Re: [RFC PATCH v2 1/2] vfio/type1: Adopt fast IOTLB flush interface when unmap IOVAs

[Suravee Suthikulpanit]

Hi Alex,

On 1/9/18 4:07 AM, Alex Williamson wrote:
> > > @@ -661,6 +705,8 @@ static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
> > >   	if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
> > >   		return 0;
> > >   
> > > +	INIT_LIST_HEAD(&unmapped_regions);
> > > +
> > >   	/*
> > >   	 * We use the IOMMU to track the physical addresses, otherwise we'd
> > >   	 * need a much more complicated tracking system.  Unfortunately that
> > > @@ -698,24 +744,36 @@ static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
> > >   				break;
> > >   		}
> > >   
> > > -		unmapped = iommu_unmap(domain->domain, iova, len);
> > > -		if (WARN_ON(!unmapped))
> > > +		entry = kzalloc(sizeof(*entry), GFP_KERNEL);
> > > +		if (!entry)
> > >   			break;
>
> Turns out this nagged at me a bit too, this function only gets called
> once to dump the vfio_dma, so bailing out here leaves pages pinned and
> IOMMU mappings in place, for a performance optimization that we could
> just skip.  We could sync&unpin anything collected up to this point and
> continue this step with a synchronous unmap/unpin.  Thanks,

Ah, that's an over look in my part also. Thanks for catching this. I'll implement
the fallback mechanism per your suggestion in v3.

Thanks,
Suravee