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 nft_set *set, const u32 *key,
  45                               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         const struct rb_node *parent;
  50         u8 genmask = nft_genmask_cur(read_pnet(&set->pnet));
  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 found:
  74                         if (!nft_set_elem_active(&rbe->ext, genmask)) {
  75                                 parent = parent->rb_left;
  76                                 continue;
  77                         }
  78                         if (nft_rbtree_interval_end(rbe))
  79                                 goto out;
  80                         spin_unlock_bh(&nft_rbtree_lock);
  81
  82                         *ext = &rbe->ext;
  83                         return true;
  84                 }
  85         }
  86
  87         if (set->flags & NFT_SET_INTERVAL && interval != NULL) {
  88                 rbe = interval;
  89                 goto found;
  90         }
  91 out:
  92         spin_unlock_bh(&nft_rbtree_lock);
  93         return false;
  94 }
  95
  96 static int __nft_rbtree_insert(const struct nft_set *set,
  97                                struct nft_rbtree_elem *new)
  98 {
  99         struct nft_rbtree *priv = nft_set_priv(set);
 100         struct nft_rbtree_elem *rbe;
 101         struct rb_node *parent, **p;
 102         u8 genmask = nft_genmask_next(read_pnet(&set->pnet));
 103         int d;
 104
 105         parent = NULL;
 106         p = &priv->root.rb_node;
 107         while (*p != NULL) {
 108                 parent = *p;
 109                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
 110                 d = memcmp(nft_set_ext_key(&rbe->ext),
 111                            nft_set_ext_key(&new->ext),
 112                            set->klen);
 113                 if (d < 0)
 114                         p = &parent->rb_left;
 115                 else if (d > 0)
 116                         p = &parent->rb_right;
 117                 else {
 118                         if (nft_set_elem_active(&rbe->ext, genmask)) {
 119                                 if (nft_rbtree_interval_end(rbe) &&
 120                                     !nft_rbtree_interval_end(new))
 121                                         p = &parent->rb_left;
 122                                 else if (!nft_rbtree_interval_end(rbe) &&
 123                                          nft_rbtree_interval_end(new))
 124                                         p = &parent->rb_right;
 125                                 else
 126                                         return -EEXIST;
 127                         }
 128                 }
 129         }
 130         rb_link_node(&new->node, parent, p);
 131         rb_insert_color(&new->node, &priv->root);
 132         return 0;
 133 }
 134
 135 static int nft_rbtree_insert(const struct nft_set *set,
 136                              const struct nft_set_elem *elem)
 137 {
 138         struct nft_rbtree_elem *rbe = elem->priv;
 139         int err;
 140
 141         spin_lock_bh(&nft_rbtree_lock);
 142         err = __nft_rbtree_insert(set, rbe);
 143         spin_unlock_bh(&nft_rbtree_lock);
 144
 145         return err;
 146 }
 147
 148 static void nft_rbtree_remove(const struct nft_set *set,
 149                               const struct nft_set_elem *elem)
 150 {
 151         struct nft_rbtree *priv = nft_set_priv(set);
 152         struct nft_rbtree_elem *rbe = elem->priv;
 153
 154         spin_lock_bh(&nft_rbtree_lock);
 155         rb_erase(&rbe->node, &priv->root);
 156         spin_unlock_bh(&nft_rbtree_lock);
 157 }
 158
 159 static void nft_rbtree_activate(const struct nft_set *set,
 160                                 const struct nft_set_elem *elem)
 161 {
 162         struct nft_rbtree_elem *rbe = elem->priv;
 163
 164         nft_set_elem_change_active(set, &rbe->ext);
 165 }
 166
 167 static void *nft_rbtree_deactivate(const struct nft_set *set,
 168                                    const struct nft_set_elem *elem)
 169 {
 170         const struct nft_rbtree *priv = nft_set_priv(set);
 171         const struct rb_node *parent = priv->root.rb_node;
 172         struct nft_rbtree_elem *rbe, *this = elem->priv;
 173         u8 genmask = nft_genmask_cur(read_pnet(&set->pnet));
 174         int d;
 175
 176         while (parent != NULL) {
 177                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
 178
 179                 d = memcmp(nft_set_ext_key(&rbe->ext), &elem->key.val,
 180                                            set->klen);
 181                 if (d < 0)
 182                         parent = parent->rb_left;
 183                 else if (d > 0)
 184                         parent = parent->rb_right;
 185                 else {
 186                         if (!nft_set_elem_active(&rbe->ext, genmask)) {
 187                                 parent = parent->rb_left;
 188                                 continue;
 189                         }
 190                         if (nft_rbtree_interval_end(rbe) &&
 191                             !nft_rbtree_interval_end(this)) {
 192                                 parent = parent->rb_left;
 193                                 continue;
 194                         } else if (!nft_rbtree_interval_end(rbe) &&
 195                                    nft_rbtree_interval_end(this)) {
 196                                 parent = parent->rb_right;
 197                                 continue;
 198                         }
 199                         nft_set_elem_change_active(set, &rbe->ext);
 200                         return rbe;
 201                 }
 202         }
 203         return NULL;
 204 }
 205
 206 static void nft_rbtree_walk(const struct nft_ctx *ctx,
 207                             const struct nft_set *set,
 208                             struct nft_set_iter *iter)
 209 {
 210         const struct nft_rbtree *priv = nft_set_priv(set);
 211         struct nft_rbtree_elem *rbe;
 212         struct nft_set_elem elem;
 213         struct rb_node *node;
 214         u8 genmask = nft_genmask_cur(read_pnet(&set->pnet));
 215
 216         spin_lock_bh(&nft_rbtree_lock);
 217         for (node = rb_first(&priv->root); node != NULL; node = rb_next(node)) {
 218                 rbe = rb_entry(node, struct nft_rbtree_elem, node);
 219
 220                 if (iter->count < iter->skip)
 221                         goto cont;
 222                 if (!nft_set_elem_active(&rbe->ext, genmask))
 223                         goto cont;
 224
 225                 elem.priv = rbe;
 226
 227                 iter->err = iter->fn(ctx, set, iter, &elem);
 228                 if (iter->err < 0) {
 229                         spin_unlock_bh(&nft_rbtree_lock);
 230                         return;
 231                 }
 232 cont:
 233                 iter->count++;
 234         }
 235         spin_unlock_bh(&nft_rbtree_lock);
 236 }
 237
 238 static unsigned int nft_rbtree_privsize(const struct nlattr * const nla[])
 239 {
 240         return sizeof(struct nft_rbtree);
 241 }
 242
 243 static int nft_rbtree_init(const struct nft_set *set,
 244                            const struct nft_set_desc *desc,
 245                            const struct nlattr * const nla[])
 246 {
 247         struct nft_rbtree *priv = nft_set_priv(set);
 248
 249         priv->root = RB_ROOT;
 250         return 0;
 251 }
 252
 253 static void nft_rbtree_destroy(const struct nft_set *set)
 254 {
 255         struct nft_rbtree *priv = nft_set_priv(set);
 256         struct nft_rbtree_elem *rbe;
 257         struct rb_node *node;
 258
 259         while ((node = priv->root.rb_node) != NULL) {
 260                 rb_erase(node, &priv->root);
 261                 rbe = rb_entry(node, struct nft_rbtree_elem, node);
 262                 nft_set_elem_destroy(set, rbe);
 263         }
 264 }
 265
 266 static bool nft_rbtree_estimate(const struct nft_set_desc *desc, u32 features,
 267                                 struct nft_set_estimate *est)
 268 {
 269         unsigned int nsize;
 270
 271         nsize = sizeof(struct nft_rbtree_elem);
 272         if (desc->size)
 273                 est->size = sizeof(struct nft_rbtree) + desc->size * nsize;
 274         else
 275                 est->size = nsize;
 276
 277         est->class = NFT_SET_CLASS_O_LOG_N;
 278
 279         return true;
 280 }
 281
 282 static struct nft_set_ops nft_rbtree_ops __read_mostly = {
 283         .privsize       = nft_rbtree_privsize,
 284         .elemsize       = offsetof(struct nft_rbtree_elem, ext),
 285         .estimate       = nft_rbtree_estimate,
 286         .init           = nft_rbtree_init,
 287         .destroy        = nft_rbtree_destroy,
 288         .insert         = nft_rbtree_insert,
 289         .remove         = nft_rbtree_remove,
 290         .deactivate     = nft_rbtree_deactivate,
 291         .activate       = nft_rbtree_activate,
 292         .lookup         = nft_rbtree_lookup,
 293         .walk           = nft_rbtree_walk,
 294         .features       = NFT_SET_INTERVAL | NFT_SET_MAP,
 295         .owner          = THIS_MODULE,
 296 };
 297
 298 static int __init nft_rbtree_module_init(void)
 299 {
 300         return nft_register_set(&nft_rbtree_ops);
 301 }
 302
 303 static void __exit nft_rbtree_module_exit(void)
 304 {
 305         nft_unregister_set(&nft_rbtree_ops);
 306 }
 307
 308 module_init(nft_rbtree_module_init);
 309 module_exit(nft_rbtree_module_exit);
 310
 311 MODULE_LICENSE("GPL");
 312 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
 313 MODULE_ALIAS_NFT_SET();