Re: [PATCH v5 3/3] efi: Add tee-based EFI variable driver

On 07.06.23 10:25, Ilias Apalodimas wrote:

Hi Sumit,

On Wed, 7 Jun 2023 at 10:25, Sumit Garg [off-list ref] wrote:

quoted

Hi Ilias,

On Wed, 7 Jun 2023 at 12:05, Ilias Apalodimas
[off-list ref] wrote:

quoted

Hi Jan,

[...]

quoted

quoted

quoted

quoted

quoted

...

quoted

quoted

I think we have a probe ordering issue with this driver:
efivarfs_fill_super() may be called before the TEE bus was probed, thus
with the default efivar ops still registered. And that means
efivar_supports_writes() will return false, and the fs declares itself
as readonly. I've seen systemd mounting it r/o initialling, and you need
to remount the fs to enable writability.

Is there anything that could be done to re-order things reliably, probe
the tee bus earlier etc.?

This driver has a dependency on user-space daemon: tee-supplicant to
be running for RPMB access. So once you start that daemon the
corresponding device will be enumerated on the TEE bus and this driver
probe will be invoked. So I would suggest you to load this daemon very
early in the boot process or better to make it a part of initramfs.

That is not the point, really.

If this dependency exists, the code should be aware of that, and made
to work correctly in spite of it. Requiring a module to be part of
initramfs is not a reasonable fix.

In fact, I've tested a non-modularized build as well, just to exclude
that issue. The daemon dependency is more likely the problem here.

quoted

IIUC, this also means that the efivar ops are updated while there is
already a client. This seems less than ideal as well

As Sumit pointed out, the 'device' won't be available from OP-TEE
until the supplicant is up and running and as a result, the module
_probe() function won't run.  Unfortunately, this isn't something we
can avoid since the supplicant is responsible for the RPMB writes.
The only thing I can think of is moving parts of the supplicant to the
kernel and wiring up the RPC calls for reading/writing data to the
eMMC subsystem.  There was another discussion here [0] requesting the
same thing for different reasons. But unless I am missing something
this won't solve the problem completely either.  You still have a
timing dependency of "when did the RT callbacks change" -- "when was
my efivarfs mounted".

With the RPMB writes wired through the kernel [1], the only dependency
left is when do you load the tee-stmm-efi driver to have real EFI
runtime variables support. IMO, tee-stmm-efi driver should be built-in
to support systems without initramfs. The distro installers may choose
to bundle it in initramfs. Do you still see a timing dependency with
this approach?

No I don't, this will work reliably without the need to remount the efivarfs.
As you point out you will still have this dependency if you end up
building them as modules and you manage to mount the efivarfs before
those get inserted.  Does anyone see a reasonable workaround?
Deceiving the kernel and making the bootloader set the RT property bit
to force the filesystem being mounted as rw is a nasty hack that we
should avoid.  Maybe adding a kernel command line parameter that says
"Ignore the RTPROP I know what I am doing"?  I don't particularly love
this either, but it's not unreasonable.

In the context of https://github.com/OP-TEE/optee_os/issues/6094,
basically this issue mapped on reboot/shutdown, I would really love to
see the unhandy tee-supplicant daemon to be overcome.

Jan

-- 
Siemens AG, Technology
Competence Center Embedded Linux


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel