On Mon, Oct 20, 2025 at 09:49:12PM +0800, Lizhi Xu wrote:

On Mon, 20 Oct 2025 21:34:56 +0800, Lizhi Xu wrote:

quoted

quoted

Task0					Task1						Task2
=====					=====						=====
[97] nr_add_node()
[113] nr_neigh_get_dev()		[97] nr_add_node()
					[214] nr_node_lock()
					[245] nr_node->routes[2].neighbour->count--
					[246] nr_neigh_put(nr_node->routes[2].neighbour);
					[248] nr_remove_neigh(nr_node->routes[2].neighbour)
					[283] nr_node_unlock()
[214] nr_node_lock()
[253] nr_node->routes[2].neighbour = nr_neigh
[254] nr_neigh_hold(nr_neigh);							[97] nr_add_node()
											[XXX] nr_neigh_put()
                                                                                        ^^^^^^^^^^^^^^^^^^^^

These charts are supposed to be chronological so [XXX] is wrong because the
use after free happens on line [248].  Do we really need three threads to
make this race work?

The UAF problem occurs in Task2. Task1 sets the refcount of nr_neigh to 1,
then Task0 adds it to routes[2]. Task2 releases routes[2].neighbour after
executing [XXX]nr_neigh_put().

Execution Order:
1 -> Task0
[113] nr_neigh_get_dev() // After execution, the refcount value is 3

2 -> Task1
[246] nr_neigh_put(nr_node->routes[2].neighbour);   // After execution, the refcount value is 2
[248] nr_remove_neigh(nr_node->routes[2].neighbour) // After execution, the refcount value is 1

3 -> Task0
[253] nr_node->routes[2].neighbour = nr_neigh       // nr_neigh's refcount value is 1 and add it to routes[2]

4 -> Task2
[XXX] nr_neigh_put(nr_node->routes[2].neighbour)    // After execution, neighhour is freed
if (nr_node->routes[2].neighbour->count == 0 && !nr_node->routes[2].neighbour->locked)  // Uaf occurs this line when accessing neighbour->count

Let's step back a bit and look at the bigger picture design.  (Which is
completely undocumented so we're just guessing).

When we put nr_neigh into nr_node->routes[] we bump the nr_neigh_hold()
reference count and nr_neigh->count++, then when we remove it from
->routes[] we drop the reference and do nr_neigh->count--.

If it's the last reference (and we are not holding ->locked) then we
remove it from the &nr_neigh_list and drop the reference count again and
free it.  So we drop the reference count twice.  This is a complicated
design with three variables: nr_neigh_hold(), nr_neigh->count and
->locked.  Why can it not just be one counter nr_neigh_hold().  So
instead of setting locked = true we would just take an extra reference?
The nr_neigh->count++ would be replaced with nr_neigh_hold() as well.

Because that's fundamentally the problem, right?  We call
nr_neigh_get_dev() so we think we're holding a reference and we're
safe, but we don't realize that calling neighbour->count-- can
result in dropping two references.

regards,
dan carpenter