net/netfilter/nft_hash.c

   1 /*
   2  * Copyright (c) 2008-2014 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/log2.h>
  16 #include <linux/jhash.h>
  17 #include <linux/netlink.h>
  18 #include <linux/workqueue.h>
  19 #include <linux/rhashtable.h>
  20 #include <linux/netfilter.h>
  21 #include <linux/netfilter/nf_tables.h>
  22 #include <net/netfilter/nf_tables.h>
  23
  24 /* We target a hash table size of 4, element hint is 75% of final size */
  25 #define NFT_HASH_ELEMENT_HINT 3
  26
  27 struct nft_hash {
  28         struct rhashtable               ht;
  29         struct delayed_work             gc_work;
  30 };
  31
  32 struct nft_hash_elem {
  33         struct rhash_head               node;
  34         struct nft_set_ext              ext;
  35 };
  36
  37 struct nft_hash_cmp_arg {
  38         const struct nft_set            *set;
  39         const u32                       *key;
  40         u8                              genmask;
  41 };
  42
  43 static const struct rhashtable_params nft_hash_params;
  44
  45 static inline u32 nft_hash_key(const void *data, u32 len, u32 seed)
  46 {
  47         const struct nft_hash_cmp_arg *arg = data;
  48
  49         return jhash(arg->key, len, seed);
  50 }
  51
  52 static inline u32 nft_hash_obj(const void *data, u32 len, u32 seed)
  53 {
  54         const struct nft_hash_elem *he = data;
  55
  56         return jhash(nft_set_ext_key(&he->ext), len, seed);
  57 }
  58
  59 static inline int nft_hash_cmp(struct rhashtable_compare_arg *arg,
  60                                const void *ptr)
  61 {
  62         const struct nft_hash_cmp_arg *x = arg->key;
  63         const struct nft_hash_elem *he = ptr;
  64
  65         if (memcmp(nft_set_ext_key(&he->ext), x->key, x->set->klen))
  66                 return 1;
  67         if (nft_set_elem_expired(&he->ext))
  68                 return 1;
  69         if (!nft_set_elem_active(&he->ext, x->genmask))
  70                 return 1;
  71         return 0;
  72 }
  73
  74 static bool nft_hash_lookup(const struct nft_set *set, const u32 *key,
  75                             const struct nft_set_ext **ext)
  76 {
  77         struct nft_hash *priv = nft_set_priv(set);
  78         const struct nft_hash_elem *he;
  79         struct nft_hash_cmp_arg arg = {
  80                 .genmask = nft_genmask_cur(read_pnet(&set->pnet)),
  81                 .set     = set,
  82                 .key     = key,
  83         };
  84
  85         he = rhashtable_lookup_fast(&priv->ht, &arg, nft_hash_params);
  86         if (he != NULL)
  87                 *ext = &he->ext;
  88
  89         return !!he;
  90 }
  91
  92 static bool nft_hash_update(struct nft_set *set, const u32 *key,
  93                             void *(*new)(struct nft_set *,
  94                                          const struct nft_expr *,
  95                                          struct nft_regs *regs),
  96                             const struct nft_expr *expr,
  97                             struct nft_regs *regs,
  98                             const struct nft_set_ext **ext)
  99 {
 100         struct nft_hash *priv = nft_set_priv(set);
 101         struct nft_hash_elem *he;
 102         struct nft_hash_cmp_arg arg = {
 103                 .genmask = NFT_GENMASK_ANY,
 104                 .set     = set,
 105                 .key     = key,
 106         };
 107
 108         he = rhashtable_lookup_fast(&priv->ht, &arg, nft_hash_params);
 109         if (he != NULL)
 110                 goto out;
 111
 112         he = new(set, expr, regs);
 113         if (he == NULL)
 114                 goto err1;
 115         if (rhashtable_lookup_insert_key(&priv->ht, &arg, &he->node,
 116                                          nft_hash_params))
 117                 goto err2;
 118 out:
 119         *ext = &he->ext;
 120         return true;
 121
 122 err2:
 123         nft_set_elem_destroy(set, he);
 124 err1:
 125         return false;
 126 }
 127
 128 static int nft_hash_insert(const struct nft_set *set,
 129                            const struct nft_set_elem *elem)
 130 {
 131         struct nft_hash *priv = nft_set_priv(set);
 132         struct nft_hash_elem *he = elem->priv;
 133         struct nft_hash_cmp_arg arg = {
 134                 .genmask = nft_genmask_next(read_pnet(&set->pnet)),
 135                 .set     = set,
 136                 .key     = elem->key.val.data,
 137         };
 138
 139         return rhashtable_lookup_insert_key(&priv->ht, &arg, &he->node,
 140                                             nft_hash_params);
 141 }
 142
 143 static void nft_hash_activate(const struct nft_set *set,
 144                               const struct nft_set_elem *elem)
 145 {
 146         struct nft_hash_elem *he = elem->priv;
 147
 148         nft_set_elem_change_active(set, &he->ext);
 149         nft_set_elem_clear_busy(&he->ext);
 150 }
 151
 152 static void *nft_hash_deactivate(const struct nft_set *set,
 153                                  const struct nft_set_elem *elem)
 154 {
 155         struct nft_hash *priv = nft_set_priv(set);
 156         struct nft_hash_elem *he;
 157         struct nft_hash_cmp_arg arg = {
 158                 .genmask = nft_genmask_next(read_pnet(&set->pnet)),
 159                 .set     = set,
 160                 .key     = elem->key.val.data,
 161         };
 162
 163         rcu_read_lock();
 164         he = rhashtable_lookup_fast(&priv->ht, &arg, nft_hash_params);
 165         if (he != NULL) {
 166                 if (!nft_set_elem_mark_busy(&he->ext))
 167                         nft_set_elem_change_active(set, &he->ext);
 168                 else
 169                         he = NULL;
 170         }
 171         rcu_read_unlock();
 172
 173         return he;
 174 }
 175
 176 static void nft_hash_remove(const struct nft_set *set,
 177                             const struct nft_set_elem *elem)
 178 {
 179         struct nft_hash *priv = nft_set_priv(set);
 180         struct nft_hash_elem *he = elem->priv;
 181
 182         rhashtable_remove_fast(&priv->ht, &he->node, nft_hash_params);
 183 }
 184
 185 static void nft_hash_walk(const struct nft_ctx *ctx, const struct nft_set *set,
 186                           struct nft_set_iter *iter)
 187 {
 188         struct nft_hash *priv = nft_set_priv(set);
 189         struct nft_hash_elem *he;
 190         struct rhashtable_iter hti;
 191         struct nft_set_elem elem;
 192         int err;
 193
 194         err = rhashtable_walk_init(&priv->ht, &hti, GFP_KERNEL);
 195         iter->err = err;
 196         if (err)
 197                 return;
 198
 199         err = rhashtable_walk_start(&hti);
 200         if (err && err != -EAGAIN) {
 201                 iter->err = err;
 202                 goto out;
 203         }
 204
 205         while ((he = rhashtable_walk_next(&hti))) {
 206                 if (IS_ERR(he)) {
 207                         err = PTR_ERR(he);
 208                         if (err != -EAGAIN) {
 209                                 iter->err = err;
 210                                 goto out;
 211                         }
 212
 213                         continue;
 214                 }
 215
 216                 if (iter->count < iter->skip)
 217                         goto cont;
 218                 if (nft_set_elem_expired(&he->ext))
 219                         goto cont;
 220                 if (!nft_set_elem_active(&he->ext, iter->genmask))
 221                         goto cont;
 222
 223                 elem.priv = he;
 224
 225                 iter->err = iter->fn(ctx, set, iter, &elem);
 226                 if (iter->err < 0)
 227                         goto out;
 228
 229 cont:
 230                 iter->count++;
 231         }
 232
 233 out:
 234         rhashtable_walk_stop(&hti);
 235         rhashtable_walk_exit(&hti);
 236 }
 237
 238 static void nft_hash_gc(struct work_struct *work)
 239 {
 240         struct nft_set *set;
 241         struct nft_hash_elem *he;
 242         struct nft_hash *priv;
 243         struct nft_set_gc_batch *gcb = NULL;
 244         struct rhashtable_iter hti;
 245         int err;
 246
 247         priv = container_of(work, struct nft_hash, gc_work.work);
 248         set  = nft_set_container_of(priv);
 249
 250         err = rhashtable_walk_init(&priv->ht, &hti, GFP_KERNEL);
 251         if (err)
 252                 goto schedule;
 253
 254         err = rhashtable_walk_start(&hti);
 255         if (err && err != -EAGAIN)
 256                 goto out;
 257
 258         while ((he = rhashtable_walk_next(&hti))) {
 259                 if (IS_ERR(he)) {
 260                         if (PTR_ERR(he) != -EAGAIN)
 261                                 goto out;
 262                         continue;
 263                 }
 264
 265                 if (!nft_set_elem_expired(&he->ext))
 266                         continue;
 267                 if (nft_set_elem_mark_busy(&he->ext))
 268                         continue;
 269
 270                 gcb = nft_set_gc_batch_check(set, gcb, GFP_ATOMIC);
 271                 if (gcb == NULL)
 272                         goto out;
 273                 rhashtable_remove_fast(&priv->ht, &he->node, nft_hash_params);
 274                 atomic_dec(&set->nelems);
 275                 nft_set_gc_batch_add(gcb, he);
 276         }
 277 out:
 278         rhashtable_walk_stop(&hti);
 279         rhashtable_walk_exit(&hti);
 280
 281         nft_set_gc_batch_complete(gcb);
 282 schedule:
 283         queue_delayed_work(system_power_efficient_wq, &priv->gc_work,
 284                            nft_set_gc_interval(set));
 285 }
 286
 287 static unsigned int nft_hash_privsize(const struct nlattr * const nla[])
 288 {
 289         return sizeof(struct nft_hash);
 290 }
 291
 292 static const struct rhashtable_params nft_hash_params = {
 293         .head_offset            = offsetof(struct nft_hash_elem, node),
 294         .hashfn                 = nft_hash_key,
 295         .obj_hashfn             = nft_hash_obj,
 296         .obj_cmpfn              = nft_hash_cmp,
 297         .automatic_shrinking    = true,
 298 };
 299
 300 static int nft_hash_init(const struct nft_set *set,
 301                          const struct nft_set_desc *desc,
 302                          const struct nlattr * const tb[])
 303 {
 304         struct nft_hash *priv = nft_set_priv(set);
 305         struct rhashtable_params params = nft_hash_params;
 306         int err;
 307
 308         params.nelem_hint = desc->size ?: NFT_HASH_ELEMENT_HINT;
 309         params.key_len    = set->klen;
 310
 311         err = rhashtable_init(&priv->ht, &params);
 312         if (err < 0)
 313                 return err;
 314
 315         INIT_DEFERRABLE_WORK(&priv->gc_work, nft_hash_gc);
 316         if (set->flags & NFT_SET_TIMEOUT)
 317                 queue_delayed_work(system_power_efficient_wq, &priv->gc_work,
 318                                    nft_set_gc_interval(set));
 319         return 0;
 320 }
 321
 322 static void nft_hash_elem_destroy(void *ptr, void *arg)
 323 {
 324         nft_set_elem_destroy((const struct nft_set *)arg, ptr);
 325 }
 326
 327 static void nft_hash_destroy(const struct nft_set *set)
 328 {
 329         struct nft_hash *priv = nft_set_priv(set);
 330
 331         cancel_delayed_work_sync(&priv->gc_work);
 332         rhashtable_free_and_destroy(&priv->ht, nft_hash_elem_destroy,
 333                                     (void *)set);
 334 }
 335
 336 static bool nft_hash_estimate(const struct nft_set_desc *desc, u32 features,
 337                               struct nft_set_estimate *est)
 338 {
 339         unsigned int esize;
 340
 341         esize = sizeof(struct nft_hash_elem);
 342         if (desc->size) {
 343                 est->size = sizeof(struct nft_hash) +
 344                             roundup_pow_of_two(desc->size * 4 / 3) *
 345                             sizeof(struct nft_hash_elem *) +
 346                             desc->size * esize;
 347         } else {
 348                 /* Resizing happens when the load drops below 30% or goes
 349                  * above 75%. The average of 52.5% load (approximated by 50%)
 350                  * is used for the size estimation of the hash buckets,
 351                  * meaning we calculate two buckets per element.
 352                  */
 353                 est->size = esize + 2 * sizeof(struct nft_hash_elem *);
 354         }
 355
 356         est->class = NFT_SET_CLASS_O_1;
 357
 358         return true;
 359 }
 360
 361 static struct nft_set_ops nft_hash_ops __read_mostly = {
 362         .privsize       = nft_hash_privsize,
 363         .elemsize       = offsetof(struct nft_hash_elem, ext),
 364         .estimate       = nft_hash_estimate,
 365         .init           = nft_hash_init,
 366         .destroy        = nft_hash_destroy,
 367         .insert         = nft_hash_insert,
 368         .activate       = nft_hash_activate,
 369         .deactivate     = nft_hash_deactivate,
 370         .remove         = nft_hash_remove,
 371         .lookup         = nft_hash_lookup,
 372         .update         = nft_hash_update,
 373         .walk           = nft_hash_walk,
 374         .features       = NFT_SET_MAP | NFT_SET_TIMEOUT,
 375         .owner          = THIS_MODULE,
 376 };
 377
 378 static int __init nft_hash_module_init(void)
 379 {
 380         return nft_register_set(&nft_hash_ops);
 381 }
 382
 383 static void __exit nft_hash_module_exit(void)
 384 {
 385         nft_unregister_set(&nft_hash_ops);
 386 }
 387
 388 module_init(nft_hash_module_init);
 389 module_exit(nft_hash_module_exit);
 390
 391 MODULE_LICENSE("GPL");
 392 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
 393 MODULE_ALIAS_NFT_SET();