Re: [PATCH 2/5] iommu/arm-smmu: Emulate bypass by using context banks

[Bjorn Andersson]

On Thu 09 Jul 11:55 PDT 2020, Rob Clark wrote:

> On Thu, Jul 9, 2020 at 9:56 AM Rob Clark <robdclark@xxxxxxxxx> wrote:
> >
> > On Thu, Jul 9, 2020 at 9:48 AM Bjorn Andersson
> > <bjorn.andersson@xxxxxxxxxx> wrote:
> > >
> > > On Thu 09 Jul 09:17 PDT 2020, Rob Clark wrote:
> > >
> > > > On Wed, Jul 8, 2020 at 10:01 PM Bjorn Andersson
> > > > <bjorn.andersson@xxxxxxxxxx> wrote:
> > > [..]
> > > > > @@ -678,7 +680,11 @@ static int arm_smmu_init_domain_context(struct iommu_domain *domain,
> > > > >         if (smmu_domain->smmu)
> > > > >                 goto out_unlock;
> > > > >
> > > > > -       if (domain->type == IOMMU_DOMAIN_IDENTITY) {
> > > > > +       /*
> > > > > +        * Nothing to do for IDENTITY domains,unless disabled context banks are
> > > > > +        * used to emulate bypass mappings on Qualcomm platforms.
> > > > > +        */
> > > > > +       if (domain->type == IOMMU_DOMAIN_IDENTITY && !smmu->qcom_bypass_quirk) {
> > > >
> > > > maybe I'm overlooking something, but I think this would put us back to
> > > > allocating pgtables (and making iommu->map/unmap() no longer no-ops),
> > > > which I don't think we want
> > > >
> > >
> > > You're right, we are allocating page tables for these contexts and
> > > map/unmap would modify the page tables. But afaict traversal is never
> > > performed, given that the banks are never enabled.
> > >
> > > But as drivers probe properly, or the direct mapped drivers sets up
> > > their iommu domains explicitly with translation this would not be used.
> > >
> > > So afaict we're just wasting some memory - for the gain of not
> > > overcomplicating this function.
> >
> > the problem is that it makes dma_map/unmap less of a no-op than it
> > should be (for the case where the driver is explicitly managing it's
> > own domain)..  I was hoping to get rid of the hacks to use dma_sync go
> > back to dma_map/unmap for cache cleaning
> 
> That said, it seems to cause less explosions than before.. either that
> or I'm not trying hard enough.  Still, I think it would probably
> result in unnecessary tlb inv's.
> 
> Previously, *somehow* we seemed to end up with dma_map/unmap trying to
> modify the domain that we had attached.
> 

I might need some help to really understand the plumbing here, but this
is what I read from the code and can see in my debugging...


The display device will upon creation be allocated a default_domain, of
type IOMMU_DOMAIN_IDENTITY - per the qcom-impl. Then you then allocate a
new iommu domain on the platform bus in the display driver and attach
this to the device. This will result in the group's domain being of type
IOMMU_DOMAIN_UNMANAGED.

But when you then invoke dma_map_single() on the display device, the map
operation will acquire the iommu_group's default_domain (not domain) and
as such attempt to map stuff on the IDENTITY domain.

__iommu_map() will however reject this, given that the type does not
have __IOMMU_DOMAIN_PAGING set. Afaict, this would happen regardless of
this patch actually setting up a page table for the default_domain or
not.


So, afaict, you can't use dma_map_single()/dma_unmap() to operate your
page table on a lately attached iommu domain.

This would also imply that the display device consumes two context
banks, which worries me more than the waste of page tables etc.

Regards,
Bjorn

> BR,
> -R
> 
> > BR,
> > -R
> >
> >
> > >
> > > Regards,
> > > Bjorn
> > >
> > > > BR,
> > > > -R
> > > >
> > > > >                 smmu_domain->stage = ARM_SMMU_DOMAIN_BYPASS;
> > > > >                 smmu_domain->smmu = smmu;
> > > > >                 goto out_unlock;
> > > > > @@ -826,6 +832,10 @@ static int arm_smmu_init_domain_context(struct iommu_domain *domain,
> > > > >         domain->geometry.aperture_end = (1UL << ias) - 1;
> > > > >         domain->geometry.force_aperture = true;
> > > > >
> > > > > +       /* Enable translation for non-identity context banks */
> > > > > +       if (domain->type != IOMMU_DOMAIN_IDENTITY)
> > > > > +               cfg->m = true;
> > > > > +
> > > > >         /* Initialise the context bank with our page table cfg */
> > > > >         arm_smmu_init_context_bank(smmu_domain, &pgtbl_cfg);
> > > > >         arm_smmu_write_context_bank(smmu, cfg->cbndx);
> > > > > diff --git a/drivers/iommu/arm-smmu.h b/drivers/iommu/arm-smmu.h
> > > > > index d172c024be61..a71d193073e4 100644
> > > > > --- a/drivers/iommu/arm-smmu.h
> > > > > +++ b/drivers/iommu/arm-smmu.h
> > > > > @@ -305,6 +305,8 @@ struct arm_smmu_device {
> > > > >
> > > > >         /* IOMMU core code handle */
> > > > >         struct iommu_device             iommu;
> > > > > +
> > > > > +       bool                            qcom_bypass_quirk;
> > > > >  };
> > > > >
> > > > >  enum arm_smmu_context_fmt {
> > > > > @@ -323,6 +325,7 @@ struct arm_smmu_cfg {
> > > > >         };
> > > > >         enum arm_smmu_cbar_type         cbar;
> > > > >         enum arm_smmu_context_fmt       fmt;
> > > > > +       bool                            m;
> > > > >  };
> > > > >  #define ARM_SMMU_INVALID_IRPTNDX       0xff
> > > > >
> > > > > --
> > > > > 2.26.2
> > > > >
> > > > > _______________________________________________
> > > > > iommu mailing list
> > > > > iommu@xxxxxxxxxxxxxxxxxxxxxxxxxx
> > > > > https://lists.linuxfoundation.org/mailman/listinfo/iommu