net/netfilter/nft_set_rbtree.c

   1 /*
   2  * Copyright (c) 2008-2009 Patrick McHardy <kaber@trash.net>
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License version 2 as
   6  * published by the Free Software Foundation.
   7  *
   8  * Development of this code funded by Astaro AG (http://www.astaro.com/)
   9  */
  10
  11 #include <linux/kernel.h>
  12 #include <linux/init.h>
  13 #include <linux/module.h>
  14 #include <linux/list.h>
  15 #include <linux/rbtree.h>
  16 #include <linux/netlink.h>
  17 #include <linux/netfilter.h>
  18 #include <linux/netfilter/nf_tables.h>
  19 #include <net/netfilter/nf_tables.h>
  20
  21 struct nft_rbtree {
  22         rwlock_t                lock;
  23         struct rb_root          root;
  24 };
  25
  26 struct nft_rbtree_elem {
  27         struct rb_node          node;
  28         struct nft_set_ext      ext;
  29 };
  30
  31 static bool nft_rbtree_interval_end(const struct nft_rbtree_elem *rbe)
  32 {
  33         return nft_set_ext_exists(&rbe->ext, NFT_SET_EXT_FLAGS) &&
  34                (*nft_set_ext_flags(&rbe->ext) & NFT_SET_ELEM_INTERVAL_END);
  35 }
  36
  37 static bool nft_rbtree_equal(const struct nft_set *set, const void *this,
  38                              const struct nft_rbtree_elem *interval)
  39 {
  40         return memcmp(this, nft_set_ext_key(&interval->ext), set->klen) == 0;
  41 }
  42
  43 static bool nft_rbtree_lookup(const struct net *net, const struct nft_set *set,
  44                               const u32 *key, const struct nft_set_ext **ext)
  45 {
  46         struct nft_rbtree *priv = nft_set_priv(set);
  47         const struct nft_rbtree_elem *rbe, *interval = NULL;
  48         u8 genmask = nft_genmask_cur(net);
  49         const struct rb_node *parent;
  50         const void *this;
  51         int d;
  52
  53         read_lock_bh(&priv->lock);
  54         parent = priv->root.rb_node;
  55         while (parent != NULL) {
  56                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
  57
  58                 this = nft_set_ext_key(&rbe->ext);
  59                 d = memcmp(this, key, set->klen);
  60                 if (d < 0) {
  61                         parent = parent->rb_left;
  62                         if (interval &&
  63                             nft_rbtree_equal(set, this, interval) &&
  64                             nft_rbtree_interval_end(this) &&
  65                             !nft_rbtree_interval_end(interval))
  66                                 continue;
  67                         interval = rbe;
  68                 } else if (d > 0)
  69                         parent = parent->rb_right;
  70                 else {
  71                         if (!nft_set_elem_active(&rbe->ext, genmask)) {
  72                                 parent = parent->rb_left;
  73                                 continue;
  74                         }
  75                         if (nft_rbtree_interval_end(rbe))
  76                                 goto out;
  77                         read_unlock_bh(&priv->lock);
  78
  79                         *ext = &rbe->ext;
  80                         return true;
  81                 }
  82         }
  83
  84         if (set->flags & NFT_SET_INTERVAL && interval != NULL &&
  85             nft_set_elem_active(&interval->ext, genmask) &&
  86             !nft_rbtree_interval_end(interval)) {
  87                 read_unlock_bh(&priv->lock);
  88                 *ext = &interval->ext;
  89                 return true;
  90         }
  91 out:
  92         read_unlock_bh(&priv->lock);
  93         return false;
  94 }
  95
  96 static int __nft_rbtree_insert(const struct net *net, const struct nft_set *set,
  97                                struct nft_rbtree_elem *new,
  98                                struct nft_set_ext **ext)
  99 {
 100         struct nft_rbtree *priv = nft_set_priv(set);
 101         u8 genmask = nft_genmask_next(net);
 102         struct nft_rbtree_elem *rbe;
 103         struct rb_node *parent, **p;
 104         int d;
 105
 106         parent = NULL;
 107         p = &priv->root.rb_node;
 108         while (*p != NULL) {
 109                 parent = *p;
 110                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
 111                 d = memcmp(nft_set_ext_key(&rbe->ext),
 112                            nft_set_ext_key(&new->ext),
 113                            set->klen);
 114                 if (d < 0)
 115                         p = &parent->rb_left;
 116                 else if (d > 0)
 117                         p = &parent->rb_right;
 118                 else {
 119                         if (nft_set_elem_active(&rbe->ext, genmask)) {
 120                                 if (nft_rbtree_interval_end(rbe) &&
 121                                     !nft_rbtree_interval_end(new))
 122                                         p = &parent->rb_left;
 123                                 else if (!nft_rbtree_interval_end(rbe) &&
 124                                          nft_rbtree_interval_end(new))
 125                                         p = &parent->rb_right;
 126                                 else {
 127                                         *ext = &rbe->ext;
 128                                         return -EEXIST;
 129                                 }
 130                         }
 131                 }
 132         }
 133         rb_link_node(&new->node, parent, p);
 134         rb_insert_color(&new->node, &priv->root);
 135         return 0;
 136 }
 137
 138 static int nft_rbtree_insert(const struct net *net, const struct nft_set *set,
 139                              const struct nft_set_elem *elem,
 140                              struct nft_set_ext **ext)
 141 {
 142         struct nft_rbtree *priv = nft_set_priv(set);
 143         struct nft_rbtree_elem *rbe = elem->priv;
 144         int err;
 145
 146         write_lock_bh(&priv->lock);
 147         err = __nft_rbtree_insert(net, set, rbe, ext);
 148         write_unlock_bh(&priv->lock);
 149
 150         return err;
 151 }
 152
 153 static void nft_rbtree_remove(const struct net *net,
 154                               const struct nft_set *set,
 155                               const struct nft_set_elem *elem)
 156 {
 157         struct nft_rbtree *priv = nft_set_priv(set);
 158         struct nft_rbtree_elem *rbe = elem->priv;
 159
 160         write_lock_bh(&priv->lock);
 161         rb_erase(&rbe->node, &priv->root);
 162         write_unlock_bh(&priv->lock);
 163 }
 164
 165 static void nft_rbtree_activate(const struct net *net,
 166                                 const struct nft_set *set,
 167                                 const struct nft_set_elem *elem)
 168 {
 169         struct nft_rbtree_elem *rbe = elem->priv;
 170
 171         nft_set_elem_change_active(net, set, &rbe->ext);
 172 }
 173
 174 static bool nft_rbtree_flush(const struct net *net,
 175                              const struct nft_set *set, void *priv)
 176 {
 177         struct nft_rbtree_elem *rbe = priv;
 178
 179         nft_set_elem_change_active(net, set, &rbe->ext);
 180         return true;
 181 }
 182
 183 static void *nft_rbtree_deactivate(const struct net *net,
 184                                    const struct nft_set *set,
 185                                    const struct nft_set_elem *elem)
 186 {
 187         const struct nft_rbtree *priv = nft_set_priv(set);
 188         const struct rb_node *parent = priv->root.rb_node;
 189         struct nft_rbtree_elem *rbe, *this = elem->priv;
 190         u8 genmask = nft_genmask_next(net);
 191         int d;
 192
 193         while (parent != NULL) {
 194                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
 195
 196                 d = memcmp(nft_set_ext_key(&rbe->ext), &elem->key.val,
 197                                            set->klen);
 198                 if (d < 0)
 199                         parent = parent->rb_left;
 200                 else if (d > 0)
 201                         parent = parent->rb_right;
 202                 else {
 203                         if (!nft_set_elem_active(&rbe->ext, genmask)) {
 204                                 parent = parent->rb_left;
 205                                 continue;
 206                         }
 207                         if (nft_rbtree_interval_end(rbe) &&
 208                             !nft_rbtree_interval_end(this)) {
 209                                 parent = parent->rb_left;
 210                                 continue;
 211                         } else if (!nft_rbtree_interval_end(rbe) &&
 212                                    nft_rbtree_interval_end(this)) {
 213                                 parent = parent->rb_right;
 214                                 continue;
 215                         }
 216                         nft_rbtree_flush(net, set, rbe);
 217                         return rbe;
 218                 }
 219         }
 220         return NULL;
 221 }
 222
 223 static void nft_rbtree_walk(const struct nft_ctx *ctx,
 224                             struct nft_set *set,
 225                             struct nft_set_iter *iter)
 226 {
 227         struct nft_rbtree *priv = nft_set_priv(set);
 228         struct nft_rbtree_elem *rbe;
 229         struct nft_set_elem elem;
 230         struct rb_node *node;
 231
 232         read_lock_bh(&priv->lock);
 233         for (node = rb_first(&priv->root); node != NULL; node = rb_next(node)) {
 234                 rbe = rb_entry(node, struct nft_rbtree_elem, node);
 235
 236                 if (iter->count < iter->skip)
 237                         goto cont;
 238                 if (!nft_set_elem_active(&rbe->ext, iter->genmask))
 239                         goto cont;
 240
 241                 elem.priv = rbe;
 242
 243                 iter->err = iter->fn(ctx, set, iter, &elem);
 244                 if (iter->err < 0) {
 245                         read_unlock_bh(&priv->lock);
 246                         return;
 247                 }
 248 cont:
 249                 iter->count++;
 250         }
 251         read_unlock_bh(&priv->lock);
 252 }
 253
 254 static unsigned int nft_rbtree_privsize(const struct nlattr * const nla[])
 255 {
 256         return sizeof(struct nft_rbtree);
 257 }
 258
 259 static int nft_rbtree_init(const struct nft_set *set,
 260                            const struct nft_set_desc *desc,
 261                            const struct nlattr * const nla[])
 262 {
 263         struct nft_rbtree *priv = nft_set_priv(set);
 264
 265         rwlock_init(&priv->lock);
 266         priv->root = RB_ROOT;
 267         return 0;
 268 }
 269
 270 static void nft_rbtree_destroy(const struct nft_set *set)
 271 {
 272         struct nft_rbtree *priv = nft_set_priv(set);
 273         struct nft_rbtree_elem *rbe;
 274         struct rb_node *node;
 275
 276         while ((node = priv->root.rb_node) != NULL) {
 277                 rb_erase(node, &priv->root);
 278                 rbe = rb_entry(node, struct nft_rbtree_elem, node);
 279                 nft_set_elem_destroy(set, rbe, true);
 280         }
 281 }
 282
 283 static bool nft_rbtree_estimate(const struct nft_set_desc *desc, u32 features,
 284                                 struct nft_set_estimate *est)
 285 {
 286         unsigned int nsize;
 287
 288         nsize = sizeof(struct nft_rbtree_elem);
 289         if (desc->size)
 290                 est->size = sizeof(struct nft_rbtree) + desc->size * nsize;
 291         else
 292                 est->size = nsize;
 293
 294         est->lookup = NFT_SET_CLASS_O_LOG_N;
 295         est->space  = NFT_SET_CLASS_O_N;
 296
 297         return true;
 298 }
 299
 300 static struct nft_set_ops nft_rbtree_ops __read_mostly = {
 301         .privsize       = nft_rbtree_privsize,
 302         .elemsize       = offsetof(struct nft_rbtree_elem, ext),
 303         .estimate       = nft_rbtree_estimate,
 304         .init           = nft_rbtree_init,
 305         .destroy        = nft_rbtree_destroy,
 306         .insert         = nft_rbtree_insert,
 307         .remove         = nft_rbtree_remove,
 308         .deactivate     = nft_rbtree_deactivate,
 309         .flush          = nft_rbtree_flush,
 310         .activate       = nft_rbtree_activate,
 311         .lookup         = nft_rbtree_lookup,
 312         .walk           = nft_rbtree_walk,
 313         .features       = NFT_SET_INTERVAL | NFT_SET_MAP | NFT_SET_OBJECT,
 314         .owner          = THIS_MODULE,
 315 };
 316
 317 static int __init nft_rbtree_module_init(void)
 318 {
 319         return nft_register_set(&nft_rbtree_ops);
 320 }
 321
 322 static void __exit nft_rbtree_module_exit(void)
 323 {
 324         nft_unregister_set(&nft_rbtree_ops);
 325 }
 326
 327 module_init(nft_rbtree_module_init);
 328 module_exit(nft_rbtree_module_exit);
 329
 330 MODULE_LICENSE("GPL");
 331 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
 332 MODULE_ALIAS_NFT_SET();