]>
Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
1ab1457c | 2 | * Linux INET6 implementation |
1da177e4 LT |
3 | * Forwarding Information Database |
4 | * | |
5 | * Authors: | |
1ab1457c | 6 | * Pedro Roque <roque@di.fc.ul.pt> |
1da177e4 | 7 | * |
1da177e4 LT |
8 | * This program is free software; you can redistribute it and/or |
9 | * modify it under the terms of the GNU General Public License | |
10 | * as published by the Free Software Foundation; either version | |
11 | * 2 of the License, or (at your option) any later version. | |
8db46f1d WY |
12 | * |
13 | * Changes: | |
14 | * Yuji SEKIYA @USAGI: Support default route on router node; | |
15 | * remove ip6_null_entry from the top of | |
16 | * routing table. | |
17 | * Ville Nuorvala: Fixed routing subtrees. | |
1da177e4 | 18 | */ |
f3213831 JP |
19 | |
20 | #define pr_fmt(fmt) "IPv6: " fmt | |
21 | ||
1da177e4 LT |
22 | #include <linux/errno.h> |
23 | #include <linux/types.h> | |
24 | #include <linux/net.h> | |
25 | #include <linux/route.h> | |
26 | #include <linux/netdevice.h> | |
27 | #include <linux/in6.h> | |
28 | #include <linux/init.h> | |
c71099ac | 29 | #include <linux/list.h> |
5a0e3ad6 | 30 | #include <linux/slab.h> |
1da177e4 | 31 | |
1da177e4 LT |
32 | #include <net/ipv6.h> |
33 | #include <net/ndisc.h> | |
34 | #include <net/addrconf.h> | |
19e42e45 | 35 | #include <net/lwtunnel.h> |
df77fe4d | 36 | #include <net/fib_notifier.h> |
1da177e4 LT |
37 | |
38 | #include <net/ip6_fib.h> | |
39 | #include <net/ip6_route.h> | |
40 | ||
437de07c | 41 | static struct kmem_cache *fib6_node_kmem __read_mostly; |
1da177e4 | 42 | |
94b2cfe0 HFS |
43 | struct fib6_cleaner { |
44 | struct fib6_walker w; | |
ec7d43c2 | 45 | struct net *net; |
1da177e4 | 46 | int (*func)(struct rt6_info *, void *arg); |
327571cb | 47 | int sernum; |
1da177e4 LT |
48 | void *arg; |
49 | }; | |
50 | ||
1da177e4 LT |
51 | #ifdef CONFIG_IPV6_SUBTREES |
52 | #define FWS_INIT FWS_S | |
1da177e4 LT |
53 | #else |
54 | #define FWS_INIT FWS_L | |
1da177e4 LT |
55 | #endif |
56 | ||
66f5d6ce WW |
57 | static struct rt6_info *fib6_find_prefix(struct net *net, |
58 | struct fib6_table *table, | |
59 | struct fib6_node *fn); | |
60 | static struct fib6_node *fib6_repair_tree(struct net *net, | |
61 | struct fib6_table *table, | |
62 | struct fib6_node *fn); | |
9a03cd8f | 63 | static int fib6_walk(struct net *net, struct fib6_walker *w); |
94b2cfe0 | 64 | static int fib6_walk_continue(struct fib6_walker *w); |
1da177e4 LT |
65 | |
66 | /* | |
67 | * A routing update causes an increase of the serial number on the | |
68 | * affected subtree. This allows for cached routes to be asynchronously | |
69 | * tested when modifications are made to the destination cache as a | |
70 | * result of redirects, path MTU changes, etc. | |
71 | */ | |
72 | ||
5b7c931d DL |
73 | static void fib6_gc_timer_cb(unsigned long arg); |
74 | ||
9a03cd8f MK |
75 | #define FOR_WALKERS(net, w) \ |
76 | list_for_each_entry(w, &(net)->ipv6.fib6_walkers, lh) | |
1da177e4 | 77 | |
9a03cd8f | 78 | static void fib6_walker_link(struct net *net, struct fib6_walker *w) |
90d41122 | 79 | { |
9a03cd8f MK |
80 | write_lock_bh(&net->ipv6.fib6_walker_lock); |
81 | list_add(&w->lh, &net->ipv6.fib6_walkers); | |
82 | write_unlock_bh(&net->ipv6.fib6_walker_lock); | |
90d41122 AB |
83 | } |
84 | ||
9a03cd8f | 85 | static void fib6_walker_unlink(struct net *net, struct fib6_walker *w) |
90d41122 | 86 | { |
9a03cd8f | 87 | write_lock_bh(&net->ipv6.fib6_walker_lock); |
bbef49da | 88 | list_del(&w->lh); |
9a03cd8f | 89 | write_unlock_bh(&net->ipv6.fib6_walker_lock); |
90d41122 | 90 | } |
94b2cfe0 | 91 | |
812918c4 | 92 | static int fib6_new_sernum(struct net *net) |
1da177e4 | 93 | { |
42b18706 HFS |
94 | int new, old; |
95 | ||
96 | do { | |
812918c4 | 97 | old = atomic_read(&net->ipv6.fib6_sernum); |
42b18706 | 98 | new = old < INT_MAX ? old + 1 : 1; |
812918c4 HFS |
99 | } while (atomic_cmpxchg(&net->ipv6.fib6_sernum, |
100 | old, new) != old); | |
42b18706 | 101 | return new; |
1da177e4 LT |
102 | } |
103 | ||
327571cb HFS |
104 | enum { |
105 | FIB6_NO_SERNUM_CHANGE = 0, | |
106 | }; | |
107 | ||
180ca444 WW |
108 | void fib6_update_sernum(struct rt6_info *rt) |
109 | { | |
110 | struct fib6_table *table = rt->rt6i_table; | |
111 | struct net *net = dev_net(rt->dst.dev); | |
112 | struct fib6_node *fn; | |
113 | ||
66f5d6ce | 114 | spin_lock_bh(&table->tb6_lock); |
180ca444 WW |
115 | fn = rcu_dereference_protected(rt->rt6i_node, |
116 | lockdep_is_held(&table->tb6_lock)); | |
117 | if (fn) | |
118 | fn->fn_sernum = fib6_new_sernum(net); | |
66f5d6ce | 119 | spin_unlock_bh(&table->tb6_lock); |
180ca444 WW |
120 | } |
121 | ||
1da177e4 LT |
122 | /* |
123 | * Auxiliary address test functions for the radix tree. | |
124 | * | |
1ab1457c | 125 | * These assume a 32bit processor (although it will work on |
1da177e4 LT |
126 | * 64bit processors) |
127 | */ | |
128 | ||
129 | /* | |
130 | * test bit | |
131 | */ | |
02cdce53 YH |
132 | #if defined(__LITTLE_ENDIAN) |
133 | # define BITOP_BE32_SWIZZLE (0x1F & ~7) | |
134 | #else | |
135 | # define BITOP_BE32_SWIZZLE 0 | |
136 | #endif | |
1da177e4 | 137 | |
94b2cfe0 | 138 | static __be32 addr_bit_set(const void *token, int fn_bit) |
1da177e4 | 139 | { |
b71d1d42 | 140 | const __be32 *addr = token; |
02cdce53 YH |
141 | /* |
142 | * Here, | |
8db46f1d | 143 | * 1 << ((~fn_bit ^ BITOP_BE32_SWIZZLE) & 0x1f) |
02cdce53 YH |
144 | * is optimized version of |
145 | * htonl(1 << ((~fn_bit)&0x1F)) | |
146 | * See include/asm-generic/bitops/le.h. | |
147 | */ | |
0eae88f3 ED |
148 | return (__force __be32)(1 << ((~fn_bit ^ BITOP_BE32_SWIZZLE) & 0x1f)) & |
149 | addr[fn_bit >> 5]; | |
1da177e4 LT |
150 | } |
151 | ||
81eb8447 | 152 | static struct fib6_node *node_alloc(struct net *net) |
1da177e4 LT |
153 | { |
154 | struct fib6_node *fn; | |
155 | ||
c3762229 | 156 | fn = kmem_cache_zalloc(fib6_node_kmem, GFP_ATOMIC); |
81eb8447 WW |
157 | if (fn) |
158 | net->ipv6.rt6_stats->fib_nodes++; | |
1da177e4 LT |
159 | |
160 | return fn; | |
161 | } | |
162 | ||
81eb8447 | 163 | static void node_free_immediate(struct net *net, struct fib6_node *fn) |
c5cff856 WW |
164 | { |
165 | kmem_cache_free(fib6_node_kmem, fn); | |
81eb8447 | 166 | net->ipv6.rt6_stats->fib_nodes--; |
c5cff856 WW |
167 | } |
168 | ||
169 | static void node_free_rcu(struct rcu_head *head) | |
1da177e4 | 170 | { |
c5cff856 WW |
171 | struct fib6_node *fn = container_of(head, struct fib6_node, rcu); |
172 | ||
1da177e4 LT |
173 | kmem_cache_free(fib6_node_kmem, fn); |
174 | } | |
175 | ||
81eb8447 | 176 | static void node_free(struct net *net, struct fib6_node *fn) |
c5cff856 WW |
177 | { |
178 | call_rcu(&fn->rcu, node_free_rcu); | |
81eb8447 | 179 | net->ipv6.rt6_stats->fib_nodes--; |
c5cff856 WW |
180 | } |
181 | ||
a460aa83 | 182 | void rt6_free_pcpu(struct rt6_info *non_pcpu_rt) |
d52d3997 MKL |
183 | { |
184 | int cpu; | |
185 | ||
186 | if (!non_pcpu_rt->rt6i_pcpu) | |
187 | return; | |
188 | ||
189 | for_each_possible_cpu(cpu) { | |
190 | struct rt6_info **ppcpu_rt; | |
191 | struct rt6_info *pcpu_rt; | |
192 | ||
193 | ppcpu_rt = per_cpu_ptr(non_pcpu_rt->rt6i_pcpu, cpu); | |
194 | pcpu_rt = *ppcpu_rt; | |
195 | if (pcpu_rt) { | |
9514528d | 196 | dst_dev_put(&pcpu_rt->dst); |
1cfb71ee | 197 | dst_release(&pcpu_rt->dst); |
d52d3997 MKL |
198 | *ppcpu_rt = NULL; |
199 | } | |
200 | } | |
201 | } | |
a460aa83 | 202 | EXPORT_SYMBOL_GPL(rt6_free_pcpu); |
1da177e4 | 203 | |
ba1cc08d SD |
204 | static void fib6_free_table(struct fib6_table *table) |
205 | { | |
206 | inetpeer_invalidate_tree(&table->tb6_peers); | |
207 | kfree(table); | |
208 | } | |
209 | ||
58f09b78 | 210 | static void fib6_link_table(struct net *net, struct fib6_table *tb) |
1b43af54 PM |
211 | { |
212 | unsigned int h; | |
213 | ||
375216ad TG |
214 | /* |
215 | * Initialize table lock at a single place to give lockdep a key, | |
216 | * tables aren't visible prior to being linked to the list. | |
217 | */ | |
66f5d6ce | 218 | spin_lock_init(&tb->tb6_lock); |
a33bc5c1 | 219 | h = tb->tb6_id & (FIB6_TABLE_HASHSZ - 1); |
1b43af54 PM |
220 | |
221 | /* | |
222 | * No protection necessary, this is the only list mutatation | |
223 | * operation, tables never disappear once they exist. | |
224 | */ | |
58f09b78 | 225 | hlist_add_head_rcu(&tb->tb6_hlist, &net->ipv6.fib_table_hash[h]); |
1b43af54 | 226 | } |
c71099ac | 227 | |
1b43af54 | 228 | #ifdef CONFIG_IPV6_MULTIPLE_TABLES |
e0b85590 | 229 | |
8ed67789 | 230 | static struct fib6_table *fib6_alloc_table(struct net *net, u32 id) |
c71099ac TG |
231 | { |
232 | struct fib6_table *table; | |
233 | ||
234 | table = kzalloc(sizeof(*table), GFP_ATOMIC); | |
507c9b1e | 235 | if (table) { |
c71099ac | 236 | table->tb6_id = id; |
66f5d6ce WW |
237 | rcu_assign_pointer(table->tb6_root.leaf, |
238 | net->ipv6.ip6_null_entry); | |
c71099ac | 239 | table->tb6_root.fn_flags = RTN_ROOT | RTN_TL_ROOT | RTN_RTINFO; |
8e773277 | 240 | inet_peer_base_init(&table->tb6_peers); |
c71099ac TG |
241 | } |
242 | ||
243 | return table; | |
244 | } | |
245 | ||
58f09b78 | 246 | struct fib6_table *fib6_new_table(struct net *net, u32 id) |
c71099ac TG |
247 | { |
248 | struct fib6_table *tb; | |
249 | ||
250 | if (id == 0) | |
251 | id = RT6_TABLE_MAIN; | |
58f09b78 | 252 | tb = fib6_get_table(net, id); |
c71099ac TG |
253 | if (tb) |
254 | return tb; | |
255 | ||
8ed67789 | 256 | tb = fib6_alloc_table(net, id); |
507c9b1e | 257 | if (tb) |
58f09b78 | 258 | fib6_link_table(net, tb); |
c71099ac TG |
259 | |
260 | return tb; | |
261 | } | |
b3b4663c | 262 | EXPORT_SYMBOL_GPL(fib6_new_table); |
c71099ac | 263 | |
58f09b78 | 264 | struct fib6_table *fib6_get_table(struct net *net, u32 id) |
c71099ac TG |
265 | { |
266 | struct fib6_table *tb; | |
58f09b78 | 267 | struct hlist_head *head; |
c71099ac TG |
268 | unsigned int h; |
269 | ||
270 | if (id == 0) | |
271 | id = RT6_TABLE_MAIN; | |
a33bc5c1 | 272 | h = id & (FIB6_TABLE_HASHSZ - 1); |
c71099ac | 273 | rcu_read_lock(); |
58f09b78 | 274 | head = &net->ipv6.fib_table_hash[h]; |
b67bfe0d | 275 | hlist_for_each_entry_rcu(tb, head, tb6_hlist) { |
c71099ac TG |
276 | if (tb->tb6_id == id) { |
277 | rcu_read_unlock(); | |
278 | return tb; | |
279 | } | |
280 | } | |
281 | rcu_read_unlock(); | |
282 | ||
283 | return NULL; | |
284 | } | |
c4850687 | 285 | EXPORT_SYMBOL_GPL(fib6_get_table); |
c71099ac | 286 | |
2c8c1e72 | 287 | static void __net_init fib6_tables_init(struct net *net) |
c71099ac | 288 | { |
58f09b78 DL |
289 | fib6_link_table(net, net->ipv6.fib6_main_tbl); |
290 | fib6_link_table(net, net->ipv6.fib6_local_tbl); | |
c71099ac | 291 | } |
c71099ac TG |
292 | #else |
293 | ||
58f09b78 | 294 | struct fib6_table *fib6_new_table(struct net *net, u32 id) |
c71099ac | 295 | { |
58f09b78 | 296 | return fib6_get_table(net, id); |
c71099ac TG |
297 | } |
298 | ||
58f09b78 | 299 | struct fib6_table *fib6_get_table(struct net *net, u32 id) |
c71099ac | 300 | { |
58f09b78 | 301 | return net->ipv6.fib6_main_tbl; |
c71099ac TG |
302 | } |
303 | ||
4c9483b2 | 304 | struct dst_entry *fib6_rule_lookup(struct net *net, struct flowi6 *fl6, |
58f09b78 | 305 | int flags, pol_lookup_t lookup) |
c71099ac | 306 | { |
ab997ad4 | 307 | struct rt6_info *rt; |
308 | ||
309 | rt = lookup(net, net->ipv6.fib6_main_tbl, fl6, flags); | |
07f61557 | 310 | if (rt->dst.error == -EAGAIN) { |
ab997ad4 | 311 | ip6_rt_put(rt); |
312 | rt = net->ipv6.ip6_null_entry; | |
313 | dst_hold(&rt->dst); | |
314 | } | |
315 | ||
316 | return &rt->dst; | |
c71099ac TG |
317 | } |
318 | ||
2c8c1e72 | 319 | static void __net_init fib6_tables_init(struct net *net) |
c71099ac | 320 | { |
58f09b78 | 321 | fib6_link_table(net, net->ipv6.fib6_main_tbl); |
c71099ac TG |
322 | } |
323 | ||
324 | #endif | |
325 | ||
e1ee0a5b IS |
326 | unsigned int fib6_tables_seq_read(struct net *net) |
327 | { | |
328 | unsigned int h, fib_seq = 0; | |
329 | ||
330 | rcu_read_lock(); | |
331 | for (h = 0; h < FIB6_TABLE_HASHSZ; h++) { | |
332 | struct hlist_head *head = &net->ipv6.fib_table_hash[h]; | |
333 | struct fib6_table *tb; | |
334 | ||
66f5d6ce | 335 | hlist_for_each_entry_rcu(tb, head, tb6_hlist) |
e1ee0a5b | 336 | fib_seq += tb->fib_seq; |
e1ee0a5b IS |
337 | } |
338 | rcu_read_unlock(); | |
339 | ||
340 | return fib_seq; | |
341 | } | |
342 | ||
343 | static int call_fib6_entry_notifier(struct notifier_block *nb, struct net *net, | |
344 | enum fib_event_type event_type, | |
345 | struct rt6_info *rt) | |
346 | { | |
347 | struct fib6_entry_notifier_info info = { | |
348 | .rt = rt, | |
349 | }; | |
350 | ||
351 | return call_fib6_notifier(nb, net, event_type, &info.info); | |
352 | } | |
353 | ||
df77fe4d IS |
354 | static int call_fib6_entry_notifiers(struct net *net, |
355 | enum fib_event_type event_type, | |
356 | struct rt6_info *rt) | |
357 | { | |
358 | struct fib6_entry_notifier_info info = { | |
359 | .rt = rt, | |
360 | }; | |
361 | ||
e1ee0a5b | 362 | rt->rt6i_table->fib_seq++; |
df77fe4d IS |
363 | return call_fib6_notifiers(net, event_type, &info.info); |
364 | } | |
365 | ||
e1ee0a5b IS |
366 | struct fib6_dump_arg { |
367 | struct net *net; | |
368 | struct notifier_block *nb; | |
369 | }; | |
370 | ||
371 | static void fib6_rt_dump(struct rt6_info *rt, struct fib6_dump_arg *arg) | |
372 | { | |
373 | if (rt == arg->net->ipv6.ip6_null_entry) | |
374 | return; | |
375 | call_fib6_entry_notifier(arg->nb, arg->net, FIB_EVENT_ENTRY_ADD, rt); | |
376 | } | |
377 | ||
378 | static int fib6_node_dump(struct fib6_walker *w) | |
379 | { | |
380 | struct rt6_info *rt; | |
381 | ||
66f5d6ce | 382 | for_each_fib6_walker_rt(w) |
e1ee0a5b IS |
383 | fib6_rt_dump(rt, w->args); |
384 | w->leaf = NULL; | |
385 | return 0; | |
386 | } | |
387 | ||
388 | static void fib6_table_dump(struct net *net, struct fib6_table *tb, | |
389 | struct fib6_walker *w) | |
390 | { | |
391 | w->root = &tb->tb6_root; | |
66f5d6ce | 392 | spin_lock_bh(&tb->tb6_lock); |
e1ee0a5b | 393 | fib6_walk(net, w); |
66f5d6ce | 394 | spin_unlock_bh(&tb->tb6_lock); |
e1ee0a5b IS |
395 | } |
396 | ||
397 | /* Called with rcu_read_lock() */ | |
398 | int fib6_tables_dump(struct net *net, struct notifier_block *nb) | |
399 | { | |
400 | struct fib6_dump_arg arg; | |
401 | struct fib6_walker *w; | |
402 | unsigned int h; | |
403 | ||
404 | w = kzalloc(sizeof(*w), GFP_ATOMIC); | |
405 | if (!w) | |
406 | return -ENOMEM; | |
407 | ||
408 | w->func = fib6_node_dump; | |
409 | arg.net = net; | |
410 | arg.nb = nb; | |
411 | w->args = &arg; | |
412 | ||
413 | for (h = 0; h < FIB6_TABLE_HASHSZ; h++) { | |
414 | struct hlist_head *head = &net->ipv6.fib_table_hash[h]; | |
415 | struct fib6_table *tb; | |
416 | ||
417 | hlist_for_each_entry_rcu(tb, head, tb6_hlist) | |
418 | fib6_table_dump(net, tb, w); | |
419 | } | |
420 | ||
421 | kfree(w); | |
422 | ||
423 | return 0; | |
424 | } | |
425 | ||
94b2cfe0 | 426 | static int fib6_dump_node(struct fib6_walker *w) |
1b43af54 PM |
427 | { |
428 | int res; | |
429 | struct rt6_info *rt; | |
430 | ||
66f5d6ce | 431 | for_each_fib6_walker_rt(w) { |
1b43af54 PM |
432 | res = rt6_dump_route(rt, w->args); |
433 | if (res < 0) { | |
434 | /* Frame is full, suspend walking */ | |
435 | w->leaf = rt; | |
436 | return 1; | |
437 | } | |
beb1afac DA |
438 | |
439 | /* Multipath routes are dumped in one route with the | |
440 | * RTA_MULTIPATH attribute. Jump 'rt' to point to the | |
441 | * last sibling of this route (no need to dump the | |
442 | * sibling routes again) | |
443 | */ | |
444 | if (rt->rt6i_nsiblings) | |
445 | rt = list_last_entry(&rt->rt6i_siblings, | |
446 | struct rt6_info, | |
447 | rt6i_siblings); | |
1b43af54 PM |
448 | } |
449 | w->leaf = NULL; | |
450 | return 0; | |
451 | } | |
452 | ||
453 | static void fib6_dump_end(struct netlink_callback *cb) | |
454 | { | |
9a03cd8f | 455 | struct net *net = sock_net(cb->skb->sk); |
94b2cfe0 | 456 | struct fib6_walker *w = (void *)cb->args[2]; |
1b43af54 PM |
457 | |
458 | if (w) { | |
7891cc81 HX |
459 | if (cb->args[4]) { |
460 | cb->args[4] = 0; | |
9a03cd8f | 461 | fib6_walker_unlink(net, w); |
7891cc81 | 462 | } |
1b43af54 PM |
463 | cb->args[2] = 0; |
464 | kfree(w); | |
465 | } | |
437de07c | 466 | cb->done = (void *)cb->args[3]; |
1b43af54 PM |
467 | cb->args[1] = 3; |
468 | } | |
469 | ||
470 | static int fib6_dump_done(struct netlink_callback *cb) | |
471 | { | |
472 | fib6_dump_end(cb); | |
473 | return cb->done ? cb->done(cb) : 0; | |
474 | } | |
475 | ||
476 | static int fib6_dump_table(struct fib6_table *table, struct sk_buff *skb, | |
477 | struct netlink_callback *cb) | |
478 | { | |
9a03cd8f | 479 | struct net *net = sock_net(skb->sk); |
94b2cfe0 | 480 | struct fib6_walker *w; |
1b43af54 PM |
481 | int res; |
482 | ||
483 | w = (void *)cb->args[2]; | |
484 | w->root = &table->tb6_root; | |
485 | ||
486 | if (cb->args[4] == 0) { | |
2bec5a36 PM |
487 | w->count = 0; |
488 | w->skip = 0; | |
489 | ||
66f5d6ce | 490 | spin_lock_bh(&table->tb6_lock); |
9a03cd8f | 491 | res = fib6_walk(net, w); |
66f5d6ce | 492 | spin_unlock_bh(&table->tb6_lock); |
2bec5a36 | 493 | if (res > 0) { |
1b43af54 | 494 | cb->args[4] = 1; |
2bec5a36 PM |
495 | cb->args[5] = w->root->fn_sernum; |
496 | } | |
1b43af54 | 497 | } else { |
2bec5a36 PM |
498 | if (cb->args[5] != w->root->fn_sernum) { |
499 | /* Begin at the root if the tree changed */ | |
500 | cb->args[5] = w->root->fn_sernum; | |
501 | w->state = FWS_INIT; | |
502 | w->node = w->root; | |
503 | w->skip = w->count; | |
504 | } else | |
505 | w->skip = 0; | |
506 | ||
66f5d6ce | 507 | spin_lock_bh(&table->tb6_lock); |
1b43af54 | 508 | res = fib6_walk_continue(w); |
66f5d6ce | 509 | spin_unlock_bh(&table->tb6_lock); |
7891cc81 | 510 | if (res <= 0) { |
9a03cd8f | 511 | fib6_walker_unlink(net, w); |
7891cc81 | 512 | cb->args[4] = 0; |
1b43af54 | 513 | } |
1b43af54 | 514 | } |
7891cc81 | 515 | |
1b43af54 PM |
516 | return res; |
517 | } | |
518 | ||
c127ea2c | 519 | static int inet6_dump_fib(struct sk_buff *skb, struct netlink_callback *cb) |
1b43af54 | 520 | { |
3b1e0a65 | 521 | struct net *net = sock_net(skb->sk); |
1b43af54 PM |
522 | unsigned int h, s_h; |
523 | unsigned int e = 0, s_e; | |
524 | struct rt6_rtnl_dump_arg arg; | |
94b2cfe0 | 525 | struct fib6_walker *w; |
1b43af54 | 526 | struct fib6_table *tb; |
58f09b78 | 527 | struct hlist_head *head; |
1b43af54 PM |
528 | int res = 0; |
529 | ||
530 | s_h = cb->args[0]; | |
531 | s_e = cb->args[1]; | |
532 | ||
533 | w = (void *)cb->args[2]; | |
507c9b1e | 534 | if (!w) { |
1b43af54 PM |
535 | /* New dump: |
536 | * | |
537 | * 1. hook callback destructor. | |
538 | */ | |
539 | cb->args[3] = (long)cb->done; | |
540 | cb->done = fib6_dump_done; | |
541 | ||
542 | /* | |
543 | * 2. allocate and initialize walker. | |
544 | */ | |
545 | w = kzalloc(sizeof(*w), GFP_ATOMIC); | |
507c9b1e | 546 | if (!w) |
1b43af54 PM |
547 | return -ENOMEM; |
548 | w->func = fib6_dump_node; | |
549 | cb->args[2] = (long)w; | |
550 | } | |
551 | ||
552 | arg.skb = skb; | |
553 | arg.cb = cb; | |
191cd582 | 554 | arg.net = net; |
1b43af54 PM |
555 | w->args = &arg; |
556 | ||
e67f88dd | 557 | rcu_read_lock(); |
a33bc5c1 | 558 | for (h = s_h; h < FIB6_TABLE_HASHSZ; h++, s_e = 0) { |
1b43af54 | 559 | e = 0; |
58f09b78 | 560 | head = &net->ipv6.fib_table_hash[h]; |
b67bfe0d | 561 | hlist_for_each_entry_rcu(tb, head, tb6_hlist) { |
1b43af54 PM |
562 | if (e < s_e) |
563 | goto next; | |
564 | res = fib6_dump_table(tb, skb, cb); | |
565 | if (res != 0) | |
566 | goto out; | |
567 | next: | |
568 | e++; | |
569 | } | |
570 | } | |
571 | out: | |
e67f88dd | 572 | rcu_read_unlock(); |
1b43af54 PM |
573 | cb->args[1] = e; |
574 | cb->args[0] = h; | |
575 | ||
576 | res = res < 0 ? res : skb->len; | |
577 | if (res <= 0) | |
578 | fib6_dump_end(cb); | |
579 | return res; | |
580 | } | |
1da177e4 LT |
581 | |
582 | /* | |
583 | * Routing Table | |
584 | * | |
585 | * return the appropriate node for a routing tree "add" operation | |
586 | * by either creating and inserting or by returning an existing | |
587 | * node. | |
588 | */ | |
589 | ||
81eb8447 WW |
590 | static struct fib6_node *fib6_add_1(struct net *net, |
591 | struct fib6_table *table, | |
66f5d6ce WW |
592 | struct fib6_node *root, |
593 | struct in6_addr *addr, int plen, | |
594 | int offset, int allow_create, | |
595 | int replace_required, | |
596 | struct netlink_ext_ack *extack) | |
1da177e4 LT |
597 | { |
598 | struct fib6_node *fn, *in, *ln; | |
599 | struct fib6_node *pn = NULL; | |
600 | struct rt6key *key; | |
601 | int bit; | |
1ab1457c | 602 | __be32 dir = 0; |
1da177e4 LT |
603 | |
604 | RT6_TRACE("fib6_add_1\n"); | |
605 | ||
606 | /* insert node in tree */ | |
607 | ||
608 | fn = root; | |
609 | ||
610 | do { | |
66f5d6ce WW |
611 | struct rt6_info *leaf = rcu_dereference_protected(fn->leaf, |
612 | lockdep_is_held(&table->tb6_lock)); | |
613 | key = (struct rt6key *)((u8 *)leaf + offset); | |
1da177e4 LT |
614 | |
615 | /* | |
616 | * Prefix match | |
617 | */ | |
618 | if (plen < fn->fn_bit || | |
4a287eba | 619 | !ipv6_prefix_equal(&key->addr, addr, fn->fn_bit)) { |
14df015b MV |
620 | if (!allow_create) { |
621 | if (replace_required) { | |
d5d531cb DA |
622 | NL_SET_ERR_MSG(extack, |
623 | "Can not replace route - no match found"); | |
f3213831 | 624 | pr_warn("Can't replace route, no match found\n"); |
14df015b MV |
625 | return ERR_PTR(-ENOENT); |
626 | } | |
f3213831 | 627 | pr_warn("NLM_F_CREATE should be set when creating new route\n"); |
14df015b | 628 | } |
1da177e4 | 629 | goto insert_above; |
4a287eba | 630 | } |
1ab1457c | 631 | |
1da177e4 LT |
632 | /* |
633 | * Exact match ? | |
634 | */ | |
1ab1457c | 635 | |
1da177e4 LT |
636 | if (plen == fn->fn_bit) { |
637 | /* clean up an intermediate node */ | |
507c9b1e | 638 | if (!(fn->fn_flags & RTN_RTINFO)) { |
66f5d6ce WW |
639 | RCU_INIT_POINTER(fn->leaf, NULL); |
640 | rt6_release(leaf); | |
1da177e4 | 641 | } |
1ab1457c | 642 | |
1da177e4 LT |
643 | return fn; |
644 | } | |
645 | ||
646 | /* | |
647 | * We have more bits to go | |
648 | */ | |
1ab1457c | 649 | |
1da177e4 | 650 | /* Try to walk down on tree. */ |
1da177e4 LT |
651 | dir = addr_bit_set(addr, fn->fn_bit); |
652 | pn = fn; | |
66f5d6ce WW |
653 | fn = dir ? |
654 | rcu_dereference_protected(fn->right, | |
655 | lockdep_is_held(&table->tb6_lock)) : | |
656 | rcu_dereference_protected(fn->left, | |
657 | lockdep_is_held(&table->tb6_lock)); | |
1da177e4 LT |
658 | } while (fn); |
659 | ||
14df015b | 660 | if (!allow_create) { |
4a287eba MV |
661 | /* We should not create new node because |
662 | * NLM_F_REPLACE was specified without NLM_F_CREATE | |
663 | * I assume it is safe to require NLM_F_CREATE when | |
664 | * REPLACE flag is used! Later we may want to remove the | |
665 | * check for replace_required, because according | |
666 | * to netlink specification, NLM_F_CREATE | |
667 | * MUST be specified if new route is created. | |
668 | * That would keep IPv6 consistent with IPv4 | |
669 | */ | |
14df015b | 670 | if (replace_required) { |
d5d531cb DA |
671 | NL_SET_ERR_MSG(extack, |
672 | "Can not replace route - no match found"); | |
f3213831 | 673 | pr_warn("Can't replace route, no match found\n"); |
14df015b MV |
674 | return ERR_PTR(-ENOENT); |
675 | } | |
f3213831 | 676 | pr_warn("NLM_F_CREATE should be set when creating new route\n"); |
4a287eba | 677 | } |
1da177e4 LT |
678 | /* |
679 | * We walked to the bottom of tree. | |
680 | * Create new leaf node without children. | |
681 | */ | |
682 | ||
81eb8447 | 683 | ln = node_alloc(net); |
1da177e4 | 684 | |
507c9b1e | 685 | if (!ln) |
188c517a | 686 | return ERR_PTR(-ENOMEM); |
1da177e4 | 687 | ln->fn_bit = plen; |
66f5d6ce | 688 | RCU_INIT_POINTER(ln->parent, pn); |
1da177e4 LT |
689 | |
690 | if (dir) | |
66f5d6ce | 691 | rcu_assign_pointer(pn->right, ln); |
1da177e4 | 692 | else |
66f5d6ce | 693 | rcu_assign_pointer(pn->left, ln); |
1da177e4 LT |
694 | |
695 | return ln; | |
696 | ||
697 | ||
698 | insert_above: | |
699 | /* | |
1ab1457c | 700 | * split since we don't have a common prefix anymore or |
1da177e4 LT |
701 | * we have a less significant route. |
702 | * we've to insert an intermediate node on the list | |
703 | * this new node will point to the one we need to create | |
704 | * and the current | |
705 | */ | |
706 | ||
66f5d6ce WW |
707 | pn = rcu_dereference_protected(fn->parent, |
708 | lockdep_is_held(&table->tb6_lock)); | |
1da177e4 LT |
709 | |
710 | /* find 1st bit in difference between the 2 addrs. | |
711 | ||
971f359d | 712 | See comment in __ipv6_addr_diff: bit may be an invalid value, |
1da177e4 LT |
713 | but if it is >= plen, the value is ignored in any case. |
714 | */ | |
1ab1457c | 715 | |
9225b230 | 716 | bit = __ipv6_addr_diff(addr, &key->addr, sizeof(*addr)); |
1da177e4 | 717 | |
1ab1457c YH |
718 | /* |
719 | * (intermediate)[in] | |
1da177e4 LT |
720 | * / \ |
721 | * (new leaf node)[ln] (old node)[fn] | |
722 | */ | |
723 | if (plen > bit) { | |
81eb8447 WW |
724 | in = node_alloc(net); |
725 | ln = node_alloc(net); | |
1ab1457c | 726 | |
507c9b1e | 727 | if (!in || !ln) { |
1da177e4 | 728 | if (in) |
81eb8447 | 729 | node_free_immediate(net, in); |
1da177e4 | 730 | if (ln) |
81eb8447 | 731 | node_free_immediate(net, ln); |
188c517a | 732 | return ERR_PTR(-ENOMEM); |
1da177e4 LT |
733 | } |
734 | ||
1ab1457c YH |
735 | /* |
736 | * new intermediate node. | |
1da177e4 LT |
737 | * RTN_RTINFO will |
738 | * be off since that an address that chooses one of | |
739 | * the branches would not match less specific routes | |
740 | * in the other branch | |
741 | */ | |
742 | ||
743 | in->fn_bit = bit; | |
744 | ||
66f5d6ce | 745 | RCU_INIT_POINTER(in->parent, pn); |
1da177e4 | 746 | in->leaf = fn->leaf; |
66f5d6ce WW |
747 | atomic_inc(&rcu_dereference_protected(in->leaf, |
748 | lockdep_is_held(&table->tb6_lock))->rt6i_ref); | |
1da177e4 | 749 | |
1da177e4 LT |
750 | /* update parent pointer */ |
751 | if (dir) | |
66f5d6ce | 752 | rcu_assign_pointer(pn->right, in); |
1da177e4 | 753 | else |
66f5d6ce | 754 | rcu_assign_pointer(pn->left, in); |
1da177e4 LT |
755 | |
756 | ln->fn_bit = plen; | |
757 | ||
66f5d6ce WW |
758 | RCU_INIT_POINTER(ln->parent, in); |
759 | rcu_assign_pointer(fn->parent, in); | |
1da177e4 | 760 | |
1da177e4 | 761 | if (addr_bit_set(addr, bit)) { |
66f5d6ce WW |
762 | rcu_assign_pointer(in->right, ln); |
763 | rcu_assign_pointer(in->left, fn); | |
1da177e4 | 764 | } else { |
66f5d6ce WW |
765 | rcu_assign_pointer(in->left, ln); |
766 | rcu_assign_pointer(in->right, fn); | |
1da177e4 LT |
767 | } |
768 | } else { /* plen <= bit */ | |
769 | ||
1ab1457c | 770 | /* |
1da177e4 LT |
771 | * (new leaf node)[ln] |
772 | * / \ | |
773 | * (old node)[fn] NULL | |
774 | */ | |
775 | ||
81eb8447 | 776 | ln = node_alloc(net); |
1da177e4 | 777 | |
507c9b1e | 778 | if (!ln) |
188c517a | 779 | return ERR_PTR(-ENOMEM); |
1da177e4 LT |
780 | |
781 | ln->fn_bit = plen; | |
782 | ||
66f5d6ce | 783 | RCU_INIT_POINTER(ln->parent, pn); |
1da177e4 LT |
784 | |
785 | if (addr_bit_set(&key->addr, plen)) | |
66f5d6ce | 786 | RCU_INIT_POINTER(ln->right, fn); |
1da177e4 | 787 | else |
66f5d6ce WW |
788 | RCU_INIT_POINTER(ln->left, fn); |
789 | ||
790 | rcu_assign_pointer(fn->parent, ln); | |
1da177e4 | 791 | |
66f5d6ce WW |
792 | if (dir) |
793 | rcu_assign_pointer(pn->right, ln); | |
794 | else | |
795 | rcu_assign_pointer(pn->left, ln); | |
1da177e4 LT |
796 | } |
797 | return ln; | |
798 | } | |
799 | ||
94b2cfe0 | 800 | static bool rt6_qualify_for_ecmp(struct rt6_info *rt) |
307f2fb9 HFS |
801 | { |
802 | return (rt->rt6i_flags & (RTF_GATEWAY|RTF_ADDRCONF|RTF_DYNAMIC)) == | |
803 | RTF_GATEWAY; | |
804 | } | |
805 | ||
e715b6d3 | 806 | static void fib6_copy_metrics(u32 *mp, const struct mx6_config *mxc) |
e5fd387a | 807 | { |
e715b6d3 FW |
808 | int i; |
809 | ||
810 | for (i = 0; i < RTAX_MAX; i++) { | |
811 | if (test_bit(i, mxc->mx_valid)) | |
812 | mp[i] = mxc->mx[i]; | |
813 | } | |
814 | } | |
815 | ||
816 | static int fib6_commit_metrics(struct dst_entry *dst, struct mx6_config *mxc) | |
817 | { | |
818 | if (!mxc->mx) | |
819 | return 0; | |
820 | ||
821 | if (dst->flags & DST_HOST) { | |
822 | u32 *mp = dst_metrics_write_ptr(dst); | |
823 | ||
824 | if (unlikely(!mp)) | |
825 | return -ENOMEM; | |
826 | ||
827 | fib6_copy_metrics(mp, mxc); | |
828 | } else { | |
829 | dst_init_metrics(dst, mxc->mx, false); | |
830 | ||
831 | /* We've stolen mx now. */ | |
832 | mxc->mx = NULL; | |
e5fd387a | 833 | } |
e715b6d3 | 834 | |
e5fd387a MK |
835 | return 0; |
836 | } | |
837 | ||
6e9e16e6 HFS |
838 | static void fib6_purge_rt(struct rt6_info *rt, struct fib6_node *fn, |
839 | struct net *net) | |
840 | { | |
66f5d6ce WW |
841 | struct fib6_table *table = rt->rt6i_table; |
842 | ||
6e9e16e6 HFS |
843 | if (atomic_read(&rt->rt6i_ref) != 1) { |
844 | /* This route is used as dummy address holder in some split | |
845 | * nodes. It is not leaked, but it still holds other resources, | |
846 | * which must be released in time. So, scan ascendant nodes | |
847 | * and replace dummy references to this route with references | |
848 | * to still alive ones. | |
849 | */ | |
850 | while (fn) { | |
66f5d6ce WW |
851 | struct rt6_info *leaf = rcu_dereference_protected(fn->leaf, |
852 | lockdep_is_held(&table->tb6_lock)); | |
853 | struct rt6_info *new_leaf; | |
854 | if (!(fn->fn_flags & RTN_RTINFO) && leaf == rt) { | |
855 | new_leaf = fib6_find_prefix(net, table, fn); | |
856 | atomic_inc(&new_leaf->rt6i_ref); | |
857 | rcu_assign_pointer(fn->leaf, new_leaf); | |
6e9e16e6 HFS |
858 | rt6_release(rt); |
859 | } | |
66f5d6ce WW |
860 | fn = rcu_dereference_protected(fn->parent, |
861 | lockdep_is_held(&table->tb6_lock)); | |
6e9e16e6 | 862 | } |
6e9e16e6 HFS |
863 | } |
864 | } | |
865 | ||
1da177e4 LT |
866 | /* |
867 | * Insert routing information in a node. | |
868 | */ | |
869 | ||
870 | static int fib6_add_rt2node(struct fib6_node *fn, struct rt6_info *rt, | |
e715b6d3 | 871 | struct nl_info *info, struct mx6_config *mxc) |
1da177e4 | 872 | { |
66f5d6ce WW |
873 | struct rt6_info *leaf = rcu_dereference_protected(fn->leaf, |
874 | lockdep_is_held(&rt->rt6i_table->tb6_lock)); | |
1da177e4 | 875 | struct rt6_info *iter = NULL; |
66f5d6ce WW |
876 | struct rt6_info __rcu **ins; |
877 | struct rt6_info __rcu **fallback_ins = NULL; | |
507c9b1e DM |
878 | int replace = (info->nlh && |
879 | (info->nlh->nlmsg_flags & NLM_F_REPLACE)); | |
880 | int add = (!info->nlh || | |
881 | (info->nlh->nlmsg_flags & NLM_F_CREATE)); | |
4a287eba | 882 | int found = 0; |
307f2fb9 | 883 | bool rt_can_ecmp = rt6_qualify_for_ecmp(rt); |
73483c12 | 884 | u16 nlflags = NLM_F_EXCL; |
e5fd387a | 885 | int err; |
1da177e4 | 886 | |
1f5e29ce DA |
887 | if (info->nlh && (info->nlh->nlmsg_flags & NLM_F_APPEND)) |
888 | nlflags |= NLM_F_APPEND; | |
889 | ||
1da177e4 LT |
890 | ins = &fn->leaf; |
891 | ||
66f5d6ce WW |
892 | for (iter = leaf; iter; |
893 | iter = rcu_dereference_protected(iter->dst.rt6_next, | |
894 | lockdep_is_held(&rt->rt6i_table->tb6_lock))) { | |
1da177e4 LT |
895 | /* |
896 | * Search for duplicates | |
897 | */ | |
898 | ||
899 | if (iter->rt6i_metric == rt->rt6i_metric) { | |
900 | /* | |
901 | * Same priority level | |
902 | */ | |
507c9b1e DM |
903 | if (info->nlh && |
904 | (info->nlh->nlmsg_flags & NLM_F_EXCL)) | |
4a287eba | 905 | return -EEXIST; |
73483c12 GN |
906 | |
907 | nlflags &= ~NLM_F_EXCL; | |
4a287eba | 908 | if (replace) { |
27596472 MK |
909 | if (rt_can_ecmp == rt6_qualify_for_ecmp(iter)) { |
910 | found++; | |
911 | break; | |
912 | } | |
913 | if (rt_can_ecmp) | |
914 | fallback_ins = fallback_ins ?: ins; | |
915 | goto next_iter; | |
4a287eba | 916 | } |
1da177e4 | 917 | |
f06b7549 | 918 | if (rt6_duplicate_nexthop(iter, rt)) { |
51ebd318 ND |
919 | if (rt->rt6i_nsiblings) |
920 | rt->rt6i_nsiblings = 0; | |
507c9b1e | 921 | if (!(iter->rt6i_flags & RTF_EXPIRES)) |
1da177e4 | 922 | return -EEXIST; |
1716a961 G |
923 | if (!(rt->rt6i_flags & RTF_EXPIRES)) |
924 | rt6_clean_expires(iter); | |
925 | else | |
926 | rt6_set_expires(iter, rt->dst.expires); | |
45e4fd26 | 927 | iter->rt6i_pmtu = rt->rt6i_pmtu; |
1da177e4 LT |
928 | return -EEXIST; |
929 | } | |
51ebd318 ND |
930 | /* If we have the same destination and the same metric, |
931 | * but not the same gateway, then the route we try to | |
932 | * add is sibling to this route, increment our counter | |
933 | * of siblings, and later we will add our route to the | |
934 | * list. | |
935 | * Only static routes (which don't have flag | |
936 | * RTF_EXPIRES) are used for ECMPv6. | |
937 | * | |
938 | * To avoid long list, we only had siblings if the | |
939 | * route have a gateway. | |
940 | */ | |
307f2fb9 HFS |
941 | if (rt_can_ecmp && |
942 | rt6_qualify_for_ecmp(iter)) | |
51ebd318 | 943 | rt->rt6i_nsiblings++; |
1da177e4 LT |
944 | } |
945 | ||
946 | if (iter->rt6i_metric > rt->rt6i_metric) | |
947 | break; | |
948 | ||
27596472 | 949 | next_iter: |
d8d1f30b | 950 | ins = &iter->dst.rt6_next; |
1da177e4 LT |
951 | } |
952 | ||
27596472 MK |
953 | if (fallback_ins && !found) { |
954 | /* No ECMP-able route found, replace first non-ECMP one */ | |
955 | ins = fallback_ins; | |
66f5d6ce WW |
956 | iter = rcu_dereference_protected(*ins, |
957 | lockdep_is_held(&rt->rt6i_table->tb6_lock)); | |
27596472 MK |
958 | found++; |
959 | } | |
960 | ||
f11e6659 DM |
961 | /* Reset round-robin state, if necessary */ |
962 | if (ins == &fn->leaf) | |
963 | fn->rr_ptr = NULL; | |
964 | ||
51ebd318 ND |
965 | /* Link this route to others same route. */ |
966 | if (rt->rt6i_nsiblings) { | |
967 | unsigned int rt6i_nsiblings; | |
968 | struct rt6_info *sibling, *temp_sibling; | |
969 | ||
970 | /* Find the first route that have the same metric */ | |
66f5d6ce | 971 | sibling = leaf; |
51ebd318 | 972 | while (sibling) { |
307f2fb9 HFS |
973 | if (sibling->rt6i_metric == rt->rt6i_metric && |
974 | rt6_qualify_for_ecmp(sibling)) { | |
51ebd318 ND |
975 | list_add_tail(&rt->rt6i_siblings, |
976 | &sibling->rt6i_siblings); | |
977 | break; | |
978 | } | |
66f5d6ce WW |
979 | sibling = rcu_dereference_protected(sibling->dst.rt6_next, |
980 | lockdep_is_held(&rt->rt6i_table->tb6_lock)); | |
51ebd318 ND |
981 | } |
982 | /* For each sibling in the list, increment the counter of | |
983 | * siblings. BUG() if counters does not match, list of siblings | |
984 | * is broken! | |
985 | */ | |
986 | rt6i_nsiblings = 0; | |
987 | list_for_each_entry_safe(sibling, temp_sibling, | |
988 | &rt->rt6i_siblings, rt6i_siblings) { | |
989 | sibling->rt6i_nsiblings++; | |
990 | BUG_ON(sibling->rt6i_nsiblings != rt->rt6i_nsiblings); | |
991 | rt6i_nsiblings++; | |
992 | } | |
993 | BUG_ON(rt6i_nsiblings != rt->rt6i_nsiblings); | |
994 | } | |
995 | ||
1da177e4 LT |
996 | /* |
997 | * insert node | |
998 | */ | |
4a287eba MV |
999 | if (!replace) { |
1000 | if (!add) | |
f3213831 | 1001 | pr_warn("NLM_F_CREATE should be set when creating new route\n"); |
4a287eba MV |
1002 | |
1003 | add: | |
73483c12 | 1004 | nlflags |= NLM_F_CREATE; |
e715b6d3 FW |
1005 | err = fib6_commit_metrics(&rt->dst, mxc); |
1006 | if (err) | |
1007 | return err; | |
1008 | ||
66f5d6ce | 1009 | rcu_assign_pointer(rt->dst.rt6_next, iter); |
4a287eba | 1010 | atomic_inc(&rt->rt6i_ref); |
66f5d6ce WW |
1011 | rcu_assign_pointer(rt->rt6i_node, fn); |
1012 | rcu_assign_pointer(*ins, rt); | |
df77fe4d IS |
1013 | call_fib6_entry_notifiers(info->nl_net, FIB_EVENT_ENTRY_ADD, |
1014 | rt); | |
3b1137fe DA |
1015 | if (!info->skip_notify) |
1016 | inet6_rt_notify(RTM_NEWROUTE, rt, info, nlflags); | |
4a287eba MV |
1017 | info->nl_net->ipv6.rt6_stats->fib_rt_entries++; |
1018 | ||
507c9b1e | 1019 | if (!(fn->fn_flags & RTN_RTINFO)) { |
4a287eba MV |
1020 | info->nl_net->ipv6.rt6_stats->fib_route_nodes++; |
1021 | fn->fn_flags |= RTN_RTINFO; | |
1022 | } | |
1da177e4 | 1023 | |
4a287eba | 1024 | } else { |
27596472 MK |
1025 | int nsiblings; |
1026 | ||
4a287eba MV |
1027 | if (!found) { |
1028 | if (add) | |
1029 | goto add; | |
f3213831 | 1030 | pr_warn("NLM_F_REPLACE set, but no existing node found!\n"); |
4a287eba MV |
1031 | return -ENOENT; |
1032 | } | |
e715b6d3 FW |
1033 | |
1034 | err = fib6_commit_metrics(&rt->dst, mxc); | |
1035 | if (err) | |
1036 | return err; | |
1037 | ||
66f5d6ce | 1038 | atomic_inc(&rt->rt6i_ref); |
4e587ea7 | 1039 | rcu_assign_pointer(rt->rt6i_node, fn); |
4a287eba | 1040 | rt->dst.rt6_next = iter->dst.rt6_next; |
66f5d6ce | 1041 | rcu_assign_pointer(*ins, rt); |
df77fe4d IS |
1042 | call_fib6_entry_notifiers(info->nl_net, FIB_EVENT_ENTRY_REPLACE, |
1043 | rt); | |
3b1137fe DA |
1044 | if (!info->skip_notify) |
1045 | inet6_rt_notify(RTM_NEWROUTE, rt, info, NLM_F_REPLACE); | |
507c9b1e | 1046 | if (!(fn->fn_flags & RTN_RTINFO)) { |
4a287eba MV |
1047 | info->nl_net->ipv6.rt6_stats->fib_route_nodes++; |
1048 | fn->fn_flags |= RTN_RTINFO; | |
1049 | } | |
27596472 | 1050 | nsiblings = iter->rt6i_nsiblings; |
7483cea7 | 1051 | iter->rt6i_node = NULL; |
6e9e16e6 | 1052 | fib6_purge_rt(iter, fn, info->nl_net); |
66f5d6ce | 1053 | if (rcu_access_pointer(fn->rr_ptr) == iter) |
383143f3 | 1054 | fn->rr_ptr = NULL; |
6e9e16e6 | 1055 | rt6_release(iter); |
27596472 MK |
1056 | |
1057 | if (nsiblings) { | |
1058 | /* Replacing an ECMP route, remove all siblings */ | |
1059 | ins = &rt->dst.rt6_next; | |
66f5d6ce WW |
1060 | iter = rcu_dereference_protected(*ins, |
1061 | lockdep_is_held(&rt->rt6i_table->tb6_lock)); | |
27596472 | 1062 | while (iter) { |
67e19400 SD |
1063 | if (iter->rt6i_metric > rt->rt6i_metric) |
1064 | break; | |
27596472 MK |
1065 | if (rt6_qualify_for_ecmp(iter)) { |
1066 | *ins = iter->dst.rt6_next; | |
7483cea7 | 1067 | iter->rt6i_node = NULL; |
27596472 | 1068 | fib6_purge_rt(iter, fn, info->nl_net); |
66f5d6ce | 1069 | if (rcu_access_pointer(fn->rr_ptr) == iter) |
383143f3 | 1070 | fn->rr_ptr = NULL; |
27596472 MK |
1071 | rt6_release(iter); |
1072 | nsiblings--; | |
81eb8447 | 1073 | info->nl_net->ipv6.rt6_stats->fib_rt_entries--; |
27596472 MK |
1074 | } else { |
1075 | ins = &iter->dst.rt6_next; | |
1076 | } | |
66f5d6ce WW |
1077 | iter = rcu_dereference_protected(*ins, |
1078 | lockdep_is_held(&rt->rt6i_table->tb6_lock)); | |
27596472 MK |
1079 | } |
1080 | WARN_ON(nsiblings != 0); | |
1081 | } | |
1da177e4 LT |
1082 | } |
1083 | ||
1084 | return 0; | |
1085 | } | |
1086 | ||
94b2cfe0 | 1087 | static void fib6_start_gc(struct net *net, struct rt6_info *rt) |
1da177e4 | 1088 | { |
417f28bb | 1089 | if (!timer_pending(&net->ipv6.ip6_fib_timer) && |
507c9b1e | 1090 | (rt->rt6i_flags & (RTF_EXPIRES | RTF_CACHE))) |
417f28bb | 1091 | mod_timer(&net->ipv6.ip6_fib_timer, |
847499ce | 1092 | jiffies + net->ipv6.sysctl.ip6_rt_gc_interval); |
1da177e4 LT |
1093 | } |
1094 | ||
63152fc0 | 1095 | void fib6_force_start_gc(struct net *net) |
1da177e4 | 1096 | { |
417f28bb SH |
1097 | if (!timer_pending(&net->ipv6.ip6_fib_timer)) |
1098 | mod_timer(&net->ipv6.ip6_fib_timer, | |
847499ce | 1099 | jiffies + net->ipv6.sysctl.ip6_rt_gc_interval); |
1da177e4 LT |
1100 | } |
1101 | ||
bbd63f06 WW |
1102 | static void fib6_update_sernum_upto_root(struct rt6_info *rt, |
1103 | int sernum) | |
1104 | { | |
1105 | struct fib6_node *fn = rcu_dereference_protected(rt->rt6i_node, | |
1106 | lockdep_is_held(&rt->rt6i_table->tb6_lock)); | |
1107 | ||
1108 | /* paired with smp_rmb() in rt6_get_cookie_safe() */ | |
1109 | smp_wmb(); | |
1110 | while (fn) { | |
1111 | fn->fn_sernum = sernum; | |
66f5d6ce WW |
1112 | fn = rcu_dereference_protected(fn->parent, |
1113 | lockdep_is_held(&rt->rt6i_table->tb6_lock)); | |
bbd63f06 WW |
1114 | } |
1115 | } | |
1116 | ||
1da177e4 LT |
1117 | /* |
1118 | * Add routing information to the routing tree. | |
1119 | * <destination addr>/<source addr> | |
1120 | * with source addr info in sub-trees | |
66f5d6ce | 1121 | * Need to own table->tb6_lock |
1da177e4 LT |
1122 | */ |
1123 | ||
e715b6d3 | 1124 | int fib6_add(struct fib6_node *root, struct rt6_info *rt, |
333c4301 DA |
1125 | struct nl_info *info, struct mx6_config *mxc, |
1126 | struct netlink_ext_ack *extack) | |
1da177e4 | 1127 | { |
66f5d6ce | 1128 | struct fib6_table *table = rt->rt6i_table; |
66729e18 | 1129 | struct fib6_node *fn, *pn = NULL; |
1da177e4 | 1130 | int err = -ENOMEM; |
4a287eba MV |
1131 | int allow_create = 1; |
1132 | int replace_required = 0; | |
812918c4 | 1133 | int sernum = fib6_new_sernum(info->nl_net); |
507c9b1e | 1134 | |
a4c2fd7f | 1135 | if (WARN_ON_ONCE(!atomic_read(&rt->dst.__refcnt))) |
8e3d5be7 | 1136 | return -EINVAL; |
2b760fcf WW |
1137 | if (WARN_ON_ONCE(rt->rt6i_flags & RTF_CACHE)) |
1138 | return -EINVAL; | |
8e3d5be7 | 1139 | |
507c9b1e DM |
1140 | if (info->nlh) { |
1141 | if (!(info->nlh->nlmsg_flags & NLM_F_CREATE)) | |
4a287eba | 1142 | allow_create = 0; |
507c9b1e | 1143 | if (info->nlh->nlmsg_flags & NLM_F_REPLACE) |
4a287eba MV |
1144 | replace_required = 1; |
1145 | } | |
1146 | if (!allow_create && !replace_required) | |
f3213831 | 1147 | pr_warn("RTM_NEWROUTE with no NLM_F_CREATE or NLM_F_REPLACE\n"); |
1da177e4 | 1148 | |
81eb8447 | 1149 | fn = fib6_add_1(info->nl_net, table, root, |
66f5d6ce | 1150 | &rt->rt6i_dst.addr, rt->rt6i_dst.plen, |
9225b230 | 1151 | offsetof(struct rt6_info, rt6i_dst), allow_create, |
bbd63f06 | 1152 | replace_required, extack); |
4a287eba MV |
1153 | if (IS_ERR(fn)) { |
1154 | err = PTR_ERR(fn); | |
ae7b4e1f | 1155 | fn = NULL; |
1da177e4 | 1156 | goto out; |
188c517a | 1157 | } |
1da177e4 | 1158 | |
66729e18 YH |
1159 | pn = fn; |
1160 | ||
1da177e4 LT |
1161 | #ifdef CONFIG_IPV6_SUBTREES |
1162 | if (rt->rt6i_src.plen) { | |
1163 | struct fib6_node *sn; | |
1164 | ||
66f5d6ce | 1165 | if (!rcu_access_pointer(fn->subtree)) { |
1da177e4 LT |
1166 | struct fib6_node *sfn; |
1167 | ||
1168 | /* | |
1169 | * Create subtree. | |
1170 | * | |
1171 | * fn[main tree] | |
1172 | * | | |
1173 | * sfn[subtree root] | |
1174 | * \ | |
1175 | * sn[new leaf node] | |
1176 | */ | |
1177 | ||
1178 | /* Create subtree root node */ | |
81eb8447 | 1179 | sfn = node_alloc(info->nl_net); |
507c9b1e | 1180 | if (!sfn) |
348a4002 | 1181 | goto failure; |
1da177e4 | 1182 | |
8ed67789 | 1183 | atomic_inc(&info->nl_net->ipv6.ip6_null_entry->rt6i_ref); |
66f5d6ce WW |
1184 | rcu_assign_pointer(sfn->leaf, |
1185 | info->nl_net->ipv6.ip6_null_entry); | |
1da177e4 | 1186 | sfn->fn_flags = RTN_ROOT; |
1da177e4 LT |
1187 | |
1188 | /* Now add the first leaf node to new subtree */ | |
1189 | ||
81eb8447 WW |
1190 | sn = fib6_add_1(info->nl_net, table, sfn, |
1191 | &rt->rt6i_src.addr, rt->rt6i_src.plen, | |
4a287eba | 1192 | offsetof(struct rt6_info, rt6i_src), |
bbd63f06 | 1193 | allow_create, replace_required, extack); |
1da177e4 | 1194 | |
f950c0ec | 1195 | if (IS_ERR(sn)) { |
1da177e4 | 1196 | /* If it is failed, discard just allocated |
348a4002 | 1197 | root, and then (in failure) stale node |
1da177e4 LT |
1198 | in main tree. |
1199 | */ | |
81eb8447 | 1200 | node_free_immediate(info->nl_net, sfn); |
188c517a | 1201 | err = PTR_ERR(sn); |
348a4002 | 1202 | goto failure; |
1da177e4 LT |
1203 | } |
1204 | ||
1205 | /* Now link new subtree to main tree */ | |
66f5d6ce WW |
1206 | rcu_assign_pointer(sfn->parent, fn); |
1207 | rcu_assign_pointer(fn->subtree, sfn); | |
1da177e4 | 1208 | } else { |
81eb8447 WW |
1209 | sn = fib6_add_1(info->nl_net, table, FIB6_SUBTREE(fn), |
1210 | &rt->rt6i_src.addr, rt->rt6i_src.plen, | |
4a287eba | 1211 | offsetof(struct rt6_info, rt6i_src), |
bbd63f06 | 1212 | allow_create, replace_required, extack); |
1da177e4 | 1213 | |
4a287eba MV |
1214 | if (IS_ERR(sn)) { |
1215 | err = PTR_ERR(sn); | |
348a4002 | 1216 | goto failure; |
188c517a | 1217 | } |
1da177e4 LT |
1218 | } |
1219 | ||
66f5d6ce | 1220 | if (!rcu_access_pointer(fn->leaf)) { |
66729e18 | 1221 | atomic_inc(&rt->rt6i_ref); |
66f5d6ce | 1222 | rcu_assign_pointer(fn->leaf, rt); |
66729e18 | 1223 | } |
1da177e4 LT |
1224 | fn = sn; |
1225 | } | |
1226 | #endif | |
1227 | ||
e715b6d3 | 1228 | err = fib6_add_rt2node(fn, rt, info, mxc); |
bbd63f06 WW |
1229 | if (!err) { |
1230 | fib6_update_sernum_upto_root(rt, sernum); | |
63152fc0 | 1231 | fib6_start_gc(info->nl_net, rt); |
bbd63f06 | 1232 | } |
1da177e4 LT |
1233 | |
1234 | out: | |
66729e18 YH |
1235 | if (err) { |
1236 | #ifdef CONFIG_IPV6_SUBTREES | |
1237 | /* | |
1238 | * If fib6_add_1 has cleared the old leaf pointer in the | |
1239 | * super-tree leaf node we have to find a new one for it. | |
1240 | */ | |
66f5d6ce WW |
1241 | struct rt6_info *pn_leaf = rcu_dereference_protected(pn->leaf, |
1242 | lockdep_is_held(&table->tb6_lock)); | |
1243 | if (pn != fn && pn_leaf == rt) { | |
1244 | pn_leaf = NULL; | |
1245 | RCU_INIT_POINTER(pn->leaf, NULL); | |
3c051235 DM |
1246 | atomic_dec(&rt->rt6i_ref); |
1247 | } | |
66f5d6ce WW |
1248 | if (pn != fn && !pn_leaf && !(pn->fn_flags & RTN_RTINFO)) { |
1249 | pn_leaf = fib6_find_prefix(info->nl_net, table, pn); | |
66729e18 | 1250 | #if RT6_DEBUG >= 2 |
66f5d6ce WW |
1251 | if (!pn_leaf) { |
1252 | WARN_ON(!pn_leaf); | |
1253 | pn_leaf = info->nl_net->ipv6.ip6_null_entry; | |
66729e18 YH |
1254 | } |
1255 | #endif | |
66f5d6ce WW |
1256 | atomic_inc(&pn_leaf->rt6i_ref); |
1257 | rcu_assign_pointer(pn->leaf, pn_leaf); | |
66729e18 YH |
1258 | } |
1259 | #endif | |
348a4002 | 1260 | goto failure; |
66729e18 | 1261 | } |
1da177e4 LT |
1262 | return err; |
1263 | ||
348a4002 WW |
1264 | failure: |
1265 | /* fn->leaf could be NULL if fn is an intermediate node and we | |
1266 | * failed to add the new route to it in both subtree creation | |
1267 | * failure and fib6_add_rt2node() failure case. | |
1268 | * In both cases, fib6_repair_tree() should be called to fix | |
1269 | * fn->leaf. | |
1da177e4 | 1270 | */ |
1da177e4 | 1271 | if (fn && !(fn->fn_flags & (RTN_RTINFO|RTN_ROOT))) |
66f5d6ce | 1272 | fib6_repair_tree(info->nl_net, table, fn); |
1cfb71ee WW |
1273 | /* Always release dst as dst->__refcnt is guaranteed |
1274 | * to be taken before entering this function | |
1275 | */ | |
587fea74 | 1276 | dst_release_immediate(&rt->dst); |
1da177e4 | 1277 | return err; |
1da177e4 LT |
1278 | } |
1279 | ||
1280 | /* | |
1281 | * Routing tree lookup | |
1282 | * | |
1283 | */ | |
1284 | ||
1285 | struct lookup_args { | |
507c9b1e | 1286 | int offset; /* key offset on rt6_info */ |
b71d1d42 | 1287 | const struct in6_addr *addr; /* search key */ |
1da177e4 LT |
1288 | }; |
1289 | ||
437de07c WY |
1290 | static struct fib6_node *fib6_lookup_1(struct fib6_node *root, |
1291 | struct lookup_args *args) | |
1da177e4 LT |
1292 | { |
1293 | struct fib6_node *fn; | |
e69a4adc | 1294 | __be32 dir; |
1da177e4 | 1295 | |
825e288e YH |
1296 | if (unlikely(args->offset == 0)) |
1297 | return NULL; | |
1298 | ||
1da177e4 LT |
1299 | /* |
1300 | * Descend on a tree | |
1301 | */ | |
1302 | ||
1303 | fn = root; | |
1304 | ||
1305 | for (;;) { | |
1306 | struct fib6_node *next; | |
1307 | ||
1308 | dir = addr_bit_set(args->addr, fn->fn_bit); | |
1309 | ||
66f5d6ce WW |
1310 | next = dir ? rcu_dereference(fn->right) : |
1311 | rcu_dereference(fn->left); | |
1da177e4 LT |
1312 | |
1313 | if (next) { | |
1314 | fn = next; | |
1315 | continue; | |
1316 | } | |
1da177e4 LT |
1317 | break; |
1318 | } | |
1319 | ||
507c9b1e | 1320 | while (fn) { |
66f5d6ce WW |
1321 | struct fib6_node *subtree = FIB6_SUBTREE(fn); |
1322 | ||
1323 | if (subtree || fn->fn_flags & RTN_RTINFO) { | |
1324 | struct rt6_info *leaf = rcu_dereference(fn->leaf); | |
1da177e4 LT |
1325 | struct rt6key *key; |
1326 | ||
8d1040e8 WW |
1327 | if (!leaf) |
1328 | goto backtrack; | |
1329 | ||
1330 | key = (struct rt6key *) ((u8 *)leaf + args->offset); | |
1da177e4 | 1331 | |
3fc5e044 YH |
1332 | if (ipv6_prefix_equal(&key->addr, args->addr, key->plen)) { |
1333 | #ifdef CONFIG_IPV6_SUBTREES | |
66f5d6ce | 1334 | if (subtree) { |
3e3be275 | 1335 | struct fib6_node *sfn; |
66f5d6ce | 1336 | sfn = fib6_lookup_1(subtree, args + 1); |
3e3be275 HFS |
1337 | if (!sfn) |
1338 | goto backtrack; | |
1339 | fn = sfn; | |
1340 | } | |
3fc5e044 | 1341 | #endif |
3e3be275 | 1342 | if (fn->fn_flags & RTN_RTINFO) |
3fc5e044 YH |
1343 | return fn; |
1344 | } | |
1da177e4 | 1345 | } |
3e3be275 | 1346 | backtrack: |
3fc5e044 YH |
1347 | if (fn->fn_flags & RTN_ROOT) |
1348 | break; | |
1349 | ||
66f5d6ce | 1350 | fn = rcu_dereference(fn->parent); |
1da177e4 LT |
1351 | } |
1352 | ||
1353 | return NULL; | |
1354 | } | |
1355 | ||
66f5d6ce WW |
1356 | /* called with rcu_read_lock() held |
1357 | */ | |
437de07c WY |
1358 | struct fib6_node *fib6_lookup(struct fib6_node *root, const struct in6_addr *daddr, |
1359 | const struct in6_addr *saddr) | |
1da177e4 | 1360 | { |
1da177e4 | 1361 | struct fib6_node *fn; |
825e288e YH |
1362 | struct lookup_args args[] = { |
1363 | { | |
1364 | .offset = offsetof(struct rt6_info, rt6i_dst), | |
1365 | .addr = daddr, | |
1366 | }, | |
1da177e4 | 1367 | #ifdef CONFIG_IPV6_SUBTREES |
825e288e YH |
1368 | { |
1369 | .offset = offsetof(struct rt6_info, rt6i_src), | |
1370 | .addr = saddr, | |
1371 | }, | |
1da177e4 | 1372 | #endif |
825e288e YH |
1373 | { |
1374 | .offset = 0, /* sentinel */ | |
1375 | } | |
1376 | }; | |
1da177e4 | 1377 | |
fefc2a6c | 1378 | fn = fib6_lookup_1(root, daddr ? args : args + 1); |
507c9b1e | 1379 | if (!fn || fn->fn_flags & RTN_TL_ROOT) |
1da177e4 LT |
1380 | fn = root; |
1381 | ||
1382 | return fn; | |
1383 | } | |
1384 | ||
1385 | /* | |
1386 | * Get node with specified destination prefix (and source prefix, | |
1387 | * if subtrees are used) | |
38fbeeee WW |
1388 | * exact_match == true means we try to find fn with exact match of |
1389 | * the passed in prefix addr | |
1390 | * exact_match == false means we try to find fn with longest prefix | |
1391 | * match of the passed in prefix addr. This is useful for finding fn | |
1392 | * for cached route as it will be stored in the exception table under | |
1393 | * the node with longest prefix length. | |
1da177e4 LT |
1394 | */ |
1395 | ||
1396 | ||
437de07c WY |
1397 | static struct fib6_node *fib6_locate_1(struct fib6_node *root, |
1398 | const struct in6_addr *addr, | |
38fbeeee WW |
1399 | int plen, int offset, |
1400 | bool exact_match) | |
1da177e4 | 1401 | { |
38fbeeee | 1402 | struct fib6_node *fn, *prev = NULL; |
1da177e4 LT |
1403 | |
1404 | for (fn = root; fn ; ) { | |
66f5d6ce | 1405 | struct rt6_info *leaf = rcu_dereference(fn->leaf); |
8d1040e8 WW |
1406 | struct rt6key *key; |
1407 | ||
1408 | /* This node is being deleted */ | |
1409 | if (!leaf) { | |
1410 | if (plen <= fn->fn_bit) | |
1411 | goto out; | |
1412 | else | |
1413 | goto next; | |
1414 | } | |
1415 | ||
1416 | key = (struct rt6key *)((u8 *)leaf + offset); | |
1da177e4 LT |
1417 | |
1418 | /* | |
1419 | * Prefix match | |
1420 | */ | |
1421 | if (plen < fn->fn_bit || | |
1422 | !ipv6_prefix_equal(&key->addr, addr, fn->fn_bit)) | |
38fbeeee | 1423 | goto out; |
1da177e4 LT |
1424 | |
1425 | if (plen == fn->fn_bit) | |
1426 | return fn; | |
1427 | ||
38fbeeee WW |
1428 | prev = fn; |
1429 | ||
8d1040e8 | 1430 | next: |
1da177e4 LT |
1431 | /* |
1432 | * We have more bits to go | |
1433 | */ | |
1434 | if (addr_bit_set(addr, fn->fn_bit)) | |
66f5d6ce | 1435 | fn = rcu_dereference(fn->right); |
1da177e4 | 1436 | else |
66f5d6ce | 1437 | fn = rcu_dereference(fn->left); |
1da177e4 | 1438 | } |
38fbeeee WW |
1439 | out: |
1440 | if (exact_match) | |
1441 | return NULL; | |
1442 | else | |
1443 | return prev; | |
1da177e4 LT |
1444 | } |
1445 | ||
437de07c WY |
1446 | struct fib6_node *fib6_locate(struct fib6_node *root, |
1447 | const struct in6_addr *daddr, int dst_len, | |
38fbeeee WW |
1448 | const struct in6_addr *saddr, int src_len, |
1449 | bool exact_match) | |
1da177e4 LT |
1450 | { |
1451 | struct fib6_node *fn; | |
1452 | ||
1453 | fn = fib6_locate_1(root, daddr, dst_len, | |
38fbeeee WW |
1454 | offsetof(struct rt6_info, rt6i_dst), |
1455 | exact_match); | |
1da177e4 LT |
1456 | |
1457 | #ifdef CONFIG_IPV6_SUBTREES | |
1458 | if (src_len) { | |
547b792c | 1459 | WARN_ON(saddr == NULL); |
0e80193b WW |
1460 | if (fn) { |
1461 | struct fib6_node *subtree = FIB6_SUBTREE(fn); | |
1462 | ||
1463 | if (subtree) { | |
1464 | fn = fib6_locate_1(subtree, saddr, src_len, | |
38fbeeee WW |
1465 | offsetof(struct rt6_info, rt6i_src), |
1466 | exact_match); | |
0e80193b WW |
1467 | } |
1468 | } | |
1da177e4 LT |
1469 | } |
1470 | #endif | |
1471 | ||
507c9b1e | 1472 | if (fn && fn->fn_flags & RTN_RTINFO) |
1da177e4 LT |
1473 | return fn; |
1474 | ||
1475 | return NULL; | |
1476 | } | |
1477 | ||
1478 | ||
1479 | /* | |
1480 | * Deletion | |
1481 | * | |
1482 | */ | |
1483 | ||
66f5d6ce WW |
1484 | static struct rt6_info *fib6_find_prefix(struct net *net, |
1485 | struct fib6_table *table, | |
1486 | struct fib6_node *fn) | |
1da177e4 | 1487 | { |
66f5d6ce WW |
1488 | struct fib6_node *child_left, *child_right; |
1489 | ||
507c9b1e | 1490 | if (fn->fn_flags & RTN_ROOT) |
8ed67789 | 1491 | return net->ipv6.ip6_null_entry; |
1da177e4 | 1492 | |
507c9b1e | 1493 | while (fn) { |
66f5d6ce WW |
1494 | child_left = rcu_dereference_protected(fn->left, |
1495 | lockdep_is_held(&table->tb6_lock)); | |
1496 | child_right = rcu_dereference_protected(fn->right, | |
1497 | lockdep_is_held(&table->tb6_lock)); | |
1498 | if (child_left) | |
1499 | return rcu_dereference_protected(child_left->leaf, | |
1500 | lockdep_is_held(&table->tb6_lock)); | |
1501 | if (child_right) | |
1502 | return rcu_dereference_protected(child_right->leaf, | |
1503 | lockdep_is_held(&table->tb6_lock)); | |
1da177e4 | 1504 | |
7fc33165 | 1505 | fn = FIB6_SUBTREE(fn); |
1da177e4 LT |
1506 | } |
1507 | return NULL; | |
1508 | } | |
1509 | ||
1510 | /* | |
1511 | * Called to trim the tree of intermediate nodes when possible. "fn" | |
1512 | * is the node we want to try and remove. | |
66f5d6ce | 1513 | * Need to own table->tb6_lock |
1da177e4 LT |
1514 | */ |
1515 | ||
8ed67789 | 1516 | static struct fib6_node *fib6_repair_tree(struct net *net, |
66f5d6ce WW |
1517 | struct fib6_table *table, |
1518 | struct fib6_node *fn) | |
1da177e4 LT |
1519 | { |
1520 | int children; | |
1521 | int nstate; | |
66f5d6ce | 1522 | struct fib6_node *child; |
94b2cfe0 | 1523 | struct fib6_walker *w; |
1da177e4 LT |
1524 | int iter = 0; |
1525 | ||
1526 | for (;;) { | |
66f5d6ce WW |
1527 | struct fib6_node *fn_r = rcu_dereference_protected(fn->right, |
1528 | lockdep_is_held(&table->tb6_lock)); | |
1529 | struct fib6_node *fn_l = rcu_dereference_protected(fn->left, | |
1530 | lockdep_is_held(&table->tb6_lock)); | |
1531 | struct fib6_node *pn = rcu_dereference_protected(fn->parent, | |
1532 | lockdep_is_held(&table->tb6_lock)); | |
1533 | struct fib6_node *pn_r = rcu_dereference_protected(pn->right, | |
1534 | lockdep_is_held(&table->tb6_lock)); | |
1535 | struct fib6_node *pn_l = rcu_dereference_protected(pn->left, | |
1536 | lockdep_is_held(&table->tb6_lock)); | |
1537 | struct rt6_info *fn_leaf = rcu_dereference_protected(fn->leaf, | |
1538 | lockdep_is_held(&table->tb6_lock)); | |
1539 | struct rt6_info *pn_leaf = rcu_dereference_protected(pn->leaf, | |
1540 | lockdep_is_held(&table->tb6_lock)); | |
1541 | struct rt6_info *new_fn_leaf; | |
1542 | ||
1da177e4 LT |
1543 | RT6_TRACE("fixing tree: plen=%d iter=%d\n", fn->fn_bit, iter); |
1544 | iter++; | |
1545 | ||
547b792c IJ |
1546 | WARN_ON(fn->fn_flags & RTN_RTINFO); |
1547 | WARN_ON(fn->fn_flags & RTN_TL_ROOT); | |
66f5d6ce | 1548 | WARN_ON(fn_leaf); |
1da177e4 LT |
1549 | |
1550 | children = 0; | |
1551 | child = NULL; | |
66f5d6ce WW |
1552 | if (fn_r) |
1553 | child = fn_r, children |= 1; | |
1554 | if (fn_l) | |
1555 | child = fn_l, children |= 2; | |
1da177e4 | 1556 | |
7fc33165 | 1557 | if (children == 3 || FIB6_SUBTREE(fn) |
1da177e4 LT |
1558 | #ifdef CONFIG_IPV6_SUBTREES |
1559 | /* Subtree root (i.e. fn) may have one child */ | |
507c9b1e | 1560 | || (children && fn->fn_flags & RTN_ROOT) |
1da177e4 LT |
1561 | #endif |
1562 | ) { | |
66f5d6ce | 1563 | new_fn_leaf = fib6_find_prefix(net, table, fn); |
1da177e4 | 1564 | #if RT6_DEBUG >= 2 |
66f5d6ce WW |
1565 | if (!new_fn_leaf) { |
1566 | WARN_ON(!new_fn_leaf); | |
1567 | new_fn_leaf = net->ipv6.ip6_null_entry; | |
1da177e4 LT |
1568 | } |
1569 | #endif | |
66f5d6ce WW |
1570 | atomic_inc(&new_fn_leaf->rt6i_ref); |
1571 | rcu_assign_pointer(fn->leaf, new_fn_leaf); | |
1572 | return pn; | |
1da177e4 LT |
1573 | } |
1574 | ||
1da177e4 | 1575 | #ifdef CONFIG_IPV6_SUBTREES |
7fc33165 | 1576 | if (FIB6_SUBTREE(pn) == fn) { |
547b792c | 1577 | WARN_ON(!(fn->fn_flags & RTN_ROOT)); |
66f5d6ce | 1578 | RCU_INIT_POINTER(pn->subtree, NULL); |
1da177e4 LT |
1579 | nstate = FWS_L; |
1580 | } else { | |
547b792c | 1581 | WARN_ON(fn->fn_flags & RTN_ROOT); |
1da177e4 | 1582 | #endif |
66f5d6ce WW |
1583 | if (pn_r == fn) |
1584 | rcu_assign_pointer(pn->right, child); | |
1585 | else if (pn_l == fn) | |
1586 | rcu_assign_pointer(pn->left, child); | |
1da177e4 | 1587 | #if RT6_DEBUG >= 2 |
547b792c IJ |
1588 | else |
1589 | WARN_ON(1); | |
1da177e4 LT |
1590 | #endif |
1591 | if (child) | |
66f5d6ce | 1592 | rcu_assign_pointer(child->parent, pn); |
1da177e4 LT |
1593 | nstate = FWS_R; |
1594 | #ifdef CONFIG_IPV6_SUBTREES | |
1595 | } | |
1596 | #endif | |
1597 | ||
9a03cd8f MK |
1598 | read_lock(&net->ipv6.fib6_walker_lock); |
1599 | FOR_WALKERS(net, w) { | |
507c9b1e | 1600 | if (!child) { |
2b760fcf | 1601 | if (w->node == fn) { |
1da177e4 LT |
1602 | RT6_TRACE("W %p adjusted by delnode 1, s=%d/%d\n", w, w->state, nstate); |
1603 | w->node = pn; | |
1604 | w->state = nstate; | |
1605 | } | |
1606 | } else { | |
1da177e4 LT |
1607 | if (w->node == fn) { |
1608 | w->node = child; | |
1609 | if (children&2) { | |
1610 | RT6_TRACE("W %p adjusted by delnode 2, s=%d\n", w, w->state); | |
8db46f1d | 1611 | w->state = w->state >= FWS_R ? FWS_U : FWS_INIT; |
1da177e4 LT |
1612 | } else { |
1613 | RT6_TRACE("W %p adjusted by delnode 2, s=%d\n", w, w->state); | |
8db46f1d | 1614 | w->state = w->state >= FWS_C ? FWS_U : FWS_INIT; |
1da177e4 LT |
1615 | } |
1616 | } | |
1617 | } | |
1618 | } | |
9a03cd8f | 1619 | read_unlock(&net->ipv6.fib6_walker_lock); |
1da177e4 | 1620 | |
81eb8447 | 1621 | node_free(net, fn); |
507c9b1e | 1622 | if (pn->fn_flags & RTN_RTINFO || FIB6_SUBTREE(pn)) |
1da177e4 LT |
1623 | return pn; |
1624 | ||
66f5d6ce WW |
1625 | RCU_INIT_POINTER(pn->leaf, NULL); |
1626 | rt6_release(pn_leaf); | |
1da177e4 LT |
1627 | fn = pn; |
1628 | } | |
1629 | } | |
1630 | ||
66f5d6ce WW |
1631 | static void fib6_del_route(struct fib6_table *table, struct fib6_node *fn, |
1632 | struct rt6_info __rcu **rtp, struct nl_info *info) | |
1da177e4 | 1633 | { |
94b2cfe0 | 1634 | struct fib6_walker *w; |
66f5d6ce WW |
1635 | struct rt6_info *rt = rcu_dereference_protected(*rtp, |
1636 | lockdep_is_held(&table->tb6_lock)); | |
c572872f | 1637 | struct net *net = info->nl_net; |
1da177e4 LT |
1638 | |
1639 | RT6_TRACE("fib6_del_route\n"); | |
1640 | ||
2b760fcf WW |
1641 | WARN_ON_ONCE(rt->rt6i_flags & RTF_CACHE); |
1642 | ||
1da177e4 | 1643 | /* Unlink it */ |
d8d1f30b | 1644 | *rtp = rt->dst.rt6_next; |
1da177e4 | 1645 | rt->rt6i_node = NULL; |
c572872f BT |
1646 | net->ipv6.rt6_stats->fib_rt_entries--; |
1647 | net->ipv6.rt6_stats->fib_discarded_routes++; | |
1da177e4 | 1648 | |
2b760fcf WW |
1649 | /* Flush all cached dst in exception table */ |
1650 | rt6_flush_exceptions(rt); | |
1651 | ||
f11e6659 | 1652 | /* Reset round-robin state, if necessary */ |
66f5d6ce | 1653 | if (rcu_access_pointer(fn->rr_ptr) == rt) |
f11e6659 DM |
1654 | fn->rr_ptr = NULL; |
1655 | ||
51ebd318 ND |
1656 | /* Remove this entry from other siblings */ |
1657 | if (rt->rt6i_nsiblings) { | |
1658 | struct rt6_info *sibling, *next_sibling; | |
1659 | ||
1660 | list_for_each_entry_safe(sibling, next_sibling, | |
1661 | &rt->rt6i_siblings, rt6i_siblings) | |
1662 | sibling->rt6i_nsiblings--; | |
1663 | rt->rt6i_nsiblings = 0; | |
1664 | list_del_init(&rt->rt6i_siblings); | |
1665 | } | |
1666 | ||
1da177e4 | 1667 | /* Adjust walkers */ |
9a03cd8f MK |
1668 | read_lock(&net->ipv6.fib6_walker_lock); |
1669 | FOR_WALKERS(net, w) { | |
1da177e4 LT |
1670 | if (w->state == FWS_C && w->leaf == rt) { |
1671 | RT6_TRACE("walker %p adjusted by delroute\n", w); | |
66f5d6ce WW |
1672 | w->leaf = rcu_dereference_protected(rt->dst.rt6_next, |
1673 | lockdep_is_held(&table->tb6_lock)); | |
507c9b1e | 1674 | if (!w->leaf) |
1da177e4 LT |
1675 | w->state = FWS_U; |
1676 | } | |
1677 | } | |
9a03cd8f | 1678 | read_unlock(&net->ipv6.fib6_walker_lock); |
1da177e4 | 1679 | |
1da177e4 | 1680 | /* If it was last route, expunge its radix tree node */ |
66f5d6ce | 1681 | if (!rcu_access_pointer(fn->leaf)) { |
1da177e4 | 1682 | fn->fn_flags &= ~RTN_RTINFO; |
c572872f | 1683 | net->ipv6.rt6_stats->fib_route_nodes--; |
66f5d6ce | 1684 | fn = fib6_repair_tree(net, table, fn); |
1da177e4 LT |
1685 | } |
1686 | ||
6e9e16e6 | 1687 | fib6_purge_rt(rt, fn, net); |
1da177e4 | 1688 | |
df77fe4d | 1689 | call_fib6_entry_notifiers(net, FIB_EVENT_ENTRY_DEL, rt); |
16a16cd3 DA |
1690 | if (!info->skip_notify) |
1691 | inet6_rt_notify(RTM_DELROUTE, rt, info, 0); | |
1da177e4 LT |
1692 | rt6_release(rt); |
1693 | } | |
1694 | ||
66f5d6ce | 1695 | /* Need to own table->tb6_lock */ |
86872cb5 | 1696 | int fib6_del(struct rt6_info *rt, struct nl_info *info) |
1da177e4 | 1697 | { |
4e587ea7 WW |
1698 | struct fib6_node *fn = rcu_dereference_protected(rt->rt6i_node, |
1699 | lockdep_is_held(&rt->rt6i_table->tb6_lock)); | |
66f5d6ce | 1700 | struct fib6_table *table = rt->rt6i_table; |
8ed67789 | 1701 | struct net *net = info->nl_net; |
66f5d6ce WW |
1702 | struct rt6_info __rcu **rtp; |
1703 | struct rt6_info __rcu **rtp_next; | |
1da177e4 LT |
1704 | |
1705 | #if RT6_DEBUG >= 2 | |
8db46f1d | 1706 | if (rt->dst.obsolete > 0) { |
53b24b8f | 1707 | WARN_ON(fn); |
1da177e4 LT |
1708 | return -ENOENT; |
1709 | } | |
1710 | #endif | |
507c9b1e | 1711 | if (!fn || rt == net->ipv6.ip6_null_entry) |
1da177e4 LT |
1712 | return -ENOENT; |
1713 | ||
547b792c | 1714 | WARN_ON(!(fn->fn_flags & RTN_RTINFO)); |
1da177e4 | 1715 | |
2b760fcf WW |
1716 | /* remove cached dst from exception table */ |
1717 | if (rt->rt6i_flags & RTF_CACHE) | |
1718 | return rt6_remove_exception_rt(rt); | |
1da177e4 LT |
1719 | |
1720 | /* | |
1721 | * Walk the leaf entries looking for ourself | |
1722 | */ | |
1723 | ||
66f5d6ce WW |
1724 | for (rtp = &fn->leaf; *rtp; rtp = rtp_next) { |
1725 | struct rt6_info *cur = rcu_dereference_protected(*rtp, | |
1726 | lockdep_is_held(&table->tb6_lock)); | |
1727 | if (rt == cur) { | |
1728 | fib6_del_route(table, fn, rtp, info); | |
1da177e4 LT |
1729 | return 0; |
1730 | } | |
66f5d6ce | 1731 | rtp_next = &cur->dst.rt6_next; |
1da177e4 LT |
1732 | } |
1733 | return -ENOENT; | |
1734 | } | |
1735 | ||
1736 | /* | |
1737 | * Tree traversal function. | |
1738 | * | |
1739 | * Certainly, it is not interrupt safe. | |
1740 | * However, it is internally reenterable wrt itself and fib6_add/fib6_del. | |
1741 | * It means, that we can modify tree during walking | |
1742 | * and use this function for garbage collection, clone pruning, | |
1ab1457c | 1743 | * cleaning tree when a device goes down etc. etc. |
1da177e4 LT |
1744 | * |
1745 | * It guarantees that every node will be traversed, | |
1746 | * and that it will be traversed only once. | |
1747 | * | |
1748 | * Callback function w->func may return: | |
1749 | * 0 -> continue walking. | |
1750 | * positive value -> walking is suspended (used by tree dumps, | |
1751 | * and probably by gc, if it will be split to several slices) | |
1752 | * negative value -> terminate walking. | |
1753 | * | |
1754 | * The function itself returns: | |
1755 | * 0 -> walk is complete. | |
1756 | * >0 -> walk is incomplete (i.e. suspended) | |
1757 | * <0 -> walk is terminated by an error. | |
66f5d6ce WW |
1758 | * |
1759 | * This function is called with tb6_lock held. | |
1da177e4 LT |
1760 | */ |
1761 | ||
94b2cfe0 | 1762 | static int fib6_walk_continue(struct fib6_walker *w) |
1da177e4 | 1763 | { |
66f5d6ce | 1764 | struct fib6_node *fn, *pn, *left, *right; |
1da177e4 | 1765 | |
2b760fcf WW |
1766 | /* w->root should always be table->tb6_root */ |
1767 | WARN_ON_ONCE(!(w->root->fn_flags & RTN_TL_ROOT)); | |
1768 | ||
1da177e4 LT |
1769 | for (;;) { |
1770 | fn = w->node; | |
507c9b1e | 1771 | if (!fn) |
1da177e4 LT |
1772 | return 0; |
1773 | ||
1da177e4 LT |
1774 | switch (w->state) { |
1775 | #ifdef CONFIG_IPV6_SUBTREES | |
1776 | case FWS_S: | |
7fc33165 YH |
1777 | if (FIB6_SUBTREE(fn)) { |
1778 | w->node = FIB6_SUBTREE(fn); | |
1da177e4 LT |
1779 | continue; |
1780 | } | |
1781 | w->state = FWS_L; | |
1ab1457c | 1782 | #endif |
275757e6 | 1783 | /* fall through */ |
1da177e4 | 1784 | case FWS_L: |
66f5d6ce WW |
1785 | left = rcu_dereference_protected(fn->left, 1); |
1786 | if (left) { | |
1787 | w->node = left; | |
1da177e4 LT |
1788 | w->state = FWS_INIT; |
1789 | continue; | |
1790 | } | |
1791 | w->state = FWS_R; | |
275757e6 | 1792 | /* fall through */ |
1da177e4 | 1793 | case FWS_R: |
66f5d6ce WW |
1794 | right = rcu_dereference_protected(fn->right, 1); |
1795 | if (right) { | |
1796 | w->node = right; | |
1da177e4 LT |
1797 | w->state = FWS_INIT; |
1798 | continue; | |
1799 | } | |
1800 | w->state = FWS_C; | |
66f5d6ce | 1801 | w->leaf = rcu_dereference_protected(fn->leaf, 1); |
275757e6 | 1802 | /* fall through */ |
1da177e4 | 1803 | case FWS_C: |
507c9b1e | 1804 | if (w->leaf && fn->fn_flags & RTN_RTINFO) { |
2bec5a36 PM |
1805 | int err; |
1806 | ||
fa809e2f ED |
1807 | if (w->skip) { |
1808 | w->skip--; | |
1c265854 | 1809 | goto skip; |
2bec5a36 PM |
1810 | } |
1811 | ||
1812 | err = w->func(w); | |
1da177e4 LT |
1813 | if (err) |
1814 | return err; | |
2bec5a36 PM |
1815 | |
1816 | w->count++; | |
1da177e4 LT |
1817 | continue; |
1818 | } | |
1c265854 | 1819 | skip: |
1da177e4 | 1820 | w->state = FWS_U; |
275757e6 | 1821 | /* fall through */ |
1da177e4 LT |
1822 | case FWS_U: |
1823 | if (fn == w->root) | |
1824 | return 0; | |
66f5d6ce WW |
1825 | pn = rcu_dereference_protected(fn->parent, 1); |
1826 | left = rcu_dereference_protected(pn->left, 1); | |
1827 | right = rcu_dereference_protected(pn->right, 1); | |
1da177e4 LT |
1828 | w->node = pn; |
1829 | #ifdef CONFIG_IPV6_SUBTREES | |
7fc33165 | 1830 | if (FIB6_SUBTREE(pn) == fn) { |
547b792c | 1831 | WARN_ON(!(fn->fn_flags & RTN_ROOT)); |
1da177e4 LT |
1832 | w->state = FWS_L; |
1833 | continue; | |
1834 | } | |
1835 | #endif | |
66f5d6ce | 1836 | if (left == fn) { |
1da177e4 LT |
1837 | w->state = FWS_R; |
1838 | continue; | |
1839 | } | |
66f5d6ce | 1840 | if (right == fn) { |
1da177e4 | 1841 | w->state = FWS_C; |
66f5d6ce | 1842 | w->leaf = rcu_dereference_protected(w->node->leaf, 1); |
1da177e4 LT |
1843 | continue; |
1844 | } | |
1845 | #if RT6_DEBUG >= 2 | |
547b792c | 1846 | WARN_ON(1); |
1da177e4 LT |
1847 | #endif |
1848 | } | |
1849 | } | |
1850 | } | |
1851 | ||
9a03cd8f | 1852 | static int fib6_walk(struct net *net, struct fib6_walker *w) |
1da177e4 LT |
1853 | { |
1854 | int res; | |
1855 | ||
1856 | w->state = FWS_INIT; | |
1857 | w->node = w->root; | |
1858 | ||
9a03cd8f | 1859 | fib6_walker_link(net, w); |
1da177e4 LT |
1860 | res = fib6_walk_continue(w); |
1861 | if (res <= 0) | |
9a03cd8f | 1862 | fib6_walker_unlink(net, w); |
1da177e4 LT |
1863 | return res; |
1864 | } | |
1865 | ||
94b2cfe0 | 1866 | static int fib6_clean_node(struct fib6_walker *w) |
1da177e4 LT |
1867 | { |
1868 | int res; | |
1869 | struct rt6_info *rt; | |
94b2cfe0 | 1870 | struct fib6_cleaner *c = container_of(w, struct fib6_cleaner, w); |
ec7d43c2 BT |
1871 | struct nl_info info = { |
1872 | .nl_net = c->net, | |
1873 | }; | |
1da177e4 | 1874 | |
327571cb HFS |
1875 | if (c->sernum != FIB6_NO_SERNUM_CHANGE && |
1876 | w->node->fn_sernum != c->sernum) | |
1877 | w->node->fn_sernum = c->sernum; | |
1878 | ||
1879 | if (!c->func) { | |
1880 | WARN_ON_ONCE(c->sernum == FIB6_NO_SERNUM_CHANGE); | |
1881 | w->leaf = NULL; | |
1882 | return 0; | |
1883 | } | |
1884 | ||
66f5d6ce | 1885 | for_each_fib6_walker_rt(w) { |
1da177e4 LT |
1886 | res = c->func(rt, c->arg); |
1887 | if (res < 0) { | |
1888 | w->leaf = rt; | |
528c4ceb | 1889 | res = fib6_del(rt, &info); |
1da177e4 LT |
1890 | if (res) { |
1891 | #if RT6_DEBUG >= 2 | |
91df42be | 1892 | pr_debug("%s: del failed: rt=%p@%p err=%d\n", |
4e587ea7 WW |
1893 | __func__, rt, |
1894 | rcu_access_pointer(rt->rt6i_node), | |
1895 | res); | |
1da177e4 LT |
1896 | #endif |
1897 | continue; | |
1898 | } | |
1899 | return 0; | |
1900 | } | |
547b792c | 1901 | WARN_ON(res != 0); |
1da177e4 LT |
1902 | } |
1903 | w->leaf = rt; | |
1904 | return 0; | |
1905 | } | |
1906 | ||
1907 | /* | |
1908 | * Convenient frontend to tree walker. | |
1ab1457c | 1909 | * |
1da177e4 LT |
1910 | * func is called on each route. |
1911 | * It may return -1 -> delete this route. | |
1912 | * 0 -> continue walking | |
1da177e4 LT |
1913 | */ |
1914 | ||
ec7d43c2 | 1915 | static void fib6_clean_tree(struct net *net, struct fib6_node *root, |
8ce11e6a | 1916 | int (*func)(struct rt6_info *, void *arg), |
2b760fcf | 1917 | int sernum, void *arg) |
1da177e4 | 1918 | { |
94b2cfe0 | 1919 | struct fib6_cleaner c; |
1da177e4 LT |
1920 | |
1921 | c.w.root = root; | |
1922 | c.w.func = fib6_clean_node; | |
2bec5a36 PM |
1923 | c.w.count = 0; |
1924 | c.w.skip = 0; | |
1da177e4 | 1925 | c.func = func; |
327571cb | 1926 | c.sernum = sernum; |
1da177e4 | 1927 | c.arg = arg; |
ec7d43c2 | 1928 | c.net = net; |
1da177e4 | 1929 | |
9a03cd8f | 1930 | fib6_walk(net, &c.w); |
1da177e4 LT |
1931 | } |
1932 | ||
327571cb HFS |
1933 | static void __fib6_clean_all(struct net *net, |
1934 | int (*func)(struct rt6_info *, void *), | |
1935 | int sernum, void *arg) | |
c71099ac | 1936 | { |
c71099ac | 1937 | struct fib6_table *table; |
58f09b78 | 1938 | struct hlist_head *head; |
1b43af54 | 1939 | unsigned int h; |
c71099ac | 1940 | |
1b43af54 | 1941 | rcu_read_lock(); |
a33bc5c1 | 1942 | for (h = 0; h < FIB6_TABLE_HASHSZ; h++) { |
f3db4851 | 1943 | head = &net->ipv6.fib_table_hash[h]; |
b67bfe0d | 1944 | hlist_for_each_entry_rcu(table, head, tb6_hlist) { |
66f5d6ce | 1945 | spin_lock_bh(&table->tb6_lock); |
ec7d43c2 | 1946 | fib6_clean_tree(net, &table->tb6_root, |
2b760fcf | 1947 | func, sernum, arg); |
66f5d6ce | 1948 | spin_unlock_bh(&table->tb6_lock); |
c71099ac TG |
1949 | } |
1950 | } | |
1b43af54 | 1951 | rcu_read_unlock(); |
c71099ac TG |
1952 | } |
1953 | ||
327571cb HFS |
1954 | void fib6_clean_all(struct net *net, int (*func)(struct rt6_info *, void *), |
1955 | void *arg) | |
1956 | { | |
1957 | __fib6_clean_all(net, func, FIB6_NO_SERNUM_CHANGE, arg); | |
1958 | } | |
1959 | ||
705f1c86 HFS |
1960 | static void fib6_flush_trees(struct net *net) |
1961 | { | |
812918c4 | 1962 | int new_sernum = fib6_new_sernum(net); |
705f1c86 | 1963 | |
327571cb | 1964 | __fib6_clean_all(net, NULL, new_sernum, NULL); |
705f1c86 HFS |
1965 | } |
1966 | ||
1da177e4 LT |
1967 | /* |
1968 | * Garbage collection | |
1969 | */ | |
1970 | ||
1da177e4 LT |
1971 | static int fib6_age(struct rt6_info *rt, void *arg) |
1972 | { | |
3570df91 | 1973 | struct fib6_gc_args *gc_args = arg; |
1da177e4 LT |
1974 | unsigned long now = jiffies; |
1975 | ||
1976 | /* | |
1977 | * check addrconf expiration here. | |
1978 | * Routes are expired even if they are in use. | |
1da177e4 LT |
1979 | */ |
1980 | ||
d1918542 DM |
1981 | if (rt->rt6i_flags & RTF_EXPIRES && rt->dst.expires) { |
1982 | if (time_after(now, rt->dst.expires)) { | |
1da177e4 | 1983 | RT6_TRACE("expiring %p\n", rt); |
1da177e4 LT |
1984 | return -1; |
1985 | } | |
3570df91 | 1986 | gc_args->more++; |
1da177e4 LT |
1987 | } |
1988 | ||
c757faa8 WW |
1989 | /* Also age clones in the exception table. |
1990 | * Note, that clones are aged out | |
1991 | * only if they are not in use now. | |
1992 | */ | |
1993 | rt6_age_exceptions(rt, gc_args, now); | |
1994 | ||
1da177e4 LT |
1995 | return 0; |
1996 | } | |
1997 | ||
2ac3ac8f | 1998 | void fib6_run_gc(unsigned long expires, struct net *net, bool force) |
1da177e4 | 1999 | { |
3570df91 | 2000 | struct fib6_gc_args gc_args; |
49a18d86 MK |
2001 | unsigned long now; |
2002 | ||
2ac3ac8f | 2003 | if (force) { |
3dc94f93 MK |
2004 | spin_lock_bh(&net->ipv6.fib6_gc_lock); |
2005 | } else if (!spin_trylock_bh(&net->ipv6.fib6_gc_lock)) { | |
2ac3ac8f MK |
2006 | mod_timer(&net->ipv6.ip6_fib_timer, jiffies + HZ); |
2007 | return; | |
1da177e4 | 2008 | } |
2ac3ac8f MK |
2009 | gc_args.timeout = expires ? (int)expires : |
2010 | net->ipv6.sysctl.ip6_rt_gc_interval; | |
db916649 | 2011 | gc_args.more = 0; |
f3db4851 | 2012 | |
3570df91 | 2013 | fib6_clean_all(net, fib6_age, &gc_args); |
49a18d86 MK |
2014 | now = jiffies; |
2015 | net->ipv6.ip6_rt_last_gc = now; | |
1da177e4 LT |
2016 | |
2017 | if (gc_args.more) | |
c8a45222 | 2018 | mod_timer(&net->ipv6.ip6_fib_timer, |
49a18d86 | 2019 | round_jiffies(now |
c8a45222 | 2020 | + net->ipv6.sysctl.ip6_rt_gc_interval)); |
417f28bb SH |
2021 | else |
2022 | del_timer(&net->ipv6.ip6_fib_timer); | |
3dc94f93 | 2023 | spin_unlock_bh(&net->ipv6.fib6_gc_lock); |
1da177e4 LT |
2024 | } |
2025 | ||
5b7c931d DL |
2026 | static void fib6_gc_timer_cb(unsigned long arg) |
2027 | { | |
2ac3ac8f | 2028 | fib6_run_gc(0, (struct net *)arg, true); |
5b7c931d DL |
2029 | } |
2030 | ||
2c8c1e72 | 2031 | static int __net_init fib6_net_init(struct net *net) |
1da177e4 | 2032 | { |
10da66f7 | 2033 | size_t size = sizeof(struct hlist_head) * FIB6_TABLE_HASHSZ; |
16ab6d7d IS |
2034 | int err; |
2035 | ||
2036 | err = fib6_notifier_init(net); | |
2037 | if (err) | |
2038 | return err; | |
10da66f7 | 2039 | |
3dc94f93 | 2040 | spin_lock_init(&net->ipv6.fib6_gc_lock); |
9a03cd8f MK |
2041 | rwlock_init(&net->ipv6.fib6_walker_lock); |
2042 | INIT_LIST_HEAD(&net->ipv6.fib6_walkers); | |
417f28bb | 2043 | setup_timer(&net->ipv6.ip6_fib_timer, fib6_gc_timer_cb, (unsigned long)net); |
63152fc0 | 2044 | |
c572872f BT |
2045 | net->ipv6.rt6_stats = kzalloc(sizeof(*net->ipv6.rt6_stats), GFP_KERNEL); |
2046 | if (!net->ipv6.rt6_stats) | |
2047 | goto out_timer; | |
2048 | ||
10da66f7 ED |
2049 | /* Avoid false sharing : Use at least a full cache line */ |
2050 | size = max_t(size_t, size, L1_CACHE_BYTES); | |
2051 | ||
2052 | net->ipv6.fib_table_hash = kzalloc(size, GFP_KERNEL); | |
58f09b78 | 2053 | if (!net->ipv6.fib_table_hash) |
c572872f | 2054 | goto out_rt6_stats; |
e0b85590 | 2055 | |
58f09b78 DL |
2056 | net->ipv6.fib6_main_tbl = kzalloc(sizeof(*net->ipv6.fib6_main_tbl), |
2057 | GFP_KERNEL); | |
2058 | if (!net->ipv6.fib6_main_tbl) | |
e0b85590 DL |
2059 | goto out_fib_table_hash; |
2060 | ||
58f09b78 | 2061 | net->ipv6.fib6_main_tbl->tb6_id = RT6_TABLE_MAIN; |
66f5d6ce WW |
2062 | rcu_assign_pointer(net->ipv6.fib6_main_tbl->tb6_root.leaf, |
2063 | net->ipv6.ip6_null_entry); | |
58f09b78 DL |
2064 | net->ipv6.fib6_main_tbl->tb6_root.fn_flags = |
2065 | RTN_ROOT | RTN_TL_ROOT | RTN_RTINFO; | |
8e773277 | 2066 | inet_peer_base_init(&net->ipv6.fib6_main_tbl->tb6_peers); |
e0b85590 DL |
2067 | |
2068 | #ifdef CONFIG_IPV6_MULTIPLE_TABLES | |
58f09b78 DL |
2069 | net->ipv6.fib6_local_tbl = kzalloc(sizeof(*net->ipv6.fib6_local_tbl), |
2070 | GFP_KERNEL); | |
2071 | if (!net->ipv6.fib6_local_tbl) | |
e0b85590 | 2072 | goto out_fib6_main_tbl; |
58f09b78 | 2073 | net->ipv6.fib6_local_tbl->tb6_id = RT6_TABLE_LOCAL; |
66f5d6ce WW |
2074 | rcu_assign_pointer(net->ipv6.fib6_local_tbl->tb6_root.leaf, |
2075 | net->ipv6.ip6_null_entry); | |
58f09b78 DL |
2076 | net->ipv6.fib6_local_tbl->tb6_root.fn_flags = |
2077 | RTN_ROOT | RTN_TL_ROOT | RTN_RTINFO; | |
8e773277 | 2078 | inet_peer_base_init(&net->ipv6.fib6_local_tbl->tb6_peers); |
e0b85590 | 2079 | #endif |
58f09b78 | 2080 | fib6_tables_init(net); |
f845ab6b | 2081 | |
417f28bb | 2082 | return 0; |
d63bddbe | 2083 | |
e0b85590 | 2084 | #ifdef CONFIG_IPV6_MULTIPLE_TABLES |
e0b85590 | 2085 | out_fib6_main_tbl: |
58f09b78 | 2086 | kfree(net->ipv6.fib6_main_tbl); |
e0b85590 | 2087 | #endif |
e0b85590 | 2088 | out_fib_table_hash: |
58f09b78 | 2089 | kfree(net->ipv6.fib_table_hash); |
c572872f BT |
2090 | out_rt6_stats: |
2091 | kfree(net->ipv6.rt6_stats); | |
63152fc0 | 2092 | out_timer: |
16ab6d7d | 2093 | fib6_notifier_exit(net); |
417f28bb | 2094 | return -ENOMEM; |
8db46f1d | 2095 | } |
58f09b78 DL |
2096 | |
2097 | static void fib6_net_exit(struct net *net) | |
2098 | { | |
ba1cc08d SD |
2099 | unsigned int i; |
2100 | ||
8ed67789 | 2101 | rt6_ifdown(net, NULL); |
417f28bb SH |
2102 | del_timer_sync(&net->ipv6.ip6_fib_timer); |
2103 | ||
32a805ba | 2104 | for (i = 0; i < FIB6_TABLE_HASHSZ; i++) { |
ba1cc08d SD |
2105 | struct hlist_head *head = &net->ipv6.fib_table_hash[i]; |
2106 | struct hlist_node *tmp; | |
2107 | struct fib6_table *tb; | |
2108 | ||
2109 | hlist_for_each_entry_safe(tb, tmp, head, tb6_hlist) { | |
2110 | hlist_del(&tb->tb6_hlist); | |
2111 | fib6_free_table(tb); | |
2112 | } | |
2113 | } | |
2114 | ||
58f09b78 | 2115 | kfree(net->ipv6.fib_table_hash); |
c572872f | 2116 | kfree(net->ipv6.rt6_stats); |
16ab6d7d | 2117 | fib6_notifier_exit(net); |
58f09b78 DL |
2118 | } |
2119 | ||
2120 | static struct pernet_operations fib6_net_ops = { | |
2121 | .init = fib6_net_init, | |
2122 | .exit = fib6_net_exit, | |
2123 | }; | |
2124 | ||
2125 | int __init fib6_init(void) | |
2126 | { | |
2127 | int ret = -ENOMEM; | |
63152fc0 | 2128 | |
58f09b78 DL |
2129 | fib6_node_kmem = kmem_cache_create("fib6_nodes", |
2130 | sizeof(struct fib6_node), | |
2131 | 0, SLAB_HWCACHE_ALIGN, | |
2132 | NULL); | |
2133 | if (!fib6_node_kmem) | |
2134 | goto out; | |
2135 | ||
2136 | ret = register_pernet_subsys(&fib6_net_ops); | |
2137 | if (ret) | |
c572872f | 2138 | goto out_kmem_cache_create; |
e8803b6c DM |
2139 | |
2140 | ret = __rtnl_register(PF_INET6, RTM_GETROUTE, NULL, inet6_dump_fib, | |
b97bac64 | 2141 | 0); |
e8803b6c DM |
2142 | if (ret) |
2143 | goto out_unregister_subsys; | |
705f1c86 HFS |
2144 | |
2145 | __fib6_flush_trees = fib6_flush_trees; | |
58f09b78 DL |
2146 | out: |
2147 | return ret; | |
2148 | ||
e8803b6c DM |
2149 | out_unregister_subsys: |
2150 | unregister_pernet_subsys(&fib6_net_ops); | |
d63bddbe DL |
2151 | out_kmem_cache_create: |
2152 | kmem_cache_destroy(fib6_node_kmem); | |
2153 | goto out; | |
1da177e4 LT |
2154 | } |
2155 | ||
2156 | void fib6_gc_cleanup(void) | |
2157 | { | |
58f09b78 | 2158 | unregister_pernet_subsys(&fib6_net_ops); |
1da177e4 LT |
2159 | kmem_cache_destroy(fib6_node_kmem); |
2160 | } | |
8d2ca1d7 HFS |
2161 | |
2162 | #ifdef CONFIG_PROC_FS | |
2163 | ||
2164 | struct ipv6_route_iter { | |
2165 | struct seq_net_private p; | |
94b2cfe0 | 2166 | struct fib6_walker w; |
8d2ca1d7 HFS |
2167 | loff_t skip; |
2168 | struct fib6_table *tbl; | |
42b18706 | 2169 | int sernum; |
8d2ca1d7 HFS |
2170 | }; |
2171 | ||
2172 | static int ipv6_route_seq_show(struct seq_file *seq, void *v) | |
2173 | { | |
2174 | struct rt6_info *rt = v; | |
2175 | struct ipv6_route_iter *iter = seq->private; | |
2176 | ||
2177 | seq_printf(seq, "%pi6 %02x ", &rt->rt6i_dst.addr, rt->rt6i_dst.plen); | |
2178 | ||
2179 | #ifdef CONFIG_IPV6_SUBTREES | |
2180 | seq_printf(seq, "%pi6 %02x ", &rt->rt6i_src.addr, rt->rt6i_src.plen); | |
2181 | #else | |
2182 | seq_puts(seq, "00000000000000000000000000000000 00 "); | |
2183 | #endif | |
2184 | if (rt->rt6i_flags & RTF_GATEWAY) | |
2185 | seq_printf(seq, "%pi6", &rt->rt6i_gateway); | |
2186 | else | |
2187 | seq_puts(seq, "00000000000000000000000000000000"); | |
2188 | ||
2189 | seq_printf(seq, " %08x %08x %08x %08x %8s\n", | |
2190 | rt->rt6i_metric, atomic_read(&rt->dst.__refcnt), | |
2191 | rt->dst.__use, rt->rt6i_flags, | |
2192 | rt->dst.dev ? rt->dst.dev->name : ""); | |
2193 | iter->w.leaf = NULL; | |
2194 | return 0; | |
2195 | } | |
2196 | ||
94b2cfe0 | 2197 | static int ipv6_route_yield(struct fib6_walker *w) |
8d2ca1d7 HFS |
2198 | { |
2199 | struct ipv6_route_iter *iter = w->args; | |
2200 | ||
2201 | if (!iter->skip) | |
2202 | return 1; | |
2203 | ||
2204 | do { | |
66f5d6ce WW |
2205 | iter->w.leaf = rcu_dereference_protected( |
2206 | iter->w.leaf->dst.rt6_next, | |
2207 | lockdep_is_held(&iter->tbl->tb6_lock)); | |
8d2ca1d7 HFS |
2208 | iter->skip--; |
2209 | if (!iter->skip && iter->w.leaf) | |
2210 | return 1; | |
2211 | } while (iter->w.leaf); | |
2212 | ||
2213 | return 0; | |
2214 | } | |
2215 | ||
9a03cd8f MK |
2216 | static void ipv6_route_seq_setup_walk(struct ipv6_route_iter *iter, |
2217 | struct net *net) | |
8d2ca1d7 HFS |
2218 | { |
2219 | memset(&iter->w, 0, sizeof(iter->w)); | |
2220 | iter->w.func = ipv6_route_yield; | |
2221 | iter->w.root = &iter->tbl->tb6_root; | |
2222 | iter->w.state = FWS_INIT; | |
2223 | iter->w.node = iter->w.root; | |
2224 | iter->w.args = iter; | |
0a67d3ef | 2225 | iter->sernum = iter->w.root->fn_sernum; |
8d2ca1d7 | 2226 | INIT_LIST_HEAD(&iter->w.lh); |
9a03cd8f | 2227 | fib6_walker_link(net, &iter->w); |
8d2ca1d7 HFS |
2228 | } |
2229 | ||
2230 | static struct fib6_table *ipv6_route_seq_next_table(struct fib6_table *tbl, | |
2231 | struct net *net) | |
2232 | { | |
2233 | unsigned int h; | |
2234 | struct hlist_node *node; | |
2235 | ||
2236 | if (tbl) { | |
2237 | h = (tbl->tb6_id & (FIB6_TABLE_HASHSZ - 1)) + 1; | |
2238 | node = rcu_dereference_bh(hlist_next_rcu(&tbl->tb6_hlist)); | |
2239 | } else { | |
2240 | h = 0; | |
2241 | node = NULL; | |
2242 | } | |
2243 | ||
2244 | while (!node && h < FIB6_TABLE_HASHSZ) { | |
2245 | node = rcu_dereference_bh( | |
2246 | hlist_first_rcu(&net->ipv6.fib_table_hash[h++])); | |
2247 | } | |
2248 | return hlist_entry_safe(node, struct fib6_table, tb6_hlist); | |
2249 | } | |
2250 | ||
0a67d3ef HFS |
2251 | static void ipv6_route_check_sernum(struct ipv6_route_iter *iter) |
2252 | { | |
2253 | if (iter->sernum != iter->w.root->fn_sernum) { | |
2254 | iter->sernum = iter->w.root->fn_sernum; | |
2255 | iter->w.state = FWS_INIT; | |
2256 | iter->w.node = iter->w.root; | |
2257 | WARN_ON(iter->w.skip); | |
2258 | iter->w.skip = iter->w.count; | |
2259 | } | |
2260 | } | |
2261 | ||
8d2ca1d7 HFS |
2262 | static void *ipv6_route_seq_next(struct seq_file *seq, void *v, loff_t *pos) |
2263 | { | |
2264 | int r; | |
2265 | struct rt6_info *n; | |
2266 | struct net *net = seq_file_net(seq); | |
2267 | struct ipv6_route_iter *iter = seq->private; | |
2268 | ||
2269 | if (!v) | |
2270 | goto iter_table; | |
2271 | ||
d0e60206 | 2272 | n = rcu_dereference_bh(((struct rt6_info *)v)->dst.rt6_next); |
8d2ca1d7 HFS |
2273 | if (n) { |
2274 | ++*pos; | |
2275 | return n; | |
2276 | } | |
2277 | ||
2278 | iter_table: | |
0a67d3ef | 2279 | ipv6_route_check_sernum(iter); |
66f5d6ce | 2280 | spin_lock_bh(&iter->tbl->tb6_lock); |
8d2ca1d7 | 2281 | r = fib6_walk_continue(&iter->w); |
66f5d6ce | 2282 | spin_unlock_bh(&iter->tbl->tb6_lock); |
8d2ca1d7 HFS |
2283 | if (r > 0) { |
2284 | if (v) | |
2285 | ++*pos; | |
2286 | return iter->w.leaf; | |
2287 | } else if (r < 0) { | |
9a03cd8f | 2288 | fib6_walker_unlink(net, &iter->w); |
8d2ca1d7 HFS |
2289 | return NULL; |
2290 | } | |
9a03cd8f | 2291 | fib6_walker_unlink(net, &iter->w); |
8d2ca1d7 HFS |
2292 | |
2293 | iter->tbl = ipv6_route_seq_next_table(iter->tbl, net); | |
2294 | if (!iter->tbl) | |
2295 | return NULL; | |
2296 | ||
9a03cd8f | 2297 | ipv6_route_seq_setup_walk(iter, net); |
8d2ca1d7 HFS |
2298 | goto iter_table; |
2299 | } | |
2300 | ||
2301 | static void *ipv6_route_seq_start(struct seq_file *seq, loff_t *pos) | |
2302 | __acquires(RCU_BH) | |
2303 | { | |
2304 | struct net *net = seq_file_net(seq); | |
2305 | struct ipv6_route_iter *iter = seq->private; | |
2306 | ||
2307 | rcu_read_lock_bh(); | |
2308 | iter->tbl = ipv6_route_seq_next_table(NULL, net); | |
2309 | iter->skip = *pos; | |
2310 | ||
2311 | if (iter->tbl) { | |
9a03cd8f | 2312 | ipv6_route_seq_setup_walk(iter, net); |
8d2ca1d7 HFS |
2313 | return ipv6_route_seq_next(seq, NULL, pos); |
2314 | } else { | |
2315 | return NULL; | |
2316 | } | |
2317 | } | |
2318 | ||
2319 | static bool ipv6_route_iter_active(struct ipv6_route_iter *iter) | |
2320 | { | |
94b2cfe0 | 2321 | struct fib6_walker *w = &iter->w; |
8d2ca1d7 HFS |
2322 | return w->node && !(w->state == FWS_U && w->node == w->root); |
2323 | } | |
2324 | ||
2325 | static void ipv6_route_seq_stop(struct seq_file *seq, void *v) | |
2326 | __releases(RCU_BH) | |
2327 | { | |
9a03cd8f | 2328 | struct net *net = seq_file_net(seq); |
8d2ca1d7 HFS |
2329 | struct ipv6_route_iter *iter = seq->private; |
2330 | ||
2331 | if (ipv6_route_iter_active(iter)) | |
9a03cd8f | 2332 | fib6_walker_unlink(net, &iter->w); |
8d2ca1d7 HFS |
2333 | |
2334 | rcu_read_unlock_bh(); | |
2335 | } | |
2336 | ||
2337 | static const struct seq_operations ipv6_route_seq_ops = { | |
2338 | .start = ipv6_route_seq_start, | |
2339 | .next = ipv6_route_seq_next, | |
2340 | .stop = ipv6_route_seq_stop, | |
2341 | .show = ipv6_route_seq_show | |
2342 | }; | |
2343 | ||
2344 | int ipv6_route_open(struct inode *inode, struct file *file) | |
2345 | { | |
2346 | return seq_open_net(inode, file, &ipv6_route_seq_ops, | |
2347 | sizeof(struct ipv6_route_iter)); | |
2348 | } | |
2349 | ||
2350 | #endif /* CONFIG_PROC_FS */ |