On Thu, Jun 04, 2020 at 04:16:31AM -0700, Stephen Boyd wrote:
> Quoting Sandeep Maheswaram (Temp) (2020-06-04 02:43:09)
> >
> > On 6/3/2020 11:06 PM, Stephen Boyd wrote:
> > > Quoting Sandeep Maheswaram (2020-03-31 22:15:43)
> > >> diff --git a/drivers/usb/dwc3/dwc3-qcom.c b/drivers/usb/dwc3/dwc3-qcom.c
> > >> index 1dfd024..d33ae86 100644
> > >> --- a/drivers/usb/dwc3/dwc3-qcom.c
> > >> +++ b/drivers/usb/dwc3/dwc3-qcom.c
> > >> @@ -285,6 +307,101 @@ static int dwc3_qcom_resume(struct dwc3_qcom *qcom)
> > >> return 0;
> > >> }
> > >>
> > >> +
> > >> +/**
> > >> + * dwc3_qcom_interconnect_init() - Get interconnect path handles
> > >> + * @qcom: Pointer to the concerned usb core.
> > >> + *
> > >> + */
> > >> +static int dwc3_qcom_interconnect_init(struct dwc3_qcom *qcom)
> > >> +{
> > >> + struct device *dev = qcom->dev;
> > >> + int ret;
> > >> +
> > >> + if (!device_is_bound(&qcom->dwc3->dev))
> > >> + return -EPROBE_DEFER;
> > > How is this supposed to work? I see that this was added in an earlier
> > > revision of this patch series but there isn't any mention of why
> > > device_is_bound() is used here. It would be great if there was a comment
> > > detailing why this is necessary. It sounds like maximum_speed is
> > > important?
> > >
> > > Furthermore, dwc3_qcom_interconnect_init() is called by
> > > dwc3_qcom_probe() which is the function that registers the device for
> > > qcom->dwc3->dev. If that device doesn't probe between the time it is
> > > registered by dwc3_qcom_probe() and this function is called then we'll
> > > fail dwc3_qcom_probe() with -EPROBE_DEFER. And that will remove the
> > > qcom->dwc3->dev device from the platform bus because we call
> > > of_platform_depopulate() on the error path of dwc3_qcom_probe().
> > >
> > > So isn't this whole thing racy and can potentially lead us to a driver
> > > probe loop where the wrapper (dwc3_qcom) and the core (dwc3) are probing
> > > and we're trying to time it just right so that driver for dwc3 binds
> > > before we setup interconnects? I don't know if dwc3 can communicate to
> > > the wrapper but that would be more of a direct way to do this. Or maybe
> > > the wrapper should try to read the DT property for maximum speed and
> > > fallback to a worst case high bandwidth value if it can't figure it out
> > > itself without help from dwc3 core.
> > >
> > This was added in V4 to address comments from Matthias in V3
> >
> > https://patchwork.kernel.org/patch/11148587/
> >
>
> Yes, that why I said:
>
> "I see that this was added in an earlier
> revision of this patch series but there isn't any mention of why
> device_is_bound() is used here. It would be great if there was a comment
> detailing why this is necessary. It sounds like maximum_speed is
> important?"
>
> Can you please respond to the rest of my email?
I agree with Stephen that using device_is_bound() isn't a good option
in this case, when I suggested it I wasn't looking at the big picture
of how probing the core driver is triggered, sorry about that.
Reading the speed from the DT with usb_get_maximum_speed() as Stephen
suggests would be an option, the inconvenient is that we then
essentially require the property to be defined, while the core driver
gets a suitable value from hardware registers. Not sure if the wrapper
driver could read from the same registers.
One option could be to poll device_is_bound() for 100 ms (or so), with
sleeps between polls. It's not elegant but would probably work if we
don't find a better solution.
