On Tue, Sep 06, 2016 at 03:18:46PM +1000, David Gibson wrote:
> > Maybe here naming the flags as IOMMU_{RW_NONE} is a little bit
> > confusing (just to leverage existing access flags), but what I was
> > trying to do is to make the two things not overlapped at all, since I
> > didn't find a mixture use case.
>
> Maybe not now, but a common use case is absolutely possible. If you
> had a single (guest) bus with a single IO address space, with vIOMMU
> and both VFIO and vhost devices on it, the same vIOMMU would need to
> send all notifications to VFIO and (at least) unmap notifications to
> vhost.
Yeah this is a good one... Thanks to point out.
Then I agree that splitting the use cases won't be enough... We may
need to exactly register IOMMU notifiers with notification type
(invalidations only, or we also need entry additions), and we just
selectively notify the consumers depending on what kind of
notification it is.
Or any smarter way to do it.
-- peterx
