net/netfilter/nft_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 static DEFINE_SPINLOCK(nft_rbtree_lock);
  22
  23 struct nft_rbtree {
  24         struct rb_root          root;
  25 };
  26
  27 struct nft_rbtree_elem {
  28         struct rb_node          node;
  29         struct nft_set_ext      ext;
  30 };
  31
  32 static bool nft_rbtree_interval_end(const struct nft_rbtree_elem *rbe)
  33 {
  34         return nft_set_ext_exists(&rbe->ext, NFT_SET_EXT_FLAGS) &&
  35                (*nft_set_ext_flags(&rbe->ext) & NFT_SET_ELEM_INTERVAL_END);
  36 }
  37
  38 static bool nft_rbtree_equal(const struct nft_set *set, const void *this,
  39                              const struct nft_rbtree_elem *interval)
  40 {
  41         return memcmp(this, nft_set_ext_key(&interval->ext), set->klen) == 0;
  42 }
  43
  44 static bool nft_rbtree_lookup(const struct net *net, const struct nft_set *set,
  45                               const u32 *key, const struct nft_set_ext **ext)
  46 {
  47         const struct nft_rbtree *priv = nft_set_priv(set);
  48         const struct nft_rbtree_elem *rbe, *interval = NULL;
  49         u8 genmask = nft_genmask_cur(net);
  50         const struct rb_node *parent;
  51         const void *this;
  52         int d;
  53
  54         spin_lock_bh(&nft_rbtree_lock);
  55         parent = priv->root.rb_node;
  56         while (parent != NULL) {
  57                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
  58
  59                 this = nft_set_ext_key(&rbe->ext);
  60                 d = memcmp(this, key, set->klen);
  61                 if (d < 0) {
  62                         parent = parent->rb_left;
  63                         /* In case of adjacent ranges, we always see the high
  64                          * part of the range in first place, before the low one.
  65                          * So don't update interval if the keys are equal.
  66                          */
  67                         if (interval && nft_rbtree_equal(set, this, interval))
  68                                 continue;
  69                         interval = rbe;
  70                 } else if (d > 0)
  71                         parent = parent->rb_right;
  72                 else {
  73                         if (!nft_set_elem_active(&rbe->ext, genmask)) {
  74                                 parent = parent->rb_left;
  75                                 continue;
  76                         }
  77                         if (nft_rbtree_interval_end(rbe))
  78                                 goto out;
  79                         spin_unlock_bh(&nft_rbtree_lock);
  80
  81                         *ext = &rbe->ext;
  82                         return true;
  83                 }
  84         }
  85
  86         if (set->flags & NFT_SET_INTERVAL && interval != NULL &&
  87             nft_set_elem_active(&interval->ext, genmask) &&
  88             !nft_rbtree_interval_end(interval)) {
  89                 spin_unlock_bh(&nft_rbtree_lock);
  90                 *ext = &interval->ext;
  91                 return true;
  92         }
  93 out:
  94         spin_unlock_bh(&nft_rbtree_lock);
  95         return false;
  96 }
  97
  98 static int __nft_rbtree_insert(const struct net *net, const struct nft_set *set,
  99                                struct nft_rbtree_elem *new)
 100 {
 101         struct nft_rbtree *priv = nft_set_priv(set);
 102         u8 genmask = nft_genmask_next(net);
 103         struct nft_rbtree_elem *rbe;
 104         struct rb_node *parent, **p;
 105         int d;
 106
 107         parent = NULL;
 108         p = &priv->root.rb_node;
 109         while (*p != NULL) {
 110                 parent = *p;
 111                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
 112                 d = memcmp(nft_set_ext_key(&rbe->ext),
 113                            nft_set_ext_key(&new->ext),
 114                            set->klen);
 115                 if (d < 0)
 116                         p = &parent->rb_left;
 117                 else if (d > 0)
 118                         p = &parent->rb_right;
 119                 else {
 120                         if (nft_set_elem_active(&rbe->ext, genmask)) {
 121                                 if (nft_rbtree_interval_end(rbe) &&
 122                                     !nft_rbtree_interval_end(new))
 123                                         p = &parent->rb_left;
 124                                 else if (!nft_rbtree_interval_end(rbe) &&
 125                                          nft_rbtree_interval_end(new))
 126                                         p = &parent->rb_right;
 127                                 else
 128                                         return -EEXIST;
 129                         }
 130                 }
 131         }
 132         rb_link_node(&new->node, parent, p);
 133         rb_insert_color(&new->node, &priv->root);
 134         return 0;
 135 }
 136
 137 static int nft_rbtree_insert(const struct net *net, const struct nft_set *set,
 138                              const struct nft_set_elem *elem)
 139 {
 140         struct nft_rbtree_elem *rbe = elem->priv;
 141         int err;
 142
 143         spin_lock_bh(&nft_rbtree_lock);
 144         err = __nft_rbtree_insert(net, set, rbe);
 145         spin_unlock_bh(&nft_rbtree_lock);
 146
 147         return err;
 148 }
 149
 150 static void nft_rbtree_remove(const struct nft_set *set,
 151                               const struct nft_set_elem *elem)
 152 {
 153         struct nft_rbtree *priv = nft_set_priv(set);
 154         struct nft_rbtree_elem *rbe = elem->priv;
 155
 156         spin_lock_bh(&nft_rbtree_lock);
 157         rb_erase(&rbe->node, &priv->root);
 158         spin_unlock_bh(&nft_rbtree_lock);
 159 }
 160
 161 static void nft_rbtree_activate(const struct net *net,
 162                                 const struct nft_set *set,
 163                                 const struct nft_set_elem *elem)
 164 {
 165         struct nft_rbtree_elem *rbe = elem->priv;
 166
 167         nft_set_elem_change_active(net, set, &rbe->ext);
 168 }
 169
 170 static void *nft_rbtree_deactivate(const struct net *net,
 171                                    const struct nft_set *set,
 172                                    const struct nft_set_elem *elem)
 173 {
 174         const struct nft_rbtree *priv = nft_set_priv(set);
 175         const struct rb_node *parent = priv->root.rb_node;
 176         struct nft_rbtree_elem *rbe, *this = elem->priv;
 177         u8 genmask = nft_genmask_next(net);
 178         int d;
 179
 180         while (parent != NULL) {
 181                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
 182
 183                 d = memcmp(nft_set_ext_key(&rbe->ext), &elem->key.val,
 184                                            set->klen);
 185                 if (d < 0)
 186                         parent = parent->rb_left;
 187                 else if (d > 0)
 188                         parent = parent->rb_right;
 189                 else {
 190                         if (!nft_set_elem_active(&rbe->ext, genmask)) {
 191                                 parent = parent->rb_left;
 192                                 continue;
 193                         }
 194                         if (nft_rbtree_interval_end(rbe) &&
 195                             !nft_rbtree_interval_end(this)) {
 196                                 parent = parent->rb_left;
 197                                 continue;
 198                         } else if (!nft_rbtree_interval_end(rbe) &&
 199                                    nft_rbtree_interval_end(this)) {
 200                                 parent = parent->rb_right;
 201                                 continue;
 202                         }
 203                         nft_set_elem_change_active(net, set, &rbe->ext);
 204                         return rbe;
 205                 }
 206         }
 207         return NULL;
 208 }
 209
 210 static void nft_rbtree_walk(const struct nft_ctx *ctx,
 211                             const struct nft_set *set,
 212                             struct nft_set_iter *iter)
 213 {
 214         const struct nft_rbtree *priv = nft_set_priv(set);
 215         struct nft_rbtree_elem *rbe;
 216         struct nft_set_elem elem;
 217         struct rb_node *node;
 218
 219         spin_lock_bh(&nft_rbtree_lock);
 220         for (node = rb_first(&priv->root); node != NULL; node = rb_next(node)) {
 221                 rbe = rb_entry(node, struct nft_rbtree_elem, node);
 222
 223                 if (iter->count < iter->skip)
 224                         goto cont;
 225                 if (!nft_set_elem_active(&rbe->ext, iter->genmask))
 226                         goto cont;
 227
 228                 elem.priv = rbe;
 229
 230                 iter->err = iter->fn(ctx, set, iter, &elem);
 231                 if (iter->err < 0) {
 232                         spin_unlock_bh(&nft_rbtree_lock);
 233                         return;
 234                 }
 235 cont:
 236                 iter->count++;
 237         }
 238         spin_unlock_bh(&nft_rbtree_lock);
 239 }
 240
 241 static unsigned int nft_rbtree_privsize(const struct nlattr * const nla[])
 242 {
 243         return sizeof(struct nft_rbtree);
 244 }
 245
 246 static int nft_rbtree_init(const struct nft_set *set,
 247                            const struct nft_set_desc *desc,
 248                            const struct nlattr * const nla[])
 249 {
 250         struct nft_rbtree *priv = nft_set_priv(set);
 251
 252         priv->root = RB_ROOT;
 253         return 0;
 254 }
 255
 256 static void nft_rbtree_destroy(const struct nft_set *set)
 257 {
 258         struct nft_rbtree *priv = nft_set_priv(set);
 259         struct nft_rbtree_elem *rbe;
 260         struct rb_node *node;
 261
 262         while ((node = priv->root.rb_node) != NULL) {
 263                 rb_erase(node, &priv->root);
 264                 rbe = rb_entry(node, struct nft_rbtree_elem, node);
 265                 nft_set_elem_destroy(set, rbe);
 266         }
 267 }
 268
 269 static bool nft_rbtree_estimate(const struct nft_set_desc *desc, u32 features,
 270                                 struct nft_set_estimate *est)
 271 {
 272         unsigned int nsize;
 273
 274         nsize = sizeof(struct nft_rbtree_elem);
 275         if (desc->size)
 276                 est->size = sizeof(struct nft_rbtree) + desc->size * nsize;
 277         else
 278                 est->size = nsize;
 279
 280         est->class = NFT_SET_CLASS_O_LOG_N;
 281
 282         return true;
 283 }
 284
 285 static struct nft_set_ops nft_rbtree_ops __read_mostly = {
 286         .privsize       = nft_rbtree_privsize,
 287         .elemsize       = offsetof(struct nft_rbtree_elem, ext),
 288         .estimate       = nft_rbtree_estimate,
 289         .init           = nft_rbtree_init,
 290         .destroy        = nft_rbtree_destroy,
 291         .insert         = nft_rbtree_insert,
 292         .remove         = nft_rbtree_remove,
 293         .deactivate     = nft_rbtree_deactivate,
 294         .activate       = nft_rbtree_activate,
 295         .lookup         = nft_rbtree_lookup,
 296         .walk           = nft_rbtree_walk,
 297         .features       = NFT_SET_INTERVAL | NFT_SET_MAP,
 298         .owner          = THIS_MODULE,
 299 };
 300
 301 static int __init nft_rbtree_module_init(void)
 302 {
 303         return nft_register_set(&nft_rbtree_ops);
 304 }
 305
 306 static void __exit nft_rbtree_module_exit(void)
 307 {
 308         nft_unregister_set(&nft_rbtree_ops);
 309 }
 310
 311 module_init(nft_rbtree_module_init);
 312 module_exit(nft_rbtree_module_exit);
 313
 314 MODULE_LICENSE("GPL");
 315 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
 316 MODULE_ALIAS_NFT_SET();