When the first child of the node being validated gets deleted during

validation, man_node_unlink() switches to MAN_NEXT_CHILD.  After
that, we have to switch back to MAN_NEXT_SIBLING after completing
validation, or subsequent parsing would add content into an already
closed node, clobbering potentially existing children, causing
information loss and a memory leak.  Bug found by kristaps@ with
valgrind in groff(7) on Mac OS X.

Note that the switch back must be conditional, for if the node being
validated itself gets deleted, we must *not* go to MAN_NEXT_SIBLING,
which would not only yield wrong results in general but also crash
in malformed manuals having an empty paragraph before the first .SH,
for example OpenBSD c++filt(1).

1 files changed, 12 insertions, 1 deletions

diff --git a/man_macro.c b/man_macro.c
index 8f12b894..7c2cc71a 100644
--- a/man_macro.c
+++ b/man_macro.c
@@ -100,7 +100,6 @@ man_unscope(struct man *man, const struct man_node *to)
 {
 	struct man_node	*n;
 
-	man->next = MAN_NEXT_SIBLING;
 	to = to->parent;
 	n = man->last;
 	while (n != to) {
@@ -139,11 +138,23 @@ man_unscope(struct man *man, const struct man_node *to)
 		 * Save a pointer to the parent such that
 		 * we know where to continue the iteration.
 		 */
+
 		man->last = n;
 		n = n->parent;
 		if ( ! man_valid_post(man))
 			return(0);
 	}
+
+	/*
+	 * If we ended up at the parent of the node we were
+	 * supposed to rewind to, that means the target node
+	 * got deleted, so add the next node we parse as a child
+	 * of the parent instead of as a sibling of the target.
+	 */
+
+	man->next = (man->last == to) ?
+	    MAN_NEXT_CHILD : MAN_NEXT_SIBLING;
+
 	return(1);
 }