Re: [PATCH net] ipv4: Fix in_device refcount resurrection in in_dev_get()

On Sun, Aug 16, 2026 at 02:20:32AM +0900, Baul Lee wrote:
> in_dev_get() reads dev->ip_ptr under RCU and then unconditionally
> increments its refcount. inetdev_destroy() clears the pointer and drops
> the last reference under RTNL, with no grace period in between, so a
> reader that fetched the pointer before the store can increment a
> refcount that has already reached zero. That resurrects an object whose
> RCU free is queued:

This part is fine.

> dropping the resurrected reference re-enters in_dev_finish_destroy()
> for a second netdev_put() and a second call_rcu() on the same
> rcu_head, and if the grace period elapses first the drop itself is a
> use-after-free.

Are you sure about this part? The reference count is set to
REFCOUNT_SATURATED when you increment from zero, so I don't think you
re-enter in_dev_finish_destroy().

> 
> inet_netconf_get_devconf() is registered RTNL_FLAG_DOIT_UNLOCKED, and
> rtnetlink_rcv_msg() exempts RTNL_KIND_GET from the CAP_NET_ADMIN check,
> so an unprivileged user can drive the reader side. Reproduced as UID
> 65534 on v7.2-rc7:
> 
>   refcount_t: addition on 0; use-after-free.
>   WARNING: lib/refcount.c:25 at refcount_warn_saturate+0x14c/0x180
>   CPU: 0 UID: 65534 PID: 655 Comm: j1_poc
>    refcount_warn_saturate+0x14c/0x180 (P)
>    inet_netconf_get_devconf+0x4b0/0x4c4
>    rtnetlink_rcv_msg+0x434/0x4d0
> 
> followed by the matching underflow when the reference is dropped.
> 
> Use refcount_inc_not_zero() and return NULL for an in_device that has
> already reached zero. All callers already handle a NULL return, which
> in_dev_get() gives today whenever dev->ip_ptr is NULL. Callers under
> RTNL see no change: ip_ptr is cleared before the last put, so a non-NULL
> ip_ptr there implies a non-zero refcount.
> 
> Discovered by XBOW, triaged by Baul Lee <[email protected]>

FYI, it was actually mentioned a few times already:

https://lore.kernel.org/netdev/20260802115639.GA270646@shredder/
https://sashiko.dev/#/patchset/20260731135202.566337-1-david.lee%40trailofbits.com

> 
> Fixes: bbcf91053bb6 ("inet: do not use RTNL in inet_netconf_get_devconf()")

I think you should blame commit 9d40c84cf5bc ("net: devinet: Reduce
refcount before grace period") instead:

1. in_dev_get() was called w/o RTNL even before bbcf91053bb6.

2. Calling in_dev_get() w/o RTNL only became unsafe after 9d40c84cf5bc.
Before that, inetdev_destroy() dropped the reference after an RCU grace
period.

The diff itself looks OK.