Re: [PATCH v10 5/7] iommu/vt-d: Make prq draining code generic

[Baolu Lu]

On 2023/7/19 13:39, Tian, Kevin wrote:
> > From: Baolu Lu <baolu.lu@xxxxxxxxxxxxxxx>
> > Sent: Friday, July 14, 2023 1:47 PM
> >
> > On 2023/7/14 11:49, Tian, Kevin wrote:
> > > > From: Baolu Lu <baolu.lu@xxxxxxxxxxxxxxx>
> > > > Sent: Friday, July 14, 2023 11:28 AM
> > > >
> > > > On 2023/7/13 15:49, Tian, Kevin wrote:
> > > > > > From: Jacob Pan <jacob.jun.pan@xxxxxxxxxxxxxxx>
> > > > > > Sent: Thursday, July 13, 2023 12:34 AM
> > > > > >
> > > > > > -	/* Domain type specific cleanup: */
> > > > > >     	domain = iommu_get_domain_for_dev_pasid(dev, pasid, 0);
> > > > > > -	if (domain) {
> > > > > > -		switch (domain->type) {
> > > > > > -		case IOMMU_DOMAIN_SVA:
> > > > > > -			intel_svm_remove_dev_pasid(dev, pasid);
> > > > > > -			break;
> > > > > > -		default:
> > > > > > -			/* should never reach here */
> > > > > > -			WARN_ON(1);
> > > > > > -			break;
> > > > > > -		}
> > > > > > +	if (!domain)
> > > > > > +		goto out_tear_down;
> > > > >
> > > > > WARN_ON()
> > > >
> > > > Why?
> > > >
> > > > My understanding is that remve_device_pasid could be call in any context
> > > > including no domain attached.
> > >
> > > oh I'm not aware of that. Can you elaborate the usage which uses a pasid
> > > w/o domain? pasid needs to point to a page table. Presumably every
> > > page table should be wrapped by a iommu domain...
> >
> > A case I can think of is error rewinding. A domain is being attached to
> > multiple pasids. When one of them is failed, remove_device_pasid should
> > be called on all pasids so that they are parked at a determinant state.
>
> Can you elaborate what is the association among those pasid's so failing
> one would lead to failing all?
>
> Just like a domain can be attached to multiple devices. I don't think there
> is an unwinding policy forcing to detach all devices just because there is
> a failure attaching the domain to a new one.

It's fine to add a check here if there's no real case. I was just
thinking that remove_device_pasid could also be invoked when there's no
domain attached.

> > On the other hand, I don't want the remove_device_pasid to be the
> > counterpart of attach_dev_pasid. remove_device_pasid simply denotes:
> >
> > - The pasid will be parked in blocking state;
> > - If any domain that has been attached to this pasid, stop reference to
> >     it any more. Otherwise, there might be use-after-free issues.
> >
> > Hence, remove_device_pasid should never fail.
>
> It should never fail. But could warn if there is a condition which shouldn't
> be hit. 😊

Okay, let's add a check here. And we can loose it later if any real use
case comes.

Best regards,
baolu