Re: [PATCH 00/12] Consolidate domain cache invalidation

[Baolu Lu]

On 3/28/24 3:59 PM, Tian, Kevin wrote:
> > From: Lu Baolu<baolu.lu@xxxxxxxxxxxxxxx>
> > Sent: Monday, March 25, 2024 10:17 AM
> >
> > The IOMMU hardware cache needs to be invalidated whenever the
> > mappings
> > in the domain are changed. Currently, domain cache invalidation is
> > scattered across different places, causing several issues:
> >
> > - IOMMU IOTLB Invalidation: This is done by iterating through the domain
> >    IDs of each domain using the following code:
> >
> >          xa_for_each(&dmar_domain->iommu_array, i, info)
> >                  iommu_flush_iotlb_psi(info->iommu, dmar_domain,
> >                                        start_pfn, nrpages,
> >                                        list_empty(&gather->freelist), 0);
> >
> >    This code could theoretically cause a use-after-free problem because
> >    there's no lock to protect the "info" pointer within the loop.
> >
> > - Inconsistent Invalidation Methods: Different domain types implement
> >    their own cache invalidation methods, making the code difficult to
> >    maintain. For example, the DMA domain, SVA domain, and nested domain
> >    have similar cache invalidation code scattered across different files.
> >
> > - SVA Domain Inconsistency: The SVA domain implementation uses a
> >    completely different data structure to track attached devices compared
> >    to other domains. This creates unnecessary differences and, even
> >    worse, leads to duplicate IOTLB invalidation when an SVA domain is
> >    attached to devices belonging to different IOMMU domains.
> can you elaborate how duplicated invalidations are caused?

Yes, sure.

Current Intel SVA implementation keeps the bond between mm and a PASID
of a device in a list of intel_svm_dev. In the mm notifier callback, it
iterates all intel_svam_dev in the list and invalidates the IOTLB and
device TLB sequentially.

If multiple devices belong to a single IOMMU, the IOTLB will be flushed
multiple times. However, since these devices share the same domain ID
and PASID, a single IOTLB cache invalidation is sufficient. The
additional flushes are redundant and negatively impact performance.

Best regards,
baolu