[PATCH 1/2] ohci-platform: Add support for devicetree instantiation

[Arnd Bergmann]

On Monday 06 January 2014, Hans de Goede wrote:
> On 01/06/2014 08:16 AM, Arnd Bergmann wrote:
> > On Monday 06 January 2014, Hans de Goede wrote:
> >> +Required properties:
> >> + - compatible: Should be "platform-ohci"
> >> + - reg: Address range of the ohci registers.
> >> + - interrupts: Should contain the ohci interrupt.
> >> +
> >> +Optional properties:
> >> + - clocks: array of clocks
> >> + - clock-names: clock names "ahb" and/or "ohci"
> >> + - phys: phy
> >> + - phy-names: "usb_phy"
> >
> > Maybe just "usb"? It obviously is a phy, so that part of the name
> > is a bit redundant. Actually, the "usb" part of the name also seems
> > redundant. Is it possible to make it an anonymous phy reference
> > without a phy-names property as we often do for clocks?
> 
> I'm not a big fan of anonymous references, IE currently the ahci-platform
> driver is using an anonymous clk reference, but for sunxi (and ie imx too)
> it needs to be extended to 2 clks, which means using names, while
> keeping compatibility with the older anonymous using dts (and board)
> files.
> 
> So I can agree to dropping the _phy, but I would like to keep the "usb"
> name I realize the chances are slim of ever needing 2 phys but never
> say never ...

Ok, fair enough.

> > Style-wise, I think I'd prefer not storing error values in the
> > ohci_platform_priv struct, but rather using NULL pointers for
> > when the clocks or phy references are unused.
> 
> This is a style I picked up from the mmc code, ie do a grep for
> !IS_ERR in drivers/mmc/host/*.c, but I agree it is not the prettiest,
> and on looking in other subsystems it is not common, so I'll convert
> this to storing NULL on error to get the phy or clk.

Ok, thanks.

> >> @@ -83,17 +171,30 @@ static int ohci_platform_probe(struct platform_device *dev)
> >>   		return -ENXIO;
> >>   	}
> >>
> >> +	hcd = usb_create_hcd(&ohci_platform_hc_driver, &dev->dev,
> >> +			dev_name(&dev->dev));
> >> +	if (!hcd)
> >> +		return -ENOMEM;
> >> +
> >> +	if (pdata == &ohci_platform_defaults) {
> >> +		struct ohci_platform_priv *priv = (struct ohci_platform_priv *)
> >> +						  hcd_to_ohci(hcd)->priv;
> >> +
> >> +		priv->phy = devm_phy_get(&dev->dev, "usb_phy");
> >> +		if (IS_ERR(priv->phy) && PTR_ERR(priv->phy) == -EPROBE_DEFER) {
> >> +			err = -EPROBE_DEFER;
> >> +			goto err_put_hcd;
> >> +		}
> >> +
> >> +		priv->ohci_clk = devm_clk_get(&dev->dev, "ohci");
> >> +		priv->ahb_clk = devm_clk_get(&dev->dev, "ahb");
> >> +	}
> >
> > I think you have to check the clocks for -EPROBE_DEFER as well here, not
> > just the PHY.
> 
> So far I've never seen clk_get return -EPROBE_DEFER, but I guess on some
> platforms it can.

Right. Most clock controllers tend to be initialized at early boot
time, but some of them are on external PMICs that are only probed after
i2c or some other subsystem is up.

> > Otherwise you don't know the difference between "no clock
> > provided", "error getting the clock reference" and "clock controller not
> > initialized yet, try again".
> 
> I guess of these 3 we really only want to continue on "no clock provided",
> so I think something like this (for both clks and the phy) would be best:
> 
> 		priv->ahb_clk = devm_clk_get(&dev->dev, "ahb");
> 		if (IS_ERR(priv->ahb_clk)) {
> 			err = PTR_ERR(priv->ahb_clk);
> 			if (err != -EINVAL && err != -ENODATA)
> 				goto err_put_hcd;
> 			priv->ahb_clk = NULL; /* No clock provided */
> 		}
> 
> To clarify -EINVAL will be returned when there is no clock-names, and
> -ENODATA when the specified name is not found in clock-names.

Ok, looks good.

> >> @@ -178,6 +277,12 @@ static int ohci_platform_resume(struct device *dev)
> >>   #define ohci_platform_resume	NULL
> >>   #endif /* CONFIG_PM */
> >>
> >> +static const struct of_device_id ohci_platform_ids[] = {
> >> +	{ .compatible = "platform-ohci", },
> >> +	{ }
> >> +};
> >> +MODULE_DEVICE_TABLE(of, ohci_platform_ids);
> >
> > I never liked the "platform-*" naming for compatible properties, the
> > name of the driver is just based on a linux implementation detail
> > (the platform bus), which could change. How about just calling the
> > generic one "ohci" or "usb-ohci"?
> 
> usb-ohci seems free but usb-ehci are already taken by
> drivers/usb/host/ehci-ppc-of.c which is as the name implies pretty ppc
> specific. And since ehci-platform.c can be build on ppc too, we could
> end up with 2 drivers claiming the same compatibility string on ppc.
> 
> Also uhci-platform.c is already using platform-uhci, so using
> ohci-platform and ehci-platform seems consistent and avoids the
> conflict with drivers/usb/host/ehci-ppc-of.c

Hmm, that file seems to do two things that ehci-platform doesn't:
It handles big-endian controllers and it has special initialization
for compatible="ibm,usb-ehci-440epx". It also uses a different way
to get to the resources because the driver dates back to before the
unification of platform_bus and of_platform_bus, but that can be
trivially changed.

I would hope that we can eventually merge the two drivers, and I'm
guessing that we will sooner or later need the big-endian support on
non-ppc machines anyway. The question is what to do about the 440epx
hack. It may be small enough that we can just leave it as a quirk
inside #ifdef CONFIG_PPC in the unified driver, or it may end up
being shared with arm64 X-Gene, which seems to share a common
ancestry with some of the ppc4xx socs.

	Arnd