[PATCH v2 1/5] drivers: of: add initialization code for reserved memory

[Grant Likely]

On Thu, 13 Feb 2014 13:48:40 -0600, Josh Cartwright <joshc at codeaurora.org> wrote:
> On Tue, Feb 11, 2014 at 09:27:36PM +0100, Tomasz Figa wrote:
> > On 11.02.2014 21:19, Josh Cartwright wrote:
> > >On Tue, Feb 11, 2014 at 09:04:21PM +0100, Tomasz Figa wrote:
> > > >On 11.02.2014 21:02, Benjamin Herrenschmidt wrote:
> > > > >On Tue, 2014-02-11 at 19:01 +0000, Grant Likely wrote:
> > > > > > > except that the former IMHO better suits the definition of memory
> > > > > > > region, which I see as a single contiguous range of memory and can be
> > > > > > > simplified to have a single reg entry per region.
> > > > > >
> > > > > > My point is rather if multiple reg tuples are found in a reserved memory
> > > > > > node, the kernel must respect them and reserve the memory. I'm not
> > > > > > arguing about whether or not that makes for a good binding.
> > > > >
> > > > > agreed.
> > > >
> > > > My point is why, if the binding defines that just a single tuple should be
> > > > provided.
> > >
> > > FWIW, the usecase I had mentioned in reply to Grant in the patch 5/5
> > > thread [1] could make use of this.  The shared memory region is split
> > > into a main chunk and several "auxiliary" chunk, but collectively these
> > > regions all share the same heap state.
> > >
> > >   Josh
> > >
> > > 1: http://lkml.kernel.org/r/20140205192502.GO20228@joshc.qualcomm.com
> >
> > The use case seems fine, but I believe it could be properly represented in
> > device tree using multiple single-reg regions as well, unless the consumer
> > can request a block of memory that crosses boundary of two sub-regions
> > specified by reg entries of single region.
> 
> I could probably make a only-one-reg-entry policy work for me, but it
> makes things a bit more awkward.  I'd lose the ability to describe
> "this set of regions need to be logically handled together" directly in
> the reserved memory node, and would need to push it up a layer.
> 
> 	reserved-memory {
> 		smem: smem {
> 			reg = <...>;
> 		};
> 		aux1: auxiliary1 {
> 			reg = <...>;
> 		};
> 		aux2: auxiliary2 {
> 			reg = <...>;
> 		};
> 		...
> 	};

If the regions are used for different purposes, it makes sense I think
to have a separate node for each. Multiple tuples would make more sense
for something like valid DMA regions for a broken device that can only
DMA into a few windows; you could have one tuple per window within a
single node.

It would be possible to collect multiple associated nodes under one
parent node which in turn has reserved-memory for its parent:

	reserved-memory {
		ranges;
		reserved-group {
			ranges;
			smem: smem {
				reg = <...>;
			};
			aux1: auxiliary1 {
				reg = <...>;
			};
			aux2: auxiliary2 {
				reg = <...>;
			};
		};
		...
	};

g.