Re: [PATCH v2 4/6] virtio: Initialize authorized attribute for confidential guest

[Dan Williams <dan.j.williams@intel.com>]

On Fri, Oct 1, 2021 at 12:03 AM Greg Kroah-Hartman
<gregkh@linuxfoundation.org> wrote:
>
> On Thu, Sep 30, 2021 at 12:04:05PM -0700, Kuppuswamy, Sathyanarayanan wrote:
> >
> >
> > On 9/30/21 8:23 AM, Greg Kroah-Hartman wrote:
> > > On Thu, Sep 30, 2021 at 08:18:18AM -0700, Kuppuswamy, Sathyanarayanan wrote:
> > > >
> > > >
> > > > On 9/30/21 6:36 AM, Dan Williams wrote:
> > > > > > And in particular, not all virtio drivers are hardened -
> > > > > > I think at this point blk and scsi drivers have been hardened - so
> > > > > > treating them all the same looks wrong.
> > > > > My understanding was that they have been audited, Sathya?
> > > >
> > > > Yes, AFAIK, it has been audited. Andi also submitted some patches
> > > > related to it. Andi, can you confirm.
> > >
> > > What is the official definition of "audited"?
> >
> >
> > In our case (Confidential Computing platform), the host is an un-trusted
> > entity. So any interaction with host from the drivers will have to be
> > protected against the possible attack from the host. For example, if we
> > are accessing a memory based on index value received from host, we have
> > to make sure it does not lead to out of bound access or when sharing the
> > memory with the host, we need to make sure only the required region is
> > shared with the host and the memory is un-shared after use properly.
>
> You have not defined the term "audited" here at all in any way that can
> be reviewed or verified by anyone from what I can tell.

Agree.

>
> You have only described a new model that you wish the kernel to run in,
> one in which it does not trust the hardware at all.  That is explicitly
> NOT what the kernel has been designed for so far, and if you wish to
> change that, lots of things need to be done outside of simply running
> some fuzzers on a few random drivers.
>
> For one example, how do you ensure that the memory you are reading from
> hasn't been modified by the host between writing to it the last time you
> did?  Do you have a list of specific drivers and kernel options that you
> feel you now "trust"?  If so, how long does that trust last for?  Until
> someonen else modifies that code?  What about modifications to functions
> that your "audited" code touches?  Who is doing this auditing?  How do
> you know the auditing has been done correctly?  Who has reviewed and
> audited the tools that are doing the auditing?  Where is the
> specification that has been agreed on how the auditing must be done?
> And so on...
>
> I feel like there are a lot of different things all being mixed up here
> into one "oh we want this to happen!" type of thread.  Please let's just
> stick to the one request that I had here, which was to move the way that
> busses are allowed to authorize the devices they wish to control into a
> generic way instead of being bus-specific logic.

I want to continue to shave this proposal down, but that last sentence
reads as self-contradictory.

Bear with me, and perhaps it's a lack of imagination on my part, but I
don't see how to get to a globally generic "authorized" sysfs ABI
given that USB and Thunderbolt want to do bus specific actions on
authorization toggle events. Certainly a default generic authorized
attribute can be defined for all the other buses that don't have
legacy here, but Thunderbolt will still require support for '2' as an
authorized value, and USB will still want to base probe decisions on
the authorization state of both the usb_device and the usb_interface.

> Any requests outside of that type of functionality are just that,
> outside the scope of this patchset and should get their own patch series
> and discussion.

A way to shave this further would be to default authorize just enough
devices to do the rest of the authorization from initramfs. That would
seem to cut out 'virtio-net' and 'virtio-storage' from default
authorized list.

The generic authorized attribute would only need to appear when
'dev_default_authorization' is set to false.