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, u32 ssid, u32 tsid,
69 u16 tclass, u32 perms,
76 struct avc_callback_node *next;
79 /* Exported via selinufs */
80 unsigned int avc_cache_threshold = AVC_DEF_CACHE_THRESHOLD;
82 #ifdef CONFIG_SECURITY_SELINUX_AVC_STATS
83 DEFINE_PER_CPU(struct avc_cache_stats, avc_cache_stats) = { 0 };
86 static struct avc_cache avc_cache;
87 static struct avc_callback_node *avc_callbacks;
88 static struct kmem_cache *avc_node_cachep;
90 static inline int avc_hash(u32 ssid, u32 tsid, u16 tclass)
92 return (ssid ^ (tsid<<2) ^ (tclass<<4)) & (AVC_CACHE_SLOTS - 1);
96 * avc_dump_av - Display an access vector in human-readable form.
97 * @tclass: target security class
100 static void avc_dump_av(struct audit_buffer *ab, u16 tclass, u32 av)
106 audit_log_format(ab, " null");
110 perms = secclass_map[tclass-1].perms;
112 audit_log_format(ab, " {");
115 while (i < (sizeof(av) * 8)) {
116 if ((perm & av) && perms[i]) {
117 audit_log_format(ab, " %s", perms[i]);
125 audit_log_format(ab, " 0x%x", av);
127 audit_log_format(ab, " }");
131 * avc_dump_query - Display a SID pair and a class in human-readable form.
132 * @ssid: source security identifier
133 * @tsid: target security identifier
134 * @tclass: target security class
136 static void avc_dump_query(struct audit_buffer *ab, u32 ssid, u32 tsid, u16 tclass)
142 rc = security_sid_to_context(ssid, &scontext, &scontext_len);
144 audit_log_format(ab, "ssid=%d", ssid);
146 audit_log_format(ab, "scontext=%s", scontext);
150 rc = security_sid_to_context(tsid, &scontext, &scontext_len);
152 audit_log_format(ab, " tsid=%d", tsid);
154 audit_log_format(ab, " tcontext=%s", scontext);
158 BUG_ON(tclass >= ARRAY_SIZE(secclass_map));
159 audit_log_format(ab, " tclass=%s", secclass_map[tclass-1].name);
163 * avc_init - Initialize the AVC.
165 * Initialize the access vector cache.
167 void __init avc_init(void)
171 for (i = 0; i < AVC_CACHE_SLOTS; i++) {
172 INIT_HLIST_HEAD(&avc_cache.slots[i]);
173 spin_lock_init(&avc_cache.slots_lock[i]);
175 atomic_set(&avc_cache.active_nodes, 0);
176 atomic_set(&avc_cache.lru_hint, 0);
178 avc_node_cachep = kmem_cache_create("avc_node", sizeof(struct avc_node),
179 0, SLAB_PANIC, NULL);
181 audit_log(current->audit_context, GFP_KERNEL, AUDIT_KERNEL, "AVC INITIALIZED\n");
184 int avc_get_hash_stats(char *page)
186 int i, chain_len, max_chain_len, slots_used;
187 struct avc_node *node;
188 struct hlist_head *head;
194 for (i = 0; i < AVC_CACHE_SLOTS; i++) {
195 head = &avc_cache.slots[i];
196 if (!hlist_empty(head)) {
197 struct hlist_node *next;
201 hlist_for_each_entry_rcu(node, next, head, list)
203 if (chain_len > max_chain_len)
204 max_chain_len = chain_len;
210 return scnprintf(page, PAGE_SIZE, "entries: %d\nbuckets used: %d/%d\n"
211 "longest chain: %d\n",
212 atomic_read(&avc_cache.active_nodes),
213 slots_used, AVC_CACHE_SLOTS, max_chain_len);
216 static void avc_node_free(struct rcu_head *rhead)
218 struct avc_node *node = container_of(rhead, struct avc_node, rhead);
219 kmem_cache_free(avc_node_cachep, node);
220 avc_cache_stats_incr(frees);
223 static void avc_node_delete(struct avc_node *node)
225 hlist_del_rcu(&node->list);
226 call_rcu(&node->rhead, avc_node_free);
227 atomic_dec(&avc_cache.active_nodes);
230 static void avc_node_kill(struct avc_node *node)
232 kmem_cache_free(avc_node_cachep, node);
233 avc_cache_stats_incr(frees);
234 atomic_dec(&avc_cache.active_nodes);
237 static void avc_node_replace(struct avc_node *new, struct avc_node *old)
239 hlist_replace_rcu(&old->list, &new->list);
240 call_rcu(&old->rhead, avc_node_free);
241 atomic_dec(&avc_cache.active_nodes);
244 static inline int avc_reclaim_node(void)
246 struct avc_node *node;
247 int hvalue, try, ecx;
249 struct hlist_head *head;
250 struct hlist_node *next;
253 for (try = 0, ecx = 0; try < AVC_CACHE_SLOTS; try++) {
254 hvalue = atomic_inc_return(&avc_cache.lru_hint) & (AVC_CACHE_SLOTS - 1);
255 head = &avc_cache.slots[hvalue];
256 lock = &avc_cache.slots_lock[hvalue];
258 if (!spin_trylock_irqsave(lock, flags))
262 hlist_for_each_entry(node, next, head, list) {
263 avc_node_delete(node);
264 avc_cache_stats_incr(reclaims);
266 if (ecx >= AVC_CACHE_RECLAIM) {
268 spin_unlock_irqrestore(lock, flags);
273 spin_unlock_irqrestore(lock, flags);
279 static struct avc_node *avc_alloc_node(void)
281 struct avc_node *node;
283 node = kmem_cache_zalloc(avc_node_cachep, GFP_ATOMIC);
287 INIT_HLIST_NODE(&node->list);
288 avc_cache_stats_incr(allocations);
290 if (atomic_inc_return(&avc_cache.active_nodes) > avc_cache_threshold)
297 static void avc_node_populate(struct avc_node *node, u32 ssid, u32 tsid, u16 tclass, struct av_decision *avd)
299 node->ae.ssid = ssid;
300 node->ae.tsid = tsid;
301 node->ae.tclass = tclass;
302 memcpy(&node->ae.avd, avd, sizeof(node->ae.avd));
305 static inline struct avc_node *avc_search_node(u32 ssid, u32 tsid, u16 tclass)
307 struct avc_node *node, *ret = NULL;
309 struct hlist_head *head;
310 struct hlist_node *next;
312 hvalue = avc_hash(ssid, tsid, tclass);
313 head = &avc_cache.slots[hvalue];
314 hlist_for_each_entry_rcu(node, next, head, list) {
315 if (ssid == node->ae.ssid &&
316 tclass == node->ae.tclass &&
317 tsid == node->ae.tsid) {
327 * avc_lookup - Look up an AVC entry.
328 * @ssid: source security identifier
329 * @tsid: target security identifier
330 * @tclass: target security class
332 * Look up an AVC entry that is valid for the
333 * (@ssid, @tsid), interpreting the permissions
334 * based on @tclass. If a valid AVC entry exists,
335 * then this function returns the avc_node.
336 * Otherwise, this function returns NULL.
338 static struct avc_node *avc_lookup(u32 ssid, u32 tsid, u16 tclass)
340 struct avc_node *node;
342 avc_cache_stats_incr(lookups);
343 node = avc_search_node(ssid, tsid, tclass);
348 avc_cache_stats_incr(misses);
352 static int avc_latest_notif_update(int seqno, int is_insert)
355 static DEFINE_SPINLOCK(notif_lock);
358 spin_lock_irqsave(¬if_lock, flag);
360 if (seqno < avc_cache.latest_notif) {
361 printk(KERN_WARNING "SELinux: avc: seqno %d < latest_notif %d\n",
362 seqno, avc_cache.latest_notif);
366 if (seqno > avc_cache.latest_notif)
367 avc_cache.latest_notif = seqno;
369 spin_unlock_irqrestore(¬if_lock, flag);
375 * avc_insert - Insert an AVC entry.
376 * @ssid: source security identifier
377 * @tsid: target security identifier
378 * @tclass: target security class
379 * @avd: resulting av decision
381 * Insert an AVC entry for the SID pair
382 * (@ssid, @tsid) and class @tclass.
383 * The access vectors and the sequence number are
384 * normally provided by the security server in
385 * response to a security_compute_av() call. If the
386 * sequence number @avd->seqno is not less than the latest
387 * revocation notification, then the function copies
388 * the access vectors into a cache entry, returns
389 * avc_node inserted. Otherwise, this function returns NULL.
391 static struct avc_node *avc_insert(u32 ssid, u32 tsid, u16 tclass, struct av_decision *avd)
393 struct avc_node *pos, *node = NULL;
397 if (avc_latest_notif_update(avd->seqno, 1))
400 node = avc_alloc_node();
402 struct hlist_head *head;
403 struct hlist_node *next;
406 hvalue = avc_hash(ssid, tsid, tclass);
407 avc_node_populate(node, ssid, tsid, tclass, avd);
409 head = &avc_cache.slots[hvalue];
410 lock = &avc_cache.slots_lock[hvalue];
412 spin_lock_irqsave(lock, flag);
413 hlist_for_each_entry(pos, next, head, list) {
414 if (pos->ae.ssid == ssid &&
415 pos->ae.tsid == tsid &&
416 pos->ae.tclass == tclass) {
417 avc_node_replace(node, pos);
421 hlist_add_head_rcu(&node->list, head);
423 spin_unlock_irqrestore(lock, flag);
430 * avc_audit_pre_callback - SELinux specific information
431 * will be called by generic audit code
432 * @ab: the audit buffer
435 static void avc_audit_pre_callback(struct audit_buffer *ab, void *a)
437 struct common_audit_data *ad = a;
438 audit_log_format(ab, "avc: %s ",
439 ad->selinux_audit_data->denied ? "denied" : "granted");
440 avc_dump_av(ab, ad->selinux_audit_data->tclass,
441 ad->selinux_audit_data->audited);
442 audit_log_format(ab, " for ");
446 * avc_audit_post_callback - SELinux specific information
447 * will be called by generic audit code
448 * @ab: the audit buffer
451 static void avc_audit_post_callback(struct audit_buffer *ab, void *a)
453 struct common_audit_data *ad = a;
454 audit_log_format(ab, " ");
455 avc_dump_query(ab, ad->selinux_audit_data->ssid,
456 ad->selinux_audit_data->tsid,
457 ad->selinux_audit_data->tclass);
460 /* This is the slow part of avc audit with big stack footprint */
461 static noinline int slow_avc_audit(u32 ssid, u32 tsid, u16 tclass,
462 u32 requested, u32 audited, u32 denied,
463 struct common_audit_data *a,
466 struct common_audit_data stack_data;
467 struct selinux_audit_data sad = {0,};
471 COMMON_AUDIT_DATA_INIT(a, NONE);
472 a->selinux_audit_data = &sad;
476 * When in a RCU walk do the audit on the RCU retry. This is because
477 * the collection of the dname in an inode audit message is not RCU
478 * safe. Note this may drop some audits when the situation changes
479 * during retry. However this is logically just as if the operation
480 * happened a little later.
482 if ((a->type == LSM_AUDIT_DATA_INODE) &&
483 (flags & MAY_NOT_BLOCK))
486 a->selinux_audit_data->tclass = tclass;
487 a->selinux_audit_data->requested = requested;
488 a->selinux_audit_data->ssid = ssid;
489 a->selinux_audit_data->tsid = tsid;
490 a->selinux_audit_data->audited = audited;
491 a->selinux_audit_data->denied = denied;
492 a->lsm_pre_audit = avc_audit_pre_callback;
493 a->lsm_post_audit = avc_audit_post_callback;
499 * avc_audit - Audit the granting or denial of permissions.
500 * @ssid: source security identifier
501 * @tsid: target security identifier
502 * @tclass: target security class
503 * @requested: requested permissions
504 * @avd: access vector decisions
505 * @result: result from avc_has_perm_noaudit
506 * @a: auxiliary audit data
507 * @flags: VFS walk flags
509 * Audit the granting or denial of permissions in accordance
510 * with the policy. This function is typically called by
511 * avc_has_perm() after a permission check, but can also be
512 * called directly by callers who use avc_has_perm_noaudit()
513 * in order to separate the permission check from the auditing.
514 * For example, this separation is useful when the permission check must
515 * be performed under a lock, to allow the lock to be released
516 * before calling the auditing code.
518 inline int avc_audit(u32 ssid, u32 tsid,
519 u16 tclass, u32 requested,
520 struct av_decision *avd, int result, struct common_audit_data *a,
524 denied = requested & ~avd->allowed;
525 if (unlikely(denied)) {
526 audited = denied & avd->auditdeny;
528 * a->selinux_audit_data->auditdeny is TRICKY! Setting a bit in
529 * this field means that ANY denials should NOT be audited if
530 * the policy contains an explicit dontaudit rule for that
531 * permission. Take notice that this is unrelated to the
532 * actual permissions that were denied. As an example lets
536 * avd.auditdeny & ACCESS == 0 (not set means explicit rule)
537 * selinux_audit_data->auditdeny & ACCESS == 1
539 * We will NOT audit the denial even though the denied
540 * permission was READ and the auditdeny checks were for
544 a->selinux_audit_data->auditdeny &&
545 !(a->selinux_audit_data->auditdeny & avd->auditdeny))
548 audited = denied = requested;
550 audited = requested & avd->auditallow;
551 if (likely(!audited))
554 return slow_avc_audit(ssid, tsid, tclass,
555 requested, audited, denied,
560 * avc_add_callback - Register a callback for security events.
561 * @callback: callback function
562 * @events: security events
563 * @ssid: source security identifier or %SECSID_WILD
564 * @tsid: target security identifier or %SECSID_WILD
565 * @tclass: target security class
566 * @perms: permissions
568 * Register a callback function for events in the set @events
569 * related to the SID pair (@ssid, @tsid)
570 * and the permissions @perms, interpreting
571 * @perms based on @tclass. Returns %0 on success or
572 * -%ENOMEM if insufficient memory exists to add the callback.
574 int avc_add_callback(int (*callback)(u32 event, u32 ssid, u32 tsid,
575 u16 tclass, u32 perms,
577 u32 events, u32 ssid, u32 tsid,
578 u16 tclass, u32 perms)
580 struct avc_callback_node *c;
583 c = kmalloc(sizeof(*c), GFP_ATOMIC);
589 c->callback = callback;
594 c->next = avc_callbacks;
600 static inline int avc_sidcmp(u32 x, u32 y)
602 return (x == y || x == SECSID_WILD || y == SECSID_WILD);
606 * avc_update_node Update an AVC entry
607 * @event : Updating event
608 * @perms : Permission mask bits
609 * @ssid,@tsid,@tclass : identifier of an AVC entry
610 * @seqno : sequence number when decision was made
612 * if a valid AVC entry doesn't exist,this function returns -ENOENT.
613 * if kmalloc() called internal returns NULL, this function returns -ENOMEM.
614 * otherwise, this function updates the AVC entry. The original AVC-entry object
615 * will release later by RCU.
617 static int avc_update_node(u32 event, u32 perms, u32 ssid, u32 tsid, u16 tclass,
622 struct avc_node *pos, *node, *orig = NULL;
623 struct hlist_head *head;
624 struct hlist_node *next;
627 node = avc_alloc_node();
633 /* Lock the target slot */
634 hvalue = avc_hash(ssid, tsid, tclass);
636 head = &avc_cache.slots[hvalue];
637 lock = &avc_cache.slots_lock[hvalue];
639 spin_lock_irqsave(lock, flag);
641 hlist_for_each_entry(pos, next, head, list) {
642 if (ssid == pos->ae.ssid &&
643 tsid == pos->ae.tsid &&
644 tclass == pos->ae.tclass &&
645 seqno == pos->ae.avd.seqno){
658 * Copy and replace original node.
661 avc_node_populate(node, ssid, tsid, tclass, &orig->ae.avd);
664 case AVC_CALLBACK_GRANT:
665 node->ae.avd.allowed |= perms;
667 case AVC_CALLBACK_TRY_REVOKE:
668 case AVC_CALLBACK_REVOKE:
669 node->ae.avd.allowed &= ~perms;
671 case AVC_CALLBACK_AUDITALLOW_ENABLE:
672 node->ae.avd.auditallow |= perms;
674 case AVC_CALLBACK_AUDITALLOW_DISABLE:
675 node->ae.avd.auditallow &= ~perms;
677 case AVC_CALLBACK_AUDITDENY_ENABLE:
678 node->ae.avd.auditdeny |= perms;
680 case AVC_CALLBACK_AUDITDENY_DISABLE:
681 node->ae.avd.auditdeny &= ~perms;
684 avc_node_replace(node, orig);
686 spin_unlock_irqrestore(lock, flag);
692 * avc_flush - Flush the cache
694 static void avc_flush(void)
696 struct hlist_head *head;
697 struct hlist_node *next;
698 struct avc_node *node;
703 for (i = 0; i < AVC_CACHE_SLOTS; i++) {
704 head = &avc_cache.slots[i];
705 lock = &avc_cache.slots_lock[i];
707 spin_lock_irqsave(lock, flag);
709 * With preemptable RCU, the outer spinlock does not
710 * prevent RCU grace periods from ending.
713 hlist_for_each_entry(node, next, head, list)
714 avc_node_delete(node);
716 spin_unlock_irqrestore(lock, flag);
721 * avc_ss_reset - Flush the cache and revalidate migrated permissions.
722 * @seqno: policy sequence number
724 int avc_ss_reset(u32 seqno)
726 struct avc_callback_node *c;
731 for (c = avc_callbacks; c; c = c->next) {
732 if (c->events & AVC_CALLBACK_RESET) {
733 tmprc = c->callback(AVC_CALLBACK_RESET,
735 /* save the first error encountered for the return
736 value and continue processing the callbacks */
742 avc_latest_notif_update(seqno, 0);
747 * Slow-path helper function for avc_has_perm_noaudit,
748 * when the avc_node lookup fails. We get called with
749 * the RCU read lock held, and need to return with it
750 * still held, but drop if for the security compute.
752 * Don't inline this, since it's the slow-path and just
753 * results in a bigger stack frame.
755 static noinline struct avc_node *avc_compute_av(u32 ssid, u32 tsid,
756 u16 tclass, struct av_decision *avd)
759 security_compute_av(ssid, tsid, tclass, avd);
761 return avc_insert(ssid, tsid, tclass, avd);
764 static noinline int avc_denied(u32 ssid, u32 tsid,
765 u16 tclass, u32 requested,
767 struct av_decision *avd)
769 if (flags & AVC_STRICT)
772 if (selinux_enforcing && !(avd->flags & AVD_FLAGS_PERMISSIVE))
775 avc_update_node(AVC_CALLBACK_GRANT, requested, ssid,
776 tsid, tclass, avd->seqno);
782 * avc_has_perm_noaudit - Check permissions but perform no auditing.
783 * @ssid: source security identifier
784 * @tsid: target security identifier
785 * @tclass: target security class
786 * @requested: requested permissions, interpreted based on @tclass
787 * @flags: AVC_STRICT or 0
788 * @avd: access vector decisions
790 * Check the AVC to determine whether the @requested permissions are granted
791 * for the SID pair (@ssid, @tsid), interpreting the permissions
792 * based on @tclass, and call the security server on a cache miss to obtain
793 * a new decision and add it to the cache. Return a copy of the decisions
794 * in @avd. Return %0 if all @requested permissions are granted,
795 * -%EACCES if any permissions are denied, or another -errno upon
796 * other errors. This function is typically called by avc_has_perm(),
797 * but may also be called directly to separate permission checking from
798 * auditing, e.g. in cases where a lock must be held for the check but
799 * should be released for the auditing.
801 inline int avc_has_perm_noaudit(u32 ssid, u32 tsid,
802 u16 tclass, u32 requested,
804 struct av_decision *avd)
806 struct avc_node *node;
814 node = avc_lookup(ssid, tsid, tclass);
815 if (unlikely(!node)) {
816 node = avc_compute_av(ssid, tsid, tclass, avd);
818 memcpy(avd, &node->ae.avd, sizeof(*avd));
822 denied = requested & ~(avd->allowed);
823 if (unlikely(denied))
824 rc = avc_denied(ssid, tsid, tclass, requested, flags, avd);
831 * avc_has_perm - Check permissions and perform any appropriate auditing.
832 * @ssid: source security identifier
833 * @tsid: target security identifier
834 * @tclass: target security class
835 * @requested: requested permissions, interpreted based on @tclass
836 * @auditdata: auxiliary audit data
837 * @flags: VFS walk flags
839 * Check the AVC to determine whether the @requested permissions are granted
840 * for the SID pair (@ssid, @tsid), interpreting the permissions
841 * based on @tclass, and call the security server on a cache miss to obtain
842 * a new decision and add it to the cache. Audit the granting or denial of
843 * permissions in accordance with the policy. Return %0 if all @requested
844 * permissions are granted, -%EACCES if any permissions are denied, or
845 * another -errno upon other errors.
847 int avc_has_perm_flags(u32 ssid, u32 tsid, u16 tclass,
848 u32 requested, struct common_audit_data *auditdata,
851 struct av_decision avd;
854 rc = avc_has_perm_noaudit(ssid, tsid, tclass, requested, 0, &avd);
856 rc2 = avc_audit(ssid, tsid, tclass, requested, &avd, rc, auditdata,
863 u32 avc_policy_seqno(void)
865 return avc_cache.latest_notif;
868 void avc_disable(void)
871 * If you are looking at this because you have realized that we are
872 * not destroying the avc_node_cachep it might be easy to fix, but
873 * I don't know the memory barrier semantics well enough to know. It's
874 * possible that some other task dereferenced security_ops when
875 * it still pointed to selinux operations. If that is the case it's
876 * possible that it is about to use the avc and is about to need the
877 * avc_node_cachep. I know I could wrap the security.c security_ops call
878 * in an rcu_lock, but seriously, it's not worth it. Instead I just flush
879 * the cache and get that memory back.
881 if (avc_node_cachep) {
883 /* kmem_cache_destroy(avc_node_cachep); */