Re: [PATCH v2 0/2] wifi: carl9170: revert broken devres conversions for input and hwrng

[Dmitry Torokhov <dmitry.torokhov@gmail.com>]

On Sun, Oct 11, 2026 at 04:41:02PM +0200, Christian Lamparter wrote:
> On 10/10/26 1:13 AM, Dmitry Torokhov wrote:
> > Hi Christian,
> > 
> > On Fri, Oct 09, 2026 at 09:37:36PM +0200, Christian Lamparter wrote:
> > > On 10/8/26 11:39 AM, Dmitry Torokhov wrote:
> > > > Commits 23de0fa0d2a0 ("carl9170: devres-ing hwrng_register usage") and
> > > > 87ddb2fc29f1 ("carl9170: devres-ing input_allocate_device") converted
> > > > the HWRNG and WPS button input device registrations in carl9170 to
> > > > devres attached to the parent struct usb_device (&ar->udev->dev) and
> > > > removed explicit unregistration from carl9170_unregister().
> > > > 
> > > > In carl9170_usb_disconnect(), the driver calls carl9170_unregister()
> > > > followed immediately by carl9170_free(), which frees struct ar9170
> > > > inside the interface .disconnect() callback before devres_release_all()
> > > > runs. Furthermore, devres on &ar->udev->dev is not released on
> > > > interface unbind or registration failure in carl9170_register().
> > > > As a result, both the WPS input device (whose input->name and
> > > > input->phys point into freed memory) and the embedded struct hwrng
> > > > remain registered after struct ar9170 has been freed, leading to
> > > > use-after-free bugs.
> > > 
> > > Ok, so you just reworded your patch? Sight...
> > 
> > Yes, as I promised I would.
> > 
> > > looking at the WPS input
> > > 
> > > |        input->name = ar->wps.name;
> > > |        input->phys = ar->wps.phys;
> > > |        input->id.bustype = BUS_USB;
> > > |        input->dev.parent = &ar->hw->wiphy->dev;
> > >            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > > 
> > > the input's dev.parent is set to ar->hw->wiphy->dev and not ar->udev->dev, right?
> > > Does this do anything at all? If not, why? The wiphy gets shutdown by
> > > ieee80211_unregister() and having the "freeing" stick around after the USB device
> > > is gone should not hurt, right?
> > 
> > devm_input_allocate_device() explicitly states:
> > 
> > 	NOTE: the owner device is set up as parent of input device and
> > 	users should not override it.
> 
> Yeah, and this is something you added in 2012... 

Yes, in the same patch where devm_input_allocate_device() was
introduced. This requirement was there from the very beginning.

> The carl9170 code driver is back from 2010.

No, the devm conversion was done in 2022, 10 years later.

> 
> So, was this changed to "break it", rather than the driver?

The conversion to devm in the driver was broken on arrival.

> 
> From my POV: What's the point of devres then?

In certain situation it allows to reduce amount of boilerplate code in
the driver. But it has be to be used properly. One of the most important
thinks it attaching resources to the right device and not mixing managed
and noon managed resources in the same driver, as that very often causes
incorrect order on unwinding.

> And why are you so "eager" to revert the changes.

Because I have to deal with syzhot reports routed to me and I would
rather not.

Humor me: why are you so against reverting the change? I have shown that
it is problematic. You can go back and re-introduce devm for input once
the rest of the driver is ready.

Thanks.

-- 
Dmitry