Re: [PATCH v2 1/3] iommu: Replace detach_dev with real blocking domains

[Baolu Lu]

On 2022/8/26 23:34, Jason Gunthorpe wrote:
> On Fri, Aug 26, 2022 at 08:30:12PM +0800, Lu Baolu wrote:
> >  From iommu core's point of view, detaching a domain from a device is
> > equal to attaching the group's blocking domain to the device. This
> > repalces all detach_dev callbacks in the IOMMU drivers with a real
> > blocking domain and handles detaching domain through it.
>
> There is a subtly here, where detach_dev() is not *always* going to a
> blocking domain. Recall we made this mistaken when building the owner
> series and it is why the blocking domain was introduced.
>
> The challenge with this work is to be sure that the code you assigned
> to the blocking domain doesn't actually set an identity domain or
> set things to use the platform DMA ops.

This looks like a defect in the current IOMMU subsystem.

> > +static int blocking_domain_attach_dev(struct iommu_domain *_domain,
> > +				      struct device *dev)
> > +{
> > +	struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
> > +
> > +	s390_iommu_detach_device(domain, dev);
> > +
> > +	return 0;
> > +}
>
> For instance, I know that this is returning the group back to the
> platform DMA ops.
>
> To really make progress on this I would suggest
>
> 1) Remove detach_dev from drivers that don't use it. This is any
>     driver that provides a default domain, if a default domain is set
>     then we always use attach_dev(default_domain) and never detach_dev
>
> 2) Of those drivers in #1 audit and try to determine if their
>     detach_dev is doing blocking domain behavior and if yes give them a
>     real blocking domain. Otherwise just delete the code.
>
> 3) The remaining drivers are not using default_domain. Rename
>     'detach_dev' to iommu_ops->set_platform_dma(dev) to make it clear
>     that what it is doing is returning control back to the platform DMA
>     ops.
>
> I would do one patch per driver on the blocking domain conversions and
> get acks from the driver owners. Do as many as can be acked, delete
> the rest.
>
> Make it clear to driver owners checking that blocking domain must halt
> all DMA, and it must not be a passthrough or identity behavior.
>
> When last I looked several of the drivers looked like they were
> setting an identity domain on the detach_dev - eg back to a boot time
> behavior where the iommu is bypassed. It is really important these not
> be converted to blocking :) Arguably if they can be found they should
> be converted to identity domains instead.

Above makes sense to me. Thanks! I will spend some time looking into
each IOMMU driver and back here after I work through all of them.

Best regards,
baolu