Re: [PATCH RFC v3 08/12] iommu/sva: Use attach/detach_pasid_dev in SVA interfaces

[Lu Baolu]

On 2022/4/13 7:36, Tian, Kevin wrote:
> > From: Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx>
> > Sent: Tuesday, April 12, 2022 8:53 PM
> >
> > > > +	if (!handle) {
> > > > +		ret = -ENOMEM;
> > > > +		goto out_put_ioas;
> > > > +	}
> > > > +
> > > > +	/* The reference to ioas will be kept until domain free. */
> > > > +	domain = iommu_sva_alloc_domain(dev, ioas);
> > >
> > > Shouldn't we first try whether existing domains are compatible to this
> > > device?
> >
> > If we think that here domain represents a hardware pagetable actually
> > used by IOMMU for a {device, pasid}, we are able to use per-{device,
> > pasid} domain without checking compatibility. Sharing a domain among
> > devices under the same IOMMU may be an optimization. That could be done
> > in the IOMMU driver just like what vt-d driver is doing for pass-through
> > DMA domains.
>
> there is only one hardware page table per mm in this case. Multiple domains
> are required only due to compatibility reason as Jason/Robin pointed out
> in SMMU case. Given all other places create multiple domains per ioas only
> upon incompatibility, probably it's more consistent to doing so in this path
> too...

Sharing domain for compatible devices is valuable when the domain
supports map/unmap operations. That can reduce the number of map/unmap
calls and the resulting synchronization of IOTLB. But for SVA case, it's
a dumb domain which only provides attach/detach operations.

A similar case could be found on pass-through DMA domains. The iommu
core allocates a default domain for each group although all the domains
represent a same page table for the compatible devices. The VT-d driver
optimizes this by exporting a static identity domain.

Anyway, I am open for this. I can add a compatible domain list if most
of you like that way. :-)

Best regards,
baolu