inetpeer: Abstract out the tree root accesses.
[pandora-kernel.git] / net / ipv4 / inetpeer.c
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 <net/ip.h>
21 #include <net/inetpeer.h>
22
23 /*
24  *  Theory of operations.
25  *  We keep one entry for each peer IP address.  The nodes contains long-living
26  *  information about the peer which doesn't depend on routes.
27  *  At this moment this information consists only of ID field for the next
28  *  outgoing IP packet.  This field is incremented with each packet as encoded
29  *  in inet_getid() function (include/net/inetpeer.h).
30  *  At the moment of writing this notes identifier of IP packets is generated
31  *  to be unpredictable using this code only for packets subjected
32  *  (actually or potentially) to defragmentation.  I.e. DF packets less than
33  *  PMTU in size uses a constant ID and do not use this code (see
34  *  ip_select_ident() in include/net/ip.h).
35  *
36  *  Route cache entries hold references to our nodes.
37  *  New cache entries get references via lookup by destination IP address in
38  *  the avl tree.  The reference is grabbed only when it's needed i.e. only
39  *  when we try to output IP packet which needs an unpredictable ID (see
40  *  __ip_select_ident() in net/ipv4/route.c).
41  *  Nodes are removed only when reference counter goes to 0.
42  *  When it's happened the node may be removed when a sufficient amount of
43  *  time has been passed since its last use.  The less-recently-used entry can
44  *  also be removed if the pool is overloaded i.e. if the total amount of
45  *  entries is greater-or-equal than the threshold.
46  *
47  *  Node pool is organised as an AVL tree.
48  *  Such an implementation has been chosen not just for fun.  It's a way to
49  *  prevent easy and efficient DoS attacks by creating hash collisions.  A huge
50  *  amount of long living nodes in a single hash slot would significantly delay
51  *  lookups performed with disabled BHs.
52  *
53  *  Serialisation issues.
54  *  1.  Nodes may appear in the tree only with the pool lock held.
55  *  2.  Nodes may disappear from the tree only with the pool lock held
56  *      AND reference count being 0.
57  *  3.  Nodes appears and disappears from unused node list only under
58  *      "inet_peer_unused_lock".
59  *  4.  Global variable peer_total is modified under the pool lock.
60  *  5.  struct inet_peer fields modification:
61  *              avl_left, avl_right, avl_parent, avl_height: pool lock
62  *              unused: unused node list lock
63  *              refcnt: atomically against modifications on other CPU;
64  *                 usually under some other lock to prevent node disappearing
65  *              dtime: unused node list lock
66  *              v4daddr: unchangeable
67  *              ip_id_count: atomic value (no lock needed)
68  */
69
70 static struct kmem_cache *peer_cachep __read_mostly;
71
72 #define node_height(x) x->avl_height
73
74 #define peer_avl_empty ((struct inet_peer *)&peer_fake_node)
75 #define peer_avl_empty_rcu ((struct inet_peer __rcu __force *)&peer_fake_node)
76 static const struct inet_peer peer_fake_node = {
77         .avl_left       = peer_avl_empty_rcu,
78         .avl_right      = peer_avl_empty_rcu,
79         .avl_height     = 0
80 };
81
82 static struct inet_peer_base {
83         struct inet_peer __rcu *root;
84         spinlock_t      lock;
85         int             total;
86 } v4_peers = {
87         .root           = peer_avl_empty_rcu,
88         .lock           = __SPIN_LOCK_UNLOCKED(v4_peers.lock),
89         .total          = 0,
90 };
91 #define PEER_MAXDEPTH 40 /* sufficient for about 2^27 nodes */
92
93 /* Exported for sysctl_net_ipv4.  */
94 int inet_peer_threshold __read_mostly = 65536 + 128;    /* start to throw entries more
95                                          * aggressively at this stage */
96 int inet_peer_minttl __read_mostly = 120 * HZ;  /* TTL under high load: 120 sec */
97 int inet_peer_maxttl __read_mostly = 10 * 60 * HZ;      /* usual time to live: 10 min */
98 int inet_peer_gc_mintime __read_mostly = 10 * HZ;
99 int inet_peer_gc_maxtime __read_mostly = 120 * HZ;
100
101 static struct {
102         struct list_head        list;
103         spinlock_t              lock;
104 } unused_peers = {
105         .list                   = LIST_HEAD_INIT(unused_peers.list),
106         .lock                   = __SPIN_LOCK_UNLOCKED(unused_peers.lock),
107 };
108
109 static void peer_check_expire(unsigned long dummy);
110 static DEFINE_TIMER(peer_periodic_timer, peer_check_expire, 0, 0);
111
112
113 /* Called from ip_output.c:ip_init  */
114 void __init inet_initpeers(void)
115 {
116         struct sysinfo si;
117
118         /* Use the straight interface to information about memory. */
119         si_meminfo(&si);
120         /* The values below were suggested by Alexey Kuznetsov
121          * <kuznet@ms2.inr.ac.ru>.  I don't have any opinion about the values
122          * myself.  --SAW
123          */
124         if (si.totalram <= (32768*1024)/PAGE_SIZE)
125                 inet_peer_threshold >>= 1; /* max pool size about 1MB on IA32 */
126         if (si.totalram <= (16384*1024)/PAGE_SIZE)
127                 inet_peer_threshold >>= 1; /* about 512KB */
128         if (si.totalram <= (8192*1024)/PAGE_SIZE)
129                 inet_peer_threshold >>= 2; /* about 128KB */
130
131         peer_cachep = kmem_cache_create("inet_peer_cache",
132                         sizeof(struct inet_peer),
133                         0, SLAB_HWCACHE_ALIGN | SLAB_PANIC,
134                         NULL);
135
136         /* All the timers, started at system startup tend
137            to synchronize. Perturb it a bit.
138          */
139         peer_periodic_timer.expires = jiffies
140                 + net_random() % inet_peer_gc_maxtime
141                 + inet_peer_gc_maxtime;
142         add_timer(&peer_periodic_timer);
143 }
144
145 /* Called with or without local BH being disabled. */
146 static void unlink_from_unused(struct inet_peer *p)
147 {
148         if (!list_empty(&p->unused)) {
149                 spin_lock_bh(&unused_peers.lock);
150                 list_del_init(&p->unused);
151                 spin_unlock_bh(&unused_peers.lock);
152         }
153 }
154
155 /*
156  * Called with local BH disabled and the pool lock held.
157  */
158 #define lookup(_daddr, _stack, _base)                           \
159 ({                                                              \
160         struct inet_peer *u;                                    \
161         struct inet_peer __rcu **v;                             \
162                                                                 \
163         stackptr = _stack;                                      \
164         *stackptr++ = &_base->root;                             \
165         for (u = rcu_dereference_protected(_base->root,         \
166                         lockdep_is_held(&_base->lock));         \
167              u != peer_avl_empty; ) {                           \
168                 if (_daddr == u->v4daddr)                       \
169                         break;                                  \
170                 if ((__force __u32)_daddr < (__force __u32)u->v4daddr)  \
171                         v = &u->avl_left;                       \
172                 else                                            \
173                         v = &u->avl_right;                      \
174                 *stackptr++ = v;                                \
175                 u = rcu_dereference_protected(*v,               \
176                         lockdep_is_held(&_base->lock));         \
177         }                                                       \
178         u;                                                      \
179 })
180
181 /*
182  * Called with rcu_read_lock_bh()
183  * Because we hold no lock against a writer, its quite possible we fall
184  * in an endless loop.
185  * But every pointer we follow is guaranteed to be valid thanks to RCU.
186  * We exit from this function if number of links exceeds PEER_MAXDEPTH
187  */
188 static struct inet_peer *lookup_rcu_bh(__be32 daddr, struct inet_peer_base *base)
189 {
190         struct inet_peer *u = rcu_dereference_bh(base->root);
191         int count = 0;
192
193         while (u != peer_avl_empty) {
194                 if (daddr == u->v4daddr) {
195                         /* Before taking a reference, check if this entry was
196                          * deleted, unlink_from_pool() sets refcnt=-1 to make
197                          * distinction between an unused entry (refcnt=0) and
198                          * a freed one.
199                          */
200                         if (unlikely(!atomic_add_unless(&u->refcnt, 1, -1)))
201                                 u = NULL;
202                         return u;
203                 }
204                 if ((__force __u32)daddr < (__force __u32)u->v4daddr)
205                         u = rcu_dereference_bh(u->avl_left);
206                 else
207                         u = rcu_dereference_bh(u->avl_right);
208                 if (unlikely(++count == PEER_MAXDEPTH))
209                         break;
210         }
211         return NULL;
212 }
213
214 /* Called with local BH disabled and the pool lock held. */
215 #define lookup_rightempty(start, base)                          \
216 ({                                                              \
217         struct inet_peer *u;                                    \
218         struct inet_peer __rcu **v;                             \
219         *stackptr++ = &start->avl_left;                         \
220         v = &start->avl_left;                                   \
221         for (u = rcu_dereference_protected(*v,                  \
222                         lockdep_is_held(&base->lock));          \
223              u->avl_right != peer_avl_empty_rcu; ) {            \
224                 v = &u->avl_right;                              \
225                 *stackptr++ = v;                                \
226                 u = rcu_dereference_protected(*v,               \
227                         lockdep_is_held(&base->lock));          \
228         }                                                       \
229         u;                                                      \
230 })
231
232 /* Called with local BH disabled and the pool lock held.
233  * Variable names are the proof of operation correctness.
234  * Look into mm/map_avl.c for more detail description of the ideas.
235  */
236 static void peer_avl_rebalance(struct inet_peer __rcu **stack[],
237                                struct inet_peer __rcu ***stackend,
238                                struct inet_peer_base *base)
239 {
240         struct inet_peer __rcu **nodep;
241         struct inet_peer *node, *l, *r;
242         int lh, rh;
243
244         while (stackend > stack) {
245                 nodep = *--stackend;
246                 node = rcu_dereference_protected(*nodep,
247                                 lockdep_is_held(&base->lock));
248                 l = rcu_dereference_protected(node->avl_left,
249                                 lockdep_is_held(&base->lock));
250                 r = rcu_dereference_protected(node->avl_right,
251                                 lockdep_is_held(&base->lock));
252                 lh = node_height(l);
253                 rh = node_height(r);
254                 if (lh > rh + 1) { /* l: RH+2 */
255                         struct inet_peer *ll, *lr, *lrl, *lrr;
256                         int lrh;
257                         ll = rcu_dereference_protected(l->avl_left,
258                                 lockdep_is_held(&base->lock));
259                         lr = rcu_dereference_protected(l->avl_right,
260                                 lockdep_is_held(&base->lock));
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_dereference_protected(lr->avl_left,
272                                         lockdep_is_held(&base->lock));  /* lrl: RH or RH-1 */
273                                 lrr = rcu_dereference_protected(lr->avl_right,
274                                         lockdep_is_held(&base->lock));  /* lrr: RH or RH-1 */
275                                 RCU_INIT_POINTER(node->avl_left, lrr);  /* lrr: RH or RH-1 */
276                                 RCU_INIT_POINTER(node->avl_right, r);   /* r: RH */
277                                 node->avl_height = rh + 1; /* node: RH+1 */
278                                 RCU_INIT_POINTER(l->avl_left, ll);      /* ll: RH */
279                                 RCU_INIT_POINTER(l->avl_right, lrl);    /* lrl: RH or RH-1 */
280                                 l->avl_height = rh + 1; /* l: RH+1 */
281                                 RCU_INIT_POINTER(lr->avl_left, l);      /* l: RH+1 */
282                                 RCU_INIT_POINTER(lr->avl_right, node);  /* node: RH+1 */
283                                 lr->avl_height = rh + 2;
284                                 RCU_INIT_POINTER(*nodep, lr);
285                         }
286                 } else if (rh > lh + 1) { /* r: LH+2 */
287                         struct inet_peer *rr, *rl, *rlr, *rll;
288                         int rlh;
289                         rr = rcu_dereference_protected(r->avl_right,
290                                 lockdep_is_held(&base->lock));
291                         rl = rcu_dereference_protected(r->avl_left,
292                                 lockdep_is_held(&base->lock));
293                         rlh = node_height(rl);
294                         if (rlh <= node_height(rr)) {   /* rr: LH+1 */
295                                 RCU_INIT_POINTER(node->avl_right, rl);  /* rl: LH or LH+1 */
296                                 RCU_INIT_POINTER(node->avl_left, l);    /* l: LH */
297                                 node->avl_height = rlh + 1; /* LH+1 or LH+2 */
298                                 RCU_INIT_POINTER(r->avl_right, rr);     /* rr: LH+1 */
299                                 RCU_INIT_POINTER(r->avl_left, node);    /* node: LH+1 or LH+2 */
300                                 r->avl_height = node->avl_height + 1;
301                                 RCU_INIT_POINTER(*nodep, r);
302                         } else { /* rr: RH, rl: RH+1 */
303                                 rlr = rcu_dereference_protected(rl->avl_right,
304                                         lockdep_is_held(&base->lock));  /* rlr: LH or LH-1 */
305                                 rll = rcu_dereference_protected(rl->avl_left,
306                                         lockdep_is_held(&base->lock));  /* rll: LH or LH-1 */
307                                 RCU_INIT_POINTER(node->avl_right, rll); /* rll: LH or LH-1 */
308                                 RCU_INIT_POINTER(node->avl_left, l);    /* l: LH */
309                                 node->avl_height = lh + 1; /* node: LH+1 */
310                                 RCU_INIT_POINTER(r->avl_right, rr);     /* rr: LH */
311                                 RCU_INIT_POINTER(r->avl_left, rlr);     /* rlr: LH or LH-1 */
312                                 r->avl_height = lh + 1; /* r: LH+1 */
313                                 RCU_INIT_POINTER(rl->avl_right, r);     /* r: LH+1 */
314                                 RCU_INIT_POINTER(rl->avl_left, node);   /* node: LH+1 */
315                                 rl->avl_height = lh + 2;
316                                 RCU_INIT_POINTER(*nodep, rl);
317                         }
318                 } else {
319                         node->avl_height = (lh > rh ? lh : rh) + 1;
320                 }
321         }
322 }
323
324 /* Called with local BH disabled and the pool lock held. */
325 #define link_to_pool(n, base)                                   \
326 do {                                                            \
327         n->avl_height = 1;                                      \
328         n->avl_left = peer_avl_empty_rcu;                       \
329         n->avl_right = peer_avl_empty_rcu;                      \
330         /* lockless readers can catch us now */                 \
331         rcu_assign_pointer(**--stackptr, n);                    \
332         peer_avl_rebalance(stack, stackptr, base);              \
333 } while (0)
334
335 static void inetpeer_free_rcu(struct rcu_head *head)
336 {
337         kmem_cache_free(peer_cachep, container_of(head, struct inet_peer, rcu));
338 }
339
340 /* May be called with local BH enabled. */
341 static void unlink_from_pool(struct inet_peer *p, struct inet_peer_base *base)
342 {
343         int do_free;
344
345         do_free = 0;
346
347         spin_lock_bh(&base->lock);
348         /* Check the reference counter.  It was artificially incremented by 1
349          * in cleanup() function to prevent sudden disappearing.  If we can
350          * atomically (because of lockless readers) take this last reference,
351          * it's safe to remove the node and free it later.
352          * We use refcnt=-1 to alert lockless readers this entry is deleted.
353          */
354         if (atomic_cmpxchg(&p->refcnt, 1, -1) == 1) {
355                 struct inet_peer __rcu **stack[PEER_MAXDEPTH];
356                 struct inet_peer __rcu ***stackptr, ***delp;
357                 if (lookup(p->v4daddr, stack, base) != p)
358                         BUG();
359                 delp = stackptr - 1; /* *delp[0] == p */
360                 if (p->avl_left == peer_avl_empty_rcu) {
361                         *delp[0] = p->avl_right;
362                         --stackptr;
363                 } else {
364                         /* look for a node to insert instead of p */
365                         struct inet_peer *t;
366                         t = lookup_rightempty(p, base);
367                         BUG_ON(rcu_dereference_protected(*stackptr[-1],
368                                         lockdep_is_held(&base->lock)) != t);
369                         **--stackptr = t->avl_left;
370                         /* t is removed, t->v4daddr > x->v4daddr for any
371                          * x in p->avl_left subtree.
372                          * Put t in the old place of p. */
373                         RCU_INIT_POINTER(*delp[0], t);
374                         t->avl_left = p->avl_left;
375                         t->avl_right = p->avl_right;
376                         t->avl_height = p->avl_height;
377                         BUG_ON(delp[1] != &p->avl_left);
378                         delp[1] = &t->avl_left; /* was &p->avl_left */
379                 }
380                 peer_avl_rebalance(stack, stackptr, base);
381                 base->total--;
382                 do_free = 1;
383         }
384         spin_unlock_bh(&base->lock);
385
386         if (do_free)
387                 call_rcu_bh(&p->rcu, inetpeer_free_rcu);
388         else
389                 /* The node is used again.  Decrease the reference counter
390                  * back.  The loop "cleanup -> unlink_from_unused
391                  *   -> unlink_from_pool -> putpeer -> link_to_unused
392                  *   -> cleanup (for the same node)"
393                  * doesn't really exist because the entry will have a
394                  * recent deletion time and will not be cleaned again soon.
395                  */
396                 inet_putpeer(p);
397 }
398
399 static struct inet_peer_base *peer_to_base(struct inet_peer *p)
400 {
401         return &v4_peers;
402 }
403
404 /* May be called with local BH enabled. */
405 static int cleanup_once(unsigned long ttl)
406 {
407         struct inet_peer *p = NULL;
408
409         /* Remove the first entry from the list of unused nodes. */
410         spin_lock_bh(&unused_peers.lock);
411         if (!list_empty(&unused_peers.list)) {
412                 __u32 delta;
413
414                 p = list_first_entry(&unused_peers.list, struct inet_peer, unused);
415                 delta = (__u32)jiffies - p->dtime;
416
417                 if (delta < ttl) {
418                         /* Do not prune fresh entries. */
419                         spin_unlock_bh(&unused_peers.lock);
420                         return -1;
421                 }
422
423                 list_del_init(&p->unused);
424
425                 /* Grab an extra reference to prevent node disappearing
426                  * before unlink_from_pool() call. */
427                 atomic_inc(&p->refcnt);
428         }
429         spin_unlock_bh(&unused_peers.lock);
430
431         if (p == NULL)
432                 /* It means that the total number of USED entries has
433                  * grown over inet_peer_threshold.  It shouldn't really
434                  * happen because of entry limits in route cache. */
435                 return -1;
436
437         unlink_from_pool(p, peer_to_base(p));
438         return 0;
439 }
440
441 static struct inet_peer_base *family_to_base(int family)
442 {
443         return &v4_peers;
444 }
445
446 /* Called with or without local BH being disabled. */
447 struct inet_peer *inet_getpeer(__be32 daddr, int create)
448 {
449         struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr;
450         struct inet_peer_base *base = family_to_base(AF_INET);
451         struct inet_peer *p;
452
453         /* Look up for the address quickly, lockless.
454          * Because of a concurrent writer, we might not find an existing entry.
455          */
456         rcu_read_lock_bh();
457         p = lookup_rcu_bh(daddr, base);
458         rcu_read_unlock_bh();
459
460         if (p) {
461                 /* The existing node has been found.
462                  * Remove the entry from unused list if it was there.
463                  */
464                 unlink_from_unused(p);
465                 return p;
466         }
467
468         /* retry an exact lookup, taking the lock before.
469          * At least, nodes should be hot in our cache.
470          */
471         spin_lock_bh(&base->lock);
472         p = lookup(daddr, stack, base);
473         if (p != peer_avl_empty) {
474                 atomic_inc(&p->refcnt);
475                 spin_unlock_bh(&base->lock);
476                 /* Remove the entry from unused list if it was there. */
477                 unlink_from_unused(p);
478                 return p;
479         }
480         p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
481         if (p) {
482                 p->v4daddr = daddr;
483                 atomic_set(&p->refcnt, 1);
484                 atomic_set(&p->rid, 0);
485                 atomic_set(&p->ip_id_count, secure_ip_id(daddr));
486                 p->tcp_ts_stamp = 0;
487                 INIT_LIST_HEAD(&p->unused);
488
489
490                 /* Link the node. */
491                 link_to_pool(p, base);
492                 base->total++;
493         }
494         spin_unlock_bh(&base->lock);
495
496         if (base->total >= inet_peer_threshold)
497                 /* Remove one less-recently-used entry. */
498                 cleanup_once(0);
499
500         return p;
501 }
502
503 static int compute_total(void)
504 {
505         return v4_peers.total;
506 }
507
508 /* Called with local BH disabled. */
509 static void peer_check_expire(unsigned long dummy)
510 {
511         unsigned long now = jiffies;
512         int ttl, total;
513
514         total = compute_total();
515         if (total >= inet_peer_threshold)
516                 ttl = inet_peer_minttl;
517         else
518                 ttl = inet_peer_maxttl
519                                 - (inet_peer_maxttl - inet_peer_minttl) / HZ *
520                                         total / inet_peer_threshold * HZ;
521         while (!cleanup_once(ttl)) {
522                 if (jiffies != now)
523                         break;
524         }
525
526         /* Trigger the timer after inet_peer_gc_mintime .. inet_peer_gc_maxtime
527          * interval depending on the total number of entries (more entries,
528          * less interval). */
529         total = compute_total();
530         if (total >= inet_peer_threshold)
531                 peer_periodic_timer.expires = jiffies + inet_peer_gc_mintime;
532         else
533                 peer_periodic_timer.expires = jiffies
534                         + inet_peer_gc_maxtime
535                         - (inet_peer_gc_maxtime - inet_peer_gc_mintime) / HZ *
536                                 total / inet_peer_threshold * HZ;
537         add_timer(&peer_periodic_timer);
538 }
539
540 void inet_putpeer(struct inet_peer *p)
541 {
542         local_bh_disable();
543
544         if (atomic_dec_and_lock(&p->refcnt, &unused_peers.lock)) {
545                 list_add_tail(&p->unused, &unused_peers.list);
546                 p->dtime = (__u32)jiffies;
547                 spin_unlock(&unused_peers.lock);
548         }
549
550         local_bh_enable();
551 }