Re: [PATCH v4 09/12] iommu/vt-d: Add iotlb flush for nested domain

[Baolu Lu]

On 2023/8/3 12:13, Tian, Kevin wrote:
> > From: Baolu Lu <baolu.lu@xxxxxxxxxxxxxxx>
> > Sent: Thursday, August 3, 2023 12:06 PM
> >
> > On 2023/8/3 12:00, Tian, Kevin wrote:
> > > > From: Baolu Lu <baolu.lu@xxxxxxxxxxxxxxx>
> > > > Sent: Thursday, August 3, 2023 11:25 AM
> > > >
> > > > On 2023/8/2 15:46, Tian, Kevin wrote:
> > > > > > From: Liu, Yi L <yi.l.liu@xxxxxxxxx>
> > > > > > Sent: Monday, July 24, 2023 7:14 PM
> > > > > >
> > > > > > +
> > > > > > +		spin_lock_irqsave(&dmar_domain->lock, flags);
> > > > > > +		list_for_each_entry(info, &dmar_domain->devices, link)
> > > > > > +			intel_nested_invalidate(info->dev, dmar_domain,
> > > > > > +						req->addr, req->npages);
> > > > > > +		spin_unlock_irqrestore(&dmar_domain->lock, flags);
> > > > >
> > > > > Disabling interrupt while invalidating iotlb is certainly unacceptable.
> > > > >
> > > > > Actually there is no need to walk devices. Under dmar_domain there
> > > > > is already a list of attached iommu's.
> > > >
> > > > Walking device is only necessary when invalidating device TLB. For iotlb
> > > > invalidation, it only needs to know the iommu's.
> > >
> > > even for device tlb we may think whether there is any better way
> > > to avoid disabling interrupt. It's a slow path, especially in a guest.
> >
> > I ever tried this. But some device drivers call iommu_unmap() in the
> > interrupt critical path. :-( So we have a long way to go.
>
> emmm... this path only comes from iommufd and the domain is
> user-managed. There won't be kernel drivers to call iommu_unmap()
> on such domain.

Probably we can use a different lock for nested domain and add a comment
around the lock with above explanation.

Best regards,
baolu