xref: /linux/net/ipv4/inetpeer.c (revision c8b90d40d5bba8e6fba457b8a7c10d3c0d467e37)
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/cache.h>
10 #include <linux/module.h>
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/timer.h>
17 #include <linux/time.h>
18 #include <linux/kernel.h>
19 #include <linux/mm.h>
20 #include <linux/net.h>
21 #include <linux/workqueue.h>
22 #include <net/ip.h>
23 #include <net/inetpeer.h>
24 #include <net/secure_seq.h>
25 
26 /*
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  *
31  *  Nodes are removed only when reference counter goes to 0.
32  *  When it's happened the node may be removed when a sufficient amount of
33  *  time has been passed since its last use.  The less-recently-used entry can
34  *  also be removed if the pool is overloaded i.e. if the total amount of
35  *  entries is greater-or-equal than the threshold.
36  *
37  *  Node pool is organised as an RB tree.
38  *  Such an implementation has been chosen not just for fun.  It's a way to
39  *  prevent easy and efficient DoS attacks by creating hash collisions.  A huge
40  *  amount of long living nodes in a single hash slot would significantly delay
41  *  lookups performed with disabled BHs.
42  *
43  *  Serialisation issues.
44  *  1.  Nodes may appear in the tree only with the pool lock held.
45  *  2.  Nodes may disappear from the tree only with the pool lock held
46  *      AND reference count being 0.
47  *  3.  Global variable peer_total is modified under the pool lock.
48  *  4.  struct inet_peer fields modification:
49  *		rb_node: pool lock
50  *		refcnt: atomically against modifications on other CPU;
51  *		   usually under some other lock to prevent node disappearing
52  *		daddr: unchangeable
53  */
54 
55 static struct kmem_cache *peer_cachep __ro_after_init;
56 
57 void inet_peer_base_init(struct inet_peer_base *bp)
58 {
59 	bp->rb_root = RB_ROOT;
60 	seqlock_init(&bp->lock);
61 	bp->total = 0;
62 }
63 EXPORT_SYMBOL_GPL(inet_peer_base_init);
64 
65 #define PEER_MAX_GC 32
66 
67 /* Exported for sysctl_net_ipv4.  */
68 int inet_peer_threshold __read_mostly;	/* start to throw entries more
69 					 * aggressively at this stage */
70 int inet_peer_minttl __read_mostly = 120 * HZ;	/* TTL under high load: 120 sec */
71 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ;	/* usual time to live: 10 min */
72 
73 /* Called from ip_output.c:ip_init  */
74 void __init inet_initpeers(void)
75 {
76 	u64 nr_entries;
77 
78 	 /* 1% of physical memory */
79 	nr_entries = div64_ul((u64)totalram_pages() << PAGE_SHIFT,
80 			      100 * L1_CACHE_ALIGN(sizeof(struct inet_peer)));
81 
82 	inet_peer_threshold = clamp_val(nr_entries, 4096, 65536 + 128);
83 
84 	peer_cachep = KMEM_CACHE(inet_peer, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
85 }
86 
87 /* Called with rcu_read_lock() or base->lock held */
88 static struct inet_peer *lookup(const struct inetpeer_addr *daddr,
89 				struct inet_peer_base *base,
90 				unsigned int seq,
91 				struct inet_peer *gc_stack[],
92 				unsigned int *gc_cnt,
93 				struct rb_node **parent_p,
94 				struct rb_node ***pp_p)
95 {
96 	struct rb_node **pp, *parent, *next;
97 	struct inet_peer *p;
98 
99 	pp = &base->rb_root.rb_node;
100 	parent = NULL;
101 	while (1) {
102 		int cmp;
103 
104 		next = rcu_dereference_raw(*pp);
105 		if (!next)
106 			break;
107 		parent = next;
108 		p = rb_entry(parent, struct inet_peer, rb_node);
109 		cmp = inetpeer_addr_cmp(daddr, &p->daddr);
110 		if (cmp == 0) {
111 			if (!refcount_inc_not_zero(&p->refcnt))
112 				break;
113 			return p;
114 		}
115 		if (gc_stack) {
116 			if (*gc_cnt < PEER_MAX_GC)
117 				gc_stack[(*gc_cnt)++] = p;
118 		} else if (unlikely(read_seqretry(&base->lock, seq))) {
119 			break;
120 		}
121 		if (cmp == -1)
122 			pp = &next->rb_left;
123 		else
124 			pp = &next->rb_right;
125 	}
126 	*parent_p = parent;
127 	*pp_p = pp;
128 	return NULL;
129 }
130 
131 /* perform garbage collect on all items stacked during a lookup */
132 static void inet_peer_gc(struct inet_peer_base *base,
133 			 struct inet_peer *gc_stack[],
134 			 unsigned int gc_cnt)
135 {
136 	int peer_threshold, peer_maxttl, peer_minttl;
137 	struct inet_peer *p;
138 	__u32 delta, ttl;
139 	int i;
140 
141 	peer_threshold = READ_ONCE(inet_peer_threshold);
142 	peer_maxttl = READ_ONCE(inet_peer_maxttl);
143 	peer_minttl = READ_ONCE(inet_peer_minttl);
144 
145 	if (base->total >= peer_threshold)
146 		ttl = 0; /* be aggressive */
147 	else
148 		ttl = peer_maxttl - (peer_maxttl - peer_minttl) / HZ *
149 			base->total / peer_threshold * HZ;
150 	for (i = 0; i < gc_cnt; i++) {
151 		p = gc_stack[i];
152 
153 		/* The READ_ONCE() pairs with the WRITE_ONCE()
154 		 * in inet_putpeer()
155 		 */
156 		delta = (__u32)jiffies - READ_ONCE(p->dtime);
157 
158 		if (delta < ttl || !refcount_dec_if_one(&p->refcnt))
159 			gc_stack[i] = NULL;
160 	}
161 	for (i = 0; i < gc_cnt; i++) {
162 		p = gc_stack[i];
163 		if (p) {
164 			rb_erase(&p->rb_node, &base->rb_root);
165 			base->total--;
166 			kfree_rcu(p, rcu);
167 		}
168 	}
169 }
170 
171 struct inet_peer *inet_getpeer(struct inet_peer_base *base,
172 			       const struct inetpeer_addr *daddr,
173 			       int create)
174 {
175 	struct inet_peer *p, *gc_stack[PEER_MAX_GC];
176 	struct rb_node **pp, *parent;
177 	unsigned int gc_cnt, seq;
178 	int invalidated;
179 
180 	/* Attempt a lockless lookup first.
181 	 * Because of a concurrent writer, we might not find an existing entry.
182 	 */
183 	rcu_read_lock();
184 	seq = read_seqbegin(&base->lock);
185 	p = lookup(daddr, base, seq, NULL, &gc_cnt, &parent, &pp);
186 	invalidated = read_seqretry(&base->lock, seq);
187 	rcu_read_unlock();
188 
189 	if (p)
190 		return p;
191 
192 	/* If no writer did a change during our lookup, we can return early. */
193 	if (!create && !invalidated)
194 		return NULL;
195 
196 	/* retry an exact lookup, taking the lock before.
197 	 * At least, nodes should be hot in our cache.
198 	 */
199 	parent = NULL;
200 	write_seqlock_bh(&base->lock);
201 
202 	gc_cnt = 0;
203 	p = lookup(daddr, base, seq, gc_stack, &gc_cnt, &parent, &pp);
204 	if (!p && create) {
205 		p = kmem_cache_alloc(peer_cachep, GFP_ATOMIC);
206 		if (p) {
207 			p->daddr = *daddr;
208 			p->dtime = (__u32)jiffies;
209 			refcount_set(&p->refcnt, 2);
210 			atomic_set(&p->rid, 0);
211 			p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
212 			p->rate_tokens = 0;
213 			p->n_redirects = 0;
214 			/* 60*HZ is arbitrary, but chosen enough high so that the first
215 			 * calculation of tokens is at its maximum.
216 			 */
217 			p->rate_last = jiffies - 60*HZ;
218 
219 			rb_link_node(&p->rb_node, parent, pp);
220 			rb_insert_color(&p->rb_node, &base->rb_root);
221 			base->total++;
222 		}
223 	}
224 	if (gc_cnt)
225 		inet_peer_gc(base, gc_stack, gc_cnt);
226 	write_sequnlock_bh(&base->lock);
227 
228 	return p;
229 }
230 EXPORT_SYMBOL_GPL(inet_getpeer);
231 
232 void inet_putpeer(struct inet_peer *p)
233 {
234 	/* The WRITE_ONCE() pairs with itself (we run lockless)
235 	 * and the READ_ONCE() in inet_peer_gc()
236 	 */
237 	WRITE_ONCE(p->dtime, (__u32)jiffies);
238 
239 	if (refcount_dec_and_test(&p->refcnt))
240 		kfree_rcu(p, rcu);
241 }
242 EXPORT_SYMBOL_GPL(inet_putpeer);
243 
244 /*
245  *	Check transmit rate limitation for given message.
246  *	The rate information is held in the inet_peer entries now.
247  *	This function is generic and could be used for other purposes
248  *	too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
249  *
250  *	Note that the same inet_peer fields are modified by functions in
251  *	route.c too, but these work for packet destinations while xrlim_allow
252  *	works for icmp destinations. This means the rate limiting information
253  *	for one "ip object" is shared - and these ICMPs are twice limited:
254  *	by source and by destination.
255  *
256  *	RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
257  *			  SHOULD allow setting of rate limits
258  *
259  * 	Shared between ICMPv4 and ICMPv6.
260  */
261 #define XRLIM_BURST_FACTOR 6
262 bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
263 {
264 	unsigned long now, token;
265 	bool rc = false;
266 
267 	if (!peer)
268 		return true;
269 
270 	token = peer->rate_tokens;
271 	now = jiffies;
272 	token += now - peer->rate_last;
273 	peer->rate_last = now;
274 	if (token > XRLIM_BURST_FACTOR * timeout)
275 		token = XRLIM_BURST_FACTOR * timeout;
276 	if (token >= timeout) {
277 		token -= timeout;
278 		rc = true;
279 	}
280 	peer->rate_tokens = token;
281 	return rc;
282 }
283 EXPORT_SYMBOL(inet_peer_xrlim_allow);
284 
285 void inetpeer_invalidate_tree(struct inet_peer_base *base)
286 {
287 	struct rb_node *p = rb_first(&base->rb_root);
288 
289 	while (p) {
290 		struct inet_peer *peer = rb_entry(p, struct inet_peer, rb_node);
291 
292 		p = rb_next(p);
293 		rb_erase(&peer->rb_node, &base->rb_root);
294 		inet_putpeer(peer);
295 		cond_resched();
296 	}
297 
298 	base->total = 0;
299 }
300 EXPORT_SYMBOL(inetpeer_invalidate_tree);
301