Re: [PATCH] driver core: fix use-after-free of driver_override via driver_match_device()

[Greg KH <gregkh@linuxfoundation.org>]

On Wed, Nov 26, 2025 at 07:30:06PM +0800, Gui-Dong Han wrote:
> driver_set_override() modifies and frees dev->driver_override while
> holding device_lock(dev). However, driver_match_device() reads
> dev->driver_override when calling bus match functions.
> 
> Currently, driver_match_device() is called from three sites. One site
> (__device_attach_driver) holds device_lock(dev), but the other two
> (bind_store and __driver_attach) do not. This allows a concurrent
> driver_set_override() to free the string while driver_match_device() is
> using it, leading to a use-after-free (UAF).
> 
> This issue affects at least 11 bus types (including PCI, AMBA, Platform)
> that rely on driver_override for matching.
> 
> Fix this by holding device_lock(dev) around the driver_match_device() calls
> in bind_store() and __driver_attach(). This ensures all access to
> dev->driver_override via driver_match_device() is protected by the device
> lock.
> 
> Tested with the PoCs from Bugzilla that trigger this UAF. Stress testing
> the two newly locked paths for 24 hours with CONFIG_PROVE_LOCKING and
> CONFIG_LOCKDEP enabled showed no UAF recurrence and no lockdep
> warnings.
> 
> Closes: https://bugzilla.kernel.org/show_bug.cgi?id=220789
> Suggested-by: Qiu-ji Chen <chenqiuji666@gmail.com>
> Signed-off-by: Gui-Dong Han <hanguidong02@gmail.com>
> ---
> The Bugzilla entry contains full KASAN reports and two PoCs that reliably
> reproduce the UAF on both unlocked paths using a standard QEMU setup
> (default e1000 device at 0000:00:03.0).
> I chose to fix this in the driver core for the following reasons:
> 1. Both racing functions are part of the driver core.
> 2. Fixing this per-driver/per-bus is tedious and would require careful
> ad-hoc locking that does not align with the existing device_lock(dev).
> 3. We cannot simply add device_lock(dev) inside bus match functions because
> one call path (__device_attach_driver) already holds this lock. Adding the
> lock inside the match callback would cause a deadlock on that path.
> ---
>  drivers/base/bus.c | 17 ++++++++++++-----
>  drivers/base/dd.c  |  3 +++
>  2 files changed, 15 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/base/bus.c b/drivers/base/bus.c
> index 5e75e1bce551..9e62d6009058 100644
> --- a/drivers/base/bus.c
> +++ b/drivers/base/bus.c
> @@ -261,13 +261,20 @@ static ssize_t bind_store(struct device_driver *drv, const char *buf,
>  	const struct bus_type *bus = bus_get(drv->bus);
>  	struct device *dev;
>  	int err = -ENODEV;
> +	int ret;
>  
>  	dev = bus_find_device_by_name(bus, NULL, buf);
> -	if (dev && driver_match_device(drv, dev)) {
> -		err = device_driver_attach(drv, dev);
> -		if (!err) {
> -			/* success */
> -			err = count;
> +	if (dev) {
> +		/* Protects against driver_set_override() races */
> +		device_lock(dev);
> +		ret = driver_match_device(drv, dev);
> +		device_unlock(dev);

Why not have driver_match_device() take the lock instead?  This way
looks like an "anti-pattern" that we will get wrong over time.

thanks,

greg k-h