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

commit 67d69e9 upstream.

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

git bisect fingered commit fb041bb	("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 8432c70 ("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 <jack@suse.cz>
Cc: Will Deacon <will@kernel.org>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Seiji Nishikawa <snishika@redhat.com>
Cc: stable@vger.kernel.org
Fixes: 8432c70 ("audit: Simplify locking around untag_chunk()")
Signed-off-by: Richard Guy Briggs <rgb@redhat.com>
Reviewed-by: Jan Kara <jack@suse.cz>
[PM: reformatted/cleaned-up the commit description]
Signed-off-by: Paul Moore <paul@paul-moore.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

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 */