[PATCH v7 13/24] iommu/arm-smmu-v3: Enable broadcast TLB maintenance

[Jean-Philippe Brucker]

[+Eric]

On Thu, May 21, 2020 at 03:38:35PM +0100, Marc Zyngier wrote:
> On 2020-05-21 15:17, Will Deacon wrote:
> > [+Marc]
> > 
> > On Tue, May 19, 2020 at 07:54:51PM +0200, Jean-Philippe Brucker wrote:
> > > The SMMUv3 can handle invalidation targeted at TLB entries with shared
> > > ASIDs. If the implementation supports broadcast TLB maintenance,
> > > enable it
> > > and keep track of it in a feature bit. The SMMU will then be
> > > affected by
> > > inner-shareable TLB invalidations from other agents.
> > > 
> > > A major side-effect of this change is that stage-2 translation
> > > contexts
> > > are now affected by all invalidations by VMID. VMIDs are all shared
> > > and
> > > the only ways to prevent over-invalidation, since the stage-2 page
> > > tables
> > > are not shared between CPU and SMMU, are to either disable BTM or
> > > allocate
> > > different VMIDs. This patch does not address the problem.
> > 
> > This sounds like a potential performance issue, particularly as we
> > expose
> > stage-2 contexts via VFIO directly.

Yes it's certainly going to affect SMMU performance, though I haven't
measured it. QEMU and kvmtool currently use stage-1 translations instead
of stage-2, so it won't be a problem until they start using nested
translation (and unless the SMMU only supports stage-2).

In the coming month I'd like to have a look at coordinating VMID
allocation between KVM and SMMU, for guest SVA. If the guest wants to
share page tables with the SMMU, the SMMU has to use the same VMIDs as the
VM to receive broadcast TLBI.

Similarly to patch 06 ("arm64: mm: Pin down ASIDs for sharing mm with
devices") the SMMU would request a VMID allocated by KVM, when setting up
a nesting VFIO container. One major downside is that the VMID is pinned
and cannot be recycled on rollover while it's being used for DMA.

I wonder if we could use this even when page tables aren't shared between
CPU and SMMU, to avoid splitting the VMID space.

> > Maybe we could reserve some portion
> > of
> > VMID space for the SMMU? Marc, what do you reckon?
> 
> Certainly doable when we have 16bits VMIDs. With smaller VMID spaces (like
> on
> v8.0), this is a bit more difficult (we do have pretty large v8.0 systems
> around).

It's only an issue if those systems have an SMMUv3 supporting DVM. With
any luck that doesn't exist?

> How many VMID bits are we talking about?

That's anyone's guess... One passed-through device per VM would halve the
VMID space. But the SMMU allocates one VMID for each device assigned to a
guest, not one per VM (well one per domain, or VFIO container, but I think
it boils down to one per device with QEMU). So with SR-IOV for example it
should be pretty easy to reach 256 VMIDs in the SMMU.

Thanks,
Jean