port more changes to make PCI work
[linux-2.4.git] / net / ipv4 / ip_options.c
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  *              The options processing module for ip.c
7  *
8  * Version:     $Id: ip_options.c,v 1.21 2001/09/01 00:31:50 davem Exp $
9  *
10  * Authors:     A.N.Kuznetsov
11  *              
12  */
13
14 #include <linux/types.h>
15 #include <asm/uaccess.h>
16 #include <linux/skbuff.h>
17 #include <linux/ip.h>
18 #include <linux/icmp.h>
19 #include <linux/netdevice.h>
20 #include <linux/rtnetlink.h>
21 #include <net/sock.h>
22 #include <net/ip.h>
23 #include <net/icmp.h>
24
25 /* 
26  * Write options to IP header, record destination address to
27  * source route option, address of outgoing interface
28  * (we should already know it, so that this  function is allowed be
29  * called only after routing decision) and timestamp,
30  * if we originate this datagram.
31  *
32  * daddr is real destination address, next hop is recorded in IP header.
33  * saddr is address of outgoing interface.
34  */
35
36 void ip_options_build(struct sk_buff * skb, struct ip_options * opt,
37                             u32 daddr, struct rtable *rt, int is_frag) 
38 {
39         unsigned char * iph = skb->nh.raw;
40
41         memcpy(&(IPCB(skb)->opt), opt, sizeof(struct ip_options));
42         memcpy(iph+sizeof(struct iphdr), opt->__data, opt->optlen);
43         opt = &(IPCB(skb)->opt);
44         opt->is_data = 0;
45
46         if (opt->srr)
47                 memcpy(iph+opt->srr+iph[opt->srr+1]-4, &daddr, 4);
48
49         if (!is_frag) {
50                 if (opt->rr_needaddr)
51                         ip_rt_get_source(iph+opt->rr+iph[opt->rr+2]-5, rt);
52                 if (opt->ts_needaddr)
53                         ip_rt_get_source(iph+opt->ts+iph[opt->ts+2]-9, rt);
54                 if (opt->ts_needtime) {
55                         struct timeval tv;
56                         __u32 midtime;
57                         do_gettimeofday(&tv);
58                         midtime = htonl((tv.tv_sec % 86400) * 1000 + tv.tv_usec / 1000);
59                         memcpy(iph+opt->ts+iph[opt->ts+2]-5, &midtime, 4);
60                 }
61                 return;
62         }
63         if (opt->rr) {
64                 memset(iph+opt->rr, IPOPT_NOP, iph[opt->rr+1]);
65                 opt->rr = 0;
66                 opt->rr_needaddr = 0;
67         }
68         if (opt->ts) {
69                 memset(iph+opt->ts, IPOPT_NOP, iph[opt->ts+1]);
70                 opt->ts = 0;
71                 opt->ts_needaddr = opt->ts_needtime = 0;
72         }
73 }
74
75 /* 
76  * Provided (sopt, skb) points to received options,
77  * build in dopt compiled option set appropriate for answering.
78  * i.e. invert SRR option, copy anothers,
79  * and grab room in RR/TS options.
80  *
81  * NOTE: dopt cannot point to skb.
82  */
83
84 int ip_options_echo(struct ip_options * dopt, struct sk_buff * skb) 
85 {
86         struct ip_options *sopt;
87         unsigned char *sptr, *dptr;
88         int soffset, doffset;
89         int     optlen;
90         u32     daddr;
91
92         memset(dopt, 0, sizeof(struct ip_options));
93
94         dopt->is_data = 1;
95
96         sopt = &(IPCB(skb)->opt);
97
98         if (sopt->optlen == 0) {
99                 dopt->optlen = 0;
100                 return 0;
101         }
102
103         sptr = skb->nh.raw;
104         dptr = dopt->__data;
105
106         if (skb->dst)
107                 daddr = ((struct rtable*)skb->dst)->rt_spec_dst;
108         else
109                 daddr = skb->nh.iph->daddr;
110
111         if (sopt->rr) {
112                 optlen  = sptr[sopt->rr+1];
113                 soffset = sptr[sopt->rr+2];
114                 dopt->rr = dopt->optlen + sizeof(struct iphdr);
115                 memcpy(dptr, sptr+sopt->rr, optlen);
116                 if (sopt->rr_needaddr && soffset <= optlen) {
117                         if (soffset + 3 > optlen)
118                                 return -EINVAL;
119                         dptr[2] = soffset + 4;
120                         dopt->rr_needaddr = 1;
121                 }
122                 dptr += optlen;
123                 dopt->optlen += optlen;
124         }
125         if (sopt->ts) {
126                 optlen = sptr[sopt->ts+1];
127                 soffset = sptr[sopt->ts+2];
128                 dopt->ts = dopt->optlen + sizeof(struct iphdr);
129                 memcpy(dptr, sptr+sopt->ts, optlen);
130                 if (soffset <= optlen) {
131                         if (sopt->ts_needaddr) {
132                                 if (soffset + 3 > optlen)
133                                         return -EINVAL;
134                                 dopt->ts_needaddr = 1;
135                                 soffset += 4;
136                         }
137                         if (sopt->ts_needtime) {
138                                 if (soffset + 3 > optlen)
139                                         return -EINVAL;
140                                 if ((dptr[3]&0xF) != IPOPT_TS_PRESPEC) {
141                                         dopt->ts_needtime = 1;
142                                         soffset += 4;
143                                 } else {
144                                         dopt->ts_needtime = 0;
145
146                                         if (soffset + 8 <= optlen) {
147                                                 __u32 addr;
148
149                                                 memcpy(&addr, sptr+soffset-1, 4);
150                                                 if (inet_addr_type(addr) != RTN_LOCAL) {
151                                                         dopt->ts_needtime = 1;
152                                                         soffset += 8;
153                                                 }
154                                         }
155                                 }
156                         }
157                         dptr[2] = soffset;
158                 }
159                 dptr += optlen;
160                 dopt->optlen += optlen;
161         }
162         if (sopt->srr) {
163                 unsigned char * start = sptr+sopt->srr;
164                 u32 faddr;
165
166                 optlen  = start[1];
167                 soffset = start[2];
168                 doffset = 0;
169                 if (soffset > optlen)
170                         soffset = optlen + 1;
171                 soffset -= 4;
172                 if (soffset > 3) {
173                         memcpy(&faddr, &start[soffset-1], 4);
174                         for (soffset-=4, doffset=4; soffset > 3; soffset-=4, doffset+=4)
175                                 memcpy(&dptr[doffset-1], &start[soffset-1], 4);
176                         /*
177                          * RFC1812 requires to fix illegal source routes.
178                          */
179                         if (memcmp(&skb->nh.iph->saddr, &start[soffset+3], 4) == 0)
180                                 doffset -= 4;
181                 }
182                 if (doffset > 3) {
183                         memcpy(&start[doffset-1], &daddr, 4);
184                         dopt->faddr = faddr;
185                         dptr[0] = start[0];
186                         dptr[1] = doffset+3;
187                         dptr[2] = 4;
188                         dptr += doffset+3;
189                         dopt->srr = dopt->optlen + sizeof(struct iphdr);
190                         dopt->optlen += doffset+3;
191                         dopt->is_strictroute = sopt->is_strictroute;
192                 }
193         }
194         while (dopt->optlen & 3) {
195                 *dptr++ = IPOPT_END;
196                 dopt->optlen++;
197         }
198         return 0;
199 }
200
201 /*
202  *      Options "fragmenting", just fill options not
203  *      allowed in fragments with NOOPs.
204  *      Simple and stupid 8), but the most efficient way.
205  */
206
207 void ip_options_fragment(struct sk_buff * skb) 
208 {
209         unsigned char * optptr = skb->nh.raw;
210         struct ip_options * opt = &(IPCB(skb)->opt);
211         int  l = opt->optlen;
212         int  optlen;
213
214         while (l > 0) {
215                 switch (*optptr) {
216                 case IPOPT_END:
217                         return;
218                 case IPOPT_NOOP:
219                         l--;
220                         optptr++;
221                         continue;
222                 }
223                 optlen = optptr[1];
224                 if (optlen<2 || optlen>l)
225                   return;
226                 if (!IPOPT_COPIED(*optptr))
227                         memset(optptr, IPOPT_NOOP, optlen);
228                 l -= optlen;
229                 optptr += optlen;
230         }
231         opt->ts = 0;
232         opt->rr = 0;
233         opt->rr_needaddr = 0;
234         opt->ts_needaddr = 0;
235         opt->ts_needtime = 0;
236         return;
237 }
238
239 /*
240  * Verify options and fill pointers in struct options.
241  * Caller should clear *opt, and set opt->data.
242  * If opt == NULL, then skb->data should point to IP header.
243  */
244
245 int ip_options_compile(struct ip_options * opt, struct sk_buff * skb)
246 {
247         int l;
248         unsigned char * iph;
249         unsigned char * optptr;
250         int optlen;
251         unsigned char * pp_ptr = NULL;
252         struct rtable *rt = skb ? (struct rtable*)skb->dst : NULL;
253
254         if (!opt) {
255                 opt = &(IPCB(skb)->opt);
256                 memset(opt, 0, sizeof(struct ip_options));
257                 iph = skb->nh.raw;
258                 opt->optlen = ((struct iphdr *)iph)->ihl*4 - sizeof(struct iphdr);
259                 optptr = iph + sizeof(struct iphdr);
260                 opt->is_data = 0;
261         } else {
262                 optptr = opt->is_data ? opt->__data : (unsigned char*)&(skb->nh.iph[1]);
263                 iph = optptr - sizeof(struct iphdr);
264         }
265
266         for (l = opt->optlen; l > 0; ) {
267                 switch (*optptr) {
268                       case IPOPT_END:
269                         for (optptr++, l--; l>0; optptr++, l--) {
270                                 if (*optptr != IPOPT_END) {
271                                         *optptr = IPOPT_END;
272                                         opt->is_changed = 1;
273                                 }
274                         }
275                         goto eol;
276                       case IPOPT_NOOP:
277                         l--;
278                         optptr++;
279                         continue;
280                 }
281                 optlen = optptr[1];
282                 if (optlen<2 || optlen>l) {
283                         pp_ptr = optptr;
284                         goto error;
285                 }
286                 switch (*optptr) {
287                       case IPOPT_SSRR:
288                       case IPOPT_LSRR:
289                         if (optlen < 3) {
290                                 pp_ptr = optptr + 1;
291                                 goto error;
292                         }
293                         if (optptr[2] < 4) {
294                                 pp_ptr = optptr + 2;
295                                 goto error;
296                         }
297                         /* NB: cf RFC-1812 5.2.4.1 */
298                         if (opt->srr) {
299                                 pp_ptr = optptr;
300                                 goto error;
301                         }
302                         if (!skb) {
303                                 if (optptr[2] != 4 || optlen < 7 || ((optlen-3) & 3)) {
304                                         pp_ptr = optptr + 1;
305                                         goto error;
306                                 }
307                                 memcpy(&opt->faddr, &optptr[3], 4);
308                                 if (optlen > 7)
309                                         memmove(&optptr[3], &optptr[7], optlen-7);
310                         }
311                         opt->is_strictroute = (optptr[0] == IPOPT_SSRR);
312                         opt->srr = optptr - iph;
313                         break;
314                       case IPOPT_RR:
315                         if (opt->rr) {
316                                 pp_ptr = optptr;
317                                 goto error;
318                         }
319                         if (optlen < 3) {
320                                 pp_ptr = optptr + 1;
321                                 goto error;
322                         }
323                         if (optptr[2] < 4) {
324                                 pp_ptr = optptr + 2;
325                                 goto error;
326                         }
327                         if (optptr[2] <= optlen) {
328                                 if (optptr[2]+3 > optlen) {
329                                         pp_ptr = optptr + 2;
330                                         goto error;
331                                 }
332                                 if (skb) {
333                                         memcpy(&optptr[optptr[2]-1], &rt->rt_spec_dst, 4);
334                                         opt->is_changed = 1;
335                                 }
336                                 optptr[2] += 4;
337                                 opt->rr_needaddr = 1;
338                         }
339                         opt->rr = optptr - iph;
340                         break;
341                       case IPOPT_TIMESTAMP:
342                         if (opt->ts) {
343                                 pp_ptr = optptr;
344                                 goto error;
345                         }
346                         if (optlen < 4) {
347                                 pp_ptr = optptr + 1;
348                                 goto error;
349                         }
350                         if (optptr[2] < 5) {
351                                 pp_ptr = optptr + 2;
352                                 goto error;
353                         }
354                         if (optptr[2] <= optlen) {
355                                 __u32 * timeptr = NULL;
356                                 if (optptr[2]+3 > optptr[1]) {
357                                         pp_ptr = optptr + 2;
358                                         goto error;
359                                 }
360                                 switch (optptr[3]&0xF) {
361                                       case IPOPT_TS_TSONLY:
362                                         opt->ts = optptr - iph;
363                                         if (skb) 
364                                                 timeptr = (__u32*)&optptr[optptr[2]-1];
365                                         opt->ts_needtime = 1;
366                                         optptr[2] += 4;
367                                         break;
368                                       case IPOPT_TS_TSANDADDR:
369                                         if (optptr[2]+7 > optptr[1]) {
370                                                 pp_ptr = optptr + 2;
371                                                 goto error;
372                                         }
373                                         opt->ts = optptr - iph;
374                                         if (skb) {
375                                                 memcpy(&optptr[optptr[2]-1], &rt->rt_spec_dst, 4);
376                                                 timeptr = (__u32*)&optptr[optptr[2]+3];
377                                         }
378                                         opt->ts_needaddr = 1;
379                                         opt->ts_needtime = 1;
380                                         optptr[2] += 8;
381                                         break;
382                                       case IPOPT_TS_PRESPEC:
383                                         if (optptr[2]+7 > optptr[1]) {
384                                                 pp_ptr = optptr + 2;
385                                                 goto error;
386                                         }
387                                         opt->ts = optptr - iph;
388                                         {
389                                                 u32 addr;
390                                                 memcpy(&addr, &optptr[optptr[2]-1], 4);
391                                                 if (inet_addr_type(addr) == RTN_UNICAST)
392                                                         break;
393                                                 if (skb)
394                                                         timeptr = (__u32*)&optptr[optptr[2]+3];
395                                         }
396                                         opt->ts_needtime = 1;
397                                         optptr[2] += 8;
398                                         break;
399                                       default:
400                                         if (!skb && !capable(CAP_NET_RAW)) {
401                                                 pp_ptr = optptr + 3;
402                                                 goto error;
403                                         }
404                                         break;
405                                 }
406                                 if (timeptr) {
407                                         struct timeval tv;
408                                         __u32  midtime;
409                                         do_gettimeofday(&tv);
410                                         midtime = htonl((tv.tv_sec % 86400) * 1000 + tv.tv_usec / 1000);
411                                         memcpy(timeptr, &midtime, sizeof(__u32));
412                                         opt->is_changed = 1;
413                                 }
414                         } else {
415                                 unsigned overflow = optptr[3]>>4;
416                                 if (overflow == 15) {
417                                         pp_ptr = optptr + 3;
418                                         goto error;
419                                 }
420                                 opt->ts = optptr - iph;
421                                 if (skb) {
422                                         optptr[3] = (optptr[3]&0xF)|((overflow+1)<<4);
423                                         opt->is_changed = 1;
424                                 }
425                         }
426                         break;
427                       case IPOPT_RA:
428                         if (optlen < 4) {
429                                 pp_ptr = optptr + 1;
430                                 goto error;
431                         }
432                         if (optptr[2] == 0 && optptr[3] == 0)
433                                 opt->router_alert = optptr - iph;
434                         break;
435                       case IPOPT_SEC:
436                       case IPOPT_SID:
437                       default:
438                         if (!skb && !capable(CAP_NET_RAW)) {
439                                 pp_ptr = optptr;
440                                 goto error;
441                         }
442                         break;
443                 }
444                 l -= optlen;
445                 optptr += optlen;
446         }
447
448 eol:
449         if (!pp_ptr)
450                 return 0;
451
452 error:
453         if (skb) {
454                 icmp_send(skb, ICMP_PARAMETERPROB, 0, htonl((pp_ptr-iph)<<24));
455         }
456         return -EINVAL;
457 }
458
459
460 /*
461  *      Undo all the changes done by ip_options_compile().
462  */
463
464 void ip_options_undo(struct ip_options * opt)
465 {
466         if (opt->srr) {
467                 unsigned  char * optptr = opt->__data+opt->srr-sizeof(struct  iphdr);
468                 memmove(optptr+7, optptr+3, optptr[1]-7);
469                 memcpy(optptr+3, &opt->faddr, 4);
470         }
471         if (opt->rr_needaddr) {
472                 unsigned  char * optptr = opt->__data+opt->rr-sizeof(struct  iphdr);
473                 optptr[2] -= 4;
474                 memset(&optptr[optptr[2]-1], 0, 4);
475         }
476         if (opt->ts) {
477                 unsigned  char * optptr = opt->__data+opt->ts-sizeof(struct  iphdr);
478                 if (opt->ts_needtime) {
479                         optptr[2] -= 4;
480                         memset(&optptr[optptr[2]-1], 0, 4);
481                         if ((optptr[3]&0xF) == IPOPT_TS_PRESPEC)
482                                 optptr[2] -= 4;
483                 }
484                 if (opt->ts_needaddr) {
485                         optptr[2] -= 4;
486                         memset(&optptr[optptr[2]-1], 0, 4);
487                 }
488         }
489 }
490
491 int ip_options_get(struct ip_options **optp, unsigned char *data, int optlen, int user)
492 {
493         struct ip_options *opt;
494
495         opt = kmalloc(sizeof(struct ip_options)+((optlen+3)&~3), GFP_KERNEL);
496         if (!opt)
497                 return -ENOMEM;
498         memset(opt, 0, sizeof(struct ip_options));
499         if (optlen) {
500                 if (user) {
501                         if (copy_from_user(opt->__data, data, optlen)) {
502                                 kfree(opt);
503                                 return -EFAULT;
504                         }
505                 } else
506                         memcpy(opt->__data, data, optlen);
507         }
508         while (optlen & 3)
509                 opt->__data[optlen++] = IPOPT_END;
510         opt->optlen = optlen;
511         opt->is_data = 1;
512         opt->is_setbyuser = 1;
513         if (optlen && ip_options_compile(opt, NULL)) {
514                 kfree(opt);
515                 return -EINVAL;
516         }
517         if (*optp)
518                 kfree(*optp);
519         *optp = opt;
520         return 0;
521 }
522
523 void ip_forward_options(struct sk_buff *skb)
524 {
525         struct   ip_options * opt       = &(IPCB(skb)->opt);
526         unsigned char * optptr;
527         struct rtable *rt = (struct rtable*)skb->dst;
528         unsigned char *raw = skb->nh.raw;
529
530         if (opt->rr_needaddr) {
531                 optptr = (unsigned char *)raw + opt->rr;
532                 ip_rt_get_source(&optptr[optptr[2]-5], rt);
533                 opt->is_changed = 1;
534         }
535         if (opt->srr_is_hit) {
536                 int srrptr, srrspace;
537
538                 optptr = raw + opt->srr;
539
540                 for ( srrptr=optptr[2], srrspace = optptr[1];
541                      srrptr <= srrspace;
542                      srrptr += 4
543                      ) {
544                         if (srrptr + 3 > srrspace)
545                                 break;
546                         if (memcmp(&rt->rt_dst, &optptr[srrptr-1], 4) == 0)
547                                 break;
548                 }
549                 if (srrptr + 3 <= srrspace) {
550                         opt->is_changed = 1;
551                         ip_rt_get_source(&optptr[srrptr-1], rt);
552                         skb->nh.iph->daddr = rt->rt_dst;
553                         optptr[2] = srrptr+4;
554                 } else if (net_ratelimit())
555                         printk(KERN_CRIT "ip_forward(): Argh! Destination lost!\n");
556                 if (opt->ts_needaddr) {
557                         optptr = raw + opt->ts;
558                         ip_rt_get_source(&optptr[optptr[2]-9], rt);
559                         opt->is_changed = 1;
560                 }
561         }
562         if (opt->is_changed) {
563                 opt->is_changed = 0;
564                 ip_send_check(skb->nh.iph);
565         }
566 }
567
568 int ip_options_rcv_srr(struct sk_buff *skb)
569 {
570         struct ip_options *opt = &(IPCB(skb)->opt);
571         int srrspace, srrptr;
572         u32 nexthop;
573         struct iphdr *iph = skb->nh.iph;
574         unsigned char * optptr = skb->nh.raw + opt->srr;
575         struct rtable *rt = (struct rtable*)skb->dst;
576         struct rtable *rt2;
577         int err;
578
579         if (!opt->srr)
580                 return 0;
581
582         if (skb->pkt_type != PACKET_HOST)
583                 return -EINVAL;
584         if (rt->rt_type == RTN_UNICAST) {
585                 if (!opt->is_strictroute)
586                         return 0;
587                 icmp_send(skb, ICMP_PARAMETERPROB, 0, htonl(16<<24));
588                 return -EINVAL;
589         }
590         if (rt->rt_type != RTN_LOCAL)
591                 return -EINVAL;
592
593         for (srrptr=optptr[2], srrspace = optptr[1]; srrptr <= srrspace; srrptr += 4) {
594                 if (srrptr + 3 > srrspace) {
595                         icmp_send(skb, ICMP_PARAMETERPROB, 0, htonl((opt->srr+2)<<24));
596                         return -EINVAL;
597                 }
598                 memcpy(&nexthop, &optptr[srrptr-1], 4);
599
600                 rt = (struct rtable*)skb->dst;
601                 skb->dst = NULL;
602                 err = ip_route_input(skb, nexthop, iph->saddr, iph->tos, skb->dev);
603                 rt2 = (struct rtable*)skb->dst;
604                 if (err || (rt2->rt_type != RTN_UNICAST && rt2->rt_type != RTN_LOCAL)) {
605                         ip_rt_put(rt2);
606                         skb->dst = &rt->u.dst;
607                         return -EINVAL;
608                 }
609                 ip_rt_put(rt);
610                 if (rt2->rt_type != RTN_LOCAL)
611                         break;
612                 /* Superfast 8) loopback forward */
613                 memcpy(&iph->daddr, &optptr[srrptr-1], 4);
614                 opt->is_changed = 1;
615         }
616         if (srrptr <= srrspace) {
617                 opt->srr_is_hit = 1;
618                 opt->is_changed = 1;
619         }
620         return 0;
621 }