Re: [Qemu-devel] userspace emulated smmu/vfio integration: how to trap updates to the table structures?

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, May 02, 2017 at 04:42:56PM +0100, Robin Murphy wrote:
> On 02/05/17 16:22, Michael S. Tsirkin wrote:
> > On Tue, May 02, 2017 at 10:17:26AM +0200, Christoffer Dall wrote:
> >> On Tue, May 2, 2017 at 10:13 AM, Auger Eric <eric.auger@redhat.com> wrote:
> >>> Hi Christoffer,
> >>>
> >>> On 02/05/2017 09:53, Christoffer Dall wrote:
> >>>> On Tue, May 2, 2017 at 9:30 AM, Auger Eric <eric.auger@redhat.com> wrote:
> >>>>> Hi Will, Robin, Jean-Philippe,
> >>>>>
> >>>>> I have been working on the integration between user-space emulated
> >>>>> SMMU-v3 and VFIO in QEMU. At the moment I fail identifying a proper easy
> >>>>> way to trap page table updates. This is requested to keep the host
> >>>>> translation structures consistent to guest translation structures.
> >>>>>
> >>>>> On Intel VTD there is a so-called "caching mode" (CM, see VTD spec
> >>>>> paragraph 6.1) that forces the OS to explicitly invalidate caches
> >>>>> whenever it updates any remapping structure (updates to not-present or
> >>>>> present entries). Those invalidation commands are used to trap and
> >>>>> update host structures. This mode was devised for virtualization. I was
> >>>>> not able to find such "caching mode" on ARM SMMU. Is there any?
> >>>>>
> >>>>> If not, do you have any other suggestion, I mean, besides the
> >>>>> virtio-based solution.
> >>>>>
> >>>>>
> >>>> Worst case, can you make the guest page tables read-only and catch the
> >>>> faults and propagate changes to SMMU translations?
> >>>
> >>> The issue I foresee is there are up to 4 level of page tables to trap.
> >>> This would lead to plenty of regions to "translate" on qemu side. Also,
> >>> besides the 1st level pointed by TTBR found in stage 1 context
> >>> descriptor, other page regions would be discovered dynamically as
> >>> mapping are built. To me this is the last resort solution if confirmed
> >>> feasible.
> >>>
> >> Completely agree, it would be a terrible solution.
> >>
> >> Thanks,
> >> -Christoffer
> > 
> > All "caching mode" is is simple a bit that the IOMMU device uses to tell
> > the OS "I might cache non-present/invalid entries, please use
> > invalidation even if you make non-present entries present".
> > 
> > So all we need from ARM is bit somewhere in one of MMU registers
> > and a promise not to use it for any other purpose preferably
> > by documenting it as caching mode bit in the next manual version.
> > 
> > Any chance of this happening?
> 
> It's unlikely as such, since the VMSAv8 translation formats which the
> SMMU architecture specifies explicitly forbid such "negative caching",
> so having an architected bit to say "I don't comply with the
> architecture" comes out looking a little nonsensical.

Same's true for AMD/intel IOMMUs FWIW.

> However, doing the same via a firmware quirk would seem a lot more
> reasonable. The io-pgtable code already has the IO_PGTABLE_TLBI_ON_MAP
> quirk to cope with other not-quite-VMSA IOMMU implementations which do
> cache invalid entries - supporting that in the io-pgtable-arm backend
> (currently only the v7s backend implements it), then having the SMMUv3
> driver detect the QEMU implementation and set that quirk on pagetables
> accordingly would end up achieving the same thing.
> 
> Robin.

Without a spec specifying it, it does however mean that non-Linux
guests aren't likely to implement this in the same way.

-- 
MST