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
66 * v4daddr: unchangeable
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 static const struct inet_peer peer_fake_node = {
76 .avl_left = peer_avl_empty,
77 .avl_right = peer_avl_empty,
82 struct inet_peer *root;
86 .root = peer_avl_empty,
87 .lock = __SPIN_LOCK_UNLOCKED(peers.lock),
90 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
92 /* Exported for sysctl_net_ipv4. */
93 int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more
94 * aggressively at this stage */
95 int inet_peer_minttl __read_mostly = 120 * HZ; /* TTL under high load: 120 sec */
96 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ; /* usual time to live: 10 min */
97 int inet_peer_gc_mintime __read_mostly = 10 * HZ;
98 int inet_peer_gc_maxtime __read_mostly = 120 * HZ;
101 struct list_head list;
104 .list = LIST_HEAD_INIT(unused_peers.list),
105 .lock = __SPIN_LOCK_UNLOCKED(unused_peers.lock),
108 static void peer_check_expire(unsigned long dummy);
109 static DEFINE_TIMER(peer_periodic_timer, peer_check_expire, 0, 0);
112 /* Called from ip_output.c:ip_init */
113 void __init inet_initpeers(void)
117 /* Use the straight interface to information about memory. */
119 /* The values below were suggested by Alexey Kuznetsov
120 * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values
123 if (si.totalram <= (32768*1024)/PAGE_SIZE)
124 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
125 if (si.totalram <= (16384*1024)/PAGE_SIZE)
126 inet_peer_threshold >>= 1; /* about 512KB */
127 if (si.totalram <= (8192*1024)/PAGE_SIZE)
128 inet_peer_threshold >>= 2; /* about 128KB */
130 peer_cachep = kmem_cache_create("inet_peer_cache",
131 sizeof(struct inet_peer),
132 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
135 /* All the timers, started at system startup tend
136 to synchronize. Perturb it a bit.
138 peer_periodic_timer.expires = jiffies
139 + net_random() % inet_peer_gc_maxtime
140 + inet_peer_gc_maxtime;
141 add_timer(&peer_periodic_timer);
144 /* Called with or without local BH being disabled. */
145 static void unlink_from_unused(struct inet_peer *p)
147 if (!list_empty(&p->unused)) {
148 spin_lock_bh(&unused_peers.lock);
149 list_del_init(&p->unused);
150 spin_unlock_bh(&unused_peers.lock);
155 * Called with local BH disabled and the pool lock held.
157 #define lookup(_daddr, _stack) \
159 struct inet_peer *u, **v; \
162 *stackptr++ = &peers.root; \
163 for (u = peers.root; u != peer_avl_empty; ) { \
164 if (_daddr == u->v4daddr) \
166 if ((__force __u32)_daddr < (__force __u32)u->v4daddr) \
177 * Called with rcu_read_lock_bh()
178 * Because we hold no lock against a writer, its quite possible we fall
179 * in an endless loop.
180 * But every pointer we follow is guaranteed to be valid thanks to RCU.
181 * We exit from this function if number of links exceeds PEER_MAXDEPTH
183 static struct inet_peer *lookup_rcu_bh(__be32 daddr)
185 struct inet_peer *u = rcu_dereference_bh(peers.root);
188 while (u != peer_avl_empty) {
189 if (daddr == u->v4daddr) {
190 /* Before taking a reference, check if this entry was
191 * deleted, unlink_from_pool() sets refcnt=-1 to make
192 * distinction between an unused entry (refcnt=0) and
195 if (unlikely(!atomic_add_unless(&u->refcnt, 1, -1)))
199 if ((__force __u32)daddr < (__force __u32)u->v4daddr)
200 u = rcu_dereference_bh(u->avl_left);
202 u = rcu_dereference_bh(u->avl_right);
203 if (unlikely(++count == PEER_MAXDEPTH))
209 /* Called with local BH disabled and the pool lock held. */
210 #define lookup_rightempty(start) \
212 struct inet_peer *u, **v; \
213 *stackptr++ = &start->avl_left; \
214 v = &start->avl_left; \
215 for (u = *v; u->avl_right != peer_avl_empty; ) { \
223 /* Called with local BH disabled and the pool lock held.
224 * Variable names are the proof of operation correctness.
225 * Look into mm/map_avl.c for more detail description of the ideas.
227 static void peer_avl_rebalance(struct inet_peer **stack[],
228 struct inet_peer ***stackend)
230 struct inet_peer **nodep, *node, *l, *r;
233 while (stackend > stack) {
240 if (lh > rh + 1) { /* l: RH+2 */
241 struct inet_peer *ll, *lr, *lrl, *lrr;
245 lrh = node_height(lr);
246 if (lrh <= node_height(ll)) { /* ll: RH+1 */
247 node->avl_left = lr; /* lr: RH or RH+1 */
248 node->avl_right = r; /* r: RH */
249 node->avl_height = lrh + 1; /* RH+1 or RH+2 */
250 l->avl_left = ll; /* ll: RH+1 */
251 l->avl_right = node; /* node: RH+1 or RH+2 */
252 l->avl_height = node->avl_height + 1;
254 } else { /* ll: RH, lr: RH+1 */
255 lrl = lr->avl_left; /* lrl: RH or RH-1 */
256 lrr = lr->avl_right; /* lrr: RH or RH-1 */
257 node->avl_left = lrr; /* lrr: RH or RH-1 */
258 node->avl_right = r; /* r: RH */
259 node->avl_height = rh + 1; /* node: RH+1 */
260 l->avl_left = ll; /* ll: RH */
261 l->avl_right = lrl; /* lrl: RH or RH-1 */
262 l->avl_height = rh + 1; /* l: RH+1 */
263 lr->avl_left = l; /* l: RH+1 */
264 lr->avl_right = node; /* node: RH+1 */
265 lr->avl_height = rh + 2;
268 } else if (rh > lh + 1) { /* r: LH+2 */
269 struct inet_peer *rr, *rl, *rlr, *rll;
273 rlh = node_height(rl);
274 if (rlh <= node_height(rr)) { /* rr: LH+1 */
275 node->avl_right = rl; /* rl: LH or LH+1 */
276 node->avl_left = l; /* l: LH */
277 node->avl_height = rlh + 1; /* LH+1 or LH+2 */
278 r->avl_right = rr; /* rr: LH+1 */
279 r->avl_left = node; /* node: LH+1 or LH+2 */
280 r->avl_height = node->avl_height + 1;
282 } else { /* rr: RH, rl: RH+1 */
283 rlr = rl->avl_right; /* rlr: LH or LH-1 */
284 rll = rl->avl_left; /* rll: LH or LH-1 */
285 node->avl_right = rll; /* rll: LH or LH-1 */
286 node->avl_left = l; /* l: LH */
287 node->avl_height = lh + 1; /* node: LH+1 */
288 r->avl_right = rr; /* rr: LH */
289 r->avl_left = rlr; /* rlr: LH or LH-1 */
290 r->avl_height = lh + 1; /* r: LH+1 */
291 rl->avl_right = r; /* r: LH+1 */
292 rl->avl_left = node; /* node: LH+1 */
293 rl->avl_height = lh + 2;
297 node->avl_height = (lh > rh ? lh : rh) + 1;
302 /* Called with local BH disabled and the pool lock held. */
303 #define link_to_pool(n) \
306 n->avl_left = peer_avl_empty; \
307 n->avl_right = peer_avl_empty; \
308 smp_wmb(); /* lockless readers can catch us now */ \
310 peer_avl_rebalance(stack, stackptr); \
313 static void inetpeer_free_rcu(struct rcu_head *head)
315 kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
318 /* May be called with local BH enabled. */
319 static void unlink_from_pool(struct inet_peer *p)
325 spin_lock_bh(&peers.lock);
326 /* Check the reference counter. It was artificially incremented by 1
327 * in cleanup() function to prevent sudden disappearing. If we can
328 * atomically (because of lockless readers) take this last reference,
329 * it's safe to remove the node and free it later.
330 * We use refcnt=-1 to alert lockless readers this entry is deleted.
332 if (atomic_cmpxchg(&p->refcnt, 1, -1) == 1) {
333 struct inet_peer **stack[PEER_MAXDEPTH];
334 struct inet_peer ***stackptr, ***delp;
335 if (lookup(p->v4daddr, stack) != p)
337 delp = stackptr - 1; /* *delp[0] == p */
338 if (p->avl_left == peer_avl_empty) {
339 *delp[0] = p->avl_right;
342 /* look for a node to insert instead of p */
344 t = lookup_rightempty(p);
345 BUG_ON(*stackptr[-1] != t);
346 **--stackptr = t->avl_left;
347 /* t is removed, t->v4daddr > x->v4daddr for any
348 * x in p->avl_left subtree.
349 * Put t in the old place of p. */
351 t->avl_left = p->avl_left;
352 t->avl_right = p->avl_right;
353 t->avl_height = p->avl_height;
354 BUG_ON(delp[1] != &p->avl_left);
355 delp[1] = &t->avl_left; /* was &p->avl_left */
357 peer_avl_rebalance(stack, stackptr);
361 spin_unlock_bh(&peers.lock);
364 call_rcu_bh(&p->rcu, inetpeer_free_rcu);
366 /* The node is used again. Decrease the reference counter
367 * back. The loop "cleanup -> unlink_from_unused
368 * -> unlink_from_pool -> putpeer -> link_to_unused
369 * -> cleanup (for the same node)"
370 * doesn't really exist because the entry will have a
371 * recent deletion time and will not be cleaned again soon.
376 /* May be called with local BH enabled. */
377 static int cleanup_once(unsigned long ttl)
379 struct inet_peer *p = NULL;
381 /* Remove the first entry from the list of unused nodes. */
382 spin_lock_bh(&unused_peers.lock);
383 if (!list_empty(&unused_peers.list)) {
386 p = list_first_entry(&unused_peers.list, struct inet_peer, unused);
387 delta = (__u32)jiffies - p->dtime;
390 /* Do not prune fresh entries. */
391 spin_unlock_bh(&unused_peers.lock);
395 list_del_init(&p->unused);
397 /* Grab an extra reference to prevent node disappearing
398 * before unlink_from_pool() call. */
399 atomic_inc(&p->refcnt);
401 spin_unlock_bh(&unused_peers.lock);
404 /* It means that the total number of USED entries has
405 * grown over inet_peer_threshold. It shouldn't really
406 * happen because of entry limits in route cache. */
413 /* Called with or without local BH being disabled. */
414 struct inet_peer *inet_getpeer(__be32 daddr, int create)
417 struct inet_peer **stack[PEER_MAXDEPTH], ***stackptr;
419 /* Look up for the address quickly, lockless.
420 * Because of a concurrent writer, we might not find an existing entry.
423 p = lookup_rcu_bh(daddr);
424 rcu_read_unlock_bh();
427 /* The existing node has been found.
428 * Remove the entry from unused list if it was there.
430 unlink_from_unused(p);
434 /* retry an exact lookup, taking the lock before.
435 * At least, nodes should be hot in our cache.
437 spin_lock_bh(&peers.lock);
438 p = lookup(daddr, stack);
439 if (p != peer_avl_empty) {
440 atomic_inc(&p->refcnt);
441 spin_unlock_bh(&peers.lock);
442 /* Remove the entry from unused list if it was there. */
443 unlink_from_unused(p);
446 p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
449 atomic_set(&p->refcnt, 1);
450 atomic_set(&p->rid, 0);
451 atomic_set(&p->ip_id_count, secure_ip_id(daddr));
453 INIT_LIST_HEAD(&p->unused);
460 spin_unlock_bh(&peers.lock);
462 if (peers.total >= inet_peer_threshold)
463 /* Remove one less-recently-used entry. */
469 /* Called with local BH disabled. */
470 static void peer_check_expire(unsigned long dummy)
472 unsigned long now = jiffies;
475 if (peers.total >= inet_peer_threshold)
476 ttl = inet_peer_minttl;
478 ttl = inet_peer_maxttl
479 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
480 peers.total / inet_peer_threshold * HZ;
481 while (!cleanup_once(ttl)) {
486 /* Trigger the timer after inet_peer_gc_mintime .. inet_peer_gc_maxtime
487 * interval depending on the total number of entries (more entries,
489 if (peers.total >= inet_peer_threshold)
490 peer_periodic_timer.expires = jiffies + inet_peer_gc_mintime;
492 peer_periodic_timer.expires = jiffies
493 + inet_peer_gc_maxtime
494 - (inet_peer_gc_maxtime - inet_peer_gc_mintime) / HZ *
495 peers.total / inet_peer_threshold * HZ;
496 add_timer(&peer_periodic_timer);
499 void inet_putpeer(struct inet_peer *p)
503 if (atomic_dec_and_lock(&p->refcnt, &unused_peers.lock)) {
504 list_add_tail(&p->unused, &unused_peers.list);
505 p->dtime = (__u32)jiffies;
506 spin_unlock(&unused_peers.lock);