Re: [RFC PATCH V1 1/6] dt-bindings: mt8183: Add binding for DIP shared memory

[Rob Herring <robh@kernel.org>]

On Thu, May 16, 2019 at 1:12 AM Tomasz Figa <tfiga@chromium.org> wrote:
>
> On Wed, May 15, 2019 at 1:19 AM Rob Herring <robh@kernel.org> wrote:
> >
> > On Tue, May 7, 2019 at 9:22 AM Frederic Chen <frederic.chen@mediatek.com> wrote:
> > >
> > > Dear Rob,
> > >
> > > I appreciate your comments.
> > >
> > > On Mon, 2019-04-29 at 20:15 -0500, Rob Herring wrote:
> > > > On Wed, Apr 17, 2019 at 06:45:06PM +0800, Frederic Chen wrote:
> > > > > This patch adds the binding for describing the shared memory
> > > > > used to exchange configuration and tuning data between the
> > > > > co-processor and Digital Image Processing (DIP) unit of the
> > > > > camera ISP system on Mediatek SoCs.
> > > > >
> > > > > Signed-off-by: Frederic Chen <frederic.chen@mediatek.com>
> > > > > ---
> > > > >  .../mediatek,reserve-memory-dip_smem.txt      | 45 +++++++++++++++++++
> > > > >  1 file changed, 45 insertions(+)
> > > > >  create mode 100644 Documentation/devicetree/bindings/reserved-memory/mediatek,reserve-memory-dip_smem.txt
> > > > >
> > > > > diff --git a/Documentation/devicetree/bindings/reserved-memory/mediatek,reserve-memory-dip_smem.txt b/Documentation/devicetree/bindings/reserved-memory/mediatek,reserve-memory-dip_smem.txt
> > > > > new file mode 100644
> > > > > index 000000000000..64c001b476b9
> > > > > --- /dev/null
> > > > > +++ b/Documentation/devicetree/bindings/reserved-memory/mediatek,reserve-memory-dip_smem.txt
> > > > > @@ -0,0 +1,45 @@
> > > > > +Mediatek DIP Shared Memory binding
> > > > > +
> > > > > +This binding describes the shared memory, which serves the purpose of
> > > > > +describing the shared memory region used to exchange data between Digital
> > > > > +Image Processing (DIP) and co-processor in Mediatek SoCs.
> > > > > +
> > > > > +The co-processor doesn't have the iommu so we need to use the physical
> > > > > +address to access the shared buffer in the firmware.
> > > > > +
> > > > > +The Digital Image Processing (DIP) can access memory through mt8183 IOMMU so
> > > > > +it can use dma address to access the memory region.
> > > > > +(See iommu/mediatek,iommu.txt for the detailed description of Mediatek IOMMU)
> > > > > +
> > > > > +
> > > > > +Required properties:
> > > > > +
> > > > > +- compatible: must be "mediatek,reserve-memory-dip_smem"
> > > >
> > > > Don't use '_'.
> > >
> > > I got it. I will use "mediatek,reserve-memory-dip-smem" instead in next
> > > version of the patch
> > >
> > > >
> > > > > +
> > > > > +- reg: required for static allocation (see reserved-memory.txt for
> > > > > +  the detailed usage)
> > > > > +
> > > > > +- alloc-range: required for dynamic allocation. The range must
> > > > > +  between 0x00000400 and 0x100000000 due to the co-processer's
> > > > > +  addressing limitation
> > > >
> > > > Generally, you should pick either static or dynamic allocation for a
> > > > given binding. Static if there's some address restriction or sharing,
> > > > dynamic if not.
> > > >
> > > > Sounds like static in this case.
> > > >
> > >
> > > DIP reserved memory has address restriction so it is the static case. I
> > > would like to remove the dynamic allocation part and modify the
> > > description as following:
> > >
> > > - reg: required for DIP. The range must be between 0x00000400 and
> > >   0x100000000 due to the co-processor's addressing limitation.
> > >   The size must be 26MB. Please see reserved-memory.txt for the
> > >   detailed usage.
> >
> > You can use dma-ranges to define addressing translations and
> > restrictions like this. That will in turn set the device's dma-mask to
> > ensure allocations are done in a region that is addressable.
> >
> > But if you have a known, fixed size, then a carve out with
> > reserved-memory is fine.
>
> There is also another aspect here. The coprocessor that we're
> allocating the memory for is behind an MPU that must be programmed
> completely in one go and locked for security reasons to ensure that
> the coprocessor itself doesn't rewrite the MPU settings. That means
> that we need to have all the allocations completed before booting that
> coprocessor.
>
> To be honest, I'd adopt a completely different design here.
>
> We're going to have a driver for that coprocessor (SCP) and IMHO any
> shared memory areas should belong to it. Then, any specific drivers
> talking to the firmware running on SCP should ask the SCP driver to
> allocate some memory from its fixed pool. WDYT?

That's more than just an address restriction, so yeah, use a
reserved-memory area.

Rob