File: /usr/src/linux/net/ipv4/inetpeer.c
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 *
6 * Version: $Id: inetpeer.c,v 1.6 2001/06/21 20:30:14 davem Exp $
7 *
8 * Authors: Andrey V. Savochkin <saw@msu.ru>
9 */
10
11 #include <linux/types.h>
12 #include <linux/slab.h>
13 #include <linux/interrupt.h>
14 #include <linux/spinlock.h>
15 #include <linux/random.h>
16 #include <linux/sched.h>
17 #include <linux/timer.h>
18 #include <linux/time.h>
19 #include <linux/kernel.h>
20 #include <linux/mm.h>
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
62 * unused_next, unused_prevp: 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: idlock
68 */
69
70 spinlock_t inet_peer_idlock = SPIN_LOCK_UNLOCKED;
71
72 static kmem_cache_t *peer_cachep;
73
74 #define node_height(x) x->avl_height
75 static struct inet_peer peer_fake_node = {
76 avl_left : &peer_fake_node,
77 avl_right : &peer_fake_node,
78 avl_height : 0
79 };
80 #define peer_avl_empty (&peer_fake_node)
81 static struct inet_peer *peer_root = peer_avl_empty;
82 static rwlock_t peer_pool_lock = RW_LOCK_UNLOCKED;
83 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
84
85 static volatile int peer_total;
86 int inet_peer_threshold = 65536 + 128; /* start to throw entries more
87 * aggressively at this stage */
88 int inet_peer_minttl = 120 * HZ; /* TTL under high load: 120 sec */
89 int inet_peer_maxttl = 10 * 60 * HZ; /* usual time to live: 10 min */
90 struct inet_peer *inet_peer_unused_head,
91 **inet_peer_unused_tailp = &inet_peer_unused_head;
92 spinlock_t inet_peer_unused_lock = SPIN_LOCK_UNLOCKED;
93 #define PEER_MAX_CLEANUP_WORK 30
94
95 static void peer_check_expire(unsigned long dummy);
96 static struct timer_list peer_periodic_timer =
97 { { NULL, NULL }, 0, 0, &peer_check_expire };
98 int inet_peer_gc_mintime = 10 * HZ,
99 inet_peer_gc_maxtime = 120 * HZ;
100
101
102 void __init inet_initpeers(void)
103 {
104 struct sysinfo si;
105
106 /* Use the straight interface to information about memory. */
107 si_meminfo(&si);
108 /* The values below were suggested by Alexey Kuznetsov
109 * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values
110 * myself. --SAW
111 */
112 if (si.totalram <= (32768*1024)/PAGE_SIZE)
113 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
114 if (si.totalram <= (16384*1024)/PAGE_SIZE)
115 inet_peer_threshold >>= 1; /* about 512KB */
116 if (si.totalram <= (8192*1024)/PAGE_SIZE)
117 inet_peer_threshold >>= 2; /* about 128KB */
118
119 peer_cachep = kmem_cache_create("inet_peer_cache",
120 sizeof(struct inet_peer),
121 0, SLAB_HWCACHE_ALIGN,
122 NULL, NULL);
123
124 /* All the timers, started at system startup tend
125 to synchronize. Perturb it a bit.
126 */
127 peer_periodic_timer.expires = jiffies
128 + net_random() % inet_peer_gc_maxtime
129 + inet_peer_gc_maxtime;
130 add_timer(&peer_periodic_timer);
131 }
132
133 /* Called with or without local BH being disabled. */
134 static void unlink_from_unused(struct inet_peer *p)
135 {
136 spin_lock_bh(&inet_peer_unused_lock);
137 if (p->unused_prevp != NULL) {
138 /* On unused list. */
139 *p->unused_prevp = p->unused_next;
140 if (p->unused_next != NULL)
141 p->unused_next->unused_prevp = p->unused_prevp;
142 else
143 inet_peer_unused_tailp = p->unused_prevp;
144 p->unused_prevp = NULL; /* mark it as removed */
145 }
146 spin_unlock_bh(&inet_peer_unused_lock);
147 }
148
149 /* Called with local BH disabled and the pool lock held. */
150 #define lookup(daddr) \
151 ({ \
152 struct inet_peer *u, **v; \
153 stackptr = stack; \
154 *stackptr++ = &peer_root; \
155 for (u = peer_root; u != peer_avl_empty; ) { \
156 if (daddr == u->v4daddr) \
157 break; \
158 if (daddr < u->v4daddr) \
159 v = &u->avl_left; \
160 else \
161 v = &u->avl_right; \
162 *stackptr++ = v; \
163 u = *v; \
164 } \
165 u; \
166 })
167
168 /* Called with local BH disabled and the pool write lock held. */
169 #define lookup_rightempty(start) \
170 ({ \
171 struct inet_peer *u, **v; \
172 *stackptr++ = &start->avl_left; \
173 v = &start->avl_left; \
174 for (u = *v; u->avl_right != peer_avl_empty; ) { \
175 v = &u->avl_right; \
176 *stackptr++ = v; \
177 u = *v; \
178 } \
179 u; \
180 })
181
182 /* Called with local BH disabled and the pool write lock held.
183 * Variable names are the proof of operation correctness.
184 * Look into mm/map_avl.c for more detail description of the ideas. */
185 static void peer_avl_rebalance(struct inet_peer **stack[],
186 struct inet_peer ***stackend)
187 {
188 struct inet_peer **nodep, *node, *l, *r;
189 int lh, rh;
190
191 while (stackend > stack) {
192 nodep = *--stackend;
193 node = *nodep;
194 l = node->avl_left;
195 r = node->avl_right;
196 lh = node_height(l);
197 rh = node_height(r);
198 if (lh > rh + 1) { /* l: RH+2 */
199 struct inet_peer *ll, *lr, *lrl, *lrr;
200 int lrh;
201 ll = l->avl_left;
202 lr = l->avl_right;
203 lrh = node_height(lr);
204 if (lrh <= node_height(ll)) { /* ll: RH+1 */
205 node->avl_left = lr; /* lr: RH or RH+1 */
206 node->avl_right = r; /* r: RH */
207 node->avl_height = lrh + 1; /* RH+1 or RH+2 */
208 l->avl_left = ll; /* ll: RH+1 */
209 l->avl_right = node; /* node: RH+1 or RH+2 */
210 l->avl_height = node->avl_height + 1;
211 *nodep = l;
212 } else { /* ll: RH, lr: RH+1 */
213 lrl = lr->avl_left; /* lrl: RH or RH-1 */
214 lrr = lr->avl_right; /* lrr: RH or RH-1 */
215 node->avl_left = lrr; /* lrr: RH or RH-1 */
216 node->avl_right = r; /* r: RH */
217 node->avl_height = rh + 1; /* node: RH+1 */
218 l->avl_left = ll; /* ll: RH */
219 l->avl_right = lrl; /* lrl: RH or RH-1 */
220 l->avl_height = rh + 1; /* l: RH+1 */
221 lr->avl_left = l; /* l: RH+1 */
222 lr->avl_right = node; /* node: RH+1 */
223 lr->avl_height = rh + 2;
224 *nodep = lr;
225 }
226 } else if (rh > lh + 1) { /* r: LH+2 */
227 struct inet_peer *rr, *rl, *rlr, *rll;
228 int rlh;
229 rr = r->avl_right;
230 rl = r->avl_left;
231 rlh = node_height(rl);
232 if (rlh <= node_height(rr)) { /* rr: LH+1 */
233 node->avl_right = rl; /* rl: LH or LH+1 */
234 node->avl_left = l; /* l: LH */
235 node->avl_height = rlh + 1; /* LH+1 or LH+2 */
236 r->avl_right = rr; /* rr: LH+1 */
237 r->avl_left = node; /* node: LH+1 or LH+2 */
238 r->avl_height = node->avl_height + 1;
239 *nodep = r;
240 } else { /* rr: RH, rl: RH+1 */
241 rlr = rl->avl_right; /* rlr: LH or LH-1 */
242 rll = rl->avl_left; /* rll: LH or LH-1 */
243 node->avl_right = rll; /* rll: LH or LH-1 */
244 node->avl_left = l; /* l: LH */
245 node->avl_height = lh + 1; /* node: LH+1 */
246 r->avl_right = rr; /* rr: LH */
247 r->avl_left = rlr; /* rlr: LH or LH-1 */
248 r->avl_height = lh + 1; /* r: LH+1 */
249 rl->avl_right = r; /* r: LH+1 */
250 rl->avl_left = node; /* node: LH+1 */
251 rl->avl_height = lh + 2;
252 *nodep = rl;
253 }
254 } else {
255 node->avl_height = (lh > rh ? lh : rh) + 1;
256 }
257 }
258 }
259
260 /* Called with local BH disabled and the pool write lock held. */
261 #define link_to_pool(n) \
262 do { \
263 n->avl_height = 1; \
264 n->avl_left = peer_avl_empty; \
265 n->avl_right = peer_avl_empty; \
266 **--stackptr = n; \
267 peer_avl_rebalance(stack, stackptr); \
268 } while(0)
269
270 /* May be called with local BH enabled. */
271 static void unlink_from_pool(struct inet_peer *p)
272 {
273 int do_free;
274
275 do_free = 0;
276
277 write_lock_bh(&peer_pool_lock);
278 /* Check the reference counter. It was artificially incremented by 1
279 * in cleanup() function to prevent sudden disappearing. If the
280 * reference count is still 1 then the node is referenced only as `p'
281 * here and from the pool. So under the exclusive pool lock it's safe
282 * to remove the node and free it later. */
283 if (atomic_read(&p->refcnt) == 1) {
284 struct inet_peer **stack[PEER_MAXDEPTH];
285 struct inet_peer ***stackptr, ***delp;
286 if (lookup(p->v4daddr) != p)
287 BUG();
288 delp = stackptr - 1; /* *delp[0] == p */
289 if (p->avl_left == peer_avl_empty) {
290 *delp[0] = p->avl_right;
291 --stackptr;
292 } else {
293 /* look for a node to insert instead of p */
294 struct inet_peer *t;
295 t = lookup_rightempty(p);
296 if (*stackptr[-1] != t)
297 BUG();
298 **--stackptr = t->avl_left;
299 /* t is removed, t->v4daddr > x->v4daddr for any
300 * x in p->avl_left subtree.
301 * Put t in the old place of p. */
302 *delp[0] = t;
303 t->avl_left = p->avl_left;
304 t->avl_right = p->avl_right;
305 t->avl_height = p->avl_height;
306 if (delp[1] != &p->avl_left)
307 BUG();
308 delp[1] = &t->avl_left; /* was &p->avl_left */
309 }
310 peer_avl_rebalance(stack, stackptr);
311 peer_total--;
312 do_free = 1;
313 }
314 write_unlock_bh(&peer_pool_lock);
315
316 if (do_free)
317 kmem_cache_free(peer_cachep, p);
318 else
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. */
325 inet_putpeer(p);
326 }
327
328 /* May be called with local BH enabled. */
329 static int cleanup_once(unsigned long ttl)
330 {
331 struct inet_peer *p;
332
333 /* Remove the first entry from the list of unused nodes. */
334 spin_lock_bh(&inet_peer_unused_lock);
335 p = inet_peer_unused_head;
336 if (p != NULL) {
337 if (time_after(p->dtime + ttl, jiffies)) {
338 /* Do not prune fresh entries. */
339 spin_unlock_bh(&inet_peer_unused_lock);
340 return -1;
341 }
342 inet_peer_unused_head = p->unused_next;
343 if (p->unused_next != NULL)
344 p->unused_next->unused_prevp = p->unused_prevp;
345 else
346 inet_peer_unused_tailp = p->unused_prevp;
347 p->unused_prevp = NULL; /* mark as not on the list */
348 /* Grab an extra reference to prevent node disappearing
349 * before unlink_from_pool() call. */
350 atomic_inc(&p->refcnt);
351 }
352 spin_unlock_bh(&inet_peer_unused_lock);
353
354 if (p == NULL)
355 /* It means that the total number of USED entries has
356 * grown over inet_peer_threshold. It shouldn't really
357 * happen because of entry limits in route cache. */
358 return -1;
359
360 unlink_from_pool(p);
361 return 0;
362 }
363
364 /* Called with or without local BH being disabled. */
365 struct inet_peer *inet_getpeer(__u32 daddr, int create)
366 {
367 struct inet_peer *p, *n;
368 struct inet_peer **stack[PEER_MAXDEPTH], ***stackptr;
369
370 /* Look up for the address quickly. */
371 read_lock_bh(&peer_pool_lock);
372 p = lookup(daddr);
373 if (p != peer_avl_empty)
374 atomic_inc(&p->refcnt);
375 read_unlock_bh(&peer_pool_lock);
376
377 if (p != peer_avl_empty) {
378 /* The existing node has been found. */
379 /* Remove the entry from unused list if it was there. */
380 unlink_from_unused(p);
381 return p;
382 }
383
384 if (!create)
385 return NULL;
386
387 /* Allocate the space outside the locked region. */
388 n = kmem_cache_alloc(peer_cachep, GFP_ATOMIC);
389 if (n == NULL)
390 return NULL;
391 n->v4daddr = daddr;
392 atomic_set(&n->refcnt, 1);
393 n->ip_id_count = secure_ip_id(daddr);
394 n->tcp_ts_stamp = 0;
395
396 write_lock_bh(&peer_pool_lock);
397 /* Check if an entry has suddenly appeared. */
398 p = lookup(daddr);
399 if (p != peer_avl_empty)
400 goto out_free;
401
402 /* Link the node. */
403 link_to_pool(n);
404 n->unused_prevp = NULL; /* not on the list */
405 peer_total++;
406 write_unlock_bh(&peer_pool_lock);
407
408 if (peer_total >= inet_peer_threshold)
409 /* Remove one less-recently-used entry. */
410 cleanup_once(0);
411
412 return n;
413
414 out_free:
415 /* The appropriate node is already in the pool. */
416 atomic_inc(&p->refcnt);
417 write_unlock_bh(&peer_pool_lock);
418 /* Remove the entry from unused list if it was there. */
419 unlink_from_unused(p);
420 /* Free preallocated the preallocated node. */
421 kmem_cache_free(peer_cachep, n);
422 return p;
423 }
424
425 /* Called with local BH disabled. */
426 static void peer_check_expire(unsigned long dummy)
427 {
428 int i;
429 int ttl;
430
431 if (peer_total >= inet_peer_threshold)
432 ttl = inet_peer_minttl;
433 else
434 ttl = inet_peer_maxttl
435 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
436 peer_total / inet_peer_threshold * HZ;
437 for (i = 0; i < PEER_MAX_CLEANUP_WORK && !cleanup_once(ttl); i++);
438
439 /* Trigger the timer after inet_peer_gc_mintime .. inet_peer_gc_maxtime
440 * interval depending on the total number of entries (more entries,
441 * less interval). */
442 peer_periodic_timer.expires = jiffies
443 + inet_peer_gc_maxtime
444 - (inet_peer_gc_maxtime - inet_peer_gc_mintime) / HZ *
445 peer_total / inet_peer_threshold * HZ;
446 add_timer(&peer_periodic_timer);
447 }
448