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.
29 * At this moment this information consists only of ID field for the next
30 * outgoing IP packet. This field is incremented with each packet as encoded
31 * in inet_getid() function (include/net/inetpeer.h).
32 * At the moment of writing this notes identifier of IP packets is generated
33 * to be unpredictable using this code only for packets subjected
34 * (actually or potentially) to defragmentation. I.e. DF packets less than
35 * PMTU in size uses a constant ID and do not use this code (see
36 * ip_select_ident() in include/net/ip.h).
38 * Route cache entries hold references to our nodes.
39 * New cache entries get references via lookup by destination IP address in
40 * the avl tree. The reference is grabbed only when it's needed i.e. only
41 * when we try to output IP packet which needs an unpredictable ID (see
42 * __ip_select_ident() in net/ipv4/route.c).
43 * Nodes are removed only when reference counter goes to 0.
44 * When it's happened the node may be removed when a sufficient amount of
45 * time has been passed since its last use. The less-recently-used entry can
46 * also be removed if the pool is overloaded i.e. if the total amount of
47 * entries is greater-or-equal than the threshold.
49 * Node pool is organised as an AVL tree.
50 * Such an implementation has been chosen not just for fun. It's a way to
51 * prevent easy and efficient DoS attacks by creating hash collisions. A huge
52 * amount of long living nodes in a single hash slot would significantly delay
53 * lookups performed with disabled BHs.
55 * Serialisation issues.
56 * 1. Nodes may appear in the tree only with the pool lock held.
57 * 2. Nodes may disappear from the tree only with the pool lock held
58 * AND reference count being 0.
59 * 3. Global variable peer_total is modified under the pool lock.
60 * 4. struct inet_peer fields modification:
61 * avl_left, avl_right, avl_parent, avl_height: pool lock
62 * refcnt: atomically against modifications on other CPU;
63 * usually under some other lock to prevent node disappearing
65 * ip_id_count: atomic value (no lock needed)
68 static struct kmem_cache *peer_cachep __read_mostly;
70 static LIST_HEAD(gc_list);
71 static const int gc_delay = 60 * HZ;
72 static struct delayed_work gc_work;
73 static DEFINE_SPINLOCK(gc_lock);
75 #define node_height(x) x->avl_height
77 #define peer_avl_empty ((struct inet_peer *)&peer_fake_node)
78 #define peer_avl_empty_rcu ((struct inet_peer __rcu __force *)&peer_fake_node)
79 static const struct inet_peer peer_fake_node = {
80 .avl_left = peer_avl_empty_rcu,
81 .avl_right = peer_avl_empty_rcu,
85 void inet_peer_base_init(struct inet_peer_base *bp)
87 bp->root = peer_avl_empty_rcu;
88 seqlock_init(&bp->lock);
92 EXPORT_SYMBOL_GPL(inet_peer_base_init);
94 static atomic_t v4_seq = ATOMIC_INIT(0);
95 static atomic_t v6_seq = ATOMIC_INIT(0);
97 static atomic_t *inetpeer_seq_ptr(int family)
99 return (family == AF_INET ? &v4_seq : &v6_seq);
102 static inline void flush_check(struct inet_peer_base *base, int family)
104 atomic_t *fp = inetpeer_seq_ptr(family);
106 if (unlikely(base->flush_seq != atomic_read(fp))) {
107 inetpeer_invalidate_tree(base);
108 base->flush_seq = atomic_read(fp);
112 void inetpeer_invalidate_family(int family)
114 atomic_t *fp = inetpeer_seq_ptr(family);
119 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
121 /* Exported for sysctl_net_ipv4. */
122 int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more
123 * aggressively at this stage */
124 int inet_peer_minttl __read_mostly = 120 * HZ; /* TTL under high load: 120 sec */
125 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ; /* usual time to live: 10 min */
127 static void inetpeer_gc_worker(struct work_struct *work)
129 struct inet_peer *p, *n, *c;
132 spin_lock_bh(&gc_lock);
133 list_replace_init(&gc_list, &list);
134 spin_unlock_bh(&gc_lock);
136 if (list_empty(&list))
139 list_for_each_entry_safe(p, n, &list, gc_list) {
144 c = rcu_dereference_protected(p->avl_left, 1);
145 if (c != peer_avl_empty) {
146 list_add_tail(&c->gc_list, &list);
147 p->avl_left = peer_avl_empty_rcu;
150 c = rcu_dereference_protected(p->avl_right, 1);
151 if (c != peer_avl_empty) {
152 list_add_tail(&c->gc_list, &list);
153 p->avl_right = peer_avl_empty_rcu;
156 n = list_entry(p->gc_list.next, struct inet_peer, gc_list);
158 if (!atomic_read(&p->refcnt)) {
159 list_del(&p->gc_list);
160 kmem_cache_free(peer_cachep, p);
164 if (list_empty(&list))
167 spin_lock_bh(&gc_lock);
168 list_splice(&list, &gc_list);
169 spin_unlock_bh(&gc_lock);
171 schedule_delayed_work(&gc_work, gc_delay);
174 /* Called from ip_output.c:ip_init */
175 void __init inet_initpeers(void)
179 /* Use the straight interface to information about memory. */
181 /* The values below were suggested by Alexey Kuznetsov
182 * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values
185 if (si.totalram <= (32768*1024)/PAGE_SIZE)
186 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
187 if (si.totalram <= (16384*1024)/PAGE_SIZE)
188 inet_peer_threshold >>= 1; /* about 512KB */
189 if (si.totalram <= (8192*1024)/PAGE_SIZE)
190 inet_peer_threshold >>= 2; /* about 128KB */
192 peer_cachep = kmem_cache_create("inet_peer_cache",
193 sizeof(struct inet_peer),
194 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
197 INIT_DEFERRABLE_WORK(&gc_work, inetpeer_gc_worker);
200 static int addr_compare(const struct inetpeer_addr *a,
201 const struct inetpeer_addr *b)
203 int i, n = (a->family == AF_INET ? 1 : 4);
205 for (i = 0; i < n; i++) {
206 if (a->addr.a6[i] == b->addr.a6[i])
208 if ((__force u32)a->addr.a6[i] < (__force u32)b->addr.a6[i])
216 #define rcu_deref_locked(X, BASE) \
217 rcu_dereference_protected(X, lockdep_is_held(&(BASE)->lock.lock))
220 * Called with local BH disabled and the pool lock held.
222 #define lookup(_daddr, _stack, _base) \
224 struct inet_peer *u; \
225 struct inet_peer __rcu **v; \
228 *stackptr++ = &_base->root; \
229 for (u = rcu_deref_locked(_base->root, _base); \
230 u != peer_avl_empty; ) { \
231 int cmp = addr_compare(_daddr, &u->daddr); \
239 u = rcu_deref_locked(*v, _base); \
245 * Called with rcu_read_lock()
246 * Because we hold no lock against a writer, its quite possible we fall
247 * in an endless loop.
248 * But every pointer we follow is guaranteed to be valid thanks to RCU.
249 * We exit from this function if number of links exceeds PEER_MAXDEPTH
251 static struct inet_peer *lookup_rcu(const struct inetpeer_addr *daddr,
252 struct inet_peer_base *base)
254 struct inet_peer *u = rcu_dereference(base->root);
257 while (u != peer_avl_empty) {
258 int cmp = addr_compare(daddr, &u->daddr);
260 /* Before taking a reference, check if this entry was
261 * deleted (refcnt=-1)
263 if (!atomic_add_unless(&u->refcnt, 1, -1))
268 u = rcu_dereference(u->avl_left);
270 u = rcu_dereference(u->avl_right);
271 if (unlikely(++count == PEER_MAXDEPTH))
277 /* Called with local BH disabled and the pool lock held. */
278 #define lookup_rightempty(start, base) \
280 struct inet_peer *u; \
281 struct inet_peer __rcu **v; \
282 *stackptr++ = &start->avl_left; \
283 v = &start->avl_left; \
284 for (u = rcu_deref_locked(*v, base); \
285 u->avl_right != peer_avl_empty_rcu; ) { \
288 u = rcu_deref_locked(*v, base); \
293 /* Called with local BH disabled and the pool lock held.
294 * Variable names are the proof of operation correctness.
295 * Look into mm/map_avl.c for more detail description of the ideas.
297 static void peer_avl_rebalance(struct inet_peer __rcu **stack[],
298 struct inet_peer __rcu ***stackend,
299 struct inet_peer_base *base)
301 struct inet_peer __rcu **nodep;
302 struct inet_peer *node, *l, *r;
305 while (stackend > stack) {
307 node = rcu_deref_locked(*nodep, base);
308 l = rcu_deref_locked(node->avl_left, base);
309 r = rcu_deref_locked(node->avl_right, base);
312 if (lh > rh + 1) { /* l: RH+2 */
313 struct inet_peer *ll, *lr, *lrl, *lrr;
315 ll = rcu_deref_locked(l->avl_left, base);
316 lr = rcu_deref_locked(l->avl_right, base);
317 lrh = node_height(lr);
318 if (lrh <= node_height(ll)) { /* ll: RH+1 */
319 RCU_INIT_POINTER(node->avl_left, lr); /* lr: RH or RH+1 */
320 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */
321 node->avl_height = lrh + 1; /* RH+1 or RH+2 */
322 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH+1 */
323 RCU_INIT_POINTER(l->avl_right, node); /* node: RH+1 or RH+2 */
324 l->avl_height = node->avl_height + 1;
325 RCU_INIT_POINTER(*nodep, l);
326 } else { /* ll: RH, lr: RH+1 */
327 lrl = rcu_deref_locked(lr->avl_left, base);/* lrl: RH or RH-1 */
328 lrr = rcu_deref_locked(lr->avl_right, base);/* lrr: RH or RH-1 */
329 RCU_INIT_POINTER(node->avl_left, lrr); /* lrr: RH or RH-1 */
330 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */
331 node->avl_height = rh + 1; /* node: RH+1 */
332 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH */
333 RCU_INIT_POINTER(l->avl_right, lrl); /* lrl: RH or RH-1 */
334 l->avl_height = rh + 1; /* l: RH+1 */
335 RCU_INIT_POINTER(lr->avl_left, l); /* l: RH+1 */
336 RCU_INIT_POINTER(lr->avl_right, node); /* node: RH+1 */
337 lr->avl_height = rh + 2;
338 RCU_INIT_POINTER(*nodep, lr);
340 } else if (rh > lh + 1) { /* r: LH+2 */
341 struct inet_peer *rr, *rl, *rlr, *rll;
343 rr = rcu_deref_locked(r->avl_right, base);
344 rl = rcu_deref_locked(r->avl_left, base);
345 rlh = node_height(rl);
346 if (rlh <= node_height(rr)) { /* rr: LH+1 */
347 RCU_INIT_POINTER(node->avl_right, rl); /* rl: LH or LH+1 */
348 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */
349 node->avl_height = rlh + 1; /* LH+1 or LH+2 */
350 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH+1 */
351 RCU_INIT_POINTER(r->avl_left, node); /* node: LH+1 or LH+2 */
352 r->avl_height = node->avl_height + 1;
353 RCU_INIT_POINTER(*nodep, r);
354 } else { /* rr: RH, rl: RH+1 */
355 rlr = rcu_deref_locked(rl->avl_right, base);/* rlr: LH or LH-1 */
356 rll = rcu_deref_locked(rl->avl_left, base);/* rll: LH or LH-1 */
357 RCU_INIT_POINTER(node->avl_right, rll); /* rll: LH or LH-1 */
358 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */
359 node->avl_height = lh + 1; /* node: LH+1 */
360 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH */
361 RCU_INIT_POINTER(r->avl_left, rlr); /* rlr: LH or LH-1 */
362 r->avl_height = lh + 1; /* r: LH+1 */
363 RCU_INIT_POINTER(rl->avl_right, r); /* r: LH+1 */
364 RCU_INIT_POINTER(rl->avl_left, node); /* node: LH+1 */
365 rl->avl_height = lh + 2;
366 RCU_INIT_POINTER(*nodep, rl);
369 node->avl_height = (lh > rh ? lh : rh) + 1;
374 /* Called with local BH disabled and the pool lock held. */
375 #define link_to_pool(n, base) \
378 n->avl_left = peer_avl_empty_rcu; \
379 n->avl_right = peer_avl_empty_rcu; \
380 /* lockless readers can catch us now */ \
381 rcu_assign_pointer(**--stackptr, n); \
382 peer_avl_rebalance(stack, stackptr, base); \
385 static void inetpeer_free_rcu(struct rcu_head *head)
387 kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
390 static void unlink_from_pool(struct inet_peer *p, struct inet_peer_base *base,
391 struct inet_peer __rcu **stack[PEER_MAXDEPTH])
393 struct inet_peer __rcu ***stackptr, ***delp;
395 if (lookup(&p->daddr, stack, base) != p)
397 delp = stackptr - 1; /* *delp[0] == p */
398 if (p->avl_left == peer_avl_empty_rcu) {
399 *delp[0] = p->avl_right;
402 /* look for a node to insert instead of p */
404 t = lookup_rightempty(p, base);
405 BUG_ON(rcu_deref_locked(*stackptr[-1], base) != t);
406 **--stackptr = t->avl_left;
407 /* t is removed, t->daddr > x->daddr for any
408 * x in p->avl_left subtree.
409 * Put t in the old place of p. */
410 RCU_INIT_POINTER(*delp[0], t);
411 t->avl_left = p->avl_left;
412 t->avl_right = p->avl_right;
413 t->avl_height = p->avl_height;
414 BUG_ON(delp[1] != &p->avl_left);
415 delp[1] = &t->avl_left; /* was &p->avl_left */
417 peer_avl_rebalance(stack, stackptr, base);
419 call_rcu(&p->rcu, inetpeer_free_rcu);
422 /* perform garbage collect on all items stacked during a lookup */
423 static int inet_peer_gc(struct inet_peer_base *base,
424 struct inet_peer __rcu **stack[PEER_MAXDEPTH],
425 struct inet_peer __rcu ***stackptr)
427 struct inet_peer *p, *gchead = NULL;
431 if (base->total >= inet_peer_threshold)
432 ttl = 0; /* be aggressive */
434 ttl = inet_peer_maxttl
435 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
436 base->total / inet_peer_threshold * HZ;
437 stackptr--; /* last stack slot is peer_avl_empty */
438 while (stackptr > stack) {
440 p = rcu_deref_locked(**stackptr, base);
441 if (atomic_read(&p->refcnt) == 0) {
443 delta = (__u32)jiffies - p->dtime;
445 atomic_cmpxchg(&p->refcnt, 0, -1) == 0) {
451 while ((p = gchead) != NULL) {
454 unlink_from_pool(p, base, stack);
459 struct inet_peer *inet_getpeer(struct inet_peer_base *base,
460 const struct inetpeer_addr *daddr,
463 struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr;
465 unsigned int sequence;
466 int invalidated, gccnt = 0;
468 flush_check(base, daddr->family);
470 /* Attempt a lockless lookup first.
471 * Because of a concurrent writer, we might not find an existing entry.
474 sequence = read_seqbegin(&base->lock);
475 p = lookup_rcu(daddr, base);
476 invalidated = read_seqretry(&base->lock, sequence);
482 /* If no writer did a change during our lookup, we can return early. */
483 if (!create && !invalidated)
486 /* retry an exact lookup, taking the lock before.
487 * At least, nodes should be hot in our cache.
489 write_seqlock_bh(&base->lock);
491 p = lookup(daddr, stack, base);
492 if (p != peer_avl_empty) {
493 atomic_inc(&p->refcnt);
494 write_sequnlock_bh(&base->lock);
498 gccnt = inet_peer_gc(base, stack, stackptr);
502 p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
505 atomic_set(&p->refcnt, 1);
506 atomic_set(&p->rid, 0);
507 atomic_set(&p->ip_id_count,
508 (daddr->family == AF_INET) ?
509 secure_ip_id(daddr->addr.a4) :
510 secure_ipv6_id(daddr->addr.a6));
511 p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
513 /* 60*HZ is arbitrary, but chosen enough high so that the first
514 * calculation of tokens is at its maximum.
516 p->rate_last = jiffies - 60*HZ;
517 INIT_LIST_HEAD(&p->gc_list);
520 link_to_pool(p, base);
523 write_sequnlock_bh(&base->lock);
527 EXPORT_SYMBOL_GPL(inet_getpeer);
529 void inet_putpeer(struct inet_peer *p)
531 p->dtime = (__u32)jiffies;
532 smp_mb__before_atomic_dec();
533 atomic_dec(&p->refcnt);
535 EXPORT_SYMBOL_GPL(inet_putpeer);
538 * Check transmit rate limitation for given message.
539 * The rate information is held in the inet_peer entries now.
540 * This function is generic and could be used for other purposes
541 * too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
543 * Note that the same inet_peer fields are modified by functions in
544 * route.c too, but these work for packet destinations while xrlim_allow
545 * works for icmp destinations. This means the rate limiting information
546 * for one "ip object" is shared - and these ICMPs are twice limited:
547 * by source and by destination.
549 * RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
550 * SHOULD allow setting of rate limits
552 * Shared between ICMPv4 and ICMPv6.
554 #define XRLIM_BURST_FACTOR 6
555 bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
557 unsigned long now, token;
563 token = peer->rate_tokens;
565 token += now - peer->rate_last;
566 peer->rate_last = now;
567 if (token > XRLIM_BURST_FACTOR * timeout)
568 token = XRLIM_BURST_FACTOR * timeout;
569 if (token >= timeout) {
573 peer->rate_tokens = token;
576 EXPORT_SYMBOL(inet_peer_xrlim_allow);
578 static void inetpeer_inval_rcu(struct rcu_head *head)
580 struct inet_peer *p = container_of(head, struct inet_peer, gc_rcu);
582 spin_lock_bh(&gc_lock);
583 list_add_tail(&p->gc_list, &gc_list);
584 spin_unlock_bh(&gc_lock);
586 schedule_delayed_work(&gc_work, gc_delay);
589 void inetpeer_invalidate_tree(struct inet_peer_base *base)
591 struct inet_peer *root;
593 write_seqlock_bh(&base->lock);
595 root = rcu_deref_locked(base->root, base);
596 if (root != peer_avl_empty) {
597 base->root = peer_avl_empty_rcu;
599 call_rcu(&root->gc_rcu, inetpeer_inval_rcu);
602 write_sequnlock_bh(&base->lock);
604 EXPORT_SYMBOL(inetpeer_invalidate_tree);