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
inet_peer_base_init(struct inet_peer_base * bp)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 */
inet_initpeers(void)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 */
lookup(const struct inetpeer_addr * daddr,struct inet_peer_base * base,unsigned int seq,struct inet_peer * gc_stack[],unsigned int * gc_cnt,struct rb_node ** parent_p,struct rb_node *** pp_p)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 u32 now;
99
100 pp = &base->rb_root.rb_node;
101 parent = NULL;
102 while (1) {
103 int cmp;
104
105 next = rcu_dereference_raw(*pp);
106 if (!next)
107 break;
108 parent = next;
109 p = rb_entry(parent, struct inet_peer, rb_node);
110 cmp = inetpeer_addr_cmp(daddr, &p->daddr);
111 if (cmp == 0) {
112 now = jiffies;
113 if (READ_ONCE(p->dtime) != now)
114 WRITE_ONCE(p->dtime, now);
115 return p;
116 }
117 if (gc_stack) {
118 if (*gc_cnt < PEER_MAX_GC)
119 gc_stack[(*gc_cnt)++] = p;
120 } else if (unlikely(read_seqretry(&base->lock, seq))) {
121 break;
122 }
123 if (cmp == -1)
124 pp = &next->rb_left;
125 else
126 pp = &next->rb_right;
127 }
128 *parent_p = parent;
129 *pp_p = pp;
130 return NULL;
131 }
132
inetpeer_free_rcu(struct rcu_head * head)133 static void inetpeer_free_rcu(struct rcu_head *head)
134 {
135 kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
136 }
137
138 /* perform garbage collect on all items stacked during a lookup */
inet_peer_gc(struct inet_peer_base * base,struct inet_peer * gc_stack[],unsigned int gc_cnt)139 static void inet_peer_gc(struct inet_peer_base *base,
140 struct inet_peer *gc_stack[],
141 unsigned int gc_cnt)
142 {
143 int peer_threshold, peer_maxttl, peer_minttl;
144 struct inet_peer *p;
145 __u32 delta, ttl;
146 int i;
147
148 peer_threshold = READ_ONCE(inet_peer_threshold);
149 peer_maxttl = READ_ONCE(inet_peer_maxttl);
150 peer_minttl = READ_ONCE(inet_peer_minttl);
151
152 if (base->total >= peer_threshold)
153 ttl = 0; /* be aggressive */
154 else
155 ttl = peer_maxttl - (peer_maxttl - peer_minttl) / HZ *
156 base->total / peer_threshold * HZ;
157 for (i = 0; i < gc_cnt; i++) {
158 p = gc_stack[i];
159
160 delta = (__u32)jiffies - READ_ONCE(p->dtime);
161
162 if (delta < ttl || !refcount_dec_if_one(&p->refcnt))
163 gc_stack[i] = NULL;
164 }
165 for (i = 0; i < gc_cnt; i++) {
166 p = gc_stack[i];
167 if (p) {
168 rb_erase(&p->rb_node, &base->rb_root);
169 base->total--;
170 call_rcu(&p->rcu, inetpeer_free_rcu);
171 }
172 }
173 }
174
175 /* Must be called under RCU : No refcount change is done here. */
inet_getpeer(struct inet_peer_base * base,const struct inetpeer_addr * daddr)176 struct inet_peer *inet_getpeer(struct inet_peer_base *base,
177 const struct inetpeer_addr *daddr)
178 {
179 struct inet_peer *p, *gc_stack[PEER_MAX_GC];
180 struct rb_node **pp, *parent;
181 unsigned int gc_cnt, seq;
182
183 /* Attempt a lockless lookup first.
184 * Because of a concurrent writer, we might not find an existing entry.
185 */
186 seq = read_seqbegin(&base->lock);
187 p = lookup(daddr, base, seq, NULL, &gc_cnt, &parent, &pp);
188
189 if (p)
190 return p;
191
192 /* retry an exact lookup, taking the lock before.
193 * At least, nodes should be hot in our cache.
194 */
195 parent = NULL;
196 write_seqlock_bh(&base->lock);
197
198 gc_cnt = 0;
199 p = lookup(daddr, base, seq, gc_stack, &gc_cnt, &parent, &pp);
200 if (!p) {
201 p = kmem_cache_alloc(peer_cachep, GFP_ATOMIC);
202 if (p) {
203 p->daddr = *daddr;
204 p->dtime = (__u32)jiffies;
205 refcount_set(&p->refcnt, 1);
206 atomic_set(&p->rid, 0);
207 p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
208 p->rate_tokens = 0;
209 p->n_redirects = 0;
210 /* 60*HZ is arbitrary, but chosen enough high so that the first
211 * calculation of tokens is at its maximum.
212 */
213 p->rate_last = jiffies - 60*HZ;
214
215 rb_link_node(&p->rb_node, parent, pp);
216 rb_insert_color(&p->rb_node, &base->rb_root);
217 base->total++;
218 }
219 }
220 if (gc_cnt)
221 inet_peer_gc(base, gc_stack, gc_cnt);
222 write_sequnlock_bh(&base->lock);
223
224 return p;
225 }
226 EXPORT_SYMBOL_GPL(inet_getpeer);
227
inet_putpeer(struct inet_peer * p)228 void inet_putpeer(struct inet_peer *p)
229 {
230 if (refcount_dec_and_test(&p->refcnt))
231 call_rcu(&p->rcu, inetpeer_free_rcu);
232 }
233
234 /*
235 * Check transmit rate limitation for given message.
236 * The rate information is held in the inet_peer entries now.
237 * This function is generic and could be used for other purposes
238 * too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
239 *
240 * Note that the same inet_peer fields are modified by functions in
241 * route.c too, but these work for packet destinations while xrlim_allow
242 * works for icmp destinations. This means the rate limiting information
243 * for one "ip object" is shared - and these ICMPs are twice limited:
244 * by source and by destination.
245 *
246 * RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
247 * SHOULD allow setting of rate limits
248 *
249 * Shared between ICMPv4 and ICMPv6.
250 */
251 #define XRLIM_BURST_FACTOR 6
inet_peer_xrlim_allow(struct inet_peer * peer,int timeout)252 bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
253 {
254 unsigned long now, token;
255 bool rc = false;
256
257 if (!peer)
258 return true;
259
260 token = peer->rate_tokens;
261 now = jiffies;
262 token += now - peer->rate_last;
263 peer->rate_last = now;
264 if (token > XRLIM_BURST_FACTOR * timeout)
265 token = XRLIM_BURST_FACTOR * timeout;
266 if (token >= timeout) {
267 token -= timeout;
268 rc = true;
269 }
270 peer->rate_tokens = token;
271 return rc;
272 }
273 EXPORT_SYMBOL(inet_peer_xrlim_allow);
274
inetpeer_invalidate_tree(struct inet_peer_base * base)275 void inetpeer_invalidate_tree(struct inet_peer_base *base)
276 {
277 struct rb_node *p = rb_first(&base->rb_root);
278
279 while (p) {
280 struct inet_peer *peer = rb_entry(p, struct inet_peer, rb_node);
281
282 p = rb_next(p);
283 rb_erase(&peer->rb_node, &base->rb_root);
284 inet_putpeer(peer);
285 cond_resched();
286 }
287
288 base->total = 0;
289 }
290 EXPORT_SYMBOL(inetpeer_invalidate_tree);
291