projects/linux·blob master · refs

317 lines · 8.6 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * INETPEER - A storage for permanent information about peers
4 *
5 * Authors: Andrey V. Savochkin <[email protected]>
6 */
7
8#include <linux/cache.h>
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#include <linux/siphash.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
55static struct kmem_cache *peer_cachep __ro_after_init;
56static siphash_aligned_key_t inetpeer_hash_key __read_mostly;
57
58static u64 inetpeer_addr_hash(const struct inetpeer_addr *a)
59{
60 net_get_random_once(&inetpeer_hash_key, sizeof(inetpeer_hash_key));
61
62 if (a->family == AF_INET)
63 return siphash_2u32((__force u32)a->a4.addr, a->a4.vif,
64 &inetpeer_hash_key);
65
66 return siphash_4u32((__force u32)a->a6.s6_addr32[0],
67 (__force u32)a->a6.s6_addr32[1],
68 (__force u32)a->a6.s6_addr32[2],
69 (__force u32)a->a6.s6_addr32[3],
70 &inetpeer_hash_key);
71}
72
73static int inetpeer_entry_cmp(u64 dhash,
74 const struct inetpeer_addr *daddr,
75 const struct inet_peer *p)
76{
77 if (dhash < p->hash)
78 return -1;
79 if (dhash > p->hash)
80 return 1;
81
82 return inetpeer_addr_cmp(daddr, &p->daddr);
83}
84
85void inet_peer_base_init(struct inet_peer_base *bp)
86{
87 bp->rb_root = RB_ROOT;
88 seqlock_init(&bp->lock);
89 bp->total = 0;
90}
91
92#define PEER_MAX_GC 32
93
94/* Exported for sysctl_net_ipv4. */
95int inet_peer_threshold __read_mostly; /* start to throw entries more
96 * aggressively at this stage */
97int inet_peer_minttl __read_mostly = 120 * HZ; /* TTL under high load: 120 sec */
98int inet_peer_maxttl __read_mostly = 10 * 60 * HZ; /* usual time to live: 10 min */
99
100/* Called from ip_output.c:ip_init */
101void __init inet_initpeers(void)
102{
103 u64 nr_entries;
104
105 /* 1% of physical memory */
106 nr_entries = div64_ul((u64)totalram_pages() << PAGE_SHIFT,
107 100 * L1_CACHE_ALIGN(sizeof(struct inet_peer)));
108
109 inet_peer_threshold = clamp_val(nr_entries, 4096, 65536 + 128);
110
111 peer_cachep = KMEM_CACHE(inet_peer, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
112}
113
114/* Called with rcu_read_lock() or base->lock held */
115static struct inet_peer *lookup(const struct inetpeer_addr *daddr,
116 u64 dhash,
117 struct inet_peer_base *base,
118 unsigned int seq,
119 struct inet_peer *gc_stack[],
120 unsigned int *gc_cnt,
121 struct rb_node **parent_p,
122 struct rb_node ***pp_p)
123{
124 struct rb_node **pp, *parent, *next;
125 struct inet_peer *p;
126 u32 now;
127
128 pp = &base->rb_root.rb_node;
129 parent = NULL;
130 while (1) {
131 int cmp;
132
133 next = rcu_dereference_raw(*pp);
134 if (!next)
135 break;
136 parent = next;
137 p = rb_entry(parent, struct inet_peer, rb_node);
138 cmp = inetpeer_entry_cmp(dhash, daddr, p);
139 if (cmp == 0) {
140 now = jiffies;
141 if (READ_ONCE(p->dtime) != now)
142 WRITE_ONCE(p->dtime, now);
143 return p;
144 }
145 if (gc_stack) {
146 if (*gc_cnt < PEER_MAX_GC)
147 gc_stack[(*gc_cnt)++] = p;
148 } else if (unlikely(read_seqretry(&base->lock, seq))) {
149 break;
150 }
151 if (cmp == -1)
152 pp = &next->rb_left;
153 else
154 pp = &next->rb_right;
155 }
156 *parent_p = parent;
157 *pp_p = pp;
158 return NULL;
159}
160
161/* perform garbage collect on all items stacked during a lookup */
162static void inet_peer_gc(struct inet_peer_base *base,
163 struct inet_peer *gc_stack[],
164 unsigned int gc_cnt)
165{
166 int peer_threshold, peer_maxttl, peer_minttl;
167 struct inet_peer *p;
168 __u32 delta, ttl;
169 int i;
170
171 peer_threshold = READ_ONCE(inet_peer_threshold);
172 peer_maxttl = READ_ONCE(inet_peer_maxttl);
173 peer_minttl = READ_ONCE(inet_peer_minttl);
174
175 if (base->total >= peer_threshold)
176 ttl = 0; /* be aggressive */
177 else
178 ttl = peer_maxttl - (peer_maxttl - peer_minttl) / HZ *
179 base->total / peer_threshold * HZ;
180 for (i = 0; i < gc_cnt; i++) {
181 p = gc_stack[i];
182
183 delta = (__u32)jiffies - READ_ONCE(p->dtime);
184
185 if (delta < ttl || !refcount_dec_if_one(&p->refcnt))
186 gc_stack[i] = NULL;
187 }
188 for (i = 0; i < gc_cnt; i++) {
189 p = gc_stack[i];
190 if (p) {
191 rb_erase(&p->rb_node, &base->rb_root);
192 base->total--;
193 kfree_rcu(p, rcu);
194 }
195 }
196}
197
198/* Must be called under RCU : No refcount change is done here. */
199struct inet_peer *inet_getpeer(struct inet_peer_base *base,
200 const struct inetpeer_addr *daddr)
201{
202 struct inet_peer *p, *gc_stack[PEER_MAX_GC];
203 u64 dhash = inetpeer_addr_hash(daddr);
204 struct rb_node **pp, *parent;
205 unsigned int gc_cnt, seq;
206
207 /* Attempt a lockless lookup first.
208 * Because of a concurrent writer, we might not find an existing entry.
209 */
210 seq = read_seqbegin(&base->lock);
211 p = lookup(daddr, dhash, base, seq, NULL, &gc_cnt, &parent, &pp);
212
213 /* Make sure tree was not modified during our lookup. */
214 if (p && !read_seqretry(&base->lock, seq))
215 return p;
216
217 /* retry an exact lookup, taking the lock before.
218 * At least, nodes should be hot in our cache.
219 */
220 parent = NULL;
221 write_seqlock_bh(&base->lock);
222
223 gc_cnt = 0;
224 p = lookup(daddr, dhash, base, seq, gc_stack, &gc_cnt, &parent, &pp);
225 if (!p) {
226 p = kmem_cache_alloc(peer_cachep, GFP_ATOMIC);
227 if (p) {
228 p->daddr = *daddr;
229 p->hash = dhash;
230 p->dtime = (__u32)jiffies;
231 refcount_set(&p->refcnt, 1);
232 atomic_set(&p->rid, 0);
233 p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
234 p->rate_tokens = 0;
235 p->n_redirects = 0;
236 /* 60*HZ is arbitrary, but chosen enough high so that the first
237 * calculation of tokens is at its maximum.
238 */
239 p->rate_last = jiffies - 60*HZ;
240
241 rb_link_node(&p->rb_node, parent, pp);
242 rb_insert_color(&p->rb_node, &base->rb_root);
243 base->total++;
244 }
245 }
246 if (gc_cnt)
247 inet_peer_gc(base, gc_stack, gc_cnt);
248 write_sequnlock_bh(&base->lock);
249
250 return p;
251}
252
253void inet_putpeer(struct inet_peer *p)
254{
255 if (refcount_dec_and_test(&p->refcnt))
256 kfree_rcu(p, rcu);
257}
258
259/*
260 * Check transmit rate limitation for given message.
261 * The rate information is held in the inet_peer entries now.
262 * This function is generic and could be used for other purposes
263 * too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
264 *
265 * Note that the same inet_peer fields are modified by functions in
266 * route.c too, but these work for packet destinations while xrlim_allow
267 * works for icmp destinations. This means the rate limiting information
268 * for one "ip object" is shared - and these ICMPs are twice limited:
269 * by source and by destination.
270 *
271 * RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
272 * SHOULD allow setting of rate limits
273 *
274 * Shared between ICMPv4 and ICMPv6.
275 */
276#define XRLIM_BURST_FACTOR 6
277bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
278{
279 unsigned long now, token, otoken, delta;
280 bool rc = false;
281
282 if (!peer)
283 return true;
284
285 token = otoken = READ_ONCE(peer->rate_tokens);
286 now = jiffies;
287 delta = now - READ_ONCE(peer->rate_last);
288 if (delta) {
289 WRITE_ONCE(peer->rate_last, now);
290 token += delta;
291 if (token > XRLIM_BURST_FACTOR * timeout)
292 token = XRLIM_BURST_FACTOR * timeout;
293 }
294 if (token >= timeout) {
295 token -= timeout;
296 rc = true;
297 }
298 if (token != otoken)
299 WRITE_ONCE(peer->rate_tokens, token);
300 return rc;
301}
302
303void inetpeer_invalidate_tree(struct inet_peer_base *base)
304{
305 struct rb_node *p = rb_first(&base->rb_root);
306
307 while (p) {
308 struct inet_peer *peer = rb_entry(p, struct inet_peer, rb_node);
309
310 p = rb_next(p);
311 rb_erase(&peer->rb_node, &base->rb_root);
312 inet_putpeer(peer);
313 cond_resched();
314 }
315
316 base->total = 0;
317}