[PATCH v4 22/28] iommu/arm-smmu-v3-kvm: Emulate CMDQ for host

[Mostafa Saleh]

On Mon, Sep 15, 2025 at 01:38:58PM -0300, Jason Gunthorpe wrote:
> On Fri, Sep 12, 2025 at 03:18:08PM +0100, Will Deacon wrote:
> > Ideally, the data structures that are shadowed by the hypervisor would
> > be mapped as normal-WB cacheable in both the host and the hypervisor so
> > we don't have to worry about coherency and we get the performance
> > benefits from the caches. Indeed, I think that's how you've mapped
> > 'host_cmdq' above _however_ I sadly don't think we can do that if the
> > actual SMMU hardware isn't coherent.
> 
> That seems like the right conclusion to me, pkvm should not be mapping
> as cachable unless it knows the IORT/IDR is marked as coherent.
> 
> This is actually something I want to fix in the SMMU driver, it should
> always be allocating cachable memory and using
> dma_sync_single_for_device() instead of non-cachable DMA coherent
> allocations. (Or perhaps better is to use
> iommu_pages_flush_incoherent())
> 
> I'm hearing about an interesting use case where we'd want to tell the
> SMMU to walk STEs non-cachable even if the HW is capable to do
> cachable. Apparently in some SOCs it gives better isochronous
> properties for realtime DMA.

Interesting, I guess that would be more noticable for the page table
walks rather than the STE, as Linux doesn't invalidate STEs that much.
>
> 
> IMHO for this series at this point pkvm should just require a coherent
> SMMU until the above revisions happen.

I think the fix for the problem Will mentioned is to just use CMOs
before accessing the host structures, so that should be simple.
If it turns to be more complicated, I am happy to drop the support
for non-coherent devices from this series and we can add it later.

Thanks,
Mostafa
> 
> Jason