[Intel-gfx] [PATCH] iommu/intel: quirk to disable DMAR for QM57 igfx

[Joonas Lahtinen]

Quoting Joerg Roedel (2019-01-22 18:51:35)
> On Tue, Jan 22, 2019 at 04:48:26PM +0200, Joonas Lahtinen wrote:
> > According to our IOMMU folks there exists some desire to be able to assign
> > the iGFX device aka have intel_iommu=on instead of intel_iommu=igfx_off
> > due to how the devices might be grouped in IOMMU groups. Even when you
> > would not be using the iGFX device.
> 
> You can force the igfx device into a SI domain, or does that also
> trigger the iommu issues on the chipset?

To be honest, we've had a mixture different issues on different SKUs
that have not been hit in the past when intel_iommu was just disabled by
default.

I know that in one group of the problems, the issue has been debugged
into the GPU having its own set of virtualization mapping translation
hardware with caching and it fails to track changes to the mapping. So
if a identity mapping was established and never changed, I'd assume that
to fix at least that class of problems.

Would just passing intel_iommu=on already cause a non-identity mapping to
possibly be used for the integrated GPU? If it did, then it would
explain quite few of the issues.

We have many reports where just having intel_iommu=on (and using the
system normally, without any virtualization stuff going on) will cause
unexplained GPU hangs. For those users, simply switching to
intel_iommu=igfx_off solves the problems, and the debug often ends
there.

Regards, Joonas

> In any case, if iommu=on breaks these systems I want to make them work
> again with opt-out, even at the cost of breaking assignability.
> 
> Regards,
> 
>         Joerg