Virtualization - Aug 2021 - [PATCH v4 11/15] pci: Add pci_iomap_shared{,_range}

> I see. Hmm. It's a bit of a random thing to do it at the map time
> though. E.g. DMA is all handled transparently behind the DMA API.
> Hardening is much more than just replacing map with map_shared
> and I suspect what you will end up with is basically
> vendors replacing map with map shared to make things work
> for their users and washing their hands.
That concept exists too. There is a separate allow list for the drivers. 
So just adding shared to a driver is not enough, until it's also added 
to the allowlist

Users can of course chose to disable the allowlist, but they need to 
understand the security implications.

>
> I would say an explicit flag in the driver that says "hardened"
> and refusing to init a non hardened one would be better.

We have that too (that's the device filtering)

But the problem is that device filtering just stops the probe functions, 
not the initcalls, and lot of legacy drivers do MMIO interactions before 
going into probe. In some cases it's unavoidable because of the device 
doesn't have a separate enumeration mechanism it needs some kind of 
probing to even check for its existence And since we don't want to 
change all of them it's far safer to make the ioremap opt-in.


-Andi

[+cc Rajat; I still don't know what "shared memory with a hypervisor
in a confidential guest" means, but now we're talking about hardened
drivers and allow lists, which Rajat is interested in]

On Tue, Aug 24, 2021 at 10:20:44AM -0700, Andi Kleen
wrote:
> 
> > I see. Hmm. It's a bit of a random thing to do it at the map time
> > though. E.g. DMA is all handled transparently behind the DMA API.
> > Hardening is much more than just replacing map with map_shared
> > and I suspect what you will end up with is basically
> > vendors replacing map with map shared to make things work
> > for their users and washing their hands.
> 
> That concept exists too. There is a separate allow list for the drivers. So
> just adding shared to a driver is not enough, until it's also added to
the
> allowlist
> 
> Users can of course chose to disable the allowlist, but they need to
> understand the security implications.
> 
> > I would say an explicit flag in the driver that says
"hardened"
> > and refusing to init a non hardened one would be better.
> 
> We have that too (that's the device filtering)
> 
> But the problem is that device filtering just stops the probe functions,
not
> the initcalls, and lot of legacy drivers do MMIO interactions before going
> into probe. In some cases it's unavoidable because of the device
doesn't
> have a separate enumeration mechanism it needs some kind of probing to even
> check for its existence And since we don't want to change all of them
it's
> far safer to make the ioremap opt-in.
> 
> 
> -Andi
>

On Tue, Aug 24, 2021 at 10:20:44AM -0700, Andi Kleen
wrote:
> 
> > I see. Hmm. It's a bit of a random thing to do it at the map time
> > though. E.g. DMA is all handled transparently behind the DMA API.
> > Hardening is much more than just replacing map with map_shared
> > and I suspect what you will end up with is basically
> > vendors replacing map with map shared to make things work
> > for their users and washing their hands.
> 
> That concept exists too. There is a separate allow list for the drivers. So
> just adding shared to a driver is not enough, until it's also added to
the
> allowlist
> 
> Users can of course chose to disable the allowlist, but they need to
> understand the security implications.
Right. So given that, why do we need to tweak a random API like the map?
If you just make all maps be shared then the user is in control.
Seems sensible to me.
> 
> > 
> > I would say an explicit flag in the driver that says
"hardened"
> > and refusing to init a non hardened one would be better.
> 
> 
> We have that too (that's the device filtering)
> 
> But the problem is that device filtering just stops the probe functions,
not
> the initcalls, and lot of legacy drivers do MMIO interactions before going
> into probe. In some cases it's unavoidable because of the device
doesn't
> have a separate enumeration mechanism it needs some kind of probing to even
> check for its existence And since we don't want to change all of them
it's
> far safer to make the ioremap opt-in.
> 
> 
> -Andi
Let's be frank, even without encryption disabling most drivers -
especially weird ones that poke at hardware before probe -
is far safer than keeping them, but one loses a bunch of features.
IOW all this hardening is nice but which security/feature tradeoff
to take it a policy decision, not something kernel should do
imho.

-- 
MST