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 * Version: $Id: inetpeer.c,v 1.7 2001/09/20 21:22:50 davem Exp $
8 * Authors: Andrey V. Savochkin <saw@msu.ru>
11 #include <linux/module.h>
12 #include <linux/types.h>
13 #include <linux/slab.h>
14 #include <linux/interrupt.h>
15 #include <linux/spinlock.h>
16 #include <linux/random.h>
17 #include <linux/timer.h>
18 #include <linux/time.h>
19 #include <linux/kernel.h>
21 #include <linux/net.h>
23 #include <net/inetpeer.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 write lock held.
57 * 2. Nodes may disappear from the tree only with the pool write lock held
58 * AND reference count being 0.
59 * 3. Nodes appears and disappears from unused node list only under
60 * "inet_peer_unused_lock".
61 * 4. Global variable peer_total is modified under the pool lock.
62 * 5. struct inet_peer fields modification:
63 * avl_left, avl_right, avl_parent, avl_height: pool lock
64 * unused: unused node list lock
65 * refcnt: atomically against modifications on other CPU;
66 * usually under some other lock to prevent node disappearing
67 * dtime: unused node list lock
68 * v4daddr: unchangeable
72 /* Exported for inet_getid inline function. */
73 DEFINE_SPINLOCK(inet_peer_idlock);
75 static struct kmem_cache *peer_cachep __read_mostly;
77 #define node_height(x) x->avl_height
78 static struct inet_peer peer_fake_node = {
79 .avl_left = &peer_fake_node,
80 .avl_right = &peer_fake_node,
83 #define peer_avl_empty (&peer_fake_node)
84 static struct inet_peer *peer_root = peer_avl_empty;
85 static DEFINE_RWLOCK(peer_pool_lock);
86 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
88 static int peer_total;
89 /* Exported for sysctl_net_ipv4. */
90 int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more
91 * aggressively at this stage */
92 int inet_peer_minttl __read_mostly = 120 * HZ; /* TTL under high load: 120 sec */
93 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ; /* usual time to live: 10 min */
94 int inet_peer_gc_mintime __read_mostly = 10 * HZ;
95 int inet_peer_gc_maxtime __read_mostly = 120 * HZ;
97 static LIST_HEAD(unused_peers);
98 static DEFINE_SPINLOCK(inet_peer_unused_lock);
100 static void peer_check_expire(unsigned long dummy);
101 static DEFINE_TIMER(peer_periodic_timer, peer_check_expire, 0, 0);
104 /* Called from ip_output.c:ip_init */
105 void __init inet_initpeers(void)
109 /* Use the straight interface to information about memory. */
111 /* The values below were suggested by Alexey Kuznetsov
112 * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values
115 if (si.totalram <= (32768*1024)/PAGE_SIZE)
116 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
117 if (si.totalram <= (16384*1024)/PAGE_SIZE)
118 inet_peer_threshold >>= 1; /* about 512KB */
119 if (si.totalram <= (8192*1024)/PAGE_SIZE)
120 inet_peer_threshold >>= 2; /* about 128KB */
122 peer_cachep = kmem_cache_create("inet_peer_cache",
123 sizeof(struct inet_peer),
124 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC,
127 /* All the timers, started at system startup tend
128 to synchronize. Perturb it a bit.
130 peer_periodic_timer.expires = jiffies
131 + net_random() % inet_peer_gc_maxtime
132 + inet_peer_gc_maxtime;
133 add_timer(&peer_periodic_timer);
136 /* Called with or without local BH being disabled. */
137 static void unlink_from_unused(struct inet_peer *p)
139 spin_lock_bh(&inet_peer_unused_lock);
140 list_del_init(&p->unused);
141 spin_unlock_bh(&inet_peer_unused_lock);
145 * Called with local BH disabled and the pool lock held.
146 * _stack is known to be NULL or not at compile time,
147 * so compiler will optimize the if (_stack) tests.
149 #define lookup(_daddr,_stack) \
151 struct inet_peer *u, **v; \
152 if (_stack != NULL) { \
154 *stackptr++ = &peer_root; \
156 for (u = peer_root; u != peer_avl_empty; ) { \
157 if (_daddr == u->v4daddr) \
159 if ((__force __u32)_daddr < (__force __u32)u->v4daddr) \
163 if (_stack != NULL) \
170 /* Called with local BH disabled and the pool write lock held. */
171 #define lookup_rightempty(start) \
173 struct inet_peer *u, **v; \
174 *stackptr++ = &start->avl_left; \
175 v = &start->avl_left; \
176 for (u = *v; u->avl_right != peer_avl_empty; ) { \
184 /* Called with local BH disabled and the pool write lock held.
185 * Variable names are the proof of operation correctness.
186 * Look into mm/map_avl.c for more detail description of the ideas. */
187 static void peer_avl_rebalance(struct inet_peer **stack[],
188 struct inet_peer ***stackend)
190 struct inet_peer **nodep, *node, *l, *r;
193 while (stackend > stack) {
200 if (lh > rh + 1) { /* l: RH+2 */
201 struct inet_peer *ll, *lr, *lrl, *lrr;
205 lrh = node_height(lr);
206 if (lrh <= node_height(ll)) { /* ll: RH+1 */
207 node->avl_left = lr; /* lr: RH or RH+1 */
208 node->avl_right = r; /* r: RH */
209 node->avl_height = lrh + 1; /* RH+1 or RH+2 */
210 l->avl_left = ll; /* ll: RH+1 */
211 l->avl_right = node; /* node: RH+1 or RH+2 */
212 l->avl_height = node->avl_height + 1;
214 } else { /* ll: RH, lr: RH+1 */
215 lrl = lr->avl_left; /* lrl: RH or RH-1 */
216 lrr = lr->avl_right; /* lrr: RH or RH-1 */
217 node->avl_left = lrr; /* lrr: RH or RH-1 */
218 node->avl_right = r; /* r: RH */
219 node->avl_height = rh + 1; /* node: RH+1 */
220 l->avl_left = ll; /* ll: RH */
221 l->avl_right = lrl; /* lrl: RH or RH-1 */
222 l->avl_height = rh + 1; /* l: RH+1 */
223 lr->avl_left = l; /* l: RH+1 */
224 lr->avl_right = node; /* node: RH+1 */
225 lr->avl_height = rh + 2;
228 } else if (rh > lh + 1) { /* r: LH+2 */
229 struct inet_peer *rr, *rl, *rlr, *rll;
233 rlh = node_height(rl);
234 if (rlh <= node_height(rr)) { /* rr: LH+1 */
235 node->avl_right = rl; /* rl: LH or LH+1 */
236 node->avl_left = l; /* l: LH */
237 node->avl_height = rlh + 1; /* LH+1 or LH+2 */
238 r->avl_right = rr; /* rr: LH+1 */
239 r->avl_left = node; /* node: LH+1 or LH+2 */
240 r->avl_height = node->avl_height + 1;
242 } else { /* rr: RH, rl: RH+1 */
243 rlr = rl->avl_right; /* rlr: LH or LH-1 */
244 rll = rl->avl_left; /* rll: LH or LH-1 */
245 node->avl_right = rll; /* rll: LH or LH-1 */
246 node->avl_left = l; /* l: LH */
247 node->avl_height = lh + 1; /* node: LH+1 */
248 r->avl_right = rr; /* rr: LH */
249 r->avl_left = rlr; /* rlr: LH or LH-1 */
250 r->avl_height = lh + 1; /* r: LH+1 */
251 rl->avl_right = r; /* r: LH+1 */
252 rl->avl_left = node; /* node: LH+1 */
253 rl->avl_height = lh + 2;
257 node->avl_height = (lh > rh ? lh : rh) + 1;
262 /* Called with local BH disabled and the pool write lock held. */
263 #define link_to_pool(n) \
266 n->avl_left = peer_avl_empty; \
267 n->avl_right = peer_avl_empty; \
269 peer_avl_rebalance(stack, stackptr); \
272 /* May be called with local BH enabled. */
273 static void unlink_from_pool(struct inet_peer *p)
279 write_lock_bh(&peer_pool_lock);
280 /* Check the reference counter. It was artificially incremented by 1
281 * in cleanup() function to prevent sudden disappearing. If the
282 * reference count is still 1 then the node is referenced only as `p'
283 * here and from the pool. So under the exclusive pool lock it's safe
284 * to remove the node and free it later. */
285 if (atomic_read(&p->refcnt) == 1) {
286 struct inet_peer **stack[PEER_MAXDEPTH];
287 struct inet_peer ***stackptr, ***delp;
288 if (lookup(p->v4daddr, stack) != p)
290 delp = stackptr - 1; /* *delp[0] == p */
291 if (p->avl_left == peer_avl_empty) {
292 *delp[0] = p->avl_right;
295 /* look for a node to insert instead of p */
297 t = lookup_rightempty(p);
298 BUG_ON(*stackptr[-1] != t);
299 **--stackptr = t->avl_left;
300 /* t is removed, t->v4daddr > x->v4daddr for any
301 * x in p->avl_left subtree.
302 * Put t in the old place of p. */
304 t->avl_left = p->avl_left;
305 t->avl_right = p->avl_right;
306 t->avl_height = p->avl_height;
307 BUG_ON(delp[1] != &p->avl_left);
308 delp[1] = &t->avl_left; /* was &p->avl_left */
310 peer_avl_rebalance(stack, stackptr);
314 write_unlock_bh(&peer_pool_lock);
317 kmem_cache_free(peer_cachep, p);
319 /* The node is used again. Decrease the reference counter
320 * back. The loop "cleanup -> unlink_from_unused
321 * -> unlink_from_pool -> putpeer -> link_to_unused
322 * -> cleanup (for the same node)"
323 * doesn't really exist because the entry will have a
324 * recent deletion time and will not be cleaned again soon. */
328 /* May be called with local BH enabled. */
329 static int cleanup_once(unsigned long ttl)
331 struct inet_peer *p = NULL;
333 /* Remove the first entry from the list of unused nodes. */
334 spin_lock_bh(&inet_peer_unused_lock);
335 if (!list_empty(&unused_peers)) {
338 p = list_first_entry(&unused_peers, struct inet_peer, unused);
339 delta = (__u32)jiffies - p->dtime;
342 /* Do not prune fresh entries. */
343 spin_unlock_bh(&inet_peer_unused_lock);
347 list_del_init(&p->unused);
349 /* Grab an extra reference to prevent node disappearing
350 * before unlink_from_pool() call. */
351 atomic_inc(&p->refcnt);
353 spin_unlock_bh(&inet_peer_unused_lock);
356 /* It means that the total number of USED entries has
357 * grown over inet_peer_threshold. It shouldn't really
358 * happen because of entry limits in route cache. */
365 /* Called with or without local BH being disabled. */
366 struct inet_peer *inet_getpeer(__be32 daddr, int create)
368 struct inet_peer *p, *n;
369 struct inet_peer **stack[PEER_MAXDEPTH], ***stackptr;
371 /* Look up for the address quickly. */
372 read_lock_bh(&peer_pool_lock);
373 p = lookup(daddr, NULL);
374 if (p != peer_avl_empty)
375 atomic_inc(&p->refcnt);
376 read_unlock_bh(&peer_pool_lock);
378 if (p != peer_avl_empty) {
379 /* The existing node has been found. */
380 /* Remove the entry from unused list if it was there. */
381 unlink_from_unused(p);
388 /* Allocate the space outside the locked region. */
389 n = kmem_cache_alloc(peer_cachep, GFP_ATOMIC);
393 atomic_set(&n->refcnt, 1);
394 atomic_set(&n->rid, 0);
395 n->ip_id_count = secure_ip_id(daddr);
398 write_lock_bh(&peer_pool_lock);
399 /* Check if an entry has suddenly appeared. */
400 p = lookup(daddr, stack);
401 if (p != peer_avl_empty)
406 INIT_LIST_HEAD(&n->unused);
408 write_unlock_bh(&peer_pool_lock);
410 if (peer_total >= inet_peer_threshold)
411 /* Remove one less-recently-used entry. */
417 /* The appropriate node is already in the pool. */
418 atomic_inc(&p->refcnt);
419 write_unlock_bh(&peer_pool_lock);
420 /* Remove the entry from unused list if it was there. */
421 unlink_from_unused(p);
422 /* Free preallocated the preallocated node. */
423 kmem_cache_free(peer_cachep, n);
427 /* Called with local BH disabled. */
428 static void peer_check_expire(unsigned long dummy)
430 unsigned long now = jiffies;
433 if (peer_total >= inet_peer_threshold)
434 ttl = inet_peer_minttl;
436 ttl = inet_peer_maxttl
437 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
438 peer_total / inet_peer_threshold * HZ;
439 while (!cleanup_once(ttl)) {
444 /* Trigger the timer after inet_peer_gc_mintime .. inet_peer_gc_maxtime
445 * interval depending on the total number of entries (more entries,
447 if (peer_total >= inet_peer_threshold)
448 peer_periodic_timer.expires = jiffies + inet_peer_gc_mintime;
450 peer_periodic_timer.expires = jiffies
451 + inet_peer_gc_maxtime
452 - (inet_peer_gc_maxtime - inet_peer_gc_mintime) / HZ *
453 peer_total / inet_peer_threshold * HZ;
454 add_timer(&peer_periodic_timer);
457 void inet_putpeer(struct inet_peer *p)
459 spin_lock_bh(&inet_peer_unused_lock);
460 if (atomic_dec_and_test(&p->refcnt)) {
461 list_add_tail(&p->unused, &unused_peers);
462 p->dtime = (__u32)jiffies;
464 spin_unlock_bh(&inet_peer_unused_lock);