Re: [PATCH v3 1/8] iommu: Add ops entry for supported default domain type

[Robin Murphy]

On 09/05/2019 03:30, Lu Baolu wrote:
> Hi Robin,
>
> On 5/7/19 6:28 PM, Robin Murphy wrote:
> > On 06/05/2019 16:32, Tom Murphy via iommu wrote:
> > > The AMD driver already solves this problem and uses the generic
> > > iommu_request_dm_for_dev function. It seems like both drivers have the
> > > same problem and could use the same solution. Is there any reason we
> > > can't have use the same solution for the intel and amd driver?
> > >
> > > Could we just copy the implementation of the AMD driver? It would be
> > > nice to have the same behavior across both drivers especially as we
> > > move to make both drivers use more generic code.
> >
> > TBH I don't think the API really needs to be involved at all here. 
> > Drivers can already not provide the requested default domain type if 
> > they don't support it, so as long as the driver can ensure that the 
> > device ends up with IOMMU or direct DMA ops as appropriate, I don't 
> > see any great problem with drivers just returning a passthrough domain 
> > when a DMA domain was requested, or vice versa (and logging a message 
> > that the requested type was overridden). The only type that we really 
> > do have to honour strictly is non-default (i.e. unmanaged) domains.
>
> I agree with you that we only have to honor strictly the non-default
> domains. But domain type saved in iommu_domain is consumed in iommu.c
> and exposed to user through sysfs. It's not clean if the iommu driver
> silently replace the default domain.

Right, I did get a bit ahead of myself there - the implicit step before 
that is to fix default domain allocation so that the core actually 
passes the relevant device which it has to hand, such that the IOMMU 
drivers *can* make the right decision up-front.

Robin.