[Intel-gfx] i915 dma faults on Xen

[Roger Pau Monné]

On Fri, Feb 19, 2021 at 12:30:23PM -0500, Jason Andryuk wrote:
> On Wed, Oct 21, 2020 at 9:59 AM Jan Beulich <jbeulich at suse.com> wrote:
> >
> > On 21.10.2020 15:36, Jason Andryuk wrote:
> > > On Wed, Oct 21, 2020 at 8:53 AM Jan Beulich <jbeulich at suse.com> wrote:
> > >>
> > >> On 21.10.2020 14:45, Jason Andryuk wrote:
> > >>> On Wed, Oct 21, 2020 at 5:58 AM Roger Pau Monné <roger.pau at citrix.com> wrote:
> > >>>> Hm, it's hard to tell what's going on. My limited experience with
> > >>>> IOMMU faults on broken systems there's a small range that initially
> > >>>> triggers those, and then the device goes wonky and starts accessing a
> > >>>> whole load of invalid addresses.
> > >>>>
> > >>>> You could try adding those manually using the rmrr Xen command line
> > >>>> option [0], maybe you can figure out which range(s) are missing?
> > >>>
> > >>> They seem to change, so it's hard to know.  Would there be harm in
> > >>> adding one to cover the end of RAM ( 0x04,7c80,0000 ) to (
> > >>> 0xff,ffff,ffff )?  Maybe that would just quiet the pointless faults
> > >>> while leaving the IOMMU enabled?
> > >>
> > >> While they may quieten the faults, I don't think those faults are
> > >> pointless. They indicate some problem with the software (less
> > >> likely the hardware, possibly the firmware) that you're using.
> > >> Also there's the question of what the overall behavior is going
> > >> to be when devices are permitted to access unpopulated address
> > >> ranges. I assume you did check already that no devices have their
> > >> BARs placed in that range?
> > >
> > > Isn't no-igfx already letting them try to read those unpopulated addresses?
> >
> > Yes, and it is for the reason that the documentation for the
> > option says "If specifying `no-igfx` fixes anything, please
> > report the problem." I imply from in in particular that one
> > better wouldn't use it for non-development purposes of whatever
> > kind.
> 
> I stopped seeing these DMA faults, but I didn't know what made them go
> away.  Then when working with an older 5.4.64 kernel, I saw them
> again.  Eric bisected down to the 5.4.y version of mainline linux
> commit:
> 
> commit 8195400f7ea95399f721ad21f4d663a62c65036f
> Author: Chris Wilson <chris at chris-wilson.co.uk>
> Date:   Mon Oct 19 11:15:23 2020 +0100
> 
>     drm/i915: Force VT'd workarounds when running as a guest OS
> 
>     If i915.ko is being used as a passthrough device, it does not know if
>     the host is using intel_iommu. Mixing the iommu and gfx causes a few
>     issues (such as scanout overfetch) which we need to workaround inside
>     the driver, so if we detect we are running under a hypervisor, also
>     assume the device access is being virtualised.

So the commit above fixes the DMA faults seen on Linux when using a
i915 gfx card?

Thanks for digging into this.

Roger.