On Fri, Oct 01, 2021 at 08:29:36AM +0200, Greg Kroah-Hartman wrote:

On Thu, Sep 30, 2021 at 12:15:16PM -0700, Andi Kleen wrote:

quoted

On 9/30/2021 10:23 AM, Greg Kroah-Hartman wrote:

quoted

On Thu, Sep 30, 2021 at 10:17:09AM -0700, Andi Kleen wrote:

quoted

The legacy drivers could be fixed, but nobody really wants to touch them
anymore and they're impossible to test.

Pointers to them?

For example if you look over old SCSI drivers in drivers/scsi/*.c there is a
substantial number that has a module init longer than just registering a
driver. As a single example look at drivers/scsi/BusLogic.c

Great, send patches to fix them up instead of adding new infrastructure
to the kernel.  It is better to remove code than add it.  You can rip
the ISA code out of that driver and then you will not have the issue
anymore.

Or again, just add that module to the deny list and never load it from
userspace.

quoted

There were also quite a few platform drivers like this.

Of course, platform drivers are horrible abusers of this.  Just like the
recent one submitted by Intel that would bind to any machine it was
loaded on and did not actually do any hardware detection assuming that
it owned the platform:
	https://lore.kernel.org/r/20210924213157.3584061-2-david.e.box@linux.intel.com (local)

So yes, some drivers are horrible, it is our job to catch that and fix
it.  If you don't want to load those drivers on your system, we have
userspace solutions for that (you can have allow/deny lists there.)

quoted

quoted

quoted

The drivers that probe something that is not enumerated in a standard way
have no choice, it cannot be implemented in a different way.

PCI devices are not enumerated in a standard way???

The pci devices are enumerated in a standard way, but typically the driver
also needs something else outside PCI that needs some other probing
mechanism.

Like what?  What PCI drivers need outside connections to control the
hardware?

quoted

quoted

quoted

So instead we're using a "firewall" the prevents these drivers from doing
bad things by not allowing ioremap access unless opted in, and also do some
filtering on the IO ports The device filter is still the primary mechanism,
the ioremap filtering is just belts and suspenders for those odd cases.

That's horrible, don't try to protect the kernel from itself.  Just fix
the drivers.

I thought we had already established this last time we discussed it.

That's completely impractical. We cannot harden thousands of drivers,
especially since it would be all wasted work since nobody will ever need
them in virtual guests. Even if we could harden them how would such a work
be maintained long term? Using a firewall and filtering mechanism is much
saner for everyone.

I agree, you can not "harden" anything here.  That is why I asked you to
use the existing process that explicitly moves the model to userspace
where a user can say "do I want this device to be controlled by the
kernel or not" which then allows you to pick and choose what drivers you
want to have in your system.

You need to trust devices, and not worry about trusting drivers as you
yourself admit :)

That isn't the way they see it.  In their approach you can never 
trust any devices at all.  Therefore devices can only be allowed to 
bind to a very small set of "hardened" drivers.  Never mind how they 
decide whether or not a driver is sufficiently "hardened".

The kernel's trust model is that once we bind to them, we trust almost
all device types almost explicitly.  If you wish to change that model,
that's great, but it is a much larger discussion than this tiny patchset
would require.

Forget about trust for the moment.  Let's say the goal is to prevent 
the kernel from creating any bindings other that those in some small 
"allowed" set.  To fully specify one of the allowed bindings, you 
would have to provide both a device ID and a driver name.  But in 
practice this isn't necessary, since a device with a given ID will 
bind to only one driver in almost all cases, and hence giving just 
the device ID is enough.

So to do what they want, all that's needed is to forbid any bindings 
except where the device ID is "allowed".  Or to put it another way, 
where the device's authorized flag (which can be initialized based on 
the device ID) is set.

(The opposite approach, in which the drivers are "allowed" rather 
than the device IDs, apparently has already been discussed and 
rejected.  I'm not convinced that was a good decision, but...)

Does this seem like a fair description of the situation?

Alan Stern
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization