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>
21 #include <net/inetpeer.h>
24 * Theory of operations.
25 * We keep one entry for each peer IP address. The nodes contains long-living
26 * information about the peer which doesn't depend on routes.
27 * At this moment this information consists only of ID field for the next
28 * outgoing IP packet. This field is incremented with each packet as encoded
29 * in inet_getid() function (include/net/inetpeer.h).
30 * At the moment of writing this notes identifier of IP packets is generated
31 * to be unpredictable using this code only for packets subjected
32 * (actually or potentially) to defragmentation. I.e. DF packets less than
33 * PMTU in size uses a constant ID and do not use this code (see
34 * ip_select_ident() in include/net/ip.h).
36 * Route cache entries hold references to our nodes.
37 * New cache entries get references via lookup by destination IP address in
38 * the avl tree. The reference is grabbed only when it's needed i.e. only
39 * when we try to output IP packet which needs an unpredictable ID (see
40 * __ip_select_ident() in net/ipv4/route.c).
41 * Nodes are removed only when reference counter goes to 0.
42 * When it's happened the node may be removed when a sufficient amount of
43 * time has been passed since its last use. The less-recently-used entry can
44 * also be removed if the pool is overloaded i.e. if the total amount of
45 * entries is greater-or-equal than the threshold.
47 * Node pool is organised as an AVL tree.
48 * Such an implementation has been chosen not just for fun. It's a way to
49 * prevent easy and efficient DoS attacks by creating hash collisions. A huge
50 * amount of long living nodes in a single hash slot would significantly delay
51 * lookups performed with disabled BHs.
53 * Serialisation issues.
54 * 1. Nodes may appear in the tree only with the pool lock held.
55 * 2. Nodes may disappear from the tree only with the pool lock held
56 * AND reference count being 0.
57 * 3. Nodes appears and disappears from unused node list only under
58 * "inet_peer_unused_lock".
59 * 4. Global variable peer_total is modified under the pool lock.
60 * 5. struct inet_peer fields modification:
61 * avl_left, avl_right, avl_parent, avl_height: pool lock
62 * unused: unused node list lock
63 * refcnt: atomically against modifications on other CPU;
64 * usually under some other lock to prevent node disappearing
65 * dtime: unused node list lock
67 * ip_id_count: atomic value (no lock needed)
70 static struct kmem_cache *peer_cachep __read_mostly;
72 #define node_height(x) x->avl_height
74 #define peer_avl_empty ((struct inet_peer *)&peer_fake_node)
75 #define peer_avl_empty_rcu ((struct inet_peer __rcu __force *)&peer_fake_node)
76 static const struct inet_peer peer_fake_node = {
77 .avl_left = peer_avl_empty_rcu,
78 .avl_right = peer_avl_empty_rcu,
82 struct inet_peer_base {
83 struct inet_peer __rcu *root;
88 static struct inet_peer_base v4_peers = {
89 .root = peer_avl_empty_rcu,
90 .lock = __SEQLOCK_UNLOCKED(v4_peers.lock),
94 static struct inet_peer_base v6_peers = {
95 .root = peer_avl_empty_rcu,
96 .lock = __SEQLOCK_UNLOCKED(v6_peers.lock),
100 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
102 /* Exported for sysctl_net_ipv4. */
103 int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more
104 * aggressively at this stage */
105 int inet_peer_minttl __read_mostly = 120 * HZ; /* TTL under high load: 120 sec */
106 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ; /* usual time to live: 10 min */
107 int inet_peer_gc_mintime __read_mostly = 10 * HZ;
108 int inet_peer_gc_maxtime __read_mostly = 120 * HZ;
111 struct list_head list;
114 .list = LIST_HEAD_INIT(unused_peers.list),
115 .lock = __SPIN_LOCK_UNLOCKED(unused_peers.lock),
118 static void peer_check_expire(unsigned long dummy);
119 static DEFINE_TIMER(peer_periodic_timer, peer_check_expire, 0, 0);
122 /* Called from ip_output.c:ip_init */
123 void __init inet_initpeers(void)
127 /* Use the straight interface to information about memory. */
129 /* The values below were suggested by Alexey Kuznetsov
130 * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values
133 if (si.totalram <= (32768*1024)/PAGE_SIZE)
134 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
135 if (si.totalram <= (16384*1024)/PAGE_SIZE)
136 inet_peer_threshold >>= 1; /* about 512KB */
137 if (si.totalram <= (8192*1024)/PAGE_SIZE)
138 inet_peer_threshold >>= 2; /* about 128KB */
140 peer_cachep = kmem_cache_create("inet_peer_cache",
141 sizeof(struct inet_peer),
142 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
145 /* All the timers, started at system startup tend
146 to synchronize. Perturb it a bit.
148 peer_periodic_timer.expires = jiffies
149 + net_random() % inet_peer_gc_maxtime
150 + inet_peer_gc_maxtime;
151 add_timer(&peer_periodic_timer);
154 /* Called with or without local BH being disabled. */
155 static void unlink_from_unused(struct inet_peer *p)
157 spin_lock_bh(&unused_peers.lock);
158 list_del_init(&p->unused);
159 spin_unlock_bh(&unused_peers.lock);
162 static int addr_compare(const struct inetpeer_addr *a,
163 const struct inetpeer_addr *b)
165 int i, n = (a->family == AF_INET ? 1 : 4);
167 for (i = 0; i < n; i++) {
168 if (a->addr.a6[i] == b->addr.a6[i])
170 if (a->addr.a6[i] < b->addr.a6[i])
178 #define rcu_deref_locked(X, BASE) \
179 rcu_dereference_protected(X, lockdep_is_held(&(BASE)->lock.lock))
182 * Called with local BH disabled and the pool lock held.
184 #define lookup(_daddr, _stack, _base) \
186 struct inet_peer *u; \
187 struct inet_peer __rcu **v; \
190 *stackptr++ = &_base->root; \
191 for (u = rcu_deref_locked(_base->root, _base); \
192 u != peer_avl_empty; ) { \
193 int cmp = addr_compare(_daddr, &u->daddr); \
201 u = rcu_deref_locked(*v, _base); \
206 static bool atomic_add_unless_return(atomic_t *ptr, int a, int u, int *newv)
208 int cur, old = atomic_read(ptr);
212 cur = atomic_cmpxchg(ptr, old, *newv);
221 * Called with rcu_read_lock()
222 * Because we hold no lock against a writer, its quite possible we fall
223 * in an endless loop.
224 * But every pointer we follow is guaranteed to be valid thanks to RCU.
225 * We exit from this function if number of links exceeds PEER_MAXDEPTH
227 static struct inet_peer *lookup_rcu(const struct inetpeer_addr *daddr,
228 struct inet_peer_base *base,
231 struct inet_peer *u = rcu_dereference(base->root);
234 while (u != peer_avl_empty) {
235 int cmp = addr_compare(daddr, &u->daddr);
237 /* Before taking a reference, check if this entry was
238 * deleted, unlink_from_pool() sets refcnt=-1 to make
239 * distinction between an unused entry (refcnt=0) and
242 if (!atomic_add_unless_return(&u->refcnt, 1, -1, newrefcnt))
247 u = rcu_dereference(u->avl_left);
249 u = rcu_dereference(u->avl_right);
250 if (unlikely(++count == PEER_MAXDEPTH))
256 /* Called with local BH disabled and the pool lock held. */
257 #define lookup_rightempty(start, base) \
259 struct inet_peer *u; \
260 struct inet_peer __rcu **v; \
261 *stackptr++ = &start->avl_left; \
262 v = &start->avl_left; \
263 for (u = rcu_deref_locked(*v, base); \
264 u->avl_right != peer_avl_empty_rcu; ) { \
267 u = rcu_deref_locked(*v, base); \
272 /* Called with local BH disabled and the pool lock held.
273 * Variable names are the proof of operation correctness.
274 * Look into mm/map_avl.c for more detail description of the ideas.
276 static void peer_avl_rebalance(struct inet_peer __rcu **stack[],
277 struct inet_peer __rcu ***stackend,
278 struct inet_peer_base *base)
280 struct inet_peer __rcu **nodep;
281 struct inet_peer *node, *l, *r;
284 while (stackend > stack) {
286 node = rcu_deref_locked(*nodep, base);
287 l = rcu_deref_locked(node->avl_left, base);
288 r = rcu_deref_locked(node->avl_right, base);
291 if (lh > rh + 1) { /* l: RH+2 */
292 struct inet_peer *ll, *lr, *lrl, *lrr;
294 ll = rcu_deref_locked(l->avl_left, base);
295 lr = rcu_deref_locked(l->avl_right, base);
296 lrh = node_height(lr);
297 if (lrh <= node_height(ll)) { /* ll: RH+1 */
298 RCU_INIT_POINTER(node->avl_left, lr); /* lr: RH or RH+1 */
299 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */
300 node->avl_height = lrh + 1; /* RH+1 or RH+2 */
301 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH+1 */
302 RCU_INIT_POINTER(l->avl_right, node); /* node: RH+1 or RH+2 */
303 l->avl_height = node->avl_height + 1;
304 RCU_INIT_POINTER(*nodep, l);
305 } else { /* ll: RH, lr: RH+1 */
306 lrl = rcu_deref_locked(lr->avl_left, base);/* lrl: RH or RH-1 */
307 lrr = rcu_deref_locked(lr->avl_right, base);/* lrr: RH or RH-1 */
308 RCU_INIT_POINTER(node->avl_left, lrr); /* lrr: RH or RH-1 */
309 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */
310 node->avl_height = rh + 1; /* node: RH+1 */
311 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH */
312 RCU_INIT_POINTER(l->avl_right, lrl); /* lrl: RH or RH-1 */
313 l->avl_height = rh + 1; /* l: RH+1 */
314 RCU_INIT_POINTER(lr->avl_left, l); /* l: RH+1 */
315 RCU_INIT_POINTER(lr->avl_right, node); /* node: RH+1 */
316 lr->avl_height = rh + 2;
317 RCU_INIT_POINTER(*nodep, lr);
319 } else if (rh > lh + 1) { /* r: LH+2 */
320 struct inet_peer *rr, *rl, *rlr, *rll;
322 rr = rcu_deref_locked(r->avl_right, base);
323 rl = rcu_deref_locked(r->avl_left, base);
324 rlh = node_height(rl);
325 if (rlh <= node_height(rr)) { /* rr: LH+1 */
326 RCU_INIT_POINTER(node->avl_right, rl); /* rl: LH or LH+1 */
327 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */
328 node->avl_height = rlh + 1; /* LH+1 or LH+2 */
329 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH+1 */
330 RCU_INIT_POINTER(r->avl_left, node); /* node: LH+1 or LH+2 */
331 r->avl_height = node->avl_height + 1;
332 RCU_INIT_POINTER(*nodep, r);
333 } else { /* rr: RH, rl: RH+1 */
334 rlr = rcu_deref_locked(rl->avl_right, base);/* rlr: LH or LH-1 */
335 rll = rcu_deref_locked(rl->avl_left, base);/* rll: LH or LH-1 */
336 RCU_INIT_POINTER(node->avl_right, rll); /* rll: LH or LH-1 */
337 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */
338 node->avl_height = lh + 1; /* node: LH+1 */
339 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH */
340 RCU_INIT_POINTER(r->avl_left, rlr); /* rlr: LH or LH-1 */
341 r->avl_height = lh + 1; /* r: LH+1 */
342 RCU_INIT_POINTER(rl->avl_right, r); /* r: LH+1 */
343 RCU_INIT_POINTER(rl->avl_left, node); /* node: LH+1 */
344 rl->avl_height = lh + 2;
345 RCU_INIT_POINTER(*nodep, rl);
348 node->avl_height = (lh > rh ? lh : rh) + 1;
353 /* Called with local BH disabled and the pool lock held. */
354 #define link_to_pool(n, base) \
357 n->avl_left = peer_avl_empty_rcu; \
358 n->avl_right = peer_avl_empty_rcu; \
359 /* lockless readers can catch us now */ \
360 rcu_assign_pointer(**--stackptr, n); \
361 peer_avl_rebalance(stack, stackptr, base); \
364 static void inetpeer_free_rcu(struct rcu_head *head)
366 kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
369 /* May be called with local BH enabled. */
370 static void unlink_from_pool(struct inet_peer *p, struct inet_peer_base *base,
371 struct inet_peer __rcu **stack[PEER_MAXDEPTH])
377 write_seqlock_bh(&base->lock);
378 /* Check the reference counter. It was artificially incremented by 1
379 * in cleanup() function to prevent sudden disappearing. If we can
380 * atomically (because of lockless readers) take this last reference,
381 * it's safe to remove the node and free it later.
382 * We use refcnt=-1 to alert lockless readers this entry is deleted.
384 if (atomic_cmpxchg(&p->refcnt, 1, -1) == 1) {
385 struct inet_peer __rcu ***stackptr, ***delp;
386 if (lookup(&p->daddr, stack, base) != p)
388 delp = stackptr - 1; /* *delp[0] == p */
389 if (p->avl_left == peer_avl_empty_rcu) {
390 *delp[0] = p->avl_right;
393 /* look for a node to insert instead of p */
395 t = lookup_rightempty(p, base);
396 BUG_ON(rcu_deref_locked(*stackptr[-1], base) != t);
397 **--stackptr = t->avl_left;
398 /* t is removed, t->daddr > x->daddr for any
399 * x in p->avl_left subtree.
400 * Put t in the old place of p. */
401 RCU_INIT_POINTER(*delp[0], t);
402 t->avl_left = p->avl_left;
403 t->avl_right = p->avl_right;
404 t->avl_height = p->avl_height;
405 BUG_ON(delp[1] != &p->avl_left);
406 delp[1] = &t->avl_left; /* was &p->avl_left */
408 peer_avl_rebalance(stack, stackptr, base);
412 write_sequnlock_bh(&base->lock);
415 call_rcu(&p->rcu, inetpeer_free_rcu);
417 /* The node is used again. Decrease the reference counter
418 * back. The loop "cleanup -> unlink_from_unused
419 * -> unlink_from_pool -> putpeer -> link_to_unused
420 * -> cleanup (for the same node)"
421 * doesn't really exist because the entry will have a
422 * recent deletion time and will not be cleaned again soon.
427 static struct inet_peer_base *family_to_base(int family)
429 return (family == AF_INET ? &v4_peers : &v6_peers);
432 static struct inet_peer_base *peer_to_base(struct inet_peer *p)
434 return family_to_base(p->daddr.family);
437 /* May be called with local BH enabled. */
438 static int cleanup_once(unsigned long ttl, struct inet_peer __rcu **stack[PEER_MAXDEPTH])
440 struct inet_peer *p = NULL;
442 /* Remove the first entry from the list of unused nodes. */
443 spin_lock_bh(&unused_peers.lock);
444 if (!list_empty(&unused_peers.list)) {
447 p = list_first_entry(&unused_peers.list, struct inet_peer, unused);
448 delta = (__u32)jiffies - p->dtime;
451 /* Do not prune fresh entries. */
452 spin_unlock_bh(&unused_peers.lock);
456 list_del_init(&p->unused);
458 /* Grab an extra reference to prevent node disappearing
459 * before unlink_from_pool() call. */
460 atomic_inc(&p->refcnt);
462 spin_unlock_bh(&unused_peers.lock);
465 /* It means that the total number of USED entries has
466 * grown over inet_peer_threshold. It shouldn't really
467 * happen because of entry limits in route cache. */
470 unlink_from_pool(p, peer_to_base(p), stack);
474 /* Called with or without local BH being disabled. */
475 struct inet_peer *inet_getpeer(struct inetpeer_addr *daddr, int create)
477 struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr;
478 struct inet_peer_base *base = family_to_base(daddr->family);
480 unsigned int sequence;
481 int invalidated, newrefcnt = 0;
483 /* Look up for the address quickly, lockless.
484 * Because of a concurrent writer, we might not find an existing entry.
487 sequence = read_seqbegin(&base->lock);
488 p = lookup_rcu(daddr, base, &newrefcnt);
489 invalidated = read_seqretry(&base->lock, sequence);
493 found: /* The existing node has been found.
494 * Remove the entry from unused list if it was there.
497 unlink_from_unused(p);
501 /* If no writer did a change during our lookup, we can return early. */
502 if (!create && !invalidated)
505 /* retry an exact lookup, taking the lock before.
506 * At least, nodes should be hot in our cache.
508 write_seqlock_bh(&base->lock);
509 p = lookup(daddr, stack, base);
510 if (p != peer_avl_empty) {
511 newrefcnt = atomic_inc_return(&p->refcnt);
512 write_sequnlock_bh(&base->lock);
515 p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
518 atomic_set(&p->refcnt, 1);
519 atomic_set(&p->rid, 0);
520 atomic_set(&p->ip_id_count, secure_ip_id(daddr->addr.a4));
522 p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
527 memset(&p->redirect_learned, 0, sizeof(p->redirect_learned));
528 INIT_LIST_HEAD(&p->unused);
532 link_to_pool(p, base);
535 write_sequnlock_bh(&base->lock);
537 if (base->total >= inet_peer_threshold)
538 /* Remove one less-recently-used entry. */
539 cleanup_once(0, stack);
544 static int compute_total(void)
546 return v4_peers.total + v6_peers.total;
548 EXPORT_SYMBOL_GPL(inet_getpeer);
550 /* Called with local BH disabled. */
551 static void peer_check_expire(unsigned long dummy)
553 unsigned long now = jiffies;
555 struct inet_peer __rcu **stack[PEER_MAXDEPTH];
557 total = compute_total();
558 if (total >= inet_peer_threshold)
559 ttl = inet_peer_minttl;
561 ttl = inet_peer_maxttl
562 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
563 total / inet_peer_threshold * HZ;
564 while (!cleanup_once(ttl, stack)) {
569 /* Trigger the timer after inet_peer_gc_mintime .. inet_peer_gc_maxtime
570 * interval depending on the total number of entries (more entries,
572 total = compute_total();
573 if (total >= inet_peer_threshold)
574 peer_periodic_timer.expires = jiffies + inet_peer_gc_mintime;
576 peer_periodic_timer.expires = jiffies
577 + inet_peer_gc_maxtime
578 - (inet_peer_gc_maxtime - inet_peer_gc_mintime) / HZ *
579 total / inet_peer_threshold * HZ;
580 add_timer(&peer_periodic_timer);
583 void inet_putpeer(struct inet_peer *p)
587 if (atomic_dec_and_lock(&p->refcnt, &unused_peers.lock)) {
588 list_add_tail(&p->unused, &unused_peers.list);
589 p->dtime = (__u32)jiffies;
590 spin_unlock(&unused_peers.lock);
595 EXPORT_SYMBOL_GPL(inet_putpeer);
598 * Check transmit rate limitation for given message.
599 * The rate information is held in the inet_peer entries now.
600 * This function is generic and could be used for other purposes
601 * too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
603 * Note that the same inet_peer fields are modified by functions in
604 * route.c too, but these work for packet destinations while xrlim_allow
605 * works for icmp destinations. This means the rate limiting information
606 * for one "ip object" is shared - and these ICMPs are twice limited:
607 * by source and by destination.
609 * RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
610 * SHOULD allow setting of rate limits
612 * Shared between ICMPv4 and ICMPv6.
614 #define XRLIM_BURST_FACTOR 6
615 bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
617 unsigned long now, token;
623 token = peer->rate_tokens;
625 token += now - peer->rate_last;
626 peer->rate_last = now;
627 if (token > XRLIM_BURST_FACTOR * timeout)
628 token = XRLIM_BURST_FACTOR * timeout;
629 if (token >= timeout) {
633 peer->rate_tokens = token;
636 EXPORT_SYMBOL(inet_peer_xrlim_allow);