Commit | Line | Data |
---|---|---|
20a69341 PM |
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 | struct rb_root root; | |
9b7e26ae FW |
23 | rwlock_t lock; |
24 | seqcount_t count; | |
8d8540c4 | 25 | struct delayed_work gc_work; |
20a69341 PM |
26 | }; |
27 | ||
28 | struct nft_rbtree_elem { | |
29 | struct rb_node node; | |
fe2811eb | 30 | struct nft_set_ext ext; |
20a69341 PM |
31 | }; |
32 | ||
ef1d20e0 PNA |
33 | static bool nft_rbtree_interval_end(const struct nft_rbtree_elem *rbe) |
34 | { | |
35 | return nft_set_ext_exists(&rbe->ext, NFT_SET_EXT_FLAGS) && | |
36 | (*nft_set_ext_flags(&rbe->ext) & NFT_SET_ELEM_INTERVAL_END); | |
37 | } | |
cc02e457 | 38 | |
e701001e PNA |
39 | static bool nft_rbtree_equal(const struct nft_set *set, const void *this, |
40 | const struct nft_rbtree_elem *interval) | |
41 | { | |
42 | return memcmp(this, nft_set_ext_key(&interval->ext), set->klen) == 0; | |
43 | } | |
44 | ||
9b7e26ae FW |
45 | static bool __nft_rbtree_lookup(const struct net *net, const struct nft_set *set, |
46 | const u32 *key, const struct nft_set_ext **ext, | |
47 | unsigned int seq) | |
20a69341 | 48 | { |
03e5fd0e | 49 | struct nft_rbtree *priv = nft_set_priv(set); |
20a69341 | 50 | const struct nft_rbtree_elem *rbe, *interval = NULL; |
42a55769 | 51 | u8 genmask = nft_genmask_cur(net); |
16c45eda | 52 | const struct rb_node *parent; |
e701001e | 53 | const void *this; |
20a69341 PM |
54 | int d; |
55 | ||
9b7e26ae | 56 | parent = rcu_dereference_raw(priv->root.rb_node); |
20a69341 | 57 | while (parent != NULL) { |
9b7e26ae FW |
58 | if (read_seqcount_retry(&priv->count, seq)) |
59 | return false; | |
60 | ||
20a69341 PM |
61 | rbe = rb_entry(parent, struct nft_rbtree_elem, node); |
62 | ||
e701001e PNA |
63 | this = nft_set_ext_key(&rbe->ext); |
64 | d = memcmp(this, key, set->klen); | |
20a69341 | 65 | if (d < 0) { |
9b7e26ae | 66 | parent = rcu_dereference_raw(parent->rb_left); |
f9121355 PNA |
67 | if (interval && |
68 | nft_rbtree_equal(set, this, interval) && | |
82e20b44 | 69 | nft_rbtree_interval_end(rbe) && |
f9121355 | 70 | !nft_rbtree_interval_end(interval)) |
e701001e | 71 | continue; |
20a69341 PM |
72 | interval = rbe; |
73 | } else if (d > 0) | |
9b7e26ae | 74 | parent = rcu_dereference_raw(parent->rb_right); |
20a69341 | 75 | else { |
cc02e457 | 76 | if (!nft_set_elem_active(&rbe->ext, genmask)) { |
9b7e26ae | 77 | parent = rcu_dereference_raw(parent->rb_left); |
cc02e457 PM |
78 | continue; |
79 | } | |
ef1d20e0 | 80 | if (nft_rbtree_interval_end(rbe)) |
20a69341 | 81 | goto out; |
b2832dd6 PM |
82 | |
83 | *ext = &rbe->ext; | |
20a69341 PM |
84 | return true; |
85 | } | |
86 | } | |
87 | ||
c1eda3c6 PNA |
88 | if (set->flags & NFT_SET_INTERVAL && interval != NULL && |
89 | nft_set_elem_active(&interval->ext, genmask) && | |
90 | !nft_rbtree_interval_end(interval)) { | |
c1eda3c6 PNA |
91 | *ext = &interval->ext; |
92 | return true; | |
20a69341 PM |
93 | } |
94 | out: | |
95 | return false; | |
96 | } | |
97 | ||
9b7e26ae FW |
98 | static bool nft_rbtree_lookup(const struct net *net, const struct nft_set *set, |
99 | const u32 *key, const struct nft_set_ext **ext) | |
100 | { | |
101 | struct nft_rbtree *priv = nft_set_priv(set); | |
102 | unsigned int seq = read_seqcount_begin(&priv->count); | |
103 | bool ret; | |
104 | ||
105 | ret = __nft_rbtree_lookup(net, set, key, ext, seq); | |
106 | if (ret || !read_seqcount_retry(&priv->count, seq)) | |
107 | return ret; | |
108 | ||
109 | read_lock_bh(&priv->lock); | |
110 | seq = read_seqcount_begin(&priv->count); | |
111 | ret = __nft_rbtree_lookup(net, set, key, ext, seq); | |
112 | read_unlock_bh(&priv->lock); | |
113 | ||
114 | return ret; | |
115 | } | |
116 | ||
ba0e4d99 PNA |
117 | static bool __nft_rbtree_get(const struct net *net, const struct nft_set *set, |
118 | const u32 *key, struct nft_rbtree_elem **elem, | |
119 | unsigned int seq, unsigned int flags, u8 genmask) | |
120 | { | |
121 | struct nft_rbtree_elem *rbe, *interval = NULL; | |
122 | struct nft_rbtree *priv = nft_set_priv(set); | |
123 | const struct rb_node *parent; | |
124 | const void *this; | |
125 | int d; | |
126 | ||
127 | parent = rcu_dereference_raw(priv->root.rb_node); | |
128 | while (parent != NULL) { | |
129 | if (read_seqcount_retry(&priv->count, seq)) | |
130 | return false; | |
131 | ||
132 | rbe = rb_entry(parent, struct nft_rbtree_elem, node); | |
133 | ||
134 | this = nft_set_ext_key(&rbe->ext); | |
135 | d = memcmp(this, key, set->klen); | |
136 | if (d < 0) { | |
137 | parent = rcu_dereference_raw(parent->rb_left); | |
138 | interval = rbe; | |
139 | } else if (d > 0) { | |
140 | parent = rcu_dereference_raw(parent->rb_right); | |
141 | } else { | |
142 | if (!nft_set_elem_active(&rbe->ext, genmask)) | |
143 | parent = rcu_dereference_raw(parent->rb_left); | |
144 | ||
145 | if (!nft_set_ext_exists(&rbe->ext, NFT_SET_EXT_FLAGS) || | |
146 | (*nft_set_ext_flags(&rbe->ext) & NFT_SET_ELEM_INTERVAL_END) == | |
147 | (flags & NFT_SET_ELEM_INTERVAL_END)) { | |
148 | *elem = rbe; | |
149 | return true; | |
150 | } | |
151 | return false; | |
152 | } | |
153 | } | |
154 | ||
155 | if (set->flags & NFT_SET_INTERVAL && interval != NULL && | |
156 | nft_set_elem_active(&interval->ext, genmask) && | |
157 | !nft_rbtree_interval_end(interval)) { | |
158 | *elem = interval; | |
159 | return true; | |
160 | } | |
161 | ||
162 | return false; | |
163 | } | |
164 | ||
165 | static void *nft_rbtree_get(const struct net *net, const struct nft_set *set, | |
166 | const struct nft_set_elem *elem, unsigned int flags) | |
167 | { | |
168 | struct nft_rbtree *priv = nft_set_priv(set); | |
169 | unsigned int seq = read_seqcount_begin(&priv->count); | |
170 | struct nft_rbtree_elem *rbe = ERR_PTR(-ENOENT); | |
171 | const u32 *key = (const u32 *)&elem->key.val; | |
172 | u8 genmask = nft_genmask_cur(net); | |
173 | bool ret; | |
174 | ||
175 | ret = __nft_rbtree_get(net, set, key, &rbe, seq, flags, genmask); | |
176 | if (ret || !read_seqcount_retry(&priv->count, seq)) | |
177 | return rbe; | |
178 | ||
179 | read_lock_bh(&priv->lock); | |
180 | seq = read_seqcount_begin(&priv->count); | |
181 | ret = __nft_rbtree_get(net, set, key, &rbe, seq, flags, genmask); | |
182 | if (!ret) | |
183 | rbe = ERR_PTR(-ENOENT); | |
184 | read_unlock_bh(&priv->lock); | |
185 | ||
186 | return rbe; | |
187 | } | |
188 | ||
42a55769 | 189 | static int __nft_rbtree_insert(const struct net *net, const struct nft_set *set, |
c016c7e4 PNA |
190 | struct nft_rbtree_elem *new, |
191 | struct nft_set_ext **ext) | |
20a69341 PM |
192 | { |
193 | struct nft_rbtree *priv = nft_set_priv(set); | |
42a55769 | 194 | u8 genmask = nft_genmask_next(net); |
20a69341 PM |
195 | struct nft_rbtree_elem *rbe; |
196 | struct rb_node *parent, **p; | |
197 | int d; | |
198 | ||
199 | parent = NULL; | |
200 | p = &priv->root.rb_node; | |
201 | while (*p != NULL) { | |
202 | parent = *p; | |
203 | rbe = rb_entry(parent, struct nft_rbtree_elem, node); | |
e562d860 PM |
204 | d = memcmp(nft_set_ext_key(&rbe->ext), |
205 | nft_set_ext_key(&new->ext), | |
206 | set->klen); | |
20a69341 PM |
207 | if (d < 0) |
208 | p = &parent->rb_left; | |
209 | else if (d > 0) | |
210 | p = &parent->rb_right; | |
cc02e457 | 211 | else { |
d2df92e9 PNA |
212 | if (nft_rbtree_interval_end(rbe) && |
213 | !nft_rbtree_interval_end(new)) { | |
214 | p = &parent->rb_left; | |
215 | } else if (!nft_rbtree_interval_end(rbe) && | |
216 | nft_rbtree_interval_end(new)) { | |
217 | p = &parent->rb_right; | |
218 | } else if (nft_set_elem_active(&rbe->ext, genmask)) { | |
219 | *ext = &rbe->ext; | |
220 | return -EEXIST; | |
221 | } else { | |
222 | p = &parent->rb_left; | |
e701001e | 223 | } |
cc02e457 | 224 | } |
20a69341 | 225 | } |
9b7e26ae | 226 | rb_link_node_rcu(&new->node, parent, p); |
20a69341 PM |
227 | rb_insert_color(&new->node, &priv->root); |
228 | return 0; | |
229 | } | |
230 | ||
42a55769 | 231 | static int nft_rbtree_insert(const struct net *net, const struct nft_set *set, |
c016c7e4 PNA |
232 | const struct nft_set_elem *elem, |
233 | struct nft_set_ext **ext) | |
20a69341 | 234 | { |
03e5fd0e | 235 | struct nft_rbtree *priv = nft_set_priv(set); |
fe2811eb | 236 | struct nft_rbtree_elem *rbe = elem->priv; |
20a69341 PM |
237 | int err; |
238 | ||
03e5fd0e | 239 | write_lock_bh(&priv->lock); |
9b7e26ae | 240 | write_seqcount_begin(&priv->count); |
c016c7e4 | 241 | err = __nft_rbtree_insert(net, set, rbe, ext); |
9b7e26ae | 242 | write_seqcount_end(&priv->count); |
03e5fd0e | 243 | write_unlock_bh(&priv->lock); |
fe2811eb | 244 | |
20a69341 PM |
245 | return err; |
246 | } | |
247 | ||
5cb82a38 PNA |
248 | static void nft_rbtree_remove(const struct net *net, |
249 | const struct nft_set *set, | |
20a69341 PM |
250 | const struct nft_set_elem *elem) |
251 | { | |
252 | struct nft_rbtree *priv = nft_set_priv(set); | |
cc02e457 | 253 | struct nft_rbtree_elem *rbe = elem->priv; |
20a69341 | 254 | |
03e5fd0e | 255 | write_lock_bh(&priv->lock); |
9b7e26ae | 256 | write_seqcount_begin(&priv->count); |
20a69341 | 257 | rb_erase(&rbe->node, &priv->root); |
9b7e26ae | 258 | write_seqcount_end(&priv->count); |
03e5fd0e | 259 | write_unlock_bh(&priv->lock); |
20a69341 PM |
260 | } |
261 | ||
42a55769 PNA |
262 | static void nft_rbtree_activate(const struct net *net, |
263 | const struct nft_set *set, | |
cc02e457 PM |
264 | const struct nft_set_elem *elem) |
265 | { | |
266 | struct nft_rbtree_elem *rbe = elem->priv; | |
267 | ||
42a55769 | 268 | nft_set_elem_change_active(net, set, &rbe->ext); |
8d8540c4 | 269 | nft_set_elem_clear_busy(&rbe->ext); |
cc02e457 PM |
270 | } |
271 | ||
1ba1c414 PNA |
272 | static bool nft_rbtree_flush(const struct net *net, |
273 | const struct nft_set *set, void *priv) | |
37df5301 PNA |
274 | { |
275 | struct nft_rbtree_elem *rbe = priv; | |
276 | ||
8d8540c4 PNA |
277 | if (!nft_set_elem_mark_busy(&rbe->ext) || |
278 | !nft_is_active(net, &rbe->ext)) { | |
279 | nft_set_elem_change_active(net, set, &rbe->ext); | |
280 | return true; | |
281 | } | |
282 | return false; | |
37df5301 PNA |
283 | } |
284 | ||
42a55769 PNA |
285 | static void *nft_rbtree_deactivate(const struct net *net, |
286 | const struct nft_set *set, | |
cc02e457 | 287 | const struct nft_set_elem *elem) |
20a69341 PM |
288 | { |
289 | const struct nft_rbtree *priv = nft_set_priv(set); | |
290 | const struct rb_node *parent = priv->root.rb_node; | |
e701001e | 291 | struct nft_rbtree_elem *rbe, *this = elem->priv; |
42a55769 | 292 | u8 genmask = nft_genmask_next(net); |
20a69341 PM |
293 | int d; |
294 | ||
295 | while (parent != NULL) { | |
296 | rbe = rb_entry(parent, struct nft_rbtree_elem, node); | |
297 | ||
7d740264 PM |
298 | d = memcmp(nft_set_ext_key(&rbe->ext), &elem->key.val, |
299 | set->klen); | |
20a69341 PM |
300 | if (d < 0) |
301 | parent = parent->rb_left; | |
302 | else if (d > 0) | |
303 | parent = parent->rb_right; | |
304 | else { | |
cc02e457 PM |
305 | if (!nft_set_elem_active(&rbe->ext, genmask)) { |
306 | parent = parent->rb_left; | |
307 | continue; | |
308 | } | |
e701001e PNA |
309 | if (nft_rbtree_interval_end(rbe) && |
310 | !nft_rbtree_interval_end(this)) { | |
311 | parent = parent->rb_left; | |
312 | continue; | |
313 | } else if (!nft_rbtree_interval_end(rbe) && | |
314 | nft_rbtree_interval_end(this)) { | |
315 | parent = parent->rb_right; | |
316 | continue; | |
317 | } | |
1ba1c414 | 318 | nft_rbtree_flush(net, set, rbe); |
cc02e457 | 319 | return rbe; |
20a69341 PM |
320 | } |
321 | } | |
cc02e457 | 322 | return NULL; |
20a69341 PM |
323 | } |
324 | ||
325 | static void nft_rbtree_walk(const struct nft_ctx *ctx, | |
de70185d | 326 | struct nft_set *set, |
20a69341 PM |
327 | struct nft_set_iter *iter) |
328 | { | |
03e5fd0e | 329 | struct nft_rbtree *priv = nft_set_priv(set); |
fe2811eb | 330 | struct nft_rbtree_elem *rbe; |
20a69341 PM |
331 | struct nft_set_elem elem; |
332 | struct rb_node *node; | |
333 | ||
03e5fd0e | 334 | read_lock_bh(&priv->lock); |
20a69341 | 335 | for (node = rb_first(&priv->root); node != NULL; node = rb_next(node)) { |
cc02e457 PM |
336 | rbe = rb_entry(node, struct nft_rbtree_elem, node); |
337 | ||
20a69341 PM |
338 | if (iter->count < iter->skip) |
339 | goto cont; | |
8588ac09 | 340 | if (!nft_set_elem_active(&rbe->ext, iter->genmask)) |
cc02e457 | 341 | goto cont; |
20a69341 | 342 | |
fe2811eb | 343 | elem.priv = rbe; |
20a69341 PM |
344 | |
345 | iter->err = iter->fn(ctx, set, iter, &elem); | |
7632667d | 346 | if (iter->err < 0) { |
03e5fd0e | 347 | read_unlock_bh(&priv->lock); |
20a69341 | 348 | return; |
7632667d | 349 | } |
20a69341 PM |
350 | cont: |
351 | iter->count++; | |
352 | } | |
03e5fd0e | 353 | read_unlock_bh(&priv->lock); |
20a69341 PM |
354 | } |
355 | ||
8d8540c4 PNA |
356 | static void nft_rbtree_gc(struct work_struct *work) |
357 | { | |
358 | struct nft_set_gc_batch *gcb = NULL; | |
359 | struct rb_node *node, *prev = NULL; | |
360 | struct nft_rbtree_elem *rbe; | |
361 | struct nft_rbtree *priv; | |
362 | struct nft_set *set; | |
363 | int i; | |
364 | ||
365 | priv = container_of(work, struct nft_rbtree, gc_work.work); | |
366 | set = nft_set_container_of(priv); | |
367 | ||
368 | write_lock_bh(&priv->lock); | |
369 | write_seqcount_begin(&priv->count); | |
370 | for (node = rb_first(&priv->root); node != NULL; node = rb_next(node)) { | |
371 | rbe = rb_entry(node, struct nft_rbtree_elem, node); | |
372 | ||
373 | if (nft_rbtree_interval_end(rbe)) { | |
374 | prev = node; | |
375 | continue; | |
376 | } | |
377 | if (!nft_set_elem_expired(&rbe->ext)) | |
378 | continue; | |
379 | if (nft_set_elem_mark_busy(&rbe->ext)) | |
380 | continue; | |
381 | ||
382 | gcb = nft_set_gc_batch_check(set, gcb, GFP_ATOMIC); | |
383 | if (!gcb) | |
c293ac95 | 384 | break; |
8d8540c4 PNA |
385 | |
386 | atomic_dec(&set->nelems); | |
387 | nft_set_gc_batch_add(gcb, rbe); | |
388 | ||
389 | if (prev) { | |
390 | rbe = rb_entry(prev, struct nft_rbtree_elem, node); | |
391 | atomic_dec(&set->nelems); | |
392 | nft_set_gc_batch_add(gcb, rbe); | |
c293ac95 | 393 | prev = NULL; |
8d8540c4 PNA |
394 | } |
395 | node = rb_next(node); | |
c293ac95 TY |
396 | if (!node) |
397 | break; | |
8d8540c4 | 398 | } |
8d8540c4 PNA |
399 | if (gcb) { |
400 | for (i = 0; i < gcb->head.cnt; i++) { | |
401 | rbe = gcb->elems[i]; | |
402 | rb_erase(&rbe->node, &priv->root); | |
403 | } | |
404 | } | |
405 | write_seqcount_end(&priv->count); | |
406 | write_unlock_bh(&priv->lock); | |
407 | ||
408 | nft_set_gc_batch_complete(gcb); | |
409 | ||
410 | queue_delayed_work(system_power_efficient_wq, &priv->gc_work, | |
411 | nft_set_gc_interval(set)); | |
412 | } | |
413 | ||
4ef360dd TY |
414 | static u64 nft_rbtree_privsize(const struct nlattr * const nla[], |
415 | const struct nft_set_desc *desc) | |
20a69341 PM |
416 | { |
417 | return sizeof(struct nft_rbtree); | |
418 | } | |
419 | ||
420 | static int nft_rbtree_init(const struct nft_set *set, | |
c50b960c | 421 | const struct nft_set_desc *desc, |
20a69341 PM |
422 | const struct nlattr * const nla[]) |
423 | { | |
424 | struct nft_rbtree *priv = nft_set_priv(set); | |
425 | ||
03e5fd0e | 426 | rwlock_init(&priv->lock); |
9b7e26ae | 427 | seqcount_init(&priv->count); |
20a69341 | 428 | priv->root = RB_ROOT; |
8d8540c4 PNA |
429 | |
430 | INIT_DEFERRABLE_WORK(&priv->gc_work, nft_rbtree_gc); | |
431 | if (set->flags & NFT_SET_TIMEOUT) | |
432 | queue_delayed_work(system_power_efficient_wq, &priv->gc_work, | |
433 | nft_set_gc_interval(set)); | |
434 | ||
20a69341 PM |
435 | return 0; |
436 | } | |
437 | ||
438 | static void nft_rbtree_destroy(const struct nft_set *set) | |
439 | { | |
440 | struct nft_rbtree *priv = nft_set_priv(set); | |
441 | struct nft_rbtree_elem *rbe; | |
442 | struct rb_node *node; | |
443 | ||
8d8540c4 | 444 | cancel_delayed_work_sync(&priv->gc_work); |
c293ac95 | 445 | rcu_barrier(); |
20a69341 PM |
446 | while ((node = priv->root.rb_node) != NULL) { |
447 | rb_erase(node, &priv->root); | |
448 | rbe = rb_entry(node, struct nft_rbtree_elem, node); | |
61f9e292 | 449 | nft_set_elem_destroy(set, rbe, true); |
20a69341 PM |
450 | } |
451 | } | |
452 | ||
c50b960c PM |
453 | static bool nft_rbtree_estimate(const struct nft_set_desc *desc, u32 features, |
454 | struct nft_set_estimate *est) | |
455 | { | |
c50b960c | 456 | if (desc->size) |
080ed636 PNA |
457 | est->size = sizeof(struct nft_rbtree) + |
458 | desc->size * sizeof(struct nft_rbtree_elem); | |
c50b960c | 459 | else |
080ed636 | 460 | est->size = ~0; |
c50b960c | 461 | |
55af753c | 462 | est->lookup = NFT_SET_CLASS_O_LOG_N; |
0b5a7874 | 463 | est->space = NFT_SET_CLASS_O_N; |
c50b960c PM |
464 | |
465 | return true; | |
466 | } | |
467 | ||
e240cd0d | 468 | struct nft_set_type nft_set_rbtree_type __read_mostly = { |
20a69341 | 469 | .owner = THIS_MODULE, |
8d8540c4 | 470 | .features = NFT_SET_INTERVAL | NFT_SET_MAP | NFT_SET_OBJECT | NFT_SET_TIMEOUT, |
71cc0873 PS |
471 | .ops = { |
472 | .privsize = nft_rbtree_privsize, | |
473 | .elemsize = offsetof(struct nft_rbtree_elem, ext), | |
474 | .estimate = nft_rbtree_estimate, | |
475 | .init = nft_rbtree_init, | |
476 | .destroy = nft_rbtree_destroy, | |
477 | .insert = nft_rbtree_insert, | |
478 | .remove = nft_rbtree_remove, | |
479 | .deactivate = nft_rbtree_deactivate, | |
480 | .flush = nft_rbtree_flush, | |
481 | .activate = nft_rbtree_activate, | |
482 | .lookup = nft_rbtree_lookup, | |
483 | .walk = nft_rbtree_walk, | |
484 | .get = nft_rbtree_get, | |
485 | }, | |
20a69341 | 486 | }; |