[PATCH v7 3/4] drivers: of: add initialization code for dma reserved memory

[Grant Likely]

On Tue, 10 Sep 2013 14:53:29 -0500, Kumar Gala <galak at codeaurora.org> wrote:
> 
> On Sep 9, 2013, at 11:01 AM, Grant Likely wrote:
> 
> > On Fri, 30 Aug 2013 15:26:24 -0500, Kumar Gala <galak at codeaurora.org> wrote:
> >> On Aug 30, 2013, at 7:39 AM, Marek Szyprowski wrote:
> >>> On 8/30/2013 12:46 AM, Grant Likely wrote:
> >>>> On Mon, 26 Aug 2013 16:39:18 +0200, Marek Szyprowski <m.szyprowski at samsung.com> wrote:
> >>>>> +	- "reserved-memory-region" - compatibility is defined, given
> >>>>> +	  region is assigned for exclusive usage for by the respective
> >>>>> +	  devices.
> >>>> 
> >>>> BTW, just so you're aware there is already a binding for marking regions
> >>>> as reserved. It was recently added to arch/powerpc/kernel/prom.c.
> >>>> Unfortunately it doesn't look like it got documented. Search for
> >>>> "reserved-ranges". However, I don't think it blocks your work here. That
> >>>> binding doesn't provide any way for matching devices to reserved ranges.
> >>>> It is only for telling the kernel "hands of that memory".
> >>> 
> >>> ok, good to know.
> >> 
> >> So I think the most generic compatible should effectively be the same
> >> as 'reserved-ranges', ie this region shouldn't be touched by the
> >> kernel.
> >> 
> >> Than we can build on top of that with more specific compatibles.
> >> 
> >> I have examples from FSL networking SoCs where we would carve out
> >> memory for backing storage managed completely by the HW and Linux
> >> wouldn't need to touch it, however we might have a
> >> "fsl,qman-backing-store" compat encase we might want some debug code
> >> in linux to look at the region of memory.
> >> 
> >> I've got examples at qualcomm where we have shared data structures in
> >> specific memory regions between different processors that aren't cache
> >> coherent, so want the memory not to be marked as cacheable by Linux
> >> when it accesses it.  Again we'd have a "qcom,smem" prop on top of the
> >> "reserved-memory-region".
> > 
> > I think that is a reasonable approach, and works really well for regions
> > associated with specific hardware because the hardware driver can be
> > responsible for wiring it up to CMA or manage the region directly.
> > Whatever the driver desires to do.
> > 
> > It still remains what to do for generic regions that any device can use.
> > As I've previously said, I don't like calling out "CMA" specifically in
> > the compatible property because that happens to be a Linux
> > implementation specific details. Two years from now someone may propose
> > a new allocator that should take over what used to be handled by CMA
> > (kind of like how CMA may end up taking over from ION)....
> > 
> > Alright, I've thought about it some more and it probably does make sense
> > to use an additional compatible string. Marek originally suggested
> > "linux,contiguous-memory-region". I later suggested "shareable-memory-region",
> > but that's actually a worse name because it doesn't have any reference
> > to what the region is for (I was fixating on the kernel being able to
> > use unallocated pages; but that is also an implementation detail). I
> > exercise my right to change my mind and agree that contiguous is
> > appropriate here. But instead of calling out the contiguous allocator,
> > the binding should specifically lay out the usage model for these
> > regions. I would change the contiguous-memory binding to state the
> > following:
> > 	Contiguous-memory is a region of general purpose memory from
> > 	which large buffers of contiguous memory can be allocated.
> > 	Contiguous buffers are often used for DMA buffers. The operating
> > 	system may use the memory for any purpose, but must immediately
> > 	release the pages if a contiguous buffer is required.
> > 
> > So the way I read things, the current proposal would be:
> > 
> > The generic form for do-not-touch memory:
> > 	compatible = "reserved-memory";
> > - I've dropped '-region' because I think it is redundant
> > - The kernel will not make use of this memory except for when a driver picks it up
> > 
> > For contiguous memory:
> > 	compatible = "contiguous-memory", "reserved-memory"
> 
> Hmm.., what's an example of a generic use of this compatible set that isn't linux/CMA specific?

My point was to decouple it from the Linux specific CMA implementation.
I was talking to Jesse Barker about a binding for ION regions (which
should eventually move to CMA). He needs pretty much the same binding.

> 
> > For contiguous memory that Linux will use by default if a specific
> > region isn't specified.
> > 	compatible = "contiguous-memory", "reserved-memory"
> > 	linux,default-contiguous-region;
> > - I stuck with a linux, prefix here because I haven't come up with a
> >   non-linux way to describe this.
> > 
> > Memory that specific hardware can pick up:
> > 	compatible = "qcom,smem", "reserved-memory"
> > 
> > Nodes with a 'size' property and without a 'reg' property need to have a
> > region allocated and reserved. The allocated region needs to be cached
> > somewhere. We could create a data structure for tracking the reserved
> > regions, or could write it in directly. While I shudder at the thought of
> > modifying the device tree at runtime by the kernel, it might just be the
> > sanest implementation at this time. I'd like to see what the
> > implementation looks like. (Since this is potentially controversial, I
> > would recommend implementing the exact regions in on patch, and add
> > dynamically allocated regions as a follow-up patch)
> 
> curious, what's the example for this situation of not having a 'reg' but having a size?

The way it was described to me was to allocate a region, but not care
where it actually existed in RAM. I don't know how important a
requirement it is.

g.