Re: [PATCH v3 6/8] iommu/arm-smmu: Add impl hook for inherit boot mappings

[Bjorn Andersson]

On Mon 21 Sep 23:08 CEST 2020, Will Deacon wrote:

> On Sat, Sep 12, 2020 at 10:25:59PM -0500, Bjorn Andersson wrote:
> > On Fri 11 Sep 12:13 CDT 2020, Robin Murphy wrote:
> > > On 2020-09-04 16:55, Bjorn Andersson wrote:
> > > > Add a new operation to allow platform implementations to inherit any
> > > > stream mappings from the boot loader.
> > > 
> > > Is there a reason we need an explicit step for this? The aim of the
> > > cfg_probe hook is that the SMMU software state should all be set up by then,
> > > and you can mess about with it however you like before arm_smmu_reset()
> > > actually commits anything to hardware. I would have thought you could
> > > permanently steal a context bank, configure it as your bypass hole, read out
> > > the previous SME configuration and tweak smmu->smrs and smmu->s2crs
> > > appropriately all together "invisibly" at that point.
> > 
> > I did this because as of 6a79a5a3842b ("iommu/arm-smmu: Call
> > configuration impl hook before consuming features") we no longer have
> > setup pgsize_bitmap as we hit cfg_probe, which means that I need to
> > replicate this logic to set up the iommu_domain.
> > 
> > If I avoid setting up an iommu_domain for the identity context, as you
> > request in patch 8, this shouldn't be needed anymore.
> > 
> > > If that can't work, I'm very curious as to what I've overlooked.
> > > 
> > 
> > I believe that will work, I will rework the patches and try it out.
> 
> Did you get a chance to rework this?
> 

Finally got a chance to dig through this properly.

Initial results where positive and with an implementation of cfg_probe
in qcom_smmu_impl I'm able to probe the arm-smmu driver just fine - and
display (e.g. efifb) stays alive.

Unfortunately as the display driver (drivers/gpu/drm/msm) is about to
probe a new iommu domain is created, which due to its match against
qcom_smmu_client_of_match[] becomes of type IOMMU_DOMAIN_IDENTITY.
This results in a S2CR of BYPASS type, which the firmware intercepts and
turns the stream into a type FAULT.

So while the cfg_probe looks very reasonable we're still in need of a
mechanism to use the fake identity context for the iommu domain
associated with the display controller.


The workings of the display driver is that it gets the iommu domain
setup for byass and then after that creates a translation context for
this same stream where it maps the framebuffer.

For testing purposes I made def_domain_type always return 0 in the qcom
impl and the result is that we get a few page faults while probing the
display driver, but these are handled somewhat gracefully and the
initialization did proceed and the system comes up nicely (but in the
case that the display driver would probe defer this leads to an storm of
faults as the screen continues to be refreshed).

TL;DR I think we still need to have a way to get the arm-smmu driver to
allow the qcom implementation to configure identity domains to use
translation - but we can make the setup of the identity context a detail
of the qcom driver.

Regards,
Bjorn