]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * INETPEER - A storage for permanent information about peers | |
3 | * | |
4 | * This source is covered by the GNU GPL, the same as all kernel sources. | |
5 | * | |
1da177e4 LT |
6 | * Authors: Andrey V. Savochkin <saw@msu.ru> |
7 | */ | |
8 | ||
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> | |
1da177e4 LT |
15 | #include <linux/timer.h> |
16 | #include <linux/time.h> | |
17 | #include <linux/kernel.h> | |
18 | #include <linux/mm.h> | |
19 | #include <linux/net.h> | |
20380731 | 20 | #include <net/ip.h> |
1da177e4 LT |
21 | #include <net/inetpeer.h> |
22 | ||
23 | /* | |
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). | |
35 | * | |
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. | |
46 | * | |
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. | |
52 | * | |
53 | * Serialisation issues. | |
54 | * 1. Nodes may appear in the tree only with the pool write lock held. | |
55 | * 2. Nodes may disappear from the tree only with the pool write 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 | |
d71209de | 62 | * unused: unused node list lock |
1da177e4 LT |
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: idlock | |
68 | */ | |
69 | ||
e18b890b | 70 | static struct kmem_cache *peer_cachep __read_mostly; |
1da177e4 LT |
71 | |
72 | #define node_height(x) x->avl_height | |
d6cc1d64 ED |
73 | |
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, | |
1da177e4 LT |
78 | .avl_height = 0 |
79 | }; | |
d6cc1d64 ED |
80 | |
81 | static struct { | |
82 | struct inet_peer *root; | |
83 | rwlock_t lock; | |
84 | int total; | |
85 | } peers = { | |
86 | .root = peer_avl_empty, | |
87 | .lock = __RW_LOCK_UNLOCKED(peers.lock), | |
88 | .total = 0, | |
89 | }; | |
1da177e4 LT |
90 | #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */ |
91 | ||
1da177e4 | 92 | /* Exported for sysctl_net_ipv4. */ |
243bbcaa | 93 | int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more |
1da177e4 | 94 | * aggressively at this stage */ |
243bbcaa ED |
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; | |
1da177e4 | 99 | |
d6cc1d64 ED |
100 | static struct { |
101 | struct list_head list; | |
102 | spinlock_t lock; | |
103 | } unused_peers = { | |
104 | .list = LIST_HEAD_INIT(unused_peers.list), | |
105 | .lock = __SPIN_LOCK_UNLOCKED(unused_peers.lock), | |
106 | }; | |
1da177e4 LT |
107 | |
108 | static void peer_check_expire(unsigned long dummy); | |
8d06afab | 109 | static DEFINE_TIMER(peer_periodic_timer, peer_check_expire, 0, 0); |
1da177e4 | 110 | |
1da177e4 LT |
111 | |
112 | /* Called from ip_output.c:ip_init */ | |
113 | void __init inet_initpeers(void) | |
114 | { | |
115 | struct sysinfo si; | |
116 | ||
117 | /* Use the straight interface to information about memory. */ | |
118 | si_meminfo(&si); | |
119 | /* The values below were suggested by Alexey Kuznetsov | |
120 | * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values | |
121 | * myself. --SAW | |
122 | */ | |
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 */ | |
129 | ||
130 | peer_cachep = kmem_cache_create("inet_peer_cache", | |
131 | sizeof(struct inet_peer), | |
e5d679f3 | 132 | 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, |
20c2df83 | 133 | NULL); |
1da177e4 | 134 | |
1da177e4 LT |
135 | /* All the timers, started at system startup tend |
136 | to synchronize. Perturb it a bit. | |
137 | */ | |
138 | peer_periodic_timer.expires = jiffies | |
139 | + net_random() % inet_peer_gc_maxtime | |
140 | + inet_peer_gc_maxtime; | |
141 | add_timer(&peer_periodic_timer); | |
142 | } | |
143 | ||
144 | /* Called with or without local BH being disabled. */ | |
145 | static void unlink_from_unused(struct inet_peer *p) | |
146 | { | |
d6cc1d64 ED |
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); | |
151 | } | |
1da177e4 LT |
152 | } |
153 | ||
243bbcaa ED |
154 | /* |
155 | * Called with local BH disabled and the pool lock held. | |
156 | * _stack is known to be NULL or not at compile time, | |
157 | * so compiler will optimize the if (_stack) tests. | |
158 | */ | |
d9319100 | 159 | #define lookup(_daddr, _stack) \ |
1da177e4 LT |
160 | ({ \ |
161 | struct inet_peer *u, **v; \ | |
fc7b9380 | 162 | if (_stack != NULL) { \ |
243bbcaa | 163 | stackptr = _stack; \ |
d6cc1d64 | 164 | *stackptr++ = &peers.root; \ |
243bbcaa | 165 | } \ |
d6cc1d64 | 166 | for (u = peers.root; u != peer_avl_empty; ) { \ |
243bbcaa | 167 | if (_daddr == u->v4daddr) \ |
1da177e4 | 168 | break; \ |
243bbcaa | 169 | if ((__force __u32)_daddr < (__force __u32)u->v4daddr) \ |
1da177e4 LT |
170 | v = &u->avl_left; \ |
171 | else \ | |
172 | v = &u->avl_right; \ | |
fc7b9380 | 173 | if (_stack != NULL) \ |
243bbcaa | 174 | *stackptr++ = v; \ |
1da177e4 LT |
175 | u = *v; \ |
176 | } \ | |
177 | u; \ | |
178 | }) | |
179 | ||
180 | /* Called with local BH disabled and the pool write lock held. */ | |
181 | #define lookup_rightempty(start) \ | |
182 | ({ \ | |
183 | struct inet_peer *u, **v; \ | |
184 | *stackptr++ = &start->avl_left; \ | |
185 | v = &start->avl_left; \ | |
186 | for (u = *v; u->avl_right != peer_avl_empty; ) { \ | |
187 | v = &u->avl_right; \ | |
188 | *stackptr++ = v; \ | |
189 | u = *v; \ | |
190 | } \ | |
191 | u; \ | |
192 | }) | |
193 | ||
194 | /* Called with local BH disabled and the pool write lock held. | |
195 | * Variable names are the proof of operation correctness. | |
196 | * Look into mm/map_avl.c for more detail description of the ideas. */ | |
197 | static void peer_avl_rebalance(struct inet_peer **stack[], | |
198 | struct inet_peer ***stackend) | |
199 | { | |
200 | struct inet_peer **nodep, *node, *l, *r; | |
201 | int lh, rh; | |
202 | ||
203 | while (stackend > stack) { | |
204 | nodep = *--stackend; | |
205 | node = *nodep; | |
206 | l = node->avl_left; | |
207 | r = node->avl_right; | |
208 | lh = node_height(l); | |
209 | rh = node_height(r); | |
210 | if (lh > rh + 1) { /* l: RH+2 */ | |
211 | struct inet_peer *ll, *lr, *lrl, *lrr; | |
212 | int lrh; | |
213 | ll = l->avl_left; | |
214 | lr = l->avl_right; | |
215 | lrh = node_height(lr); | |
216 | if (lrh <= node_height(ll)) { /* ll: RH+1 */ | |
217 | node->avl_left = lr; /* lr: RH or RH+1 */ | |
218 | node->avl_right = r; /* r: RH */ | |
219 | node->avl_height = lrh + 1; /* RH+1 or RH+2 */ | |
220 | l->avl_left = ll; /* ll: RH+1 */ | |
221 | l->avl_right = node; /* node: RH+1 or RH+2 */ | |
222 | l->avl_height = node->avl_height + 1; | |
223 | *nodep = l; | |
224 | } else { /* ll: RH, lr: RH+1 */ | |
225 | lrl = lr->avl_left; /* lrl: RH or RH-1 */ | |
226 | lrr = lr->avl_right; /* lrr: RH or RH-1 */ | |
227 | node->avl_left = lrr; /* lrr: RH or RH-1 */ | |
228 | node->avl_right = r; /* r: RH */ | |
229 | node->avl_height = rh + 1; /* node: RH+1 */ | |
230 | l->avl_left = ll; /* ll: RH */ | |
231 | l->avl_right = lrl; /* lrl: RH or RH-1 */ | |
232 | l->avl_height = rh + 1; /* l: RH+1 */ | |
233 | lr->avl_left = l; /* l: RH+1 */ | |
234 | lr->avl_right = node; /* node: RH+1 */ | |
235 | lr->avl_height = rh + 2; | |
236 | *nodep = lr; | |
237 | } | |
238 | } else if (rh > lh + 1) { /* r: LH+2 */ | |
239 | struct inet_peer *rr, *rl, *rlr, *rll; | |
240 | int rlh; | |
241 | rr = r->avl_right; | |
242 | rl = r->avl_left; | |
243 | rlh = node_height(rl); | |
244 | if (rlh <= node_height(rr)) { /* rr: LH+1 */ | |
245 | node->avl_right = rl; /* rl: LH or LH+1 */ | |
246 | node->avl_left = l; /* l: LH */ | |
247 | node->avl_height = rlh + 1; /* LH+1 or LH+2 */ | |
248 | r->avl_right = rr; /* rr: LH+1 */ | |
249 | r->avl_left = node; /* node: LH+1 or LH+2 */ | |
250 | r->avl_height = node->avl_height + 1; | |
251 | *nodep = r; | |
252 | } else { /* rr: RH, rl: RH+1 */ | |
253 | rlr = rl->avl_right; /* rlr: LH or LH-1 */ | |
254 | rll = rl->avl_left; /* rll: LH or LH-1 */ | |
255 | node->avl_right = rll; /* rll: LH or LH-1 */ | |
256 | node->avl_left = l; /* l: LH */ | |
257 | node->avl_height = lh + 1; /* node: LH+1 */ | |
258 | r->avl_right = rr; /* rr: LH */ | |
259 | r->avl_left = rlr; /* rlr: LH or LH-1 */ | |
260 | r->avl_height = lh + 1; /* r: LH+1 */ | |
261 | rl->avl_right = r; /* r: LH+1 */ | |
262 | rl->avl_left = node; /* node: LH+1 */ | |
263 | rl->avl_height = lh + 2; | |
264 | *nodep = rl; | |
265 | } | |
266 | } else { | |
267 | node->avl_height = (lh > rh ? lh : rh) + 1; | |
268 | } | |
269 | } | |
270 | } | |
271 | ||
272 | /* Called with local BH disabled and the pool write lock held. */ | |
273 | #define link_to_pool(n) \ | |
274 | do { \ | |
275 | n->avl_height = 1; \ | |
276 | n->avl_left = peer_avl_empty; \ | |
277 | n->avl_right = peer_avl_empty; \ | |
278 | **--stackptr = n; \ | |
279 | peer_avl_rebalance(stack, stackptr); \ | |
d6cc1d64 | 280 | } while (0) |
1da177e4 LT |
281 | |
282 | /* May be called with local BH enabled. */ | |
283 | static void unlink_from_pool(struct inet_peer *p) | |
284 | { | |
285 | int do_free; | |
286 | ||
287 | do_free = 0; | |
288 | ||
d6cc1d64 | 289 | write_lock_bh(&peers.lock); |
1da177e4 LT |
290 | /* Check the reference counter. It was artificially incremented by 1 |
291 | * in cleanup() function to prevent sudden disappearing. If the | |
292 | * reference count is still 1 then the node is referenced only as `p' | |
293 | * here and from the pool. So under the exclusive pool lock it's safe | |
294 | * to remove the node and free it later. */ | |
295 | if (atomic_read(&p->refcnt) == 1) { | |
296 | struct inet_peer **stack[PEER_MAXDEPTH]; | |
297 | struct inet_peer ***stackptr, ***delp; | |
243bbcaa | 298 | if (lookup(p->v4daddr, stack) != p) |
1da177e4 LT |
299 | BUG(); |
300 | delp = stackptr - 1; /* *delp[0] == p */ | |
301 | if (p->avl_left == peer_avl_empty) { | |
302 | *delp[0] = p->avl_right; | |
303 | --stackptr; | |
304 | } else { | |
305 | /* look for a node to insert instead of p */ | |
306 | struct inet_peer *t; | |
307 | t = lookup_rightempty(p); | |
09a62660 | 308 | BUG_ON(*stackptr[-1] != t); |
1da177e4 LT |
309 | **--stackptr = t->avl_left; |
310 | /* t is removed, t->v4daddr > x->v4daddr for any | |
311 | * x in p->avl_left subtree. | |
312 | * Put t in the old place of p. */ | |
313 | *delp[0] = t; | |
314 | t->avl_left = p->avl_left; | |
315 | t->avl_right = p->avl_right; | |
316 | t->avl_height = p->avl_height; | |
09a62660 | 317 | BUG_ON(delp[1] != &p->avl_left); |
1da177e4 LT |
318 | delp[1] = &t->avl_left; /* was &p->avl_left */ |
319 | } | |
320 | peer_avl_rebalance(stack, stackptr); | |
d6cc1d64 | 321 | peers.total--; |
1da177e4 LT |
322 | do_free = 1; |
323 | } | |
d6cc1d64 | 324 | write_unlock_bh(&peers.lock); |
1da177e4 LT |
325 | |
326 | if (do_free) | |
327 | kmem_cache_free(peer_cachep, p); | |
328 | else | |
329 | /* The node is used again. Decrease the reference counter | |
330 | * back. The loop "cleanup -> unlink_from_unused | |
331 | * -> unlink_from_pool -> putpeer -> link_to_unused | |
332 | * -> cleanup (for the same node)" | |
333 | * doesn't really exist because the entry will have a | |
334 | * recent deletion time and will not be cleaned again soon. */ | |
335 | inet_putpeer(p); | |
336 | } | |
337 | ||
338 | /* May be called with local BH enabled. */ | |
339 | static int cleanup_once(unsigned long ttl) | |
340 | { | |
d71209de | 341 | struct inet_peer *p = NULL; |
1da177e4 LT |
342 | |
343 | /* Remove the first entry from the list of unused nodes. */ | |
d6cc1d64 ED |
344 | spin_lock_bh(&unused_peers.lock); |
345 | if (!list_empty(&unused_peers.list)) { | |
d71209de PE |
346 | __u32 delta; |
347 | ||
d6cc1d64 | 348 | p = list_first_entry(&unused_peers.list, struct inet_peer, unused); |
d71209de PE |
349 | delta = (__u32)jiffies - p->dtime; |
350 | ||
4663afe2 | 351 | if (delta < ttl) { |
1da177e4 | 352 | /* Do not prune fresh entries. */ |
d6cc1d64 | 353 | spin_unlock_bh(&unused_peers.lock); |
1da177e4 LT |
354 | return -1; |
355 | } | |
d71209de PE |
356 | |
357 | list_del_init(&p->unused); | |
358 | ||
1da177e4 LT |
359 | /* Grab an extra reference to prevent node disappearing |
360 | * before unlink_from_pool() call. */ | |
361 | atomic_inc(&p->refcnt); | |
362 | } | |
d6cc1d64 | 363 | spin_unlock_bh(&unused_peers.lock); |
1da177e4 LT |
364 | |
365 | if (p == NULL) | |
366 | /* It means that the total number of USED entries has | |
367 | * grown over inet_peer_threshold. It shouldn't really | |
368 | * happen because of entry limits in route cache. */ | |
369 | return -1; | |
370 | ||
371 | unlink_from_pool(p); | |
372 | return 0; | |
373 | } | |
374 | ||
375 | /* Called with or without local BH being disabled. */ | |
53576d9b | 376 | struct inet_peer *inet_getpeer(__be32 daddr, int create) |
1da177e4 LT |
377 | { |
378 | struct inet_peer *p, *n; | |
379 | struct inet_peer **stack[PEER_MAXDEPTH], ***stackptr; | |
380 | ||
381 | /* Look up for the address quickly. */ | |
d6cc1d64 | 382 | read_lock_bh(&peers.lock); |
243bbcaa | 383 | p = lookup(daddr, NULL); |
1da177e4 LT |
384 | if (p != peer_avl_empty) |
385 | atomic_inc(&p->refcnt); | |
d6cc1d64 | 386 | read_unlock_bh(&peers.lock); |
1da177e4 LT |
387 | |
388 | if (p != peer_avl_empty) { | |
389 | /* The existing node has been found. */ | |
390 | /* Remove the entry from unused list if it was there. */ | |
391 | unlink_from_unused(p); | |
392 | return p; | |
393 | } | |
394 | ||
395 | if (!create) | |
396 | return NULL; | |
397 | ||
398 | /* Allocate the space outside the locked region. */ | |
399 | n = kmem_cache_alloc(peer_cachep, GFP_ATOMIC); | |
400 | if (n == NULL) | |
401 | return NULL; | |
402 | n->v4daddr = daddr; | |
403 | atomic_set(&n->refcnt, 1); | |
89cee8b1 | 404 | atomic_set(&n->rid, 0); |
2c1409a0 | 405 | atomic_set(&n->ip_id_count, secure_ip_id(daddr)); |
1da177e4 LT |
406 | n->tcp_ts_stamp = 0; |
407 | ||
d6cc1d64 | 408 | write_lock_bh(&peers.lock); |
1da177e4 | 409 | /* Check if an entry has suddenly appeared. */ |
243bbcaa | 410 | p = lookup(daddr, stack); |
1da177e4 LT |
411 | if (p != peer_avl_empty) |
412 | goto out_free; | |
413 | ||
414 | /* Link the node. */ | |
415 | link_to_pool(n); | |
d71209de | 416 | INIT_LIST_HEAD(&n->unused); |
d6cc1d64 ED |
417 | peers.total++; |
418 | write_unlock_bh(&peers.lock); | |
1da177e4 | 419 | |
d6cc1d64 | 420 | if (peers.total >= inet_peer_threshold) |
1da177e4 LT |
421 | /* Remove one less-recently-used entry. */ |
422 | cleanup_once(0); | |
423 | ||
424 | return n; | |
425 | ||
426 | out_free: | |
427 | /* The appropriate node is already in the pool. */ | |
428 | atomic_inc(&p->refcnt); | |
d6cc1d64 | 429 | write_unlock_bh(&peers.lock); |
1da177e4 LT |
430 | /* Remove the entry from unused list if it was there. */ |
431 | unlink_from_unused(p); | |
432 | /* Free preallocated the preallocated node. */ | |
433 | kmem_cache_free(peer_cachep, n); | |
434 | return p; | |
435 | } | |
436 | ||
437 | /* Called with local BH disabled. */ | |
438 | static void peer_check_expire(unsigned long dummy) | |
439 | { | |
4663afe2 | 440 | unsigned long now = jiffies; |
1da177e4 LT |
441 | int ttl; |
442 | ||
d6cc1d64 | 443 | if (peers.total >= inet_peer_threshold) |
1da177e4 LT |
444 | ttl = inet_peer_minttl; |
445 | else | |
446 | ttl = inet_peer_maxttl | |
447 | - (inet_peer_maxttl - inet_peer_minttl) / HZ * | |
d6cc1d64 | 448 | peers.total / inet_peer_threshold * HZ; |
4663afe2 ED |
449 | while (!cleanup_once(ttl)) { |
450 | if (jiffies != now) | |
451 | break; | |
452 | } | |
1da177e4 LT |
453 | |
454 | /* Trigger the timer after inet_peer_gc_mintime .. inet_peer_gc_maxtime | |
455 | * interval depending on the total number of entries (more entries, | |
456 | * less interval). */ | |
d6cc1d64 | 457 | if (peers.total >= inet_peer_threshold) |
1344a416 DJ |
458 | peer_periodic_timer.expires = jiffies + inet_peer_gc_mintime; |
459 | else | |
460 | peer_periodic_timer.expires = jiffies | |
461 | + inet_peer_gc_maxtime | |
462 | - (inet_peer_gc_maxtime - inet_peer_gc_mintime) / HZ * | |
d6cc1d64 | 463 | peers.total / inet_peer_threshold * HZ; |
1da177e4 LT |
464 | add_timer(&peer_periodic_timer); |
465 | } | |
4663afe2 ED |
466 | |
467 | void inet_putpeer(struct inet_peer *p) | |
468 | { | |
d6cc1d64 ED |
469 | local_bh_disable(); |
470 | ||
471 | if (atomic_dec_and_lock(&p->refcnt, &unused_peers.lock)) { | |
472 | list_add_tail(&p->unused, &unused_peers.list); | |
4663afe2 | 473 | p->dtime = (__u32)jiffies; |
d6cc1d64 | 474 | spin_unlock(&unused_peers.lock); |
4663afe2 | 475 | } |
d6cc1d64 ED |
476 | |
477 | local_bh_enable(); | |
4663afe2 | 478 | } |