Loading...
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/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
25/*
26 * Theory of operations.
27 * We keep one entry for each peer IP address. The nodes contains long-living
28 * information about the peer which doesn't depend on routes.
29 * At this moment this information consists only of ID field for the next
30 * outgoing IP packet. This field is incremented with each packet as encoded
31 * in inet_getid() function (include/net/inetpeer.h).
32 * At the moment of writing this notes identifier of IP packets is generated
33 * to be unpredictable using this code only for packets subjected
34 * (actually or potentially) to defragmentation. I.e. DF packets less than
35 * PMTU in size uses a constant ID and do not use this code (see
36 * ip_select_ident() in include/net/ip.h).
37 *
38 * Route cache entries hold references to our nodes.
39 * New cache entries get references via lookup by destination IP address in
40 * the avl tree. The reference is grabbed only when it's needed i.e. only
41 * when we try to output IP packet which needs an unpredictable ID (see
42 * __ip_select_ident() in net/ipv4/route.c).
43 * Nodes are removed only when reference counter goes to 0.
44 * When it's happened the node may be removed when a sufficient amount of
45 * time has been passed since its last use. The less-recently-used entry can
46 * also be removed if the pool is overloaded i.e. if the total amount of
47 * entries is greater-or-equal than the threshold.
48 *
49 * Node pool is organised as an AVL tree.
50 * Such an implementation has been chosen not just for fun. It's a way to
51 * prevent easy and efficient DoS attacks by creating hash collisions. A huge
52 * amount of long living nodes in a single hash slot would significantly delay
53 * lookups performed with disabled BHs.
54 *
55 * Serialisation issues.
56 * 1. Nodes may appear in the tree only with the pool lock held.
57 * 2. Nodes may disappear from the tree only with the pool lock held
58 * AND reference count being 0.
59 * 3. Global variable peer_total is modified under the pool lock.
60 * 4. struct inet_peer fields modification:
61 * avl_left, avl_right, avl_parent, avl_height: pool lock
62 * refcnt: atomically against modifications on other CPU;
63 * usually under some other lock to prevent node disappearing
64 * daddr: unchangeable
65 * ip_id_count: atomic value (no lock needed)
66 */
67
68static struct kmem_cache *peer_cachep __read_mostly;
69
70static LIST_HEAD(gc_list);
71static const int gc_delay = 60 * HZ;
72static struct delayed_work gc_work;
73static DEFINE_SPINLOCK(gc_lock);
74
75#define node_height(x) x->avl_height
76
77#define peer_avl_empty ((struct inet_peer *)&peer_fake_node)
78#define peer_avl_empty_rcu ((struct inet_peer __rcu __force *)&peer_fake_node)
79static const struct inet_peer peer_fake_node = {
80 .avl_left = peer_avl_empty_rcu,
81 .avl_right = peer_avl_empty_rcu,
82 .avl_height = 0
83};
84
85struct inet_peer_base {
86 struct inet_peer __rcu *root;
87 seqlock_t lock;
88 int total;
89};
90
91static struct inet_peer_base v4_peers = {
92 .root = peer_avl_empty_rcu,
93 .lock = __SEQLOCK_UNLOCKED(v4_peers.lock),
94 .total = 0,
95};
96
97static struct inet_peer_base v6_peers = {
98 .root = peer_avl_empty_rcu,
99 .lock = __SEQLOCK_UNLOCKED(v6_peers.lock),
100 .total = 0,
101};
102
103#define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
104
105/* Exported for sysctl_net_ipv4. */
106int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more
107 * aggressively at this stage */
108int inet_peer_minttl __read_mostly = 120 * HZ; /* TTL under high load: 120 sec */
109int inet_peer_maxttl __read_mostly = 10 * 60 * HZ; /* usual time to live: 10 min */
110
111static void inetpeer_gc_worker(struct work_struct *work)
112{
113 struct inet_peer *p, *n;
114 LIST_HEAD(list);
115
116 spin_lock_bh(&gc_lock);
117 list_replace_init(&gc_list, &list);
118 spin_unlock_bh(&gc_lock);
119
120 if (list_empty(&list))
121 return;
122
123 list_for_each_entry_safe(p, n, &list, gc_list) {
124
125 if(need_resched())
126 cond_resched();
127
128 if (p->avl_left != peer_avl_empty) {
129 list_add_tail(&p->avl_left->gc_list, &list);
130 p->avl_left = peer_avl_empty;
131 }
132
133 if (p->avl_right != peer_avl_empty) {
134 list_add_tail(&p->avl_right->gc_list, &list);
135 p->avl_right = peer_avl_empty;
136 }
137
138 n = list_entry(p->gc_list.next, struct inet_peer, gc_list);
139
140 if (!atomic_read(&p->refcnt)) {
141 list_del(&p->gc_list);
142 kmem_cache_free(peer_cachep, p);
143 }
144 }
145
146 if (list_empty(&list))
147 return;
148
149 spin_lock_bh(&gc_lock);
150 list_splice(&list, &gc_list);
151 spin_unlock_bh(&gc_lock);
152
153 schedule_delayed_work(&gc_work, gc_delay);
154}
155
156/* Called from ip_output.c:ip_init */
157void __init inet_initpeers(void)
158{
159 struct sysinfo si;
160
161 /* Use the straight interface to information about memory. */
162 si_meminfo(&si);
163 /* The values below were suggested by Alexey Kuznetsov
164 * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values
165 * myself. --SAW
166 */
167 if (si.totalram <= (32768*1024)/PAGE_SIZE)
168 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
169 if (si.totalram <= (16384*1024)/PAGE_SIZE)
170 inet_peer_threshold >>= 1; /* about 512KB */
171 if (si.totalram <= (8192*1024)/PAGE_SIZE)
172 inet_peer_threshold >>= 2; /* about 128KB */
173
174 peer_cachep = kmem_cache_create("inet_peer_cache",
175 sizeof(struct inet_peer),
176 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
177 NULL);
178
179 INIT_DELAYED_WORK_DEFERRABLE(&gc_work, inetpeer_gc_worker);
180}
181
182static int addr_compare(const struct inetpeer_addr *a,
183 const struct inetpeer_addr *b)
184{
185 int i, n = (a->family == AF_INET ? 1 : 4);
186
187 for (i = 0; i < n; i++) {
188 if (a->addr.a6[i] == b->addr.a6[i])
189 continue;
190 if ((__force u32)a->addr.a6[i] < (__force u32)b->addr.a6[i])
191 return -1;
192 return 1;
193 }
194
195 return 0;
196}
197
198#define rcu_deref_locked(X, BASE) \
199 rcu_dereference_protected(X, lockdep_is_held(&(BASE)->lock.lock))
200
201/*
202 * Called with local BH disabled and the pool lock held.
203 */
204#define lookup(_daddr, _stack, _base) \
205({ \
206 struct inet_peer *u; \
207 struct inet_peer __rcu **v; \
208 \
209 stackptr = _stack; \
210 *stackptr++ = &_base->root; \
211 for (u = rcu_deref_locked(_base->root, _base); \
212 u != peer_avl_empty; ) { \
213 int cmp = addr_compare(_daddr, &u->daddr); \
214 if (cmp == 0) \
215 break; \
216 if (cmp == -1) \
217 v = &u->avl_left; \
218 else \
219 v = &u->avl_right; \
220 *stackptr++ = v; \
221 u = rcu_deref_locked(*v, _base); \
222 } \
223 u; \
224})
225
226/*
227 * Called with rcu_read_lock()
228 * Because we hold no lock against a writer, its quite possible we fall
229 * in an endless loop.
230 * But every pointer we follow is guaranteed to be valid thanks to RCU.
231 * We exit from this function if number of links exceeds PEER_MAXDEPTH
232 */
233static struct inet_peer *lookup_rcu(const struct inetpeer_addr *daddr,
234 struct inet_peer_base *base)
235{
236 struct inet_peer *u = rcu_dereference(base->root);
237 int count = 0;
238
239 while (u != peer_avl_empty) {
240 int cmp = addr_compare(daddr, &u->daddr);
241 if (cmp == 0) {
242 /* Before taking a reference, check if this entry was
243 * deleted (refcnt=-1)
244 */
245 if (!atomic_add_unless(&u->refcnt, 1, -1))
246 u = NULL;
247 return u;
248 }
249 if (cmp == -1)
250 u = rcu_dereference(u->avl_left);
251 else
252 u = rcu_dereference(u->avl_right);
253 if (unlikely(++count == PEER_MAXDEPTH))
254 break;
255 }
256 return NULL;
257}
258
259/* Called with local BH disabled and the pool lock held. */
260#define lookup_rightempty(start, base) \
261({ \
262 struct inet_peer *u; \
263 struct inet_peer __rcu **v; \
264 *stackptr++ = &start->avl_left; \
265 v = &start->avl_left; \
266 for (u = rcu_deref_locked(*v, base); \
267 u->avl_right != peer_avl_empty_rcu; ) { \
268 v = &u->avl_right; \
269 *stackptr++ = v; \
270 u = rcu_deref_locked(*v, base); \
271 } \
272 u; \
273})
274
275/* Called with local BH disabled and the pool lock held.
276 * Variable names are the proof of operation correctness.
277 * Look into mm/map_avl.c for more detail description of the ideas.
278 */
279static void peer_avl_rebalance(struct inet_peer __rcu **stack[],
280 struct inet_peer __rcu ***stackend,
281 struct inet_peer_base *base)
282{
283 struct inet_peer __rcu **nodep;
284 struct inet_peer *node, *l, *r;
285 int lh, rh;
286
287 while (stackend > stack) {
288 nodep = *--stackend;
289 node = rcu_deref_locked(*nodep, base);
290 l = rcu_deref_locked(node->avl_left, base);
291 r = rcu_deref_locked(node->avl_right, base);
292 lh = node_height(l);
293 rh = node_height(r);
294 if (lh > rh + 1) { /* l: RH+2 */
295 struct inet_peer *ll, *lr, *lrl, *lrr;
296 int lrh;
297 ll = rcu_deref_locked(l->avl_left, base);
298 lr = rcu_deref_locked(l->avl_right, base);
299 lrh = node_height(lr);
300 if (lrh <= node_height(ll)) { /* ll: RH+1 */
301 RCU_INIT_POINTER(node->avl_left, lr); /* lr: RH or RH+1 */
302 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */
303 node->avl_height = lrh + 1; /* RH+1 or RH+2 */
304 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH+1 */
305 RCU_INIT_POINTER(l->avl_right, node); /* node: RH+1 or RH+2 */
306 l->avl_height = node->avl_height + 1;
307 RCU_INIT_POINTER(*nodep, l);
308 } else { /* ll: RH, lr: RH+1 */
309 lrl = rcu_deref_locked(lr->avl_left, base);/* lrl: RH or RH-1 */
310 lrr = rcu_deref_locked(lr->avl_right, base);/* lrr: RH or RH-1 */
311 RCU_INIT_POINTER(node->avl_left, lrr); /* lrr: RH or RH-1 */
312 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */
313 node->avl_height = rh + 1; /* node: RH+1 */
314 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH */
315 RCU_INIT_POINTER(l->avl_right, lrl); /* lrl: RH or RH-1 */
316 l->avl_height = rh + 1; /* l: RH+1 */
317 RCU_INIT_POINTER(lr->avl_left, l); /* l: RH+1 */
318 RCU_INIT_POINTER(lr->avl_right, node); /* node: RH+1 */
319 lr->avl_height = rh + 2;
320 RCU_INIT_POINTER(*nodep, lr);
321 }
322 } else if (rh > lh + 1) { /* r: LH+2 */
323 struct inet_peer *rr, *rl, *rlr, *rll;
324 int rlh;
325 rr = rcu_deref_locked(r->avl_right, base);
326 rl = rcu_deref_locked(r->avl_left, base);
327 rlh = node_height(rl);
328 if (rlh <= node_height(rr)) { /* rr: LH+1 */
329 RCU_INIT_POINTER(node->avl_right, rl); /* rl: LH or LH+1 */
330 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */
331 node->avl_height = rlh + 1; /* LH+1 or LH+2 */
332 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH+1 */
333 RCU_INIT_POINTER(r->avl_left, node); /* node: LH+1 or LH+2 */
334 r->avl_height = node->avl_height + 1;
335 RCU_INIT_POINTER(*nodep, r);
336 } else { /* rr: RH, rl: RH+1 */
337 rlr = rcu_deref_locked(rl->avl_right, base);/* rlr: LH or LH-1 */
338 rll = rcu_deref_locked(rl->avl_left, base);/* rll: LH or LH-1 */
339 RCU_INIT_POINTER(node->avl_right, rll); /* rll: LH or LH-1 */
340 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */
341 node->avl_height = lh + 1; /* node: LH+1 */
342 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH */
343 RCU_INIT_POINTER(r->avl_left, rlr); /* rlr: LH or LH-1 */
344 r->avl_height = lh + 1; /* r: LH+1 */
345 RCU_INIT_POINTER(rl->avl_right, r); /* r: LH+1 */
346 RCU_INIT_POINTER(rl->avl_left, node); /* node: LH+1 */
347 rl->avl_height = lh + 2;
348 RCU_INIT_POINTER(*nodep, rl);
349 }
350 } else {
351 node->avl_height = (lh > rh ? lh : rh) + 1;
352 }
353 }
354}
355
356/* Called with local BH disabled and the pool lock held. */
357#define link_to_pool(n, base) \
358do { \
359 n->avl_height = 1; \
360 n->avl_left = peer_avl_empty_rcu; \
361 n->avl_right = peer_avl_empty_rcu; \
362 /* lockless readers can catch us now */ \
363 rcu_assign_pointer(**--stackptr, n); \
364 peer_avl_rebalance(stack, stackptr, base); \
365} while (0)
366
367static void inetpeer_free_rcu(struct rcu_head *head)
368{
369 kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
370}
371
372static void unlink_from_pool(struct inet_peer *p, struct inet_peer_base *base,
373 struct inet_peer __rcu **stack[PEER_MAXDEPTH])
374{
375 struct inet_peer __rcu ***stackptr, ***delp;
376
377 if (lookup(&p->daddr, stack, base) != p)
378 BUG();
379 delp = stackptr - 1; /* *delp[0] == p */
380 if (p->avl_left == peer_avl_empty_rcu) {
381 *delp[0] = p->avl_right;
382 --stackptr;
383 } else {
384 /* look for a node to insert instead of p */
385 struct inet_peer *t;
386 t = lookup_rightempty(p, base);
387 BUG_ON(rcu_deref_locked(*stackptr[-1], base) != t);
388 **--stackptr = t->avl_left;
389 /* t is removed, t->daddr > x->daddr for any
390 * x in p->avl_left subtree.
391 * Put t in the old place of p. */
392 RCU_INIT_POINTER(*delp[0], t);
393 t->avl_left = p->avl_left;
394 t->avl_right = p->avl_right;
395 t->avl_height = p->avl_height;
396 BUG_ON(delp[1] != &p->avl_left);
397 delp[1] = &t->avl_left; /* was &p->avl_left */
398 }
399 peer_avl_rebalance(stack, stackptr, base);
400 base->total--;
401 call_rcu(&p->rcu, inetpeer_free_rcu);
402}
403
404static struct inet_peer_base *family_to_base(int family)
405{
406 return family == AF_INET ? &v4_peers : &v6_peers;
407}
408
409/* perform garbage collect on all items stacked during a lookup */
410static int inet_peer_gc(struct inet_peer_base *base,
411 struct inet_peer __rcu **stack[PEER_MAXDEPTH],
412 struct inet_peer __rcu ***stackptr)
413{
414 struct inet_peer *p, *gchead = NULL;
415 __u32 delta, ttl;
416 int cnt = 0;
417
418 if (base->total >= inet_peer_threshold)
419 ttl = 0; /* be aggressive */
420 else
421 ttl = inet_peer_maxttl
422 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
423 base->total / inet_peer_threshold * HZ;
424 stackptr--; /* last stack slot is peer_avl_empty */
425 while (stackptr > stack) {
426 stackptr--;
427 p = rcu_deref_locked(**stackptr, base);
428 if (atomic_read(&p->refcnt) == 0) {
429 smp_rmb();
430 delta = (__u32)jiffies - p->dtime;
431 if (delta >= ttl &&
432 atomic_cmpxchg(&p->refcnt, 0, -1) == 0) {
433 p->gc_next = gchead;
434 gchead = p;
435 }
436 }
437 }
438 while ((p = gchead) != NULL) {
439 gchead = p->gc_next;
440 cnt++;
441 unlink_from_pool(p, base, stack);
442 }
443 return cnt;
444}
445
446struct inet_peer *inet_getpeer(const struct inetpeer_addr *daddr, int create)
447{
448 struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr;
449 struct inet_peer_base *base = family_to_base(daddr->family);
450 struct inet_peer *p;
451 unsigned int sequence;
452 int invalidated, gccnt = 0;
453
454 /* Attempt a lockless lookup first.
455 * Because of a concurrent writer, we might not find an existing entry.
456 */
457 rcu_read_lock();
458 sequence = read_seqbegin(&base->lock);
459 p = lookup_rcu(daddr, base);
460 invalidated = read_seqretry(&base->lock, sequence);
461 rcu_read_unlock();
462
463 if (p)
464 return p;
465
466 /* If no writer did a change during our lookup, we can return early. */
467 if (!create && !invalidated)
468 return NULL;
469
470 /* retry an exact lookup, taking the lock before.
471 * At least, nodes should be hot in our cache.
472 */
473 write_seqlock_bh(&base->lock);
474relookup:
475 p = lookup(daddr, stack, base);
476 if (p != peer_avl_empty) {
477 atomic_inc(&p->refcnt);
478 write_sequnlock_bh(&base->lock);
479 return p;
480 }
481 if (!gccnt) {
482 gccnt = inet_peer_gc(base, stack, stackptr);
483 if (gccnt && create)
484 goto relookup;
485 }
486 p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
487 if (p) {
488 p->daddr = *daddr;
489 atomic_set(&p->refcnt, 1);
490 atomic_set(&p->rid, 0);
491 atomic_set(&p->ip_id_count,
492 (daddr->family == AF_INET) ?
493 secure_ip_id(daddr->addr.a4) :
494 secure_ipv6_id(daddr->addr.a6));
495 p->tcp_ts_stamp = 0;
496 p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
497 p->rate_tokens = 0;
498 p->rate_last = 0;
499 p->pmtu_expires = 0;
500 p->pmtu_orig = 0;
501 memset(&p->redirect_learned, 0, sizeof(p->redirect_learned));
502 INIT_LIST_HEAD(&p->gc_list);
503
504 /* Link the node. */
505 link_to_pool(p, base);
506 base->total++;
507 }
508 write_sequnlock_bh(&base->lock);
509
510 return p;
511}
512EXPORT_SYMBOL_GPL(inet_getpeer);
513
514void inet_putpeer(struct inet_peer *p)
515{
516 p->dtime = (__u32)jiffies;
517 smp_mb__before_atomic_dec();
518 atomic_dec(&p->refcnt);
519}
520EXPORT_SYMBOL_GPL(inet_putpeer);
521
522/*
523 * Check transmit rate limitation for given message.
524 * The rate information is held in the inet_peer entries now.
525 * This function is generic and could be used for other purposes
526 * too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
527 *
528 * Note that the same inet_peer fields are modified by functions in
529 * route.c too, but these work for packet destinations while xrlim_allow
530 * works for icmp destinations. This means the rate limiting information
531 * for one "ip object" is shared - and these ICMPs are twice limited:
532 * by source and by destination.
533 *
534 * RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
535 * SHOULD allow setting of rate limits
536 *
537 * Shared between ICMPv4 and ICMPv6.
538 */
539#define XRLIM_BURST_FACTOR 6
540bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
541{
542 unsigned long now, token;
543 bool rc = false;
544
545 if (!peer)
546 return true;
547
548 token = peer->rate_tokens;
549 now = jiffies;
550 token += now - peer->rate_last;
551 peer->rate_last = now;
552 if (token > XRLIM_BURST_FACTOR * timeout)
553 token = XRLIM_BURST_FACTOR * timeout;
554 if (token >= timeout) {
555 token -= timeout;
556 rc = true;
557 }
558 peer->rate_tokens = token;
559 return rc;
560}
561EXPORT_SYMBOL(inet_peer_xrlim_allow);
562
563static void inetpeer_inval_rcu(struct rcu_head *head)
564{
565 struct inet_peer *p = container_of(head, struct inet_peer, gc_rcu);
566
567 spin_lock_bh(&gc_lock);
568 list_add_tail(&p->gc_list, &gc_list);
569 spin_unlock_bh(&gc_lock);
570
571 schedule_delayed_work(&gc_work, gc_delay);
572}
573
574void inetpeer_invalidate_tree(int family)
575{
576 struct inet_peer *old, *new, *prev;
577 struct inet_peer_base *base = family_to_base(family);
578
579 write_seqlock_bh(&base->lock);
580
581 old = base->root;
582 if (old == peer_avl_empty_rcu)
583 goto out;
584
585 new = peer_avl_empty_rcu;
586
587 prev = cmpxchg(&base->root, old, new);
588 if (prev == old) {
589 base->total = 0;
590 call_rcu(&prev->gc_rcu, inetpeer_inval_rcu);
591 }
592
593out:
594 write_sequnlock_bh(&base->lock);
595}
596EXPORT_SYMBOL(inetpeer_invalidate_tree);
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/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
25/*
26 * Theory of operations.
27 * We keep one entry for each peer IP address. The nodes contains long-living
28 * information about the peer which doesn't depend on routes.
29 *
30 * Nodes are removed only when reference counter goes to 0.
31 * When it's happened the node may be removed when a sufficient amount of
32 * time has been passed since its last use. The less-recently-used entry can
33 * also be removed if the pool is overloaded i.e. if the total amount of
34 * entries is greater-or-equal than the threshold.
35 *
36 * Node pool is organised as an AVL tree.
37 * Such an implementation has been chosen not just for fun. It's a way to
38 * prevent easy and efficient DoS attacks by creating hash collisions. A huge
39 * amount of long living nodes in a single hash slot would significantly delay
40 * lookups performed with disabled BHs.
41 *
42 * Serialisation issues.
43 * 1. Nodes may appear in the tree only with the pool lock held.
44 * 2. Nodes may disappear from the tree only with the pool lock held
45 * AND reference count being 0.
46 * 3. Global variable peer_total is modified under the pool lock.
47 * 4. struct inet_peer fields modification:
48 * avl_left, avl_right, avl_parent, avl_height: pool lock
49 * refcnt: atomically against modifications on other CPU;
50 * usually under some other lock to prevent node disappearing
51 * daddr: unchangeable
52 */
53
54static struct kmem_cache *peer_cachep __read_mostly;
55
56static LIST_HEAD(gc_list);
57static const int gc_delay = 60 * HZ;
58static struct delayed_work gc_work;
59static DEFINE_SPINLOCK(gc_lock);
60
61#define node_height(x) x->avl_height
62
63#define peer_avl_empty ((struct inet_peer *)&peer_fake_node)
64#define peer_avl_empty_rcu ((struct inet_peer __rcu __force *)&peer_fake_node)
65static const struct inet_peer peer_fake_node = {
66 .avl_left = peer_avl_empty_rcu,
67 .avl_right = peer_avl_empty_rcu,
68 .avl_height = 0
69};
70
71void inet_peer_base_init(struct inet_peer_base *bp)
72{
73 bp->root = peer_avl_empty_rcu;
74 seqlock_init(&bp->lock);
75 bp->total = 0;
76}
77EXPORT_SYMBOL_GPL(inet_peer_base_init);
78
79#define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
80
81/* Exported for sysctl_net_ipv4. */
82int inet_peer_threshold __read_mostly = 65536 + 128; /* start to throw entries more
83 * aggressively at this stage */
84int inet_peer_minttl __read_mostly = 120 * HZ; /* TTL under high load: 120 sec */
85int inet_peer_maxttl __read_mostly = 10 * 60 * HZ; /* usual time to live: 10 min */
86
87static void inetpeer_gc_worker(struct work_struct *work)
88{
89 struct inet_peer *p, *n, *c;
90 struct list_head list;
91
92 spin_lock_bh(&gc_lock);
93 list_replace_init(&gc_list, &list);
94 spin_unlock_bh(&gc_lock);
95
96 if (list_empty(&list))
97 return;
98
99 list_for_each_entry_safe(p, n, &list, gc_list) {
100
101 if (need_resched())
102 cond_resched();
103
104 c = rcu_dereference_protected(p->avl_left, 1);
105 if (c != peer_avl_empty) {
106 list_add_tail(&c->gc_list, &list);
107 p->avl_left = peer_avl_empty_rcu;
108 }
109
110 c = rcu_dereference_protected(p->avl_right, 1);
111 if (c != peer_avl_empty) {
112 list_add_tail(&c->gc_list, &list);
113 p->avl_right = peer_avl_empty_rcu;
114 }
115
116 n = list_entry(p->gc_list.next, struct inet_peer, gc_list);
117
118 if (!atomic_read(&p->refcnt)) {
119 list_del(&p->gc_list);
120 kmem_cache_free(peer_cachep, p);
121 }
122 }
123
124 if (list_empty(&list))
125 return;
126
127 spin_lock_bh(&gc_lock);
128 list_splice(&list, &gc_list);
129 spin_unlock_bh(&gc_lock);
130
131 schedule_delayed_work(&gc_work, gc_delay);
132}
133
134/* Called from ip_output.c:ip_init */
135void __init inet_initpeers(void)
136{
137 struct sysinfo si;
138
139 /* Use the straight interface to information about memory. */
140 si_meminfo(&si);
141 /* The values below were suggested by Alexey Kuznetsov
142 * <kuznet@ms2.inr.ac.ru>. I don't have any opinion about the values
143 * myself. --SAW
144 */
145 if (si.totalram <= (32768*1024)/PAGE_SIZE)
146 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
147 if (si.totalram <= (16384*1024)/PAGE_SIZE)
148 inet_peer_threshold >>= 1; /* about 512KB */
149 if (si.totalram <= (8192*1024)/PAGE_SIZE)
150 inet_peer_threshold >>= 2; /* about 128KB */
151
152 peer_cachep = kmem_cache_create("inet_peer_cache",
153 sizeof(struct inet_peer),
154 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
155 NULL);
156
157 INIT_DEFERRABLE_WORK(&gc_work, inetpeer_gc_worker);
158}
159
160#define rcu_deref_locked(X, BASE) \
161 rcu_dereference_protected(X, lockdep_is_held(&(BASE)->lock.lock))
162
163/*
164 * Called with local BH disabled and the pool lock held.
165 */
166#define lookup(_daddr, _stack, _base) \
167({ \
168 struct inet_peer *u; \
169 struct inet_peer __rcu **v; \
170 \
171 stackptr = _stack; \
172 *stackptr++ = &_base->root; \
173 for (u = rcu_deref_locked(_base->root, _base); \
174 u != peer_avl_empty;) { \
175 int cmp = inetpeer_addr_cmp(_daddr, &u->daddr); \
176 if (cmp == 0) \
177 break; \
178 if (cmp == -1) \
179 v = &u->avl_left; \
180 else \
181 v = &u->avl_right; \
182 *stackptr++ = v; \
183 u = rcu_deref_locked(*v, _base); \
184 } \
185 u; \
186})
187
188/*
189 * Called with rcu_read_lock()
190 * Because we hold no lock against a writer, its quite possible we fall
191 * in an endless loop.
192 * But every pointer we follow is guaranteed to be valid thanks to RCU.
193 * We exit from this function if number of links exceeds PEER_MAXDEPTH
194 */
195static struct inet_peer *lookup_rcu(const struct inetpeer_addr *daddr,
196 struct inet_peer_base *base)
197{
198 struct inet_peer *u = rcu_dereference(base->root);
199 int count = 0;
200
201 while (u != peer_avl_empty) {
202 int cmp = inetpeer_addr_cmp(daddr, &u->daddr);
203 if (cmp == 0) {
204 /* Before taking a reference, check if this entry was
205 * deleted (refcnt=-1)
206 */
207 if (!atomic_add_unless(&u->refcnt, 1, -1))
208 u = NULL;
209 return u;
210 }
211 if (cmp == -1)
212 u = rcu_dereference(u->avl_left);
213 else
214 u = rcu_dereference(u->avl_right);
215 if (unlikely(++count == PEER_MAXDEPTH))
216 break;
217 }
218 return NULL;
219}
220
221/* Called with local BH disabled and the pool lock held. */
222#define lookup_rightempty(start, base) \
223({ \
224 struct inet_peer *u; \
225 struct inet_peer __rcu **v; \
226 *stackptr++ = &start->avl_left; \
227 v = &start->avl_left; \
228 for (u = rcu_deref_locked(*v, base); \
229 u->avl_right != peer_avl_empty_rcu;) { \
230 v = &u->avl_right; \
231 *stackptr++ = v; \
232 u = rcu_deref_locked(*v, base); \
233 } \
234 u; \
235})
236
237/* Called with local BH disabled and the pool lock held.
238 * Variable names are the proof of operation correctness.
239 * Look into mm/map_avl.c for more detail description of the ideas.
240 */
241static void peer_avl_rebalance(struct inet_peer __rcu **stack[],
242 struct inet_peer __rcu ***stackend,
243 struct inet_peer_base *base)
244{
245 struct inet_peer __rcu **nodep;
246 struct inet_peer *node, *l, *r;
247 int lh, rh;
248
249 while (stackend > stack) {
250 nodep = *--stackend;
251 node = rcu_deref_locked(*nodep, base);
252 l = rcu_deref_locked(node->avl_left, base);
253 r = rcu_deref_locked(node->avl_right, base);
254 lh = node_height(l);
255 rh = node_height(r);
256 if (lh > rh + 1) { /* l: RH+2 */
257 struct inet_peer *ll, *lr, *lrl, *lrr;
258 int lrh;
259 ll = rcu_deref_locked(l->avl_left, base);
260 lr = rcu_deref_locked(l->avl_right, base);
261 lrh = node_height(lr);
262 if (lrh <= node_height(ll)) { /* ll: RH+1 */
263 RCU_INIT_POINTER(node->avl_left, lr); /* lr: RH or RH+1 */
264 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */
265 node->avl_height = lrh + 1; /* RH+1 or RH+2 */
266 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH+1 */
267 RCU_INIT_POINTER(l->avl_right, node); /* node: RH+1 or RH+2 */
268 l->avl_height = node->avl_height + 1;
269 RCU_INIT_POINTER(*nodep, l);
270 } else { /* ll: RH, lr: RH+1 */
271 lrl = rcu_deref_locked(lr->avl_left, base);/* lrl: RH or RH-1 */
272 lrr = rcu_deref_locked(lr->avl_right, base);/* lrr: RH or RH-1 */
273 RCU_INIT_POINTER(node->avl_left, lrr); /* lrr: RH or RH-1 */
274 RCU_INIT_POINTER(node->avl_right, r); /* r: RH */
275 node->avl_height = rh + 1; /* node: RH+1 */
276 RCU_INIT_POINTER(l->avl_left, ll); /* ll: RH */
277 RCU_INIT_POINTER(l->avl_right, lrl); /* lrl: RH or RH-1 */
278 l->avl_height = rh + 1; /* l: RH+1 */
279 RCU_INIT_POINTER(lr->avl_left, l); /* l: RH+1 */
280 RCU_INIT_POINTER(lr->avl_right, node); /* node: RH+1 */
281 lr->avl_height = rh + 2;
282 RCU_INIT_POINTER(*nodep, lr);
283 }
284 } else if (rh > lh + 1) { /* r: LH+2 */
285 struct inet_peer *rr, *rl, *rlr, *rll;
286 int rlh;
287 rr = rcu_deref_locked(r->avl_right, base);
288 rl = rcu_deref_locked(r->avl_left, base);
289 rlh = node_height(rl);
290 if (rlh <= node_height(rr)) { /* rr: LH+1 */
291 RCU_INIT_POINTER(node->avl_right, rl); /* rl: LH or LH+1 */
292 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */
293 node->avl_height = rlh + 1; /* LH+1 or LH+2 */
294 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH+1 */
295 RCU_INIT_POINTER(r->avl_left, node); /* node: LH+1 or LH+2 */
296 r->avl_height = node->avl_height + 1;
297 RCU_INIT_POINTER(*nodep, r);
298 } else { /* rr: RH, rl: RH+1 */
299 rlr = rcu_deref_locked(rl->avl_right, base);/* rlr: LH or LH-1 */
300 rll = rcu_deref_locked(rl->avl_left, base);/* rll: LH or LH-1 */
301 RCU_INIT_POINTER(node->avl_right, rll); /* rll: LH or LH-1 */
302 RCU_INIT_POINTER(node->avl_left, l); /* l: LH */
303 node->avl_height = lh + 1; /* node: LH+1 */
304 RCU_INIT_POINTER(r->avl_right, rr); /* rr: LH */
305 RCU_INIT_POINTER(r->avl_left, rlr); /* rlr: LH or LH-1 */
306 r->avl_height = lh + 1; /* r: LH+1 */
307 RCU_INIT_POINTER(rl->avl_right, r); /* r: LH+1 */
308 RCU_INIT_POINTER(rl->avl_left, node); /* node: LH+1 */
309 rl->avl_height = lh + 2;
310 RCU_INIT_POINTER(*nodep, rl);
311 }
312 } else {
313 node->avl_height = (lh > rh ? lh : rh) + 1;
314 }
315 }
316}
317
318/* Called with local BH disabled and the pool lock held. */
319#define link_to_pool(n, base) \
320do { \
321 n->avl_height = 1; \
322 n->avl_left = peer_avl_empty_rcu; \
323 n->avl_right = peer_avl_empty_rcu; \
324 /* lockless readers can catch us now */ \
325 rcu_assign_pointer(**--stackptr, n); \
326 peer_avl_rebalance(stack, stackptr, base); \
327} while (0)
328
329static void inetpeer_free_rcu(struct rcu_head *head)
330{
331 kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
332}
333
334static void unlink_from_pool(struct inet_peer *p, struct inet_peer_base *base,
335 struct inet_peer __rcu **stack[PEER_MAXDEPTH])
336{
337 struct inet_peer __rcu ***stackptr, ***delp;
338
339 if (lookup(&p->daddr, stack, base) != p)
340 BUG();
341 delp = stackptr - 1; /* *delp[0] == p */
342 if (p->avl_left == peer_avl_empty_rcu) {
343 *delp[0] = p->avl_right;
344 --stackptr;
345 } else {
346 /* look for a node to insert instead of p */
347 struct inet_peer *t;
348 t = lookup_rightempty(p, base);
349 BUG_ON(rcu_deref_locked(*stackptr[-1], base) != t);
350 **--stackptr = t->avl_left;
351 /* t is removed, t->daddr > x->daddr for any
352 * x in p->avl_left subtree.
353 * Put t in the old place of p. */
354 RCU_INIT_POINTER(*delp[0], t);
355 t->avl_left = p->avl_left;
356 t->avl_right = p->avl_right;
357 t->avl_height = p->avl_height;
358 BUG_ON(delp[1] != &p->avl_left);
359 delp[1] = &t->avl_left; /* was &p->avl_left */
360 }
361 peer_avl_rebalance(stack, stackptr, base);
362 base->total--;
363 call_rcu(&p->rcu, inetpeer_free_rcu);
364}
365
366/* perform garbage collect on all items stacked during a lookup */
367static int inet_peer_gc(struct inet_peer_base *base,
368 struct inet_peer __rcu **stack[PEER_MAXDEPTH],
369 struct inet_peer __rcu ***stackptr)
370{
371 struct inet_peer *p, *gchead = NULL;
372 __u32 delta, ttl;
373 int cnt = 0;
374
375 if (base->total >= inet_peer_threshold)
376 ttl = 0; /* be aggressive */
377 else
378 ttl = inet_peer_maxttl
379 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
380 base->total / inet_peer_threshold * HZ;
381 stackptr--; /* last stack slot is peer_avl_empty */
382 while (stackptr > stack) {
383 stackptr--;
384 p = rcu_deref_locked(**stackptr, base);
385 if (atomic_read(&p->refcnt) == 0) {
386 smp_rmb();
387 delta = (__u32)jiffies - p->dtime;
388 if (delta >= ttl &&
389 atomic_cmpxchg(&p->refcnt, 0, -1) == 0) {
390 p->gc_next = gchead;
391 gchead = p;
392 }
393 }
394 }
395 while ((p = gchead) != NULL) {
396 gchead = p->gc_next;
397 cnt++;
398 unlink_from_pool(p, base, stack);
399 }
400 return cnt;
401}
402
403struct inet_peer *inet_getpeer(struct inet_peer_base *base,
404 const struct inetpeer_addr *daddr,
405 int create)
406{
407 struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr;
408 struct inet_peer *p;
409 unsigned int sequence;
410 int invalidated, gccnt = 0;
411
412 /* Attempt a lockless lookup first.
413 * Because of a concurrent writer, we might not find an existing entry.
414 */
415 rcu_read_lock();
416 sequence = read_seqbegin(&base->lock);
417 p = lookup_rcu(daddr, base);
418 invalidated = read_seqretry(&base->lock, sequence);
419 rcu_read_unlock();
420
421 if (p)
422 return p;
423
424 /* If no writer did a change during our lookup, we can return early. */
425 if (!create && !invalidated)
426 return NULL;
427
428 /* retry an exact lookup, taking the lock before.
429 * At least, nodes should be hot in our cache.
430 */
431 write_seqlock_bh(&base->lock);
432relookup:
433 p = lookup(daddr, stack, base);
434 if (p != peer_avl_empty) {
435 atomic_inc(&p->refcnt);
436 write_sequnlock_bh(&base->lock);
437 return p;
438 }
439 if (!gccnt) {
440 gccnt = inet_peer_gc(base, stack, stackptr);
441 if (gccnt && create)
442 goto relookup;
443 }
444 p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
445 if (p) {
446 p->daddr = *daddr;
447 atomic_set(&p->refcnt, 1);
448 atomic_set(&p->rid, 0);
449 p->metrics[RTAX_LOCK-1] = INETPEER_METRICS_NEW;
450 p->rate_tokens = 0;
451 /* 60*HZ is arbitrary, but chosen enough high so that the first
452 * calculation of tokens is at its maximum.
453 */
454 p->rate_last = jiffies - 60*HZ;
455 INIT_LIST_HEAD(&p->gc_list);
456
457 /* Link the node. */
458 link_to_pool(p, base);
459 base->total++;
460 }
461 write_sequnlock_bh(&base->lock);
462
463 return p;
464}
465EXPORT_SYMBOL_GPL(inet_getpeer);
466
467void inet_putpeer(struct inet_peer *p)
468{
469 p->dtime = (__u32)jiffies;
470 smp_mb__before_atomic();
471 atomic_dec(&p->refcnt);
472}
473EXPORT_SYMBOL_GPL(inet_putpeer);
474
475/*
476 * Check transmit rate limitation for given message.
477 * The rate information is held in the inet_peer entries now.
478 * This function is generic and could be used for other purposes
479 * too. It uses a Token bucket filter as suggested by Alexey Kuznetsov.
480 *
481 * Note that the same inet_peer fields are modified by functions in
482 * route.c too, but these work for packet destinations while xrlim_allow
483 * works for icmp destinations. This means the rate limiting information
484 * for one "ip object" is shared - and these ICMPs are twice limited:
485 * by source and by destination.
486 *
487 * RFC 1812: 4.3.2.8 SHOULD be able to limit error message rate
488 * SHOULD allow setting of rate limits
489 *
490 * Shared between ICMPv4 and ICMPv6.
491 */
492#define XRLIM_BURST_FACTOR 6
493bool inet_peer_xrlim_allow(struct inet_peer *peer, int timeout)
494{
495 unsigned long now, token;
496 bool rc = false;
497
498 if (!peer)
499 return true;
500
501 token = peer->rate_tokens;
502 now = jiffies;
503 token += now - peer->rate_last;
504 peer->rate_last = now;
505 if (token > XRLIM_BURST_FACTOR * timeout)
506 token = XRLIM_BURST_FACTOR * timeout;
507 if (token >= timeout) {
508 token -= timeout;
509 rc = true;
510 }
511 peer->rate_tokens = token;
512 return rc;
513}
514EXPORT_SYMBOL(inet_peer_xrlim_allow);
515
516static void inetpeer_inval_rcu(struct rcu_head *head)
517{
518 struct inet_peer *p = container_of(head, struct inet_peer, gc_rcu);
519
520 spin_lock_bh(&gc_lock);
521 list_add_tail(&p->gc_list, &gc_list);
522 spin_unlock_bh(&gc_lock);
523
524 schedule_delayed_work(&gc_work, gc_delay);
525}
526
527void inetpeer_invalidate_tree(struct inet_peer_base *base)
528{
529 struct inet_peer *root;
530
531 write_seqlock_bh(&base->lock);
532
533 root = rcu_deref_locked(base->root, base);
534 if (root != peer_avl_empty) {
535 base->root = peer_avl_empty_rcu;
536 base->total = 0;
537 call_rcu(&root->gc_rcu, inetpeer_inval_rcu);
538 }
539
540 write_sequnlock_bh(&base->lock);
541}
542EXPORT_SYMBOL(inetpeer_invalidate_tree);