Re: [PATCH] usb: dwc3: qcom: Make sure core device is fully initialized before it is used

[Matthias Kaehlcke <mka@chromium.org>]

Hi Stephen,

On Wed, Jun 17, 2020 at 12:49:13PM -0700, Stephen Boyd wrote:
> Quoting Matthias Kaehlcke (2020-06-16 13:37:37)
> > dwc3_qcom_of_register_core() uses of_platform_populate() to add
> > the dwc3 core device. The driver core will try to probe the device,
> > however this might fail (e.g. due to deferred probing) and
> > of_platform_populate() would still return 0 if the device was
> > successully added to the platform bus. Verify that the core device
> > is actually bound to its driver before using it, defer probing of the
> > dwc3_qcom device if the core device isn't ready (yet).
> > 
> > Fixes: a4333c3a6ba9 ("usb: dwc3: Add Qualcomm DWC3 glue driver").
> > Signed-off-by: Matthias Kaehlcke <mka@chromium.org>
> > ---
> > depends on:
> >   https://lore.kernel.org/patchwork/patch/1251661/ ("driver core:Export
> >     the symbol device_is_bound")
> > 
> >  drivers/usb/dwc3/dwc3-qcom.c | 10 ++++++++++
> >  1 file changed, 10 insertions(+)
> > 
> > diff --git a/drivers/usb/dwc3/dwc3-qcom.c b/drivers/usb/dwc3/dwc3-qcom.c
> > index 1dfd024cd06b..5a9036b050c6 100644
> > --- a/drivers/usb/dwc3/dwc3-qcom.c
> > +++ b/drivers/usb/dwc3/dwc3-qcom.c
> > @@ -537,6 +537,16 @@ static int dwc3_qcom_of_register_core(struct platform_device *pdev)
> >                 return -ENODEV;
> >         }
> >  
> > +       /*
> > +        * A successful return from of_platform_populate() only guarantees that
> > +        * the core device was added to the platform bus, however it might not
> > +        * be bound to its driver (e.g. due to deferred probing). This driver
> > +        * requires the core device to be fully initialized, so defer probing
> > +        * if it isn't ready (yet).
> > +        */
> > +       if (!device_is_bound(&qcom->dwc3->dev))
> > +               return -EPROBE_DEFER;
> 
> Isn't this still broken? i.e. the dwc3 core driver may bind much later
> and then device_is_bound() will return an error here and then we'll
> return to the caller, dwc3_qcom_probe(), and that will depopulate the
> device with of_platform_depopulate(). It seems like we need to run some
> sort of wait for driver to be bound function instead of a one-shot check
> for the driver being bound.

My understanding is that the probing is done synchronously and either done,
failed or deferred when returning from of_platform_populate(). Ideally we
would be able to differentiate between a failure and deferral, and not defer
probing in case of an error, however I'm not aware of a way to do this with
the current driver framework.

The call flow is:

  of_platform_populate
    of_platform_bus_create
      of_platform_device_create_pdata
        of_device_add
	  device_add
	    bus_probe_device
	      device_initial_probe
	        __device_attach
	          __device_attach_driver
	            driver_probe_device
                      really_probe
                        ->probe()


> Also, what about acpi? That has the same problem but it isn't covered by
> the dwc3_qcom_of_register_core() function.

I wouldn't be surprised if it had the same problem. I'm not familiar with ACPI
though and don't have a device for testing, hence I limited the patch to the
platform device.