[RFC PATCH 1/3] of/device: manage resources similar to platform_device_add

[Grant Likely]

On Thu, 22 Jan 2015 15:49:59 -0600
, Suman Anna <s-anna at ti.com>
 wrote:
> Hi Grant,
> 
> On 01/13/2015 05:04 PM, Suman Anna wrote:
> > On 01/13/2015 04:00 PM, Rob Herring wrote:
> >> On Tue, Jan 13, 2015 at 3:25 PM, Suman Anna <s-anna at ti.com> wrote:
> >>> Hi Rob,
> >>>
> >>> On 01/13/2015 02:38 PM, Rob Herring wrote:
> >>>> On Wed, Jan 7, 2015 at 11:30 AM, Suman Anna <s-anna at ti.com> wrote:
> >>>>> Drivers can use of_platform_populate() to create platform devices
> >>>>> for children of the device main node, and a complementary API
> >>>>> of_platform_depopulate() is provided to delete these child platform
> >>>>> devices. The of_platform_depopulate() leverages the platform API
> >>>>> for performing the cleanup of these devices.
> >>>>>
> >>>>> The platform device resources are managed differently between
> >>>>> of_device_add and platform_device_add, and this asymmetry causes
> >>>>> a kernel oops in platform_device_del during removal of the resources.
> >>>>> Manage the platform device resources similar to platform_device_add
> >>>>> to fix this kernel oops.
> >>>>
> >>>> This is a known issue and has been attempted to be fixed before (I
> >>>> believe there is a revert in mainline). The problem is there are known
> >>>> devicetrees which have overlapping resources and they will break with
> >>>> your change.
> >>>
> >>> Are you referring to 02bbde7849e6 (Revert "of: use
> >>> platform_device_add")?
> >>
> >> I believe that's the one.
> >>
> >>> That one seems to be in registration path, and
> >>> this crash is in the unregistration path. If so, to fix the crash,
> >>> should we be skipping the release_resource() for now in
> >>> platform_device_del for DT nodes, or replace platform_device_unregister
> >>> with of_device_unregister in of_platform_device_destroy()?
> >>
> >> IIRC, the problem is inserting a resource twice on add from 2
> >> different nodes, not the removal path. Perhaps we could make a
> >> collision non-fatal for in the DT case.
> > 
> > We may be talking two different things here, I understand that this
> > patch would create an issue with inserting a resource twice in the
> > devicetrees with overlapping resources (just like the commit that was
> > reverted above), but the crash is on devices with resources whose
> > parent, child, sibling pointers have never been initialized (the
> > of_device_add path does not touch these at all), and get dereferenced in
> > platform_device_del()->release_resource(). See the following that has a
> > better explanation [1].
> > 
> > regards
> > Suman
> > 
> > [1]
> > http://lists.infradead.org/pipermail/linux-arm-kernel/2014-July/274412.html
> > 
> > 
> >> Grant may have some ideas on
> >> what's needed here.
> 
> Ping, any suggestions here? Do we ought to replace
> platform_device_unregister() with of_device_unregister() similar to the
> approach taken in 02bbde7849e6 (Revert "of: use platform_device_add")?

Hi Suman,

Yes, I think the solution to both problems is to create an
of_device_unregister() function. It's not the prettiest thing, but I
think it is for the best.

g.