audit: move put_tree() to avoid trim_trees refcount underflow and UAF

commit 67d69e9d1a6c889d98951c1d74b19332ce0565af upstream.

AUDIT_TRIM is expected to be idempotent, but multiple executions resulted
in a refcount underflow and use-after-free.

git bisect fingered commit fb041bb7c0a9	("locking/refcount: Consolidate
implementations of refcount_t") but this patch with its more thorough
checking that wasn't in the x86 assembly code merely exposed a previously
existing tree refcount imbalance in the case of tree trimming code that
was refactored with prune_one() to remove a tree introduced in
commit 8432c7006297 ("audit: Simplify locking around untag_chunk()")

Move the put_tree() to cover only the prune_one() case.

Passes audit-testsuite and 3 passes of "auditctl -t" with at least one
directory watch.

Cc: Jan Kara <[email protected]>
Cc: Will Deacon <[email protected]>
Cc: Alexander Viro <[email protected]>
Cc: Seiji Nishikawa <[email protected]>
Cc: [email protected]
Fixes: 8432c7006297 ("audit: Simplify locking around untag_chunk()")
Signed-off-by: Richard Guy Briggs <[email protected]>
Reviewed-by: Jan Kara <[email protected]>
[PM: reformatted/cleaned-up the commit description]
Signed-off-by: Paul Moore <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>

kernel/audit_tree.c

@@ -593,7 +593,6 @@ static void prune_tree_chunks(struct audit_tree *victim, bool tagged)
 		spin_lock(&hash_lock);
 	}
 	spin_unlock(&hash_lock);
-	put_tree(victim);
 }
 
 /*
@@ -602,6 +601,7 @@ static void prune_tree_chunks(struct audit_tree *victim, bool tagged)
 static void prune_one(struct audit_tree *victim)
 {
 	prune_tree_chunks(victim, false);
+	put_tree(victim);
 }
 
 /* trim the uncommitted chunks from tree */