2 * Implementation of the kernel access vector cache (AVC).
4 * Authors: Stephen Smalley, <sds@epoch.ncsc.mil>
5 * James Morris <jmorris@redhat.com>
7 * Update: KaiGai, Kohei <kaigai@ak.jp.nec.com>
8 * Replaced the avc_lock spinlock by RCU.
10 * Copyright (C) 2003 Red Hat, Inc., James Morris <jmorris@redhat.com>
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License version 2,
14 * as published by the Free Software Foundation.
16 #include <linux/types.h>
17 #include <linux/stddef.h>
18 #include <linux/kernel.h>
19 #include <linux/slab.h>
21 #include <linux/dcache.h>
22 #include <linux/init.h>
23 #include <linux/skbuff.h>
24 #include <linux/percpu.h>
27 #include <net/af_unix.h>
29 #include <linux/audit.h>
30 #include <linux/ipv6.h>
36 #define AVC_CACHE_SLOTS 512
37 #define AVC_DEF_CACHE_THRESHOLD 512
38 #define AVC_CACHE_RECLAIM 16
40 #ifdef CONFIG_SECURITY_SELINUX_AVC_STATS
41 #define avc_cache_stats_incr(field) this_cpu_inc(avc_cache_stats.field)
43 #define avc_cache_stats_incr(field) do {} while (0)
50 struct av_decision avd;
55 struct hlist_node list; /* anchored in avc_cache->slots[i] */
56 struct rcu_head rhead;
60 struct hlist_head slots[AVC_CACHE_SLOTS]; /* head for avc_node->list */
61 spinlock_t slots_lock[AVC_CACHE_SLOTS]; /* lock for writes */
62 atomic_t lru_hint; /* LRU hint for reclaim scan */
63 atomic_t active_nodes;
64 u32 latest_notif; /* latest revocation notification */
67 struct avc_callback_node {
68 int (*callback) (u32 event);
70 struct avc_callback_node *next;
73 /* Exported via selinufs */
74 unsigned int avc_cache_threshold = AVC_DEF_CACHE_THRESHOLD;
76 #ifdef CONFIG_SECURITY_SELINUX_AVC_STATS
77 DEFINE_PER_CPU(struct avc_cache_stats, avc_cache_stats) = { 0 };
80 static struct avc_cache avc_cache;
81 static struct avc_callback_node *avc_callbacks;
82 static struct kmem_cache *avc_node_cachep;
84 static inline int avc_hash(u32 ssid, u32 tsid, u16 tclass)
86 return (ssid ^ (tsid<<2) ^ (tclass<<4)) & (AVC_CACHE_SLOTS - 1);
90 * avc_dump_av - Display an access vector in human-readable form.
91 * @tclass: target security class
94 static void avc_dump_av(struct audit_buffer *ab, u16 tclass, u32 av)
100 audit_log_format(ab, " null");
104 perms = secclass_map[tclass-1].perms;
106 audit_log_format(ab, " {");
109 while (i < (sizeof(av) * 8)) {
110 if ((perm & av) && perms[i]) {
111 audit_log_format(ab, " %s", perms[i]);
119 audit_log_format(ab, " 0x%x", av);
121 audit_log_format(ab, " }");
125 * avc_dump_query - Display a SID pair and a class in human-readable form.
126 * @ssid: source security identifier
127 * @tsid: target security identifier
128 * @tclass: target security class
130 static void avc_dump_query(struct audit_buffer *ab, u32 ssid, u32 tsid, u16 tclass)
136 rc = security_sid_to_context(ssid, &scontext, &scontext_len);
138 audit_log_format(ab, "ssid=%d", ssid);
140 audit_log_format(ab, "scontext=%s", scontext);
144 rc = security_sid_to_context(tsid, &scontext, &scontext_len);
146 audit_log_format(ab, " tsid=%d", tsid);
148 audit_log_format(ab, " tcontext=%s", scontext);
152 BUG_ON(tclass >= ARRAY_SIZE(secclass_map));
153 audit_log_format(ab, " tclass=%s", secclass_map[tclass-1].name);
157 * avc_init - Initialize the AVC.
159 * Initialize the access vector cache.
161 void __init avc_init(void)
165 for (i = 0; i < AVC_CACHE_SLOTS; i++) {
166 INIT_HLIST_HEAD(&avc_cache.slots[i]);
167 spin_lock_init(&avc_cache.slots_lock[i]);
169 atomic_set(&avc_cache.active_nodes, 0);
170 atomic_set(&avc_cache.lru_hint, 0);
172 avc_node_cachep = kmem_cache_create("avc_node", sizeof(struct avc_node),
173 0, SLAB_PANIC, NULL);
175 audit_log(current->audit_context, GFP_KERNEL, AUDIT_KERNEL, "AVC INITIALIZED\n");
178 int avc_get_hash_stats(char *page)
180 int i, chain_len, max_chain_len, slots_used;
181 struct avc_node *node;
182 struct hlist_head *head;
188 for (i = 0; i < AVC_CACHE_SLOTS; i++) {
189 head = &avc_cache.slots[i];
190 if (!hlist_empty(head)) {
193 hlist_for_each_entry_rcu(node, head, list)
195 if (chain_len > max_chain_len)
196 max_chain_len = chain_len;
202 return scnprintf(page, PAGE_SIZE, "entries: %d\nbuckets used: %d/%d\n"
203 "longest chain: %d\n",
204 atomic_read(&avc_cache.active_nodes),
205 slots_used, AVC_CACHE_SLOTS, max_chain_len);
208 static void avc_node_free(struct rcu_head *rhead)
210 struct avc_node *node = container_of(rhead, struct avc_node, rhead);
211 kmem_cache_free(avc_node_cachep, node);
212 avc_cache_stats_incr(frees);
215 static void avc_node_delete(struct avc_node *node)
217 hlist_del_rcu(&node->list);
218 call_rcu(&node->rhead, avc_node_free);
219 atomic_dec(&avc_cache.active_nodes);
222 static void avc_node_kill(struct avc_node *node)
224 kmem_cache_free(avc_node_cachep, node);
225 avc_cache_stats_incr(frees);
226 atomic_dec(&avc_cache.active_nodes);
229 static void avc_node_replace(struct avc_node *new, struct avc_node *old)
231 hlist_replace_rcu(&old->list, &new->list);
232 call_rcu(&old->rhead, avc_node_free);
233 atomic_dec(&avc_cache.active_nodes);
236 static inline int avc_reclaim_node(void)
238 struct avc_node *node;
239 int hvalue, try, ecx;
241 struct hlist_head *head;
244 for (try = 0, ecx = 0; try < AVC_CACHE_SLOTS; try++) {
245 hvalue = atomic_inc_return(&avc_cache.lru_hint) & (AVC_CACHE_SLOTS - 1);
246 head = &avc_cache.slots[hvalue];
247 lock = &avc_cache.slots_lock[hvalue];
249 if (!spin_trylock_irqsave(lock, flags))
253 hlist_for_each_entry(node, head, list) {
254 avc_node_delete(node);
255 avc_cache_stats_incr(reclaims);
257 if (ecx >= AVC_CACHE_RECLAIM) {
259 spin_unlock_irqrestore(lock, flags);
264 spin_unlock_irqrestore(lock, flags);
270 static struct avc_node *avc_alloc_node(void)
272 struct avc_node *node;
274 node = kmem_cache_zalloc(avc_node_cachep, GFP_ATOMIC|__GFP_NOMEMALLOC);
278 INIT_HLIST_NODE(&node->list);
279 avc_cache_stats_incr(allocations);
281 if (atomic_inc_return(&avc_cache.active_nodes) > avc_cache_threshold)
288 static void avc_node_populate(struct avc_node *node, u32 ssid, u32 tsid, u16 tclass, struct av_decision *avd)
290 node->ae.ssid = ssid;
291 node->ae.tsid = tsid;
292 node->ae.tclass = tclass;
293 memcpy(&node->ae.avd, avd, sizeof(node->ae.avd));
296 static inline struct avc_node *avc_search_node(u32 ssid, u32 tsid, u16 tclass)
298 struct avc_node *node, *ret = NULL;
300 struct hlist_head *head;
302 hvalue = avc_hash(ssid, tsid, tclass);
303 head = &avc_cache.slots[hvalue];
304 hlist_for_each_entry_rcu(node, head, list) {
305 if (ssid == node->ae.ssid &&
306 tclass == node->ae.tclass &&
307 tsid == node->ae.tsid) {
317 * avc_lookup - Look up an AVC entry.
318 * @ssid: source security identifier
319 * @tsid: target security identifier
320 * @tclass: target security class
322 * Look up an AVC entry that is valid for the
323 * (@ssid, @tsid), interpreting the permissions
324 * based on @tclass. If a valid AVC entry exists,
325 * then this function returns the avc_node.
326 * Otherwise, this function returns NULL.
328 static struct avc_node *avc_lookup(u32 ssid, u32 tsid, u16 tclass)
330 struct avc_node *node;
332 avc_cache_stats_incr(lookups);
333 node = avc_search_node(ssid, tsid, tclass);
338 avc_cache_stats_incr(misses);
342 static int avc_latest_notif_update(int seqno, int is_insert)
345 static DEFINE_SPINLOCK(notif_lock);
348 spin_lock_irqsave(¬if_lock, flag);
350 if (seqno < avc_cache.latest_notif) {
351 printk(KERN_WARNING "SELinux: avc: seqno %d < latest_notif %d\n",
352 seqno, avc_cache.latest_notif);
356 if (seqno > avc_cache.latest_notif)
357 avc_cache.latest_notif = seqno;
359 spin_unlock_irqrestore(¬if_lock, flag);
365 * avc_insert - Insert an AVC entry.
366 * @ssid: source security identifier
367 * @tsid: target security identifier
368 * @tclass: target security class
369 * @avd: resulting av decision
371 * Insert an AVC entry for the SID pair
372 * (@ssid, @tsid) and class @tclass.
373 * The access vectors and the sequence number are
374 * normally provided by the security server in
375 * response to a security_compute_av() call. If the
376 * sequence number @avd->seqno is not less than the latest
377 * revocation notification, then the function copies
378 * the access vectors into a cache entry, returns
379 * avc_node inserted. Otherwise, this function returns NULL.
381 static struct avc_node *avc_insert(u32 ssid, u32 tsid, u16 tclass, struct av_decision *avd)
383 struct avc_node *pos, *node = NULL;
387 if (avc_latest_notif_update(avd->seqno, 1))
390 node = avc_alloc_node();
392 struct hlist_head *head;
395 hvalue = avc_hash(ssid, tsid, tclass);
396 avc_node_populate(node, ssid, tsid, tclass, avd);
398 head = &avc_cache.slots[hvalue];
399 lock = &avc_cache.slots_lock[hvalue];
401 spin_lock_irqsave(lock, flag);
402 hlist_for_each_entry(pos, head, list) {
403 if (pos->ae.ssid == ssid &&
404 pos->ae.tsid == tsid &&
405 pos->ae.tclass == tclass) {
406 avc_node_replace(node, pos);
410 hlist_add_head_rcu(&node->list, head);
412 spin_unlock_irqrestore(lock, flag);
419 * avc_audit_pre_callback - SELinux specific information
420 * will be called by generic audit code
421 * @ab: the audit buffer
424 static void avc_audit_pre_callback(struct audit_buffer *ab, void *a)
426 struct common_audit_data *ad = a;
427 audit_log_format(ab, "avc: %s ",
428 ad->selinux_audit_data->denied ? "denied" : "granted");
429 avc_dump_av(ab, ad->selinux_audit_data->tclass,
430 ad->selinux_audit_data->audited);
431 audit_log_format(ab, " for ");
435 * avc_audit_post_callback - SELinux specific information
436 * will be called by generic audit code
437 * @ab: the audit buffer
440 static void avc_audit_post_callback(struct audit_buffer *ab, void *a)
442 struct common_audit_data *ad = a;
443 audit_log_format(ab, " ");
444 avc_dump_query(ab, ad->selinux_audit_data->ssid,
445 ad->selinux_audit_data->tsid,
446 ad->selinux_audit_data->tclass);
449 /* This is the slow part of avc audit with big stack footprint */
450 noinline int slow_avc_audit(u32 ssid, u32 tsid, u16 tclass,
451 u32 requested, u32 audited, u32 denied,
452 struct common_audit_data *a,
455 struct common_audit_data stack_data;
456 struct selinux_audit_data sad;
460 a->type = LSM_AUDIT_DATA_NONE;
464 * When in a RCU walk do the audit on the RCU retry. This is because
465 * the collection of the dname in an inode audit message is not RCU
466 * safe. Note this may drop some audits when the situation changes
467 * during retry. However this is logically just as if the operation
468 * happened a little later.
470 if ((a->type == LSM_AUDIT_DATA_INODE) &&
471 (flags & MAY_NOT_BLOCK))
475 sad.requested = requested;
478 sad.audited = audited;
481 a->selinux_audit_data = &sad;
483 common_lsm_audit(a, avc_audit_pre_callback, avc_audit_post_callback);
488 * avc_add_callback - Register a callback for security events.
489 * @callback: callback function
490 * @events: security events
492 * Register a callback function for events in the set @events.
493 * Returns %0 on success or -%ENOMEM if insufficient memory
494 * exists to add the callback.
496 int __init avc_add_callback(int (*callback)(u32 event), u32 events)
498 struct avc_callback_node *c;
501 c = kmalloc(sizeof(*c), GFP_KERNEL);
507 c->callback = callback;
509 c->next = avc_callbacks;
515 static inline int avc_sidcmp(u32 x, u32 y)
517 return (x == y || x == SECSID_WILD || y == SECSID_WILD);
521 * avc_update_node Update an AVC entry
522 * @event : Updating event
523 * @perms : Permission mask bits
524 * @ssid,@tsid,@tclass : identifier of an AVC entry
525 * @seqno : sequence number when decision was made
527 * if a valid AVC entry doesn't exist,this function returns -ENOENT.
528 * if kmalloc() called internal returns NULL, this function returns -ENOMEM.
529 * otherwise, this function updates the AVC entry. The original AVC-entry object
530 * will release later by RCU.
532 static int avc_update_node(u32 event, u32 perms, u32 ssid, u32 tsid, u16 tclass,
537 struct avc_node *pos, *node, *orig = NULL;
538 struct hlist_head *head;
541 node = avc_alloc_node();
547 /* Lock the target slot */
548 hvalue = avc_hash(ssid, tsid, tclass);
550 head = &avc_cache.slots[hvalue];
551 lock = &avc_cache.slots_lock[hvalue];
553 spin_lock_irqsave(lock, flag);
555 hlist_for_each_entry(pos, head, list) {
556 if (ssid == pos->ae.ssid &&
557 tsid == pos->ae.tsid &&
558 tclass == pos->ae.tclass &&
559 seqno == pos->ae.avd.seqno){
572 * Copy and replace original node.
575 avc_node_populate(node, ssid, tsid, tclass, &orig->ae.avd);
578 case AVC_CALLBACK_GRANT:
579 node->ae.avd.allowed |= perms;
581 case AVC_CALLBACK_TRY_REVOKE:
582 case AVC_CALLBACK_REVOKE:
583 node->ae.avd.allowed &= ~perms;
585 case AVC_CALLBACK_AUDITALLOW_ENABLE:
586 node->ae.avd.auditallow |= perms;
588 case AVC_CALLBACK_AUDITALLOW_DISABLE:
589 node->ae.avd.auditallow &= ~perms;
591 case AVC_CALLBACK_AUDITDENY_ENABLE:
592 node->ae.avd.auditdeny |= perms;
594 case AVC_CALLBACK_AUDITDENY_DISABLE:
595 node->ae.avd.auditdeny &= ~perms;
598 avc_node_replace(node, orig);
600 spin_unlock_irqrestore(lock, flag);
606 * avc_flush - Flush the cache
608 static void avc_flush(void)
610 struct hlist_head *head;
611 struct avc_node *node;
616 for (i = 0; i < AVC_CACHE_SLOTS; i++) {
617 head = &avc_cache.slots[i];
618 lock = &avc_cache.slots_lock[i];
620 spin_lock_irqsave(lock, flag);
622 * With preemptable RCU, the outer spinlock does not
623 * prevent RCU grace periods from ending.
626 hlist_for_each_entry(node, head, list)
627 avc_node_delete(node);
629 spin_unlock_irqrestore(lock, flag);
634 * avc_ss_reset - Flush the cache and revalidate migrated permissions.
635 * @seqno: policy sequence number
637 int avc_ss_reset(u32 seqno)
639 struct avc_callback_node *c;
644 for (c = avc_callbacks; c; c = c->next) {
645 if (c->events & AVC_CALLBACK_RESET) {
646 tmprc = c->callback(AVC_CALLBACK_RESET);
647 /* save the first error encountered for the return
648 value and continue processing the callbacks */
654 avc_latest_notif_update(seqno, 0);
659 * Slow-path helper function for avc_has_perm_noaudit,
660 * when the avc_node lookup fails. We get called with
661 * the RCU read lock held, and need to return with it
662 * still held, but drop if for the security compute.
664 * Don't inline this, since it's the slow-path and just
665 * results in a bigger stack frame.
667 static noinline struct avc_node *avc_compute_av(u32 ssid, u32 tsid,
668 u16 tclass, struct av_decision *avd)
671 security_compute_av(ssid, tsid, tclass, avd);
673 return avc_insert(ssid, tsid, tclass, avd);
676 static noinline int avc_denied(u32 ssid, u32 tsid,
677 u16 tclass, u32 requested,
679 struct av_decision *avd)
681 if (flags & AVC_STRICT)
684 if (selinux_enforcing && !(avd->flags & AVD_FLAGS_PERMISSIVE))
687 avc_update_node(AVC_CALLBACK_GRANT, requested, ssid,
688 tsid, tclass, avd->seqno);
694 * avc_has_perm_noaudit - Check permissions but perform no auditing.
695 * @ssid: source security identifier
696 * @tsid: target security identifier
697 * @tclass: target security class
698 * @requested: requested permissions, interpreted based on @tclass
699 * @flags: AVC_STRICT or 0
700 * @avd: access vector decisions
702 * Check the AVC to determine whether the @requested permissions are granted
703 * for the SID pair (@ssid, @tsid), interpreting the permissions
704 * based on @tclass, and call the security server on a cache miss to obtain
705 * a new decision and add it to the cache. Return a copy of the decisions
706 * in @avd. Return %0 if all @requested permissions are granted,
707 * -%EACCES if any permissions are denied, or another -errno upon
708 * other errors. This function is typically called by avc_has_perm(),
709 * but may also be called directly to separate permission checking from
710 * auditing, e.g. in cases where a lock must be held for the check but
711 * should be released for the auditing.
713 inline int avc_has_perm_noaudit(u32 ssid, u32 tsid,
714 u16 tclass, u32 requested,
716 struct av_decision *avd)
718 struct avc_node *node;
726 node = avc_lookup(ssid, tsid, tclass);
727 if (unlikely(!node)) {
728 node = avc_compute_av(ssid, tsid, tclass, avd);
730 memcpy(avd, &node->ae.avd, sizeof(*avd));
734 denied = requested & ~(avd->allowed);
735 if (unlikely(denied))
736 rc = avc_denied(ssid, tsid, tclass, requested, flags, avd);
743 * avc_has_perm - Check permissions and perform any appropriate auditing.
744 * @ssid: source security identifier
745 * @tsid: target security identifier
746 * @tclass: target security class
747 * @requested: requested permissions, interpreted based on @tclass
748 * @auditdata: auxiliary audit data
750 * Check the AVC to determine whether the @requested permissions are granted
751 * for the SID pair (@ssid, @tsid), interpreting the permissions
752 * based on @tclass, and call the security server on a cache miss to obtain
753 * a new decision and add it to the cache. Audit the granting or denial of
754 * permissions in accordance with the policy. Return %0 if all @requested
755 * permissions are granted, -%EACCES if any permissions are denied, or
756 * another -errno upon other errors.
758 int avc_has_perm(u32 ssid, u32 tsid, u16 tclass,
759 u32 requested, struct common_audit_data *auditdata)
761 struct av_decision avd;
764 rc = avc_has_perm_noaudit(ssid, tsid, tclass, requested, 0, &avd);
766 rc2 = avc_audit(ssid, tsid, tclass, requested, &avd, rc, auditdata);
772 u32 avc_policy_seqno(void)
774 return avc_cache.latest_notif;
777 void avc_disable(void)
780 * If you are looking at this because you have realized that we are
781 * not destroying the avc_node_cachep it might be easy to fix, but
782 * I don't know the memory barrier semantics well enough to know. It's
783 * possible that some other task dereferenced security_ops when
784 * it still pointed to selinux operations. If that is the case it's
785 * possible that it is about to use the avc and is about to need the
786 * avc_node_cachep. I know I could wrap the security.c security_ops call
787 * in an rcu_lock, but seriously, it's not worth it. Instead I just flush
788 * the cache and get that memory back.
790 if (avc_node_cachep) {
792 /* kmem_cache_destroy(avc_node_cachep); */