Re: [PATCH v5 04/11] clocksource/drivers: Add HPE GXP timer

[Rob Herring <robh@kernel.org>]

On Tue, Apr 26, 2022 at 08:00:20AM +0200, Arnd Bergmann wrote:
> On Mon, Apr 25, 2022 at 10:38 PM Linus Walleij <linus.walleij@linaro.org> wrote:
> > On Fri, Apr 22, 2022 at 3:16 PM Arnd Bergmann <arnd@arndb.de> wrote:
> > > On Thu, Apr 21, 2022 at 9:21 PM <nick.hawkins@hpe.com> wrote:
> > >
> > > > +
> > > > +static struct platform_device gxp_watchdog_device = {
> > > > +       .name = "gxp-wdt",
> > > > +       .id = -1,
> > > > +};
> > > > +/*
> > > > + * This probe gets called after the timer is already up and running. This will create
> > > > + * the watchdog device as a child since the registers are shared.
> > > > + */
> > > > +
> > > > +static int gxp_timer_probe(struct platform_device *pdev)
> > > > +{
> > > > +       struct device *dev = &pdev->dev;
> > > > +
> > > > +       /* Pass the base address (counter) as platform data and nothing else */
> > > > +       gxp_watchdog_device.dev.platform_data = local_gxp_timer->counter;
> > > > +       gxp_watchdog_device.dev.parent = dev;
> > > > +       return platform_device_register(&gxp_watchdog_device);
> > > > +}
> > >
> > > I don't understand what this is about: the device should be created from
> > > DT, not defined statically in the code. There are multiple ways of creating
> > > a platform_device from a DT node, or you can allocate one here, but static
> > > definitions are generally a mistake.
> > >
> > > I see that you copied this from the ixp4xx driver, so I think we should fix this
> > > there as well.
> >
> > The ixp4xx driver looks like that because the register range used for
> > the timer and the watchdog is combined, i.e. it is a single IP block:
> >
> >                 timer@c8005000 {
> >                         compatible = "intel,ixp4xx-timer";
> >                         reg = <0xc8005000 0x100>;
> >                         interrupts = <5 IRQ_TYPE_LEVEL_HIGH>;
> >                 };
> >
> > Device tree probing does not allow two devices to probe from the same
> > DT node, so this was solved by letting the (less important) watchdog
> > be spawn as a platform device from the timer.
> >
> > I don't know if double-probing for the same register range can be fixed,
> > but I was assuming that the one-compatible-to-one-driver assumption
> > was pretty hard-coded into the abstractions. Maybe it isn't?
> 
> Having a child device is fine, my objection was about the way
> the device is created from a 'static platform_device ...' definition
> rather than having the device structure allocated at probe time.
> 
> > Another way is of course to introduce an MFD. That becomes
> > problematic in another way: MFD abstractions are supposed to
> > be inbetween the resource and the devices it spawns, and with
> > timers/clocksources this creates a horrible special-casing since the
> > MFD bus (the parent may be providing e.g. an MMIO regmap)
> > then need to be early-populated and searched by the timer core
> > from TIMER_OF_DECLARE() early in boot.
> >
> > So this solution was the lesser evil that I could think about.
> 
> There are multiple ways of doing this that we already discussed
> in the thread. The easiest is probably to have a child node without
> custom registers in the DT and then use the DT helpers to
> populate the linux devices with the correct data.

I think that's what the wdt binding is doing, but I don't like that. 
Maybe it's not a child node, I can't tell.

Bindings should not be decided on the *current* driver split on one 
particular OS. This looks like 1 block, so 1 node. If that doesn't work 
well or easy for Linux, then we should fix Linux.

Rob