Re: [PATCH] uio_pci_generic does not export memory resources

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

On Sun, Jun 10, 2012 at 11:38:25AM -0600, Alex Williamson wrote:
> On Sun, 2012-06-10 at 19:44 +0300, Michael S. Tsirkin wrote:
> > On Sun, Jun 10, 2012 at 10:09:26AM -0600, Alex Williamson wrote:
> > > On Sun, 2012-06-10 at 17:18 +0300, Michael S. Tsirkin wrote:
> > > > On Fri, Jun 08, 2012 at 11:11:16AM -0600, Alex Williamson wrote:
> > > > > On Fri, 2012-06-08 at 18:44 +0200, Hans J. Koch wrote:
> > > > > > On Fri, Jun 08, 2012 at 06:16:18PM +0200, Andreas Hartmann wrote:
> > > > > > > Hi Dominic,
> > > > > > > 
> > > > > > > Dominic Eschweiler wrote:
> > > > > > > > Am Freitag, den 08.06.2012, 08:16 -0600 schrieb Alex Williamson:
> > > > > > > >> Yes, thanks Jan.  This is exactly what VFIO does.  VFIO provides
> > > > > > > >> secure config space access, resource access, DMA mapping services, and
> > > > > > > >> full interrupt support to userspace.  
> > > > > > 
> > > > > > VFIO is not a "better UIO". It *requires* an IOMMU. Dominic didn't say on
> > > > > > what CPU he's working, so it's not clear if he can use VFIO at all.
> > > > > > 
> > > > > > UIO is intended for general use with devices that have mappable registers
> > > > > > and don't fit into any other subsystem. No more, no less.
> > > > > 
> > > > > VFIO is a secure UIO.
> > > > 
> > > > A secure UIO *for VFs*. I think that's why it's called VFIO :).
> > > > Other stuff sometimes also works but no real guarantees, though
> > > > VFIO tries to make sure you don't burn yourself too badly
> > > > if it breaks.
> > > 
> > > We do a little better than that.  Multifunction devices that don't
> > > explicitly report ACS support are grouped together, so we have security
> > > for multifunction devices as well.
> > 
> > How can you get security with insecure hardware?
> > 
> > So you prevent the device from writing to host memory? Cool.
> > Now guest puts a virus on an on-card flash, the
> > moment device is assigned to another VM it will own that,
> > or host if it's enabled in host.
> > 
> > I can make up more silliness.  Buggy userspace can brick the device,
> > e.g. by damaging the internal eeprom memory, and these things were known
> > to happen even by accident.
> > 
> > Simply put if you want secure userspace drivers you must be able to
> > trust your hardware for security and the only hardware that promises you
> > security is a VF in an SRIOV device.
> 
> Next I suppose you're going to say assigning a NIC to a guest is
> insecure because it could host a malicious OS that infects other systems
> on the network.

*Of course* it is less secure than a firewalled guest
with a virtual NIC. You argue this is not true?

But at least there are ways to contain a NIC on a network.
So it depends on the setup.

Not so for an assigned PF. It depends on the internals of the PF
which you have no idea about.

> So to clarify, by secure, I mean that users of VFIO
> devices don't have access to the host.

Yes. And since you can't guarantee it for PFs, it's insecure.

> The host still needs to be
> suspicious of any data the user might have tainted

That's not the only point. Host data might also leak to guest
when device is assigned.

> after a device is returned.

For a VF you have a way to validate what the VF does.
For a PF there is no way to be suspicious of the device state.

> > > Either single of multifunction PFs
> > > can have an option ROM, but since there's no defined mechanism to
> > > program the ROM, we can't protect it.  Secure boot actually helps us
> > > here since the ROM loaded by the host BIOS or drivers would need to
> > > verify the ROM before using it.  Note that secure boot will likely close
> > > off the pci-sysfs path uio_pci and KVM device assignment use to get
> > > resources since it allows unprotected access to the system.  VFIO
> > > provides an interface where we control secure access, so should be
> > > compatible with secure boot.  Thanks,
> > > 
> > > Alex
> > 
> > IMHO all this means VFIO *works* not just for VFs.
> > Not that it's secure.
> 
> By your argument above, not even VFs are "secure".

VFs can be secure if PF hardware and driver are secure.

There's no sure way to secure a PF.

>  A user could just as
> easily taint a disk attached to an HBA VF...

But if I don't run stuff from this disk I am safe.

What is the way to guarantee security with an assigned PF?

-- 
MST