Re: [PATCH v5 1/2] dt-bindings: reserved-memory: Introduce carved-out memory region binding

[Maxime Ripard <mripard@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 4369 bytes --]

Hi Rob,

On Tue, Jul 01, 2025 at 09:12:18AM +0200, Maxime Ripard wrote:
> On Mon, Jun 30, 2025 at 05:08:19PM -0500, Rob Herring wrote:
> > On Mon, Jun 30, 2025 at 06:41:38PM +0200, Maxime Ripard wrote:
> > > Hi Rob,
> > > 
> > > On Fri, Jun 27, 2025 at 02:31:32PM -0500, Rob Herring wrote:
> > > > On Tue, Jun 17, 2025 at 02:25:40PM +0200, Maxime Ripard wrote:
> > > > > Some parts of the memory can be dedicated to specific purposes and
> > > > > exposed as a dedicated memory allocator.
> > > > > 
> > > > > This is especially useful if that particular region has a particular
> > > > > properties the rest of the memory doesn't have. For example, some
> > > > > platforms have their entire RAM covered by ECC but for a small area
> > > > > meant to be used by applications that don't need ECC, and its associated
> > > > > overhead.
> > > > > 
> > > > > Let's introduce a binding to describe such a region and allow the OS to
> > > > > create a dedicated memory allocator for it.
> > > > > 
> > > > > Signed-off-by: Maxime Ripard <mripard@kernel.org>
> > > > > ---
> > > > >  .../bindings/reserved-memory/carved-out.yaml       | 49 ++++++++++++++++++++++
> > > > >  1 file changed, 49 insertions(+)
> > > > > 
> > > > > diff --git a/Documentation/devicetree/bindings/reserved-memory/carved-out.yaml b/Documentation/devicetree/bindings/reserved-memory/carved-out.yaml
> > > > > new file mode 100644
> > > > > index 0000000000000000000000000000000000000000..9ab5d1ebd9ebd9111b7c064fabe1c45e752da83b
> > > > > --- /dev/null
> > > > > +++ b/Documentation/devicetree/bindings/reserved-memory/carved-out.yaml
> > > > > @@ -0,0 +1,49 @@
> > > > > +# SPDX-License-Identifier: GPL-2.0-only OR BSD-2-Clause
> > > > > +%YAML 1.2
> > > > > +---
> > > > > +$id: http://devicetree.org/schemas/reserved-memory/carved-out.yaml#
> > > > > +$schema: http://devicetree.org/meta-schemas/core.yaml#
> > > > > +
> > > > > +title: Carved-out Memory Region
> > > > > +
> > > > > +description: |
> > > > 
> > > > Don't need '|'.
> > > > 
> > > > > +  Specifies that the reserved memory region has been carved out of the
> > > > > +  main memory allocator, and is intended to be used by the OS as a
> > > > > +  dedicated memory allocator.
> > > > 
> > > > Other than the commit msg, it is completely lost that this is for 
> > > > ECC-less memory.
> > > 
> > > Because it's not. One of the first feedback I got was that the way to
> > > identify what a heap provides was the heap name.
> > > 
> > > So, as far as the binding go, a heap just exposes a chunk of memory the
> > > memory allocator wouldn't use. The actual semantics of that chunk of
> > > memory don't matter.
> > 
> > But they do because you use one carve out for one thing and another 
> > carve out for another purpose and they probably aren't interchangeable.
> 
> That was also my initial thought, but it was then discussed that the
> name of the region is enough of a discriminant. And it makes sense too,
> it's a sufficient discriminant for the device tree to uniquely identify
> a given memory region on a given platform already, so we don't really
> need anything else.
> 
> > For the most part, everything in /reserved-memory is a carve out from 
> > regular memory though we failed to enforce that.
> > 
> > > > This description applies to CMA area as well. So what's the difference?
> > > 
> > > Yeah, I kind of agree, which is why I initially started with a property,
> > > and you then asked for a compatible.
> > 
> > My issues with properties is we have to support N factorial cases for 
> > combinations of N properties. It's already fragile. Whereas a compatible 
> > is (hopefully) well defined as to what's needed and is only 1 more case 
> > to support.
> 
> I think that's also what John especially wanted to avoid. If we have a
> generic compatible, but the attributes/properties/whatever of the
> buffers allocated from that region differ (like ecc vs non-ecc,
> protected vs non-protected, etc.) we will need properties in the device
> tree to describe them too.

I thought about it some more over the weekend, but if a compatible isn't
a great solution either, would it make sense to just create a heap for
each shared-dma-pool region, and thus we just wouldn't need any extra
compatible or property?

Maxime

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 273 bytes --]