net/ipv4/xfrm4_policy.c

   1 /*
   2  * xfrm4_policy.c
   3  *
   4  * Changes:
   5  *      Kazunori MIYAZAWA @USAGI
   6  *      YOSHIFUJI Hideaki @USAGI
   7  *              Split up af-specific portion
   8  *
   9  */
  10
  11 #include <linux/compiler.h>
  12 #include <linux/inetdevice.h>
  13 #include <net/xfrm.h>
  14 #include <net/ip.h>
  15
  16 static struct dst_ops xfrm4_dst_ops;
  17 static struct xfrm_policy_afinfo xfrm4_policy_afinfo;
  18
  19 static int xfrm4_dst_lookup(struct xfrm_dst **dst, struct flowi *fl)
  20 {
  21         return __ip_route_output_key((struct rtable**)dst, fl);
  22 }
  23
  24 static int xfrm4_get_saddr(xfrm_address_t *saddr, xfrm_address_t *daddr)
  25 {
  26         struct rtable *rt;
  27         struct flowi fl_tunnel = {
  28                 .nl_u = {
  29                         .ip4_u = {
  30                                 .daddr = daddr->a4,
  31                         },
  32                 },
  33         };
  34
  35         if (!xfrm4_dst_lookup((struct xfrm_dst **)&rt, &fl_tunnel)) {
  36                 saddr->a4 = rt->rt_src;
  37                 dst_release(&rt->u.dst);
  38                 return 0;
  39         }
  40         return -EHOSTUNREACH;
  41 }
  42
  43 static struct dst_entry *
  44 __xfrm4_find_bundle(struct flowi *fl, struct xfrm_policy *policy)
  45 {
  46         struct dst_entry *dst;
  47
  48         read_lock_bh(&policy->lock);
  49         for (dst = policy->bundles; dst; dst = dst->next) {
  50                 struct xfrm_dst *xdst = (struct xfrm_dst*)dst;
  51                 if (xdst->u.rt.fl.oif == fl->oif &&     /*XXX*/
  52                     xdst->u.rt.fl.fl4_dst == fl->fl4_dst &&
  53                     xdst->u.rt.fl.fl4_src == fl->fl4_src &&
  54                     xdst->u.rt.fl.fl4_tos == fl->fl4_tos &&
  55                     xfrm_bundle_ok(policy, xdst, fl, AF_INET, 0)) {
  56                         dst_clone(dst);
  57                         break;
  58                 }
  59         }
  60         read_unlock_bh(&policy->lock);
  61         return dst;
  62 }
  63
  64 /* Allocate chain of dst_entry's, attach known xfrm's, calculate
  65  * all the metrics... Shortly, bundle a bundle.
  66  */
  67
  68 static int
  69 __xfrm4_bundle_create(struct xfrm_policy *policy, struct xfrm_state **xfrm, int nx,
  70                       struct flowi *fl, struct dst_entry **dst_p)
  71 {
  72         struct dst_entry *dst, *dst_prev;
  73         struct rtable *rt0 = (struct rtable*)(*dst_p);
  74         struct rtable *rt = rt0;
  75         __be32 remote = fl->fl4_dst;
  76         __be32 local  = fl->fl4_src;
  77         struct flowi fl_tunnel = {
  78                 .nl_u = {
  79                         .ip4_u = {
  80                                 .saddr = local,
  81                                 .daddr = remote,
  82                                 .tos = fl->fl4_tos
  83                         }
  84                 }
  85         };
  86         int i;
  87         int err;
  88         int header_len = 0;
  89         int trailer_len = 0;
  90
  91         dst = dst_prev = NULL;
  92         dst_hold(&rt->u.dst);
  93
  94         for (i = 0; i < nx; i++) {
  95                 struct dst_entry *dst1 = dst_alloc(&xfrm4_dst_ops);
  96                 struct xfrm_dst *xdst;
  97                 int tunnel = 0;
  98
  99                 if (unlikely(dst1 == NULL)) {
 100                         err = -ENOBUFS;
 101                         dst_release(&rt->u.dst);
 102                         goto error;
 103                 }
 104
 105                 if (!dst)
 106                         dst = dst1;
 107                 else {
 108                         dst_prev->child = dst1;
 109                         dst1->flags |= DST_NOHASH;
 110                         dst_clone(dst1);
 111                 }
 112
 113                 xdst = (struct xfrm_dst *)dst1;
 114                 xdst->route = &rt->u.dst;
 115                 xdst->genid = xfrm[i]->genid;
 116
 117                 dst1->next = dst_prev;
 118                 dst_prev = dst1;
 119                 if (xfrm[i]->props.mode != XFRM_MODE_TRANSPORT) {
 120                         remote = xfrm[i]->id.daddr.a4;
 121                         local  = xfrm[i]->props.saddr.a4;
 122                         tunnel = 1;
 123                 }
 124                 header_len += xfrm[i]->props.header_len;
 125                 trailer_len += xfrm[i]->props.trailer_len;
 126
 127                 if (tunnel) {
 128                         fl_tunnel.fl4_src = local;
 129                         fl_tunnel.fl4_dst = remote;
 130                         err = xfrm_dst_lookup((struct xfrm_dst **)&rt,
 131                                               &fl_tunnel, AF_INET);
 132                         if (err)
 133                                 goto error;
 134                 } else
 135                         dst_hold(&rt->u.dst);
 136         }
 137
 138         dst_prev->child = &rt->u.dst;
 139         dst->path = &rt->u.dst;
 140
 141         *dst_p = dst;
 142         dst = dst_prev;
 143
 144         dst_prev = *dst_p;
 145         i = 0;
 146         for (; dst_prev != &rt->u.dst; dst_prev = dst_prev->child) {
 147                 struct xfrm_dst *x = (struct xfrm_dst*)dst_prev;
 148                 x->u.rt.fl = *fl;
 149
 150                 dst_prev->xfrm = xfrm[i++];
 151                 dst_prev->dev = rt->u.dst.dev;
 152                 if (rt->u.dst.dev)
 153                         dev_hold(rt->u.dst.dev);
 154                 dst_prev->obsolete      = -1;
 155                 dst_prev->flags        |= DST_HOST;
 156                 dst_prev->lastuse       = jiffies;
 157                 dst_prev->header_len    = header_len;
 158                 dst_prev->nfheader_len  = 0;
 159                 dst_prev->trailer_len   = trailer_len;
 160                 memcpy(&dst_prev->metrics, &x->route->metrics, sizeof(dst_prev->metrics));
 161
 162                 /* Copy neighbout for reachability confirmation */
 163                 dst_prev->neighbour     = neigh_clone(rt->u.dst.neighbour);
 164                 dst_prev->input         = rt->u.dst.input;
 165                 dst_prev->output        = xfrm4_output;
 166                 if (rt->peer)
 167                         atomic_inc(&rt->peer->refcnt);
 168                 x->u.rt.peer = rt->peer;
 169                 /* Sheit... I remember I did this right. Apparently,
 170                  * it was magically lost, so this code needs audit */
 171                 x->u.rt.rt_flags = rt0->rt_flags&(RTCF_BROADCAST|RTCF_MULTICAST|RTCF_LOCAL);
 172                 x->u.rt.rt_type = rt->rt_type;
 173                 x->u.rt.rt_src = rt0->rt_src;
 174                 x->u.rt.rt_dst = rt0->rt_dst;
 175                 x->u.rt.rt_gateway = rt->rt_gateway;
 176                 x->u.rt.rt_spec_dst = rt0->rt_spec_dst;
 177                 x->u.rt.idev = rt0->idev;
 178                 in_dev_hold(rt0->idev);
 179                 header_len -= x->u.dst.xfrm->props.header_len;
 180                 trailer_len -= x->u.dst.xfrm->props.trailer_len;
 181         }
 182
 183         xfrm_init_pmtu(dst);
 184         return 0;
 185
 186 error:
 187         if (dst)
 188                 dst_free(dst);
 189         return err;
 190 }
 191
 192 static void
 193 _decode_session4(struct sk_buff *skb, struct flowi *fl)
 194 {
 195         struct iphdr *iph = skb->nh.iph;
 196         u8 *xprth = skb->nh.raw + iph->ihl*4;
 197
 198         memset(fl, 0, sizeof(struct flowi));
 199         if (!(iph->frag_off & htons(IP_MF | IP_OFFSET))) {
 200                 switch (iph->protocol) {
 201                 case IPPROTO_UDP:
 202                 case IPPROTO_UDPLITE:
 203                 case IPPROTO_TCP:
 204                 case IPPROTO_SCTP:
 205                 case IPPROTO_DCCP:
 206                         if (pskb_may_pull(skb, xprth + 4 - skb->data)) {
 207                                 __be16 *ports = (__be16 *)xprth;
 208
 209                                 fl->fl_ip_sport = ports[0];
 210                                 fl->fl_ip_dport = ports[1];
 211                         }
 212                         break;
 213
 214                 case IPPROTO_ICMP:
 215                         if (pskb_may_pull(skb, xprth + 2 - skb->data)) {
 216                                 u8 *icmp = xprth;
 217
 218                                 fl->fl_icmp_type = icmp[0];
 219                                 fl->fl_icmp_code = icmp[1];
 220                         }
 221                         break;
 222
 223                 case IPPROTO_ESP:
 224                         if (pskb_may_pull(skb, xprth + 4 - skb->data)) {
 225                                 __be32 *ehdr = (__be32 *)xprth;
 226
 227                                 fl->fl_ipsec_spi = ehdr[0];
 228                         }
 229                         break;
 230
 231                 case IPPROTO_AH:
 232                         if (pskb_may_pull(skb, xprth + 8 - skb->data)) {
 233                                 __be32 *ah_hdr = (__be32*)xprth;
 234
 235                                 fl->fl_ipsec_spi = ah_hdr[1];
 236                         }
 237                         break;
 238
 239                 case IPPROTO_COMP:
 240                         if (pskb_may_pull(skb, xprth + 4 - skb->data)) {
 241                                 __be16 *ipcomp_hdr = (__be16 *)xprth;
 242
 243                                 fl->fl_ipsec_spi = htonl(ntohs(ipcomp_hdr[1]));
 244                         }
 245                         break;
 246                 default:
 247                         fl->fl_ipsec_spi = 0;
 248                         break;
 249                 };
 250         }
 251         fl->proto = iph->protocol;
 252         fl->fl4_dst = iph->daddr;
 253         fl->fl4_src = iph->saddr;
 254         fl->fl4_tos = iph->tos;
 255 }
 256
 257 static inline int xfrm4_garbage_collect(void)
 258 {
 259         xfrm4_policy_afinfo.garbage_collect();
 260         return (atomic_read(&xfrm4_dst_ops.entries) > xfrm4_dst_ops.gc_thresh*2);
 261 }
 262
 263 static void xfrm4_update_pmtu(struct dst_entry *dst, u32 mtu)
 264 {
 265         struct xfrm_dst *xdst = (struct xfrm_dst *)dst;
 266         struct dst_entry *path = xdst->route;
 267
 268         path->ops->update_pmtu(path, mtu);
 269 }
 270
 271 static void xfrm4_dst_destroy(struct dst_entry *dst)
 272 {
 273         struct xfrm_dst *xdst = (struct xfrm_dst *)dst;
 274
 275         if (likely(xdst->u.rt.idev))
 276                 in_dev_put(xdst->u.rt.idev);
 277         if (likely(xdst->u.rt.peer))
 278                 inet_putpeer(xdst->u.rt.peer);
 279         xfrm_dst_destroy(xdst);
 280 }
 281
 282 static void xfrm4_dst_ifdown(struct dst_entry *dst, struct net_device *dev,
 283                              int unregister)
 284 {
 285         struct xfrm_dst *xdst;
 286
 287         if (!unregister)
 288                 return;
 289
 290         xdst = (struct xfrm_dst *)dst;
 291         if (xdst->u.rt.idev->dev == dev) {
 292                 struct in_device *loopback_idev = in_dev_get(&loopback_dev);
 293                 BUG_ON(!loopback_idev);
 294
 295                 do {
 296                         in_dev_put(xdst->u.rt.idev);
 297                         xdst->u.rt.idev = loopback_idev;
 298                         in_dev_hold(loopback_idev);
 299                         xdst = (struct xfrm_dst *)xdst->u.dst.child;
 300                 } while (xdst->u.dst.xfrm);
 301
 302                 __in_dev_put(loopback_idev);
 303         }
 304
 305         xfrm_dst_ifdown(dst, dev);
 306 }
 307
 308 static struct dst_ops xfrm4_dst_ops = {
 309         .family =               AF_INET,
 310         .protocol =             __constant_htons(ETH_P_IP),
 311         .gc =                   xfrm4_garbage_collect,
 312         .update_pmtu =          xfrm4_update_pmtu,
 313         .destroy =              xfrm4_dst_destroy,
 314         .ifdown =               xfrm4_dst_ifdown,
 315         .gc_thresh =            1024,
 316         .entry_size =           sizeof(struct xfrm_dst),
 317 };
 318
 319 static struct xfrm_policy_afinfo xfrm4_policy_afinfo = {
 320         .family =               AF_INET,
 321         .dst_ops =              &xfrm4_dst_ops,
 322         .dst_lookup =           xfrm4_dst_lookup,
 323         .get_saddr =            xfrm4_get_saddr,
 324         .find_bundle =          __xfrm4_find_bundle,
 325         .bundle_create =        __xfrm4_bundle_create,
 326         .decode_session =       _decode_session4,
 327 };
 328
 329 static void __init xfrm4_policy_init(void)
 330 {
 331         xfrm_policy_register_afinfo(&xfrm4_policy_afinfo);
 332 }
 333
 334 static void __exit xfrm4_policy_fini(void)
 335 {
 336         xfrm_policy_unregister_afinfo(&xfrm4_policy_afinfo);
 337 }
 338
 339 void __init xfrm4_init(void)
 340 {
 341         xfrm4_state_init();
 342         xfrm4_policy_init();
 343 }
 344