Hello,

On Tue, Dec 21, 2021 at 01:04:50PM +0100, Lars-Peter Clausen wrote:
> On 12/21/21 12:35 PM, Uwe Kleine-König wrote:
> > On Tue, Dec 21, 2021 at 12:12:12PM +0100, Lars-Peter Clausen wrote:
> > > On 12/21/21 11:45 AM, Uwe Kleine-König wrote:
> > > > similar to patch
> > > > https://lore.kernel.org/r/4bde7cbd9e43a5909208102094444219d3154466.1640072891.git.vilhelm.gray@gmail.com
> > > > the usage of struct counter_device::priv can be replaced by
> > > > container_of which improves type safety and code size.
> > > > 
> > > > This series depends on above patch, converts the remaining drivers and
> > > > finally drops struct counter_device::priv.
> > > Not sure if this is such a good idea. struct counter_device should not be
> > > embedded in the drivers state struct in the first place.
> > Just to mention it: My patch series didn't change this, this was already
> > broken before.
> I know, but this series has to be reverted when the framework is fixed.
> > 
> > > struct counter_device contains a struct device, which is a reference counted
> > > object. But by embedding it in the driver state struct the life time of both
> > > the struct counter_device and and struct device are bound to the life time
> > > of the driver state struct.
> > > 
> > > Which means the struct device memory can get freed before the last reference
> > > is dropped, which leads to a use-after-free and undefined behavior.
> > Well, the driver struct is allocated using devm_kzalloc for all drivers.
> 
> devm_kzalloc() doesn't make a difference. The managed memory is freed when
> the parent device is unbound/removed. There may very well be reference to
> the counter_device at this point.
> 
> > So I think it's not *very* urgent to fix. Still you're right, this
> > should be addressed.
> 
> Yes and no, this can trivially be used for privilege escalation, but then
> again on systems with a counter_device probably everything runs as root
> anyway.

I could provoke an oops with the following shell command:


	{ sleep 5; echo bang; } > /dev/counter0 & sleep 1; echo 40000000.timer:counter > /sys/bus/platform/drivers/stm32-timer-counter/unbind

I have a protype here to split counter_register() into counter_alloc() +
counter_add(), but I didn't convert a driver to it yet. If you want to
take a look, it's currently available from

	https://git.pengutronix.de/git/ukl/linux counter-dev-livetime

(or if you prefer a webif:

	https://git.pengutronix.de/cgit/ukl/linux/log/?h=counter-dev-livetime

). I planned to just invest a two or so hours to fix this. But the plan
failed (one reason is that v5.16-rc6 failed to boot on the stm32mp1 I
work on and I bisected that first.)

Maybe I find some time between the years to get this forward a bit.

Best regards
Uwe

-- 
Pengutronix e.K.                           | Uwe Kleine-König            |
Industrial Linux Solutions                 | https://www.pengutronix.de/ |