Re: [tpmdd-devel] [PATCH] base/platform: fix panic when probe function is NULL

[Wilck, Martin]

On Do, 2015-11-26 at 13:30 -0700, Jason Gunthorpe wrote:
> On Thu, Nov 26, 2015 at 08:01:34PM +0100, martin.wilck@xxxxxxxxxxxxxx wrote:
> > From: Martin Wilck <Martin.Wilck@xxxxxxxxxxxxxx>
> > 
> > Since b8b2c7d845d5, platform_drv_probe() is called for all platform
> > devices. If drv->probe is NULL, and dev_pm_domain_attach() fails,
> > platform_drv_probe() will return the error code from dev_pm_domain_attach().
> > 
> > This causes real_probe() to enter the "probe_failed" path and set
> > dev->driver to NULL. Before b8b2c7d845d5, real_probe() would assume
> > success if both dev->bus->probe and drv->probe are missing.
> > 
> > This may cause a panic later. For example, inserting the tpm_tis
> > driver with parameter "force=1" (i.e. registering tpm_tis as a platform
> > driver) will panic in tpmm_chip_alloc() because dev->driver is NULL:
> > 
> >      chip->cdev.owner = chip->pdev->driver->owner;
> 
> Is this happening because tpm_tis is not creating the platform device
> properly? ie it just calls platform_device_register_simple and then
> force initializes it via tpm_tis_init, which expects to be called from
> a probe function with an attached driver.
> 
> Instead we should setup a proper platform device with the default
> IO range for x86 and let the driver core call tpm_tis_init via
> tis_drv.probe.
> 
> Would changing things in this way fix the problem you've observed?

I think so. Nonetheless, patch b8b2c7d845d5 introduced a change in the
way platform device registration behaves. The platform device code seems
to be prepared for cases where platform_driver->probe == NULL, so that
case should be handled gracefully. Otherwise, failure should occur
earlier, e.g. when platform_driver_register() is called with 
platform_driver->probe == NULL. tpm_tis may not be the only driver that
uses platform_device_register_simple() in this way.

> I have some patches to do this that are part of my OF enablement
> series, but I can make something simpler that would deal with this
> fairly quickly if you can test.

Let's first wait what the platform guys say.

Martin