> From: Jason Gunthorpe <jgg@xxxxxxxxxx>
> Sent: Thursday, March 24, 2022 4:34 AM
>
> On Wed, Mar 23, 2022 at 02:04:46PM -0600, Alex Williamson wrote:
> > On Wed, 23 Mar 2022 16:34:39 -0300
> > Jason Gunthorpe <jgg@xxxxxxxxxx> wrote:
> >
> > > On Wed, Mar 23, 2022 at 01:10:38PM -0600, Alex Williamson wrote:
> > > > On Fri, 18 Mar 2022 14:27:33 -0300
> > > > Jason Gunthorpe <jgg@xxxxxxxxxx> wrote:
> > > >
> > > > > +static int conv_iommu_prot(u32 map_flags)
> > > > > +{
> > > > > + int iommu_prot;
> > > > > +
> > > > > + /*
> > > > > + * We provide no manual cache coherency ioctls to userspace
> and most
> > > > > + * architectures make the CPU ops for cache flushing
> privileged.
> > > > > + * Therefore we require the underlying IOMMU to support
> CPU coherent
> > > > > + * operation.
> > > > > + */
> > > > > + iommu_prot = IOMMU_CACHE;
> > > >
> > > > Where is this requirement enforced? AIUI we'd need to test
> > > > IOMMU_CAP_CACHE_COHERENCY somewhere since functions like
> > > > intel_iommu_map() simply drop the flag when not supported by HW.
> > >
> > > You are right, the correct thing to do is to fail device
> > > binding/attach entirely if IOMMU_CAP_CACHE_COHERENCY is not there,
> > > however we can't do that because Intel abuses the meaning of
> > > IOMMU_CAP_CACHE_COHERENCY to mean their special no-snoop
> behavior is
> > > supported.
> > >
> > > I want Intel to split out their special no-snoop from IOMMU_CACHE and
> > > IOMMU_CAP_CACHE_COHERENCY so these things have a consisent
> meaning in
> > > all iommu drivers. Once this is done vfio and iommufd should both
> > > always set IOMMU_CACHE and refuse to work without
> > > IOMMU_CAP_CACHE_COHERENCY. (unless someone knows of
> an !IOMMU_CACHE
> > > arch that does in fact work today with vfio, somehow, but I don't..)
> >
> > IIRC, the DMAR on Intel CPUs dedicated to IGD was where we'd often see
> > lack of snoop-control support, causing us to have mixed coherent and
> > non-coherent domains. I don't recall if you go back far enough in VT-d
> > history if the primary IOMMU might have lacked this support. So I
> > think there are systems we care about with IOMMUs that can't enforce
> > DMA coherency.
> >
> > As it is today, if the IOMMU reports IOMMU_CAP_CACHE_COHERENCY and
> all
> > mappings make use of IOMMU_CACHE, then all DMA is coherent. Are you
> > suggesting IOMMU_CAP_CACHE_COHERENCY should indicate that all
> mappings
> > are coherent regardless of mapping protection flags? What's the point
> > of IOMMU_CACHE at that point?
>
> IOMMU_CAP_CACHE_COHERENCY should return to what it was before Intel's
> change.
One nit (as I explained in previous v1 discussion). It is not that Intel
abuses this capability as it was firstly introduced for Intel's force
snoop requirement. It is just that when later its meaning was changed
to match what you described below the original use of Intel was not
caught and fixed properly. 😊
>
> It only means normal DMAs issued in a normal way are coherent with the
> CPU and do not require special cache flushing instructions. ie DMA
> issued by a kernel driver using the DMA API.
>
> It does not mean that non-coherence DMA does not exist, or that
> platform or device specific ways to trigger non-coherence are blocked.
>
> Stated another way, any platform that wires dev_is_dma_coherent() to
> true, like all x86 does, must support IOMMU_CACHE and report
> IOMMU_CAP_CACHE_COHERENCY for every iommu_domain the platform
> supports. The platform obviously declares it support this in order to
> support the in-kernel DMA API.
This is a good explanation of IOMMU_CACHE. From that intel-iommu
driver should always report this capability and do nothing with
IOMMU_CACHE since it's already guaranteed by the arch. Actually
this is exactly what AMD iommu driver does today.
Then we'll introduce a new cap/prot in particular for enforcing snoop
as you suggested below.
>
> Thus, a new cap indicating that 'all dma is coherent' or 'no-snoop
> blocking' should be created to cover Intel's special need. From what I
> know it is only implemented in the Intel driver and apparently only
> for some IOMMUs connected to IGD.
>
> > > Yes, it was missed in the notes for vfio compat that Intel no-snoop is
> > > not working currently, I fixed it.
> >
> > Right, I see it in the comments relative to extensions, but missed in
> > the commit log. Thanks,
>
> Oh good, I remembered it was someplace..
>
Thanks
Kevin
