xref: /openbmc/linux/net/ipv4/inetpeer.c (revision 827634ad)
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  *  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>
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>
20 #include <linux/workqueue.h>
21 #include <net/ip.h>
22 #include <net/inetpeer.h>
23 #include <net/secure_seq.h>
24 
25 /*
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  *
30  *  Nodes are removed only when reference counter goes to 0.
31  *  When it's happened the node may be removed when a sufficient amount of
32  *  time has been passed since its last use.  The less-recently-used entry can
33  *  also be removed if the pool is overloaded i.e. if the total amount of
34  *  entries is greater-or-equal than the threshold.
35  *
36  *  Node pool is organised as an AVL tree.
37  *  Such an implementation has been chosen not just for fun.  It's a way to
38  *  prevent easy and efficient DoS attacks by creating hash collisions.  A huge
39  *  amount of long living nodes in a single hash slot would significantly delay
40  *  lookups performed with disabled BHs.
41  *
42  *  Serialisation issues.
43  *  1.  Nodes may appear in the tree only with the pool lock held.
44  *  2.  Nodes may disappear from the tree only with the pool lock held
45  *      AND reference count being 0.
46  *  3.  Global variable peer_total is modified under the pool lock.
47  *  4.  struct inet_peer fields modification:
48  *		avl_left, avl_right, avl_parent, avl_height: pool lock
49  *		refcnt: atomically against modifications on other CPU;
50  *		   usually under some other lock to prevent node disappearing
51  *		daddr: unchangeable
52  */
53 
54 static struct kmem_cache *peer_cachep __read_mostly;
55 
56 static LIST_HEAD(gc_list);
57 static const int gc_delay = 60 * HZ;
58 static struct delayed_work gc_work;
59 static DEFINE_SPINLOCK(gc_lock);
60 
61 #define node_height(x) x->avl_height
62 
63 #define peer_avl_empty ((struct inet_peer *)&peer_fake_node)
64 #define peer_avl_empty_rcu ((struct inet_peer __rcu __force *)&peer_fake_node)
65 static const struct inet_peer peer_fake_node = {
66 	.avl_left	= peer_avl_empty_rcu,
67 	.avl_right	= peer_avl_empty_rcu,
68 	.avl_height	= 0
69 };
70 
71 void inet_peer_base_init(struct inet_peer_base *bp)
72 {
73 	bp->root = peer_avl_empty_rcu;
74 	seqlock_init(&bp->lock);
75 	bp->total = 0;
76 }
77 EXPORT_SYMBOL_GPL(inet_peer_base_init);
78 
79 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
80 
81 /* Exported for sysctl_net_ipv4.  */
82 int inet_peer_threshold __read_mostly = 65536 + 128;	/* start to throw entries more
83 					 * aggressively at this stage */
84 int inet_peer_minttl __read_mostly = 120 * HZ;	/* TTL under high load: 120 sec */
85 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ;	/* usual time to live: 10 min */
86 
87 static void inetpeer_gc_worker(struct work_struct *work)
88 {
89 	struct inet_peer *p, *n, *c;
90 	struct list_head list;
91 
92 	spin_lock_bh(&gc_lock);
93 	list_replace_init(&gc_list, &list);
94 	spin_unlock_bh(&gc_lock);
95 
96 	if (list_empty(&list))
97 		return;
98 
99 	list_for_each_entry_safe(p, n, &list, gc_list) {
100 
101 		if (need_resched())
102 			cond_resched();
103 
104 		c = rcu_dereference_protected(p->avl_left, 1);
105 		if (c != peer_avl_empty) {
106 			list_add_tail(&c->gc_list, &list);
107 			p->avl_left = peer_avl_empty_rcu;
108 		}
109 
110 		c = rcu_dereference_protected(p->avl_right, 1);
111 		if (c != peer_avl_empty) {
112 			list_add_tail(&c->gc_list, &list);
113 			p->avl_right = peer_avl_empty_rcu;
114 		}
115 
116 		n = list_entry(p->gc_list.next, struct inet_peer, gc_list);
117 
118 		if (!atomic_read(&p->refcnt)) {
119 			list_del(&p->gc_list);
120 			kmem_cache_free(peer_cachep, p);
121 		}
122 	}
123 
124 	if (list_empty(&list))
125 		return;
126 
127 	spin_lock_bh(&gc_lock);
128 	list_splice(&list, &gc_list);
129 	spin_unlock_bh(&gc_lock);
130 
131 	schedule_delayed_work(&gc_work, gc_delay);
132 }
133 
134 /* Called from ip_output.c:ip_init  */
135 void __init inet_initpeers(void)
136 {
137 	struct sysinfo si;
138 
139 	/* Use the straight interface to information about memory. */
140 	si_meminfo(&si);
141 	/* The values below were suggested by Alexey Kuznetsov
142 	 * <kuznet@ms2.inr.ac.ru>.  I don't have any opinion about the values
143 	 * myself.  --SAW
144 	 */
145 	if (si.totalram <= (32768*1024)/PAGE_SIZE)
146 		inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
147 	if (si.totalram <= (16384*1024)/PAGE_SIZE)
148 		inet_peer_threshold >>= 1; /* about 512KB */
149 	if (si.totalram <= (8192*1024)/PAGE_SIZE)
150 		inet_peer_threshold >>= 2; /* about 128KB */
151 
152 	peer_cachep = kmem_cache_create("inet_peer_cache",
153 			sizeof(struct inet_peer),
154 			0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
155 			NULL);
156 
157 	INIT_DEFERRABLE_WORK(&gc_work, inetpeer_gc_worker);
158 }
159 
160 static int addr_compare(const struct inetpeer_addr *a,
161 			const struct inetpeer_addr *b)
162 {
163 	int i, n = (a->family == AF_INET ? 1 : 4);
164 
165 	for (i = 0; i < n; i++) {
166 		if (a->addr.a6[i] == b->addr.a6[i])
167 			continue;
168 		if ((__force u32)a->addr.a6[i] < (__force u32)b->addr.a6[i])
169 			return -1;
170 		return 1;
171 	}
172 
173 	return 0;
174 }
175 
176 #define rcu_deref_locked(X, BASE)				\
177 	rcu_dereference_protected(X, lockdep_is_held(&(BASE)->lock.lock))
178 
179 /*
180  * Called with local BH disabled and the pool lock held.
181  */
182 #define lookup(_daddr, _stack, _base)				\
183 ({								\
184 	struct inet_peer *u;					\
185 	struct inet_peer __rcu **v;				\
186 								\
187 	stackptr = _stack;					\
188 	*stackptr++ = &_base->root;				\
189 	for (u = rcu_deref_locked(_base->root, _base);		\
190 	     u != peer_avl_empty;) {				\
191 		int cmp = addr_compare(_daddr, &u->daddr);	\
192 		if (cmp == 0)					\
193 			break;					\
194 		if (cmp == -1)					\
195 			v = &u->avl_left;			\
196 		else						\
197 			v = &u->avl_right;			\
198 		*stackptr++ = v;				\
199 		u = rcu_deref_locked(*v, _base);		\
200 	}							\
201 	u;							\
202 })
203 
204 /*
205  * Called with rcu_read_lock()
206  * Because we hold no lock against a writer, its quite possible we fall
207  * in an endless loop.
208  * But every pointer we follow is guaranteed to be valid thanks to RCU.
209  * We exit from this function if number of links exceeds PEER_MAXDEPTH
210  */
211 static struct inet_peer *lookup_rcu(const struct inetpeer_addr *daddr,
212 				    struct inet_peer_base *base)
213 {
214 	struct inet_peer *u = rcu_dereference(base->root);
215 	int count = 0;
216 
217 	while (u != peer_avl_empty) {
218 		int cmp = addr_compare(daddr, &u->daddr);
219 		if (cmp == 0) {
220 			/* Before taking a reference, check if this entry was
221 			 * deleted (refcnt=-1)
222 			 */
223 			if (!atomic_add_unless(&u->refcnt, 1, -1))
224 				u = NULL;
225 			return u;
226 		}
227 		if (cmp == -1)
228 			u = rcu_dereference(u->avl_left);
229 		else
230 			u = rcu_dereference(u->avl_right);
231 		if (unlikely(++count == PEER_MAXDEPTH))
232 			break;
233 	}
234 	return NULL;
235 }
236 
237 /* Called with local BH disabled and the pool lock held. */
238 #define lookup_rightempty(start, base)				\
239 ({								\
240 	struct inet_peer *u;					\
241 	struct inet_peer __rcu **v;				\
242 	*stackptr++ = &start->avl_left;				\
243 	v = &start->avl_left;					\
244 	for (u = rcu_deref_locked(*v, base);			\
245 	     u->avl_right != peer_avl_empty_rcu;) {		\
246 		v = &u->avl_right;				\
247 		*stackptr++ = v;				\
248 		u = rcu_deref_locked(*v, base);			\
249 	}							\
250 	u;							\
251 })
252 
253 /* Called with local BH disabled and the pool lock held.
254  * Variable names are the proof of operation correctness.
255  * Look into mm/map_avl.c for more detail description of the ideas.
256  */
257 static void peer_avl_rebalance(struct inet_peer __rcu **stack[],
258 			       struct inet_peer __rcu ***stackend,
259 			       struct inet_peer_base *base)
260 {
261 	struct inet_peer __rcu **nodep;
262 	struct inet_peer *node, *l, *r;
263 	int lh, rh;
264 
265 	while (stackend > stack) {
266 		nodep = *--stackend;
267 		node = rcu_deref_locked(*nodep, base);
268 		l = rcu_deref_locked(node->avl_left, base);
269 		r = rcu_deref_locked(node->avl_right, base);
270 		lh = node_height(l);
271 		rh = node_height(r);
272 		if (lh > rh + 1) { /* l: RH+2 */
273 			struct inet_peer *ll, *lr, *lrl, *lrr;
274 			int lrh;
275 			ll = rcu_deref_locked(l->avl_left, base);
276 			lr = rcu_deref_locked(l->avl_right, base);
277 			lrh = node_height(lr);
278 			if (lrh <= node_height(ll)) {	/* ll: RH+1 */
279 				RCU_INIT_POINTER(node->avl_left, lr);	/* lr: RH or RH+1 */
280 				RCU_INIT_POINTER(node->avl_right, r);	/* r: RH */
281 				node->avl_height = lrh + 1; /* RH+1 or RH+2 */
282 				RCU_INIT_POINTER(l->avl_left, ll);       /* ll: RH+1 */
283 				RCU_INIT_POINTER(l->avl_right, node);	/* node: RH+1 or RH+2 */
284 				l->avl_height = node->avl_height + 1;
285 				RCU_INIT_POINTER(*nodep, l);
286 			} else { /* ll: RH, lr: RH+1 */
287 				lrl = rcu_deref_locked(lr->avl_left, base);/* lrl: RH or RH-1 */
288 				lrr = rcu_deref_locked(lr->avl_right, base);/* lrr: RH or RH-1 */
289 				RCU_INIT_POINTER(node->avl_left, lrr);	/* lrr: RH or RH-1 */
290 				RCU_INIT_POINTER(node->avl_right, r);	/* r: RH */
291 				node->avl_height = rh + 1; /* node: RH+1 */
292 				RCU_INIT_POINTER(l->avl_left, ll);	/* ll: RH */
293 				RCU_INIT_POINTER(l->avl_right, lrl);	/* lrl: RH or RH-1 */
294 				l->avl_height = rh + 1;	/* l: RH+1 */
295 				RCU_INIT_POINTER(lr->avl_left, l);	/* l: RH+1 */
296 				RCU_INIT_POINTER(lr->avl_right, node);	/* node: RH+1 */
297 				lr->avl_height = rh + 2;
298 				RCU_INIT_POINTER(*nodep, lr);
299 			}
300 		} else if (rh > lh + 1) { /* r: LH+2 */
301 			struct inet_peer *rr, *rl, *rlr, *rll;
302 			int rlh;
303 			rr = rcu_deref_locked(r->avl_right, base);
304 			rl = rcu_deref_locked(r->avl_left, base);
305 			rlh = node_height(rl);
306 			if (rlh <= node_height(rr)) {	/* rr: LH+1 */
307 				RCU_INIT_POINTER(node->avl_right, rl);	/* rl: LH or LH+1 */
308 				RCU_INIT_POINTER(node->avl_left, l);	/* l: LH */
309 				node->avl_height = rlh + 1; /* LH+1 or LH+2 */
310 				RCU_INIT_POINTER(r->avl_right, rr);	/* rr: LH+1 */
311 				RCU_INIT_POINTER(r->avl_left, node);	/* node: LH+1 or LH+2 */
312 				r->avl_height = node->avl_height + 1;
313 				RCU_INIT_POINTER(*nodep, r);
314 			} else { /* rr: RH, rl: RH+1 */
315 				rlr = rcu_deref_locked(rl->avl_right, base);/* rlr: LH or LH-1 */
316 				rll = rcu_deref_locked(rl->avl_left, base);/* rll: LH or LH-1 */
317 				RCU_INIT_POINTER(node->avl_right, rll);	/* rll: LH or LH-1 */
318 				RCU_INIT_POINTER(node->avl_left, l);	/* l: LH */
319 				node->avl_height = lh + 1; /* node: LH+1 */
320 				RCU_INIT_POINTER(r->avl_right, rr);	/* rr: LH */
321 				RCU_INIT_POINTER(r->avl_left, rlr);	/* rlr: LH or LH-1 */
322 				r->avl_height = lh + 1;	/* r: LH+1 */
323 				RCU_INIT_POINTER(rl->avl_right, r);	/* r: LH+1 */
324 				RCU_INIT_POINTER(rl->avl_left, node);	/* node: LH+1 */
325 				rl->avl_height = lh + 2;
326 				RCU_INIT_POINTER(*nodep, rl);
327 			}
328 		} else {
329 			node->avl_height = (lh > rh ? lh : rh) + 1;
330 		}
331 	}
332 }
333 
334 /* Called with local BH disabled and the pool lock held. */
335 #define link_to_pool(n, base)					\
336 do {								\
337 	n->avl_height = 1;					\
338 	n->avl_left = peer_avl_empty_rcu;			\
339 	n->avl_right = peer_avl_empty_rcu;			\
340 	/* lockless readers can catch us now */			\
341 	rcu_assign_pointer(**--stackptr, n);			\
342 	peer_avl_rebalance(stack, stackptr, base);		\
343 } while (0)
344 
345 static void inetpeer_free_rcu(struct rcu_head *head)
346 {
347 	kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
348 }
349 
350 static void unlink_from_pool(struct inet_peer *p, struct inet_peer_base *base,
351 			     struct inet_peer __rcu **stack[PEER_MAXDEPTH])
352 {
353 	struct inet_peer __rcu ***stackptr, ***delp;
354 
355 	if (lookup(&p->daddr, stack, base) != p)
356 		BUG();
357 	delp = stackptr - 1; /* *delp[0] == p */
358 	if (p->avl_left == peer_avl_empty_rcu) {
359 		*delp[0] = p->avl_right;
360 		--stackptr;
361 	} else {
362 		/* look for a node to insert instead of p */
363 		struct inet_peer *t;
364 		t = lookup_rightempty(p, base);
365 		BUG_ON(rcu_deref_locked(*stackptr[-1], base) != t);
366 		**--stackptr = t->avl_left;
367 		/* t is removed, t->daddr > x->daddr for any
368 		 * x in p->avl_left subtree.
369 		 * Put t in the old place of p. */
370 		RCU_INIT_POINTER(*delp[0], t);
371 		t->avl_left = p->avl_left;
372 		t->avl_right = p->avl_right;
373 		t->avl_height = p->avl_height;
374 		BUG_ON(delp[1] != &p->avl_left);
375 		delp[1] = &t->avl_left; /* was &p->avl_left */
376 	}
377 	peer_avl_rebalance(stack, stackptr, base);
378 	base->total--;
379 	call_rcu(&p->rcu, inetpeer_free_rcu);
380 }
381 
382 /* perform garbage collect on all items stacked during a lookup */
383 static int inet_peer_gc(struct inet_peer_base *base,
384 			struct inet_peer __rcu **stack[PEER_MAXDEPTH],
385 			struct inet_peer __rcu ***stackptr)
386 {
387 	struct inet_peer *p, *gchead = NULL;
388 	__u32 delta, ttl;
389 	int cnt = 0;
390 
391 	if (base->total >= inet_peer_threshold)
392 		ttl = 0; /* be aggressive */
393 	else
394 		ttl = inet_peer_maxttl
395 				- (inet_peer_maxttl - inet_peer_minttl) / HZ *
396 					base->total / inet_peer_threshold * HZ;
397 	stackptr--; /* last stack slot is peer_avl_empty */
398 	while (stackptr > stack) {
399 		stackptr--;
400 		p = rcu_deref_locked(**stackptr, base);
401 		if (atomic_read(&p->refcnt) == 0) {
402 			smp_rmb();
403 			delta = (__u32)jiffies - p->dtime;
404 			if (delta >= ttl &&
405 			    atomic_cmpxchg(&p->refcnt, 0, -1) == 0) {
406 				p->gc_next = gchead;
407 				gchead = p;
408 			}
409 		}
410 	}
411 	while ((p = gchead) != NULL) {
412 		gchead = p->gc_next;
413 		cnt++;
414 		unlink_from_pool(p, base, stack);
415 	}
416 	return cnt;
417 }
418 
419 struct inet_peer *inet_getpeer(struct inet_peer_base *base,
420 			       const struct inetpeer_addr *daddr,
421 			       int create)
422 {
423 	struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr;
424 	struct inet_peer *p;
425 	unsigned int sequence;
426 	int invalidated, gccnt = 0;
427 
428 	/* Attempt a lockless lookup first.
429 	 * Because of a concurrent writer, we might not find an existing entry.
430 	 */
431 	rcu_read_lock();
432 	sequence = read_seqbegin(&base->lock);
433 	p = lookup_rcu(daddr, base);
434 	invalidated = read_seqretry(&base->lock, sequence);
435 	rcu_read_unlock();
436 
437 	if (p)
438 		return p;
439 
440 	/* If no writer did a change during our lookup, we can return early. */
441 	if (!create && !invalidated)
442 		return NULL;
443 
444 	/* retry an exact lookup, taking the lock before.
445 	 * At least, nodes should be hot in our cache.
446 	 */
447 	write_seqlock_bh(&base->lock);
448 relookup:
449 	p = lookup(daddr, stack, base);
450 	if (p != peer_avl_empty) {
451 		atomic_inc(&p->refcnt);
452 		write_sequnlock_bh(&base->lock);
453 		return p;
454 	}
455 	if (!gccnt) {
456 		gccnt = inet_peer_gc(base, stack, stackptr);
457 		if (gccnt && create)
458 			goto relookup;
459 	}
460 	p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
461 	if (p) {
462 		p->daddr = *daddr;
463 		atomic_set(&p->refcnt, 1);
464 		atomic_set(&p->rid, 0);
465 		p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
466 		p->rate_tokens = 0;
467 		/* 60*HZ is arbitrary, but chosen enough high so that the first
468 		 * calculation of tokens is at its maximum.
469 		 */
470 		p->rate_last = jiffies - 60*HZ;
471 		INIT_LIST_HEAD(&p->gc_list);
472 
473 		/* Link the node. */
474 		link_to_pool(p, base);
475 		base->total++;
476 	}
477 	write_sequnlock_bh(&base->lock);
478 
479 	return p;
480 }
481 EXPORT_SYMBOL_GPL(inet_getpeer);
482 
483 void inet_putpeer(struct inet_peer *p)
484 {
485 	p->dtime = (__u32)jiffies;
486 	smp_mb__before_atomic();
487 	atomic_dec(&p->refcnt);
488 }
489 EXPORT_SYMBOL_GPL(inet_putpeer);
490 
491 /*
492  *	Check transmit rate limitation for given message.
493  *	The rate information is held in the inet_peer entries now.
494  *	This function is generic and could be used for other purposes
495  *	too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
496  *
497  *	Note that the same inet_peer fields are modified by functions in
498  *	route.c too, but these work for packet destinations while xrlim_allow
499  *	works for icmp destinations. This means the rate limiting information
500  *	for one "ip object" is shared - and these ICMPs are twice limited:
501  *	by source and by destination.
502  *
503  *	RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
504  *			  SHOULD allow setting of rate limits
505  *
506  * 	Shared between ICMPv4 and ICMPv6.
507  */
508 #define XRLIM_BURST_FACTOR 6
509 bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
510 {
511 	unsigned long now, token;
512 	bool rc = false;
513 
514 	if (!peer)
515 		return true;
516 
517 	token = peer->rate_tokens;
518 	now = jiffies;
519 	token += now - peer->rate_last;
520 	peer->rate_last = now;
521 	if (token > XRLIM_BURST_FACTOR * timeout)
522 		token = XRLIM_BURST_FACTOR * timeout;
523 	if (token >= timeout) {
524 		token -= timeout;
525 		rc = true;
526 	}
527 	peer->rate_tokens = token;
528 	return rc;
529 }
530 EXPORT_SYMBOL(inet_peer_xrlim_allow);
531 
532 static void inetpeer_inval_rcu(struct rcu_head *head)
533 {
534 	struct inet_peer *p = container_of(head, struct inet_peer, gc_rcu);
535 
536 	spin_lock_bh(&gc_lock);
537 	list_add_tail(&p->gc_list, &gc_list);
538 	spin_unlock_bh(&gc_lock);
539 
540 	schedule_delayed_work(&gc_work, gc_delay);
541 }
542 
543 void inetpeer_invalidate_tree(struct inet_peer_base *base)
544 {
545 	struct inet_peer *root;
546 
547 	write_seqlock_bh(&base->lock);
548 
549 	root = rcu_deref_locked(base->root, base);
550 	if (root != peer_avl_empty) {
551 		base->root = peer_avl_empty_rcu;
552 		base->total = 0;
553 		call_rcu(&root->gc_rcu, inetpeer_inval_rcu);
554 	}
555 
556 	write_sequnlock_bh(&base->lock);
557 }
558 EXPORT_SYMBOL(inetpeer_invalidate_tree);
559