On Thu, Jul 16, 2020 at 09:50:53AM +0100, Will Deacon wrote:
> On Mon, Jul 13, 2020 at 11:19:17AM -0600, Jordan Crouse wrote:
> > On Mon, Jul 13, 2020 at 04:09:02PM +0100, Will Deacon wrote:
> > > On Fri, Jun 26, 2020 at 02:00:38PM -0600, Jordan Crouse wrote:
> > > > diff --git a/drivers/iommu/arm-smmu.h b/drivers/iommu/arm-smmu.h
> > > > index 5f2de20e883b..d33cfe26b2f5 100644
> > > > --- a/drivers/iommu/arm-smmu.h
> > > > +++ b/drivers/iommu/arm-smmu.h
> > > > @@ -345,6 +345,7 @@ struct arm_smmu_domain {
> > > > struct mutex init_mutex; /* Protects smmu pointer */
> > > > spinlock_t cb_lock; /* Serialises ATS1* ops and TLB syncs */
> > > > struct iommu_domain domain;
> > > > + struct device *dev; /* Device attached to this domain */
> > >
> > > This really doesn't feel right to me -- you can generally have multiple
> > > devices attached to a domain and they can come and go without the domain
> > > being destroyed. Perhaps you could instead identify the GPU during
> > > cfg_probe() and squirrel that information away somewhere?
> >
> > I need some help here. The SMMU device (qcom,adreno-smmu) will have at least two
> > stream ids from two different platform devices (GPU and GMU) and I need to
> > configure split-pagetable and stall/terminate differently on the two domains.
>
> Hmm. How does the GPU driver know which context bank is assigned to the GPU
> and which one is assigned to the GMU? I assume it needs this information so
> that it can play its nasty tricks with the TTBR registers?
>
> I ask because if we need to guarantee stability of the context-bank
> assignment, then you could match on that in the ->init_context() callback,
> but now I worry that it currently works by luck :/
Its more like "educated" luck. If we assign the domains in the right order we
know that the GPU will be on context bank 0 but you are entirely right that if
everything doesn't go exactly the way we need things go badly.
Some targets do have the ability to reprogram which context bank is used but
that would require a domain attribute from the SMMU driver to communicate that
information back to the GPU driver.
> Do we need to add an extra callback to allocate the context bank?
That seems like a reasonable option. That seems like it would work with legacy
targets and rely less on luck.
Jordan
--
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
