On Wed, Feb 19, 2020 at 02:09:54PM +0000, Peter Chen wrote:
> On 20-02-19 15:38:15, Heikki Krogerus wrote:
> > On Wed, Feb 19, 2020 at 01:58:38AM +0000, Peter Chen wrote:
> > > On 20-02-18 14:25:45, Heikki Krogerus wrote:
> > > > Hi,
> > > >
> > > > On Tue, Feb 18, 2020 at 07:23:41AM +0000, Peter Chen wrote:
> > > > > > > @@ -1118,6 +1119,7 @@ static int ci_hdrc_probe(struct platform_device *pdev)
> > > > > > > }
> > > > > > >
> > > > > > > if (ci_role_switch.fwnode) {
> > > > > > > + ci_role_switch.driver_data = ci;
> > > > > > > ci->role_switch = usb_role_switch_register(dev,
> > > > > > > &ci_role_switch);
> > > > >
> > > > > Why the struct usb_role_switch_desc needs drvdata, the struct
> > > > > usb_role_switch has already one?
> > > >
> > > > I'm assuming that you are asking why not just register the switch,
> > > > and then call usb_role_switch_set_drvdata(), right?
> > >
> > > Yes.
> > >
> > > >
> > > > That may create a race condition where the switch is accessed before
> > > > the driver data is available. That can happen for example if the
> > > > switch is exposed to the user space.
> > > >
> > > > To play it safe, supplying the driver data as part of the descriptor.
> > > > That way we can be sure that the driver data is always available
> > > > the moment the switch is registered.
> > > >
> > >
> > > Then, you may use the uniform way for the driver. Some may have
> > > race condition like you said.
> >
> > Uniform way for the driver?
> >
>
> Sorry, unify way. Take chipidea and cdns3 as an example, at chipidea you
> use struct usb_role_switch_desc
>
> + ci_role_switch.driver_data = ci;
> ci->role_switch = usb_role_switch_register(dev,
> &ci_role_switch);
>
> But at cdns3 side, you set usb_role_switch drvdata directly.
> + usb_role_switch_set_drvdata(cdns->role_sw, cdns);
>
> But according to your comments, all the driver needs to use chipidea's
> way to avoid race condition.
OK, I'll fix that.
thanks,
--
heikki
