1136 break;
1137 case TCP_TIME_WAIT:
1138 /*
1139 * received a retransmission of the FIN,
1140 * restart the TIME_WAIT timer.
1141 */
1142 tcp_reset_msl_timer(sk, TIME_CLOSE, TCP_TIMEWAIT_LEN);
1143 return(0);
1144 case TCP_FIN_WAIT1:
1145 /*
1146 * This case occurs when a simultaneous close
1147 * happens, we must ack the received FIN and
1148 * enter the CLOSING state.
1149 *
1150 * This causes a WRITE timeout, which will either
1151 * move on to TIME_WAIT when we timeout, or resend
1152 * the FIN properly (maybe we get rid of that annoying
1153 * FIN lost hang). The TIME_WRITE code is already correct
1154 * for handling this timeout.
1155 */
1156
1157 if(sk->ip_xmit_timeout != TIME_WRITE)
1158 tcp_reset_xmit_timer(sk, TIME_WRITE, sk->rto);
1159 tcp_set_state(sk,TCP_CLOSING);
1160 break;
1161 case TCP_FIN_WAIT2:
1162 /*
1163 * received a FIN -- send ACK and enter TIME_WAIT
1164 */
1165 tcp_reset_msl_timer(sk, TIME_CLOSE, TCP_TIMEWAIT_LEN);
1166 sk->shutdown|=SHUTDOWN_MASK;
1167 tcp_set_state(sk,TCP_TIME_WAIT);
1168 break;
1169 case TCP_CLOSE:
1170 /*
1171 * already in CLOSE
1172 */
1173 break;
1174 default:
1175 tcp_set_state(sk,TCP_LAST_ACK);
1176
1177 /* Start the timers. */
1178 tcp_reset_msl_timer(sk, TIME_CLOSE, TCP_TIMEWAIT_LEN);
1179 return(0);
1180 }
1181
1182 return(0);
1183 }
1184
1185 /*
1186 * Called for each packet when we find a new ACK endpoint sequence in it
1187 */
1188 static inline u32 tcp_queue_ack(struct sk_buff * skb, struct sock * sk)
/* */
1189 {
1190 /*
1191 * When we ack the fin, we do the FIN
1192 * processing.
1193 */
1194 skb->acked = 1;
1195 if (skb->h.th->fin)
1196 tcp_fin(skb,sk,skb->h.th);
1197 return skb->end_seq;
1198 }
1199
1200
1201 /*
1202 * Add a sk_buff to the TCP receive queue, calculating
1203 * the ACK sequence as we go..
1204 */
1205 static void tcp_queue(struct sk_buff * skb, struct sock * sk,
/* */
1206 struct tcphdr *th, unsigned long saddr)
1207 {
1208 struct sk_buff_head * list = &sk->receive_queue;
1209 struct sk_buff * next;
1210 u32 ack_seq;
1211
1212 /*
1213 * Find where the new skb goes.. (This goes backwards,
1214 * on the assumption that we get the packets in order)
1215 */
1216 next = list->prev;
1217 while (next != (struct sk_buff *) list) {
1218 if (!after(next->seq, skb->seq))
1219 break;
1220 next = next->prev;
1221 }
1222 /*
1223 * put it after the packet we found (which
1224 * may be the list-head, but that's fine).
1225 */
1226 __skb_append(next, skb, list);
1227 next = skb->next;
1228
1229 /*
1230 * Did we get anything new to ack?
1231 */
1232 ack_seq = sk->acked_seq;
1233 if (!after(skb->seq, ack_seq) && after(skb->end_seq, ack_seq)) {
1234 ack_seq = tcp_queue_ack(skb, sk);
1235
1236 /*
1237 * Do we have any old packets to ack that the above
1238 * made visible? (Go forward from skb)
1239 */
1240 while (next != (struct sk_buff *) list) {
1241 if (after(next->seq, ack_seq))
1242 break;
1243 if (after(next->end_seq, ack_seq))
1244 ack_seq = tcp_queue_ack(next, sk);
1245 next = next->next;
1246 }
1247
1248 /*
1249 * Ok, we found new data, update acked_seq as
1250 * necessary (and possibly send the actual
1251 * ACK packet).
1252 *
1253 * rules for delaying an ack:
1254 * - delay time <= 0.5 HZ
1255 * - we don't have a window update to send
1256 * - must send at least every 2 full sized packets
1257 */
1258 sk->acked_seq = ack_seq;
1259 if (!sk->delay_acks ||
1260 /* sk->ack_backlog >= sk->max_ack_backlog || */
1261 sk->bytes_rcv > sk->max_unacked || th->fin ||
1262 sk->ato > HZ/2) {
1263 tcp_send_ack(sk->sent_seq, sk->acked_seq, sk, th, saddr);
1264 }
1265 else
1266 {
1267 sk->ack_backlog++;
1268 if(sk->debug)
1269 printk("Ack queued.\n");
1270 tcp_reset_xmit_timer(sk, TIME_WRITE, sk->ato);
1271 }
1272 }
1273 }
1274
1275
1276 /*
1277 * This routine handles the data. If there is room in the buffer,
1278 * it will be have already been moved into it. If there is no
1279 * room, then we will just have to discard the packet.
1280 */
1281
1282 static int tcp_data(struct sk_buff *skb, struct sock *sk,
/* */
1283 unsigned long saddr, unsigned short len)
1284 {
1285 struct tcphdr *th;
1286 u32 new_seq, shut_seq;
1287
1288 th = skb->h.th;
1289 skb_pull(skb,th->doff*4);
1290 skb_trim(skb,len-(th->doff*4));
1291
1292 /*
1293 * The bytes in the receive read/assembly queue has increased. Needed for the
1294 * low memory discard algorithm
1295 */
1296
1297 sk->bytes_rcv += skb->len;
1298
1299 if (skb->len == 0 && !th->fin)
1300 {
1301 /*
1302 * Don't want to keep passing ack's back and forth.
1303 * (someone sent us dataless, boring frame)
1304 */
1305 if (!th->ack)
1306 tcp_send_ack(sk->sent_seq, sk->acked_seq,sk, th, saddr);
1307 kfree_skb(skb, FREE_READ);
1308 return(0);
1309 }
1310
1311 /*
1312 * We no longer have anyone receiving data on this connection.
1313 */
1314
1315 #ifndef TCP_DONT_RST_SHUTDOWN
1316
1317 if(sk->shutdown & RCV_SHUTDOWN)
1318 {
1319 /*
1320 * FIXME: BSD has some magic to avoid sending resets to
1321 * broken 4.2 BSD keepalives. Much to my surprise a few non
1322 * BSD stacks still have broken keepalives so we want to
1323 * cope with it.
1324 */
1325
1326 if(skb->len) /* We don't care if it's just an ack or
1327 a keepalive/window probe */
1328 {
1329 new_seq = skb->seq + skb->len + th->syn; /* Right edge of _data_ part of frame */
1330
1331 /* Do this the way 4.4BSD treats it. Not what I'd
1332 regard as the meaning of the spec but it's what BSD
1333 does and clearly they know everything 8) */
1334
1335 /*
1336 * This is valid because of two things
1337 *
1338 * a) The way tcp_data behaves at the bottom.
1339 * b) A fin takes effect when read not when received.
1340 */
1341
1342 shut_seq = sk->acked_seq+1; /* Last byte */
1343
1344 if(after(new_seq,shut_seq))
1345 {
1346 if(sk->debug)
1347 printk("Data arrived on %p after close [Data right edge %X, Socket shut on %X] %d\n",
1348 sk, new_seq, shut_seq, sk->blog);
1349 if(sk->dead)
1350 {
1351 sk->acked_seq = new_seq + th->fin;
1352 tcp_send_reset(sk->saddr, sk->daddr, skb->h.th,
1353 sk->prot, NULL, skb->dev, sk->ip_tos, sk->ip_ttl);
1354 tcp_statistics.TcpEstabResets++;
1355 sk->err = EPIPE;
1356 sk->error_report(sk);
1357 sk->shutdown = SHUTDOWN_MASK;
1358 tcp_set_state(sk,TCP_CLOSE);
1359 kfree_skb(skb, FREE_READ);
1360 return 0;
1361 }
1362 }
1363 }
1364 }
1365
1366 #endif
1367
1368 tcp_queue(skb, sk, th, saddr);
1369
1370 /*
1371 * If we've missed a packet, send an ack.
1372 * Also start a timer to send another.
1373 */
1374
1375 if (!skb->acked)
1376 {
1377
1378 /*
1379 * This is important. If we don't have much room left,
1380 * we need to throw out a few packets so we have a good
1381 * window. Note that mtu is used, not mss, because mss is really
1382 * for the send side. He could be sending us stuff as large as mtu.
1383 */
1384
1385 while (sock_rspace(sk) < sk->mtu)
1386 {
1387 struct sk_buff * skb1 = skb_peek(&sk->receive_queue);
1388 if (skb1 == NULL)
1389 {
1390 printk("INET: tcp.c:tcp_data memory leak detected.\n");
1391 break;
1392 }
1393
1394 /*
1395 * Don't throw out something that has been acked.
1396 */
1397
1398 if (skb1->acked)
1399 {
1400 break;
1401 }
1402
1403 skb_unlink(skb1);
1404 kfree_skb(skb1, FREE_READ);
1405 }
1406 tcp_send_ack(sk->sent_seq, sk->acked_seq, sk, th, saddr);
1407 sk->ack_backlog++;
1408 tcp_reset_xmit_timer(sk, TIME_WRITE, min(sk->ato, HZ/2));
1409 }
1410
1411 /*
1412 * Now tell the user we may have some data.
1413 */
1414
1415 if (!sk->dead)
1416 {
1417 if(sk->debug)
1418 printk("Data wakeup.\n");
1419 sk->data_ready(sk,0);
1420 }
1421 return(0);
1422 }
1423
1424
1425 /*
1426 * This routine is only called when we have urgent data
1427 * signalled. Its the 'slow' part of tcp_urg. It could be
1428 * moved inline now as tcp_urg is only called from one
1429 * place. We handle URGent data wrong. We have to - as
1430 * BSD still doesn't use the correction from RFC961.
1431 */
1432
1433 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
/* */
1434 {
1435 u32 ptr = ntohs(th->urg_ptr);
1436
1437 if (ptr)
1438 ptr--;
1439 ptr += ntohl(th->seq);
1440
1441 /* ignore urgent data that we've already seen and read */
1442 if (after(sk->copied_seq, ptr))
1443 return;
1444
1445 /* do we already have a newer (or duplicate) urgent pointer? */
1446 if (sk->urg_data && !after(ptr, sk->urg_seq))
1447 return;
1448
1449 /* tell the world about our new urgent pointer */
1450 if (sk->proc != 0) {
1451 if (sk->proc > 0) {
1452 kill_proc(sk->proc, SIGURG, 1);
1453 } else {
1454 kill_pg(-sk->proc, SIGURG, 1);
1455 }
1456 }
1457 sk->urg_data = URG_NOTYET;
1458 sk->urg_seq = ptr;
1459 }
1460
1461 /*
1462 * This is the 'fast' part of urgent handling.
1463 */
1464
1465 static inline void tcp_urg(struct sock *sk, struct tcphdr *th, unsigned long len)
/* */
1466 {
1467 /*
1468 * Check if we get a new urgent pointer - normally not
1469 */
1470
1471 if (th->urg)
1472 tcp_check_urg(sk,th);
1473
1474 /*
1475 * Do we wait for any urgent data? - normally not
1476 */
1477
1478 if (sk->urg_data == URG_NOTYET) {
1479 u32 ptr;
1480
1481 /*
1482 * Is the urgent pointer pointing into this packet?
1483 */
1484 ptr = sk->urg_seq - ntohl(th->seq) + th->doff*4;
1485 if (ptr < len) {
1486 sk->urg_data = URG_VALID | *(ptr + (unsigned char *) th);
1487 if (!sk->dead)
1488 sk->data_ready(sk,0);
1489 }
1490 }
1491 }
1492
1493
1494 /*
1495 * A TCP packet has arrived.
1496 * skb->h.raw is the TCP header.
1497 */
1498
1499 int tcp_rcv(struct sk_buff *skb, struct device *dev, struct options *opt,
/* ![[last]](../icons/n_last.png)
*/
1500 __u32 daddr, unsigned short len,
1501 __u32 saddr, int redo, struct inet_protocol * protocol)
1502 {
1503 struct tcphdr *th;
1504 struct sock *sk;
1505 int syn_ok=0;
1506
1507 /*
1508 * "redo" is 1 if we have already seen this skb but couldn't
1509 * use it at that time (the socket was locked). In that case
1510 * we have already done a lot of the work (looked up the socket
1511 * etc).
1512 */
1513 th = skb->h.th;
1514 sk = skb->sk;
1515 if (!redo) {
1516 tcp_statistics.TcpInSegs++;
1517 if (skb->pkt_type!=PACKET_HOST)
1518 goto discard_it;
1519
1520 /*
1521 * Pull up the IP header.
1522 */
1523
1524 skb_pull(skb, skb->h.raw-skb->data);
1525
1526 /*
1527 * Try to use the device checksum if provided.
1528 */
1529 switch (skb->ip_summed)
1530 {
1531 case CHECKSUM_NONE:
1532 skb->csum = csum_partial((char *)th, len, 0);
1533 case CHECKSUM_HW:
1534 if (tcp_check(th, len, saddr, daddr, skb->csum))
1535 goto discard_it;
1536 default:
1537 /* CHECKSUM_UNNECESSARY */
1538 }
1539 sk = get_tcp_sock(saddr, th->source, daddr, th->dest);
1540 if (!sk)
1541 goto no_tcp_socket;
1542 skb->sk = sk;
1543 skb->seq = ntohl(th->seq);
1544 skb->end_seq = skb->seq + th->syn + th->fin + len - th->doff*4;
1545 skb->ack_seq = ntohl(th->ack_seq);
1546
1547 skb->acked = 0;
1548 skb->used = 0;
1549 skb->free = 1;
1550 skb->saddr = daddr;
1551 skb->daddr = saddr;
1552
1553 /*
1554 * We may need to add it to the backlog here.
1555 */
1556 if (sk->users)
1557 {
1558 __skb_queue_tail(&sk->back_log, skb);
1559 return(0);
1560 }
1561 }
1562
1563 /*
1564 * If this socket has got a reset it's to all intents and purposes
1565 * really dead. Count closed sockets as dead.
1566 *
1567 * Note: BSD appears to have a bug here. A 'closed' TCP in BSD
1568 * simply drops data. This seems incorrect as a 'closed' TCP doesn't
1569 * exist so should cause resets as if the port was unreachable.
1570 */
1571
1572 if (sk->zapped || sk->state==TCP_CLOSE)
1573 goto no_tcp_socket;
1574
1575 if (!sk->prot)
1576 {
1577 printk("IMPOSSIBLE 3\n");
1578 return(0);
1579 }
1580
1581
1582 /*
1583 * Charge the memory to the socket.
1584 */
1585
1586 skb->sk=sk;
1587 sk->rmem_alloc += skb->truesize;
1588
1589 /*
1590 * We should now do header prediction.
1591 */
1592
1593 /*
1594 * This basically follows the flow suggested by RFC793, with the corrections in RFC1122. We
1595 * don't implement precedence and we process URG incorrectly (deliberately so) for BSD bug
1596 * compatibility. We also set up variables more thoroughly [Karn notes in the
1597 * KA9Q code the RFC793 incoming segment rules don't initialise the variables for all paths].
1598 */
1599
1600 if(sk->state!=TCP_ESTABLISHED) /* Skip this lot for normal flow */
1601 {
1602
1603 /*
1604 * Now deal with unusual cases.
1605 */
1606
1607 if(sk->state==TCP_LISTEN)
1608 {
1609 if(th->ack) /* These use the socket TOS.. might want to be the received TOS */
1610 tcp_send_reset(daddr,saddr,th,sk->prot,opt,dev,sk->ip_tos, sk->ip_ttl);
1611
1612 /*
1613 * We don't care for RST, and non SYN are absorbed (old segments)
1614 * Broadcast/multicast SYN isn't allowed. Note - bug if you change the
1615 * netmask on a running connection it can go broadcast. Even Sun's have
1616 * this problem so I'm ignoring it
1617 */
1618
1619 if(th->rst || !th->syn || th->ack || ip_chk_addr(daddr)!=IS_MYADDR)
1620 {
1621 kfree_skb(skb, FREE_READ);
1622 return 0;
1623 }
1624
1625 /*
1626 * Guess we need to make a new socket up
1627 */
1628
1629 tcp_conn_request(sk, skb, daddr, saddr, opt, dev, tcp_init_seq());
1630
1631 /*
1632 * Now we have several options: In theory there is nothing else
1633 * in the frame. KA9Q has an option to send data with the syn,
1634 * BSD accepts data with the syn up to the [to be] advertised window
1635 * and Solaris 2.1 gives you a protocol error. For now we just ignore
1636 * it, that fits the spec precisely and avoids incompatibilities. It
1637 * would be nice in future to drop through and process the data.
1638 *
1639 * Now TTCP is starting to use we ought to queue this data.
1640 */
1641
1642 return 0;
1643 }
1644
1645 /*
1646 * Retransmitted SYN for our socket. This is uninteresting. If sk->state==TCP_LISTEN
1647 * then its a new connection
1648 */
1649
1650 if (sk->state == TCP_SYN_RECV && th->syn && skb->seq+1 == sk->acked_seq)
1651 {
1652 kfree_skb(skb, FREE_READ);
1653 return 0;
1654 }
1655
1656 /*
1657 * SYN sent means we have to look for a suitable ack and either reset
1658 * for bad matches or go to connected. The SYN_SENT case is unusual and should
1659 * not be in line code. [AC]
1660 */
1661
1662 if(sk->state==TCP_SYN_SENT)
1663 {
1664 /* Crossed SYN or previous junk segment */
1665 if(th->ack)
1666 {
1667 /* We got an ack, but it's not a good ack */
1668 if(!tcp_ack(sk,th,skb->ack_seq,len))
1669 {
1670 /* Reset the ack - its an ack from a
1671 different connection [ th->rst is checked in tcp_send_reset()] */
1672 tcp_statistics.TcpAttemptFails++;
1673 tcp_send_reset(daddr, saddr, th,
1674 sk->prot, opt,dev,sk->ip_tos,sk->ip_ttl);
1675 kfree_skb(skb, FREE_READ);
1676 return(0);
1677 }
1678 if(th->rst)
1679 return tcp_reset(sk,skb);
1680 if(!th->syn)
1681 {
1682 /* A valid ack from a different connection
1683 start. Shouldn't happen but cover it */
1684 tcp_statistics.TcpAttemptFails++;
1685 tcp_send_reset(daddr, saddr, th,
1686 sk->prot, opt,dev,sk->ip_tos,sk->ip_ttl);
1687 kfree_skb(skb, FREE_READ);
1688 return 0;
1689 }
1690 /*
1691 * Ok.. it's good. Set up sequence numbers and
1692 * move to established.
1693 */
1694 syn_ok=1; /* Don't reset this connection for the syn */
1695 sk->acked_seq = skb->seq+1;
1696 sk->lastwin_seq = skb->seq+1;
1697 sk->fin_seq = skb->seq;
1698 tcp_send_ack(sk->sent_seq,sk->acked_seq,sk,th,sk->daddr);
1699 tcp_set_state(sk, TCP_ESTABLISHED);
1700 tcp_options(sk,th);
1701 sk->dummy_th.dest=th->source;
1702 sk->copied_seq = sk->acked_seq;
1703 if(!sk->dead)
1704 {
1705 sk->state_change(sk);
1706 sock_wake_async(sk->socket, 0);
1707 }
1708 if(sk->max_window==0)
1709 {
1710 sk->max_window = 32;
1711 sk->mss = min(sk->max_window, sk->mtu);
1712 }
1713 }
1714 else
1715 {
1716 /* See if SYN's cross. Drop if boring */
1717 if(th->syn && !th->rst)
1718 {
1719 /* Crossed SYN's are fine - but talking to
1720 yourself is right out... */
1721 if(sk->saddr==saddr && sk->daddr==daddr &&
1722 sk->dummy_th.source==th->source &&
1723 sk->dummy_th.dest==th->dest)
1724 {
1725 tcp_statistics.TcpAttemptFails++;
1726 return tcp_reset(sk,skb);
1727 }
1728 tcp_set_state(sk,TCP_SYN_RECV);
1729
1730 /*
1731 * FIXME:
1732 * Must send SYN|ACK here
1733 */
1734 }
1735 /* Discard junk segment */
1736 kfree_skb(skb, FREE_READ);
1737 return 0;
1738 }
1739 /*
1740 * SYN_RECV with data maybe.. drop through
1741 */
1742 goto rfc_step6;
1743 }
1744
1745 /*
1746 * BSD has a funny hack with TIME_WAIT and fast reuse of a port. There is
1747 * a more complex suggestion for fixing these reuse issues in RFC1644
1748 * but not yet ready for general use. Also see RFC1379.
1749 *
1750 * Note the funny way we go back to the top of this function for
1751 * this case ("goto try_next_socket"). That also takes care of
1752 * checking "sk->users" for the new socket as well as doing all
1753 * the normal tests on the packet.
1754 */
1755
1756 #define BSD_TIME_WAIT
1757 #ifdef BSD_TIME_WAIT
1758 if (sk->state == TCP_TIME_WAIT && th->syn && sk->dead &&
1759 after(skb->seq, sk->acked_seq) && !th->rst)
1760 {
1761 u32 seq = sk->write_seq;
1762 if(sk->debug)
1763 printk("Doing a BSD time wait\n");
1764 tcp_statistics.TcpEstabResets++;
1765 sk->rmem_alloc -= skb->truesize;
1766 skb->sk = NULL;
1767 sk->err=ECONNRESET;
1768 tcp_set_state(sk, TCP_CLOSE);
1769 sk->shutdown = SHUTDOWN_MASK;
1770 sk=get_sock(&tcp_prot, th->dest, saddr, th->source, daddr);
1771 /* this is not really correct: we should check sk->users */
1772 if (sk && sk->state==TCP_LISTEN)
1773 {
1774 skb->sk = sk;
1775 sk->rmem_alloc += skb->truesize;
1776 tcp_conn_request(sk, skb, daddr, saddr,opt, dev,seq+128000);
1777 return 0;
1778 }
1779 kfree_skb(skb, FREE_READ);
1780 return 0;
1781 }
1782 #endif
1783 }
1784
1785 /*
1786 * We are now in normal data flow (see the step list in the RFC)
1787 * Note most of these are inline now. I'll inline the lot when
1788 * I have time to test it hard and look at what gcc outputs
1789 */
1790
1791 if (!tcp_sequence(sk, skb->seq, skb->end_seq-th->syn))
1792 {
1793 bad_tcp_sequence(sk, th, len, opt, saddr, dev);
1794 kfree_skb(skb, FREE_READ);
1795 return 0;
1796 }
1797
1798 if(th->rst)
1799 return tcp_reset(sk,skb);
1800
1801 /*
1802 * !syn_ok is effectively the state test in RFC793.
1803 */
1804
1805 if(th->syn && !syn_ok)
1806 {
1807 tcp_send_reset(daddr,saddr,th, &tcp_prot, opt, dev, skb->ip_hdr->tos, 255);
1808 return tcp_reset(sk,skb);
1809 }
1810
1811 tcp_delack_estimator(sk);
1812
1813 /*
1814 * Process the ACK
1815 */
1816
1817
1818 if(th->ack && !tcp_ack(sk,th,skb->ack_seq,len))
1819 {
1820 /*
1821 * Our three way handshake failed.
1822 */
1823
1824 if(sk->state==TCP_SYN_RECV)
1825 {
1826 tcp_send_reset(daddr, saddr, th,sk->prot, opt, dev,sk->ip_tos,sk->ip_ttl);
1827 }
1828 kfree_skb(skb, FREE_READ);
1829 return 0;
1830 }
1831
1832 rfc_step6: /* I'll clean this up later */
1833
1834 /*
1835 * If the accepted buffer put us over our queue size we
1836 * now drop it (we must process the ack first to avoid
1837 * deadlock cases).
1838 */
1839
1840 if (sk->rmem_alloc >= sk->rcvbuf)
1841 {
1842 kfree_skb(skb, FREE_READ);
1843 return(0);
1844 }
1845
1846
1847 /*
1848 * Process urgent data
1849 */
1850
1851 tcp_urg(sk, th, len);
1852
1853 /*
1854 * Process the encapsulated data
1855 */
1856
1857 if(tcp_data(skb,sk, saddr, len))
1858 kfree_skb(skb, FREE_READ);
1859
1860 /*
1861 * And done
1862 */
1863
1864 return 0;
1865
1866 no_tcp_socket:
1867 /*
1868 * No such TCB. If th->rst is 0 send a reset (checked in tcp_send_reset)
1869 */
1870 tcp_send_reset(daddr, saddr, th, &tcp_prot, opt,dev,skb->ip_hdr->tos,255);
1871
1872 discard_it:
1873 /*
1874 * Discard frame
1875 */
1876 skb->sk = NULL;
1877 kfree_skb(skb, FREE_READ);
1878 return 0;
1879 }
![[bottom]](../icons/n_bottom.png){kind=icon}
*/