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/sched.h>
18 #include <linux/timer.h>
19 #include <linux/time.h>
20 #include <linux/kernel.h>
22 #include <linux/net.h>
24 #include <net/inetpeer.h>
27 * Theory of operations.
28 * We keep one entry for each peer IP address. The nodes contains long-living
29 * information about the peer which doesn't depend on routes.
30 * At this moment this information consists only of ID field for the next
31 * outgoing IP packet. This field is incremented with each packet as encoded
32 * in inet_getid() function (include/net/inetpeer.h).
33 * At the moment of writing this notes identifier of IP packets is generated
34 * to be unpredictable using this code only for packets subjected
35 * (actually or potentially) to defragmentation. I.e. DF packets less than
36 * PMTU in size uses a constant ID and do not use this code (see
37 * ip_select_ident() in include/net/ip.h).
39 * Route cache entries hold references to our nodes.
40 * New cache entries get references via lookup by destination IP address in
41 * the avl tree. The reference is grabbed only when it's needed i.e. only
42 * when we try to output IP packet which needs an unpredictable ID (see
43 * __ip_select_ident() in net/ipv4/route.c).
44 * Nodes are removed only when reference counter goes to 0.
45 * When it's happened the node may be removed when a sufficient amount of
46 * time has been passed since its last use. The less-recently-used entry can
47 * also be removed if the pool is overloaded i.e. if the total amount of
48 * entries is greater-or-equal than the threshold.
50 * Node pool is organised as an AVL tree.
51 * Such an implementation has been chosen not just for fun. It's a way to
52 * prevent easy and efficient DoS attacks by creating hash collisions. A huge
53 * amount of long living nodes in a single hash slot would significantly delay
54 * lookups performed with disabled BHs.
56 * Serialisation issues.
57 * 1. Nodes may appear in the tree only with the pool write lock held.
58 * 2. Nodes may disappear from the tree only with the pool write lock held
59 * AND reference count being 0.
60 * 3. Nodes appears and disappears from unused node list only under
61 * "inet_peer_unused_lock".
62 * 4. Global variable peer_total is modified under the pool lock.
63 * 5. struct inet_peer fields modification:
64 * avl_left, avl_right, avl_parent, avl_height: pool lock
65 * unused_next, unused_prevp: unused node list lock
66 * refcnt: atomically against modifications on other CPU;
67 * usually under some other lock to prevent node disappearing
68 * dtime: unused node list lock
69 * v4daddr: unchangeable
73 /* Exported for inet_getid inline function. */
74 DEFINE_SPINLOCK(inet_peer_idlock);
76 static kmem_cache_t *peer_cachep __read_mostly;
78 #define node_height(x) x->avl_height
79 static struct inet_peer peer_fake_node = {
80 .avl_left = &peer_fake_node,
81 .avl_right = &peer_fake_node,
84 #define peer_avl_empty (&peer_fake_node)
85 static struct inet_peer *peer_root = peer_avl_empty;
86 static DEFINE_RWLOCK(peer_pool_lock);
87 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
89 static int peer_total;
90 /* Exported for sysctl_net_ipv4. */
91 int inet_peer_threshold = 65536 + 128; /* start to throw entries more
92 * aggressively at this stage */
93 int inet_peer_minttl = 120 * HZ; /* TTL under high load: 120 sec */
94 int inet_peer_maxttl = 10 * 60 * HZ; /* usual time to live: 10 min */
96 static struct inet_peer *inet_peer_unused_head;
97 static struct inet_peer **inet_peer_unused_tailp = &inet_peer_unused_head;
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);
103 /* Exported for sysctl_net_ipv4. */
104 int inet_peer_gc_mintime = 10 * HZ,
105 inet_peer_gc_maxtime = 120 * HZ;
107 /* Called from ip_output.c:ip_init */
108 void __init inet_initpeers(void)
112 /* Use the straight interface to information about memory. */
114 /* The values below were suggested by Alexey Kuznetsov
115 * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values
118 if (si.totalram <= (32768*1024)/PAGE_SIZE)
119 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
120 if (si.totalram <= (16384*1024)/PAGE_SIZE)
121 inet_peer_threshold >>= 1; /* about 512KB */
122 if (si.totalram <= (8192*1024)/PAGE_SIZE)
123 inet_peer_threshold >>= 2; /* about 128KB */
125 peer_cachep = kmem_cache_create("inet_peer_cache",
126 sizeof(struct inet_peer),
127 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC,
130 /* All the timers, started at system startup tend
131 to synchronize. Perturb it a bit.
133 peer_periodic_timer.expires = jiffies
134 + net_random() % inet_peer_gc_maxtime
135 + inet_peer_gc_maxtime;
136 add_timer(&peer_periodic_timer);
139 /* Called with or without local BH being disabled. */
140 static void unlink_from_unused(struct inet_peer *p)
142 spin_lock_bh(&inet_peer_unused_lock);
143 if (p->unused_prevp != NULL) {
144 /* On unused list. */
145 *p->unused_prevp = p->unused_next;
146 if (p->unused_next != NULL)
147 p->unused_next->unused_prevp = p->unused_prevp;
149 inet_peer_unused_tailp = p->unused_prevp;
150 p->unused_prevp = NULL; /* mark it as removed */
152 spin_unlock_bh(&inet_peer_unused_lock);
155 /* Called with local BH disabled and the pool lock held. */
156 #define lookup(daddr) \
158 struct inet_peer *u, **v; \
160 *stackptr++ = &peer_root; \
161 for (u = peer_root; u != peer_avl_empty; ) { \
162 if (daddr == u->v4daddr) \
164 if ((__force __u32)daddr < (__force __u32)u->v4daddr) \
174 /* Called with local BH disabled and the pool write lock held. */
175 #define lookup_rightempty(start) \
177 struct inet_peer *u, **v; \
178 *stackptr++ = &start->avl_left; \
179 v = &start->avl_left; \
180 for (u = *v; u->avl_right != peer_avl_empty; ) { \
188 /* Called with local BH disabled and the pool write lock held.
189 * Variable names are the proof of operation correctness.
190 * Look into mm/map_avl.c for more detail description of the ideas. */
191 static void peer_avl_rebalance(struct inet_peer **stack[],
192 struct inet_peer ***stackend)
194 struct inet_peer **nodep, *node, *l, *r;
197 while (stackend > stack) {
204 if (lh > rh + 1) { /* l: RH+2 */
205 struct inet_peer *ll, *lr, *lrl, *lrr;
209 lrh = node_height(lr);
210 if (lrh <= node_height(ll)) { /* ll: RH+1 */
211 node->avl_left = lr; /* lr: RH or RH+1 */
212 node->avl_right = r; /* r: RH */
213 node->avl_height = lrh + 1; /* RH+1 or RH+2 */
214 l->avl_left = ll; /* ll: RH+1 */
215 l->avl_right = node; /* node: RH+1 or RH+2 */
216 l->avl_height = node->avl_height + 1;
218 } else { /* ll: RH, lr: RH+1 */
219 lrl = lr->avl_left; /* lrl: RH or RH-1 */
220 lrr = lr->avl_right; /* lrr: RH or RH-1 */
221 node->avl_left = lrr; /* lrr: RH or RH-1 */
222 node->avl_right = r; /* r: RH */
223 node->avl_height = rh + 1; /* node: RH+1 */
224 l->avl_left = ll; /* ll: RH */
225 l->avl_right = lrl; /* lrl: RH or RH-1 */
226 l->avl_height = rh + 1; /* l: RH+1 */
227 lr->avl_left = l; /* l: RH+1 */
228 lr->avl_right = node; /* node: RH+1 */
229 lr->avl_height = rh + 2;
232 } else if (rh > lh + 1) { /* r: LH+2 */
233 struct inet_peer *rr, *rl, *rlr, *rll;
237 rlh = node_height(rl);
238 if (rlh <= node_height(rr)) { /* rr: LH+1 */
239 node->avl_right = rl; /* rl: LH or LH+1 */
240 node->avl_left = l; /* l: LH */
241 node->avl_height = rlh + 1; /* LH+1 or LH+2 */
242 r->avl_right = rr; /* rr: LH+1 */
243 r->avl_left = node; /* node: LH+1 or LH+2 */
244 r->avl_height = node->avl_height + 1;
246 } else { /* rr: RH, rl: RH+1 */
247 rlr = rl->avl_right; /* rlr: LH or LH-1 */
248 rll = rl->avl_left; /* rll: LH or LH-1 */
249 node->avl_right = rll; /* rll: LH or LH-1 */
250 node->avl_left = l; /* l: LH */
251 node->avl_height = lh + 1; /* node: LH+1 */
252 r->avl_right = rr; /* rr: LH */
253 r->avl_left = rlr; /* rlr: LH or LH-1 */
254 r->avl_height = lh + 1; /* r: LH+1 */
255 rl->avl_right = r; /* r: LH+1 */
256 rl->avl_left = node; /* node: LH+1 */
257 rl->avl_height = lh + 2;
261 node->avl_height = (lh > rh ? lh : rh) + 1;
266 /* Called with local BH disabled and the pool write lock held. */
267 #define link_to_pool(n) \
270 n->avl_left = peer_avl_empty; \
271 n->avl_right = peer_avl_empty; \
273 peer_avl_rebalance(stack, stackptr); \
276 /* May be called with local BH enabled. */
277 static void unlink_from_pool(struct inet_peer *p)
283 write_lock_bh(&peer_pool_lock);
284 /* Check the reference counter. It was artificially incremented by 1
285 * in cleanup() function to prevent sudden disappearing. If the
286 * reference count is still 1 then the node is referenced only as `p'
287 * here and from the pool. So under the exclusive pool lock it's safe
288 * to remove the node and free it later. */
289 if (atomic_read(&p->refcnt) == 1) {
290 struct inet_peer **stack[PEER_MAXDEPTH];
291 struct inet_peer ***stackptr, ***delp;
292 if (lookup(p->v4daddr) != p)
294 delp = stackptr - 1; /* *delp[0] == p */
295 if (p->avl_left == peer_avl_empty) {
296 *delp[0] = p->avl_right;
299 /* look for a node to insert instead of p */
301 t = lookup_rightempty(p);
302 BUG_ON(*stackptr[-1] != t);
303 **--stackptr = t->avl_left;
304 /* t is removed, t->v4daddr > x->v4daddr for any
305 * x in p->avl_left subtree.
306 * Put t in the old place of p. */
308 t->avl_left = p->avl_left;
309 t->avl_right = p->avl_right;
310 t->avl_height = p->avl_height;
311 BUG_ON(delp[1] != &p->avl_left);
312 delp[1] = &t->avl_left; /* was &p->avl_left */
314 peer_avl_rebalance(stack, stackptr);
318 write_unlock_bh(&peer_pool_lock);
321 kmem_cache_free(peer_cachep, p);
323 /* The node is used again. Decrease the reference counter
324 * back. The loop "cleanup -> unlink_from_unused
325 * -> unlink_from_pool -> putpeer -> link_to_unused
326 * -> cleanup (for the same node)"
327 * doesn't really exist because the entry will have a
328 * recent deletion time and will not be cleaned again soon. */
332 /* May be called with local BH enabled. */
333 static int cleanup_once(unsigned long ttl)
337 /* Remove the first entry from the list of unused nodes. */
338 spin_lock_bh(&inet_peer_unused_lock);
339 p = inet_peer_unused_head;
341 __u32 delta = (__u32)jiffies - p->dtime;
343 /* Do not prune fresh entries. */
344 spin_unlock_bh(&inet_peer_unused_lock);
347 inet_peer_unused_head = p->unused_next;
348 if (p->unused_next != NULL)
349 p->unused_next->unused_prevp = p->unused_prevp;
351 inet_peer_unused_tailp = p->unused_prevp;
352 p->unused_prevp = NULL; /* mark as not on the list */
353 /* Grab an extra reference to prevent node disappearing
354 * before unlink_from_pool() call. */
355 atomic_inc(&p->refcnt);
357 spin_unlock_bh(&inet_peer_unused_lock);
360 /* It means that the total number of USED entries has
361 * grown over inet_peer_threshold. It shouldn't really
362 * happen because of entry limits in route cache. */
369 /* Called with or without local BH being disabled. */
370 struct inet_peer *inet_getpeer(__be32 daddr, int create)
372 struct inet_peer *p, *n;
373 struct inet_peer **stack[PEER_MAXDEPTH], ***stackptr;
375 /* Look up for the address quickly. */
376 read_lock_bh(&peer_pool_lock);
378 if (p != peer_avl_empty)
379 atomic_inc(&p->refcnt);
380 read_unlock_bh(&peer_pool_lock);
382 if (p != peer_avl_empty) {
383 /* The existing node has been found. */
384 /* Remove the entry from unused list if it was there. */
385 unlink_from_unused(p);
392 /* Allocate the space outside the locked region. */
393 n = kmem_cache_alloc(peer_cachep, GFP_ATOMIC);
397 atomic_set(&n->refcnt, 1);
398 atomic_set(&n->rid, 0);
399 n->ip_id_count = secure_ip_id(daddr);
402 write_lock_bh(&peer_pool_lock);
403 /* Check if an entry has suddenly appeared. */
405 if (p != peer_avl_empty)
410 n->unused_prevp = NULL; /* not on the list */
412 write_unlock_bh(&peer_pool_lock);
414 if (peer_total >= inet_peer_threshold)
415 /* Remove one less-recently-used entry. */
421 /* The appropriate node is already in the pool. */
422 atomic_inc(&p->refcnt);
423 write_unlock_bh(&peer_pool_lock);
424 /* Remove the entry from unused list if it was there. */
425 unlink_from_unused(p);
426 /* Free preallocated the preallocated node. */
427 kmem_cache_free(peer_cachep, n);
431 /* Called with local BH disabled. */
432 static void peer_check_expire(unsigned long dummy)
434 unsigned long now = jiffies;
437 if (peer_total >= inet_peer_threshold)
438 ttl = inet_peer_minttl;
440 ttl = inet_peer_maxttl
441 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
442 peer_total / inet_peer_threshold * HZ;
443 while (!cleanup_once(ttl)) {
448 /* Trigger the timer after inet_peer_gc_mintime .. inet_peer_gc_maxtime
449 * interval depending on the total number of entries (more entries,
451 if (peer_total >= inet_peer_threshold)
452 peer_periodic_timer.expires = jiffies + inet_peer_gc_mintime;
454 peer_periodic_timer.expires = jiffies
455 + inet_peer_gc_maxtime
456 - (inet_peer_gc_maxtime - inet_peer_gc_mintime) / HZ *
457 peer_total / inet_peer_threshold * HZ;
458 add_timer(&peer_periodic_timer);
461 void inet_putpeer(struct inet_peer *p)
463 spin_lock_bh(&inet_peer_unused_lock);
464 if (atomic_dec_and_test(&p->refcnt)) {
465 p->unused_prevp = inet_peer_unused_tailp;
466 p->unused_next = NULL;
467 *inet_peer_unused_tailp = p;
468 inet_peer_unused_tailp = &p->unused_next;
469 p->dtime = (__u32)jiffies;
471 spin_unlock_bh(&inet_peer_unused_lock);