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_create("inet_peer_cache",
85 sizeof(struct inet_peer),
86 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
87 NULL);
88 }
89
90 /* 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)91 static struct inet_peer *lookup(const struct inetpeer_addr *daddr,
92 struct inet_peer_base *base,
93 unsigned int seq,
94 struct inet_peer *gc_stack[],
95 unsigned int *gc_cnt,
96 struct rb_node **parent_p,
97 struct rb_node ***pp_p)
98 {
99 struct rb_node **pp, *parent, *next;
100 struct inet_peer *p;
101 u32 now;
102
103 pp = &base->rb_root.rb_node;
104 parent = NULL;
105 while (1) {
106 int cmp;
107
108 next = rcu_dereference_raw(*pp);
109 if (!next)
110 break;
111 parent = next;
112 p = rb_entry(parent, struct inet_peer, rb_node);
113 cmp = inetpeer_addr_cmp(daddr, &p->daddr);
114 if (cmp == 0) {
115 now = jiffies;
116 if (READ_ONCE(p->dtime) != now)
117 WRITE_ONCE(p->dtime, now);
118 return p;
119 }
120 if (gc_stack) {
121 if (*gc_cnt < PEER_MAX_GC)
122 gc_stack[(*gc_cnt)++] = p;
123 } else if (unlikely(read_seqretry(&base->lock, seq))) {
124 break;
125 }
126 if (cmp == -1)
127 pp = &next->rb_left;
128 else
129 pp = &next->rb_right;
130 }
131 *parent_p = parent;
132 *pp_p = pp;
133 return NULL;
134 }
135
inetpeer_free_rcu(struct rcu_head * head)136 static void inetpeer_free_rcu(struct rcu_head *head)
137 {
138 kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
139 }
140
141 /* 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)142 static void inet_peer_gc(struct inet_peer_base *base,
143 struct inet_peer *gc_stack[],
144 unsigned int gc_cnt)
145 {
146 int peer_threshold, peer_maxttl, peer_minttl;
147 struct inet_peer *p;
148 __u32 delta, ttl;
149 int i;
150
151 peer_threshold = READ_ONCE(inet_peer_threshold);
152 peer_maxttl = READ_ONCE(inet_peer_maxttl);
153 peer_minttl = READ_ONCE(inet_peer_minttl);
154
155 if (base->total >= peer_threshold)
156 ttl = 0; /* be aggressive */
157 else
158 ttl = peer_maxttl - (peer_maxttl - peer_minttl) / HZ *
159 base->total / peer_threshold * HZ;
160 for (i = 0; i < gc_cnt; i++) {
161 p = gc_stack[i];
162
163 delta = (__u32)jiffies - READ_ONCE(p->dtime);
164
165 if (delta < ttl || !refcount_dec_if_one(&p->refcnt))
166 gc_stack[i] = NULL;
167 }
168 for (i = 0; i < gc_cnt; i++) {
169 p = gc_stack[i];
170 if (p) {
171 rb_erase(&p->rb_node, &base->rb_root);
172 base->total--;
173 call_rcu(&p->rcu, inetpeer_free_rcu);
174 }
175 }
176 }
177
178 /* Must be called under RCU : No refcount change is done here. */
inet_getpeer(struct inet_peer_base * base,const struct inetpeer_addr * daddr)179 struct inet_peer *inet_getpeer(struct inet_peer_base *base,
180 const struct inetpeer_addr *daddr)
181 {
182 struct inet_peer *p, *gc_stack[PEER_MAX_GC];
183 struct rb_node **pp, *parent;
184 unsigned int gc_cnt, seq;
185
186 /* Attempt a lockless lookup first.
187 * Because of a concurrent writer, we might not find an existing entry.
188 */
189 seq = read_seqbegin(&base->lock);
190 p = lookup(daddr, base, seq, NULL, &gc_cnt, &parent, &pp);
191
192 if (p)
193 return p;
194
195 /* retry an exact lookup, taking the lock before.
196 * At least, nodes should be hot in our cache.
197 */
198 parent = NULL;
199 write_seqlock_bh(&base->lock);
200
201 gc_cnt = 0;
202 p = lookup(daddr, base, seq, gc_stack, &gc_cnt, &parent, &pp);
203 if (!p) {
204 p = kmem_cache_alloc(peer_cachep, GFP_ATOMIC);
205 if (p) {
206 p->daddr = *daddr;
207 p->dtime = (__u32)jiffies;
208 refcount_set(&p->refcnt, 1);
209 atomic_set(&p->rid, 0);
210 p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
211 p->rate_tokens = 0;
212 p->n_redirects = 0;
213 /* 60*HZ is arbitrary, but chosen enough high so that the first
214 * calculation of tokens is at its maximum.
215 */
216 p->rate_last = jiffies - 60*HZ;
217
218 rb_link_node(&p->rb_node, parent, pp);
219 rb_insert_color(&p->rb_node, &base->rb_root);
220 base->total++;
221 }
222 }
223 if (gc_cnt)
224 inet_peer_gc(base, gc_stack, gc_cnt);
225 write_sequnlock_bh(&base->lock);
226
227 return p;
228 }
229 EXPORT_SYMBOL_GPL(inet_getpeer);
230
inet_putpeer(struct inet_peer * p)231 void inet_putpeer(struct inet_peer *p)
232 {
233 if (refcount_dec_and_test(&p->refcnt))
234 call_rcu(&p->rcu, inetpeer_free_rcu);
235 }
236
237 /*
238 * Check transmit rate limitation for given message.
239 * The rate information is held in the inet_peer entries now.
240 * This function is generic and could be used for other purposes
241 * too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
242 *
243 * Note that the same inet_peer fields are modified by functions in
244 * route.c too, but these work for packet destinations while xrlim_allow
245 * works for icmp destinations. This means the rate limiting information
246 * for one "ip object" is shared - and these ICMPs are twice limited:
247 * by source and by destination.
248 *
249 * RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
250 * SHOULD allow setting of rate limits
251 *
252 * Shared between ICMPv4 and ICMPv6.
253 */
254 #define XRLIM_BURST_FACTOR 6
inet_peer_xrlim_allow(struct inet_peer * peer,int timeout)255 bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
256 {
257 unsigned long now, token;
258 bool rc = false;
259
260 if (!peer)
261 return true;
262
263 token = peer->rate_tokens;
264 now = jiffies;
265 token += now - peer->rate_last;
266 peer->rate_last = now;
267 if (token > XRLIM_BURST_FACTOR * timeout)
268 token = XRLIM_BURST_FACTOR * timeout;
269 if (token >= timeout) {
270 token -= timeout;
271 rc = true;
272 }
273 peer->rate_tokens = token;
274 return rc;
275 }
276 EXPORT_SYMBOL(inet_peer_xrlim_allow);
277
inetpeer_invalidate_tree(struct inet_peer_base * base)278 void inetpeer_invalidate_tree(struct inet_peer_base *base)
279 {
280 struct rb_node *p = rb_first(&base->rb_root);
281
282 while (p) {
283 struct inet_peer *peer = rb_entry(p, struct inet_peer, rb_node);
284
285 p = rb_next(p);
286 rb_erase(&peer->rb_node, &base->rb_root);
287 inet_putpeer(peer);
288 cond_resched();
289 }
290
291 base->total = 0;
292 }
293 EXPORT_SYMBOL(inetpeer_invalidate_tree);
294