Re: [RFC] /dev/ioasid uAPI proposal

[Alex Williamson <alex.williamson@redhat.com>]

On Thu, 3 Jun 2021 09:40:36 -0300
Jason Gunthorpe <jgg@nvidia.com> wrote:

> On Thu, Jun 03, 2021 at 03:22:27AM +0000, Tian, Kevin wrote:
> > > From: Alex Williamson <alex.williamson@redhat.com>
> > > Sent: Thursday, June 3, 2021 10:51 AM
> > > 
> > > On Wed, 2 Jun 2021 19:45:36 -0300
> > > Jason Gunthorpe <jgg@nvidia.com> wrote:
> > >   
> > > > On Wed, Jun 02, 2021 at 02:37:34PM -0600, Alex Williamson wrote:
> > > >  
> > > > > Right.  I don't follow where you're jumping to relaying DMA_PTE_SNP
> > > > > from the guest page table... what page table?  
> > > >
> > > > I see my confusion now, the phrasing in your earlier remark led me
> > > > think this was about allowing the no-snoop performance enhancement in
> > > > some restricted way.
> > > >
> > > > It is really about blocking no-snoop 100% of the time and then
> > > > disabling the dangerous wbinvd when the block is successful.
> > > >
> > > > Didn't closely read the kvm code :\
> > > >
> > > > If it was about allowing the optimization then I'd expect the guest to
> > > > enable no-snoopable regions via it's vIOMMU and realize them to the
> > > > hypervisor and plumb the whole thing through. Hence my remark about
> > > > the guest page tables..
> > > >
> > > > So really the test is just 'were we able to block it' ?  
> > > 
> > > Yup.  Do we really still consider that there's some performance benefit
> > > to be had by enabling a device to use no-snoop?  This seems largely a
> > > legacy thing.  
> > 
> > Yes, there is indeed performance benefit for device to use no-snoop,
> > e.g. 8K display and some imaging processing path, etc. The problem is
> > that the IOMMU for such devices is typically a different one from the
> > default IOMMU for most devices. This special IOMMU may not have
> > the ability of enforcing snoop on no-snoop PCI traffic then this fact
> > must be understood by KVM to do proper mtrr/pat/wbinvd virtualization 
> > for such devices to work correctly.  
> 
> Or stated another way:
> 
> We in Linux don't have a way to control if the VFIO IO page table will
> be snoop or no snoop from userspace so Intel has forced the platform's
> IOMMU path for the integrated GPU to be unable to enforce snoop, thus
> "solving" the problem.

That's giving vfio a lot of credit for influencing VT-d design.

> I don't think that is sustainable in the oveall ecosystem though.

Our current behavior is a reasonable default IMO, but I agree more
control will probably benefit us in the long run.

> 'qemu --allow-no-snoop' makes more sense to me

I'd be tempted to attach it to the -device vfio-pci option, it's
specific drivers for specific devices that are going to want this and
those devices may not be permanently attached to the VM.  But I see in
the other thread you're trying to optimize IOMMU page table sharing.

There's a usability question in either case though and I'm not sure how
to get around it other than QEMU or the kernel knowing a list of
devices (explicit IDs or vendor+class) to select per device defaults.

> > When discussing I/O page fault support in another thread, the consensus
> > is that an device handle will be registered (by user) or allocated (return
> > to user) in /dev/ioasid when binding the device to ioasid fd. From this 
> > angle we can register {ioasid_fd, device_handle} to KVM and then call 
> > something like ioasidfd_device_is_coherent() to get the property. 
> > Anyway the coherency is a per-device property which is not changed 
> > by how many I/O page tables are attached to it.  
> 
> It is not device specific, it is driver specific
> 
> As I said before, the question is if the IOASID itself can enforce
> snoop, or not. AND if the device will issue no-snoop or not.
> 
> Devices that are hard wired to never issue no-snoop are safe even with
> an IOASID that cannot enforce snoop. AFAIK really only GPUs use this
> feature. Eg I would be comfortable to say mlx5 never uses the no-snoop
> TLP flag.
> 
> Only the vfio_driver could know this.

Could you clarify "vfio_driver"?  The existing vfio-pci driver can't
know this, beyond perhaps probing if the Enable No-snoop bit is
hardwired to zero.  It's the driver running on top of vfio that
ultimately controls whether a capable device actually issues no-snoop
TLPs, but that can't be known to us.  A vendor variant of vfio-pci
might certainly know more about how its device is used by those
userspace/VM drivers.  Thanks,

Alex