net/netfilter/nft_set_rbtree.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2008-2009 Patrick McHardy <kaber@trash.net>
   4  *
   5  * Development of this code funded by Astaro AG (http://www.astaro.com/)
   6  */
   7
   8 #include <linux/kernel.h>
   9 #include <linux/init.h>
  10 #include <linux/module.h>
  11 #include <linux/list.h>
  12 #include <linux/rbtree.h>
  13 #include <linux/netlink.h>
  14 #include <linux/netfilter.h>
  15 #include <linux/netfilter/nf_tables.h>
  16 #include <net/netfilter/nf_tables_core.h>
  17
  18 struct nft_rbtree {
  19         struct rb_root          root;
  20         rwlock_t                lock;
  21         seqcount_rwlock_t       count;
  22         struct delayed_work     gc_work;
  23 };
  24
  25 struct nft_rbtree_elem {
  26         struct rb_node          node;
  27         struct nft_set_ext      ext;
  28 };
  29
  30 static bool nft_rbtree_interval_end(const struct nft_rbtree_elem *rbe)
  31 {
  32         return nft_set_ext_exists(&rbe->ext, NFT_SET_EXT_FLAGS) &&
  33                (*nft_set_ext_flags(&rbe->ext) & NFT_SET_ELEM_INTERVAL_END);
  34 }
  35
  36 static bool nft_rbtree_interval_start(const struct nft_rbtree_elem *rbe)
  37 {
  38         return !nft_rbtree_interval_end(rbe);
  39 }
  40
  41 static int nft_rbtree_cmp(const struct nft_set *set,
  42                           const struct nft_rbtree_elem *e1,
  43                           const struct nft_rbtree_elem *e2)
  44 {
  45         return memcmp(nft_set_ext_key(&e1->ext), nft_set_ext_key(&e2->ext),
  46                       set->klen);
  47 }
  48
  49 static bool nft_rbtree_elem_expired(const struct nft_rbtree_elem *rbe)
  50 {
  51         return nft_set_elem_expired(&rbe->ext) ||
  52                nft_set_elem_is_dead(&rbe->ext);
  53 }
  54
  55 static bool __nft_rbtree_lookup(const struct net *net, const struct nft_set *set,
  56                                 const u32 *key, const struct nft_set_ext **ext,
  57                                 unsigned int seq)
  58 {
  59         struct nft_rbtree *priv = nft_set_priv(set);
  60         const struct nft_rbtree_elem *rbe, *interval = NULL;
  61         u8 genmask = nft_genmask_cur(net);
  62         const struct rb_node *parent;
  63         int d;
  64
  65         parent = rcu_dereference_raw(priv->root.rb_node);
  66         while (parent != NULL) {
  67                 if (read_seqcount_retry(&priv->count, seq))
  68                         return false;
  69
  70                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
  71
  72                 d = memcmp(nft_set_ext_key(&rbe->ext), key, set->klen);
  73                 if (d < 0) {
  74                         parent = rcu_dereference_raw(parent->rb_left);
  75                         if (interval &&
  76                             !nft_rbtree_cmp(set, rbe, interval) &&
  77                             nft_rbtree_interval_end(rbe) &&
  78                             nft_rbtree_interval_start(interval))
  79                                 continue;
  80                         interval = rbe;
  81                 } else if (d > 0)
  82                         parent = rcu_dereference_raw(parent->rb_right);
  83                 else {
  84                         if (!nft_set_elem_active(&rbe->ext, genmask)) {
  85                                 parent = rcu_dereference_raw(parent->rb_left);
  86                                 continue;
  87                         }
  88
  89                         if (nft_rbtree_elem_expired(rbe))
  90                                 return false;
  91
  92                         if (nft_rbtree_interval_end(rbe)) {
  93                                 if (nft_set_is_anonymous(set))
  94                                         return false;
  95                                 parent = rcu_dereference_raw(parent->rb_left);
  96                                 interval = NULL;
  97                                 continue;
  98                         }
  99
 100                         *ext = &rbe->ext;
 101                         return true;
 102                 }
 103         }
 104
 105         if (set->flags & NFT_SET_INTERVAL && interval != NULL &&
 106             nft_set_elem_active(&interval->ext, genmask) &&
 107             !nft_rbtree_elem_expired(interval) &&
 108             nft_rbtree_interval_start(interval)) {
 109                 *ext = &interval->ext;
 110                 return true;
 111         }
 112
 113         return false;
 114 }
 115
 116 INDIRECT_CALLABLE_SCOPE
 117 bool nft_rbtree_lookup(const struct net *net, const struct nft_set *set,
 118                        const u32 *key, const struct nft_set_ext **ext)
 119 {
 120         struct nft_rbtree *priv = nft_set_priv(set);
 121         unsigned int seq = read_seqcount_begin(&priv->count);
 122         bool ret;
 123
 124         ret = __nft_rbtree_lookup(net, set, key, ext, seq);
 125         if (ret || !read_seqcount_retry(&priv->count, seq))
 126                 return ret;
 127
 128         read_lock_bh(&priv->lock);
 129         seq = read_seqcount_begin(&priv->count);
 130         ret = __nft_rbtree_lookup(net, set, key, ext, seq);
 131         read_unlock_bh(&priv->lock);
 132
 133         return ret;
 134 }
 135
 136 static bool __nft_rbtree_get(const struct net *net, const struct nft_set *set,
 137                              const u32 *key, struct nft_rbtree_elem **elem,
 138                              unsigned int seq, unsigned int flags, u8 genmask)
 139 {
 140         struct nft_rbtree_elem *rbe, *interval = NULL;
 141         struct nft_rbtree *priv = nft_set_priv(set);
 142         const struct rb_node *parent;
 143         const void *this;
 144         int d;
 145
 146         parent = rcu_dereference_raw(priv->root.rb_node);
 147         while (parent != NULL) {
 148                 if (read_seqcount_retry(&priv->count, seq))
 149                         return false;
 150
 151                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
 152
 153                 this = nft_set_ext_key(&rbe->ext);
 154                 d = memcmp(this, key, set->klen);
 155                 if (d < 0) {
 156                         parent = rcu_dereference_raw(parent->rb_left);
 157                         if (!(flags & NFT_SET_ELEM_INTERVAL_END))
 158                                 interval = rbe;
 159                 } else if (d > 0) {
 160                         parent = rcu_dereference_raw(parent->rb_right);
 161                         if (flags & NFT_SET_ELEM_INTERVAL_END)
 162                                 interval = rbe;
 163                 } else {
 164                         if (!nft_set_elem_active(&rbe->ext, genmask)) {
 165                                 parent = rcu_dereference_raw(parent->rb_left);
 166                                 continue;
 167                         }
 168
 169                         if (nft_set_elem_expired(&rbe->ext))
 170                                 return false;
 171
 172                         if (!nft_set_ext_exists(&rbe->ext, NFT_SET_EXT_FLAGS) ||
 173                             (*nft_set_ext_flags(&rbe->ext) & NFT_SET_ELEM_INTERVAL_END) ==
 174                             (flags & NFT_SET_ELEM_INTERVAL_END)) {
 175                                 *elem = rbe;
 176                                 return true;
 177                         }
 178
 179                         if (nft_rbtree_interval_end(rbe))
 180                                 interval = NULL;
 181
 182                         parent = rcu_dereference_raw(parent->rb_left);
 183                 }
 184         }
 185
 186         if (set->flags & NFT_SET_INTERVAL && interval != NULL &&
 187             nft_set_elem_active(&interval->ext, genmask) &&
 188             !nft_set_elem_expired(&interval->ext) &&
 189             ((!nft_rbtree_interval_end(interval) &&
 190               !(flags & NFT_SET_ELEM_INTERVAL_END)) ||
 191              (nft_rbtree_interval_end(interval) &&
 192               (flags & NFT_SET_ELEM_INTERVAL_END)))) {
 193                 *elem = interval;
 194                 return true;
 195         }
 196
 197         return false;
 198 }
 199
 200 static void *nft_rbtree_get(const struct net *net, const struct nft_set *set,
 201                             const struct nft_set_elem *elem, unsigned int flags)
 202 {
 203         struct nft_rbtree *priv = nft_set_priv(set);
 204         unsigned int seq = read_seqcount_begin(&priv->count);
 205         struct nft_rbtree_elem *rbe = ERR_PTR(-ENOENT);
 206         const u32 *key = (const u32 *)&elem->key.val;
 207         u8 genmask = nft_genmask_cur(net);
 208         bool ret;
 209
 210         ret = __nft_rbtree_get(net, set, key, &rbe, seq, flags, genmask);
 211         if (ret || !read_seqcount_retry(&priv->count, seq))
 212                 return rbe;
 213
 214         read_lock_bh(&priv->lock);
 215         seq = read_seqcount_begin(&priv->count);
 216         ret = __nft_rbtree_get(net, set, key, &rbe, seq, flags, genmask);
 217         if (!ret)
 218                 rbe = ERR_PTR(-ENOENT);
 219         read_unlock_bh(&priv->lock);
 220
 221         return rbe;
 222 }
 223
 224 static void nft_rbtree_gc_remove(struct net *net, struct nft_set *set,
 225                                  struct nft_rbtree *priv,
 226                                  struct nft_rbtree_elem *rbe)
 227 {
 228         struct nft_set_elem elem = {
 229                 .priv   = rbe,
 230         };
 231
 232         nft_setelem_data_deactivate(net, set, &elem);
 233         rb_erase(&rbe->node, &priv->root);
 234 }
 235
 236 static int nft_rbtree_gc_elem(const struct nft_set *__set,
 237                               struct nft_rbtree *priv,
 238                               struct nft_rbtree_elem *rbe,
 239                               u8 genmask)
 240 {
 241         struct nft_set *set = (struct nft_set *)__set;
 242         struct rb_node *prev = rb_prev(&rbe->node);
 243         struct net *net = read_pnet(&set->net);
 244         struct nft_rbtree_elem *rbe_prev;
 245         struct nft_trans_gc *gc;
 246
 247         gc = nft_trans_gc_alloc(set, 0, GFP_ATOMIC);
 248         if (!gc)
 249                 return -ENOMEM;
 250
 251         /* search for end interval coming before this element.
 252          * end intervals don't carry a timeout extension, they
 253          * are coupled with the interval start element.
 254          */
 255         while (prev) {
 256                 rbe_prev = rb_entry(prev, struct nft_rbtree_elem, node);
 257                 if (nft_rbtree_interval_end(rbe_prev) &&
 258                     nft_set_elem_active(&rbe_prev->ext, genmask))
 259                         break;
 260
 261                 prev = rb_prev(prev);
 262         }
 263
 264         if (prev) {
 265                 rbe_prev = rb_entry(prev, struct nft_rbtree_elem, node);
 266                 nft_rbtree_gc_remove(net, set, priv, rbe_prev);
 267
 268                 /* There is always room in this trans gc for this element,
 269                  * memory allocation never actually happens, hence, the warning
 270                  * splat in such case. No need to set NFT_SET_ELEM_DEAD_BIT,
 271                  * this is synchronous gc which never fails.
 272                  */
 273                 gc = nft_trans_gc_queue_sync(gc, GFP_ATOMIC);
 274                 if (WARN_ON_ONCE(!gc))
 275                         return -ENOMEM;
 276
 277                 nft_trans_gc_elem_add(gc, rbe_prev);
 278         }
 279
 280         nft_rbtree_gc_remove(net, set, priv, rbe);
 281         gc = nft_trans_gc_queue_sync(gc, GFP_ATOMIC);
 282         if (WARN_ON_ONCE(!gc))
 283                 return -ENOMEM;
 284
 285         nft_trans_gc_elem_add(gc, rbe);
 286
 287         nft_trans_gc_queue_sync_done(gc);
 288
 289         return 0;
 290 }
 291
 292 static bool nft_rbtree_update_first(const struct nft_set *set,
 293                                     struct nft_rbtree_elem *rbe,
 294                                     struct rb_node *first)
 295 {
 296         struct nft_rbtree_elem *first_elem;
 297
 298         first_elem = rb_entry(first, struct nft_rbtree_elem, node);
 299         /* this element is closest to where the new element is to be inserted:
 300          * update the first element for the node list path.
 301          */
 302         if (nft_rbtree_cmp(set, rbe, first_elem) < 0)
 303                 return true;
 304
 305         return false;
 306 }
 307
 308 static int __nft_rbtree_insert(const struct net *net, const struct nft_set *set,
 309                                struct nft_rbtree_elem *new,
 310                                struct nft_set_ext **ext)
 311 {
 312         struct nft_rbtree_elem *rbe, *rbe_le = NULL, *rbe_ge = NULL;
 313         struct rb_node *node, *next, *parent, **p, *first = NULL;
 314         struct nft_rbtree *priv = nft_set_priv(set);
 315         u8 genmask = nft_genmask_next(net);
 316         int d, err;
 317
 318         /* Descend the tree to search for an existing element greater than the
 319          * key value to insert that is greater than the new element. This is the
 320          * first element to walk the ordered elements to find possible overlap.
 321          */
 322         parent = NULL;
 323         p = &priv->root.rb_node;
 324         while (*p != NULL) {
 325                 parent = *p;
 326                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
 327                 d = nft_rbtree_cmp(set, rbe, new);
 328
 329                 if (d < 0) {
 330                         p = &parent->rb_left;
 331                 } else if (d > 0) {
 332                         if (!first ||
 333                             nft_rbtree_update_first(set, rbe, first))
 334                                 first = &rbe->node;
 335
 336                         p = &parent->rb_right;
 337                 } else {
 338                         if (nft_rbtree_interval_end(rbe))
 339                                 p = &parent->rb_left;
 340                         else
 341                                 p = &parent->rb_right;
 342                 }
 343         }
 344
 345         if (!first)
 346                 first = rb_first(&priv->root);
 347
 348         /* Detect overlap by going through the list of valid tree nodes.
 349          * Values stored in the tree are in reversed order, starting from
 350          * highest to lowest value.
 351          */
 352         for (node = first; node != NULL; node = next) {
 353                 next = rb_next(node);
 354
 355                 rbe = rb_entry(node, struct nft_rbtree_elem, node);
 356
 357                 if (!nft_set_elem_active(&rbe->ext, genmask))
 358                         continue;
 359
 360                 /* perform garbage collection to avoid bogus overlap reports. */
 361                 if (nft_set_elem_expired(&rbe->ext)) {
 362                         err = nft_rbtree_gc_elem(set, priv, rbe, genmask);
 363                         if (err < 0)
 364                                 return err;
 365
 366                         continue;
 367                 }
 368
 369                 d = nft_rbtree_cmp(set, rbe, new);
 370                 if (d == 0) {
 371                         /* Matching end element: no need to look for an
 372                          * overlapping greater or equal element.
 373                          */
 374                         if (nft_rbtree_interval_end(rbe)) {
 375                                 rbe_le = rbe;
 376                                 break;
 377                         }
 378
 379                         /* first element that is greater or equal to key value. */
 380                         if (!rbe_ge) {
 381                                 rbe_ge = rbe;
 382                                 continue;
 383                         }
 384
 385                         /* this is a closer more or equal element, update it. */
 386                         if (nft_rbtree_cmp(set, rbe_ge, new) != 0) {
 387                                 rbe_ge = rbe;
 388                                 continue;
 389                         }
 390
 391                         /* element is equal to key value, make sure flags are
 392                          * the same, an existing more or equal start element
 393                          * must not be replaced by more or equal end element.
 394                          */
 395                         if ((nft_rbtree_interval_start(new) &&
 396                              nft_rbtree_interval_start(rbe_ge)) ||
 397                             (nft_rbtree_interval_end(new) &&
 398                              nft_rbtree_interval_end(rbe_ge))) {
 399                                 rbe_ge = rbe;
 400                                 continue;
 401                         }
 402                 } else if (d > 0) {
 403                         /* annotate element greater than the new element. */
 404                         rbe_ge = rbe;
 405                         continue;
 406                 } else if (d < 0) {
 407                         /* annotate element less than the new element. */
 408                         rbe_le = rbe;
 409                         break;
 410                 }
 411         }
 412
 413         /* - new start element matching existing start element: full overlap
 414          *   reported as -EEXIST, cleared by caller if NLM_F_EXCL is not given.
 415          */
 416         if (rbe_ge && !nft_rbtree_cmp(set, new, rbe_ge) &&
 417             nft_rbtree_interval_start(rbe_ge) == nft_rbtree_interval_start(new)) {
 418                 *ext = &rbe_ge->ext;
 419                 return -EEXIST;
 420         }
 421
 422         /* - new end element matching existing end element: full overlap
 423          *   reported as -EEXIST, cleared by caller if NLM_F_EXCL is not given.
 424          */
 425         if (rbe_le && !nft_rbtree_cmp(set, new, rbe_le) &&
 426             nft_rbtree_interval_end(rbe_le) == nft_rbtree_interval_end(new)) {
 427                 *ext = &rbe_le->ext;
 428                 return -EEXIST;
 429         }
 430
 431         /* - new start element with existing closest, less or equal key value
 432          *   being a start element: partial overlap, reported as -ENOTEMPTY.
 433          *   Anonymous sets allow for two consecutive start element since they
 434          *   are constant, skip them to avoid bogus overlap reports.
 435          */
 436         if (!nft_set_is_anonymous(set) && rbe_le &&
 437             nft_rbtree_interval_start(rbe_le) && nft_rbtree_interval_start(new))
 438                 return -ENOTEMPTY;
 439
 440         /* - new end element with existing closest, less or equal key value
 441          *   being a end element: partial overlap, reported as -ENOTEMPTY.
 442          */
 443         if (rbe_le &&
 444             nft_rbtree_interval_end(rbe_le) && nft_rbtree_interval_end(new))
 445                 return -ENOTEMPTY;
 446
 447         /* - new end element with existing closest, greater or equal key value
 448          *   being an end element: partial overlap, reported as -ENOTEMPTY
 449          */
 450         if (rbe_ge &&
 451             nft_rbtree_interval_end(rbe_ge) && nft_rbtree_interval_end(new))
 452                 return -ENOTEMPTY;
 453
 454         /* Accepted element: pick insertion point depending on key value */
 455         parent = NULL;
 456         p = &priv->root.rb_node;
 457         while (*p != NULL) {
 458                 parent = *p;
 459                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
 460                 d = nft_rbtree_cmp(set, rbe, new);
 461
 462                 if (d < 0)
 463                         p = &parent->rb_left;
 464                 else if (d > 0)
 465                         p = &parent->rb_right;
 466                 else if (nft_rbtree_interval_end(rbe))
 467                         p = &parent->rb_left;
 468                 else
 469                         p = &parent->rb_right;
 470         }
 471
 472         rb_link_node_rcu(&new->node, parent, p);
 473         rb_insert_color(&new->node, &priv->root);
 474         return 0;
 475 }
 476
 477 static int nft_rbtree_insert(const struct net *net, const struct nft_set *set,
 478                              const struct nft_set_elem *elem,
 479                              struct nft_set_ext **ext)
 480 {
 481         struct nft_rbtree *priv = nft_set_priv(set);
 482         struct nft_rbtree_elem *rbe = elem->priv;
 483         int err;
 484
 485         write_lock_bh(&priv->lock);
 486         write_seqcount_begin(&priv->count);
 487         err = __nft_rbtree_insert(net, set, rbe, ext);
 488         write_seqcount_end(&priv->count);
 489         write_unlock_bh(&priv->lock);
 490
 491         return err;
 492 }
 493
 494 static void nft_rbtree_remove(const struct net *net,
 495                               const struct nft_set *set,
 496                               const struct nft_set_elem *elem)
 497 {
 498         struct nft_rbtree *priv = nft_set_priv(set);
 499         struct nft_rbtree_elem *rbe = elem->priv;
 500
 501         write_lock_bh(&priv->lock);
 502         write_seqcount_begin(&priv->count);
 503         rb_erase(&rbe->node, &priv->root);
 504         write_seqcount_end(&priv->count);
 505         write_unlock_bh(&priv->lock);
 506 }
 507
 508 static void nft_rbtree_activate(const struct net *net,
 509                                 const struct nft_set *set,
 510                                 const struct nft_set_elem *elem)
 511 {
 512         struct nft_rbtree_elem *rbe = elem->priv;
 513
 514         nft_set_elem_change_active(net, set, &rbe->ext);
 515 }
 516
 517 static bool nft_rbtree_flush(const struct net *net,
 518                              const struct nft_set *set, void *priv)
 519 {
 520         struct nft_rbtree_elem *rbe = priv;
 521
 522         nft_set_elem_change_active(net, set, &rbe->ext);
 523
 524         return true;
 525 }
 526
 527 static void *nft_rbtree_deactivate(const struct net *net,
 528                                    const struct nft_set *set,
 529                                    const struct nft_set_elem *elem)
 530 {
 531         const struct nft_rbtree *priv = nft_set_priv(set);
 532         const struct rb_node *parent = priv->root.rb_node;
 533         struct nft_rbtree_elem *rbe, *this = elem->priv;
 534         u8 genmask = nft_genmask_next(net);
 535         int d;
 536
 537         while (parent != NULL) {
 538                 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
 539
 540                 d = memcmp(nft_set_ext_key(&rbe->ext), &elem->key.val,
 541                                            set->klen);
 542                 if (d < 0)
 543                         parent = parent->rb_left;
 544                 else if (d > 0)
 545                         parent = parent->rb_right;
 546                 else {
 547                         if (nft_rbtree_interval_end(rbe) &&
 548                             nft_rbtree_interval_start(this)) {
 549                                 parent = parent->rb_left;
 550                                 continue;
 551                         } else if (nft_rbtree_interval_start(rbe) &&
 552                                    nft_rbtree_interval_end(this)) {
 553                                 parent = parent->rb_right;
 554                                 continue;
 555                         } else if (!nft_set_elem_active(&rbe->ext, genmask)) {
 556                                 parent = parent->rb_left;
 557                                 continue;
 558                         }
 559                         nft_rbtree_flush(net, set, rbe);
 560                         return rbe;
 561                 }
 562         }
 563         return NULL;
 564 }
 565
 566 static void nft_rbtree_walk(const struct nft_ctx *ctx,
 567                             struct nft_set *set,
 568                             struct nft_set_iter *iter)
 569 {
 570         struct nft_rbtree *priv = nft_set_priv(set);
 571         struct nft_rbtree_elem *rbe;
 572         struct nft_set_elem elem;
 573         struct rb_node *node;
 574
 575         read_lock_bh(&priv->lock);
 576         for (node = rb_first(&priv->root); node != NULL; node = rb_next(node)) {
 577                 rbe = rb_entry(node, struct nft_rbtree_elem, node);
 578
 579                 if (iter->count < iter->skip)
 580                         goto cont;
 581                 if (!nft_set_elem_active(&rbe->ext, iter->genmask))
 582                         goto cont;
 583
 584                 elem.priv = rbe;
 585
 586                 iter->err = iter->fn(ctx, set, iter, &elem);
 587                 if (iter->err < 0) {
 588                         read_unlock_bh(&priv->lock);
 589                         return;
 590                 }
 591 cont:
 592                 iter->count++;
 593         }
 594         read_unlock_bh(&priv->lock);
 595 }
 596
 597 static void nft_rbtree_gc(struct work_struct *work)
 598 {
 599         struct nft_rbtree_elem *rbe, *rbe_end = NULL;
 600         struct nftables_pernet *nft_net;
 601         struct nft_rbtree *priv;
 602         struct nft_trans_gc *gc;
 603         struct rb_node *node;
 604         struct nft_set *set;
 605         unsigned int gc_seq;
 606         struct net *net;
 607
 608         priv = container_of(work, struct nft_rbtree, gc_work.work);
 609         set  = nft_set_container_of(priv);
 610         net  = read_pnet(&set->net);
 611         nft_net = nft_pernet(net);
 612         gc_seq  = READ_ONCE(nft_net->gc_seq);
 613
 614         if (nft_set_gc_is_pending(set))
 615                 goto done;
 616
 617         gc = nft_trans_gc_alloc(set, gc_seq, GFP_KERNEL);
 618         if (!gc)
 619                 goto done;
 620
 621         write_lock_bh(&priv->lock);
 622         write_seqcount_begin(&priv->count);
 623         for (node = rb_first(&priv->root); node != NULL; node = rb_next(node)) {
 624
 625                 /* Ruleset has been updated, try later. */
 626                 if (READ_ONCE(nft_net->gc_seq) != gc_seq) {
 627                         nft_trans_gc_destroy(gc);
 628                         gc = NULL;
 629                         goto try_later;
 630                 }
 631
 632                 rbe = rb_entry(node, struct nft_rbtree_elem, node);
 633
 634                 if (nft_set_elem_is_dead(&rbe->ext))
 635                         goto dead_elem;
 636
 637                 /* elements are reversed in the rbtree for historical reasons,
 638                  * from highest to lowest value, that is why end element is
 639                  * always visited before the start element.
 640                  */
 641                 if (nft_rbtree_interval_end(rbe)) {
 642                         rbe_end = rbe;
 643                         continue;
 644                 }
 645                 if (!nft_set_elem_expired(&rbe->ext))
 646                         continue;
 647
 648                 nft_set_elem_dead(&rbe->ext);
 649
 650                 if (!rbe_end)
 651                         continue;
 652
 653                 nft_set_elem_dead(&rbe_end->ext);
 654
 655                 gc = nft_trans_gc_queue_async(gc, gc_seq, GFP_ATOMIC);
 656                 if (!gc)
 657                         goto try_later;
 658
 659                 nft_trans_gc_elem_add(gc, rbe_end);
 660                 rbe_end = NULL;
 661 dead_elem:
 662                 gc = nft_trans_gc_queue_async(gc, gc_seq, GFP_ATOMIC);
 663                 if (!gc)
 664                         goto try_later;
 665
 666                 nft_trans_gc_elem_add(gc, rbe);
 667         }
 668
 669         gc = nft_trans_gc_catchall(gc, gc_seq);
 670
 671 try_later:
 672         write_seqcount_end(&priv->count);
 673         write_unlock_bh(&priv->lock);
 674
 675         if (gc)
 676                 nft_trans_gc_queue_async_done(gc);
 677 done:
 678         queue_delayed_work(system_power_efficient_wq, &priv->gc_work,
 679                            nft_set_gc_interval(set));
 680 }
 681
 682 static u64 nft_rbtree_privsize(const struct nlattr * const nla[],
 683                                const struct nft_set_desc *desc)
 684 {
 685         return sizeof(struct nft_rbtree);
 686 }
 687
 688 static int nft_rbtree_init(const struct nft_set *set,
 689                            const struct nft_set_desc *desc,
 690                            const struct nlattr * const nla[])
 691 {
 692         struct nft_rbtree *priv = nft_set_priv(set);
 693
 694         rwlock_init(&priv->lock);
 695         seqcount_rwlock_init(&priv->count, &priv->lock);
 696         priv->root = RB_ROOT;
 697
 698         INIT_DEFERRABLE_WORK(&priv->gc_work, nft_rbtree_gc);
 699         if (set->flags & NFT_SET_TIMEOUT)
 700                 queue_delayed_work(system_power_efficient_wq, &priv->gc_work,
 701                                    nft_set_gc_interval(set));
 702
 703         return 0;
 704 }
 705
 706 static void nft_rbtree_destroy(const struct nft_ctx *ctx,
 707                                const struct nft_set *set)
 708 {
 709         struct nft_rbtree *priv = nft_set_priv(set);
 710         struct nft_rbtree_elem *rbe;
 711         struct rb_node *node;
 712
 713         cancel_delayed_work_sync(&priv->gc_work);
 714         rcu_barrier();
 715         while ((node = priv->root.rb_node) != NULL) {
 716                 rb_erase(node, &priv->root);
 717                 rbe = rb_entry(node, struct nft_rbtree_elem, node);
 718                 nf_tables_set_elem_destroy(ctx, set, rbe);
 719         }
 720 }
 721
 722 static bool nft_rbtree_estimate(const struct nft_set_desc *desc, u32 features,
 723                                 struct nft_set_estimate *est)
 724 {
 725         if (desc->field_count > 1)
 726                 return false;
 727
 728         if (desc->size)
 729                 est->size = sizeof(struct nft_rbtree) +
 730                             desc->size * sizeof(struct nft_rbtree_elem);
 731         else
 732                 est->size = ~0;
 733
 734         est->lookup = NFT_SET_CLASS_O_LOG_N;
 735         est->space  = NFT_SET_CLASS_O_N;
 736
 737         return true;
 738 }
 739
 740 const struct nft_set_type nft_set_rbtree_type = {
 741         .features       = NFT_SET_INTERVAL | NFT_SET_MAP | NFT_SET_OBJECT | NFT_SET_TIMEOUT,
 742         .ops            = {
 743                 .privsize       = nft_rbtree_privsize,
 744                 .elemsize       = offsetof(struct nft_rbtree_elem, ext),
 745                 .estimate       = nft_rbtree_estimate,
 746                 .init           = nft_rbtree_init,
 747                 .destroy        = nft_rbtree_destroy,
 748                 .insert         = nft_rbtree_insert,
 749                 .remove         = nft_rbtree_remove,
 750                 .deactivate     = nft_rbtree_deactivate,
 751                 .flush          = nft_rbtree_flush,
 752                 .activate       = nft_rbtree_activate,
 753                 .lookup         = nft_rbtree_lookup,
 754                 .walk           = nft_rbtree_walk,
 755                 .get            = nft_rbtree_get,
 756         },
 757 };