2 * INETPEER - A storage for permanent information about peers
4 * This source is covered by the GNU GPL, the same as all kernel sources.
6 * Authors: Andrey V. Savochkin <saw@msu.ru>
9 #include <linux/module.h>
10 #include <linux/types.h>
11 #include <linux/slab.h>
12 #include <linux/interrupt.h>
13 #include <linux/spinlock.h>
14 #include <linux/random.h>
15 #include <linux/timer.h>
16 #include <linux/time.h>
17 #include <linux/kernel.h>
19 #include <linux/net.h>
20 #include <linux/workqueue.h>
22 #include <net/inetpeer.h>
23 #include <net/secure_seq.h>
26 * Theory of operations.
27 * We keep one entry for each peer IP address. The nodes contains long-living
28 * information about the peer which doesn't depend on routes.
30 * Nodes are removed only when reference counter goes to 0.
31 * When it's happened the node may be removed when a sufficient amount of
32 * time has been passed since its last use. The less-recently-used entry can
33 * also be removed if the pool is overloaded i.e. if the total amount of
34 * entries is greater-or-equal than the threshold.
36 * Node pool is organised as an AVL tree.
37 * Such an implementation has been chosen not just for fun. It's a way to
38 * prevent easy and efficient DoS attacks by creating hash collisions. A huge
39 * amount of long living nodes in a single hash slot would significantly delay
40 * lookups performed with disabled BHs.
42 * Serialisation issues.
43 * 1. Nodes may appear in the tree only with the pool lock held.
44 * 2. Nodes may disappear from the tree only with the pool lock held
45 * AND reference count being 0.
46 * 3. Global variable peer_total is modified under the pool lock.
47 * 4. struct inet_peer fields modification:
48 * avl_left, avl_right, avl_parent, avl_height: pool lock
49 * refcnt: atomically against modifications on other CPU;
50 * usually under some other lock to prevent node disappearing
54 static struct kmem_cache *peer_cachep __read_mostly;
56 static LIST_HEAD(gc_list);
57 static const int gc_delay = 60 * HZ;
58 static struct delayed_work gc_work;
59 static DEFINE_SPINLOCK(gc_lock);
61 #define node_height(x) x->avl_height
63 #define peer_avl_empty ((struct inet_peer *)&peer_fake_node)
64 #define peer_avl_empty_rcu ((struct inet_peer __rcu __force *)&peer_fake_node)
65 static const struct inet_peer peer_fake_node = {
66 .avl_left = peer_avl_empty_rcu,
67 .avl_right = peer_avl_empty_rcu,
71 void inet_peer_base_init(struct inet_peer_base *bp)
73 bp->root = peer_avl_empty_rcu;
74 seqlock_init(&bp->lock);
78 EXPORT_SYMBOL_GPL(inet_peer_base_init);
80 static atomic_t v4_seq = ATOMIC_INIT(0);
81 static atomic_t v6_seq = ATOMIC_INIT(0);
83 static atomic_t *inetpeer_seq_ptr(int family)
85 return (family == AF_INET ? &v4_seq : &v6_seq);
88 static inline void flush_check(struct inet_peer_base *base, int family)
90 atomic_t *fp = inetpeer_seq_ptr(family);
92 if (unlikely(base->flush_seq != atomic_read(fp))) {
93 inetpeer_invalidate_tree(base);
94 base->flush_seq = atomic_read(fp);
98 void inetpeer_invalidate_family(int family)
100 atomic_t *fp = inetpeer_seq_ptr(family);
105 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
107 /* Exported for sysctl_net_ipv4. */
108 int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more
109 * aggressively at this stage */
110 int inet_peer_minttl __read_mostly = 120 * HZ; /* TTL under high load: 120 sec */
111 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ; /* usual time to live: 10 min */
113 static void inetpeer_gc_worker(struct work_struct *work)
115 struct inet_peer *p, *n, *c;
118 spin_lock_bh(&gc_lock);
119 list_replace_init(&gc_list, &list);
120 spin_unlock_bh(&gc_lock);
122 if (list_empty(&list))
125 list_for_each_entry_safe(p, n, &list, gc_list) {
130 c = rcu_dereference_protected(p->avl_left, 1);
131 if (c != peer_avl_empty) {
132 list_add_tail(&c->gc_list, &list);
133 p->avl_left = peer_avl_empty_rcu;
136 c = rcu_dereference_protected(p->avl_right, 1);
137 if (c != peer_avl_empty) {
138 list_add_tail(&c->gc_list, &list);
139 p->avl_right = peer_avl_empty_rcu;
142 n = list_entry(p->gc_list.next, struct inet_peer, gc_list);
144 if (!atomic_read(&p->refcnt)) {
145 list_del(&p->gc_list);
146 kmem_cache_free(peer_cachep, p);
150 if (list_empty(&list))
153 spin_lock_bh(&gc_lock);
154 list_splice(&list, &gc_list);
155 spin_unlock_bh(&gc_lock);
157 schedule_delayed_work(&gc_work, gc_delay);
160 /* Called from ip_output.c:ip_init */
161 void __init inet_initpeers(void)
165 /* Use the straight interface to information about memory. */
167 /* The values below were suggested by Alexey Kuznetsov
168 * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values
171 if (si.totalram <= (32768*1024)/PAGE_SIZE)
172 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
173 if (si.totalram <= (16384*1024)/PAGE_SIZE)
174 inet_peer_threshold >>= 1; /* about 512KB */
175 if (si.totalram <= (8192*1024)/PAGE_SIZE)
176 inet_peer_threshold >>= 2; /* about 128KB */
178 peer_cachep = kmem_cache_create("inet_peer_cache",
179 sizeof(struct inet_peer),
180 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
183 INIT_DEFERRABLE_WORK(&gc_work, inetpeer_gc_worker);
186 static int addr_compare(const struct inetpeer_addr *a,
187 const struct inetpeer_addr *b)
189 int i, n = (a->family == AF_INET ? 1 : 4);
191 for (i = 0; i < n; i++) {
192 if (a->addr.a6[i] == b->addr.a6[i])
194 if ((__force u32)a->addr.a6[i] < (__force u32)b->addr.a6[i])
202 #define rcu_deref_locked(X, BASE) \
203 rcu_dereference_protected(X, lockdep_is_held(&(BASE)->lock.lock))
206 * Called with local BH disabled and the pool lock held.
208 #define lookup(_daddr, _stack, _base) \
210 struct inet_peer *u; \
211 struct inet_peer __rcu **v; \
214 *stackptr++ = &_base->root; \
215 for (u = rcu_deref_locked(_base->root, _base); \
216 u != peer_avl_empty; ) { \
217 int cmp = addr_compare(_daddr, &u->daddr); \
225 u = rcu_deref_locked(*v, _base); \
231 * Called with rcu_read_lock()
232 * Because we hold no lock against a writer, its quite possible we fall
233 * in an endless loop.
234 * But every pointer we follow is guaranteed to be valid thanks to RCU.
235 * We exit from this function if number of links exceeds PEER_MAXDEPTH
237 static struct inet_peer *lookup_rcu(const struct inetpeer_addr *daddr,
238 struct inet_peer_base *base)
240 struct inet_peer *u = rcu_dereference(base->root);
243 while (u != peer_avl_empty) {
244 int cmp = addr_compare(daddr, &u->daddr);
246 /* Before taking a reference, check if this entry was
247 * deleted (refcnt=-1)
249 if (!atomic_add_unless(&u->refcnt, 1, -1))
254 u = rcu_dereference(u->avl_left);
256 u = rcu_dereference(u->avl_right);
257 if (unlikely(++count == PEER_MAXDEPTH))
263 /* Called with local BH disabled and the pool lock held. */
264 #define lookup_rightempty(start, base) \
266 struct inet_peer *u; \
267 struct inet_peer __rcu **v; \
268 *stackptr++ = &start->avl_left; \
269 v = &start->avl_left; \
270 for (u = rcu_deref_locked(*v, base); \
271 u->avl_right != peer_avl_empty_rcu; ) { \
274 u = rcu_deref_locked(*v, base); \
279 /* Called with local BH disabled and the pool lock held.
280 * Variable names are the proof of operation correctness.
281 * Look into mm/map_avl.c for more detail description of the ideas.
283 static void peer_avl_rebalance(struct inet_peer __rcu **stack[],
284 struct inet_peer __rcu ***stackend,
285 struct inet_peer_base *base)
287 struct inet_peer __rcu **nodep;
288 struct inet_peer *node, *l, *r;
291 while (stackend > stack) {
293 node = rcu_deref_locked(*nodep, base);
294 l = rcu_deref_locked(node->avl_left, base);
295 r = rcu_deref_locked(node->avl_right, base);
298 if (lh > rh + 1) { /* l: RH+2 */
299 struct inet_peer *ll, *lr, *lrl, *lrr;
301 ll = rcu_deref_locked(l->avl_left, base);
302 lr = rcu_deref_locked(l->avl_right, base);
303 lrh = node_height(lr);
304 if (lrh <= node_height(ll)) { /* ll: RH+1 */
305 RCU_INIT_POINTER(node->avl_left, lr); /* lr: RH or RH+1 */
306 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */
307 node->avl_height = lrh + 1; /* RH+1 or RH+2 */
308 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH+1 */
309 RCU_INIT_POINTER(l->avl_right, node); /* node: RH+1 or RH+2 */
310 l->avl_height = node->avl_height + 1;
311 RCU_INIT_POINTER(*nodep, l);
312 } else { /* ll: RH, lr: RH+1 */
313 lrl = rcu_deref_locked(lr->avl_left, base);/* lrl: RH or RH-1 */
314 lrr = rcu_deref_locked(lr->avl_right, base);/* lrr: RH or RH-1 */
315 RCU_INIT_POINTER(node->avl_left, lrr); /* lrr: RH or RH-1 */
316 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */
317 node->avl_height = rh + 1; /* node: RH+1 */
318 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH */
319 RCU_INIT_POINTER(l->avl_right, lrl); /* lrl: RH or RH-1 */
320 l->avl_height = rh + 1; /* l: RH+1 */
321 RCU_INIT_POINTER(lr->avl_left, l); /* l: RH+1 */
322 RCU_INIT_POINTER(lr->avl_right, node); /* node: RH+1 */
323 lr->avl_height = rh + 2;
324 RCU_INIT_POINTER(*nodep, lr);
326 } else if (rh > lh + 1) { /* r: LH+2 */
327 struct inet_peer *rr, *rl, *rlr, *rll;
329 rr = rcu_deref_locked(r->avl_right, base);
330 rl = rcu_deref_locked(r->avl_left, base);
331 rlh = node_height(rl);
332 if (rlh <= node_height(rr)) { /* rr: LH+1 */
333 RCU_INIT_POINTER(node->avl_right, rl); /* rl: LH or LH+1 */
334 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */
335 node->avl_height = rlh + 1; /* LH+1 or LH+2 */
336 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH+1 */
337 RCU_INIT_POINTER(r->avl_left, node); /* node: LH+1 or LH+2 */
338 r->avl_height = node->avl_height + 1;
339 RCU_INIT_POINTER(*nodep, r);
340 } else { /* rr: RH, rl: RH+1 */
341 rlr = rcu_deref_locked(rl->avl_right, base);/* rlr: LH or LH-1 */
342 rll = rcu_deref_locked(rl->avl_left, base);/* rll: LH or LH-1 */
343 RCU_INIT_POINTER(node->avl_right, rll); /* rll: LH or LH-1 */
344 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */
345 node->avl_height = lh + 1; /* node: LH+1 */
346 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH */
347 RCU_INIT_POINTER(r->avl_left, rlr); /* rlr: LH or LH-1 */
348 r->avl_height = lh + 1; /* r: LH+1 */
349 RCU_INIT_POINTER(rl->avl_right, r); /* r: LH+1 */
350 RCU_INIT_POINTER(rl->avl_left, node); /* node: LH+1 */
351 rl->avl_height = lh + 2;
352 RCU_INIT_POINTER(*nodep, rl);
355 node->avl_height = (lh > rh ? lh : rh) + 1;
360 /* Called with local BH disabled and the pool lock held. */
361 #define link_to_pool(n, base) \
364 n->avl_left = peer_avl_empty_rcu; \
365 n->avl_right = peer_avl_empty_rcu; \
366 /* lockless readers can catch us now */ \
367 rcu_assign_pointer(**--stackptr, n); \
368 peer_avl_rebalance(stack, stackptr, base); \
371 static void inetpeer_free_rcu(struct rcu_head *head)
373 kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
376 static void unlink_from_pool(struct inet_peer *p, struct inet_peer_base *base,
377 struct inet_peer __rcu **stack[PEER_MAXDEPTH])
379 struct inet_peer __rcu ***stackptr, ***delp;
381 if (lookup(&p->daddr, stack, base) != p)
383 delp = stackptr - 1; /* *delp[0] == p */
384 if (p->avl_left == peer_avl_empty_rcu) {
385 *delp[0] = p->avl_right;
388 /* look for a node to insert instead of p */
390 t = lookup_rightempty(p, base);
391 BUG_ON(rcu_deref_locked(*stackptr[-1], base) != t);
392 **--stackptr = t->avl_left;
393 /* t is removed, t->daddr > x->daddr for any
394 * x in p->avl_left subtree.
395 * Put t in the old place of p. */
396 RCU_INIT_POINTER(*delp[0], t);
397 t->avl_left = p->avl_left;
398 t->avl_right = p->avl_right;
399 t->avl_height = p->avl_height;
400 BUG_ON(delp[1] != &p->avl_left);
401 delp[1] = &t->avl_left; /* was &p->avl_left */
403 peer_avl_rebalance(stack, stackptr, base);
405 call_rcu(&p->rcu, inetpeer_free_rcu);
408 /* perform garbage collect on all items stacked during a lookup */
409 static int inet_peer_gc(struct inet_peer_base *base,
410 struct inet_peer __rcu **stack[PEER_MAXDEPTH],
411 struct inet_peer __rcu ***stackptr)
413 struct inet_peer *p, *gchead = NULL;
417 if (base->total >= inet_peer_threshold)
418 ttl = 0; /* be aggressive */
420 ttl = inet_peer_maxttl
421 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
422 base->total / inet_peer_threshold * HZ;
423 stackptr--; /* last stack slot is peer_avl_empty */
424 while (stackptr > stack) {
426 p = rcu_deref_locked(**stackptr, base);
427 if (atomic_read(&p->refcnt) == 0) {
429 delta = (__u32)jiffies - p->dtime;
431 atomic_cmpxchg(&p->refcnt, 0, -1) == 0) {
437 while ((p = gchead) != NULL) {
440 unlink_from_pool(p, base, stack);
445 struct inet_peer *inet_getpeer(struct inet_peer_base *base,
446 const struct inetpeer_addr *daddr,
449 struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr;
451 unsigned int sequence;
452 int invalidated, gccnt = 0;
454 flush_check(base, daddr->family);
456 /* Attempt a lockless lookup first.
457 * Because of a concurrent writer, we might not find an existing entry.
460 sequence = read_seqbegin(&base->lock);
461 p = lookup_rcu(daddr, base);
462 invalidated = read_seqretry(&base->lock, sequence);
468 /* If no writer did a change during our lookup, we can return early. */
469 if (!create && !invalidated)
472 /* retry an exact lookup, taking the lock before.
473 * At least, nodes should be hot in our cache.
475 write_seqlock_bh(&base->lock);
477 p = lookup(daddr, stack, base);
478 if (p != peer_avl_empty) {
479 atomic_inc(&p->refcnt);
480 write_sequnlock_bh(&base->lock);
484 gccnt = inet_peer_gc(base, stack, stackptr);
488 p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
491 atomic_set(&p->refcnt, 1);
492 atomic_set(&p->rid, 0);
493 p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
495 /* 60*HZ is arbitrary, but chosen enough high so that the first
496 * calculation of tokens is at its maximum.
498 p->rate_last = jiffies - 60*HZ;
499 INIT_LIST_HEAD(&p->gc_list);
502 link_to_pool(p, base);
505 write_sequnlock_bh(&base->lock);
509 EXPORT_SYMBOL_GPL(inet_getpeer);
511 void inet_putpeer(struct inet_peer *p)
513 p->dtime = (__u32)jiffies;
514 smp_mb__before_atomic_dec();
515 atomic_dec(&p->refcnt);
517 EXPORT_SYMBOL_GPL(inet_putpeer);
520 * Check transmit rate limitation for given message.
521 * The rate information is held in the inet_peer entries now.
522 * This function is generic and could be used for other purposes
523 * too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
525 * Note that the same inet_peer fields are modified by functions in
526 * route.c too, but these work for packet destinations while xrlim_allow
527 * works for icmp destinations. This means the rate limiting information
528 * for one "ip object" is shared - and these ICMPs are twice limited:
529 * by source and by destination.
531 * RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
532 * SHOULD allow setting of rate limits
534 * Shared between ICMPv4 and ICMPv6.
536 #define XRLIM_BURST_FACTOR 6
537 bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
539 unsigned long now, token;
545 token = peer->rate_tokens;
547 token += now - peer->rate_last;
548 peer->rate_last = now;
549 if (token > XRLIM_BURST_FACTOR * timeout)
550 token = XRLIM_BURST_FACTOR * timeout;
551 if (token >= timeout) {
555 peer->rate_tokens = token;
558 EXPORT_SYMBOL(inet_peer_xrlim_allow);
560 static void inetpeer_inval_rcu(struct rcu_head *head)
562 struct inet_peer *p = container_of(head, struct inet_peer, gc_rcu);
564 spin_lock_bh(&gc_lock);
565 list_add_tail(&p->gc_list, &gc_list);
566 spin_unlock_bh(&gc_lock);
568 schedule_delayed_work(&gc_work, gc_delay);
571 void inetpeer_invalidate_tree(struct inet_peer_base *base)
573 struct inet_peer *root;
575 write_seqlock_bh(&base->lock);
577 root = rcu_deref_locked(base->root, base);
578 if (root != peer_avl_empty) {
579 base->root = peer_avl_empty_rcu;
581 call_rcu(&root->gc_rcu, inetpeer_inval_rcu);
584 write_sequnlock_bh(&base->lock);
586 EXPORT_SYMBOL(inetpeer_invalidate_tree);