"magic" handling of memory nodes

"magic" handling of memory nodes

[Leif Lindholm]

Hi,

Following on the special handling of nodes called memory@0, I went to
have a look at the various platforms that do not actually declare a
device_type = "memory" for their "memory" nodes.

Firstly, we currently have 162(ish, I did a sloppy grep) such .dts{i}
files in the kernel tree.

Secondly, the only reason these platforms could ever have worked is
because they include .dtsi files that define a memory node with a
type explicitly set. Since this node already exists, its contents get
overridden, but the type tag remains. Of course, this only happens
with nodes called explicitly "memory" - but it happens regardless of
what other things they contain.

In the ARM tree, most of these seem to stem from the inclusion of
skeleton.dtsi.

I don't really know what could/should be done about this, but it
does not feel optimal.

/
	Leif
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

"magic" handling of memory nodes

[Leif Lindholm]

Hi,

Following on the special handling of nodes called memory at 0, I went to
have a look at the various platforms that do not actually declare a
device_type = "memory" for their "memory" nodes.

Firstly, we currently have 162(ish, I did a sloppy grep) such .dts{i}
files in the kernel tree.

Secondly, the only reason these platforms could ever have worked is
because they include .dtsi files that define a memory node with a
type explicitly set. Since this node already exists, its contents get
overridden, but the type tag remains. Of course, this only happens
with nodes called explicitly "memory" - but it happens regardless of
what other things they contain.

In the ARM tree, most of these seem to stem from the inclusion of
skeleton.dtsi.

I don't really know what could/should be done about this, but it
does not feel optimal.

/
	Leif

Re: "magic" handling of memory nodes

[Stephen Warren]

On 04/24/2014 05:33 AM, Leif Lindholm wrote:
> Hi,
> 
> Following on the special handling of nodes called memory@0, I went to
> have a look at the various platforms that do not actually declare a
> device_type = "memory" for their "memory" nodes.
> 
> Firstly, we currently have 162(ish, I did a sloppy grep) such .dts{i}
> files in the kernel tree.
> 
> Secondly, the only reason these platforms could ever have worked is
> because they include .dtsi files that define a memory node with a
> type explicitly set. Since this node already exists, its contents get
> overridden, but the type tag remains. Of course, this only happens
> with nodes called explicitly "memory" - but it happens regardless of
> what other things they contain.

That's precisely how DT includes/overrides are supposed to work, and is
entirely expected and normal.

Since skeleton.dtsi already says:

        memory { device_type = "memory"; reg = <0 0>; };

... then any .dts which includes that already has the device_type
property set, so there's no need to repeat that property. Subsequent
changes to /memory/reg have no impact on /memory/device_type; any new
node definitions simply over-write any previous definitions of a
redefined property, but leave unmentioned properties unchanged (unless
you /delete-property/).

If skeleton.dtsi were changed to remove that property then yes a lot of
files would then need to set it, but why would it be removed?
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

"magic" handling of memory nodes

[Stephen Warren]

On 04/24/2014 05:33 AM, Leif Lindholm wrote:
> Hi,
> 
> Following on the special handling of nodes called memory at 0, I went to
> have a look at the various platforms that do not actually declare a
> device_type = "memory" for their "memory" nodes.
> 
> Firstly, we currently have 162(ish, I did a sloppy grep) such .dts{i}
> files in the kernel tree.
> 
> Secondly, the only reason these platforms could ever have worked is
> because they include .dtsi files that define a memory node with a
> type explicitly set. Since this node already exists, its contents get
> overridden, but the type tag remains. Of course, this only happens
> with nodes called explicitly "memory" - but it happens regardless of
> what other things they contain.

That's precisely how DT includes/overrides are supposed to work, and is
entirely expected and normal.

Since skeleton.dtsi already says:

        memory { device_type = "memory"; reg = <0 0>; };

... then any .dts which includes that already has the device_type
property set, so there's no need to repeat that property. Subsequent
changes to /memory/reg have no impact on /memory/device_type; any new
node definitions simply over-write any previous definitions of a
redefined property, but leave unmentioned properties unchanged (unless
you /delete-property/).

If skeleton.dtsi were changed to remove that property then yes a lot of
files would then need to set it, but why would it be removed?

Re: "magic" handling of memory nodes

[Leif Lindholm]

On Thu, Apr 24, 2014 at 10:57:03AM -0600, Stephen Warren wrote:
> > Secondly, the only reason these platforms could ever have worked is
> > because they include .dtsi files that define a memory node with a
> > type explicitly set. Since this node already exists, its contents get
> > overridden, but the type tag remains. Of course, this only happens
> > with nodes called explicitly "memory" - but it happens regardless of
> > what other things they contain.
> 
> That's precisely how DT includes/overrides are supposed to work, and is
> entirely expected and normal.

Understood.

> Since skeleton.dtsi already says:
> 
>         memory { device_type = "memory"; reg = <0 0>; };
> 
> ... then any .dts which includes that already has the device_type
> property set, so there's no need to repeat that property. Subsequent
> changes to /memory/reg have no impact on /memory/device_type; any new
> node definitions simply over-write any previous definitions of a
> redefined property, but leave unmentioned properties unchanged (unless
> you /delete-property/).
> 
> If skeleton.dtsi were changed to remove that property then yes a lot of
> files would then need to set it, but why would it be removed?

Thank you for the explanation.

All of the device trees with memory@ nodes do explicitly specify the
device_type.

I do still find the setup (as opposed to the mechanism) somewhat
counterintuituve. The effect is pretty much that of a potentially
architecture-specific magic keyword.

/
    Leif
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

"magic" handling of memory nodes

[Leif Lindholm]

On Thu, Apr 24, 2014 at 10:57:03AM -0600, Stephen Warren wrote:
> > Secondly, the only reason these platforms could ever have worked is
> > because they include .dtsi files that define a memory node with a
> > type explicitly set. Since this node already exists, its contents get
> > overridden, but the type tag remains. Of course, this only happens
> > with nodes called explicitly "memory" - but it happens regardless of
> > what other things they contain.
> 
> That's precisely how DT includes/overrides are supposed to work, and is
> entirely expected and normal.

Understood.

> Since skeleton.dtsi already says:
> 
>         memory { device_type = "memory"; reg = <0 0>; };
> 
> ... then any .dts which includes that already has the device_type
> property set, so there's no need to repeat that property. Subsequent
> changes to /memory/reg have no impact on /memory/device_type; any new
> node definitions simply over-write any previous definitions of a
> redefined property, but leave unmentioned properties unchanged (unless
> you /delete-property/).
> 
> If skeleton.dtsi were changed to remove that property then yes a lot of
> files would then need to set it, but why would it be removed?

Thank you for the explanation.

All of the device trees with memory@ nodes do explicitly specify the
device_type.

I do still find the setup (as opposed to the mechanism) somewhat
counterintuituve. The effect is pretty much that of a potentially
architecture-specific magic keyword.

/
    Leif