Re: [PATCH v6 03/12] iommu: Add attach/detach_dev_pasid domain ops

[Jean-Philippe Brucker]

On Wed, May 11, 2022 at 09:02:40AM -0300, Jason Gunthorpe wrote:
> On Wed, May 11, 2022 at 08:54:39AM +0100, Jean-Philippe Brucker wrote:
> > > > > Then 'detach pasid' is:
> > > > >
> > > > > iommu_ops->blocking_domain->ops->attach_dev_pasid(domain, dev,
> > > > pasid);
> > > > >
> > > > > And we move away from the notion of 'detach' and in the direction that
> > > > > everything continuously has a domain set. PASID would logically
> > > > > default to blocking_domain, though we wouldn't track this anywhere.
> > > > 
> > > > I am not sure whether we still need to keep the blocking domain concept
> > > > when we are entering the new PASID world. Please allow me to wait and
> > > > listen to more opinions.
> > > > 
> > > 
> > > I'm with Jason on this direction. In concept after a PASID is detached it's
> > > essentially blocked. Implementation-wise it doesn't prevent the iommu
> > > driver from marking the PASID entry as non-present as doing in this
> > > series instead of actually pointing to the empty page table of the block
> > > domain. But api-wise it does make the entire semantics more consistent.
> > 
> > This is all internal to IOMMU so I don't think we should be concerned
> > about API consistency. I prefer a straighforward detach() operation
> > because that way IOMMU drivers don't have to keep track of which domain is
> > attached to which PASID. That code can be factored into the IOMMU core.
> 
> Why would a driver need to keep additional tracking?
> 
> > In addition to clearing contexts, detach() also needs to invalidate TLBs,
> > and for that the SMMU driver needs to know the old ASID (!= PASID) that
> > was used by the context descriptor. We can certainly work around a missing
> > detach() to implement this, but it will be convoluted.
> 
> It is not "missing" it is just renamed to blocking_domain->ops->set_dev_pasid()
> 
> The implementation of that function would be identical to
> detach_dev_pasid.

  attach(dev, pasid, sva_domain)
  detach(dev, pasid, sva_domain)

versus

  set_dev_pasid(dev, pasid, sva_domain)
  set_dev_pasid(dev, pasid, blocking)

we loose the information of the domain previously attached, and the SMMU
driver has to retrieve it to find the ASID corresponding to the mm. 

Thanks,
Jean