Merge branch 'release' of git://git.kernel.org/pub/scm/linux/kernel/git/aegl/linux-2.6
[linux-2.6] / 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/module.h>
15 #include <linux/types.h>
16 #include <asm/uaccess.h>
17 #include <linux/skbuff.h>
18 #include <linux/ip.h>
19 #include <linux/icmp.h>
20 #include <linux/netdevice.h>
21 #include <linux/rtnetlink.h>
22 #include <net/sock.h>
23 #include <net/ip.h>
24 #include <net/icmp.h>
25
26 /* 
27  * Write options to IP header, record destination address to
28  * source route option, address of outgoing interface
29  * (we should already know it, so that this  function is allowed be
30  * called only after routing decision) and timestamp,
31  * if we originate this datagram.
32  *
33  * daddr is real destination address, next hop is recorded in IP header.
34  * saddr is address of outgoing interface.
35  */
36
37 void ip_options_build(struct sk_buff * skb, struct ip_options * opt,
38                             u32 daddr, struct rtable *rt, int is_frag) 
39 {
40         unsigned char * iph = skb->nh.raw;
41
42         memcpy(&(IPCB(skb)->opt), opt, sizeof(struct ip_options));
43         memcpy(iph+sizeof(struct iphdr), opt->__data, opt->optlen);
44         opt = &(IPCB(skb)->opt);
45         opt->is_data = 0;
46
47         if (opt->srr)
48                 memcpy(iph+opt->srr+iph[opt->srr+1]-4, &daddr, 4);
49
50         if (!is_frag) {
51                 if (opt->rr_needaddr)
52                         ip_rt_get_source(iph+opt->rr+iph[opt->rr+2]-5, rt);
53                 if (opt->ts_needaddr)
54                         ip_rt_get_source(iph+opt->ts+iph[opt->ts+2]-9, rt);
55                 if (opt->ts_needtime) {
56                         struct timeval tv;
57                         __u32 midtime;
58                         do_gettimeofday(&tv);
59                         midtime = htonl((tv.tv_sec % 86400) * 1000 + tv.tv_usec / 1000);
60                         memcpy(iph+opt->ts+iph[opt->ts+2]-5, &midtime, 4);
61                 }
62                 return;
63         }
64         if (opt->rr) {
65                 memset(iph+opt->rr, IPOPT_NOP, iph[opt->rr+1]);
66                 opt->rr = 0;
67                 opt->rr_needaddr = 0;
68         }
69         if (opt->ts) {
70                 memset(iph+opt->ts, IPOPT_NOP, iph[opt->ts+1]);
71                 opt->ts = 0;
72                 opt->ts_needaddr = opt->ts_needtime = 0;
73         }
74 }
75
76 /* 
77  * Provided (sopt, skb) points to received options,
78  * build in dopt compiled option set appropriate for answering.
79  * i.e. invert SRR option, copy anothers,
80  * and grab room in RR/TS options.
81  *
82  * NOTE: dopt cannot point to skb.
83  */
84
85 int ip_options_echo(struct ip_options * dopt, struct sk_buff * skb) 
86 {
87         struct ip_options *sopt;
88         unsigned char *sptr, *dptr;
89         int soffset, doffset;
90         int     optlen;
91         u32     daddr;
92
93         memset(dopt, 0, sizeof(struct ip_options));
94
95         dopt->is_data = 1;
96
97         sopt = &(IPCB(skb)->opt);
98
99         if (sopt->optlen == 0) {
100                 dopt->optlen = 0;
101                 return 0;
102         }
103
104         sptr = skb->nh.raw;
105         dptr = dopt->__data;
106
107         if (skb->dst)
108                 daddr = ((struct rtable*)skb->dst)->rt_spec_dst;
109         else
110                 daddr = skb->nh.iph->daddr;
111
112         if (sopt->rr) {
113                 optlen  = sptr[sopt->rr+1];
114                 soffset = sptr[sopt->rr+2];
115                 dopt->rr = dopt->optlen + sizeof(struct iphdr);
116                 memcpy(dptr, sptr+sopt->rr, optlen);
117                 if (sopt->rr_needaddr && soffset <= optlen) {
118                         if (soffset + 3 > optlen)
119                                 return -EINVAL;
120                         dptr[2] = soffset + 4;
121                         dopt->rr_needaddr = 1;
122                 }
123                 dptr += optlen;
124                 dopt->optlen += optlen;
125         }
126         if (sopt->ts) {
127                 optlen = sptr[sopt->ts+1];
128                 soffset = sptr[sopt->ts+2];
129                 dopt->ts = dopt->optlen + sizeof(struct iphdr);
130                 memcpy(dptr, sptr+sopt->ts, optlen);
131                 if (soffset <= optlen) {
132                         if (sopt->ts_needaddr) {
133                                 if (soffset + 3 > optlen)
134                                         return -EINVAL;
135                                 dopt->ts_needaddr = 1;
136                                 soffset += 4;
137                         }
138                         if (sopt->ts_needtime) {
139                                 if (soffset + 3 > optlen)
140                                         return -EINVAL;
141                                 if ((dptr[3]&0xF) != IPOPT_TS_PRESPEC) {
142                                         dopt->ts_needtime = 1;
143                                         soffset += 4;
144                                 } else {
145                                         dopt->ts_needtime = 0;
146
147                                         if (soffset + 8 <= optlen) {
148                                                 __u32 addr;
149
150                                                 memcpy(&addr, sptr+soffset-1, 4);
151                                                 if (inet_addr_type(addr) != RTN_LOCAL) {
152                                                         dopt->ts_needtime = 1;
153                                                         soffset += 8;
154                                                 }
155                                         }
156                                 }
157                         }
158                         dptr[2] = soffset;
159                 }
160                 dptr += optlen;
161                 dopt->optlen += optlen;
162         }
163         if (sopt->srr) {
164                 unsigned char * start = sptr+sopt->srr;
165                 u32 faddr;
166
167                 optlen  = start[1];
168                 soffset = start[2];
169                 doffset = 0;
170                 if (soffset > optlen)
171                         soffset = optlen + 1;
172                 soffset -= 4;
173                 if (soffset > 3) {
174                         memcpy(&faddr, &start[soffset-1], 4);
175                         for (soffset-=4, doffset=4; soffset > 3; soffset-=4, doffset+=4)
176                                 memcpy(&dptr[doffset-1], &start[soffset-1], 4);
177                         /*
178                          * RFC1812 requires to fix illegal source routes.
179                          */
180                         if (memcmp(&skb->nh.iph->saddr, &start[soffset+3], 4) == 0)
181                                 doffset -= 4;
182                 }
183                 if (doffset > 3) {
184                         memcpy(&start[doffset-1], &daddr, 4);
185                         dopt->faddr = faddr;
186                         dptr[0] = start[0];
187                         dptr[1] = doffset+3;
188                         dptr[2] = 4;
189                         dptr += doffset+3;
190                         dopt->srr = dopt->optlen + sizeof(struct iphdr);
191                         dopt->optlen += doffset+3;
192                         dopt->is_strictroute = sopt->is_strictroute;
193                 }
194         }
195         while (dopt->optlen & 3) {
196                 *dptr++ = IPOPT_END;
197                 dopt->optlen++;
198         }
199         return 0;
200 }
201
202 /*
203  *      Options "fragmenting", just fill options not
204  *      allowed in fragments with NOOPs.
205  *      Simple and stupid 8), but the most efficient way.
206  */
207
208 void ip_options_fragment(struct sk_buff * skb) 
209 {
210         unsigned char * optptr = skb->nh.raw;
211         struct ip_options * opt = &(IPCB(skb)->opt);
212         int  l = opt->optlen;
213         int  optlen;
214
215         while (l > 0) {
216                 switch (*optptr) {
217                 case IPOPT_END:
218                         return;
219                 case IPOPT_NOOP:
220                         l--;
221                         optptr++;
222                         continue;
223                 }
224                 optlen = optptr[1];
225                 if (optlen<2 || optlen>l)
226                   return;
227                 if (!IPOPT_COPIED(*optptr))
228                         memset(optptr, IPOPT_NOOP, optlen);
229                 l -= optlen;
230                 optptr += optlen;
231         }
232         opt->ts = 0;
233         opt->rr = 0;
234         opt->rr_needaddr = 0;
235         opt->ts_needaddr = 0;
236         opt->ts_needtime = 0;
237         return;
238 }
239
240 /*
241  * Verify options and fill pointers in struct options.
242  * Caller should clear *opt, and set opt->data.
243  * If opt == NULL, then skb->data should point to IP header.
244  */
245
246 int ip_options_compile(struct ip_options * opt, struct sk_buff * skb)
247 {
248         int l;
249         unsigned char * iph;
250         unsigned char * optptr;
251         int optlen;
252         unsigned char * pp_ptr = NULL;
253         struct rtable *rt = skb ? (struct rtable*)skb->dst : NULL;
254
255         if (!opt) {
256                 opt = &(IPCB(skb)->opt);
257                 memset(opt, 0, sizeof(struct ip_options));
258                 iph = skb->nh.raw;
259                 opt->optlen = ((struct iphdr *)iph)->ihl*4 - sizeof(struct iphdr);
260                 optptr = iph + sizeof(struct iphdr);
261                 opt->is_data = 0;
262         } else {
263                 optptr = opt->is_data ? opt->__data : (unsigned char*)&(skb->nh.iph[1]);
264                 iph = optptr - sizeof(struct iphdr);
265         }
266
267         for (l = opt->optlen; l > 0; ) {
268                 switch (*optptr) {
269                       case IPOPT_END:
270                         for (optptr++, l--; l>0; optptr++, l--) {
271                                 if (*optptr != IPOPT_END) {
272                                         *optptr = IPOPT_END;
273                                         opt->is_changed = 1;
274                                 }
275                         }
276                         goto eol;
277                       case IPOPT_NOOP:
278                         l--;
279                         optptr++;
280                         continue;
281                 }
282                 optlen = optptr[1];
283                 if (optlen<2 || optlen>l) {
284                         pp_ptr = optptr;
285                         goto error;
286                 }
287                 switch (*optptr) {
288                       case IPOPT_SSRR:
289                       case IPOPT_LSRR:
290                         if (optlen < 3) {
291                                 pp_ptr = optptr + 1;
292                                 goto error;
293                         }
294                         if (optptr[2] < 4) {
295                                 pp_ptr = optptr + 2;
296                                 goto error;
297                         }
298                         /* NB: cf RFC-1812 5.2.4.1 */
299                         if (opt->srr) {
300                                 pp_ptr = optptr;
301                                 goto error;
302                         }
303                         if (!skb) {
304                                 if (optptr[2] != 4 || optlen < 7 || ((optlen-3) & 3)) {
305                                         pp_ptr = optptr + 1;
306                                         goto error;
307                                 }
308                                 memcpy(&opt->faddr, &optptr[3], 4);
309                                 if (optlen > 7)
310                                         memmove(&optptr[3], &optptr[7], optlen-7);
311                         }
312                         opt->is_strictroute = (optptr[0] == IPOPT_SSRR);
313                         opt->srr = optptr - iph;
314                         break;
315                       case IPOPT_RR:
316                         if (opt->rr) {
317                                 pp_ptr = optptr;
318                                 goto error;
319                         }
320                         if (optlen < 3) {
321                                 pp_ptr = optptr + 1;
322                                 goto error;
323                         }
324                         if (optptr[2] < 4) {
325                                 pp_ptr = optptr + 2;
326                                 goto error;
327                         }
328                         if (optptr[2] <= optlen) {
329                                 if (optptr[2]+3 > optlen) {
330                                         pp_ptr = optptr + 2;
331                                         goto error;
332                                 }
333                                 if (skb) {
334                                         memcpy(&optptr[optptr[2]-1], &rt->rt_spec_dst, 4);
335                                         opt->is_changed = 1;
336                                 }
337                                 optptr[2] += 4;
338                                 opt->rr_needaddr = 1;
339                         }
340                         opt->rr = optptr - iph;
341                         break;
342                       case IPOPT_TIMESTAMP:
343                         if (opt->ts) {
344                                 pp_ptr = optptr;
345                                 goto error;
346                         }
347                         if (optlen < 4) {
348                                 pp_ptr = optptr + 1;
349                                 goto error;
350                         }
351                         if (optptr[2] < 5) {
352                                 pp_ptr = optptr + 2;
353                                 goto error;
354                         }
355                         if (optptr[2] <= optlen) {
356                                 __u32 * timeptr = NULL;
357                                 if (optptr[2]+3 > optptr[1]) {
358                                         pp_ptr = optptr + 2;
359                                         goto error;
360                                 }
361                                 switch (optptr[3]&0xF) {
362                                       case IPOPT_TS_TSONLY:
363                                         opt->ts = optptr - iph;
364                                         if (skb) 
365                                                 timeptr = (__u32*)&optptr[optptr[2]-1];
366                                         opt->ts_needtime = 1;
367                                         optptr[2] += 4;
368                                         break;
369                                       case IPOPT_TS_TSANDADDR:
370                                         if (optptr[2]+7 > optptr[1]) {
371                                                 pp_ptr = optptr + 2;
372                                                 goto error;
373                                         }
374                                         opt->ts = optptr - iph;
375                                         if (skb) {
376                                                 memcpy(&optptr[optptr[2]-1], &rt->rt_spec_dst, 4);
377                                                 timeptr = (__u32*)&optptr[optptr[2]+3];
378                                         }
379                                         opt->ts_needaddr = 1;
380                                         opt->ts_needtime = 1;
381                                         optptr[2] += 8;
382                                         break;
383                                       case IPOPT_TS_PRESPEC:
384                                         if (optptr[2]+7 > optptr[1]) {
385                                                 pp_ptr = optptr + 2;
386                                                 goto error;
387                                         }
388                                         opt->ts = optptr - iph;
389                                         {
390                                                 u32 addr;
391                                                 memcpy(&addr, &optptr[optptr[2]-1], 4);
392                                                 if (inet_addr_type(addr) == RTN_UNICAST)
393                                                         break;
394                                                 if (skb)
395                                                         timeptr = (__u32*)&optptr[optptr[2]+3];
396                                         }
397                                         opt->ts_needtime = 1;
398                                         optptr[2] += 8;
399                                         break;
400                                       default:
401                                         if (!skb && !capable(CAP_NET_RAW)) {
402                                                 pp_ptr = optptr + 3;
403                                                 goto error;
404                                         }
405                                         break;
406                                 }
407                                 if (timeptr) {
408                                         struct timeval tv;
409                                         __u32  midtime;
410                                         do_gettimeofday(&tv);
411                                         midtime = htonl((tv.tv_sec % 86400) * 1000 + tv.tv_usec / 1000);
412                                         memcpy(timeptr, &midtime, sizeof(__u32));
413                                         opt->is_changed = 1;
414                                 }
415                         } else {
416                                 unsigned overflow = optptr[3]>>4;
417                                 if (overflow == 15) {
418                                         pp_ptr = optptr + 3;
419                                         goto error;
420                                 }
421                                 opt->ts = optptr - iph;
422                                 if (skb) {
423                                         optptr[3] = (optptr[3]&0xF)|((overflow+1)<<4);
424                                         opt->is_changed = 1;
425                                 }
426                         }
427                         break;
428                       case IPOPT_RA:
429                         if (optlen < 4) {
430                                 pp_ptr = optptr + 1;
431                                 goto error;
432                         }
433                         if (optptr[2] == 0 && optptr[3] == 0)
434                                 opt->router_alert = optptr - iph;
435                         break;
436                       case IPOPT_SEC:
437                       case IPOPT_SID:
438                       default:
439                         if (!skb && !capable(CAP_NET_RAW)) {
440                                 pp_ptr = optptr;
441                                 goto error;
442                         }
443                         break;
444                 }
445                 l -= optlen;
446                 optptr += optlen;
447         }
448
449 eol:
450         if (!pp_ptr)
451                 return 0;
452
453 error:
454         if (skb) {
455                 icmp_send(skb, ICMP_PARAMETERPROB, 0, htonl((pp_ptr-iph)<<24));
456         }
457         return -EINVAL;
458 }
459
460
461 /*
462  *      Undo all the changes done by ip_options_compile().
463  */
464
465 void ip_options_undo(struct ip_options * opt)
466 {
467         if (opt->srr) {
468                 unsigned  char * optptr = opt->__data+opt->srr-sizeof(struct  iphdr);
469                 memmove(optptr+7, optptr+3, optptr[1]-7);
470                 memcpy(optptr+3, &opt->faddr, 4);
471         }
472         if (opt->rr_needaddr) {
473                 unsigned  char * optptr = opt->__data+opt->rr-sizeof(struct  iphdr);
474                 optptr[2] -= 4;
475                 memset(&optptr[optptr[2]-1], 0, 4);
476         }
477         if (opt->ts) {
478                 unsigned  char * optptr = opt->__data+opt->ts-sizeof(struct  iphdr);
479                 if (opt->ts_needtime) {
480                         optptr[2] -= 4;
481                         memset(&optptr[optptr[2]-1], 0, 4);
482                         if ((optptr[3]&0xF) == IPOPT_TS_PRESPEC)
483                                 optptr[2] -= 4;
484                 }
485                 if (opt->ts_needaddr) {
486                         optptr[2] -= 4;
487                         memset(&optptr[optptr[2]-1], 0, 4);
488                 }
489         }
490 }
491
492 static struct ip_options *ip_options_get_alloc(const int optlen)
493 {
494         struct ip_options *opt = kmalloc(sizeof(*opt) + ((optlen + 3) & ~3),
495                                          GFP_KERNEL);
496         if (opt)
497                 memset(opt, 0, sizeof(*opt));
498         return opt;
499 }
500
501 static int ip_options_get_finish(struct ip_options **optp,
502                                  struct ip_options *opt, int optlen)
503 {
504         while (optlen & 3)
505                 opt->__data[optlen++] = IPOPT_END;
506         opt->optlen = optlen;
507         opt->is_data = 1;
508         opt->is_setbyuser = 1;
509         if (optlen && ip_options_compile(opt, NULL)) {
510                 kfree(opt);
511                 return -EINVAL;
512         }
513         kfree(*optp);
514         *optp = opt;
515         return 0;
516 }
517
518 int ip_options_get_from_user(struct ip_options **optp, unsigned char __user *data, int optlen)
519 {
520         struct ip_options *opt = ip_options_get_alloc(optlen);
521
522         if (!opt)
523                 return -ENOMEM;
524         if (optlen && copy_from_user(opt->__data, data, optlen)) {
525                 kfree(opt);
526                 return -EFAULT;
527         }
528         return ip_options_get_finish(optp, opt, optlen);
529 }
530
531 int ip_options_get(struct ip_options **optp, unsigned char *data, int optlen)
532 {
533         struct ip_options *opt = ip_options_get_alloc(optlen);
534
535         if (!opt)
536                 return -ENOMEM;
537         if (optlen)
538                 memcpy(opt->__data, data, optlen);
539         return ip_options_get_finish(optp, opt, optlen);
540 }
541
542 void ip_forward_options(struct sk_buff *skb)
543 {
544         struct   ip_options * opt       = &(IPCB(skb)->opt);
545         unsigned char * optptr;
546         struct rtable *rt = (struct rtable*)skb->dst;
547         unsigned char *raw = skb->nh.raw;
548
549         if (opt->rr_needaddr) {
550                 optptr = (unsigned char *)raw + opt->rr;
551                 ip_rt_get_source(&optptr[optptr[2]-5], rt);
552                 opt->is_changed = 1;
553         }
554         if (opt->srr_is_hit) {
555                 int srrptr, srrspace;
556
557                 optptr = raw + opt->srr;
558
559                 for ( srrptr=optptr[2], srrspace = optptr[1];
560                      srrptr <= srrspace;
561                      srrptr += 4
562                      ) {
563                         if (srrptr + 3 > srrspace)
564                                 break;
565                         if (memcmp(&rt->rt_dst, &optptr[srrptr-1], 4) == 0)
566                                 break;
567                 }
568                 if (srrptr + 3 <= srrspace) {
569                         opt->is_changed = 1;
570                         ip_rt_get_source(&optptr[srrptr-1], rt);
571                         skb->nh.iph->daddr = rt->rt_dst;
572                         optptr[2] = srrptr+4;
573                 } else if (net_ratelimit())
574                         printk(KERN_CRIT "ip_forward(): Argh! Destination lost!\n");
575                 if (opt->ts_needaddr) {
576                         optptr = raw + opt->ts;
577                         ip_rt_get_source(&optptr[optptr[2]-9], rt);
578                         opt->is_changed = 1;
579                 }
580         }
581         if (opt->is_changed) {
582                 opt->is_changed = 0;
583                 ip_send_check(skb->nh.iph);
584         }
585 }
586
587 int ip_options_rcv_srr(struct sk_buff *skb)
588 {
589         struct ip_options *opt = &(IPCB(skb)->opt);
590         int srrspace, srrptr;
591         u32 nexthop;
592         struct iphdr *iph = skb->nh.iph;
593         unsigned char * optptr = skb->nh.raw + opt->srr;
594         struct rtable *rt = (struct rtable*)skb->dst;
595         struct rtable *rt2;
596         int err;
597
598         if (!opt->srr)
599                 return 0;
600
601         if (skb->pkt_type != PACKET_HOST)
602                 return -EINVAL;
603         if (rt->rt_type == RTN_UNICAST) {
604                 if (!opt->is_strictroute)
605                         return 0;
606                 icmp_send(skb, ICMP_PARAMETERPROB, 0, htonl(16<<24));
607                 return -EINVAL;
608         }
609         if (rt->rt_type != RTN_LOCAL)
610                 return -EINVAL;
611
612         for (srrptr=optptr[2], srrspace = optptr[1]; srrptr <= srrspace; srrptr += 4) {
613                 if (srrptr + 3 > srrspace) {
614                         icmp_send(skb, ICMP_PARAMETERPROB, 0, htonl((opt->srr+2)<<24));
615                         return -EINVAL;
616                 }
617                 memcpy(&nexthop, &optptr[srrptr-1], 4);
618
619                 rt = (struct rtable*)skb->dst;
620                 skb->dst = NULL;
621                 err = ip_route_input(skb, nexthop, iph->saddr, iph->tos, skb->dev);
622                 rt2 = (struct rtable*)skb->dst;
623                 if (err || (rt2->rt_type != RTN_UNICAST && rt2->rt_type != RTN_LOCAL)) {
624                         ip_rt_put(rt2);
625                         skb->dst = &rt->u.dst;
626                         return -EINVAL;
627                 }
628                 ip_rt_put(rt);
629                 if (rt2->rt_type != RTN_LOCAL)
630                         break;
631                 /* Superfast 8) loopback forward */
632                 memcpy(&iph->daddr, &optptr[srrptr-1], 4);
633                 opt->is_changed = 1;
634         }
635         if (srrptr <= srrspace) {
636                 opt->srr_is_hit = 1;
637                 opt->is_changed = 1;
638         }
639         return 0;
640 }