Merge http://oss.oracle.com/git/ocfs2
[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 #include <net/route.h>
26
27 /* 
28  * Write options to IP header, record destination address to
29  * source route option, address of outgoing interface
30  * (we should already know it, so that this  function is allowed be
31  * called only after routing decision) and timestamp,
32  * if we originate this datagram.
33  *
34  * daddr is real destination address, next hop is recorded in IP header.
35  * saddr is address of outgoing interface.
36  */
37
38 void ip_options_build(struct sk_buff * skb, struct ip_options * opt,
39                             u32 daddr, struct rtable *rt, int is_frag) 
40 {
41         unsigned char * iph = skb->nh.raw;
42
43         memcpy(&(IPCB(skb)->opt), opt, sizeof(struct ip_options));
44         memcpy(iph+sizeof(struct iphdr), opt->__data, opt->optlen);
45         opt = &(IPCB(skb)->opt);
46         opt->is_data = 0;
47
48         if (opt->srr)
49                 memcpy(iph+opt->srr+iph[opt->srr+1]-4, &daddr, 4);
50
51         if (!is_frag) {
52                 if (opt->rr_needaddr)
53                         ip_rt_get_source(iph+opt->rr+iph[opt->rr+2]-5, rt);
54                 if (opt->ts_needaddr)
55                         ip_rt_get_source(iph+opt->ts+iph[opt->ts+2]-9, rt);
56                 if (opt->ts_needtime) {
57                         struct timeval tv;
58                         __u32 midtime;
59                         do_gettimeofday(&tv);
60                         midtime = htonl((tv.tv_sec % 86400) * 1000 + tv.tv_usec / 1000);
61                         memcpy(iph+opt->ts+iph[opt->ts+2]-5, &midtime, 4);
62                 }
63                 return;
64         }
65         if (opt->rr) {
66                 memset(iph+opt->rr, IPOPT_NOP, iph[opt->rr+1]);
67                 opt->rr = 0;
68                 opt->rr_needaddr = 0;
69         }
70         if (opt->ts) {
71                 memset(iph+opt->ts, IPOPT_NOP, iph[opt->ts+1]);
72                 opt->ts = 0;
73                 opt->ts_needaddr = opt->ts_needtime = 0;
74         }
75 }
76
77 /* 
78  * Provided (sopt, skb) points to received options,
79  * build in dopt compiled option set appropriate for answering.
80  * i.e. invert SRR option, copy anothers,
81  * and grab room in RR/TS options.
82  *
83  * NOTE: dopt cannot point to skb.
84  */
85
86 int ip_options_echo(struct ip_options * dopt, struct sk_buff * skb) 
87 {
88         struct ip_options *sopt;
89         unsigned char *sptr, *dptr;
90         int soffset, doffset;
91         int     optlen;
92         u32     daddr;
93
94         memset(dopt, 0, sizeof(struct ip_options));
95
96         dopt->is_data = 1;
97
98         sopt = &(IPCB(skb)->opt);
99
100         if (sopt->optlen == 0) {
101                 dopt->optlen = 0;
102                 return 0;
103         }
104
105         sptr = skb->nh.raw;
106         dptr = dopt->__data;
107
108         if (skb->dst)
109                 daddr = ((struct rtable*)skb->dst)->rt_spec_dst;
110         else
111                 daddr = skb->nh.iph->daddr;
112
113         if (sopt->rr) {
114                 optlen  = sptr[sopt->rr+1];
115                 soffset = sptr[sopt->rr+2];
116                 dopt->rr = dopt->optlen + sizeof(struct iphdr);
117                 memcpy(dptr, sptr+sopt->rr, optlen);
118                 if (sopt->rr_needaddr && soffset <= optlen) {
119                         if (soffset + 3 > optlen)
120                                 return -EINVAL;
121                         dptr[2] = soffset + 4;
122                         dopt->rr_needaddr = 1;
123                 }
124                 dptr += optlen;
125                 dopt->optlen += optlen;
126         }
127         if (sopt->ts) {
128                 optlen = sptr[sopt->ts+1];
129                 soffset = sptr[sopt->ts+2];
130                 dopt->ts = dopt->optlen + sizeof(struct iphdr);
131                 memcpy(dptr, sptr+sopt->ts, optlen);
132                 if (soffset <= optlen) {
133                         if (sopt->ts_needaddr) {
134                                 if (soffset + 3 > optlen)
135                                         return -EINVAL;
136                                 dopt->ts_needaddr = 1;
137                                 soffset += 4;
138                         }
139                         if (sopt->ts_needtime) {
140                                 if (soffset + 3 > optlen)
141                                         return -EINVAL;
142                                 if ((dptr[3]&0xF) != IPOPT_TS_PRESPEC) {
143                                         dopt->ts_needtime = 1;
144                                         soffset += 4;
145                                 } else {
146                                         dopt->ts_needtime = 0;
147
148                                         if (soffset + 8 <= optlen) {
149                                                 __u32 addr;
150
151                                                 memcpy(&addr, sptr+soffset-1, 4);
152                                                 if (inet_addr_type(addr) != RTN_LOCAL) {
153                                                         dopt->ts_needtime = 1;
154                                                         soffset += 8;
155                                                 }
156                                         }
157                                 }
158                         }
159                         dptr[2] = soffset;
160                 }
161                 dptr += optlen;
162                 dopt->optlen += optlen;
163         }
164         if (sopt->srr) {
165                 unsigned char * start = sptr+sopt->srr;
166                 u32 faddr;
167
168                 optlen  = start[1];
169                 soffset = start[2];
170                 doffset = 0;
171                 if (soffset > optlen)
172                         soffset = optlen + 1;
173                 soffset -= 4;
174                 if (soffset > 3) {
175                         memcpy(&faddr, &start[soffset-1], 4);
176                         for (soffset-=4, doffset=4; soffset > 3; soffset-=4, doffset+=4)
177                                 memcpy(&dptr[doffset-1], &start[soffset-1], 4);
178                         /*
179                          * RFC1812 requires to fix illegal source routes.
180                          */
181                         if (memcmp(&skb->nh.iph->saddr, &start[soffset+3], 4) == 0)
182                                 doffset -= 4;
183                 }
184                 if (doffset > 3) {
185                         memcpy(&start[doffset-1], &daddr, 4);
186                         dopt->faddr = faddr;
187                         dptr[0] = start[0];
188                         dptr[1] = doffset+3;
189                         dptr[2] = 4;
190                         dptr += doffset+3;
191                         dopt->srr = dopt->optlen + sizeof(struct iphdr);
192                         dopt->optlen += doffset+3;
193                         dopt->is_strictroute = sopt->is_strictroute;
194                 }
195         }
196         while (dopt->optlen & 3) {
197                 *dptr++ = IPOPT_END;
198                 dopt->optlen++;
199         }
200         return 0;
201 }
202
203 /*
204  *      Options "fragmenting", just fill options not
205  *      allowed in fragments with NOOPs.
206  *      Simple and stupid 8), but the most efficient way.
207  */
208
209 void ip_options_fragment(struct sk_buff * skb) 
210 {
211         unsigned char * optptr = skb->nh.raw;
212         struct ip_options * opt = &(IPCB(skb)->opt);
213         int  l = opt->optlen;
214         int  optlen;
215
216         while (l > 0) {
217                 switch (*optptr) {
218                 case IPOPT_END:
219                         return;
220                 case IPOPT_NOOP:
221                         l--;
222                         optptr++;
223                         continue;
224                 }
225                 optlen = optptr[1];
226                 if (optlen<2 || optlen>l)
227                   return;
228                 if (!IPOPT_COPIED(*optptr))
229                         memset(optptr, IPOPT_NOOP, optlen);
230                 l -= optlen;
231                 optptr += optlen;
232         }
233         opt->ts = 0;
234         opt->rr = 0;
235         opt->rr_needaddr = 0;
236         opt->ts_needaddr = 0;
237         opt->ts_needtime = 0;
238         return;
239 }
240
241 /*
242  * Verify options and fill pointers in struct options.
243  * Caller should clear *opt, and set opt->data.
244  * If opt == NULL, then skb->data should point to IP header.
245  */
246
247 int ip_options_compile(struct ip_options * opt, struct sk_buff * skb)
248 {
249         int l;
250         unsigned char * iph;
251         unsigned char * optptr;
252         int optlen;
253         unsigned char * pp_ptr = NULL;
254         struct rtable *rt = skb ? (struct rtable*)skb->dst : NULL;
255
256         if (!opt) {
257                 opt = &(IPCB(skb)->opt);
258                 memset(opt, 0, sizeof(struct ip_options));
259                 iph = skb->nh.raw;
260                 opt->optlen = ((struct iphdr *)iph)->ihl*4 - sizeof(struct iphdr);
261                 optptr = iph + sizeof(struct iphdr);
262                 opt->is_data = 0;
263         } else {
264                 optptr = opt->is_data ? opt->__data : (unsigned char*)&(skb->nh.iph[1]);
265                 iph = optptr - sizeof(struct iphdr);
266         }
267
268         for (l = opt->optlen; l > 0; ) {
269                 switch (*optptr) {
270                       case IPOPT_END:
271                         for (optptr++, l--; l>0; optptr++, l--) {
272                                 if (*optptr != IPOPT_END) {
273                                         *optptr = IPOPT_END;
274                                         opt->is_changed = 1;
275                                 }
276                         }
277                         goto eol;
278                       case IPOPT_NOOP:
279                         l--;
280                         optptr++;
281                         continue;
282                 }
283                 optlen = optptr[1];
284                 if (optlen<2 || optlen>l) {
285                         pp_ptr = optptr;
286                         goto error;
287                 }
288                 switch (*optptr) {
289                       case IPOPT_SSRR:
290                       case IPOPT_LSRR:
291                         if (optlen < 3) {
292                                 pp_ptr = optptr + 1;
293                                 goto error;
294                         }
295                         if (optptr[2] < 4) {
296                                 pp_ptr = optptr + 2;
297                                 goto error;
298                         }
299                         /* NB: cf RFC-1812 5.2.4.1 */
300                         if (opt->srr) {
301                                 pp_ptr = optptr;
302                                 goto error;
303                         }
304                         if (!skb) {
305                                 if (optptr[2] != 4 || optlen < 7 || ((optlen-3) & 3)) {
306                                         pp_ptr = optptr + 1;
307                                         goto error;
308                                 }
309                                 memcpy(&opt->faddr, &optptr[3], 4);
310                                 if (optlen > 7)
311                                         memmove(&optptr[3], &optptr[7], optlen-7);
312                         }
313                         opt->is_strictroute = (optptr[0] == IPOPT_SSRR);
314                         opt->srr = optptr - iph;
315                         break;
316                       case IPOPT_RR:
317                         if (opt->rr) {
318                                 pp_ptr = optptr;
319                                 goto error;
320                         }
321                         if (optlen < 3) {
322                                 pp_ptr = optptr + 1;
323                                 goto error;
324                         }
325                         if (optptr[2] < 4) {
326                                 pp_ptr = optptr + 2;
327                                 goto error;
328                         }
329                         if (optptr[2] <= optlen) {
330                                 if (optptr[2]+3 > optlen) {
331                                         pp_ptr = optptr + 2;
332                                         goto error;
333                                 }
334                                 if (skb) {
335                                         memcpy(&optptr[optptr[2]-1], &rt->rt_spec_dst, 4);
336                                         opt->is_changed = 1;
337                                 }
338                                 optptr[2] += 4;
339                                 opt->rr_needaddr = 1;
340                         }
341                         opt->rr = optptr - iph;
342                         break;
343                       case IPOPT_TIMESTAMP:
344                         if (opt->ts) {
345                                 pp_ptr = optptr;
346                                 goto error;
347                         }
348                         if (optlen < 4) {
349                                 pp_ptr = optptr + 1;
350                                 goto error;
351                         }
352                         if (optptr[2] < 5) {
353                                 pp_ptr = optptr + 2;
354                                 goto error;
355                         }
356                         if (optptr[2] <= optlen) {
357                                 __u32 * timeptr = NULL;
358                                 if (optptr[2]+3 > optptr[1]) {
359                                         pp_ptr = optptr + 2;
360                                         goto error;
361                                 }
362                                 switch (optptr[3]&0xF) {
363                                       case IPOPT_TS_TSONLY:
364                                         opt->ts = optptr - iph;
365                                         if (skb) 
366                                                 timeptr = (__u32*)&optptr[optptr[2]-1];
367                                         opt->ts_needtime = 1;
368                                         optptr[2] += 4;
369                                         break;
370                                       case IPOPT_TS_TSANDADDR:
371                                         if (optptr[2]+7 > optptr[1]) {
372                                                 pp_ptr = optptr + 2;
373                                                 goto error;
374                                         }
375                                         opt->ts = optptr - iph;
376                                         if (skb) {
377                                                 memcpy(&optptr[optptr[2]-1], &rt->rt_spec_dst, 4);
378                                                 timeptr = (__u32*)&optptr[optptr[2]+3];
379                                         }
380                                         opt->ts_needaddr = 1;
381                                         opt->ts_needtime = 1;
382                                         optptr[2] += 8;
383                                         break;
384                                       case IPOPT_TS_PRESPEC:
385                                         if (optptr[2]+7 > optptr[1]) {
386                                                 pp_ptr = optptr + 2;
387                                                 goto error;
388                                         }
389                                         opt->ts = optptr - iph;
390                                         {
391                                                 u32 addr;
392                                                 memcpy(&addr, &optptr[optptr[2]-1], 4);
393                                                 if (inet_addr_type(addr) == RTN_UNICAST)
394                                                         break;
395                                                 if (skb)
396                                                         timeptr = (__u32*)&optptr[optptr[2]+3];
397                                         }
398                                         opt->ts_needtime = 1;
399                                         optptr[2] += 8;
400                                         break;
401                                       default:
402                                         if (!skb && !capable(CAP_NET_RAW)) {
403                                                 pp_ptr = optptr + 3;
404                                                 goto error;
405                                         }
406                                         break;
407                                 }
408                                 if (timeptr) {
409                                         struct timeval tv;
410                                         __u32  midtime;
411                                         do_gettimeofday(&tv);
412                                         midtime = htonl((tv.tv_sec % 86400) * 1000 + tv.tv_usec / 1000);
413                                         memcpy(timeptr, &midtime, sizeof(__u32));
414                                         opt->is_changed = 1;
415                                 }
416                         } else {
417                                 unsigned overflow = optptr[3]>>4;
418                                 if (overflow == 15) {
419                                         pp_ptr = optptr + 3;
420                                         goto error;
421                                 }
422                                 opt->ts = optptr - iph;
423                                 if (skb) {
424                                         optptr[3] = (optptr[3]&0xF)|((overflow+1)<<4);
425                                         opt->is_changed = 1;
426                                 }
427                         }
428                         break;
429                       case IPOPT_RA:
430                         if (optlen < 4) {
431                                 pp_ptr = optptr + 1;
432                                 goto error;
433                         }
434                         if (optptr[2] == 0 && optptr[3] == 0)
435                                 opt->router_alert = optptr - iph;
436                         break;
437                       case IPOPT_SEC:
438                       case IPOPT_SID:
439                       default:
440                         if (!skb && !capable(CAP_NET_RAW)) {
441                                 pp_ptr = optptr;
442                                 goto error;
443                         }
444                         break;
445                 }
446                 l -= optlen;
447                 optptr += optlen;
448         }
449
450 eol:
451         if (!pp_ptr)
452                 return 0;
453
454 error:
455         if (skb) {
456                 icmp_send(skb, ICMP_PARAMETERPROB, 0, htonl((pp_ptr-iph)<<24));
457         }
458         return -EINVAL;
459 }
460
461
462 /*
463  *      Undo all the changes done by ip_options_compile().
464  */
465
466 void ip_options_undo(struct ip_options * opt)
467 {
468         if (opt->srr) {
469                 unsigned  char * optptr = opt->__data+opt->srr-sizeof(struct  iphdr);
470                 memmove(optptr+7, optptr+3, optptr[1]-7);
471                 memcpy(optptr+3, &opt->faddr, 4);
472         }
473         if (opt->rr_needaddr) {
474                 unsigned  char * optptr = opt->__data+opt->rr-sizeof(struct  iphdr);
475                 optptr[2] -= 4;
476                 memset(&optptr[optptr[2]-1], 0, 4);
477         }
478         if (opt->ts) {
479                 unsigned  char * optptr = opt->__data+opt->ts-sizeof(struct  iphdr);
480                 if (opt->ts_needtime) {
481                         optptr[2] -= 4;
482                         memset(&optptr[optptr[2]-1], 0, 4);
483                         if ((optptr[3]&0xF) == IPOPT_TS_PRESPEC)
484                                 optptr[2] -= 4;
485                 }
486                 if (opt->ts_needaddr) {
487                         optptr[2] -= 4;
488                         memset(&optptr[optptr[2]-1], 0, 4);
489                 }
490         }
491 }
492
493 static struct ip_options *ip_options_get_alloc(const int optlen)
494 {
495         struct ip_options *opt = kmalloc(sizeof(*opt) + ((optlen + 3) & ~3),
496                                          GFP_KERNEL);
497         if (opt)
498                 memset(opt, 0, sizeof(*opt));
499         return opt;
500 }
501
502 static int ip_options_get_finish(struct ip_options **optp,
503                                  struct ip_options *opt, int optlen)
504 {
505         while (optlen & 3)
506                 opt->__data[optlen++] = IPOPT_END;
507         opt->optlen = optlen;
508         opt->is_data = 1;
509         opt->is_setbyuser = 1;
510         if (optlen && ip_options_compile(opt, NULL)) {
511                 kfree(opt);
512                 return -EINVAL;
513         }
514         kfree(*optp);
515         *optp = opt;
516         return 0;
517 }
518
519 int ip_options_get_from_user(struct ip_options **optp, unsigned char __user *data, int optlen)
520 {
521         struct ip_options *opt = ip_options_get_alloc(optlen);
522
523         if (!opt)
524                 return -ENOMEM;
525         if (optlen && copy_from_user(opt->__data, data, optlen)) {
526                 kfree(opt);
527                 return -EFAULT;
528         }
529         return ip_options_get_finish(optp, opt, optlen);
530 }
531
532 int ip_options_get(struct ip_options **optp, unsigned char *data, int optlen)
533 {
534         struct ip_options *opt = ip_options_get_alloc(optlen);
535
536         if (!opt)
537                 return -ENOMEM;
538         if (optlen)
539                 memcpy(opt->__data, data, optlen);
540         return ip_options_get_finish(optp, opt, optlen);
541 }
542
543 void ip_forward_options(struct sk_buff *skb)
544 {
545         struct   ip_options * opt       = &(IPCB(skb)->opt);
546         unsigned char * optptr;
547         struct rtable *rt = (struct rtable*)skb->dst;
548         unsigned char *raw = skb->nh.raw;
549
550         if (opt->rr_needaddr) {
551                 optptr = (unsigned char *)raw + opt->rr;
552                 ip_rt_get_source(&optptr[optptr[2]-5], rt);
553                 opt->is_changed = 1;
554         }
555         if (opt->srr_is_hit) {
556                 int srrptr, srrspace;
557
558                 optptr = raw + opt->srr;
559
560                 for ( srrptr=optptr[2], srrspace = optptr[1];
561                      srrptr <= srrspace;
562                      srrptr += 4
563                      ) {
564                         if (srrptr + 3 > srrspace)
565                                 break;
566                         if (memcmp(&rt->rt_dst, &optptr[srrptr-1], 4) == 0)
567                                 break;
568                 }
569                 if (srrptr + 3 <= srrspace) {
570                         opt->is_changed = 1;
571                         ip_rt_get_source(&optptr[srrptr-1], rt);
572                         skb->nh.iph->daddr = rt->rt_dst;
573                         optptr[2] = srrptr+4;
574                 } else if (net_ratelimit())
575                         printk(KERN_CRIT "ip_forward(): Argh! Destination lost!\n");
576                 if (opt->ts_needaddr) {
577                         optptr = raw + opt->ts;
578                         ip_rt_get_source(&optptr[optptr[2]-9], rt);
579                         opt->is_changed = 1;
580                 }
581         }
582         if (opt->is_changed) {
583                 opt->is_changed = 0;
584                 ip_send_check(skb->nh.iph);
585         }
586 }
587
588 int ip_options_rcv_srr(struct sk_buff *skb)
589 {
590         struct ip_options *opt = &(IPCB(skb)->opt);
591         int srrspace, srrptr;
592         u32 nexthop;
593         struct iphdr *iph = skb->nh.iph;
594         unsigned char * optptr = skb->nh.raw + opt->srr;
595         struct rtable *rt = (struct rtable*)skb->dst;
596         struct rtable *rt2;
597         int err;
598
599         if (!opt->srr)
600                 return 0;
601
602         if (skb->pkt_type != PACKET_HOST)
603                 return -EINVAL;
604         if (rt->rt_type == RTN_UNICAST) {
605                 if (!opt->is_strictroute)
606                         return 0;
607                 icmp_send(skb, ICMP_PARAMETERPROB, 0, htonl(16<<24));
608                 return -EINVAL;
609         }
610         if (rt->rt_type != RTN_LOCAL)
611                 return -EINVAL;
612
613         for (srrptr=optptr[2], srrspace = optptr[1]; srrptr <= srrspace; srrptr += 4) {
614                 if (srrptr + 3 > srrspace) {
615                         icmp_send(skb, ICMP_PARAMETERPROB, 0, htonl((opt->srr+2)<<24));
616                         return -EINVAL;
617                 }
618                 memcpy(&nexthop, &optptr[srrptr-1], 4);
619
620                 rt = (struct rtable*)skb->dst;
621                 skb->dst = NULL;
622                 err = ip_route_input(skb, nexthop, iph->saddr, iph->tos, skb->dev);
623                 rt2 = (struct rtable*)skb->dst;
624                 if (err || (rt2->rt_type != RTN_UNICAST && rt2->rt_type != RTN_LOCAL)) {
625                         ip_rt_put(rt2);
626                         skb->dst = &rt->u.dst;
627                         return -EINVAL;
628                 }
629                 ip_rt_put(rt);
630                 if (rt2->rt_type != RTN_LOCAL)
631                         break;
632                 /* Superfast 8) loopback forward */
633                 memcpy(&iph->daddr, &optptr[srrptr-1], 4);
634                 opt->is_changed = 1;
635         }
636         if (srrptr <= srrspace) {
637                 opt->srr_is_hit = 1;
638                 opt->is_changed = 1;
639         }
640         return 0;
641 }