Re: [PATCH v2 1/4] dt-bindings: reserved-memory: Document "active" property

[Thierry Reding]

On Thu, Sep 24, 2020 at 04:23:59PM +0300, Dmitry Osipenko wrote:
> 04.09.2020 15:59, Thierry Reding пишет:
> > From: Thierry Reding <treding@xxxxxxxxxx>
> > 
> > Reserved memory regions can be marked as "active" if hardware is
> > expected to access the regions during boot and before the operating
> > system can take control. One example where this is useful is for the
> > operating system to infer whether the region needs to be identity-
> > mapped through an IOMMU.
> > 
> > Signed-off-by: Thierry Reding <treding@xxxxxxxxxx>
> > ---
> >  .../bindings/reserved-memory/reserved-memory.txt           | 7 +++++++
> >  1 file changed, 7 insertions(+)
> > 
> > diff --git a/Documentation/devicetree/bindings/reserved-memory/reserved-memory.txt b/Documentation/devicetree/bindings/reserved-memory/reserved-memory.txt
> > index 4dd20de6977f..163d2927e4fc 100644
> > --- a/Documentation/devicetree/bindings/reserved-memory/reserved-memory.txt
> > +++ b/Documentation/devicetree/bindings/reserved-memory/reserved-memory.txt
> > @@ -63,6 +63,13 @@ reusable (optional) - empty property
> >        able to reclaim it back. Typically that means that the operating
> >        system can use that region to store volatile or cached data that
> >        can be otherwise regenerated or migrated elsewhere.
> > +active (optional) - empty property
> > +    - If this property is set for a reserved memory region, it indicates
> > +      that some piece of hardware may be actively accessing this region.
> > +      Should the operating system want to enable IOMMU protection for a
> > +      device, all active memory regions must have been identity-mapped
> > +      in order to ensure that non-quiescent hardware during boot can
> > +      continue to access the memory.
> >  
> >  Linux implementation note:
> >  - If a "linux,cma-default" property is present, then Linux will use the
> > 
> 
> Hi,
> 
> Could you please explain what devices need this quirk? I see that you're
> targeting Tegra SMMU driver, which means that it should be some pre-T186
> device.

Primarily I'm looking at Tegra210 and later, because on earlier devices
the bootloader doesn't consistently initialize display. I know that it
does on some devices, but not all of them. This same code should also
work on Tegra186 and later (with an ARM SMMU) although the situation is
slightly more complicated there because IOMMU translations will fault by
default long before these identity mappings can be established.

> Is this reservation needed for some device that has display
> hardwired to a very specific IOMMU domain at the boot time?

No, this is only used to convey information about the active framebuffer
to the kernel. In practice the DMA/IOMMU code will use this information
to establish a 1:1 mapping on whatever IOMMU domain that was picked for
display.

> If you're targeting devices that don't have IOMMU enabled by default at
> the boot time, then this approach won't work for the existing devices
> which won't ever get an updated bootloader.

If the devices don't use an IOMMU, then there should be no problem. The
extra reserved-memory nodes would still be necessary to ensure that the
kernel doesn't reuse the framebuffer memory for the slab allocator, but
if no IOMMU is used, then the display controller accessing the memory
isn't going to cause problems other than perhaps scanning out data that
is no longer a framebuffer.

There should also be no problem for devices with an old bootloader
because this code is triggered by the presence of a reserved-memory node
referenced via the memory-region property. Devices with an old
bootloader should continue to work as they did before. Although I
suppose they would start faulting once we enable DMA/IOMMU integration
for Tegra SMMU if they have a bootloader that does initialize display to
actively scan out during boot.

> I think Robin Murphy already suggested that we should simply create
> a dummy "identity" IOMMU domain by default for the DRM/VDE devices and
> then replace it with an explicitly created domain within the drivers.

I don't recall reading about that suggestion. So does this mean that for
certain devices we'd want to basically passthrough by default and then
at some point during boot take over with a properly managed IOMMU
domain?

The primary goal here is to move towards using the DMA API rather than
the IOMMU API directly, so we don't really have the option of replacing
with an explicitly created domain. Unless we have code in the DMA/IOMMU
code that does this somehow.

But I'm not sure what would be a good way to mark certain devices as
needing an identity domain by default. Do we still use the reserved-
memory node for that? That would still require some sort of flag to
specify which reserved-memory regions would need this identity mapping
because, as was pointed out in earlier review, some devices may have
reserved-memory regions that are not meant to be identity mapped.

> Secondly, all NVIDIA bootloaders are passing tegra_fbmem=... via
> kernel's cmdline with the physical location of the framebuffer in
> memory. Maybe we could support this option?

I'm not a big fan of that command-line option, but I also realize that
for older bootloaders that's probably the only option we have. I don't
suppose all of the devices support U-Boot? Because ideally we'd just
translate from tegra_fbmem=... to reserved-memory region there so that
we don't have to carry backwards-compatibility code for these purely
downstream bootloaders.

Thierry

Attachment: signature.asc
Description: PGP signature