usb: dwc3: of-simple: fix use-after-free on remove

[Felipe Balbi <balbi@kernel.org>]

Hi,

Roger Quadros <rogerq@ti.com> writes:
> On 18/06/18 12:51, Felipe Balbi wrote:
>> 
>> Hi,
>> 
>> Johan Hovold <johan@kernel.org> writes:
>>> On Mon, Jun 18, 2018 at 12:33:44PM +0300, Felipe Balbi wrote:
>>>
>>>> Johan Hovold <johan@kernel.org> writes:
>>>
>>>>> I suggest merging this fix for 4.18-rc, and then Roger can rework the
>>>>> driver so that it works also on OMAP.
>>>>
>>>> omap has its own glue layer for several reasons. If you're talking about
>>>> Keystone devices, then okay, I understand. But in that case, this would
>>>> mean Keystone is copying the same arguably broken PM domain design from
>>>> OMAP and it would be best not to propagate that idea.
>>>
>>> Maybe so. I'm not sure what Roger's use case is, but perhaps the omap
>>> glue driver could be used instead.
>> 
>> unlikely. Keystone devices are very different from OMAP family. But
>> we'll see what Roger says.
>> 
>
> Well, I was considering to use of-simple for the AM654 SoC [1] but now
> I'm of the opinion that it might be better to add a new glue layer driver

why isn't dwc3-keystone.c enough?

> for that because
> - it needs to poke a few registers in the wrapper region

dwc3-keystone.c does that already

> - it doesn't really need the driver to enable any clock

Seems to me you're trying to port omap_device to arm64...

> - it needs a pm_runtime_get_sync() to be done in probe

this really shouldn't be necessary. Keystone doesn't rely on all the
omap_device legacy. At least it didn't use to. Could it be that you're
just missing a struct dev_pm_domain definition for arm64?

I haven't seen how you guys implemented your PM for arm64 (is there a
publically accessible version somewhere?), but I'd say you should take
the opportunity to remove this relying on pm_runtime_get_sync() calls
from probe and just do what everybody else does; namely: enable clocks
on probe, pm_runtime_set_active, etc.

This helps drivers being able to make assumptions about devices being
enabled during probe. pm_runtime becomes easier to implement generically
too.