Re: [PATCH v2 1/2] arm64: dts: ti: k3-am65: disable optional peripherals by default

[Matthias Schiffer]

On Mon, 2022-02-07 at 13:25 +0200, Tony Lindgren wrote:
> * Matthias Schiffer <matthias.schiffer@xxxxxxxxxxxxxxx> [220207
> 08:45]:
> > Generally I think that it's a bootloader's responsiblity to disable
> > unneeded devices - the kernel may not even have a driver for some
> > peripherals, leading to the same behaviour as a "disabled" status.
> > For
> > this reason I believe that it should always be okay to set unneeded
> > devices to "disabled", and it should be considered a safe default.
> 
> Not possible, think kexec for example :) How would the previous
> kernel
> even know what to disable if Linux has no idea about the devices?

Well, optimally, bootloader and all kernels would agree on the devices
that are actually available, but I get your point.

> 
> If there are issues you're seeing, it's likely a bug in some of the
> device drivers for not checking for the necessary resources like
> pinctrl for i2c lines.

I don't think it's common for individual drivers to care about pinctrl
unless switching between different pin settings is required at runtime.
Many drivers can be used on different hardware, some of which may
require pinmuxing, while others don't.

Also, what is the expected behavior of a driver that is probed for an
unusable device? Wouldn't this require some as-of-yet nonexisting
status between "okay" and "disabled" that conveys something like "probe
this device, initialize (and disable) PM, but don't register anything",
so no unusable devices become visible to userspace (and possibly other
kernel drivers)?

> 
> > I'm not sure what the consensus on these issues is. I'm more
> > familiar
> > with NXP's i.MX and Layerscape SoCs, where it's common to have all
> > muxable peripherals set to "disabled" in the base DTSI, and a quick
> > grep through a few dts directories gives me the impression that
> > this is
> > the case for most other vendors as well.
> 
> This approach only works for SoCs that don't need the kernel to idle
> devices for runtime PM.

I'm pretty sure that most modern SoCs I looked at have runtime PM, and
it is simply expected that unusable devices are never enabled in the
first place, so there is no need for the kernel to know about them.

Regards,
Matthias




> Regards,
> 
> Tony