Re: [PATCH] intel-iommu: ignore SNP bit in scalable mode

[Peter Xu]

On Mon, Nov 29, 2021 at 10:28:42AM +0800, Jason Wang wrote:
> > I think we can still have Jason's patch continued because the kernel commit 
> > to
> > apply SNP bit is merged in v5.13, so we may need the qemu change to let it
> > still work with v5.13-v5.15+ guest kernels.  We'll loose the resv bit check 
> > a
> > bit, but looks worthwhile.  Jason?
> 
> Yes, I agree. The only thing that may worry me is the migration
> compatibility. If we migrate from new to old we may break the guests,
> we probably need compatibility props for that.

Hmm.. How important is new->old migrations?  Is that normally for the
old->new->old migration so that users can always fallback to the old hosts?

If that's the case then IMHO we're fine here, since the new binary check less
on resv bits than the old binary, then if the guest code can work with the old
binary already then migrating back to old binary should work too.  Changing
guest OS during migration of new->old can have a problem but hopefully rare.

OTOH - do you know any of the real enterprise user that uses scalable mode yet?
To my own understanding it's still mostly "experimental", then hopefully we can
avoid worrying on that too much?

> 
> And in the future, it could be even more troublesome,e.g there's one
> day we found another bit that needs not to be checked. Maybe we should
> even remove all the rsvd bits checks?

When a real hardware sees any of the reserved bits set, it'll bail out and
raise an error, right?

If that's the case, I'm wondering whether we should always follow the hardware
behavior as an emulator.

Now I'm trying to remember normally how a spec could re-use a bit that was used
to be reserved: should the hardware bumps the version of the version reg in so
that softwares will know what to expect?

So I'm thinking whether the emulator code can identify the version bump by
"scalable mode enabled", if so we know some resved bits are "ignored" now, and
IIUC that's mostly the original proposal to add a quirk when scalable mode in
vtd_init().

But again, I really think it should be the spec owner who should have
considered all these.. e.g. explicitly document "this bit was used to reserved,
but when scalable mode enabled it's ignored and programmable by the guest
driver", or something like that.

Thanks,

-- 
Peter Xu