Re: [PATCH 0/5] Share sva domains with all devices bound to a mm

[Jason Gunthorpe]

On Thu, Aug 10, 2023 at 07:49:11AM +0000, Tian, Kevin wrote:
> > From: Zhang, Tina <tina.zhang@xxxxxxxxx>
> > Sent: Thursday, August 10, 2023 9:32 AM
> > 
> > Hi,
> > 
> > On 8/9/23 17:41, Tian, Kevin wrote:
> > >> From: Zhang, Tina <tina.zhang@xxxxxxxxx>
> > >> Sent: Tuesday, August 8, 2023 3:50 PM
> > >>
> > >> A sva domain's lifetime begins with binding a device to a mm and ends
> > >> by releasing all the bound devices from that sva domain. Technically,
> > >> there could be more than one sva domain identified by the mm PASID for
> > >> the use of bound devices issuing DMA transactions.
> > >
> > > Could you elaborate it with some concrete examples which motivate
> > > this change?
> > The motivation is to remove the superfluous IOTLB invalidation in
> > current VT-d driver.
> > 
> > Currently, in VT-d driver, due to lacking shared sva domain info, in
> > intel_flush_svm_range(), both iotlb and dev-tlb invalidation operations
> > are performed per-device. However, difference devices could be behind
> > one IOMMU (e.g., four devices are behind one IOMMU) and invoking iotlb
> > per-device gives us more iotlb invalidation than necessary (4 iotlb
> > invalidation instead of 1). This issue may give more performance impact
> > when in a virtual machine guest, as currently we have one virtual VT-d
> > for in front of those virtual devices.
> > 
> > 
> > This patch fixes this issue by attaching shared sva domain information
> > to mm, so that it can be utilized in the mm_notifier_ops callbacks.
> > 
> 
> that is one of the motivations. e.g. another one as Jason suggested
> is to cleanup to decouple the common sva logic from enqcmd. Both
> should be mentioned in next version cover letter.

I also want to purge all the de-duplication and refcounting code
around mm's and sva_binds from the drivers. Eg see the mess this makes
of SMMUv3.

Core code provides a single iommu_domain per-mm for SVA. Driver can
rely on this optimization and does not need to de-duplicate.

Single domain tracks all attachments. Driver can optimize using that
information by de-duplicating (eg ASID invalidation vs ATC
invalidation)

After this we need to fix the domain allocation op to add a
'alloc_domain_sva(dev, mm_struct)' op so that the drivers can setup
their SVA domains fully in a nice lock-safe environment.

Jason