On Thu, 23 Oct 2025 13:44:18 +0200, Paolo Abeni wrote:

quoted

The root cause of the problem is that multiple different tasks initiate
SIOCADDRT & NETROM_NODE commands to add new routes, there is no lock
between them to protect the same nr_neigh.

Task0 can add the nr_neigh.refcount value of 1 on Task1 to routes[2].
When Task2 executes nr_neigh_put(nr_node->routes[2].neighbour), it will
release the neighbour because its refcount value is 1.

In this case, the following situation causes a UAF on Task2:

Task0					Task1						Task2
=====					=====						=====
nr_add_node()
nr_neigh_get_dev()			nr_add_node()
					nr_node_lock()
					nr_node->routes[2].neighbour->count--
					nr_neigh_put(nr_node->routes[2].neighbour);
					nr_remove_neigh(nr_node->routes[2].neighbour)
					nr_node_unlock()
nr_node_lock()
nr_node->routes[2].neighbour = nr_neigh
nr_neigh_hold(nr_neigh);								nr_add_node()
											nr_neigh_put()
											if (nr_node->routes[2].neighbour->count
Description of the UAF triggering process:
First, Task 0 executes nr_neigh_get_dev() to set neighbor refcount to 3.
Then, Task 1 puts the same neighbor from its routes[2] and executes
nr_remove_neigh() because the count is 0. After these two operations,
the neighbor's refcount becomes 1. Then, Task 0 acquires the nr node
lock and writes it to its routes[2].neighbour.
Finally, Task 2 executes nr_neigh_put(nr_node->routes[2].neighbour) to
release the neighbor. The subsequent execution of the neighbor->count
check triggers a UAF.

The solution to the problem is to use a lock to synchronize each add a
route to node, but for rigor, I'll add locks to related ioctl and route
frame operations to maintain synchronization.

I think that adding another locking mechanism on top of an already
complex and not well understood locking and reference infra is not the
right direction.

Why reordering the statements as:

	if (nr_node->routes[2].neighbour->count == 0 &&
!nr_node->routes[2].neighbour->locked)
		nr_remove_neigh(nr_node->routes[2].neighbour);
	nr_neigh_put(nr_node->routes[2].neighbour);

is not enough?

This is not enough, the same uaf will appear, nr_remove_neigh() will also
execute nr_neigh_put(), and then executing nr_neigh_put() again will
trigger the uaf.

quoted

syzbot reported:
BUG: KASAN: slab-use-after-free in nr_add_node+0x25db/0x2c00 net/netrom/nr_route.c:248
Read of size 4 at addr ffff888051e6e9b0 by task syz.1.2539/8741

Call Trace:
 <TASK>
 nr_add_node+0x25db/0x2c00 net/netrom/nr_route.c:248

Reported-by: syzbot+2860e75836a08b172755@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=2860e75836a08b172755
Signed-off-by: Lizhi Xu <redacted>

quoted

---
V1 -> V2: update comments for cause uaf
V2 -> V3: sync neighbor operations in ioctl and route frame, update comments

 net/netrom/nr_route.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/net/netrom/nr_route.c b/net/netrom/nr_route.c
index b94cb2ffbaf8..debe3e925338 100644
--- a/net/netrom/nr_route.c
+++ b/net/netrom/nr_route.c

@@ -40,6 +40,7 @@ static HLIST_HEAD(nr_node_list);
 static DEFINE_SPINLOCK(nr_node_list_lock);
 static HLIST_HEAD(nr_neigh_list);
 static DEFINE_SPINLOCK(nr_neigh_list_lock);
+static DEFINE_MUTEX(neighbor_lock);

 static struct nr_node *nr_node_get(ax25_address *callsign)
 {

@@ -633,6 +634,8 @@ int nr_rt_ioctl(unsigned int cmd, void __user *arg)
 	ax25_digi digi;
 	int ret;

+	guard(mutex)(&neighbor_lock);

See:

https://elixir.bootlin.com/linux/v6.18-rc1/source/Documentation/process/maintainer-netdev.rst#L395

Using guard is not recommended. I'll reconsider.

BR,
Lizhi