net/sched/act_nat.c

   1 /*
   2  * Stateless NAT actions
   3  *
   4  * Copyright (c) 2007 Herbert Xu <herbert@gondor.apana.org.au>
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License as published by the Free
   8  * Software Foundation; either version 2 of the License, or (at your option)
   9  * any later version.
  10  */
  11
  12 #include <linux/errno.h>
  13 #include <linux/init.h>
  14 #include <linux/kernel.h>
  15 #include <linux/module.h>
  16 #include <linux/netfilter.h>
  17 #include <linux/rtnetlink.h>
  18 #include <linux/skbuff.h>
  19 #include <linux/slab.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/string.h>
  22 #include <linux/tc_act/tc_nat.h>
  23 #include <net/act_api.h>
  24 #include <net/icmp.h>
  25 #include <net/ip.h>
  26 #include <net/netlink.h>
  27 #include <net/tc_act/tc_nat.h>
  28 #include <net/tcp.h>
  29 #include <net/udp.h>
  30
  31
  32 #define NAT_TAB_MASK    15
  33 static struct tcf_common *tcf_nat_ht[NAT_TAB_MASK + 1];
  34 static u32 nat_idx_gen;
  35 static DEFINE_RWLOCK(nat_lock);
  36
  37 static struct tcf_hashinfo nat_hash_info = {
  38         .htab   =       tcf_nat_ht,
  39         .hmask  =       NAT_TAB_MASK,
  40         .lock   =       &nat_lock,
  41 };
  42
  43 static const struct nla_policy nat_policy[TCA_NAT_MAX + 1] = {
  44         [TCA_NAT_PARMS] = { .len = sizeof(struct tc_nat) },
  45 };
  46
  47 static int tcf_nat_init(struct nlattr *nla, struct nlattr *est,
  48                         struct tc_action *a, int ovr, int bind)
  49 {
  50         struct nlattr *tb[TCA_NAT_MAX + 1];
  51         struct tc_nat *parm;
  52         int ret = 0, err;
  53         struct tcf_nat *p;
  54         struct tcf_common *pc;
  55
  56         if (nla == NULL)
  57                 return -EINVAL;
  58
  59         err = nla_parse_nested(tb, TCA_NAT_MAX, nla, nat_policy);
  60         if (err < 0)
  61                 return err;
  62
  63         if (tb[TCA_NAT_PARMS] == NULL)
  64                 return -EINVAL;
  65         parm = nla_data(tb[TCA_NAT_PARMS]);
  66
  67         pc = tcf_hash_check(parm->index, a, bind, &nat_hash_info);
  68         if (!pc) {
  69                 pc = tcf_hash_create(parm->index, est, a, sizeof(*p), bind,
  70                                      &nat_idx_gen, &nat_hash_info);
  71                 if (IS_ERR(pc))
  72                         return PTR_ERR(pc);
  73                 p = to_tcf_nat(pc);
  74                 ret = ACT_P_CREATED;
  75         } else {
  76                 p = to_tcf_nat(pc);
  77                 if (!ovr) {
  78                         tcf_hash_release(pc, bind, &nat_hash_info);
  79                         return -EEXIST;
  80                 }
  81         }
  82
  83         spin_lock_bh(&p->tcf_lock);
  84         p->old_addr = parm->old_addr;
  85         p->new_addr = parm->new_addr;
  86         p->mask = parm->mask;
  87         p->flags = parm->flags;
  88
  89         p->tcf_action = parm->action;
  90         spin_unlock_bh(&p->tcf_lock);
  91
  92         if (ret == ACT_P_CREATED)
  93                 tcf_hash_insert(pc, &nat_hash_info);
  94
  95         return ret;
  96 }
  97
  98 static int tcf_nat_cleanup(struct tc_action *a, int bind)
  99 {
 100         struct tcf_nat *p = a->priv;
 101
 102         return tcf_hash_release(&p->common, bind, &nat_hash_info);
 103 }
 104
 105 static int tcf_nat(struct sk_buff *skb, const struct tc_action *a,
 106                    struct tcf_result *res)
 107 {
 108         struct tcf_nat *p = a->priv;
 109         struct iphdr *iph;
 110         __be32 old_addr;
 111         __be32 new_addr;
 112         __be32 mask;
 113         __be32 addr;
 114         int egress;
 115         int action;
 116         int ihl;
 117         int noff;
 118
 119         spin_lock(&p->tcf_lock);
 120
 121         p->tcf_tm.lastuse = jiffies;
 122         old_addr = p->old_addr;
 123         new_addr = p->new_addr;
 124         mask = p->mask;
 125         egress = p->flags & TCA_NAT_FLAG_EGRESS;
 126         action = p->tcf_action;
 127
 128         bstats_update(&p->tcf_bstats, skb);
 129
 130         spin_unlock(&p->tcf_lock);
 131
 132         if (unlikely(action == TC_ACT_SHOT))
 133                 goto drop;
 134
 135         noff = skb_network_offset(skb);
 136         if (!pskb_may_pull(skb, sizeof(*iph) + noff))
 137                 goto drop;
 138
 139         iph = ip_hdr(skb);
 140
 141         if (egress)
 142                 addr = iph->saddr;
 143         else
 144                 addr = iph->daddr;
 145
 146         if (!((old_addr ^ addr) & mask)) {
 147                 if (skb_try_make_writable(skb, sizeof(*iph) + noff))
 148                         goto drop;
 149
 150                 new_addr &= mask;
 151                 new_addr |= addr & ~mask;
 152
 153                 /* Rewrite IP header */
 154                 iph = ip_hdr(skb);
 155                 if (egress)
 156                         iph->saddr = new_addr;
 157                 else
 158                         iph->daddr = new_addr;
 159
 160                 csum_replace4(&iph->check, addr, new_addr);
 161         } else if ((iph->frag_off & htons(IP_OFFSET)) ||
 162                    iph->protocol != IPPROTO_ICMP) {
 163                 goto out;
 164         }
 165
 166         ihl = iph->ihl * 4;
 167
 168         /* It would be nice to share code with stateful NAT. */
 169         switch (iph->frag_off & htons(IP_OFFSET) ? 0 : iph->protocol) {
 170         case IPPROTO_TCP:
 171         {
 172                 struct tcphdr *tcph;
 173
 174                 if (!pskb_may_pull(skb, ihl + sizeof(*tcph) + noff) ||
 175                     skb_try_make_writable(skb, ihl + sizeof(*tcph) + noff))
 176                         goto drop;
 177
 178                 tcph = (void *)(skb_network_header(skb) + ihl);
 179                 inet_proto_csum_replace4(&tcph->check, skb, addr, new_addr, 1);
 180                 break;
 181         }
 182         case IPPROTO_UDP:
 183         {
 184                 struct udphdr *udph;
 185
 186                 if (!pskb_may_pull(skb, ihl + sizeof(*udph) + noff) ||
 187                     skb_try_make_writable(skb, ihl + sizeof(*udph) + noff))
 188                         goto drop;
 189
 190                 udph = (void *)(skb_network_header(skb) + ihl);
 191                 if (udph->check || skb->ip_summed == CHECKSUM_PARTIAL) {
 192                         inet_proto_csum_replace4(&udph->check, skb, addr,
 193                                                  new_addr, 1);
 194                         if (!udph->check)
 195                                 udph->check = CSUM_MANGLED_0;
 196                 }
 197                 break;
 198         }
 199         case IPPROTO_ICMP:
 200         {
 201                 struct icmphdr *icmph;
 202
 203                 if (!pskb_may_pull(skb, ihl + sizeof(*icmph) + noff))
 204                         goto drop;
 205
 206                 icmph = (void *)(skb_network_header(skb) + ihl);
 207
 208                 if ((icmph->type != ICMP_DEST_UNREACH) &&
 209                     (icmph->type != ICMP_TIME_EXCEEDED) &&
 210                     (icmph->type != ICMP_PARAMETERPROB))
 211                         break;
 212
 213                 if (!pskb_may_pull(skb, ihl + sizeof(*icmph) + sizeof(*iph) +
 214                                         noff))
 215                         goto drop;
 216
 217                 icmph = (void *)(skb_network_header(skb) + ihl);
 218                 iph = (void *)(icmph + 1);
 219                 if (egress)
 220                         addr = iph->daddr;
 221                 else
 222                         addr = iph->saddr;
 223
 224                 if ((old_addr ^ addr) & mask)
 225                         break;
 226
 227                 if (skb_try_make_writable(skb, ihl + sizeof(*icmph) +
 228                                           sizeof(*iph) + noff))
 229                         goto drop;
 230
 231                 icmph = (void *)(skb_network_header(skb) + ihl);
 232                 iph = (void *)(icmph + 1);
 233
 234                 new_addr &= mask;
 235                 new_addr |= addr & ~mask;
 236
 237                 /* XXX Fix up the inner checksums. */
 238                 if (egress)
 239                         iph->daddr = new_addr;
 240                 else
 241                         iph->saddr = new_addr;
 242
 243                 inet_proto_csum_replace4(&icmph->checksum, skb, addr, new_addr,
 244                                          0);
 245                 break;
 246         }
 247         default:
 248                 break;
 249         }
 250
 251 out:
 252         return action;
 253
 254 drop:
 255         spin_lock(&p->tcf_lock);
 256         p->tcf_qstats.drops++;
 257         spin_unlock(&p->tcf_lock);
 258         return TC_ACT_SHOT;
 259 }
 260
 261 static int tcf_nat_dump(struct sk_buff *skb, struct tc_action *a,
 262                         int bind, int ref)
 263 {
 264         unsigned char *b = skb_tail_pointer(skb);
 265         struct tcf_nat *p = a->priv;
 266         struct tc_nat opt = {
 267                 .old_addr = p->old_addr,
 268                 .new_addr = p->new_addr,
 269                 .mask     = p->mask,
 270                 .flags    = p->flags,
 271
 272                 .index    = p->tcf_index,
 273                 .action   = p->tcf_action,
 274                 .refcnt   = p->tcf_refcnt - ref,
 275                 .bindcnt  = p->tcf_bindcnt - bind,
 276         };
 277         struct tcf_t t;
 278
 279         NLA_PUT(skb, TCA_NAT_PARMS, sizeof(opt), &opt);
 280         t.install = jiffies_to_clock_t(jiffies - p->tcf_tm.install);
 281         t.lastuse = jiffies_to_clock_t(jiffies - p->tcf_tm.lastuse);
 282         t.expires = jiffies_to_clock_t(p->tcf_tm.expires);
 283         NLA_PUT(skb, TCA_NAT_TM, sizeof(t), &t);
 284
 285         return skb->len;
 286
 287 nla_put_failure:
 288         nlmsg_trim(skb, b);
 289         return -1;
 290 }
 291
 292 static struct tc_action_ops act_nat_ops = {
 293         .kind           =       "nat",
 294         .hinfo          =       &nat_hash_info,
 295         .type           =       TCA_ACT_NAT,
 296         .capab          =       TCA_CAP_NONE,
 297         .owner          =       THIS_MODULE,
 298         .act            =       tcf_nat,
 299         .dump           =       tcf_nat_dump,
 300         .cleanup        =       tcf_nat_cleanup,
 301         .lookup         =       tcf_hash_search,
 302         .init           =       tcf_nat_init,
 303         .walk           =       tcf_generic_walker
 304 };
 305
 306 MODULE_DESCRIPTION("Stateless NAT actions");
 307 MODULE_LICENSE("GPL");
 308
 309 static int __init nat_init_module(void)
 310 {
 311         return tcf_register_action(&act_nat_ops);
 312 }
 313
 314 static void __exit nat_cleanup_module(void)
 315 {
 316         tcf_unregister_action(&act_nat_ops);
 317 }
 318
 319 module_init(nat_init_module);
 320 module_exit(nat_cleanup_module);