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
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
inetpeer_free_rcu(struct rcu_head * head)131 static void inetpeer_free_rcu(struct rcu_head *head)
132 {
133 kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
134 }
135
136 /* 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)137 static void inet_peer_gc(struct inet_peer_base *base,
138 struct inet_peer *gc_stack[],
139 unsigned int gc_cnt)
140 {
141 int peer_threshold, peer_maxttl, peer_minttl;
142 struct inet_peer *p;
143 __u32 delta, ttl;
144 int i;
145
146 peer_threshold = READ_ONCE(inet_peer_threshold);
147 peer_maxttl = READ_ONCE(inet_peer_maxttl);
148 peer_minttl = READ_ONCE(inet_peer_minttl);
149
150 if (base->total >= peer_threshold)
151 ttl = 0; /* be aggressive */
152 else
153 ttl = peer_maxttl - (peer_maxttl - peer_minttl) / HZ *
154 base->total / peer_threshold * HZ;
155 for (i = 0; i < gc_cnt; i++) {
156 p = gc_stack[i];
157
158 /* The READ_ONCE() pairs with the WRITE_ONCE()
159 * in inet_putpeer()
160 */
161 delta = (__u32)jiffies - READ_ONCE(p->dtime);
162
163 if (delta < ttl || !refcount_dec_if_one(&p->refcnt))
164 gc_stack[i] = NULL;
165 }
166 for (i = 0; i < gc_cnt; i++) {
167 p = gc_stack[i];
168 if (p) {
169 rb_erase(&p->rb_node, &base->rb_root);
170 base->total--;
171 call_rcu(&p->rcu, inetpeer_free_rcu);
172 }
173 }
174 }
175
inet_getpeer(struct inet_peer_base * base,const struct inetpeer_addr * daddr,int create)176 struct inet_peer *inet_getpeer(struct inet_peer_base *base,
177 const struct inetpeer_addr *daddr,
178 int create)
179 {
180 struct inet_peer *p, *gc_stack[PEER_MAX_GC];
181 struct rb_node **pp, *parent;
182 unsigned int gc_cnt, seq;
183 int invalidated;
184
185 /* Attempt a lockless lookup first.
186 * Because of a concurrent writer, we might not find an existing entry.
187 */
188 rcu_read_lock();
189 seq = read_seqbegin(&base->lock);
190 p = lookup(daddr, base, seq, NULL, &gc_cnt, &parent, &pp);
191 invalidated = read_seqretry(&base->lock, seq);
192 rcu_read_unlock();
193
194 if (p)
195 return p;
196
197 /* If no writer did a change during our lookup, we can return early. */
198 if (!create && !invalidated)
199 return NULL;
200
201 /* retry an exact lookup, taking the lock before.
202 * At least, nodes should be hot in our cache.
203 */
204 parent = NULL;
205 write_seqlock_bh(&base->lock);
206
207 gc_cnt = 0;
208 p = lookup(daddr, base, seq, gc_stack, &gc_cnt, &parent, &pp);
209 if (!p && create) {
210 p = kmem_cache_alloc(peer_cachep, GFP_ATOMIC);
211 if (p) {
212 p->daddr = *daddr;
213 p->dtime = (__u32)jiffies;
214 refcount_set(&p->refcnt, 2);
215 atomic_set(&p->rid, 0);
216 p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
217 p->rate_tokens = 0;
218 p->n_redirects = 0;
219 /* 60*HZ is arbitrary, but chosen enough high so that the first
220 * calculation of tokens is at its maximum.
221 */
222 p->rate_last = jiffies - 60*HZ;
223
224 rb_link_node(&p->rb_node, parent, pp);
225 rb_insert_color(&p->rb_node, &base->rb_root);
226 base->total++;
227 }
228 }
229 if (gc_cnt)
230 inet_peer_gc(base, gc_stack, gc_cnt);
231 write_sequnlock_bh(&base->lock);
232
233 return p;
234 }
235 EXPORT_SYMBOL_GPL(inet_getpeer);
236
inet_putpeer(struct inet_peer * p)237 void inet_putpeer(struct inet_peer *p)
238 {
239 /* The WRITE_ONCE() pairs with itself (we run lockless)
240 * and the READ_ONCE() in inet_peer_gc()
241 */
242 WRITE_ONCE(p->dtime, (__u32)jiffies);
243
244 if (refcount_dec_and_test(&p->refcnt))
245 call_rcu(&p->rcu, inetpeer_free_rcu);
246 }
247 EXPORT_SYMBOL_GPL(inet_putpeer);
248
249 /*
250 * Check transmit rate limitation for given message.
251 * The rate information is held in the inet_peer entries now.
252 * This function is generic and could be used for other purposes
253 * too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
254 *
255 * Note that the same inet_peer fields are modified by functions in
256 * route.c too, but these work for packet destinations while xrlim_allow
257 * works for icmp destinations. This means the rate limiting information
258 * for one "ip object" is shared - and these ICMPs are twice limited:
259 * by source and by destination.
260 *
261 * RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
262 * SHOULD allow setting of rate limits
263 *
264 * Shared between ICMPv4 and ICMPv6.
265 */
266 #define XRLIM_BURST_FACTOR 6
inet_peer_xrlim_allow(struct inet_peer * peer,int timeout)267 bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
268 {
269 unsigned long now, token;
270 bool rc = false;
271
272 if (!peer)
273 return true;
274
275 token = peer->rate_tokens;
276 now = jiffies;
277 token += now - peer->rate_last;
278 peer->rate_last = now;
279 if (token > XRLIM_BURST_FACTOR * timeout)
280 token = XRLIM_BURST_FACTOR * timeout;
281 if (token >= timeout) {
282 token -= timeout;
283 rc = true;
284 }
285 peer->rate_tokens = token;
286 return rc;
287 }
288 EXPORT_SYMBOL(inet_peer_xrlim_allow);
289
inetpeer_invalidate_tree(struct inet_peer_base * base)290 void inetpeer_invalidate_tree(struct inet_peer_base *base)
291 {
292 struct rb_node *p = rb_first(&base->rb_root);
293
294 while (p) {
295 struct inet_peer *peer = rb_entry(p, struct inet_peer, rb_node);
296
297 p = rb_next(p);
298 rb_erase(&peer->rb_node, &base->rb_root);
299 inet_putpeer(peer);
300 cond_resched();
301 }
302
303 base->total = 0;
304 }
305 EXPORT_SYMBOL(inetpeer_invalidate_tree);
306